diff options
Diffstat (limited to 'pkg/tcpip/link')
25 files changed, 359 insertions, 260 deletions
diff --git a/pkg/tcpip/link/channel/channel.go b/pkg/tcpip/link/channel/channel.go index c40744b8e..18adb2085 100644 --- a/pkg/tcpip/link/channel/channel.go +++ b/pkg/tcpip/link/channel/channel.go @@ -44,14 +44,12 @@ type Endpoint struct { } // New creates a new channel endpoint. -func New(size int, mtu uint32, linkAddr tcpip.LinkAddress) (tcpip.LinkEndpointID, *Endpoint) { - e := &Endpoint{ +func New(size int, mtu uint32, linkAddr tcpip.LinkAddress) *Endpoint { + return &Endpoint{ C: make(chan PacketInfo, size), mtu: mtu, linkAddr: linkAddr, } - - return stack.RegisterLinkEndpoint(e), e } // Drain removes all outbound packets from the channel and counts them. @@ -135,3 +133,6 @@ func (e *Endpoint) WritePacket(_ *stack.Route, gso *stack.GSO, hdr buffer.Prepen return nil } + +// Wait implements stack.LinkEndpoint.Wait. +func (*Endpoint) Wait() {} diff --git a/pkg/tcpip/link/fdbased/BUILD b/pkg/tcpip/link/fdbased/BUILD index d786d8fdf..8fa9e3984 100644 --- a/pkg/tcpip/link/fdbased/BUILD +++ b/pkg/tcpip/link/fdbased/BUILD @@ -1,4 +1,5 @@ -load("//tools/go_stateify:defs.bzl", "go_library", "go_test") +load("//tools/go_stateify:defs.bzl", "go_library") +load("@io_bazel_rules_go//go:def.bzl", "go_test") package(licenses = ["notice"]) @@ -8,8 +9,8 @@ go_library( "endpoint.go", "endpoint_unsafe.go", "mmap.go", - "mmap_amd64.go", - "mmap_amd64_unsafe.go", + "mmap_stub.go", + "mmap_unsafe.go", "packet_dispatchers.go", ], importpath = "gvisor.dev/gvisor/pkg/tcpip/link/fdbased", diff --git a/pkg/tcpip/link/fdbased/endpoint.go b/pkg/tcpip/link/fdbased/endpoint.go index 77f988b9f..7636418b1 100644 --- a/pkg/tcpip/link/fdbased/endpoint.go +++ b/pkg/tcpip/link/fdbased/endpoint.go @@ -41,6 +41,7 @@ package fdbased import ( "fmt" + "sync" "syscall" "golang.org/x/sys/unix" @@ -81,6 +82,19 @@ const ( PacketMMap ) +func (p PacketDispatchMode) String() string { + switch p { + case Readv: + return "Readv" + case RecvMMsg: + return "RecvMMsg" + case PacketMMap: + return "PacketMMap" + default: + return fmt.Sprintf("unknown packet dispatch mode %v", p) + } +} + type endpoint struct { // fds is the set of file descriptors each identifying one inbound/outbound // channel. The endpoint will dispatch from all inbound channels as well as @@ -114,6 +128,9 @@ type endpoint struct { // gsoMaxSize is the maximum GSO packet size. It is zero if GSO is // disabled. gsoMaxSize uint32 + + // wg keeps track of running goroutines. + wg sync.WaitGroup } // Options specify the details about the fd-based endpoint to be created. @@ -164,8 +181,9 @@ type Options struct { // New creates a new fd-based endpoint. // // Makes fd non-blocking, but does not take ownership of fd, which must remain -// open for the lifetime of the returned endpoint. -func New(opts *Options) (tcpip.LinkEndpointID, error) { +// open for the lifetime of the returned endpoint (until after the endpoint has +// stopped being using and Wait returns). +func New(opts *Options) (stack.LinkEndpoint, error) { caps := stack.LinkEndpointCapabilities(0) if opts.RXChecksumOffload { caps |= stack.CapabilityRXChecksumOffload @@ -190,7 +208,7 @@ func New(opts *Options) (tcpip.LinkEndpointID, error) { } if len(opts.FDs) == 0 { - return 0, fmt.Errorf("opts.FD is empty, at least one FD must be specified") + return nil, fmt.Errorf("opts.FD is empty, at least one FD must be specified") } e := &endpoint{ @@ -207,12 +225,12 @@ func New(opts *Options) (tcpip.LinkEndpointID, error) { for i := 0; i < len(e.fds); i++ { fd := e.fds[i] if err := syscall.SetNonblock(fd, true); err != nil { - return 0, fmt.Errorf("syscall.SetNonblock(%v) failed: %v", fd, err) + return nil, fmt.Errorf("syscall.SetNonblock(%v) failed: %v", fd, err) } isSocket, err := isSocketFD(fd) if err != nil { - return 0, err + return nil, err } if isSocket { if opts.GSOMaxSize != 0 { @@ -222,12 +240,12 @@ func New(opts *Options) (tcpip.LinkEndpointID, error) { } inboundDispatcher, err := createInboundDispatcher(e, fd, isSocket) if err != nil { - return 0, fmt.Errorf("createInboundDispatcher(...) = %v", err) + return nil, fmt.Errorf("createInboundDispatcher(...) = %v", err) } e.inboundDispatchers = append(e.inboundDispatchers, inboundDispatcher) } - return stack.RegisterLinkEndpoint(e), nil + return e, nil } func createInboundDispatcher(e *endpoint, fd int, isSocket bool) (linkDispatcher, error) { @@ -290,7 +308,11 @@ func (e *endpoint) Attach(dispatcher stack.NetworkDispatcher) { // saved, they stop sending outgoing packets and all incoming packets // are rejected. for i := range e.inboundDispatchers { - go e.dispatchLoop(e.inboundDispatchers[i]) // S/R-SAFE: See above. + e.wg.Add(1) + go func(i int) { // S/R-SAFE: See above. + e.dispatchLoop(e.inboundDispatchers[i]) + e.wg.Done() + }(i) } } @@ -320,6 +342,12 @@ func (e *endpoint) LinkAddress() tcpip.LinkAddress { return e.addr } +// Wait implements stack.LinkEndpoint.Wait. It waits for the endpoint to stop +// reading from its FD. +func (e *endpoint) Wait() { + e.wg.Wait() +} + // virtioNetHdr is declared in linux/virtio_net.h. type virtioNetHdr struct { flags uint8 @@ -435,14 +463,12 @@ func (e *InjectableEndpoint) Inject(protocol tcpip.NetworkProtocolNumber, vv buf } // NewInjectable creates a new fd-based InjectableEndpoint. -func NewInjectable(fd int, mtu uint32, capabilities stack.LinkEndpointCapabilities) (tcpip.LinkEndpointID, *InjectableEndpoint) { +func NewInjectable(fd int, mtu uint32, capabilities stack.LinkEndpointCapabilities) *InjectableEndpoint { syscall.SetNonblock(fd, true) - e := &InjectableEndpoint{endpoint: endpoint{ + return &InjectableEndpoint{endpoint: endpoint{ fds: []int{fd}, mtu: mtu, caps: capabilities, }} - - return stack.RegisterLinkEndpoint(e), e } diff --git a/pkg/tcpip/link/fdbased/endpoint_test.go b/pkg/tcpip/link/fdbased/endpoint_test.go index e305252d6..04406bc9a 100644 --- a/pkg/tcpip/link/fdbased/endpoint_test.go +++ b/pkg/tcpip/link/fdbased/endpoint_test.go @@ -68,11 +68,10 @@ func newContext(t *testing.T, opt *Options) *context { } opt.FDs = []int{fds[1]} - epID, err := New(opt) + ep, err := New(opt) if err != nil { t.Fatalf("Failed to create FD endpoint: %v", err) } - ep := stack.FindLinkEndpoint(epID).(*endpoint) c := &context{ t: t, diff --git a/pkg/tcpip/link/fdbased/mmap.go b/pkg/tcpip/link/fdbased/mmap.go index 2dca173c2..8bfeb97e4 100644 --- a/pkg/tcpip/link/fdbased/mmap.go +++ b/pkg/tcpip/link/fdbased/mmap.go @@ -12,12 +12,183 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build !linux !amd64 +// +build linux,amd64 linux,arm64 package fdbased -// Stubbed out version for non-linux/non-amd64 platforms. +import ( + "encoding/binary" + "syscall" -func newPacketMMapDispatcher(fd int, e *endpoint) (linkDispatcher, error) { - return nil, nil + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/buffer" + "gvisor.dev/gvisor/pkg/tcpip/header" + "gvisor.dev/gvisor/pkg/tcpip/link/rawfile" +) + +const ( + tPacketAlignment = uintptr(16) + tpStatusKernel = 0 + tpStatusUser = 1 + tpStatusCopy = 2 + tpStatusLosing = 4 +) + +// We overallocate the frame size to accommodate space for the +// TPacketHdr+RawSockAddrLinkLayer+MAC header and any padding. +// +// Memory allocated for the ring buffer: tpBlockSize * tpBlockNR = 2 MiB +// +// NOTE: +// Frames need to be aligned at 16 byte boundaries. +// BlockSize needs to be page aligned. +// +// For details see PACKET_MMAP setting constraints in +// https://www.kernel.org/doc/Documentation/networking/packet_mmap.txt +const ( + tpFrameSize = 65536 + 128 + tpBlockSize = tpFrameSize * 32 + tpBlockNR = 1 + tpFrameNR = (tpBlockSize * tpBlockNR) / tpFrameSize +) + +// tPacketAlign aligns the pointer v at a tPacketAlignment boundary. Direct +// translation of the TPACKET_ALIGN macro in <linux/if_packet.h>. +func tPacketAlign(v uintptr) uintptr { + return (v + tPacketAlignment - 1) & uintptr(^(tPacketAlignment - 1)) +} + +// tPacketReq is the tpacket_req structure as described in +// https://www.kernel.org/doc/Documentation/networking/packet_mmap.txt +type tPacketReq struct { + tpBlockSize uint32 + tpBlockNR uint32 + tpFrameSize uint32 + tpFrameNR uint32 +} + +// tPacketHdr is tpacket_hdr structure as described in <linux/if_packet.h> +type tPacketHdr []byte + +const ( + tpStatusOffset = 0 + tpLenOffset = 8 + tpSnapLenOffset = 12 + tpMacOffset = 16 + tpNetOffset = 18 + tpSecOffset = 20 + tpUSecOffset = 24 +) + +func (t tPacketHdr) tpLen() uint32 { + return binary.LittleEndian.Uint32(t[tpLenOffset:]) +} + +func (t tPacketHdr) tpSnapLen() uint32 { + return binary.LittleEndian.Uint32(t[tpSnapLenOffset:]) +} + +func (t tPacketHdr) tpMac() uint16 { + return binary.LittleEndian.Uint16(t[tpMacOffset:]) +} + +func (t tPacketHdr) tpNet() uint16 { + return binary.LittleEndian.Uint16(t[tpNetOffset:]) +} + +func (t tPacketHdr) tpSec() uint32 { + return binary.LittleEndian.Uint32(t[tpSecOffset:]) +} + +func (t tPacketHdr) tpUSec() uint32 { + return binary.LittleEndian.Uint32(t[tpUSecOffset:]) +} + +func (t tPacketHdr) Payload() []byte { + return t[uint32(t.tpMac()) : uint32(t.tpMac())+t.tpSnapLen()] +} + +// packetMMapDispatcher uses PACKET_RX_RING's to read/dispatch inbound packets. +// See: mmap_amd64_unsafe.go for implementation details. +type packetMMapDispatcher struct { + // fd is the file descriptor used to send and receive packets. + fd int + + // e is the endpoint this dispatcher is attached to. + e *endpoint + + // ringBuffer is only used when PacketMMap dispatcher is used and points + // to the start of the mmapped PACKET_RX_RING buffer. + ringBuffer []byte + + // ringOffset is the current offset into the ring buffer where the next + // inbound packet will be placed by the kernel. + ringOffset int +} + +func (d *packetMMapDispatcher) readMMappedPacket() ([]byte, *tcpip.Error) { + hdr := tPacketHdr(d.ringBuffer[d.ringOffset*tpFrameSize:]) + for hdr.tpStatus()&tpStatusUser == 0 { + event := rawfile.PollEvent{ + FD: int32(d.fd), + Events: unix.POLLIN | unix.POLLERR, + } + if _, errno := rawfile.BlockingPoll(&event, 1, nil); errno != 0 { + if errno == syscall.EINTR { + continue + } + return nil, rawfile.TranslateErrno(errno) + } + if hdr.tpStatus()&tpStatusCopy != 0 { + // This frame is truncated so skip it after flipping the + // buffer to the kernel. + hdr.setTPStatus(tpStatusKernel) + d.ringOffset = (d.ringOffset + 1) % tpFrameNR + hdr = (tPacketHdr)(d.ringBuffer[d.ringOffset*tpFrameSize:]) + continue + } + } + + // Copy out the packet from the mmapped frame to a locally owned buffer. + pkt := make([]byte, hdr.tpSnapLen()) + copy(pkt, hdr.Payload()) + // Release packet to kernel. + hdr.setTPStatus(tpStatusKernel) + d.ringOffset = (d.ringOffset + 1) % tpFrameNR + return pkt, nil +} + +// dispatch reads packets from an mmaped ring buffer and dispatches them to the +// network stack. +func (d *packetMMapDispatcher) dispatch() (bool, *tcpip.Error) { + pkt, err := d.readMMappedPacket() + if err != nil { + return false, err + } + var ( + p tcpip.NetworkProtocolNumber + remote, local tcpip.LinkAddress + ) + if d.e.hdrSize > 0 { + eth := header.Ethernet(pkt) + p = eth.Type() + remote = eth.SourceAddress() + local = eth.DestinationAddress() + } else { + // We don't get any indication of what the packet is, so try to guess + // if it's an IPv4 or IPv6 packet. + switch header.IPVersion(pkt) { + case header.IPv4Version: + p = header.IPv4ProtocolNumber + case header.IPv6Version: + p = header.IPv6ProtocolNumber + default: + return true, nil + } + } + + pkt = pkt[d.e.hdrSize:] + d.e.dispatcher.DeliverNetworkPacket(d.e, remote, local, p, buffer.NewVectorisedView(len(pkt), []buffer.View{buffer.View(pkt)})) + return true, nil } diff --git a/pkg/tcpip/link/fdbased/mmap_amd64.go b/pkg/tcpip/link/fdbased/mmap_amd64.go deleted file mode 100644 index 029f86a18..000000000 --- a/pkg/tcpip/link/fdbased/mmap_amd64.go +++ /dev/null @@ -1,194 +0,0 @@ -// 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. - -// +build linux,amd64 - -package fdbased - -import ( - "encoding/binary" - "syscall" - - "golang.org/x/sys/unix" - "gvisor.dev/gvisor/pkg/tcpip" - "gvisor.dev/gvisor/pkg/tcpip/buffer" - "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/link/rawfile" -) - -const ( - tPacketAlignment = uintptr(16) - tpStatusKernel = 0 - tpStatusUser = 1 - tpStatusCopy = 2 - tpStatusLosing = 4 -) - -// We overallocate the frame size to accommodate space for the -// TPacketHdr+RawSockAddrLinkLayer+MAC header and any padding. -// -// Memory allocated for the ring buffer: tpBlockSize * tpBlockNR = 2 MiB -// -// NOTE: -// Frames need to be aligned at 16 byte boundaries. -// BlockSize needs to be page aligned. -// -// For details see PACKET_MMAP setting constraints in -// https://www.kernel.org/doc/Documentation/networking/packet_mmap.txt -const ( - tpFrameSize = 65536 + 128 - tpBlockSize = tpFrameSize * 32 - tpBlockNR = 1 - tpFrameNR = (tpBlockSize * tpBlockNR) / tpFrameSize -) - -// tPacketAlign aligns the pointer v at a tPacketAlignment boundary. Direct -// translation of the TPACKET_ALIGN macro in <linux/if_packet.h>. -func tPacketAlign(v uintptr) uintptr { - return (v + tPacketAlignment - 1) & uintptr(^(tPacketAlignment - 1)) -} - -// tPacketReq is the tpacket_req structure as described in -// https://www.kernel.org/doc/Documentation/networking/packet_mmap.txt -type tPacketReq struct { - tpBlockSize uint32 - tpBlockNR uint32 - tpFrameSize uint32 - tpFrameNR uint32 -} - -// tPacketHdr is tpacket_hdr structure as described in <linux/if_packet.h> -type tPacketHdr []byte - -const ( - tpStatusOffset = 0 - tpLenOffset = 8 - tpSnapLenOffset = 12 - tpMacOffset = 16 - tpNetOffset = 18 - tpSecOffset = 20 - tpUSecOffset = 24 -) - -func (t tPacketHdr) tpLen() uint32 { - return binary.LittleEndian.Uint32(t[tpLenOffset:]) -} - -func (t tPacketHdr) tpSnapLen() uint32 { - return binary.LittleEndian.Uint32(t[tpSnapLenOffset:]) -} - -func (t tPacketHdr) tpMac() uint16 { - return binary.LittleEndian.Uint16(t[tpMacOffset:]) -} - -func (t tPacketHdr) tpNet() uint16 { - return binary.LittleEndian.Uint16(t[tpNetOffset:]) -} - -func (t tPacketHdr) tpSec() uint32 { - return binary.LittleEndian.Uint32(t[tpSecOffset:]) -} - -func (t tPacketHdr) tpUSec() uint32 { - return binary.LittleEndian.Uint32(t[tpUSecOffset:]) -} - -func (t tPacketHdr) Payload() []byte { - return t[uint32(t.tpMac()) : uint32(t.tpMac())+t.tpSnapLen()] -} - -// packetMMapDispatcher uses PACKET_RX_RING's to read/dispatch inbound packets. -// See: mmap_amd64_unsafe.go for implementation details. -type packetMMapDispatcher struct { - // fd is the file descriptor used to send and receive packets. - fd int - - // e is the endpoint this dispatcher is attached to. - e *endpoint - - // ringBuffer is only used when PacketMMap dispatcher is used and points - // to the start of the mmapped PACKET_RX_RING buffer. - ringBuffer []byte - - // ringOffset is the current offset into the ring buffer where the next - // inbound packet will be placed by the kernel. - ringOffset int -} - -func (d *packetMMapDispatcher) readMMappedPacket() ([]byte, *tcpip.Error) { - hdr := tPacketHdr(d.ringBuffer[d.ringOffset*tpFrameSize:]) - for hdr.tpStatus()&tpStatusUser == 0 { - event := rawfile.PollEvent{ - FD: int32(d.fd), - Events: unix.POLLIN | unix.POLLERR, - } - if _, errno := rawfile.BlockingPoll(&event, 1, nil); errno != 0 { - if errno == syscall.EINTR { - continue - } - return nil, rawfile.TranslateErrno(errno) - } - if hdr.tpStatus()&tpStatusCopy != 0 { - // This frame is truncated so skip it after flipping the - // buffer to the kernel. - hdr.setTPStatus(tpStatusKernel) - d.ringOffset = (d.ringOffset + 1) % tpFrameNR - hdr = (tPacketHdr)(d.ringBuffer[d.ringOffset*tpFrameSize:]) - continue - } - } - - // Copy out the packet from the mmapped frame to a locally owned buffer. - pkt := make([]byte, hdr.tpSnapLen()) - copy(pkt, hdr.Payload()) - // Release packet to kernel. - hdr.setTPStatus(tpStatusKernel) - d.ringOffset = (d.ringOffset + 1) % tpFrameNR - return pkt, nil -} - -// dispatch reads packets from an mmaped ring buffer and dispatches them to the -// network stack. -func (d *packetMMapDispatcher) dispatch() (bool, *tcpip.Error) { - pkt, err := d.readMMappedPacket() - if err != nil { - return false, err - } - var ( - p tcpip.NetworkProtocolNumber - remote, local tcpip.LinkAddress - ) - if d.e.hdrSize > 0 { - eth := header.Ethernet(pkt) - p = eth.Type() - remote = eth.SourceAddress() - local = eth.DestinationAddress() - } else { - // We don't get any indication of what the packet is, so try to guess - // if it's an IPv4 or IPv6 packet. - switch header.IPVersion(pkt) { - case header.IPv4Version: - p = header.IPv4ProtocolNumber - case header.IPv6Version: - p = header.IPv6ProtocolNumber - default: - return true, nil - } - } - - pkt = pkt[d.e.hdrSize:] - d.e.dispatcher.DeliverNetworkPacket(d.e, remote, local, p, buffer.NewVectorisedView(len(pkt), []buffer.View{buffer.View(pkt)})) - return true, nil -} diff --git a/pkg/tcpip/link/fdbased/mmap_stub.go b/pkg/tcpip/link/fdbased/mmap_stub.go new file mode 100644 index 000000000..67be52d67 --- /dev/null +++ b/pkg/tcpip/link/fdbased/mmap_stub.go @@ -0,0 +1,23 @@ +// 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. + +// +build !linux !amd64,!arm64 + +package fdbased + +// Stubbed out version for non-linux/non-amd64/non-arm64 platforms. + +func newPacketMMapDispatcher(fd int, e *endpoint) (linkDispatcher, error) { + return nil, nil +} diff --git a/pkg/tcpip/link/fdbased/mmap_amd64_unsafe.go b/pkg/tcpip/link/fdbased/mmap_unsafe.go index 47cb1d1cc..3894185ae 100644 --- a/pkg/tcpip/link/fdbased/mmap_amd64_unsafe.go +++ b/pkg/tcpip/link/fdbased/mmap_unsafe.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build linux,amd64 +// +build linux,amd64 linux,arm64 package fdbased diff --git a/pkg/tcpip/link/loopback/loopback.go b/pkg/tcpip/link/loopback/loopback.go index ab6a53988..b36629d2c 100644 --- a/pkg/tcpip/link/loopback/loopback.go +++ b/pkg/tcpip/link/loopback/loopback.go @@ -32,8 +32,8 @@ type endpoint struct { // New creates a new loopback endpoint. This link-layer endpoint just turns // outbound packets into inbound packets. -func New() tcpip.LinkEndpointID { - return stack.RegisterLinkEndpoint(&endpoint{}) +func New() stack.LinkEndpoint { + return &endpoint{} } // Attach implements stack.LinkEndpoint.Attach. It just saves the stack network- @@ -85,3 +85,6 @@ func (e *endpoint) WritePacket(_ *stack.Route, _ *stack.GSO, hdr buffer.Prependa return nil } + +// Wait implements stack.LinkEndpoint.Wait. +func (*endpoint) Wait() {} diff --git a/pkg/tcpip/link/muxed/BUILD b/pkg/tcpip/link/muxed/BUILD index ea12ef1ac..1bab380b0 100644 --- a/pkg/tcpip/link/muxed/BUILD +++ b/pkg/tcpip/link/muxed/BUILD @@ -1,4 +1,5 @@ -load("//tools/go_stateify:defs.bzl", "go_library", "go_test") +load("//tools/go_stateify:defs.bzl", "go_library") +load("@io_bazel_rules_go//go:def.bzl", "go_test") package(licenses = ["notice"]) diff --git a/pkg/tcpip/link/muxed/injectable.go b/pkg/tcpip/link/muxed/injectable.go index a577a3d52..7c946101d 100644 --- a/pkg/tcpip/link/muxed/injectable.go +++ b/pkg/tcpip/link/muxed/injectable.go @@ -104,10 +104,16 @@ func (m *InjectableEndpoint) WriteRawPacket(dest tcpip.Address, packet []byte) * return endpoint.WriteRawPacket(dest, packet) } +// Wait implements stack.LinkEndpoint.Wait. +func (m *InjectableEndpoint) Wait() { + for _, ep := range m.routes { + ep.Wait() + } +} + // NewInjectableEndpoint creates a new multi-endpoint injectable endpoint. -func NewInjectableEndpoint(routes map[tcpip.Address]stack.InjectableLinkEndpoint) (tcpip.LinkEndpointID, *InjectableEndpoint) { - e := &InjectableEndpoint{ +func NewInjectableEndpoint(routes map[tcpip.Address]stack.InjectableLinkEndpoint) *InjectableEndpoint { + return &InjectableEndpoint{ routes: routes, } - return stack.RegisterLinkEndpoint(e), e } diff --git a/pkg/tcpip/link/muxed/injectable_test.go b/pkg/tcpip/link/muxed/injectable_test.go index 174b9330f..3086fec00 100644 --- a/pkg/tcpip/link/muxed/injectable_test.go +++ b/pkg/tcpip/link/muxed/injectable_test.go @@ -87,8 +87,8 @@ func makeTestInjectableEndpoint(t *testing.T) (*InjectableEndpoint, *os.File, tc if err != nil { t.Fatal("Failed to create socket pair:", err) } - _, underlyingEndpoint := fdbased.NewInjectable(pair[1], 6500, stack.CapabilityNone) + underlyingEndpoint := fdbased.NewInjectable(pair[1], 6500, stack.CapabilityNone) routes := map[tcpip.Address]stack.InjectableLinkEndpoint{dstIP: underlyingEndpoint} - _, endpoint := NewInjectableEndpoint(routes) + endpoint := NewInjectableEndpoint(routes) return endpoint, os.NewFile(uintptr(pair[0]), "test route end"), dstIP } diff --git a/pkg/tcpip/link/rawfile/BUILD b/pkg/tcpip/link/rawfile/BUILD index 6e3a7a9d7..2e8bc772a 100644 --- a/pkg/tcpip/link/rawfile/BUILD +++ b/pkg/tcpip/link/rawfile/BUILD @@ -6,8 +6,9 @@ go_library( name = "rawfile", srcs = [ "blockingpoll_amd64.s", - "blockingpoll_amd64_unsafe.go", - "blockingpoll_unsafe.go", + "blockingpoll_arm64.s", + "blockingpoll_noyield_unsafe.go", + "blockingpoll_yield_unsafe.go", "errors.go", "rawfile_unsafe.go", ], diff --git a/pkg/tcpip/link/rawfile/blockingpoll_arm64.s b/pkg/tcpip/link/rawfile/blockingpoll_arm64.s new file mode 100644 index 000000000..b62888b93 --- /dev/null +++ b/pkg/tcpip/link/rawfile/blockingpoll_arm64.s @@ -0,0 +1,42 @@ +// 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. + +#include "textflag.h" + +// BlockingPoll makes the ppoll() syscall while calling the version of +// entersyscall that relinquishes the P so that other Gs can run. This is meant +// to be called in cases when the syscall is expected to block. +// +// func BlockingPoll(fds *PollEvent, nfds int, timeout *syscall.Timespec) (n int, err syscall.Errno) +TEXT ·BlockingPoll(SB),NOSPLIT,$0-40 + BL ·callEntersyscallblock(SB) + MOVD fds+0(FP), R0 + MOVD nfds+8(FP), R1 + MOVD timeout+16(FP), R2 + MOVD $0x0, R3 // sigmask parameter which isn't used here + MOVD $0x49, R8 // SYS_PPOLL + SVC + CMP $0xfffffffffffff001, R0 + BLS ok + MOVD $-1, R1 + MOVD R1, n+24(FP) + NEG R0, R0 + MOVD R0, err+32(FP) + BL ·callExitsyscall(SB) + RET +ok: + MOVD R0, n+24(FP) + MOVD $0, err+32(FP) + BL ·callExitsyscall(SB) + RET diff --git a/pkg/tcpip/link/rawfile/blockingpoll_unsafe.go b/pkg/tcpip/link/rawfile/blockingpoll_noyield_unsafe.go index 84dc0e918..621ab8d29 100644 --- a/pkg/tcpip/link/rawfile/blockingpoll_unsafe.go +++ b/pkg/tcpip/link/rawfile/blockingpoll_noyield_unsafe.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build linux,!amd64 +// +build linux,!amd64,!arm64 package rawfile @@ -22,7 +22,7 @@ import ( ) // BlockingPoll is just a stub function that forwards to the ppoll() system call -// on non-amd64 platforms. +// on non-amd64 and non-arm64 platforms. func BlockingPoll(fds *PollEvent, nfds int, timeout *syscall.Timespec) (int, syscall.Errno) { n, _, e := syscall.Syscall6(syscall.SYS_PPOLL, uintptr(unsafe.Pointer(fds)), uintptr(nfds), uintptr(unsafe.Pointer(timeout)), 0, 0, 0) diff --git a/pkg/tcpip/link/rawfile/blockingpoll_amd64_unsafe.go b/pkg/tcpip/link/rawfile/blockingpoll_yield_unsafe.go index 47039a446..dda3b10a6 100644 --- a/pkg/tcpip/link/rawfile/blockingpoll_amd64_unsafe.go +++ b/pkg/tcpip/link/rawfile/blockingpoll_yield_unsafe.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build linux,amd64 +// +build linux,amd64 linux,arm64 // +build go1.12 // +build !go1.14 @@ -25,6 +25,12 @@ import ( _ "unsafe" // for go:linkname ) +// BlockingPoll on amd64/arm64 makes the ppoll() syscall while calling the +// version of entersyscall that relinquishes the P so that other Gs can +// run. This is meant to be called in cases when the syscall is expected to +// block. On non amd64/arm64 platforms it just forwards to the ppoll() system +// call. +// //go:noescape func BlockingPoll(fds *PollEvent, nfds int, timeout *syscall.Timespec) (int, syscall.Errno) diff --git a/pkg/tcpip/link/sharedmem/BUILD b/pkg/tcpip/link/sharedmem/BUILD index f2998aa98..0a5ea3dc4 100644 --- a/pkg/tcpip/link/sharedmem/BUILD +++ b/pkg/tcpip/link/sharedmem/BUILD @@ -1,4 +1,5 @@ -load("//tools/go_stateify:defs.bzl", "go_library", "go_test") +load("//tools/go_stateify:defs.bzl", "go_library") +load("@io_bazel_rules_go//go:def.bzl", "go_test") package(licenses = ["notice"]) diff --git a/pkg/tcpip/link/sharedmem/pipe/BUILD b/pkg/tcpip/link/sharedmem/pipe/BUILD index 94725cb11..330ed5e94 100644 --- a/pkg/tcpip/link/sharedmem/pipe/BUILD +++ b/pkg/tcpip/link/sharedmem/pipe/BUILD @@ -1,4 +1,5 @@ -load("//tools/go_stateify:defs.bzl", "go_library", "go_test") +load("//tools/go_stateify:defs.bzl", "go_library") +load("@io_bazel_rules_go//go:def.bzl", "go_test") package(licenses = ["notice"]) diff --git a/pkg/tcpip/link/sharedmem/queue/BUILD b/pkg/tcpip/link/sharedmem/queue/BUILD index 160a8f864..de1ce043d 100644 --- a/pkg/tcpip/link/sharedmem/queue/BUILD +++ b/pkg/tcpip/link/sharedmem/queue/BUILD @@ -1,4 +1,5 @@ -load("//tools/go_stateify:defs.bzl", "go_library", "go_test") +load("//tools/go_stateify:defs.bzl", "go_library") +load("@io_bazel_rules_go//go:def.bzl", "go_test") package(licenses = ["notice"]) diff --git a/pkg/tcpip/link/sharedmem/sharedmem.go b/pkg/tcpip/link/sharedmem/sharedmem.go index 834ea5c40..9e71d4edf 100644 --- a/pkg/tcpip/link/sharedmem/sharedmem.go +++ b/pkg/tcpip/link/sharedmem/sharedmem.go @@ -94,7 +94,7 @@ type endpoint struct { // New creates a new shared-memory-based endpoint. Buffers will be broken up // into buffers of "bufferSize" bytes. -func New(mtu, bufferSize uint32, addr tcpip.LinkAddress, tx, rx QueueConfig) (tcpip.LinkEndpointID, error) { +func New(mtu, bufferSize uint32, addr tcpip.LinkAddress, tx, rx QueueConfig) (stack.LinkEndpoint, error) { e := &endpoint{ mtu: mtu, bufferSize: bufferSize, @@ -102,15 +102,15 @@ func New(mtu, bufferSize uint32, addr tcpip.LinkAddress, tx, rx QueueConfig) (tc } if err := e.tx.init(bufferSize, &tx); err != nil { - return 0, err + return nil, err } if err := e.rx.init(bufferSize, &rx); err != nil { e.tx.cleanup() - return 0, err + return nil, err } - return stack.RegisterLinkEndpoint(e), nil + return e, nil } // Close frees all resources associated with the endpoint. @@ -132,7 +132,8 @@ func (e *endpoint) Close() { } } -// Wait waits until all workers have stopped after a Close() call. +// Wait implements stack.LinkEndpoint.Wait. It waits until all workers have +// stopped after a Close() call. func (e *endpoint) Wait() { e.completed.Wait() } diff --git a/pkg/tcpip/link/sharedmem/sharedmem_test.go b/pkg/tcpip/link/sharedmem/sharedmem_test.go index 98036f367..0e9ba0846 100644 --- a/pkg/tcpip/link/sharedmem/sharedmem_test.go +++ b/pkg/tcpip/link/sharedmem/sharedmem_test.go @@ -119,12 +119,12 @@ func newTestContext(t *testing.T, mtu, bufferSize uint32, addr tcpip.LinkAddress initQueue(t, &c.txq, &c.txCfg) initQueue(t, &c.rxq, &c.rxCfg) - id, err := New(mtu, bufferSize, addr, c.txCfg, c.rxCfg) + ep, err := New(mtu, bufferSize, addr, c.txCfg, c.rxCfg) if err != nil { t.Fatalf("New failed: %v", err) } - c.ep = stack.FindLinkEndpoint(id).(*endpoint) + c.ep = ep.(*endpoint) c.ep.Attach(c) return c diff --git a/pkg/tcpip/link/sniffer/sniffer.go b/pkg/tcpip/link/sniffer/sniffer.go index 36c8c46fc..e401dce44 100644 --- a/pkg/tcpip/link/sniffer/sniffer.go +++ b/pkg/tcpip/link/sniffer/sniffer.go @@ -58,10 +58,10 @@ type endpoint struct { // New creates a new sniffer link-layer endpoint. It wraps around another // endpoint and logs packets and they traverse the endpoint. -func New(lower tcpip.LinkEndpointID) tcpip.LinkEndpointID { - return stack.RegisterLinkEndpoint(&endpoint{ - lower: stack.FindLinkEndpoint(lower), - }) +func New(lower stack.LinkEndpoint) stack.LinkEndpoint { + return &endpoint{ + lower: lower, + } } func zoneOffset() (int32, error) { @@ -102,15 +102,15 @@ func writePCAPHeader(w io.Writer, maxLen uint32) error { // snapLen is the maximum amount of a packet to be saved. Packets with a length // less than or equal too snapLen will be saved in their entirety. Longer // packets will be truncated to snapLen. -func NewWithFile(lower tcpip.LinkEndpointID, file *os.File, snapLen uint32) (tcpip.LinkEndpointID, error) { +func NewWithFile(lower stack.LinkEndpoint, file *os.File, snapLen uint32) (stack.LinkEndpoint, error) { if err := writePCAPHeader(file, snapLen); err != nil { - return 0, err + return nil, err } - return stack.RegisterLinkEndpoint(&endpoint{ - lower: stack.FindLinkEndpoint(lower), + return &endpoint{ + lower: lower, file: file, maxPCAPLen: snapLen, - }), nil + }, nil } // DeliverNetworkPacket implements the stack.NetworkDispatcher interface. It is @@ -240,6 +240,9 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen return e.lower.WritePacket(r, gso, hdr, payload, protocol) } +// Wait implements stack.LinkEndpoint.Wait. +func (*endpoint) Wait() {} + func logPacket(prefix string, protocol tcpip.NetworkProtocolNumber, b buffer.View, gso *stack.GSO) { // Figure out the network layer info. var transProto uint8 diff --git a/pkg/tcpip/link/waitable/BUILD b/pkg/tcpip/link/waitable/BUILD index 2597d4b3e..0746dc8ec 100644 --- a/pkg/tcpip/link/waitable/BUILD +++ b/pkg/tcpip/link/waitable/BUILD @@ -1,4 +1,5 @@ -load("//tools/go_stateify:defs.bzl", "go_library", "go_test") +load("//tools/go_stateify:defs.bzl", "go_library") +load("@io_bazel_rules_go//go:def.bzl", "go_test") package(licenses = ["notice"]) diff --git a/pkg/tcpip/link/waitable/waitable.go b/pkg/tcpip/link/waitable/waitable.go index 3b6ac2ff7..5a1791cb5 100644 --- a/pkg/tcpip/link/waitable/waitable.go +++ b/pkg/tcpip/link/waitable/waitable.go @@ -40,11 +40,10 @@ type Endpoint struct { // New creates a new waitable link-layer endpoint. It wraps around another // endpoint and allows the caller to block new write/dispatch calls and wait for // the inflight ones to finish before returning. -func New(lower tcpip.LinkEndpointID) (tcpip.LinkEndpointID, *Endpoint) { - e := &Endpoint{ - lower: stack.FindLinkEndpoint(lower), +func New(lower stack.LinkEndpoint) *Endpoint { + return &Endpoint{ + lower: lower, } - return stack.RegisterLinkEndpoint(e), e } // DeliverNetworkPacket implements stack.NetworkDispatcher.DeliverNetworkPacket. @@ -121,3 +120,6 @@ func (e *Endpoint) WaitWrite() { func (e *Endpoint) WaitDispatch() { e.dispatchGate.Close() } + +// Wait implements stack.LinkEndpoint.Wait. +func (e *Endpoint) Wait() {} diff --git a/pkg/tcpip/link/waitable/waitable_test.go b/pkg/tcpip/link/waitable/waitable_test.go index 56e18ecb0..ae23c96b7 100644 --- a/pkg/tcpip/link/waitable/waitable_test.go +++ b/pkg/tcpip/link/waitable/waitable_test.go @@ -70,9 +70,12 @@ func (e *countedEndpoint) WritePacket(r *stack.Route, _ *stack.GSO, hdr buffer.P return nil } +// Wait implements stack.LinkEndpoint.Wait. +func (*countedEndpoint) Wait() {} + func TestWaitWrite(t *testing.T) { ep := &countedEndpoint{} - _, wep := New(stack.RegisterLinkEndpoint(ep)) + wep := New(ep) // Write and check that it goes through. wep.WritePacket(nil, nil /* gso */, buffer.Prependable{}, buffer.VectorisedView{}, 0) @@ -97,7 +100,7 @@ func TestWaitWrite(t *testing.T) { func TestWaitDispatch(t *testing.T) { ep := &countedEndpoint{} - _, wep := New(stack.RegisterLinkEndpoint(ep)) + wep := New(ep) // Check that attach happens. wep.Attach(ep) @@ -139,7 +142,7 @@ func TestOtherMethods(t *testing.T) { hdrLen: hdrLen, linkAddr: linkAddr, } - _, wep := New(stack.RegisterLinkEndpoint(ep)) + wep := New(ep) if v := wep.MTU(); v != mtu { t.Fatalf("Unexpected mtu: got=%v, want=%v", v, mtu) |