Merge branch 'drm-etnaviv-next' of git://git.pengutronix.de/git/lst/linux into drm-next

Notable changes:
- correctness fixes to the GPU cache flushing when switching execution
state and when powering down the GPU
- reduction of time spent in hardirq-off context
- placement improvements to the GPU DMA linear window, allowing the
driver to properly work on i.MX6 systems with more than 2GB of RAM

* 'drm-etnaviv-next' of git://git.pengutronix.de/git/lst/linux:
  drm: etnaviv: clean up submit_bo()
  drm: etnaviv: clean up vram_mapping submission/retire path
  drm: etnaviv: improve readability of command insertion to ring buffer
  drm: etnaviv: clean up GPU command submission
  drm: etnaviv: use previous GPU pipe state when pipe switching
  drm: etnaviv: flush all GPU caches when stopping GPU
  drm: etnaviv: track current execution state
  drm: etnaviv: extract arming of semaphore
  drm: etnaviv: extract replacement of WAIT command
  drm: etnaviv: extract command ring reservation
  drm/etnaviv: move GPU linear window to end of DMA window
  drm/etnaviv: move runtime PM balance into retire worker

drivers/gpu/drm/etnaviv/etnaviv_buffer.c

@@ -21,6 +21,7 @@
 
 #include "common.xml.h"
 #include "state.xml.h"
+#include "state_3d.xml.h"
 #include "cmdstream.xml.h"
 
 /*
@@ -85,10 +86,17 @@ static inline void CMD_STALL(struct etnaviv_cmdbuf *buffer,
 	OUT(buffer, VIV_FE_STALL_TOKEN_FROM(from) | VIV_FE_STALL_TOKEN_TO(to));
 }
 
-static void etnaviv_cmd_select_pipe(struct etnaviv_cmdbuf *buffer, u8 pipe)
+static inline void CMD_SEM(struct etnaviv_cmdbuf *buffer, u32 from, u32 to)
 {
-	u32 flush;
-	u32 stall;
+	CMD_LOAD_STATE(buffer, VIVS_GL_SEMAPHORE_TOKEN,
+		       VIVS_GL_SEMAPHORE_TOKEN_FROM(from) |
+		       VIVS_GL_SEMAPHORE_TOKEN_TO(to));
+}
+
+static void etnaviv_cmd_select_pipe(struct etnaviv_gpu *gpu,
+	struct etnaviv_cmdbuf *buffer, u8 pipe)
+{
+	u32 flush = 0;
 
 	/*
 	 * This assumes that if we're switching to 2D, we're switching
@@ -96,17 +104,13 @@ static void etnaviv_cmd_select_pipe(struct etnaviv_cmdbuf *buffer, u8 pipe)
 	 * the 2D core, we need to flush the 3D depth and color caches,
 	 * otherwise we need to flush the 2D pixel engine cache.
 	 */
-	if (pipe == ETNA_PIPE_2D)
-		flush = VIVS_GL_FLUSH_CACHE_DEPTH | VIVS_GL_FLUSH_CACHE_COLOR;
-	else
+	if (gpu->exec_state == ETNA_PIPE_2D)
 		flush = VIVS_GL_FLUSH_CACHE_PE2D;
-
-	stall = VIVS_GL_SEMAPHORE_TOKEN_FROM(SYNC_RECIPIENT_FE) |
-		VIVS_GL_SEMAPHORE_TOKEN_TO(SYNC_RECIPIENT_PE);
+	else if (gpu->exec_state == ETNA_PIPE_3D)
+		flush = VIVS_GL_FLUSH_CACHE_DEPTH | VIVS_GL_FLUSH_CACHE_COLOR;
 
 	CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE, flush);
-	CMD_LOAD_STATE(buffer, VIVS_GL_SEMAPHORE_TOKEN, stall);
-
+	CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 	CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 
 	CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
@@ -131,6 +135,36 @@ static void etnaviv_buffer_dump(struct etnaviv_gpu *gpu,
 			ptr, len * 4, 0);
 }
 
+/*
+ * Safely replace the WAIT of a waitlink with a new command and argument.
+ * The GPU may be executing this WAIT while we're modifying it, so we have
+ * to write it in a specific order to avoid the GPU branching to somewhere
+ * else.  'wl_offset' is the offset to the first byte of the WAIT command.
+ */
+static void etnaviv_buffer_replace_wait(struct etnaviv_cmdbuf *buffer,
+	unsigned int wl_offset, u32 cmd, u32 arg)
+{
+	u32 *lw = buffer->vaddr + wl_offset;
+
+	lw[1] = arg;
+	mb();
+	lw[0] = cmd;
+	mb();
+}
+
+/*
+ * Ensure that there is space in the command buffer to contiguously write
+ * 'cmd_dwords' 64-bit words into the buffer, wrapping if necessary.
+ */
+static u32 etnaviv_buffer_reserve(struct etnaviv_gpu *gpu,
+	struct etnaviv_cmdbuf *buffer, unsigned int cmd_dwords)
+{
+	if (buffer->user_size + cmd_dwords * sizeof(u64) > buffer->size)
+		buffer->user_size = 0;
+
+	return gpu_va(gpu, buffer) + buffer->user_size;
+}
+
 u16 etnaviv_buffer_init(struct etnaviv_gpu *gpu)
 {
 	struct etnaviv_cmdbuf *buffer = gpu->buffer;
@@ -147,81 +181,79 @@ u16 etnaviv_buffer_init(struct etnaviv_gpu *gpu)
 void etnaviv_buffer_end(struct etnaviv_gpu *gpu)
 {
 	struct etnaviv_cmdbuf *buffer = gpu->buffer;
+	unsigned int waitlink_offset = buffer->user_size - 16;
+	u32 link_target, flush = 0;
 
-	/* Replace the last WAIT with an END */
-	buffer->user_size -= 16;
-
-	CMD_END(buffer);
-	mb();
+	if (gpu->exec_state == ETNA_PIPE_2D)
+		flush = VIVS_GL_FLUSH_CACHE_PE2D;
+	else if (gpu->exec_state == ETNA_PIPE_3D)
+		flush = VIVS_GL_FLUSH_CACHE_DEPTH |
+			VIVS_GL_FLUSH_CACHE_COLOR |
+			VIVS_GL_FLUSH_CACHE_TEXTURE |
+			VIVS_GL_FLUSH_CACHE_TEXTUREVS |
+			VIVS_GL_FLUSH_CACHE_SHADER_L2;
+
+	if (flush) {
+		unsigned int dwords = 7;
+
+		link_target = etnaviv_buffer_reserve(gpu, buffer, dwords);
+
+		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
+		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
+		CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE, flush);
+		if (gpu->exec_state == ETNA_PIPE_3D)
+			CMD_LOAD_STATE(buffer, VIVS_TS_FLUSH_CACHE,
+				       VIVS_TS_FLUSH_CACHE_FLUSH);
+		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
+		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
+		CMD_END(buffer);
+
+		etnaviv_buffer_replace_wait(buffer, waitlink_offset,
+					    VIV_FE_LINK_HEADER_OP_LINK |
+					    VIV_FE_LINK_HEADER_PREFETCH(dwords),
+					    link_target);
+	} else {
+		/* Replace the last link-wait with an "END" command */
+		etnaviv_buffer_replace_wait(buffer, waitlink_offset,
+					    VIV_FE_END_HEADER_OP_END, 0);
+	}
 }
 
+/* Append a command buffer to the ring buffer. */
 void etnaviv_buffer_queue(struct etnaviv_gpu *gpu, unsigned int event,
 	struct etnaviv_cmdbuf *cmdbuf)
 {
 	struct etnaviv_cmdbuf *buffer = gpu->buffer;
-	u32 *lw = buffer->vaddr + buffer->user_size - 16;
-	u32 back, link_target, link_size, reserve_size, extra_size = 0;
+	unsigned int waitlink_offset = buffer->user_size - 16;
+	u32 return_target, return_dwords;
+	u32 link_target, link_dwords;
 
 	if (drm_debug & DRM_UT_DRIVER)
 		etnaviv_buffer_dump(gpu, buffer, 0, 0x50);
 
+	link_target = gpu_va(gpu, cmdbuf);
+	link_dwords = cmdbuf->size / 8;
+
 	/*
-	 * If we need to flush the MMU prior to submitting this buffer, we
-	 * will need to append a mmu flush load state, followed by a new
+	 * If we need maintanence prior to submitting this buffer, we will
+	 * need to append a mmu flush load state, followed by a new
 	 * link to this buffer - a total of four additional words.
 	 */
 	if (gpu->mmu->need_flush || gpu->switch_context) {
+		u32 target, extra_dwords;
+
 		/* link command */
-		extra_size += 2;
+		extra_dwords = 1;
+
 		/* flush command */
 		if (gpu->mmu->need_flush)
-			extra_size += 2;
+			extra_dwords += 1;
+
 		/* pipe switch commands */
 		if (gpu->switch_context)
-			extra_size += 8;
-	}
+			extra_dwords += 4;
 
-	reserve_size = (6 + extra_size) * 4;
-
-	/*
-	 * if we are going to completely overflow the buffer, we need to wrap.
-	 */
-	if (buffer->user_size + reserve_size > buffer->size)
-		buffer->user_size = 0;
-
-	/* save offset back into main buffer */
-	back = buffer->user_size + reserve_size - 6 * 4;
-	link_target = gpu_va(gpu, buffer) + buffer->user_size;
-	link_size = 6;
-
-	/* Skip over any extra instructions */
-	link_target += extra_size * sizeof(u32);
-
-	if (drm_debug & DRM_UT_DRIVER)
-		pr_info("stream link to 0x%08x @ 0x%08x %p\n",
-			link_target, gpu_va(gpu, cmdbuf), cmdbuf->vaddr);
-
-	/* jump back from cmd to main buffer */
-	CMD_LINK(cmdbuf, link_size, link_target);
-
-	link_target = gpu_va(gpu, cmdbuf);
-	link_size = cmdbuf->size / 8;
-
-
-
-	if (drm_debug & DRM_UT_DRIVER) {
-		print_hex_dump(KERN_INFO, "cmd ", DUMP_PREFIX_OFFSET, 16, 4,
-			       cmdbuf->vaddr, cmdbuf->size, 0);
-
-		pr_info("link op: %p\n", lw);
-		pr_info("link addr: %p\n", lw + 1);
-		pr_info("addr: 0x%08x\n", link_target);
-		pr_info("back: 0x%08x\n", gpu_va(gpu, buffer) + back);
-		pr_info("event: %d\n", event);
-	}
-
-	if (gpu->mmu->need_flush || gpu->switch_context) {
-		u32 new_target = gpu_va(gpu, buffer) + buffer->user_size;
+		target = etnaviv_buffer_reserve(gpu, buffer, extra_dwords);
 
 		if (gpu->mmu->need_flush) {
 			/* Add the MMU flush */
@@ -236,32 +268,59 @@ void etnaviv_buffer_queue(struct etnaviv_gpu *gpu, unsigned int event,
 		}
 
 		if (gpu->switch_context) {
-			etnaviv_cmd_select_pipe(buffer, cmdbuf->exec_state);
+			etnaviv_cmd_select_pipe(gpu, buffer, cmdbuf->exec_state);
+			gpu->exec_state = cmdbuf->exec_state;
 			gpu->switch_context = false;
 		}
 
-		/* And the link to the first buffer */
-		CMD_LINK(buffer, link_size, link_target);
+		/* And the link to the submitted buffer */
+		CMD_LINK(buffer, link_dwords, link_target);
 
 		/* Update the link target to point to above instructions */
-		link_target = new_target;
-		link_size = extra_size;
+		link_target = target;
+		link_dwords = extra_dwords;
 	}
 
-	/* trigger event */
+	/*
+	 * Append a LINK to the submitted command buffer to return to
+	 * the ring buffer.  return_target is the ring target address.
+	 * We need three dwords: event, wait, link.
+	 */
+	return_dwords = 3;
+	return_target = etnaviv_buffer_reserve(gpu, buffer, return_dwords);
+	CMD_LINK(cmdbuf, return_dwords, return_target);
+
+	/*
+	 * Append event, wait and link pointing back to the wait
+	 * command to the ring buffer.
+	 */
 	CMD_LOAD_STATE(buffer, VIVS_GL_EVENT, VIVS_GL_EVENT_EVENT_ID(event) |
 		       VIVS_GL_EVENT_FROM_PE);
-
-	/* append WAIT/LINK to main buffer */
 	CMD_WAIT(buffer);
-	CMD_LINK(buffer, 2, gpu_va(gpu, buffer) + (buffer->user_size - 4));
+	CMD_LINK(buffer, 2, return_target + 8);
 
-	/* Change WAIT into a LINK command; write the address first. */
-	*(lw + 1) = link_target;
-	mb();
-	*(lw) = VIV_FE_LINK_HEADER_OP_LINK |
-		VIV_FE_LINK_HEADER_PREFETCH(link_size);
-	mb();
+	if (drm_debug & DRM_UT_DRIVER)
+		pr_info("stream link to 0x%08x @ 0x%08x %p\n",
+			return_target, gpu_va(gpu, cmdbuf), cmdbuf->vaddr);
+
+	if (drm_debug & DRM_UT_DRIVER) {
+		print_hex_dump(KERN_INFO, "cmd ", DUMP_PREFIX_OFFSET, 16, 4,
+			       cmdbuf->vaddr, cmdbuf->size, 0);
+
+		pr_info("link op: %p\n", buffer->vaddr + waitlink_offset);
+		pr_info("addr: 0x%08x\n", link_target);
+		pr_info("back: 0x%08x\n", return_target);
+		pr_info("event: %d\n", event);
+	}
+
+	/*
+	 * Kick off the submitted command by replacing the previous
+	 * WAIT with a link to the address in the ring buffer.
+	 */
+	etnaviv_buffer_replace_wait(buffer, waitlink_offset,
+				    VIV_FE_LINK_HEADER_OP_LINK |
+				    VIV_FE_LINK_HEADER_PREFETCH(link_dwords),
+				    link_target);
 
 	if (drm_debug & DRM_UT_DRIVER)
 		etnaviv_buffer_dump(gpu, buffer, 0, 0x50);

drivers/gpu/drm/etnaviv/etnaviv_drv.h

@@ -75,9 +75,6 @@ int etnaviv_ioctl_gem_submit(struct drm_device *dev, void *data,
 int etnaviv_gem_mmap(struct file *filp, struct vm_area_struct *vma);
 int etnaviv_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf);
 int etnaviv_gem_mmap_offset(struct drm_gem_object *obj, u64 *offset);
-int etnaviv_gem_get_iova(struct etnaviv_gpu *gpu,
-	struct drm_gem_object *obj, u32 *iova);
-void etnaviv_gem_put_iova(struct etnaviv_gpu *gpu, struct drm_gem_object *obj);
 struct sg_table *etnaviv_gem_prime_get_sg_table(struct drm_gem_object *obj);
 void *etnaviv_gem_prime_vmap(struct drm_gem_object *obj);
 void etnaviv_gem_prime_vunmap(struct drm_gem_object *obj, void *vaddr);

drivers/gpu/drm/etnaviv/etnaviv_gem.c

@@ -260,8 +260,32 @@ etnaviv_gem_get_vram_mapping(struct etnaviv_gem_object *obj,
 	return NULL;
 }
 
-int etnaviv_gem_get_iova(struct etnaviv_gpu *gpu,
-	struct drm_gem_object *obj, u32 *iova)
+void etnaviv_gem_mapping_reference(struct etnaviv_vram_mapping *mapping)
+{
+	struct etnaviv_gem_object *etnaviv_obj = mapping->object;
+
+	drm_gem_object_reference(&etnaviv_obj->base);
+
+	mutex_lock(&etnaviv_obj->lock);
+	WARN_ON(mapping->use == 0);
+	mapping->use += 1;
+	mutex_unlock(&etnaviv_obj->lock);
+}
+
+void etnaviv_gem_mapping_unreference(struct etnaviv_vram_mapping *mapping)
+{
+	struct etnaviv_gem_object *etnaviv_obj = mapping->object;
+
+	mutex_lock(&etnaviv_obj->lock);
+	WARN_ON(mapping->use == 0);
+	mapping->use -= 1;
+	mutex_unlock(&etnaviv_obj->lock);
+
+	drm_gem_object_unreference_unlocked(&etnaviv_obj->base);
+}
+
+struct etnaviv_vram_mapping *etnaviv_gem_mapping_get(
+	struct drm_gem_object *obj, struct etnaviv_gpu *gpu)
 {
 	struct etnaviv_gem_object *etnaviv_obj = to_etnaviv_bo(obj);
 	struct etnaviv_vram_mapping *mapping;
@@ -329,28 +353,12 @@ int etnaviv_gem_get_iova(struct etnaviv_gpu *gpu,
 out:
 	mutex_unlock(&etnaviv_obj->lock);
 
-	if (!ret) {
-		/* Take a reference on the object */
-		drm_gem_object_reference(obj);
-		*iova = mapping->iova;
-	}
-
-	return ret;
-}
-
-void etnaviv_gem_put_iova(struct etnaviv_gpu *gpu, struct drm_gem_object *obj)
-{
-	struct etnaviv_gem_object *etnaviv_obj = to_etnaviv_bo(obj);
-	struct etnaviv_vram_mapping *mapping;
-
-	mutex_lock(&etnaviv_obj->lock);
-	mapping = etnaviv_gem_get_vram_mapping(etnaviv_obj, gpu->mmu);
-
-	WARN_ON(mapping->use == 0);
-	mapping->use -= 1;
-	mutex_unlock(&etnaviv_obj->lock);
+	if (ret)
+		return ERR_PTR(ret);
 
-	drm_gem_object_unreference_unlocked(obj);
+	/* Take a reference on the object */
+	drm_gem_object_reference(obj);
+	return mapping;
 }
 
 void *etnaviv_gem_vmap(struct drm_gem_object *obj)

drivers/gpu/drm/etnaviv/etnaviv_gem.h

@@ -88,6 +88,12 @@ static inline bool is_active(struct etnaviv_gem_object *etnaviv_obj)
 
 #define MAX_CMDS 4
 
+struct etnaviv_gem_submit_bo {
+	u32 flags;
+	struct etnaviv_gem_object *obj;
+	struct etnaviv_vram_mapping *mapping;
+};
+
 /* 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
@@ -99,11 +105,7 @@ struct etnaviv_gem_submit {
 	struct ww_acquire_ctx ticket;
 	u32 fence;
 	unsigned int nr_bos;
-	struct {
-		u32 flags;
-		struct etnaviv_gem_object *obj;
-		u32 iova;
-	} bos[0];
+	struct etnaviv_gem_submit_bo bos[0];
 };
 
 int etnaviv_gem_wait_bo(struct etnaviv_gpu *gpu, struct drm_gem_object *obj,
@@ -115,4 +117,9 @@ int etnaviv_gem_obj_add(struct drm_device *dev, struct drm_gem_object *obj);
 struct page **etnaviv_gem_get_pages(struct etnaviv_gem_object *obj);
 void etnaviv_gem_put_pages(struct etnaviv_gem_object *obj);
 
+struct etnaviv_vram_mapping *etnaviv_gem_mapping_get(
+	struct drm_gem_object *obj, struct etnaviv_gpu *gpu);
+void etnaviv_gem_mapping_reference(struct etnaviv_vram_mapping *mapping);
+void etnaviv_gem_mapping_unreference(struct etnaviv_vram_mapping *mapping);
+
 #endif /* __ETNAVIV_GEM_H__ */

drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c

@@ -187,12 +187,10 @@ static void submit_unpin_objects(struct etnaviv_gem_submit *submit)
 	int i;
 
 	for (i = 0; i < submit->nr_bos; i++) {
-		struct etnaviv_gem_object *etnaviv_obj = submit->bos[i].obj;
-
 		if (submit->bos[i].flags & BO_PINNED)
-			etnaviv_gem_put_iova(submit->gpu, &etnaviv_obj->base);
+			etnaviv_gem_mapping_unreference(submit->bos[i].mapping);
 
-		submit->bos[i].iova = 0;
+		submit->bos[i].mapping = NULL;
 		submit->bos[i].flags &= ~BO_PINNED;
 	}
 }
@@ -203,22 +201,24 @@ static int submit_pin_objects(struct etnaviv_gem_submit *submit)
 
 	for (i = 0; i < submit->nr_bos; i++) {
 		struct etnaviv_gem_object *etnaviv_obj = submit->bos[i].obj;
-		u32 iova;
+		struct etnaviv_vram_mapping *mapping;
 
-		ret = etnaviv_gem_get_iova(submit->gpu, &etnaviv_obj->base,
-					   &iova);
-		if (ret)
+		mapping = etnaviv_gem_mapping_get(&etnaviv_obj->base,
+						  submit->gpu);
+		if (IS_ERR(mapping)) {
+			ret = PTR_ERR(mapping);
 			break;
+		}
 
 		submit->bos[i].flags |= BO_PINNED;
-		submit->bos[i].iova = iova;
+		submit->bos[i].mapping = mapping;
 	}
 
 	return ret;
 }
 
 static int submit_bo(struct etnaviv_gem_submit *submit, u32 idx,
-		struct etnaviv_gem_object **obj, u32 *iova)
+	struct etnaviv_gem_submit_bo **bo)
 {
 	if (idx >= submit->nr_bos) {
 		DRM_ERROR("invalid buffer index: %u (out of %u)\n",
@@ -226,10 +226,7 @@ static int submit_bo(struct etnaviv_gem_submit *submit, u32 idx,
 		return -EINVAL;
 	}
 
-	if (obj)
-		*obj = submit->bos[idx].obj;
-	if (iova)
-		*iova = submit->bos[idx].iova;
+	*bo = &submit->bos[idx];
 
 	return 0;
 }
@@ -245,8 +242,8 @@ static int submit_reloc(struct etnaviv_gem_submit *submit, void *stream,
 
 	for (i = 0; i < nr_relocs; i++) {
 		const struct drm_etnaviv_gem_submit_reloc *r = relocs + i;
-		struct etnaviv_gem_object *bobj;
-		u32 iova, off;
+		struct etnaviv_gem_submit_bo *bo;
+		u32 off;
 
 		if (unlikely(r->flags)) {
 			DRM_ERROR("invalid reloc flags\n");
@@ -268,17 +265,16 @@ static int submit_reloc(struct etnaviv_gem_submit *submit, void *stream,
 			return -EINVAL;
 		}
 
-		ret = submit_bo(submit, r->reloc_idx, &bobj, &iova);
+		ret = submit_bo(submit, r->reloc_idx, &bo);
 		if (ret)
 			return ret;
 
-		if (r->reloc_offset >=
-		    bobj->base.size - sizeof(*ptr)) {
+		if (r->reloc_offset >= bo->obj->base.size - sizeof(*ptr)) {
 			DRM_ERROR("relocation %u outside object", i);
 			return -EINVAL;
 		}
 
-		ptr[off] = iova + r->reloc_offset;
+		ptr[off] = bo->mapping->iova + r->reloc_offset;
 
 		last_offset = off;
 	}

drivers/gpu/drm/etnaviv/etnaviv_gpu.c

@@ -628,6 +628,7 @@ int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
 	/* Now program the hardware */
 	mutex_lock(&gpu->lock);
 	etnaviv_gpu_hw_init(gpu);
+	gpu->exec_state = -1;
 	mutex_unlock(&gpu->lock);
 
 	pm_runtime_mark_last_busy(gpu->dev);
@@ -871,17 +872,13 @@ static void recover_worker(struct work_struct *work)
 		gpu->event[i].fence = NULL;
 		gpu->event[i].used = false;
 		complete(&gpu->event_free);
-		/*
-		 * Decrement the PM count for each stuck event. This is safe
-		 * even in atomic context as we use ASYNC RPM here.
-		 */
-		pm_runtime_put_autosuspend(gpu->dev);
 	}
 	spin_unlock_irqrestore(&gpu->event_spinlock, flags);
 	gpu->completed_fence = gpu->active_fence;
 
 	etnaviv_gpu_hw_init(gpu);
 	gpu->switch_context = true;
+	gpu->exec_state = -1;
 
 	mutex_unlock(&gpu->lock);
 	pm_runtime_mark_last_busy(gpu->dev);
@@ -1106,7 +1103,7 @@ struct etnaviv_cmdbuf *etnaviv_gpu_cmdbuf_new(struct etnaviv_gpu *gpu, u32 size,
 	size_t nr_bos)
 {
 	struct etnaviv_cmdbuf *cmdbuf;
-	size_t sz = size_vstruct(nr_bos, sizeof(cmdbuf->bo[0]),
+	size_t sz = size_vstruct(nr_bos, sizeof(cmdbuf->bo_map[0]),
 				 sizeof(*cmdbuf));
 
 	cmdbuf = kzalloc(sz, GFP_KERNEL);
@@ -1150,14 +1147,23 @@ static void retire_worker(struct work_struct *work)
 		fence_put(cmdbuf->fence);
 
 		for (i = 0; i < cmdbuf->nr_bos; i++) {
-			struct etnaviv_gem_object *etnaviv_obj = cmdbuf->bo[i];
+			struct etnaviv_vram_mapping *mapping = cmdbuf->bo_map[i];
+			struct etnaviv_gem_object *etnaviv_obj = mapping->object;
 
 			atomic_dec(&etnaviv_obj->gpu_active);
 			/* drop the refcount taken in etnaviv_gpu_submit */
-			etnaviv_gem_put_iova(gpu, &etnaviv_obj->base);
+			etnaviv_gem_mapping_unreference(mapping);
 		}
 
 		etnaviv_gpu_cmdbuf_free(cmdbuf);
+		/*
+		 * We need to balance the runtime PM count caused by
+		 * each submission.  Upon submission, we increment
+		 * the runtime PM counter, and allocate one event.
+		 * So here, we put the runtime PM count for each
+		 * completed event.
+		 */
+		pm_runtime_put_autosuspend(gpu->dev);
 	}
 
 	gpu->retired_fence = fence;
@@ -1304,11 +1310,10 @@ int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
 
 	for (i = 0; i < submit->nr_bos; i++) {
 		struct etnaviv_gem_object *etnaviv_obj = submit->bos[i].obj;
-		u32 iova;
 
-		/* Each cmdbuf takes a refcount on the iova */
-		etnaviv_gem_get_iova(gpu, &etnaviv_obj->base, &iova);
-		cmdbuf->bo[i] = etnaviv_obj;
+		/* Each cmdbuf takes a refcount on the mapping */
+		etnaviv_gem_mapping_reference(submit->bos[i].mapping);
+		cmdbuf->bo_map[i] = submit->bos[i].mapping;
 		atomic_inc(&etnaviv_obj->gpu_active);
 
 		if (submit->bos[i].flags & ETNA_SUBMIT_BO_WRITE)
@@ -1378,15 +1383,6 @@ static irqreturn_t irq_handler(int irq, void *data)
 				gpu->completed_fence = fence->seqno;
 
 			event_free(gpu, event);
-
-			/*
-			 * We need to balance the runtime PM count caused by
-			 * each submission.  Upon submission, we increment
-			 * the runtime PM counter, and allocate one event.
-			 * So here, we put the runtime PM count for each
-			 * completed event.
-			 */
-			pm_runtime_put_autosuspend(gpu->dev);
 		}
 
 		/* Retire the buffer objects in a work */
@@ -1481,6 +1477,7 @@ static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
 	etnaviv_gpu_hw_init(gpu);
 
 	gpu->switch_context = true;
+	gpu->exec_state = -1;
 
 	mutex_unlock(&gpu->lock);
 
@@ -1569,6 +1566,7 @@ static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
 {
 	struct device *dev = &pdev->dev;
 	struct etnaviv_gpu *gpu;
+	u32 dma_mask;
 	int err = 0;
 
 	gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
@@ -1579,12 +1577,16 @@ static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
 	mutex_init(&gpu->lock);
 
 	/*
-	 * Set the GPU base address to the start of physical memory.  This
-	 * ensures that if we have up to 2GB, the v1 MMU can address the
-	 * highest memory.  This is important as command buffers may be
-	 * allocated outside of this limit.
+	 * Set the GPU linear window to be at the end of the DMA window, where
+	 * the CMA area is likely to reside. This ensures that we are able to
+	 * map the command buffers while having the linear window overlap as
+	 * much RAM as possible, so we can optimize mappings for other buffers.
 	 */
-	gpu->memory_base = PHYS_OFFSET;
+	dma_mask = (u32)dma_get_required_mask(dev);
+	if (dma_mask < PHYS_OFFSET + SZ_2G)
+		gpu->memory_base = PHYS_OFFSET;
+	else
+		gpu->memory_base = dma_mask - SZ_2G + 1;
 
 	/* Map registers: */
 	gpu->mmio = etnaviv_ioremap(pdev, NULL, dev_name(gpu->dev));

drivers/gpu/drm/etnaviv/etnaviv_gpu.h

@@ -23,6 +23,7 @@
 #include "etnaviv_drv.h"
 
 struct etnaviv_gem_submit;
+struct etnaviv_vram_mapping;
 
 struct etnaviv_chip_identity {
 	/* Chip model. */
@@ -103,6 +104,7 @@ struct etnaviv_gpu {
 
 	/* 'ring'-buffer: */
 	struct etnaviv_cmdbuf *buffer;
+	int exec_state;
 
 	/* bus base address of memory  */
 	u32 memory_base;
@@ -166,7 +168,7 @@ struct etnaviv_cmdbuf {
 	struct list_head node;
 	/* BOs attached to this command buffer */
 	unsigned int nr_bos;
-	struct etnaviv_gem_object *bo[0];
+	struct etnaviv_vram_mapping *bo_map[0];
 };
 
 static inline void gpu_write(struct etnaviv_gpu *gpu, u32 reg, u32 data)

drivers/gpu/drm/etnaviv/etnaviv_mmu.c

@@ -193,7 +193,7 @@ int etnaviv_iommu_map_gem(struct etnaviv_iommu *mmu,
 
 		/*
 		 * Unmap the blocks which need to be reaped from the MMU.
-		 * Clear the mmu pointer to prevent the get_iova finding
+		 * Clear the mmu pointer to prevent the mapping_get finding
 		 * this mapping.
 		 */
 		list_for_each_entry_safe(m, n, &list, scan_node) {

drivers/gpu/drm/etnaviv/state_3d.xml.h

@@ -0,0 +1,9 @@
+#ifndef STATE_3D_XML
+#define STATE_3D_XML
+
+/* This is a cut-down version of the state_3d.xml.h file */
+
+#define VIVS_TS_FLUSH_CACHE					0x00001650
+#define VIVS_TS_FLUSH_CACHE_FLUSH				0x00000001
+
+#endif /* STATE_3D_XML */