On 06.10.2025 17:07, Steven Price wrote:
> On 01/10/2025 03:20, Adrián Larumbe wrote:
> > Avoid waiting for the DRM scheduler job timedout handler, and instead, let
> > the DRM scheduler core signal the error fence immediately when HW job
> > submission fails.
> >
> > That means we must also decrement the runtime-PM refcnt for the device,
> > because the job will never be enqueued or inflight.
> >
> > Reviewed-by: Boris Brezillon <[email protected]>
> > Signed-off-by: Adrián Larumbe <[email protected]>
> > ---
> > drivers/gpu/drm/panfrost/panfrost_job.c | 20 ++++++++++++++++----
> > 1 file changed, 16 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c
> > b/drivers/gpu/drm/panfrost/panfrost_job.c
> > index a0123d0a1b7d..3f60adc9b69d 100644
> > --- a/drivers/gpu/drm/panfrost/panfrost_job.c
> > +++ b/drivers/gpu/drm/panfrost/panfrost_job.c
> > @@ -196,7 +196,7 @@ panfrost_enqueue_job(struct panfrost_device *pfdev, int
> > slot,
> > return 1;
> > }
> >
> > -static void panfrost_job_hw_submit(struct panfrost_job *job, int js)
> > +static int panfrost_job_hw_submit(struct panfrost_job *job, int js)
> > {
> > struct panfrost_device *pfdev = job->pfdev;
> > unsigned int subslot;
> > @@ -208,10 +208,11 @@ static void panfrost_job_hw_submit(struct
> > panfrost_job *job, int js)
> >
> > ret = pm_runtime_get_sync(pfdev->base.dev);
> > if (ret < 0)
> > - return;
> > + goto err_hwsubmit;
> >
> > if (WARN_ON(job_read(pfdev, JS_COMMAND_NEXT(js)))) {
> > - return;
> > + ret = -EINVAL;
> > + goto err_hwsubmit;
> > }
> >
> > cfg = panfrost_mmu_as_get(pfdev, job->mmu);
> > @@ -262,6 +263,12 @@ static void panfrost_job_hw_submit(struct panfrost_job
> > *job, int js)
> > job, js, subslot, jc_head, cfg & 0xf);
> > }
> > spin_unlock(&pfdev->js->job_lock);
> > +
> > + return 0;
> > +
> > +err_hwsubmit:
> > + pm_runtime_put_autosuspend(pfdev->base.dev);
>
> I think you're missing something here. You've put a call to
> pm_runtime_put_autosuspend() here which matches the call to
> pm_runtime_get_sync() that we do earlier in the function. But there's no
> corresponding panfrost_devfreq_record_idle() (but the first thing this
> function does is panfrost_devfreq_record_busy()).
>
> So unless I'm missing something (very possible) then this is going to
> mess up the devfreq accounting. A simple fix would be just to move the
> panfrost_devfreq_record_busy() call down in the function.
You didn't miss anything, I completely forgot to keep the devfreq busy
count balanced after this change.
I've moved panfrost_devfreq_record_busy() right after the point the function
can no longer result in an error, as you suggested.
> Thanks,
> Steve
>
> > + return ret;
> > }
> >
> > static int panfrost_acquire_object_fences(struct drm_gem_object **bos,
> > @@ -384,6 +391,7 @@ static struct dma_fence *panfrost_job_run(struct
> > drm_sched_job *sched_job)
> > struct panfrost_device *pfdev = job->pfdev;
> > int slot = panfrost_job_get_slot(job);
> > struct dma_fence *fence = NULL;
> > + int ret;
> >
> > if (job->ctx->destroyed)
> > return ERR_PTR(-ECANCELED);
> > @@ -405,7 +413,11 @@ static struct dma_fence *panfrost_job_run(struct
> > drm_sched_job *sched_job)
> > dma_fence_put(job->done_fence);
> > job->done_fence = dma_fence_get(fence);
> >
> > - panfrost_job_hw_submit(job, slot);
> > + ret = panfrost_job_hw_submit(job, slot);
> > + if (ret) {
> > + dma_fence_put(fence);
> > + return ERR_PTR(ret);
> > + }
> >
> > return fence;
> > }
Adrian Larumbe
