From a551add5d8a5bf631cd9859c761e579fdb33ec82 Mon Sep 17 00:00:00 2001 From: Kevin Krakauer Date: Mon, 13 Apr 2020 17:37:21 -0700 Subject: Remove View.First() and View.RemoveFirst() These methods let users eaily break the VectorisedView abstraction, and allowed netstack to slip into pseudo-enforcement of the "all headers are in the first View" invariant. Removing them and replacing with PullUp(n) breaks this reliance and will make it easier to add iptables support and rework network buffer management. The new View.PullUp(n) method is low cost in the common case, when when all the headers fit in the first View. --- pkg/tcpip/stack/forwarder_test.go | 13 ++++++++----- pkg/tcpip/stack/iptables.go | 22 ++++++++++++++++++---- pkg/tcpip/stack/iptables_targets.go | 23 +++++++++++++++++------ pkg/tcpip/stack/nic.go | 34 +++++++++++----------------------- pkg/tcpip/stack/packet_buffer.go | 4 +++- pkg/tcpip/stack/stack_test.go | 10 ++++++---- pkg/tcpip/stack/transport_test.go | 5 +++-- 7 files changed, 66 insertions(+), 45 deletions(-) (limited to 'pkg/tcpip/stack') diff --git a/pkg/tcpip/stack/forwarder_test.go b/pkg/tcpip/stack/forwarder_test.go index e9c652042..c7c663498 100644 --- a/pkg/tcpip/stack/forwarder_test.go +++ b/pkg/tcpip/stack/forwarder_test.go @@ -70,7 +70,10 @@ func (f *fwdTestNetworkEndpoint) ID() *NetworkEndpointID { func (f *fwdTestNetworkEndpoint) HandlePacket(r *Route, pkt PacketBuffer) { // Consume the network header. - b := pkt.Data.First() + b, ok := pkt.Data.PullUp(fwdTestNetHeaderLen) + if !ok { + return + } pkt.Data.TrimFront(fwdTestNetHeaderLen) // Dispatch the packet to the transport protocol. @@ -473,7 +476,7 @@ func TestForwardingWithFakeResolverPartialTimeout(t *testing.T) { t.Fatal("packet not forwarded") } - b := p.Pkt.Header.View() + b := p.Pkt.Data.ToView() if b[0] != 3 { t.Fatalf("got b[0] = %d, want = 3", b[0]) } @@ -517,7 +520,7 @@ func TestForwardingWithFakeResolverTwoPackets(t *testing.T) { t.Fatal("packet not forwarded") } - b := p.Pkt.Header.View() + b := p.Pkt.Data.ToView() if b[0] != 3 { t.Fatalf("got b[0] = %d, want = 3", b[0]) } @@ -564,7 +567,7 @@ func TestForwardingWithFakeResolverManyPackets(t *testing.T) { t.Fatal("packet not forwarded") } - b := p.Pkt.Header.View() + b := p.Pkt.Data.ToView() if b[0] != 3 { t.Fatalf("got b[0] = %d, want = 3", b[0]) } @@ -619,7 +622,7 @@ func TestForwardingWithFakeResolverManyResolutions(t *testing.T) { // The first 5 packets (address 3 to 7) should not be forwarded // because their address resolutions are interrupted. - b := p.Pkt.Header.View() + b := p.Pkt.Data.ToView() if b[0] < 8 { t.Fatalf("got b[0] = %d, want b[0] >= 8", b[0]) } diff --git a/pkg/tcpip/stack/iptables.go b/pkg/tcpip/stack/iptables.go index 6c0a4b24d..6b91159d4 100644 --- a/pkg/tcpip/stack/iptables.go +++ b/pkg/tcpip/stack/iptables.go @@ -212,6 +212,11 @@ func (it *IPTables) Check(hook Hook, pkt PacketBuffer) bool { // CheckPackets runs pkts through the rules for hook and returns a map of packets that // should not go forward. // +// Precondition: pkt is a IPv4 packet of at least length header.IPv4MinimumSize. +// +// TODO(gvisor.dev/issue/170): pk.NetworkHeader will always be set as a +// precondition. +// // NOTE: unlike the Check API the returned map contains packets that should be // dropped. func (it *IPTables) CheckPackets(hook Hook, pkts PacketBufferList) (drop map[*PacketBuffer]struct{}) { @@ -226,7 +231,9 @@ func (it *IPTables) CheckPackets(hook Hook, pkts PacketBufferList) (drop map[*Pa return drop } -// Precondition: pkt.NetworkHeader is set. +// Precondition: pkt is a IPv4 packet of at least length header.IPv4MinimumSize. +// TODO(gvisor.dev/issue/170): pk.NetworkHeader will always be set as a +// precondition. func (it *IPTables) checkChain(hook Hook, pkt PacketBuffer, table Table, ruleIdx int) chainVerdict { // Start from ruleIdx and walk the list of rules until a rule gives us // a verdict. @@ -271,14 +278,21 @@ func (it *IPTables) checkChain(hook Hook, pkt PacketBuffer, table Table, ruleIdx return chainDrop } -// Precondition: pk.NetworkHeader is set. +// Precondition: pkt is a IPv4 packet of at least length header.IPv4MinimumSize. +// TODO(gvisor.dev/issue/170): pk.NetworkHeader will always be set as a +// precondition. func (it *IPTables) checkRule(hook Hook, pkt PacketBuffer, table Table, ruleIdx int) (RuleVerdict, int) { rule := table.Rules[ruleIdx] // If pkt.NetworkHeader hasn't been set yet, it will be contained in - // pkt.Data.First(). + // pkt.Data. if pkt.NetworkHeader == nil { - pkt.NetworkHeader = pkt.Data.First() + var ok bool + pkt.NetworkHeader, ok = pkt.Data.PullUp(header.IPv4MinimumSize) + if !ok { + // Precondition has been violated. + panic(fmt.Sprintf("iptables checks require IPv4 headers of at least %d bytes", header.IPv4MinimumSize)) + } } // Check whether the packet matches the IP header filter. diff --git a/pkg/tcpip/stack/iptables_targets.go b/pkg/tcpip/stack/iptables_targets.go index 7b4543caf..8be61f4b1 100644 --- a/pkg/tcpip/stack/iptables_targets.go +++ b/pkg/tcpip/stack/iptables_targets.go @@ -96,9 +96,12 @@ func (rt RedirectTarget) Action(pkt PacketBuffer) (RuleVerdict, int) { newPkt := pkt.Clone() // Set network header. - headerView := newPkt.Data.First() + headerView, ok := newPkt.Data.PullUp(header.IPv4MinimumSize) + if !ok { + return RuleDrop, 0 + } netHeader := header.IPv4(headerView) - newPkt.NetworkHeader = headerView[:header.IPv4MinimumSize] + newPkt.NetworkHeader = headerView hlen := int(netHeader.HeaderLength()) tlen := int(netHeader.TotalLength()) @@ -117,10 +120,14 @@ func (rt RedirectTarget) Action(pkt PacketBuffer) (RuleVerdict, int) { if newPkt.TransportHeader != nil { udpHeader = header.UDP(newPkt.TransportHeader) } else { - if len(pkt.Data.First()) < header.UDPMinimumSize { + if pkt.Data.Size() < header.UDPMinimumSize { + return RuleDrop, 0 + } + hdr, ok := newPkt.Data.PullUp(header.UDPMinimumSize) + if !ok { return RuleDrop, 0 } - udpHeader = header.UDP(newPkt.Data.First()) + udpHeader = header.UDP(hdr) } udpHeader.SetDestinationPort(rt.MinPort) case header.TCPProtocolNumber: @@ -128,10 +135,14 @@ func (rt RedirectTarget) Action(pkt PacketBuffer) (RuleVerdict, int) { if newPkt.TransportHeader != nil { tcpHeader = header.TCP(newPkt.TransportHeader) } else { - if len(pkt.Data.First()) < header.TCPMinimumSize { + if pkt.Data.Size() < header.TCPMinimumSize { return RuleDrop, 0 } - tcpHeader = header.TCP(newPkt.TransportHeader) + hdr, ok := newPkt.Data.PullUp(header.TCPMinimumSize) + if !ok { + return RuleDrop, 0 + } + tcpHeader = header.TCP(hdr) } // TODO(gvisor.dev/issue/170): Need to recompute checksum // and implement nat connection tracking to support TCP. diff --git a/pkg/tcpip/stack/nic.go b/pkg/tcpip/stack/nic.go index 016dbe15e..0c2b1f36a 100644 --- a/pkg/tcpip/stack/nic.go +++ b/pkg/tcpip/stack/nic.go @@ -1203,12 +1203,12 @@ func (n *NIC) DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.Link n.stack.stats.IP.PacketsReceived.Increment() } - if len(pkt.Data.First()) < netProto.MinimumPacketSize() { + netHeader, ok := pkt.Data.PullUp(netProto.MinimumPacketSize()) + if !ok { n.stack.stats.MalformedRcvdPackets.Increment() return } - - src, dst := netProto.ParseAddresses(pkt.Data.First()) + src, dst := netProto.ParseAddresses(netHeader) if n.stack.handleLocal && !n.isLoopback() && n.getRef(protocol, src) != nil { // The source address is one of our own, so we never should have gotten a @@ -1289,22 +1289,8 @@ func (n *NIC) DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.Link func (n *NIC) forwardPacket(r *Route, protocol tcpip.NetworkProtocolNumber, pkt PacketBuffer) { // TODO(b/143425874) Decrease the TTL field in forwarded packets. - - firstData := pkt.Data.First() - pkt.Data.RemoveFirst() - - if linkHeaderLen := int(n.linkEP.MaxHeaderLength()); linkHeaderLen == 0 { - pkt.Header = buffer.NewPrependableFromView(firstData) - } else { - firstDataLen := len(firstData) - - // pkt.Header should have enough capacity to hold n.linkEP's headers. - pkt.Header = buffer.NewPrependable(firstDataLen + linkHeaderLen) - - // TODO(b/151227689): avoid copying the packet when forwarding - if n := copy(pkt.Header.Prepend(firstDataLen), firstData); n != firstDataLen { - panic(fmt.Sprintf("copied %d bytes, expected %d", n, firstDataLen)) - } + if linkHeaderLen := int(n.linkEP.MaxHeaderLength()); linkHeaderLen != 0 { + pkt.Header = buffer.NewPrependable(linkHeaderLen) } if err := n.linkEP.WritePacket(r, nil /* gso */, protocol, pkt); err != nil { @@ -1332,12 +1318,13 @@ func (n *NIC) DeliverTransportPacket(r *Route, protocol tcpip.TransportProtocolN // validly formed. n.stack.demux.deliverRawPacket(r, protocol, pkt) - if len(pkt.Data.First()) < transProto.MinimumPacketSize() { + transHeader, ok := pkt.Data.PullUp(transProto.MinimumPacketSize()) + if !ok { n.stack.stats.MalformedRcvdPackets.Increment() return } - srcPort, dstPort, err := transProto.ParsePorts(pkt.Data.First()) + srcPort, dstPort, err := transProto.ParsePorts(transHeader) if err != nil { n.stack.stats.MalformedRcvdPackets.Increment() return @@ -1375,11 +1362,12 @@ func (n *NIC) DeliverTransportControlPacket(local, remote tcpip.Address, net tcp // ICMPv4 only guarantees that 8 bytes of the transport protocol will // be present in the payload. We know that the ports are within the // first 8 bytes for all known transport protocols. - if len(pkt.Data.First()) < 8 { + transHeader, ok := pkt.Data.PullUp(8) + if !ok { return } - srcPort, dstPort, err := transProto.ParsePorts(pkt.Data.First()) + srcPort, dstPort, err := transProto.ParsePorts(transHeader) if err != nil { return } diff --git a/pkg/tcpip/stack/packet_buffer.go b/pkg/tcpip/stack/packet_buffer.go index dc125f25e..e954a8b7e 100644 --- a/pkg/tcpip/stack/packet_buffer.go +++ b/pkg/tcpip/stack/packet_buffer.go @@ -37,7 +37,9 @@ type PacketBuffer struct { Data buffer.VectorisedView // Header holds the headers of outbound packets. As a packet is passed - // down the stack, each layer adds to Header. + // down the stack, each layer adds to Header. Note that forwarded + // packets don't populate Headers on their way out -- their headers and + // payload are never parsed out and remain in Data. Header buffer.Prependable // These fields are used by both inbound and outbound packets. They diff --git a/pkg/tcpip/stack/stack_test.go b/pkg/tcpip/stack/stack_test.go index c7634ceb1..d45d2cc1f 100644 --- a/pkg/tcpip/stack/stack_test.go +++ b/pkg/tcpip/stack/stack_test.go @@ -95,16 +95,18 @@ func (f *fakeNetworkEndpoint) HandlePacket(r *stack.Route, pkt stack.PacketBuffe f.proto.packetCount[int(f.id.LocalAddress[0])%len(f.proto.packetCount)]++ // Consume the network header. - b := pkt.Data.First() + b, ok := pkt.Data.PullUp(fakeNetHeaderLen) + if !ok { + return + } pkt.Data.TrimFront(fakeNetHeaderLen) // Handle control packets. if b[2] == uint8(fakeControlProtocol) { - nb := pkt.Data.First() - if len(nb) < fakeNetHeaderLen { + nb, ok := pkt.Data.PullUp(fakeNetHeaderLen) + if !ok { return } - pkt.Data.TrimFront(fakeNetHeaderLen) f.dispatcher.DeliverTransportControlPacket(tcpip.Address(nb[1:2]), tcpip.Address(nb[0:1]), fakeNetNumber, tcpip.TransportProtocolNumber(nb[2]), stack.ControlPortUnreachable, 0, pkt) return diff --git a/pkg/tcpip/stack/transport_test.go b/pkg/tcpip/stack/transport_test.go index 3084e6593..a611e44ab 100644 --- a/pkg/tcpip/stack/transport_test.go +++ b/pkg/tcpip/stack/transport_test.go @@ -642,10 +642,11 @@ func TestTransportForwarding(t *testing.T) { t.Fatal("Response packet not forwarded") } - if dst := p.Pkt.Header.View()[0]; dst != 3 { + hdrs := p.Pkt.Data.ToView() + if dst := hdrs[0]; dst != 3 { t.Errorf("Response packet has incorrect destination addresss: got = %d, want = 3", dst) } - if src := p.Pkt.Header.View()[1]; src != 1 { + if src := hdrs[1]; src != 1 { t.Errorf("Response packet has incorrect source addresss: got = %d, want = 3", src) } } -- cgit v1.2.3