drm/etnaviv: move PMRs to submit object

To make them available to the event worker even after the actual
command stream execution has finished.
Signed-off-by: Lucas Stach <l.stach@pengutronix.de>

drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c

@@ -88,10 +88,9 @@ void etnaviv_cmdbuf_suballoc_destroy(struct etnaviv_cmdbuf_suballoc *suballoc)
 
 struct etnaviv_cmdbuf *
 etnaviv_cmdbuf_new(struct etnaviv_cmdbuf_suballoc *suballoc, u32 size,
-		   size_t nr_bos, size_t nr_pmrs)
+		   size_t nr_bos)
 {
 	struct etnaviv_cmdbuf *cmdbuf;
-	struct etnaviv_perfmon_request *pmrs;
 	size_t sz = size_vstruct(nr_bos, sizeof(cmdbuf->bo_map[0]),
 				 sizeof(*cmdbuf));
 	int granule_offs, order, ret;
@@ -100,12 +99,6 @@ etnaviv_cmdbuf_new(struct etnaviv_cmdbuf_suballoc *suballoc, u32 size,
 	if (!cmdbuf)
 		return NULL;
 
-	sz = sizeof(*pmrs) * nr_pmrs;
-	pmrs = kzalloc(sz, GFP_KERNEL);
-	if (!pmrs)
-		goto out_free_cmdbuf;
-
-	cmdbuf->pmrs = pmrs;
 	cmdbuf->suballoc = suballoc;
 	cmdbuf->size = size;
 
@@ -132,10 +125,6 @@ etnaviv_cmdbuf_new(struct etnaviv_cmdbuf_suballoc *suballoc, u32 size,
 	cmdbuf->vaddr = suballoc->vaddr + cmdbuf->suballoc_offset;
 
 	return cmdbuf;
-
-out_free_cmdbuf:
-	kfree(cmdbuf);
-	return NULL;
 }
 
 void etnaviv_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf)
@@ -151,7 +140,6 @@ void etnaviv_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf)
 	suballoc->free_space = 1;
 	mutex_unlock(&suballoc->lock);
 	wake_up_all(&suballoc->free_event);
-	kfree(cmdbuf->pmrs);
 	kfree(cmdbuf);
 }
 

drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h

@@ -39,9 +39,6 @@ struct etnaviv_cmdbuf {
 	u32 exec_state;
 	/* per GPU in-flight list */
 	struct list_head node;
-	/* perfmon requests */
-	unsigned int nr_pmrs;
-	struct etnaviv_perfmon_request *pmrs;
 	/* BOs attached to this command buffer */
 	unsigned int nr_bos;
 	struct etnaviv_vram_mapping *bo_map[0];
@@ -53,7 +50,7 @@ void etnaviv_cmdbuf_suballoc_destroy(struct etnaviv_cmdbuf_suballoc *suballoc);
 
 struct etnaviv_cmdbuf *
 etnaviv_cmdbuf_new(struct etnaviv_cmdbuf_suballoc *suballoc, u32 size,
-		   size_t nr_bos, size_t nr_pmrs);
+		   size_t nr_bos);
 void etnaviv_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf);
 
 u32 etnaviv_cmdbuf_get_va(struct etnaviv_cmdbuf *buf);

drivers/gpu/drm/etnaviv/etnaviv_gem.h

@@ -97,14 +97,15 @@ struct etnaviv_gem_submit_bo {
 
 /* Created per submit-ioctl, to track bo's and cmdstream bufs, etc,
  * associated with the cmdstream submission for synchronization (and
- * make it easier to unwind when things go wrong, etc).  This only
- * lasts for the duration of the submit-ioctl.
+ * make it easier to unwind when things go wrong, etc).
  */
 struct etnaviv_gem_submit {
 	struct kref refcount;
 	struct etnaviv_gpu *gpu;
 	struct dma_fence *out_fence, *in_fence;
 	u32 flags;
+	unsigned int nr_pmrs;
+	struct etnaviv_perfmon_request *pmrs;
 	unsigned int nr_bos;
 	struct etnaviv_gem_submit_bo bos[0];
 	/* No new members here, the previous one is variable-length! */

drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c

@@ -33,15 +33,23 @@
 #define BO_PINNED   0x2000
 
 static struct etnaviv_gem_submit *submit_create(struct drm_device *dev,
-		struct etnaviv_gpu *gpu, size_t nr)
+		struct etnaviv_gpu *gpu, size_t nr_bos, size_t nr_pmrs)
 {
 	struct etnaviv_gem_submit *submit;
-	size_t sz = size_vstruct(nr, sizeof(submit->bos[0]), sizeof(*submit));
+	size_t sz = size_vstruct(nr_bos, sizeof(submit->bos[0]), sizeof(*submit));
 
 	submit = kzalloc(sz, GFP_KERNEL);
 	if (!submit)
 		return NULL;
 
+	submit->pmrs = kcalloc(nr_pmrs, sizeof(struct etnaviv_perfmon_request),
+			       GFP_KERNEL);
+	if (!submit->pmrs) {
+		kfree(submit);
+		return NULL;
+	}
+	submit->nr_pmrs = nr_pmrs;
+
 	submit->gpu = gpu;
 	kref_init(&submit->refcount);
 
@@ -295,13 +303,11 @@ static int submit_reloc(struct etnaviv_gem_submit *submit, void *stream,
 }
 
 static int submit_perfmon_validate(struct etnaviv_gem_submit *submit,
-		struct etnaviv_cmdbuf *cmdbuf,
-		const struct drm_etnaviv_gem_submit_pmr *pmrs,
-		u32 nr_pms)
+		u32 exec_state, const struct drm_etnaviv_gem_submit_pmr *pmrs)
 {
 	u32 i;
 
-	for (i = 0; i < nr_pms; i++) {
+	for (i = 0; i < submit->nr_pmrs; i++) {
 		const struct drm_etnaviv_gem_submit_pmr *r = pmrs + i;
 		struct etnaviv_gem_submit_bo *bo;
 		int ret;
@@ -326,17 +332,17 @@ static int submit_perfmon_validate(struct etnaviv_gem_submit *submit,
 			return -EINVAL;
 		}
 
-		if (etnaviv_pm_req_validate(r, cmdbuf->exec_state)) {
+		if (etnaviv_pm_req_validate(r, exec_state)) {
 			DRM_ERROR("perfmon request: domain or signal not valid");
 			return -EINVAL;
 		}
 
-		cmdbuf->pmrs[i].flags = r->flags;
-		cmdbuf->pmrs[i].domain = r->domain;
-		cmdbuf->pmrs[i].signal = r->signal;
-		cmdbuf->pmrs[i].sequence = r->sequence;
-		cmdbuf->pmrs[i].offset = r->read_offset;
-		cmdbuf->pmrs[i].bo_vma = etnaviv_gem_vmap(&bo->obj->base);
+		submit->pmrs[i].flags = r->flags;
+		submit->pmrs[i].domain = r->domain;
+		submit->pmrs[i].signal = r->signal;
+		submit->pmrs[i].sequence = r->sequence;
+		submit->pmrs[i].offset = r->read_offset;
+		submit->pmrs[i].bo_vma = etnaviv_gem_vmap(&bo->obj->base);
 	}
 
 	return 0;
@@ -367,6 +373,7 @@ static void submit_cleanup(struct kref *kref)
 		dma_fence_put(submit->in_fence);
 	if (submit->out_fence)
 		dma_fence_put(submit->out_fence);
+	kfree(submit->pmrs);
 	kfree(submit);
 }
 
@@ -427,7 +434,7 @@ int etnaviv_ioctl_gem_submit(struct drm_device *dev, void *data,
 	stream = kvmalloc_array(1, args->stream_size, GFP_KERNEL);
 	cmdbuf = etnaviv_cmdbuf_new(gpu->cmdbuf_suballoc,
 				    ALIGN(args->stream_size, 8) + 8,
-				    args->nr_bos, args->nr_pmrs);
+				    args->nr_bos);
 	if (!bos || !relocs || !pmrs || !stream || !cmdbuf) {
 		ret = -ENOMEM;
 		goto err_submit_cmds;
@@ -456,7 +463,6 @@ int etnaviv_ioctl_gem_submit(struct drm_device *dev, void *data,
 		ret = -EFAULT;
 		goto err_submit_cmds;
 	}
-	cmdbuf->nr_pmrs = args->nr_pmrs;
 
 	ret = copy_from_user(stream, u64_to_user_ptr(args->stream),
 			     args->stream_size);
@@ -475,7 +481,7 @@ int etnaviv_ioctl_gem_submit(struct drm_device *dev, void *data,
 
 	ww_acquire_init(&ticket, &reservation_ww_class);
 
-	submit = submit_create(dev, gpu, args->nr_bos);
+	submit = submit_create(dev, gpu, args->nr_bos, args->nr_pmrs);
 	if (!submit) {
 		ret = -ENOMEM;
 		goto err_submit_ww_acquire;
@@ -518,7 +524,7 @@ int etnaviv_ioctl_gem_submit(struct drm_device *dev, void *data,
 	if (ret)
 		goto err_submit_objects;
 
-	ret = submit_perfmon_validate(submit, cmdbuf, pmrs, args->nr_pmrs);
+	ret = submit_perfmon_validate(submit, args->exec_state, pmrs);
 	if (ret)
 		goto err_submit_objects;
 

drivers/gpu/drm/etnaviv/etnaviv_gpu.c

@@ -717,7 +717,7 @@ int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
 	}
 
 	/* Create buffer: */
-	gpu->buffer = etnaviv_cmdbuf_new(gpu->cmdbuf_suballoc, PAGE_SIZE, 0, 0);
+	gpu->buffer = etnaviv_cmdbuf_new(gpu->cmdbuf_suballoc, PAGE_SIZE, 0);
 	if (!gpu->buffer) {
 		ret = -ENOMEM;
 		dev_err(gpu->dev, "could not create command buffer\n");
@@ -1317,11 +1317,11 @@ void etnaviv_gpu_pm_put(struct etnaviv_gpu *gpu)
 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
 	struct etnaviv_event *event, unsigned int flags)
 {
-	const struct etnaviv_cmdbuf *cmdbuf = event->cmdbuf;
+	const struct etnaviv_gem_submit *submit = event->submit;
 	unsigned int i;
 
-	for (i = 0; i < cmdbuf->nr_pmrs; i++) {
-		const struct etnaviv_perfmon_request *pmr = cmdbuf->pmrs + i;
+	for (i = 0; i < submit->nr_pmrs; i++) {
+		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
 
 		if (pmr->flags == flags)
 			etnaviv_perfmon_process(gpu, pmr);
@@ -1349,14 +1349,14 @@ static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
 	struct etnaviv_event *event)
 {
-	const struct etnaviv_cmdbuf *cmdbuf = event->cmdbuf;
+	const struct etnaviv_gem_submit *submit = event->submit;
 	unsigned int i;
 	u32 val;
 
 	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;
+	for (i = 0; i < submit->nr_pmrs; i++) {
+		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
 
 		*pmr->bo_vma = pmr->sequence;
 	}
@@ -1392,7 +1392,7 @@ int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
 	 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
 	 *   and update the sequence number for userspace.
 	 */
-	if (cmdbuf->nr_pmrs)
+	if (submit->nr_pmrs)
 		nr_events = 3;
 
 	ret = event_alloc(gpu, nr_events, event);
@@ -1416,17 +1416,19 @@ int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
 	submit->out_fence = dma_fence_get(fence);
 	gpu->active_fence = submit->out_fence->seqno;
 
-	if (cmdbuf->nr_pmrs) {
+	if (submit->nr_pmrs) {
 		gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
-		gpu->event[event[1]].cmdbuf = cmdbuf;
+		kref_get(&submit->refcount);
+		gpu->event[event[1]].submit = submit;
 		etnaviv_sync_point_queue(gpu, event[1]);
 	}
 
 	etnaviv_buffer_queue(gpu, event[0], cmdbuf);
 
-	if (cmdbuf->nr_pmrs) {
+	if (submit->nr_pmrs) {
 		gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
-		gpu->event[event[2]].cmdbuf = cmdbuf;
+		kref_get(&submit->refcount);
+		gpu->event[event[2]].submit = submit;
 		etnaviv_sync_point_queue(gpu, event[2]);
 	}
 
@@ -1465,6 +1467,7 @@ static void sync_point_worker(struct work_struct *work)
 	u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
 
 	event->sync_point(gpu, event);
+	etnaviv_submit_put(event->submit);
 	event_free(gpu, gpu->sync_point_event);
 
 	/* restart FE last to avoid GPU and IRQ racing against this worker */

drivers/gpu/drm/etnaviv/etnaviv_gpu.h

@@ -89,7 +89,7 @@ struct etnaviv_chip_identity {
 
 struct etnaviv_event {
 	struct dma_fence *fence;
-	struct etnaviv_cmdbuf *cmdbuf;
+	struct etnaviv_gem_submit *submit;
 
 	void (*sync_point)(struct etnaviv_gpu *gpu, struct etnaviv_event *event);
 };