[PATCH 10/21] drm/etnaviv: use 'sync points' for performance monitor requests

[Lucas Stach]

Am Freitag, den 09.06.2017, 12:26 +0200 schrieb Christian Gmeiner:
> With 'sync points' we can sample the reqeustes perform signals
> before and/or after the submited command buffer.
> 
> Signed-off-by: Christian Gmeiner <christian.gmeiner at gmail.com>
> ---
>  drivers/gpu/drm/etnaviv/etnaviv_gpu.c | 112
> +++++++++++++++++++++++++++++-----
>  drivers/gpu/drm/etnaviv/etnaviv_gpu.h |   4 ++
>  2 files changed, 102 insertions(+), 14 deletions(-)
> 
> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> index 0766861..2e9f031 100644
> --- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> +++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> @@ -1313,12 +1313,47 @@ void etnaviv_gpu_pm_put(struct etnaviv_gpu
> *gpu)
>  	pm_runtime_put_autosuspend(gpu->dev);
>  }
>  
> +static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
> +	struct etnaviv_event *event, unsigned int flags)
> +{
> +	unsigned int i;
> +
> +	for (i = 0; i < event->nr_pmrs; i++) {
> +		const struct etnaviv_perfmon_request *pmr = event-
> >pmrs + i;
> +
> +		if (pmr->flags == flags)
> +			etnaviv_perfmon_process(gpu, pmr);
> +	}
> +}
> +
> +static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
> +	struct etnaviv_event *event)
> +{
> +	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
> +}
> +
> +static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
> +	struct etnaviv_event *event)
> +{
> +	unsigned int i;
> +
> +	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
> +
> +	for (i = 0; i < event->nr_pmrs; i++) {
> +		const struct etnaviv_perfmon_request *pmr = event-
> >pmrs + i;
> +
> +		*pmr->bo_vma = pmr->sequence;
> +	}
> +}
> +
> +
>  /* add bo's to gpu's ring, and kick gpu: */
>  int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
>  	struct etnaviv_gem_submit *submit, struct etnaviv_cmdbuf
> *cmdbuf)
>  {
>  	struct dma_fence *fence;
>  	unsigned int event, i;
> +	unsigned int sync[2] = { ~0U, ~0U };
>  	int ret;
>  
>  	ret = etnaviv_gpu_pm_get_sync(gpu);
> @@ -1341,6 +1376,39 @@ int etnaviv_gpu_submit(struct etnaviv_gpu
> *gpu,
>  		goto out_pm_put;
>  	}
>  
> +	/*
> +	 * if there are performance monitor requests we need to have
> a sync point to
> +	 * re-configure gpu and process ETNA_PM_PROCESS_PRE
> requests.
> +	 */
> +	if (cmdbuf->nr_pmrs) {
> +		sync[0] = event_alloc(gpu);
> +
> +		if (unlikely(sync[0] == ~0U)) {
> +			DRM_ERROR("no free events for sync point
> 0\n");
> +			event_free(gpu, event);
> +			ret = -EBUSY;
> +			goto out_pm_put;
> +		}
> +	}
> +
> +	/*
> +	 * if there are performance monitor requests we need to have
> sync point to
> +	 * re-configure gpu, process ETNA_PM_PROCESS_POST requests
> and update the
> +	 * sequence number for userspace.
> +	 */
> +	if (cmdbuf->nr_pmrs) {
> +		sync[1] = event_alloc(gpu);
> +
> +		if (unlikely(sync[1] == ~0U)) {
> +			DRM_ERROR("no free events for sync point
> 1\n");
> +			event_free(gpu, event);
> +			if (unlikely(sync[0] == ~0U))
> +				event_free(gpu, sync[0]);
> +			ret = -EBUSY;
> +			goto out_pm_put;
> +		}
> +	}
> +

This is dangerous. We aren't holding the GPU lock at this point (and we
can't because of livelocks with the GPU hangchecker), so given enough
parallel submits with PMRs all the submits might abort as they can't
allocate enough events, as each one might hold one out of the available
events.

I think what we need here is to extend the event_alloc API to take the
number of events we need and grab them all at once under the event
spinlock.

>  	mutex_lock(&gpu->lock);
>  
>  	fence = etnaviv_gpu_fence_alloc(gpu);
> @@ -1360,8 +1428,22 @@ int etnaviv_gpu_submit(struct etnaviv_gpu
> *gpu,
>  		gpu->lastctx = cmdbuf->ctx;
>  	}
>  
> +	if (sync[0] != ~0U) {
> +		gpu->event[sync[0]].sync_point =
> &sync_point_perfmon_sample_pre;
> +		gpu->event[sync[0]].nr_pmrs = cmdbuf->nr_pmrs;
> +		gpu->event[sync[0]].pmrs = cmdbuf->pmrs;
> +		etnaviv_sync_point_queue(gpu, sync[0]);
> +	}
> +
>  	etnaviv_buffer_queue(gpu, event, cmdbuf);
>  
> +	if (sync[1] != ~0U) {
> +		gpu->event[sync[1]].sync_point =
> &sync_point_perfmon_sample_post;
> +		gpu->event[sync[1]].nr_pmrs = cmdbuf->nr_pmrs;
> +		gpu->event[sync[1]].pmrs = cmdbuf->pmrs;
> +		etnaviv_sync_point_queue(gpu, sync[1]);
> +	}
> +
>  	cmdbuf->fence = fence;
>  	list_add_tail(&cmdbuf->node, &gpu->active_cmd_list);
>  
> @@ -1455,20 +1537,22 @@ static irqreturn_t irq_handler(int irq, void
> *data)
>  				etnaviv_process_sync_point(gpu,
> &gpu->event[event]);
>  
>  			fence = gpu->event[event].fence;
> -			gpu->event[event].fence = NULL;
> -			dma_fence_signal(fence);
> -
> -			/*
> -			 * Events can be processed out of
> order.  Eg,
> -			 * - allocate and queue event 0
> -			 * - allocate event 1
> -			 * - event 0 completes, we process it
> -			 * - allocate and queue event 0
> -			 * - event 1 and event 0 complete
> -			 * we can end up processing event 0 first,
> then 1.
> -			 */
> -			if (fence_after(fence->seqno, gpu-
> >completed_fence))
> -				gpu->completed_fence = fence->seqno;
> +			if (fence) {
> +				gpu->event[event].fence = NULL;
> +				dma_fence_signal(fence);
> +
> +				/*
> +				 * Events can be processed out of
> order.  Eg,
> +				 * - allocate and queue event 0
> +				 * - allocate event 1
> +				 * - event 0 completes, we process
> it
> +				 * - allocate and queue event 0
> +				 * - event 1 and event 0 complete
> +				 * we can end up processing event 0
> first, then 1.
> +				 */
> +				if (fence_after(fence->seqno, gpu-
> >completed_fence))
> +					gpu->completed_fence =
> fence->seqno;
> +			}
>  
>  			event_free(gpu, event);
>  		}
> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
> b/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
> index fee6ed9..71375ab 100644
> --- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
> +++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
> @@ -92,6 +92,10 @@ struct etnaviv_event {
>  	struct dma_fence *fence;
>  
>  	void (*sync_point)(struct etnaviv_gpu *gpu, struct
> etnaviv_event *event);
> +
> +	/* performance monitor requests */
> +	unsigned int nr_pmrs;
> +	struct etnaviv_perfmon_request *pmrs;

This should be a pointer to the cmdbuf itself, so we don't copy the
information to various places.
>  };
>  
>  struct etnaviv_cmdbuf_suballoc;