On 23/05/2026 07:06, Boris Brezillon wrote:
On Fri, 22 May 2026 17:25:18 +0100
Tvrtko Ursulin <[email protected]> wrote:

On 22/05/2026 17:08, Boris Brezillon wrote:
On Fri, 22 May 2026 12:38:17 +0100
Tvrtko Ursulin <[email protected]> wrote:
Currently an unordered workqueue is used for the DRM scheduler which means
its concurrency is externally managed, and given there is one scheduler
instance per userspace queue, that means workqueue management logic is
within its rights to spawn many kernel threads to submit their respective
jobs.

Problem there is that all run job callbacks are serialized on the device
global mutex, making the potential thread storm just causing lock
contention.

Yeah, so initially we were not supposed to take the lock over the whole
run_job() function. We should normally be able to queue things to the
ring buffer, and only briefly take the lock to check if the context is
still resident and kick the group scheduler if it's not. I agree that
in practice it turned in a huge synchronization point. I guess we should
consider turning that mutex into a rwsem that's taken in write mode in
the tick path, and read-mode elsewhere.

There is some software state modified too, so I am not sure how easy or
hard it would be to make run job only hold the read lock?

Yeah, it's probably not as easy as it sounds.


If we add a separate ordered workqueue for the DRM scheduler integration
we can avoid this problem, since the ordered property directly expresses
the nature of the submission backend implementation.

I don't see alloc_ordered_workqueue() being used for the sched_wq
workqueue, is that intended (according to this comment, you seem to
want an ordered wq).

Yeah, it seems I mistyped the wq allocation below.

And considering the other user of this workqueue, the free job callback,
which is not globally serialized in this manner so could be thought to
potentially regress with this change, it should not be the case since
commit
a58f317c1ca0 ("drm/sched: Free all finished jobs at once")
made the DRM scheduler handle the cleanup of finished jobs more promptly.

Signed-off-by: Tvrtko Ursulin <[email protected]>
Cc: Boris Brezillon <[email protected]>
Cc: Liviu Dudau <[email protected]>
Cc: Steven Price <[email protected]>
---
   drivers/gpu/drm/panthor/panthor_sched.c | 27 ++++++++++++++++---------
   1 file changed, 17 insertions(+), 10 deletions(-)

diff --git a/drivers/gpu/drm/panthor/panthor_sched.c 
b/drivers/gpu/drm/panthor/panthor_sched.c
index 2bee1c92fb9e..cc6b3e2b015a 100644
--- a/drivers/gpu/drm/panthor/panthor_sched.c
+++ b/drivers/gpu/drm/panthor/panthor_sched.c
@@ -147,13 +147,11 @@ struct panthor_scheduler {
        struct panthor_device *ptdev;
/**
-        * @wq: Workqueue used by our internal scheduler logic and
-        * drm_gpu_scheduler.
+        * @wq: Workqueue used by our internal scheduler logic.
         *
         * Used for the scheduler tick, group update or other kind of FW
         * event processing that can't be handled in the threaded interrupt
-        * path. Also passed to the drm_gpu_scheduler instances embedded
-        * in panthor_queue.
+        * path.
         */
        struct workqueue_struct *wq;
@@ -166,6 +164,14 @@ struct panthor_scheduler {
         */
        struct workqueue_struct *heap_alloc_wq;
+ /**
+        * @sched_wq: Workqueue used for the DRM scheduler.
+        *
+        * Workqueue used for drm_gpu_scheduler instances embedded in
+        * panthor_queue.
+        */
+       struct workqueue_struct *sched_wq;
+
        /** @tick_work: Work executed on a scheduling tick. */
        struct delayed_work tick_work;
@@ -3488,7 +3494,7 @@ group_create_queue(struct panthor_group *group,
   {
        struct drm_sched_init_args sched_args = {
                .ops = &panthor_queue_sched_ops,
-               .submit_wq = group->ptdev->scheduler->wq,
+               .submit_wq = group->ptdev->scheduler->sched_wq,
                /*
                 * The credit limit argument tells us the total number of
                 * instructions across all CS slots in the ringbuffer, with
@@ -4078,6 +4084,9 @@ static void panthor_sched_fini(struct drm_device *ddev, 
void *res)
        if (sched->heap_alloc_wq)
                destroy_workqueue(sched->heap_alloc_wq);
+ if (sched->sched_wq)
+               destroy_workqueue(sched->sched_wq);
+
        for (prio = PANTHOR_CSG_PRIORITY_COUNT - 1; prio >= 0; prio--) {
                drm_WARN_ON(ddev, !list_empty(&sched->groups.runnable[prio]));
                drm_WARN_ON(ddev, !list_empty(&sched->groups.idle[prio]));
@@ -4167,13 +4176,11 @@ int panthor_sched_init(struct panthor_device *ptdev)
         * FW is smart enough to fall back on other methods if the kernel can't
         * allocate memory, and fail the tiling job if none of these
         * countermeasures worked.
-        *
-        * Set WQ_MEM_RECLAIM on sched->wq to unblock the situation when the
-        * system is running out of memory.
         */
        sched->heap_alloc_wq = alloc_workqueue("panthor-heap-alloc", 
WQ_UNBOUND, 0);
-       sched->wq = alloc_workqueue("panthor-csf-sched", WQ_MEM_RECLAIM | 
WQ_UNBOUND, 0);
-       if (!sched->wq || !sched->heap_alloc_wq) {
+       sched->wq = alloc_workqueue("panthor-csf-sched", WQ_UNBOUND, 0);
+       sched->sched_wq = alloc_workqueue("panthor-drm-sched", WQ_MEM_RECLAIM, 
0);

The other one also needs MEM_RECLAIM, because you need work items
queued to sched->wq to run to guarantee forward progress, and you need
to guarantee forward progress to reclaim GPU mem.

Ack, I had a feeling that might be the case.

I will respin next week or so. Or if you tell me the global lock can be
easily dropped from .run_job I can drop and forget about it. Wider
context is that I am experimenting with kthread_worker conversion and
trying to polish a
somewhat-broken-but-showing-great-latency-improvements branch.

Yep, I know, your experimental branch is actually on my list of things
to look at/test ;-).

Hold off a few days at least, there is a UAF in the branch Chia-I tested and I am in the process of reworking some details to eliminate that.

For that
I can kind of take either one global worker, or one worker per client
route for the RFC, no big deal either way for the prototype.

We probably want a worker per-prio (and possibly per-cpu), but certainly
not one per-client, or you'll end up with the thread explosion that was
addressed by the kthread -> workqueue transition.

Yeah I know, it is non-trivial, especially when different drivers are considered. There are multiple angles I am attacking that from and we can discuss it once I have something polished enough for sharing.

Just one of the things being the fact both panthor and I think xe currently serialize on the device global lock (per GuC CT lock in case of xe) so the whole point of per sched workers is questionable. I am yet to look at the other drivers to gain a fuller picture, but there is also VM BIND in panthor which is a problem.

Also, I didn't look at your panthor changes in this branch yet, but if
we're switching drm_sched to kthread workers, we probably want to
transition most existing panthor works to kthread_work, because some FW
events might need to processed for the GPU context to be unblocked, and
if we keep queuing those to a regular workqueue, they will be lagging
behing the HI_PRIO thread you have for HI_PRIO contexts.

Possibly yes. I glanced over some tick based machine so it is plausible that could be changed completely. For now the goal was to do a proof of concept and see if the numbers support it. And it seems the numbers are dramatically strong so I am cleaning it all up for an RFC.

Regards,

Tvrtko

Reply via email to