UNIX-SIMPLE-POLL restarts timeouts for EINTR

[Michał "phoe" Herda]

The current sb-unix:unix-simple-poll implementation is buggy wrt. EINTR handling; the timeout is reset on every EINTR, meaning that on a busy system, where EINTR happens regularly, longer timeouts will never trigger.

This is what was causing Keepit queues on production (and to some extent on test) to fail to time out, whereas on dev and build systems this worked great.

The bug has been reported to the sbcl-help mailing list at https://sourceforge.net/p/sbcl/mailman/message/35854538/ but AFAIK this was never followed up on. I just discovered that Keepit, for years, has been using a never-upstreamed fix that meddles with SBCL internals, so I'm tying the loose threads in hope of removing the hack from our codebase.

[Christophe Rhodes (csr21-cantab)]

The patch looks OK to me but it would be nice to fix the #-os-provides-poll case further down, I think? It looks like it has the same issue.

[Michał "phoe" Herda (phoe-krk)]

OK. I edited and loaded the modified UNIX-SIMPLE-POLL and my SBCL still seems to work, and the timeout logic seems OK. It's the first time in my career when I touch syscalls though, so please review before merging.

[Wojciech S. Gac (wsgac)]

In the first hunk you have `(let ((timeout (if (= minusp to-msec)' instead of the intended `(let ((timeout (if (minusp to-msec)'.

[Michał "phoe" Herda (phoe-krk)]

Thanks.