diff options
Diffstat (limited to 'pkg/tcpip/link/fdbased/endpoint.go')
| -rw-r--r-- | pkg/tcpip/link/fdbased/endpoint.go | 63 |
1 files changed, 50 insertions, 13 deletions
diff --git a/pkg/tcpip/link/fdbased/endpoint.go b/pkg/tcpip/link/fdbased/endpoint.go index 77f988b9f..f80ac3435 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. @@ -161,11 +178,20 @@ type Options struct { RXChecksumOffload bool } +// fanoutID is used for AF_PACKET based endpoints to enable PACKET_FANOUT +// support in the host kernel. This allows us to use multiple FD's to receive +// from the same underlying NIC. The fanoutID needs to be the same for a given +// set of FD's that point to the same NIC. Trying to set the PACKET_FANOUT +// option for an FD with a fanoutID already in use by another FD for a different +// NIC will return an EINVAL. +var fanoutID = 1 + // 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 +216,7 @@ 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 +233,12 @@ 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 +248,16 @@ 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 + // Increment fanoutID to ensure that we don't re-use the same fanoutID for + // the next endpoint. + fanoutID++ + + return e, nil } func createInboundDispatcher(e *endpoint, fd int, isSocket bool) (linkDispatcher, error) { @@ -247,7 +277,6 @@ func createInboundDispatcher(e *endpoint, fd int, isSocket bool) (linkDispatcher case *unix.SockaddrLinklayer: // enable PACKET_FANOUT mode is the underlying socket is // of type AF_PACKET. - const fanoutID = 1 const fanoutType = 0x8000 // PACKET_FANOUT_HASH | PACKET_FANOUT_FLAG_DEFRAG fanoutArg := fanoutID | fanoutType<<16 if err := syscall.SetsockoptInt(fd, syscall.SOL_PACKET, unix.PACKET_FANOUT, fanoutArg); err != nil { @@ -290,7 +319,11 @@ 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 +353,12 @@ 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 +474,12 @@ 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 } |