more fixes!

overlay/tio/blockon_linux.go

@@ -0,0 +1,50 @@
+//go:build linux && !android
+// +build linux,!android
+
+package tio
+
+import (
+	"os"
+
+	"golang.org/x/sys/unix"
+)
+
+// blockOn parks the calling goroutine until fd is ready (events is POLLIN for
+// reads, POLLOUT for writes) or shutdownFd signals teardown. It builds the
+// pollfd array on the stack every call, so concurrent callers on the same
+// Queue never share Revents storage: the previous shared-array implementation
+// was a genuine Go data race when two writers parked in poll(2) at once (the
+// kernel writing Revents while another goroutine zeroed it). Level-triggered
+// events kept it from deadlocking, but it was still a race.
+//
+// Poll(2) is looped over EINTR. err is checked before the Revents bits are
+// trusted, since a failed poll may leave them bogus. Returns os.ErrClosed when
+// shutdown was signaled (POLLIN on shutdownFd) or either fd reported a problem
+// condition (POLLHUP|POLLNVAL|POLLERR).
+func blockOn(fd, shutdownFd int32, events int16) error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	pfds := [2]unix.PollFd{
+		{Fd: fd, Events: events},
+		{Fd: shutdownFd, Events: unix.POLLIN},
+	}
+	var err error
+	for {
+		_, err = unix.Poll(pfds[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	tunEvents := pfds[0].Revents
+	shutdownEvents := pfds[1].Revents
+	// Check err before trusting the potentially bogus bits we just got.
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	}
+	if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}

overlay/tio/tio.go

@@ -26,8 +26,9 @@ type Capabilities struct {
 	USO bool
 }
 
-// Queue is a readable/writable Poll queue. One Queue is driven by a single
-// read goroutine plus a single writer (see Write below).
+// Queue is a readable/writable Poll queue. Concurrency contract: a single
+// read goroutine drives Read; plain Write is safe for concurrent callers;
+// WriteGSO (on Queues that implement GSOWriter) is single-writer per queue.
 type Queue interface {
 	io.Closer
 
@@ -37,11 +38,12 @@ type Queue interface {
 	// or copy each slice before the next call. A Packet may carry a
 	// GSO/USO superpacket (see GSOInfo); when GSO.IsSuperpacket() is
 	// true the caller must segment Bytes before treating it as a single
-	// IP datagram. Not safe for concurrent Reads.
+	// IP datagram. Single-reader only: not safe for concurrent Reads (it
+	// reuses per-queue rx scratch each call).
 	Read() ([]Packet, error)
 
 	// Write emits a single packet on the plaintext (outside→inside)
-	// delivery path. Not safe for concurrent Writes.
+	// delivery path. Safe for concurrent use.
 	Write(p []byte) (int, error)
 }
 

overlay/tio/tio_gso_linux.go

@@ -8,7 +8,6 @@ import (
 	"io"
 	"log/slog"
 	"os"
-	"sync"
 	"sync/atomic"
 	"syscall"
 	"unsafe"
@@ -62,17 +61,10 @@ var validVnetHdr = [virtio.Size]byte{unix.VIRTIO_NET_HDR_F_DATA_VALID}
 type Offload struct {
 	fd         int
 	shutdownFd int
-	readPoll   [2]unix.PollFd
-	writePoll  [2]unix.PollFd
-	// writeLock serializes blockOnWrite's read+clear of writePoll[*].Revents.
-	// Any goroutine that calls Write may end up parked in poll(2); without
-	// the lock concurrent waiters could race the Revents reset and lose
-	// events.
-	writeLock sync.Mutex
-	closed    atomic.Bool
-	rxBuf     []byte   // backing store for kernel-handed packets read this drain
-	rxOff     int      // cursor into rxBuf for the current Read drain
-	pending   []Packet // packets returned from the most recent Read
+	closed     atomic.Bool
+	rxBuf      []byte   // backing store for kernel-handed packets read this drain
+	rxOff      int      // cursor into rxBuf for the current Read drain
+	pending    []Packet // packets returned from the most recent Read
 
 	// readVnetScratch holds the 10-byte virtio_net_hdr split off the front of
 	// every TUN read via readv(2). Decoupling the header from the packet body
@@ -109,15 +101,6 @@ func newOffload(fd int, shutdownFd int, usoEnabled bool) (*Offload, error) {
 		shutdownFd: shutdownFd,
 		usoEnabled: usoEnabled,
 		closed:     atomic.Bool{},
-		readPoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLIN},
-			{Fd: int32(shutdownFd), Events: unix.POLLIN},
-		},
-		writePoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLOUT},
-			{Fd: int32(shutdownFd), Events: unix.POLLIN},
-		},
-		writeLock: sync.Mutex{},
 
 		rxBuf:   make([]byte, tunRxBufCap),
 		gsoIovs: make([]unix.Iovec, 2, gsoMaxIovs),
@@ -135,57 +118,11 @@ func newOffload(fd int, shutdownFd int, usoEnabled bool) (*Offload, error) {
 }
 
 func (r *Offload) blockOnRead() error {
-	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
-	var err error
-	for {
-		_, err = unix.Poll(r.readPoll[:], -1)
-		if err != unix.EINTR {
-			break
-		}
-	}
-	//always reset these!
-	tunEvents := r.readPoll[0].Revents
-	shutdownEvents := r.readPoll[1].Revents
-	r.readPoll[0].Revents = 0
-	r.readPoll[1].Revents = 0
-	//do the err check before trusting the potentially bogus bits we just got
-	if err != nil {
-		return err
-	}
-	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
-		return os.ErrClosed
-	} else if tunEvents&problemFlags != 0 {
-		return os.ErrClosed
-	}
-	return nil
+	return blockOn(int32(r.fd), int32(r.shutdownFd), unix.POLLIN)
 }
 
 func (r *Offload) blockOnWrite() error {
-	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
-	var err error
-	for {
-		_, err = unix.Poll(r.writePoll[:], -1)
-		if err != unix.EINTR {
-			break
-		}
-	}
-	//always reset these!
-	r.writeLock.Lock()
-	tunEvents := r.writePoll[0].Revents
-	shutdownEvents := r.writePoll[1].Revents
-	r.writePoll[0].Revents = 0
-	r.writePoll[1].Revents = 0
-	r.writeLock.Unlock()
-	//do the err check before trusting the potentially bogus bits we just got
-	if err != nil {
-		return err
-	}
-	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
-		return os.ErrClosed
-	} else if tunEvents&problemFlags != 0 {
-		return os.ErrClosed
-	}
-	return nil
+	return blockOn(int32(r.fd), int32(r.shutdownFd), unix.POLLOUT)
 }
 
 // readPacket issues a single readv(2) splitting the virtio_net_hdr off

overlay/tio/tio_poll_linux.go

@@ -17,11 +17,9 @@ import (
 const tunReadBufSize = 65535
 
 type Poll struct {
-	fd int
-
-	readPoll  [2]unix.PollFd
-	writePoll [2]unix.PollFd
-	closed    atomic.Bool
+	fd         int
+	shutdownFd int
+	closed     atomic.Bool
 
 	readBuf  []byte
 	batchRet [1]Packet
@@ -37,16 +35,9 @@ func newPoll(fd int, shutdownFd int) (*Poll, error) {
 	}
 
 	out := &Poll{
-		fd:      fd,
-		readBuf: make([]byte, tunReadBufSize),
-		readPoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLIN},
-			{Fd: int32(shutdownFd), Events: unix.POLLIN},
-		},
-		writePoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLOUT},
-			{Fd: int32(shutdownFd), Events: unix.POLLIN},
-		},
+		fd:         fd,
+		shutdownFd: shutdownFd,
+		readBuf:    make([]byte, tunReadBufSize),
 	}
 	return out, nil
 }
@@ -54,53 +45,11 @@ func newPoll(fd int, shutdownFd int) (*Poll, error) {
 // blockOnRead waits until the Poll fd is readable or shutdown has been signaled.
 // Returns os.ErrClosed if Close was called.
 func (t *Poll) blockOnRead() error {
-	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
-	var err error
-	for {
-		_, err = unix.Poll(t.readPoll[:], -1)
-		if err != unix.EINTR {
-			break
-		}
-	}
-	tunEvents := t.readPoll[0].Revents
-	shutdownEvents := t.readPoll[1].Revents
-	t.readPoll[0].Revents = 0
-	t.readPoll[1].Revents = 0
-	if err != nil {
-		return err
-	}
-	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
-		return os.ErrClosed
-	}
-	if tunEvents&problemFlags != 0 {
-		return os.ErrClosed
-	}
-	return nil
+	return blockOn(int32(t.fd), int32(t.shutdownFd), unix.POLLIN)
 }
 
 func (t *Poll) blockOnWrite() error {
-	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
-	var err error
-	for {
-		_, err = unix.Poll(t.writePoll[:], -1)
-		if err != unix.EINTR {
-			break
-		}
-	}
-	tunEvents := t.writePoll[0].Revents
-	shutdownEvents := t.writePoll[1].Revents
-	t.writePoll[0].Revents = 0
-	t.writePoll[1].Revents = 0
-	if err != nil {
-		return err
-	}
-	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
-		return os.ErrClosed
-	}
-	if tunEvents&problemFlags != 0 {
-		return os.ErrClosed
-	}
-	return nil
+	return blockOn(int32(t.fd), int32(t.shutdownFd), unix.POLLOUT)
 }
 
 func (t *Poll) Read() ([]Packet, error) {
@@ -133,7 +82,7 @@ func (t *Poll) readOne(to []byte) (int, error) {
 	}
 }
 
-// Write is only valid for single threaded use
+// Write is safe for concurrent use
 func (t *Poll) Write(from []byte) (int, error) {
 	for {
 		n, errno := unix.Write(t.fd, from)

overlay/tio/tun_file_linux_test.go

@@ -70,6 +70,76 @@ func TestPoll_WakeForShutdown_WakesFriends(t *testing.T) {
 	}
 }
 
+// TestPoll_ConcurrentWrite_NoRace hammers a single Poll queue from two writer
+// goroutines while a reader drains the other end of the pipe. The writers
+// overflow the pipe buffer, so both repeatedly park in blockOnWrite at the same
+// time — the exact scenario that raced on the old shared writePoll member
+// array. Run under -race; a shared-array regression trips the detector here.
+func TestPoll_ConcurrentWrite_NoRace(t *testing.T) {
+	var fds [2]int
+	require.NoError(t, unix.Pipe2(fds[:], unix.O_CLOEXEC))
+	readFd, writeFd := fds[0], fds[1]
+
+	shutdownFd, err := unix.Eventfd(0, unix.EFD_NONBLOCK|unix.EFD_CLOEXEC)
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = unix.Close(shutdownFd) })
+
+	p, err := newPoll(writeFd, shutdownFd)
+	require.NoError(t, err)
+
+	const writers = 2
+	const perWriter = 4000
+	payload := make([]byte, 100)
+	total := writers * perWriter * len(payload)
+
+	// Reader: drain the read end (blocking) until every writer's bytes are
+	// consumed, so the writers keep making progress rather than wedging on a
+	// permanently full pipe.
+	readDone := make(chan struct{})
+	go func() {
+		defer close(readDone)
+		buf := make([]byte, 4096)
+		got := 0
+		for got < total {
+			n, rerr := unix.Read(readFd, buf)
+			got += n
+			if rerr != nil {
+				if rerr == unix.EINTR {
+					continue
+				}
+				return
+			}
+			if n == 0 { // EOF
+				return
+			}
+		}
+	}()
+
+	var wg sync.WaitGroup
+	for w := 0; w < writers; w++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			for i := 0; i < perWriter; i++ {
+				if _, werr := p.Write(payload); werr != nil {
+					t.Errorf("write: %v", werr)
+					return
+				}
+			}
+		}()
+	}
+	wg.Wait()
+
+	select {
+	case <-readDone:
+	case <-time.After(10 * time.Second):
+		t.Fatal("reader did not drain")
+	}
+
+	require.NoError(t, p.Close())
+	_ = unix.Close(readFd)
+}
+
 // TestPoll_NewPoll_DoesNotCloseFdOnFailure pins the ownership rule: when
 // newPoll fails, it must leave fd open so the caller (pollQueueSet.Add's
 // callers in tun_linux.go) is the sole closer. If newPoll also closed fd,

overlay/tio/virtio/segment_linux.go

@@ -143,7 +143,7 @@ func CorrectHdrLen(pkt []byte, hdr *Hdr) error {
 	if hdr.HdrLen < hdr.CsumStart {
 		return fmt.Errorf("virtioNetHdr.HdrLen (%d) < virtioNetHdr.CsumStart (%d)", hdr.HdrLen, hdr.CsumStart)
 	}
-	cSumAt := int(hdr.CsumStart + hdr.CsumStart)
+	cSumAt := int(hdr.CsumStart + hdr.CsumOffset)
 	if cSumAt+1 >= len(pkt) {
 		return fmt.Errorf("end of checksum offset (%d) exceeds packet length (%d)", cSumAt+1, len(pkt))
 	}

overlay/tio/virtio/segment_linux_test.go

@@ -211,6 +211,55 @@ func TestSegmentTCPHeaderNotCorrupted(t *testing.T) {
 	}
 }
 
+// TestCorrectHdrLenChecksumBound guards the checksum-field bounds check in
+// CorrectHdrLen. The checksum field sits at CsumStart+CsumOffset, so the check
+// must be computed from CsumStart+CsumOffset — NOT CsumStart+CsumStart, a
+// regression that doubled CsumStart and thus over-tightened the bound (since
+// CsumOffset, 6 for UDP / 16 for TCP, is always < CsumStart >= 20). That bogus
+// bound spuriously rejected valid small USO superpackets in decodeRead.
+func TestCorrectHdrLenChecksumBound(t *testing.T) {
+	// A valid IPv4 USO superpacket: 20B IPv4 + 8B UDP + two 6-byte segments
+	// (payload 12) = 40 bytes total. CsumStart=20, CsumOffset=6, so the UDP
+	// checksum field lives at bytes 26..27, comfortably inside the 40-byte
+	// packet. The OLD formula computed cSumAt = CsumStart+CsumStart = 40 and
+	// rejected on cSumAt+1 (41) >= len(pkt) (40); the fix (CsumStart+CsumOffset
+	// = 26) accepts. This case FAILS against the CsumStart+CsumStart regression.
+	t.Run("valid-small-uso-accepted", func(t *testing.T) {
+		pkt, _, csumStart := buildUDPv4Super(12) // total len 40
+		hdr := Hdr{
+			Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+			GSOType:    unix.VIRTIO_NET_HDR_GSO_UDP_L4,
+			GSOSize:    6, // two 6-byte segments
+			CsumStart:  csumStart,
+			CsumOffset: 6,
+		}
+		if err := CorrectHdrLen(pkt, &hdr); err != nil {
+			t.Fatalf("CorrectHdrLen rejected a valid 40-byte USO superpacket: %v", err)
+		}
+		if hdr.HdrLen != csumStart+udpHeaderLen {
+			t.Errorf("HdrLen = %d, want %d", hdr.HdrLen, csumStart+udpHeaderLen)
+		}
+	})
+
+	// A genuinely-too-short packet: CsumStart=20, CsumOffset=6 means the
+	// checksum field would end at byte 27, but the packet is only 25 bytes
+	// (CsumStart+CsumOffset+2 = 28 > 25). CorrectHdrLen must still reject it.
+	t.Run("too-short-rejected", func(t *testing.T) {
+		pkt := make([]byte, 25)
+		pkt[0] = 0x45 // IPv4, IHL 5
+		hdr := Hdr{
+			Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+			GSOType:    unix.VIRTIO_NET_HDR_GSO_UDP_L4,
+			GSOSize:    6,
+			CsumStart:  20,
+			CsumOffset: 6,
+		}
+		if err := CorrectHdrLen(pkt, &hdr); err == nil {
+			t.Fatalf("CorrectHdrLen accepted a too-short (25-byte) packet")
+		}
+	})
+}
+
 // TestSegmentUDPHeaderNotCorrupted is the USO counterpart: SegmentUDP performs
 // the same header stamp and must be correct when gsoSize < headerLen.
 func TestSegmentUDPHeaderNotCorrupted(t *testing.T) {