video: tegra: host: add prefence signal support

[1] Adds support for prefence signal support
[2] Facilitates handling of SOF feature request from higher
    levels
[3] Update ioctl structure for emu_submit to allow prefence
    with signal action

Jira DLA-1992
Jira DLA-1993

Change-Id: Id0fe8a7a3ff46227121ec5342197b1da2a3f346d
Signed-off-by: Arvind M <am@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/2121530
Reviewed-by: Shridhar Rasal <srasal@nvidia.com>
Reviewed-by: Bharat Nihalani <bnihalani@nvidia.com>
GVS: Gerrit_Virtual_Submit
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

drivers/video/tegra/host/nvdla/nvdla.h

@@ -99,6 +99,8 @@
  */
 #define MAX_NUM_NVDLA_PREFENCES		32
 #define MAX_NUM_NVDLA_POSTFENCES	32
+#define MAX_NUM_NVDLA_EMU_PREFENCES	16
+#define MAX_NUM_NVDLA_EMU_POSTFENCES	16
 #define MAX_NUM_NVDLA_IN_TASK_STATUS	MAX_NUM_NVDLA_PREFENCES
 #define MAX_NUM_NVDLA_OUT_TASK_STATUS	MAX_NUM_NVDLA_POSTFENCES
 #define NUM_PROFILING_POSTACTION	1
@@ -219,7 +221,9 @@ struct nvdla_device {
  *
  * @queue		Queue in which task submitted
  * @sp			pointer to syncpt
+ * @prefences		pointer to pre fences
  * @postfences		pointer to post fences
+ * @num_prefences	Number of prefences in task
  * @num_postfences	Number of postfences in task
  * @fence		Fence tracking for current task
  * @fence_counter	Counter used to track fence value
@@ -228,7 +232,9 @@ struct nvdla_device {
 struct nvdla_emu_task {
 	struct nvdla_queue *queue;
 	struct nvhost_syncpt *sp;
-	struct nvdev_fence postfences[MAX_NUM_NVDLA_POSTFENCES];
+	struct nvdev_fence prefences[MAX_NUM_NVDLA_EMU_PREFENCES];
+	struct nvdev_fence postfences[MAX_NUM_NVDLA_EMU_POSTFENCES];
+	u32 num_prefences;
 	u32 num_postfences;
 	u32 fence;
 	u32 fence_counter;
@@ -402,7 +408,7 @@ int nvdla_free_gcov_region(struct platform_device *pdev, bool update_region);
 int nvdla_emulator_submit(struct nvdla_queue *queue,
 				struct nvdla_emu_task *task);
 void task_free(struct kref *ref);
-int nvdla_get_postfences(struct nvdla_queue *queue, void *in_task);
+int nvdla_get_signal_fences(struct nvdla_queue *queue, void *in_task);
 int nvdla_send_gos_region(struct platform_device *pdev);
 
 #endif /* End of __NVHOST_NVDLA_H__ */

drivers/video/tegra/host/nvdla/nvdla_ioctl.c

@@ -392,20 +392,66 @@ fail:
 	return err;
 }
 
-static int nvdla_send_emu_postfences(struct nvdla_emu_task *task,
+static int nvdla_send_emu_signal_fences(struct nvdla_emu_task *task,
 			struct nvdla_ioctl_emu_submit_task *user_task)
 {
 	int err = 0, i;
 	struct platform_device *dla_pdev = task->queue->pool->pdev;
 	struct platform_device *host_pdev =
 				to_platform_device(dla_pdev->dev.parent);
+	struct nvdev_fence __user *prefences =
+		(struct nvdev_fence __user *)(uintptr_t)user_task->prefences;
 	struct nvdev_fence __user *postfences =
 		(struct nvdev_fence __user *)(uintptr_t)user_task->postfences;
 	char fence_name[32];
 
-	nvdla_dbg_fn(dla_pdev, "sending post fences");
+	nvdla_dbg_fn(dla_pdev, "sending signal fences");
+
+	for (i = 0; i < task->num_prefences; i++) {
+		if (task->prefences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
+		if (task->prefences[i].type == NVDEV_FENCE_TYPE_SYNC_FD) {
+			struct nvhost_ctrl_sync_fence_info info;
+
+			info.id = task->prefences[i].syncpoint_index;
+			info.thresh = task->prefences[i].syncpoint_value;
+
+			nvdla_dbg_info(dla_pdev,
+					"creating pre sync fd [%d]:[%d]\n",
+					info.id, info.thresh);
+
+			/* create fence name format example: nvdla0_1_fence */
+			snprintf(fence_name, sizeof(fence_name),
+				"%s_%d_%d_prefence", dev_name(&dla_pdev->dev),
+				task->prefences[i].syncpoint_index, i);
+
+			err = nvhost_sync_create_fence_fd(host_pdev,
+				&info, 1, fence_name,
+				&task->prefences[i].sync_fd);
+
+			if (err) {
+				nvdla_dbg_err(dla_pdev,
+					"fail to create prefence syncfd\n");
+				goto fail;
+			}
+		}
+	}
+
+	nvdla_dbg_fn(dla_pdev, "copy prefences to user");
+	/* send pre fences */
+	if (copy_to_user(prefences, task->prefences,
+		(task->num_prefences * sizeof(struct nvdev_fence)))) {
+		err = -EFAULT;
+		nvdla_dbg_err(dla_pdev, "failed to send prefences");
+		goto fail;
+	}
+	nvdla_dbg_info(dla_pdev, "prefences sent");
 
 	for (i = 0; i < task->num_postfences; i++) {
+		if (task->postfences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
 		if (task->postfences[i].type == NVDEV_FENCE_TYPE_SYNC_FD) {
 			struct nvhost_ctrl_sync_fence_info info;
 
@@ -418,8 +464,8 @@ static int nvdla_send_emu_postfences(struct nvdla_emu_task *task,
 
 			/* create fence name format example: nvdla0_1_fence */
 			snprintf(fence_name, sizeof(fence_name),
-				"%s_%d_fence", dev_name(&dla_pdev->dev),
-				task->postfences[i].syncpoint_index);
+				"%s_%d_%d_postfence", dev_name(&dla_pdev->dev),
+				task->postfences[i].syncpoint_index, i);
 
 			err = nvhost_sync_create_fence_fd(host_pdev,
 				&info, 1, fence_name,
@@ -447,20 +493,67 @@ fail:
 	return err;
 }
 
-static int nvdla_update_postfences(struct nvdla_task *task,
+static int nvdla_update_signal_fences(struct nvdla_task *task,
 			struct nvdla_ioctl_submit_task *user_task)
 {
 	int err = 0, i;
 	struct platform_device *dla_pdev = task->queue->pool->pdev;
 	struct platform_device *host_pdev =
 				to_platform_device(dla_pdev->dev.parent);
+	struct nvdev_fence __user *prefences =
+		(struct nvdev_fence __user *)(uintptr_t)user_task->prefences;
 	struct nvdev_fence __user *postfences =
 		(struct nvdev_fence __user *)(uintptr_t)user_task->postfences;
 	char fence_name[32];
 
-	nvdla_dbg_fn(dla_pdev, "copy post fences for user");
+	nvdla_dbg_fn(dla_pdev, "copy fences for user");
 
+	/* update pre fence signals to users */
+	for (i = 0; i < task->num_prefences; i++) {
+		if (task->prefences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
+		if (task->prefences[i].type == NVDEV_FENCE_TYPE_SYNC_FD) {
+			struct nvhost_ctrl_sync_fence_info info;
+
+			info.id = task->prefences[i].syncpoint_index;
+			info.thresh = task->prefences[i].syncpoint_value;
+
+			nvdla_dbg_info(dla_pdev,
+					"creating pre sync fd [%d]:[%d]\n",
+					info.id, info.thresh);
+
+			/* create fence name format example: nvdla0_1_fence */
+			snprintf(fence_name, sizeof(fence_name),
+				"%s_%d_%d_prefence", dev_name(&dla_pdev->dev),
+				task->prefences[i].syncpoint_index, i);
+
+			err = nvhost_sync_create_fence_fd(host_pdev,
+				&info, 1, fence_name,
+				&task->prefences[i].sync_fd);
+
+			if (err) {
+				nvdla_dbg_err(dla_pdev,
+					"fail to create prefence syncfd\n");
+				goto fail;
+			}
+		}
+	}
+
+	nvdla_dbg_fn(dla_pdev, "copy prefences to user");
+	/* copy pre fences */
+	if (copy_to_user(prefences, task->prefences,
+				(task->num_prefences * sizeof(struct nvdev_fence)))) {
+		err = -EFAULT;
+		nvdla_dbg_err(dla_pdev, "failed to copy prefences");
+		goto fail;
+	}
+
+	/* update post fence signals to user */
 	for (i = 0; i < task->num_postfences; i++) {
+		if (task->postfences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
 		if (task->postfences[i].type == NVDEV_FENCE_TYPE_SYNC_FD) {
 			struct nvhost_ctrl_sync_fence_info info;
 
@@ -473,8 +566,8 @@ static int nvdla_update_postfences(struct nvdla_task *task,
 
 			/* create fence name format example: nvdla0_1_fence */
 			snprintf(fence_name, sizeof(fence_name),
-				"%s_%d_fence", dev_name(&dla_pdev->dev),
-				task->postfences[i].syncpoint_index);
+				"%s_%d_%d_postfence", dev_name(&dla_pdev->dev),
+				task->postfences[i].syncpoint_index, i);
 
 			err = nvhost_sync_create_fence_fd(host_pdev,
 				&info, 1, fence_name,
@@ -531,9 +624,9 @@ static int nvdla_val_task_submit_input(struct nvdla_ioctl_submit_task *in_task)
 			MAX_NUM_NVDLA_OUT_TASK_STATUS);
 		return -EINVAL;
 	}
-        if (in_task->num_addresses < 1) {
+	if (in_task->num_addresses < 1) {
 		pr_err("num addresses[%u] should be min one\n",
-			in_task->num_addresses);
+				in_task->num_addresses);
 		return -EINVAL;
 	}
 	if (in_task->num_addresses > NVDLA_MAX_BUFFERS_PER_TASK) {
@@ -718,8 +811,18 @@ static int nvdla_emu_task_submit(struct nvdla_private *priv, void *arg)
 
 		nvdla_dbg_info(pdev, "submit [%d]th task", i + 1);
 
+		task.num_prefences = local_tasks[i].num_prefences;
 		task.num_postfences = local_tasks[i].num_postfences;
 
+		/* get pre fences */
+		if (copy_from_user(task.prefences,
+			(void __user *)local_tasks[i].prefences,
+			(task.num_prefences * sizeof(struct nvdev_fence)))) {
+			err = -EFAULT;
+			nvdla_dbg_err(pdev, "failed to copy prefences");
+			goto exit;
+		}
+
 		/* get post fences */
 		if (copy_from_user(task.postfences,
 			(void __user *)local_tasks[i].postfences,
@@ -736,13 +839,13 @@ static int nvdla_emu_task_submit(struct nvdla_private *priv, void *arg)
 		}
 		nvdla_dbg_info(pdev, "task[%d] submitted", i + 1);
 
-		/* send fences to user */
-		err = nvdla_send_emu_postfences(&task, local_tasks + i);
+		/* send signal fences to user */
+		err = nvdla_send_emu_signal_fences(&task, local_tasks + i);
 		if (err) {
-			nvdla_dbg_err(pdev, "fail to send postfence%d", i + 1);
+			nvdla_dbg_err(pdev, "fail to send sig fence%d", i + 1);
 			goto exit;
 		}
-		nvdla_dbg_info(pdev, "postfences of task[%d] sent", i + 1);
+		nvdla_dbg_info(pdev, "signal fences of task[%d] sent", i + 1);
 	}
 	nvdla_dbg_fn(pdev, "Emulator task submitted, done!");
 
@@ -825,8 +928,8 @@ static int nvdla_submit(struct nvdla_private *priv, void *arg)
 		}
 		nvdla_dbg_info(pdev, "task[%d] desc filled", i + 1);
 
-		/* get expected postfences prior to submit */
-		err = nvdla_get_postfences(queue, task);
+		/* get expected signal fences prior to submit */
+		err = nvdla_get_signal_fences(queue, task);
 		if (err) {
 			nvdla_dbg_err(pdev, "fail to get fences%d", i + 1);
 			goto fail_to_get_fences;
@@ -834,7 +937,7 @@ static int nvdla_submit(struct nvdla_private *priv, void *arg)
 		nvdla_dbg_info(pdev, "task[%d] got fences", i + 1);
 
 		/* update fences to user */
-		err = nvdla_update_postfences(task, local_tasks + i);
+		err = nvdla_update_signal_fences(task, local_tasks + i);
 		if (err) {
 			nvdla_dbg_err(pdev, "fail update postfence%d", i + 1);
 			goto fail_to_update_postfences;

drivers/video/tegra/host/nvdla/nvdla_queue.c

@@ -651,7 +651,6 @@ static int nvdla_fill_wait_fence_action(struct nvdla_task *task,
 					fence->syncpoint_value);
 		} else {
 			dma_addr_t syncpt_addr;
-
 			nvdla_dbg_info(pdev, "GoS missing");
 
 			syncpt_addr = nvhost_syncpt_address(
@@ -945,9 +944,6 @@ static int nvdla_fill_postactions(struct nvdla_task *task)
 		}
 	}
 
-	/* reset fence counter */
-	task->fence_counter = 0;
-
 	/* fill all postactions */
 	for (i = 0; i < task->num_postfences; i++) {
 		/* update action */
@@ -992,8 +988,12 @@ static int nvdla_fill_preactions(struct nvdla_task *task)
 
 	start = next = (u8 *)task_desc + preactionlist_of;
 
-	/* fill all preactions */
+	/* fill all preactions wait */
 	for (i = 0; i < task->num_prefences; i++) {
+		if (task->prefences[i].action != NVDEV_FENCE_WAIT)
+			continue;
+
+		/* update action */
 		err = nvdla_fill_wait_fence_action(task,
 				&task->prefences[i],
 				&task->prefences_sem_dmabuf[i],
@@ -1016,6 +1016,24 @@ static int nvdla_fill_preactions(struct nvdla_task *task)
 		}
 	}
 
+	/* fill all preactions signals */
+	for (i = 0; i < task->num_prefences; i++) {
+		/* update action */
+		if (task->prefences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
+		err = nvdla_fill_signal_fence_action(task,
+				&task->prefences[i],
+				&task->prefences_sem_dmabuf[i],
+				&next);
+		if (err < 0) {
+			nvdla_dbg_err(pdev,
+					"fail to fill fence sig action [%d]",
+					i);
+			goto fail;
+		}
+	}
+
 	/* update end of action list */
 	next = add_opcode(next, ACTION_TERMINATE);
 
@@ -1077,6 +1095,9 @@ int nvdla_fill_task_desc(struct nvdla_task *task)
 
 	nvdla_update_gos(pdev);
 
+	/* reset fence counter */
+	task->fence_counter = 0;
+
 	/* fill pre actions */
 	nvdla_fill_preactions(task);
 
@@ -1187,8 +1208,30 @@ int nvdla_emulator_submit(struct nvdla_queue *queue, struct nvdla_emu_task *task
 
 	/* reset fence counter */
 	task->fence_counter = 0;
+
+	/* fill all preactions */
+	for (i = 0; i < task->num_prefences; i++) {
+		if (task->prefences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
+		/* update action */
+		switch (task->prefences[i].type) {
+		case NVDEV_FENCE_TYPE_SYNCPT:
+		case NVDEV_FENCE_TYPE_SYNC_FD: {
+			task->fence_counter = task->fence_counter + 1;
+			break;
+		}
+		default:
+			nvdla_dbg_err(pdev, "Invalid prefence sync type[%d]",
+				task->prefences[i].type);
+			return -EINVAL;
+		}
+	}
+
 	/* fill all postactions */
 	for (i = 0; i < task->num_postfences; i++) {
+		if (task->postfences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
 
 		/* update action */
 		switch (task->postfences[i].type) {
@@ -1212,9 +1255,31 @@ int nvdla_emulator_submit(struct nvdla_queue *queue, struct nvdla_emu_task *task
 				queue->syncpt_id, task->fence,
 				task, task->fence_counter);
 
-	/* Update postfences for all */
+	/* Update signal fences for all */
 	counter = task->fence_counter - 1;
+	for (i = 0; i < task->num_prefences; i++) {
+		if (task->prefences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
+		if ((task->prefences[i].type == NVDEV_FENCE_TYPE_SYNCPT) ||
+		    (task->prefences[i].type == NVDEV_FENCE_TYPE_SYNC_FD)) {
+			task->prefences[i].syncpoint_index =
+					queue->syncpt_id;
+			task->prefences[i].syncpoint_value =
+					task->fence - counter;
+
+			nvdla_dbg_info(pdev, "[%d] prefence set[%u]:[%u]",
+				i, task->prefences[i].syncpoint_index,
+				task->prefences[i].syncpoint_value);
+
+			counter = counter - 1;
+		}
+	}
+
 	for (i = 0; i < task->num_postfences; i++) {
+		if (task->postfences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
 		if ((task->postfences[i].type == NVDEV_FENCE_TYPE_SYNCPT) ||
 		    (task->postfences[i].type == NVDEV_FENCE_TYPE_SYNC_FD)) {
 			task->postfences[i].syncpoint_index =
@@ -1233,7 +1298,7 @@ int nvdla_emulator_submit(struct nvdla_queue *queue, struct nvdla_emu_task *task
 	return 0;
 }
 
-int nvdla_get_postfences(struct nvdla_queue *queue, void *in_task)
+int nvdla_get_signal_fences(struct nvdla_queue *queue, void *in_task)
 {
 	struct nvdla_task *task = (struct nvdla_task *)in_task;
 	struct platform_device *pdev = queue->pool->pdev;
@@ -1251,9 +1316,31 @@ int nvdla_get_postfences(struct nvdla_queue *queue, void *in_task)
 	task_fence = nvhost_syncpt_read_maxval(pdev, queue->syncpt_id) +
 			task->fence_counter;
 
-	/* Update postfences for all */
+	/* Update fences signal updates for both prefence and postfence */
 	counter = task->fence_counter - 1;
+	for (i = 0; i < task->num_prefences; i++) {
+		if (task->prefences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
+		if ((task->prefences[i].type == NVDEV_FENCE_TYPE_SYNCPT) ||
+		    (task->prefences[i].type == NVDEV_FENCE_TYPE_SYNC_FD)) {
+			task->prefences[i].syncpoint_index =
+					queue->syncpt_id;
+			task->prefences[i].syncpoint_value =
+					task_fence - counter;
+
+			nvdla_dbg_info(pdev, "[%d] prefence set[%u]:[%u]",
+				i, task->prefences[i].syncpoint_index,
+				task->prefences[i].syncpoint_value);
+
+			counter = counter - 1;
+		}
+	}
+
 	for (i = 0; i < task->num_postfences; i++) {
+		if (task->postfences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
 		if ((task->postfences[i].type == NVDEV_FENCE_TYPE_SYNCPT) ||
 		    (task->postfences[i].type == NVDEV_FENCE_TYPE_SYNC_FD)) {
 			task->postfences[i].syncpoint_index =

include/uapi/linux/nvhost_nvdla_ioctl.h

@@ -3,7 +3,7 @@
  *
  * Tegra NvDLA Driver
  *
- * Copyright (c) 2016-2018, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2016-2019, NVIDIA CORPORATION.  All rights reserved.
  *
  * This program is free software; you can redistribute it and/or modify it
  * under the terms and conditions of the GNU General Public License,
@@ -160,15 +160,17 @@ struct nvdla_ioctl_submit_task {
  * struct nvdla_ioctl_emu_submit_task structure for single emulator task
  * information
  *
+ * @num_prefences 		number of pre-fences in task
  * @num_postfences		number of post-fences in task
- * @reserved			reserved for padding and future use
- * @postfences			pointer to post-fence struct table
+ * @prefences     		pointer to pre-fence struct table
+ * @postfences    		pointer to post-fence struct table
  *
  */
 struct nvdla_ioctl_emu_submit_task {
+	__u32 num_prefences;
 	__u32 num_postfences;
-	__u32 reserved;
 
+	__u64 prefences;
 	__u64 postfences;
 };