Allow setting socket error

[sef]

Summary

getsockopt() allows you to get the socket error; this allows setsockopt to set the error. This is primarily for testing, although I've got a use on a different platform with a network proxy.

Test Plan

Build & install. I wrote a small server that does

static void
set_error(int connection, int errnum)
{
	int kr = setsockopt(connection, SOL_SOCKET, SO_ERROR, &errnum, sizeof(int));
	if (kr == -1) 
	{
		warn("Could not set socket error to %d", errnum);
	}
}

on demand, and ensures the socket gets the error.

[sef]

I forgot to edit the man page.

[markj]

I'm sure this is useful, but I'd be a bit nervous about kernel code that somehow assumes that so_error can't be set. That is, it might become possible for userspace to trigger unexpected behaviour. I don't have anything concrete to point to though.

I also note that so_error is not locked consistently among different protocols (there are lots of unlocked reads, etc.), so it's hard to see whether taking SO_LOCK() here is sufficient.

[sef]

I was paranoid about the lock, as is only correct.

I did a quick glance through the kernel code and wasn't too worried about so_error changing -- getsockopt will clear it if you get the error -- but one thing this _does_ allow you to do is change the error that would be reported if it hasn't been checked yet.

[sef]

Any comments or reviews? 😄

[markj]

In D45602#1042628, @sef wrote:

Any comments or reviews? 😄

I'm still skeptical that this is safe, especially given that it can be used to clear an existing error. I would, at the very least, disallow the use of this option to set so_error = 0.

[sef]

If you call getsockopt with SO_ERROR, it clears out so_error, so the ability to asynchronously clear the error already exists, and hasn't broken things.

[glebius]

There are protocols that would never set so_error. As we go away from sosend_generic/soreceive_generic, number of such protocols will grow. So the protocol itself asserts that its so_error is 0 always. Handling magic setting from the userland will complicate things. I understand this can be useful for some debugging or instrumenting fail scenarios, but I'm really reluctant on having that in the main branch.

[sef]

My actual use case (on a different OS) was for a proxy.

What protocols require that so_error always be 0?