Re: sem_wait and EINTR

[Rich Felker <dalias@libc.org>]

On Wed, Dec 12, 2018 at 06:15:59AM +0100, Markus Wichmann wrote:
> On Tue, Dec 11, 2018 at 07:32:38PM -0500, Rich Felker wrote:
> > One other thought: would it be preferable for the EINTR suppression in
> > the absence of interruptible signal handlers to be in __timedwait
> > rather than sem_timedwait? Then any code using __timedwait would
> > benefit from it. I'm not sure if there are other callers where it
> > would help but it wouldn't hurt either.
> 
> Nope, that would not help. I had a look at all users of SYS_futex that

Perhaps help was the wrong word; I think you're right that there's
nowhere else it matters and that all other callers already ignore
EINTR unconditionally because they're supposed to. The only plausible
improvement is avoiding spurious dec/inc cycle on the waiter count in
some places. On the other hand it might be a nicer factorization (less
ugly and linux-bug-specific logic in high level code, i.e.
sem_timedwait) if the workaround were buried in low-level stuff
(__timedwait).

x> might be impacted by the kernel bug you mentioned (and followed them
> back to the public interfaces that use them). Only __timedwait does
> anything with the return value at all, and of all the users of
> __timedwait(), sem_timedwait() is the only function even specified to
> return EINTR. All others are specified to *never* return EINTR. At least
> according to the manpages I have here (manpages-posix).
> 
> So only sem_timedwait() needs this patch. For the other users it would
> hurt conformance.

I don't see how it could hurt conformance.

Rich