On 28/06/2021 08:42, Boris Brezillon wrote:
> Now that we can pass our own workqueue to drm_sched_init(), we can use
> an ordered workqueue on for both the scheduler timeout tdr and our own
> reset work (which we use when the reset is not caused by a fault/timeout
> on a specific job, like when we have AS_ACTIVE bit stuck). This
> guarantees that the timeout handlers and reset handler can't run
> concurrently which drastically simplifies the locking.
>
> v4:
> * Actually pass the reset workqueue to drm_sched_init()
> * Don't call cancel_work_sync() in panfrost_reset(). It will deadlock
> since it might be called from the reset work, which is executing and
> cancel_work_sync() will wait for the handler to return. Checking the
> reset pending status should avoid spurious resets
>
> v3:
> * New patch
>
> Suggested-by: Daniel Vetter <daniel.vetter@xxxxxxxx>
> Signed-off-by: Boris Brezillon <boris.brezillon@xxxxxxxxxxxxx>
While I'm very much not keen on the hack to drop out of the
dma_fence_signalling section - I don't believe this is introducing the
bug just making it more visible. What we really need is someone to find
time to actually fix up the lock ordering. But until then:
Reviewed-by: Steven Price <steven.price@xxxxxxx>
> ---
> drivers/gpu/drm/panfrost/panfrost_device.h | 6 +-
> drivers/gpu/drm/panfrost/panfrost_job.c | 187 ++++++++-------------
> 2 files changed, 72 insertions(+), 121 deletions(-)
>
> diff --git a/drivers/gpu/drm/panfrost/panfrost_device.h b/drivers/gpu/drm/panfrost/panfrost_device.h
> index f2190f90be75..59a487e8aba3 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_device.h
> +++ b/drivers/gpu/drm/panfrost/panfrost_device.h
> @@ -108,6 +108,7 @@ struct panfrost_device {
> struct mutex sched_lock;
>
> struct {
> + struct workqueue_struct *wq;
> struct work_struct work;
> atomic_t pending;
> } reset;
> @@ -246,9 +247,8 @@ const char *panfrost_exception_name(u32 exception_code);
> static inline void
> panfrost_device_schedule_reset(struct panfrost_device *pfdev)
> {
> - /* Schedule a reset if there's no reset in progress. */
> - if (!atomic_xchg(&pfdev->reset.pending, 1))
> - schedule_work(&pfdev->reset.work);
> + atomic_set(&pfdev->reset.pending, 1);
> + queue_work(pfdev->reset.wq, &pfdev->reset.work);
> }
>
> #endif
> diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c b/drivers/gpu/drm/panfrost/panfrost_job.c
> index e0c479e67304..98193a557a2d 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_job.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_job.c
> @@ -25,17 +25,8 @@
> #define job_write(dev, reg, data) writel(data, dev->iomem + (reg))
> #define job_read(dev, reg) readl(dev->iomem + (reg))
>
> -enum panfrost_queue_status {
> - PANFROST_QUEUE_STATUS_ACTIVE,
> - PANFROST_QUEUE_STATUS_STOPPED,
> - PANFROST_QUEUE_STATUS_STARTING,
> - PANFROST_QUEUE_STATUS_FAULT_PENDING,
> -};
> -
> struct panfrost_queue_state {
> struct drm_gpu_scheduler sched;
> - atomic_t status;
> - struct mutex lock;
> u64 fence_context;
> u64 emit_seqno;
> };
> @@ -379,57 +370,72 @@ void panfrost_job_enable_interrupts(struct panfrost_device *pfdev)
> job_write(pfdev, JOB_INT_MASK, irq_mask);
> }
>
> -static bool panfrost_scheduler_stop(struct panfrost_queue_state *queue,
> - struct drm_sched_job *bad)
> +static void panfrost_reset(struct panfrost_device *pfdev,
> + struct drm_sched_job *bad)
> {
> - enum panfrost_queue_status old_status;
> - bool stopped = false;
> + unsigned int i;
> + bool cookie;
>
> - mutex_lock(&queue->lock);
> - old_status = atomic_xchg(&queue->status,
> - PANFROST_QUEUE_STATUS_STOPPED);
> - if (old_status == PANFROST_QUEUE_STATUS_STOPPED)
> - goto out;
> + if (!atomic_read(&pfdev->reset.pending))
> + return;
> +
> + /* Stop the schedulers.
> + *
> + * FIXME: We temporarily get out of the dma_fence_signalling section
> + * because the cleanup path generate lockdep splats when taking locks
> + * to release job resources. We should rework the code to follow this
> + * pattern:
> + *
> + * try_lock
> + * if (locked)
> + * release
> + * else
> + * schedule_work_to_release_later
> + */
> + for (i = 0; i < NUM_JOB_SLOTS; i++)
> + drm_sched_stop(&pfdev->js->queue[i].sched, bad);
> +
> + cookie = dma_fence_begin_signalling();
>
> - WARN_ON(old_status != PANFROST_QUEUE_STATUS_ACTIVE);
> - drm_sched_stop(&queue->sched, bad);
> if (bad)
> drm_sched_increase_karma(bad);
>
> - stopped = true;
> + spin_lock(&pfdev->js->job_lock);
> + for (i = 0; i < NUM_JOB_SLOTS; i++) {
> + if (pfdev->jobs[i]) {
> + pm_runtime_put_noidle(pfdev->dev);
> + panfrost_devfreq_record_idle(&pfdev->pfdevfreq);
> + pfdev->jobs[i] = NULL;
> + }
> + }
> + spin_unlock(&pfdev->js->job_lock);
>
> - /*
> - * Set the timeout to max so the timer doesn't get started
> - * when we return from the timeout handler (restored in
> - * panfrost_scheduler_start()).
> + panfrost_device_reset(pfdev);
> +
> + /* GPU has been reset, we can cancel timeout/fault work that may have
> + * been queued in the meantime and clear the reset pending bit.
> */
> - queue->sched.timeout = MAX_SCHEDULE_TIMEOUT;
> + atomic_set(&pfdev->reset.pending, 0);
> + for (i = 0; i < NUM_JOB_SLOTS; i++)
> + cancel_delayed_work(&pfdev->js->queue[i].sched.work_tdr);
>
> -out:
> - mutex_unlock(&queue->lock);
>
> - return stopped;
> -}
> + /* Now resubmit jobs that were previously queued but didn't have a
> + * chance to finish.
> + * FIXME: We temporarily get out of the DMA fence signalling section
> + * while resubmitting jobs because the job submission logic will
> + * allocate memory with the GFP_KERNEL flag which can trigger memory
> + * reclaim and exposes a lock ordering issue.
> + */
> + dma_fence_end_signalling(cookie);
> + for (i = 0; i < NUM_JOB_SLOTS; i++)
> + drm_sched_resubmit_jobs(&pfdev->js->queue[i].sched);
> + cookie = dma_fence_begin_signalling();
>
> -static void panfrost_scheduler_start(struct panfrost_queue_state *queue)
> -{
> - enum panfrost_queue_status old_status;
> + for (i = 0; i < NUM_JOB_SLOTS; i++)
> + drm_sched_start(&pfdev->js->queue[i].sched, true);
>
> - mutex_lock(&queue->lock);
> - old_status = atomic_xchg(&queue->status,
> - PANFROST_QUEUE_STATUS_STARTING);
> - WARN_ON(old_status != PANFROST_QUEUE_STATUS_STOPPED);
> -
> - /* Restore the original timeout before starting the scheduler. */
> - queue->sched.timeout = msecs_to_jiffies(JOB_TIMEOUT_MS);
> - drm_sched_resubmit_jobs(&queue->sched);
> - drm_sched_start(&queue->sched, true);
> - old_status = atomic_xchg(&queue->status,
> - PANFROST_QUEUE_STATUS_ACTIVE);
> - if (old_status == PANFROST_QUEUE_STATUS_FAULT_PENDING)
> - drm_sched_fault(&queue->sched);
> -
> - mutex_unlock(&queue->lock);
> + dma_fence_end_signalling(cookie);
> }
>
> static enum drm_gpu_sched_stat panfrost_job_timedout(struct drm_sched_job
> @@ -454,11 +460,8 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct drm_sched_job
> job_read(pfdev, JS_TAIL_LO(js)),
> sched_job);
>
> - /* Scheduler is already stopped, nothing to do. */
> - if (!panfrost_scheduler_stop(&pfdev->js->queue[js], sched_job))
> - return DRM_GPU_SCHED_STAT_NOMINAL;
> -
> - panfrost_device_schedule_reset(pfdev);
> + atomic_set(&pfdev->reset.pending, 1);
> + panfrost_reset(pfdev, sched_job);
>
> return DRM_GPU_SCHED_STAT_NOMINAL;
> }
> @@ -485,8 +488,6 @@ static void panfrost_job_handle_irq(struct panfrost_device *pfdev, u32 status)
> job_write(pfdev, JOB_INT_CLEAR, mask);
>
> if (status & JOB_INT_MASK_ERR(j)) {
> - enum panfrost_queue_status old_status;
> -
> job_write(pfdev, JS_COMMAND_NEXT(j), JS_COMMAND_NOP);
>
> dev_err(pfdev->dev, "js fault, js=%d, status=%s, head=0x%x, tail=0x%x",
> @@ -494,19 +495,7 @@ static void panfrost_job_handle_irq(struct panfrost_device *pfdev, u32 status)
> panfrost_exception_name(job_read(pfdev, JS_STATUS(j))),
> job_read(pfdev, JS_HEAD_LO(j)),
> job_read(pfdev, JS_TAIL_LO(j)));
> -
> - /*
> - * When the queue is being restarted we don't report
> - * faults directly to avoid races between the timeout
> - * and reset handlers. panfrost_scheduler_start() will
> - * call drm_sched_fault() after the queue has been
> - * started if status == FAULT_PENDING.
> - */
> - old_status = atomic_cmpxchg(&pfdev->js->queue[j].status,
> - PANFROST_QUEUE_STATUS_STARTING,
> - PANFROST_QUEUE_STATUS_FAULT_PENDING);
> - if (old_status == PANFROST_QUEUE_STATUS_ACTIVE)
> - drm_sched_fault(&pfdev->js->queue[j].sched);
> + drm_sched_fault(&pfdev->js->queue[j].sched);
> }
>
> if (status & JOB_INT_MASK_DONE(j)) {
> @@ -561,56 +550,13 @@ static irqreturn_t panfrost_job_irq_handler(int irq, void *data)
> return IRQ_WAKE_THREAD;
> }
>
> -static void panfrost_reset(struct work_struct *work)
> +static void panfrost_reset_work(struct work_struct *work)
> {
> struct panfrost_device *pfdev = container_of(work,
> struct panfrost_device,
> reset.work);
> - unsigned int i;
> - bool cookie;
>
> - cookie = dma_fence_begin_signalling();
> - for (i = 0; i < NUM_JOB_SLOTS; i++) {
> - /*
> - * We want pending timeouts to be handled before we attempt
> - * to stop the scheduler. If we don't do that and the timeout
> - * handler is in flight, it might have removed the bad job
> - * from the list, and we'll lose this job if the reset handler
> - * enters the critical section in panfrost_scheduler_stop()
> - * before the timeout handler.
> - *
> - * Timeout is set to MAX_SCHEDULE_TIMEOUT - 1 because we need
> - * something big enough to make sure the timer will not expire
> - * before we manage to stop the scheduler, but we can't use
> - * MAX_SCHEDULE_TIMEOUT because drm_sched_get_cleanup_job()
> - * considers that as 'timer is not running' and will dequeue
> - * the job without making sure the timeout handler is not
> - * running.
> - */
> - pfdev->js->queue[i].sched.timeout = MAX_SCHEDULE_TIMEOUT - 1;
> - cancel_delayed_work_sync(&pfdev->js->queue[i].sched.work_tdr);
> - panfrost_scheduler_stop(&pfdev->js->queue[i], NULL);
> - }
> -
> - /* All timers have been stopped, we can safely reset the pending state. */
> - atomic_set(&pfdev->reset.pending, 0);
> -
> - spin_lock(&pfdev->js->job_lock);
> - for (i = 0; i < NUM_JOB_SLOTS; i++) {
> - if (pfdev->jobs[i]) {
> - pm_runtime_put_noidle(pfdev->dev);
> - panfrost_devfreq_record_idle(&pfdev->pfdevfreq);
> - pfdev->jobs[i] = NULL;
> - }
> - }
> - spin_unlock(&pfdev->js->job_lock);
> -
> - panfrost_device_reset(pfdev);
> -
> - for (i = 0; i < NUM_JOB_SLOTS; i++)
> - panfrost_scheduler_start(&pfdev->js->queue[i]);
> -
> - dma_fence_end_signalling(cookie);
> + panfrost_reset(pfdev, NULL);
> }
>
> int panfrost_job_init(struct panfrost_device *pfdev)
> @@ -618,7 +564,7 @@ int panfrost_job_init(struct panfrost_device *pfdev)
> struct panfrost_job_slot *js;
> int ret, j, irq;
>
> - INIT_WORK(&pfdev->reset.work, panfrost_reset);
> + INIT_WORK(&pfdev->reset.work, panfrost_reset_work);
>
> pfdev->js = js = devm_kzalloc(pfdev->dev, sizeof(*js), GFP_KERNEL);
> if (!js)
> @@ -640,15 +586,18 @@ int panfrost_job_init(struct panfrost_device *pfdev)
> return ret;
> }
>
> - for (j = 0; j < NUM_JOB_SLOTS; j++) {
> - mutex_init(&js->queue[j].lock);
> + pfdev->reset.wq = alloc_ordered_workqueue("panfrost-reset", 0);
> + if (!pfdev->reset.wq)
> + return -ENOMEM;
>
> + for (j = 0; j < NUM_JOB_SLOTS; j++) {
> js->queue[j].fence_context = dma_fence_context_alloc(1);
>
> ret = drm_sched_init(&js->queue[j].sched,
> &panfrost_sched_ops,
> 1, 0,
> - msecs_to_jiffies(JOB_TIMEOUT_MS), NULL,
> + msecs_to_jiffies(JOB_TIMEOUT_MS),
> + pfdev->reset.wq,
> NULL, "pan_js");
> if (ret) {
> dev_err(pfdev->dev, "Failed to create scheduler: %d.", ret);
> @@ -664,6 +613,7 @@ int panfrost_job_init(struct panfrost_device *pfdev)
> for (j--; j >= 0; j--)
> drm_sched_fini(&js->queue[j].sched);
>
> + destroy_workqueue(pfdev->reset.wq);
> return ret;
> }
>
> @@ -676,9 +626,10 @@ void panfrost_job_fini(struct panfrost_device *pfdev)
>
> for (j = 0; j < NUM_JOB_SLOTS; j++) {
> drm_sched_fini(&js->queue[j].sched);
> - mutex_destroy(&js->queue[j].lock);
> }
>
> + cancel_work_sync(&pfdev->reset.work);
> + destroy_workqueue(pfdev->reset.wq);
> }
>
> int panfrost_job_open(struct panfrost_file_priv *panfrost_priv)
>
