[PATCH] locking/osq_lock: Ensure proper locking semantics for osq_lock/osq_unlock()

[PATCH] locking/osq_lock: Ensure proper locking semantics for osq_lock/osq_unlock()

[Waiman Long]

The osq_lock is special in the sense that lock transfer from one CPU to
the next can happen either over the common optimistic_spin_queue.tail
value with uncontended lock or over a lock waiter's own percpu
optimistic_spin_node.locked flag when the lock is contended.

To ensure proper lock synchronization, we need to provide
the acquire/release semantics for the osq_lock/osq_unlock()
functions in both cases. This is currently the case for the
common optimistic_spin_queue.tail value, but not for the percpu
optimistic_spin_node.locked flag as the proper barriers are missing in
some places. Fix that by adding the needed barriers in those places.

Note that the two percpu optimistic_spin_node.locked setting in
osq_unlock() are proceeded by a full barrier xchg() call, but the
contended cachelines are different. This should probably work in most
cases except in some exotic architectures where the barrier semantics
may be cacheline specific. Nevertheless a release barrier is still added
for safety reason as we may opt to relax the xchg() calls in the future.

The "node->locked" read in osq_lock() was relaxed by commit 036cc30c6b6a
("locking/osq: No need for load/acquire when acquire-polling") a while
ago as the smp_load_acquire() loop was causing a performance hit due to
the repeated acquire barriers in the loop and it argued that an earlier
atomic_xchg() call could provide the needed barrier. That may not be
enough especially if we have to loop for a while before the lock is
released. Now with the new smp_cond_load_acquire() helper, only one
acquire barrier is added at the end of the loop. So it shouldn't have
the performance hit noted in that commit.

Currently osq_lock is used only by mutex and rw_semaphore code for queuing
purpose. As a result, the imperfect lock synchronization support does
not cause harmful consequence as the new osq_lock owner of a contended
osq_lock will still have to wait for the real mutex and rwsem lock to
be released by the pervious osq_lock owner before it can acquire it and
go into its critical section. For correctness, we still have to fix it
in case it is used elsewhere which doesn't have this inherent protection.

Fixes: 036cc30c6b6a ("locking/osq: No need for load/acquire when acquire-polling")
Tested-by: Håkon Bugge <haakon.bugge@oracle.com>
Signed-off-by: Waiman Long <longman@redhat.com>
---
 kernel/locking/osq_lock.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/kernel/locking/osq_lock.c b/kernel/locking/osq_lock.c
index b4233dc2c2b0..ef1bbd914917 100644
--- a/kernel/locking/osq_lock.c
+++ b/kernel/locking/osq_lock.c
@@ -143,7 +143,7 @@ bool osq_lock(struct optimistic_spin_queue *lock)
 	 * is implemented with a monitor-wait. vcpu_is_preempted() relies on
 	 * polling, be careful.
 	 */
-	if (smp_cond_load_relaxed(&node->locked, VAL || need_resched() ||
+	if (smp_cond_load_acquire(&node->locked, VAL || need_resched() ||
 				  vcpu_is_preempted(node_cpu(node->prev))))
 		return true;
 
@@ -224,11 +224,14 @@ void osq_unlock(struct optimistic_spin_queue *lock)
 	node = this_cpu_ptr(&osq_node);
 	next = xchg(&node->next, NULL);
 	if (next) {
-		WRITE_ONCE(next->locked, 1);
+		/* Provide release barrier for unlock */
+		smp_store_release(&next->locked, 1);
 		return;
 	}
 
 	next = osq_wait_next(lock, node, OSQ_UNLOCKED_VAL);
-	if (next)
-		WRITE_ONCE(next->locked, 1);
+	if (next) {
+		/* Provide release barrier for unlock */
+		smp_store_release(&next->locked, 1);
+	}
 }
-- 
2.55.0

Re: [PATCH] p for osq_lock/osq_unlock()

[Haakon Bugge]

> On 10 Sep 2026, at 16:19, Waiman Long <longman@redhat.com> wrote:
> 
> The osq_lock is special in the sense that lock transfer from one CPU to
> the next can happen either over the common optimistic_spin_queue.tail
> value with uncontended lock or over a lock waiter's own percpu
> optimistic_spin_node.locked flag when the lock is contended.
> 
> To ensure proper lock synchronization, we need to provide
> the acquire/release semantics for the osq_lock/osq_unlock()
> functions in both cases. This is currently the case for the
> common optimistic_spin_queue.tail value, but not for the percpu
> optimistic_spin_node.locked flag as the proper barriers are missing in
> some places. Fix that by adding the needed barriers in those places.
> 
> Note that the two percpu optimistic_spin_node.locked setting in
> osq_unlock() are proceeded by a full barrier xchg() call, but the

s/proceeded/preceded/

> contended cachelines are different. This should probably work in most
> cases except in some exotic architectures where the barrier semantics
> may be cacheline specific.

As of today, doesn't atomic_xchg() provide full memory barrier?
From the doc: "RMW operations that have a return value are fully
ordered". I assume this applies to both the intra- and inter-
cacheline cases.

> Nevertheless a release barrier is still added
> for safety reason as we may opt to relax the xchg() calls in the future.
> 
> The "node->locked" read in osq_lock() was relaxed by commit 036cc30c6b6a
> ("locking/osq: No need for load/acquire when acquire-polling") a while
> ago as the smp_load_acquire() loop was causing a performance hit due to
> the repeated acquire barriers in the loop and it argued that an earlier
> atomic_xchg() call could provide the needed barrier. That may not be
> enough especially if we have to loop for a while before the lock is
> released. Now with the new smp_cond_load_acquire() helper, only one
> acquire barrier is added at the end of the loop. So it shouldn't have
> the performance hit noted in that commit.
> 
> Currently osq_lock is used only by mutex and rw_semaphore code for queuing
> purpose. As a result, the imperfect lock synchronization support does
> not cause harmful consequence as the new osq_lock owner of a contended
> osq_lock will still have to wait for the real mutex and rwsem lock to
> be released by the pervious osq_lock owner before it can acquire it and

s/pervious/previous/

> go into its critical section. For correctness, we still have to fix it
> in case it is used elsewhere which doesn't have this inherent protection.
> 
> Fixes: 036cc30c6b6a ("locking/osq: No need for load/acquire when acquire-polling")
> Tested-by: Håkon Bugge <haakon.bugge@oracle.com>
> Signed-off-by: Waiman Long <longman@redhat.com>

With the nits fixed and the memory ordering of atomic_xchg() clarified,

Acked-by: Håkon Bugge <haakon.bugge@oracle.com>


Thxs, Håkon


> ---
>  kernel/locking/osq_lock.c | 11 +++++++----
>  1 file changed, 7 insertions(+), 4 deletions(-)
> 
> diff --git a/kernel/locking/osq_lock.c b/kernel/locking/osq_lock.c
> index b4233dc2c2b0..ef1bbd914917 100644
> --- a/kernel/locking/osq_lock.c
> +++ b/kernel/locking/osq_lock.c
> @@ -143,7 +143,7 @@ bool osq_lock(struct optimistic_spin_queue *lock)
>  	 * is implemented with a monitor-wait. vcpu_is_preempted() relies on
>  	 * polling, be careful.
>  	 */
> -	if (smp_cond_load_relaxed(&node->locked, VAL || need_resched() ||
> +	if (smp_cond_load_acquire(&node->locked, VAL || need_resched() ||
>  				  vcpu_is_preempted(node_cpu(node->prev))))
>  		return true;
>  
> @@ -224,11 +224,14 @@ void osq_unlock(struct optimistic_spin_queue *lock)
>  	node = this_cpu_ptr(&osq_node);
>  	next = xchg(&node->next, NULL);
>  	if (next) {
> -		WRITE_ONCE(next->locked, 1);
> +		/* Provide release barrier for unlock */
> +		smp_store_release(&next->locked, 1);
>  		return;
>  	}
>  
>  	next = osq_wait_next(lock, node, OSQ_UNLOCKED_VAL);
> -	if (next)
> -		WRITE_ONCE(next->locked, 1);
> +	if (next) {
> +		/* Provide release barrier for unlock */
> +		smp_store_release(&next->locked, 1);
> +	}
>  }
> -- 
> 2.55.0
> 

Re: [PATCH] locking/osq_lock: Ensure proper locking semantics for osq_lock/osq_unlock()

[David Laight]

On Thu, 10 Sep 2026 10:19:08 -0400
Waiman Long <longman@redhat.com> wrote:

> The osq_lock is special in the sense that lock transfer from one CPU to
> the next can happen either over the common optimistic_spin_queue.tail
> value with uncontended lock or over a lock waiter's own percpu
> optimistic_spin_node.locked flag when the lock is contended.
> 
> To ensure proper lock synchronization, we need to provide
> the acquire/release semantics for the osq_lock/osq_unlock()
> functions in both cases. This is currently the case for the
> common optimistic_spin_queue.tail value, but not for the percpu
> optimistic_spin_node.locked flag as the proper barriers are missing in
> some places. Fix that by adding the needed barriers in those places.
> 
> Note that the two percpu optimistic_spin_node.locked setting in
> osq_unlock() are proceeded by a full barrier xchg() call, but the
> contended cachelines are different. This should probably work in most
> cases except in some exotic architectures where the barrier semantics
> may be cacheline specific. Nevertheless a release barrier is still added
> for safety reason as we may opt to relax the xchg() calls in the future.
> 
> The "node->locked" read in osq_lock() was relaxed by commit 036cc30c6b6a
> ("locking/osq: No need for load/acquire when acquire-polling") a while
> ago as the smp_load_acquire() loop was causing a performance hit due to
> the repeated acquire barriers in the loop and it argued that an earlier
> atomic_xchg() call could provide the needed barrier. That may not be
> enough especially if we have to loop for a while before the lock is
> released. Now with the new smp_cond_load_acquire() helper, only one
> acquire barrier is added at the end of the loop. So it shouldn't have
> the performance hit noted in that commit.
> 
> Currently osq_lock is used only by mutex and rw_semaphore code for queuing
> purpose. As a result, the imperfect lock synchronization support does
> not cause harmful consequence as the new osq_lock owner of a contended
> osq_lock will still have to wait for the real mutex and rwsem lock to
> be released by the pervious osq_lock owner before it can acquire it and
> go into its critical section. For correctness, we still have to fix it
> in case it is used elsewhere which doesn't have this inherent protection.
> 
> Fixes: 036cc30c6b6a ("locking/osq: No need for load/acquire when acquire-polling")
> Tested-by: Håkon Bugge <haakon.bugge@oracle.com>
> Signed-off-by: Waiman Long <longman@redhat.com>
> ---
>  kernel/locking/osq_lock.c | 11 +++++++----
>  1 file changed, 7 insertions(+), 4 deletions(-)
> 
> diff --git a/kernel/locking/osq_lock.c b/kernel/locking/osq_lock.c
> index b4233dc2c2b0..ef1bbd914917 100644
> --- a/kernel/locking/osq_lock.c
> +++ b/kernel/locking/osq_lock.c
> @@ -143,7 +143,7 @@ bool osq_lock(struct optimistic_spin_queue *lock)
>  	 * is implemented with a monitor-wait. vcpu_is_preempted() relies on
>  	 * polling, be careful.
>  	 */
> -	if (smp_cond_load_relaxed(&node->locked, VAL || need_resched() ||
> +	if (smp_cond_load_acquire(&node->locked, VAL || need_resched() ||
>  				  vcpu_is_preempted(node_cpu(node->prev))))

The comment above needs changing to match.

David

>  		return true;
>  
> @@ -224,11 +224,14 @@ void osq_unlock(struct optimistic_spin_queue *lock)
>  	node = this_cpu_ptr(&osq_node);
>  	next = xchg(&node->next, NULL);
>  	if (next) {
> -		WRITE_ONCE(next->locked, 1);
> +		/* Provide release barrier for unlock */
> +		smp_store_release(&next->locked, 1);
>  		return;
>  	}
>  
>  	next = osq_wait_next(lock, node, OSQ_UNLOCKED_VAL);
> -	if (next)
> -		WRITE_ONCE(next->locked, 1);
> +	if (next) {
> +		/* Provide release barrier for unlock */
> +		smp_store_release(&next->locked, 1);
> +	}
>  }

Re: [PATCH] locking/osq_lock: Ensure proper locking semantics for osq_lock/osq_unlock()

[Waiman Long]

On 9/14/26 5:53 AM, David Laight wrote:
> On Thu, 10 Sep 2026 10:19:08 -0400
> Waiman Long <longman@redhat.com> wrote:
>
>> The osq_lock is special in the sense that lock transfer from one CPU to
>> the next can happen either over the common optimistic_spin_queue.tail
>> value with uncontended lock or over a lock waiter's own percpu
>> optimistic_spin_node.locked flag when the lock is contended.
>>
>> To ensure proper lock synchronization, we need to provide
>> the acquire/release semantics for the osq_lock/osq_unlock()
>> functions in both cases. This is currently the case for the
>> common optimistic_spin_queue.tail value, but not for the percpu
>> optimistic_spin_node.locked flag as the proper barriers are missing in
>> some places. Fix that by adding the needed barriers in those places.
>>
>> Note that the two percpu optimistic_spin_node.locked setting in
>> osq_unlock() are proceeded by a full barrier xchg() call, but the
>> contended cachelines are different. This should probably work in most
>> cases except in some exotic architectures where the barrier semantics
>> may be cacheline specific. Nevertheless a release barrier is still added
>> for safety reason as we may opt to relax the xchg() calls in the future.
>>
>> The "node->locked" read in osq_lock() was relaxed by commit 036cc30c6b6a
>> ("locking/osq: No need for load/acquire when acquire-polling") a while
>> ago as the smp_load_acquire() loop was causing a performance hit due to
>> the repeated acquire barriers in the loop and it argued that an earlier
>> atomic_xchg() call could provide the needed barrier. That may not be
>> enough especially if we have to loop for a while before the lock is
>> released. Now with the new smp_cond_load_acquire() helper, only one
>> acquire barrier is added at the end of the loop. So it shouldn't have
>> the performance hit noted in that commit.
>>
>> Currently osq_lock is used only by mutex and rw_semaphore code for queuing
>> purpose. As a result, the imperfect lock synchronization support does
>> not cause harmful consequence as the new osq_lock owner of a contended
>> osq_lock will still have to wait for the real mutex and rwsem lock to
>> be released by the pervious osq_lock owner before it can acquire it and
>> go into its critical section. For correctness, we still have to fix it
>> in case it is used elsewhere which doesn't have this inherent protection.
>>
>> Fixes: 036cc30c6b6a ("locking/osq: No need for load/acquire when acquire-polling")
>> Tested-by: Håkon Bugge <haakon.bugge@oracle.com>
>> Signed-off-by: Waiman Long <longman@redhat.com>
>> ---
>>   kernel/locking/osq_lock.c | 11 +++++++----
>>   1 file changed, 7 insertions(+), 4 deletions(-)
>>
>> diff --git a/kernel/locking/osq_lock.c b/kernel/locking/osq_lock.c
>> index b4233dc2c2b0..ef1bbd914917 100644
>> --- a/kernel/locking/osq_lock.c
>> +++ b/kernel/locking/osq_lock.c
>> @@ -143,7 +143,7 @@ bool osq_lock(struct optimistic_spin_queue *lock)
>>   	 * is implemented with a monitor-wait. vcpu_is_preempted() relies on
>>   	 * polling, be careful.
>>   	 */
>> -	if (smp_cond_load_relaxed(&node->locked, VAL || need_resched() ||
>> +	if (smp_cond_load_acquire(&node->locked, VAL || need_resched() ||
>>   				  vcpu_is_preempted(node_cpu(node->prev))))
> The comment above needs changing to match.
>
> David

Thank for noticing that. I update the comment in v3 patch.

Cheers,
Longman

Re: [PATCH] locking/osq_lock: Ensure proper locking semantics for osq_lock/osq_unlock()

[Peter Zijlstra]

On Thu, Sep 10, 2026 at 10:19:08AM -0400, Waiman Long wrote:
> The osq_lock is special in the sense that lock transfer from one CPU to
> the next can happen either over the common optimistic_spin_queue.tail
> value with uncontended lock or over a lock waiter's own percpu
> optimistic_spin_node.locked flag when the lock is contended.
> 
> To ensure proper lock synchronization, we need to provide
> the acquire/release semantics for the osq_lock/osq_unlock()
> functions in both cases. This is currently the case for the
> common optimistic_spin_queue.tail value, but not for the percpu
> optimistic_spin_node.locked flag as the proper barriers are missing in
> some places. Fix that by adding the needed barriers in those places.
> 
> Note that the two percpu optimistic_spin_node.locked setting in
> osq_unlock() are proceeded by a full barrier xchg() call, but the
> contended cachelines are different. This should probably work in most
> cases except in some exotic architectures where the barrier semantics
> may be cacheline specific. Nevertheless a release barrier is still added
> for safety reason as we may opt to relax the xchg() calls in the future.
> 
> The "node->locked" read in osq_lock() was relaxed by commit 036cc30c6b6a
> ("locking/osq: No need for load/acquire when acquire-polling") a while
> ago as the smp_load_acquire() loop was causing a performance hit due to
> the repeated acquire barriers in the loop and it argued that an earlier
> atomic_xchg() call could provide the needed barrier. That may not be
> enough especially if we have to loop for a while before the lock is
> released. Now with the new smp_cond_load_acquire() helper, only one
> acquire barrier is added at the end of the loop. So it shouldn't have
> the performance hit noted in that commit.

You need to substantiate this *should*.

> Currently osq_lock is used only by mutex and rw_semaphore code for queuing
> purpose. As a result, the imperfect lock synchronization support does
> not cause harmful consequence as the new osq_lock owner of a contended
> osq_lock will still have to wait for the real mutex and rwsem lock to
> be released by the pervious osq_lock owner before it can acquire it and
> go into its critical section. For correctness, we still have to fix it
> in case it is used elsewhere which doesn't have this inherent protection.
> 
> Fixes: 036cc30c6b6a ("locking/osq: No need for load/acquire when acquire-polling")

This doesn't make sense. You cannot argue that the code is correct as is
and still add Fixes.

Re: [PATCH] locking/osq_lock: Ensure proper locking semantics for osq_lock/osq_unlock()

[Waiman Long]

On 9/14/26 7:22 AM, Peter Zijlstra wrote:
> On Thu, Sep 10, 2026 at 10:19:08AM -0400, Waiman Long wrote:
>> The osq_lock is special in the sense that lock transfer from one CPU to
>> the next can happen either over the common optimistic_spin_queue.tail
>> value with uncontended lock or over a lock waiter's own percpu
>> optimistic_spin_node.locked flag when the lock is contended.
>>
>> To ensure proper lock synchronization, we need to provide
>> the acquire/release semantics for the osq_lock/osq_unlock()
>> functions in both cases. This is currently the case for the
>> common optimistic_spin_queue.tail value, but not for the percpu
>> optimistic_spin_node.locked flag as the proper barriers are missing in
>> some places. Fix that by adding the needed barriers in those places.
>>
>> Note that the two percpu optimistic_spin_node.locked setting in
>> osq_unlock() are proceeded by a full barrier xchg() call, but the
>> contended cachelines are different. This should probably work in most
>> cases except in some exotic architectures where the barrier semantics
>> may be cacheline specific. Nevertheless a release barrier is still added
>> for safety reason as we may opt to relax the xchg() calls in the future.
>>
>> The "node->locked" read in osq_lock() was relaxed by commit 036cc30c6b6a
>> ("locking/osq: No need for load/acquire when acquire-polling") a while
>> ago as the smp_load_acquire() loop was causing a performance hit due to
>> the repeated acquire barriers in the loop and it argued that an earlier
>> atomic_xchg() call could provide the needed barrier. That may not be
>> enough especially if we have to loop for a while before the lock is
>> released. Now with the new smp_cond_load_acquire() helper, only one
>> acquire barrier is added at the end of the loop. So it shouldn't have
>> the performance hit noted in that commit.
> You need to substantiate this *should*.
I don't have a good benchmark that can show any noticeable performance 
difference by adding an acquire barrier. My suspicion is that a 
repeating acquire barrier can be expensive depending on the actual 
processor used. Anyway the other osq_lock/osq_unlock() entry and exit 
points all have suitable barrier. I doubt adding one more acquire 
barrier will have a noticeable impact given that it is one-off at the 
end instead of a repeating one like the original osq_lock() code.
>> Currently osq_lock is used only by mutex and rw_semaphore code for queuing
>> purpose. As a result, the imperfect lock synchronization support does
>> not cause harmful consequence as the new osq_lock owner of a contended
>> osq_lock will still have to wait for the real mutex and rwsem lock to
>> be released by the pervious osq_lock owner before it can acquire it and
>> go into its critical section. For correctness, we still have to fix it
>> in case it is used elsewhere which doesn't have this inherent protection.
>>
>> Fixes: 036cc30c6b6a ("locking/osq: No need for load/acquire when acquire-polling")
> This doesn't make sense. You cannot argue that the code is correct as is
> and still add Fixes.

That is true. I added this paragraph to show that it is not a serious 
bug, but I still think it is better to fix it or we have to explicitly 
say that it is not a real lock in term of how it can be used. I have 
taken it out to avoid confusion.

Cheers,
Longman


>

Re: [PATCH] locking/osq_lock: Ensure proper locking semantics for osq_lock/osq_unlock()

[Peter Zijlstra]

On Mon, Sep 14, 2026 at 10:30:12PM -0400, Waiman Long wrote:
> On 9/14/26 7:22 AM, Peter Zijlstra wrote:

> > This doesn't make sense. You cannot argue that the code is correct as is
> > and still add Fixes.
> 
> That is true. I added this paragraph to show that it is not a serious bug,
> but I still think it is better to fix it or we have to explicitly say that
> it is not a real lock in term of how it can be used. I have taken it out to
> avoid confusion.

"It is not a real lock" will do. It has always been meant as a optimistic
spin queue -- very much not a lock.

Re: [PATCH] locking/osq_lock: Ensure proper locking semantics for osq_lock/osq_unlock()

[David Laight]

On Tue, 15 Sep 2026 07:19:05 +0200
Peter Zijlstra <peterz@infradead.org> wrote:

> On Mon, Sep 14, 2026 at 10:30:12PM -0400, Waiman Long wrote:
> > On 9/14/26 7:22 AM, Peter Zijlstra wrote:  
> 
> > > This doesn't make sense. You cannot argue that the code is correct as is
> > > and still add Fixes.  
> > 
> > That is true. I added this paragraph to show that it is not a serious bug,
> > but I still think it is better to fix it or we have to explicitly say that
> > it is not a real lock in term of how it can be used. I have taken it out to
> > avoid confusion.  
> 
> "It is not a real lock" will do. It has always been meant as a optimistic
> spin queue -- very much not a lock.
> 
> 

I'm not so sure, assuming that is probably short sighted and it will
cause grief sometime in the future.

I did start trying to locate the actual race.
If a read barrier fixes it there must be two reads that get swapped,
so it should be possible to find them.
Unfortunately I don't have a system that is likely to reproduce the error.

David

Re: [PATCH] locking/osq_lock: Ensure proper locking semantics for osq_lock/osq_unlock()

[Peter Zijlstra]

On Mon, Sep 14, 2026 at 10:30:12PM -0400, Waiman Long wrote:
> On 9/14/26 7:22 AM, Peter Zijlstra wrote:

> > > The "node->locked" read in osq_lock() was relaxed by commit 036cc30c6ba
> > > ("locking/osq: No need for load/acquire when acquire-polling") a while
> > > ago as the smp_load_acquire() loop was causing a performance hit due to
> > > the repeated acquire barriers in the loop and it argued that an earlier
> > > atomic_xchg() call could provide the needed barrier. That may not be
> > > enough especially if we have to loop for a while before the lock is
> > > released. Now with the new smp_cond_load_acquire() helper, only one
> > > acquire barrier is added at the end of the loop. So it shouldn't have
> > > the performance hit noted in that commit.

> > You need to substantiate this *should*.

> I don't have a good benchmark that can show any noticeable performance
> difference by adding an acquire barrier. My suspicion is that a repeating
> acquire barrier can be expensive depending on the actual processor used.
> Anyway the other osq_lock/osq_unlock() entry and exit points all have
> suitable barrier. I doubt adding one more acquire barrier will have a
> noticeable impact given that it is one-off at the end instead of a repeating
> one like the original osq_lock() code.

The obvious benchmark is the one that got used for 036cc30c6ba, no?

Re: [PATCH] locking/osq_lock: Ensure proper locking semantics for osq_lock/osq_unlock()

[Haakon Bugge]

> On 15 Sep 2026, at 10:26, Peter Zijlstra <peterz@infradead.org> wrote:
> 
> On Mon, Sep 14, 2026 at 10:30:12PM -0400, Waiman Long wrote:
> > On 9/14/26 7:22 AM, Peter Zijlstra wrote:
> 
> > > > The "node->locked" read in osq_lock() was relaxed by commit 036cc30c6ba
> > > > ("locking/osq: No need for load/acquire when acquire-polling") a while
> > > > ago as the smp_load_acquire() loop was causing a performance hit due to
> > > > the repeated acquire barriers in the loop and it argued that an earlier
> > > > atomic_xchg() call could provide the needed barrier. That may not be
> > > > enough especially if we have to loop for a while before the lock is
> > > > released. Now with the new smp_cond_load_acquire() helper, only one
> > > > acquire barrier is added at the end of the loop. So it shouldn't have
> > > > the performance hit noted in that commit.
> 
> > > You need to substantiate this *should*.
> 
> > I don't have a good benchmark that can show any noticeable performance
> > difference by adding an acquire barrier. My suspicion is that a repeating
> > acquire barrier can be expensive depending on the actual processor used.
> > Anyway the other osq_lock/osq_unlock() entry and exit points all have
> > suitable barrier. I doubt adding one more acquire barrier will have a
> > noticeable impact given that it is one-off at the end instead of a repeating
> > one like the original osq_lock() code.
> 
> The obvious benchmark is the one that got used for 036cc30c6ba, no?

The code in 036cc30c6b was quite different. The
"while (!ACCESS_ONCE(node->locked))" loop doesn't exist
today. I did a quick test, by adding a debug print of the accumulated
number of iterations for my mx_test testing mutex's. So, bigger is better.
Average of 100 runs of 10 seconds on a 160 CPU Arm BM system:

    _relaxed	   __acquire
3978996981,9  	3986106562,2

So, Waiman's _acquire variant is a modest 0,18% faster on this system.


Thxs, Håkon

Re: [PATCH] locking/osq_lock: Ensure proper locking semantics for osq_lock/osq_unlock()

[Waiman Long]

On 9/14/26 6:30 AM, Haakon Bugge wrote:
> [$Subject fixed]
>
>> On 10 Sep 2026, at 16:19, Waiman Long <longman@redhat.com> wrote:
>>
>> The osq_lock is special in the sense that lock transfer from one CPU to
>> the next can happen either over the common optimistic_spin_queue.tail
>> value with uncontended lock or over a lock waiter's own percpu
>> optimistic_spin_node.locked flag when the lock is contended.
>>
>> To ensure proper lock synchronization, we need to provide
>> the acquire/release semantics for the osq_lock/osq_unlock()
>> functions in both cases. This is currently the case for the
>> common optimistic_spin_queue.tail value, but not for the percpu
>> optimistic_spin_node.locked flag as the proper barriers are missing in
>> some places. Fix that by adding the needed barriers in those places.
>>
>> Note that the two percpu optimistic_spin_node.locked setting in
>> osq_unlock() are proceeded by a full barrier xchg() call, but the
>
> s/proceeded/preceded/
>
>> contended cachelines are different. This should probably work in most
>> cases except in some exotic architectures where the barrier semantics
>> may be cacheline specific.
>
> As of today, doesn't atomic_xchg() provide full memory barrier?
> From the doc: "RMW operations that have a return value are fully
> ordered". I assume this applies to both the intra- and inter-
> cacheline cases.

Yes, you are right. That is why I decide to keep the WRITE_ONCE() in v2/v3.

Cheers,
Longman