diff mbox series

[v6,4/4] Avoid extra load with CAS in __pthread_mutex_clocklock_common [BZ #28537]

Message ID 20211111162428.2286605-5-hjl.tools@gmail.com
State Committed
Commit 49302b8fdf9103b6fc0a398678668a22fa19574c
Headers
Series Optimize CAS [BZ #28537] |

Checks

Context Check Description
dj/TryBot-apply_patch success Patch applied to master at the time it was sent
dj/TryBot-32bit success Build for i686

Commit Message

H.J. Lu Nov. 11, 2021, 4:24 p.m. UTC 
  Replace boolean CAS with value CAS to avoid the extra load.
---
 nptl/pthread_mutex_timedlock.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

Comments

Szabolcs Nagy Nov. 12, 2021, 4:32 p.m. UTC | #1 
The 11/11/2021 08:24, H.J. Lu via Libc-alpha wrote:
> Replace boolean CAS with value CAS to avoid the extra load.

this looks good in general.

Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>

> ---
>  nptl/pthread_mutex_timedlock.c | 10 +++++-----
>  1 file changed, 5 insertions(+), 5 deletions(-)
> 
> diff --git a/nptl/pthread_mutex_timedlock.c b/nptl/pthread_mutex_timedlock.c
> index 57f3f28869..f763cfc7fa 100644
> --- a/nptl/pthread_mutex_timedlock.c
> +++ b/nptl/pthread_mutex_timedlock.c
> @@ -233,12 +233,12 @@ __pthread_mutex_clocklock_common (pthread_mutex_t *mutex,
>  	     meantime.  */
>  	  if ((oldval & FUTEX_WAITERS) == 0)
>  	    {
> -	      if (atomic_compare_and_exchange_bool_acq (&mutex->__data.__lock,
> -							oldval | FUTEX_WAITERS,
> -							oldval)
> -		  != 0)
> +	      int val;
> +	      if ((val = atomic_compare_and_exchange_val_acq
> +		   (&mutex->__data.__lock, oldval | FUTEX_WAITERS,
> +		    oldval)) != oldval)
>  		{
> -		  oldval = mutex->__data.__lock;
> +		  oldval = val;
>  		  continue;
>  		}
>  	      oldval |= FUTEX_WAITERS;
> -- 
> 2.33.1
>
Andreas Schwab Nov. 12, 2021, 6:51 p.m. UTC | #2 
On Nov 11 2021, H.J. Lu wrote:

> diff --git a/nptl/pthread_mutex_timedlock.c b/nptl/pthread_mutex_timedlock.c
> index 57f3f28869..f763cfc7fa 100644
> --- a/nptl/pthread_mutex_timedlock.c
> +++ b/nptl/pthread_mutex_timedlock.c
> @@ -233,12 +233,12 @@ __pthread_mutex_clocklock_common (pthread_mutex_t *mutex,
>  	     meantime.  */
>  	  if ((oldval & FUTEX_WAITERS) == 0)
>  	    {
> -	      if (atomic_compare_and_exchange_bool_acq (&mutex->__data.__lock,
> -							oldval | FUTEX_WAITERS,
> -							oldval)
> -		  != 0)
> +	      int val;
> +	      if ((val = atomic_compare_and_exchange_val_acq
> +		   (&mutex->__data.__lock, oldval | FUTEX_WAITERS,
> +		    oldval)) != oldval)

Please move the assignment out of the condition.

Andreas.
Sunil Pandey Sept. 11, 2022, 8:12 p.m. UTC | #3 
On Fri, Nov 12, 2021 at 10:53 AM Andreas Schwab <schwab@linux-m68k.org> wrote:
>
> On Nov 11 2021, H.J. Lu wrote:
>
> > diff --git a/nptl/pthread_mutex_timedlock.c b/nptl/pthread_mutex_timedlock.c
> > index 57f3f28869..f763cfc7fa 100644
> > --- a/nptl/pthread_mutex_timedlock.c
> > +++ b/nptl/pthread_mutex_timedlock.c
> > @@ -233,12 +233,12 @@ __pthread_mutex_clocklock_common (pthread_mutex_t *mutex,
> >            meantime.  */
> >         if ((oldval & FUTEX_WAITERS) == 0)
> >           {
> > -           if (atomic_compare_and_exchange_bool_acq (&mutex->__data.__lock,
> > -                                                     oldval | FUTEX_WAITERS,
> > -                                                     oldval)
> > -               != 0)
> > +           int val;
> > +           if ((val = atomic_compare_and_exchange_val_acq
> > +                (&mutex->__data.__lock, oldval | FUTEX_WAITERS,
> > +                 oldval)) != oldval)
>
> Please move the assignment out of the condition.
>
> Andreas.
>
> --
> Andreas Schwab, schwab@linux-m68k.org
> GPG Key fingerprint = 7578 EB47 D4E5 4D69 2510  2552 DF73 E780 A9DA AEC1
> "And now for something completely different."

I would like to backport this patch to release branch 2.33 and 2.34

Any comments/suggestions or objections on this.

commit 0b82747dc48d5bf0871bdc6da8cb6eec1256355f
Author: H.J. Lu <hjl.tools@gmail.com>
Date:   Thu Nov 11 06:31:51 2021 -0800

    Avoid extra load with CAS in __pthread_mutex_lock_full [BZ #28537]

    Replace boolean CAS with value CAS to avoid the extra load.

    Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
Arjan van de Ven Sept. 11, 2022, 8:15 p.m. UTC | #4 
On 9/11/2022 1:12 PM, Sunil Pandey wrote:
> On Fri, Nov 12, 2021 at 10:53 AM Andreas Schwab <schwab@linux-m68k.org> wrote:
>>
>> On Nov 11 2021, H.J. Lu wrote:
>>
>>> diff --git a/nptl/pthread_mutex_timedlock.c b/nptl/pthread_mutex_timedlock.c
>>> index 57f3f28869..f763cfc7fa 100644
>>> --- a/nptl/pthread_mutex_timedlock.c
>>> +++ b/nptl/pthread_mutex_timedlock.c
>>> @@ -233,12 +233,12 @@ __pthread_mutex_clocklock_common (pthread_mutex_t *mutex,
>>>             meantime.  */
>>>          if ((oldval & FUTEX_WAITERS) == 0)
>>>            {
>>> -           if (atomic_compare_and_exchange_bool_acq (&mutex->__data.__lock,
>>> -                                                     oldval | FUTEX_WAITERS,
>>> -                                                     oldval)
>>> -               != 0)
>>> +           int val;
>>> +           if ((val = atomic_compare_and_exchange_val_acq
>>> +                (&mutex->__data.__lock, oldval | FUTEX_WAITERS,
>>> +                 oldval)) != oldval)
>>
>> Please move the assignment out of the condition.
>>
>> Andreas.
>>
>> --
>> Andreas Schwab, schwab@linux-m68k.org
>> GPG Key fingerprint = 7578 EB47 D4E5 4D69 2510  2552 DF73 E780 A9DA AEC1
>> "And now for something completely different."
> 
> I would like to backport this patch to release branch 2.33 and 2.34
> 

what exactly is the stable branch policy that would suggest to backport performance improvements like this ?
(most projects are sticking to "strict bugfixes and other gross oversights" as much as possible)
Florian Weimer Sept. 11, 2022, 9:26 p.m. UTC | #5 
* Arjan van de Ven via Libc-alpha:

> what exactly is the stable branch policy that would suggest to
> backport performance improvements like this ?  (most projects are
> sticking to "strict bugfixes and other gross oversights" as much as
> possible)

We occasionally backport safe (and not-so-safe) performance
optimizations.  We draw the line at ABI changes, those aren't possible.
It's also very desirable that all future releases have a superset of the
backports.

But anything else depends on who is willing to do the work.

Thanks,
Florian
Noah Goldstein Sept. 29, 2022, 12:09 a.m. UTC | #6 
On Sun, Sep 11, 2022 at 4:13 PM Sunil Pandey via Libc-stable
<libc-stable@sourceware.org> wrote:
>
> On Fri, Nov 12, 2021 at 10:53 AM Andreas Schwab <schwab@linux-m68k.org> wrote:
> >
> > On Nov 11 2021, H.J. Lu wrote:
> >
> > > diff --git a/nptl/pthread_mutex_timedlock.c b/nptl/pthread_mutex_timedlock.c
> > > index 57f3f28869..f763cfc7fa 100644
> > > --- a/nptl/pthread_mutex_timedlock.c
> > > +++ b/nptl/pthread_mutex_timedlock.c
> > > @@ -233,12 +233,12 @@ __pthread_mutex_clocklock_common (pthread_mutex_t *mutex,
> > >            meantime.  */
> > >         if ((oldval & FUTEX_WAITERS) == 0)
> > >           {
> > > -           if (atomic_compare_and_exchange_bool_acq (&mutex->__data.__lock,
> > > -                                                     oldval | FUTEX_WAITERS,
> > > -                                                     oldval)
> > > -               != 0)
> > > +           int val;
> > > +           if ((val = atomic_compare_and_exchange_val_acq
> > > +                (&mutex->__data.__lock, oldval | FUTEX_WAITERS,
> > > +                 oldval)) != oldval)
> >
> > Please move the assignment out of the condition.
> >
> > Andreas.
> >
> > --
> > Andreas Schwab, schwab@linux-m68k.org
> > GPG Key fingerprint = 7578 EB47 D4E5 4D69 2510  2552 DF73 E780 A9DA AEC1
> > "And now for something completely different."
>
> I would like to backport this patch to release branch 2.33 and 2.34
>
> Any comments/suggestions or objections on this.
>
> commit 0b82747dc48d5bf0871bdc6da8cb6eec1256355f
> Author: H.J. Lu <hjl.tools@gmail.com>
> Date:   Thu Nov 11 06:31:51 2021 -0800
>
>     Avoid extra load with CAS in __pthread_mutex_lock_full [BZ #28537]
>
>     Replace boolean CAS with value CAS to avoid the extra load.
>
>     Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>

Fine by me.
diff mbox series

Patch

diff --git a/nptl/pthread_mutex_timedlock.c b/nptl/pthread_mutex_timedlock.c
index 57f3f28869..f763cfc7fa 100644
--- a/nptl/pthread_mutex_timedlock.c
+++ b/nptl/pthread_mutex_timedlock.c
@@ -233,12 +233,12 @@  __pthread_mutex_clocklock_common (pthread_mutex_t *mutex,
 	     meantime.  */
 	  if ((oldval & FUTEX_WAITERS) == 0)
 	    {
-	      if (atomic_compare_and_exchange_bool_acq (&mutex->__data.__lock,
-							oldval | FUTEX_WAITERS,
-							oldval)
-		  != 0)
+	      int val;
+	      if ((val = atomic_compare_and_exchange_val_acq
+		   (&mutex->__data.__lock, oldval | FUTEX_WAITERS,
+		    oldval)) != oldval)
 		{
-		  oldval = mutex->__data.__lock;
+		  oldval = val;
 		  continue;
 		}
 	      oldval |= FUTEX_WAITERS;