From 9143fcd7fd38243dd40f927dafaeb75f6ef8ef49 Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Tue, 21 Jan 2020 14:47:17 -0800 Subject: Add UDP matchers. --- pkg/sentry/socket/netfilter/netfilter.go | 164 ++++++++++++++++++++++++++----- 1 file changed, 141 insertions(+), 23 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index e1f2bacce..45296b339 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -131,6 +131,7 @@ func FillDefaultIPTables(stack *stack.Stack) { stack.SetIPTables(ipt) } +// TODO: Return proto. // convertNetstackToBinary converts the iptables as stored in netstack to the // format expected by the iptables tool. Linux stores each table as a binary // blob that can only be traversed by parsing a bit, reading some offsets, @@ -318,10 +319,12 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { } var entry linux.IPTEntry buf := optVal[:linux.SizeOfIPTEntry] - optVal = optVal[linux.SizeOfIPTEntry:] binary.Unmarshal(buf, usermem.ByteOrder, &entry) - if entry.TargetOffset != linux.SizeOfIPTEntry { - // TODO(gvisor.dev/issue/170): Support matchers. + initialOptValLen := len(optVal) + optVal = optVal[linux.SizeOfIPTEntry:] + + if entry.TargetOffset < linux.SizeOfIPTEntry { + log.Warningf("netfilter: entry has too-small target offset %d", entry.TargetOffset) return syserr.ErrInvalidArgument } @@ -332,19 +335,41 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { return err } + // TODO: Matchers (and maybe targets) can specify that they only work for certiain protocols, hooks, tables. + // Get matchers. + matchersSize := entry.TargetOffset - linux.SizeOfIPTEntry + if len(optVal) < int(matchersSize) { + log.Warningf("netfilter: entry doesn't have enough room for its matchers (only %d bytes remain)", len(optVal)) + } + matchers, err := parseMatchers(filter, optVal[:matchersSize]) + if err != nil { + log.Warningf("netfilter: failed to parse matchers: %v", err) + return err + } + optVal = optVal[matchersSize:] + // Get the target of the rule. - target, consumed, err := parseTarget(optVal) + targetSize := entry.NextOffset - entry.TargetOffset + if len(optVal) < int(targetSize) { + log.Warningf("netfilter: entry doesn't have enough room for its target (only %d bytes remain)", len(optVal)) + } + target, err := parseTarget(optVal[:targetSize]) if err != nil { return err } - optVal = optVal[consumed:] + optVal = optVal[targetSize:] table.Rules = append(table.Rules, iptables.Rule{ - Filter: filter, - Target: target, + Filter: filter, + Target: target, + Matchers: matchers, }) offsets = append(offsets, offset) - offset += linux.SizeOfIPTEntry + consumed + offset += uint32(entry.NextOffset) + + if initialOptValLen-len(optVal) != int(entry.NextOffset) { + log.Warningf("netfilter: entry NextOffset is %d, but entry took up %d bytes", entry.NextOffset, initialOptValLen-len(optVal)) + } } // Go through the list of supported hooks for this table and, for each @@ -401,12 +426,105 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { return nil } -// parseTarget parses a target from the start of optVal and returns the target -// along with the number of bytes it occupies in optVal. -func parseTarget(optVal []byte) (iptables.Target, uint32, *syserr.Error) { +// parseMatchers parses 0 or more matchers from optVal. optVal should contain +// only the matchers. +func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Matcher, *syserr.Error) { + var matchers []iptables.Matcher + for len(optVal) > 0 { + log.Infof("parseMatchers: optVal has len %d", len(optVal)) + // Get the XTEntryMatch. + if len(optVal) < linux.SizeOfXTEntryMatch { + log.Warningf("netfilter: optVal has insufficient size for entry match: %d", len(optVal)) + return nil, syserr.ErrInvalidArgument + } + var match linux.XTEntryMatch + buf := optVal[:linux.SizeOfXTEntryMatch] + binary.Unmarshal(buf, usermem.ByteOrder, &match) + log.Infof("parseMatchers: parsed entry match %q: %+v", match.Name.String(), match) + + // Check some invariants. + if match.MatchSize < linux.SizeOfXTEntryMatch { + log.Warningf("netfilter: match size is too small, must be at least %d", linux.SizeOfXTEntryMatch) + return nil, syserr.ErrInvalidArgument + } + if len(optVal) < int(match.MatchSize) { + log.Warningf("netfilter: optVal has insufficient size for match: %d", len(optVal)) + return nil, syserr.ErrInvalidArgument + } + + buf = optVal[linux.SizeOfXTEntryMatch:match.MatchSize] + var matcher iptables.Matcher + var err error + switch match.Name.String() { + case "tcp": + if len(buf) < linux.SizeOfXTTCP { + log.Warningf("netfilter: optVal has insufficient size for TCP match: %d", len(optVal)) + return nil, syserr.ErrInvalidArgument + } + var matchData linux.XTTCP + // For alignment reasons, the match's total size may exceed what's + // strictly necessary to hold matchData. + binary.Unmarshal(buf[:linux.SizeOfXTUDP], usermem.ByteOrder, &matchData) + log.Infof("parseMatchers: parsed XTTCP: %+v", matchData) + matcher, err = iptables.NewTCPMatcher(filter, iptables.TCPMatcherData{ + SourcePortStart: matchData.SourcePortStart, + SourcePortEnd: matchData.SourcePortEnd, + DestinationPortStart: matchData.DestinationPortStart, + DestinationPortEnd: matchData.DestinationPortEnd, + Option: matchData.Option, + FlagMask: matchData.FlagMask, + FlagCompare: matchData.FlagCompare, + InverseFlags: matchData.InverseFlags, + }) + if err != nil { + log.Warningf("netfilter: failed to create TCP matcher: %v", err) + return nil, syserr.ErrInvalidArgument + } + + case "udp": + if len(buf) < linux.SizeOfXTUDP { + log.Warningf("netfilter: optVal has insufficient size for UDP match: %d", len(optVal)) + return nil, syserr.ErrInvalidArgument + } + var matchData linux.XTUDP + // For alignment reasons, the match's total size may exceed what's + // strictly necessary to hold matchData. + binary.Unmarshal(buf[:linux.SizeOfXTUDP], usermem.ByteOrder, &matchData) + log.Infof("parseMatchers: parsed XTUDP: %+v", matchData) + matcher, err = iptables.NewUDPMatcher(filter, iptables.UDPMatcherData{ + SourcePortStart: matchData.SourcePortStart, + SourcePortEnd: matchData.SourcePortEnd, + DestinationPortStart: matchData.DestinationPortStart, + DestinationPortEnd: matchData.DestinationPortEnd, + InverseFlags: matchData.InverseFlags, + }) + if err != nil { + log.Warningf("netfilter: failed to create UDP matcher: %v", err) + return nil, syserr.ErrInvalidArgument + } + + default: + log.Warningf("netfilter: unsupported matcher with name %q", match.Name.String()) + return nil, syserr.ErrInvalidArgument + } + + matchers = append(matchers, matcher) + + // TODO: Support revision. + // TODO: Support proto -- matchers usually specify which proto(s) they work with. + optVal = optVal[match.MatchSize:] + } + + // TODO: Check that optVal is exhausted. + return matchers, nil +} + +// parseTarget parses a target from optVal. optVal should contain only the +// target. +func parseTarget(optVal []byte) (iptables.Target, *syserr.Error) { if len(optVal) < linux.SizeOfXTEntryTarget { log.Warningf("netfilter: optVal has insufficient size for entry target %d", len(optVal)) - return nil, 0, syserr.ErrInvalidArgument + return nil, syserr.ErrInvalidArgument } var target linux.XTEntryTarget buf := optVal[:linux.SizeOfXTEntryTarget] @@ -414,9 +532,9 @@ func parseTarget(optVal []byte) (iptables.Target, uint32, *syserr.Error) { switch target.Name.String() { case "": // Standard target. - if len(optVal) < linux.SizeOfXTStandardTarget { - log.Warningf("netfilter.SetEntries: optVal has insufficient size for standard target %d", len(optVal)) - return nil, 0, syserr.ErrInvalidArgument + if len(optVal) != linux.SizeOfXTStandardTarget { + log.Warningf("netfilter.SetEntries: optVal has wrong size for standard target %d", len(optVal)) + return nil, syserr.ErrInvalidArgument } var standardTarget linux.XTStandardTarget buf = optVal[:linux.SizeOfXTStandardTarget] @@ -424,22 +542,22 @@ func parseTarget(optVal []byte) (iptables.Target, uint32, *syserr.Error) { verdict, err := translateToStandardVerdict(standardTarget.Verdict) if err != nil { - return nil, 0, err + return nil, err } switch verdict { case iptables.Accept: - return iptables.UnconditionalAcceptTarget{}, linux.SizeOfXTStandardTarget, nil + return iptables.UnconditionalAcceptTarget{}, nil case iptables.Drop: - return iptables.UnconditionalDropTarget{}, linux.SizeOfXTStandardTarget, nil + return iptables.UnconditionalDropTarget{}, nil default: panic(fmt.Sprintf("Unknown verdict: %v", verdict)) } case errorTargetName: // Error target. - if len(optVal) < linux.SizeOfXTErrorTarget { + if len(optVal) != linux.SizeOfXTErrorTarget { log.Infof("netfilter.SetEntries: optVal has insufficient size for error target %d", len(optVal)) - return nil, 0, syserr.ErrInvalidArgument + return nil, syserr.ErrInvalidArgument } var errorTarget linux.XTErrorTarget buf = optVal[:linux.SizeOfXTErrorTarget] @@ -454,16 +572,16 @@ func parseTarget(optVal []byte) (iptables.Target, uint32, *syserr.Error) { // rules have an error with the name of the chain. switch errorTarget.Name.String() { case errorTargetName: - return iptables.ErrorTarget{}, linux.SizeOfXTErrorTarget, nil + return iptables.ErrorTarget{}, nil default: log.Infof("Unknown error target %q doesn't exist or isn't supported yet.", errorTarget.Name.String()) - return nil, 0, syserr.ErrInvalidArgument + return nil, syserr.ErrInvalidArgument } } // Unknown target. log.Infof("Unknown target %q doesn't exist or isn't supported yet.", target.Name.String()) - return nil, 0, syserr.ErrInvalidArgument + return nil, syserr.ErrInvalidArgument } func filterFromIPTIP(iptip linux.IPTIP) (iptables.IPHeaderFilter, *syserr.Error) { -- cgit v1.2.3 From 2661101ad470548cb15dce0afc694296668d780a Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Tue, 21 Jan 2020 14:51:28 -0800 Subject: Removed TCP work (saved in ipt-tcp-match). --- pkg/abi/linux/netfilter.go | 52 ------------- pkg/sentry/socket/netfilter/netfilter.go | 26 ------- pkg/tcpip/iptables/tcp_matcher.go | 122 ------------------------------- 3 files changed, 200 deletions(-) delete mode 100644 pkg/tcpip/iptables/tcp_matcher.go (limited to 'pkg/sentry/socket') diff --git a/pkg/abi/linux/netfilter.go b/pkg/abi/linux/netfilter.go index fb4588272..f0e544f9c 100644 --- a/pkg/abi/linux/netfilter.go +++ b/pkg/abi/linux/netfilter.go @@ -341,58 +341,6 @@ func goString(cstring []byte) string { return string(cstring) } -// XTTCP holds data for matching TCP packets. It corresponds to struct xt_tcp -// in include/uapi/linux/netfilter/xt_tcpudp.h. -type XTTCP struct { - // SourcePortStart specifies the inclusive start of the range of source - // ports to which the matcher applies. - SourcePortStart uint16 - - // SourcePortEnd specifies the inclusive end of the range of source ports - // to which the matcher applies. - SourcePortEnd uint16 - - // DestinationPortStart specifies the start of the destination port - // range to which the matcher applies. - DestinationPortStart uint16 - - // DestinationPortEnd specifies the start of the destination port - // range to which the matcher applies. - DestinationPortEnd uint16 - - // Option specifies that a particular TCP option must be set. - Option uint8 - - // FlagMask masks the FlagCompare byte when comparing to the TCP flag - // fields. - FlagMask uint8 - - // FlagCompare is binary and-ed with the TCP flag fields. - FlagCompare uint8 - - // InverseFlags flips the meaning of certain fields. See the - // TX_TCP_INV_* flags. - InverseFlags uint8 -} - -// SizeOfXTTCP is the size of an XTTCP. -const SizeOfXTTCP = 12 - -// Flags in XTTCP.InverseFlags. Corresponding constants are in -// include/uapi/linux/netfilter/xt_tcpudp.h. -const ( - // Invert the meaning of SourcePortStart/End. - XT_TCP_INV_SRCPT = 0x01 - // Invert the meaning of DestinationPortStart/End. - XT_TCP_INV_DSTPT = 0x02 - // Invert the meaning of FlagCompare. - XT_TCP_INV_FLAGS = 0x04 - // Invert the meaning of Option. - XT_TCP_INV_OPTION = 0x08 - // Enable all flags. - XT_TCP_INV_MASK = 0x0F -) - // XTUDP holds data for matching UDP packets. It corresponds to struct xt_udp // in include/uapi/linux/netfilter/xt_tcpudp.h. type XTUDP struct { diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index 45296b339..f8ed1acbc 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -131,7 +131,6 @@ func FillDefaultIPTables(stack *stack.Stack) { stack.SetIPTables(ipt) } -// TODO: Return proto. // convertNetstackToBinary converts the iptables as stored in netstack to the // format expected by the iptables tool. Linux stores each table as a binary // blob that can only be traversed by parsing a bit, reading some offsets, @@ -456,31 +455,6 @@ func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Ma var matcher iptables.Matcher var err error switch match.Name.String() { - case "tcp": - if len(buf) < linux.SizeOfXTTCP { - log.Warningf("netfilter: optVal has insufficient size for TCP match: %d", len(optVal)) - return nil, syserr.ErrInvalidArgument - } - var matchData linux.XTTCP - // For alignment reasons, the match's total size may exceed what's - // strictly necessary to hold matchData. - binary.Unmarshal(buf[:linux.SizeOfXTUDP], usermem.ByteOrder, &matchData) - log.Infof("parseMatchers: parsed XTTCP: %+v", matchData) - matcher, err = iptables.NewTCPMatcher(filter, iptables.TCPMatcherData{ - SourcePortStart: matchData.SourcePortStart, - SourcePortEnd: matchData.SourcePortEnd, - DestinationPortStart: matchData.DestinationPortStart, - DestinationPortEnd: matchData.DestinationPortEnd, - Option: matchData.Option, - FlagMask: matchData.FlagMask, - FlagCompare: matchData.FlagCompare, - InverseFlags: matchData.InverseFlags, - }) - if err != nil { - log.Warningf("netfilter: failed to create TCP matcher: %v", err) - return nil, syserr.ErrInvalidArgument - } - case "udp": if len(buf) < linux.SizeOfXTUDP { log.Warningf("netfilter: optVal has insufficient size for UDP match: %d", len(optVal)) diff --git a/pkg/tcpip/iptables/tcp_matcher.go b/pkg/tcpip/iptables/tcp_matcher.go deleted file mode 100644 index 6acbd6eb9..000000000 --- a/pkg/tcpip/iptables/tcp_matcher.go +++ /dev/null @@ -1,122 +0,0 @@ -// Copyright 2020 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package iptables - -import ( - "fmt" - - "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/tcpip" - "gvisor.dev/gvisor/pkg/tcpip/header" -) - -type TCPMatcher struct { - data TCPMatcherData - - // tablename string - // unsigned int matchsize; - // unsigned int usersize; - // #ifdef CONFIG_COMPAT - // unsigned int compatsize; - // #endif - // unsigned int hooks; - // unsigned short proto; - // unsigned short family; -} - -// TODO: Delete? -// MatchCheckEntryParams - -type TCPMatcherData struct { - // Filter IPHeaderFilter - - SourcePortStart uint16 - SourcePortEnd uint16 - DestinationPortStart uint16 - DestinationPortEnd uint16 - Option uint8 - FlagMask uint8 - FlagCompare uint8 - InverseFlags uint8 -} - -func NewTCPMatcher(filter IPHeaderFilter, data TCPMatcherData) (Matcher, error) { - // TODO: We currently only support source port and destination port. - log.Infof("Adding rule with TCPMatcherData: %+v", data) - - if data.Option != 0 || - data.FlagMask != 0 || - data.FlagCompare != 0 || - data.InverseFlags != 0 { - return nil, fmt.Errorf("unsupported TCP matcher flags set") - } - - if filter.Protocol != header.TCPProtocolNumber { - log.Warningf("TCP matching is only valid for protocol %d.", header.TCPProtocolNumber) - } - - return &TCPMatcher{data: data}, nil -} - -// TODO: Check xt_tcpudp.c. Need to check for same things (e.g. fragments). -func (tm *TCPMatcher) Match(hook Hook, pkt tcpip.PacketBuffer, interfaceName string) (bool, bool) { - netHeader := header.IPv4(pkt.NetworkHeader) - - // TODO: Do we check proto here or elsewhere? I think elsewhere (check - // codesearch). - if netHeader.TransportProtocol() != header.TCPProtocolNumber { - return false, false - } - - // We dont't match fragments. - if frag := netHeader.FragmentOffset(); frag != 0 { - if frag == 1 { - log.Warningf("Dropping TCP packet: malicious packet with fragment with fragment offest of 1.") - return false, true - } - return false, false - } - - // Now we need the transport header. However, this may not have been set - // yet. - // TODO - var tcpHeader header.TCP - if pkt.TransportHeader != nil { - tcpHeader = header.TCP(pkt.TransportHeader) - } else { - // The TCP header hasn't been parsed yet. We have to do it here. - if len(pkt.Data.First()) < header.TCPMinimumSize { - // There's no valid TCP header here, so we hotdrop the - // packet. - // TODO: Stats. - log.Warningf("Dropping TCP packet: size to small.") - return false, true - } - tcpHeader = header.TCP(pkt.Data.First()) - } - - // Check whether the source and destination ports are within the - // matching range. - sourcePort := tcpHeader.SourcePort() - destinationPort := tcpHeader.DestinationPort() - if sourcePort < tm.data.SourcePortStart || tm.data.SourcePortEnd < sourcePort { - return false, false - } - if destinationPort < tm.data.DestinationPortStart || tm.data.DestinationPortEnd < destinationPort { - return false, false - } - - return true, false -} -- cgit v1.2.3 From 538053538dfb378aa8bc512d484ea305177e617b Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Tue, 21 Jan 2020 16:51:17 -0800 Subject: Adding serialization. --- pkg/sentry/socket/netfilter/netfilter.go | 29 ++++++++++++++++++++++++++++- pkg/tcpip/iptables/udp_matcher.go | 14 +++++++------- 2 files changed, 35 insertions(+), 8 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index f8ed1acbc..3caabca9a 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -196,7 +196,9 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern } func marshalMatcher(matcher iptables.Matcher) []byte { - switch matcher.(type) { + switch m := matcher.(type) { + case *iptables.UDPMatcher: + return marshalUDPMatcher(m) default: // TODO(gvisor.dev/issue/170): We don't support any matchers // yet, so any call to marshalMatcher will panic. @@ -204,6 +206,31 @@ func marshalMatcher(matcher iptables.Matcher) []byte { } } +func marshalUDPMatcher(matcher *iptables.UDPMatcher) []byte { + type udpMatch struct { + linux.XTEntryMatch + linux.XTUDP + } + linuxMatcher := udpMatch{ + XTEntryMatch: linux.XTEntryMatch{ + MatchSize: linux.SizeOfXTEntryMatch + linux.SizeOfXTUDP, + // Name: "udp", + }, + XTUDP: linux.XTUDP{ + SourcePortStart: matcher.Data.SourcePortStart, + SourcePortEnd: matcher.Data.SourcePortEnd, + DestinationPortStart: matcher.Data.DestinationPortStart, + DestinationPortEnd: matcher.Data.DestinationPortEnd, + InverseFlags: matcher.Data.InverseFlags, + }, + } + copy(linuxMatcher.Name[:], "udp") + + var buf [linux.SizeOfXTEntryMatch + linux.SizeOfXTUDP]byte + binary.Marshal(buf[:], usermem.ByteOrder, linuxMatcher) + return buf[:] +} + func marshalTarget(target iptables.Target) []byte { switch target.(type) { case iptables.UnconditionalAcceptTarget: diff --git a/pkg/tcpip/iptables/udp_matcher.go b/pkg/tcpip/iptables/udp_matcher.go index ce4368a3d..fca457199 100644 --- a/pkg/tcpip/iptables/udp_matcher.go +++ b/pkg/tcpip/iptables/udp_matcher.go @@ -24,7 +24,7 @@ import ( ) type UDPMatcher struct { - data UDPMatcherData + Data UDPMatcherData // tablename string // unsigned int matchsize; @@ -62,11 +62,11 @@ func NewUDPMatcher(filter IPHeaderFilter, data UDPMatcherData) (Matcher, error) log.Warningf("UDP matching is only valid for protocol %d.", header.UDPProtocolNumber) } - return &UDPMatcher{data: data}, nil + return &UDPMatcher{Data: data}, nil } // TODO: Check xt_tcpudp.c. Need to check for same things (e.g. fragments). -func (tm *UDPMatcher) Match(hook Hook, pkt tcpip.PacketBuffer, interfaceName string) (bool, bool) { +func (um *UDPMatcher) Match(hook Hook, pkt tcpip.PacketBuffer, interfaceName string) (bool, bool) { log.Infof("UDPMatcher called from: %s", string(debug.Stack())) netHeader := header.IPv4(pkt.NetworkHeader) @@ -114,12 +114,12 @@ func (tm *UDPMatcher) Match(hook Hook, pkt tcpip.PacketBuffer, interfaceName str destinationPort := udpHeader.DestinationPort() log.Infof("UDPMatcher: sport and dport are %d and %d. sports and dport start and end are (%d, %d) and (%d, %d)", udpHeader.SourcePort(), udpHeader.DestinationPort(), - tm.data.SourcePortStart, tm.data.SourcePortEnd, - tm.data.DestinationPortStart, tm.data.DestinationPortEnd) - if sourcePort < tm.data.SourcePortStart || tm.data.SourcePortEnd < sourcePort { + um.Data.SourcePortStart, um.Data.SourcePortEnd, + um.Data.DestinationPortStart, um.Data.DestinationPortEnd) + if sourcePort < um.Data.SourcePortStart || um.Data.SourcePortEnd < sourcePort { return false, false } - if destinationPort < tm.data.DestinationPortStart || tm.data.DestinationPortEnd < destinationPort { + if destinationPort < um.Data.DestinationPortStart || um.Data.DestinationPortEnd < destinationPort { return false, false } -- cgit v1.2.3 From b7853f688b4bcd3465c0c3087fcbd8d53bdf26ae Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Wed, 22 Jan 2020 14:46:15 -0800 Subject: Error marshalling the matcher. The iptables binary is looking for libxt_.so when it should be looking for libxt_udp.so, so it's having an issue reading the data in xt_match_entry. I think it may be an alignment issue. Trying to fix this is leading to me fighting with the metadata struct, so I'm gonna go kill that. --- pkg/abi/linux/netfilter.go | 5 +++++ pkg/sentry/socket/netfilter/netfilter.go | 35 ++++++++++++++++++++------------ pkg/tcpip/iptables/udp_matcher.go | 2 +- 3 files changed, 28 insertions(+), 14 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/abi/linux/netfilter.go b/pkg/abi/linux/netfilter.go index f0e544f9c..effed7976 100644 --- a/pkg/abi/linux/netfilter.go +++ b/pkg/abi/linux/netfilter.go @@ -198,6 +198,11 @@ type XTEntryMatch struct { // SizeOfXTEntryMatch is the size of an XTEntryMatch. const SizeOfXTEntryMatch = 32 +type KernelXTEntryMatch struct { + XTEntryMatch + Data []byte +} + // XTEntryTarget holds a target for a rule. For example, it can specify that // packets matching the rule should DROP, ACCEPT, or use an extension target. // iptables-extension(8) has a list of possible targets. diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index 3caabca9a..b49fe5b3e 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -207,26 +207,34 @@ func marshalMatcher(matcher iptables.Matcher) []byte { } func marshalUDPMatcher(matcher *iptables.UDPMatcher) []byte { - type udpMatch struct { - linux.XTEntryMatch - linux.XTUDP - } - linuxMatcher := udpMatch{ + linuxMatcher := linux.KernelXTEntryMatch{ XTEntryMatch: linux.XTEntryMatch{ MatchSize: linux.SizeOfXTEntryMatch + linux.SizeOfXTUDP, // Name: "udp", }, - XTUDP: linux.XTUDP{ - SourcePortStart: matcher.Data.SourcePortStart, - SourcePortEnd: matcher.Data.SourcePortEnd, - DestinationPortStart: matcher.Data.DestinationPortStart, - DestinationPortEnd: matcher.Data.DestinationPortEnd, - InverseFlags: matcher.Data.InverseFlags, - }, + Data: make([]byte, linux.SizeOfXTUDP+22), } + // copy(linuxMatcher.Name[:], "udp") copy(linuxMatcher.Name[:], "udp") - var buf [linux.SizeOfXTEntryMatch + linux.SizeOfXTUDP]byte + // TODO: Must be aligned. + xtudp := linux.XTUDP{ + SourcePortStart: matcher.Data.SourcePortStart, + SourcePortEnd: matcher.Data.SourcePortEnd, + DestinationPortStart: matcher.Data.DestinationPortStart, + DestinationPortEnd: matcher.Data.DestinationPortEnd, + InverseFlags: matcher.Data.InverseFlags, + } + binary.Marshal(linuxMatcher.Data[:linux.SizeOfXTUDP], usermem.ByteOrder, xtudp) + + if binary.Size(linuxMatcher)%64 != 0 { + panic(fmt.Sprintf("size is actually: %d", binary.Size(linuxMatcher))) + } + + var buf [linux.SizeOfXTEntryMatch + linux.SizeOfXTUDP + 22]byte + if len(buf)%64 != 0 { + panic(fmt.Sprintf("len is actually: %d", len(buf))) + } binary.Marshal(buf[:], usermem.ByteOrder, linuxMatcher) return buf[:] } @@ -245,6 +253,7 @@ func marshalTarget(target iptables.Target) []byte { } func marshalStandardTarget(verdict iptables.Verdict) []byte { + // TODO: Must be aligned. // The target's name will be the empty string. target := linux.XTStandardTarget{ Target: linux.XTEntryTarget{ diff --git a/pkg/tcpip/iptables/udp_matcher.go b/pkg/tcpip/iptables/udp_matcher.go index fca457199..65ae7f9e0 100644 --- a/pkg/tcpip/iptables/udp_matcher.go +++ b/pkg/tcpip/iptables/udp_matcher.go @@ -59,7 +59,7 @@ func NewUDPMatcher(filter IPHeaderFilter, data UDPMatcherData) (Matcher, error) } if filter.Protocol != header.UDPProtocolNumber { - log.Warningf("UDP matching is only valid for protocol %d.", header.UDPProtocolNumber) + return nil, fmt.Errorf("UDP matching is only valid for protocol %d.", header.UDPProtocolNumber) } return &UDPMatcher{Data: data}, nil -- cgit v1.2.3 From 2946fe81627afa223853769ed736e2a56e0144b7 Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Fri, 24 Jan 2020 17:12:03 -0800 Subject: We can now actually write out the udp matcher. --- pkg/sentry/socket/netfilter/netfilter.go | 78 ++++++++++++++++++++++++-------- 1 file changed, 58 insertions(+), 20 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index 3ca22932d..6c88a50a6 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -36,7 +36,7 @@ const errorTargetName = "ERROR" // metadata is opaque to netstack. It holds data that we need to translate // between Linux's and netstack's iptables representations. -// TODO(gvisor.dev/issue/170): This might be removable. +// TODO(gvisor.dev/issue/170): Use metadata to check correctness. type metadata struct { HookEntry [linux.NF_INET_NUMHOOKS]uint32 Underflow [linux.NF_INET_NUMHOOKS]uint32 @@ -44,6 +44,14 @@ type metadata struct { Size uint32 } +const enableDebugLog = true + +func nflog(format string, args ...interface{}) { + if enableDebugLog { + log.Infof("netfilter: "+format, args...) + } +} + // GetInfo returns information about iptables. func GetInfo(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr) (linux.IPTGetinfo, *syserr.Error) { // Read in the struct and table name. @@ -72,6 +80,8 @@ func GetInfo(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr) (linux.IPT info.NumEntries = metadata.NumEntries info.Size = metadata.Size + nflog("GetInfo returning info: %+v", info) + return info, nil } @@ -80,21 +90,26 @@ func GetEntries(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr, outLen // Read in the struct and table name. var userEntries linux.IPTGetEntries if _, err := t.CopyIn(outPtr, &userEntries); err != nil { + log.Warningf("netfilter: couldn't copy in entries %q", userEntries.Name) return linux.KernelIPTGetEntries{}, syserr.FromError(err) } // Find the appropriate table. table, err := findTable(stack, userEntries.Name) if err != nil { + log.Warningf("netfilter: couldn't find table %q", userEntries.Name) return linux.KernelIPTGetEntries{}, err } // Convert netstack's iptables rules to something that the iptables // tool can understand. - entries, _, err := convertNetstackToBinary(userEntries.Name.String(), table) + entries, meta, err := convertNetstackToBinary(userEntries.Name.String(), table) if err != nil { return linux.KernelIPTGetEntries{}, err } + if meta != table.Metadata().(metadata) { + panic(fmt.Sprintf("Table %q metadata changed between writing and reading. Was saved as %+v, but is now %+v", userEntries.Name.String(), table.Metadata().(metadata), meta)) + } if binary.Size(entries) > uintptr(outLen) { log.Warningf("Insufficient GetEntries output size: %d", uintptr(outLen)) return linux.KernelIPTGetEntries{}, syserr.ErrInvalidArgument @@ -148,15 +163,19 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern copy(entries.Name[:], tablename) for ruleIdx, rule := range table.Rules { + nflog("Current offset: %d", entries.Size) + // Is this a chain entry point? for hook, hookRuleIdx := range table.BuiltinChains { if hookRuleIdx == ruleIdx { + nflog("Found hook %d at offset %d", hook, entries.Size) meta.HookEntry[hook] = entries.Size } } // Is this a chain underflow point? for underflow, underflowRuleIdx := range table.Underflows { if underflowRuleIdx == ruleIdx { + nflog("Found underflow %d at offset %d", underflow, entries.Size) meta.Underflow[underflow] = entries.Size } } @@ -176,6 +195,10 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern // Serialize the matcher and add it to the // entry. serialized := marshalMatcher(matcher) + nflog("matcher serialized as: %v", serialized) + if len(serialized)%8 != 0 { + panic(fmt.Sprintf("matcher %T is not 64-bit aligned", matcher)) + } entry.Elems = append(entry.Elems, serialized...) entry.NextOffset += uint16(len(serialized)) entry.TargetOffset += uint16(len(serialized)) @@ -183,18 +206,25 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern // Serialize and append the target. serialized := marshalTarget(rule.Target) + if len(serialized)%8 != 0 { + panic(fmt.Sprintf("target %T is not 64-bit aligned", rule.Target)) + } entry.Elems = append(entry.Elems, serialized...) entry.NextOffset += uint16(len(serialized)) + nflog("Adding entry: %+v", entry) + entries.Size += uint32(entry.NextOffset) entries.Entrytable = append(entries.Entrytable, entry) meta.NumEntries++ } + nflog("Finished with an marshalled size of %d", meta.Size) meta.Size = entries.Size return entries, meta, nil } +// TODO: SOMEHOW THIS IS NOT GETTING APPENDED! func marshalMatcher(matcher iptables.Matcher) []byte { switch m := matcher.(type) { case *iptables.UDPMatcher: @@ -207,17 +237,17 @@ func marshalMatcher(matcher iptables.Matcher) []byte { } func marshalUDPMatcher(matcher *iptables.UDPMatcher) []byte { + nflog("Marshalling UDP matcher: %+v", matcher) + linuxMatcher := linux.KernelXTEntryMatch{ XTEntryMatch: linux.XTEntryMatch{ - MatchSize: linux.SizeOfXTEntryMatch + linux.SizeOfXTUDP, + MatchSize: linux.SizeOfXTEntryMatch + linux.SizeOfXTUDP + 6, // Name: "udp", }, - Data: make([]byte, linux.SizeOfXTUDP+22), + Data: make([]byte, 0, linux.SizeOfXTUDP), } - // copy(linuxMatcher.Name[:], "udp") copy(linuxMatcher.Name[:], "udp") - // TODO: Must be aligned. xtudp := linux.XTUDP{ SourcePortStart: matcher.Data.SourcePortStart, SourcePortEnd: matcher.Data.SourcePortEnd, @@ -225,17 +255,17 @@ func marshalUDPMatcher(matcher *iptables.UDPMatcher) []byte { DestinationPortEnd: matcher.Data.DestinationPortEnd, InverseFlags: matcher.Data.InverseFlags, } - binary.Marshal(linuxMatcher.Data[:linux.SizeOfXTUDP], usermem.ByteOrder, xtudp) - - if binary.Size(linuxMatcher)%64 != 0 { - panic(fmt.Sprintf("size is actually: %d", binary.Size(linuxMatcher))) - } - - var buf [linux.SizeOfXTEntryMatch + linux.SizeOfXTUDP + 22]byte - if len(buf)%64 != 0 { - panic(fmt.Sprintf("len is actually: %d", len(buf))) - } - binary.Marshal(buf[:], usermem.ByteOrder, linuxMatcher) + nflog("marshalUDPMatcher: xtudp: %+v", xtudp) + linuxMatcher.Data = binary.Marshal(linuxMatcher.Data, usermem.ByteOrder, xtudp) + nflog("marshalUDPMatcher: linuxMatcher: %+v", linuxMatcher) + + // We have to pad this struct size to a multiple of 8 bytes, so we make + // this a little longer than it needs to be. + buf := make([]byte, 0, linux.SizeOfXTEntryMatch+linux.SizeOfXTUDP+6) + buf = binary.Marshal(buf, usermem.ByteOrder, linuxMatcher) + buf = append(buf, []byte{0, 0, 0, 0, 0, 0}...) + nflog("Marshalled into matcher of size %d", len(buf)) + nflog("marshalUDPMatcher: buf is: %v", buf) return buf[:] } @@ -253,6 +283,8 @@ func marshalTarget(target iptables.Target) []byte { } func marshalStandardTarget(verdict iptables.Verdict) []byte { + nflog("Marshalling standard target with size %d", linux.SizeOfXTStandardTarget) + // TODO: Must be aligned. // The target's name will be the empty string. target := linux.XTStandardTarget{ @@ -321,7 +353,7 @@ func translateToStandardVerdict(val int32) (iptables.Verdict, *syserr.Error) { // SetEntries sets iptables rules for a single table. See // net/ipv4/netfilter/ip_tables.c:translate_table for reference. func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { - printReplace(optVal) + // printReplace(optVal) // Get the basic rules data (struct ipt_replace). if len(optVal) < linux.SizeOfIPTReplace { @@ -343,10 +375,14 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { return syserr.ErrInvalidArgument } + nflog("Setting entries in table %q", replace.Name.String()) + // Convert input into a list of rules and their offsets. var offset uint32 var offsets []uint32 for entryIdx := uint32(0); entryIdx < replace.NumEntries; entryIdx++ { + nflog("Processing entry at offset %d", offset) + // Get the struct ipt_entry. if len(optVal) < linux.SizeOfIPTEntry { log.Warningf("netfilter: optVal has insufficient size for entry %d", len(optVal)) @@ -464,9 +500,10 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // parseMatchers parses 0 or more matchers from optVal. optVal should contain // only the matchers. func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Matcher, *syserr.Error) { + nflog("Parsing matchers of size %d", len(optVal)) var matchers []iptables.Matcher for len(optVal) > 0 { - log.Infof("parseMatchers: optVal has len %d", len(optVal)) + nflog("parseMatchers: optVal has len %d", len(optVal)) // Get the XTEntryMatch. if len(optVal) < linux.SizeOfXTEntryMatch { log.Warningf("netfilter: optVal has insufficient size for entry match: %d", len(optVal)) @@ -475,7 +512,7 @@ func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Ma var match linux.XTEntryMatch buf := optVal[:linux.SizeOfXTEntryMatch] binary.Unmarshal(buf, usermem.ByteOrder, &match) - log.Infof("parseMatchers: parsed entry match %q: %+v", match.Name.String(), match) + nflog("parseMatchers: parsed entry match %q: %+v", match.Name.String(), match) // Check some invariants. if match.MatchSize < linux.SizeOfXTEntryMatch { @@ -532,6 +569,7 @@ func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Ma // parseTarget parses a target from optVal. optVal should contain only the // target. func parseTarget(optVal []byte) (iptables.Target, *syserr.Error) { + nflog("Parsing target of size %d", len(optVal)) if len(optVal) < linux.SizeOfXTEntryTarget { log.Warningf("netfilter: optVal has insufficient size for entry target %d", len(optVal)) return nil, syserr.ErrInvalidArgument -- cgit v1.2.3 From 29316e66adfc49c158425554761e34c12338f1d9 Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Mon, 27 Jan 2020 12:27:04 -0800 Subject: Cleanup for GH review. --- pkg/abi/linux/netfilter.go | 14 +-- pkg/sentry/socket/netfilter/netfilter.go | 144 ++++++++++++------------------- pkg/tcpip/iptables/types.go | 15 ---- pkg/tcpip/iptables/udp_matcher.go | 62 ++++++------- test/iptables/filter_input.go | 6 +- 5 files changed, 88 insertions(+), 153 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/abi/linux/netfilter.go b/pkg/abi/linux/netfilter.go index effed7976..8e40bcc62 100644 --- a/pkg/abi/linux/netfilter.go +++ b/pkg/abi/linux/netfilter.go @@ -198,6 +198,8 @@ type XTEntryMatch struct { // SizeOfXTEntryMatch is the size of an XTEntryMatch. const SizeOfXTEntryMatch = 32 +// KernelXTEntryMatch is identical to XTEntryMatch, but contains +// variable-length Data field. type KernelXTEntryMatch struct { XTEntryMatch Data []byte @@ -349,19 +351,19 @@ func goString(cstring []byte) string { // XTUDP holds data for matching UDP packets. It corresponds to struct xt_udp // in include/uapi/linux/netfilter/xt_tcpudp.h. type XTUDP struct { - // SourcePortStart specifies the inclusive start of the range of source - // ports to which the matcher applies. + // SourcePortStart is the inclusive start of the range of source ports + // to which the matcher applies. SourcePortStart uint16 - // SourcePortEnd specifies the inclusive end of the range of source ports - // to which the matcher applies. + // SourcePortEnd is the inclusive end of the range of source ports to + // which the matcher applies. SourcePortEnd uint16 - // DestinationPortStart specifies the start of the destination port + // DestinationPortStart is the inclusive start of the destination port // range to which the matcher applies. DestinationPortStart uint16 - // DestinationPortEnd specifies the start of the destination port + // DestinationPortEnd is the inclusive end of the destination port // range to which the matcher applies. DestinationPortEnd uint16 diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index 6c88a50a6..b8848f08a 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -34,9 +34,16 @@ import ( // shouldn't be reached - an error has occurred if we fall through to one. const errorTargetName = "ERROR" -// metadata is opaque to netstack. It holds data that we need to translate -// between Linux's and netstack's iptables representations. -// TODO(gvisor.dev/issue/170): Use metadata to check correctness. +const ( + matcherNameUDP = "udp" +) + +// Metadata is used to verify that we are correctly serializing and +// deserializing iptables into structs consumable by the iptables tool. We save +// a metadata struct when the tables are written, and when they are read out we +// verify that certain fields are the same. +// +// metadata is opaque to netstack. type metadata struct { HookEntry [linux.NF_INET_NUMHOOKS]uint32 Underflow [linux.NF_INET_NUMHOOKS]uint32 @@ -44,10 +51,12 @@ type metadata struct { Size uint32 } -const enableDebugLog = true +const enableDebug = false +// nflog logs messages related to the writing and reading of iptables, but only +// when enableDebug is true. func nflog(format string, args ...interface{}) { - if enableDebugLog { + if enableDebug { log.Infof("netfilter: "+format, args...) } } @@ -80,7 +89,7 @@ func GetInfo(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr) (linux.IPT info.NumEntries = metadata.NumEntries info.Size = metadata.Size - nflog("GetInfo returning info: %+v", info) + nflog("returning info: %+v", info) return info, nil } @@ -163,19 +172,19 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern copy(entries.Name[:], tablename) for ruleIdx, rule := range table.Rules { - nflog("Current offset: %d", entries.Size) + nflog("convert to binary: current offset: %d", entries.Size) // Is this a chain entry point? for hook, hookRuleIdx := range table.BuiltinChains { if hookRuleIdx == ruleIdx { - nflog("Found hook %d at offset %d", hook, entries.Size) + nflog("convert to binary: found hook %d at offset %d", hook, entries.Size) meta.HookEntry[hook] = entries.Size } } // Is this a chain underflow point? for underflow, underflowRuleIdx := range table.Underflows { if underflowRuleIdx == ruleIdx { - nflog("Found underflow %d at offset %d", underflow, entries.Size) + nflog("convert to binary: found underflow %d at offset %d", underflow, entries.Size) meta.Underflow[underflow] = entries.Size } } @@ -195,7 +204,7 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern // Serialize the matcher and add it to the // entry. serialized := marshalMatcher(matcher) - nflog("matcher serialized as: %v", serialized) + nflog("convert to binary: matcher serialized as: %v", serialized) if len(serialized)%8 != 0 { panic(fmt.Sprintf("matcher %T is not 64-bit aligned", matcher)) } @@ -212,14 +221,14 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern entry.Elems = append(entry.Elems, serialized...) entry.NextOffset += uint16(len(serialized)) - nflog("Adding entry: %+v", entry) + nflog("convert to binary: adding entry: %+v", entry) entries.Size += uint32(entry.NextOffset) entries.Entrytable = append(entries.Entrytable, entry) meta.NumEntries++ } - nflog("Finished with an marshalled size of %d", meta.Size) + nflog("convert to binary: finished with an marshalled size of %d", meta.Size) meta.Size = entries.Size return entries, meta, nil } @@ -237,16 +246,18 @@ func marshalMatcher(matcher iptables.Matcher) []byte { } func marshalUDPMatcher(matcher *iptables.UDPMatcher) []byte { - nflog("Marshalling UDP matcher: %+v", matcher) + nflog("convert to binary: marshalling UDP matcher: %+v", matcher) + + // We have to pad this struct size to a multiple of 8 bytes. + const size = linux.SizeOfXTEntryMatch + linux.SizeOfXTUDP + 6 linuxMatcher := linux.KernelXTEntryMatch{ XTEntryMatch: linux.XTEntryMatch{ - MatchSize: linux.SizeOfXTEntryMatch + linux.SizeOfXTUDP + 6, - // Name: "udp", + MatchSize: size, }, Data: make([]byte, 0, linux.SizeOfXTUDP), } - copy(linuxMatcher.Name[:], "udp") + copy(linuxMatcher.Name[:], matcherNameUDP) xtudp := linux.XTUDP{ SourcePortStart: matcher.Data.SourcePortStart, @@ -255,17 +266,12 @@ func marshalUDPMatcher(matcher *iptables.UDPMatcher) []byte { DestinationPortEnd: matcher.Data.DestinationPortEnd, InverseFlags: matcher.Data.InverseFlags, } - nflog("marshalUDPMatcher: xtudp: %+v", xtudp) linuxMatcher.Data = binary.Marshal(linuxMatcher.Data, usermem.ByteOrder, xtudp) - nflog("marshalUDPMatcher: linuxMatcher: %+v", linuxMatcher) - // We have to pad this struct size to a multiple of 8 bytes, so we make - // this a little longer than it needs to be. - buf := make([]byte, 0, linux.SizeOfXTEntryMatch+linux.SizeOfXTUDP+6) + buf := make([]byte, 0, size) buf = binary.Marshal(buf, usermem.ByteOrder, linuxMatcher) buf = append(buf, []byte{0, 0, 0, 0, 0, 0}...) - nflog("Marshalled into matcher of size %d", len(buf)) - nflog("marshalUDPMatcher: buf is: %v", buf) + nflog("convert to binary: marshalled UDP matcher into %v", buf) return buf[:] } @@ -283,9 +289,8 @@ func marshalTarget(target iptables.Target) []byte { } func marshalStandardTarget(verdict iptables.Verdict) []byte { - nflog("Marshalling standard target with size %d", linux.SizeOfXTStandardTarget) + nflog("convert to binary: marshalling standard target with size %d", linux.SizeOfXTStandardTarget) - // TODO: Must be aligned. // The target's name will be the empty string. target := linux.XTStandardTarget{ Target: linux.XTEntryTarget{ @@ -353,8 +358,6 @@ func translateToStandardVerdict(val int32) (iptables.Verdict, *syserr.Error) { // SetEntries sets iptables rules for a single table. See // net/ipv4/netfilter/ip_tables.c:translate_table for reference. func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { - // printReplace(optVal) - // Get the basic rules data (struct ipt_replace). if len(optVal) < linux.SizeOfIPTReplace { log.Warningf("netfilter.SetEntries: optVal has insufficient size for replace %d", len(optVal)) @@ -375,13 +378,13 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { return syserr.ErrInvalidArgument } - nflog("Setting entries in table %q", replace.Name.String()) + nflog("set entries: setting entries in table %q", replace.Name.String()) // Convert input into a list of rules and their offsets. var offset uint32 var offsets []uint32 for entryIdx := uint32(0); entryIdx < replace.NumEntries; entryIdx++ { - nflog("Processing entry at offset %d", offset) + nflog("set entries: processing entry at offset %d", offset) // Get the struct ipt_entry. if len(optVal) < linux.SizeOfIPTEntry { @@ -406,11 +409,13 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { return err } - // TODO: Matchers (and maybe targets) can specify that they only work for certiain protocols, hooks, tables. + // TODO(gvisor.dev/issue/170): Matchers and targets can specify + // that they only work for certiain protocols, hooks, tables. // Get matchers. matchersSize := entry.TargetOffset - linux.SizeOfIPTEntry if len(optVal) < int(matchersSize) { log.Warningf("netfilter: entry doesn't have enough room for its matchers (only %d bytes remain)", len(optVal)) + return syserr.ErrInvalidArgument } matchers, err := parseMatchers(filter, optVal[:matchersSize]) if err != nil { @@ -423,6 +428,7 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { targetSize := entry.NextOffset - entry.TargetOffset if len(optVal) < int(targetSize) { log.Warningf("netfilter: entry doesn't have enough room for its target (only %d bytes remain)", len(optVal)) + return syserr.ErrInvalidArgument } target, err := parseTarget(optVal[:targetSize]) if err != nil { @@ -500,10 +506,11 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // parseMatchers parses 0 or more matchers from optVal. optVal should contain // only the matchers. func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Matcher, *syserr.Error) { - nflog("Parsing matchers of size %d", len(optVal)) + nflog("set entries: parsing matchers of size %d", len(optVal)) var matchers []iptables.Matcher for len(optVal) > 0 { - nflog("parseMatchers: optVal has len %d", len(optVal)) + nflog("set entries: optVal has len %d", len(optVal)) + // Get the XTEntryMatch. if len(optVal) < linux.SizeOfXTEntryMatch { log.Warningf("netfilter: optVal has insufficient size for entry match: %d", len(optVal)) @@ -512,7 +519,7 @@ func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Ma var match linux.XTEntryMatch buf := optVal[:linux.SizeOfXTEntryMatch] binary.Unmarshal(buf, usermem.ByteOrder, &match) - nflog("parseMatchers: parsed entry match %q: %+v", match.Name.String(), match) + nflog("set entries: parsed entry match %q: %+v", match.Name.String(), match) // Check some invariants. if match.MatchSize < linux.SizeOfXTEntryMatch { @@ -528,17 +535,17 @@ func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Ma var matcher iptables.Matcher var err error switch match.Name.String() { - case "udp": + case matcherNameUDP: if len(buf) < linux.SizeOfXTUDP { log.Warningf("netfilter: optVal has insufficient size for UDP match: %d", len(optVal)) return nil, syserr.ErrInvalidArgument } + // For alignment reasons, the match's total size may + // exceed what's strictly necessary to hold matchData. var matchData linux.XTUDP - // For alignment reasons, the match's total size may exceed what's - // strictly necessary to hold matchData. binary.Unmarshal(buf[:linux.SizeOfXTUDP], usermem.ByteOrder, &matchData) log.Infof("parseMatchers: parsed XTUDP: %+v", matchData) - matcher, err = iptables.NewUDPMatcher(filter, iptables.UDPMatcherData{ + matcher, err = iptables.NewUDPMatcher(filter, iptables.UDPMatcherParams{ SourcePortStart: matchData.SourcePortStart, SourcePortEnd: matchData.SourcePortEnd, DestinationPortStart: matchData.DestinationPortStart, @@ -557,19 +564,22 @@ func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Ma matchers = append(matchers, matcher) - // TODO: Support revision. - // TODO: Support proto -- matchers usually specify which proto(s) they work with. + // TODO(gvisor.dev/issue/170): Check the revision field. optVal = optVal[match.MatchSize:] } - // TODO: Check that optVal is exhausted. + if len(optVal) != 0 { + log.Warningf("netfilter: optVal should be exhausted after parsing matchers") + return nil, syserr.ErrInvalidArgument + } + return matchers, nil } // parseTarget parses a target from optVal. optVal should contain only the // target. func parseTarget(optVal []byte) (iptables.Target, *syserr.Error) { - nflog("Parsing target of size %d", len(optVal)) + nflog("set entries: parsing target of size %d", len(optVal)) if len(optVal) < linux.SizeOfXTEntryTarget { log.Warningf("netfilter: optVal has insufficient size for entry target %d", len(optVal)) return nil, syserr.ErrInvalidArgument @@ -598,7 +608,8 @@ func parseTarget(optVal []byte) (iptables.Target, *syserr.Error) { case iptables.Drop: return iptables.UnconditionalDropTarget{}, nil default: - panic(fmt.Sprintf("Unknown verdict: %v", verdict)) + log.Warningf("Unknown verdict: %v", verdict) + return nil, syserr.ErrInvalidArgument } case errorTargetName: @@ -673,52 +684,3 @@ func hookFromLinux(hook int) iptables.Hook { } panic(fmt.Sprintf("Unknown hook %d does not correspond to a builtin chain", hook)) } - -// printReplace prints information about the struct ipt_replace in optVal. It -// is only for debugging. -func printReplace(optVal []byte) { - // Basic replace info. - var replace linux.IPTReplace - replaceBuf := optVal[:linux.SizeOfIPTReplace] - optVal = optVal[linux.SizeOfIPTReplace:] - binary.Unmarshal(replaceBuf, usermem.ByteOrder, &replace) - log.Infof("Replacing table %q: %+v", replace.Name.String(), replace) - - // Read in the list of entries at the end of replace. - var totalOffset uint16 - for entryIdx := uint32(0); entryIdx < replace.NumEntries; entryIdx++ { - var entry linux.IPTEntry - entryBuf := optVal[:linux.SizeOfIPTEntry] - binary.Unmarshal(entryBuf, usermem.ByteOrder, &entry) - log.Infof("Entry %d (total offset %d): %+v", entryIdx, totalOffset, entry) - - totalOffset += entry.NextOffset - if entry.TargetOffset == linux.SizeOfIPTEntry { - log.Infof("Entry has no matches.") - } else { - log.Infof("Entry has matches.") - } - - var target linux.XTEntryTarget - targetBuf := optVal[entry.TargetOffset : entry.TargetOffset+linux.SizeOfXTEntryTarget] - binary.Unmarshal(targetBuf, usermem.ByteOrder, &target) - log.Infof("Target named %q: %+v", target.Name.String(), target) - - switch target.Name.String() { - case "": - var standardTarget linux.XTStandardTarget - stBuf := optVal[entry.TargetOffset : entry.TargetOffset+linux.SizeOfXTStandardTarget] - binary.Unmarshal(stBuf, usermem.ByteOrder, &standardTarget) - log.Infof("Standard target with verdict %q (%d).", linux.VerdictStrings[standardTarget.Verdict], standardTarget.Verdict) - case errorTargetName: - var errorTarget linux.XTErrorTarget - etBuf := optVal[entry.TargetOffset : entry.TargetOffset+linux.SizeOfXTErrorTarget] - binary.Unmarshal(etBuf, usermem.ByteOrder, &errorTarget) - log.Infof("Error target with name %q.", errorTarget.Name.String()) - default: - log.Infof("Unknown target type.") - } - - optVal = optVal[entry.NextOffset:] - } -} diff --git a/pkg/tcpip/iptables/types.go b/pkg/tcpip/iptables/types.go index d47447d40..ba5ed75b4 100644 --- a/pkg/tcpip/iptables/types.go +++ b/pkg/tcpip/iptables/types.go @@ -169,8 +169,6 @@ type IPHeaderFilter struct { Protocol tcpip.TransportProtocolNumber } -// TODO: Should these be able to marshal/unmarshal themselves? -// TODO: Something has to map the name to the matcher. // A Matcher is the interface for matching packets. type Matcher interface { // Match returns whether the packet matches and whether the packet @@ -179,19 +177,6 @@ type Matcher interface { // // Precondition: packet.NetworkHeader is set. Match(hook Hook, packet tcpip.PacketBuffer, interfaceName string) (matches bool, hotdrop bool) - - // TODO: Make this typesafe by having each Matcher have their own, typed CheckEntry? - // CheckEntry(params MatchCheckEntryParams) bool -} - -// TODO: Unused? -type MatchCheckEntryParams struct { - Table string // TODO: Tables should be an enum... - Filter IPHeaderFilter - Info interface{} // TODO: Type unsafe. - // HookMask uint8 - // Family uint8 - // NFTCompat bool } // A Target is the interface for taking an action for a packet. diff --git a/pkg/tcpip/iptables/udp_matcher.go b/pkg/tcpip/iptables/udp_matcher.go index 65ae7f9e0..f59ca2027 100644 --- a/pkg/tcpip/iptables/udp_matcher.go +++ b/pkg/tcpip/iptables/udp_matcher.go @@ -16,33 +16,28 @@ package iptables import ( "fmt" - "runtime/debug" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" ) +// TODO(gvisor.dev/issue/170): The following per-matcher params should be +// supported: +// - Table name +// - Match size +// - User size +// - Hooks +// - Proto +// - Family + +// UDPMatcher matches UDP packets and their headers. It implements Matcher. type UDPMatcher struct { - Data UDPMatcherData - - // tablename string - // unsigned int matchsize; - // unsigned int usersize; - // #ifdef CONFIG_COMPAT - // unsigned int compatsize; - // #endif - // unsigned int hooks; - // unsigned short proto; - // unsigned short family; + Data UDPMatcherParams } -// TODO: Delete? -// MatchCheckEntryParams - -type UDPMatcherData struct { - // Filter IPHeaderFilter - +// UDPMatcherParams are the parameters used to create a UDPMatcher. +type UDPMatcherParams struct { SourcePortStart uint16 SourcePortEnd uint16 DestinationPortStart uint16 @@ -50,12 +45,12 @@ type UDPMatcherData struct { InverseFlags uint8 } -func NewUDPMatcher(filter IPHeaderFilter, data UDPMatcherData) (Matcher, error) { - // TODO: We currently only support source port and destination port. - log.Infof("Adding rule with UDPMatcherData: %+v", data) +// NewUDPMatcher returns a new instance of UDPMatcher. +func NewUDPMatcher(filter IPHeaderFilter, data UDPMatcherParams) (Matcher, error) { + log.Infof("Adding rule with UDPMatcherParams: %+v", data) if data.InverseFlags != 0 { - return nil, fmt.Errorf("unsupported UDP matcher flags set") + return nil, fmt.Errorf("unsupported UDP matcher inverse flags set") } if filter.Protocol != header.UDPProtocolNumber { @@ -65,21 +60,18 @@ func NewUDPMatcher(filter IPHeaderFilter, data UDPMatcherData) (Matcher, error) return &UDPMatcher{Data: data}, nil } -// TODO: Check xt_tcpudp.c. Need to check for same things (e.g. fragments). +// Match implements Matcher.Match. func (um *UDPMatcher) Match(hook Hook, pkt tcpip.PacketBuffer, interfaceName string) (bool, bool) { - log.Infof("UDPMatcher called from: %s", string(debug.Stack())) netHeader := header.IPv4(pkt.NetworkHeader) - // TODO: Do we check proto here or elsewhere? I think elsewhere (check - // codesearch). + // TODO(gvisor.dev/issue/170): Proto checks should ultimately be moved + // into the iptables.Check codepath as matchers are added. if netHeader.TransportProtocol() != header.UDPProtocolNumber { - log.Infof("UDPMatcher: wrong protocol number") return false, false } // We dont't match fragments. if frag := netHeader.FragmentOffset(); frag != 0 { - log.Infof("UDPMatcher: it's a fragment") if frag == 1 { return false, true } @@ -89,20 +81,18 @@ func (um *UDPMatcher) Match(hook Hook, pkt tcpip.PacketBuffer, interfaceName str // Now we need the transport header. However, this may not have been set // yet. - // TODO + // TODO(gvisor.dev/issue/170): Parsing the transport header should + // ultimately be moved into the iptables.Check codepath as matchers are + // added. var udpHeader header.UDP if pkt.TransportHeader != nil { - log.Infof("UDPMatcher: transport header is not nil") udpHeader = header.UDP(pkt.TransportHeader) } else { - log.Infof("UDPMatcher: transport header is nil") - log.Infof("UDPMatcher: is network header nil: %t", pkt.NetworkHeader == nil) // The UDP header hasn't been parsed yet. We have to do it here. if len(pkt.Data.First()) < header.UDPMinimumSize { // There's no valid UDP header here, so we hotdrop the // packet. - // TODO: Stats. - log.Warningf("Dropping UDP packet: size to small.") + log.Warningf("Dropping UDP packet: size too small.") return false, true } udpHeader = header.UDP(pkt.Data.First()) @@ -112,10 +102,6 @@ func (um *UDPMatcher) Match(hook Hook, pkt tcpip.PacketBuffer, interfaceName str // matching range. sourcePort := udpHeader.SourcePort() destinationPort := udpHeader.DestinationPort() - log.Infof("UDPMatcher: sport and dport are %d and %d. sports and dport start and end are (%d, %d) and (%d, %d)", - udpHeader.SourcePort(), udpHeader.DestinationPort(), - um.Data.SourcePortStart, um.Data.SourcePortEnd, - um.Data.DestinationPortStart, um.Data.DestinationPortEnd) if sourcePort < um.Data.SourcePortStart || um.Data.SourcePortEnd < sourcePort { return false, false } diff --git a/test/iptables/filter_input.go b/test/iptables/filter_input.go index bc963d40e..e9f0978eb 100644 --- a/test/iptables/filter_input.go +++ b/test/iptables/filter_input.go @@ -264,9 +264,9 @@ func (FilterInputMultiUDPRules) ContainerAction(ip net.IP) error { if err := filterTable("-A", "INPUT", "-p", "udp", "-m", "udp", "--destination-port", fmt.Sprintf("%d", dropPort), "-j", "DROP"); err != nil { return err } - // if err := filterTable("-A", "INPUT", "-p", "udp", "-m", "udp", "--destination-port", fmt.Sprintf("%d", acceptPort), "-j", "ACCEPT"); err != nil { - // return err - // } + if err := filterTable("-A", "INPUT", "-p", "udp", "-m", "udp", "--destination-port", fmt.Sprintf("%d", acceptPort), "-j", "ACCEPT"); err != nil { + return err + } return filterTable("-L") } -- cgit v1.2.3 From d6a2e01d3e57e0837c7e5cfda3b56c4dcfbb4627 Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Mon, 27 Jan 2020 16:40:46 -0800 Subject: Address GH comments. --- pkg/sentry/socket/netfilter/netfilter.go | 25 ++++++++++++------------- 1 file changed, 12 insertions(+), 13 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index b8848f08a..a06562743 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -43,7 +43,7 @@ const ( // a metadata struct when the tables are written, and when they are read out we // verify that certain fields are the same. // -// metadata is opaque to netstack. +// metadata is used by this serialization/deserializing code, not netstack. type metadata struct { HookEntry [linux.NF_INET_NUMHOOKS]uint32 Underflow [linux.NF_INET_NUMHOOKS]uint32 @@ -51,14 +51,10 @@ type metadata struct { Size uint32 } -const enableDebug = false - // nflog logs messages related to the writing and reading of iptables, but only // when enableDebug is true. func nflog(format string, args ...interface{}) { - if enableDebug { - log.Infof("netfilter: "+format, args...) - } + log.Infof("netfilter: "+format, args...) } // GetInfo returns information about iptables. @@ -233,14 +229,12 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern return entries, meta, nil } -// TODO: SOMEHOW THIS IS NOT GETTING APPENDED! func marshalMatcher(matcher iptables.Matcher) []byte { switch m := matcher.(type) { case *iptables.UDPMatcher: return marshalUDPMatcher(m) default: - // TODO(gvisor.dev/issue/170): We don't support any matchers - // yet, so any call to marshalMatcher will panic. + // TODO(gvisor.dev/issue/170): Support other matchers. panic(fmt.Errorf("unknown matcher of type %T", matcher)) } } @@ -249,11 +243,11 @@ func marshalUDPMatcher(matcher *iptables.UDPMatcher) []byte { nflog("convert to binary: marshalling UDP matcher: %+v", matcher) // We have to pad this struct size to a multiple of 8 bytes. - const size = linux.SizeOfXTEntryMatch + linux.SizeOfXTUDP + 6 + size := alignUp(linux.SizeOfXTEntryMatch+linux.SizeOfXTUDP, 8) linuxMatcher := linux.KernelXTEntryMatch{ XTEntryMatch: linux.XTEntryMatch{ - MatchSize: size, + MatchSize: uint16(size), }, Data: make([]byte, 0, linux.SizeOfXTUDP), } @@ -270,7 +264,7 @@ func marshalUDPMatcher(matcher *iptables.UDPMatcher) []byte { buf := make([]byte, 0, size) buf = binary.Marshal(buf, usermem.ByteOrder, linuxMatcher) - buf = append(buf, []byte{0, 0, 0, 0, 0, 0}...) + buf = append(buf, make([]byte, size-len(buf))...) nflog("convert to binary: marshalled UDP matcher into %v", buf) return buf[:] } @@ -410,7 +404,7 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { } // TODO(gvisor.dev/issue/170): Matchers and targets can specify - // that they only work for certiain protocols, hooks, tables. + // that they only work for certain protocols, hooks, tables. // Get matchers. matchersSize := entry.TargetOffset - linux.SizeOfIPTEntry if len(optVal) < int(matchersSize) { @@ -684,3 +678,8 @@ func hookFromLinux(hook int) iptables.Hook { } panic(fmt.Sprintf("Unknown hook %d does not correspond to a builtin chain", hook)) } + +// alignUp rounds a length up to an alignment. align must be a power of 2. +func alignUp(length int, align uint) int { + return (length + int(align) - 1) & ^(int(align) - 1) +} -- cgit v1.2.3 From 51b783505b1ec164b02b48a0fd234509fba01a73 Mon Sep 17 00:00:00 2001 From: Bhasker Hariharan Date: Wed, 29 Jan 2020 15:41:51 -0800 Subject: Add support for TCP_DEFER_ACCEPT. PiperOrigin-RevId: 292233574 --- pkg/sentry/socket/netstack/netstack.go | 22 ++++ pkg/tcpip/tcpip.go | 6 ++ pkg/tcpip/transport/tcp/BUILD | 1 + pkg/tcpip/transport/tcp/accept.go | 25 ++--- pkg/tcpip/transport/tcp/connect.go | 53 +++++++++- pkg/tcpip/transport/tcp/endpoint.go | 26 ++++- pkg/tcpip/transport/tcp/forwarder.go | 4 +- pkg/tcpip/transport/tcp/tcp_test.go | 126 ++++++++++++++++++++++ test/syscalls/linux/socket_inet_loopback.cc | 158 ++++++++++++++++++++++++++++ test/syscalls/linux/tcp_socket.cc | 53 ++++++++++ 10 files changed, 451 insertions(+), 23 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index 8619cc506..049d04bf2 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -1260,6 +1260,18 @@ func getSockOptTCP(t *kernel.Task, ep commonEndpoint, name, outLen int) (interfa return int32(time.Duration(v) / time.Second), nil + case linux.TCP_DEFER_ACCEPT: + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + + var v tcpip.TCPDeferAcceptOption + if err := ep.GetSockOpt(&v); err != nil { + return nil, syserr.TranslateNetstackError(err) + } + + return int32(time.Duration(v) / time.Second), nil + default: emitUnimplementedEventTCP(t, name) } @@ -1713,6 +1725,16 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * v := usermem.ByteOrder.Uint32(optVal) return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.TCPLingerTimeoutOption(time.Second * time.Duration(v)))) + case linux.TCP_DEFER_ACCEPT: + if len(optVal) < sizeOfInt32 { + return syserr.ErrInvalidArgument + } + v := int32(usermem.ByteOrder.Uint32(optVal)) + if v < 0 { + v = 0 + } + return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.TCPDeferAcceptOption(time.Second * time.Duration(v)))) + case linux.TCP_REPAIR_OPTIONS: t.Kernel().EmitUnimplementedEvent(t) diff --git a/pkg/tcpip/tcpip.go b/pkg/tcpip/tcpip.go index 59c9b3fb0..0fa141d58 100644 --- a/pkg/tcpip/tcpip.go +++ b/pkg/tcpip/tcpip.go @@ -626,6 +626,12 @@ type TCPLingerTimeoutOption time.Duration // before being marked closed. type TCPTimeWaitTimeoutOption time.Duration +// TCPDeferAcceptOption is used by SetSockOpt/GetSockOpt to allow a +// accept to return a completed connection only when there is data to be +// read. This usually means the listening socket will drop the final ACK +// for a handshake till the specified timeout until a segment with data arrives. +type TCPDeferAcceptOption time.Duration + // MulticastTTLOption is used by SetSockOpt/GetSockOpt to control the default // TTL value for multicast messages. The default is 1. type MulticastTTLOption uint8 diff --git a/pkg/tcpip/transport/tcp/BUILD b/pkg/tcpip/transport/tcp/BUILD index 4acd9fb9a..7b4a87a2d 100644 --- a/pkg/tcpip/transport/tcp/BUILD +++ b/pkg/tcpip/transport/tcp/BUILD @@ -57,6 +57,7 @@ go_library( imports = ["gvisor.dev/gvisor/pkg/tcpip/buffer"], visibility = ["//visibility:public"], deps = [ + "//pkg/log", "//pkg/rand", "//pkg/sleep", "//pkg/sync", diff --git a/pkg/tcpip/transport/tcp/accept.go b/pkg/tcpip/transport/tcp/accept.go index d469758eb..6101f2945 100644 --- a/pkg/tcpip/transport/tcp/accept.go +++ b/pkg/tcpip/transport/tcp/accept.go @@ -222,13 +222,13 @@ func (l *listenContext) isCookieValid(id stack.TransportEndpointID, cookie seqnu // createConnectingEndpoint creates a new endpoint in a connecting state, with // the connection parameters given by the arguments. -func (l *listenContext) createConnectingEndpoint(s *segment, iss seqnum.Value, irs seqnum.Value, rcvdSynOpts *header.TCPSynOptions) (*endpoint, *tcpip.Error) { +func (l *listenContext) createConnectingEndpoint(s *segment, iss seqnum.Value, irs seqnum.Value, rcvdSynOpts *header.TCPSynOptions, queue *waiter.Queue) (*endpoint, *tcpip.Error) { // Create a new endpoint. netProto := l.netProto if netProto == 0 { netProto = s.route.NetProto } - n := newEndpoint(l.stack, netProto, nil) + n := newEndpoint(l.stack, netProto, queue) n.v6only = l.v6only n.ID = s.id n.boundNICID = s.route.NICID() @@ -273,16 +273,17 @@ func (l *listenContext) createConnectingEndpoint(s *segment, iss seqnum.Value, i // createEndpoint creates a new endpoint in connected state and then performs // the TCP 3-way handshake. -func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *header.TCPSynOptions) (*endpoint, *tcpip.Error) { +func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *header.TCPSynOptions, queue *waiter.Queue) (*endpoint, *tcpip.Error) { // Create new endpoint. irs := s.sequenceNumber isn := generateSecureISN(s.id, l.stack.Seed()) - ep, err := l.createConnectingEndpoint(s, isn, irs, opts) + ep, err := l.createConnectingEndpoint(s, isn, irs, opts, queue) if err != nil { return nil, err } // listenEP is nil when listenContext is used by tcp.Forwarder. + deferAccept := time.Duration(0) if l.listenEP != nil { l.listenEP.mu.Lock() if l.listenEP.EndpointState() != StateListen { @@ -290,13 +291,12 @@ func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *head return nil, tcpip.ErrConnectionAborted } l.addPendingEndpoint(ep) + deferAccept = l.listenEP.deferAccept l.listenEP.mu.Unlock() } // Perform the 3-way handshake. - h := newHandshake(ep, seqnum.Size(ep.initialReceiveWindow())) - - h.resetToSynRcvd(isn, irs, opts) + h := newPassiveHandshake(ep, seqnum.Size(ep.initialReceiveWindow()), isn, irs, opts, deferAccept) if err := h.execute(); err != nil { ep.Close() if l.listenEP != nil { @@ -377,16 +377,14 @@ func (e *endpoint) handleSynSegment(ctx *listenContext, s *segment, opts *header defer e.decSynRcvdCount() defer s.decRef() - n, err := ctx.createEndpointAndPerformHandshake(s, opts) + n, err := ctx.createEndpointAndPerformHandshake(s, opts, &waiter.Queue{}) if err != nil { e.stack.Stats().TCP.FailedConnectionAttempts.Increment() e.stats.FailedConnectionAttempts.Increment() return } ctx.removePendingEndpoint(n) - // Start the protocol goroutine. - wq := &waiter.Queue{} - n.startAcceptedLoop(wq) + n.startAcceptedLoop() e.stack.Stats().TCP.PassiveConnectionOpenings.Increment() e.deliverAccepted(n) @@ -546,7 +544,7 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) { rcvdSynOptions.TSEcr = s.parsedOptions.TSEcr } - n, err := ctx.createConnectingEndpoint(s, s.ackNumber-1, s.sequenceNumber-1, rcvdSynOptions) + n, err := ctx.createConnectingEndpoint(s, s.ackNumber-1, s.sequenceNumber-1, rcvdSynOptions, &waiter.Queue{}) if err != nil { e.stack.Stats().TCP.FailedConnectionAttempts.Increment() e.stats.FailedConnectionAttempts.Increment() @@ -576,8 +574,7 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) { // space available in the backlog. // Start the protocol goroutine. - wq := &waiter.Queue{} - n.startAcceptedLoop(wq) + n.startAcceptedLoop() e.stack.Stats().TCP.PassiveConnectionOpenings.Increment() go e.deliverAccepted(n) } diff --git a/pkg/tcpip/transport/tcp/connect.go b/pkg/tcpip/transport/tcp/connect.go index 4e3c5419c..9ff7ac261 100644 --- a/pkg/tcpip/transport/tcp/connect.go +++ b/pkg/tcpip/transport/tcp/connect.go @@ -86,6 +86,19 @@ type handshake struct { // rcvWndScale is the receive window scale, as defined in RFC 1323. rcvWndScale int + + // startTime is the time at which the first SYN/SYN-ACK was sent. + startTime time.Time + + // deferAccept if non-zero will drop the final ACK for a passive + // handshake till an ACK segment with data is received or the timeout is + // hit. + deferAccept time.Duration + + // acked is true if the the final ACK for a 3-way handshake has + // been received. This is required to stop retransmitting the + // original SYN-ACK when deferAccept is enabled. + acked bool } func newHandshake(ep *endpoint, rcvWnd seqnum.Size) handshake { @@ -112,6 +125,12 @@ func newHandshake(ep *endpoint, rcvWnd seqnum.Size) handshake { return h } +func newPassiveHandshake(ep *endpoint, rcvWnd seqnum.Size, isn, irs seqnum.Value, opts *header.TCPSynOptions, deferAccept time.Duration) handshake { + h := newHandshake(ep, rcvWnd) + h.resetToSynRcvd(isn, irs, opts, deferAccept) + return h +} + // FindWndScale determines the window scale to use for the given maximum window // size. func FindWndScale(wnd seqnum.Size) int { @@ -181,7 +200,7 @@ func (h *handshake) effectiveRcvWndScale() uint8 { // resetToSynRcvd resets the state of the handshake object to the SYN-RCVD // state. -func (h *handshake) resetToSynRcvd(iss seqnum.Value, irs seqnum.Value, opts *header.TCPSynOptions) { +func (h *handshake) resetToSynRcvd(iss seqnum.Value, irs seqnum.Value, opts *header.TCPSynOptions, deferAccept time.Duration) { h.active = false h.state = handshakeSynRcvd h.flags = header.TCPFlagSyn | header.TCPFlagAck @@ -189,6 +208,7 @@ func (h *handshake) resetToSynRcvd(iss seqnum.Value, irs seqnum.Value, opts *hea h.ackNum = irs + 1 h.mss = opts.MSS h.sndWndScale = opts.WS + h.deferAccept = deferAccept h.ep.mu.Lock() h.ep.setEndpointState(StateSynRecv) h.ep.mu.Unlock() @@ -352,6 +372,14 @@ func (h *handshake) synRcvdState(s *segment) *tcpip.Error { // We have previously received (and acknowledged) the peer's SYN. If the // peer acknowledges our SYN, the handshake is completed. if s.flagIsSet(header.TCPFlagAck) { + // If deferAccept is not zero and this is a bare ACK and the + // timeout is not hit then drop the ACK. + if h.deferAccept != 0 && s.data.Size() == 0 && time.Since(h.startTime) < h.deferAccept { + h.acked = true + h.ep.stack.Stats().DroppedPackets.Increment() + return nil + } + // If the timestamp option is negotiated and the segment does // not carry a timestamp option then the segment must be dropped // as per https://tools.ietf.org/html/rfc7323#section-3.2. @@ -365,10 +393,16 @@ func (h *handshake) synRcvdState(s *segment) *tcpip.Error { h.ep.updateRecentTimestamp(s.parsedOptions.TSVal, h.ackNum, s.sequenceNumber) } h.state = handshakeCompleted + h.ep.mu.Lock() h.ep.transitionToStateEstablishedLocked(h) + // If the segment has data then requeue it for the receiver + // to process it again once main loop is started. + if s.data.Size() > 0 { + s.incRef() + h.ep.enqueueSegment(s) + } h.ep.mu.Unlock() - return nil } @@ -471,6 +505,7 @@ func (h *handshake) execute() *tcpip.Error { } } + h.startTime = time.Now() // Initialize the resend timer. resendWaker := sleep.Waker{} timeOut := time.Duration(time.Second) @@ -524,11 +559,21 @@ func (h *handshake) execute() *tcpip.Error { switch index, _ := s.Fetch(true); index { case wakerForResend: timeOut *= 2 - if timeOut > 60*time.Second { + if timeOut > MaxRTO { return tcpip.ErrTimeout } rt.Reset(timeOut) - h.ep.sendSynTCP(&h.ep.route, h.ep.ID, h.ep.ttl, h.ep.sendTOS, h.flags, h.iss, h.ackNum, h.rcvWnd, synOpts) + // Resend the SYN/SYN-ACK only if the following conditions hold. + // - It's an active handshake (deferAccept does not apply) + // - It's a passive handshake and we have not yet got the final-ACK. + // - It's a passive handshake and we got an ACK but deferAccept is + // enabled and we are now past the deferAccept duration. + // The last is required to provide a way for the peer to complete + // the connection with another ACK or data (as ACKs are never + // retransmitted on their own). + if h.active || !h.acked || h.deferAccept != 0 && time.Since(h.startTime) > h.deferAccept { + h.ep.sendSynTCP(&h.ep.route, h.ep.ID, h.ep.ttl, h.ep.sendTOS, h.flags, h.iss, h.ackNum, h.rcvWnd, synOpts) + } case wakerForNotification: n := h.ep.fetchNotifications() diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go index 13718ff55..8d52414b7 100644 --- a/pkg/tcpip/transport/tcp/endpoint.go +++ b/pkg/tcpip/transport/tcp/endpoint.go @@ -498,6 +498,13 @@ type endpoint struct { // without any data being acked. userTimeout time.Duration + // deferAccept if non-zero specifies a user specified time during + // which the final ACK of a handshake will be dropped provided the + // ACK is a bare ACK and carries no data. If the timeout is crossed then + // the bare ACK is accepted and the connection is delivered to the + // listener. + deferAccept time.Duration + // pendingAccepted is a synchronization primitive used to track number // of connections that are queued up to be delivered to the accepted // channel. We use this to ensure that all goroutines blocked on writing @@ -1574,6 +1581,15 @@ func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { e.mu.Unlock() return nil + case tcpip.TCPDeferAcceptOption: + e.mu.Lock() + if time.Duration(v) > MaxRTO { + v = tcpip.TCPDeferAcceptOption(MaxRTO) + } + e.deferAccept = time.Duration(v) + e.mu.Unlock() + return nil + default: return nil } @@ -1798,6 +1814,12 @@ func (e *endpoint) GetSockOpt(opt interface{}) *tcpip.Error { e.mu.Unlock() return nil + case *tcpip.TCPDeferAcceptOption: + e.mu.Lock() + *o = tcpip.TCPDeferAcceptOption(e.deferAccept) + e.mu.Unlock() + return nil + default: return tcpip.ErrUnknownProtocolOption } @@ -2149,9 +2171,8 @@ func (e *endpoint) listen(backlog int) *tcpip.Error { // startAcceptedLoop sets up required state and starts a goroutine with the // main loop for accepted connections. -func (e *endpoint) startAcceptedLoop(waiterQueue *waiter.Queue) { +func (e *endpoint) startAcceptedLoop() { e.mu.Lock() - e.waiterQueue = waiterQueue e.workerRunning = true e.mu.Unlock() wakerInitDone := make(chan struct{}) @@ -2177,7 +2198,6 @@ func (e *endpoint) Accept() (tcpip.Endpoint, *waiter.Queue, *tcpip.Error) { default: return nil, nil, tcpip.ErrWouldBlock } - return n, n.waiterQueue, nil } diff --git a/pkg/tcpip/transport/tcp/forwarder.go b/pkg/tcpip/transport/tcp/forwarder.go index 7eb613be5..c9ee5bf06 100644 --- a/pkg/tcpip/transport/tcp/forwarder.go +++ b/pkg/tcpip/transport/tcp/forwarder.go @@ -157,13 +157,13 @@ func (r *ForwarderRequest) CreateEndpoint(queue *waiter.Queue) (tcpip.Endpoint, TSVal: r.synOptions.TSVal, TSEcr: r.synOptions.TSEcr, SACKPermitted: r.synOptions.SACKPermitted, - }) + }, queue) if err != nil { return nil, err } // Start the protocol goroutine. - ep.startAcceptedLoop(queue) + ep.startAcceptedLoop() return ep, nil } diff --git a/pkg/tcpip/transport/tcp/tcp_test.go b/pkg/tcpip/transport/tcp/tcp_test.go index df2fb1071..a12336d47 100644 --- a/pkg/tcpip/transport/tcp/tcp_test.go +++ b/pkg/tcpip/transport/tcp/tcp_test.go @@ -6787,3 +6787,129 @@ func TestIncreaseWindowOnBufferResize(t *testing.T) { ), ) } + +func TestTCPDeferAccept(t *testing.T) { + c := context.New(t, defaultMTU) + defer c.Cleanup() + + c.Create(-1) + + if err := c.EP.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { + t.Fatal("Bind failed:", err) + } + + if err := c.EP.Listen(10); err != nil { + t.Fatal("Listen failed:", err) + } + + const tcpDeferAccept = 1 * time.Second + if err := c.EP.SetSockOpt(tcpip.TCPDeferAcceptOption(tcpDeferAccept)); err != nil { + t.Fatalf("c.EP.SetSockOpt(TCPDeferAcceptOption(%s) failed: %v", tcpDeferAccept, err) + } + + irs, iss := executeHandshake(t, c, context.TestPort, false /* synCookiesInUse */) + + if _, _, err := c.EP.Accept(); err != tcpip.ErrWouldBlock { + t.Fatalf("c.EP.Accept() returned unexpected error got: %v, want: %s", err, tcpip.ErrWouldBlock) + } + + // Send data. This should result in an acceptable endpoint. + c.SendPacket([]byte{1, 2, 3, 4}, &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck, + SeqNum: irs + 1, + AckNum: iss + 1, + }) + + // Receive ACK for the data we sent. + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.DstPort(context.TestPort), + checker.TCPFlags(header.TCPFlagAck), + checker.SeqNum(uint32(iss+1)), + checker.AckNum(uint32(irs+5)))) + + // Give a bit of time for the socket to be delivered to the accept queue. + time.Sleep(50 * time.Millisecond) + aep, _, err := c.EP.Accept() + if err != nil { + t.Fatalf("c.EP.Accept() returned unexpected error got: %v, want: nil", err) + } + + aep.Close() + // Closing aep without reading the data should trigger a RST. + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.DstPort(context.TestPort), + checker.TCPFlags(header.TCPFlagRst|header.TCPFlagAck), + checker.SeqNum(uint32(iss+1)), + checker.AckNum(uint32(irs+5)))) +} + +func TestTCPDeferAcceptTimeout(t *testing.T) { + c := context.New(t, defaultMTU) + defer c.Cleanup() + + c.Create(-1) + + if err := c.EP.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { + t.Fatal("Bind failed:", err) + } + + if err := c.EP.Listen(10); err != nil { + t.Fatal("Listen failed:", err) + } + + const tcpDeferAccept = 1 * time.Second + if err := c.EP.SetSockOpt(tcpip.TCPDeferAcceptOption(tcpDeferAccept)); err != nil { + t.Fatalf("c.EP.SetSockOpt(TCPDeferAcceptOption(%s) failed: %v", tcpDeferAccept, err) + } + + irs, iss := executeHandshake(t, c, context.TestPort, false /* synCookiesInUse */) + + if _, _, err := c.EP.Accept(); err != tcpip.ErrWouldBlock { + t.Fatalf("c.EP.Accept() returned unexpected error got: %v, want: %s", err, tcpip.ErrWouldBlock) + } + + // Sleep for a little of the tcpDeferAccept timeout. + time.Sleep(tcpDeferAccept + 100*time.Millisecond) + + // On timeout expiry we should get a SYN-ACK retransmission. + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.TCPFlags(header.TCPFlagAck|header.TCPFlagSyn), + checker.AckNum(uint32(irs)+1))) + + // Send data. This should result in an acceptable endpoint. + c.SendPacket([]byte{1, 2, 3, 4}, &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck, + SeqNum: irs + 1, + AckNum: iss + 1, + }) + + // Receive ACK for the data we sent. + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.TCPFlags(header.TCPFlagAck), + checker.SeqNum(uint32(iss+1)), + checker.AckNum(uint32(irs+5)))) + + // Give sometime for the endpoint to be delivered to the accept queue. + time.Sleep(50 * time.Millisecond) + aep, _, err := c.EP.Accept() + if err != nil { + t.Fatalf("c.EP.Accept() returned unexpected error got: %v, want: nil", err) + } + + aep.Close() + // Closing aep without reading the data should trigger a RST. + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.TCPFlags(header.TCPFlagRst|header.TCPFlagAck), + checker.SeqNum(uint32(iss+1)), + checker.AckNum(uint32(irs+5)))) +} diff --git a/test/syscalls/linux/socket_inet_loopback.cc b/test/syscalls/linux/socket_inet_loopback.cc index 2f9821555..3bf7081b9 100644 --- a/test/syscalls/linux/socket_inet_loopback.cc +++ b/test/syscalls/linux/socket_inet_loopback.cc @@ -828,6 +828,164 @@ TEST_P(SocketInetLoopbackTest, AcceptedInheritsTCPUserTimeout) { EXPECT_EQ(get, kUserTimeout); } +// TODO(gvisor.dev/issue/1688): Partially completed passive endpoints are not +// saved. Enable S/R once issue is fixed. +TEST_P(SocketInetLoopbackTest, TCPDeferAccept_NoRandomSave) { + // TODO(gvisor.dev/issue/1688): Partially completed passive endpoints are not + // saved. Enable S/R issue is fixed. + DisableSave ds; + + auto const& param = GetParam(); + TestAddress const& listener = param.listener; + TestAddress const& connector = param.connector; + + // Create the listening socket. + const FileDescriptor listen_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(listener.family(), SOCK_STREAM, IPPROTO_TCP)); + sockaddr_storage listen_addr = listener.addr; + ASSERT_THAT(bind(listen_fd.get(), reinterpret_cast(&listen_addr), + listener.addr_len), + SyscallSucceeds()); + ASSERT_THAT(listen(listen_fd.get(), SOMAXCONN), SyscallSucceeds()); + + // Get the port bound by the listening socket. + socklen_t addrlen = listener.addr_len; + ASSERT_THAT(getsockname(listen_fd.get(), + reinterpret_cast(&listen_addr), &addrlen), + SyscallSucceeds()); + + const uint16_t port = + ASSERT_NO_ERRNO_AND_VALUE(AddrPort(listener.family(), listen_addr)); + + // Set the TCP_DEFER_ACCEPT on the listening socket. + constexpr int kTCPDeferAccept = 3; + ASSERT_THAT(setsockopt(listen_fd.get(), IPPROTO_TCP, TCP_DEFER_ACCEPT, + &kTCPDeferAccept, sizeof(kTCPDeferAccept)), + SyscallSucceeds()); + + // Connect to the listening socket. + FileDescriptor conn_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + sockaddr_storage conn_addr = connector.addr; + ASSERT_NO_ERRNO(SetAddrPort(connector.family(), &conn_addr, port)); + ASSERT_THAT(RetryEINTR(connect)(conn_fd.get(), + reinterpret_cast(&conn_addr), + connector.addr_len), + SyscallSucceeds()); + + // Set the listening socket to nonblock so that we can verify that there is no + // connection in queue despite the connect above succeeding since the peer has + // sent no data and TCP_DEFER_ACCEPT is set on the listening socket. Set the + // FD to O_NONBLOCK. + int opts; + ASSERT_THAT(opts = fcntl(listen_fd.get(), F_GETFL), SyscallSucceeds()); + opts |= O_NONBLOCK; + ASSERT_THAT(fcntl(listen_fd.get(), F_SETFL, opts), SyscallSucceeds()); + + ASSERT_THAT(accept(listen_fd.get(), nullptr, nullptr), + SyscallFailsWithErrno(EWOULDBLOCK)); + + // Set FD back to blocking. + opts &= ~O_NONBLOCK; + ASSERT_THAT(fcntl(listen_fd.get(), F_SETFL, opts), SyscallSucceeds()); + + // Now write some data to the socket. + int data = 0; + ASSERT_THAT(RetryEINTR(write)(conn_fd.get(), &data, sizeof(data)), + SyscallSucceedsWithValue(sizeof(data))); + + // This should now cause the connection to complete and be delivered to the + // accept socket. + + // Accept the connection. + auto accepted = + ASSERT_NO_ERRNO_AND_VALUE(Accept(listen_fd.get(), nullptr, nullptr)); + + // Verify that the accepted socket returns the data written. + int get = -1; + ASSERT_THAT(RetryEINTR(recv)(accepted.get(), &get, sizeof(get), 0), + SyscallSucceedsWithValue(sizeof(get))); + + EXPECT_EQ(get, data); +} + +// TODO(gvisor.dev/issue/1688): Partially completed passive endpoints are not +// saved. Enable S/R once issue is fixed. +TEST_P(SocketInetLoopbackTest, TCPDeferAcceptTimeout_NoRandomSave) { + // TODO(gvisor.dev/issue/1688): Partially completed passive endpoints are not + // saved. Enable S/R once issue is fixed. + DisableSave ds; + + auto const& param = GetParam(); + TestAddress const& listener = param.listener; + TestAddress const& connector = param.connector; + + // Create the listening socket. + const FileDescriptor listen_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(listener.family(), SOCK_STREAM, IPPROTO_TCP)); + sockaddr_storage listen_addr = listener.addr; + ASSERT_THAT(bind(listen_fd.get(), reinterpret_cast(&listen_addr), + listener.addr_len), + SyscallSucceeds()); + ASSERT_THAT(listen(listen_fd.get(), SOMAXCONN), SyscallSucceeds()); + + // Get the port bound by the listening socket. + socklen_t addrlen = listener.addr_len; + ASSERT_THAT(getsockname(listen_fd.get(), + reinterpret_cast(&listen_addr), &addrlen), + SyscallSucceeds()); + + const uint16_t port = + ASSERT_NO_ERRNO_AND_VALUE(AddrPort(listener.family(), listen_addr)); + + // Set the TCP_DEFER_ACCEPT on the listening socket. + constexpr int kTCPDeferAccept = 3; + ASSERT_THAT(setsockopt(listen_fd.get(), IPPROTO_TCP, TCP_DEFER_ACCEPT, + &kTCPDeferAccept, sizeof(kTCPDeferAccept)), + SyscallSucceeds()); + + // Connect to the listening socket. + FileDescriptor conn_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + sockaddr_storage conn_addr = connector.addr; + ASSERT_NO_ERRNO(SetAddrPort(connector.family(), &conn_addr, port)); + ASSERT_THAT(RetryEINTR(connect)(conn_fd.get(), + reinterpret_cast(&conn_addr), + connector.addr_len), + SyscallSucceeds()); + + // Set the listening socket to nonblock so that we can verify that there is no + // connection in queue despite the connect above succeeding since the peer has + // sent no data and TCP_DEFER_ACCEPT is set on the listening socket. Set the + // FD to O_NONBLOCK. + int opts; + ASSERT_THAT(opts = fcntl(listen_fd.get(), F_GETFL), SyscallSucceeds()); + opts |= O_NONBLOCK; + ASSERT_THAT(fcntl(listen_fd.get(), F_SETFL, opts), SyscallSucceeds()); + + // Verify that there is no acceptable connection before TCP_DEFER_ACCEPT + // timeout is hit. + absl::SleepFor(absl::Seconds(kTCPDeferAccept - 1)); + ASSERT_THAT(accept(listen_fd.get(), nullptr, nullptr), + SyscallFailsWithErrno(EWOULDBLOCK)); + + // Set FD back to blocking. + opts &= ~O_NONBLOCK; + ASSERT_THAT(fcntl(listen_fd.get(), F_SETFL, opts), SyscallSucceeds()); + + // Now sleep for a little over the TCP_DEFER_ACCEPT duration. When the timeout + // is hit a SYN-ACK should be retransmitted by the listener as a last ditch + // attempt to complete the connection with or without data. + absl::SleepFor(absl::Seconds(2)); + + // Verify that we have a connection that can be accepted even though no + // data was written. + auto accepted = + ASSERT_NO_ERRNO_AND_VALUE(Accept(listen_fd.get(), nullptr, nullptr)); +} + INSTANTIATE_TEST_SUITE_P( All, SocketInetLoopbackTest, ::testing::Values( diff --git a/test/syscalls/linux/tcp_socket.cc b/test/syscalls/linux/tcp_socket.cc index 33a5ac66c..525ccbd88 100644 --- a/test/syscalls/linux/tcp_socket.cc +++ b/test/syscalls/linux/tcp_socket.cc @@ -1286,6 +1286,59 @@ TEST_P(SimpleTcpSocketTest, SetTCPUserTimeout) { EXPECT_EQ(get, kTCPUserTimeout); } +TEST_P(SimpleTcpSocketTest, SetTCPDeferAcceptNeg) { + FileDescriptor s = + ASSERT_NO_ERRNO_AND_VALUE(Socket(GetParam(), SOCK_STREAM, IPPROTO_TCP)); + + // -ve TCP_DEFER_ACCEPT is same as setting it to zero. + constexpr int kNeg = -1; + EXPECT_THAT( + setsockopt(s.get(), IPPROTO_TCP, TCP_DEFER_ACCEPT, &kNeg, sizeof(kNeg)), + SyscallSucceeds()); + int get = -1; + socklen_t get_len = sizeof(get); + ASSERT_THAT( + getsockopt(s.get(), IPPROTO_TCP, TCP_USER_TIMEOUT, &get, &get_len), + SyscallSucceedsWithValue(0)); + EXPECT_EQ(get_len, sizeof(get)); + EXPECT_EQ(get, 0); +} + +TEST_P(SimpleTcpSocketTest, GetTCPDeferAcceptDefault) { + FileDescriptor s = + ASSERT_NO_ERRNO_AND_VALUE(Socket(GetParam(), SOCK_STREAM, IPPROTO_TCP)); + + int get = -1; + socklen_t get_len = sizeof(get); + ASSERT_THAT( + getsockopt(s.get(), IPPROTO_TCP, TCP_USER_TIMEOUT, &get, &get_len), + SyscallSucceedsWithValue(0)); + EXPECT_EQ(get_len, sizeof(get)); + EXPECT_EQ(get, 0); +} + +TEST_P(SimpleTcpSocketTest, SetTCPDeferAcceptGreaterThanZero) { + FileDescriptor s = + ASSERT_NO_ERRNO_AND_VALUE(Socket(GetParam(), SOCK_STREAM, IPPROTO_TCP)); + // kTCPDeferAccept is in seconds. + // NOTE: linux translates seconds to # of retries and back from + // #of retries to seconds. Which means only certain values + // translate back exactly. That's why we use 3 here, a value of + // 5 will result in us getting back 7 instead of 5 in the + // getsockopt. + constexpr int kTCPDeferAccept = 3; + ASSERT_THAT(setsockopt(s.get(), IPPROTO_TCP, TCP_DEFER_ACCEPT, + &kTCPDeferAccept, sizeof(kTCPDeferAccept)), + SyscallSucceeds()); + int get = -1; + socklen_t get_len = sizeof(get); + ASSERT_THAT( + getsockopt(s.get(), IPPROTO_TCP, TCP_DEFER_ACCEPT, &get, &get_len), + SyscallSucceeds()); + EXPECT_EQ(get_len, sizeof(get)); + EXPECT_EQ(get, kTCPDeferAccept); +} + INSTANTIATE_TEST_SUITE_P(AllInetTests, SimpleTcpSocketTest, ::testing::Values(AF_INET, AF_INET6)); -- cgit v1.2.3 From eba7bdc24d31388ca81eeab251ed2db108f785dc Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Fri, 31 Jan 2020 13:46:13 -0800 Subject: iptables: enable TCP matching with "-m tcp". A couple other things that changed: - There's a proper extension registration system for matchers. Anyone adding another matcher can use tcp_matcher.go or udp_matcher.go as a template. - All logging and use of syserr.Error in the netfilter package happens at the highest possible level (public functions). Lower-level functions just return normal, descriptive golang errors. --- pkg/abi/linux/netfilter.go | 52 ++++++++ pkg/sentry/socket/netfilter/BUILD | 4 + pkg/sentry/socket/netfilter/extensions.go | 98 +++++++++++++++ pkg/sentry/socket/netfilter/netfilter.go | 187 ++++++++--------------------- pkg/sentry/socket/netfilter/tcp_matcher.go | 143 ++++++++++++++++++++++ pkg/sentry/socket/netfilter/udp_matcher.go | 142 ++++++++++++++++++++++ pkg/tcpip/iptables/BUILD | 1 - pkg/tcpip/iptables/types.go | 3 + pkg/tcpip/iptables/udp_matcher.go | 113 ----------------- 9 files changed, 495 insertions(+), 248 deletions(-) create mode 100644 pkg/sentry/socket/netfilter/extensions.go create mode 100644 pkg/sentry/socket/netfilter/tcp_matcher.go create mode 100644 pkg/sentry/socket/netfilter/udp_matcher.go delete mode 100644 pkg/tcpip/iptables/udp_matcher.go (limited to 'pkg/sentry/socket') diff --git a/pkg/abi/linux/netfilter.go b/pkg/abi/linux/netfilter.go index 8e40bcc62..e4aabb6bb 100644 --- a/pkg/abi/linux/netfilter.go +++ b/pkg/abi/linux/netfilter.go @@ -348,6 +348,58 @@ func goString(cstring []byte) string { return string(cstring) } +// XTTCP holds data for matching TCP packets. It corresponds to struct xt_tcp +// in include/uapi/linux/netfilter/xt_tcpudp.h. +type XTTCP struct { + // SourcePortStart specifies the inclusive start of the range of source + // ports to which the matcher applies. + SourcePortStart uint16 + + // SourcePortEnd specifies the inclusive end of the range of source ports + // to which the matcher applies. + SourcePortEnd uint16 + + // DestinationPortStart specifies the start of the destination port + // range to which the matcher applies. + DestinationPortStart uint16 + + // DestinationPortEnd specifies the start of the destination port + // range to which the matcher applies. + DestinationPortEnd uint16 + + // Option specifies that a particular TCP option must be set. + Option uint8 + + // FlagMask masks the FlagCompare byte when comparing to the TCP flag + // fields. + FlagMask uint8 + + // FlagCompare is binary and-ed with the TCP flag fields. + FlagCompare uint8 + + // InverseFlags flips the meaning of certain fields. See the + // TX_TCP_INV_* flags. + InverseFlags uint8 +} + +// SizeOfXTTCP is the size of an XTTCP. +const SizeOfXTTCP = 12 + +// Flags in XTTCP.InverseFlags. Corresponding constants are in +// include/uapi/linux/netfilter/xt_tcpudp.h. +const ( + // Invert the meaning of SourcePortStart/End. + XT_TCP_INV_SRCPT = 0x01 + // Invert the meaning of DestinationPortStart/End. + XT_TCP_INV_DSTPT = 0x02 + // Invert the meaning of FlagCompare. + XT_TCP_INV_FLAGS = 0x04 + // Invert the meaning of Option. + XT_TCP_INV_OPTION = 0x08 + // Enable all flags. + XT_TCP_INV_MASK = 0x0F +) + // XTUDP holds data for matching UDP packets. It corresponds to struct xt_udp // in include/uapi/linux/netfilter/xt_tcpudp.h. type XTUDP struct { diff --git a/pkg/sentry/socket/netfilter/BUILD b/pkg/sentry/socket/netfilter/BUILD index fa2a2cb66..c91ec7494 100644 --- a/pkg/sentry/socket/netfilter/BUILD +++ b/pkg/sentry/socket/netfilter/BUILD @@ -5,7 +5,10 @@ package(licenses = ["notice"]) go_library( name = "netfilter", srcs = [ + "extensions.go", "netfilter.go", + "tcp_matcher.go", + "udp_matcher.go", ], # This target depends on netstack and should only be used by epsocket, # which is allowed to depend on netstack. @@ -17,6 +20,7 @@ go_library( "//pkg/sentry/kernel", "//pkg/syserr", "//pkg/tcpip", + "//pkg/tcpip/header", "//pkg/tcpip/iptables", "//pkg/tcpip/stack", "//pkg/usermem", diff --git a/pkg/sentry/socket/netfilter/extensions.go b/pkg/sentry/socket/netfilter/extensions.go new file mode 100644 index 000000000..5a4cac84c --- /dev/null +++ b/pkg/sentry/socket/netfilter/extensions.go @@ -0,0 +1,98 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netfilter + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/usermem" +) + +// TODO(gvisor.dev/issue/170): The following per-matcher params should be +// supported: +// - Table name +// - Match size +// - User size +// - Hooks +// - Proto +// - Family + +// matchMarshaler knows how to (un)marshal the matcher named name(). +type matchMarshaler interface { + // name is the matcher name as stored in the xt_entry_match struct. + name() string + + // marshal converts from an iptables.Matcher to an ABI struct. + marshal(matcher iptables.Matcher) []byte + + // unmarshal converts from the ABI matcher struct to an + // iptables.Matcher. + unmarshal(buf []byte, filter iptables.IPHeaderFilter) (iptables.Matcher, error) +} + +var matchMarshalers = map[string]matchMarshaler{} + +// registerMatchMarshaler should be called by match extensions to register them +// with the netfilter package. +func registerMatchMarshaler(mm matchMarshaler) { + if _, ok := matchMarshalers[mm.name()]; ok { + panic(fmt.Sprintf("Multiple matches registered with name %q.", mm.name())) + } + matchMarshalers[mm.name()] = mm +} + +func marshalMatcher(matcher iptables.Matcher) []byte { + matchMaker, ok := matchMarshalers[matcher.Name()] + if !ok { + panic(fmt.Errorf("Unknown matcher of type %T.", matcher)) + } + return matchMaker.marshal(matcher) +} + +// marshalEntryMatch creates a marshalled XTEntryMatch with the given name and +// data appended at the end. +func marshalEntryMatch(name string, data []byte) []byte { + nflog("marshaling matcher %q", name) + + // We have to pad this struct size to a multiple of 8 bytes. + size := alignUp(linux.SizeOfXTEntryMatch+len(data), 8) + matcher := linux.KernelXTEntryMatch{ + XTEntryMatch: linux.XTEntryMatch{ + MatchSize: uint16(size), + }, + Data: data, + } + copy(matcher.Name[:], name) + + buf := make([]byte, 0, size) + buf = binary.Marshal(buf, usermem.ByteOrder, matcher) + return append(buf, make([]byte, size-len(buf))...) +} + +func unmarshalMatcher(match linux.XTEntryMatch, filter iptables.IPHeaderFilter, buf []byte) (iptables.Matcher, error) { + matchMaker, ok := matchMarshalers[match.Name.String()] + if !ok { + return nil, fmt.Errorf("unsupported matcher with name %q", match.Name.String()) + } + return matchMaker.unmarshal(buf, filter) +} + +// alignUp rounds a length up to an alignment. align must be a power of 2. +func alignUp(length int, align uint) int { + return (length + int(align) - 1) & ^(int(align) - 1) +} diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index 3dda6c7a1..8f14643b0 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -17,6 +17,7 @@ package netfilter import ( + "errors" "fmt" "gvisor.dev/gvisor/pkg/abi/linux" @@ -34,10 +35,6 @@ import ( // shouldn't be reached - an error has occurred if we fall through to one. const errorTargetName = "ERROR" -const ( - matcherNameUDP = "udp" -) - // Metadata is used to verify that we are correctly serializing and // deserializing iptables into structs consumable by the iptables tool. We save // a metadata struct when the tables are written, and when they are read out we @@ -68,7 +65,8 @@ func GetInfo(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr) (linux.IPT // Find the appropriate table. table, err := findTable(stack, info.Name) if err != nil { - return linux.IPTGetinfo{}, err + nflog("%v", err) + return linux.IPTGetinfo{}, syserr.ErrInvalidArgument } // Get the hooks that apply to this table. @@ -95,39 +93,40 @@ func GetEntries(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr, outLen // Read in the struct and table name. var userEntries linux.IPTGetEntries if _, err := t.CopyIn(outPtr, &userEntries); err != nil { - log.Warningf("netfilter: couldn't copy in entries %q", userEntries.Name) + nflog("couldn't copy in entries %q", userEntries.Name) return linux.KernelIPTGetEntries{}, syserr.FromError(err) } // Find the appropriate table. table, err := findTable(stack, userEntries.Name) if err != nil { - log.Warningf("netfilter: couldn't find table %q", userEntries.Name) - return linux.KernelIPTGetEntries{}, err + nflog("%v", err) + return linux.KernelIPTGetEntries{}, syserr.ErrInvalidArgument } // Convert netstack's iptables rules to something that the iptables // tool can understand. entries, meta, err := convertNetstackToBinary(userEntries.Name.String(), table) if err != nil { - return linux.KernelIPTGetEntries{}, err + nflog("couldn't read entries: %v", err) + return linux.KernelIPTGetEntries{}, syserr.ErrInvalidArgument } if meta != table.Metadata().(metadata) { panic(fmt.Sprintf("Table %q metadata changed between writing and reading. Was saved as %+v, but is now %+v", userEntries.Name.String(), table.Metadata().(metadata), meta)) } if binary.Size(entries) > uintptr(outLen) { - log.Warningf("Insufficient GetEntries output size: %d", uintptr(outLen)) + nflog("insufficient GetEntries output size: %d", uintptr(outLen)) return linux.KernelIPTGetEntries{}, syserr.ErrInvalidArgument } return entries, nil } -func findTable(stack *stack.Stack, tablename linux.TableName) (iptables.Table, *syserr.Error) { +func findTable(stack *stack.Stack, tablename linux.TableName) (iptables.Table, error) { ipt := stack.IPTables() table, ok := ipt.Tables[tablename.String()] if !ok { - return iptables.Table{}, syserr.ErrInvalidArgument + return iptables.Table{}, fmt.Errorf("couldn't find table %q", tablename) } return table, nil } @@ -151,19 +150,19 @@ func FillDefaultIPTables(stack *stack.Stack) { stack.SetIPTables(ipt) } +// TODO: Return proto. // convertNetstackToBinary converts the iptables as stored in netstack to the // format expected by the iptables tool. Linux stores each table as a binary // blob that can only be traversed by parsing a bit, reading some offsets, // jumping to those offsets, parsing again, etc. -func convertNetstackToBinary(tablename string, table iptables.Table) (linux.KernelIPTGetEntries, metadata, *syserr.Error) { +func convertNetstackToBinary(tablename string, table iptables.Table) (linux.KernelIPTGetEntries, metadata, error) { // Return values. var entries linux.KernelIPTGetEntries var meta metadata // The table name has to fit in the struct. if linux.XT_TABLE_MAXNAMELEN < len(tablename) { - log.Warningf("Table name %q too long.", tablename) - return linux.KernelIPTGetEntries{}, metadata{}, syserr.ErrInvalidArgument + return linux.KernelIPTGetEntries{}, metadata{}, fmt.Errorf("Table name %q too long.", tablename) } copy(entries.Name[:], tablename) @@ -229,46 +228,6 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern return entries, meta, nil } -func marshalMatcher(matcher iptables.Matcher) []byte { - switch m := matcher.(type) { - case *iptables.UDPMatcher: - return marshalUDPMatcher(m) - default: - // TODO(gvisor.dev/issue/170): Support other matchers. - panic(fmt.Errorf("unknown matcher of type %T", matcher)) - } -} - -func marshalUDPMatcher(matcher *iptables.UDPMatcher) []byte { - nflog("convert to binary: marshalling UDP matcher: %+v", matcher) - - // We have to pad this struct size to a multiple of 8 bytes. - size := alignUp(linux.SizeOfXTEntryMatch+linux.SizeOfXTUDP, 8) - - linuxMatcher := linux.KernelXTEntryMatch{ - XTEntryMatch: linux.XTEntryMatch{ - MatchSize: uint16(size), - }, - Data: make([]byte, 0, linux.SizeOfXTUDP), - } - copy(linuxMatcher.Name[:], matcherNameUDP) - - xtudp := linux.XTUDP{ - SourcePortStart: matcher.Data.SourcePortStart, - SourcePortEnd: matcher.Data.SourcePortEnd, - DestinationPortStart: matcher.Data.DestinationPortStart, - DestinationPortEnd: matcher.Data.DestinationPortEnd, - InverseFlags: matcher.Data.InverseFlags, - } - linuxMatcher.Data = binary.Marshal(linuxMatcher.Data, usermem.ByteOrder, xtudp) - - buf := make([]byte, 0, size) - buf = binary.Marshal(buf, usermem.ByteOrder, linuxMatcher) - buf = append(buf, make([]byte, size-len(buf))...) - nflog("convert to binary: marshalled UDP matcher into %v", buf) - return buf[:] -} - func marshalTarget(target iptables.Target) []byte { switch target.(type) { case iptables.UnconditionalAcceptTarget: @@ -332,7 +291,7 @@ func translateFromStandardVerdict(verdict iptables.Verdict) int32 { // translateToStandardVerdict translates from the value in a // linux.XTStandardTarget to an iptables.Verdict. -func translateToStandardVerdict(val int32) (iptables.Verdict, *syserr.Error) { +func translateToStandardVerdict(val int32) (iptables.Verdict, error) { // TODO(gvisor.dev/issue/170): Support other verdicts. switch val { case -linux.NF_ACCEPT - 1: @@ -340,13 +299,12 @@ func translateToStandardVerdict(val int32) (iptables.Verdict, *syserr.Error) { case -linux.NF_DROP - 1: return iptables.Drop, nil case -linux.NF_QUEUE - 1: - log.Warningf("Unsupported iptables verdict QUEUE.") + return iptables.Invalid, errors.New("unsupported iptables verdict QUEUE") case linux.NF_RETURN: - log.Warningf("Unsupported iptables verdict RETURN.") + return iptables.Invalid, errors.New("unsupported iptables verdict RETURN") default: - log.Warningf("Unknown iptables verdict %d.", val) + return iptables.Invalid, fmt.Errorf("unknown iptables verdict %d.", val) } - return iptables.Invalid, syserr.ErrInvalidArgument } // SetEntries sets iptables rules for a single table. See @@ -354,7 +312,7 @@ func translateToStandardVerdict(val int32) (iptables.Verdict, *syserr.Error) { func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // Get the basic rules data (struct ipt_replace). if len(optVal) < linux.SizeOfIPTReplace { - log.Warningf("netfilter.SetEntries: optVal has insufficient size for replace %d", len(optVal)) + nflog("optVal has insufficient size for replace %d", len(optVal)) return syserr.ErrInvalidArgument } var replace linux.IPTReplace @@ -368,7 +326,7 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { case iptables.TablenameFilter: table = iptables.EmptyFilterTable() default: - log.Warningf("We don't yet support writing to the %q table (gvisor.dev/issue/170)", replace.Name.String()) + nflog("we don't yet support writing to the %q table (gvisor.dev/issue/170)", replace.Name.String()) return syserr.ErrInvalidArgument } @@ -382,7 +340,7 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // Get the struct ipt_entry. if len(optVal) < linux.SizeOfIPTEntry { - log.Warningf("netfilter: optVal has insufficient size for entry %d", len(optVal)) + nflog("optVal has insufficient size for entry %d", len(optVal)) return syserr.ErrInvalidArgument } var entry linux.IPTEntry @@ -392,7 +350,7 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { optVal = optVal[linux.SizeOfIPTEntry:] if entry.TargetOffset < linux.SizeOfIPTEntry { - log.Warningf("netfilter: entry has too-small target offset %d", entry.TargetOffset) + nflog("entry has too-small target offset %d", entry.TargetOffset) return syserr.ErrInvalidArgument } @@ -400,7 +358,8 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // filtering fields. filter, err := filterFromIPTIP(entry.IP) if err != nil { - return err + nflog("bad iptip: %v", err) + return syserr.ErrInvalidArgument } // TODO(gvisor.dev/issue/170): Matchers and targets can specify @@ -408,25 +367,26 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // Get matchers. matchersSize := entry.TargetOffset - linux.SizeOfIPTEntry if len(optVal) < int(matchersSize) { - log.Warningf("netfilter: entry doesn't have enough room for its matchers (only %d bytes remain)", len(optVal)) + nflog("entry doesn't have enough room for its matchers (only %d bytes remain)", len(optVal)) return syserr.ErrInvalidArgument } matchers, err := parseMatchers(filter, optVal[:matchersSize]) if err != nil { - log.Warningf("netfilter: failed to parse matchers: %v", err) - return err + nflog("failed to parse matchers: %v", err) + return syserr.ErrInvalidArgument } optVal = optVal[matchersSize:] // Get the target of the rule. targetSize := entry.NextOffset - entry.TargetOffset if len(optVal) < int(targetSize) { - log.Warningf("netfilter: entry doesn't have enough room for its target (only %d bytes remain)", len(optVal)) + nflog("entry doesn't have enough room for its target (only %d bytes remain)", len(optVal)) return syserr.ErrInvalidArgument } target, err := parseTarget(optVal[:targetSize]) if err != nil { - return err + nflog("failed to parse target: %v", err) + return syserr.ErrInvalidArgument } optVal = optVal[targetSize:] @@ -439,7 +399,7 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { offset += uint32(entry.NextOffset) if initialOptValLen-len(optVal) != int(entry.NextOffset) { - log.Warningf("netfilter: entry NextOffset is %d, but entry took up %d bytes", entry.NextOffset, initialOptValLen-len(optVal)) + nflog("entry NextOffset is %d, but entry took up %d bytes", entry.NextOffset, initialOptValLen-len(optVal)) } } @@ -457,11 +417,11 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { } } if ruleIdx := table.BuiltinChains[hk]; ruleIdx == iptables.HookUnset { - log.Warningf("Hook %v is unset.", hk) + nflog("hook %v is unset.", hk) return syserr.ErrInvalidArgument } if ruleIdx := table.Underflows[hk]; ruleIdx == iptables.HookUnset { - log.Warningf("Underflow %v is unset.", hk) + nflog("underflow %v is unset.", hk) return syserr.ErrInvalidArgument } } @@ -473,7 +433,7 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { for hook, ruleIdx := range table.BuiltinChains { if hook != iptables.Input { if _, ok := table.Rules[ruleIdx].Target.(iptables.UnconditionalAcceptTarget); !ok { - log.Warningf("Hook %d is unsupported.", hook) + nflog("hook %d is unsupported.", hook) return syserr.ErrInvalidArgument } } @@ -499,7 +459,7 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // parseMatchers parses 0 or more matchers from optVal. optVal should contain // only the matchers. -func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Matcher, *syserr.Error) { +func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Matcher, error) { nflog("set entries: parsing matchers of size %d", len(optVal)) var matchers []iptables.Matcher for len(optVal) > 0 { @@ -507,8 +467,7 @@ func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Ma // Get the XTEntryMatch. if len(optVal) < linux.SizeOfXTEntryMatch { - log.Warningf("netfilter: optVal has insufficient size for entry match: %d", len(optVal)) - return nil, syserr.ErrInvalidArgument + return nil, fmt.Errorf("optVal has insufficient size for entry match: %d", len(optVal)) } var match linux.XTEntryMatch buf := optVal[:linux.SizeOfXTEntryMatch] @@ -517,45 +476,18 @@ func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Ma // Check some invariants. if match.MatchSize < linux.SizeOfXTEntryMatch { - log.Warningf("netfilter: match size is too small, must be at least %d", linux.SizeOfXTEntryMatch) - return nil, syserr.ErrInvalidArgument + + return nil, fmt.Errorf("match size is too small, must be at least %d", linux.SizeOfXTEntryMatch) } if len(optVal) < int(match.MatchSize) { - log.Warningf("netfilter: optVal has insufficient size for match: %d", len(optVal)) - return nil, syserr.ErrInvalidArgument + return nil, fmt.Errorf("optVal has insufficient size for match: %d", len(optVal)) } - buf = optVal[linux.SizeOfXTEntryMatch:match.MatchSize] - var matcher iptables.Matcher - var err error - switch match.Name.String() { - case matcherNameUDP: - if len(buf) < linux.SizeOfXTUDP { - log.Warningf("netfilter: optVal has insufficient size for UDP match: %d", len(optVal)) - return nil, syserr.ErrInvalidArgument - } - // For alignment reasons, the match's total size may - // exceed what's strictly necessary to hold matchData. - var matchData linux.XTUDP - binary.Unmarshal(buf[:linux.SizeOfXTUDP], usermem.ByteOrder, &matchData) - log.Infof("parseMatchers: parsed XTUDP: %+v", matchData) - matcher, err = iptables.NewUDPMatcher(filter, iptables.UDPMatcherParams{ - SourcePortStart: matchData.SourcePortStart, - SourcePortEnd: matchData.SourcePortEnd, - DestinationPortStart: matchData.DestinationPortStart, - DestinationPortEnd: matchData.DestinationPortEnd, - InverseFlags: matchData.InverseFlags, - }) - if err != nil { - log.Warningf("netfilter: failed to create UDP matcher: %v", err) - return nil, syserr.ErrInvalidArgument - } - - default: - log.Warningf("netfilter: unsupported matcher with name %q", match.Name.String()) - return nil, syserr.ErrInvalidArgument + // Parse the specific matcher. + matcher, err := unmarshalMatcher(match, filter, optVal[linux.SizeOfXTEntryMatch:match.MatchSize]) + if err != nil { + return nil, fmt.Errorf("failed to create matcher: %v", err) } - matchers = append(matchers, matcher) // TODO(gvisor.dev/issue/170): Check the revision field. @@ -563,8 +495,7 @@ func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Ma } if len(optVal) != 0 { - log.Warningf("netfilter: optVal should be exhausted after parsing matchers") - return nil, syserr.ErrInvalidArgument + return nil, errors.New("optVal should be exhausted after parsing matchers") } return matchers, nil @@ -572,11 +503,10 @@ func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Ma // parseTarget parses a target from optVal. optVal should contain only the // target. -func parseTarget(optVal []byte) (iptables.Target, *syserr.Error) { +func parseTarget(optVal []byte) (iptables.Target, error) { nflog("set entries: parsing target of size %d", len(optVal)) if len(optVal) < linux.SizeOfXTEntryTarget { - log.Warningf("netfilter: optVal has insufficient size for entry target %d", len(optVal)) - return nil, syserr.ErrInvalidArgument + return nil, fmt.Errorf("optVal has insufficient size for entry target %d", len(optVal)) } var target linux.XTEntryTarget buf := optVal[:linux.SizeOfXTEntryTarget] @@ -585,8 +515,7 @@ func parseTarget(optVal []byte) (iptables.Target, *syserr.Error) { case "": // Standard target. if len(optVal) != linux.SizeOfXTStandardTarget { - log.Warningf("netfilter.SetEntries: optVal has wrong size for standard target %d", len(optVal)) - return nil, syserr.ErrInvalidArgument + return nil, fmt.Errorf("optVal has wrong size for standard target %d", len(optVal)) } var standardTarget linux.XTStandardTarget buf = optVal[:linux.SizeOfXTStandardTarget] @@ -602,15 +531,13 @@ func parseTarget(optVal []byte) (iptables.Target, *syserr.Error) { case iptables.Drop: return iptables.UnconditionalDropTarget{}, nil default: - log.Warningf("Unknown verdict: %v", verdict) - return nil, syserr.ErrInvalidArgument + return nil, fmt.Errorf("Unknown verdict: %v", verdict) } case errorTargetName: // Error target. if len(optVal) != linux.SizeOfXTErrorTarget { - log.Infof("netfilter.SetEntries: optVal has insufficient size for error target %d", len(optVal)) - return nil, syserr.ErrInvalidArgument + return nil, fmt.Errorf("optVal has insufficient size for error target %d", len(optVal)) } var errorTarget linux.XTErrorTarget buf = optVal[:linux.SizeOfXTErrorTarget] @@ -627,20 +554,17 @@ func parseTarget(optVal []byte) (iptables.Target, *syserr.Error) { case errorTargetName: return iptables.ErrorTarget{}, nil default: - log.Infof("Unknown error target %q doesn't exist or isn't supported yet.", errorTarget.Name.String()) - return nil, syserr.ErrInvalidArgument + return nil, fmt.Errorf("Unknown error target %q doesn't exist or isn't supported yet.", errorTarget.Name.String()) } } // Unknown target. - log.Infof("Unknown target %q doesn't exist or isn't supported yet.", target.Name.String()) - return nil, syserr.ErrInvalidArgument + return nil, fmt.Errorf("Unknown target %q doesn't exist or isn't supported yet.", target.Name.String()) } -func filterFromIPTIP(iptip linux.IPTIP) (iptables.IPHeaderFilter, *syserr.Error) { +func filterFromIPTIP(iptip linux.IPTIP) (iptables.IPHeaderFilter, error) { if containsUnsupportedFields(iptip) { - log.Warningf("netfilter: unsupported fields in struct iptip: %+v", iptip) - return iptables.IPHeaderFilter{}, syserr.ErrInvalidArgument + return iptables.IPHeaderFilter{}, fmt.Errorf("unsupported fields in struct iptip: %+v", iptip) } return iptables.IPHeaderFilter{ Protocol: tcpip.TransportProtocolNumber(iptip.Protocol), @@ -678,8 +602,3 @@ func hookFromLinux(hook int) iptables.Hook { } panic(fmt.Sprintf("Unknown hook %d does not correspond to a builtin chain", hook)) } - -// alignUp rounds a length up to an alignment. align must be a power of 2. -func alignUp(length int, align uint) int { - return (length + int(align) - 1) & ^(int(align) - 1) -} diff --git a/pkg/sentry/socket/netfilter/tcp_matcher.go b/pkg/sentry/socket/netfilter/tcp_matcher.go new file mode 100644 index 000000000..1646d22f7 --- /dev/null +++ b/pkg/sentry/socket/netfilter/tcp_matcher.go @@ -0,0 +1,143 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netfilter + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/header" + "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/usermem" +) + +const matcherNameTCP = "tcp" + +func init() { + registerMatchMarshaler(tcpMarshaler{}) +} + +// tcpMarshaler implements matchMarshaler for TCP matching. +type tcpMarshaler struct{} + +// name implements matchMarshaler.name. +func (tcpMarshaler) name() string { + return matcherNameTCP +} + +// marshal implements matchMarshaler.marshal. +func (tcpMarshaler) marshal(mr iptables.Matcher) []byte { + matcher := mr.(*TCPMatcher) + xttcp := linux.XTTCP{ + SourcePortStart: matcher.sourcePortStart, + SourcePortEnd: matcher.sourcePortEnd, + DestinationPortStart: matcher.destinationPortStart, + DestinationPortEnd: matcher.destinationPortEnd, + } + buf := make([]byte, 0, linux.SizeOfXTUDP) + return marshalEntryMatch(matcherNameTCP, binary.Marshal(buf, usermem.ByteOrder, xttcp)) +} + +// unmarshal implements matchMarshaler.unmarshal. +func (tcpMarshaler) unmarshal(buf []byte, filter iptables.IPHeaderFilter) (iptables.Matcher, error) { + if len(buf) < linux.SizeOfXTTCP { + return nil, fmt.Errorf("buf has insufficient size for TCP match: %d", len(buf)) + } + + // For alignment reasons, the match's total size may + // exceed what's strictly necessary to hold matchData. + var matchData linux.XTTCP + binary.Unmarshal(buf[:linux.SizeOfXTTCP], usermem.ByteOrder, &matchData) + nflog("parseMatchers: parsed XTTCP: %+v", matchData) + + if matchData.Option != 0 || + matchData.FlagMask != 0 || + matchData.FlagCompare != 0 || + matchData.InverseFlags != 0 { + return nil, fmt.Errorf("unsupported TCP matcher flags set") + } + + if filter.Protocol != header.TCPProtocolNumber { + return nil, fmt.Errorf("TCP matching is only valid for protocol %d.", header.TCPProtocolNumber) + } + + return &TCPMatcher{ + sourcePortStart: matchData.SourcePortStart, + sourcePortEnd: matchData.SourcePortEnd, + destinationPortStart: matchData.DestinationPortStart, + destinationPortEnd: matchData.DestinationPortEnd, + }, nil +} + +// TCPMatcher matches TCP packets and their headers. It implements Matcher. +type TCPMatcher struct { + sourcePortStart uint16 + sourcePortEnd uint16 + destinationPortStart uint16 + destinationPortEnd uint16 +} + +// Name implements Matcher.Name. +func (*TCPMatcher) Name() string { + return matcherNameTCP +} + +// Match implements Matcher.Match. +func (tm *TCPMatcher) Match(hook iptables.Hook, pkt tcpip.PacketBuffer, interfaceName string) (bool, bool) { + netHeader := header.IPv4(pkt.NetworkHeader) + + if netHeader.TransportProtocol() != header.TCPProtocolNumber { + return false, false + } + + // We dont't match fragments. + if frag := netHeader.FragmentOffset(); frag != 0 { + if frag == 1 { + return false, true + } + return false, false + } + + // Now we need the transport header. However, this may not have been set + // yet. + // TODO(gvisor.dev/issue/170): Parsing the transport header should + // ultimately be moved into the iptables.Check codepath as matchers are + // added. + var tcpHeader header.TCP + if pkt.TransportHeader != nil { + tcpHeader = header.TCP(pkt.TransportHeader) + } else { + // The TCP header hasn't been parsed yet. We have to do it here. + if len(pkt.Data.First()) < header.TCPMinimumSize { + // There's no valid TCP header here, so we hotdrop the + // packet. + return false, true + } + tcpHeader = header.TCP(pkt.Data.First()) + } + + // Check whether the source and destination ports are within the + // matching range. + if sourcePort := tcpHeader.SourcePort(); sourcePort < tm.sourcePortStart || tm.sourcePortEnd < sourcePort { + return false, false + } + if destinationPort := tcpHeader.DestinationPort(); destinationPort < tm.destinationPortStart || tm.destinationPortEnd < destinationPort { + return false, false + } + + return true, false +} diff --git a/pkg/sentry/socket/netfilter/udp_matcher.go b/pkg/sentry/socket/netfilter/udp_matcher.go new file mode 100644 index 000000000..b6e95bbc5 --- /dev/null +++ b/pkg/sentry/socket/netfilter/udp_matcher.go @@ -0,0 +1,142 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netfilter + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/header" + "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/usermem" +) + +const matcherNameUDP = "udp" + +func init() { + registerMatchMarshaler(udpMarshaler{}) +} + +// udpMarshaler implements matchMarshaler for UDP matching. +type udpMarshaler struct{} + +// name implements matchMarshaler.name. +func (udpMarshaler) name() string { + return matcherNameUDP +} + +// marshal implements matchMarshaler.marshal. +func (udpMarshaler) marshal(mr iptables.Matcher) []byte { + matcher := mr.(*UDPMatcher) + xtudp := linux.XTUDP{ + SourcePortStart: matcher.sourcePortStart, + SourcePortEnd: matcher.sourcePortEnd, + DestinationPortStart: matcher.destinationPortStart, + DestinationPortEnd: matcher.destinationPortEnd, + } + buf := make([]byte, 0, linux.SizeOfXTUDP) + return marshalEntryMatch(matcherNameUDP, binary.Marshal(buf, usermem.ByteOrder, xtudp)) +} + +// unmarshal implements matchMarshaler.unmarshal. +func (udpMarshaler) unmarshal(buf []byte, filter iptables.IPHeaderFilter) (iptables.Matcher, error) { + if len(buf) < linux.SizeOfXTUDP { + return nil, fmt.Errorf("buf has insufficient size for UDP match: %d", len(buf)) + } + + // For alignment reasons, the match's total size may exceed what's + // strictly necessary to hold matchData. + var matchData linux.XTUDP + binary.Unmarshal(buf[:linux.SizeOfXTUDP], usermem.ByteOrder, &matchData) + nflog("parseMatchers: parsed XTUDP: %+v", matchData) + + if matchData.InverseFlags != 0 { + return nil, fmt.Errorf("unsupported UDP matcher inverse flags set") + } + + if filter.Protocol != header.UDPProtocolNumber { + return nil, fmt.Errorf("UDP matching is only valid for protocol %d.", header.UDPProtocolNumber) + } + + return &UDPMatcher{ + sourcePortStart: matchData.SourcePortStart, + sourcePortEnd: matchData.SourcePortEnd, + destinationPortStart: matchData.DestinationPortStart, + destinationPortEnd: matchData.DestinationPortEnd, + }, nil +} + +// UDPMatcher matches UDP packets and their headers. It implements Matcher. +type UDPMatcher struct { + sourcePortStart uint16 + sourcePortEnd uint16 + destinationPortStart uint16 + destinationPortEnd uint16 +} + +// Name implements Matcher.Name. +func (*UDPMatcher) Name() string { + return matcherNameUDP +} + +// Match implements Matcher.Match. +func (um *UDPMatcher) Match(hook iptables.Hook, pkt tcpip.PacketBuffer, interfaceName string) (bool, bool) { + netHeader := header.IPv4(pkt.NetworkHeader) + + // TODO(gvisor.dev/issue/170): Proto checks should ultimately be moved + // into the iptables.Check codepath as matchers are added. + if netHeader.TransportProtocol() != header.UDPProtocolNumber { + return false, false + } + + // We dont't match fragments. + if frag := netHeader.FragmentOffset(); frag != 0 { + if frag == 1 { + return false, true + } + return false, false + } + + // Now we need the transport header. However, this may not have been set + // yet. + // TODO(gvisor.dev/issue/170): Parsing the transport header should + // ultimately be moved into the iptables.Check codepath as matchers are + // added. + var udpHeader header.UDP + if pkt.TransportHeader != nil { + udpHeader = header.UDP(pkt.TransportHeader) + } else { + // The UDP header hasn't been parsed yet. We have to do it here. + if len(pkt.Data.First()) < header.UDPMinimumSize { + // There's no valid UDP header here, so we hotdrop the + // packet. + return false, true + } + udpHeader = header.UDP(pkt.Data.First()) + } + + // Check whether the source and destination ports are within the + // matching range. + if sourcePort := udpHeader.SourcePort(); sourcePort < um.sourcePortStart || um.sourcePortEnd < sourcePort { + return false, false + } + if destinationPort := udpHeader.DestinationPort(); destinationPort < um.destinationPortStart || um.destinationPortEnd < destinationPort { + return false, false + } + + return true, false +} diff --git a/pkg/tcpip/iptables/BUILD b/pkg/tcpip/iptables/BUILD index bab26580b..d1b73cfdf 100644 --- a/pkg/tcpip/iptables/BUILD +++ b/pkg/tcpip/iptables/BUILD @@ -8,7 +8,6 @@ go_library( "iptables.go", "targets.go", "types.go", - "udp_matcher.go", ], visibility = ["//visibility:public"], deps = [ diff --git a/pkg/tcpip/iptables/types.go b/pkg/tcpip/iptables/types.go index 7f77802a0..d660aab04 100644 --- a/pkg/tcpip/iptables/types.go +++ b/pkg/tcpip/iptables/types.go @@ -171,6 +171,9 @@ type IPHeaderFilter struct { // A Matcher is the interface for matching packets. type Matcher interface { + // Name returns the name of the Matcher. + Name() string + // Match returns whether the packet matches and whether the packet // should be "hotdropped", i.e. dropped immediately. This is usually // used for suspicious packets. diff --git a/pkg/tcpip/iptables/udp_matcher.go b/pkg/tcpip/iptables/udp_matcher.go deleted file mode 100644 index 3bb076f9c..000000000 --- a/pkg/tcpip/iptables/udp_matcher.go +++ /dev/null @@ -1,113 +0,0 @@ -// Copyright 2020 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package iptables - -import ( - "fmt" - - "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/tcpip" - "gvisor.dev/gvisor/pkg/tcpip/header" -) - -// TODO(gvisor.dev/issue/170): The following per-matcher params should be -// supported: -// - Table name -// - Match size -// - User size -// - Hooks -// - Proto -// - Family - -// UDPMatcher matches UDP packets and their headers. It implements Matcher. -type UDPMatcher struct { - Data UDPMatcherParams -} - -// UDPMatcherParams are the parameters used to create a UDPMatcher. -type UDPMatcherParams struct { - SourcePortStart uint16 - SourcePortEnd uint16 - DestinationPortStart uint16 - DestinationPortEnd uint16 - InverseFlags uint8 -} - -// NewUDPMatcher returns a new instance of UDPMatcher. -func NewUDPMatcher(filter IPHeaderFilter, data UDPMatcherParams) (Matcher, error) { - log.Infof("Adding rule with UDPMatcherParams: %+v", data) - - if data.InverseFlags != 0 { - return nil, fmt.Errorf("unsupported UDP matcher inverse flags set") - } - - if filter.Protocol != header.UDPProtocolNumber { - return nil, fmt.Errorf("UDP matching is only valid for protocol %d.", header.UDPProtocolNumber) - } - - return &UDPMatcher{Data: data}, nil -} - -// Match implements Matcher.Match. -func (um *UDPMatcher) Match(hook Hook, pkt tcpip.PacketBuffer, interfaceName string) (bool, bool) { - netHeader := header.IPv4(pkt.NetworkHeader) - - // TODO(gvisor.dev/issue/170): Proto checks should ultimately be moved - // into the iptables.Check codepath as matchers are added. - if netHeader.TransportProtocol() != header.UDPProtocolNumber { - return false, false - } - - // We dont't match fragments. - if frag := netHeader.FragmentOffset(); frag != 0 { - if frag == 1 { - log.Warningf("Dropping UDP packet: malicious fragmented packet.") - return false, true - } - return false, false - } - - // Now we need the transport header. However, this may not have been set - // yet. - // TODO(gvisor.dev/issue/170): Parsing the transport header should - // ultimately be moved into the iptables.Check codepath as matchers are - // added. - var udpHeader header.UDP - if pkt.TransportHeader != nil { - udpHeader = header.UDP(pkt.TransportHeader) - } else { - // The UDP header hasn't been parsed yet. We have to do it here. - if len(pkt.Data.First()) < header.UDPMinimumSize { - // There's no valid UDP header here, so we hotdrop the - // packet. - log.Warningf("Dropping UDP packet: size too small.") - return false, true - } - udpHeader = header.UDP(pkt.Data.First()) - } - - // Check whether the source and destination ports are within the - // matching range. - sourcePort := udpHeader.SourcePort() - destinationPort := udpHeader.DestinationPort() - if sourcePort < um.Data.SourcePortStart || um.Data.SourcePortEnd < sourcePort { - return false, false - } - if destinationPort < um.Data.DestinationPortStart || um.Data.DestinationPortEnd < destinationPort { - return false, false - } - - return true, false -} -- cgit v1.2.3 From 29ad5762e4549d961f48c65292cfdeb7256524f6 Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Fri, 31 Jan 2020 13:53:58 -0800 Subject: Spelling --- pkg/sentry/socket/netfilter/extensions.go | 18 +++++++++--------- pkg/sentry/socket/netfilter/tcp_matcher.go | 10 +++++----- pkg/sentry/socket/netfilter/udp_matcher.go | 10 +++++----- 3 files changed, 19 insertions(+), 19 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/netfilter/extensions.go b/pkg/sentry/socket/netfilter/extensions.go index 5a4cac84c..b5fbb52e4 100644 --- a/pkg/sentry/socket/netfilter/extensions.go +++ b/pkg/sentry/socket/netfilter/extensions.go @@ -32,8 +32,8 @@ import ( // - Proto // - Family -// matchMarshaler knows how to (un)marshal the matcher named name(). -type matchMarshaler interface { +// matchMaker knows how to (un)marshal the matcher named name(). +type matchMaker interface { // name is the matcher name as stored in the xt_entry_match struct. name() string @@ -45,19 +45,19 @@ type matchMarshaler interface { unmarshal(buf []byte, filter iptables.IPHeaderFilter) (iptables.Matcher, error) } -var matchMarshalers = map[string]matchMarshaler{} +var matchMakers = map[string]matchMaker{} -// registerMatchMarshaler should be called by match extensions to register them +// registermatchMaker should be called by match extensions to register them // with the netfilter package. -func registerMatchMarshaler(mm matchMarshaler) { - if _, ok := matchMarshalers[mm.name()]; ok { +func registerMatchMaker(mm matchMaker) { + if _, ok := matchMakers[mm.name()]; ok { panic(fmt.Sprintf("Multiple matches registered with name %q.", mm.name())) } - matchMarshalers[mm.name()] = mm + matchMakers[mm.name()] = mm } func marshalMatcher(matcher iptables.Matcher) []byte { - matchMaker, ok := matchMarshalers[matcher.Name()] + matchMaker, ok := matchMakers[matcher.Name()] if !ok { panic(fmt.Errorf("Unknown matcher of type %T.", matcher)) } @@ -85,7 +85,7 @@ func marshalEntryMatch(name string, data []byte) []byte { } func unmarshalMatcher(match linux.XTEntryMatch, filter iptables.IPHeaderFilter, buf []byte) (iptables.Matcher, error) { - matchMaker, ok := matchMarshalers[match.Name.String()] + matchMaker, ok := matchMakers[match.Name.String()] if !ok { return nil, fmt.Errorf("unsupported matcher with name %q", match.Name.String()) } diff --git a/pkg/sentry/socket/netfilter/tcp_matcher.go b/pkg/sentry/socket/netfilter/tcp_matcher.go index 1646d22f7..6b2f4c31a 100644 --- a/pkg/sentry/socket/netfilter/tcp_matcher.go +++ b/pkg/sentry/socket/netfilter/tcp_matcher.go @@ -28,18 +28,18 @@ import ( const matcherNameTCP = "tcp" func init() { - registerMatchMarshaler(tcpMarshaler{}) + registerMatchMaker(tcpMarshaler{}) } -// tcpMarshaler implements matchMarshaler for TCP matching. +// tcpMarshaler implements matchMaker for TCP matching. type tcpMarshaler struct{} -// name implements matchMarshaler.name. +// name implements matchMaker.name. func (tcpMarshaler) name() string { return matcherNameTCP } -// marshal implements matchMarshaler.marshal. +// marshal implements matchMaker.marshal. func (tcpMarshaler) marshal(mr iptables.Matcher) []byte { matcher := mr.(*TCPMatcher) xttcp := linux.XTTCP{ @@ -52,7 +52,7 @@ func (tcpMarshaler) marshal(mr iptables.Matcher) []byte { return marshalEntryMatch(matcherNameTCP, binary.Marshal(buf, usermem.ByteOrder, xttcp)) } -// unmarshal implements matchMarshaler.unmarshal. +// unmarshal implements matchMaker.unmarshal. func (tcpMarshaler) unmarshal(buf []byte, filter iptables.IPHeaderFilter) (iptables.Matcher, error) { if len(buf) < linux.SizeOfXTTCP { return nil, fmt.Errorf("buf has insufficient size for TCP match: %d", len(buf)) diff --git a/pkg/sentry/socket/netfilter/udp_matcher.go b/pkg/sentry/socket/netfilter/udp_matcher.go index b6e95bbc5..86aa11696 100644 --- a/pkg/sentry/socket/netfilter/udp_matcher.go +++ b/pkg/sentry/socket/netfilter/udp_matcher.go @@ -28,18 +28,18 @@ import ( const matcherNameUDP = "udp" func init() { - registerMatchMarshaler(udpMarshaler{}) + registerMatchMaker(udpMarshaler{}) } -// udpMarshaler implements matchMarshaler for UDP matching. +// udpMarshaler implements matchMaker for UDP matching. type udpMarshaler struct{} -// name implements matchMarshaler.name. +// name implements matchMaker.name. func (udpMarshaler) name() string { return matcherNameUDP } -// marshal implements matchMarshaler.marshal. +// marshal implements matchMaker.marshal. func (udpMarshaler) marshal(mr iptables.Matcher) []byte { matcher := mr.(*UDPMatcher) xtudp := linux.XTUDP{ @@ -52,7 +52,7 @@ func (udpMarshaler) marshal(mr iptables.Matcher) []byte { return marshalEntryMatch(matcherNameUDP, binary.Marshal(buf, usermem.ByteOrder, xtudp)) } -// unmarshal implements matchMarshaler.unmarshal. +// unmarshal implements matchMaker.unmarshal. func (udpMarshaler) unmarshal(buf []byte, filter iptables.IPHeaderFilter) (iptables.Matcher, error) { if len(buf) < linux.SizeOfXTUDP { return nil, fmt.Errorf("buf has insufficient size for UDP match: %d", len(buf)) -- cgit v1.2.3 From 492229d0176c1af2ab4ea4cf91bf211e940b5b12 Mon Sep 17 00:00:00 2001 From: Jamie Liu Date: Tue, 4 Feb 2020 11:28:36 -0800 Subject: VFS2 gofer client Updates #1198 Opening host pipes (by spinning in fdpipe) and host sockets is not yet complete, and will be done in a future CL. Major differences from VFS1 gofer client (sentry/fs/gofer), with varying levels of backportability: - "Cache policies" are replaced by InteropMode, which control the behavior of timestamps in addition to caching. Under InteropModeExclusive (analogous to cacheAll) and InteropModeWritethrough (analogous to cacheAllWritethrough), client timestamps are *not* written back to the server (it is not possible in 9P or Linux for clients to set ctime, so writing back client-authoritative timestamps results in incoherence between atime/mtime and ctime). Under InteropModeShared (analogous to cacheRemoteRevalidating), client timestamps are not used at all (remote filesystem clocks are authoritative). cacheNone is translated to InteropModeShared + new option filesystemOptions.specialRegularFiles. - Under InteropModeShared, "unstable attribute" reloading for permission checks, lookup, and revalidation are fused, which is feasible in VFS2 since gofer.filesystem controls path resolution. This results in a ~33% reduction in RPCs for filesystem operations compared to cacheRemoteRevalidating. For example, consider stat("/foo/bar/baz") where "/foo/bar/baz" fails revalidation, resulting in the instantiation of a new dentry: VFS1 RPCs: getattr("/") // fs.MountNamespace.FindLink() => fs.Inode.CheckPermission() => gofer.inodeOperations.check() => gofer.inodeOperations.UnstableAttr() walkgetattr("/", "foo") = fid1 // fs.Dirent.walk() => gofer.session.Revalidate() => gofer.cachePolicy.Revalidate() clunk(fid1) getattr("/foo") // CheckPermission walkgetattr("/foo", "bar") = fid2 // Revalidate clunk(fid2) getattr("/foo/bar") // CheckPermission walkgetattr("/foo/bar", "baz") = fid3 // Revalidate clunk(fid3) walkgetattr("/foo/bar", "baz") = fid4 // fs.Dirent.walk() => gofer.inodeOperations.Lookup getattr("/foo/bar/baz") // linux.stat() => gofer.inodeOperations.UnstableAttr() VFS2 RPCs: getattr("/") // gofer.filesystem.walkExistingLocked() walkgetattr("/", "foo") = fid1 // gofer.filesystem.stepExistingLocked() clunk(fid1) // No getattr: walkgetattr already updated metadata for permission check walkgetattr("/foo", "bar") = fid2 clunk(fid2) walkgetattr("/foo/bar", "baz") = fid3 // No clunk: fid3 used for new gofer.dentry // No getattr: walkgetattr already updated metadata for stat() - gofer.filesystem.unlinkAt() does not require instantiation of a dentry that represents the file to be deleted. Updates #898. - gofer.regularFileFD.OnClose() skips Tflushf for regular files under InteropModeExclusive, as it's nonsensical to request a remote file flush without flushing locally-buffered writes to that remote file first. - Symlink targets are cached when InteropModeShared is not in effect. - p9.QID.Path (which is already required to be unique for each file within a server, and is accordingly already synthesized from device/inode numbers in all known gofers) is used as-is for inode numbers, rather than being mapped along with attr.RDev in the client to yet another synthetic inode number. - Relevant parts of fsutil.CachingInodeOperations are inlined directly into gofer package code. This avoids having to duplicate part of its functionality in fsutil.HostMappable. PiperOrigin-RevId: 293190213 --- pkg/safemem/seq_unsafe.go | 17 + pkg/sentry/fs/fsutil/BUILD | 4 +- pkg/sentry/fs/fsutil/frame_ref_set.go | 13 +- pkg/sentry/fs/fsutil/inode_cached.go | 2 +- pkg/sentry/fsimpl/gofer/BUILD | 55 ++ pkg/sentry/fsimpl/gofer/directory.go | 190 +++++ pkg/sentry/fsimpl/gofer/filesystem.go | 1087 ++++++++++++++++++++++++++++ pkg/sentry/fsimpl/gofer/gofer.go | 1147 ++++++++++++++++++++++++++++++ pkg/sentry/fsimpl/gofer/handle.go | 135 ++++ pkg/sentry/fsimpl/gofer/handle_unsafe.go | 66 ++ pkg/sentry/fsimpl/gofer/p9file.go | 219 ++++++ pkg/sentry/fsimpl/gofer/pagemath.go | 31 + pkg/sentry/fsimpl/gofer/regular_file.go | 860 ++++++++++++++++++++++ pkg/sentry/fsimpl/gofer/special_file.go | 159 +++++ pkg/sentry/fsimpl/gofer/symlink.go | 47 ++ pkg/sentry/fsimpl/gofer/time.go | 75 ++ pkg/sentry/fsimpl/tmpfs/filesystem.go | 2 +- pkg/sentry/socket/hostinet/socket.go | 23 +- 18 files changed, 4103 insertions(+), 29 deletions(-) create mode 100644 pkg/sentry/fsimpl/gofer/BUILD create mode 100644 pkg/sentry/fsimpl/gofer/directory.go create mode 100644 pkg/sentry/fsimpl/gofer/filesystem.go create mode 100644 pkg/sentry/fsimpl/gofer/gofer.go create mode 100644 pkg/sentry/fsimpl/gofer/handle.go create mode 100644 pkg/sentry/fsimpl/gofer/handle_unsafe.go create mode 100644 pkg/sentry/fsimpl/gofer/p9file.go create mode 100644 pkg/sentry/fsimpl/gofer/pagemath.go create mode 100644 pkg/sentry/fsimpl/gofer/regular_file.go create mode 100644 pkg/sentry/fsimpl/gofer/special_file.go create mode 100644 pkg/sentry/fsimpl/gofer/symlink.go create mode 100644 pkg/sentry/fsimpl/gofer/time.go (limited to 'pkg/sentry/socket') diff --git a/pkg/safemem/seq_unsafe.go b/pkg/safemem/seq_unsafe.go index 354a95dde..dcdfc9600 100644 --- a/pkg/safemem/seq_unsafe.go +++ b/pkg/safemem/seq_unsafe.go @@ -18,6 +18,7 @@ import ( "bytes" "fmt" "reflect" + "syscall" "unsafe" ) @@ -297,3 +298,19 @@ func ZeroSeq(dsts BlockSeq) (uint64, error) { } return done, nil } + +// IovecsFromBlockSeq returns a []syscall.Iovec representing seq. +func IovecsFromBlockSeq(bs BlockSeq) []syscall.Iovec { + iovs := make([]syscall.Iovec, 0, bs.NumBlocks()) + for ; !bs.IsEmpty(); bs = bs.Tail() { + b := bs.Head() + iovs = append(iovs, syscall.Iovec{ + Base: &b.ToSlice()[0], + Len: uint64(b.Len()), + }) + // We don't need to care about b.NeedSafecopy(), because the host + // kernel will handle such address ranges just fine (by returning + // EFAULT). + } + return iovs +} diff --git a/pkg/sentry/fs/fsutil/BUILD b/pkg/sentry/fs/fsutil/BUILD index 4ab2a384f..789369220 100644 --- a/pkg/sentry/fs/fsutil/BUILD +++ b/pkg/sentry/fs/fsutil/BUILD @@ -28,13 +28,13 @@ go_template_instance( "platform": "gvisor.dev/gvisor/pkg/sentry/platform", }, package = "fsutil", - prefix = "frameRef", + prefix = "FrameRef", template = "//pkg/segment:generic_set", types = { "Key": "uint64", "Range": "platform.FileRange", "Value": "uint64", - "Functions": "frameRefSetFunctions", + "Functions": "FrameRefSetFunctions", }, ) diff --git a/pkg/sentry/fs/fsutil/frame_ref_set.go b/pkg/sentry/fs/fsutil/frame_ref_set.go index dd63db32b..6564fd0c6 100644 --- a/pkg/sentry/fs/fsutil/frame_ref_set.go +++ b/pkg/sentry/fs/fsutil/frame_ref_set.go @@ -20,24 +20,25 @@ import ( "gvisor.dev/gvisor/pkg/sentry/platform" ) -type frameRefSetFunctions struct{} +// FrameRefSetFunctions implements segment.Functions for FrameRefSet. +type FrameRefSetFunctions struct{} // MinKey implements segment.Functions.MinKey. -func (frameRefSetFunctions) MinKey() uint64 { +func (FrameRefSetFunctions) MinKey() uint64 { return 0 } // MaxKey implements segment.Functions.MaxKey. -func (frameRefSetFunctions) MaxKey() uint64 { +func (FrameRefSetFunctions) MaxKey() uint64 { return math.MaxUint64 } // ClearValue implements segment.Functions.ClearValue. -func (frameRefSetFunctions) ClearValue(val *uint64) { +func (FrameRefSetFunctions) ClearValue(val *uint64) { } // Merge implements segment.Functions.Merge. -func (frameRefSetFunctions) Merge(_ platform.FileRange, val1 uint64, _ platform.FileRange, val2 uint64) (uint64, bool) { +func (FrameRefSetFunctions) Merge(_ platform.FileRange, val1 uint64, _ platform.FileRange, val2 uint64) (uint64, bool) { if val1 != val2 { return 0, false } @@ -45,6 +46,6 @@ func (frameRefSetFunctions) Merge(_ platform.FileRange, val1 uint64, _ platform. } // Split implements segment.Functions.Split. -func (frameRefSetFunctions) Split(_ platform.FileRange, val uint64, _ uint64) (uint64, uint64) { +func (FrameRefSetFunctions) Split(_ platform.FileRange, val uint64, _ uint64) (uint64, uint64) { return val, val } diff --git a/pkg/sentry/fs/fsutil/inode_cached.go b/pkg/sentry/fs/fsutil/inode_cached.go index 573b8586e..800c8b4e1 100644 --- a/pkg/sentry/fs/fsutil/inode_cached.go +++ b/pkg/sentry/fs/fsutil/inode_cached.go @@ -111,7 +111,7 @@ type CachingInodeOperations struct { // refs tracks active references to data in the cache. // // refs is protected by dataMu. - refs frameRefSet + refs FrameRefSet } // CachingInodeOperationsOptions configures a CachingInodeOperations. diff --git a/pkg/sentry/fsimpl/gofer/BUILD b/pkg/sentry/fsimpl/gofer/BUILD new file mode 100644 index 000000000..4ba76a1e8 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/BUILD @@ -0,0 +1,55 @@ +load("//tools:defs.bzl", "go_library") +load("//tools/go_generics:defs.bzl", "go_template_instance") + +licenses(["notice"]) + +go_template_instance( + name = "dentry_list", + out = "dentry_list.go", + package = "gofer", + prefix = "dentry", + template = "//pkg/ilist:generic_list", + types = { + "Element": "*dentry", + "Linker": "*dentry", + }, +) + +go_library( + name = "gofer", + srcs = [ + "dentry_list.go", + "directory.go", + "filesystem.go", + "gofer.go", + "handle.go", + "handle_unsafe.go", + "p9file.go", + "pagemath.go", + "regular_file.go", + "special_file.go", + "symlink.go", + "time.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fd", + "//pkg/fspath", + "//pkg/log", + "//pkg/p9", + "//pkg/safemem", + "//pkg/sentry/fs/fsutil", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/time", + "//pkg/sentry/memmap", + "//pkg/sentry/pgalloc", + "//pkg/sentry/platform", + "//pkg/sentry/usage", + "//pkg/sentry/vfs", + "//pkg/syserror", + "//pkg/unet", + "//pkg/usermem", + ], +) diff --git a/pkg/sentry/fsimpl/gofer/directory.go b/pkg/sentry/fsimpl/gofer/directory.go new file mode 100644 index 000000000..baa2cdd8e --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/directory.go @@ -0,0 +1,190 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "sync" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +func (d *dentry) isDir() bool { + return d.fileType() == linux.S_IFDIR +} + +// Preconditions: d.dirMu must be locked. d.isDir(). fs.opts.interop != +// InteropModeShared. +func (d *dentry) cacheNegativeChildLocked(name string) { + if d.negativeChildren == nil { + d.negativeChildren = make(map[string]struct{}) + } + d.negativeChildren[name] = struct{}{} +} + +type directoryFD struct { + fileDescription + vfs.DirectoryFileDescriptionDefaultImpl + + mu sync.Mutex + off int64 + dirents []vfs.Dirent +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (fd *directoryFD) Release() { +} + +// IterDirents implements vfs.FileDescriptionImpl.IterDirents. +func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error { + fd.mu.Lock() + defer fd.mu.Unlock() + + if fd.dirents == nil { + ds, err := fd.dentry().getDirents(ctx) + if err != nil { + return err + } + fd.dirents = ds + } + + for fd.off < int64(len(fd.dirents)) { + if !cb.Handle(fd.dirents[fd.off]) { + return nil + } + fd.off++ + } + return nil +} + +// Preconditions: d.isDir(). There exists at least one directoryFD representing d. +func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { + // 9P2000.L's readdir does not specify behavior in the presence of + // concurrent mutation of an iterated directory, so implementations may + // duplicate or omit entries in this case, which violates POSIX semantics. + // Thus we read all directory entries while holding d.dirMu to exclude + // directory mutations. (Note that it is impossible for the client to + // exclude concurrent mutation from other remote filesystem users. Since + // there is no way to detect if the server has incorrectly omitted + // directory entries, we simply assume that the server is well-behaved + // under InteropModeShared.) This is inconsistent with Linux (which appears + // to assume that directory fids have the correct semantics, and translates + // struct file_operations::readdir calls directly to readdir RPCs), but is + // consistent with VFS1. + + d.fs.renameMu.RLock() + defer d.fs.renameMu.RUnlock() + d.dirMu.Lock() + defer d.dirMu.Unlock() + if d.dirents != nil { + return d.dirents, nil + } + + // It's not clear if 9P2000.L's readdir is expected to return "." and "..", + // so we generate them here. + parent := d.vfsd.ParentOrSelf().Impl().(*dentry) + dirents := []vfs.Dirent{ + { + Name: ".", + Type: linux.DT_DIR, + Ino: d.ino, + NextOff: 1, + }, + { + Name: "..", + Type: uint8(atomic.LoadUint32(&parent.mode) >> 12), + Ino: parent.ino, + NextOff: 2, + }, + } + off := uint64(0) + const count = 64 * 1024 // for consistency with the vfs1 client + d.handleMu.RLock() + defer d.handleMu.RUnlock() + if !d.handleReadable { + // This should not be possible because a readable handle should have + // been opened when the calling directoryFD was opened. + panic("gofer.dentry.getDirents called without a readable handle") + } + for { + p9ds, err := d.handle.file.readdir(ctx, off, count) + if err != nil { + return nil, err + } + if len(p9ds) == 0 { + // Cache dirents for future directoryFDs if permitted. + if d.fs.opts.interop != InteropModeShared { + d.dirents = dirents + } + return dirents, nil + } + for _, p9d := range p9ds { + if p9d.Name == "." || p9d.Name == ".." { + continue + } + dirent := vfs.Dirent{ + Name: p9d.Name, + Ino: p9d.QID.Path, + NextOff: int64(len(dirents) + 1), + } + // p9 does not expose 9P2000.U's DMDEVICE, DMNAMEDPIPE, or + // DMSOCKET. + switch p9d.Type { + case p9.TypeSymlink: + dirent.Type = linux.DT_LNK + case p9.TypeDir: + dirent.Type = linux.DT_DIR + default: + dirent.Type = linux.DT_REG + } + dirents = append(dirents, dirent) + } + off = p9ds[len(p9ds)-1].Offset + } +} + +// Seek implements vfs.FileDescriptionImpl.Seek. +func (fd *directoryFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + fd.mu.Lock() + defer fd.mu.Unlock() + + switch whence { + case linux.SEEK_SET: + if offset < 0 { + return 0, syserror.EINVAL + } + if offset == 0 { + // Ensure that the next call to fd.IterDirents() calls + // fd.dentry().getDirents(). + fd.dirents = nil + } + fd.off = offset + return fd.off, nil + case linux.SEEK_CUR: + offset += fd.off + if offset < 0 { + return 0, syserror.EINVAL + } + // Don't clear fd.dirents in this case, even if offset == 0. + fd.off = offset + return fd.off, nil + default: + return 0, syserror.EINVAL + } +} diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go new file mode 100644 index 000000000..8eb61debf --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/filesystem.go @@ -0,0 +1,1087 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "sync" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Sync implements vfs.FilesystemImpl.Sync. +func (fs *filesystem) Sync(ctx context.Context) error { + // Snapshot current dentries and special files. + fs.syncMu.Lock() + ds := make([]*dentry, 0, len(fs.dentries)) + for d := range fs.dentries { + ds = append(ds, d) + } + sffds := make([]*specialFileFD, 0, len(fs.specialFileFDs)) + for sffd := range fs.specialFileFDs { + sffds = append(sffds, sffd) + } + fs.syncMu.Unlock() + + // Return the first error we encounter, but sync everything we can + // regardless. + var retErr error + + // Sync regular files. + for _, d := range ds { + if !d.TryIncRef() { + continue + } + err := d.syncSharedHandle(ctx) + d.DecRef() + if err != nil && retErr == nil { + retErr = err + } + } + + // Sync special files, which may be writable but do not use dentry shared + // handles (so they won't be synced by the above). + for _, sffd := range sffds { + if !sffd.vfsfd.TryIncRef() { + continue + } + err := sffd.Sync(ctx) + sffd.vfsfd.DecRef() + if err != nil && retErr == nil { + retErr = err + } + } + + return retErr +} + +// maxFilenameLen is the maximum length of a filename. This is dictated by 9P's +// encoding of strings, which uses 2 bytes for the length prefix. +const maxFilenameLen = (1 << 16) - 1 + +// dentrySlicePool is a pool of *[]*dentry used to store dentries for which +// dentry.checkCachingLocked() must be called. The pool holds pointers to +// slices because Go lacks generics, so sync.Pool operates on interface{}, so +// every call to (what should be) sync.Pool<[]*dentry>.Put() allocates a copy +// of the slice header on the heap. +var dentrySlicePool = sync.Pool{ + New: func() interface{} { + ds := make([]*dentry, 0, 4) // arbitrary non-zero initial capacity + return &ds + }, +} + +func appendDentry(ds *[]*dentry, d *dentry) *[]*dentry { + if ds == nil { + ds = dentrySlicePool.Get().(*[]*dentry) + } + *ds = append(*ds, d) + return ds +} + +// Preconditions: ds != nil. +func putDentrySlice(ds *[]*dentry) { + // Allow dentries to be GC'd. + for i := range *ds { + (*ds)[i] = nil + } + *ds = (*ds)[:0] + dentrySlicePool.Put(ds) +} + +// stepLocked resolves rp.Component() to an existing file, starting from the +// given directory. +// +// Dentries which may become cached as a result of the traversal are appended +// to *ds. +// +// Preconditions: fs.renameMu must be locked. d.dirMu must be locked. +// !rp.Done(). If fs.opts.interop == InteropModeShared, then d's cached +// metadata must be up to date. +func (fs *filesystem) stepLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) { + if !d.isDir() { + return nil, syserror.ENOTDIR + } + if err := d.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + return nil, err + } +afterSymlink: + name := rp.Component() + if name == "." { + rp.Advance() + return d, nil + } + if name == ".." { + parentVFSD, err := rp.ResolveParent(&d.vfsd) + if err != nil { + return nil, err + } + parent := parentVFSD.Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + // We must assume that parentVFSD is correct, because if d has been + // moved elsewhere in the remote filesystem so that its parent has + // changed, we have no way of determining its new parent's location + // in the filesystem. Get updated metadata for parentVFSD. + _, attrMask, attr, err := parent.file.getAttr(ctx, dentryAttrMask()) + if err != nil { + return nil, err + } + parent.updateFromP9Attrs(attrMask, &attr) + } + rp.Advance() + return parent, nil + } + childVFSD, err := rp.ResolveChild(&d.vfsd, name) + if err != nil { + return nil, err + } + // FIXME(jamieliu): Linux performs revalidation before mount lookup + // (fs/namei.c:lookup_fast() => __d_lookup_rcu(), d_revalidate(), + // __follow_mount_rcu()). + child, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), d, name, childVFSD, ds) + if err != nil { + return nil, err + } + if child == nil { + return nil, syserror.ENOENT + } + if child.isSymlink() && rp.ShouldFollowSymlink() { + target, err := child.readlink(ctx, rp.Mount()) + if err != nil { + return nil, err + } + if err := rp.HandleSymlink(target); err != nil { + return nil, err + } + goto afterSymlink // don't check the current directory again + } + rp.Advance() + return child, nil +} + +// revalidateChildLocked must be called after a call to parent.vfsd.Child(name) +// or vfs.ResolvingPath.ResolveChild(name) returns childVFSD (which may be +// nil) to verify that the returned child (or lack thereof) is correct. If no file +// exists at name, revalidateChildLocked returns (nil, nil). +// +// Preconditions: fs.renameMu must be locked. parent.dirMu must be locked. +// parent.isDir(). name is not "." or "..". +// +// Postconditions: If revalidateChildLocked returns a non-nil dentry, its +// cached metadata is up to date. +func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *dentry, name string, childVFSD *vfs.Dentry, ds **[]*dentry) (*dentry, error) { + if childVFSD != nil && fs.opts.interop != InteropModeShared { + // We have a cached dentry that is assumed to be correct. + return childVFSD.Impl().(*dentry), nil + } + // We either don't have a cached dentry or need to verify that it's still + // correct, either of which requires a remote lookup. Check if this name is + // valid before performing the lookup. + if len(name) > maxFilenameLen { + return nil, syserror.ENAMETOOLONG + } + // Check if we've already cached this lookup with a negative result. + if _, ok := parent.negativeChildren[name]; ok { + return nil, nil + } + // Perform the remote lookup. + qid, file, attrMask, attr, err := parent.file.walkGetAttrOne(ctx, name) + if err != nil && err != syserror.ENOENT { + return nil, err + } + if childVFSD != nil { + child := childVFSD.Impl().(*dentry) + if !file.isNil() && qid.Path == child.ino { + // The file at this path hasn't changed. Just update cached + // metadata. + file.close(ctx) + child.updateFromP9Attrs(attrMask, &attr) + return child, nil + } + // The file at this path has changed or no longer exists. Remove + // the stale dentry from the tree, and re-evaluate its caching + // status (i.e. if it has 0 references, drop it). + vfsObj.ForceDeleteDentry(childVFSD) + *ds = appendDentry(*ds, child) + childVFSD = nil + } + if file.isNil() { + // No file exists at this path now. Cache the negative lookup if + // allowed. + if fs.opts.interop != InteropModeShared { + parent.cacheNegativeChildLocked(name) + } + return nil, nil + } + // Create a new dentry representing the file. + child, err := fs.newDentry(ctx, file, qid, attrMask, &attr) + if err != nil { + file.close(ctx) + return nil, err + } + parent.IncRef() // reference held by child on its parent + parent.vfsd.InsertChild(&child.vfsd, name) + // For now, child has 0 references, so our caller should call + // child.checkCachingLocked(). + *ds = appendDentry(*ds, child) + return child, nil +} + +// walkParentDirLocked resolves all but the last path component of rp to an +// existing directory, starting from the given directory (which is usually +// rp.Start().Impl().(*dentry)). It does not check that the returned directory +// is searchable by the provider of rp. +// +// Preconditions: fs.renameMu must be locked. !rp.Done(). If fs.opts.interop == +// InteropModeShared, then d's cached metadata must be up to date. +func (fs *filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) { + for !rp.Final() { + d.dirMu.Lock() + next, err := fs.stepLocked(ctx, rp, d, ds) + d.dirMu.Unlock() + if err != nil { + return nil, err + } + d = next + } + if !d.isDir() { + return nil, syserror.ENOTDIR + } + return d, nil +} + +// resolveLocked resolves rp to an existing file. +// +// Preconditions: fs.renameMu must be locked. +func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, ds **[]*dentry) (*dentry, error) { + d := rp.Start().Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + // Get updated metadata for rp.Start() as required by fs.stepLocked(). + if err := d.updateFromGetattr(ctx); err != nil { + return nil, err + } + } + for !rp.Done() { + d.dirMu.Lock() + next, err := fs.stepLocked(ctx, rp, d, ds) + d.dirMu.Unlock() + if err != nil { + return nil, err + } + d = next + } + if rp.MustBeDir() && !d.isDir() { + return nil, syserror.ENOTDIR + } + return d, nil +} + +// doCreateAt checks that creating a file at rp is permitted, then invokes +// create to do so. +// +// Preconditions: !rp.Done(). For the final path component in rp, +// !rp.ShouldFollowSymlink(). +func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, create func(parent *dentry, name string) error) error { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + start := rp.Start().Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + // Get updated metadata for start as required by + // fs.walkParentDirLocked(). + if err := start.updateFromGetattr(ctx); err != nil { + return err + } + } + parent, err := fs.walkParentDirLocked(ctx, rp, start, &ds) + if err != nil { + return err + } + if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + return err + } + if parent.isDeleted() { + return syserror.ENOENT + } + name := rp.Component() + if name == "." || name == ".." { + return syserror.EEXIST + } + if len(name) > maxFilenameLen { + return syserror.ENAMETOOLONG + } + if !dir && rp.MustBeDir() { + return syserror.ENOENT + } + mnt := rp.Mount() + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + if fs.opts.interop == InteropModeShared { + // The existence of a dentry at name would be inconclusive because the + // file it represents may have been deleted from the remote filesystem, + // so we would need to make an RPC to revalidate the dentry. Just + // attempt the file creation RPC instead. If a file does exist, the RPC + // will fail with EEXIST like we would have. If the RPC succeeds, and a + // stale dentry exists, the dentry will fail revalidation next time + // it's used. + return create(parent, name) + } + if parent.vfsd.Child(name) != nil { + return syserror.EEXIST + } + // No cached dentry exists; however, there might still be an existing file + // at name. As above, we attempt the file creation RPC anyway. + if err := create(parent, name); err != nil { + return err + } + parent.touchCMtime(ctx) + delete(parent.negativeChildren, name) + parent.dirents = nil + return nil +} + +// Preconditions: !rp.Done(). +func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool) error { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + start := rp.Start().Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + // Get updated metadata for start as required by + // fs.walkParentDirLocked(). + if err := start.updateFromGetattr(ctx); err != nil { + return err + } + } + parent, err := fs.walkParentDirLocked(ctx, rp, start, &ds) + if err != nil { + return err + } + if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + return err + } + if err := rp.Mount().CheckBeginWrite(); err != nil { + return err + } + defer rp.Mount().EndWrite() + + name := rp.Component() + if dir { + if name == "." { + return syserror.EINVAL + } + if name == ".." { + return syserror.ENOTEMPTY + } + } else { + if name == "." || name == ".." { + return syserror.EISDIR + } + } + vfsObj := rp.VirtualFilesystem() + mntns := vfs.MountNamespaceFromContext(ctx) + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + childVFSD := parent.vfsd.Child(name) + var child *dentry + // We only need a dentry representing the file at name if it can be a mount + // point. If childVFSD is nil, then it can't be a mount point. If childVFSD + // is non-nil but stale, the actual file can't be a mount point either; we + // detect this case by just speculatively calling PrepareDeleteDentry and + // only revalidating the dentry if that fails (indicating that the existing + // dentry is a mount point). + if childVFSD != nil { + child = childVFSD.Impl().(*dentry) + if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil { + child, err = fs.revalidateChildLocked(ctx, vfsObj, parent, name, childVFSD, &ds) + if err != nil { + return err + } + if child != nil { + childVFSD = &child.vfsd + if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil { + return err + } + } else { + childVFSD = nil + } + } + } else if _, ok := parent.negativeChildren[name]; ok { + return syserror.ENOENT + } + flags := uint32(0) + if dir { + if child != nil && !child.isDir() { + return syserror.ENOTDIR + } + flags = linux.AT_REMOVEDIR + } else { + if child != nil && child.isDir() { + return syserror.EISDIR + } + if rp.MustBeDir() { + return syserror.ENOTDIR + } + } + err = parent.file.unlinkAt(ctx, name, flags) + if err != nil { + if childVFSD != nil { + vfsObj.AbortDeleteDentry(childVFSD) + } + return err + } + if fs.opts.interop != InteropModeShared { + parent.touchCMtime(ctx) + parent.cacheNegativeChildLocked(name) + parent.dirents = nil + } + if child != nil { + child.setDeleted() + vfsObj.CommitDeleteDentry(childVFSD) + ds = appendDentry(ds, child) + } + return nil +} + +// renameMuRUnlockAndCheckCaching calls fs.renameMu.RUnlock(), then calls +// dentry.checkCachingLocked on all dentries in *ds with fs.renameMu locked for +// writing. +// +// ds is a pointer-to-pointer since defer evaluates its arguments immediately, +// but dentry slices are allocated lazily, and it's much easier to say "defer +// fs.renameMuRUnlockAndCheckCaching(&ds)" than "defer func() { +// fs.renameMuRUnlockAndCheckCaching(ds) }()" to work around this. +func (fs *filesystem) renameMuRUnlockAndCheckCaching(ds **[]*dentry) { + fs.renameMu.RUnlock() + if *ds == nil { + return + } + if len(**ds) != 0 { + fs.renameMu.Lock() + for _, d := range **ds { + d.checkCachingLocked() + } + fs.renameMu.Unlock() + } + putDentrySlice(*ds) +} + +func (fs *filesystem) renameMuUnlockAndCheckCaching(ds **[]*dentry) { + if *ds == nil { + fs.renameMu.Unlock() + return + } + for _, d := range **ds { + d.checkCachingLocked() + } + fs.renameMu.Unlock() + putDentrySlice(*ds) +} + +// GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. +func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return nil, err + } + if opts.CheckSearchable { + if !d.isDir() { + return nil, syserror.ENOTDIR + } + if err := d.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + return nil, err + } + } + d.IncRef() + return &d.vfsd, nil +} + +// GetParentDentryAt implements vfs.FilesystemImpl.GetParentDentryAt. +func (fs *filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + start := rp.Start().Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + // Get updated metadata for start as required by + // fs.walkParentDirLocked(). + if err := start.updateFromGetattr(ctx); err != nil { + return nil, err + } + } + d, err := fs.walkParentDirLocked(ctx, rp, start, &ds) + if err != nil { + return nil, err + } + d.IncRef() + return &d.vfsd, nil +} + +// LinkAt implements vfs.FilesystemImpl.LinkAt. +func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry) error { + return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, childName string) error { + if rp.Mount() != vd.Mount() { + return syserror.EXDEV + } + // 9P2000.L supports hard links, but we don't. + return syserror.EPERM + }) +} + +// MkdirAt implements vfs.FilesystemImpl.MkdirAt. +func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { + return fs.doCreateAt(ctx, rp, true /* dir */, func(parent *dentry, name string) error { + creds := rp.Credentials() + _, err := parent.file.mkdir(ctx, name, (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) + return err + }) +} + +// MknodAt implements vfs.FilesystemImpl.MknodAt. +func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error { + return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string) error { + creds := rp.Credentials() + _, err := parent.file.mknod(ctx, name, (p9.FileMode)(opts.Mode), opts.DevMajor, opts.DevMinor, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) + return err + }) +} + +// OpenAt implements vfs.FilesystemImpl.OpenAt. +func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + // Reject O_TMPFILE, which is not supported; supporting it correctly in the + // presence of other remote filesystem users requires remote filesystem + // support, and it isn't clear that there's any way to implement this in + // 9P. + if opts.Flags&linux.O_TMPFILE != 0 { + return nil, syserror.EOPNOTSUPP + } + mayCreate := opts.Flags&linux.O_CREAT != 0 + mustCreate := opts.Flags&(linux.O_CREAT|linux.O_EXCL) == (linux.O_CREAT | linux.O_EXCL) + + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + + start := rp.Start().Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + // Get updated metadata for start as required by fs.stepLocked(). + if err := start.updateFromGetattr(ctx); err != nil { + return nil, err + } + } + if rp.Done() { + return start.openLocked(ctx, rp, opts.Flags) + } + +afterTrailingSymlink: + parent, err := fs.walkParentDirLocked(ctx, rp, start, &ds) + if err != nil { + return nil, err + } + // Check for search permission in the parent directory. + if err := parent.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + return nil, err + } + // Determine whether or not we need to create a file. + parent.dirMu.Lock() + child, err := fs.stepLocked(ctx, rp, parent, &ds) + if err == syserror.ENOENT && mayCreate { + fd, err := parent.createAndOpenChildLocked(ctx, rp, &opts) + parent.dirMu.Unlock() + return fd, err + } + if err != nil { + parent.dirMu.Unlock() + return nil, err + } + // Open existing child or follow symlink. + parent.dirMu.Unlock() + if mustCreate { + return nil, syserror.EEXIST + } + if child.isSymlink() && rp.ShouldFollowSymlink() { + target, err := child.readlink(ctx, rp.Mount()) + if err != nil { + return nil, err + } + if err := rp.HandleSymlink(target); err != nil { + return nil, err + } + start = parent + goto afterTrailingSymlink + } + return child.openLocked(ctx, rp, opts.Flags) +} + +// Preconditions: fs.renameMu must be locked. +func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, flags uint32) (*vfs.FileDescription, error) { + ats := vfs.AccessTypesForOpenFlags(flags) + if err := d.checkPermissions(rp.Credentials(), ats, d.isDir()); err != nil { + return nil, err + } + mnt := rp.Mount() + filetype := d.fileType() + switch { + case filetype == linux.S_IFREG && !d.fs.opts.regularFilesUseSpecialFileFD: + if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, flags&linux.O_TRUNC != 0); err != nil { + return nil, err + } + fd := ®ularFileFD{} + if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{ + AllowDirectIO: true, + }); err != nil { + return nil, err + } + return &fd.vfsfd, nil + case filetype == linux.S_IFDIR: + // Can't open directories with O_CREAT. + if flags&linux.O_CREAT != 0 { + return nil, syserror.EISDIR + } + // Can't open directories writably. + if ats&vfs.MayWrite != 0 { + return nil, syserror.EISDIR + } + if flags&linux.O_DIRECT != 0 { + return nil, syserror.EINVAL + } + if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, false /* write */, false /* trunc */); err != nil { + return nil, err + } + fd := &directoryFD{} + if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return &fd.vfsfd, nil + case filetype == linux.S_IFLNK: + // Can't open symlinks without O_PATH (which is unimplemented). + return nil, syserror.ELOOP + default: + if flags&linux.O_DIRECT != 0 { + return nil, syserror.EINVAL + } + h, err := openHandle(ctx, d.file, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, flags&linux.O_TRUNC != 0) + if err != nil { + return nil, err + } + fd := &specialFileFD{ + handle: h, + } + if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { + h.close(ctx) + return nil, err + } + return &fd.vfsfd, nil + } +} + +// Preconditions: d.fs.renameMu must be locked. d.dirMu must be locked. +func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions) (*vfs.FileDescription, error) { + if err := d.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { + return nil, err + } + if d.isDeleted() { + return nil, syserror.ENOENT + } + mnt := rp.Mount() + if err := mnt.CheckBeginWrite(); err != nil { + return nil, err + } + defer mnt.EndWrite() + + // 9P2000.L's lcreate takes a fid representing the parent directory, and + // converts it into an open fid representing the created file, so we need + // to duplicate the directory fid first. + _, dirfile, err := d.file.walk(ctx, nil) + if err != nil { + return nil, err + } + creds := rp.Credentials() + name := rp.Component() + fdobj, openFile, createQID, _, err := dirfile.create(ctx, name, (p9.OpenFlags)(opts.Flags), (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) + if err != nil { + dirfile.close(ctx) + return nil, err + } + // Then we need to walk to the file we just created to get a non-open fid + // representing it, and to get its metadata. This must use d.file since, as + // explained above, dirfile was invalidated by dirfile.Create(). + walkQID, nonOpenFile, attrMask, attr, err := d.file.walkGetAttrOne(ctx, name) + if err != nil { + openFile.close(ctx) + if fdobj != nil { + fdobj.Close() + } + return nil, err + } + // Sanity-check that we walked to the file we created. + if createQID.Path != walkQID.Path { + // Probably due to concurrent remote filesystem mutation? + ctx.Warningf("gofer.dentry.createAndOpenChildLocked: created file has QID %v before walk, QID %v after (interop=%v)", createQID, walkQID, d.fs.opts.interop) + nonOpenFile.close(ctx) + openFile.close(ctx) + if fdobj != nil { + fdobj.Close() + } + return nil, syserror.EAGAIN + } + + // Construct the new dentry. + child, err := d.fs.newDentry(ctx, nonOpenFile, createQID, attrMask, &attr) + if err != nil { + nonOpenFile.close(ctx) + openFile.close(ctx) + if fdobj != nil { + fdobj.Close() + } + return nil, err + } + // Incorporate the fid that was opened by lcreate. + useRegularFileFD := child.fileType() == linux.S_IFREG && !d.fs.opts.regularFilesUseSpecialFileFD + if useRegularFileFD { + child.handleMu.Lock() + child.handle.file = openFile + if fdobj != nil { + child.handle.fd = int32(fdobj.Release()) + } + child.handleReadable = vfs.MayReadFileWithOpenFlags(opts.Flags) + child.handleWritable = vfs.MayWriteFileWithOpenFlags(opts.Flags) + child.handleMu.Unlock() + } + // Take a reference on the new dentry to be held by the new file + // description. (This reference also means that the new dentry is not + // eligible for caching yet, so we don't need to append to a dentry slice.) + child.refs = 1 + // Insert the dentry into the tree. + d.IncRef() // reference held by child on its parent d + d.vfsd.InsertChild(&child.vfsd, name) + if d.fs.opts.interop != InteropModeShared { + d.touchCMtime(ctx) + delete(d.negativeChildren, name) + d.dirents = nil + } + + // Finally, construct a file description representing the created file. + var childVFSFD *vfs.FileDescription + mnt.IncRef() + if useRegularFileFD { + fd := ®ularFileFD{} + if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &child.vfsd, &vfs.FileDescriptionOptions{ + AllowDirectIO: true, + }); err != nil { + return nil, err + } + childVFSFD = &fd.vfsfd + } else { + fd := &specialFileFD{ + handle: handle{ + file: openFile, + fd: -1, + }, + } + if fdobj != nil { + fd.handle.fd = int32(fdobj.Release()) + } + if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &child.vfsd, &vfs.FileDescriptionOptions{}); err != nil { + fd.handle.close(ctx) + return nil, err + } + childVFSFD = &fd.vfsfd + } + return childVFSFD, nil +} + +// ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. +func (fs *filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (string, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return "", err + } + if !d.isSymlink() { + return "", syserror.EINVAL + } + return d.readlink(ctx, rp.Mount()) +} + +// RenameAt implements vfs.FilesystemImpl.RenameAt. +func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldParentVD vfs.VirtualDentry, oldName string, opts vfs.RenameOptions) error { + if opts.Flags != 0 { + // Requires 9P support. + return syserror.EINVAL + } + + var ds *[]*dentry + fs.renameMu.Lock() + defer fs.renameMuUnlockAndCheckCaching(&ds) + newParent, err := fs.walkParentDirLocked(ctx, rp, rp.Start().Impl().(*dentry), &ds) + if err != nil { + return err + } + newName := rp.Component() + if newName == "." || newName == ".." { + return syserror.EBUSY + } + mnt := rp.Mount() + if mnt != oldParentVD.Mount() { + return syserror.EXDEV + } + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + + oldParent := oldParentVD.Dentry().Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + if err := oldParent.updateFromGetattr(ctx); err != nil { + return err + } + } + if err := oldParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + return err + } + vfsObj := rp.VirtualFilesystem() + // We need a dentry representing the renamed file since, if it's a + // directory, we need to check for write permission on it. + oldParent.dirMu.Lock() + defer oldParent.dirMu.Unlock() + renamed, err := fs.revalidateChildLocked(ctx, vfsObj, oldParent, oldName, oldParent.vfsd.Child(oldName), &ds) + if err != nil { + return err + } + if renamed == nil { + return syserror.ENOENT + } + if renamed.isDir() { + if renamed == newParent || renamed.vfsd.IsAncestorOf(&newParent.vfsd) { + return syserror.EINVAL + } + if oldParent != newParent { + if err := renamed.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { + return err + } + } + } else { + if opts.MustBeDir || rp.MustBeDir() { + return syserror.ENOTDIR + } + } + + if oldParent != newParent { + if err := newParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + return err + } + newParent.dirMu.Lock() + defer newParent.dirMu.Unlock() + } + if newParent.isDeleted() { + return syserror.ENOENT + } + replacedVFSD := newParent.vfsd.Child(newName) + var replaced *dentry + // This is similar to unlinkAt, except: + // + // - We revalidate the replaced dentry unconditionally for simplicity. + // + // - If rp.MustBeDir(), then we need a dentry representing the replaced + // file regardless to confirm that it's a directory. + if replacedVFSD != nil || rp.MustBeDir() { + replaced, err = fs.revalidateChildLocked(ctx, vfsObj, newParent, newName, replacedVFSD, &ds) + if err != nil { + return err + } + if replaced != nil { + if replaced.isDir() { + if !renamed.isDir() { + return syserror.EISDIR + } + } else { + if rp.MustBeDir() || renamed.isDir() { + return syserror.ENOTDIR + } + } + replacedVFSD = &replaced.vfsd + } else { + replacedVFSD = nil + } + } + + if oldParent == newParent && oldName == newName { + return nil + } + if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), &renamed.vfsd, replacedVFSD); err != nil { + return err + } + if err := renamed.file.rename(ctx, newParent.file, newName); err != nil { + vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) + return err + } + if fs.opts.interop != InteropModeShared { + oldParent.cacheNegativeChildLocked(oldName) + oldParent.dirents = nil + delete(newParent.negativeChildren, newName) + newParent.dirents = nil + } + vfsObj.CommitRenameReplaceDentry(&renamed.vfsd, &newParent.vfsd, newName, replacedVFSD) + return nil +} + +// RmdirAt implements vfs.FilesystemImpl.RmdirAt. +func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error { + return fs.unlinkAt(ctx, rp, true /* dir */) +} + +// SetStatAt implements vfs.FilesystemImpl.SetStatAt. +func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return err + } + return d.setStat(ctx, rp.Credentials(), &opts.Stat, rp.Mount()) +} + +// StatAt implements vfs.FilesystemImpl.StatAt. +func (fs *filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.StatOptions) (linux.Statx, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return linux.Statx{}, err + } + // Since walking updates metadata for all traversed dentries under + // InteropModeShared, including the returned one, we can return cached + // metadata here regardless of fs.opts.interop. + var stat linux.Statx + d.statTo(&stat) + return stat, nil +} + +// StatFSAt implements vfs.FilesystemImpl.StatFSAt. +func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linux.Statfs, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return linux.Statfs{}, err + } + fsstat, err := d.file.statFS(ctx) + if err != nil { + return linux.Statfs{}, err + } + nameLen := uint64(fsstat.NameLength) + if nameLen > maxFilenameLen { + nameLen = maxFilenameLen + } + return linux.Statfs{ + // This is primarily for distinguishing a gofer file system in + // tests. Testing is important, so instead of defining + // something completely random, use a standard value. + Type: linux.V9FS_MAGIC, + BlockSize: int64(fsstat.BlockSize), + Blocks: fsstat.Blocks, + BlocksFree: fsstat.BlocksFree, + BlocksAvailable: fsstat.BlocksAvailable, + Files: fsstat.Files, + FilesFree: fsstat.FilesFree, + NameLength: nameLen, + }, nil +} + +// SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. +func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error { + return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string) error { + creds := rp.Credentials() + _, err := parent.file.symlink(ctx, target, name, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) + return err + }) +} + +// UnlinkAt implements vfs.FilesystemImpl.UnlinkAt. +func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error { + return fs.unlinkAt(ctx, rp, false /* dir */) +} + +// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. +func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return nil, err + } + return d.listxattr(ctx) +} + +// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. +func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return "", err + } + return d.getxattr(ctx, name) +} + +// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. +func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return err + } + return d.setxattr(ctx, &opts) +} + +// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. +func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return err + } + return d.removexattr(ctx, name) +} + +// PrependPath implements vfs.FilesystemImpl.PrependPath. +func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { + fs.renameMu.RLock() + defer fs.renameMu.RUnlock() + return vfs.GenericPrependPath(vfsroot, vd, b) +} diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go new file mode 100644 index 000000000..d0552bd99 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/gofer.go @@ -0,0 +1,1147 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package gofer provides a filesystem implementation that is backed by a 9p +// server, interchangably referred to as "gofers" throughout this package. +// +// Lock order: +// regularFileFD/directoryFD.mu +// filesystem.renameMu +// dentry.dirMu +// filesystem.syncMu +// dentry.metadataMu +// *** "memmap.Mappable locks" below this point +// dentry.mapsMu +// *** "memmap.Mappable locks taken by Translate" below this point +// dentry.handleMu +// dentry.dataMu +// +// Locking dentry.dirMu in multiple dentries requires holding +// filesystem.renameMu for writing. +package gofer + +import ( + "fmt" + "strconv" + "sync" + "sync/atomic" + "syscall" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/pgalloc" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/unet" + "gvisor.dev/gvisor/pkg/usermem" +) + +// FilesystemType implements vfs.FilesystemType. +type FilesystemType struct{} + +// filesystem implements vfs.FilesystemImpl. +type filesystem struct { + vfsfs vfs.Filesystem + + // mfp is used to allocate memory that caches regular file contents. mfp is + // immutable. + mfp pgalloc.MemoryFileProvider + + // Immutable options. + opts filesystemOptions + + // client is the client used by this filesystem. client is immutable. + client *p9.Client + + // uid and gid are the effective KUID and KGID of the filesystem's creator, + // and are used as the owner and group for files that don't specify one. + // uid and gid are immutable. + uid auth.KUID + gid auth.KGID + + // renameMu serves two purposes: + // + // - It synchronizes path resolution with renaming initiated by this + // client. + // + // - It is held by path resolution to ensure that reachable dentries remain + // valid. A dentry is reachable by path resolution if it has a non-zero + // reference count (such that it is usable as vfs.ResolvingPath.Start() or + // is reachable from its children), or if it is a child dentry (such that + // it is reachable from its parent). + renameMu sync.RWMutex + + // cachedDentries contains all dentries with 0 references. (Due to race + // conditions, it may also contain dentries with non-zero references.) + // cachedDentriesLen is the number of dentries in cachedDentries. These + // fields are protected by renameMu. + cachedDentries dentryList + cachedDentriesLen uint64 + + // dentries contains all dentries in this filesystem. specialFileFDs + // contains all open specialFileFDs. These fields are protected by syncMu. + syncMu sync.Mutex + dentries map[*dentry]struct{} + specialFileFDs map[*specialFileFD]struct{} +} + +type filesystemOptions struct { + // "Standard" 9P options. + fd int + aname string + interop InteropMode // derived from the "cache" mount option + msize uint32 + version string + + // maxCachedDentries is the maximum number of dentries with 0 references + // retained by the client. + maxCachedDentries uint64 + + // If forcePageCache is true, host FDs may not be used for application + // memory mappings even if available; instead, the client must perform its + // own caching of regular file pages. This is primarily useful for testing. + forcePageCache bool + + // If limitHostFDTranslation is true, apply maxFillRange() constraints to + // host FD mappings returned by dentry.(memmap.Mappable).Translate(). This + // makes memory accounting behavior more consistent between cases where + // host FDs are / are not available, but may increase the frequency of + // sentry-handled page faults on files for which a host FD is available. + limitHostFDTranslation bool + + // If overlayfsStaleRead is true, O_RDONLY host FDs provided by the remote + // filesystem may not be coherent with writable host FDs opened later, so + // mappings of the former must be replaced by mappings of the latter. This + // is usually only the case when the remote filesystem is an overlayfs + // mount on Linux < 4.19. + overlayfsStaleRead bool + + // If regularFilesUseSpecialFileFD is true, application FDs representing + // regular files will use distinct file handles for each FD, in the same + // way that application FDs representing "special files" such as sockets + // do. Note that this disables client caching and mmap for regular files. + regularFilesUseSpecialFileFD bool +} + +// InteropMode controls the client's interaction with other remote filesystem +// users. +type InteropMode uint32 + +const ( + // InteropModeExclusive is appropriate when the filesystem client is the + // only user of the remote filesystem. + // + // - The client may cache arbitrary filesystem state (file data, metadata, + // filesystem structure, etc.). + // + // - Client changes to filesystem state may be sent to the remote + // filesystem asynchronously, except when server permission checks are + // necessary. + // + // - File timestamps are based on client clocks. This ensures that users of + // the client observe timestamps that are coherent with their own clocks + // and consistent with Linux's semantics. However, since it is not always + // possible for clients to set arbitrary atimes and mtimes, and never + // possible for clients to set arbitrary ctimes, file timestamp changes are + // stored in the client only and never sent to the remote filesystem. + InteropModeExclusive InteropMode = iota + + // InteropModeWritethrough is appropriate when there are read-only users of + // the remote filesystem that expect to observe changes made by the + // filesystem client. + // + // - The client may cache arbitrary filesystem state. + // + // - Client changes to filesystem state must be sent to the remote + // filesystem synchronously. + // + // - File timestamps are based on client clocks. As a corollary, access + // timestamp changes from other remote filesystem users will not be visible + // to the client. + InteropModeWritethrough + + // InteropModeShared is appropriate when there are users of the remote + // filesystem that may mutate its state other than the client. + // + // - The client must verify cached filesystem state before using it. + // + // - Client changes to filesystem state must be sent to the remote + // filesystem synchronously. + // + // - File timestamps are based on server clocks. This is necessary to + // ensure that timestamp changes are synchronized between remote filesystem + // users. + // + // Note that the correctness of InteropModeShared depends on the server + // correctly implementing 9P fids (i.e. each fid immutably represents a + // single filesystem object), even in the presence of remote filesystem + // mutations from other users. If this is violated, the behavior of the + // client is undefined. + InteropModeShared +) + +// GetFilesystem implements vfs.FilesystemType.GetFilesystem. +func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + mfp := pgalloc.MemoryFileProviderFromContext(ctx) + if mfp == nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: context does not provide a pgalloc.MemoryFileProvider") + return nil, nil, syserror.EINVAL + } + + mopts := vfs.GenericParseMountOptions(opts.Data) + var fsopts filesystemOptions + + // Check that the transport is "fd". + trans, ok := mopts["trans"] + if !ok { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: transport must be specified as 'trans=fd'") + return nil, nil, syserror.EINVAL + } + delete(mopts, "trans") + if trans != "fd" { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: unsupported transport: trans=%s", trans) + return nil, nil, syserror.EINVAL + } + + // Check that read and write FDs are provided and identical. + rfdstr, ok := mopts["rfdno"] + if !ok { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: read FD must be specified as 'rfdno=") + return nil, nil, syserror.EINVAL + } + delete(mopts, "rfdno") + rfd, err := strconv.Atoi(rfdstr) + if err != nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid read FD: rfdno=%s", rfdstr) + return nil, nil, syserror.EINVAL + } + wfdstr, ok := mopts["wfdno"] + if !ok { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: write FD must be specified as 'wfdno=") + return nil, nil, syserror.EINVAL + } + delete(mopts, "wfdno") + wfd, err := strconv.Atoi(wfdstr) + if err != nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid write FD: wfdno=%s", wfdstr) + return nil, nil, syserror.EINVAL + } + if rfd != wfd { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: read FD (%d) and write FD (%d) must be equal", rfd, wfd) + return nil, nil, syserror.EINVAL + } + fsopts.fd = rfd + + // Get the attach name. + fsopts.aname = "/" + if aname, ok := mopts["aname"]; ok { + delete(mopts, "aname") + fsopts.aname = aname + } + + // Parse the cache policy. For historical reasons, this defaults to the + // least generally-applicable option, InteropModeExclusive. + fsopts.interop = InteropModeExclusive + if cache, ok := mopts["cache"]; ok { + delete(mopts, "cache") + switch cache { + case "fscache": + fsopts.interop = InteropModeExclusive + case "fscache_writethrough": + fsopts.interop = InteropModeWritethrough + case "none": + fsopts.regularFilesUseSpecialFileFD = true + fallthrough + case "remote_revalidating": + fsopts.interop = InteropModeShared + default: + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid cache policy: cache=%s", cache) + return nil, nil, syserror.EINVAL + } + } + + // Parse the 9P message size. + fsopts.msize = 1024 * 1024 // 1M, tested to give good enough performance up to 64M + if msizestr, ok := mopts["msize"]; ok { + delete(mopts, "msize") + msize, err := strconv.ParseUint(msizestr, 10, 32) + if err != nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid message size: msize=%s", msizestr) + return nil, nil, syserror.EINVAL + } + fsopts.msize = uint32(msize) + } + + // Parse the 9P protocol version. + fsopts.version = p9.HighestVersionString() + if version, ok := mopts["version"]; ok { + delete(mopts, "version") + fsopts.version = version + } + + // Parse the dentry cache limit. + fsopts.maxCachedDentries = 1000 + if str, ok := mopts["dentry_cache_limit"]; ok { + delete(mopts, "dentry_cache_limit") + maxCachedDentries, err := strconv.ParseUint(str, 10, 64) + if err != nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid dentry cache limit: dentry_cache_limit=%s", str) + return nil, nil, syserror.EINVAL + } + fsopts.maxCachedDentries = maxCachedDentries + } + + // Handle simple flags. + if _, ok := mopts["force_page_cache"]; ok { + delete(mopts, "force_page_cache") + fsopts.forcePageCache = true + } + if _, ok := mopts["limit_host_fd_translation"]; ok { + delete(mopts, "limit_host_fd_translation") + fsopts.limitHostFDTranslation = true + } + if _, ok := mopts["overlayfs_stale_read"]; ok { + delete(mopts, "overlayfs_stale_read") + fsopts.overlayfsStaleRead = true + } + // fsopts.regularFilesUseSpecialFileFD can only be enabled by specifying + // "cache=none". + + // Check for unparsed options. + if len(mopts) != 0 { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: unknown options: %v", mopts) + return nil, nil, syserror.EINVAL + } + + // Establish a connection with the server. + conn, err := unet.NewSocket(fsopts.fd) + if err != nil { + return nil, nil, err + } + + // Perform version negotiation with the server. + ctx.UninterruptibleSleepStart(false) + client, err := p9.NewClient(conn, fsopts.msize, fsopts.version) + ctx.UninterruptibleSleepFinish(false) + if err != nil { + conn.Close() + return nil, nil, err + } + // Ownership of conn has been transferred to client. + + // Perform attach to obtain the filesystem root. + ctx.UninterruptibleSleepStart(false) + attached, err := client.Attach(fsopts.aname) + ctx.UninterruptibleSleepFinish(false) + if err != nil { + client.Close() + return nil, nil, err + } + attachFile := p9file{attached} + qid, attrMask, attr, err := attachFile.getAttr(ctx, dentryAttrMask()) + if err != nil { + attachFile.close(ctx) + client.Close() + return nil, nil, err + } + + // Construct the filesystem object. + fs := &filesystem{ + mfp: mfp, + opts: fsopts, + uid: creds.EffectiveKUID, + gid: creds.EffectiveKGID, + client: client, + dentries: make(map[*dentry]struct{}), + specialFileFDs: make(map[*specialFileFD]struct{}), + } + fs.vfsfs.Init(vfsObj, fs) + + // Construct the root dentry. + root, err := fs.newDentry(ctx, attachFile, qid, attrMask, &attr) + if err != nil { + attachFile.close(ctx) + fs.vfsfs.DecRef() + return nil, nil, err + } + // Set the root's reference count to 2. One reference is returned to the + // caller, and the other is deliberately leaked to prevent the root from + // being "cached" and subsequently evicted. Its resources will still be + // cleaned up by fs.Release(). + root.refs = 2 + + return &fs.vfsfs, &root.vfsd, nil +} + +// Release implements vfs.FilesystemImpl.Release. +func (fs *filesystem) Release() { + ctx := context.Background() + mf := fs.mfp.MemoryFile() + + fs.syncMu.Lock() + for d := range fs.dentries { + d.handleMu.Lock() + d.dataMu.Lock() + if d.handleWritable { + // Write dirty cached data to the remote file. + if err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, fs.mfp.MemoryFile(), d.handle.writeFromBlocksAt); err != nil { + log.Warningf("gofer.filesystem.Release: failed to flush dentry: %v", err) + } + // TODO(jamieliu): Do we need to flushf/fsync d? + } + // Discard cached pages. + d.cache.DropAll(mf) + d.dirty.RemoveAll() + d.dataMu.Unlock() + // Close the host fd if one exists. + if d.handle.fd >= 0 { + syscall.Close(int(d.handle.fd)) + d.handle.fd = -1 + } + d.handleMu.Unlock() + } + // There can't be any specialFileFDs still using fs, since each such + // FileDescription would hold a reference on a Mount holding a reference on + // fs. + fs.syncMu.Unlock() + + // Close the connection to the server. This implicitly clunks all fids. + fs.client.Close() +} + +// dentry implements vfs.DentryImpl. +type dentry struct { + vfsd vfs.Dentry + + // refs is the reference count. Each dentry holds a reference on its + // parent, even if disowned. refs is accessed using atomic memory + // operations. + refs int64 + + // fs is the owning filesystem. fs is immutable. + fs *filesystem + + // We don't support hard links, so each dentry maps 1:1 to an inode. + + // file is the unopened p9.File that backs this dentry. file is immutable. + file p9file + + // If deleted is non-zero, the file represented by this dentry has been + // deleted. deleted is accessed using atomic memory operations. + deleted uint32 + + // If cached is true, dentryEntry links dentry into + // filesystem.cachedDentries. cached and dentryEntry are protected by + // filesystem.renameMu. + cached bool + dentryEntry + + dirMu sync.Mutex + + // If this dentry represents a directory, and InteropModeShared is not in + // effect, negativeChildren is a set of child names in this directory that + // are known not to exist. negativeChildren is protected by dirMu. + negativeChildren map[string]struct{} + + // If this dentry represents a directory, InteropModeShared is not in + // effect, and dirents is not nil, it is a cache of all entries in the + // directory, in the order they were returned by the server. dirents is + // protected by dirMu. + dirents []vfs.Dirent + + // Cached metadata; protected by metadataMu and accessed using atomic + // memory operations unless otherwise specified. + metadataMu sync.Mutex + ino uint64 // immutable + mode uint32 // type is immutable, perms are mutable + uid uint32 // auth.KUID, but stored as raw uint32 for sync/atomic + gid uint32 // auth.KGID, but ... + blockSize uint32 // 0 if unknown + // Timestamps, all nsecs from the Unix epoch. + atime int64 + mtime int64 + ctime int64 + btime int64 + // File size, protected by both metadataMu and dataMu (i.e. both must be + // locked to mutate it). + size uint64 + + mapsMu sync.Mutex + + // If this dentry represents a regular file, mappings tracks mappings of + // the file into memmap.MappingSpaces. mappings is protected by mapsMu. + mappings memmap.MappingSet + + // If this dentry represents a regular file or directory: + // + // - handle is the I/O handle used by all regularFileFDs/directoryFDs + // representing this dentry. + // + // - handleReadable is true if handle is readable. + // + // - handleWritable is true if handle is writable. + // + // Invariants: + // + // - If handleReadable == handleWritable == false, then handle.file == nil + // (i.e. there is no open handle). Conversely, if handleReadable || + // handleWritable == true, then handle.file != nil (i.e. there is an open + // handle). + // + // - handleReadable and handleWritable cannot transition from true to false + // (i.e. handles may not be downgraded). + // + // These fields are protected by handleMu. + handleMu sync.RWMutex + handle handle + handleReadable bool + handleWritable bool + + dataMu sync.RWMutex + + // If this dentry represents a regular file that is client-cached, cache + // maps offsets into the cached file to offsets into + // filesystem.mfp.MemoryFile() that store the file's data. cache is + // protected by dataMu. + cache fsutil.FileRangeSet + + // If this dentry represents a regular file that is client-cached, dirty + // tracks dirty segments in cache. dirty is protected by dataMu. + dirty fsutil.DirtySet + + // pf implements platform.File for mappings of handle.fd. + pf dentryPlatformFile + + // If this dentry represents a symbolic link, InteropModeShared is not in + // effect, and haveTarget is true, target is the symlink target. haveTarget + // and target are protected by dataMu. + haveTarget bool + target string +} + +// dentryAttrMask returns a p9.AttrMask enabling all attributes used by the +// gofer client. +func dentryAttrMask() p9.AttrMask { + return p9.AttrMask{ + Mode: true, + UID: true, + GID: true, + ATime: true, + MTime: true, + CTime: true, + Size: true, + BTime: true, + } +} + +// newDentry creates a new dentry representing the given file. The dentry +// initially has no references, but is not cached; it is the caller's +// responsibility to set the dentry's reference count and/or call +// dentry.checkCachingLocked() as appropriate. +func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, mask p9.AttrMask, attr *p9.Attr) (*dentry, error) { + if !mask.Mode { + ctx.Warningf("can't create gofer.dentry without file type") + return nil, syserror.EIO + } + if attr.Mode.FileType() == p9.ModeRegular && !mask.Size { + ctx.Warningf("can't create regular file gofer.dentry without file size") + return nil, syserror.EIO + } + + d := &dentry{ + fs: fs, + file: file, + ino: qid.Path, + mode: uint32(attr.Mode), + uid: uint32(fs.uid), + gid: uint32(fs.gid), + blockSize: usermem.PageSize, + handle: handle{ + fd: -1, + }, + } + d.pf.dentry = d + if mask.UID { + d.uid = uint32(attr.UID) + } + if mask.GID { + d.gid = uint32(attr.GID) + } + if mask.Size { + d.size = attr.Size + } + if attr.BlockSize != 0 { + d.blockSize = uint32(attr.BlockSize) + } + if mask.ATime { + d.atime = dentryTimestampFromP9(attr.ATimeSeconds, attr.ATimeNanoSeconds) + } + if mask.MTime { + d.mtime = dentryTimestampFromP9(attr.MTimeSeconds, attr.MTimeNanoSeconds) + } + if mask.CTime { + d.ctime = dentryTimestampFromP9(attr.CTimeSeconds, attr.CTimeNanoSeconds) + } + if mask.BTime { + d.btime = dentryTimestampFromP9(attr.BTimeSeconds, attr.BTimeNanoSeconds) + } + d.vfsd.Init(d) + + fs.syncMu.Lock() + fs.dentries[d] = struct{}{} + fs.syncMu.Unlock() + return d, nil +} + +// updateFromP9Attrs is called to update d's metadata after an update from the +// remote filesystem. +func (d *dentry) updateFromP9Attrs(mask p9.AttrMask, attr *p9.Attr) { + d.metadataMu.Lock() + if mask.Mode { + if got, want := uint32(attr.Mode.FileType()), d.fileType(); got != want { + d.metadataMu.Unlock() + panic(fmt.Sprintf("gofer.dentry file type changed from %#o to %#o", want, got)) + } + atomic.StoreUint32(&d.mode, uint32(attr.Mode)) + } + if mask.UID { + atomic.StoreUint32(&d.uid, uint32(attr.UID)) + } + if mask.GID { + atomic.StoreUint32(&d.gid, uint32(attr.GID)) + } + // There is no P9_GETATTR_* bit for I/O block size. + if attr.BlockSize != 0 { + atomic.StoreUint32(&d.blockSize, uint32(attr.BlockSize)) + } + if mask.ATime { + atomic.StoreInt64(&d.atime, dentryTimestampFromP9(attr.ATimeSeconds, attr.ATimeNanoSeconds)) + } + if mask.MTime { + atomic.StoreInt64(&d.mtime, dentryTimestampFromP9(attr.MTimeSeconds, attr.MTimeNanoSeconds)) + } + if mask.CTime { + atomic.StoreInt64(&d.ctime, dentryTimestampFromP9(attr.CTimeSeconds, attr.CTimeNanoSeconds)) + } + if mask.BTime { + atomic.StoreInt64(&d.btime, dentryTimestampFromP9(attr.BTimeSeconds, attr.BTimeNanoSeconds)) + } + if mask.Size { + d.dataMu.Lock() + atomic.StoreUint64(&d.size, attr.Size) + d.dataMu.Unlock() + } + d.metadataMu.Unlock() +} + +func (d *dentry) updateFromGetattr(ctx context.Context) error { + // Use d.handle.file, which represents a 9P fid that has been opened, in + // preference to d.file, which represents a 9P fid that has not. This may + // be significantly more efficient in some implementations. + var ( + file p9file + handleMuRLocked bool + ) + d.handleMu.RLock() + if !d.handle.file.isNil() { + file = d.handle.file + handleMuRLocked = true + } else { + file = d.file + d.handleMu.RUnlock() + } + _, attrMask, attr, err := file.getAttr(ctx, dentryAttrMask()) + if handleMuRLocked { + d.handleMu.RUnlock() + } + if err != nil { + return err + } + d.updateFromP9Attrs(attrMask, &attr) + return nil +} + +func (d *dentry) fileType() uint32 { + return atomic.LoadUint32(&d.mode) & linux.S_IFMT +} + +func (d *dentry) statTo(stat *linux.Statx) { + stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | linux.STATX_UID | linux.STATX_GID | linux.STATX_ATIME | linux.STATX_MTIME | linux.STATX_CTIME | linux.STATX_INO | linux.STATX_SIZE | linux.STATX_BLOCKS | linux.STATX_BTIME + stat.Blksize = atomic.LoadUint32(&d.blockSize) + stat.Nlink = 1 + if d.isDir() { + stat.Nlink = 2 + } + stat.UID = atomic.LoadUint32(&d.uid) + stat.GID = atomic.LoadUint32(&d.gid) + stat.Mode = uint16(atomic.LoadUint32(&d.mode)) + stat.Ino = d.ino + stat.Size = atomic.LoadUint64(&d.size) + // This is consistent with regularFileFD.Seek(), which treats regular files + // as having no holes. + stat.Blocks = (stat.Size + 511) / 512 + stat.Atime = statxTimestampFromDentry(atomic.LoadInt64(&d.atime)) + stat.Btime = statxTimestampFromDentry(atomic.LoadInt64(&d.btime)) + stat.Ctime = statxTimestampFromDentry(atomic.LoadInt64(&d.ctime)) + stat.Mtime = statxTimestampFromDentry(atomic.LoadInt64(&d.mtime)) + // TODO(jamieliu): device number +} + +func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *linux.Statx, mnt *vfs.Mount) error { + if stat.Mask == 0 { + return nil + } + if stat.Mask&^(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_SIZE) != 0 { + return syserror.EPERM + } + if err := vfs.CheckSetStat(creds, stat, uint16(atomic.LoadUint32(&d.mode))&^linux.S_IFMT, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))); err != nil { + return err + } + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + setLocalAtime := false + setLocalMtime := false + if d.fs.opts.interop != InteropModeShared { + // Timestamp updates will be handled locally. + setLocalAtime = stat.Mask&linux.STATX_ATIME != 0 + setLocalMtime = stat.Mask&linux.STATX_MTIME != 0 + stat.Mask &^= linux.STATX_ATIME | linux.STATX_MTIME + if !setLocalMtime && (stat.Mask&linux.STATX_SIZE != 0) { + // Truncate updates mtime. + setLocalMtime = true + stat.Mtime.Nsec = linux.UTIME_NOW + } + } + d.metadataMu.Lock() + defer d.metadataMu.Unlock() + if stat.Mask != 0 { + if err := d.file.setAttr(ctx, p9.SetAttrMask{ + Permissions: stat.Mask&linux.STATX_MODE != 0, + UID: stat.Mask&linux.STATX_UID != 0, + GID: stat.Mask&linux.STATX_GID != 0, + Size: stat.Mask&linux.STATX_SIZE != 0, + ATime: stat.Mask&linux.STATX_ATIME != 0, + MTime: stat.Mask&linux.STATX_MTIME != 0, + ATimeNotSystemTime: stat.Atime.Nsec != linux.UTIME_NOW, + MTimeNotSystemTime: stat.Mtime.Nsec != linux.UTIME_NOW, + }, p9.SetAttr{ + Permissions: p9.FileMode(stat.Mode), + UID: p9.UID(stat.UID), + GID: p9.GID(stat.GID), + Size: stat.Size, + ATimeSeconds: uint64(stat.Atime.Sec), + ATimeNanoSeconds: uint64(stat.Atime.Nsec), + MTimeSeconds: uint64(stat.Mtime.Sec), + MTimeNanoSeconds: uint64(stat.Mtime.Nsec), + }); err != nil { + return err + } + } + if d.fs.opts.interop == InteropModeShared { + // There's no point to updating d's metadata in this case since it'll + // be overwritten by revalidation before the next time it's used + // anyway. (InteropModeShared inhibits client caching of regular file + // data, so there's no cache to truncate either.) + return nil + } + now, haveNow := nowFromContext(ctx) + if !haveNow { + ctx.Warningf("gofer.dentry.setStat: current time not available") + } + if stat.Mask&linux.STATX_MODE != 0 { + atomic.StoreUint32(&d.mode, d.fileType()|uint32(stat.Mode)) + } + if stat.Mask&linux.STATX_UID != 0 { + atomic.StoreUint32(&d.uid, stat.UID) + } + if stat.Mask&linux.STATX_GID != 0 { + atomic.StoreUint32(&d.gid, stat.GID) + } + if setLocalAtime { + if stat.Atime.Nsec == linux.UTIME_NOW { + if haveNow { + atomic.StoreInt64(&d.atime, now) + } + } else { + atomic.StoreInt64(&d.atime, dentryTimestampFromStatx(stat.Atime)) + } + } + if setLocalMtime { + if stat.Mtime.Nsec == linux.UTIME_NOW { + if haveNow { + atomic.StoreInt64(&d.mtime, now) + } + } else { + atomic.StoreInt64(&d.mtime, dentryTimestampFromStatx(stat.Mtime)) + } + } + if haveNow { + atomic.StoreInt64(&d.ctime, now) + } + if stat.Mask&linux.STATX_SIZE != 0 { + d.dataMu.Lock() + oldSize := d.size + d.size = stat.Size + // d.dataMu must be unlocked to lock d.mapsMu and invalidate mappings + // below. This allows concurrent calls to Read/Translate/etc. These + // functions synchronize with truncation by refusing to use cache + // contents beyond the new d.size. (We are still holding d.metadataMu, + // so we can't race with Write or another truncate.) + d.dataMu.Unlock() + if d.size < oldSize { + oldpgend := pageRoundUp(oldSize) + newpgend := pageRoundUp(d.size) + if oldpgend != newpgend { + d.mapsMu.Lock() + d.mappings.Invalidate(memmap.MappableRange{newpgend, oldpgend}, memmap.InvalidateOpts{ + // Compare Linux's mm/truncate.c:truncate_setsize() => + // truncate_pagecache() => + // mm/memory.c:unmap_mapping_range(evencows=1). + InvalidatePrivate: true, + }) + d.mapsMu.Unlock() + } + // We are now guaranteed that there are no translations of + // truncated pages, and can remove them from the cache. Since + // truncated pages have been removed from the remote file, they + // should be dropped without being written back. + d.dataMu.Lock() + d.cache.Truncate(d.size, d.fs.mfp.MemoryFile()) + d.dirty.KeepClean(memmap.MappableRange{d.size, oldpgend}) + d.dataMu.Unlock() + } + } + return nil +} + +func (d *dentry) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes, isDir bool) error { + return vfs.GenericCheckPermissions(creds, ats, isDir, uint16(atomic.LoadUint32(&d.mode))&0777, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))) +} + +// IncRef implements vfs.DentryImpl.IncRef. +func (d *dentry) IncRef() { + // d.refs may be 0 if d.fs.renameMu is locked, which serializes against + // d.checkCachingLocked(). + atomic.AddInt64(&d.refs, 1) +} + +// TryIncRef implements vfs.DentryImpl.TryIncRef. +func (d *dentry) TryIncRef() bool { + for { + refs := atomic.LoadInt64(&d.refs) + if refs == 0 { + return false + } + if atomic.CompareAndSwapInt64(&d.refs, refs, refs+1) { + return true + } + } +} + +// DecRef implements vfs.DentryImpl.DecRef. +func (d *dentry) DecRef() { + if refs := atomic.AddInt64(&d.refs, -1); refs == 0 { + d.fs.renameMu.Lock() + d.checkCachingLocked() + d.fs.renameMu.Unlock() + } else if refs < 0 { + panic("gofer.dentry.DecRef() called without holding a reference") + } +} + +// checkCachingLocked should be called after d's reference count becomes 0 or it +// becomes disowned. +// +// Preconditions: d.fs.renameMu must be locked for writing. +func (d *dentry) checkCachingLocked() { + // Dentries with a non-zero reference count must be retained. (The only way + // to obtain a reference on a dentry with zero references is via path + // resolution, which requires renameMu, so if d.refs is zero then it will + // remain zero while we hold renameMu for writing.) + if atomic.LoadInt64(&d.refs) != 0 { + if d.cached { + d.fs.cachedDentries.Remove(d) + d.fs.cachedDentriesLen-- + d.cached = false + } + return + } + // Non-child dentries with zero references are no longer reachable by path + // resolution and should be dropped immediately. + if d.vfsd.Parent() == nil || d.vfsd.IsDisowned() { + if d.cached { + d.fs.cachedDentries.Remove(d) + d.fs.cachedDentriesLen-- + d.cached = false + } + d.destroyLocked() + return + } + // If d is already cached, just move it to the front of the LRU. + if d.cached { + d.fs.cachedDentries.Remove(d) + d.fs.cachedDentries.PushFront(d) + return + } + // Cache the dentry, then evict the least recently used cached dentry if + // the cache becomes over-full. + d.fs.cachedDentries.PushFront(d) + d.fs.cachedDentriesLen++ + d.cached = true + if d.fs.cachedDentriesLen > d.fs.opts.maxCachedDentries { + victim := d.fs.cachedDentries.Back() + d.fs.cachedDentries.Remove(victim) + d.fs.cachedDentriesLen-- + victim.cached = false + // victim.refs may have become non-zero from an earlier path + // resolution since it was inserted into fs.cachedDentries; see + // dentry.incRefLocked(). Either way, we brought + // fs.cachedDentriesLen back down to fs.opts.maxCachedDentries, so + // we don't loop. + if atomic.LoadInt64(&victim.refs) == 0 { + if victimParentVFSD := victim.vfsd.Parent(); victimParentVFSD != nil { + victimParent := victimParentVFSD.Impl().(*dentry) + victimParent.dirMu.Lock() + if !victim.vfsd.IsDisowned() { + // victim can't be a mount point (in any mount + // namespace), since VFS holds references on mount + // points. + d.fs.vfsfs.VirtualFilesystem().ForceDeleteDentry(&victim.vfsd) + // We're only deleting the dentry, not the file it + // represents, so we don't need to update + // victimParent.dirents etc. + } + victimParent.dirMu.Unlock() + } + victim.destroyLocked() + } + } +} + +// Preconditions: d.fs.renameMu must be locked for writing. d.refs == 0. d is +// not a child dentry. +func (d *dentry) destroyLocked() { + ctx := context.Background() + d.handleMu.Lock() + if !d.handle.file.isNil() { + mf := d.fs.mfp.MemoryFile() + d.dataMu.Lock() + // Write dirty pages back to the remote filesystem. + if d.handleWritable { + if err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, mf, d.handle.writeFromBlocksAt); err != nil { + log.Warningf("gofer.dentry.DecRef: failed to write dirty data back: %v", err) + } + } + // Discard cached data. + d.cache.DropAll(mf) + d.dirty.RemoveAll() + d.dataMu.Unlock() + // Clunk open fids and close open host FDs. + d.handle.close(ctx) + } + d.handleMu.Unlock() + d.file.close(ctx) + // Remove d from the set of all dentries. + d.fs.syncMu.Lock() + delete(d.fs.dentries, d) + d.fs.syncMu.Unlock() + // Drop the reference held by d on its parent. + if parentVFSD := d.vfsd.Parent(); parentVFSD != nil { + parent := parentVFSD.Impl().(*dentry) + // This is parent.DecRef() without recursive locking of d.fs.renameMu. + if refs := atomic.AddInt64(&parent.refs, -1); refs == 0 { + parent.checkCachingLocked() + } else if refs < 0 { + panic("gofer.dentry.DecRef() called without holding a reference") + } + } +} + +func (d *dentry) isDeleted() bool { + return atomic.LoadUint32(&d.deleted) != 0 +} + +func (d *dentry) setDeleted() { + atomic.StoreUint32(&d.deleted, 1) +} + +func (d *dentry) listxattr(ctx context.Context) ([]string, error) { + return nil, syserror.ENOTSUP +} + +func (d *dentry) getxattr(ctx context.Context, name string) (string, error) { + // TODO(jamieliu): add vfs.GetxattrOptions.Size + return d.file.getXattr(ctx, name, linux.XATTR_SIZE_MAX) +} + +func (d *dentry) setxattr(ctx context.Context, opts *vfs.SetxattrOptions) error { + return d.file.setXattr(ctx, opts.Name, opts.Value, opts.Flags) +} + +func (d *dentry) removexattr(ctx context.Context, name string) error { + return syserror.ENOTSUP +} + +// Preconditions: d.isRegularFile() || d.isDirectory(). +func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool) error { + // O_TRUNC unconditionally requires us to obtain a new handle (opened with + // O_TRUNC). + if !trunc { + d.handleMu.RLock() + if (!read || d.handleReadable) && (!write || d.handleWritable) { + // The current handle is sufficient. + d.handleMu.RUnlock() + return nil + } + d.handleMu.RUnlock() + } + + haveOldFD := false + d.handleMu.Lock() + if (read && !d.handleReadable) || (write && !d.handleWritable) || trunc { + // Get a new handle. + wantReadable := d.handleReadable || read + wantWritable := d.handleWritable || write + h, err := openHandle(ctx, d.file, wantReadable, wantWritable, trunc) + if err != nil { + d.handleMu.Unlock() + return err + } + if !d.handle.file.isNil() { + // Check that old and new handles are compatible: If the old handle + // includes a host file descriptor but the new one does not, or + // vice versa, old and new memory mappings may be incoherent. + haveOldFD = d.handle.fd >= 0 + haveNewFD := h.fd >= 0 + if haveOldFD != haveNewFD { + d.handleMu.Unlock() + ctx.Warningf("gofer.dentry.ensureSharedHandle: can't change host FD availability from %v to %v across dentry handle upgrade", haveOldFD, haveNewFD) + h.close(ctx) + return syserror.EIO + } + if haveOldFD { + // We may have raced with callers of d.pf.FD() that are now + // using the old file descriptor, preventing us from safely + // closing it. We could handle this by invalidating existing + // memmap.Translations, but this is expensive. Instead, use + // dup2() to make the old file descriptor refer to the new file + // description, then close the new file descriptor (which is no + // longer needed). Racing callers may use the old or new file + // description, but this doesn't matter since they refer to the + // same file (unless d.fs.opts.overlayfsStaleRead is true, + // which we handle separately). + if err := syscall.Dup2(int(h.fd), int(d.handle.fd)); err != nil { + d.handleMu.Unlock() + ctx.Warningf("gofer.dentry.ensureSharedHandle: failed to dup fd %d to fd %d: %v", h.fd, d.handle.fd, err) + h.close(ctx) + return err + } + syscall.Close(int(h.fd)) + h.fd = d.handle.fd + if d.fs.opts.overlayfsStaleRead { + // Replace sentry mappings of the old FD with mappings of + // the new FD, since the two are not necessarily coherent. + if err := d.pf.hostFileMapper.RegenerateMappings(int(h.fd)); err != nil { + d.handleMu.Unlock() + ctx.Warningf("gofer.dentry.ensureSharedHandle: failed to replace sentry mappings of old FD with mappings of new FD: %v", err) + h.close(ctx) + return err + } + } + // Clunk the old fid before making the new handle visible (by + // unlocking d.handleMu). + d.handle.file.close(ctx) + } + } + // Switch to the new handle. + d.handle = h + d.handleReadable = wantReadable + d.handleWritable = wantWritable + } + d.handleMu.Unlock() + + if d.fs.opts.overlayfsStaleRead && haveOldFD { + // Invalidate application mappings that may be using the old FD; they + // will be replaced with mappings using the new FD after future calls + // to d.Translate(). This requires holding d.mapsMu, which precedes + // d.handleMu in the lock order. + d.mapsMu.Lock() + d.mappings.InvalidateAll(memmap.InvalidateOpts{}) + d.mapsMu.Unlock() + } + + return nil +} + +// fileDescription is embedded by gofer implementations of +// vfs.FileDescriptionImpl. +type fileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl +} + +func (fd *fileDescription) filesystem() *filesystem { + return fd.vfsfd.Mount().Filesystem().Impl().(*filesystem) +} + +func (fd *fileDescription) dentry() *dentry { + return fd.vfsfd.Dentry().Impl().(*dentry) +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + d := fd.dentry() + if d.fs.opts.interop == InteropModeShared && opts.Mask&(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_CTIME|linux.STATX_SIZE|linux.STATX_BLOCKS|linux.STATX_BTIME) != 0 && opts.Sync != linux.AT_STATX_DONT_SYNC { + // TODO(jamieliu): Use specialFileFD.handle.file for the getattr if + // available? + if err := d.updateFromGetattr(ctx); err != nil { + return linux.Statx{}, err + } + } + var stat linux.Statx + d.statTo(&stat) + return stat, nil +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + return fd.dentry().setStat(ctx, auth.CredentialsFromContext(ctx), &opts.Stat, fd.vfsfd.Mount()) +} + +// Listxattr implements vfs.FileDescriptionImpl.Listxattr. +func (fd *fileDescription) Listxattr(ctx context.Context) ([]string, error) { + return fd.dentry().listxattr(ctx) +} + +// Getxattr implements vfs.FileDescriptionImpl.Getxattr. +func (fd *fileDescription) Getxattr(ctx context.Context, name string) (string, error) { + return fd.dentry().getxattr(ctx, name) +} + +// Setxattr implements vfs.FileDescriptionImpl.Setxattr. +func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOptions) error { + return fd.dentry().setxattr(ctx, &opts) +} + +// Removexattr implements vfs.FileDescriptionImpl.Removexattr. +func (fd *fileDescription) Removexattr(ctx context.Context, name string) error { + return fd.dentry().removexattr(ctx, name) +} diff --git a/pkg/sentry/fsimpl/gofer/handle.go b/pkg/sentry/fsimpl/gofer/handle.go new file mode 100644 index 000000000..cfe66f797 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/handle.go @@ -0,0 +1,135 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "syscall" + + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/safemem" +) + +// handle represents a remote "open file descriptor", consisting of an opened +// fid (p9.File) and optionally a host file descriptor. +type handle struct { + file p9file + fd int32 // -1 if unavailable +} + +// Preconditions: read || write. +func openHandle(ctx context.Context, file p9file, read, write, trunc bool) (handle, error) { + _, newfile, err := file.walk(ctx, nil) + if err != nil { + return handle{fd: -1}, err + } + var flags p9.OpenFlags + switch { + case read && !write: + flags = p9.ReadOnly + case !read && write: + flags = p9.WriteOnly + case read && write: + flags = p9.ReadWrite + } + if trunc { + flags |= p9.OpenTruncate + } + fdobj, _, _, err := newfile.open(ctx, flags) + if err != nil { + newfile.close(ctx) + return handle{fd: -1}, err + } + fd := int32(-1) + if fdobj != nil { + fd = int32(fdobj.Release()) + } + return handle{ + file: newfile, + fd: fd, + }, nil +} + +func (h *handle) close(ctx context.Context) { + h.file.close(ctx) + h.file = p9file{} + if h.fd >= 0 { + syscall.Close(int(h.fd)) + h.fd = -1 + } +} + +func (h *handle) readToBlocksAt(ctx context.Context, dsts safemem.BlockSeq, offset uint64) (uint64, error) { + if dsts.IsEmpty() { + return 0, nil + } + if h.fd >= 0 { + ctx.UninterruptibleSleepStart(false) + n, err := hostPreadv(h.fd, dsts, int64(offset)) + ctx.UninterruptibleSleepFinish(false) + return n, err + } + if dsts.NumBlocks() == 1 && !dsts.Head().NeedSafecopy() { + n, err := h.file.readAt(ctx, dsts.Head().ToSlice(), offset) + return uint64(n), err + } + // Buffer the read since p9.File.ReadAt() takes []byte. + buf := make([]byte, dsts.NumBytes()) + n, err := h.file.readAt(ctx, buf, offset) + if n == 0 { + return 0, err + } + if cp, cperr := safemem.CopySeq(dsts, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf[:n]))); cperr != nil { + return cp, cperr + } + return uint64(n), err +} + +func (h *handle) writeFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error) { + if srcs.IsEmpty() { + return 0, nil + } + if h.fd >= 0 { + ctx.UninterruptibleSleepStart(false) + n, err := hostPwritev(h.fd, srcs, int64(offset)) + ctx.UninterruptibleSleepFinish(false) + return n, err + } + if srcs.NumBlocks() == 1 && !srcs.Head().NeedSafecopy() { + n, err := h.file.writeAt(ctx, srcs.Head().ToSlice(), offset) + return uint64(n), err + } + // Buffer the write since p9.File.WriteAt() takes []byte. + buf := make([]byte, srcs.NumBytes()) + cp, cperr := safemem.CopySeq(safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), srcs) + if cp == 0 { + return 0, cperr + } + n, err := h.file.writeAt(ctx, buf[:cp], offset) + if err != nil { + return uint64(n), err + } + return cp, cperr +} + +func (h *handle) sync(ctx context.Context) error { + if h.fd >= 0 { + ctx.UninterruptibleSleepStart(false) + err := syscall.Fsync(int(h.fd)) + ctx.UninterruptibleSleepFinish(false) + return err + } + return h.file.fsync(ctx) +} diff --git a/pkg/sentry/fsimpl/gofer/handle_unsafe.go b/pkg/sentry/fsimpl/gofer/handle_unsafe.go new file mode 100644 index 000000000..19560ab26 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/handle_unsafe.go @@ -0,0 +1,66 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "syscall" + "unsafe" + + "gvisor.dev/gvisor/pkg/safemem" +) + +// Preconditions: !dsts.IsEmpty(). +func hostPreadv(fd int32, dsts safemem.BlockSeq, off int64) (uint64, error) { + // No buffering is necessary regardless of safecopy; host syscalls will + // return EFAULT if appropriate, instead of raising SIGBUS. + if dsts.NumBlocks() == 1 { + // Use pread() instead of preadv() to avoid iovec allocation and + // copying. + dst := dsts.Head() + n, _, e := syscall.Syscall6(syscall.SYS_PREAD64, uintptr(fd), dst.Addr(), uintptr(dst.Len()), uintptr(off), 0, 0) + if e != 0 { + return 0, e + } + return uint64(n), nil + } + iovs := safemem.IovecsFromBlockSeq(dsts) + n, _, e := syscall.Syscall6(syscall.SYS_PREADV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(off), 0, 0) + if e != 0 { + return 0, e + } + return uint64(n), nil +} + +// Preconditions: !srcs.IsEmpty(). +func hostPwritev(fd int32, srcs safemem.BlockSeq, off int64) (uint64, error) { + // No buffering is necessary regardless of safecopy; host syscalls will + // return EFAULT if appropriate, instead of raising SIGBUS. + if srcs.NumBlocks() == 1 { + // Use pwrite() instead of pwritev() to avoid iovec allocation and + // copying. + src := srcs.Head() + n, _, e := syscall.Syscall6(syscall.SYS_PWRITE64, uintptr(fd), src.Addr(), uintptr(src.Len()), uintptr(off), 0, 0) + if e != 0 { + return 0, e + } + return uint64(n), nil + } + iovs := safemem.IovecsFromBlockSeq(srcs) + n, _, e := syscall.Syscall6(syscall.SYS_PWRITEV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(off), 0, 0) + if e != 0 { + return 0, e + } + return uint64(n), nil +} diff --git a/pkg/sentry/fsimpl/gofer/p9file.go b/pkg/sentry/fsimpl/gofer/p9file.go new file mode 100644 index 000000000..755ac2985 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/p9file.go @@ -0,0 +1,219 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fd" + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/syserror" +) + +// p9file is a wrapper around p9.File that provides methods that are +// Context-aware. +type p9file struct { + file p9.File +} + +func (f p9file) isNil() bool { + return f.file == nil +} + +func (f p9file) walk(ctx context.Context, names []string) ([]p9.QID, p9file, error) { + ctx.UninterruptibleSleepStart(false) + qids, newfile, err := f.file.Walk(names) + ctx.UninterruptibleSleepFinish(false) + return qids, p9file{newfile}, err +} + +func (f p9file) walkGetAttr(ctx context.Context, names []string) ([]p9.QID, p9file, p9.AttrMask, p9.Attr, error) { + ctx.UninterruptibleSleepStart(false) + qids, newfile, attrMask, attr, err := f.file.WalkGetAttr(names) + ctx.UninterruptibleSleepFinish(false) + return qids, p9file{newfile}, attrMask, attr, err +} + +// walkGetAttrOne is a wrapper around p9.File.WalkGetAttr that takes a single +// path component and returns a single qid. +func (f p9file) walkGetAttrOne(ctx context.Context, name string) (p9.QID, p9file, p9.AttrMask, p9.Attr, error) { + ctx.UninterruptibleSleepStart(false) + qids, newfile, attrMask, attr, err := f.file.WalkGetAttr([]string{name}) + ctx.UninterruptibleSleepFinish(false) + if err != nil { + return p9.QID{}, p9file{}, p9.AttrMask{}, p9.Attr{}, err + } + if len(qids) != 1 { + ctx.Warningf("p9.File.WalkGetAttr returned %d qids (%v), wanted 1", len(qids), qids) + if newfile != nil { + p9file{newfile}.close(ctx) + } + return p9.QID{}, p9file{}, p9.AttrMask{}, p9.Attr{}, syserror.EIO + } + return qids[0], p9file{newfile}, attrMask, attr, nil +} + +func (f p9file) statFS(ctx context.Context) (p9.FSStat, error) { + ctx.UninterruptibleSleepStart(false) + fsstat, err := f.file.StatFS() + ctx.UninterruptibleSleepFinish(false) + return fsstat, err +} + +func (f p9file) getAttr(ctx context.Context, req p9.AttrMask) (p9.QID, p9.AttrMask, p9.Attr, error) { + ctx.UninterruptibleSleepStart(false) + qid, attrMask, attr, err := f.file.GetAttr(req) + ctx.UninterruptibleSleepFinish(false) + return qid, attrMask, attr, err +} + +func (f p9file) setAttr(ctx context.Context, valid p9.SetAttrMask, attr p9.SetAttr) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.SetAttr(valid, attr) + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) getXattr(ctx context.Context, name string, size uint64) (string, error) { + ctx.UninterruptibleSleepStart(false) + val, err := f.file.GetXattr(name, size) + ctx.UninterruptibleSleepFinish(false) + return val, err +} + +func (f p9file) setXattr(ctx context.Context, name, value string, flags uint32) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.SetXattr(name, value, flags) + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) allocate(ctx context.Context, mode p9.AllocateMode, offset, length uint64) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.Allocate(mode, offset, length) + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) close(ctx context.Context) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.Close() + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) open(ctx context.Context, flags p9.OpenFlags) (*fd.FD, p9.QID, uint32, error) { + ctx.UninterruptibleSleepStart(false) + fdobj, qid, iounit, err := f.file.Open(flags) + ctx.UninterruptibleSleepFinish(false) + return fdobj, qid, iounit, err +} + +func (f p9file) readAt(ctx context.Context, p []byte, offset uint64) (int, error) { + ctx.UninterruptibleSleepStart(false) + n, err := f.file.ReadAt(p, offset) + ctx.UninterruptibleSleepFinish(false) + return n, err +} + +func (f p9file) writeAt(ctx context.Context, p []byte, offset uint64) (int, error) { + ctx.UninterruptibleSleepStart(false) + n, err := f.file.WriteAt(p, offset) + ctx.UninterruptibleSleepFinish(false) + return n, err +} + +func (f p9file) fsync(ctx context.Context) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.FSync() + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) create(ctx context.Context, name string, flags p9.OpenFlags, permissions p9.FileMode, uid p9.UID, gid p9.GID) (*fd.FD, p9file, p9.QID, uint32, error) { + ctx.UninterruptibleSleepStart(false) + fdobj, newfile, qid, iounit, err := f.file.Create(name, flags, permissions, uid, gid) + ctx.UninterruptibleSleepFinish(false) + return fdobj, p9file{newfile}, qid, iounit, err +} + +func (f p9file) mkdir(ctx context.Context, name string, permissions p9.FileMode, uid p9.UID, gid p9.GID) (p9.QID, error) { + ctx.UninterruptibleSleepStart(false) + qid, err := f.file.Mkdir(name, permissions, uid, gid) + ctx.UninterruptibleSleepFinish(false) + return qid, err +} + +func (f p9file) symlink(ctx context.Context, oldName string, newName string, uid p9.UID, gid p9.GID) (p9.QID, error) { + ctx.UninterruptibleSleepStart(false) + qid, err := f.file.Symlink(oldName, newName, uid, gid) + ctx.UninterruptibleSleepFinish(false) + return qid, err +} + +func (f p9file) link(ctx context.Context, target p9file, newName string) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.Link(target.file, newName) + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) mknod(ctx context.Context, name string, mode p9.FileMode, major uint32, minor uint32, uid p9.UID, gid p9.GID) (p9.QID, error) { + ctx.UninterruptibleSleepStart(false) + qid, err := f.file.Mknod(name, mode, major, minor, uid, gid) + ctx.UninterruptibleSleepFinish(false) + return qid, err +} + +func (f p9file) rename(ctx context.Context, newDir p9file, newName string) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.Rename(newDir.file, newName) + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) unlinkAt(ctx context.Context, name string, flags uint32) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.UnlinkAt(name, flags) + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) readdir(ctx context.Context, offset uint64, count uint32) ([]p9.Dirent, error) { + ctx.UninterruptibleSleepStart(false) + dirents, err := f.file.Readdir(offset, count) + ctx.UninterruptibleSleepFinish(false) + return dirents, err +} + +func (f p9file) readlink(ctx context.Context) (string, error) { + ctx.UninterruptibleSleepStart(false) + target, err := f.file.Readlink() + ctx.UninterruptibleSleepFinish(false) + return target, err +} + +func (f p9file) flush(ctx context.Context) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.Flush() + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) connect(ctx context.Context, flags p9.ConnectFlags) (*fd.FD, error) { + ctx.UninterruptibleSleepStart(false) + fdobj, err := f.file.Connect(flags) + ctx.UninterruptibleSleepFinish(false) + return fdobj, err +} diff --git a/pkg/sentry/fsimpl/gofer/pagemath.go b/pkg/sentry/fsimpl/gofer/pagemath.go new file mode 100644 index 000000000..847cb0784 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/pagemath.go @@ -0,0 +1,31 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "gvisor.dev/gvisor/pkg/usermem" +) + +// This are equivalent to usermem.Addr.RoundDown/Up, but without the +// potentially truncating conversion to usermem.Addr. This is necessary because +// there is no way to define generic "PageRoundDown/Up" functions in Go. + +func pageRoundDown(x uint64) uint64 { + return x &^ (usermem.PageSize - 1) +} + +func pageRoundUp(x uint64) uint64 { + return pageRoundDown(x + usermem.PageSize - 1) +} diff --git a/pkg/sentry/fsimpl/gofer/regular_file.go b/pkg/sentry/fsimpl/gofer/regular_file.go new file mode 100644 index 000000000..8e11e06b3 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/regular_file.go @@ -0,0 +1,860 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "fmt" + "io" + "math" + "sync" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/pgalloc" + "gvisor.dev/gvisor/pkg/sentry/platform" + "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +func (d *dentry) isRegularFile() bool { + return d.fileType() == linux.S_IFREG +} + +type regularFileFD struct { + fileDescription + + // off is the file offset. off is protected by mu. + mu sync.Mutex + off int64 +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (fd *regularFileFD) Release() { +} + +// OnClose implements vfs.FileDescriptionImpl.OnClose. +func (fd *regularFileFD) OnClose(ctx context.Context) error { + if !fd.vfsfd.IsWritable() { + return nil + } + // Skip flushing if writes may be buffered by the client, since (as with + // the VFS1 client) we don't flush buffered writes on close anyway. + d := fd.dentry() + if d.fs.opts.interop == InteropModeExclusive { + return nil + } + d.handleMu.RLock() + defer d.handleMu.RUnlock() + return d.handle.file.flush(ctx) +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + if offset < 0 { + return 0, syserror.EINVAL + } + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + // Check for reading at EOF before calling into MM (but not under + // InteropModeShared, which makes d.size unreliable). + d := fd.dentry() + if d.fs.opts.interop != InteropModeShared && uint64(offset) >= atomic.LoadUint64(&d.size) { + return 0, io.EOF + } + + if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 { + // Lock d.metadataMu for the rest of the read to prevent d.size from + // changing. + d.metadataMu.Lock() + defer d.metadataMu.Unlock() + // Write dirty cached pages that will be touched by the read back to + // the remote file. + if err := d.writeback(ctx, offset, dst.NumBytes()); err != nil { + return 0, err + } + } + + rw := getDentryReadWriter(ctx, d, offset) + if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 { + // Require the read to go to the remote file. + rw.direct = true + } + n, err := dst.CopyOutFrom(ctx, rw) + putDentryReadWriter(rw) + if d.fs.opts.interop != InteropModeShared { + // Compare Linux's mm/filemap.c:do_generic_file_read() => file_accessed(). + d.touchAtime(ctx, fd.vfsfd.Mount()) + } + return n, err +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (fd *regularFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + fd.mu.Lock() + n, err := fd.PRead(ctx, dst, fd.off, opts) + fd.off += n + fd.mu.Unlock() + return n, err +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + if offset < 0 { + return 0, syserror.EINVAL + } + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + d := fd.dentry() + d.metadataMu.Lock() + defer d.metadataMu.Unlock() + if d.fs.opts.interop != InteropModeShared { + // Compare Linux's mm/filemap.c:__generic_file_write_iter() => + // file_update_time(). This is d.touchCMtime(), but without locking + // d.metadataMu (recursively). + if now, ok := nowFromContext(ctx); ok { + atomic.StoreInt64(&d.mtime, now) + atomic.StoreInt64(&d.ctime, now) + } + } + if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 { + // Write dirty cached pages that will be touched by the write back to + // the remote file. + if err := d.writeback(ctx, offset, src.NumBytes()); err != nil { + return 0, err + } + // Remove touched pages from the cache. + pgstart := pageRoundDown(uint64(offset)) + pgend := pageRoundUp(uint64(offset + src.NumBytes())) + if pgend < pgstart { + return 0, syserror.EINVAL + } + mr := memmap.MappableRange{pgstart, pgend} + var freed []platform.FileRange + d.dataMu.Lock() + cseg := d.cache.LowerBoundSegment(mr.Start) + for cseg.Ok() && cseg.Start() < mr.End { + cseg = d.cache.Isolate(cseg, mr) + freed = append(freed, platform.FileRange{cseg.Value(), cseg.Value() + cseg.Range().Length()}) + cseg = d.cache.Remove(cseg).NextSegment() + } + d.dataMu.Unlock() + // Invalidate mappings of removed pages. + d.mapsMu.Lock() + d.mappings.Invalidate(mr, memmap.InvalidateOpts{}) + d.mapsMu.Unlock() + // Finally free pages removed from the cache. + mf := d.fs.mfp.MemoryFile() + for _, freedFR := range freed { + mf.DecRef(freedFR) + } + } + rw := getDentryReadWriter(ctx, d, offset) + if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 { + // Require the write to go to the remote file. + rw.direct = true + } + n, err := src.CopyInTo(ctx, rw) + putDentryReadWriter(rw) + if n != 0 && fd.vfsfd.StatusFlags()&(linux.O_DSYNC|linux.O_SYNC) != 0 { + // Write dirty cached pages touched by the write back to the remote + // file. + if err := d.writeback(ctx, offset, src.NumBytes()); err != nil { + return 0, err + } + // Request the remote filesystem to sync the remote file. + if err := d.handle.file.fsync(ctx); err != nil { + return 0, err + } + } + return n, err +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *regularFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + fd.mu.Lock() + n, err := fd.PWrite(ctx, src, fd.off, opts) + fd.off += n + fd.mu.Unlock() + return n, err +} + +type dentryReadWriter struct { + ctx context.Context + d *dentry + off uint64 + direct bool +} + +var dentryReadWriterPool = sync.Pool{ + New: func() interface{} { + return &dentryReadWriter{} + }, +} + +func getDentryReadWriter(ctx context.Context, d *dentry, offset int64) *dentryReadWriter { + rw := dentryReadWriterPool.Get().(*dentryReadWriter) + rw.ctx = ctx + rw.d = d + rw.off = uint64(offset) + rw.direct = false + return rw +} + +func putDentryReadWriter(rw *dentryReadWriter) { + rw.ctx = nil + rw.d = nil + dentryReadWriterPool.Put(rw) +} + +// ReadToBlocks implements safemem.Reader.ReadToBlocks. +func (rw *dentryReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { + if dsts.IsEmpty() { + return 0, nil + } + + // If we have a mmappable host FD (which must be used here to ensure + // coherence with memory-mapped I/O), or if InteropModeShared is in effect + // (which prevents us from caching file contents and makes dentry.size + // unreliable), or if the file was opened O_DIRECT, read directly from + // dentry.handle without locking dentry.dataMu. + rw.d.handleMu.RLock() + if (rw.d.handle.fd >= 0 && !rw.d.fs.opts.forcePageCache) || rw.d.fs.opts.interop == InteropModeShared || rw.direct { + n, err := rw.d.handle.readToBlocksAt(rw.ctx, dsts, rw.off) + rw.d.handleMu.RUnlock() + rw.off += n + return n, err + } + + // Otherwise read from/through the cache. + mf := rw.d.fs.mfp.MemoryFile() + fillCache := mf.ShouldCacheEvictable() + var dataMuUnlock func() + if fillCache { + rw.d.dataMu.Lock() + dataMuUnlock = rw.d.dataMu.Unlock + } else { + rw.d.dataMu.RLock() + dataMuUnlock = rw.d.dataMu.RUnlock + } + + // Compute the range to read (limited by file size and overflow-checked). + if rw.off >= rw.d.size { + dataMuUnlock() + rw.d.handleMu.RUnlock() + return 0, io.EOF + } + end := rw.d.size + if rend := rw.off + dsts.NumBytes(); rend > rw.off && rend < end { + end = rend + } + + var done uint64 + seg, gap := rw.d.cache.Find(rw.off) + for rw.off < end { + mr := memmap.MappableRange{rw.off, end} + switch { + case seg.Ok(): + // Get internal mappings from the cache. + ims, err := mf.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Read) + if err != nil { + dataMuUnlock() + rw.d.handleMu.RUnlock() + return done, err + } + + // Copy from internal mappings. + n, err := safemem.CopySeq(dsts, ims) + done += n + rw.off += n + dsts = dsts.DropFirst64(n) + if err != nil { + dataMuUnlock() + rw.d.handleMu.RUnlock() + return done, err + } + + // Continue. + seg, gap = seg.NextNonEmpty() + + case gap.Ok(): + gapMR := gap.Range().Intersect(mr) + if fillCache { + // Read into the cache, then re-enter the loop to read from the + // cache. + reqMR := memmap.MappableRange{ + Start: pageRoundDown(gapMR.Start), + End: pageRoundUp(gapMR.End), + } + optMR := gap.Range() + err := rw.d.cache.Fill(rw.ctx, reqMR, maxFillRange(reqMR, optMR), mf, usage.PageCache, rw.d.handle.readToBlocksAt) + mf.MarkEvictable(rw.d, pgalloc.EvictableRange{optMR.Start, optMR.End}) + seg, gap = rw.d.cache.Find(rw.off) + if !seg.Ok() { + dataMuUnlock() + rw.d.handleMu.RUnlock() + return done, err + } + // err might have occurred in part of gap.Range() outside + // gapMR. Forget about it for now; if the error matters and + // persists, we'll run into it again in a later iteration of + // this loop. + } else { + // Read directly from the file. + gapDsts := dsts.TakeFirst64(gapMR.Length()) + n, err := rw.d.handle.readToBlocksAt(rw.ctx, gapDsts, gapMR.Start) + done += n + rw.off += n + dsts = dsts.DropFirst64(n) + // Partial reads are fine. But we must stop reading. + if n != gapDsts.NumBytes() || err != nil { + dataMuUnlock() + rw.d.handleMu.RUnlock() + return done, err + } + + // Continue. + seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{} + } + } + } + dataMuUnlock() + rw.d.handleMu.RUnlock() + return done, nil +} + +// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. +// +// Preconditions: rw.d.metadataMu must be locked. +func (rw *dentryReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { + if srcs.IsEmpty() { + return 0, nil + } + + // If we have a mmappable host FD (which must be used here to ensure + // coherence with memory-mapped I/O), or if InteropModeShared is in effect + // (which prevents us from caching file contents), or if the file was + // opened with O_DIRECT, write directly to dentry.handle without locking + // dentry.dataMu. + rw.d.handleMu.RLock() + if (rw.d.handle.fd >= 0 && !rw.d.fs.opts.forcePageCache) || rw.d.fs.opts.interop == InteropModeShared || rw.direct { + n, err := rw.d.handle.writeFromBlocksAt(rw.ctx, srcs, rw.off) + rw.d.handleMu.RUnlock() + rw.off += n + return n, err + } + + // Otherwise write to/through the cache. + mf := rw.d.fs.mfp.MemoryFile() + rw.d.dataMu.Lock() + + // Compute the range to write (overflow-checked). + start := rw.off + end := rw.off + srcs.NumBytes() + if end <= rw.off { + end = math.MaxInt64 + } + + var ( + done uint64 + retErr error + ) + seg, gap := rw.d.cache.Find(rw.off) + for rw.off < end { + mr := memmap.MappableRange{rw.off, end} + switch { + case seg.Ok(): + // Get internal mappings from the cache. + segMR := seg.Range().Intersect(mr) + ims, err := mf.MapInternal(seg.FileRangeOf(segMR), usermem.Write) + if err != nil { + retErr = err + goto exitLoop + } + + // Copy to internal mappings. + n, err := safemem.CopySeq(ims, srcs) + done += n + rw.off += n + srcs = srcs.DropFirst64(n) + rw.d.dirty.MarkDirty(segMR) + if err != nil { + retErr = err + goto exitLoop + } + + // Continue. + seg, gap = seg.NextNonEmpty() + + case gap.Ok(): + // Write directly to the file. At present, we never fill the cache + // when writing, since doing so can convert small writes into + // inefficient read-modify-write cycles, and we have no mechanism + // for detecting or avoiding this. + gapMR := gap.Range().Intersect(mr) + gapSrcs := srcs.TakeFirst64(gapMR.Length()) + n, err := rw.d.handle.writeFromBlocksAt(rw.ctx, gapSrcs, gapMR.Start) + done += n + rw.off += n + srcs = srcs.DropFirst64(n) + // Partial writes are fine. But we must stop writing. + if n != gapSrcs.NumBytes() || err != nil { + retErr = err + goto exitLoop + } + + // Continue. + seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{} + } + } +exitLoop: + if rw.off > rw.d.size { + atomic.StoreUint64(&rw.d.size, rw.off) + // The remote file's size will implicitly be extended to the correct + // value when we write back to it. + } + // If InteropModeWritethrough is in effect, flush written data back to the + // remote filesystem. + if rw.d.fs.opts.interop == InteropModeWritethrough && done != 0 { + if err := fsutil.SyncDirty(rw.ctx, memmap.MappableRange{ + Start: start, + End: rw.off, + }, &rw.d.cache, &rw.d.dirty, rw.d.size, mf, rw.d.handle.writeFromBlocksAt); err != nil { + // We have no idea how many bytes were actually flushed. + rw.off = start + done = 0 + retErr = err + } + } + rw.d.dataMu.Unlock() + rw.d.handleMu.RUnlock() + return done, retErr +} + +func (d *dentry) writeback(ctx context.Context, offset, size int64) error { + if size == 0 { + return nil + } + d.handleMu.RLock() + defer d.handleMu.RUnlock() + d.dataMu.Lock() + defer d.dataMu.Unlock() + // Compute the range of valid bytes (overflow-checked). + if uint64(offset) >= d.size { + return nil + } + end := int64(d.size) + if rend := offset + size; rend > offset && rend < end { + end = rend + } + return fsutil.SyncDirty(ctx, memmap.MappableRange{ + Start: uint64(offset), + End: uint64(end), + }, &d.cache, &d.dirty, d.size, d.fs.mfp.MemoryFile(), d.handle.writeFromBlocksAt) +} + +// Seek implements vfs.FileDescriptionImpl.Seek. +func (fd *regularFileFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + fd.mu.Lock() + defer fd.mu.Unlock() + switch whence { + case linux.SEEK_SET: + // Use offset as specified. + case linux.SEEK_CUR: + offset += fd.off + case linux.SEEK_END, linux.SEEK_DATA, linux.SEEK_HOLE: + // Ensure file size is up to date. + d := fd.dentry() + if fd.filesystem().opts.interop == InteropModeShared { + if err := d.updateFromGetattr(ctx); err != nil { + return 0, err + } + } + size := int64(atomic.LoadUint64(&d.size)) + // For SEEK_DATA and SEEK_HOLE, treat the file as a single contiguous + // block of data. + switch whence { + case linux.SEEK_END: + offset += size + case linux.SEEK_DATA: + if offset > size { + return 0, syserror.ENXIO + } + // Use offset as specified. + case linux.SEEK_HOLE: + if offset > size { + return 0, syserror.ENXIO + } + offset = size + } + default: + return 0, syserror.EINVAL + } + if offset < 0 { + return 0, syserror.EINVAL + } + fd.off = offset + return offset, nil +} + +// Sync implements vfs.FileDescriptionImpl.Sync. +func (fd *regularFileFD) Sync(ctx context.Context) error { + return fd.dentry().syncSharedHandle(ctx) +} + +func (d *dentry) syncSharedHandle(ctx context.Context) error { + d.handleMu.RLock() + if !d.handleWritable { + d.handleMu.RUnlock() + return nil + } + d.dataMu.Lock() + // Write dirty cached data to the remote file. + err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, d.fs.mfp.MemoryFile(), d.handle.writeFromBlocksAt) + d.dataMu.Unlock() + if err == nil { + // Sync the remote file. + err = d.handle.sync(ctx) + } + d.handleMu.RUnlock() + return err +} + +// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. +func (fd *regularFileFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + d := fd.dentry() + switch d.fs.opts.interop { + case InteropModeExclusive: + // Any mapping is fine. + case InteropModeWritethrough: + // Shared writable mappings require a host FD, since otherwise we can't + // synchronously flush memory-mapped writes to the remote file. + if opts.Private || !opts.MaxPerms.Write { + break + } + fallthrough + case InteropModeShared: + // All mappings require a host FD to be coherent with other filesystem + // users. + if d.fs.opts.forcePageCache { + // Whether or not we have a host FD, we're not allowed to use it. + return syserror.ENODEV + } + d.handleMu.RLock() + haveFD := d.handle.fd >= 0 + d.handleMu.RUnlock() + if !haveFD { + return syserror.ENODEV + } + default: + panic(fmt.Sprintf("unknown InteropMode %v", d.fs.opts.interop)) + } + return vfs.GenericConfigureMMap(&fd.vfsfd, d, opts) +} + +func (d *dentry) mayCachePages() bool { + if d.fs.opts.interop == InteropModeShared { + return false + } + if d.fs.opts.forcePageCache { + return true + } + d.handleMu.RLock() + haveFD := d.handle.fd >= 0 + d.handleMu.RUnlock() + return haveFD +} + +// AddMapping implements memmap.Mappable.AddMapping. +func (d *dentry) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) error { + d.mapsMu.Lock() + mapped := d.mappings.AddMapping(ms, ar, offset, writable) + // Do this unconditionally since whether we have a host FD can change + // across save/restore. + for _, r := range mapped { + d.pf.hostFileMapper.IncRefOn(r) + } + if d.mayCachePages() { + // d.Evict() will refuse to evict memory-mapped pages, so tell the + // MemoryFile to not bother trying. + mf := d.fs.mfp.MemoryFile() + for _, r := range mapped { + mf.MarkUnevictable(d, pgalloc.EvictableRange{r.Start, r.End}) + } + } + d.mapsMu.Unlock() + return nil +} + +// RemoveMapping implements memmap.Mappable.RemoveMapping. +func (d *dentry) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) { + d.mapsMu.Lock() + unmapped := d.mappings.RemoveMapping(ms, ar, offset, writable) + for _, r := range unmapped { + d.pf.hostFileMapper.DecRefOn(r) + } + if d.mayCachePages() { + // Pages that are no longer referenced by any application memory + // mappings are now considered unused; allow MemoryFile to evict them + // when necessary. + mf := d.fs.mfp.MemoryFile() + d.dataMu.Lock() + for _, r := range unmapped { + // Since these pages are no longer mapped, they are no longer + // concurrently dirtyable by a writable memory mapping. + d.dirty.AllowClean(r) + mf.MarkEvictable(d, pgalloc.EvictableRange{r.Start, r.End}) + } + d.dataMu.Unlock() + } + d.mapsMu.Unlock() +} + +// CopyMapping implements memmap.Mappable.CopyMapping. +func (d *dentry) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64, writable bool) error { + return d.AddMapping(ctx, ms, dstAR, offset, writable) +} + +// Translate implements memmap.Mappable.Translate. +func (d *dentry) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) { + d.handleMu.RLock() + if d.handle.fd >= 0 && !d.fs.opts.forcePageCache { + d.handleMu.RUnlock() + mr := optional + if d.fs.opts.limitHostFDTranslation { + mr = maxFillRange(required, optional) + } + return []memmap.Translation{ + { + Source: mr, + File: &d.pf, + Offset: mr.Start, + Perms: usermem.AnyAccess, + }, + }, nil + } + + d.dataMu.Lock() + + // Constrain translations to d.size (rounded up) to prevent translation to + // pages that may be concurrently truncated. + pgend := pageRoundUp(d.size) + var beyondEOF bool + if required.End > pgend { + if required.Start >= pgend { + d.dataMu.Unlock() + d.handleMu.RUnlock() + return nil, &memmap.BusError{io.EOF} + } + beyondEOF = true + required.End = pgend + } + if optional.End > pgend { + optional.End = pgend + } + + mf := d.fs.mfp.MemoryFile() + cerr := d.cache.Fill(ctx, required, maxFillRange(required, optional), mf, usage.PageCache, d.handle.readToBlocksAt) + + var ts []memmap.Translation + var translatedEnd uint64 + for seg := d.cache.FindSegment(required.Start); seg.Ok() && seg.Start() < required.End; seg, _ = seg.NextNonEmpty() { + segMR := seg.Range().Intersect(optional) + // TODO(jamieliu): Make Translations writable even if writability is + // not required if already kept-dirty by another writable translation. + perms := usermem.AccessType{ + Read: true, + Execute: true, + } + if at.Write { + // From this point forward, this memory can be dirtied through the + // mapping at any time. + d.dirty.KeepDirty(segMR) + perms.Write = true + } + ts = append(ts, memmap.Translation{ + Source: segMR, + File: mf, + Offset: seg.FileRangeOf(segMR).Start, + Perms: perms, + }) + translatedEnd = segMR.End + } + + d.dataMu.Unlock() + d.handleMu.RUnlock() + + // Don't return the error returned by c.cache.Fill if it occurred outside + // of required. + if translatedEnd < required.End && cerr != nil { + return ts, &memmap.BusError{cerr} + } + if beyondEOF { + return ts, &memmap.BusError{io.EOF} + } + return ts, nil +} + +func maxFillRange(required, optional memmap.MappableRange) memmap.MappableRange { + const maxReadahead = 64 << 10 // 64 KB, chosen arbitrarily + if required.Length() >= maxReadahead { + return required + } + if optional.Length() <= maxReadahead { + return optional + } + optional.Start = required.Start + if optional.Length() <= maxReadahead { + return optional + } + optional.End = optional.Start + maxReadahead + return optional +} + +// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable. +func (d *dentry) InvalidateUnsavable(ctx context.Context) error { + // Whether we have a host fd (and consequently what platform.File is + // mapped) can change across save/restore, so invalidate all translations + // unconditionally. + d.mapsMu.Lock() + defer d.mapsMu.Unlock() + d.mappings.InvalidateAll(memmap.InvalidateOpts{}) + + // Write the cache's contents back to the remote file so that if we have a + // host fd after restore, the remote file's contents are coherent. + mf := d.fs.mfp.MemoryFile() + d.dataMu.Lock() + defer d.dataMu.Unlock() + if err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, mf, d.handle.writeFromBlocksAt); err != nil { + return err + } + + // Discard the cache so that it's not stored in saved state. This is safe + // because per InvalidateUnsavable invariants, no new translations can have + // been returned after we invalidated all existing translations above. + d.cache.DropAll(mf) + d.dirty.RemoveAll() + + return nil +} + +// Evict implements pgalloc.EvictableMemoryUser.Evict. +func (d *dentry) Evict(ctx context.Context, er pgalloc.EvictableRange) { + d.mapsMu.Lock() + defer d.mapsMu.Unlock() + d.dataMu.Lock() + defer d.dataMu.Unlock() + + mr := memmap.MappableRange{er.Start, er.End} + mf := d.fs.mfp.MemoryFile() + // Only allow pages that are no longer memory-mapped to be evicted. + for mgap := d.mappings.LowerBoundGap(mr.Start); mgap.Ok() && mgap.Start() < mr.End; mgap = mgap.NextGap() { + mgapMR := mgap.Range().Intersect(mr) + if mgapMR.Length() == 0 { + continue + } + if err := fsutil.SyncDirty(ctx, mgapMR, &d.cache, &d.dirty, d.size, mf, d.handle.writeFromBlocksAt); err != nil { + log.Warningf("Failed to writeback cached data %v: %v", mgapMR, err) + } + d.cache.Drop(mgapMR, mf) + d.dirty.KeepClean(mgapMR) + } +} + +// dentryPlatformFile implements platform.File. It exists solely because dentry +// cannot implement both vfs.DentryImpl.IncRef and platform.File.IncRef. +// +// dentryPlatformFile is only used when a host FD representing the remote file +// is available (i.e. dentry.handle.fd >= 0), and that FD is used for +// application memory mappings (i.e. !filesystem.opts.forcePageCache). +type dentryPlatformFile struct { + *dentry + + // fdRefs counts references on platform.File offsets. fdRefs is protected + // by dentry.dataMu. + fdRefs fsutil.FrameRefSet + + // If this dentry represents a regular file, and handle.fd >= 0, + // hostFileMapper caches mappings of handle.fd. + hostFileMapper fsutil.HostFileMapper +} + +// IncRef implements platform.File.IncRef. +func (d *dentryPlatformFile) IncRef(fr platform.FileRange) { + d.dataMu.Lock() + seg, gap := d.fdRefs.Find(fr.Start) + for { + switch { + case seg.Ok() && seg.Start() < fr.End: + seg = d.fdRefs.Isolate(seg, fr) + seg.SetValue(seg.Value() + 1) + seg, gap = seg.NextNonEmpty() + case gap.Ok() && gap.Start() < fr.End: + newRange := gap.Range().Intersect(fr) + usage.MemoryAccounting.Inc(newRange.Length(), usage.Mapped) + seg, gap = d.fdRefs.InsertWithoutMerging(gap, newRange, 1).NextNonEmpty() + default: + d.fdRefs.MergeAdjacent(fr) + d.dataMu.Unlock() + return + } + } +} + +// DecRef implements platform.File.DecRef. +func (d *dentryPlatformFile) DecRef(fr platform.FileRange) { + d.dataMu.Lock() + seg := d.fdRefs.FindSegment(fr.Start) + + for seg.Ok() && seg.Start() < fr.End { + seg = d.fdRefs.Isolate(seg, fr) + if old := seg.Value(); old == 1 { + usage.MemoryAccounting.Dec(seg.Range().Length(), usage.Mapped) + seg = d.fdRefs.Remove(seg).NextSegment() + } else { + seg.SetValue(old - 1) + seg = seg.NextSegment() + } + } + d.fdRefs.MergeAdjacent(fr) + d.dataMu.Unlock() + +} + +// MapInternal implements platform.File.MapInternal. +func (d *dentryPlatformFile) MapInternal(fr platform.FileRange, at usermem.AccessType) (safemem.BlockSeq, error) { + d.handleMu.RLock() + bs, err := d.hostFileMapper.MapInternal(fr, int(d.handle.fd), at.Write) + d.handleMu.RUnlock() + return bs, err +} + +// FD implements platform.File.FD. +func (d *dentryPlatformFile) FD() int { + d.handleMu.RLock() + fd := d.handle.fd + d.handleMu.RUnlock() + return int(fd) +} diff --git a/pkg/sentry/fsimpl/gofer/special_file.go b/pkg/sentry/fsimpl/gofer/special_file.go new file mode 100644 index 000000000..08c691c47 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/special_file.go @@ -0,0 +1,159 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "sync" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// specialFileFD implements vfs.FileDescriptionImpl for files other than +// regular files, directories, and symlinks: pipes, sockets, etc. It is also +// used for regular files when filesystemOptions.specialRegularFiles is in +// effect. specialFileFD differs from regularFileFD by using per-FD handles +// instead of shared per-dentry handles, and never buffering I/O. +type specialFileFD struct { + fileDescription + + // handle is immutable. + handle handle + + // off is the file offset. off is protected by mu. (POSIX 2.9.7 only + // requires operations using the file offset to be atomic for regular files + // and symlinks; however, since specialFileFD may be used for regular + // files, we apply this atomicity unconditionally.) + mu sync.Mutex + off int64 +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (fd *specialFileFD) Release() { + fd.handle.close(context.Background()) + fs := fd.vfsfd.Mount().Filesystem().Impl().(*filesystem) + fs.syncMu.Lock() + delete(fs.specialFileFDs, fd) + fs.syncMu.Unlock() +} + +// OnClose implements vfs.FileDescriptionImpl.OnClose. +func (fd *specialFileFD) OnClose(ctx context.Context) error { + if !fd.vfsfd.IsWritable() { + return nil + } + return fd.handle.file.flush(ctx) +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + if offset < 0 { + return 0, syserror.EINVAL + } + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + // Going through dst.CopyOutFrom() holds MM locks around file operations of + // unknown duration. For regularFileFD, doing so is necessary to support + // mmap due to lock ordering; MM locks precede dentry.dataMu. That doesn't + // hold here since specialFileFD doesn't client-cache data. Just buffer the + // read instead. + if d := fd.dentry(); d.fs.opts.interop != InteropModeShared { + d.touchAtime(ctx, fd.vfsfd.Mount()) + } + buf := make([]byte, dst.NumBytes()) + n, err := fd.handle.readToBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset)) + if n == 0 { + return 0, err + } + if cp, cperr := dst.CopyOut(ctx, buf[:n]); cperr != nil { + return int64(cp), cperr + } + return int64(n), err +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (fd *specialFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + fd.mu.Lock() + n, err := fd.PRead(ctx, dst, fd.off, opts) + fd.off += n + fd.mu.Unlock() + return n, err +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *specialFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + if offset < 0 { + return 0, syserror.EINVAL + } + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + // Do a buffered write. See rationale in PRead. + if d := fd.dentry(); d.fs.opts.interop != InteropModeShared { + d.touchCMtime(ctx) + } + buf := make([]byte, src.NumBytes()) + // Don't do partial writes if we get a partial read from src. + if _, err := src.CopyIn(ctx, buf); err != nil { + return 0, err + } + n, err := fd.handle.writeFromBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset)) + return int64(n), err +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *specialFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + fd.mu.Lock() + n, err := fd.PWrite(ctx, src, fd.off, opts) + fd.off += n + fd.mu.Unlock() + return n, err +} + +// Seek implements vfs.FileDescriptionImpl.Seek. +func (fd *specialFileFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + fd.mu.Lock() + defer fd.mu.Unlock() + switch whence { + case linux.SEEK_SET: + // Use offset as given. + case linux.SEEK_CUR: + offset += fd.off + default: + // SEEK_END, SEEK_DATA, and SEEK_HOLE aren't supported since it's not + // clear that file size is even meaningful for these files. + return 0, syserror.EINVAL + } + if offset < 0 { + return 0, syserror.EINVAL + } + fd.off = offset + return offset, nil +} + +// Sync implements vfs.FileDescriptionImpl.Sync. +func (fd *specialFileFD) Sync(ctx context.Context) error { + if !fd.vfsfd.IsWritable() { + return nil + } + return fd.handle.sync(ctx) +} diff --git a/pkg/sentry/fsimpl/gofer/symlink.go b/pkg/sentry/fsimpl/gofer/symlink.go new file mode 100644 index 000000000..adf43be60 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/symlink.go @@ -0,0 +1,47 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +func (d *dentry) isSymlink() bool { + return d.fileType() == linux.S_IFLNK +} + +// Precondition: d.isSymlink(). +func (d *dentry) readlink(ctx context.Context, mnt *vfs.Mount) (string, error) { + if d.fs.opts.interop != InteropModeShared { + d.touchAtime(ctx, mnt) + d.dataMu.Lock() + if d.haveTarget { + target := d.target + d.dataMu.Unlock() + return target, nil + } + } + target, err := d.file.readlink(ctx) + if d.fs.opts.interop != InteropModeShared { + if err == nil { + d.haveTarget = true + d.target = target + } + d.dataMu.Unlock() + } + return target, err +} diff --git a/pkg/sentry/fsimpl/gofer/time.go b/pkg/sentry/fsimpl/gofer/time.go new file mode 100644 index 000000000..7598ec6a8 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/time.go @@ -0,0 +1,75 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +func dentryTimestampFromP9(s, ns uint64) int64 { + return int64(s*1e9 + ns) +} + +func dentryTimestampFromStatx(ts linux.StatxTimestamp) int64 { + return ts.Sec*1e9 + int64(ts.Nsec) +} + +func statxTimestampFromDentry(ns int64) linux.StatxTimestamp { + return linux.StatxTimestamp{ + Sec: ns / 1e9, + Nsec: uint32(ns % 1e9), + } +} + +func nowFromContext(ctx context.Context) (int64, bool) { + if clock := ktime.RealtimeClockFromContext(ctx); clock != nil { + return clock.Now().Nanoseconds(), true + } + return 0, false +} + +// Preconditions: fs.interop != InteropModeShared. +func (d *dentry) touchAtime(ctx context.Context, mnt *vfs.Mount) { + if err := mnt.CheckBeginWrite(); err != nil { + return + } + now, ok := nowFromContext(ctx) + if !ok { + mnt.EndWrite() + return + } + d.metadataMu.Lock() + atomic.StoreInt64(&d.atime, now) + d.metadataMu.Unlock() + mnt.EndWrite() +} + +// Preconditions: fs.interop != InteropModeShared. The caller has successfully +// called vfs.Mount.CheckBeginWrite(). +func (d *dentry) touchCMtime(ctx context.Context) { + now, ok := nowFromContext(ctx) + if !ok { + return + } + d.metadataMu.Lock() + atomic.StoreInt64(&d.mtime, now) + atomic.StoreInt64(&d.ctime, now) + d.metadataMu.Unlock() +} diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go index 5ee9cf1e9..72bc15264 100644 --- a/pkg/sentry/fsimpl/tmpfs/filesystem.go +++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go @@ -622,7 +622,7 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error if child.inode.isDir() { return syserror.EISDIR } - if !rp.MustBeDir() { + if rp.MustBeDir() { return syserror.ENOTDIR } mnt := rp.Mount() diff --git a/pkg/sentry/socket/hostinet/socket.go b/pkg/sentry/socket/hostinet/socket.go index bde4c7a1e..34f63986f 100644 --- a/pkg/sentry/socket/hostinet/socket.go +++ b/pkg/sentry/socket/hostinet/socket.go @@ -126,7 +126,7 @@ func (s *socketOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOS } return uint64(n), nil } - return readv(s.fd, iovecsFromBlockSeq(dsts)) + return readv(s.fd, safemem.IovecsFromBlockSeq(dsts)) })) return int64(n), err } @@ -149,7 +149,7 @@ func (s *socketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IO } return uint64(n), nil } - return writev(s.fd, iovecsFromBlockSeq(srcs)) + return writev(s.fd, safemem.IovecsFromBlockSeq(srcs)) })) return int64(n), err } @@ -402,7 +402,7 @@ func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags // We always do a non-blocking recv*(). sysflags := flags | syscall.MSG_DONTWAIT - iovs := iovecsFromBlockSeq(dsts) + iovs := safemem.IovecsFromBlockSeq(dsts) msg := syscall.Msghdr{ Iov: &iovs[0], Iovlen: uint64(len(iovs)), @@ -522,7 +522,7 @@ func (s *socketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to [] return uint64(n), nil } - iovs := iovecsFromBlockSeq(srcs) + iovs := safemem.IovecsFromBlockSeq(srcs) msg := syscall.Msghdr{ Iov: &iovs[0], Iovlen: uint64(len(iovs)), @@ -567,21 +567,6 @@ func (s *socketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to [] return int(n), syserr.FromError(err) } -func iovecsFromBlockSeq(bs safemem.BlockSeq) []syscall.Iovec { - iovs := make([]syscall.Iovec, 0, bs.NumBlocks()) - for ; !bs.IsEmpty(); bs = bs.Tail() { - b := bs.Head() - iovs = append(iovs, syscall.Iovec{ - Base: &b.ToSlice()[0], - Len: uint64(b.Len()), - }) - // We don't need to care about b.NeedSafecopy(), because the host - // kernel will handle such address ranges just fine (by returning - // EFAULT). - } - return iovs -} - func translateIOSyscallError(err error) error { if err == syscall.EAGAIN || err == syscall.EWOULDBLOCK { return syserror.ErrWouldBlock -- cgit v1.2.3 From 665b614e4a6e715bac25bea15c5c29184016e549 Mon Sep 17 00:00:00 2001 From: Ting-Yu Wang Date: Tue, 4 Feb 2020 18:04:26 -0800 Subject: Support RTM_NEWADDR and RTM_GETLINK in (rt)netlink. PiperOrigin-RevId: 293271055 --- pkg/sentry/inet/inet.go | 4 + pkg/sentry/inet/test_stack.go | 6 + pkg/sentry/socket/hostinet/stack.go | 5 + pkg/sentry/socket/netlink/BUILD | 14 +- pkg/sentry/socket/netlink/message.go | 129 +++++++++++ pkg/sentry/socket/netlink/message_test.go | 312 +++++++++++++++++++++++++++ pkg/sentry/socket/netlink/provider.go | 2 +- pkg/sentry/socket/netlink/route/BUILD | 2 - pkg/sentry/socket/netlink/route/protocol.go | 238 ++++++++++++++------ pkg/sentry/socket/netlink/socket.go | 54 ++--- pkg/sentry/socket/netlink/uevent/protocol.go | 2 +- pkg/sentry/socket/netstack/stack.go | 55 +++++ pkg/tcpip/stack/stack.go | 9 + test/syscalls/linux/BUILD | 2 + test/syscalls/linux/socket_netlink_route.cc | 296 ++++++++++++++++++++----- test/syscalls/linux/socket_netlink_util.cc | 45 +++- test/syscalls/linux/socket_netlink_util.h | 9 + 17 files changed, 1022 insertions(+), 162 deletions(-) create mode 100644 pkg/sentry/socket/netlink/message_test.go (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/inet/inet.go b/pkg/sentry/inet/inet.go index a7dfb78a7..2916a0644 100644 --- a/pkg/sentry/inet/inet.go +++ b/pkg/sentry/inet/inet.go @@ -28,6 +28,10 @@ type Stack interface { // interface indexes to a slice of associated interface address properties. InterfaceAddrs() map[int32][]InterfaceAddr + // AddInterfaceAddr adds an address to the network interface identified by + // index. + AddInterfaceAddr(idx int32, addr InterfaceAddr) error + // SupportsIPv6 returns true if the stack supports IPv6 connectivity. SupportsIPv6() bool diff --git a/pkg/sentry/inet/test_stack.go b/pkg/sentry/inet/test_stack.go index dcfcbd97e..d8961fc94 100644 --- a/pkg/sentry/inet/test_stack.go +++ b/pkg/sentry/inet/test_stack.go @@ -47,6 +47,12 @@ func (s *TestStack) InterfaceAddrs() map[int32][]InterfaceAddr { return s.InterfaceAddrsMap } +// AddInterfaceAddr implements Stack.AddInterfaceAddr. +func (s *TestStack) AddInterfaceAddr(idx int32, addr InterfaceAddr) error { + s.InterfaceAddrsMap[idx] = append(s.InterfaceAddrsMap[idx], addr) + return nil +} + // SupportsIPv6 implements Stack.SupportsIPv6. func (s *TestStack) SupportsIPv6() bool { return s.SupportsIPv6Flag diff --git a/pkg/sentry/socket/hostinet/stack.go b/pkg/sentry/socket/hostinet/stack.go index 034eca676..a48082631 100644 --- a/pkg/sentry/socket/hostinet/stack.go +++ b/pkg/sentry/socket/hostinet/stack.go @@ -310,6 +310,11 @@ func (s *Stack) InterfaceAddrs() map[int32][]inet.InterfaceAddr { return addrs } +// AddInterfaceAddr implements inet.Stack.AddInterfaceAddr. +func (s *Stack) AddInterfaceAddr(idx int32, addr inet.InterfaceAddr) error { + return syserror.EACCES +} + // SupportsIPv6 implements inet.Stack.SupportsIPv6. func (s *Stack) SupportsIPv6() bool { return s.supportsIPv6 diff --git a/pkg/sentry/socket/netlink/BUILD b/pkg/sentry/socket/netlink/BUILD index f8b8e467d..1911cd9b8 100644 --- a/pkg/sentry/socket/netlink/BUILD +++ b/pkg/sentry/socket/netlink/BUILD @@ -1,4 +1,4 @@ -load("//tools:defs.bzl", "go_library") +load("//tools:defs.bzl", "go_library", "go_test") package(licenses = ["notice"]) @@ -33,3 +33,15 @@ go_library( "//pkg/waiter", ], ) + +go_test( + name = "netlink_test", + size = "small", + srcs = [ + "message_test.go", + ], + deps = [ + ":netlink", + "//pkg/abi/linux", + ], +) diff --git a/pkg/sentry/socket/netlink/message.go b/pkg/sentry/socket/netlink/message.go index b21e0ca4b..4ea252ccb 100644 --- a/pkg/sentry/socket/netlink/message.go +++ b/pkg/sentry/socket/netlink/message.go @@ -30,8 +30,16 @@ func alignUp(length int, align uint) int { return (length + int(align) - 1) &^ (int(align) - 1) } +// alignPad returns the length of padding required for alignment. +// +// Preconditions: align is a power of two. +func alignPad(length int, align uint) int { + return alignUp(length, align) - length +} + // Message contains a complete serialized netlink message. type Message struct { + hdr linux.NetlinkMessageHeader buf []byte } @@ -40,10 +48,86 @@ type Message struct { // The header length will be updated by Finalize. func NewMessage(hdr linux.NetlinkMessageHeader) *Message { return &Message{ + hdr: hdr, buf: binary.Marshal(nil, usermem.ByteOrder, hdr), } } +// ParseMessage parses the first message seen at buf, returning the rest of the +// buffer. If message is malformed, ok of false is returned. For last message, +// padding check is loose, if there isn't enought padding, whole buf is consumed +// and ok is set to true. +func ParseMessage(buf []byte) (msg *Message, rest []byte, ok bool) { + b := BytesView(buf) + + hdrBytes, ok := b.Extract(linux.NetlinkMessageHeaderSize) + if !ok { + return + } + var hdr linux.NetlinkMessageHeader + binary.Unmarshal(hdrBytes, usermem.ByteOrder, &hdr) + + // Msg portion. + totalMsgLen := int(hdr.Length) + _, ok = b.Extract(totalMsgLen - linux.NetlinkMessageHeaderSize) + if !ok { + return + } + + // Padding. + numPad := alignPad(totalMsgLen, linux.NLMSG_ALIGNTO) + // Linux permits the last message not being aligned, just consume all of it. + // Ref: net/netlink/af_netlink.c:netlink_rcv_skb + if numPad > len(b) { + numPad = len(b) + } + _, ok = b.Extract(numPad) + if !ok { + return + } + + return &Message{ + hdr: hdr, + buf: buf[:totalMsgLen], + }, []byte(b), true +} + +// Header returns the header of this message. +func (m *Message) Header() linux.NetlinkMessageHeader { + return m.hdr +} + +// GetData unmarshals the payload message header from this netlink message, and +// returns the attributes portion. +func (m *Message) GetData(msg interface{}) (AttrsView, bool) { + b := BytesView(m.buf) + + _, ok := b.Extract(linux.NetlinkMessageHeaderSize) + if !ok { + return nil, false + } + + size := int(binary.Size(msg)) + msgBytes, ok := b.Extract(size) + if !ok { + return nil, false + } + binary.Unmarshal(msgBytes, usermem.ByteOrder, msg) + + numPad := alignPad(linux.NetlinkMessageHeaderSize+size, linux.NLMSG_ALIGNTO) + // Linux permits the last message not being aligned, just consume all of it. + // Ref: net/netlink/af_netlink.c:netlink_rcv_skb + if numPad > len(b) { + numPad = len(b) + } + _, ok = b.Extract(numPad) + if !ok { + return nil, false + } + + return AttrsView(b), true +} + // Finalize returns the []byte containing the entire message, with the total // length set in the message header. The Message must not be modified after // calling Finalize. @@ -157,3 +241,48 @@ func (ms *MessageSet) AddMessage(hdr linux.NetlinkMessageHeader) *Message { ms.Messages = append(ms.Messages, m) return m } + +// AttrsView is a view into the attributes portion of a netlink message. +type AttrsView []byte + +// Empty returns whether there is no attribute left in v. +func (v AttrsView) Empty() bool { + return len(v) == 0 +} + +// ParseFirst parses first netlink attribute at the beginning of v. +func (v AttrsView) ParseFirst() (hdr linux.NetlinkAttrHeader, value []byte, rest AttrsView, ok bool) { + b := BytesView(v) + + hdrBytes, ok := b.Extract(linux.NetlinkAttrHeaderSize) + if !ok { + return + } + binary.Unmarshal(hdrBytes, usermem.ByteOrder, &hdr) + + value, ok = b.Extract(int(hdr.Length) - linux.NetlinkAttrHeaderSize) + if !ok { + return + } + + _, ok = b.Extract(alignPad(int(hdr.Length), linux.NLA_ALIGNTO)) + if !ok { + return + } + + return hdr, value, AttrsView(b), ok +} + +// BytesView supports extracting data from a byte slice with bounds checking. +type BytesView []byte + +// Extract removes the first n bytes from v and returns it. If n is out of +// bounds, it returns false. +func (v *BytesView) Extract(n int) ([]byte, bool) { + if n < 0 || n > len(*v) { + return nil, false + } + extracted := (*v)[:n] + *v = (*v)[n:] + return extracted, true +} diff --git a/pkg/sentry/socket/netlink/message_test.go b/pkg/sentry/socket/netlink/message_test.go new file mode 100644 index 000000000..ef13d9386 --- /dev/null +++ b/pkg/sentry/socket/netlink/message_test.go @@ -0,0 +1,312 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package message_test + +import ( + "bytes" + "reflect" + "testing" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/socket/netlink" +) + +type dummyNetlinkMsg struct { + Foo uint16 +} + +func TestParseMessage(t *testing.T) { + tests := []struct { + desc string + input []byte + + header linux.NetlinkMessageHeader + dataMsg *dummyNetlinkMsg + restLen int + ok bool + }{ + { + desc: "valid", + input: []byte{ + 0x14, 0x00, 0x00, 0x00, // Length + 0x01, 0x00, // Type + 0x02, 0x00, // Flags + 0x03, 0x00, 0x00, 0x00, // Seq + 0x04, 0x00, 0x00, 0x00, // PortID + 0x30, 0x31, 0x00, 0x00, // Data message with 2 bytes padding + }, + header: linux.NetlinkMessageHeader{ + Length: 20, + Type: 1, + Flags: 2, + Seq: 3, + PortID: 4, + }, + dataMsg: &dummyNetlinkMsg{ + Foo: 0x3130, + }, + restLen: 0, + ok: true, + }, + { + desc: "valid with next message", + input: []byte{ + 0x14, 0x00, 0x00, 0x00, // Length + 0x01, 0x00, // Type + 0x02, 0x00, // Flags + 0x03, 0x00, 0x00, 0x00, // Seq + 0x04, 0x00, 0x00, 0x00, // PortID + 0x30, 0x31, 0x00, 0x00, // Data message with 2 bytes padding + 0xFF, // Next message (rest) + }, + header: linux.NetlinkMessageHeader{ + Length: 20, + Type: 1, + Flags: 2, + Seq: 3, + PortID: 4, + }, + dataMsg: &dummyNetlinkMsg{ + Foo: 0x3130, + }, + restLen: 1, + ok: true, + }, + { + desc: "valid for last message without padding", + input: []byte{ + 0x12, 0x00, 0x00, 0x00, // Length + 0x01, 0x00, // Type + 0x02, 0x00, // Flags + 0x03, 0x00, 0x00, 0x00, // Seq + 0x04, 0x00, 0x00, 0x00, // PortID + 0x30, 0x31, // Data message + }, + header: linux.NetlinkMessageHeader{ + Length: 18, + Type: 1, + Flags: 2, + Seq: 3, + PortID: 4, + }, + dataMsg: &dummyNetlinkMsg{ + Foo: 0x3130, + }, + restLen: 0, + ok: true, + }, + { + desc: "valid for last message not to be aligned", + input: []byte{ + 0x13, 0x00, 0x00, 0x00, // Length + 0x01, 0x00, // Type + 0x02, 0x00, // Flags + 0x03, 0x00, 0x00, 0x00, // Seq + 0x04, 0x00, 0x00, 0x00, // PortID + 0x30, 0x31, // Data message + 0x00, // Excessive 1 byte permitted at end + }, + header: linux.NetlinkMessageHeader{ + Length: 19, + Type: 1, + Flags: 2, + Seq: 3, + PortID: 4, + }, + dataMsg: &dummyNetlinkMsg{ + Foo: 0x3130, + }, + restLen: 0, + ok: true, + }, + { + desc: "header.Length too short", + input: []byte{ + 0x04, 0x00, 0x00, 0x00, // Length + 0x01, 0x00, // Type + 0x02, 0x00, // Flags + 0x03, 0x00, 0x00, 0x00, // Seq + 0x04, 0x00, 0x00, 0x00, // PortID + 0x30, 0x31, 0x00, 0x00, // Data message with 2 bytes padding + }, + ok: false, + }, + { + desc: "header.Length too long", + input: []byte{ + 0xFF, 0xFF, 0x00, 0x00, // Length + 0x01, 0x00, // Type + 0x02, 0x00, // Flags + 0x03, 0x00, 0x00, 0x00, // Seq + 0x04, 0x00, 0x00, 0x00, // PortID + 0x30, 0x31, 0x00, 0x00, // Data message with 2 bytes padding + }, + ok: false, + }, + { + desc: "header incomplete", + input: []byte{ + 0x04, 0x00, 0x00, 0x00, // Length + }, + ok: false, + }, + { + desc: "empty message", + input: []byte{}, + ok: false, + }, + } + for _, test := range tests { + msg, rest, ok := netlink.ParseMessage(test.input) + if ok != test.ok { + t.Errorf("%v: got ok = %v, want = %v", test.desc, ok, test.ok) + continue + } + if !test.ok { + continue + } + if !reflect.DeepEqual(msg.Header(), test.header) { + t.Errorf("%v: got hdr = %+v, want = %+v", test.desc, msg.Header(), test.header) + } + + dataMsg := &dummyNetlinkMsg{} + _, dataOk := msg.GetData(dataMsg) + if !dataOk { + t.Errorf("%v: GetData.ok = %v, want = true", test.desc, dataOk) + } else if !reflect.DeepEqual(dataMsg, test.dataMsg) { + t.Errorf("%v: GetData.msg = %+v, want = %+v", test.desc, dataMsg, test.dataMsg) + } + + if got, want := rest, test.input[len(test.input)-test.restLen:]; !bytes.Equal(got, want) { + t.Errorf("%v: got rest = %v, want = %v", test.desc, got, want) + } + } +} + +func TestAttrView(t *testing.T) { + tests := []struct { + desc string + input []byte + + // Outputs for ParseFirst. + hdr linux.NetlinkAttrHeader + value []byte + restLen int + ok bool + + // Outputs for Empty. + isEmpty bool + }{ + { + desc: "valid", + input: []byte{ + 0x06, 0x00, // Length + 0x01, 0x00, // Type + 0x30, 0x31, 0x00, 0x00, // Data with 2 bytes padding + }, + hdr: linux.NetlinkAttrHeader{ + Length: 6, + Type: 1, + }, + value: []byte{0x30, 0x31}, + restLen: 0, + ok: true, + isEmpty: false, + }, + { + desc: "at alignment", + input: []byte{ + 0x08, 0x00, // Length + 0x01, 0x00, // Type + 0x30, 0x31, 0x32, 0x33, // Data + }, + hdr: linux.NetlinkAttrHeader{ + Length: 8, + Type: 1, + }, + value: []byte{0x30, 0x31, 0x32, 0x33}, + restLen: 0, + ok: true, + isEmpty: false, + }, + { + desc: "at alignment with rest data", + input: []byte{ + 0x08, 0x00, // Length + 0x01, 0x00, // Type + 0x30, 0x31, 0x32, 0x33, // Data + 0xFF, 0xFE, // Rest data + }, + hdr: linux.NetlinkAttrHeader{ + Length: 8, + Type: 1, + }, + value: []byte{0x30, 0x31, 0x32, 0x33}, + restLen: 2, + ok: true, + isEmpty: false, + }, + { + desc: "hdr.Length too long", + input: []byte{ + 0xFF, 0x00, // Length + 0x01, 0x00, // Type + 0x30, 0x31, 0x32, 0x33, // Data + }, + ok: false, + isEmpty: false, + }, + { + desc: "hdr.Length too short", + input: []byte{ + 0x01, 0x00, // Length + 0x01, 0x00, // Type + 0x30, 0x31, 0x32, 0x33, // Data + }, + ok: false, + isEmpty: false, + }, + { + desc: "empty", + input: []byte{}, + ok: false, + isEmpty: true, + }, + } + for _, test := range tests { + attrs := netlink.AttrsView(test.input) + + // Test ParseFirst(). + hdr, value, rest, ok := attrs.ParseFirst() + if ok != test.ok { + t.Errorf("%v: got ok = %v, want = %v", test.desc, ok, test.ok) + } else if test.ok { + if !reflect.DeepEqual(hdr, test.hdr) { + t.Errorf("%v: got hdr = %+v, want = %+v", test.desc, hdr, test.hdr) + } + if !bytes.Equal(value, test.value) { + t.Errorf("%v: got value = %v, want = %v", test.desc, value, test.value) + } + if wantRest := test.input[len(test.input)-test.restLen:]; !bytes.Equal(rest, wantRest) { + t.Errorf("%v: got rest = %v, want = %v", test.desc, rest, wantRest) + } + } + + // Test Empty(). + if got, want := attrs.Empty(), test.isEmpty; got != want { + t.Errorf("%v: got empty = %v, want = %v", test.desc, got, want) + } + } +} diff --git a/pkg/sentry/socket/netlink/provider.go b/pkg/sentry/socket/netlink/provider.go index 07f860a49..b0dc70e5c 100644 --- a/pkg/sentry/socket/netlink/provider.go +++ b/pkg/sentry/socket/netlink/provider.go @@ -42,7 +42,7 @@ type Protocol interface { // If err == nil, any messages added to ms will be sent back to the // other end of the socket. Setting ms.Multi will cause an NLMSG_DONE // message to be sent even if ms contains no messages. - ProcessMessage(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *MessageSet) *syserr.Error + ProcessMessage(ctx context.Context, msg *Message, ms *MessageSet) *syserr.Error } // Provider is a function that creates a new Protocol for a specific netlink diff --git a/pkg/sentry/socket/netlink/route/BUILD b/pkg/sentry/socket/netlink/route/BUILD index 622a1eafc..93127398d 100644 --- a/pkg/sentry/socket/netlink/route/BUILD +++ b/pkg/sentry/socket/netlink/route/BUILD @@ -10,13 +10,11 @@ go_library( visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", - "//pkg/binary", "//pkg/context", "//pkg/sentry/inet", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/socket/netlink", "//pkg/syserr", - "//pkg/usermem", ], ) diff --git a/pkg/sentry/socket/netlink/route/protocol.go b/pkg/sentry/socket/netlink/route/protocol.go index 2b3c7f5b3..c84d8bd7c 100644 --- a/pkg/sentry/socket/netlink/route/protocol.go +++ b/pkg/sentry/socket/netlink/route/protocol.go @@ -17,16 +17,15 @@ package route import ( "bytes" + "syscall" "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/socket/netlink" "gvisor.dev/gvisor/pkg/syserr" - "gvisor.dev/gvisor/pkg/usermem" ) // commandKind describes the operational class of a message type. @@ -69,13 +68,7 @@ func (p *Protocol) CanSend() bool { } // dumpLinks handles RTM_GETLINK dump requests. -func (p *Protocol) dumpLinks(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *netlink.MessageSet) *syserr.Error { - // TODO(b/68878065): Only the dump variant of the types below are - // supported. - if hdr.Flags&linux.NLM_F_DUMP != linux.NLM_F_DUMP { - return syserr.ErrNotSupported - } - +func (p *Protocol) dumpLinks(ctx context.Context, msg *netlink.Message, ms *netlink.MessageSet) *syserr.Error { // NLM_F_DUMP + RTM_GETLINK messages are supposed to include an // ifinfomsg. However, Linux <3.9 only checked for rtgenmsg, and some // userspace applications (including glibc) still include rtgenmsg. @@ -99,44 +92,105 @@ func (p *Protocol) dumpLinks(ctx context.Context, hdr linux.NetlinkMessageHeader return nil } - for id, i := range stack.Interfaces() { - m := ms.AddMessage(linux.NetlinkMessageHeader{ - Type: linux.RTM_NEWLINK, - }) + for idx, i := range stack.Interfaces() { + addNewLinkMessage(ms, idx, i) + } - m.Put(linux.InterfaceInfoMessage{ - Family: linux.AF_UNSPEC, - Type: i.DeviceType, - Index: id, - Flags: i.Flags, - }) + return nil +} - m.PutAttrString(linux.IFLA_IFNAME, i.Name) - m.PutAttr(linux.IFLA_MTU, i.MTU) +// getLinks handles RTM_GETLINK requests. +func (p *Protocol) getLink(ctx context.Context, msg *netlink.Message, ms *netlink.MessageSet) *syserr.Error { + stack := inet.StackFromContext(ctx) + if stack == nil { + // No network devices. + return nil + } - mac := make([]byte, 6) - brd := mac - if len(i.Addr) > 0 { - mac = i.Addr - brd = bytes.Repeat([]byte{0xff}, len(i.Addr)) + // Parse message. + var ifi linux.InterfaceInfoMessage + attrs, ok := msg.GetData(&ifi) + if !ok { + return syserr.ErrInvalidArgument + } + + // Parse attributes. + var byName []byte + for !attrs.Empty() { + ahdr, value, rest, ok := attrs.ParseFirst() + if !ok { + return syserr.ErrInvalidArgument } - m.PutAttr(linux.IFLA_ADDRESS, mac) - m.PutAttr(linux.IFLA_BROADCAST, brd) + attrs = rest - // TODO(gvisor.dev/issue/578): There are many more attributes. + switch ahdr.Type { + case linux.IFLA_IFNAME: + if len(value) < 1 { + return syserr.ErrInvalidArgument + } + byName = value[:len(value)-1] + + // TODO(gvisor.dev/issue/578): Support IFLA_EXT_MASK. + } } + found := false + for idx, i := range stack.Interfaces() { + switch { + case ifi.Index > 0: + if idx != ifi.Index { + continue + } + case byName != nil: + if string(byName) != i.Name { + continue + } + default: + // Criteria not specified. + return syserr.ErrInvalidArgument + } + + addNewLinkMessage(ms, idx, i) + found = true + break + } + if !found { + return syserr.ErrNoDevice + } return nil } -// dumpAddrs handles RTM_GETADDR dump requests. -func (p *Protocol) dumpAddrs(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *netlink.MessageSet) *syserr.Error { - // TODO(b/68878065): Only the dump variant of the types below are - // supported. - if hdr.Flags&linux.NLM_F_DUMP != linux.NLM_F_DUMP { - return syserr.ErrNotSupported +// addNewLinkMessage appends RTM_NEWLINK message for the given interface into +// the message set. +func addNewLinkMessage(ms *netlink.MessageSet, idx int32, i inet.Interface) { + m := ms.AddMessage(linux.NetlinkMessageHeader{ + Type: linux.RTM_NEWLINK, + }) + + m.Put(linux.InterfaceInfoMessage{ + Family: linux.AF_UNSPEC, + Type: i.DeviceType, + Index: idx, + Flags: i.Flags, + }) + + m.PutAttrString(linux.IFLA_IFNAME, i.Name) + m.PutAttr(linux.IFLA_MTU, i.MTU) + + mac := make([]byte, 6) + brd := mac + if len(i.Addr) > 0 { + mac = i.Addr + brd = bytes.Repeat([]byte{0xff}, len(i.Addr)) } + m.PutAttr(linux.IFLA_ADDRESS, mac) + m.PutAttr(linux.IFLA_BROADCAST, brd) + + // TODO(gvisor.dev/issue/578): There are many more attributes. +} +// dumpAddrs handles RTM_GETADDR dump requests. +func (p *Protocol) dumpAddrs(ctx context.Context, msg *netlink.Message, ms *netlink.MessageSet) *syserr.Error { // RTM_GETADDR dump requests need not contain anything more than the // netlink header and 1 byte protocol family common to all // NETLINK_ROUTE requests. @@ -168,6 +222,7 @@ func (p *Protocol) dumpAddrs(ctx context.Context, hdr linux.NetlinkMessageHeader Index: uint32(id), }) + m.PutAttr(linux.IFA_LOCAL, []byte(a.Addr)) m.PutAttr(linux.IFA_ADDRESS, []byte(a.Addr)) // TODO(gvisor.dev/issue/578): There are many more attributes. @@ -252,12 +307,12 @@ func fillRoute(routes []inet.Route, addr []byte) (inet.Route, *syserr.Error) { } // parseForDestination parses a message as format of RouteMessage-RtAttr-dst. -func parseForDestination(data []byte) ([]byte, *syserr.Error) { +func parseForDestination(msg *netlink.Message) ([]byte, *syserr.Error) { var rtMsg linux.RouteMessage - if len(data) < linux.SizeOfRouteMessage { + attrs, ok := msg.GetData(&rtMsg) + if !ok { return nil, syserr.ErrInvalidArgument } - binary.Unmarshal(data[:linux.SizeOfRouteMessage], usermem.ByteOrder, &rtMsg) // iproute2 added the RTM_F_LOOKUP_TABLE flag in version v4.4.0. See // commit bc234301af12. Note we don't check this flag for backward // compatibility. @@ -265,26 +320,15 @@ func parseForDestination(data []byte) ([]byte, *syserr.Error) { return nil, syserr.ErrNotSupported } - data = data[linux.SizeOfRouteMessage:] - - // TODO(gvisor.dev/issue/1611): Add generic attribute parsing. - var rtAttr linux.RtAttr - if len(data) < linux.SizeOfRtAttr { - return nil, syserr.ErrInvalidArgument + // Expect first attribute is RTA_DST. + if hdr, value, _, ok := attrs.ParseFirst(); ok && hdr.Type == linux.RTA_DST { + return value, nil } - binary.Unmarshal(data[:linux.SizeOfRtAttr], usermem.ByteOrder, &rtAttr) - if rtAttr.Type != linux.RTA_DST { - return nil, syserr.ErrInvalidArgument - } - - if len(data) < int(rtAttr.Len) { - return nil, syserr.ErrInvalidArgument - } - return data[linux.SizeOfRtAttr:rtAttr.Len], nil + return nil, syserr.ErrInvalidArgument } // dumpRoutes handles RTM_GETROUTE requests. -func (p *Protocol) dumpRoutes(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *netlink.MessageSet) *syserr.Error { +func (p *Protocol) dumpRoutes(ctx context.Context, msg *netlink.Message, ms *netlink.MessageSet) *syserr.Error { // RTM_GETROUTE dump requests need not contain anything more than the // netlink header and 1 byte protocol family common to all // NETLINK_ROUTE requests. @@ -295,10 +339,11 @@ func (p *Protocol) dumpRoutes(ctx context.Context, hdr linux.NetlinkMessageHeade return nil } + hdr := msg.Header() routeTables := stack.RouteTable() if hdr.Flags == linux.NLM_F_REQUEST { - dst, err := parseForDestination(data) + dst, err := parseForDestination(msg) if err != nil { return err } @@ -357,10 +402,55 @@ func (p *Protocol) dumpRoutes(ctx context.Context, hdr linux.NetlinkMessageHeade return nil } +// newAddr handles RTM_NEWADDR requests. +func (p *Protocol) newAddr(ctx context.Context, msg *netlink.Message, ms *netlink.MessageSet) *syserr.Error { + stack := inet.StackFromContext(ctx) + if stack == nil { + // No network stack. + return syserr.ErrProtocolNotSupported + } + + var ifa linux.InterfaceAddrMessage + attrs, ok := msg.GetData(&ifa) + if !ok { + return syserr.ErrInvalidArgument + } + + for !attrs.Empty() { + ahdr, value, rest, ok := attrs.ParseFirst() + if !ok { + return syserr.ErrInvalidArgument + } + attrs = rest + + switch ahdr.Type { + case linux.IFA_LOCAL: + err := stack.AddInterfaceAddr(int32(ifa.Index), inet.InterfaceAddr{ + Family: ifa.Family, + PrefixLen: ifa.PrefixLen, + Flags: ifa.Flags, + Addr: value, + }) + if err == syscall.EEXIST { + flags := msg.Header().Flags + if flags&linux.NLM_F_EXCL != 0 { + return syserr.ErrExists + } + } else if err != nil { + return syserr.ErrInvalidArgument + } + } + } + return nil +} + // ProcessMessage implements netlink.Protocol.ProcessMessage. -func (p *Protocol) ProcessMessage(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *netlink.MessageSet) *syserr.Error { +func (p *Protocol) ProcessMessage(ctx context.Context, msg *netlink.Message, ms *netlink.MessageSet) *syserr.Error { + hdr := msg.Header() + // All messages start with a 1 byte protocol family. - if len(data) < 1 { + var family uint8 + if _, ok := msg.GetData(&family); !ok { // Linux ignores messages missing the protocol family. See // net/core/rtnetlink.c:rtnetlink_rcv_msg. return nil @@ -374,16 +464,32 @@ func (p *Protocol) ProcessMessage(ctx context.Context, hdr linux.NetlinkMessageH } } - switch hdr.Type { - case linux.RTM_GETLINK: - return p.dumpLinks(ctx, hdr, data, ms) - case linux.RTM_GETADDR: - return p.dumpAddrs(ctx, hdr, data, ms) - case linux.RTM_GETROUTE: - return p.dumpRoutes(ctx, hdr, data, ms) - default: - return syserr.ErrNotSupported + if hdr.Flags&linux.NLM_F_DUMP == linux.NLM_F_DUMP { + // TODO(b/68878065): Only the dump variant of the types below are + // supported. + switch hdr.Type { + case linux.RTM_GETLINK: + return p.dumpLinks(ctx, msg, ms) + case linux.RTM_GETADDR: + return p.dumpAddrs(ctx, msg, ms) + case linux.RTM_GETROUTE: + return p.dumpRoutes(ctx, msg, ms) + default: + return syserr.ErrNotSupported + } + } else if hdr.Flags&linux.NLM_F_REQUEST == linux.NLM_F_REQUEST { + switch hdr.Type { + case linux.RTM_GETLINK: + return p.getLink(ctx, msg, ms) + case linux.RTM_GETROUTE: + return p.dumpRoutes(ctx, msg, ms) + case linux.RTM_NEWADDR: + return p.newAddr(ctx, msg, ms) + default: + return syserr.ErrNotSupported + } } + return syserr.ErrNotSupported } // init registers the NETLINK_ROUTE provider. diff --git a/pkg/sentry/socket/netlink/socket.go b/pkg/sentry/socket/netlink/socket.go index c4b95debb..2ca02567d 100644 --- a/pkg/sentry/socket/netlink/socket.go +++ b/pkg/sentry/socket/netlink/socket.go @@ -644,47 +644,38 @@ func (s *Socket) sendResponse(ctx context.Context, ms *MessageSet) *syserr.Error return nil } -func (s *Socket) dumpErrorMesage(ctx context.Context, hdr linux.NetlinkMessageHeader, ms *MessageSet, err *syserr.Error) *syserr.Error { +func dumpErrorMesage(hdr linux.NetlinkMessageHeader, ms *MessageSet, err *syserr.Error) { m := ms.AddMessage(linux.NetlinkMessageHeader{ Type: linux.NLMSG_ERROR, }) - m.Put(linux.NetlinkErrorMessage{ Error: int32(-err.ToLinux().Number()), Header: hdr, }) - return nil +} +func dumpAckMesage(hdr linux.NetlinkMessageHeader, ms *MessageSet) { + m := ms.AddMessage(linux.NetlinkMessageHeader{ + Type: linux.NLMSG_ERROR, + }) + m.Put(linux.NetlinkErrorMessage{ + Error: 0, + Header: hdr, + }) } // processMessages handles each message in buf, passing it to the protocol // handler for final handling. func (s *Socket) processMessages(ctx context.Context, buf []byte) *syserr.Error { for len(buf) > 0 { - if len(buf) < linux.NetlinkMessageHeaderSize { + msg, rest, ok := ParseMessage(buf) + if !ok { // Linux ignores messages that are too short. See // net/netlink/af_netlink.c:netlink_rcv_skb. break } - - var hdr linux.NetlinkMessageHeader - binary.Unmarshal(buf[:linux.NetlinkMessageHeaderSize], usermem.ByteOrder, &hdr) - - if hdr.Length < linux.NetlinkMessageHeaderSize || uint64(hdr.Length) > uint64(len(buf)) { - // Linux ignores malformed messages. See - // net/netlink/af_netlink.c:netlink_rcv_skb. - break - } - - // Data from this message. - data := buf[linux.NetlinkMessageHeaderSize:hdr.Length] - - // Advance to the next message. - next := alignUp(int(hdr.Length), linux.NLMSG_ALIGNTO) - if next >= len(buf)-1 { - next = len(buf) - 1 - } - buf = buf[next:] + buf = rest + hdr := msg.Header() // Ignore control messages. if hdr.Type < linux.NLMSG_MIN_TYPE { @@ -692,19 +683,10 @@ func (s *Socket) processMessages(ctx context.Context, buf []byte) *syserr.Error } ms := NewMessageSet(s.portID, hdr.Seq) - var err *syserr.Error - // TODO(b/68877377): ACKs not supported yet. - if hdr.Flags&linux.NLM_F_ACK == linux.NLM_F_ACK { - err = syserr.ErrNotSupported - } else { - - err = s.protocol.ProcessMessage(ctx, hdr, data, ms) - } - if err != nil { - ms = NewMessageSet(s.portID, hdr.Seq) - if err := s.dumpErrorMesage(ctx, hdr, ms, err); err != nil { - return err - } + if err := s.protocol.ProcessMessage(ctx, msg, ms); err != nil { + dumpErrorMesage(hdr, ms, err) + } else if hdr.Flags&linux.NLM_F_ACK == linux.NLM_F_ACK { + dumpAckMesage(hdr, ms) } if err := s.sendResponse(ctx, ms); err != nil { diff --git a/pkg/sentry/socket/netlink/uevent/protocol.go b/pkg/sentry/socket/netlink/uevent/protocol.go index 1ee4296bc..029ba21b5 100644 --- a/pkg/sentry/socket/netlink/uevent/protocol.go +++ b/pkg/sentry/socket/netlink/uevent/protocol.go @@ -49,7 +49,7 @@ func (p *Protocol) CanSend() bool { } // ProcessMessage implements netlink.Protocol.ProcessMessage. -func (p *Protocol) ProcessMessage(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *netlink.MessageSet) *syserr.Error { +func (p *Protocol) ProcessMessage(ctx context.Context, msg *netlink.Message, ms *netlink.MessageSet) *syserr.Error { // Silently ignore all messages. return nil } diff --git a/pkg/sentry/socket/netstack/stack.go b/pkg/sentry/socket/netstack/stack.go index 31ea66eca..0692482e9 100644 --- a/pkg/sentry/socket/netstack/stack.go +++ b/pkg/sentry/socket/netstack/stack.go @@ -20,6 +20,8 @@ import ( "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/socket/netfilter" "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/iptables" "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" @@ -88,6 +90,59 @@ func (s *Stack) InterfaceAddrs() map[int32][]inet.InterfaceAddr { return nicAddrs } +// AddInterfaceAddr implements inet.Stack.AddInterfaceAddr. +func (s *Stack) AddInterfaceAddr(idx int32, addr inet.InterfaceAddr) error { + var ( + protocol tcpip.NetworkProtocolNumber + address tcpip.Address + ) + switch addr.Family { + case linux.AF_INET: + if len(addr.Addr) < header.IPv4AddressSize { + return syserror.EINVAL + } + if addr.PrefixLen > header.IPv4AddressSize*8 { + return syserror.EINVAL + } + protocol = ipv4.ProtocolNumber + address = tcpip.Address(addr.Addr[:header.IPv4AddressSize]) + + case linux.AF_INET6: + if len(addr.Addr) < header.IPv6AddressSize { + return syserror.EINVAL + } + if addr.PrefixLen > header.IPv6AddressSize*8 { + return syserror.EINVAL + } + protocol = ipv6.ProtocolNumber + address = tcpip.Address(addr.Addr[:header.IPv6AddressSize]) + + default: + return syserror.ENOTSUP + } + + protocolAddress := tcpip.ProtocolAddress{ + Protocol: protocol, + AddressWithPrefix: tcpip.AddressWithPrefix{ + Address: address, + PrefixLen: int(addr.PrefixLen), + }, + } + + // Attach address to interface. + if err := s.Stack.AddProtocolAddressWithOptions(tcpip.NICID(idx), protocolAddress, stack.CanBePrimaryEndpoint); err != nil { + return syserr.TranslateNetstackError(err).ToError() + } + + // Add route for local network. + s.Stack.AddRoute(tcpip.Route{ + Destination: protocolAddress.AddressWithPrefix.Subnet(), + Gateway: "", // No gateway for local network. + NIC: tcpip.NICID(idx), + }) + return nil +} + // TCPReceiveBufferSize implements inet.Stack.TCPReceiveBufferSize. func (s *Stack) TCPReceiveBufferSize() (inet.TCPBufferSize, error) { var rs tcp.ReceiveBufferSizeOption diff --git a/pkg/tcpip/stack/stack.go b/pkg/tcpip/stack/stack.go index 7057b110e..b793f1d74 100644 --- a/pkg/tcpip/stack/stack.go +++ b/pkg/tcpip/stack/stack.go @@ -795,6 +795,8 @@ func (s *Stack) Forwarding() bool { // SetRouteTable assigns the route table to be used by this stack. It // specifies which NIC to use for given destination address ranges. +// +// This method takes ownership of the table. func (s *Stack) SetRouteTable(table []tcpip.Route) { s.mu.Lock() defer s.mu.Unlock() @@ -809,6 +811,13 @@ func (s *Stack) GetRouteTable() []tcpip.Route { return append([]tcpip.Route(nil), s.routeTable...) } +// AddRoute appends a route to the route table. +func (s *Stack) AddRoute(route tcpip.Route) { + s.mu.Lock() + defer s.mu.Unlock() + s.routeTable = append(s.routeTable, route) +} + // NewEndpoint creates a new transport layer endpoint of the given protocol. func (s *Stack) NewEndpoint(transport tcpip.TransportProtocolNumber, network tcpip.NetworkProtocolNumber, waiterQueue *waiter.Queue) (tcpip.Endpoint, *tcpip.Error) { t, ok := s.transportProtocols[transport] diff --git a/test/syscalls/linux/BUILD b/test/syscalls/linux/BUILD index 273b014d6..f2e3c7072 100644 --- a/test/syscalls/linux/BUILD +++ b/test/syscalls/linux/BUILD @@ -2769,9 +2769,11 @@ cc_binary( deps = [ ":socket_netlink_util", ":socket_test_util", + "//test/util:capability_util", "//test/util:cleanup", "//test/util:file_descriptor", "@com_google_absl//absl/strings:str_format", + "@com_google_absl//absl/types:optional", gtest, "//test/util:test_main", "//test/util:test_util", diff --git a/test/syscalls/linux/socket_netlink_route.cc b/test/syscalls/linux/socket_netlink_route.cc index 1e28e658d..e5aed1eec 100644 --- a/test/syscalls/linux/socket_netlink_route.cc +++ b/test/syscalls/linux/socket_netlink_route.cc @@ -14,6 +14,7 @@ #include #include +#include #include #include #include @@ -25,8 +26,10 @@ #include "gtest/gtest.h" #include "absl/strings/str_format.h" +#include "absl/types/optional.h" #include "test/syscalls/linux/socket_netlink_util.h" #include "test/syscalls/linux/socket_test_util.h" +#include "test/util/capability_util.h" #include "test/util/cleanup.h" #include "test/util/file_descriptor.h" #include "test/util/test_util.h" @@ -38,6 +41,8 @@ namespace testing { namespace { +constexpr uint32_t kSeq = 12345; + using ::testing::AnyOf; using ::testing::Eq; @@ -113,58 +118,224 @@ void CheckGetLinkResponse(const struct nlmsghdr* hdr, int seq, int port) { // TODO(mpratt): Check ifinfomsg contents and following attrs. } +PosixError DumpLinks( + const FileDescriptor& fd, uint32_t seq, + const std::function& fn) { + struct request { + struct nlmsghdr hdr; + struct ifinfomsg ifm; + }; + + struct request req = {}; + req.hdr.nlmsg_len = sizeof(req); + req.hdr.nlmsg_type = RTM_GETLINK; + req.hdr.nlmsg_flags = NLM_F_REQUEST | NLM_F_DUMP; + req.hdr.nlmsg_seq = seq; + req.ifm.ifi_family = AF_UNSPEC; + + return NetlinkRequestResponse(fd, &req, sizeof(req), fn, false); +} + TEST(NetlinkRouteTest, GetLinkDump) { FileDescriptor fd = ASSERT_NO_ERRNO_AND_VALUE(NetlinkBoundSocket(NETLINK_ROUTE)); uint32_t port = ASSERT_NO_ERRNO_AND_VALUE(NetlinkPortID(fd.get())); + // Loopback is common among all tests, check that it's found. + bool loopbackFound = false; + ASSERT_NO_ERRNO(DumpLinks(fd, kSeq, [&](const struct nlmsghdr* hdr) { + CheckGetLinkResponse(hdr, kSeq, port); + if (hdr->nlmsg_type != RTM_NEWLINK) { + return; + } + ASSERT_GE(hdr->nlmsg_len, NLMSG_SPACE(sizeof(struct ifinfomsg))); + const struct ifinfomsg* msg = + reinterpret_cast(NLMSG_DATA(hdr)); + std::cout << "Found interface idx=" << msg->ifi_index + << ", type=" << std::hex << msg->ifi_type; + if (msg->ifi_type == ARPHRD_LOOPBACK) { + loopbackFound = true; + EXPECT_NE(msg->ifi_flags & IFF_LOOPBACK, 0); + } + })); + EXPECT_TRUE(loopbackFound); +} + +struct Link { + int index; + std::string name; +}; + +PosixErrorOr> FindLoopbackLink() { + ASSIGN_OR_RETURN_ERRNO(FileDescriptor fd, NetlinkBoundSocket(NETLINK_ROUTE)); + + absl::optional link; + RETURN_IF_ERRNO(DumpLinks(fd, kSeq, [&](const struct nlmsghdr* hdr) { + if (hdr->nlmsg_type != RTM_NEWLINK || + hdr->nlmsg_len < NLMSG_SPACE(sizeof(struct ifinfomsg))) { + return; + } + const struct ifinfomsg* msg = + reinterpret_cast(NLMSG_DATA(hdr)); + if (msg->ifi_type == ARPHRD_LOOPBACK) { + const auto* rta = FindRtAttr(hdr, msg, IFLA_IFNAME); + if (rta == nullptr) { + // Ignore links that do not have a name. + return; + } + + link = Link(); + link->index = msg->ifi_index; + link->name = std::string(reinterpret_cast(RTA_DATA(rta))); + } + })); + return link; +} + +// CheckLinkMsg checks a netlink message against an expected link. +void CheckLinkMsg(const struct nlmsghdr* hdr, const Link& link) { + ASSERT_THAT(hdr->nlmsg_type, Eq(RTM_NEWLINK)); + ASSERT_GE(hdr->nlmsg_len, NLMSG_SPACE(sizeof(struct ifinfomsg))); + const struct ifinfomsg* msg = + reinterpret_cast(NLMSG_DATA(hdr)); + EXPECT_EQ(msg->ifi_index, link.index); + + const struct rtattr* rta = FindRtAttr(hdr, msg, IFLA_IFNAME); + EXPECT_NE(nullptr, rta) << "IFLA_IFNAME not found in message."; + if (rta != nullptr) { + std::string name(reinterpret_cast(RTA_DATA(rta))); + EXPECT_EQ(name, link.name); + } +} + +TEST(NetlinkRouteTest, GetLinkByIndex) { + absl::optional loopback_link = + ASSERT_NO_ERRNO_AND_VALUE(FindLoopbackLink()); + ASSERT_TRUE(loopback_link.has_value()); + + FileDescriptor fd = + ASSERT_NO_ERRNO_AND_VALUE(NetlinkBoundSocket(NETLINK_ROUTE)); + struct request { struct nlmsghdr hdr; struct ifinfomsg ifm; }; - constexpr uint32_t kSeq = 12345; - struct request req = {}; req.hdr.nlmsg_len = sizeof(req); req.hdr.nlmsg_type = RTM_GETLINK; - req.hdr.nlmsg_flags = NLM_F_REQUEST | NLM_F_DUMP; + req.hdr.nlmsg_flags = NLM_F_REQUEST; req.hdr.nlmsg_seq = kSeq; req.ifm.ifi_family = AF_UNSPEC; + req.ifm.ifi_index = loopback_link->index; - // Loopback is common among all tests, check that it's found. - bool loopbackFound = false; + bool found = false; ASSERT_NO_ERRNO(NetlinkRequestResponse( fd, &req, sizeof(req), [&](const struct nlmsghdr* hdr) { - CheckGetLinkResponse(hdr, kSeq, port); - if (hdr->nlmsg_type != RTM_NEWLINK) { - return; - } - ASSERT_GE(hdr->nlmsg_len, NLMSG_SPACE(sizeof(struct ifinfomsg))); - const struct ifinfomsg* msg = - reinterpret_cast(NLMSG_DATA(hdr)); - std::cout << "Found interface idx=" << msg->ifi_index - << ", type=" << std::hex << msg->ifi_type; - if (msg->ifi_type == ARPHRD_LOOPBACK) { - loopbackFound = true; - EXPECT_NE(msg->ifi_flags & IFF_LOOPBACK, 0); - } + CheckLinkMsg(hdr, *loopback_link); + found = true; }, false)); - EXPECT_TRUE(loopbackFound); + EXPECT_TRUE(found) << "Netlink response does not contain any links."; } -TEST(NetlinkRouteTest, MsgHdrMsgUnsuppType) { +TEST(NetlinkRouteTest, GetLinkByName) { + absl::optional loopback_link = + ASSERT_NO_ERRNO_AND_VALUE(FindLoopbackLink()); + ASSERT_TRUE(loopback_link.has_value()); + FileDescriptor fd = ASSERT_NO_ERRNO_AND_VALUE(NetlinkBoundSocket(NETLINK_ROUTE)); struct request { struct nlmsghdr hdr; struct ifinfomsg ifm; + struct rtattr rtattr; + char ifname[IFNAMSIZ]; + char pad[NLMSG_ALIGNTO + RTA_ALIGNTO]; }; - constexpr uint32_t kSeq = 12345; + struct request req = {}; + req.hdr.nlmsg_type = RTM_GETLINK; + req.hdr.nlmsg_flags = NLM_F_REQUEST; + req.hdr.nlmsg_seq = kSeq; + req.ifm.ifi_family = AF_UNSPEC; + req.rtattr.rta_type = IFLA_IFNAME; + req.rtattr.rta_len = RTA_LENGTH(loopback_link->name.size() + 1); + strncpy(req.ifname, loopback_link->name.c_str(), sizeof(req.ifname)); + req.hdr.nlmsg_len = + NLMSG_LENGTH(sizeof(req.ifm)) + NLMSG_ALIGN(req.rtattr.rta_len); + + bool found = false; + ASSERT_NO_ERRNO(NetlinkRequestResponse( + fd, &req, sizeof(req), + [&](const struct nlmsghdr* hdr) { + CheckLinkMsg(hdr, *loopback_link); + found = true; + }, + false)); + EXPECT_TRUE(found) << "Netlink response does not contain any links."; +} + +TEST(NetlinkRouteTest, GetLinkByIndexNotFound) { + FileDescriptor fd = + ASSERT_NO_ERRNO_AND_VALUE(NetlinkBoundSocket(NETLINK_ROUTE)); + + struct request { + struct nlmsghdr hdr; + struct ifinfomsg ifm; + }; + + struct request req = {}; + req.hdr.nlmsg_len = sizeof(req); + req.hdr.nlmsg_type = RTM_GETLINK; + req.hdr.nlmsg_flags = NLM_F_REQUEST; + req.hdr.nlmsg_seq = kSeq; + req.ifm.ifi_family = AF_UNSPEC; + req.ifm.ifi_index = 1234590; + + EXPECT_THAT(NetlinkRequestAckOrError(fd, kSeq, &req, sizeof(req)), + PosixErrorIs(ENODEV, ::testing::_)); +} + +TEST(NetlinkRouteTest, GetLinkByNameNotFound) { + const std::string name = "nodevice?!"; + + FileDescriptor fd = + ASSERT_NO_ERRNO_AND_VALUE(NetlinkBoundSocket(NETLINK_ROUTE)); + + struct request { + struct nlmsghdr hdr; + struct ifinfomsg ifm; + struct rtattr rtattr; + char ifname[IFNAMSIZ]; + char pad[NLMSG_ALIGNTO + RTA_ALIGNTO]; + }; + + struct request req = {}; + req.hdr.nlmsg_type = RTM_GETLINK; + req.hdr.nlmsg_flags = NLM_F_REQUEST; + req.hdr.nlmsg_seq = kSeq; + req.ifm.ifi_family = AF_UNSPEC; + req.rtattr.rta_type = IFLA_IFNAME; + req.rtattr.rta_len = RTA_LENGTH(name.size() + 1); + strncpy(req.ifname, name.c_str(), sizeof(req.ifname)); + req.hdr.nlmsg_len = + NLMSG_LENGTH(sizeof(req.ifm)) + NLMSG_ALIGN(req.rtattr.rta_len); + + EXPECT_THAT(NetlinkRequestAckOrError(fd, kSeq, &req, sizeof(req)), + PosixErrorIs(ENODEV, ::testing::_)); +} + +TEST(NetlinkRouteTest, MsgHdrMsgUnsuppType) { + FileDescriptor fd = + ASSERT_NO_ERRNO_AND_VALUE(NetlinkBoundSocket(NETLINK_ROUTE)); + + struct request { + struct nlmsghdr hdr; + struct ifinfomsg ifm; + }; struct request req = {}; req.hdr.nlmsg_len = sizeof(req); @@ -175,18 +346,8 @@ TEST(NetlinkRouteTest, MsgHdrMsgUnsuppType) { req.hdr.nlmsg_seq = kSeq; req.ifm.ifi_family = AF_UNSPEC; - ASSERT_NO_ERRNO(NetlinkRequestResponse( - fd, &req, sizeof(req), - [&](const struct nlmsghdr* hdr) { - EXPECT_THAT(hdr->nlmsg_type, Eq(NLMSG_ERROR)); - EXPECT_EQ(hdr->nlmsg_seq, kSeq); - EXPECT_GE(hdr->nlmsg_len, sizeof(*hdr) + sizeof(struct nlmsgerr)); - - const struct nlmsgerr* msg = - reinterpret_cast(NLMSG_DATA(hdr)); - EXPECT_EQ(msg->error, -EOPNOTSUPP); - }, - true)); + EXPECT_THAT(NetlinkRequestAckOrError(fd, kSeq, &req, sizeof(req)), + PosixErrorIs(EOPNOTSUPP, ::testing::_)); } TEST(NetlinkRouteTest, MsgHdrMsgTrunc) { @@ -198,8 +359,6 @@ TEST(NetlinkRouteTest, MsgHdrMsgTrunc) { struct ifinfomsg ifm; }; - constexpr uint32_t kSeq = 12345; - struct request req = {}; req.hdr.nlmsg_len = sizeof(req); req.hdr.nlmsg_type = RTM_GETLINK; @@ -238,8 +397,6 @@ TEST(NetlinkRouteTest, MsgTruncMsgHdrMsgTrunc) { struct ifinfomsg ifm; }; - constexpr uint32_t kSeq = 12345; - struct request req = {}; req.hdr.nlmsg_len = sizeof(req); req.hdr.nlmsg_type = RTM_GETLINK; @@ -282,8 +439,6 @@ TEST(NetlinkRouteTest, ControlMessageIgnored) { struct ifinfomsg ifm; }; - constexpr uint32_t kSeq = 12345; - struct request req = {}; // This control message is ignored. We still receive a response for the @@ -317,8 +472,6 @@ TEST(NetlinkRouteTest, GetAddrDump) { struct rtgenmsg rgm; }; - constexpr uint32_t kSeq = 12345; - struct request req; req.hdr.nlmsg_len = sizeof(req); req.hdr.nlmsg_type = RTM_GETADDR; @@ -367,6 +520,57 @@ TEST(NetlinkRouteTest, LookupAll) { ASSERT_GT(count, 0); } +TEST(NetlinkRouteTest, AddAddr) { + SKIP_IF(!ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_NET_ADMIN))); + + absl::optional loopback_link = + ASSERT_NO_ERRNO_AND_VALUE(FindLoopbackLink()); + ASSERT_TRUE(loopback_link.has_value()); + + FileDescriptor fd = + ASSERT_NO_ERRNO_AND_VALUE(NetlinkBoundSocket(NETLINK_ROUTE)); + + struct request { + struct nlmsghdr hdr; + struct ifaddrmsg ifa; + struct rtattr rtattr; + struct in_addr addr; + char pad[NLMSG_ALIGNTO + RTA_ALIGNTO]; + }; + + struct request req = {}; + req.hdr.nlmsg_type = RTM_NEWADDR; + req.hdr.nlmsg_seq = kSeq; + req.ifa.ifa_family = AF_INET; + req.ifa.ifa_prefixlen = 24; + req.ifa.ifa_flags = 0; + req.ifa.ifa_scope = 0; + req.ifa.ifa_index = loopback_link->index; + req.rtattr.rta_type = IFA_LOCAL; + req.rtattr.rta_len = RTA_LENGTH(sizeof(req.addr)); + inet_pton(AF_INET, "10.0.0.1", &req.addr); + req.hdr.nlmsg_len = + NLMSG_LENGTH(sizeof(req.ifa)) + NLMSG_ALIGN(req.rtattr.rta_len); + + // Create should succeed, as no such address in kernel. + req.hdr.nlmsg_flags = NLM_F_REQUEST | NLM_F_CREATE | NLM_F_ACK; + EXPECT_NO_ERRNO( + NetlinkRequestAckOrError(fd, req.hdr.nlmsg_seq, &req, req.hdr.nlmsg_len)); + + // Replace an existing address should succeed. + req.hdr.nlmsg_flags = NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_ACK; + req.hdr.nlmsg_seq++; + EXPECT_NO_ERRNO( + NetlinkRequestAckOrError(fd, req.hdr.nlmsg_seq, &req, req.hdr.nlmsg_len)); + + // Create exclusive should fail, as we created the address above. + req.hdr.nlmsg_flags = NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL | NLM_F_ACK; + req.hdr.nlmsg_seq++; + EXPECT_THAT( + NetlinkRequestAckOrError(fd, req.hdr.nlmsg_seq, &req, req.hdr.nlmsg_len), + PosixErrorIs(EEXIST, ::testing::_)); +} + // GetRouteDump tests a RTM_GETROUTE + NLM_F_DUMP request. TEST(NetlinkRouteTest, GetRouteDump) { FileDescriptor fd = @@ -378,8 +582,6 @@ TEST(NetlinkRouteTest, GetRouteDump) { struct rtmsg rtm; }; - constexpr uint32_t kSeq = 12345; - struct request req = {}; req.hdr.nlmsg_len = sizeof(req); req.hdr.nlmsg_type = RTM_GETROUTE; @@ -538,8 +740,6 @@ TEST(NetlinkRouteTest, RecvmsgTrunc) { struct rtgenmsg rgm; }; - constexpr uint32_t kSeq = 12345; - struct request req; req.hdr.nlmsg_len = sizeof(req); req.hdr.nlmsg_type = RTM_GETADDR; @@ -615,8 +815,6 @@ TEST(NetlinkRouteTest, RecvmsgTruncPeek) { struct rtgenmsg rgm; }; - constexpr uint32_t kSeq = 12345; - struct request req; req.hdr.nlmsg_len = sizeof(req); req.hdr.nlmsg_type = RTM_GETADDR; @@ -695,8 +893,6 @@ TEST(NetlinkRouteTest, NoPasscredNoCreds) { struct rtgenmsg rgm; }; - constexpr uint32_t kSeq = 12345; - struct request req; req.hdr.nlmsg_len = sizeof(req); req.hdr.nlmsg_type = RTM_GETADDR; @@ -743,8 +939,6 @@ TEST(NetlinkRouteTest, PasscredCreds) { struct rtgenmsg rgm; }; - constexpr uint32_t kSeq = 12345; - struct request req; req.hdr.nlmsg_len = sizeof(req); req.hdr.nlmsg_type = RTM_GETADDR; diff --git a/test/syscalls/linux/socket_netlink_util.cc b/test/syscalls/linux/socket_netlink_util.cc index cd2212a1a..952eecfe8 100644 --- a/test/syscalls/linux/socket_netlink_util.cc +++ b/test/syscalls/linux/socket_netlink_util.cc @@ -16,6 +16,7 @@ #include #include +#include #include #include @@ -71,9 +72,10 @@ PosixError NetlinkRequestResponse( iov.iov_base = buf.data(); iov.iov_len = buf.size(); - // Response is a series of NLM_F_MULTI messages, ending with a NLMSG_DONE - // message. + // If NLM_F_MULTI is set, response is a series of messages that ends with a + // NLMSG_DONE message. int type = -1; + int flags = 0; do { int len; RETURN_ERROR_IF_SYSCALL_FAIL(len = RetryEINTR(recvmsg)(fd.get(), &msg, 0)); @@ -89,6 +91,7 @@ PosixError NetlinkRequestResponse( for (struct nlmsghdr* hdr = reinterpret_cast(buf.data()); NLMSG_OK(hdr, len); hdr = NLMSG_NEXT(hdr, len)) { fn(hdr); + flags = hdr->nlmsg_flags; type = hdr->nlmsg_type; // Done should include an integer payload for dump_done_errno. // See net/netlink/af_netlink.c:netlink_dump @@ -98,11 +101,11 @@ PosixError NetlinkRequestResponse( EXPECT_GE(hdr->nlmsg_len, NLMSG_LENGTH(sizeof(int))); } } - } while (type != NLMSG_DONE && type != NLMSG_ERROR); + } while ((flags & NLM_F_MULTI) && type != NLMSG_DONE && type != NLMSG_ERROR); if (expect_nlmsgerr) { EXPECT_EQ(type, NLMSG_ERROR); - } else { + } else if (flags & NLM_F_MULTI) { EXPECT_EQ(type, NLMSG_DONE); } return NoError(); @@ -146,5 +149,39 @@ PosixError NetlinkRequestResponseSingle( return NoError(); } +PosixError NetlinkRequestAckOrError(const FileDescriptor& fd, uint32_t seq, + void* request, size_t len) { + // Dummy negative number for no error message received. + // We won't get a negative error number so there will be no confusion. + int err = -42; + RETURN_IF_ERRNO(NetlinkRequestResponse( + fd, request, len, + [&](const struct nlmsghdr* hdr) { + EXPECT_EQ(NLMSG_ERROR, hdr->nlmsg_type); + EXPECT_EQ(hdr->nlmsg_seq, seq); + EXPECT_GE(hdr->nlmsg_len, sizeof(*hdr) + sizeof(struct nlmsgerr)); + + const struct nlmsgerr* msg = + reinterpret_cast(NLMSG_DATA(hdr)); + err = -msg->error; + }, + true)); + return PosixError(err); +} + +const struct rtattr* FindRtAttr(const struct nlmsghdr* hdr, + const struct ifinfomsg* msg, int16_t attr) { + const int ifi_space = NLMSG_SPACE(sizeof(*msg)); + int attrlen = hdr->nlmsg_len - ifi_space; + const struct rtattr* rta = reinterpret_cast( + reinterpret_cast(hdr) + NLMSG_ALIGN(ifi_space)); + for (; RTA_OK(rta, attrlen); rta = RTA_NEXT(rta, attrlen)) { + if (rta->rta_type == attr) { + return rta; + } + } + return nullptr; +} + } // namespace testing } // namespace gvisor diff --git a/test/syscalls/linux/socket_netlink_util.h b/test/syscalls/linux/socket_netlink_util.h index 3678c0599..e13ead406 100644 --- a/test/syscalls/linux/socket_netlink_util.h +++ b/test/syscalls/linux/socket_netlink_util.h @@ -19,6 +19,7 @@ // socket.h has to be included before if_arp.h. #include #include +#include #include "test/util/file_descriptor.h" #include "test/util/posix_error.h" @@ -47,6 +48,14 @@ PosixError NetlinkRequestResponseSingle( const FileDescriptor& fd, void* request, size_t len, const std::function& fn); +// Send the passed request then expect and return an ack or error. +PosixError NetlinkRequestAckOrError(const FileDescriptor& fd, uint32_t seq, + void* request, size_t len); + +// Find rtnetlink attribute in message. +const struct rtattr* FindRtAttr(const struct nlmsghdr* hdr, + const struct ifinfomsg* msg, int16_t attr); + } // namespace testing } // namespace gvisor -- cgit v1.2.3 From f3d95607036b8a502c65aa7b3e8145227274dbbc Mon Sep 17 00:00:00 2001 From: Eyal Soha Date: Wed, 5 Feb 2020 17:56:00 -0800 Subject: recv() on a closed TCP socket returns ENOTCONN From RFC 793 s3.9 p58 Event Processing: If RECEIVE Call arrives in CLOSED state and the user has access to such a connection, the return should be "error: connection does not exist" Fixes #1598 PiperOrigin-RevId: 293494287 --- pkg/sentry/socket/netstack/netstack.go | 7 ++++++- pkg/tcpip/tcpip.go | 4 ++++ pkg/tcpip/transport/tcp/endpoint.go | 4 ++-- pkg/tcpip/transport/tcp/tcp_test.go | 9 ++++----- test/syscalls/linux/tcp_socket.cc | 9 +++++++++ 5 files changed, 25 insertions(+), 8 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index 049d04bf2..ed2fbcceb 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -2229,11 +2229,16 @@ func (s *SocketOperations) coalescingRead(ctx context.Context, dst usermem.IOSeq var copied int // Copy as many views as possible into the user-provided buffer. - for dst.NumBytes() != 0 { + for { + // Always do at least one fetchReadView, even if the number of bytes to + // read is 0. err = s.fetchReadView() if err != nil { break } + if dst.NumBytes() == 0 { + break + } var n int var e error diff --git a/pkg/tcpip/tcpip.go b/pkg/tcpip/tcpip.go index 0fa141d58..d29d9a704 100644 --- a/pkg/tcpip/tcpip.go +++ b/pkg/tcpip/tcpip.go @@ -1124,6 +1124,10 @@ type ReadErrors struct { // InvalidEndpointState is the number of times we found the endpoint state // to be unexpected. InvalidEndpointState StatCounter + + // NotConnected is the number of times we tried to read but found that the + // endpoint was not connected. + NotConnected StatCounter } // WriteErrors collects packet write errors from an endpoint write call. diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go index b5a8e15ee..e4a6b1b8b 100644 --- a/pkg/tcpip/transport/tcp/endpoint.go +++ b/pkg/tcpip/transport/tcp/endpoint.go @@ -1003,8 +1003,8 @@ func (e *endpoint) Read(*tcpip.FullAddress) (buffer.View, tcpip.ControlMessages, if s == StateError { return buffer.View{}, tcpip.ControlMessages{}, he } - e.stats.ReadErrors.InvalidEndpointState.Increment() - return buffer.View{}, tcpip.ControlMessages{}, tcpip.ErrInvalidEndpointState + e.stats.ReadErrors.NotConnected.Increment() + return buffer.View{}, tcpip.ControlMessages{}, tcpip.ErrNotConnected } v, err := e.readLocked() diff --git a/pkg/tcpip/transport/tcp/tcp_test.go b/pkg/tcpip/transport/tcp/tcp_test.go index 2c1505067..cc118c993 100644 --- a/pkg/tcpip/transport/tcp/tcp_test.go +++ b/pkg/tcpip/transport/tcp/tcp_test.go @@ -5405,12 +5405,11 @@ func TestEndpointBindListenAcceptState(t *testing.T) { t.Errorf("Unexpected endpoint state: want %v, got %v", want, got) } - // Expect InvalidEndpointState errors on a read at this point. - if _, _, err := ep.Read(nil); err != tcpip.ErrInvalidEndpointState { - t.Fatalf("got c.EP.Read(nil) = %v, want = %v", err, tcpip.ErrInvalidEndpointState) + if _, _, err := ep.Read(nil); err != tcpip.ErrNotConnected { + t.Errorf("got c.EP.Read(nil) = %v, want = %v", err, tcpip.ErrNotConnected) } - if got := ep.Stats().(*tcp.Stats).ReadErrors.InvalidEndpointState.Value(); got != 1 { - t.Fatalf("got EP stats Stats.ReadErrors.InvalidEndpointState got %v want %v", got, 1) + if got := ep.Stats().(*tcp.Stats).ReadErrors.NotConnected.Value(); got != 1 { + t.Errorf("got EP stats Stats.ReadErrors.NotConnected got %v want %v", got, 1) } if err := ep.Listen(10); err != nil { diff --git a/test/syscalls/linux/tcp_socket.cc b/test/syscalls/linux/tcp_socket.cc index 525ccbd88..8a8b68e75 100644 --- a/test/syscalls/linux/tcp_socket.cc +++ b/test/syscalls/linux/tcp_socket.cc @@ -1339,6 +1339,15 @@ TEST_P(SimpleTcpSocketTest, SetTCPDeferAcceptGreaterThanZero) { EXPECT_EQ(get, kTCPDeferAccept); } +TEST_P(SimpleTcpSocketTest, RecvOnClosedSocket) { + auto s = + ASSERT_NO_ERRNO_AND_VALUE(Socket(GetParam(), SOCK_STREAM, IPPROTO_TCP)); + char buf[1]; + EXPECT_THAT(recv(s.get(), buf, 0, 0), SyscallFailsWithErrno(ENOTCONN)); + EXPECT_THAT(recv(s.get(), buf, sizeof(buf), 0), + SyscallFailsWithErrno(ENOTCONN)); +} + INSTANTIATE_TEST_SUITE_P(AllInetTests, SimpleTcpSocketTest, ::testing::Values(AF_INET, AF_INET6)); -- cgit v1.2.3 From 1b6a12a768216a99a5e0428c42ea4faf79cf3b50 Mon Sep 17 00:00:00 2001 From: Adin Scannell Date: Wed, 5 Feb 2020 22:45:44 -0800 Subject: Add notes to relevant tests. These were out-of-band notes that can help provide additional context and simplify automated imports. PiperOrigin-RevId: 293525915 --- pkg/metric/metric.go | 1 - pkg/sentry/arch/arch_x86.go | 4 ++ pkg/sentry/arch/signal_amd64.go | 2 +- pkg/sentry/fs/file_overlay_test.go | 1 + pkg/sentry/fs/proc/README.md | 4 ++ pkg/sentry/kernel/BUILD | 1 + pkg/sentry/kernel/kernel.go | 3 ++ pkg/sentry/kernel/kernel_opts.go | 20 +++++++ pkg/sentry/socket/hostinet/BUILD | 1 + pkg/sentry/socket/hostinet/socket.go | 5 +- pkg/sentry/socket/hostinet/sockopt_impl.go | 27 ++++++++++ pkg/tcpip/transport/tcp/endpoint.go | 3 ++ runsc/boot/filter/BUILD | 1 + runsc/boot/filter/config.go | 13 ----- runsc/boot/filter/config_profile.go | 34 ++++++++++++ runsc/container/console_test.go | 5 +- runsc/dockerutil/dockerutil.go | 11 ++-- runsc/testutil/BUILD | 5 +- runsc/testutil/testutil.go | 54 ------------------- runsc/testutil/testutil_runfiles.go | 75 +++++++++++++++++++++++++++ test/image/image_test.go | 8 +-- test/syscalls/build_defs.bzl | 35 +++++++++++-- test/syscalls/linux/chroot.cc | 2 +- test/syscalls/linux/concurrency.cc | 3 +- test/syscalls/linux/exec_proc_exe_workload.cc | 6 +++ test/syscalls/linux/fork.cc | 5 +- test/syscalls/linux/mmap.cc | 8 +-- test/syscalls/linux/open_create.cc | 1 + test/syscalls/linux/preadv.cc | 1 + test/syscalls/linux/proc.cc | 46 +++++++++++++--- test/syscalls/linux/readv.cc | 4 +- test/syscalls/linux/rseq.cc | 2 +- test/syscalls/linux/select.cc | 2 +- test/syscalls/linux/shm.cc | 2 +- test/syscalls/linux/sigprocmask.cc | 2 +- test/syscalls/linux/socket_unix_non_stream.cc | 4 +- test/syscalls/linux/symlink.cc | 2 +- test/syscalls/linux/tcp_socket.cc | 3 +- test/syscalls/linux/time.cc | 1 + test/syscalls/linux/tkill.cc | 2 +- test/util/temp_path.cc | 1 + tools/build/tags.bzl | 4 ++ tools/defs.bzl | 17 +++++- 43 files changed, 318 insertions(+), 113 deletions(-) create mode 100644 pkg/sentry/kernel/kernel_opts.go create mode 100644 pkg/sentry/socket/hostinet/sockopt_impl.go create mode 100644 runsc/boot/filter/config_profile.go create mode 100644 runsc/testutil/testutil_runfiles.go (limited to 'pkg/sentry/socket') diff --git a/pkg/metric/metric.go b/pkg/metric/metric.go index 93d4f2b8c..006fcd9ab 100644 --- a/pkg/metric/metric.go +++ b/pkg/metric/metric.go @@ -46,7 +46,6 @@ var ( // // TODO(b/67298402): Support non-cumulative metrics. // TODO(b/67298427): Support metric fields. -// type Uint64Metric struct { // value is the actual value of the metric. It must be accessed // atomically. diff --git a/pkg/sentry/arch/arch_x86.go b/pkg/sentry/arch/arch_x86.go index a18093155..3db8bd34b 100644 --- a/pkg/sentry/arch/arch_x86.go +++ b/pkg/sentry/arch/arch_x86.go @@ -114,6 +114,10 @@ func newX86FPStateSlice() []byte { size, align := cpuid.HostFeatureSet().ExtendedStateSize() capacity := size // Always use at least 4096 bytes. + // + // For the KVM platform, this state is a fixed 4096 bytes, so make sure + // that the underlying array is at _least_ that size otherwise we will + // corrupt random memory. This is not a pleasant thing to debug. if capacity < 4096 { capacity = 4096 } diff --git a/pkg/sentry/arch/signal_amd64.go b/pkg/sentry/arch/signal_amd64.go index 81b92bb43..6fb756f0e 100644 --- a/pkg/sentry/arch/signal_amd64.go +++ b/pkg/sentry/arch/signal_amd64.go @@ -55,7 +55,7 @@ type SignalContext64 struct { Trapno uint64 Oldmask linux.SignalSet Cr2 uint64 - // Pointer to a struct _fpstate. + // Pointer to a struct _fpstate. See b/33003106#comment8. Fpstate uint64 Reserved [8]uint64 } diff --git a/pkg/sentry/fs/file_overlay_test.go b/pkg/sentry/fs/file_overlay_test.go index 02538bb4f..a76d87e3a 100644 --- a/pkg/sentry/fs/file_overlay_test.go +++ b/pkg/sentry/fs/file_overlay_test.go @@ -177,6 +177,7 @@ func TestReaddirRevalidation(t *testing.T) { // TestReaddirOverlayFrozen tests that calling Readdir on an overlay file with // a frozen dirent tree does not make Readdir calls to the underlying files. +// This is a regression test for b/114808269. func TestReaddirOverlayFrozen(t *testing.T) { ctx := contexttest.Context(t) diff --git a/pkg/sentry/fs/proc/README.md b/pkg/sentry/fs/proc/README.md index 5d4ec6c7b..6667a0916 100644 --- a/pkg/sentry/fs/proc/README.md +++ b/pkg/sentry/fs/proc/README.md @@ -11,6 +11,8 @@ inconsistency, please file a bug. The following files are implemented: + + | File /proc/ | Content | | :------------------------ | :---------------------------------------------------- | | [cpuinfo](#cpuinfo) | Info about the CPU | @@ -22,6 +24,8 @@ The following files are implemented: | [uptime](#uptime) | Wall clock since boot, combined idle time of all cpus | | [version](#version) | Kernel version | + + ### cpuinfo ```bash diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD index a27628c0a..2231d6973 100644 --- a/pkg/sentry/kernel/BUILD +++ b/pkg/sentry/kernel/BUILD @@ -91,6 +91,7 @@ go_library( "fs_context.go", "ipc_namespace.go", "kernel.go", + "kernel_opts.go", "kernel_state.go", "pending_signals.go", "pending_signals_list.go", diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index dcd6e91c4..3ee760ba2 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -235,6 +235,9 @@ type Kernel struct { // events. This is initialized lazily on the first unimplemented // syscall. unimplementedSyscallEmitter eventchannel.Emitter `state:"nosave"` + + // SpecialOpts contains special kernel options. + SpecialOpts } // InitKernelArgs holds arguments to Init. diff --git a/pkg/sentry/kernel/kernel_opts.go b/pkg/sentry/kernel/kernel_opts.go new file mode 100644 index 000000000..2e66ec587 --- /dev/null +++ b/pkg/sentry/kernel/kernel_opts.go @@ -0,0 +1,20 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package kernel + +// SpecialOpts contains non-standard options for the kernel. +// +// +stateify savable +type SpecialOpts struct{} diff --git a/pkg/sentry/socket/hostinet/BUILD b/pkg/sentry/socket/hostinet/BUILD index 5a07d5d0e..023bad156 100644 --- a/pkg/sentry/socket/hostinet/BUILD +++ b/pkg/sentry/socket/hostinet/BUILD @@ -10,6 +10,7 @@ go_library( "save_restore.go", "socket.go", "socket_unsafe.go", + "sockopt_impl.go", "stack.go", ], visibility = ["//pkg/sentry:internal"], diff --git a/pkg/sentry/socket/hostinet/socket.go b/pkg/sentry/socket/hostinet/socket.go index 34f63986f..de76388ac 100644 --- a/pkg/sentry/socket/hostinet/socket.go +++ b/pkg/sentry/socket/hostinet/socket.go @@ -285,7 +285,7 @@ func (s *socketOperations) GetSockOpt(t *kernel.Task, level int, name int, outPt } // Whitelist options and constrain option length. - var optlen int + optlen := getSockOptLen(t, level, name) switch level { case linux.SOL_IP: switch name { @@ -330,7 +330,7 @@ func (s *socketOperations) GetSockOpt(t *kernel.Task, level int, name int, outPt // SetSockOpt implements socket.Socket.SetSockOpt. func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *syserr.Error { // Whitelist options and constrain option length. - var optlen int + optlen := setSockOptLen(t, level, name) switch level { case linux.SOL_IP: switch name { @@ -353,6 +353,7 @@ func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt [ optlen = sizeofInt32 } } + if optlen == 0 { // Pretend to accept socket options we don't understand. This seems // dangerous, but it's what netstack does... diff --git a/pkg/sentry/socket/hostinet/sockopt_impl.go b/pkg/sentry/socket/hostinet/sockopt_impl.go new file mode 100644 index 000000000..8a783712e --- /dev/null +++ b/pkg/sentry/socket/hostinet/sockopt_impl.go @@ -0,0 +1,27 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package hostinet + +import ( + "gvisor.dev/gvisor/pkg/sentry/kernel" +) + +func getSockOptLen(t *kernel.Task, level, name int) int { + return 0 // No custom options. +} + +func setSockOptLen(t *kernel.Task, level, name int) int { + return 0 // No custom options. +} diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go index e4a6b1b8b..f2be0e651 100644 --- a/pkg/tcpip/transport/tcp/endpoint.go +++ b/pkg/tcpip/transport/tcp/endpoint.go @@ -2166,6 +2166,9 @@ func (e *endpoint) listen(backlog int) *tcpip.Error { e.isRegistered = true e.setEndpointState(StateListen) + // The channel may be non-nil when we're restoring the endpoint, and it + // may be pre-populated with some previously accepted (but not Accepted) + // endpoints. if e.acceptedChan == nil { e.acceptedChan = make(chan *endpoint, backlog) } diff --git a/runsc/boot/filter/BUILD b/runsc/boot/filter/BUILD index ce30f6c53..ed18f0047 100644 --- a/runsc/boot/filter/BUILD +++ b/runsc/boot/filter/BUILD @@ -8,6 +8,7 @@ go_library( "config.go", "config_amd64.go", "config_arm64.go", + "config_profile.go", "extra_filters.go", "extra_filters_msan.go", "extra_filters_race.go", diff --git a/runsc/boot/filter/config.go b/runsc/boot/filter/config.go index f8d351c7b..c69f4c602 100644 --- a/runsc/boot/filter/config.go +++ b/runsc/boot/filter/config.go @@ -536,16 +536,3 @@ func controlServerFilters(fd int) seccomp.SyscallRules { }, } } - -// profileFilters returns extra syscalls made by runtime/pprof package. -func profileFilters() seccomp.SyscallRules { - return seccomp.SyscallRules{ - syscall.SYS_OPENAT: []seccomp.Rule{ - { - seccomp.AllowAny{}, - seccomp.AllowAny{}, - seccomp.AllowValue(syscall.O_RDONLY | syscall.O_LARGEFILE | syscall.O_CLOEXEC), - }, - }, - } -} diff --git a/runsc/boot/filter/config_profile.go b/runsc/boot/filter/config_profile.go new file mode 100644 index 000000000..194952a7b --- /dev/null +++ b/runsc/boot/filter/config_profile.go @@ -0,0 +1,34 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package filter + +import ( + "syscall" + + "gvisor.dev/gvisor/pkg/seccomp" +) + +// profileFilters returns extra syscalls made by runtime/pprof package. +func profileFilters() seccomp.SyscallRules { + return seccomp.SyscallRules{ + syscall.SYS_OPENAT: []seccomp.Rule{ + { + seccomp.AllowAny{}, + seccomp.AllowAny{}, + seccomp.AllowValue(syscall.O_RDONLY | syscall.O_LARGEFILE | syscall.O_CLOEXEC), + }, + }, + } +} diff --git a/runsc/container/console_test.go b/runsc/container/console_test.go index 060b63bf3..c2518d52b 100644 --- a/runsc/container/console_test.go +++ b/runsc/container/console_test.go @@ -196,7 +196,10 @@ func TestJobControlSignalExec(t *testing.T) { defer ptyMaster.Close() defer ptySlave.Close() - // Exec bash and attach a terminal. + // Exec bash and attach a terminal. Note that occasionally /bin/sh + // may be a different shell or have a different configuration (such + // as disabling interactive mode and job control). Since we want to + // explicitly test interactive mode, use /bin/bash. See b/116981926. execArgs := &control.ExecArgs{ Filename: "/bin/bash", // Don't let bash execute from profile or rc files, otherwise diff --git a/runsc/dockerutil/dockerutil.go b/runsc/dockerutil/dockerutil.go index 9b6346ca2..1ff5e8cc3 100644 --- a/runsc/dockerutil/dockerutil.go +++ b/runsc/dockerutil/dockerutil.go @@ -143,8 +143,11 @@ func PrepareFiles(names ...string) (string, error) { return "", fmt.Errorf("os.Chmod(%q, 0777) failed: %v", dir, err) } for _, name := range names { - src := getLocalPath(name) - dst := path.Join(dir, name) + src, err := testutil.FindFile(name) + if err != nil { + return "", fmt.Errorf("testutil.Preparefiles(%q) failed: %v", name, err) + } + dst := path.Join(dir, path.Base(name)) if err := testutil.Copy(src, dst); err != nil { return "", fmt.Errorf("testutil.Copy(%q, %q) failed: %v", src, dst, err) } @@ -152,10 +155,6 @@ func PrepareFiles(names ...string) (string, error) { return dir, nil } -func getLocalPath(file string) string { - return path.Join(".", file) -} - // do executes docker command. func do(args ...string) (string, error) { log.Printf("Running: docker %s\n", args) diff --git a/runsc/testutil/BUILD b/runsc/testutil/BUILD index f845120b0..945405303 100644 --- a/runsc/testutil/BUILD +++ b/runsc/testutil/BUILD @@ -5,7 +5,10 @@ package(licenses = ["notice"]) go_library( name = "testutil", testonly = 1, - srcs = ["testutil.go"], + srcs = [ + "testutil.go", + "testutil_runfiles.go", + ], visibility = ["//:sandbox"], deps = [ "//pkg/log", diff --git a/runsc/testutil/testutil.go b/runsc/testutil/testutil.go index edf2e809a..80c2c9680 100644 --- a/runsc/testutil/testutil.go +++ b/runsc/testutil/testutil.go @@ -79,60 +79,6 @@ func ConfigureExePath() error { return nil } -// FindFile searchs for a file inside the test run environment. It returns the -// full path to the file. It fails if none or more than one file is found. -func FindFile(path string) (string, error) { - wd, err := os.Getwd() - if err != nil { - return "", err - } - - // The test root is demarcated by a path element called "__main__". Search for - // it backwards from the working directory. - root := wd - for { - dir, name := filepath.Split(root) - if name == "__main__" { - break - } - if len(dir) == 0 { - return "", fmt.Errorf("directory __main__ not found in %q", wd) - } - // Remove ending slash to loop around. - root = dir[:len(dir)-1] - } - - // Annoyingly, bazel adds the build type to the directory path for go - // binaries, but not for c++ binaries. We use two different patterns to - // to find our file. - patterns := []string{ - // Try the obvious path first. - filepath.Join(root, path), - // If it was a go binary, use a wildcard to match the build - // type. The pattern is: /test-path/__main__/directories/*/file. - filepath.Join(root, filepath.Dir(path), "*", filepath.Base(path)), - } - - for _, p := range patterns { - matches, err := filepath.Glob(p) - if err != nil { - // "The only possible returned error is ErrBadPattern, - // when pattern is malformed." -godoc - return "", fmt.Errorf("error globbing %q: %v", p, err) - } - switch len(matches) { - case 0: - // Try the next pattern. - case 1: - // We found it. - return matches[0], nil - default: - return "", fmt.Errorf("more than one match found for %q: %s", path, matches) - } - } - return "", fmt.Errorf("file %q not found", path) -} - // TestConfig returns the default configuration to use in tests. Note that // 'RootDir' must be set by caller if required. func TestConfig() *boot.Config { diff --git a/runsc/testutil/testutil_runfiles.go b/runsc/testutil/testutil_runfiles.go new file mode 100644 index 000000000..ece9ea9a1 --- /dev/null +++ b/runsc/testutil/testutil_runfiles.go @@ -0,0 +1,75 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package testutil + +import ( + "fmt" + "os" + "path/filepath" +) + +// FindFile searchs for a file inside the test run environment. It returns the +// full path to the file. It fails if none or more than one file is found. +func FindFile(path string) (string, error) { + wd, err := os.Getwd() + if err != nil { + return "", err + } + + // The test root is demarcated by a path element called "__main__". Search for + // it backwards from the working directory. + root := wd + for { + dir, name := filepath.Split(root) + if name == "__main__" { + break + } + if len(dir) == 0 { + return "", fmt.Errorf("directory __main__ not found in %q", wd) + } + // Remove ending slash to loop around. + root = dir[:len(dir)-1] + } + + // Annoyingly, bazel adds the build type to the directory path for go + // binaries, but not for c++ binaries. We use two different patterns to + // to find our file. + patterns := []string{ + // Try the obvious path first. + filepath.Join(root, path), + // If it was a go binary, use a wildcard to match the build + // type. The pattern is: /test-path/__main__/directories/*/file. + filepath.Join(root, filepath.Dir(path), "*", filepath.Base(path)), + } + + for _, p := range patterns { + matches, err := filepath.Glob(p) + if err != nil { + // "The only possible returned error is ErrBadPattern, + // when pattern is malformed." -godoc + return "", fmt.Errorf("error globbing %q: %v", p, err) + } + switch len(matches) { + case 0: + // Try the next pattern. + case 1: + // We found it. + return matches[0], nil + default: + return "", fmt.Errorf("more than one match found for %q: %s", path, matches) + } + } + return "", fmt.Errorf("file %q not found", path) +} diff --git a/test/image/image_test.go b/test/image/image_test.go index d0dcb1861..0a1e19d6f 100644 --- a/test/image/image_test.go +++ b/test/image/image_test.go @@ -107,7 +107,7 @@ func TestHttpd(t *testing.T) { } d := dockerutil.MakeDocker("http-test") - dir, err := dockerutil.PrepareFiles("latin10k.txt") + dir, err := dockerutil.PrepareFiles("test/image/latin10k.txt") if err != nil { t.Fatalf("PrepareFiles() failed: %v", err) } @@ -139,7 +139,7 @@ func TestNginx(t *testing.T) { } d := dockerutil.MakeDocker("net-test") - dir, err := dockerutil.PrepareFiles("latin10k.txt") + dir, err := dockerutil.PrepareFiles("test/image/latin10k.txt") if err != nil { t.Fatalf("PrepareFiles() failed: %v", err) } @@ -183,7 +183,7 @@ func TestMysql(t *testing.T) { } client := dockerutil.MakeDocker("mysql-client-test") - dir, err := dockerutil.PrepareFiles("mysql.sql") + dir, err := dockerutil.PrepareFiles("test/image/mysql.sql") if err != nil { t.Fatalf("PrepareFiles() failed: %v", err) } @@ -283,7 +283,7 @@ func TestRuby(t *testing.T) { } d := dockerutil.MakeDocker("ruby-test") - dir, err := dockerutil.PrepareFiles("ruby.rb", "ruby.sh") + dir, err := dockerutil.PrepareFiles("test/image/ruby.rb", "test/image/ruby.sh") if err != nil { t.Fatalf("PrepareFiles() failed: %v", err) } diff --git a/test/syscalls/build_defs.bzl b/test/syscalls/build_defs.bzl index 1df761dd0..cbab85ef7 100644 --- a/test/syscalls/build_defs.bzl +++ b/test/syscalls/build_defs.bzl @@ -2,8 +2,6 @@ load("//tools:defs.bzl", "loopback") -# syscall_test is a macro that will create targets to run the given test target -# on the host (native) and runsc. def syscall_test( test, shard_count = 5, @@ -13,6 +11,19 @@ def syscall_test( add_uds_tree = False, add_hostinet = False, tags = None): + """syscall_test is a macro that will create targets for all platforms. + + Args: + test: the test target. + shard_count: shards for defined tests. + size: the defined test size. + use_tmpfs: use tmpfs in the defined tests. + add_overlay: add an overlay test. + add_uds_tree: add a UDS test. + add_hostinet: add a hostinet test. + tags: starting test tags. + """ + _syscall_test( test = test, shard_count = shard_count, @@ -111,6 +122,19 @@ def _syscall_test( # all the tests on a specific flavor. Use --test_tag_filters=ptrace,file_shared. tags += [full_platform, "file_" + file_access] + # Hash this target into one of 15 buckets. This can be used to + # randomly split targets between different workflows. + hash15 = hash(native.package_name() + name) % 15 + tags.append("hash15:" + str(hash15)) + + # TODO(b/139838000): Tests using hostinet must be disabled on Guitar until + # we figure out how to request ipv4 sockets on Guitar machines. + if network == "host": + tags.append("noguitar") + + # Disable off-host networking. + tags.append("requires-net:loopback") + # Add tag to prevent the tests from running in a Bazel sandbox. # TODO(b/120560048): Make the tests run without this tag. tags.append("no-sandbox") @@ -118,8 +142,11 @@ def _syscall_test( # TODO(b/112165693): KVM tests are tagged "manual" to until the platform is # more stable. if platform == "kvm": - tags += ["manual"] - tags += ["requires-kvm"] + tags.append("manual") + tags.append("requires-kvm") + + # TODO(b/112165693): Remove when tests pass reliably. + tags.append("notap") args = [ # Arguments are passed directly to syscall_test_runner binary. diff --git a/test/syscalls/linux/chroot.cc b/test/syscalls/linux/chroot.cc index 0a2d44a2c..85ec013d5 100644 --- a/test/syscalls/linux/chroot.cc +++ b/test/syscalls/linux/chroot.cc @@ -167,7 +167,7 @@ TEST(ChrootTest, DotDotFromOpenFD) { } // Test that link resolution in a chroot can escape the root by following an -// open proc fd. +// open proc fd. Regression test for b/32316719. TEST(ChrootTest, ProcFdLinkResolutionInChroot) { SKIP_IF(!ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_SYS_CHROOT))); diff --git a/test/syscalls/linux/concurrency.cc b/test/syscalls/linux/concurrency.cc index f41f99900..7cd6a75bd 100644 --- a/test/syscalls/linux/concurrency.cc +++ b/test/syscalls/linux/concurrency.cc @@ -46,7 +46,8 @@ TEST(ConcurrencyTest, SingleProcessMultithreaded) { } // Test that multiple threads in this process continue to execute in parallel, -// even if an unrelated second process is spawned. +// even if an unrelated second process is spawned. Regression test for +// b/32119508. TEST(ConcurrencyTest, MultiProcessMultithreaded) { // In PID 1, start TIDs 1 and 2, and put both to sleep. // diff --git a/test/syscalls/linux/exec_proc_exe_workload.cc b/test/syscalls/linux/exec_proc_exe_workload.cc index b790fe5be..2989379b7 100644 --- a/test/syscalls/linux/exec_proc_exe_workload.cc +++ b/test/syscalls/linux/exec_proc_exe_workload.cc @@ -21,6 +21,12 @@ #include "test/util/posix_error.h" int main(int argc, char** argv, char** envp) { + // This is annoying. Because remote build systems may put these binaries + // in a content-addressable-store, you may wind up with /proc/self/exe + // pointing to some random path (but with a sensible argv[0]). + // + // Therefore, this test simply checks that the /proc/self/exe + // is absolute and *doesn't* match argv[1]. std::string exe = gvisor::testing::ProcessExePath(getpid()).ValueOrDie(); if (exe[0] != '/') { diff --git a/test/syscalls/linux/fork.cc b/test/syscalls/linux/fork.cc index 906f3358d..ff8bdfeb0 100644 --- a/test/syscalls/linux/fork.cc +++ b/test/syscalls/linux/fork.cc @@ -271,7 +271,7 @@ TEST_F(ForkTest, Alarm) { EXPECT_EQ(0, alarmed); } -// Child cannot affect parent private memory. +// Child cannot affect parent private memory. Regression test for b/24137240. TEST_F(ForkTest, PrivateMemory) { std::atomic local(0); @@ -298,6 +298,9 @@ TEST_F(ForkTest, PrivateMemory) { } // Kernel-accessed buffers should remain coherent across COW. +// +// The buffer must be >= usermem.ZeroCopyMinBytes, as UnsafeAccess operates +// differently. Regression test for b/33811887. TEST_F(ForkTest, COWSegment) { constexpr int kBufSize = 1024; char* read_buf = private_; diff --git a/test/syscalls/linux/mmap.cc b/test/syscalls/linux/mmap.cc index 1c4d9f1c7..11fb1b457 100644 --- a/test/syscalls/linux/mmap.cc +++ b/test/syscalls/linux/mmap.cc @@ -1418,7 +1418,7 @@ TEST_P(MMapFileParamTest, NoSigBusOnPageContainingEOF) { // // On most platforms this is trivial, but when the file is mapped via the sentry // page cache (which does not yet support writing to shared mappings), a bug -// caused reads to fail unnecessarily on such mappings. +// caused reads to fail unnecessarily on such mappings. See b/28913513. TEST_F(MMapFileTest, ReadingWritableSharedFilePageSucceeds) { uintptr_t addr; size_t len = strlen(kFileContents); @@ -1435,7 +1435,7 @@ TEST_F(MMapFileTest, ReadingWritableSharedFilePageSucceeds) { // Tests that EFAULT is returned when invoking a syscall that requires the OS to // read past end of file (resulting in a fault in sentry context in the gVisor -// case). +// case). See b/28913513. TEST_F(MMapFileTest, InternalSigBus) { uintptr_t addr; ASSERT_THAT(addr = Map(0, 2 * kPageSize, PROT_READ | PROT_WRITE, MAP_PRIVATE, @@ -1578,7 +1578,7 @@ TEST_F(MMapFileTest, Bug38498194) { } // Tests that reading from a file to a memory mapping of the same file does not -// deadlock. +// deadlock. See b/34813270. TEST_F(MMapFileTest, SelfRead) { uintptr_t addr; ASSERT_THAT(addr = Map(0, kPageSize, PROT_READ | PROT_WRITE, MAP_SHARED, @@ -1590,7 +1590,7 @@ TEST_F(MMapFileTest, SelfRead) { } // Tests that writing to a file from a memory mapping of the same file does not -// deadlock. +// deadlock. Regression test for b/34813270. TEST_F(MMapFileTest, SelfWrite) { uintptr_t addr; ASSERT_THAT(addr = Map(0, kPageSize, PROT_READ, MAP_SHARED, fd_.get(), 0), diff --git a/test/syscalls/linux/open_create.cc b/test/syscalls/linux/open_create.cc index 431733dbe..902d0a0dc 100644 --- a/test/syscalls/linux/open_create.cc +++ b/test/syscalls/linux/open_create.cc @@ -132,6 +132,7 @@ TEST(CreateTest, CreateFailsOnDirWithoutWritePerms) { } // A file originally created RW, but opened RO can later be opened RW. +// Regression test for b/65385065. TEST(CreateTest, OpenCreateROThenRW) { TempPath file(NewTempAbsPath()); diff --git a/test/syscalls/linux/preadv.cc b/test/syscalls/linux/preadv.cc index f7ea44054..5b0743fe9 100644 --- a/test/syscalls/linux/preadv.cc +++ b/test/syscalls/linux/preadv.cc @@ -37,6 +37,7 @@ namespace testing { namespace { +// Stress copy-on-write. Attempts to reproduce b/38430174. TEST(PreadvTest, MMConcurrencyStress) { // Fill a one-page file with zeroes (the contents don't really matter). const auto f = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateFileWith( diff --git a/test/syscalls/linux/proc.cc b/test/syscalls/linux/proc.cc index 169b723eb..a23fdb58d 100644 --- a/test/syscalls/linux/proc.cc +++ b/test/syscalls/linux/proc.cc @@ -1352,13 +1352,19 @@ TEST(ProcPidSymlink, SubprocessZombied) { // FIXME(gvisor.dev/issue/164): Inconsistent behavior between gVisor and linux // on proc files. - // 4.17 & gVisor: Syscall succeeds and returns 1 + // + // ~4.3: Syscall fails with EACCES. + // 4.17 & gVisor: Syscall succeeds and returns 1. + // // EXPECT_THAT(ReadlinkWhileZombied("ns/pid", buf, sizeof(buf)), // SyscallFailsWithErrno(EACCES)); // FIXME(gvisor.dev/issue/164): Inconsistent behavior between gVisor and linux // on proc files. - // 4.17 & gVisor: Syscall succeeds and returns 1. + // + // ~4.3: Syscall fails with EACCES. + // 4.17 & gVisor: Syscall succeeds and returns 1. + // // EXPECT_THAT(ReadlinkWhileZombied("ns/user", buf, sizeof(buf)), // SyscallFailsWithErrno(EACCES)); } @@ -1431,8 +1437,12 @@ TEST(ProcPidFile, SubprocessRunning) { TEST(ProcPidFile, SubprocessZombie) { char buf[1]; - // 4.17: Succeeds and returns 1 - // gVisor: Succeeds and returns 0 + // FIXME(gvisor.dev/issue/164): Loosen requirement due to inconsistent + // behavior on different kernels. + // + // ~4.3: Succeds and returns 0. + // 4.17: Succeeds and returns 1. + // gVisor: Succeeds and returns 0. EXPECT_THAT(ReadWhileZombied("auxv", buf, sizeof(buf)), SyscallSucceeds()); EXPECT_THAT(ReadWhileZombied("cmdline", buf, sizeof(buf)), @@ -1458,7 +1468,10 @@ TEST(ProcPidFile, SubprocessZombie) { // FIXME(gvisor.dev/issue/164): Inconsistent behavior between gVisor and linux // on proc files. + // + // ~4.3: Fails and returns EACCES. // gVisor & 4.17: Succeeds and returns 1. + // // EXPECT_THAT(ReadWhileZombied("io", buf, sizeof(buf)), // SyscallFailsWithErrno(EACCES)); } @@ -1467,9 +1480,12 @@ TEST(ProcPidFile, SubprocessZombie) { TEST(ProcPidFile, SubprocessExited) { char buf[1]; - // FIXME(gvisor.dev/issue/164): Inconsistent behavior between kernels + // FIXME(gvisor.dev/issue/164): Inconsistent behavior between kernels. + // + // ~4.3: Fails and returns ESRCH. // gVisor: Fails with ESRCH. // 4.17: Succeeds and returns 1. + // // EXPECT_THAT(ReadWhileExited("auxv", buf, sizeof(buf)), // SyscallFailsWithErrno(ESRCH)); @@ -1641,7 +1657,7 @@ TEST(ProcTask, KilledThreadsDisappear) { EXPECT_NO_ERRNO(DirContainsExactly("/proc/self/task", TaskFiles(initial, {child1.Tid()}))); - // Stat child1's task file. + // Stat child1's task file. Regression test for b/32097707. struct stat statbuf; const std::string child1_task_file = absl::StrCat("/proc/self/task/", child1.Tid()); @@ -1669,7 +1685,7 @@ TEST(ProcTask, KilledThreadsDisappear) { EXPECT_NO_ERRNO(EventuallyDirContainsExactly( "/proc/self/task", TaskFiles(initial, {child3.Tid(), child5.Tid()}))); - // Stat child1's task file again. This time it should fail. + // Stat child1's task file again. This time it should fail. See b/32097707. EXPECT_THAT(stat(child1_task_file.c_str(), &statbuf), SyscallFailsWithErrno(ENOENT)); @@ -1824,7 +1840,7 @@ TEST(ProcSysVmOvercommitMemory, HasNumericValue) { } // Check that link for proc fd entries point the target node, not the -// symlink itself. +// symlink itself. Regression test for b/31155070. TEST(ProcTaskFd, FstatatFollowsSymlink) { const TempPath file = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateFile()); const FileDescriptor fd = @@ -1883,6 +1899,20 @@ TEST(ProcMounts, IsSymlink) { EXPECT_EQ(link, "self/mounts"); } +TEST(ProcSelfMountinfo, RequiredFieldsArePresent) { + auto mountinfo = + ASSERT_NO_ERRNO_AND_VALUE(GetContents("/proc/self/mountinfo")); + EXPECT_THAT( + mountinfo, + AllOf( + // Root mount. + ContainsRegex( + R"([0-9]+ [0-9]+ [0-9]+:[0-9]+ / / (rw|ro).*- \S+ \S+ (rw|ro)\S*)"), + // Proc mount - always rw. + ContainsRegex( + R"([0-9]+ [0-9]+ [0-9]+:[0-9]+ / /proc rw.*- \S+ \S+ rw\S*)"))); +} + // Check that /proc/self/mounts looks something like a real mounts file. TEST(ProcSelfMounts, RequiredFieldsArePresent) { auto mounts = ASSERT_NO_ERRNO_AND_VALUE(GetContents("/proc/self/mounts")); diff --git a/test/syscalls/linux/readv.cc b/test/syscalls/linux/readv.cc index 4069cbc7e..baaf9f757 100644 --- a/test/syscalls/linux/readv.cc +++ b/test/syscalls/linux/readv.cc @@ -254,7 +254,9 @@ TEST_F(ReadvTest, IovecOutsideTaskAddressRangeInNonemptyArray) { // This test depends on the maximum extent of a single readv() syscall, so // we can't tolerate interruption from saving. TEST(ReadvTestNoFixture, TruncatedAtMax_NoRandomSave) { - // Ensure that we won't be interrupted by ITIMER_PROF. + // Ensure that we won't be interrupted by ITIMER_PROF. This is particularly + // important in environments where automated profiling tools may start + // ITIMER_PROF automatically. struct itimerval itv = {}; auto const cleanup_itimer = ASSERT_NO_ERRNO_AND_VALUE(ScopedItimer(ITIMER_PROF, itv)); diff --git a/test/syscalls/linux/rseq.cc b/test/syscalls/linux/rseq.cc index 106c045e3..4bfb1ff56 100644 --- a/test/syscalls/linux/rseq.cc +++ b/test/syscalls/linux/rseq.cc @@ -36,7 +36,7 @@ namespace { // We must be very careful about how these tests are written. Each thread may // only have one struct rseq registration, which may be done automatically at // thread start (as of 2019-11-13, glibc does *not* support rseq and thus does -// not do so). +// not do so, but other libraries do). // // Testing of rseq is thus done primarily in a child process with no // registration. This means exec'ing a nostdlib binary, as rseq registration can diff --git a/test/syscalls/linux/select.cc b/test/syscalls/linux/select.cc index 424e2a67f..be2364fb8 100644 --- a/test/syscalls/linux/select.cc +++ b/test/syscalls/linux/select.cc @@ -146,7 +146,7 @@ TEST_F(SelectTest, IgnoreBitsAboveNfds) { // This test illustrates Linux's behavior of 'select' calls passing after // setrlimit RLIMIT_NOFILE is called. In particular, versions of sshd rely on -// this behavior. +// this behavior. See b/122318458. TEST_F(SelectTest, SetrlimitCallNOFILE) { fd_set read_set; FD_ZERO(&read_set); diff --git a/test/syscalls/linux/shm.cc b/test/syscalls/linux/shm.cc index 7ba752599..c7fdbb924 100644 --- a/test/syscalls/linux/shm.cc +++ b/test/syscalls/linux/shm.cc @@ -473,7 +473,7 @@ TEST(ShmTest, PartialUnmap) { } // Check that sentry does not panic when asked for a zero-length private shm -// segment. +// segment. Regression test for b/110694797. TEST(ShmTest, GracefullyFailOnZeroLenSegmentCreation) { EXPECT_THAT(Shmget(IPC_PRIVATE, 0, 0), PosixErrorIs(EINVAL, _)); } diff --git a/test/syscalls/linux/sigprocmask.cc b/test/syscalls/linux/sigprocmask.cc index 654c6a47f..a603fc1d1 100644 --- a/test/syscalls/linux/sigprocmask.cc +++ b/test/syscalls/linux/sigprocmask.cc @@ -237,7 +237,7 @@ TEST_F(SigProcMaskTest, SignalHandler) { } // Check that sigprocmask correctly handles aliasing of the set and oldset -// pointers. +// pointers. Regression test for b/30502311. TEST_F(SigProcMaskTest, AliasedSets) { sigset_t mask; diff --git a/test/syscalls/linux/socket_unix_non_stream.cc b/test/syscalls/linux/socket_unix_non_stream.cc index 276a94eb8..884319e1d 100644 --- a/test/syscalls/linux/socket_unix_non_stream.cc +++ b/test/syscalls/linux/socket_unix_non_stream.cc @@ -109,7 +109,7 @@ PosixErrorOr> CreateFragmentedRegion(const int size, } // A contiguous iov that is heavily fragmented in FileMem can still be sent -// successfully. +// successfully. See b/115833655. TEST_P(UnixNonStreamSocketPairTest, FragmentedSendMsg) { auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); @@ -165,7 +165,7 @@ TEST_P(UnixNonStreamSocketPairTest, FragmentedSendMsg) { } // A contiguous iov that is heavily fragmented in FileMem can still be received -// into successfully. +// into successfully. Regression test for b/115833655. TEST_P(UnixNonStreamSocketPairTest, FragmentedRecvMsg) { auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); diff --git a/test/syscalls/linux/symlink.cc b/test/syscalls/linux/symlink.cc index b249ff91f..03ee1250d 100644 --- a/test/syscalls/linux/symlink.cc +++ b/test/syscalls/linux/symlink.cc @@ -38,7 +38,7 @@ mode_t FilePermission(const std::string& path) { } // Test that name collisions are checked on the new link path, not the source -// path. +// path. Regression test for b/31782115. TEST(SymlinkTest, CanCreateSymlinkWithCachedSourceDirent) { const std::string srcname = NewTempAbsPath(); const std::string newname = NewTempAbsPath(); diff --git a/test/syscalls/linux/tcp_socket.cc b/test/syscalls/linux/tcp_socket.cc index 8a8b68e75..c4591a3b9 100644 --- a/test/syscalls/linux/tcp_socket.cc +++ b/test/syscalls/linux/tcp_socket.cc @@ -244,7 +244,8 @@ TEST_P(TcpSocketTest, ZeroWriteAllowed) { } // Test that a non-blocking write with a buffer that is larger than the send -// buffer size will not actually write the whole thing at once. +// buffer size will not actually write the whole thing at once. Regression test +// for b/64438887. TEST_P(TcpSocketTest, NonblockingLargeWrite) { // Set the FD to O_NONBLOCK. int opts; diff --git a/test/syscalls/linux/time.cc b/test/syscalls/linux/time.cc index c7eead17e..1ccb95733 100644 --- a/test/syscalls/linux/time.cc +++ b/test/syscalls/linux/time.cc @@ -62,6 +62,7 @@ TEST(TimeTest, VsyscallTime_InvalidAddressSIGSEGV) { ::testing::KilledBySignal(SIGSEGV), ""); } +// Mimics the gettimeofday(2) wrapper from the Go runtime <= 1.2. int vsyscall_gettimeofday(struct timeval* tv, struct timezone* tz) { constexpr uint64_t kVsyscallGettimeofdayEntry = 0xffffffffff600000; return reinterpret_cast( diff --git a/test/syscalls/linux/tkill.cc b/test/syscalls/linux/tkill.cc index bae377c69..8d8ebbb24 100644 --- a/test/syscalls/linux/tkill.cc +++ b/test/syscalls/linux/tkill.cc @@ -54,7 +54,7 @@ void SigHandler(int sig, siginfo_t* info, void* context) { TEST_CHECK(info->si_code == SI_TKILL); } -// Test with a real signal. +// Test with a real signal. Regression test for b/24790092. TEST(TkillTest, ValidTIDAndRealSignal) { struct sigaction sa; sa.sa_sigaction = SigHandler; diff --git a/test/util/temp_path.cc b/test/util/temp_path.cc index 35aacb172..9c10b6674 100644 --- a/test/util/temp_path.cc +++ b/test/util/temp_path.cc @@ -77,6 +77,7 @@ std::string NewTempAbsPath() { std::string NewTempRelPath() { return NextTempBasename(); } std::string GetAbsoluteTestTmpdir() { + // Note that TEST_TMPDIR is guaranteed to be set. char* env_tmpdir = getenv("TEST_TMPDIR"); std::string tmp_dir = env_tmpdir != nullptr ? std::string(env_tmpdir) : "/tmp"; diff --git a/tools/build/tags.bzl b/tools/build/tags.bzl index e99c87f81..a6db44e47 100644 --- a/tools/build/tags.bzl +++ b/tools/build/tags.bzl @@ -33,4 +33,8 @@ go_suffixes = [ "_wasm_unsafe", "_linux", "_linux_unsafe", + "_opts", + "_opts_unsafe", + "_impl", + "_impl_unsafe", ] diff --git a/tools/defs.bzl b/tools/defs.bzl index 5d5fa134a..c03b557ae 100644 --- a/tools/defs.bzl +++ b/tools/defs.bzl @@ -73,6 +73,16 @@ def calculate_sets(srcs): result[target].append(file) return result +def go_imports(name, src, out): + """Simplify a single Go source file by eliminating unused imports.""" + native.genrule( + name = name, + srcs = [src], + outs = [out], + tools = ["@org_golang_x_tools//cmd/goimports:goimports"], + cmd = ("$(location @org_golang_x_tools//cmd/goimports:goimports) $(SRCS) > $@"), + ) + def go_library(name, srcs, deps = [], imports = [], stateify = True, marshal = False, **kwargs): """Wraps the standard go_library and does stateification and marshalling. @@ -107,10 +117,15 @@ def go_library(name, srcs, deps = [], imports = [], stateify = True, marshal = F state_sets = calculate_sets(srcs) for (suffix, srcs) in state_sets.items(): go_stateify( - name = name + suffix + "_state_autogen", + name = name + suffix + "_state_autogen_with_imports", srcs = srcs, imports = imports, package = name, + out = name + suffix + "_state_autogen_with_imports.go", + ) + go_imports( + name = name + suffix + "_state_autogen", + src = name + suffix + "_state_autogen_with_imports.go", out = name + suffix + "_state_autogen.go", ) all_srcs = all_srcs + [ -- cgit v1.2.3 From c141eb5f430dc50f6bf90232c369b7b3a542155e Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Fri, 7 Feb 2020 13:47:57 -0800 Subject: Address GH comments. --- pkg/abi/linux/netfilter.go | 2 +- pkg/sentry/socket/netfilter/extensions.go | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/abi/linux/netfilter.go b/pkg/abi/linux/netfilter.go index e4aabb6bb..7363185b7 100644 --- a/pkg/abi/linux/netfilter.go +++ b/pkg/abi/linux/netfilter.go @@ -363,7 +363,7 @@ type XTTCP struct { // range to which the matcher applies. DestinationPortStart uint16 - // DestinationPortEnd specifies the start of the destination port + // DestinationPortEnd specifies the end of the destination port // range to which the matcher applies. DestinationPortEnd uint16 diff --git a/pkg/sentry/socket/netfilter/extensions.go b/pkg/sentry/socket/netfilter/extensions.go index b5fbb52e4..3082976cd 100644 --- a/pkg/sentry/socket/netfilter/extensions.go +++ b/pkg/sentry/socket/netfilter/extensions.go @@ -1,4 +1,4 @@ -// Copyright 2019 The gVisor Authors. +// Copyright 2020 The gVisor Authors. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. -- cgit v1.2.3 From 762e4761cc4edd92108f6836ad1933c7158b8be8 Mon Sep 17 00:00:00 2001 From: gVisor bot Date: Tue, 11 Feb 2020 11:08:28 -0800 Subject: Move Align{Up,Down} into binary package. PiperOrigin-RevId: 294477647 --- pkg/binary/binary.go | 10 ++++++++++ pkg/sentry/socket/control/control.go | 26 ++++++++------------------ pkg/sentry/socket/netfilter/extensions.go | 7 +------ pkg/sentry/socket/netlink/message.go | 15 ++++----------- pkg/sentry/strace/BUILD | 1 - pkg/sentry/strace/socket.go | 7 +++---- 6 files changed, 26 insertions(+), 40 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/binary/binary.go b/pkg/binary/binary.go index 631785f7b..25065aef9 100644 --- a/pkg/binary/binary.go +++ b/pkg/binary/binary.go @@ -254,3 +254,13 @@ func WriteUint64(w io.Writer, order binary.ByteOrder, num uint64) error { _, err := w.Write(buf) return err } + +// AlignUp rounds a length up to an alignment. align must be a power of 2. +func AlignUp(length int, align uint) int { + return (length + int(align) - 1) & ^(int(align) - 1) +} + +// AlignDown rounds a length down to an alignment. align must be a power of 2. +func AlignDown(length int, align uint) int { + return length & ^(int(align) - 1) +} diff --git a/pkg/sentry/socket/control/control.go b/pkg/sentry/socket/control/control.go index 00265f15b..6145a7fc3 100644 --- a/pkg/sentry/socket/control/control.go +++ b/pkg/sentry/socket/control/control.go @@ -189,7 +189,7 @@ func putUint32(buf []byte, n uint32) []byte { // putCmsg writes a control message header and as much data as will fit into // the unused capacity of a buffer. func putCmsg(buf []byte, flags int, msgType uint32, align uint, data []int32) ([]byte, int) { - space := AlignDown(cap(buf)-len(buf), 4) + space := binary.AlignDown(cap(buf)-len(buf), 4) // We can't write to space that doesn't exist, so if we are going to align // the available space, we must align down. @@ -282,19 +282,9 @@ func PackCredentials(t *kernel.Task, creds SCMCredentials, buf []byte, flags int return putCmsg(buf, flags, linux.SCM_CREDENTIALS, align, c) } -// AlignUp rounds a length up to an alignment. align must be a power of 2. -func AlignUp(length int, align uint) int { - return (length + int(align) - 1) & ^(int(align) - 1) -} - -// AlignDown rounds a down to an alignment. align must be a power of 2. -func AlignDown(length int, align uint) int { - return length & ^(int(align) - 1) -} - // alignSlice extends a slice's length (up to the capacity) to align it. func alignSlice(buf []byte, align uint) []byte { - aligned := AlignUp(len(buf), align) + aligned := binary.AlignUp(len(buf), align) if aligned > cap(buf) { // Linux allows unaligned data if there isn't room for alignment. // Since there isn't room for alignment, there isn't room for any @@ -377,7 +367,7 @@ func PackControlMessages(t *kernel.Task, cmsgs socket.ControlMessages, buf []byt // cmsgSpace is equivalent to CMSG_SPACE in Linux. func cmsgSpace(t *kernel.Task, dataLen int) int { - return linux.SizeOfControlMessageHeader + AlignUp(dataLen, t.Arch().Width()) + return linux.SizeOfControlMessageHeader + binary.AlignUp(dataLen, t.Arch().Width()) } // CmsgsSpace returns the number of bytes needed to fit the control messages @@ -437,7 +427,7 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con case linux.SOL_SOCKET: switch h.Type { case linux.SCM_RIGHTS: - rightsSize := AlignDown(length, linux.SizeOfControlMessageRight) + rightsSize := binary.AlignDown(length, linux.SizeOfControlMessageRight) numRights := rightsSize / linux.SizeOfControlMessageRight if len(fds)+numRights > linux.SCM_MAX_FD { @@ -448,7 +438,7 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con fds = append(fds, int32(usermem.ByteOrder.Uint32(buf[j:j+linux.SizeOfControlMessageRight]))) } - i += AlignUp(length, width) + i += binary.AlignUp(length, width) case linux.SCM_CREDENTIALS: if length < linux.SizeOfControlMessageCredentials { @@ -462,7 +452,7 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con return socket.ControlMessages{}, err } cmsgs.Unix.Credentials = scmCreds - i += AlignUp(length, width) + i += binary.AlignUp(length, width) default: // Unknown message type. @@ -476,7 +466,7 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con } cmsgs.IP.HasTOS = true binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageTOS], usermem.ByteOrder, &cmsgs.IP.TOS) - i += AlignUp(length, width) + i += binary.AlignUp(length, width) default: return socket.ControlMessages{}, syserror.EINVAL @@ -489,7 +479,7 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con } cmsgs.IP.HasTClass = true binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageTClass], usermem.ByteOrder, &cmsgs.IP.TClass) - i += AlignUp(length, width) + i += binary.AlignUp(length, width) default: return socket.ControlMessages{}, syserror.EINVAL diff --git a/pkg/sentry/socket/netfilter/extensions.go b/pkg/sentry/socket/netfilter/extensions.go index 22fd0ebe7..b4b244abf 100644 --- a/pkg/sentry/socket/netfilter/extensions.go +++ b/pkg/sentry/socket/netfilter/extensions.go @@ -72,7 +72,7 @@ func marshalEntryMatch(name string, data []byte) []byte { nflog("marshaling matcher %q", name) // We have to pad this struct size to a multiple of 8 bytes. - size := alignUp(linux.SizeOfXTEntryMatch+len(data), 8) + size := binary.AlignUp(linux.SizeOfXTEntryMatch+len(data), 8) matcher := linux.KernelXTEntryMatch{ XTEntryMatch: linux.XTEntryMatch{ MatchSize: uint16(size), @@ -93,8 +93,3 @@ func unmarshalMatcher(match linux.XTEntryMatch, filter iptables.IPHeaderFilter, } return matchMaker.unmarshal(buf, filter) } - -// alignUp rounds a length up to an alignment. align must be a power of 2. -func alignUp(length int, align uint) int { - return (length + int(align) - 1) & ^(int(align) - 1) -} diff --git a/pkg/sentry/socket/netlink/message.go b/pkg/sentry/socket/netlink/message.go index 4ea252ccb..0899c61d1 100644 --- a/pkg/sentry/socket/netlink/message.go +++ b/pkg/sentry/socket/netlink/message.go @@ -23,18 +23,11 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -// alignUp rounds a length up to an alignment. -// -// Preconditions: align is a power of two. -func alignUp(length int, align uint) int { - return (length + int(align) - 1) &^ (int(align) - 1) -} - // alignPad returns the length of padding required for alignment. // // Preconditions: align is a power of two. func alignPad(length int, align uint) int { - return alignUp(length, align) - length + return binary.AlignUp(length, align) - length } // Message contains a complete serialized netlink message. @@ -138,7 +131,7 @@ func (m *Message) Finalize() []byte { // Align the message. Note that the message length in the header (set // above) is the useful length of the message, not the total aligned // length. See net/netlink/af_netlink.c:__nlmsg_put. - aligned := alignUp(len(m.buf), linux.NLMSG_ALIGNTO) + aligned := binary.AlignUp(len(m.buf), linux.NLMSG_ALIGNTO) m.putZeros(aligned - len(m.buf)) return m.buf } @@ -173,7 +166,7 @@ func (m *Message) PutAttr(atype uint16, v interface{}) { m.Put(v) // Align the attribute. - aligned := alignUp(l, linux.NLA_ALIGNTO) + aligned := binary.AlignUp(l, linux.NLA_ALIGNTO) m.putZeros(aligned - l) } @@ -190,7 +183,7 @@ func (m *Message) PutAttrString(atype uint16, s string) { m.putZeros(1) // Align the attribute. - aligned := alignUp(l, linux.NLA_ALIGNTO) + aligned := binary.AlignUp(l, linux.NLA_ALIGNTO) m.putZeros(aligned - l) } diff --git a/pkg/sentry/strace/BUILD b/pkg/sentry/strace/BUILD index 762a946fe..2f39a6f2b 100644 --- a/pkg/sentry/strace/BUILD +++ b/pkg/sentry/strace/BUILD @@ -30,7 +30,6 @@ go_library( "//pkg/seccomp", "//pkg/sentry/arch", "//pkg/sentry/kernel", - "//pkg/sentry/socket/control", "//pkg/sentry/socket/netlink", "//pkg/sentry/socket/netstack", "//pkg/sentry/syscalls/linux", diff --git a/pkg/sentry/strace/socket.go b/pkg/sentry/strace/socket.go index f7ff4573e..51e6d81b2 100644 --- a/pkg/sentry/strace/socket.go +++ b/pkg/sentry/strace/socket.go @@ -22,7 +22,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/socket/control" "gvisor.dev/gvisor/pkg/sentry/socket/netlink" "gvisor.dev/gvisor/pkg/sentry/socket/netstack" slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" @@ -220,13 +219,13 @@ func cmsghdr(t *kernel.Task, addr usermem.Addr, length uint64, maxBytes uint64) if skipData { strs = append(strs, fmt.Sprintf("{level=%s, type=%s, length=%d}", level, typ, h.Length)) - i += control.AlignUp(length, width) + i += binary.AlignUp(length, width) continue } switch h.Type { case linux.SCM_RIGHTS: - rightsSize := control.AlignDown(length, linux.SizeOfControlMessageRight) + rightsSize := binary.AlignDown(length, linux.SizeOfControlMessageRight) numRights := rightsSize / linux.SizeOfControlMessageRight fds := make(linux.ControlMessageRights, numRights) @@ -295,7 +294,7 @@ func cmsghdr(t *kernel.Task, addr usermem.Addr, length uint64, maxBytes uint64) default: panic("unreachable") } - i += control.AlignUp(length, width) + i += binary.AlignUp(length, width) } return fmt.Sprintf("%#x %s", addr, strings.Join(strs, ", ")) -- cgit v1.2.3 From 6fdf2c53a1d084b70602170b660242036fd8fe4f Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Fri, 7 Feb 2020 11:21:07 -0800 Subject: iptables: User chains - Adds creation of user chains via `-N ` - Adds `-j RETURN` support for built-in chains, which triggers the chain's underflow rule (usually the default policy). - Adds tests for chain creation, default policies, and `-j RETURN' from built-in chains. --- pkg/sentry/socket/netfilter/netfilter.go | 115 +++++++++++++++++++----------- pkg/tcpip/iptables/iptables.go | 74 ++++++++++++------- pkg/tcpip/iptables/targets.go | 41 ++++++++--- pkg/tcpip/iptables/types.go | 50 +++++-------- test/iptables/filter_input.go | 117 ++++++++++++++++++++++++++++++- test/iptables/iptables_test.go | 24 +++++++ 6 files changed, 310 insertions(+), 111 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index ea02627de..3fc80e0de 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -50,7 +50,9 @@ type metadata struct { // nflog logs messages related to the writing and reading of iptables. func nflog(format string, args ...interface{}) { - log.Infof("netfilter: "+format, args...) + if log.IsLogging(log.Debug) { + log.Debugf("netfilter: "+format, args...) + } } // GetInfo returns information about iptables. @@ -227,19 +229,23 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern } func marshalTarget(target iptables.Target) []byte { - switch target.(type) { - case iptables.UnconditionalAcceptTarget: - return marshalStandardTarget(iptables.Accept) - case iptables.UnconditionalDropTarget: - return marshalStandardTarget(iptables.Drop) + switch tg := target.(type) { + case iptables.AcceptTarget: + return marshalStandardTarget(iptables.RuleAccept) + case iptables.DropTarget: + return marshalStandardTarget(iptables.RuleDrop) case iptables.ErrorTarget: - return marshalErrorTarget() + return marshalErrorTarget(errorTargetName) + case iptables.UserChainTarget: + return marshalErrorTarget(tg.Name) + case iptables.ReturnTarget: + return marshalStandardTarget(iptables.RuleReturn) default: panic(fmt.Errorf("unknown target of type %T", target)) } } -func marshalStandardTarget(verdict iptables.Verdict) []byte { +func marshalStandardTarget(verdict iptables.RuleVerdict) []byte { nflog("convert to binary: marshalling standard target with size %d", linux.SizeOfXTStandardTarget) // The target's name will be the empty string. @@ -254,14 +260,14 @@ func marshalStandardTarget(verdict iptables.Verdict) []byte { return binary.Marshal(ret, usermem.ByteOrder, target) } -func marshalErrorTarget() []byte { +func marshalErrorTarget(errorName string) []byte { // This is an error target named error target := linux.XTErrorTarget{ Target: linux.XTEntryTarget{ TargetSize: linux.SizeOfXTErrorTarget, }, } - copy(target.Name[:], errorTargetName) + copy(target.Name[:], errorName) copy(target.Target.Name[:], errorTargetName) ret := make([]byte, 0, linux.SizeOfXTErrorTarget) @@ -270,38 +276,35 @@ func marshalErrorTarget() []byte { // translateFromStandardVerdict translates verdicts the same way as the iptables // tool. -func translateFromStandardVerdict(verdict iptables.Verdict) int32 { +func translateFromStandardVerdict(verdict iptables.RuleVerdict) int32 { switch verdict { - case iptables.Accept: + case iptables.RuleAccept: return -linux.NF_ACCEPT - 1 - case iptables.Drop: + case iptables.RuleDrop: return -linux.NF_DROP - 1 - case iptables.Queue: - return -linux.NF_QUEUE - 1 - case iptables.Return: + case iptables.RuleReturn: return linux.NF_RETURN - case iptables.Jump: + default: // TODO(gvisor.dev/issue/170): Support Jump. - panic("Jump isn't supported yet") + panic(fmt.Sprintf("unknown standard verdict: %d", verdict)) } - panic(fmt.Sprintf("unknown standard verdict: %d", verdict)) } -// translateToStandardVerdict translates from the value in a +// translateToStandardTarget translates from the value in a // linux.XTStandardTarget to an iptables.Verdict. -func translateToStandardVerdict(val int32) (iptables.Verdict, error) { +func translateToStandardTarget(val int32) (iptables.Target, error) { // TODO(gvisor.dev/issue/170): Support other verdicts. switch val { case -linux.NF_ACCEPT - 1: - return iptables.Accept, nil + return iptables.AcceptTarget{}, nil case -linux.NF_DROP - 1: - return iptables.Drop, nil + return iptables.DropTarget{}, nil case -linux.NF_QUEUE - 1: - return iptables.Invalid, errors.New("unsupported iptables verdict QUEUE") + return nil, errors.New("unsupported iptables verdict QUEUE") case linux.NF_RETURN: - return iptables.Invalid, errors.New("unsupported iptables verdict RETURN") + return iptables.ReturnTarget{}, nil default: - return iptables.Invalid, fmt.Errorf("unknown iptables verdict %d", val) + return nil, fmt.Errorf("unknown iptables verdict %d", val) } } @@ -411,6 +414,10 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { table.BuiltinChains[hk] = ruleIdx } if offset == replace.Underflow[hook] { + if !validUnderflow(table.Rules[ruleIdx]) { + nflog("underflow for hook %d isn't an unconditional ACCEPT or DROP.") + return syserr.ErrInvalidArgument + } table.Underflows[hk] = ruleIdx } } @@ -425,12 +432,34 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { } } + // Add the user chains. + for ruleIdx, rule := range table.Rules { + target, ok := rule.Target.(iptables.UserChainTarget) + if !ok { + continue + } + + // We found a user chain. Before inserting it into the table, + // check that: + // - There's some other rule after it. + // - There are no matchers. + if ruleIdx == len(table.Rules)-1 { + nflog("user chain must have a rule or default policy.") + return syserr.ErrInvalidArgument + } + if len(table.Rules[ruleIdx].Matchers) != 0 { + nflog("user chain's first node must have no matcheres.") + return syserr.ErrInvalidArgument + } + table.UserChains[target.Name] = ruleIdx + 1 + } + // TODO(gvisor.dev/issue/170): Support other chains. // Since we only support modifying the INPUT chain right now, make sure // all other chains point to ACCEPT rules. for hook, ruleIdx := range table.BuiltinChains { if hook != iptables.Input { - if _, ok := table.Rules[ruleIdx].Target.(iptables.UnconditionalAcceptTarget); !ok { + if _, ok := table.Rules[ruleIdx].Target.(iptables.AcceptTarget); !ok { nflog("hook %d is unsupported.", hook) return syserr.ErrInvalidArgument } @@ -519,18 +548,7 @@ func parseTarget(optVal []byte) (iptables.Target, error) { buf = optVal[:linux.SizeOfXTStandardTarget] binary.Unmarshal(buf, usermem.ByteOrder, &standardTarget) - verdict, err := translateToStandardVerdict(standardTarget.Verdict) - if err != nil { - return nil, err - } - switch verdict { - case iptables.Accept: - return iptables.UnconditionalAcceptTarget{}, nil - case iptables.Drop: - return iptables.UnconditionalDropTarget{}, nil - default: - return nil, fmt.Errorf("Unknown verdict: %v", verdict) - } + return translateToStandardTarget(standardTarget.Verdict) case errorTargetName: // Error target. @@ -548,11 +566,14 @@ func parseTarget(optVal []byte) (iptables.Target, error) { // somehow fall through every rule. // * To mark the start of a user defined chain. These // rules have an error with the name of the chain. - switch errorTarget.Name.String() { + switch name := errorTarget.Name.String(); name { case errorTargetName: + nflog("set entries: error target") return iptables.ErrorTarget{}, nil default: - return nil, fmt.Errorf("unknown error target %q doesn't exist or isn't supported yet.", errorTarget.Name.String()) + // User defined chain. + nflog("set entries: user-defined target %q", name) + return iptables.UserChainTarget{Name: name}, nil } } @@ -585,6 +606,18 @@ func containsUnsupportedFields(iptip linux.IPTIP) bool { iptip.InverseFlags != 0 } +func validUnderflow(rule iptables.Rule) bool { + if len(rule.Matchers) != 0 { + return false + } + switch rule.Target.(type) { + case iptables.AcceptTarget, iptables.DropTarget: + return true + default: + return false + } +} + func hookFromLinux(hook int) iptables.Hook { switch hook { case linux.NF_INET_PRE_ROUTING: diff --git a/pkg/tcpip/iptables/iptables.go b/pkg/tcpip/iptables/iptables.go index 1b9485bbd..75a433a3b 100644 --- a/pkg/tcpip/iptables/iptables.go +++ b/pkg/tcpip/iptables/iptables.go @@ -52,10 +52,10 @@ func DefaultTables() IPTables { Tables: map[string]Table{ TablenameNat: Table{ Rules: []Rule{ - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: UnconditionalAcceptTarget{}}, + Rule{Target: AcceptTarget{}}, + Rule{Target: AcceptTarget{}}, + Rule{Target: AcceptTarget{}}, + Rule{Target: AcceptTarget{}}, Rule{Target: ErrorTarget{}}, }, BuiltinChains: map[Hook]int{ @@ -74,8 +74,8 @@ func DefaultTables() IPTables { }, TablenameMangle: Table{ Rules: []Rule{ - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: UnconditionalAcceptTarget{}}, + Rule{Target: AcceptTarget{}}, + Rule{Target: AcceptTarget{}}, Rule{Target: ErrorTarget{}}, }, BuiltinChains: map[Hook]int{ @@ -90,9 +90,9 @@ func DefaultTables() IPTables { }, TablenameFilter: Table{ Rules: []Rule{ - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: UnconditionalAcceptTarget{}}, + Rule{Target: AcceptTarget{}}, + Rule{Target: AcceptTarget{}}, + Rule{Target: AcceptTarget{}}, Rule{Target: ErrorTarget{}}, }, BuiltinChains: map[Hook]int{ @@ -149,13 +149,11 @@ func (it *IPTables) Check(hook Hook, pkt tcpip.PacketBuffer) bool { for _, tablename := range it.Priorities[hook] { switch verdict := it.checkTable(hook, pkt, tablename); verdict { // If the table returns Accept, move on to the next table. - case Accept: + case TableAccept: continue // The Drop verdict is final. - case Drop: + case TableDrop: return false - case Stolen, Queue, Repeat, None, Jump, Return, Continue: - panic(fmt.Sprintf("Unimplemented verdict %v.", verdict)) default: panic(fmt.Sprintf("Unknown verdict %v.", verdict)) } @@ -166,36 +164,58 @@ func (it *IPTables) Check(hook Hook, pkt tcpip.PacketBuffer) bool { } // Precondition: pkt.NetworkHeader is set. -func (it *IPTables) checkTable(hook Hook, pkt tcpip.PacketBuffer, tablename string) Verdict { +func (it *IPTables) checkTable(hook Hook, pkt tcpip.PacketBuffer, tablename string) TableVerdict { // Start from ruleIdx and walk the list of rules until a rule gives us // a verdict. table := it.Tables[tablename] for ruleIdx := table.BuiltinChains[hook]; ruleIdx < len(table.Rules); ruleIdx++ { switch verdict := it.checkRule(hook, pkt, table, ruleIdx); verdict { - // In either of these cases, this table is done with the packet. - case Accept, Drop: - return verdict - // Continue traversing the rules of the table. - case Continue: + case RuleAccept: + return TableAccept + + case RuleDrop: + return TableDrop + + case RuleContinue: continue - case Stolen, Queue, Repeat, None, Jump, Return: - panic(fmt.Sprintf("Unimplemented verdict %v.", verdict)) + + case RuleReturn: + // TODO(gvisor.dev/issue/170): We don't implement jump + // yet, so any Return is from a built-in chain. That + // means we have to to call the underflow. + underflow := table.Rules[table.Underflows[hook]] + // Underflow is guaranteed to be an unconditional + // ACCEPT or DROP. + switch v, _ := underflow.Target.Action(pkt); v { + case RuleAccept: + return TableAccept + case RuleDrop: + return TableDrop + case RuleContinue, RuleReturn: + panic("Underflows should only return RuleAccept or RuleDrop.") + default: + panic(fmt.Sprintf("Unknown verdict: %d", v)) + } + default: - panic(fmt.Sprintf("Unknown verdict %v.", verdict)) + panic(fmt.Sprintf("Unknown verdict: %d", verdict)) } + } - panic(fmt.Sprintf("Traversed past the entire list of iptables rules in table %q.", tablename)) + // We got through the entire table without a decision. Default to DROP + // for safety. + return TableDrop } // Precondition: pk.NetworkHeader is set. -func (it *IPTables) checkRule(hook Hook, pkt tcpip.PacketBuffer, table Table, ruleIdx int) Verdict { +func (it *IPTables) checkRule(hook Hook, pkt tcpip.PacketBuffer, table Table, ruleIdx int) RuleVerdict { rule := table.Rules[ruleIdx] // First check whether the packet matches the IP header filter. // TODO(gvisor.dev/issue/170): Support other fields of the filter. if rule.Filter.Protocol != 0 && rule.Filter.Protocol != header.IPv4(pkt.NetworkHeader).TransportProtocol() { - return Continue + return RuleContinue } // Go through each rule matcher. If they all match, run @@ -203,10 +223,10 @@ func (it *IPTables) checkRule(hook Hook, pkt tcpip.PacketBuffer, table Table, ru for _, matcher := range rule.Matchers { matches, hotdrop := matcher.Match(hook, pkt, "") if hotdrop { - return Drop + return RuleDrop } if !matches { - return Continue + return RuleContinue } } diff --git a/pkg/tcpip/iptables/targets.go b/pkg/tcpip/iptables/targets.go index 4dd281371..9fc60cfad 100644 --- a/pkg/tcpip/iptables/targets.go +++ b/pkg/tcpip/iptables/targets.go @@ -21,20 +21,20 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" ) -// UnconditionalAcceptTarget accepts all packets. -type UnconditionalAcceptTarget struct{} +// AcceptTarget accepts packets. +type AcceptTarget struct{} // Action implements Target.Action. -func (UnconditionalAcceptTarget) Action(packet tcpip.PacketBuffer) (Verdict, string) { - return Accept, "" +func (AcceptTarget) Action(packet tcpip.PacketBuffer) (RuleVerdict, string) { + return RuleAccept, "" } -// UnconditionalDropTarget denies all packets. -type UnconditionalDropTarget struct{} +// DropTarget drops packets. +type DropTarget struct{} // Action implements Target.Action. -func (UnconditionalDropTarget) Action(packet tcpip.PacketBuffer) (Verdict, string) { - return Drop, "" +func (DropTarget) Action(packet tcpip.PacketBuffer) (RuleVerdict, string) { + return RuleDrop, "" } // ErrorTarget logs an error and drops the packet. It represents a target that @@ -42,7 +42,26 @@ func (UnconditionalDropTarget) Action(packet tcpip.PacketBuffer) (Verdict, strin type ErrorTarget struct{} // Action implements Target.Action. -func (ErrorTarget) Action(packet tcpip.PacketBuffer) (Verdict, string) { - log.Warningf("ErrorTarget triggered.") - return Drop, "" +func (ErrorTarget) Action(packet tcpip.PacketBuffer) (RuleVerdict, string) { + log.Debugf("ErrorTarget triggered.") + return RuleDrop, "" +} + +// UserChainTarget marks a rule as the beginning of a user chain. +type UserChainTarget struct { + Name string +} + +// Action implements Target.Action. +func (UserChainTarget) Action(tcpip.PacketBuffer) (RuleVerdict, string) { + panic("UserChainTarget should never be called.") +} + +// ReturnTarget returns from the current chain. If the chain is a built-in, the +// hook's underflow should be called. +type ReturnTarget struct{} + +// Action implements Target.Action. +func (ReturnTarget) Action(tcpip.PacketBuffer) (RuleVerdict, string) { + return RuleReturn, "" } diff --git a/pkg/tcpip/iptables/types.go b/pkg/tcpip/iptables/types.go index 7d593c35c..5735d001b 100644 --- a/pkg/tcpip/iptables/types.go +++ b/pkg/tcpip/iptables/types.go @@ -56,44 +56,32 @@ const ( NumHooks ) -// A Verdict is returned by a rule's target to indicate how traversal of rules -// should (or should not) continue. -type Verdict int +// A TableVerdict is what a table decides should be done with a packet. +type TableVerdict int const ( - // Invalid indicates an unkonwn or erroneous verdict. - Invalid Verdict = iota + // TableAccept indicates the packet should continue through netstack. + TableAccept TableVerdict = iota - // Accept indicates the packet should continue traversing netstack as - // normal. - Accept - - // Drop inicates the packet should be dropped, stopping traversing - // netstack. - Drop - - // Stolen indicates the packet was co-opted by the target and should - // stop traversing netstack. - Stolen - - // Queue indicates the packet should be queued for userspace processing. - Queue + // TableAccept indicates the packet should be dropped. + TableDrop +) - // Repeat indicates the packet should re-traverse the chains for the - // current hook. - Repeat +// A RuleVerdict is what a rule decides should be done with a packet. +type RuleVerdict int - // None indicates no verdict was reached. - None +const ( + // RuleAccept indicates the packet should continue through netstack. + RuleAccept RuleVerdict = iota - // Jump indicates a jump to another chain. - Jump + // RuleContinue indicates the packet should continue to the next rule. + RuleContinue - // Continue indicates that traversal should continue at the next rule. - Continue + // RuleDrop indicates the packet should be dropped. + RuleDrop - // Return indicates that traversal should return to the calling chain. - Return + // RuleReturn indicates the packet should return to the previous chain. + RuleReturn ) // IPTables holds all the tables for a netstack. @@ -187,5 +175,5 @@ type Target interface { // Action takes an action on the packet and returns a verdict on how // traversal should (or should not) continue. If the return value is // Jump, it also returns the name of the chain to jump to. - Action(packet tcpip.PacketBuffer) (Verdict, string) + Action(packet tcpip.PacketBuffer) (RuleVerdict, string) } diff --git a/test/iptables/filter_input.go b/test/iptables/filter_input.go index bd6059921..e26d6a7d2 100644 --- a/test/iptables/filter_input.go +++ b/test/iptables/filter_input.go @@ -36,6 +36,10 @@ func init() { RegisterTestCase(FilterInputDropTCPSrcPort{}) RegisterTestCase(FilterInputDropUDPPort{}) RegisterTestCase(FilterInputDropUDP{}) + RegisterTestCase(FilterInputCreateUserChain{}) + RegisterTestCase(FilterInputDefaultPolicyAccept{}) + RegisterTestCase(FilterInputDefaultPolicyDrop{}) + RegisterTestCase(FilterInputReturnUnderflow{}) } // FilterInputDropUDP tests that we can drop UDP traffic. @@ -295,8 +299,119 @@ func (FilterInputRequireProtocolUDP) ContainerAction(ip net.IP) error { return nil } -// LocalAction implements TestCase.LocalAction. func (FilterInputRequireProtocolUDP) LocalAction(ip net.IP) error { // No-op. return nil } + +// FilterInputCreateUserChain tests chain creation. +type FilterInputCreateUserChain struct{} + +// Name implements TestCase.Name. +func (FilterInputCreateUserChain) Name() string { + return "FilterInputCreateUserChain" +} + +// ContainerAction implements TestCase.ContainerAction. +func (FilterInputCreateUserChain) ContainerAction(ip net.IP) error { + // Create a chain. + const chainName = "foochain" + if err := filterTable("-N", chainName); err != nil { + return err + } + + // Add a simple rule to the chain. + return filterTable("-A", chainName, "-j", "DROP") +} + +// LocalAction implements TestCase.LocalAction. +func (FilterInputCreateUserChain) LocalAction(ip net.IP) error { + // No-op. + return nil +} + +// FilterInputDefaultPolicyAccept tests the default ACCEPT policy. +type FilterInputDefaultPolicyAccept struct{} + +// Name implements TestCase.Name. +func (FilterInputDefaultPolicyAccept) Name() string { + return "FilterInputDefaultPolicyAccept" +} + +// ContainerAction implements TestCase.ContainerAction. +func (FilterInputDefaultPolicyAccept) ContainerAction(ip net.IP) error { + // Set the default policy to accept, then receive a packet. + if err := filterTable("-P", "INPUT", "ACCEPT"); err != nil { + return err + } + return listenUDP(acceptPort, sendloopDuration) +} + +// LocalAction implements TestCase.LocalAction. +func (FilterInputDefaultPolicyAccept) LocalAction(ip net.IP) error { + return sendUDPLoop(ip, acceptPort, sendloopDuration) +} + +// FilterInputDefaultPolicyDrop tests the default DROP policy. +type FilterInputDefaultPolicyDrop struct{} + +// Name implements TestCase.Name. +func (FilterInputDefaultPolicyDrop) Name() string { + return "FilterInputDefaultPolicyDrop" +} + +// ContainerAction implements TestCase.ContainerAction. +func (FilterInputDefaultPolicyDrop) ContainerAction(ip net.IP) error { + if err := filterTable("-P", "INPUT", "DROP"); err != nil { + return err + } + + // Listen for UDP packets on dropPort. + if err := listenUDP(dropPort, sendloopDuration); err == nil { + return fmt.Errorf("packets on port %d should have been dropped, but got a packet", dropPort) + } else if netErr, ok := err.(net.Error); !ok || !netErr.Timeout() { + return fmt.Errorf("error reading: %v", err) + } + + // At this point we know that reading timed out and never received a + // packet. + return nil +} + +// LocalAction implements TestCase.LocalAction. +func (FilterInputDefaultPolicyDrop) LocalAction(ip net.IP) error { + return sendUDPLoop(ip, acceptPort, sendloopDuration) +} + +// FilterInputReturnUnderflow tests that -j RETURN in a built-in chain causes +// the underflow rule (i.e. default policy) to be executed. +type FilterInputReturnUnderflow struct{} + +// Name implements TestCase.Name. +func (FilterInputReturnUnderflow) Name() string { + return "FilterInputReturnUnderflow" +} + +// ContainerAction implements TestCase.ContainerAction. +func (FilterInputReturnUnderflow) ContainerAction(ip net.IP) error { + // Add a RETURN rule followed by an unconditional accept, and set the + // default policy to DROP. + if err := filterTable("-A", "INPUT", "-j", "RETURN"); err != nil { + return err + } + if err := filterTable("-A", "INPUT", "-j", "DROP"); err != nil { + return err + } + if err := filterTable("-P", "INPUT", "ACCEPT"); err != nil { + return err + } + + // We should receive packets, as the RETURN rule will trigger the default + // ACCEPT policy. + return listenUDP(acceptPort, sendloopDuration) +} + +// LocalAction implements TestCase.LocalAction. +func (FilterInputReturnUnderflow) LocalAction(ip net.IP) error { + return sendUDPLoop(ip, acceptPort, sendloopDuration) +} diff --git a/test/iptables/iptables_test.go b/test/iptables/iptables_test.go index 41909582a..46a7c99b0 100644 --- a/test/iptables/iptables_test.go +++ b/test/iptables/iptables_test.go @@ -214,6 +214,30 @@ func TestFilterInputDropTCPSrcPort(t *testing.T) { } } +func TestFilterInputCreateUserChain(t *testing.T) { + if err := singleTest(FilterInputCreateUserChain{}); err != nil { + t.Fatal(err) + } +} + +func TestFilterInputDefaultPolicyAccept(t *testing.T) { + if err := singleTest(FilterInputDefaultPolicyAccept{}); err != nil { + t.Fatal(err) + } +} + +func TestFilterInputDefaultPolicyDrop(t *testing.T) { + if err := singleTest(FilterInputDefaultPolicyDrop{}); err != nil { + t.Fatal(err) + } +} + +func TestFilterInputReturnUnderflow(t *testing.T) { + if err := singleTest(FilterInputReturnUnderflow{}); err != nil { + t.Fatal(err) + } +} + func TestFilterOutputDropTCPDestPort(t *testing.T) { if err := singleTest(FilterOutputDropTCPDestPort{}); err != nil { t.Fatal(err) -- cgit v1.2.3 From 69bf39e8a47d3b4dcbbd04d2e8df476cdfab5e74 Mon Sep 17 00:00:00 2001 From: gVisor bot Date: Thu, 13 Feb 2020 10:58:47 -0800 Subject: Internal change. PiperOrigin-RevId: 294952610 --- pkg/abi/linux/socket.go | 13 ++++ pkg/sentry/socket/control/BUILD | 1 + pkg/sentry/socket/control/control.go | 43 +++++++++++++ pkg/sentry/socket/hostinet/socket.go | 11 +++- pkg/sentry/socket/netstack/netstack.go | 37 ++++++++++-- pkg/tcpip/tcpip.go | 25 ++++++++ pkg/tcpip/transport/udp/endpoint.go | 26 ++++++++ test/syscalls/linux/socket_ip_udp_generic.cc | 44 ++++++++++++++ test/syscalls/linux/socket_ipv4_udp_unbound.cc | 84 ++++++++++++++++++++++++++ test/syscalls/linux/udp_socket_test_cases.cc | 1 - 10 files changed, 278 insertions(+), 7 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/abi/linux/socket.go b/pkg/abi/linux/socket.go index 766ee4014..4a14ef691 100644 --- a/pkg/abi/linux/socket.go +++ b/pkg/abi/linux/socket.go @@ -411,6 +411,15 @@ type ControlMessageCredentials struct { GID uint32 } +// A ControlMessageIPPacketInfo is IP_PKTINFO socket control message. +// +// ControlMessageIPPacketInfo represents struct in_pktinfo from linux/in.h. +type ControlMessageIPPacketInfo struct { + NIC int32 + LocalAddr InetAddr + DestinationAddr InetAddr +} + // SizeOfControlMessageCredentials is the binary size of a // ControlMessageCredentials struct. var SizeOfControlMessageCredentials = int(binary.Size(ControlMessageCredentials{})) @@ -431,6 +440,10 @@ const SizeOfControlMessageTOS = 1 // SizeOfControlMessageTClass is the size of an IPV6_TCLASS control message. const SizeOfControlMessageTClass = 4 +// SizeOfControlMessageIPPacketInfo is the size of an IP_PKTINFO +// control message. +const SizeOfControlMessageIPPacketInfo = 12 + // SCM_MAX_FD is the maximum number of FDs accepted in a single sendmsg call. // From net/scm.h. const SCM_MAX_FD = 253 diff --git a/pkg/sentry/socket/control/BUILD b/pkg/sentry/socket/control/BUILD index 79e16d6e8..4d42d29cb 100644 --- a/pkg/sentry/socket/control/BUILD +++ b/pkg/sentry/socket/control/BUILD @@ -19,6 +19,7 @@ go_library( "//pkg/sentry/socket", "//pkg/sentry/socket/unix/transport", "//pkg/syserror", + "//pkg/tcpip", "//pkg/usermem", ], ) diff --git a/pkg/sentry/socket/control/control.go b/pkg/sentry/socket/control/control.go index 6145a7fc3..4667373d2 100644 --- a/pkg/sentry/socket/control/control.go +++ b/pkg/sentry/socket/control/control.go @@ -26,6 +26,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/socket" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/usermem" ) @@ -338,6 +339,22 @@ func PackTClass(t *kernel.Task, tClass int32, buf []byte) []byte { ) } +// PackIPPacketInfo packs an IP_PKTINFO socket control message. +func PackIPPacketInfo(t *kernel.Task, packetInfo tcpip.IPPacketInfo, buf []byte) []byte { + var p linux.ControlMessageIPPacketInfo + p.NIC = int32(packetInfo.NIC) + copy(p.LocalAddr[:], []byte(packetInfo.LocalAddr)) + copy(p.DestinationAddr[:], []byte(packetInfo.DestinationAddr)) + + return putCmsgStruct( + buf, + linux.SOL_IP, + linux.IP_PKTINFO, + t.Arch().Width(), + p, + ) +} + // PackControlMessages packs control messages into the given buffer. // // We skip control messages specific to Unix domain sockets. @@ -362,6 +379,10 @@ func PackControlMessages(t *kernel.Task, cmsgs socket.ControlMessages, buf []byt buf = PackTClass(t, cmsgs.IP.TClass, buf) } + if cmsgs.IP.HasIPPacketInfo { + buf = PackIPPacketInfo(t, cmsgs.IP.PacketInfo, buf) + } + return buf } @@ -394,6 +415,16 @@ func CmsgsSpace(t *kernel.Task, cmsgs socket.ControlMessages) int { return space } +// NewIPPacketInfo returns the IPPacketInfo struct. +func NewIPPacketInfo(packetInfo linux.ControlMessageIPPacketInfo) tcpip.IPPacketInfo { + var p tcpip.IPPacketInfo + p.NIC = tcpip.NICID(packetInfo.NIC) + copy([]byte(p.LocalAddr), packetInfo.LocalAddr[:]) + copy([]byte(p.DestinationAddr), packetInfo.DestinationAddr[:]) + + return p +} + // Parse parses a raw socket control message into portable objects. func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.ControlMessages, error) { var ( @@ -468,6 +499,18 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageTOS], usermem.ByteOrder, &cmsgs.IP.TOS) i += binary.AlignUp(length, width) + case linux.IP_PKTINFO: + if length < linux.SizeOfControlMessageIPPacketInfo { + return socket.ControlMessages{}, syserror.EINVAL + } + + cmsgs.IP.HasIPPacketInfo = true + var packetInfo linux.ControlMessageIPPacketInfo + binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageIPPacketInfo], usermem.ByteOrder, &packetInfo) + + cmsgs.IP.PacketInfo = NewIPPacketInfo(packetInfo) + i += binary.AlignUp(length, width) + default: return socket.ControlMessages{}, syserror.EINVAL } diff --git a/pkg/sentry/socket/hostinet/socket.go b/pkg/sentry/socket/hostinet/socket.go index de76388ac..22f78d2e2 100644 --- a/pkg/sentry/socket/hostinet/socket.go +++ b/pkg/sentry/socket/hostinet/socket.go @@ -289,7 +289,7 @@ func (s *socketOperations) GetSockOpt(t *kernel.Task, level int, name int, outPt switch level { case linux.SOL_IP: switch name { - case linux.IP_TOS, linux.IP_RECVTOS: + case linux.IP_TOS, linux.IP_RECVTOS, linux.IP_PKTINFO: optlen = sizeofInt32 } case linux.SOL_IPV6: @@ -336,6 +336,8 @@ func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt [ switch name { case linux.IP_TOS, linux.IP_RECVTOS: optlen = sizeofInt32 + case linux.IP_PKTINFO: + optlen = linux.SizeOfControlMessageIPPacketInfo } case linux.SOL_IPV6: switch name { @@ -473,7 +475,14 @@ func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags case syscall.IP_TOS: controlMessages.IP.HasTOS = true binary.Unmarshal(unixCmsg.Data[:linux.SizeOfControlMessageTOS], usermem.ByteOrder, &controlMessages.IP.TOS) + + case syscall.IP_PKTINFO: + controlMessages.IP.HasIPPacketInfo = true + var packetInfo linux.ControlMessageIPPacketInfo + binary.Unmarshal(unixCmsg.Data[:linux.SizeOfControlMessageIPPacketInfo], usermem.ByteOrder, &packetInfo) + controlMessages.IP.PacketInfo = control.NewIPPacketInfo(packetInfo) } + case syscall.SOL_IPV6: switch unixCmsg.Header.Type { case syscall.IPV6_TCLASS: diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index ed2fbcceb..9757fbfba 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -1414,6 +1414,21 @@ func getSockOptIP(t *kernel.Task, ep commonEndpoint, name, outLen int, family in } return o, nil + case linux.IP_PKTINFO: + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + + v, err := ep.GetSockOptBool(tcpip.ReceiveIPPacketInfoOption) + if err != nil { + return nil, syserr.TranslateNetstackError(err) + } + var o int32 + if v { + o = 1 + } + return o, nil + default: emitUnimplementedEventIP(t, name) } @@ -1762,6 +1777,7 @@ func setSockOptIPv6(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) linux.IPV6_IPSEC_POLICY, linux.IPV6_JOIN_ANYCAST, linux.IPV6_LEAVE_ANYCAST, + // TODO(b/148887420): Add support for IPV6_PKTINFO. linux.IPV6_PKTINFO, linux.IPV6_ROUTER_ALERT, linux.IPV6_XFRM_POLICY, @@ -1949,6 +1965,16 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s } return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.ReceiveTOSOption, v != 0)) + case linux.IP_PKTINFO: + if len(optVal) == 0 { + return nil + } + v, err := parseIntOrChar(optVal) + if err != nil { + return err + } + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.ReceiveIPPacketInfoOption, v != 0)) + case linux.IP_ADD_SOURCE_MEMBERSHIP, linux.IP_BIND_ADDRESS_NO_PORT, linux.IP_BLOCK_SOURCE, @@ -1964,7 +1990,6 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s linux.IP_NODEFRAG, linux.IP_OPTIONS, linux.IP_PASSSEC, - linux.IP_PKTINFO, linux.IP_RECVERR, linux.IP_RECVFRAGSIZE, linux.IP_RECVOPTS, @@ -2395,10 +2420,12 @@ func (s *SocketOperations) nonBlockingRead(ctx context.Context, dst usermem.IOSe func (s *SocketOperations) controlMessages() socket.ControlMessages { return socket.ControlMessages{ IP: tcpip.ControlMessages{ - HasTimestamp: s.readCM.HasTimestamp && s.sockOptTimestamp, - Timestamp: s.readCM.Timestamp, - HasTOS: s.readCM.HasTOS, - TOS: s.readCM.TOS, + HasTimestamp: s.readCM.HasTimestamp && s.sockOptTimestamp, + Timestamp: s.readCM.Timestamp, + HasTOS: s.readCM.HasTOS, + TOS: s.readCM.TOS, + HasIPPacketInfo: s.readCM.HasIPPacketInfo, + PacketInfo: s.readCM.PacketInfo, }, } } diff --git a/pkg/tcpip/tcpip.go b/pkg/tcpip/tcpip.go index 0e944712f..9ca39ce40 100644 --- a/pkg/tcpip/tcpip.go +++ b/pkg/tcpip/tcpip.go @@ -328,6 +328,12 @@ type ControlMessages struct { // Tclass is the IPv6 traffic class of the associated packet. TClass int32 + + // HasIPPacketInfo indicates whether PacketInfo is set. + HasIPPacketInfo bool + + // PacketInfo holds interface and address data on an incoming packet. + PacketInfo IPPacketInfo } // Endpoint is the interface implemented by transport protocols (e.g., tcp, udp) @@ -503,6 +509,11 @@ const ( // V6OnlyOption is used by {G,S}etSockOptBool to specify whether an IPv6 // socket is to be restricted to sending and receiving IPv6 packets only. V6OnlyOption + + // ReceiveIPPacketInfoOption is used by {G,S}etSockOptBool to specify + // if more inforamtion is provided with incoming packets such + // as interface index and address. + ReceiveIPPacketInfoOption ) // SockOptInt represents socket options which values have the int type. @@ -685,6 +696,20 @@ type IPv4TOSOption uint8 // for all subsequent outgoing IPv6 packets from the endpoint. type IPv6TrafficClassOption uint8 +// IPPacketInfo is the message struture for IP_PKTINFO. +// +// +stateify savable +type IPPacketInfo struct { + // NIC is the ID of the NIC to be used. + NIC NICID + + // LocalAddr is the local address. + LocalAddr Address + + // DestinationAddr is the destination address. + DestinationAddr Address +} + // Route is a row in the routing table. It specifies through which NIC (and // gateway) sets of packets should be routed. A row is considered viable if the // masked target address matches the destination address in the row. diff --git a/pkg/tcpip/transport/udp/endpoint.go b/pkg/tcpip/transport/udp/endpoint.go index c9cbed8f4..3fe91cac2 100644 --- a/pkg/tcpip/transport/udp/endpoint.go +++ b/pkg/tcpip/transport/udp/endpoint.go @@ -29,6 +29,7 @@ import ( type udpPacket struct { udpPacketEntry senderAddress tcpip.FullAddress + packetInfo tcpip.IPPacketInfo data buffer.VectorisedView `state:".(buffer.VectorisedView)"` timestamp int64 tos uint8 @@ -118,6 +119,9 @@ type endpoint struct { // as ancillary data to ControlMessages on Read. receiveTOS bool + // receiveIPPacketInfo determines if the packet info is returned by Read. + receiveIPPacketInfo bool + // shutdownFlags represent the current shutdown state of the endpoint. shutdownFlags tcpip.ShutdownFlags @@ -254,11 +258,17 @@ func (e *endpoint) Read(addr *tcpip.FullAddress) (buffer.View, tcpip.ControlMess } e.mu.RLock() receiveTOS := e.receiveTOS + receiveIPPacketInfo := e.receiveIPPacketInfo e.mu.RUnlock() if receiveTOS { cm.HasTOS = true cm.TOS = p.tos } + + if receiveIPPacketInfo { + cm.HasIPPacketInfo = true + cm.PacketInfo = p.packetInfo + } return p.data.ToView(), cm, nil } @@ -495,6 +505,13 @@ func (e *endpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error { } e.v6only = v + return nil + + case tcpip.ReceiveIPPacketInfoOption: + e.mu.Lock() + e.receiveIPPacketInfo = v + e.mu.Unlock() + return nil } return nil @@ -703,6 +720,12 @@ func (e *endpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { e.mu.RUnlock() return v, nil + + case tcpip.ReceiveIPPacketInfoOption: + e.mu.RLock() + v := e.receiveIPPacketInfo + e.mu.RUnlock() + return v, nil } return false, tcpip.ErrUnknownProtocolOption @@ -1247,6 +1270,9 @@ func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pk switch r.NetProto { case header.IPv4ProtocolNumber: packet.tos, _ = header.IPv4(pkt.NetworkHeader).TOS() + packet.packetInfo.LocalAddr = r.LocalAddress + packet.packetInfo.DestinationAddr = r.RemoteAddress + packet.packetInfo.NIC = r.NICID() } packet.timestamp = e.stack.NowNanoseconds() diff --git a/test/syscalls/linux/socket_ip_udp_generic.cc b/test/syscalls/linux/socket_ip_udp_generic.cc index 53290bed7..db5663ecd 100644 --- a/test/syscalls/linux/socket_ip_udp_generic.cc +++ b/test/syscalls/linux/socket_ip_udp_generic.cc @@ -357,5 +357,49 @@ TEST_P(UDPSocketPairTest, SetReuseAddrReusePort) { EXPECT_EQ(get, kSockOptOn); } +// Test getsockopt for a socket which is not set with IP_PKTINFO option. +TEST_P(UDPSocketPairTest, IPPKTINFODefault) { + auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); + + int get = -1; + socklen_t get_len = sizeof(get); + + ASSERT_THAT( + getsockopt(sockets->first_fd(), SOL_IP, IP_PKTINFO, &get, &get_len), + SyscallSucceedsWithValue(0)); + EXPECT_EQ(get_len, sizeof(get)); + EXPECT_EQ(get, kSockOptOff); +} + +// Test setsockopt and getsockopt for a socket with IP_PKTINFO option. +TEST_P(UDPSocketPairTest, SetAndGetIPPKTINFO) { + auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); + + int level = SOL_IP; + int type = IP_PKTINFO; + + // Check getsockopt before IP_PKTINFO is set. + int get = -1; + socklen_t get_len = sizeof(get); + + ASSERT_THAT(setsockopt(sockets->first_fd(), level, type, &kSockOptOn, + sizeof(kSockOptOn)), + SyscallSucceedsWithValue(0)); + + ASSERT_THAT(getsockopt(sockets->first_fd(), level, type, &get, &get_len), + SyscallSucceedsWithValue(0)); + EXPECT_EQ(get, kSockOptOn); + EXPECT_EQ(get_len, sizeof(get)); + + ASSERT_THAT(setsockopt(sockets->first_fd(), level, type, &kSockOptOff, + sizeof(kSockOptOff)), + SyscallSucceedsWithValue(0)); + + ASSERT_THAT(getsockopt(sockets->first_fd(), level, type, &get, &get_len), + SyscallSucceedsWithValue(0)); + EXPECT_EQ(get, kSockOptOff); + EXPECT_EQ(get_len, sizeof(get)); +} + } // namespace testing } // namespace gvisor diff --git a/test/syscalls/linux/socket_ipv4_udp_unbound.cc b/test/syscalls/linux/socket_ipv4_udp_unbound.cc index 990ccf23c..bc4b07a62 100644 --- a/test/syscalls/linux/socket_ipv4_udp_unbound.cc +++ b/test/syscalls/linux/socket_ipv4_udp_unbound.cc @@ -15,6 +15,7 @@ #include "test/syscalls/linux/socket_ipv4_udp_unbound.h" #include +#include #include #include #include @@ -2128,5 +2129,88 @@ TEST_P(IPv4UDPUnboundSocketTest, ReuseAddrReusePortDistribution) { SyscallSucceedsWithValue(kMessageSize)); } +// Test that socket will receive packet info control message. +TEST_P(IPv4UDPUnboundSocketTest, SetAndReceiveIPPKTINFO) { + // TODO(gvisor.dev/issue/1202): ioctl() is not supported by hostinet. + SKIP_IF((IsRunningWithHostinet())); + + auto sender = ASSERT_NO_ERRNO_AND_VALUE(NewSocket()); + auto receiver = ASSERT_NO_ERRNO_AND_VALUE(NewSocket()); + auto sender_addr = V4Loopback(); + int level = SOL_IP; + int type = IP_PKTINFO; + + ASSERT_THAT( + bind(receiver->get(), reinterpret_cast(&sender_addr.addr), + sender_addr.addr_len), + SyscallSucceeds()); + socklen_t sender_addr_len = sender_addr.addr_len; + ASSERT_THAT(getsockname(receiver->get(), + reinterpret_cast(&sender_addr.addr), + &sender_addr_len), + SyscallSucceeds()); + EXPECT_EQ(sender_addr_len, sender_addr.addr_len); + + auto receiver_addr = V4Loopback(); + reinterpret_cast(&receiver_addr.addr)->sin_port = + reinterpret_cast(&sender_addr.addr)->sin_port; + ASSERT_THAT( + connect(sender->get(), reinterpret_cast(&receiver_addr.addr), + receiver_addr.addr_len), + SyscallSucceeds()); + + // Allow socket to receive control message. + ASSERT_THAT( + setsockopt(receiver->get(), level, type, &kSockOptOn, sizeof(kSockOptOn)), + SyscallSucceeds()); + + // Prepare message to send. + constexpr size_t kDataLength = 1024; + msghdr sent_msg = {}; + iovec sent_iov = {}; + char sent_data[kDataLength]; + sent_iov.iov_base = sent_data; + sent_iov.iov_len = kDataLength; + sent_msg.msg_iov = &sent_iov; + sent_msg.msg_iovlen = 1; + sent_msg.msg_flags = 0; + + ASSERT_THAT(RetryEINTR(sendmsg)(sender->get(), &sent_msg, 0), + SyscallSucceedsWithValue(kDataLength)); + + msghdr received_msg = {}; + iovec received_iov = {}; + char received_data[kDataLength]; + char received_cmsg_buf[CMSG_SPACE(sizeof(in_pktinfo))] = {}; + size_t cmsg_data_len = sizeof(in_pktinfo); + received_iov.iov_base = received_data; + received_iov.iov_len = kDataLength; + received_msg.msg_iov = &received_iov; + received_msg.msg_iovlen = 1; + received_msg.msg_controllen = CMSG_LEN(cmsg_data_len); + received_msg.msg_control = received_cmsg_buf; + + ASSERT_THAT(RetryEINTR(recvmsg)(receiver->get(), &received_msg, 0), + SyscallSucceedsWithValue(kDataLength)); + + cmsghdr* cmsg = CMSG_FIRSTHDR(&received_msg); + ASSERT_NE(cmsg, nullptr); + EXPECT_EQ(cmsg->cmsg_len, CMSG_LEN(cmsg_data_len)); + EXPECT_EQ(cmsg->cmsg_level, level); + EXPECT_EQ(cmsg->cmsg_type, type); + + // Get loopback index. + ifreq ifr = {}; + absl::SNPrintF(ifr.ifr_name, IFNAMSIZ, "lo"); + ASSERT_THAT(ioctl(sender->get(), SIOCGIFINDEX, &ifr), SyscallSucceeds()); + ASSERT_NE(ifr.ifr_ifindex, 0); + + // Check the data + in_pktinfo received_pktinfo = {}; + memcpy(&received_pktinfo, CMSG_DATA(cmsg), sizeof(in_pktinfo)); + EXPECT_EQ(received_pktinfo.ipi_ifindex, ifr.ifr_ifindex); + EXPECT_EQ(received_pktinfo.ipi_spec_dst.s_addr, htonl(INADDR_LOOPBACK)); + EXPECT_EQ(received_pktinfo.ipi_addr.s_addr, htonl(INADDR_LOOPBACK)); +} } // namespace testing } // namespace gvisor diff --git a/test/syscalls/linux/udp_socket_test_cases.cc b/test/syscalls/linux/udp_socket_test_cases.cc index a2f6ef8cc..9f8de6b48 100644 --- a/test/syscalls/linux/udp_socket_test_cases.cc +++ b/test/syscalls/linux/udp_socket_test_cases.cc @@ -1495,6 +1495,5 @@ TEST_P(UdpSocketTest, SendAndReceiveTOS) { memcpy(&received_tos, CMSG_DATA(cmsg), sizeof(received_tos)); EXPECT_EQ(received_tos, sent_tos); } - } // namespace testing } // namespace gvisor -- cgit v1.2.3 From 6ef63cd7da107d487fda7c48af50fa9802913cd9 Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Wed, 12 Feb 2020 16:19:06 -0800 Subject: We can now create and jump in iptables. For example: $ iptables -N foochain $ iptables -A INPUT -j foochain --- pkg/abi/linux/netfilter.go | 9 +- pkg/sentry/socket/netfilter/BUILD | 1 + pkg/sentry/socket/netfilter/netfilter.go | 62 +++++++-- pkg/sentry/socket/netfilter/targets.go | 35 +++++ pkg/tcpip/iptables/iptables.go | 103 +++++++++------ pkg/tcpip/iptables/targets.go | 20 ++- pkg/tcpip/iptables/types.go | 21 +-- test/iptables/filter_input.go | 217 ++++++++++++++++++++++++++++--- test/iptables/iptables_test.go | 36 +++++ test/iptables/iptables_util.go | 10 ++ 10 files changed, 420 insertions(+), 94 deletions(-) create mode 100644 pkg/sentry/socket/netfilter/targets.go (limited to 'pkg/sentry/socket') diff --git a/pkg/abi/linux/netfilter.go b/pkg/abi/linux/netfilter.go index bbc4df74c..bd2e13ba1 100644 --- a/pkg/abi/linux/netfilter.go +++ b/pkg/abi/linux/netfilter.go @@ -225,11 +225,14 @@ type XTEntryTarget struct { // SizeOfXTEntryTarget is the size of an XTEntryTarget. const SizeOfXTEntryTarget = 32 -// XTStandardTarget is a builtin target, one of ACCEPT, DROP, JUMP, QUEUE, or -// RETURN. It corresponds to struct xt_standard_target in +// XTStandardTarget is a built-in target, one of ACCEPT, DROP, JUMP, QUEUE, +// RETURN, or jump. It corresponds to struct xt_standard_target in // include/uapi/linux/netfilter/x_tables.h. type XTStandardTarget struct { - Target XTEntryTarget + Target XTEntryTarget + // A positive verdict indicates a jump, and is the offset from the + // start of the table to jump to. A negative value means one of the + // other built-in targets. Verdict int32 _ [4]byte } diff --git a/pkg/sentry/socket/netfilter/BUILD b/pkg/sentry/socket/netfilter/BUILD index c91ec7494..7cd2ce55b 100644 --- a/pkg/sentry/socket/netfilter/BUILD +++ b/pkg/sentry/socket/netfilter/BUILD @@ -7,6 +7,7 @@ go_library( srcs = [ "extensions.go", "netfilter.go", + "targets.go", "tcp_matcher.go", "udp_matcher.go", ], diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index 3fc80e0de..d322e4144 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -240,13 +240,15 @@ func marshalTarget(target iptables.Target) []byte { return marshalErrorTarget(tg.Name) case iptables.ReturnTarget: return marshalStandardTarget(iptables.RuleReturn) + case JumpTarget: + return marshalJumpTarget(tg) default: panic(fmt.Errorf("unknown target of type %T", target)) } } func marshalStandardTarget(verdict iptables.RuleVerdict) []byte { - nflog("convert to binary: marshalling standard target with size %d", linux.SizeOfXTStandardTarget) + nflog("convert to binary: marshalling standard target") // The target's name will be the empty string. target := linux.XTStandardTarget{ @@ -274,6 +276,23 @@ func marshalErrorTarget(errorName string) []byte { return binary.Marshal(ret, usermem.ByteOrder, target) } +func marshalJumpTarget(jt JumpTarget) []byte { + nflog("convert to binary: marshalling jump target") + + // The target's name will be the empty string. + target := linux.XTStandardTarget{ + Target: linux.XTEntryTarget{ + TargetSize: linux.SizeOfXTStandardTarget, + }, + // Verdict is overloaded by the ABI. When positive, it holds + // the jump offset from the start of the table. + Verdict: int32(jt.Offset), + } + + ret := make([]byte, 0, linux.SizeOfXTStandardTarget) + return binary.Marshal(ret, usermem.ByteOrder, target) +} + // translateFromStandardVerdict translates verdicts the same way as the iptables // tool. func translateFromStandardVerdict(verdict iptables.RuleVerdict) int32 { @@ -335,7 +354,8 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // Convert input into a list of rules and their offsets. var offset uint32 - var offsets []uint32 + // offsets maps rule byte offsets to their position in table.Rules. + offsets := map[uint32]int{} for entryIdx := uint32(0); entryIdx < replace.NumEntries; entryIdx++ { nflog("set entries: processing entry at offset %d", offset) @@ -396,11 +416,12 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { Target: target, Matchers: matchers, }) - offsets = append(offsets, offset) + offsets[offset] = int(entryIdx) offset += uint32(entry.NextOffset) if initialOptValLen-len(optVal) != int(entry.NextOffset) { nflog("entry NextOffset is %d, but entry took up %d bytes", entry.NextOffset, initialOptValLen-len(optVal)) + return syserr.ErrInvalidArgument } } @@ -409,13 +430,13 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { for hook, _ := range replace.HookEntry { if table.ValidHooks()&(1<= 0 indicates a jump. + return JumpTarget{Offset: uint32(standardTarget.Verdict)}, nil + } case errorTargetName: // Error target. diff --git a/pkg/sentry/socket/netfilter/targets.go b/pkg/sentry/socket/netfilter/targets.go new file mode 100644 index 000000000..c421b87cf --- /dev/null +++ b/pkg/sentry/socket/netfilter/targets.go @@ -0,0 +1,35 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netfilter + +import ( + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/iptables" +) + +// JumpTarget implements iptables.Target. +type JumpTarget struct { + // Offset is the byte offset of the rule to jump to. It is used for + // marshaling and unmarshaling. + Offset uint32 + + // RuleNum is the rule to jump to. + RuleNum int +} + +// Action implements iptables.Target.Action. +func (jt JumpTarget) Action(tcpip.PacketBuffer) (iptables.RuleVerdict, int) { + return iptables.RuleJump, jt.RuleNum +} diff --git a/pkg/tcpip/iptables/iptables.go b/pkg/tcpip/iptables/iptables.go index 75a433a3b..dbaccbb36 100644 --- a/pkg/tcpip/iptables/iptables.go +++ b/pkg/tcpip/iptables/iptables.go @@ -135,25 +135,53 @@ func EmptyFilterTable() Table { } } +// A chainVerdict is what a table decides should be done with a packet. +type chainVerdict int + +const ( + // chainAccept indicates the packet should continue through netstack. + chainAccept chainVerdict = iota + + // chainAccept indicates the packet should be dropped. + chainDrop + + // chainReturn indicates the packet should return to the calling chain + // or the underflow rule of a builtin chain. + chainReturn +) + // Check runs pkt through the rules for hook. It returns true when the packet // should continue traversing the network stack and false when it should be // dropped. // // Precondition: pkt.NetworkHeader is set. func (it *IPTables) Check(hook Hook, pkt tcpip.PacketBuffer) bool { - // TODO(gvisor.dev/issue/170): A lot of this is uncomplicated because - // we're missing features. Jumps, the call stack, etc. aren't checked - // for yet because we're yet to support them. - // Go through each table containing the hook. for _, tablename := range it.Priorities[hook] { - switch verdict := it.checkTable(hook, pkt, tablename); verdict { + table := it.Tables[tablename] + ruleIdx := table.BuiltinChains[hook] + switch verdict := it.checkChain(hook, pkt, table, ruleIdx); verdict { // If the table returns Accept, move on to the next table. - case TableAccept: + case chainAccept: continue // The Drop verdict is final. - case TableDrop: + case chainDrop: return false + case chainReturn: + // Any Return from a built-in chain means we have to + // call the underflow. + underflow := table.Rules[table.Underflows[hook]] + switch v, _ := underflow.Target.Action(pkt); v { + case RuleAccept: + continue + case RuleDrop: + return false + case RuleJump, RuleReturn: + panic("Underflows should only return RuleAccept or RuleDrop.") + default: + panic(fmt.Sprintf("Unknown verdict: %d", v)) + } + default: panic(fmt.Sprintf("Unknown verdict %v.", verdict)) } @@ -164,37 +192,37 @@ func (it *IPTables) Check(hook Hook, pkt tcpip.PacketBuffer) bool { } // Precondition: pkt.NetworkHeader is set. -func (it *IPTables) checkTable(hook Hook, pkt tcpip.PacketBuffer, tablename string) TableVerdict { +func (it *IPTables) checkChain(hook Hook, pkt tcpip.PacketBuffer, table Table, ruleIdx int) chainVerdict { // Start from ruleIdx and walk the list of rules until a rule gives us // a verdict. - table := it.Tables[tablename] - for ruleIdx := table.BuiltinChains[hook]; ruleIdx < len(table.Rules); ruleIdx++ { - switch verdict := it.checkRule(hook, pkt, table, ruleIdx); verdict { + for ruleIdx < len(table.Rules) { + switch verdict, jumpTo := it.checkRule(hook, pkt, table, ruleIdx); verdict { case RuleAccept: - return TableAccept + return chainAccept case RuleDrop: - return TableDrop - - case RuleContinue: - continue + return chainDrop case RuleReturn: - // TODO(gvisor.dev/issue/170): We don't implement jump - // yet, so any Return is from a built-in chain. That - // means we have to to call the underflow. - underflow := table.Rules[table.Underflows[hook]] - // Underflow is guaranteed to be an unconditional - // ACCEPT or DROP. - switch v, _ := underflow.Target.Action(pkt); v { - case RuleAccept: - return TableAccept - case RuleDrop: - return TableDrop - case RuleContinue, RuleReturn: - panic("Underflows should only return RuleAccept or RuleDrop.") + return chainReturn + + case RuleJump: + // "Jumping" to the next rule just means we're + // continuing on down the list. + if jumpTo == ruleIdx+1 { + ruleIdx++ + continue + } + switch verdict := it.checkChain(hook, pkt, table, jumpTo); verdict { + case chainAccept: + return chainAccept + case chainDrop: + return chainDrop + case chainReturn: + ruleIdx++ + continue default: - panic(fmt.Sprintf("Unknown verdict: %d", v)) + panic(fmt.Sprintf("Unknown verdict: %d", verdict)) } default: @@ -205,17 +233,18 @@ func (it *IPTables) checkTable(hook Hook, pkt tcpip.PacketBuffer, tablename stri // We got through the entire table without a decision. Default to DROP // for safety. - return TableDrop + return chainDrop } // Precondition: pk.NetworkHeader is set. -func (it *IPTables) checkRule(hook Hook, pkt tcpip.PacketBuffer, table Table, ruleIdx int) RuleVerdict { +func (it *IPTables) checkRule(hook Hook, pkt tcpip.PacketBuffer, table Table, ruleIdx int) (RuleVerdict, int) { rule := table.Rules[ruleIdx] // First check whether the packet matches the IP header filter. // TODO(gvisor.dev/issue/170): Support other fields of the filter. if rule.Filter.Protocol != 0 && rule.Filter.Protocol != header.IPv4(pkt.NetworkHeader).TransportProtocol() { - return RuleContinue + // Continue on to the next rule. + return RuleJump, ruleIdx + 1 } // Go through each rule matcher. If they all match, run @@ -223,14 +252,14 @@ func (it *IPTables) checkRule(hook Hook, pkt tcpip.PacketBuffer, table Table, ru for _, matcher := range rule.Matchers { matches, hotdrop := matcher.Match(hook, pkt, "") if hotdrop { - return RuleDrop + return RuleDrop, 0 } if !matches { - return RuleContinue + // Continue on to the next rule. + return RuleJump, ruleIdx + 1 } } // All the matchers matched, so run the target. - verdict, _ := rule.Target.Action(pkt) - return verdict + return rule.Target.Action(pkt) } diff --git a/pkg/tcpip/iptables/targets.go b/pkg/tcpip/iptables/targets.go index 9fc60cfad..81a2e39a2 100644 --- a/pkg/tcpip/iptables/targets.go +++ b/pkg/tcpip/iptables/targets.go @@ -12,8 +12,6 @@ // See the License for the specific language governing permissions and // limitations under the License. -// This file contains various Targets. - package iptables import ( @@ -25,16 +23,16 @@ import ( type AcceptTarget struct{} // Action implements Target.Action. -func (AcceptTarget) Action(packet tcpip.PacketBuffer) (RuleVerdict, string) { - return RuleAccept, "" +func (AcceptTarget) Action(packet tcpip.PacketBuffer) (RuleVerdict, int) { + return RuleAccept, 0 } // DropTarget drops packets. type DropTarget struct{} // Action implements Target.Action. -func (DropTarget) Action(packet tcpip.PacketBuffer) (RuleVerdict, string) { - return RuleDrop, "" +func (DropTarget) Action(packet tcpip.PacketBuffer) (RuleVerdict, int) { + return RuleDrop, 0 } // ErrorTarget logs an error and drops the packet. It represents a target that @@ -42,9 +40,9 @@ func (DropTarget) Action(packet tcpip.PacketBuffer) (RuleVerdict, string) { type ErrorTarget struct{} // Action implements Target.Action. -func (ErrorTarget) Action(packet tcpip.PacketBuffer) (RuleVerdict, string) { +func (ErrorTarget) Action(packet tcpip.PacketBuffer) (RuleVerdict, int) { log.Debugf("ErrorTarget triggered.") - return RuleDrop, "" + return RuleDrop, 0 } // UserChainTarget marks a rule as the beginning of a user chain. @@ -53,7 +51,7 @@ type UserChainTarget struct { } // Action implements Target.Action. -func (UserChainTarget) Action(tcpip.PacketBuffer) (RuleVerdict, string) { +func (UserChainTarget) Action(tcpip.PacketBuffer) (RuleVerdict, int) { panic("UserChainTarget should never be called.") } @@ -62,6 +60,6 @@ func (UserChainTarget) Action(tcpip.PacketBuffer) (RuleVerdict, string) { type ReturnTarget struct{} // Action implements Target.Action. -func (ReturnTarget) Action(tcpip.PacketBuffer) (RuleVerdict, string) { - return RuleReturn, "" +func (ReturnTarget) Action(tcpip.PacketBuffer) (RuleVerdict, int) { + return RuleReturn, 0 } diff --git a/pkg/tcpip/iptables/types.go b/pkg/tcpip/iptables/types.go index 5735d001b..7d032fd23 100644 --- a/pkg/tcpip/iptables/types.go +++ b/pkg/tcpip/iptables/types.go @@ -56,17 +56,6 @@ const ( NumHooks ) -// A TableVerdict is what a table decides should be done with a packet. -type TableVerdict int - -const ( - // TableAccept indicates the packet should continue through netstack. - TableAccept TableVerdict = iota - - // TableAccept indicates the packet should be dropped. - TableDrop -) - // A RuleVerdict is what a rule decides should be done with a packet. type RuleVerdict int @@ -74,12 +63,12 @@ const ( // RuleAccept indicates the packet should continue through netstack. RuleAccept RuleVerdict = iota - // RuleContinue indicates the packet should continue to the next rule. - RuleContinue - // RuleDrop indicates the packet should be dropped. RuleDrop + // RuleJump indicates the packet should jump to another chain. + RuleJump + // RuleReturn indicates the packet should return to the previous chain. RuleReturn ) @@ -174,6 +163,6 @@ type Matcher interface { type Target interface { // Action takes an action on the packet and returns a verdict on how // traversal should (or should not) continue. If the return value is - // Jump, it also returns the name of the chain to jump to. - Action(packet tcpip.PacketBuffer) (RuleVerdict, string) + // Jump, it also returns the index of the rule to jump to. + Action(packet tcpip.PacketBuffer) (RuleVerdict, int) } diff --git a/test/iptables/filter_input.go b/test/iptables/filter_input.go index e26d6a7d2..706c09cea 100644 --- a/test/iptables/filter_input.go +++ b/test/iptables/filter_input.go @@ -26,6 +26,7 @@ const ( acceptPort = 2402 sendloopDuration = 2 * time.Second network = "udp4" + chainName = "foochain" ) func init() { @@ -40,6 +41,12 @@ func init() { RegisterTestCase(FilterInputDefaultPolicyAccept{}) RegisterTestCase(FilterInputDefaultPolicyDrop{}) RegisterTestCase(FilterInputReturnUnderflow{}) + RegisterTestCase(FilterInputSerializeJump{}) + RegisterTestCase(FilterInputJumpBasic{}) + RegisterTestCase(FilterInputJumpReturn{}) + RegisterTestCase(FilterInputJumpReturnDrop{}) + RegisterTestCase(FilterInputJumpBuiltin{}) + RegisterTestCase(FilterInputJumpTwice{}) } // FilterInputDropUDP tests that we can drop UDP traffic. @@ -267,13 +274,12 @@ func (FilterInputMultiUDPRules) Name() string { // ContainerAction implements TestCase.ContainerAction. func (FilterInputMultiUDPRules) ContainerAction(ip net.IP) error { - if err := filterTable("-A", "INPUT", "-p", "udp", "-m", "udp", "--destination-port", fmt.Sprintf("%d", dropPort), "-j", "DROP"); err != nil { - return err - } - if err := filterTable("-A", "INPUT", "-p", "udp", "-m", "udp", "--destination-port", fmt.Sprintf("%d", acceptPort), "-j", "ACCEPT"); err != nil { - return err + rules := [][]string{ + {"-A", "INPUT", "-p", "udp", "-m", "udp", "--destination-port", fmt.Sprintf("%d", dropPort), "-j", "DROP"}, + {"-A", "INPUT", "-p", "udp", "-m", "udp", "--destination-port", fmt.Sprintf("%d", acceptPort), "-j", "ACCEPT"}, + {"-L"}, } - return filterTable("-L") + return filterTableRules(rules) } // LocalAction implements TestCase.LocalAction. @@ -314,14 +320,13 @@ func (FilterInputCreateUserChain) Name() string { // ContainerAction implements TestCase.ContainerAction. func (FilterInputCreateUserChain) ContainerAction(ip net.IP) error { - // Create a chain. - const chainName = "foochain" - if err := filterTable("-N", chainName); err != nil { - return err + rules := [][]string{ + // Create a chain. + {"-N", chainName}, + // Add a simple rule to the chain. + {"-A", chainName, "-j", "DROP"}, } - - // Add a simple rule to the chain. - return filterTable("-A", chainName, "-j", "DROP") + return filterTableRules(rules) } // LocalAction implements TestCase.LocalAction. @@ -396,13 +401,12 @@ func (FilterInputReturnUnderflow) Name() string { func (FilterInputReturnUnderflow) ContainerAction(ip net.IP) error { // Add a RETURN rule followed by an unconditional accept, and set the // default policy to DROP. - if err := filterTable("-A", "INPUT", "-j", "RETURN"); err != nil { - return err + rules := [][]string{ + {"-A", "INPUT", "-j", "RETURN"}, + {"-A", "INPUT", "-j", "DROP"}, + {"-P", "INPUT", "ACCEPT"}, } - if err := filterTable("-A", "INPUT", "-j", "DROP"); err != nil { - return err - } - if err := filterTable("-P", "INPUT", "ACCEPT"); err != nil { + if err := filterTableRules(rules); err != nil { return err } @@ -415,3 +419,178 @@ func (FilterInputReturnUnderflow) ContainerAction(ip net.IP) error { func (FilterInputReturnUnderflow) LocalAction(ip net.IP) error { return sendUDPLoop(ip, acceptPort, sendloopDuration) } + +// Verify that we can serialize jumps. +type FilterInputSerializeJump struct{} + +// Name implements TestCase.Name. +func (FilterInputSerializeJump) Name() string { + return "FilterInputSerializeJump" +} + +// ContainerAction implements TestCase.ContainerAction. +func (FilterInputSerializeJump) ContainerAction(ip net.IP) error { + // Write a JUMP rule, the serialize it with `-L`. + rules := [][]string{ + {"-N", chainName}, + {"-A", "INPUT", "-j", chainName}, + {"-L"}, + } + return filterTableRules(rules) +} + +// LocalAction implements TestCase.LocalAction. +func (FilterInputSerializeJump) LocalAction(ip net.IP) error { + // No-op. + return nil +} + +// Jump to a chain and execute a rule there. +type FilterInputJumpBasic struct{} + +// Name implements TestCase.Name. +func (FilterInputJumpBasic) Name() string { + return "FilterInputJumpBasic" +} + +// ContainerAction implements TestCase.ContainerAction. +func (FilterInputJumpBasic) ContainerAction(ip net.IP) error { + rules := [][]string{ + {"-P", "INPUT", "DROP"}, + {"-N", chainName}, + {"-A", "INPUT", "-j", chainName}, + {"-A", chainName, "-j", "ACCEPT"}, + } + if err := filterTableRules(rules); err != nil { + return err + } + + // Listen for UDP packets on acceptPort. + return listenUDP(acceptPort, sendloopDuration) +} + +// LocalAction implements TestCase.LocalAction. +func (FilterInputJumpBasic) LocalAction(ip net.IP) error { + return sendUDPLoop(ip, acceptPort, sendloopDuration) +} + +// Jump, return, and execute a rule. +type FilterInputJumpReturn struct{} + +// Name implements TestCase.Name. +func (FilterInputJumpReturn) Name() string { + return "FilterInputJumpReturn" +} + +// ContainerAction implements TestCase.ContainerAction. +func (FilterInputJumpReturn) ContainerAction(ip net.IP) error { + rules := [][]string{ + {"-N", chainName}, + {"-P", "INPUT", "ACCEPT"}, + {"-A", "INPUT", "-j", chainName}, + {"-A", chainName, "-j", "RETURN"}, + {"-A", chainName, "-j", "DROP"}, + } + if err := filterTableRules(rules); err != nil { + return err + } + + // Listen for UDP packets on acceptPort. + return listenUDP(acceptPort, sendloopDuration) +} + +// LocalAction implements TestCase.LocalAction. +func (FilterInputJumpReturn) LocalAction(ip net.IP) error { + return sendUDPLoop(ip, acceptPort, sendloopDuration) +} + +type FilterInputJumpReturnDrop struct{} + +// Name implements TestCase.Name. +func (FilterInputJumpReturnDrop) Name() string { + return "FilterInputJumpReturnDrop" +} + +// ContainerAction implements TestCase.ContainerAction. +func (FilterInputJumpReturnDrop) ContainerAction(ip net.IP) error { + rules := [][]string{ + {"-N", chainName}, + {"-A", "INPUT", "-j", chainName}, + {"-A", "INPUT", "-j", "DROP"}, + {"-A", chainName, "-j", "RETURN"}, + } + if err := filterTableRules(rules); err != nil { + return err + } + + // Listen for UDP packets on dropPort. + if err := listenUDP(dropPort, sendloopDuration); err == nil { + return fmt.Errorf("packets on port %d should have been dropped, but got a packet", dropPort) + } else if netErr, ok := err.(net.Error); !ok || !netErr.Timeout() { + return fmt.Errorf("error reading: %v", err) + } + + // At this point we know that reading timed out and never received a + // packet. + return nil +} + +// LocalAction implements TestCase.LocalAction. +func (FilterInputJumpReturnDrop) LocalAction(ip net.IP) error { + return sendUDPLoop(ip, dropPort, sendloopDuration) +} + +// Jumping to a top-levl chain is illegal. +type FilterInputJumpBuiltin struct{} + +// Name implements TestCase.Name. +func (FilterInputJumpBuiltin) Name() string { + return "FilterInputJumpBuiltin" +} + +// ContainerAction implements TestCase.ContainerAction. +func (FilterInputJumpBuiltin) ContainerAction(ip net.IP) error { + if err := filterTable("-A", "INPUT", "-j", "OUTPUT"); err == nil { + return fmt.Errorf("iptables should be unable to jump to a built-in chain") + } + return nil +} + +// LocalAction implements TestCase.LocalAction. +func (FilterInputJumpBuiltin) LocalAction(ip net.IP) error { + // No-op. + return nil +} + +// Jump twice, then return twice and execute a rule. +type FilterInputJumpTwice struct{} + +// Name implements TestCase.Name. +func (FilterInputJumpTwice) Name() string { + return "FilterInputJumpTwice" +} + +// ContainerAction implements TestCase.ContainerAction. +func (FilterInputJumpTwice) ContainerAction(ip net.IP) error { + const chainName2 = chainName + "2" + rules := [][]string{ + {"-P", "INPUT", "DROP"}, + {"-N", chainName}, + {"-N", chainName2}, + {"-A", "INPUT", "-j", chainName}, + {"-A", chainName, "-j", chainName2}, + {"-A", "INPUT", "-j", "ACCEPT"}, + } + if err := filterTableRules(rules); err != nil { + return err + } + + // UDP packets should jump and return twice, eventually hitting the + // ACCEPT rule. + return listenUDP(acceptPort, sendloopDuration) +} + +// LocalAction implements TestCase.LocalAction. +func (FilterInputJumpTwice) LocalAction(ip net.IP) error { + return sendUDPLoop(ip, acceptPort, sendloopDuration) +} diff --git a/test/iptables/iptables_test.go b/test/iptables/iptables_test.go index 46a7c99b0..0621861eb 100644 --- a/test/iptables/iptables_test.go +++ b/test/iptables/iptables_test.go @@ -249,3 +249,39 @@ func TestFilterOutputDropTCPSrcPort(t *testing.T) { t.Fatal(err) } } + +func TestJumpSerialize(t *testing.T) { + if err := singleTest(FilterInputSerializeJump{}); err != nil { + t.Fatal(err) + } +} + +func TestJumpBasic(t *testing.T) { + if err := singleTest(FilterInputJumpBasic{}); err != nil { + t.Fatal(err) + } +} + +func TestJumpReturn(t *testing.T) { + if err := singleTest(FilterInputJumpReturn{}); err != nil { + t.Fatal(err) + } +} + +func TestJumpReturnDrop(t *testing.T) { + if err := singleTest(FilterInputJumpReturnDrop{}); err != nil { + t.Fatal(err) + } +} + +func TestJumpBuiltin(t *testing.T) { + if err := singleTest(FilterInputJumpBuiltin{}); err != nil { + t.Fatal(err) + } +} + +func TestJumpTwice(t *testing.T) { + if err := singleTest(FilterInputJumpTwice{}); err != nil { + t.Fatal(err) + } +} diff --git a/test/iptables/iptables_util.go b/test/iptables/iptables_util.go index 043114c78..293c4e6ed 100644 --- a/test/iptables/iptables_util.go +++ b/test/iptables/iptables_util.go @@ -35,6 +35,16 @@ func filterTable(args ...string) error { return nil } +// filterTableRules is like filterTable, but runs multiple iptables commands. +func filterTableRules(argsList [][]string) error { + for _, args := range argsList { + if err := filterTable(args...); err != nil { + return err + } + } + return nil +} + // listenUDP listens on a UDP port and returns the value of net.Conn.Read() for // the first read on that port. func listenUDP(port int, timeout time.Duration) error { -- cgit v1.2.3 From 56fd9504aab44a738d3df164cbee8e572b309f28 Mon Sep 17 00:00:00 2001 From: gVisor bot Date: Tue, 18 Feb 2020 15:44:22 -0800 Subject: Enable IPV6_RECVTCLASS socket option for datagram sockets Added the ability to get/set the IP_RECVTCLASS socket option on UDP endpoints. If enabled, traffic class from the incoming Network Header passed as ancillary data in the ControlMessages. Adding Get/SetSockOptBool to decrease the overhead of getting/setting simple options. (This was absorbed in a CL that will be landing before this one). Test: * Added unit test to udp_test.go that tests getting/setting as well as verifying that we receive expected TOS from incoming packet. * Added a syscall test for verifying getting/setting * Removed test skip for existing syscall test to enable end to end test. PiperOrigin-RevId: 295840218 --- pkg/sentry/socket/control/control.go | 2 +- pkg/sentry/socket/netstack/netstack.go | 27 +++++- pkg/tcpip/checker/checker.go | 14 +++ pkg/tcpip/tcpip.go | 15 ++- pkg/tcpip/transport/udp/endpoint.go | 38 +++++++- pkg/tcpip/transport/udp/udp_test.go | 120 ++++++++++++++---------- test/syscalls/linux/ip_socket_test_util.h | 16 ++-- test/syscalls/linux/socket_ip_udp_generic.cc | 133 +++++++++++++++++++-------- test/syscalls/linux/udp_socket_test_cases.cc | 4 - 9 files changed, 260 insertions(+), 109 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/control/control.go b/pkg/sentry/socket/control/control.go index 4667373d2..8834a1e1a 100644 --- a/pkg/sentry/socket/control/control.go +++ b/pkg/sentry/socket/control/control.go @@ -329,7 +329,7 @@ func PackTOS(t *kernel.Task, tos uint8, buf []byte) []byte { } // PackTClass packs an IPV6_TCLASS socket control message. -func PackTClass(t *kernel.Task, tClass int32, buf []byte) []byte { +func PackTClass(t *kernel.Task, tClass uint32, buf []byte) []byte { return putCmsgStruct( buf, linux.SOL_IPV6, diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index 9757fbfba..e187276c5 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -1318,6 +1318,22 @@ func getSockOptIPv6(t *kernel.Task, ep commonEndpoint, name, outLen int) (interf } return ib, nil + case linux.IPV6_RECVTCLASS: + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + + v, err := ep.GetSockOptBool(tcpip.ReceiveTClassOption) + if err != nil { + return nil, syserr.TranslateNetstackError(err) + } + + var o int32 + if v { + o = 1 + } + return o, nil + default: emitUnimplementedEventIPv6(t, name) } @@ -1803,6 +1819,14 @@ func setSockOptIPv6(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) } return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.IPv6TrafficClassOption(v))) + case linux.IPV6_RECVTCLASS: + v, err := parseIntOrChar(optVal) + if err != nil { + return err + } + + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.ReceiveTClassOption, v != 0)) + default: emitUnimplementedEventIPv6(t, name) } @@ -2086,7 +2110,6 @@ func emitUnimplementedEventIPv6(t *kernel.Task, name int) { linux.IPV6_RECVPATHMTU, linux.IPV6_RECVPKTINFO, linux.IPV6_RECVRTHDR, - linux.IPV6_RECVTCLASS, linux.IPV6_RTHDR, linux.IPV6_RTHDRDSTOPTS, linux.IPV6_TCLASS, @@ -2424,6 +2447,8 @@ func (s *SocketOperations) controlMessages() socket.ControlMessages { Timestamp: s.readCM.Timestamp, HasTOS: s.readCM.HasTOS, TOS: s.readCM.TOS, + HasTClass: s.readCM.HasTClass, + TClass: s.readCM.TClass, HasIPPacketInfo: s.readCM.HasIPPacketInfo, PacketInfo: s.readCM.PacketInfo, }, diff --git a/pkg/tcpip/checker/checker.go b/pkg/tcpip/checker/checker.go index 4d6ae0871..c6c160dfc 100644 --- a/pkg/tcpip/checker/checker.go +++ b/pkg/tcpip/checker/checker.go @@ -161,6 +161,20 @@ func FragmentFlags(flags uint8) NetworkChecker { } } +// ReceiveTClass creates a checker that checks the TCLASS field in +// ControlMessages. +func ReceiveTClass(want uint32) ControlMessagesChecker { + return func(t *testing.T, cm tcpip.ControlMessages) { + t.Helper() + if !cm.HasTClass { + t.Fatalf("got cm.HasTClass = %t, want cm.TClass = %d", cm.HasTClass, want) + } + if got := cm.TClass; got != want { + t.Fatalf("got cm.TClass = %d, want %d", got, want) + } + } +} + // ReceiveTOS creates a checker that checks the TOS field in ControlMessages. func ReceiveTOS(want uint8) ControlMessagesChecker { return func(t *testing.T, cm tcpip.ControlMessages) { diff --git a/pkg/tcpip/tcpip.go b/pkg/tcpip/tcpip.go index 9ca39ce40..ce5527391 100644 --- a/pkg/tcpip/tcpip.go +++ b/pkg/tcpip/tcpip.go @@ -323,11 +323,11 @@ type ControlMessages struct { // TOS is the IPv4 type of service of the associated packet. TOS uint8 - // HasTClass indicates whether Tclass is valid/set. + // HasTClass indicates whether TClass is valid/set. HasTClass bool - // Tclass is the IPv6 traffic class of the associated packet. - TClass int32 + // TClass is the IPv6 traffic class of the associated packet. + TClass uint32 // HasIPPacketInfo indicates whether PacketInfo is set. HasIPPacketInfo bool @@ -502,9 +502,13 @@ type WriteOptions struct { type SockOptBool int const ( + // ReceiveTClassOption is used by SetSockOpt/GetSockOpt to specify if the + // IPV6_TCLASS ancillary message is passed with incoming packets. + ReceiveTClassOption SockOptBool = iota + // ReceiveTOSOption is used by SetSockOpt/GetSockOpt to specify if the TOS // ancillary message is passed with incoming packets. - ReceiveTOSOption SockOptBool = iota + ReceiveTOSOption // V6OnlyOption is used by {G,S}etSockOptBool to specify whether an IPv6 // socket is to be restricted to sending and receiving IPv6 packets only. @@ -514,6 +518,9 @@ const ( // if more inforamtion is provided with incoming packets such // as interface index and address. ReceiveIPPacketInfoOption + + // TODO(b/146901447): convert existing bool socket options to be handled via + // Get/SetSockOptBool ) // SockOptInt represents socket options which values have the int type. diff --git a/pkg/tcpip/transport/udp/endpoint.go b/pkg/tcpip/transport/udp/endpoint.go index 3fe91cac2..eff7f3600 100644 --- a/pkg/tcpip/transport/udp/endpoint.go +++ b/pkg/tcpip/transport/udp/endpoint.go @@ -32,7 +32,8 @@ type udpPacket struct { packetInfo tcpip.IPPacketInfo data buffer.VectorisedView `state:".(buffer.VectorisedView)"` timestamp int64 - tos uint8 + // tos stores either the receiveTOS or receiveTClass value. + tos uint8 } // EndpointState represents the state of a UDP endpoint. @@ -119,6 +120,10 @@ type endpoint struct { // as ancillary data to ControlMessages on Read. receiveTOS bool + // receiveTClass determines if the incoming IPv6 TClass header field is + // passed as ancillary data to ControlMessages on Read. + receiveTClass bool + // receiveIPPacketInfo determines if the packet info is returned by Read. receiveIPPacketInfo bool @@ -258,13 +263,18 @@ func (e *endpoint) Read(addr *tcpip.FullAddress) (buffer.View, tcpip.ControlMess } e.mu.RLock() receiveTOS := e.receiveTOS + receiveTClass := e.receiveTClass receiveIPPacketInfo := e.receiveIPPacketInfo e.mu.RUnlock() if receiveTOS { cm.HasTOS = true cm.TOS = p.tos } - + if receiveTClass { + cm.HasTClass = true + // Although TClass is an 8-bit value it's read in the CMsg as a uint32. + cm.TClass = uint32(p.tos) + } if receiveIPPacketInfo { cm.HasIPPacketInfo = true cm.PacketInfo = p.packetInfo @@ -490,6 +500,17 @@ func (e *endpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error { e.mu.Unlock() return nil + case tcpip.ReceiveTClassOption: + // We only support this option on v6 endpoints. + if e.NetProto != header.IPv6ProtocolNumber { + return tcpip.ErrNotSupported + } + + e.mu.Lock() + e.receiveTClass = v + e.mu.Unlock() + return nil + case tcpip.V6OnlyOption: // We only recognize this option on v6 endpoints. if e.NetProto != header.IPv6ProtocolNumber { @@ -709,6 +730,17 @@ func (e *endpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { e.mu.RUnlock() return v, nil + case tcpip.ReceiveTClassOption: + // We only support this option on v6 endpoints. + if e.NetProto != header.IPv6ProtocolNumber { + return false, tcpip.ErrNotSupported + } + + e.mu.RLock() + v := e.receiveTClass + e.mu.RUnlock() + return v, nil + case tcpip.V6OnlyOption: // We only recognize this option on v6 endpoints. if e.NetProto != header.IPv6ProtocolNumber { @@ -1273,6 +1305,8 @@ func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pk packet.packetInfo.LocalAddr = r.LocalAddress packet.packetInfo.DestinationAddr = r.RemoteAddress packet.packetInfo.NIC = r.NICID() + case header.IPv6ProtocolNumber: + packet.tos, _ = header.IPv6(pkt.NetworkHeader).TOS() } packet.timestamp = e.stack.NowNanoseconds() diff --git a/pkg/tcpip/transport/udp/udp_test.go b/pkg/tcpip/transport/udp/udp_test.go index f0ff3fe71..34b7c2360 100644 --- a/pkg/tcpip/transport/udp/udp_test.go +++ b/pkg/tcpip/transport/udp/udp_test.go @@ -409,6 +409,7 @@ func (c *testContext) injectV6Packet(payload []byte, h *header4Tuple, valid bool // Initialize the IP header. ip := header.IPv6(buf) ip.Encode(&header.IPv6Fields{ + TrafficClass: testTOS, PayloadLength: uint16(header.UDPMinimumSize + len(payload)), NextHeader: uint8(udp.ProtocolNumber), HopLimit: 65, @@ -1336,7 +1337,7 @@ func TestSetTTL(t *testing.T) { } } -func TestTOSV4(t *testing.T) { +func TestSetTOS(t *testing.T) { for _, flow := range []testFlow{unicastV4, multicastV4, broadcast} { t.Run(fmt.Sprintf("flow:%s", flow), func(t *testing.T) { c := newDualTestContext(t, defaultMTU) @@ -1347,23 +1348,23 @@ func TestTOSV4(t *testing.T) { const tos = testTOS var v tcpip.IPv4TOSOption if err := c.ep.GetSockOpt(&v); err != nil { - c.t.Errorf("GetSockopt failed: %s", err) + c.t.Errorf("GetSockopt(%T) failed: %s", v, err) } // Test for expected default value. if v != 0 { - c.t.Errorf("got GetSockOpt(...) = %#v, want = %#v", v, 0) + c.t.Errorf("got GetSockOpt(%T) = 0x%x, want = 0x%x", v, v, 0) } if err := c.ep.SetSockOpt(tcpip.IPv4TOSOption(tos)); err != nil { - c.t.Errorf("SetSockOpt(%#v) failed: %s", tcpip.IPv4TOSOption(tos), err) + c.t.Errorf("SetSockOpt(%T, 0x%x) failed: %s", v, tcpip.IPv4TOSOption(tos), err) } if err := c.ep.GetSockOpt(&v); err != nil { - c.t.Errorf("GetSockopt failed: %s", err) + c.t.Errorf("GetSockopt(%T) failed: %s", v, err) } if want := tcpip.IPv4TOSOption(tos); v != want { - c.t.Errorf("got GetSockOpt(...) = %#v, want = %#v", v, want) + c.t.Errorf("got GetSockOpt(%T) = 0x%x, want = 0x%x", v, v, want) } testWrite(c, flow, checker.TOS(tos, 0)) @@ -1371,7 +1372,7 @@ func TestTOSV4(t *testing.T) { } } -func TestTOSV6(t *testing.T) { +func TestSetTClass(t *testing.T) { for _, flow := range []testFlow{unicastV4in6, unicastV6, unicastV6Only, multicastV4in6, multicastV6, broadcastIn6} { t.Run(fmt.Sprintf("flow:%s", flow), func(t *testing.T) { c := newDualTestContext(t, defaultMTU) @@ -1379,71 +1380,92 @@ func TestTOSV6(t *testing.T) { c.createEndpointForFlow(flow) - const tos = testTOS + const tClass = testTOS var v tcpip.IPv6TrafficClassOption if err := c.ep.GetSockOpt(&v); err != nil { - c.t.Errorf("GetSockopt failed: %s", err) + c.t.Errorf("GetSockopt(%T) failed: %s", v, err) } // Test for expected default value. if v != 0 { - c.t.Errorf("got GetSockOpt(...) = %#v, want = %#v", v, 0) + c.t.Errorf("got GetSockOpt(%T) = 0x%x, want = 0x%x", v, v, 0) } - if err := c.ep.SetSockOpt(tcpip.IPv6TrafficClassOption(tos)); err != nil { - c.t.Errorf("SetSockOpt failed: %s", err) + if err := c.ep.SetSockOpt(tcpip.IPv6TrafficClassOption(tClass)); err != nil { + c.t.Errorf("SetSockOpt(%T, 0x%x) failed: %s", v, tcpip.IPv6TrafficClassOption(tClass), err) } if err := c.ep.GetSockOpt(&v); err != nil { - c.t.Errorf("GetSockopt failed: %s", err) + c.t.Errorf("GetSockopt(%T) failed: %s", v, err) } - if want := tcpip.IPv6TrafficClassOption(tos); v != want { - c.t.Errorf("got GetSockOpt(...) = %#v, want = %#v", v, want) + if want := tcpip.IPv6TrafficClassOption(tClass); v != want { + c.t.Errorf("got GetSockOpt(%T) = 0x%x, want = 0x%x", v, v, want) } - testWrite(c, flow, checker.TOS(tos, 0)) + // The header getter for TClass is called TOS, so use that checker. + testWrite(c, flow, checker.TOS(tClass, 0)) }) } } -func TestReceiveTOSV4(t *testing.T) { - for _, flow := range []testFlow{unicastV4, broadcast} { - t.Run(fmt.Sprintf("flow:%s", flow), func(t *testing.T) { - c := newDualTestContext(t, defaultMTU) - defer c.cleanup() +func TestReceiveTosTClass(t *testing.T) { + testCases := []struct { + name string + getReceiveOption tcpip.SockOptBool + tests []testFlow + }{ + {"ReceiveTosOption", tcpip.ReceiveTOSOption, []testFlow{unicastV4, broadcast}}, + {"ReceiveTClassOption", tcpip.ReceiveTClassOption, []testFlow{unicastV4in6, unicastV6, unicastV6Only, broadcastIn6}}, + } + for _, testCase := range testCases { + for _, flow := range testCase.tests { + t.Run(fmt.Sprintf("%s:flow:%s", testCase.name, flow), func(t *testing.T) { + c := newDualTestContext(t, defaultMTU) + defer c.cleanup() - c.createEndpointForFlow(flow) + c.createEndpointForFlow(flow) + option := testCase.getReceiveOption + name := testCase.name - // Verify that setting and reading the option works. - v, err := c.ep.GetSockOptBool(tcpip.ReceiveTOSOption) - if err != nil { - c.t.Fatal("GetSockOptBool(tcpip.ReceiveTOSOption) failed:", err) - } - // Test for expected default value. - if v != false { - c.t.Errorf("got GetSockOptBool(tcpip.ReceiveTOSOption) = %t, want = %t", v, false) - } + // Verify that setting and reading the option works. + v, err := c.ep.GetSockOptBool(option) + if err != nil { + c.t.Errorf("GetSockoptBool(%s) failed: %s", name, err) + } + // Test for expected default value. + if v != false { + c.t.Errorf("got GetSockOptBool(%s) = %t, want = %t", name, v, false) + } - want := true - if err := c.ep.SetSockOptBool(tcpip.ReceiveTOSOption, want); err != nil { - c.t.Fatalf("SetSockOptBool(tcpip.ReceiveTOSOption, %t) failed: %s", want, err) - } + want := true + if err := c.ep.SetSockOptBool(option, want); err != nil { + c.t.Fatalf("SetSockOptBool(%s, %t) failed: %s", name, want, err) + } - got, err := c.ep.GetSockOptBool(tcpip.ReceiveTOSOption) - if err != nil { - c.t.Fatal("GetSockOptBool(tcpip.ReceiveTOSOption) failed:", err) - } - if got != want { - c.t.Fatalf("got GetSockOptBool(tcpip.ReceiveTOSOption) = %t, want = %t", got, want) - } + got, err := c.ep.GetSockOptBool(option) + if err != nil { + c.t.Errorf("GetSockoptBool(%s) failed: %s", name, err) + } - // Verify that the correct received TOS is handed through as - // ancillary data to the ControlMessages struct. - if err := c.ep.Bind(tcpip.FullAddress{Port: stackPort}); err != nil { - c.t.Fatal("Bind failed:", err) - } - testRead(c, flow, checker.ReceiveTOS(testTOS)) - }) + if got != want { + c.t.Errorf("got GetSockOptBool(%s) = %t, want = %t", name, got, want) + } + + // Verify that the correct received TOS or TClass is handed through as + // ancillary data to the ControlMessages struct. + if err := c.ep.Bind(tcpip.FullAddress{Port: stackPort}); err != nil { + c.t.Fatalf("Bind failed: %s", err) + } + switch option { + case tcpip.ReceiveTClassOption: + testRead(c, flow, checker.ReceiveTClass(testTOS)) + case tcpip.ReceiveTOSOption: + testRead(c, flow, checker.ReceiveTOS(testTOS)) + default: + t.Fatalf("unknown test variant: %s", name) + } + }) + } } } diff --git a/test/syscalls/linux/ip_socket_test_util.h b/test/syscalls/linux/ip_socket_test_util.h index 083ebbcf0..39fd6709d 100644 --- a/test/syscalls/linux/ip_socket_test_util.h +++ b/test/syscalls/linux/ip_socket_test_util.h @@ -84,20 +84,20 @@ SocketPairKind DualStackUDPBidirectionalBindSocketPair(int type); // SocketPairs created with AF_INET and the given type. SocketPairKind IPv4UDPUnboundSocketPair(int type); -// IPv4UDPUnboundSocketPair returns a SocketKind that represents -// a SimpleSocket created with AF_INET, SOCK_DGRAM, and the given type. +// IPv4UDPUnboundSocket returns a SocketKind that represents a SimpleSocket +// created with AF_INET, SOCK_DGRAM, and the given type. SocketKind IPv4UDPUnboundSocket(int type); -// IPv6UDPUnboundSocketPair returns a SocketKind that represents -// a SimpleSocket created with AF_INET6, SOCK_DGRAM, and the given type. +// IPv6UDPUnboundSocket returns a SocketKind that represents a SimpleSocket +// created with AF_INET6, SOCK_DGRAM, and the given type. SocketKind IPv6UDPUnboundSocket(int type); -// IPv4TCPUnboundSocketPair returns a SocketKind that represents -// a SimpleSocket created with AF_INET, SOCK_STREAM and the given type. +// IPv4TCPUnboundSocket returns a SocketKind that represents a SimpleSocket +// created with AF_INET, SOCK_STREAM and the given type. SocketKind IPv4TCPUnboundSocket(int type); -// IPv6TCPUnboundSocketPair returns a SocketKind that represents -// a SimpleSocket created with AF_INET6, SOCK_STREAM and the given type. +// IPv6TCPUnboundSocket returns a SocketKind that represents a SimpleSocket +// created with AF_INET6, SOCK_STREAM and the given type. SocketKind IPv6TCPUnboundSocket(int type); // IfAddrHelper is a helper class that determines the local interfaces present diff --git a/test/syscalls/linux/socket_ip_udp_generic.cc b/test/syscalls/linux/socket_ip_udp_generic.cc index db5663ecd..1c533fdf2 100644 --- a/test/syscalls/linux/socket_ip_udp_generic.cc +++ b/test/syscalls/linux/socket_ip_udp_generic.cc @@ -14,6 +14,7 @@ #include "test/syscalls/linux/socket_ip_udp_generic.h" +#include #include #include #include @@ -209,46 +210,6 @@ TEST_P(UDPSocketPairTest, SetMulticastLoopChar) { EXPECT_EQ(get, kSockOptOn); } -// Ensure that Receiving TOS is off by default. -TEST_P(UDPSocketPairTest, RecvTosDefault) { - auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); - - int get = -1; - socklen_t get_len = sizeof(get); - ASSERT_THAT( - getsockopt(sockets->first_fd(), IPPROTO_IP, IP_RECVTOS, &get, &get_len), - SyscallSucceedsWithValue(0)); - EXPECT_EQ(get_len, sizeof(get)); - EXPECT_EQ(get, kSockOptOff); -} - -// Test that setting and getting IP_RECVTOS works as expected. -TEST_P(UDPSocketPairTest, SetRecvTos) { - auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); - - ASSERT_THAT(setsockopt(sockets->first_fd(), IPPROTO_IP, IP_RECVTOS, - &kSockOptOff, sizeof(kSockOptOff)), - SyscallSucceeds()); - - int get = -1; - socklen_t get_len = sizeof(get); - ASSERT_THAT( - getsockopt(sockets->first_fd(), IPPROTO_IP, IP_RECVTOS, &get, &get_len), - SyscallSucceedsWithValue(0)); - EXPECT_EQ(get_len, sizeof(get)); - EXPECT_EQ(get, kSockOptOff); - - ASSERT_THAT(setsockopt(sockets->first_fd(), IPPROTO_IP, IP_RECVTOS, - &kSockOptOn, sizeof(kSockOptOn)), - SyscallSucceeds()); - - ASSERT_THAT( - getsockopt(sockets->first_fd(), IPPROTO_IP, IP_RECVTOS, &get, &get_len), - SyscallSucceedsWithValue(0)); - EXPECT_EQ(get_len, sizeof(get)); - EXPECT_EQ(get, kSockOptOn); -} - TEST_P(UDPSocketPairTest, ReuseAddrDefault) { auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); @@ -401,5 +362,97 @@ TEST_P(UDPSocketPairTest, SetAndGetIPPKTINFO) { EXPECT_EQ(get_len, sizeof(get)); } +// Holds TOS or TClass information for IPv4 or IPv6 respectively. +struct RecvTosOption { + int level; + int option; +}; + +RecvTosOption GetRecvTosOption(int domain) { + TEST_CHECK(domain == AF_INET || domain == AF_INET6); + RecvTosOption opt; + switch (domain) { + case AF_INET: + opt.level = IPPROTO_IP; + opt.option = IP_RECVTOS; + break; + case AF_INET6: + opt.level = IPPROTO_IPV6; + opt.option = IPV6_RECVTCLASS; + break; + } + return opt; +} + +// Ensure that Receiving TOS or TCLASS is off by default. +TEST_P(UDPSocketPairTest, RecvTosDefault) { + auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); + RecvTosOption t = GetRecvTosOption(GetParam().domain); + int get = -1; + socklen_t get_len = sizeof(get); + ASSERT_THAT( + getsockopt(sockets->first_fd(), t.level, t.option, &get, &get_len), + SyscallSucceedsWithValue(0)); + EXPECT_EQ(get_len, sizeof(get)); + EXPECT_EQ(get, kSockOptOff); +} + +// Test that setting and getting IP_RECVTOS or IPV6_RECVTCLASS works as +// expected. +TEST_P(UDPSocketPairTest, SetRecvTos) { + auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); + RecvTosOption t = GetRecvTosOption(GetParam().domain); + + ASSERT_THAT(setsockopt(sockets->first_fd(), t.level, t.option, &kSockOptOff, + sizeof(kSockOptOff)), + SyscallSucceeds()); + + int get = -1; + socklen_t get_len = sizeof(get); + ASSERT_THAT( + getsockopt(sockets->first_fd(), t.level, t.option, &get, &get_len), + SyscallSucceedsWithValue(0)); + EXPECT_EQ(get_len, sizeof(get)); + EXPECT_EQ(get, kSockOptOff); + + ASSERT_THAT(setsockopt(sockets->first_fd(), t.level, t.option, &kSockOptOn, + sizeof(kSockOptOn)), + SyscallSucceeds()); + + ASSERT_THAT( + getsockopt(sockets->first_fd(), t.level, t.option, &get, &get_len), + SyscallSucceedsWithValue(0)); + EXPECT_EQ(get_len, sizeof(get)); + EXPECT_EQ(get, kSockOptOn); +} + +// Test that any socket (including IPv6 only) accepts the IPv4 TOS option: this +// mirrors behavior in linux. +TEST_P(UDPSocketPairTest, TOSRecvMismatch) { + auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); + RecvTosOption t = GetRecvTosOption(AF_INET); + int get = -1; + socklen_t get_len = sizeof(get); + + ASSERT_THAT( + getsockopt(sockets->first_fd(), t.level, t.option, &get, &get_len), + SyscallSucceedsWithValue(0)); +} + +// Test that an IPv4 socket does not support the IPv6 TClass option. +TEST_P(UDPSocketPairTest, TClassRecvMismatch) { + // This should only test AF_INET sockets for the mismatch behavior. + SKIP_IF(GetParam().domain != AF_INET); + + auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); + + int get = -1; + socklen_t get_len = sizeof(get); + + ASSERT_THAT(getsockopt(sockets->first_fd(), IPPROTO_IPV6, IPV6_RECVTCLASS, + &get, &get_len), + SyscallFailsWithErrno(EOPNOTSUPP)); +} + } // namespace testing } // namespace gvisor diff --git a/test/syscalls/linux/udp_socket_test_cases.cc b/test/syscalls/linux/udp_socket_test_cases.cc index 9f8de6b48..57b1a357c 100644 --- a/test/syscalls/linux/udp_socket_test_cases.cc +++ b/test/syscalls/linux/udp_socket_test_cases.cc @@ -1349,9 +1349,6 @@ TEST_P(UdpSocketTest, TimestampIoctlPersistence) { // outgoing packets, and that a receiving socket with IP_RECVTOS or // IPV6_RECVTCLASS will create the corresponding control message. TEST_P(UdpSocketTest, SetAndReceiveTOS) { - // TODO(b/144868438): IPV6_RECVTCLASS not supported for netstack. - SKIP_IF((GetParam() != AddressFamily::kIpv4) && IsRunningOnGvisor() && - !IsRunningWithHostinet()); ASSERT_THAT(bind(s_, addr_[0], addrlen_), SyscallSucceeds()); ASSERT_THAT(connect(t_, addr_[0], addrlen_), SyscallSucceeds()); @@ -1422,7 +1419,6 @@ TEST_P(UdpSocketTest, SetAndReceiveTOS) { // TOS byte on outgoing packets, and that a receiving socket with IP_RECVTOS or // IPV6_RECVTCLASS will create the corresponding control message. TEST_P(UdpSocketTest, SendAndReceiveTOS) { - // TODO(b/144868438): IPV6_RECVTCLASS not supported for netstack. // TODO(b/146661005): Setting TOS via cmsg not supported for netstack. SKIP_IF(IsRunningOnGvisor() && !IsRunningWithHostinet()); ASSERT_THAT(bind(s_, addr_[0], addrlen_), SyscallSucceeds()); -- cgit v1.2.3 From 2daa21e4d73f2297a8bca32c76100333e9ac4af4 Mon Sep 17 00:00:00 2001 From: gVisor bot Date: Wed, 19 Feb 2020 16:47:58 -0800 Subject: Internal change. PiperOrigin-RevId: 296088213 --- pkg/sentry/socket/netstack/provider.go | 2 ++ 1 file changed, 2 insertions(+) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/netstack/provider.go b/pkg/sentry/socket/netstack/provider.go index 5afff2564..5f181f017 100644 --- a/pkg/sentry/socket/netstack/provider.go +++ b/pkg/sentry/socket/netstack/provider.go @@ -75,6 +75,8 @@ func getTransportProtocol(ctx context.Context, stype linux.SockType, protocol in switch protocol { case syscall.IPPROTO_ICMP: return header.ICMPv4ProtocolNumber, true, nil + case syscall.IPPROTO_ICMPV6: + return header.ICMPv6ProtocolNumber, true, nil case syscall.IPPROTO_UDP: return header.UDPProtocolNumber, true, nil case syscall.IPPROTO_TCP: -- cgit v1.2.3 From abf7ebcd38e8c2750f4542f29115140bb2b44a9b Mon Sep 17 00:00:00 2001 From: Nayana Bidari Date: Thu, 27 Feb 2020 10:59:32 -0800 Subject: Internal change. PiperOrigin-RevId: 297638665 --- pkg/sentry/socket/netstack/netstack.go | 40 +++++++++-- pkg/tcpip/transport/packet/endpoint.go | 21 +++++- test/syscalls/linux/packet_socket.cc | 124 ++++++++++++++++++++++++++++++--- 3 files changed, 167 insertions(+), 18 deletions(-) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index e187276c5..48c268bfa 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -712,14 +712,40 @@ func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo // Bind implements the linux syscall bind(2) for sockets backed by // tcpip.Endpoint. func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { - addr, family, err := AddressAndFamily(sockaddr) - if err != nil { - return err - } - if err := s.checkFamily(family, true /* exact */); err != nil { - return err + family := usermem.ByteOrder.Uint16(sockaddr) + var addr tcpip.FullAddress + + // Bind for AF_PACKET requires only family, protocol and ifindex. + // In function AddressAndFamily, we check the address length which is + // not needed for AF_PACKET bind. + if family == linux.AF_PACKET { + var a linux.SockAddrLink + if len(sockaddr) < sockAddrLinkSize { + return syserr.ErrInvalidArgument + } + binary.Unmarshal(sockaddr[:sockAddrLinkSize], usermem.ByteOrder, &a) + + if a.Protocol != uint16(s.protocol) { + return syserr.ErrInvalidArgument + } + + addr = tcpip.FullAddress{ + NIC: tcpip.NICID(a.InterfaceIndex), + Addr: tcpip.Address(a.HardwareAddr[:header.EthernetAddressSize]), + } + } else { + var err *syserr.Error + addr, family, err = AddressAndFamily(sockaddr) + if err != nil { + return err + } + + if err = s.checkFamily(family, true /* exact */); err != nil { + return err + } + + addr = s.mapFamily(addr, family) } - addr = s.mapFamily(addr, family) // Issue the bind request to the endpoint. return syserr.TranslateNetstackError(s.Endpoint.Bind(addr)) diff --git a/pkg/tcpip/transport/packet/endpoint.go b/pkg/tcpip/transport/packet/endpoint.go index 5722815e9..09a1cd436 100644 --- a/pkg/tcpip/transport/packet/endpoint.go +++ b/pkg/tcpip/transport/packet/endpoint.go @@ -76,6 +76,7 @@ type endpoint struct { sndBufSize int closed bool stats tcpip.TransportEndpointStats `state:"nosave"` + bound bool } // NewEndpoint returns a new packet endpoint. @@ -125,6 +126,7 @@ func (ep *endpoint) Close() { } ep.closed = true + ep.bound = false ep.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) } @@ -216,7 +218,24 @@ func (ep *endpoint) Bind(addr tcpip.FullAddress) *tcpip.Error { // sll_family (should be AF_PACKET), sll_protocol, and sll_ifindex." // - packet(7). - return tcpip.ErrNotSupported + ep.mu.Lock() + defer ep.mu.Unlock() + + if ep.bound { + return tcpip.ErrAlreadyBound + } + + // Unregister endpoint with all the nics. + ep.stack.UnregisterPacketEndpoint(0, ep.netProto, ep) + + // Bind endpoint to receive packets from specific interface. + if err := ep.stack.RegisterPacketEndpoint(addr.NIC, ep.netProto, ep); err != nil { + return err + } + + ep.bound = true + + return nil } // GetLocalAddress implements tcpip.Endpoint.GetLocalAddress. diff --git a/test/syscalls/linux/packet_socket.cc b/test/syscalls/linux/packet_socket.cc index 92ae55eec..bc22de788 100644 --- a/test/syscalls/linux/packet_socket.cc +++ b/test/syscalls/linux/packet_socket.cc @@ -13,6 +13,7 @@ // limitations under the License. #include +#include #include #include #include @@ -163,16 +164,11 @@ int CookedPacketTest::GetLoopbackIndex() { return ifr.ifr_ifindex; } -// Receive via a packet socket. -TEST_P(CookedPacketTest, Receive) { - // Let's use a simple IP payload: a UDP datagram. - FileDescriptor udp_sock = - ASSERT_NO_ERRNO_AND_VALUE(Socket(AF_INET, SOCK_DGRAM, 0)); - SendUDPMessage(udp_sock.get()); - +// Receive and verify the message via packet socket on interface. +void ReceiveMessage(int sock, int ifindex) { // Wait for the socket to become readable. struct pollfd pfd = {}; - pfd.fd = socket_; + pfd.fd = sock; pfd.events = POLLIN; EXPECT_THAT(RetryEINTR(poll)(&pfd, 1, 2000), SyscallSucceedsWithValue(1)); @@ -182,9 +178,10 @@ TEST_P(CookedPacketTest, Receive) { char buf[64]; struct sockaddr_ll src = {}; socklen_t src_len = sizeof(src); - ASSERT_THAT(recvfrom(socket_, buf, sizeof(buf), 0, + ASSERT_THAT(recvfrom(sock, buf, sizeof(buf), 0, reinterpret_cast(&src), &src_len), SyscallSucceedsWithValue(packet_size)); + // sockaddr_ll ends with an 8 byte physical address field, but ethernet // addresses only use 6 bytes. Linux used to return sizeof(sockaddr_ll)-2 // here, but since commit b2cf86e1563e33a14a1c69b3e508d15dc12f804c returns @@ -194,7 +191,7 @@ TEST_P(CookedPacketTest, Receive) { // TODO(b/129292371): Verify protocol once we return it. // Verify the source address. EXPECT_EQ(src.sll_family, AF_PACKET); - EXPECT_EQ(src.sll_ifindex, GetLoopbackIndex()); + EXPECT_EQ(src.sll_ifindex, ifindex); EXPECT_EQ(src.sll_halen, ETH_ALEN); // This came from the loopback device, so the address is all 0s. for (int i = 0; i < src.sll_halen; i++) { @@ -222,6 +219,18 @@ TEST_P(CookedPacketTest, Receive) { EXPECT_EQ(strncmp(payload, kMessage, sizeof(kMessage)), 0); } +// Receive via a packet socket. +TEST_P(CookedPacketTest, Receive) { + // Let's use a simple IP payload: a UDP datagram. + FileDescriptor udp_sock = + ASSERT_NO_ERRNO_AND_VALUE(Socket(AF_INET, SOCK_DGRAM, 0)); + SendUDPMessage(udp_sock.get()); + + // Receive and verify the data. + int loopback_index = GetLoopbackIndex(); + ReceiveMessage(socket_, loopback_index); +} + // Send via a packet socket. TEST_P(CookedPacketTest, Send) { // TODO(b/129292371): Remove once we support packet socket writing. @@ -313,6 +322,101 @@ TEST_P(CookedPacketTest, Send) { EXPECT_EQ(src.sin_addr.s_addr, htonl(INADDR_LOOPBACK)); } +// Bind and receive via packet socket. +TEST_P(CookedPacketTest, BindReceive) { + struct sockaddr_ll bind_addr = {}; + bind_addr.sll_family = AF_PACKET; + bind_addr.sll_protocol = htons(GetParam()); + bind_addr.sll_ifindex = GetLoopbackIndex(); + + ASSERT_THAT(bind(socket_, reinterpret_cast(&bind_addr), + sizeof(bind_addr)), + SyscallSucceeds()); + + // Let's use a simple IP payload: a UDP datagram. + FileDescriptor udp_sock = + ASSERT_NO_ERRNO_AND_VALUE(Socket(AF_INET, SOCK_DGRAM, 0)); + SendUDPMessage(udp_sock.get()); + + // Receive and verify the data. + ReceiveMessage(socket_, bind_addr.sll_ifindex); +} + +// Double Bind socket. +TEST_P(CookedPacketTest, DoubleBind) { + struct sockaddr_ll bind_addr = {}; + bind_addr.sll_family = AF_PACKET; + bind_addr.sll_protocol = htons(GetParam()); + bind_addr.sll_ifindex = GetLoopbackIndex(); + + ASSERT_THAT(bind(socket_, reinterpret_cast(&bind_addr), + sizeof(bind_addr)), + SyscallSucceeds()); + + // Binding socket again should fail. + ASSERT_THAT( + bind(socket_, reinterpret_cast(&bind_addr), + sizeof(bind_addr)), + // Linux 4.09 returns EINVAL here, but some time before 4.19 it switched + // to EADDRINUSE. + AnyOf(SyscallFailsWithErrno(EADDRINUSE), SyscallFailsWithErrno(EINVAL))); +} + +// Bind and verify we do not receive data on interface which is not bound +TEST_P(CookedPacketTest, BindDrop) { + // Let's use a simple IP payload: a UDP datagram. + FileDescriptor udp_sock = + ASSERT_NO_ERRNO_AND_VALUE(Socket(AF_INET, SOCK_DGRAM, 0)); + + struct ifaddrs* if_addr_list = nullptr; + auto cleanup = Cleanup([&if_addr_list]() { freeifaddrs(if_addr_list); }); + + ASSERT_THAT(getifaddrs(&if_addr_list), SyscallSucceeds()); + + // Get interface other than loopback. + struct ifreq ifr = {}; + for (struct ifaddrs* i = if_addr_list; i; i = i->ifa_next) { + if (strcmp(i->ifa_name, "lo") != 0) { + strncpy(ifr.ifr_name, i->ifa_name, sizeof(ifr.ifr_name)); + break; + } + } + + // Skip if no interface is available other than loopback. + if (strlen(ifr.ifr_name) == 0) { + GTEST_SKIP(); + } + + // Get interface index. + EXPECT_THAT(ioctl(socket_, SIOCGIFINDEX, &ifr), SyscallSucceeds()); + EXPECT_NE(ifr.ifr_ifindex, 0); + + // Bind to packet socket requires only family, protocol and ifindex. + struct sockaddr_ll bind_addr = {}; + bind_addr.sll_family = AF_PACKET; + bind_addr.sll_protocol = htons(GetParam()); + bind_addr.sll_ifindex = ifr.ifr_ifindex; + + ASSERT_THAT(bind(socket_, reinterpret_cast(&bind_addr), + sizeof(bind_addr)), + SyscallSucceeds()); + + // Send to loopback interface. + struct sockaddr_in dest = {}; + dest.sin_addr.s_addr = htonl(INADDR_LOOPBACK); + dest.sin_family = AF_INET; + dest.sin_port = kPort; + EXPECT_THAT(sendto(udp_sock.get(), kMessage, sizeof(kMessage), 0, + reinterpret_cast(&dest), sizeof(dest)), + SyscallSucceedsWithValue(sizeof(kMessage))); + + // Wait and make sure the socket never receives any data. + struct pollfd pfd = {}; + pfd.fd = socket_; + pfd.events = POLLIN; + EXPECT_THAT(RetryEINTR(poll)(&pfd, 1, 1000), SyscallSucceedsWithValue(0)); +} + INSTANTIATE_TEST_SUITE_P(AllInetTests, CookedPacketTest, ::testing::Values(ETH_P_IP, ETH_P_ALL)); -- cgit v1.2.3 From 42fb7d349137bd8847e7c3df6493fde3bc8e6e89 Mon Sep 17 00:00:00 2001 From: Andrei Vagin Date: Mon, 2 Mar 2020 10:32:20 -0800 Subject: socket: take readMu to access readView DATA RACE in netstack.(*SocketOperations).fetchReadView Write at 0x00c001dca138 by goroutine 1001: gvisor.dev/gvisor/pkg/sentry/socket/netstack.(*SocketOperations).fetchReadView() pkg/sentry/socket/netstack/netstack.go:418 +0x85 gvisor.dev/gvisor/pkg/sentry/socket/netstack.(*SocketOperations).coalescingRead() pkg/sentry/socket/netstack/netstack.go:2309 +0x67 gvisor.dev/gvisor/pkg/sentry/socket/netstack.(*SocketOperations).nonBlockingRead() pkg/sentry/socket/netstack/netstack.go:2378 +0x183d Previous read at 0x00c001dca138 by goroutine 1111: gvisor.dev/gvisor/pkg/sentry/socket/netstack.(*SocketOperations).Ioctl() pkg/sentry/socket/netstack/netstack.go:2666 +0x533 gvisor.dev/gvisor/pkg/sentry/syscalls/linux.Ioctl() Reported-by: syzbot+d4c3885fcc346f08deb6@syzkaller.appspotmail.com PiperOrigin-RevId: 298387377 --- pkg/sentry/socket/netstack/netstack.go | 2 ++ 1 file changed, 2 insertions(+) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index 48c268bfa..1eeb37446 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -2663,7 +2663,9 @@ func (s *SocketOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, } // Add bytes removed from the endpoint but not yet sent to the caller. + s.readMu.Lock() v += len(s.readView) + s.readMu.Unlock() if v > math.MaxInt32 { v = math.MaxInt32 -- cgit v1.2.3 From 43abb24657e737dee1108ff0d512b2e1b6d8a3f6 Mon Sep 17 00:00:00 2001 From: Nayana Bidari Date: Mon, 2 Mar 2020 16:30:51 -0800 Subject: Fix panic caused by invalid address for Bind in packet sockets. PiperOrigin-RevId: 298476533 --- pkg/sentry/socket/netstack/netstack.go | 4 ++++ test/syscalls/linux/packet_socket.cc | 13 +++++++++++++ 2 files changed, 17 insertions(+) (limited to 'pkg/sentry/socket') diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index 1eeb37446..13a9a60b4 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -712,6 +712,10 @@ func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo // Bind implements the linux syscall bind(2) for sockets backed by // tcpip.Endpoint. func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { + if len(sockaddr) < 2 { + return syserr.ErrInvalidArgument + } + family := usermem.ByteOrder.Uint16(sockaddr) var addr tcpip.FullAddress diff --git a/test/syscalls/linux/packet_socket.cc b/test/syscalls/linux/packet_socket.cc index bc22de788..248762ca9 100644 --- a/test/syscalls/linux/packet_socket.cc +++ b/test/syscalls/linux/packet_socket.cc @@ -417,6 +417,19 @@ TEST_P(CookedPacketTest, BindDrop) { EXPECT_THAT(RetryEINTR(poll)(&pfd, 1, 1000), SyscallSucceedsWithValue(0)); } +// Bind with invalid address. +TEST_P(CookedPacketTest, BindFail) { + // Null address. + ASSERT_THAT(bind(socket_, nullptr, sizeof(struct sockaddr)), + SyscallFailsWithErrno(EFAULT)); + + // Address of size 1. + uint8_t addr = 0; + ASSERT_THAT( + bind(socket_, reinterpret_cast(&addr), sizeof(addr)), + SyscallFailsWithErrno(EINVAL)); +} + INSTANTIATE_TEST_SUITE_P(AllInetTests, CookedPacketTest, ::testing::Values(ETH_P_IP, ETH_P_ALL)); -- cgit v1.2.3