[PATCH RFC 0/3] nptl: Introduce and use FUTEX_LOCK_PI2

[Kurt Kanzenbach]

On Mon Jun 21 2021, Adhemerval Zanella wrote:
> Currently we check for PI mutex support on pthread_mutex_init which basically
> check for futex_cmpxchg_enabled within kernel (so it fails only on a handful
> configurations).
>
> For FUTEX_LOCK_PI2 I think we will need to rework it, since we are moving
> the futex PI failure from pthread_mutex_init to pthread_mutex_{timed}lock.
> It means that we will need to remove the prio_inherit_missing() test on
> pthread_mutex_init and make the pthread_mutex_{timed}lock fail instead
> (not sure if we should use ENOTSUP or keep with current EINVAL).
>
> The proposed futex_lockpi2_supported() incurs in an extra syscall on every
> mutex slow path, we should avoid it.

Yes, sure.

> I would like also to avoid the CRIU issue as well, so I think it would
> be better to avoid any caching (as done by prio_inherit_missing()),
> and optimize the FUTEX_LOCK_PI2 to be used only when the timeout for
> clock different than CLOCK_REALTIME is required.

OK.

>
> So instead it would be better to move the logic solely on futex_lock_pi() 
> (I am assuming FUTEX_LOCK_PI2 would be only added for futex_time64):

The kernel also adds FUTEX_LOCK_PI2 for the old system call interface.

>
>   static __always_inline int
>   futex_lock_pi2_64 (int *futex_word, const struct __timespec64 *abtime,
>                      int private)
>   {
>   # if __ASSUME_FUTEX_LOCK_PI2
>     return INTERNAL_SYSCALL_CALL (futex_time64, futex_word,
>                                   __lll_private_flag (FUTEX_LOCK_PI2, private), 0,
>                                   abstime);
>   # else
>     if (abstime != NULL && clockid != CLOCK_REALTIME)
>       return INTERNAL_SYSCALL_CALL (futex_time64, futex_word,
> 		  		    __lll_private_flag (FUTEX_LOCK_PI2, private), 0,
> 				    abstime);

At this point __ASSUME_FUTEX_LOCK_PI2 is false meaning the kernel does
not have FUTEX_LOCK_PI2 support. But, it calls FUTEX_LOCK_PI2 for
clockid monotonic. This will result in ENOSYS unless it's an old kernel
which is patched. Is that intended?

>     else
>       {
>         int err =  INTERNAL_SYSCALL_CALL (futex_time64, futex_word,
>                                           __lll_private_flag (FUTEX_LOCK_PI, private), 0,
>                                           abstime);
>         if (err == -ENOSYS)
>           err = -EOVERFLOW;
>       }
>   # endif /* __ASSUME_FUTEX_LOCK_PI2  */
>   }
>
>   static __always_inline int
>   futex_lock_pi64 (int *futex_word, const struct __timespec64 *abstime,
>                    int private)
>   {
>     int err;
>   #ifdef __ASSUME_TIME64_SYSCALLS
>     err = futex_lock_pi2_64 (futex_word, abstime, private);

Makes sense.

>   #else /* __ASSUME_TIME64_SYSCALLS  */
>     bool need_time64 = abstime != NULL && !in_time_t_range (abstime->tv_sec)
>     if (need_time64)
>       {
>         err = futex_lock_pi2_64 (futex_word, abstime, private);
>       }
>     else
>       {
>         struct timespec ts32;
>         if (abstime != NULL)
>           ts32 = valid_timespec64_to_timespec (*abstime);
>
>         err = INTERNAL_SYSCALL_CALL (futex, futex_word, __lll_private_flag
>                                      (FUTEX_LOCK_PI, private), 0,
>                                      abstime != NULL ? &ts32 : NULL);
>       }
>   #endif /* __ASSUME_TIME64_SYSCALLS  */
>    [...]
>   }
>
> It would make the changes on pthread_mutex code minimal, it would be only to
> remove the extra check for clockid and adjust the comment.

Well, that's an interesting point. I think the current check has to
stay, because there are two locking paths. Only the slow path calls
futex_lock_pi_64() which may result in ENOSYS for clock monotonic. But,
the fast path which doesn't call futex_lock_pi64() would succeed if the
check is removed.

Maybe something like this:

sysdeps/nptl/futex-internal.h:
|static __always_inline bool
|futex_lockpi2_supported (void)
|{
|  return __ASSUME_FUTEX_LOCK_PI2;
|}

nptl/pthread_mutex_timedlock.c:
|	if (__glibc_unlikely (! futex_lockpi2_supported () &&
|			      clockid != CLOCK_REALTIME))
|	  return EINVAL;

Or did I get something wrong?

>
> Also, as Joseph has noted the __ASSUME_FUTEX_LOCK_PI2 should be the first
> patch.

OK.

> It also does not make sense to add the __ASSUME_FUTEX_LOCK_PI2 on
> tests, they need to be kernel agnostic so you will need to handle a
> possible EINVAL/ENOSUP failure instead.

Agreed.

Thanks for your feedback. I'll rework it accordingly.

Thanks,
Kurt
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 861 bytes
Desc: not available
URL: <https://sourceware.org/pipermail/libc-alpha/attachments/20210622/9e08232e/attachment.sig>