tcpip/link/fdbased: socket operation on non-socket

[xjasonlyu]

Description

Since commit e511fc9 removed the WritePacket method from tcpip/link/fdbased package, this error may occur under some certain circumstances.

e.g. When we use tcpip/link/tun to open a TUN device:

fd, err := tun.Open(t.name)
if err != nil {
	return nil, fmt.Errorf("create tun: %w", err)
}

Then pass the TUN fd as a fdbased.Option and add the endpoint to stack:

ep, err := fdbased.New(&fdbased.Options{
	FDs: []int{fd},
	...
})

The WritePackets method would be called to write back packets:

gvisor/pkg/tcpip/link/fdbased/endpoint.go

Line 671 in e511fc9

func (e *endpoint) WritePackets(_ stack.RouteInfo, pkts stack.PacketBufferList, _ tcpip.NetworkProtocolNumber) (int, tcpip.Error) {

Then the sendBatch:

gvisor/pkg/tcpip/link/fdbased/endpoint.go

Line 689 in e511fc9

n, err := e.sendBatch(batchFD, batch)

However, in the the sendBatch, the else would be hit and call the rawfile.NonBlockingSendMMsg(batchFD, mmsgHdrs):

gvisor/pkg/tcpip/link/fdbased/endpoint.go

Lines 637 to 657 in e511fc9

	if len(mmsgHdrs) == 0 {
	// We can't fit batch[0] into a mmsghdr while staying under
	// e.maxSyscallHeaderBytes. Use WritePacket, which will avoid the
	// mmsghdr (by using writev) and re-buffer iovecs more aggressively
	// if necessary (by using e.writevMaxIovs instead of
	// rawfile.MaxIovs).
	pkt := batch[0]
	if err := e.writePacket(pkt.EgressRoute, pkt.NetworkProtocolNumber, pkt); err != nil {
	return packets, err
	}
	packets++
	} else {
	for len(mmsgHdrs) > 0 {
	sent, err := rawfile.NonBlockingSendMMsg(batchFD, mmsgHdrs)
	if err != nil {
	return packets, err
	}
	packets += sent
	mmsgHdrs = mmsgHdrs[sent:]
	}
	}

So the issue happens here, the SENDMMSG syscall is used on a non-socket fd (TUN fd instead) and cause the socket operation on non-socket error:

gvisor/pkg/tcpip/link/rawfile/rawfile_unsafe.go

Lines 135 to 142 in e511fc9

	func NonBlockingSendMMsg(fd int, msgHdrs []MMsgHdr) (int, tcpip.Error) {
	n, _, e := unix.RawSyscall6(unix.SYS_SENDMMSG, uintptr(fd), uintptr(unsafe.Pointer(&msgHdrs[0])), uintptr(len(msgHdrs)), unix.MSG_DONTWAIT, 0, 0)
	if e != 0 {
	return 0, TranslateErrno(e)
	}
	return int(n), nil
	}

Steps to reproduce

Mentioned above.

runsc version

null

docker version (if using docker)

null

uname

null

kubectl (if using Kubernetes)

null

repo state (if built from source)

null

runsc debug logs (if available)

null

[hbhasker]

Thanks for the detailed report. I will take a look.

[hbhasker]

We just need to store that the underlying FD is not a socket and degrade WritePackets to writePacket always.

We do the socket check fdbased.New() but its not plumbed through to the write path. We should do that.