[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <alpine.LNX.2.00.1408280026590.6544@monopod.intra.ispras.ru>
Date: Thu, 28 Aug 2014 01:06:36 +0400 (MSK)
From: Alexander Monakov <amonakov@...ras.ru>
To: musl@...ts.openwall.com
Subject: Re: sem_getvalue conformance considerations
On Wed, 27 Aug 2014, Alexander Monakov wrote:
> On IRC Rich noted that I've too conveniently elided cancellation, so here's how
> I think cancellation handler should look like:
s/cancellation handler/"unwait" procedure/: returning with EINTR is botched in
my earlier mail and should employ the same procedure to undo state change
> val = sem->val[0];
> while (val < 0) {
> /* We believe we are a waiter that no sem_post has "seen". */
> oldval = val;
> if ((val = a_cas(sem->val, val, val+1)) == oldval)
> return;
> }
> /* We are a waiter, and a non-negative val[0] indicates that one sem_post
> * noticed us. Consume the corresponding wake count increment. */
> a_dec(sem->val+1);
A plain atomic decrement on wake count like above is not correct. However:
1. I think it's correct as long as no new waiters appear while we are
performing unwait. If new waiters are matched by new posts it's ok; we only
care if the amount of new waiters is higher than the amount of new posts, as
then we are risking to cause a missing wake for one of those, or setting wake
count to -1 if they all consume wakes before we do the increment.
2. Thus it needs a fancy atomic op: decrement val[1] if val[0] is still equal
to "val" we retrieved earlier. Otherwise, retry from the beginning.
Futhermore, doing the above in a true atomic fashion might not be required!
Isn't it okay to decrement wake count, and if we observe new waiters,
increment it back and cause a futex wake? Thus:
retry:
val = sem->val[0];
while (val < 0) {
/* We believe we are a waiter that no sem_post has "seen". */
oldval = val;
if ((val = a_cas(sem->val, val, val+1)) == oldval)
return;
}
/* We are a waiter, and a non-negative val[0] indicates that one sem_post
* noticed us. Consume the corresponding wake count increment. */
a_dec(sem->val+1);
if (val > sem->val[0]) {
/* New waiters appeared. Avoid causing a missing wake and restart. We
* don't enter here if more new posts are posted than new waiters come. */
a_inc(sem->val+1);
futex_wake(sem->val+1, 1);
goto retry;
}
Alexander
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
