netstack/tcp: software segmentation offload

Right now, we send each tcp packet separately, we call one system
call per-packet. This patch allows to generate multiple tcp packets
and send them by sendmmsg.

The arguable part of this CL is a way how to handle multiple headers.
This CL adds the next field to the Prepandable buffer.

Nginx test results:

Server Software:        nginx/1.15.9
Server Hostname:        10.138.0.2
Server Port:            8080

Document Path:          /10m.txt
Document Length:        10485760 bytes

w/o gso:
Concurrency Level:      5
Time taken for tests:   5.491 seconds
Complete requests:      100
Failed requests:        0
Total transferred:      1048600200 bytes
HTML transferred:       1048576000 bytes
Requests per second:    18.21 [#/sec] (mean)
Time per request:       274.525 [ms] (mean)
Time per request:       54.905 [ms] (mean, across all concurrent requests)
Transfer rate:          186508.03 [Kbytes/sec] received

sw-gso:

Concurrency Level:      5
Time taken for tests:   3.852 seconds
Complete requests:      100
Failed requests:        0
Total transferred:      1048600200 bytes
HTML transferred:       1048576000 bytes
Requests per second:    25.96 [#/sec] (mean)
Time per request:       192.576 [ms] (mean)
Time per request:       38.515 [ms] (mean, across all concurrent requests)
Transfer rate:          265874.92 [Kbytes/sec] received

w/o gso:
$ ./tcp_benchmark --client --duration 15  --ideal
[SUM]  0.0-15.1 sec  2.20 GBytes  1.25 Gbits/sec

software gso:
$ tcp_benchmark --client --duration 15  --ideal --gso $((1<<16)) --swgso
[SUM]  0.0-15.1 sec  3.99 GBytes  2.26 Gbits/sec

PiperOrigin-RevId: 276112677

11 files changed, 236 insertions, 16 deletions

diff --git a/pkg/tcpip/link/channel/channel.go b/pkg/tcpip/link/channel/channel.go
index 95c6f0828..14f197a77 100644
--- a/pkg/tcpip/link/channel/channel.go
+++ b/pkg/tcpip/link/channel/channel.go
@@ -96,7 +96,7 @@ func (e *Endpoint) MTU() uint32 {
 func (e *Endpoint) Capabilities() stack.LinkEndpointCapabilities {
 	caps := stack.LinkEndpointCapabilities(0)
 	if e.GSO {
-		caps |= stack.CapabilityGSO
+		caps |= stack.CapabilityHardwareGSO
 	}
 	return caps
 }
@@ -134,6 +134,33 @@ func (e *Endpoint) WritePacket(_ *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 	return nil
 }
 
+// WritePackets stores outbound packets into the channel.
+func (e *Endpoint) WritePackets(_ *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	payloadView := payload.ToView()
+	n := 0
+packetLoop:
+	for i := range hdrs {
+		hdr := &hdrs[i].Hdr
+		off := hdrs[i].Off
+		size := hdrs[i].Size
+		p := PacketInfo{
+			Header:  hdr.View(),
+			Proto:   protocol,
+			Payload: buffer.NewViewFromBytes(payloadView[off : off+size]),
+			GSO:     gso,
+		}
+
+		select {
+		case e.C <- p:
+			n++
+		default:
+			break packetLoop
+		}
+	}
+
+	return n, nil
+}
+
 // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
 func (e *Endpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	p := PacketInfo{
diff --git a/pkg/tcpip/link/fdbased/endpoint.go b/pkg/tcpip/link/fdbased/endpoint.go
index 255220d1a..ae4858529 100644
--- a/pkg/tcpip/link/fdbased/endpoint.go
+++ b/pkg/tcpip/link/fdbased/endpoint.go
@@ -165,6 +165,9 @@ type Options struct {
 	// disabled.
 	GSOMaxSize uint32
 
+	// SoftwareGSOEnabled indicates whether software GSO is enabled or not.
+	SoftwareGSOEnabled bool
+
 	// PacketDispatchMode specifies the type of inbound dispatcher to be
 	// used for this endpoint.
 	PacketDispatchMode PacketDispatchMode
@@ -242,7 +245,11 @@ func New(opts *Options) (stack.LinkEndpoint, error) {
 		}
 		if isSocket {
 			if opts.GSOMaxSize != 0 {
-				e.caps |= stack.CapabilityGSO
+				if opts.SoftwareGSOEnabled {
+					e.caps |= stack.CapabilitySoftwareGSO
+				} else {
+					e.caps |= stack.CapabilityHardwareGSO
+				}
 				e.gsoMaxSize = opts.GSOMaxSize
 			}
 		}
@@ -397,7 +404,7 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 		eth.Encode(ethHdr)
 	}
 
-	if e.Capabilities()&stack.CapabilityGSO != 0 {
+	if e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 		vnetHdr := virtioNetHdr{}
 		vnetHdrBuf := vnetHdrToByteSlice(&vnetHdr)
 		if gso != nil {
@@ -430,6 +437,123 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 	return rawfile.NonBlockingWrite3(e.fds[0], hdr.View(), payload.ToView(), nil)
 }
 
+// WritePackets writes outbound packets to the file descriptor. If it is not
+// currently writable, the packet is dropped.
+func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	var ethHdrBuf []byte
+	// hdr + data
+	iovLen := 2
+	if e.hdrSize > 0 {
+		// Add ethernet header if needed.
+		ethHdrBuf = make([]byte, header.EthernetMinimumSize)
+		eth := header.Ethernet(ethHdrBuf)
+		ethHdr := &header.EthernetFields{
+			DstAddr: r.RemoteLinkAddress,
+			Type:    protocol,
+		}
+
+		// Preserve the src address if it's set in the route.
+		if r.LocalLinkAddress != "" {
+			ethHdr.SrcAddr = r.LocalLinkAddress
+		} else {
+			ethHdr.SrcAddr = e.addr
+		}
+		eth.Encode(ethHdr)
+		iovLen++
+	}
+
+	n := len(hdrs)
+
+	views := payload.Views()
+	/*
+	 * Each bondary in views can add one more iovec.
+	 *
+	 * payload |      |          |         |
+	 *         -----------------------------
+	 * packets |    |    |    |    |    |  |
+	 *         -----------------------------
+	 * iovecs  |    | |  |    |  | |    |  |
+	 */
+	iovec := make([]syscall.Iovec, n*iovLen+len(views)-1)
+	mmsgHdrs := make([]rawfile.MMsgHdr, n)
+
+	iovecIdx := 0
+	viewIdx := 0
+	viewOff := 0
+	off := 0
+	nextOff := 0
+	for i := range hdrs {
+		prevIovecIdx := iovecIdx
+		mmsgHdr := &mmsgHdrs[i]
+		mmsgHdr.Msg.Iov = &iovec[iovecIdx]
+		packetSize := hdrs[i].Size
+		hdr := &hdrs[i].Hdr
+
+		off = hdrs[i].Off
+		if off != nextOff {
+			// We stop in a different point last time.
+			size := packetSize
+			viewIdx = 0
+			viewOff = 0
+			for size > 0 {
+				if size >= len(views[viewIdx]) {
+					viewIdx++
+					viewOff = 0
+					size -= len(views[viewIdx])
+				} else {
+					viewOff = size
+					size = 0
+				}
+			}
+		}
+		nextOff = off + packetSize
+
+		if ethHdrBuf != nil {
+			v := &iovec[iovecIdx]
+			v.Base = &ethHdrBuf[0]
+			v.Len = uint64(len(ethHdrBuf))
+			iovecIdx++
+		}
+
+		v := &iovec[iovecIdx]
+		hdrView := hdr.View()
+		v.Base = &hdrView[0]
+		v.Len = uint64(len(hdrView))
+		iovecIdx++
+
+		for packetSize > 0 {
+			vec := &iovec[iovecIdx]
+			iovecIdx++
+
+			v := views[viewIdx]
+			vec.Base = &v[viewOff]
+			s := len(v) - viewOff
+			if s <= packetSize {
+				viewIdx++
+				viewOff = 0
+			} else {
+				s = packetSize
+				viewOff += s
+			}
+			vec.Len = uint64(s)
+			packetSize -= s
+		}
+
+		mmsgHdr.Msg.Iovlen = uint64(iovecIdx - prevIovecIdx)
+	}
+
+	packets := 0
+	for packets < n {
+		sent, err := rawfile.NonBlockingSendMMsg(e.fds[0], mmsgHdrs)
+		if err != nil {
+			return packets, err
+		}
+		packets += sent
+		mmsgHdrs = mmsgHdrs[sent:]
+	}
+	return packets, nil
+}
+
 // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
 func (e *endpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	return rawfile.NonBlockingWrite(e.fds[0], packet.ToView())
diff --git a/pkg/tcpip/link/fdbased/packet_dispatchers.go b/pkg/tcpip/link/fdbased/packet_dispatchers.go
index 9fb760891..12168a1dc 100644
--- a/pkg/tcpip/link/fdbased/packet_dispatchers.go
+++ b/pkg/tcpip/link/fdbased/packet_dispatchers.go
@@ -53,7 +53,7 @@ func newReadVDispatcher(fd int, e *endpoint) (linkDispatcher, error) {
 	d := &readVDispatcher{fd: fd, e: e}
 	d.views = make([]buffer.View, len(BufConfig))
 	iovLen := len(BufConfig)
-	if d.e.Capabilities()&stack.CapabilityGSO != 0 {
+	if d.e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 		iovLen++
 	}
 	d.iovecs = make([]syscall.Iovec, iovLen)
@@ -63,7 +63,7 @@ func newReadVDispatcher(fd int, e *endpoint) (linkDispatcher, error) {
 func (d *readVDispatcher) allocateViews(bufConfig []int) {
 	var vnetHdr [virtioNetHdrSize]byte
 	vnetHdrOff := 0
-	if d.e.Capabilities()&stack.CapabilityGSO != 0 {
+	if d.e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 		// The kernel adds virtioNetHdr before each packet, but
 		// we don't use it, so so we allocate a buffer for it,
 		// add it in iovecs but don't add it in a view.
@@ -106,7 +106,7 @@ func (d *readVDispatcher) dispatch() (bool, *tcpip.Error) {
 	if err != nil {
 		return false, err
 	}
-	if d.e.Capabilities()&stack.CapabilityGSO != 0 {
+	if d.e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 		// Skip virtioNetHdr which is added before each packet, it
 		// isn't used and it isn't in a view.
 		n -= virtioNetHdrSize
@@ -195,7 +195,7 @@ func newRecvMMsgDispatcher(fd int, e *endpoint) (linkDispatcher, error) {
 	}
 	d.iovecs = make([][]syscall.Iovec, MaxMsgsPerRecv)
 	iovLen := len(BufConfig)
-	if d.e.Capabilities()&stack.CapabilityGSO != 0 {
+	if d.e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 		// virtioNetHdr is prepended before each packet.
 		iovLen++
 	}
@@ -226,7 +226,7 @@ func (d *recvMMsgDispatcher) allocateViews(bufConfig []int) {
 	for k := 0; k < len(d.views); k++ {
 		var vnetHdr [virtioNetHdrSize]byte
 		vnetHdrOff := 0
-		if d.e.Capabilities()&stack.CapabilityGSO != 0 {
+		if d.e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 			// The kernel adds virtioNetHdr before each packet, but
 			// we don't use it, so so we allocate a buffer for it,
 			// add it in iovecs but don't add it in a view.
@@ -262,7 +262,7 @@ func (d *recvMMsgDispatcher) dispatch() (bool, *tcpip.Error) {
 	// Process each of received packets.
 	for k := 0; k < nMsgs; k++ {
 		n := int(d.msgHdrs[k].Len)
-		if d.e.Capabilities()&stack.CapabilityGSO != 0 {
+		if d.e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 			n -= virtioNetHdrSize
 		}
 		if n <= d.e.hdrSize {
diff --git a/pkg/tcpip/link/loopback/loopback.go b/pkg/tcpip/link/loopback/loopback.go
index 5f5bacf27..a3b48fa73 100644
--- a/pkg/tcpip/link/loopback/loopback.go
+++ b/pkg/tcpip/link/loopback/loopback.go
@@ -90,6 +90,11 @@ func (e *endpoint) WritePacket(_ *stack.Route, _ *stack.GSO, hdr buffer.Prependa
 	return nil
 }
 
+// WritePackets implements stack.LinkEndpoint.WritePackets.
+func (e *endpoint) WritePackets(_ *stack.Route, _ *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	panic("not implemented")
+}
+
 // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
 func (e *endpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	// Reject the packet if it's shorter than an ethernet header.
diff --git a/pkg/tcpip/link/muxed/injectable.go b/pkg/tcpip/link/muxed/injectable.go
index 41524e271..682b60291 100644
--- a/pkg/tcpip/link/muxed/injectable.go
+++ b/pkg/tcpip/link/muxed/injectable.go
@@ -84,12 +84,23 @@ func (m *InjectableEndpoint) InjectInbound(protocol tcpip.NetworkProtocolNumber,
 	m.dispatcher.DeliverNetworkPacket(m, "" /* remote */, "" /* local */, protocol, vv, nil /* linkHeader */)
 }
 
+// WritePackets writes outbound packets to the appropriate
+// LinkInjectableEndpoint based on the RemoteAddress. HandleLocal only works if
+// r.RemoteAddress has a route registered in this endpoint.
+func (m *InjectableEndpoint) WritePackets(r *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	endpoint, ok := m.routes[r.RemoteAddress]
+	if !ok {
+		return 0, tcpip.ErrNoRoute
+	}
+	return endpoint.WritePackets(r, gso, hdrs, payload, protocol)
+}
+
 // WritePacket writes outbound packets to the appropriate LinkInjectableEndpoint
 // based on the RemoteAddress. HandleLocal only works if r.RemoteAddress has a
 // route registered in this endpoint.
-func (m *InjectableEndpoint) WritePacket(r *stack.Route, _ *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) *tcpip.Error {
+func (m *InjectableEndpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) *tcpip.Error {
 	if endpoint, ok := m.routes[r.RemoteAddress]; ok {
-		return endpoint.WritePacket(r, nil /* gso */, hdr, payload, protocol)
+		return endpoint.WritePacket(r, gso, hdr, payload, protocol)
 	}
 	return tcpip.ErrNoRoute
 }
diff --git a/pkg/tcpip/link/rawfile/BUILD b/pkg/tcpip/link/rawfile/BUILD
index 2e8bc772a..05c7b8024 100644
--- a/pkg/tcpip/link/rawfile/BUILD
+++ b/pkg/tcpip/link/rawfile/BUILD
@@ -16,5 +16,8 @@ go_library(
     visibility = [
         "//visibility:public",
     ],
-    deps = ["//pkg/tcpip"],
+    deps = [
+        "//pkg/tcpip",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
 )
diff --git a/pkg/tcpip/link/rawfile/rawfile_unsafe.go b/pkg/tcpip/link/rawfile/rawfile_unsafe.go
index 7e286a3a6..44e25d475 100644
--- a/pkg/tcpip/link/rawfile/rawfile_unsafe.go
+++ b/pkg/tcpip/link/rawfile/rawfile_unsafe.go
@@ -22,6 +22,7 @@ import (
 	"syscall"
 	"unsafe"
 
+	"golang.org/x/sys/unix"
 	"gvisor.dev/gvisor/pkg/tcpip"
 )
 
@@ -101,6 +102,16 @@ func NonBlockingWrite3(fd int, b1, b2, b3 []byte) *tcpip.Error {
 	return nil
 }
 
+// NonBlockingSendMMsg sends multiple messages on a socket.
+func NonBlockingSendMMsg(fd int, msgHdrs []MMsgHdr) (int, *tcpip.Error) {
+	n, _, e := syscall.RawSyscall6(unix.SYS_SENDMMSG, uintptr(fd), uintptr(unsafe.Pointer(&msgHdrs[0])), uintptr(len(msgHdrs)), syscall.MSG_DONTWAIT, 0, 0)
+	if e != 0 {
+		return 0, TranslateErrno(e)
+	}
+
+	return int(n), nil
+}
+
 // PollEvent represents the pollfd structure passed to a poll() system call.
 type PollEvent struct {
 	FD      int32
diff --git a/pkg/tcpip/link/sharedmem/sharedmem.go b/pkg/tcpip/link/sharedmem/sharedmem.go
index 4393403c5..279e2b457 100644
--- a/pkg/tcpip/link/sharedmem/sharedmem.go
+++ b/pkg/tcpip/link/sharedmem/sharedmem.go
@@ -212,6 +212,11 @@ func (e *endpoint) WritePacket(r *stack.Route, _ *stack.GSO, hdr buffer.Prependa
 	return nil
 }
 
+// WritePackets implements stack.LinkEndpoint.WritePackets.
+func (e *endpoint) WritePackets(r *stack.Route, _ *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	panic("not implemented")
+}
+
 // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
 func (e *endpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	v := packet.ToView()
diff --git a/pkg/tcpip/link/sniffer/sniffer.go b/pkg/tcpip/link/sniffer/sniffer.go
index de9ff98cb..39757ea2a 100644
--- a/pkg/tcpip/link/sniffer/sniffer.go
+++ b/pkg/tcpip/link/sniffer/sniffer.go
@@ -193,10 +193,7 @@ func (e *endpoint) GSOMaxSize() uint32 {
 	return 0
 }
 
-// WritePacket implements the stack.LinkEndpoint interface. It is called by
-// higher-level protocols to write packets; it just logs the packet and forwards
-// the request to the lower endpoint.
-func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) *tcpip.Error {
+func (e *endpoint) dumpPacket(gso *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) {
 	if atomic.LoadUint32(&LogPackets) == 1 && e.file == nil {
 		logPacket("send", protocol, hdr.View(), gso)
 	}
@@ -223,9 +220,27 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 			panic(err)
 		}
 	}
+}
+
+// WritePacket implements the stack.LinkEndpoint interface. It is called by
+// higher-level protocols to write packets; it just logs the packet and
+// forwards the request to the lower endpoint.
+func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) *tcpip.Error {
+	e.dumpPacket(gso, hdr, payload, protocol)
 	return e.lower.WritePacket(r, gso, hdr, payload, protocol)
 }
 
+// WritePackets implements the stack.LinkEndpoint interface. It is called by
+// higher-level protocols to write packets; it just logs the packet and
+// forwards the request to the lower endpoint.
+func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	view := payload.ToView()
+	for _, d := range hdrs {
+		e.dumpPacket(gso, d.Hdr, buffer.NewVectorisedView(d.Size, []buffer.View{view[d.Off:][:d.Size]}), protocol)
+	}
+	return e.lower.WritePackets(r, gso, hdrs, payload, protocol)
+}
+
 // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
 func (e *endpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	if atomic.LoadUint32(&LogPackets) == 1 && e.file == nil {
diff --git a/pkg/tcpip/link/waitable/waitable.go b/pkg/tcpip/link/waitable/waitable.go
index e9b50b802..a04fc1062 100644
--- a/pkg/tcpip/link/waitable/waitable.go
+++ b/pkg/tcpip/link/waitable/waitable.go
@@ -109,6 +109,19 @@ func (e *Endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 	return err
 }
 
+// WritePackets implements stack.LinkEndpoint.WritePackets. It is called by
+// higher-level protocols to write packets. It only forwards packets to the
+// lower endpoint if Wait or WaitWrite haven't been called.
+func (e *Endpoint) WritePackets(r *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	if !e.writeGate.Enter() {
+		return len(hdrs), nil
+	}
+
+	n, err := e.lower.WritePackets(r, gso, hdrs, payload, protocol)
+	e.writeGate.Leave()
+	return n, err
+}
+
 // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
 func (e *Endpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	if !e.writeGate.Enter() {
diff --git a/pkg/tcpip/link/waitable/waitable_test.go b/pkg/tcpip/link/waitable/waitable_test.go
index f03bbceec..5f0f8fa2d 100644
--- a/pkg/tcpip/link/waitable/waitable_test.go
+++ b/pkg/tcpip/link/waitable/waitable_test.go
@@ -70,6 +70,12 @@ func (e *countedEndpoint) WritePacket(r *stack.Route, _ *stack.GSO, hdr buffer.P
 	return nil
 }
 
+// WritePackets implements stack.LinkEndpoint.WritePackets.
+func (e *countedEndpoint) WritePackets(r *stack.Route, _ *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	e.writeCount += len(hdrs)
+	return len(hdrs), nil
+}
+
 func (e *countedEndpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	e.writeCount++
 	return nil