drm/etnaviv: use 'sync points' for performance monitor requests
authorChristian Gmeiner <christian.gmeiner@gmail.com>
Sun, 24 Sep 2017 13:15:30 +0000 (15:15 +0200)
committerLucas Stach <l.stach@pengutronix.de>
Tue, 10 Oct 2017 09:45:47 +0000 (11:45 +0200)
With 'sync points' we can sample the reqeustes perform signals
before and/or after the submited command buffer.

Changes v2 -> v3:
- fixed indentation and init nr_events to 1

Changes v4 -> v5:
- simplify logic around fence handling.

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
drivers/gpu/drm/etnaviv/etnaviv_gpu.c
drivers/gpu/drm/etnaviv/etnaviv_gpu.h

index 8aabdac8d00bf71c3ac2ad8bf7929b88deff21ee..56764b4b94aa2a3b9727dbcbcef98f4ebf62feb4 100644 (file)
@@ -1329,12 +1329,48 @@ 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)
+{
+       const struct etnaviv_cmdbuf *cmdbuf = event->cmdbuf;
+       unsigned int i;
+
+       for (i = 0; i < cmdbuf->nr_pmrs; i++) {
+               const struct etnaviv_perfmon_request *pmr = cmdbuf->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)
+{
+       const struct etnaviv_cmdbuf *cmdbuf = event->cmdbuf;
+       unsigned int i;
+
+       sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
+
+       for (i = 0; i < cmdbuf->nr_pmrs; i++) {
+               const struct etnaviv_perfmon_request *pmr = cmdbuf->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 i, nr_events = 1, event[3];
        int ret;
 
        ret = etnaviv_gpu_pm_get_sync(gpu);
@@ -1350,9 +1386,19 @@ int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
         *
         */
 
-       ret = event_alloc(gpu, 1, &event);
+       /*
+        * if there are performance monitor requests we need to have
+        * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
+        *   requests.
+        * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
+        *   and update the sequence number for userspace.
+        */
+       if (cmdbuf->nr_pmrs)
+               nr_events = 3;
+
+       ret = event_alloc(gpu, nr_events, event);
        if (ret) {
-               DRM_ERROR("no free event\n");
+               DRM_ERROR("no free events\n");
                goto out_pm_put;
        }
 
@@ -1360,12 +1406,14 @@ int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
 
        fence = etnaviv_gpu_fence_alloc(gpu);
        if (!fence) {
-               event_free(gpu, event);
+               for (i = 0; i < nr_events; i++)
+                       event_free(gpu, event[i]);
+
                ret = -ENOMEM;
                goto out_unlock;
        }
 
-       gpu->event[event].fence = fence;
+       gpu->event[event[0]].fence = fence;
        submit->fence = dma_fence_get(fence);
        gpu->active_fence = submit->fence->seqno;
 
@@ -1375,7 +1423,19 @@ int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
                gpu->lastctx = cmdbuf->ctx;
        }
 
-       etnaviv_buffer_queue(gpu, event, cmdbuf);
+       if (cmdbuf->nr_pmrs) {
+               gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
+               gpu->event[event[1]].cmdbuf = cmdbuf;
+               etnaviv_sync_point_queue(gpu, event[1]);
+       }
+
+       etnaviv_buffer_queue(gpu, event[0], cmdbuf);
+
+       if (cmdbuf->nr_pmrs) {
+               gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
+               gpu->event[event[2]].cmdbuf = cmdbuf;
+               etnaviv_sync_point_queue(gpu, event[2]);
+       }
 
        cmdbuf->fence = fence;
        list_add_tail(&cmdbuf->node, &gpu->active_cmd_list);
@@ -1481,6 +1541,9 @@ static irqreturn_t irq_handler(int irq, void *data)
                        }
 
                        fence = gpu->event[event].fence;
+                       if (!fence)
+                               continue;
+
                        gpu->event[event].fence = NULL;
                        dma_fence_signal(fence);
 
index 3be5cb53e89f62ebdc9756739a199483e20db5df..4f10f147297ab25cb233e0032e3042b862f98442 100644 (file)
@@ -89,6 +89,7 @@ struct etnaviv_chip_identity {
 
 struct etnaviv_event {
        struct dma_fence *fence;
+       struct etnaviv_cmdbuf *cmdbuf;
 
        void (*sync_point)(struct etnaviv_gpu *gpu, struct etnaviv_event *event);
 };