udp: retry failed GSO sendmsg on EIO without GSO

[mxinden]

We have multiple reports of Linux systems that might:

1. Pass our GSO support check at runtime
   

   quinn/quinn-udp/src/unix.rs

   Lines 823 to 825 in 93b6d01

   	/// Checks whether GSO support is available by checking the kernel version followed by setting
   	/// the UDP_SEGMENT option on a socket
   	pub(crate) fn max_gso_segments() -> usize {

2. But then fail with EIO on the first GSO sendmsg
   

   quinn/quinn-udp/src/unix.rs

   Lines 345 to 349 in 93b6d01

   	// Some network adapters and drivers do not support GSO. Unfortunately, Linux
   	// offers no easy way for us to detect this short of an EIO or sometimes EINVAL
   	// when we try to actually send datagrams using it.
   	#[cfg(any(target_os = "linux", target_os = "android"))]
   	if let Some(libc::EIO) | Some(libc::EINVAL) = e.raw_os_error() {

Reports:

• Disable GSO after encountering EIO on send, redux #1210
• https://bugzilla.mozilla.org/show_bug.cgi?id=1989895

Currently quinn-udp simply returns an error.

Instead, quinn-udp could attempt sending each segment individually.

Note that we already do a retry on other errors:

quinn/quinn-udp/src/unix.rs

Lines 362 to 376 in 93b6d01

	// Some arguments to `sendmsg` are not supported. Switch to
	// fallback mode and retry if we haven't already.
	if e.raw_os_error() == Some(libc::EINVAL) && !state.sendmsg_einval() {
	state.set_sendmsg_einval();
	prepare_msg(
	transmit,
	&dst_addr,
	&mut msg_hdr,
	&mut iovec,
	&mut cmsgs,
	encode_src_ip,
	state.sendmsg_einval(),
	);
	continue;
	}

I am planning to land a hot-fix in Firefox for now, Relying on the QUIC layer to resend. relying on our QUIC state machine adds unnecessary delays.

Documenting this optimization here. I am not sure I will get to it any time soon.

[Ralith]

Retrying immediately makes sense to me. I think the only reason we haven't done this is because it's fiddlier than just dropping a flag.

[thomaseizinger]

I think this can be reproduced by using ethtool -K eth0 tx off to disable segementation offloading. We had to patch that in Firezone too, see firezone/firezone#10279.

Also relevant here: #2201