Re: [PATCH] drm/xe: Limit number of jobs per exec queue

[Matthew Brost <matthew.brost@intel.com>]

On Thu, Oct 23, 2025 at 09:51:27AM -0600, Lin, Shuicheng wrote:
> On Wed, Oct 22, 2025 11:48 AM Matthew Brost wrote:
> > On Wed, Oct 22, 2025 at 06:10:37PM +0000, Shuicheng Lin wrote:
> > > Add a limit to the number of jobs that can be queued in a single exec
> > > queue to avoid potential resource exhaustion.
> > >
> > > A new field `job_cnt` is introduced in `struct xe_exec_queue` to track
> > > the number of active DRM jobs, along with a maximum limit
> > > `XE_MAX_JOB_COUNT_PER_EXEC_QUEUE` set to 0x1000.
> > >
> > > If the job count exceeds this threshold, `xe_exec_ioctl()` now returns
> > > `-EAGAIN` to signal that the caller should retry later.
> > >
> > > Suggested-by: Matthew Brost <matthew.brost@intel.com>
> > > Signed-off-by: Shuicheng Lin <shuicheng.lin@intel.com>
> > > ---
> > >  drivers/gpu/drm/xe/xe_exec.c             | 5 +++++
> > >  drivers/gpu/drm/xe/xe_exec_queue_types.h | 5 +++++
> > >  drivers/gpu/drm/xe/xe_sched_job.c        | 2 ++
> > >  3 files changed, 12 insertions(+)
> > >
> > > diff --git a/drivers/gpu/drm/xe/xe_exec.c
> > > b/drivers/gpu/drm/xe/xe_exec.c index 0dc27476832b..722a5ac0200a
> > 100644
> > > --- a/drivers/gpu/drm/xe/xe_exec.c
> > > +++ b/drivers/gpu/drm/xe/xe_exec.c
> > > @@ -154,6 +154,11 @@ int xe_exec_ioctl(struct drm_device *dev, void
> > *data, struct drm_file *file)
> > >  		goto err_exec_queue;
> > >  	}
> > >
> > > +	if (q->job_cnt >= XE_MAX_JOB_COUNT_PER_EXEC_QUEUE) {
> > > +		err = -EAGAIN;
> > 
> > How about an ftrace point if this occurs? Should help us detect if users
> > somehow hit this during normal operation.
> 
> How about use " xe_err_once" log to detect it? err log in dmesg is more noticeable.
> 

It is not an error. -EAGAIN is a specific return code in DRM that means
"try again." We want to avoid returning -EAGAIN to well-behaved
applications, while still being able to detect it. At the same time, we
want to prevent misbehaving applications from spamming dmesg.

Matt

> > 
> > > +		goto err_exec_queue;
> > > +	}
> > > +
> > >  	if (args->num_syncs) {
> > >  		syncs = kcalloc(args->num_syncs, sizeof(*syncs),
> > GFP_KERNEL);
> > >  		if (!syncs) {
> > > diff --git a/drivers/gpu/drm/xe/xe_exec_queue_types.h
> > > b/drivers/gpu/drm/xe/xe_exec_queue_types.h
> > > index 282505fa1377..5f3219acec3c 100644
> > > --- a/drivers/gpu/drm/xe/xe_exec_queue_types.h
> > > +++ b/drivers/gpu/drm/xe/xe_exec_queue_types.h
> > > @@ -162,6 +162,11 @@ struct xe_exec_queue {
> > >  	const struct xe_ring_ops *ring_ops;
> > >  	/** @entity: DRM sched entity for this exec queue (1 to 1 relationship)
> > */
> > >  	struct drm_sched_entity *entity;
> > > +
> > > +#define XE_MAX_JOB_COUNT_PER_EXEC_QUEUE	0x1000
> > 
> > Hmm, I'm trying to think of a reasonable limit. 4K seems high. Maybe we
> > should check with the UMDs to understand the deepest reasonable pipeline of
> > jobs they build — to avoid blocking them here while still limiting the DoS
> > surface. My initial thought that 512 or 1k should be a reasonable limit for now.
> > Can always adjust going forward.
> 
> Yes. Let me change it to 1k. And adjust it later if needed.
> 
> > 
> > > +
> > > +	/** @job_cnt: number of drm jobs in this exec queue */
> > > +	u32 job_cnt;
> > 
> > This needs to be atomic. xe_sched_job_destroy can run in parallel with the
> > exec IOCTL.
> 
> Thanks. Yes, it should be atomic.
> 
> Shuicheng
> 
> > 
> > Everything else looks correct.
> > 
> > Matt
> > 
> > >  	/**
> > >  	 * @tlb_flush_seqno: The seqno of the last rebind tlb flush performed
> > >  	 * Protected by @vm's resv. Unused if @vm == NULL.
> > > diff --git a/drivers/gpu/drm/xe/xe_sched_job.c
> > > b/drivers/gpu/drm/xe/xe_sched_job.c
> > > index d21bf8f26964..37f58be7cbcc 100644
> > > --- a/drivers/gpu/drm/xe/xe_sched_job.c
> > > +++ b/drivers/gpu/drm/xe/xe_sched_job.c
> > > @@ -146,6 +146,7 @@ struct xe_sched_job *xe_sched_job_create(struct
> > xe_exec_queue *q,
> > >  	for (i = 0; i < width; ++i)
> > >  		job->ptrs[i].batch_addr = batch_addr[i];
> > >
> > > +	q->job_cnt++;
> > >  	xe_pm_runtime_get_noresume(job_to_xe(job));
> > >  	trace_xe_sched_job_create(job);
> > >  	return job;
> > > @@ -177,6 +178,7 @@ void xe_sched_job_destroy(struct kref *ref)
> > >  	dma_fence_put(job->fence);
> > >  	drm_sched_job_cleanup(&job->drm);
> > >  	job_free(job);
> > > +	q->job_cnt--;
> > >  	xe_exec_queue_put(q);
> > >  	xe_pm_runtime_put(xe);
> > >  }
> > > --
> > > 2.49.0
> > >