From: Joakim Sindholt <opensource@zhasha.com>
To: musl@lists.openwall.com
Subject: Re: New optimized normal-type mutex?
Date: Wed, 29 Jul 2015 14:09:27 +0200 [thread overview]
Message-ID: <8c49d81e.dNq.dMV.21.hNiSfA@mailjet.com> (raw)
In-Reply-To: <20150521234402.GA25373@brightrain.aerifal.cx>
On Fri, May 22, 2015 at 1:44 AM, Rich Felker <dalias@libc.org> wrote:
> I realized (thanks to conversations with wm4) that it's possible to
> optimize normal-type mutexes a lot to eliminate spurious futex wakes
> and reduce the number of atomic ops/barriers under contention.
>
> The concept is to store the full mutex state in a single int:
>
> bits 0-29: waiter count
> bit 30: waiter flag
> bit 31: lock bit
>
> Then unlock is:
>
> old = a_fetch_and(p, 0x3fffffff); // new op
> if (old & 0x40000000) wake(p,1);
>
> Trylock is:
>
> old = *p;
> if (old < 0) {
> a_barrier();
> old = *p;
> }
> if (old >= 0 && a_cas(p, old, old|INT_MIN)) return 0;
>
> Lock is:
>
> while (spins--) {
> if (!trylock()) return 0;
> a_spin();
> }
> while ((old = *p) < 0) {
> new = old+1 | 0x40000000;
> if (a_cas(p, old, new)==old) break;
> }
> for(;;) {
> while ((old = *p) < 0) {
> futex_wait(p, old);
> }
> new = old-1;
> if (new) new |= 0x40000000; // set waiters flag if any other waiters
> new |= INT_MIN;
> if (a_cas(p, old, new)==old) return 0;
> }
I implemented this lock in my own library and I even did some
preliminary race tests. I couldn't find fault with it. However amonakov
did and made me aware of it on IRC.
Thread A takes the lock
Thread B tries to take the lock but ends up in futex_wait
Thread A unlocks the lock
Thread A futex_wakes Thread B
Thread A steals the lock
Thread B goes back into futex_wait
Thread A unlocks the lock
Thread B hangs
Because:
The first unlock clears the 0x40000000 flag and it does not get set
again at any point.
So he went on and suggested that a cas-less lock was possible with
a_fetch_add however I can't make it work and I don't think he can
either. His idea however is sound: the one who flips the sign bit takes
the lock. Based on that I've cobbled together a different lock that will
probably perform worse than this approach but none-the-less be correct
as far as I can tell.
The difference is that we consider the lock owner a waiter as well, thus
requiring a cas loop in the unlock function to remove itself, so to
speak, from the waiter count. a_fetch_and also turns into a cas loop so
I consider this fairly minor.
This makes the wait loop a little simpler while still maintaining a
waiter count and still only using one int.
void lock(volatile int *p)
{
int val, spins = 200;
while (spins--) {
if (trylock(p)) { return; }
spin();
}
while (1) {
if ((val = a_load(p)) < 0) {
if (a_cas(p, val, val - 1) == val) { break; }
} else {
if (a_cas(p, val, -val - 1) == val) { return; }
}
}
while (1) {
while ((val = a_load(p)) < 0) {
futex_wait(p, val);
}
if (a_cas(p, val, -val) == val) { return; }
}
}
int trylock(volatile int *p)
{
int val = a_load(p);
if (val < 0) {
a_barrier();
val = a_load(p);
}
return (val >= 0 && a_cas(p, val, -val - 1) == val);
}
void unlock(volatile int *p)
{
int new, val = a_load(p);
while ((new = a_cas(p, val, -(val + 1))) != val) { val = new; }
if (val + 1 != 0) { futex_wake(p, 1); }
}
-- Joakim
next prev parent reply other threads:[~2015-07-29 12:09 UTC|newest]
Thread overview: 20+ messages / expand[flat|nested] mbox.gz Atom feed top
2015-05-21 23:44 Rich Felker
2015-05-22 7:30 ` Jens Gustedt
2015-05-22 7:51 ` Rich Felker
2015-07-29 12:09 ` Joakim Sindholt [this message]
2015-07-29 22:11 ` Jens Gustedt
2015-07-29 23:30 ` Rich Felker
2015-07-29 23:49 ` Jens Gustedt
2015-07-30 0:10 ` Rich Felker
2015-07-30 8:07 ` Jens Gustedt
2015-07-30 9:10 ` Jens Gustedt
2015-07-30 9:36 ` Alexander Monakov
2015-07-30 10:00 ` Jens Gustedt
2015-07-30 11:37 ` Alexander Monakov
2015-07-30 13:46 ` Rich Felker
2015-07-30 16:07 ` Jens Gustedt
2015-08-03 16:36 ` Alexander Monakov
2015-08-03 19:43 ` Jens Gustedt
2015-08-03 20:05 ` Isaac Dunham
2015-08-04 5:49 ` Jens Gustedt
2015-07-30 13:45 ` Rich Felker
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=8c49d81e.dNq.dMV.21.hNiSfA@mailjet.com \
--to=opensource@zhasha.com \
--cc=musl@lists.openwall.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
Code repositories for project(s) associated with this public inbox
https://git.vuxu.org/mirror/musl/
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).