[tailscale] net: add TCP socket creation/close hooks to SockTrace API

Extends the hooks added by #45 to also expose when TCP sockets are
created or closed (meant to allow TCP stats to be read from them). We
don't do this for all socket types since stats are not available for
UDP sockets, and they tend to be short-lived, thus invoking the hooks
would be useless overhead.

Also fixes read/write hooks to not count out-of-band data, since that's
usually not sent over the wire.

Updates tailscale/corp#9230
Updates #58

Signed-off-by: Jenny Zhang <jz@tailscale.com>
(Cherry-picked from db4dc90)
(cherry picked from commit 51a96ad)
(cherry picked from commit aff7d04)

Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>

api/go1.99999.txt

@@ -1,2 +1,10 @@
+pkg net, func ContextSockTrace(context.Context) *SockTrace
 pkg net, func SetDialEnforcer(func(context.Context, []Addr) error) #55
 pkg net, func SetResolveEnforcer(func(context.Context, string, string, string, Addr) error) #55
+pkg net, func WithSockTrace(context.Context, *SockTrace) context.Context
+pkg net, type SockTrace struct
+pkg net, type SockTrace struct, DidCreateTCPConn func(syscall.RawConn)
+pkg net, type SockTrace struct, DidRead func(int)
+pkg net, type SockTrace struct, DidWrite func(int)
+pkg net, type SockTrace struct, WillCloseTCPConn func(syscall.RawConn)
+pkg net, type SockTrace struct, WillOverwrite func(*SockTrace)

src/net/fd_posix.go

@@ -24,6 +24,12 @@ type netFD struct {
 	net         string
 	laddr       Addr
 	raddr       Addr
+
+	// hooks (if provided) are called after successful reads or writes with the
+	// number of bytes transferred.
+	readHook  func(int)
+	writeHook func(int)
+	closeHook func()
 }
 
 func (fd *netFD) setAddr(laddr, raddr Addr) {
@@ -34,6 +40,9 @@ func (fd *netFD) setAddr(laddr, raddr Addr) {
 
 func (fd *netFD) Close() error {
 	runtime.SetFinalizer(fd, nil)
+	if fd.closeHook != nil {
+		fd.closeHook()
+	}
 	return fd.pfd.Close()
 }
 
@@ -44,10 +53,16 @@ func (fd *netFD) shutdown(how int) error {
 }
 
 func (fd *netFD) closeRead() error {
+	if fd.closeHook != nil {
+		fd.closeHook()
+	}
 	return fd.shutdown(syscall.SHUT_RD)
 }
 
 func (fd *netFD) closeWrite() error {
+	if fd.closeHook != nil {
+		fd.closeHook()
+	}
 	return fd.shutdown(syscall.SHUT_WR)
 }
 
@@ -76,18 +91,27 @@ func (fd *netFD) readFromInet6(p []byte, from *syscall.SockaddrInet6) (n int, er
 
 func (fd *netFD) readMsg(p []byte, oob []byte, flags int) (n, oobn, retflags int, sa syscall.Sockaddr, err error) {
 	n, oobn, retflags, sa, err = fd.pfd.ReadMsg(p, oob, flags)
+	if fd.readHook != nil && err == nil {
+		fd.readHook(n)
+	}
 	runtime.KeepAlive(fd)
 	return n, oobn, retflags, sa, wrapSyscallError(readMsgSyscallName, err)
 }
 
 func (fd *netFD) readMsgInet4(p []byte, oob []byte, flags int, sa *syscall.SockaddrInet4) (n, oobn, retflags int, err error) {
 	n, oobn, retflags, err = fd.pfd.ReadMsgInet4(p, oob, flags, sa)
+	if fd.readHook != nil && err == nil {
+		fd.readHook(n)
+	}
 	runtime.KeepAlive(fd)
 	return n, oobn, retflags, wrapSyscallError(readMsgSyscallName, err)
 }
 
 func (fd *netFD) readMsgInet6(p []byte, oob []byte, flags int, sa *syscall.SockaddrInet6) (n, oobn, retflags int, err error) {
 	n, oobn, retflags, err = fd.pfd.ReadMsgInet6(p, oob, flags, sa)
+	if fd.readHook != nil && err == nil {
+		fd.readHook(n)
+	}
 	runtime.KeepAlive(fd)
 	return n, oobn, retflags, wrapSyscallError(readMsgSyscallName, err)
 }
@@ -118,18 +142,27 @@ func (fd *netFD) writeToInet6(p []byte, sa *syscall.SockaddrInet6) (n int, err e
 
 func (fd *netFD) writeMsg(p []byte, oob []byte, sa syscall.Sockaddr) (n int, oobn int, err error) {
 	n, oobn, err = fd.pfd.WriteMsg(p, oob, sa)
+	if fd.writeHook != nil && err == nil {
+		fd.writeHook(n)
+	}
 	runtime.KeepAlive(fd)
 	return n, oobn, wrapSyscallError(writeMsgSyscallName, err)
 }
 
 func (fd *netFD) writeMsgInet4(p []byte, oob []byte, sa *syscall.SockaddrInet4) (n int, oobn int, err error) {
 	n, oobn, err = fd.pfd.WriteMsgInet4(p, oob, sa)
+	if fd.writeHook != nil && err == nil {
+		fd.writeHook(n)
+	}
 	runtime.KeepAlive(fd)
 	return n, oobn, wrapSyscallError(writeMsgSyscallName, err)
 }
 
 func (fd *netFD) writeMsgInet6(p []byte, oob []byte, sa *syscall.SockaddrInet6) (n int, oobn int, err error) {
 	n, oobn, err = fd.pfd.WriteMsgInet6(p, oob, sa)
+	if fd.writeHook != nil && err == nil {
+		fd.writeHook(n)
+	}
 	runtime.KeepAlive(fd)
 	return n, oobn, wrapSyscallError(writeMsgSyscallName, err)
 }

src/net/sock_posix.go

@@ -28,6 +28,21 @@ func socket(ctx context.Context, net string, family, sotype, proto int, ipv6only
 		poll.CloseFunc(s)
 		return nil, err
 	}
+	if trace := ContextSockTrace(ctx); trace != nil {
+		fd.readHook = trace.DidRead
+		fd.writeHook = trace.DidWrite
+		if (trace.DidCreateTCPConn != nil || trace.WillCloseTCPConn != nil) && len(net) >= 3 && net[0:3] == "tcp" {
+			c := newRawConn(fd)
+			if trace.DidCreateTCPConn != nil {
+				trace.DidCreateTCPConn(c)
+			}
+			if trace.WillCloseTCPConn != nil {
+				fd.closeHook = func() {
+					trace.WillCloseTCPConn(c)
+				}
+			}
+		}
+	}
 
 	// This function makes a network file descriptor for the
 	// following applications:

src/net/socktrace.go

@@ -0,0 +1,53 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package net
+
+import (
+	"context"
+	"syscall"
+)
+
+// SockTrace is a set of hooks to run at various operations on a network socket.
+// Any particular hook may be nil. Functions may be called concurrently from
+// different goroutines.
+type SockTrace struct {
+	// DidOpenTCPConn is called when a TCP socket was created. The
+	// underlying raw network connection that was created is provided.
+	DidCreateTCPConn func(c syscall.RawConn)
+	// DidRead is called after a successful read from the socket, where n bytes
+	// were read.
+	DidRead func(n int)
+	// DidWrite is called after a successful write to the socket, where n bytes
+	// were written.
+	DidWrite func(n int)
+	// WillOverwrite is called when the registered trace is overwritten by a
+	// subsequent call to WithSockTrace. The provided trace is the new trace
+	// that will be used.
+	WillOverwrite func(trace *SockTrace)
+	// WillCloseTCPConn is called when a TCP socket is about to be closed. The
+	// underlying raw network connection that is being closed is provided.
+	WillCloseTCPConn func(c syscall.RawConn)
+}
+
+// WithSockTrace returns a new context based on the provided parent
+// ctx. Socket reads and writes made with the returned context will use
+// the provided trace hooks. Any previous hooks registered with ctx are
+// ovewritten (their WillOverwrite hook will be called).
+func WithSockTrace(ctx context.Context, trace *SockTrace) context.Context {
+	if previous := ContextSockTrace(ctx); previous != nil && previous.WillOverwrite != nil {
+		previous.WillOverwrite(trace)
+	}
+	return context.WithValue(ctx, sockTraceKey{}, trace)
+}
+
+// ContextSockTrace returns the SockTrace associated with the
+// provided context. If none, it returns nil.
+func ContextSockTrace(ctx context.Context) *SockTrace {
+	trace, _ := ctx.Value(sockTraceKey{}).(*SockTrace)
+	return trace
+}
+
+// unique type to prevent assignment.
+type sockTraceKey struct{}