| Message ID | 20200313080517.28728-3-jgross@suse.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | xen/locks: fix preempt disabling in lock handling | expand |
On 13.03.2020 09:05, Juergen Gross wrote: > @@ -199,10 +199,10 @@ unsigned long _spin_lock_irqsave(spinlock_t *lock) > void _spin_unlock(spinlock_t *lock) > { > arch_lock_release_barrier(); > - preempt_enable(); > LOCK_PROFILE_REL; > rel_lock(&lock->debug); > add_sized(&lock->tickets.head, 1); > + preempt_enable(); > arch_lock_signal(); > } arch_lock_signal() is a barrier on Arm, hence just like for patch 1 I wonder whether the insertion wouldn't better come after it. Jan
On 13.03.20 09:55, Jan Beulich wrote: > On 13.03.2020 09:05, Juergen Gross wrote: >> @@ -199,10 +199,10 @@ unsigned long _spin_lock_irqsave(spinlock_t *lock) >> void _spin_unlock(spinlock_t *lock) >> { >> arch_lock_release_barrier(); >> - preempt_enable(); >> LOCK_PROFILE_REL; >> rel_lock(&lock->debug); >> add_sized(&lock->tickets.head, 1); >> + preempt_enable(); >> arch_lock_signal(); >> } > > arch_lock_signal() is a barrier on Arm, hence just like for patch 1 > I wonder whether the insertion wouldn't better come after it. Either way is fine for me. It should be noted that preemption is only relevant on the local cpu. So this is about trading lock state visibility against preemption disabled time, and I agree the visible time of the lock held should be minimized at higher priority than the preemption disabled time. I'll modify my patches accordingly, adding a note in this regard to the commit message. Juergen
Hi, On 13/03/2020 09:00, Jürgen Groß wrote: > On 13.03.20 09:55, Jan Beulich wrote: >> On 13.03.2020 09:05, Juergen Gross wrote: >>> @@ -199,10 +199,10 @@ unsigned long _spin_lock_irqsave(spinlock_t *lock) >>> void _spin_unlock(spinlock_t *lock) >>> { >>> arch_lock_release_barrier(); >>> - preempt_enable(); >>> LOCK_PROFILE_REL; >>> rel_lock(&lock->debug); >>> add_sized(&lock->tickets.head, 1); >>> + preempt_enable(); >>> arch_lock_signal(); >>> } >> >> arch_lock_signal() is a barrier on Arm, hence just like for patch 1 >> I wonder whether the insertion wouldn't better come after it. The important barrier in spin_unlock() is arch_lock_release_barrier(). The one in arch_lock_signal() is just to ensure that waking up the other CPUs will not happen before the unlock is seen. The barrier would not have been necessary if the we didn't use 'sev'. > > Either way is fine for me. It should be noted that preemption is only > relevant on the local cpu. So this is about trading lock state > visibility against preemption disabled time, and I agree the visible > time of the lock held should be minimized at higher priority than the > preemption disabled time. I don't think the rationale is about "performance" here. The rationale is you don't know the implementation of arch_lock_signal(). If you get preempted by a thread trying to acquire the same lock, then it may not do the right thing. Linux will also re-enable preemption only after the unlock has been completed. So it would be best to follow the same pattern. Cheers,
diff --git a/xen/common/spinlock.c b/xen/common/spinlock.c index 344981c54a..f05fb068cd 100644 --- a/xen/common/spinlock.c +++ b/xen/common/spinlock.c @@ -160,6 +160,7 @@ void inline _spin_lock_cb(spinlock_t *lock, void (*cb)(void *), void *data) LOCK_PROFILE_VAR; check_lock(&lock->debug); + preempt_disable(); tickets.head_tail = arch_fetch_and_add(&lock->tickets.head_tail, tickets.head_tail); while ( tickets.tail != observe_head(&lock->tickets) ) @@ -171,7 +172,6 @@ void inline _spin_lock_cb(spinlock_t *lock, void (*cb)(void *), void *data) } got_lock(&lock->debug); LOCK_PROFILE_GOT; - preempt_disable(); arch_lock_acquire_barrier(); } @@ -199,10 +199,10 @@ unsigned long _spin_lock_irqsave(spinlock_t *lock) void _spin_unlock(spinlock_t *lock) { arch_lock_release_barrier(); - preempt_enable(); LOCK_PROFILE_REL; rel_lock(&lock->debug); add_sized(&lock->tickets.head, 1); + preempt_enable(); arch_lock_signal(); } @@ -242,15 +242,18 @@ int _spin_trylock(spinlock_t *lock) return 0; new = old; new.tail++; + preempt_disable(); if ( cmpxchg(&lock->tickets.head_tail, old.head_tail, new.head_tail) != old.head_tail ) + { + preempt_enable(); return 0; + } got_lock(&lock->debug); #ifdef CONFIG_DEBUG_LOCK_PROFILE if (lock->profile) lock->profile->time_locked = NOW(); #endif - preempt_disable(); /* * cmpxchg() is a full barrier so no need for an * arch_lock_acquire_barrier().
In case Xen ever gains preemption support the spinlock coding's placement of preempt_disable() and preempt_enable() should be outside of the locked section. Signed-off-by: Juergen Gross <jgross@suse.com> --- xen/common/spinlock.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)