Libc.errno() and GC safepoints

[vchuravy]

errno is a peculiar function since one must guarantee that no other code was executed on the thread between the function that set errno and the query.

As far as I know we currently don't have a way to express that kind of atomicity, and we currently rely on the fact that in most cases the Libc.errno() function is inlined.

If that function were not inlined there is the possibility for it to hit the safepoint on function entry. As far as I know errno is not interrupt safe in general.

This does not seem to be a problem in practice, but we maybe need a function annotation that blocks safepoint on entry, or at the very least mark

julia/base/libc.jl

Line 293 in 6f3fdf7

errno() = ccall(:jl_errno, Cint, ())

as @inline.

We could also consider making sigatomic more user-friendly and recommending it's usage in the context of errno

[vtjnash]

errno is async and safepoint safe (and codegen, etc). Did you run into any counterexample that someone broke recently?

[gbaraldi]

Wouldn't errno potentially get changed if we run a gc in between and somehow that thread does a syscall?

[vtjnash]

Only if someone has introduced a bug into the GC recently (which may be possible, as Windows is known to fairly eagerly smash errno, unlike posix which is only supposed to smash it on failures):

julia/src/gc-stock.c

Lines 3308 to 3311 in e4b29f7

	int last_errno = errno;
	#ifdef _OS_WINDOWS_
	DWORD last_error = GetLastError();
	#endif

[vchuravy]

I don't have a concrete failure in mind, this was more prompted by a recent Slack conversation where I realized I didn't know the rules of the Game.

I see that the usr2 handler is restoring errno, but I didn't see the same logic in the segv handler for the safepoint. Now maybe all that code doesn't touch errno xD