drm/vmwgfx: Restart command buffers after errors

Previously we skipped the command buffer and added an extra fence to
avoid hangs due to skipped fence commands.
Now we instead restart the command buffer after the failing command,
if there are any commands left.
In addition we print out some information about the failing command
and its location in the command buffer.

Testing Done: ran glxgears using mesa modified to send the NOP_ERROR
command before each 10th clear and verified that we detected the device
error properly and that there were no other device errors caused by
incorrectly ordered command buffers. Also ran the piglit "quick" test
suite which generates a couple of device errors and verified that
they were handled as intended.

Signed-off-by: Thomas Hellstrom <[e-mail]>
Reviewed-by: Brian Paul <[e-mail]>
Reviewed-by: Sinclair Yeh <[e-mail]>

drivers/gpu/drm/vmwgfx/vmwgfx_cmdbuf.c

···
 	struct list_head hw_submitted;
 	struct list_head preempted;
 	unsigned num_hw_submitted;
+	bool block_submission;
 };
 
 /**
···
  * kernel command submissions, @cur.
  * @space_mutex: Mutex to protect against starvation when we allocate
  * main pool buffer space.
+ * @error_mutex: Mutex to serialize the work queue error handling.
+ * Note this is not needed if the same workqueue handler
+ * can't race with itself...
  * @work: A struct work_struct implementeing command buffer error handling.
  * Immutable.
  * @dev_priv: Pointer to the device private struct. Immutable.
···
 struct vmw_cmdbuf_man {
 	struct mutex cur_mutex;
 	struct mutex space_mutex;
+	struct mutex error_mutex;
 	struct work_struct work;
 	struct vmw_private *dev_priv;
 	struct vmw_cmdbuf_context ctx[SVGA_CB_CONTEXT_MAX];
···
 };
 
 /* Loop over each context in the command buffer manager. */
-#define for_each_cmdbuf_ctx(_man, _i, _ctx) \
+#define for_each_cmdbuf_ctx(_man, _i, _ctx)				\
 	for (_i = 0, _ctx = &(_man)->ctx[0]; (_i) < SVGA_CB_CONTEXT_MAX; \
 	     ++(_i), ++(_ctx))
 
-static int vmw_cmdbuf_startstop(struct vmw_cmdbuf_man *man, bool enable);
-
+static int vmw_cmdbuf_startstop(struct vmw_cmdbuf_man *man, u32 context,
+				bool enable);
+static int vmw_cmdbuf_preempt(struct vmw_cmdbuf_man *man, u32 context);
 
 /**
  * vmw_cmdbuf_cur_lock - Helper to lock the cur_mutex.
···
 				  struct vmw_cmdbuf_context *ctx)
 {
 	while (ctx->num_hw_submitted < man->max_hw_submitted &&
-	      !list_empty(&ctx->submitted)) {
+	       !list_empty(&ctx->submitted) &&
+	       !ctx->block_submission) {
 		struct vmw_cmdbuf_header *entry;
 		SVGACBStatus status;
 
···
 			__vmw_cmdbuf_header_free(entry);
 			break;
 		case SVGA_CB_STATUS_COMMAND_ERROR:
-		case SVGA_CB_STATUS_CB_HEADER_ERROR:
+			entry->cb_header->status = SVGA_CB_STATUS_NONE;
 			list_add_tail(&entry->list, &man->error);
 			schedule_work(&man->work);
 			break;
 		case SVGA_CB_STATUS_PREEMPTED:
-			list_add(&entry->list, &ctx->preempted);
+			entry->cb_header->status = SVGA_CB_STATUS_NONE;
+			list_add_tail(&entry->list, &ctx->preempted);
+			break;
+		case SVGA_CB_STATUS_CB_HEADER_ERROR:
+			WARN_ONCE(true, "Command buffer header error.\n");
+			__vmw_cmdbuf_header_free(entry);
 			break;
 		default:
 			WARN_ONCE(true, "Undefined command buffer status.\n");
···
 		container_of(work, struct vmw_cmdbuf_man, work);
 	struct vmw_cmdbuf_header *entry, *next;
 	uint32_t dummy;
-	bool restart = false;
+	bool restart[SVGA_CB_CONTEXT_MAX];
+	bool send_fence = false;
+	struct list_head restart_head[SVGA_CB_CONTEXT_MAX];
+	int i;
+	struct vmw_cmdbuf_context *ctx;
 
+	for_each_cmdbuf_ctx(man, i, ctx) {
+		INIT_LIST_HEAD(&restart_head[i]);
+		restart[i] = false;
+	}
+
+	mutex_lock(&man->error_mutex);
 	spin_lock(&man->lock);
 	list_for_each_entry_safe(entry, next, &man->error, list) {
-		restart = true;
-		DRM_ERROR("Command buffer error.\n");
+		SVGACBHeader *cb_hdr = entry->cb_header;
+		SVGA3dCmdHeader *header = (SVGA3dCmdHeader *)
+			(entry->cmd + cb_hdr->errorOffset);
+		u32 error_cmd_size, new_start_offset;
+		const char *cmd_name;
 
-		list_del(&entry->list);
-		__vmw_cmdbuf_header_free(entry);
-		wake_up_all(&man->idle_queue);
+		list_del_init(&entry->list);
+		restart[entry->cb_context] = true;
+
+		if (!vmw_cmd_describe(header, &error_cmd_size, &cmd_name)) {
+			DRM_ERROR("Unknown command causing device error.\n");
+			DRM_ERROR("Command buffer offset is %lu\n",
+				  (unsigned long) cb_hdr->errorOffset);
+			__vmw_cmdbuf_header_free(entry);
+			send_fence = true;
+			continue;
+		}
+
+		DRM_ERROR("Command \"%s\" causing device error.\n", cmd_name);
+		DRM_ERROR("Command buffer offset is %lu\n",
+			  (unsigned long) cb_hdr->errorOffset);
+		DRM_ERROR("Command size is %lu\n",
+			  (unsigned long) error_cmd_size);
+
+		new_start_offset = cb_hdr->errorOffset + error_cmd_size;
+
+		if (new_start_offset >= cb_hdr->length) {
+			__vmw_cmdbuf_header_free(entry);
+			send_fence = true;
+			continue;
+		}
+
+		if (man->using_mob)
+			cb_hdr->ptr.mob.mobOffset += new_start_offset;
+		else
+			cb_hdr->ptr.pa += (u64) new_start_offset;
+
+		entry->cmd += new_start_offset;
+		cb_hdr->length -= new_start_offset;
+		cb_hdr->errorOffset = 0;
+		list_add_tail(&entry->list, &restart_head[entry->cb_context]);
+		man->ctx[entry->cb_context].block_submission = true;
 	}
 	spin_unlock(&man->lock);
 
-	if (restart && vmw_cmdbuf_startstop(man, true))
-		DRM_ERROR("Failed restarting command buffer context 0.\n");
+	/* Preempt all contexts with errors */
+	for_each_cmdbuf_ctx(man, i, ctx) {
+		if (ctx->block_submission && vmw_cmdbuf_preempt(man, i))
+			DRM_ERROR("Failed preempting command buffer "
+				  "context %u.\n", i);
+	}
+
+	spin_lock(&man->lock);
+	for_each_cmdbuf_ctx(man, i, ctx) {
+		if (!ctx->block_submission)
+			continue;
+
+		/* Move preempted command buffers to the preempted queue. */
+		vmw_cmdbuf_ctx_process(man, ctx, &dummy);
+
+		/*
+		 * Add the preempted queue after the command buffer
+		 * that caused an error.
+		 */
+		list_splice_init(&ctx->preempted, restart_head[i].prev);
+
+		/*
+		 * Finally add all command buffers first in the submitted
+		 * queue, to rerun them.
+		 */
+		list_splice_init(&restart_head[i], &ctx->submitted);
+
+		ctx->block_submission = false;
+	}
+
+	vmw_cmdbuf_man_process(man);
+	spin_unlock(&man->lock);
+
+	for_each_cmdbuf_ctx(man, i, ctx) {
+		if (restart[i] && vmw_cmdbuf_startstop(man, i, true))
+			DRM_ERROR("Failed restarting command buffer "
+				  "context %u.\n", i);
+	}
 
 	/* Send a new fence in case one was removed */
-	vmw_fifo_send_fence(man->dev_priv, &dummy);
+	if (send_fence) {
+		vmw_fifo_send_fence(man->dev_priv, &dummy);
+		wake_up_all(&man->idle_queue);
+	}
+
+	mutex_unlock(&man->error_mutex);
 }
 
 /**
···
 }
 
 /**
+ * vmw_cmdbuf_preempt - Send a preempt command through the device
+ * context.
+ *
+ * @man: The command buffer manager.
+ *
+ * Synchronously sends a preempt command.
+ */
+static int vmw_cmdbuf_preempt(struct vmw_cmdbuf_man *man, u32 context)
+{
+	struct {
+		uint32 id;
+		SVGADCCmdPreempt body;
+	} __packed cmd;
+
+	cmd.id = SVGA_DC_CMD_PREEMPT;
+	cmd.body.context = SVGA_CB_CONTEXT_0 + context;
+	cmd.body.ignoreIDZero = 0;
+
+	return vmw_cmdbuf_send_device_command(man, &cmd, sizeof(cmd));
+}
+
+
+/**
  * vmw_cmdbuf_startstop - Send a start / stop command through the device
  * context.
  *
···
  *
  * Synchronously sends a device start / stop context command.
  */
-static int vmw_cmdbuf_startstop(struct vmw_cmdbuf_man *man,
+static int vmw_cmdbuf_startstop(struct vmw_cmdbuf_man *man, u32 context,
 				bool enable)
 {
 	struct {
···
 
 	cmd.id = SVGA_DC_CMD_START_STOP_CONTEXT;
 	cmd.body.enable = (enable) ? 1 : 0;
-	cmd.body.context = SVGA_CB_CONTEXT_0;
+	cmd.body.context = SVGA_CB_CONTEXT_0 + context;
 
 	return vmw_cmdbuf_send_device_command(man, &cmd, sizeof(cmd));
 }
···
 {
 	struct vmw_cmdbuf_man *man;
 	struct vmw_cmdbuf_context *ctx;
-	int i;
+	unsigned int i;
 	int ret;
 
 	if (!(dev_priv->capabilities & SVGA_CAP_COMMAND_BUFFERS))
···
 	spin_lock_init(&man->lock);
 	mutex_init(&man->cur_mutex);
 	mutex_init(&man->space_mutex);
+	mutex_init(&man->error_mutex);
 	man->default_size = VMW_CMDBUF_INLINE_SIZE;
 	init_waitqueue_head(&man->alloc_queue);
 	init_waitqueue_head(&man->idle_queue);
···
 	INIT_WORK(&man->work, &vmw_cmdbuf_work_func);
 	vmw_generic_waiter_add(dev_priv, SVGA_IRQFLAG_ERROR,
 			       &dev_priv->error_waiters);
-	ret = vmw_cmdbuf_startstop(man, true);
-	if (ret) {
-		DRM_ERROR("Failed starting command buffer context 0.\n");
-		vmw_cmdbuf_man_destroy(man);
-		return ERR_PTR(ret);
+	for_each_cmdbuf_ctx(man, i, ctx) {
+		ret = vmw_cmdbuf_startstop(man, i, true);
+		if (ret) {
+			DRM_ERROR("Failed starting command buffer "
+				  "context %u.\n", i);
+			vmw_cmdbuf_man_destroy(man);
+			return ERR_PTR(ret);
+		}
 	}
 
 	return man;
···
  */
 void vmw_cmdbuf_man_destroy(struct vmw_cmdbuf_man *man)
 {
+	struct vmw_cmdbuf_context *ctx;
+	unsigned int i;
+
 	WARN_ON_ONCE(man->has_pool);
 	(void) vmw_cmdbuf_idle(man, false, 10*HZ);
-	if (vmw_cmdbuf_startstop(man, false))
-		DRM_ERROR("Failed stopping command buffer context 0.\n");
+
+	for_each_cmdbuf_ctx(man, i, ctx)
+		if (vmw_cmdbuf_startstop(man, i, false))
+			DRM_ERROR("Failed stopping command buffer "
+				  "context %u.\n", i);
 
 	vmw_generic_waiter_remove(man->dev_priv, SVGA_IRQFLAG_ERROR,
 				  &man->dev_priv->error_waiters);
···
 	dma_pool_destroy(man->headers);
 	mutex_destroy(&man->cur_mutex);
 	mutex_destroy(&man->space_mutex);
+	mutex_destroy(&man->error_mutex);
 	kfree(man);
 }

drivers/gpu/drm/vmwgfx/vmwgfx_drv.h

···
 				      struct ttm_buffer_object *bo,
 				      bool interruptible,
 				      bool validate_as_mob);
-
+bool vmw_cmd_describe(const void *buf, u32 *size, char const **cmd);
 
 /**
  * IRQs and wating - vmwgfx_irq.c

drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c

···
 	bool user_allow;
 	bool gb_disable;
 	bool gb_enable;
+	const char *cmd_name;
 };
 
 #define VMW_CMD_DEF(_cmd, _func, _user_allow, _gb_disable, _gb_enable)	\
 	[(_cmd) - SVGA_3D_CMD_BASE] = {(_func), (_user_allow),\
-				       (_gb_disable), (_gb_enable)}
+				       (_gb_disable), (_gb_enable), #_cmd}
 
 static int vmw_resource_context_res_add(struct vmw_private *dev_priv,
 					struct vmw_sw_context *sw_context,
···
 		    &vmw_cmd_dx_transfer_from_buffer,
 		    true, false, true),
 };
+
+bool vmw_cmd_describe(const void *buf, u32 *size, char const **cmd)
+{
+	u32 cmd_id = ((u32 *) buf)[0];
+
+	if (cmd_id >= SVGA_CMD_MAX) {
+		SVGA3dCmdHeader *header = (SVGA3dCmdHeader *) buf;
+		const struct vmw_cmd_entry *entry;
+
+		*size = header->size + sizeof(SVGA3dCmdHeader);
+		cmd_id = header->id;
+		if (cmd_id >= SVGA_3D_CMD_MAX)
+			return false;
+
+		cmd_id -= SVGA_3D_CMD_BASE;
+		entry = &vmw_cmd_entries[cmd_id];
+		*cmd = entry->cmd_name;
+		return true;
+	}
+
+	switch (cmd_id) {
+	case SVGA_CMD_UPDATE:
+		*cmd = "SVGA_CMD_UPDATE";
+		*size = sizeof(u32) + sizeof(SVGAFifoCmdUpdate);
+		break;
+	case SVGA_CMD_DEFINE_GMRFB:
+		*cmd = "SVGA_CMD_DEFINE_GMRFB";
+		*size = sizeof(u32) + sizeof(SVGAFifoCmdDefineGMRFB);
+		break;
+	case SVGA_CMD_BLIT_GMRFB_TO_SCREEN:
+		*cmd = "SVGA_CMD_BLIT_GMRFB_TO_SCREEN";
+		*size = sizeof(u32) + sizeof(SVGAFifoCmdBlitGMRFBToScreen);
+		break;
+	case SVGA_CMD_BLIT_SCREEN_TO_GMRFB:
+		*cmd = "SVGA_CMD_BLIT_SCREEN_TO_GMRFB";
+		*size = sizeof(u32) + sizeof(SVGAFifoCmdBlitGMRFBToScreen);
+		break;
+	default:
+		*cmd = "UNKNOWN";
+		*size = 0;
+		return false;
+	}
+
+	return true;
+}
 
 static int vmw_cmd_check(struct vmw_private *dev_priv,
 			 struct vmw_sw_context *sw_context,