On Tue, 2026-06-30 at 10:23 +0100, Tvrtko Ursulin wrote:

On 26/06/2026 09:19, Philipp Stanner wrote:

...

The entity->last_scheduled field has always been set and read with special RCU functions in addition to memory barriers. There is no obvious reason for that, since the entity lock is available and taken at all places that evaluate the last_scheduled field. The only exception is drm_sched_entity_error(), which is not performance critical in any way.

I agree this looks odd since all call sites apart from drm_sched_entity_error() use "rcu_dereference_check(entity->last_scheduled, true);" ie. "ignore" the RCU.

Btw this was added in:

commit 70102d77ff22dd88a0111b1c3bac5099ac5d0425 Author: Christian König christian.koenig@amd.com Date:   Mon Apr 17 17:32:11 2023 +0200

drm/scheduler: add drm_sched_entity_error and use rcu for last_scheduled

You may want to add this as a reference in the commit message.

I did git-blame for that commit. It looks like this:

drm/scheduler: add drm_sched_entity_error and use rcu for last_scheduled

Switch to using RCU handling for the last scheduled job and add a function to return the error code of it.

It's a good example of why I think it's so vital to write verbose commit messages. The only way to find out why this was added is to ask the author, if he's still around [which is the case in this case].

I can't see the value of adding a link? That commit says "add foo" and my commit says "remove foo because it achieves nothing".

I guess it relied on dma-fence RCU destruction to enable lockless lookups from the AMD submit path. Given how many other locks we have in those paths it is probably noise to have one more so maybe it is a win to remove some barriers and those rcu_dereference_check-true lines. I think Christian will need to comment.

My argument is more that locks are the right tool to use unless there is proof to the contrary.

...

Improve robustness, readability and maintainability by replacing RCU and barriers with the lock.

As a preparational step, while at it, also guard spsc_queue_pop() with the lock, since spsc_queue is deprecated and supposed to be replaced with a locked list.

You would have said to split the logical changes into separate patches.

Me? :D

In this case, a lock that did not exist is added from nowhere. But I tend to think that you are right. We could leave spsc_queue lockless for now. That's cleaner.

...

[…]

...

struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)   {

• /* Helper to avoid dropping the reference while the entity lock is held,

• * just to have some more robustness.
  

• */
  

• struct dma_fence *prev_last_scheduled;

struct drm_sched_job *sched_job;      sched_job = drm_sched_entity_queue_peek(entity); @@ -523,19 +532,20 @@ struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)    if (entity->guilty && atomic_read(entity->guilty))    dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);

• dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));
• rcu_assign_pointer(entity->last_scheduled,

• 	   dma_fence_get(&sched_job->s_fence->finished));
  

• spin_lock(&entity->lock);
• prev_last_scheduled = entity->last_scheduled;
• entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);

• /*

• * If the queue is empty we allow drm_sched_entity_select_rq() to
  

• * locklessly access ->last_scheduled. This only works if we set the
  

• * pointer before we dequeue and if we a write barrier here.
  

• /* A recent rework required taking the spinlock above. Since spsc_queue

• * is scheduled for removal as per the DRM-TODO-list, we access it here
  

• * locked already to prepare for that cleanup.
  

• *
  

• * TODO: Fully replace spsc_queue with a locked (h)list.
  

*/

• smp_wmb();

spsc_queue_pop(&entity->job_queue);

• spin_unlock(&entity->lock);

• dma_fence_put(prev_last_scheduled);

drm_sched_rq_pop_entity(entity);

Notice the entity->lock ends up cycled twice for no good reason (second

Getting rid of hard to understand barriers + RCU *is* a _very_ good reason.

is in drm_sched_rq_pop_entity()). So I would suggest you somehow reduce that to once. Probably just pull out entity->lock out of the drm_sched_rq_pop_entity() to drm_sched_entity_pop_job()?

Can you see danger in sense of a significant performance regression because of that?

I guess if you do that then the "while at it" part of the commit message can be "upgraded" to "spsc_queue_pop() being under the lock as a consequence of the rework" and then no need to split it.

I agree with you that it should be *downgraded* instead.

...

/* Jobs and entities might have different lifecycles. Since we're @@ -561,21 +571,15 @@ void drm_sched_entity_select_rq(struct drm_sched_entity *entity)    if (spsc_queue_count(&entity->job_queue))    return;

• /*

• * Only when the queue is empty are we guaranteed that
  

• * drm_sched_run_job_work() cannot change entity->last_scheduled. To
  

• * enforce ordering we need a read barrier here. See
  

• * drm_sched_entity_pop_job() for the other side.
  

• */
  

• smp_rmb();
• fence = rcu_dereference_check(entity->last_scheduled, true);

• spin_lock(&entity->lock);
• fence = entity->last_scheduled;

/* stay on the same engine if the previous job hasn't finished */

• if (fence && !dma_fence_is_signaled(fence))

• if (fence && !dma_fence_is_signaled(fence)) {

• spin_unlock(&entity->lock);
  

Have you tried with lockdep to see if there are any hidden lock inversions with this?

As far as I could grep really no one touches the entity lock (which is not surprising, since the entire drm_sched design resolves around the central philosophy: "NEVER use a spinlock unless you absolutely have to". When you look at the old code and documentation, you see that locks were really only ever used to protect lists.

Anyways. This is the scheduler's fence. It can never implement any callback to someone who might interfere with the entity lock, can it?

I also wonder if we could demote this to a flag check only and remove any doubt. I don't think opportunistic signalling matter in this code path.

With the new fence API, where we can bypass the ops, that would probably be the more canonical code. But that's then indeed something for a separate patch.

P.