drivers: pva: add support for exe_id2

- update submission path to accept exe_id2
 - submit additional bin info to FW

JIRA PVAAS-12334

Change-Id: Ie7b314c4cf76ecfdabd40de05a9b9f1e3d3cc049
Signed-off-by: Krish Agarwal <krisha@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvidia/+/2860864
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nv-oot/+/2999153
Reviewed-by: Amruta Sai Anusha Bhamidipati <abhamidipati@nvidia.com>
Reviewed-by: Omar Nemri <onemri@nvidia.com>
Tested-by: Omar Nemri <onemri@nvidia.com>
GVS: Gerrit_Virtual_Submit <buildbot_gerritrpt@nvidia.com>

drivers/video/tegra/host/pva/fw_include/pva-task.h

@@ -211,7 +211,7 @@ struct PVA_PACKED pva_td_s {
 	/** @brief IOVA pointer to a pva_bin_info_t structure */
 	pva_iova			bin_info;
 	/** @brief IOVA pointer to a pva_bin_info_t structure */
-	pva_iova			ppe_bin_info;
+	pva_iova			bin_info2;
 	/** @brief IOVA pointer to a pva_dma_info_t structure */
 	pva_iova			dma_info;
 	/** IOVA pointer to a pva_circular_info_t structure */

drivers/video/tegra/host/pva/pva_dma.c

@@ -263,6 +263,19 @@ out:
 	return err;
 }
 
+static uint16_t
+get_sym_exe_id(struct pva_submit_task *task)
+{
+	u16 exe_id = NVPVA_NOOP_EXE_ID;
+
+	if (task->exe_id2 != NVPVA_NOOP_EXE_ID)
+		exe_id = task->exe_id2;
+	if (task->exe_id1 != NVPVA_NOOP_EXE_ID)
+		exe_id = task->exe_id1;
+
+	return exe_id;
+}
+
 static int32_t
 patch_dma_desc_address(struct pva_submit_task *task,
 		      struct nvpva_dma_descriptor *umd_dma_desc,
@@ -271,11 +284,13 @@ patch_dma_desc_address(struct pva_submit_task *task,
 {
 	int32_t err = 0;
 	uint64_t addr_base = 0;
+	uint16_t exe_id = 0;
 	struct pva_dma_task_buffer_info_s *buff_info = &task->task_buff_info[desc_id];
 	int hwgen = task->pva->version;
 
 	nvpva_dbg_fn(task->pva, "");
 
+	exe_id = get_sym_exe_id(task);
 	switch (umd_dma_desc->srcTransferMode) {
 	case DMA_DESC_SRC_XFER_L2RAM:
 		/*
@@ -329,7 +344,7 @@ patch_dma_desc_address(struct pva_submit_task *task,
 			goto out;
 		}
 
-		err = pva_get_sym_offset(&task->client->elf_ctx, task->exe_id1,
+		err = pva_get_sym_offset(&task->client->elf_ctx, exe_id,
 					 umd_dma_desc->srcPtr, &addr, &size);
 		if (err) {
 			err = -EINVAL;
@@ -371,7 +386,7 @@ patch_dma_desc_address(struct pva_submit_task *task,
 
 		/* calculate symbol address */
 		/* TODO: check VPUC handling in ELF segment */
-		err = pva_get_sym_offset(&task->client->elf_ctx, task->exe_id1,
+		err = pva_get_sym_offset(&task->client->elf_ctx, exe_id,
 					 umd_dma_desc->srcPtr, &addr, &size);
 		if (err) {
 			task_err(task, "ERROR: Invalid offset or address");
@@ -516,7 +531,7 @@ patch_dma_desc_address(struct pva_submit_task *task,
 			goto out;
 		}
 
-		err = pva_get_sym_offset(&task->client->elf_ctx, task->exe_id1,
+		err = pva_get_sym_offset(&task->client->elf_ctx, exe_id,
 					 umd_dma_desc->dstPtr, &addr, &size);
 		if (err) {
 			err = -EINVAL;
@@ -528,7 +543,7 @@ patch_dma_desc_address(struct pva_submit_task *task,
 
 		if (umd_dma_desc->dst2Ptr != NVPVA_INVALID_SYMBOL_ID) {
 			err = pva_get_sym_offset(&task->client->elf_ctx,
-						 task->exe_id1,
+						 exe_id,
 						 umd_dma_desc->dst2Ptr,
 						 &addr2,
 						 &size2);
@@ -763,6 +778,8 @@ static int32_t nvpva_task_dma_desc_mapping(struct pva_submit_task *task,
 	bool is_misr;
 	u8 num_dma_descriptors = task->num_dma_descriptors;
 	u8 *num_dma_desc_processed = &task->num_dma_desc_processed;
+	uint16_t exe_id = 0;
+
 	nvpva_dbg_fn(task->pva, "");
 
 	desc_num = *did;
@@ -885,9 +902,10 @@ static int32_t nvpva_task_dma_desc_mapping(struct pva_submit_task *task,
 		/* DMA_DESC_PY */
 		dma_desc->py = (uint8_t)umd_dma_desc->py;
 		/* DMA_DESC_FRDA */
+		exe_id = get_sym_exe_id(task);
 		if (umd_dma_desc->dst2Ptr != NVPVA_INVALID_SYMBOL_ID) {
 			err = pva_get_sym_offset(&task->client->elf_ctx,
-						 task->exe_id1,
+						 exe_id,
 						 umd_dma_desc->dst2Ptr,
 						 &addr,
 						 &size);

drivers/video/tegra/host/pva/pva_ioctl.c

@@ -170,6 +170,7 @@ static int pva_copy_task(struct nvpva_ioctl_task *ioctl_task,
 	 * copy them.
 	 */
 	task->exe_id1 = ioctl_task->exe_id1;
+	task->exe_id2 = ioctl_task->exe_id2;
 	task->stream_id = ioctl_task->stream_id;
 	task->prog_id = ioctl_task->prog_id;
 	task->l2_alloc_size = ioctl_task->l2_alloc_size;

drivers/video/tegra/host/pva/pva_queue.c

@@ -87,6 +87,7 @@ static void pva_task_dump(struct pva_submit_task *task)
 	int i;
 
 	nvpva_dbg_info(task->pva, "task=%p, exe_id1=%u", task, task->exe_id1);
+	nvpva_dbg_info(task->pva, "task=%p, exe_id2=%u", task, task->exe_id2);
 
 	for (i = 0; i < task->num_input_task_status; i++)
 		nvpva_dbg_info(task->pva, "input task status %d: pin_id=%u, offset=%u", i,
@@ -562,11 +563,19 @@ pva_task_write_vpu_parameter(struct pva_submit_task *task,
 	u32 tail_count = 0U;
 	struct pva_vpu_parameters_s *hw_task_param_list;
 
-	if ((task->exe_id1 == NVPVA_NOOP_EXE_ID) || (task->num_symbols == 0U))
+	if (task->num_symbols == 0U)
+		goto out;
+
+	if ((task->exe_id1 == NVPVA_NOOP_EXE_ID) && (task->exe_id2 == NVPVA_NOOP_EXE_ID))
 		goto out;
 
 	tail_index = ((u32)task->num_symbols - 1U);
+
+	if (task->exe_id1 != NVPVA_NOOP_EXE_ID) {
 		elf = get_elf_image(&task->client->elf_ctx, task->exe_id1);
+	} else if (task->exe_id2 != NVPVA_NOOP_EXE_ID)
+		elf = get_elf_image(&task->client->elf_ctx, task->exe_id2);
+
 	if (task->num_symbols > elf->num_symbols) {
 		task_err(task, "invalid number of symbols");
 		err = -EINVAL;
@@ -706,6 +715,8 @@ pva_task_write_vpu_parameter(struct pva_submit_task *task,
 	hw_task->task.parameter_info_base = task->dma_addr
 					    + offsetof(struct pva_hw_task, param_info);
 
+	if (task->exe_id1 != NVPVA_NOOP_EXE_ID) {
+
 		err = pva_task_acquire_ref_vpu_app(&task->client->elf_ctx,
 				task->exe_id1);
 		if (err) {
@@ -713,6 +724,18 @@ pva_task_write_vpu_parameter(struct pva_submit_task *task,
 				"unable to acquire ref count for app with id = %u",
 				task->exe_id1);
 		}
+	}
+
+	if (task->exe_id2 != NVPVA_NOOP_EXE_ID) {
+
+		err = pva_task_acquire_ref_vpu_app(&task->client->elf_ctx,
+				task->exe_id2);
+		if (err) {
+			task_err(task,
+				"unable to acquire ref count for app with id = %u",
+				task->exe_id2);
+		}
+	}
 
 	task->pinned_app = true;
 out:
@@ -760,7 +783,13 @@ static int pva_task_write(struct pva_submit_task *task)
 
 	if (!pva_vpu_elf_is_registered(&task->client->elf_ctx, task->exe_id1) &&
 	    (task->exe_id1 != NVPVA_NOOP_EXE_ID)) {
-		task_err(task, "invalid exe id: %d", task->exe_id1);
+		task_err(task, "invalid VPU exe id: %d", task->exe_id1);
+		return -EINVAL;
+	}
+
+	if (!pva_vpu_elf_is_registered(&task->client->elf_ctx, task->exe_id2) &&
+	    (task->exe_id2 != NVPVA_NOOP_EXE_ID)) {
+		task_err(task, "invalid exe id: %d", task->exe_id2);
 		return -EINVAL;
 	}
 	/* Task start from the memory base */
@@ -810,6 +839,9 @@ static int pva_task_write(struct pva_submit_task *task)
 	hw_task->task.bin_info =
 	    phys_get_bin_info(&task->client->elf_ctx, task->exe_id1);
 
+	hw_task->task.bin_info2 =
+	    phys_get_bin_info(&task->client->elf_ctx, task->exe_id2);
+
 	if (task->stdout) {
 		hw_task->stdout_cb_info.buffer = task->stdout->buffer_addr;
 		hw_task->stdout_cb_info.head = task->stdout->head_addr;
@@ -927,9 +959,12 @@ void pva_task_free(struct kref *ref)
 	mutex_unlock(&my_queue->tail_lock);
 
 	pva_task_unpin_mem(task);
-	if (task->pinned_app)
+	if (task->pinned_app) {
 		pva_task_release_ref_vpu_app(&task->client->elf_ctx,
 						     task->exe_id1);
+		pva_task_release_ref_vpu_app(&task->client->elf_ctx,
+						     task->exe_id2);
+	}
 
 	nvhost_module_idle(task->pva->pdev);
 	nvpva_client_context_put(task->client);

drivers/video/tegra/host/pva/pva_queue.h

@@ -110,6 +110,7 @@ struct pva_submit_task {
 	u32 exe_id1;
 	u64 stream_id;
 	u64 prog_id;
+	u32 exe_id2;
 
 	u32 l2_alloc_size; /* Not applicable for Xavier */
 	struct pva_cb *stdout;

drivers/video/tegra/host/pva/pva_vpu_exe.c

@@ -909,7 +909,7 @@ pva_get_sym_tab(struct nvpva_elf_context *d,
 
 	return 0;
 }
-int32_t pva_get_sym_info(struct nvpva_elf_context *d, uint16_t vpu_exe_id,
+int32_t pva_get_sym_info(struct nvpva_elf_context *d, uint16_t exe_id,
 		       const char *sym_name, struct pva_elf_symbol *symbol)
 {
 	struct pva_elf_image *elf;
@@ -917,7 +917,7 @@ int32_t pva_get_sym_info(struct nvpva_elf_context *d, uint16_t vpu_exe_id,
 	int32_t err = 0;
 	size_t strSize = strnlen(sym_name, ELF_MAX_SYMBOL_LENGTH);
 
-	elf = get_elf_image(d, vpu_exe_id);
+	elf = get_elf_image(d, exe_id);
 	for (i = 0; i < elf->num_symbols; i++) {
 		if (strncmp(sym_name, elf->sym[i].symbol_name, strSize) == 0) {
 			symbol->symbolID = elf->sym[i].symbolID;

drivers/video/tegra/host/pva/pva_vpu_exe.h

@@ -151,7 +151,7 @@ struct nvpva_elf_context {
  * Get Symbol info given the symbol name from a vpu app
  *
  * @param d		Pointer to Elf Context
- * @param vpu_exe_id	ID of the VPU app
+ * @param exe_id	ID of the VPU app
  * @param *sym_name	String containing Name of the symbol
  * @param *symbol	symbol information
 
@@ -159,7 +159,7 @@ struct nvpva_elf_context {
  *			When -EINVAL is returned, ignore values in id and
  *			sym_size
  */
-int32_t pva_get_sym_info(struct nvpva_elf_context *d, uint16_t vpu_exe_id,
+int32_t pva_get_sym_info(struct nvpva_elf_context *d, uint16_t exe_id,
 		       const char *sym_name, struct pva_elf_symbol *symbol);
 
 /**

include/uapi/linux/nvpva_ioctl.h

@@ -475,6 +475,7 @@ struct nvpva_ioctl_task {
 	uint64_t prog_id;
 	uint64_t stream_id;
 	uint16_t exe_id1;
+	uint16_t exe_id2;
 	uint32_t flags;
 	uint32_t l2_alloc_size; /* Not applicable for Xavier */
 	struct nvpva_ioctl_part prefences;