pva: mirror from gitlab cv/pva-sys-sw

Gitlab commit 689815042dacd ("Fix syncpt related crash durin...")

Changes since last deployment:

- Fix syncpt related crash during SC7 resume
- Add SC7 simulation and test
- tests: docs: fix mat_add test specs
- compat: bugfix: prevent fence override
- Add AI code reviewer to CI

Change-Id: I781ac9379787adb9dbdde55cd3fdec5729b08a95
Signed-off-by: nanwa <nanwa@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nv-oot/+/3358344
GVS: buildbot_gerritrpt <buildbot_gerritrpt@nvidia.com>
Reviewed-by: Mohnish Jain <mohnishj@nvidia.com>

drivers/video/tegra/host/pva/src/kmd/common/pva_kmd_context.c

@@ -96,7 +96,8 @@ static enum pva_error notify_fw_context_init(struct pva_kmd_context *ctx)
 		ctx->ctx_resource_table.table_mem->iova,
 		ctx->ctx_resource_table.n_entries);
 
-	syncpt_info = pva_kmd_queue_get_rw_syncpt_info(ctx, ctx->ccq_id);
+	syncpt_info = pva_kmd_queue_get_rw_syncpt_info(
+		ctx->pva, PVA_PRIV_CCQ_ID, ctx->ccq_id);
 	pva_kmd_set_cmd_init_queue(
 		queue_cmd, PVA_PRIV_CCQ_ID,
 		ctx->ccq_id, /* For privileged queues, queue ID == user CCQ ID*/

drivers/video/tegra/host/pva/src/kmd/common/pva_kmd_debugfs.c

@@ -319,6 +319,60 @@ static int64_t get_fw_debug_log_level(struct pva_kmd_device *dev,
 					(uint64_t)formatted_len);
 }
 
+static int64_t write_simulate_sc7(struct pva_kmd_device *pva, void *file_data,
+				  const uint8_t *in_buffer, uint64_t offset,
+				  uint64_t size)
+{
+	uint8_t buf = 0;
+	enum pva_error err;
+	unsigned long ret;
+
+	if ((offset != 0) || (size < 1)) {
+		return -EINVAL;
+	}
+
+	ret = pva_kmd_copy_data_from_user(&buf, in_buffer, 1);
+	if (ret != 0) {
+		pva_kmd_log_err(
+			"SC7 simulation: failed to copy data from user");
+		return -EFAULT;
+	}
+
+	if (buf == '1') {
+		if (pva->debugfs_context.entered_sc7 == 0) {
+			err = pva_kmd_simulate_enter_sc7(pva);
+			if (err != PVA_SUCCESS) {
+				return -EFAULT;
+			}
+			pva->debugfs_context.entered_sc7 = 1;
+		}
+	} else if (buf == '0') {
+		if (pva->debugfs_context.entered_sc7 == 1) {
+			err = pva_kmd_simulate_exit_sc7(pva);
+			if (err != PVA_SUCCESS) {
+				return -EFAULT;
+			}
+			pva->debugfs_context.entered_sc7 = 0;
+		}
+	} else {
+		pva_kmd_log_err(
+			"SC7 simulation: invalid input; Must be 0 or 1");
+		return -EINVAL;
+	}
+
+	return size;
+}
+
+static int64_t read_simulate_sc7(struct pva_kmd_device *pva, void *file_data,
+				 uint8_t *out_buffer, uint64_t offset,
+				 uint64_t size)
+{
+	char buf;
+	buf = pva->debugfs_context.entered_sc7 ? '1' : '0';
+
+	return read_from_buffer_to_user(out_buffer, size, offset, &buf, 1);
+}
+
 enum pva_error pva_kmd_debugfs_create_nodes(struct pva_kmd_device *pva)
 {
 	static const char *vpu_ocd_names[NUM_VPU_BLOCKS] = { "ocd_vpu0_v3",
@@ -420,6 +474,16 @@ enum pva_error pva_kmd_debugfs_create_nodes(struct pva_kmd_device *pva)
 	pva_kmd_device_init_profiler(pva);
 	pva_kmd_device_init_tegra_stats(pva);
 
+	pva->debugfs_context.simulate_sc7_fops.read = &read_simulate_sc7;
+	pva->debugfs_context.simulate_sc7_fops.write = &write_simulate_sc7;
+	pva->debugfs_context.simulate_sc7_fops.pdev = pva;
+	err = pva_kmd_debugfs_create_file(
+		pva, "simulate_sc7", &pva->debugfs_context.simulate_sc7_fops);
+	if (err != PVA_SUCCESS) {
+		pva_kmd_log_err("Failed to create simulate_sc7 debugfs file");
+		return err;
+	}
+
 	return PVA_SUCCESS;
 }
 

drivers/video/tegra/host/pva/src/kmd/common/pva_kmd_debugfs.h

@@ -24,6 +24,7 @@ struct pva_kmd_debugfs_context {
 	bool stats_enable;
 	bool vpu_debug;
 	bool vpu_print_enable;
+	bool entered_sc7;
 	char *allowlist_path;
 	uint32_t profiling_level;
 	struct pva_kmd_file_ops vpu_fops;
@@ -35,6 +36,7 @@ struct pva_kmd_debugfs_context {
 	struct pva_kmd_file_ops vpu_ocd_fops[NUM_VPU_BLOCKS];
 	struct pva_kmd_fw_profiling_config g_fw_profiling_config;
 	struct pva_kmd_file_ops fw_debug_log_level_fops;
+	struct pva_kmd_file_ops simulate_sc7_fops;
 };
 
 enum pva_error pva_kmd_debugfs_create_nodes(struct pva_kmd_device *dev);

drivers/video/tegra/host/pva/src/kmd/common/pva_kmd_op_handler.c

@@ -479,7 +479,8 @@ pva_kmd_op_queue_create(struct pva_kmd_context *ctx, const void *input_buffer,
 		goto out;
 	}
 
-	syncpt_info = pva_kmd_queue_get_rw_syncpt_info(ctx, queue_id);
+	syncpt_info = pva_kmd_queue_get_rw_syncpt_info(ctx->pva, ctx->ccq_id,
+						       queue_id);
 	queue_out_args.error = err;
 	queue_out_args.queue_id = queue_id;
 	queue_out_args.syncpt_id = syncpt_info->syncpt_id;

drivers/video/tegra/host/pva/src/kmd/common/pva_kmd_pm.c

@@ -138,7 +138,7 @@ enum pva_error pva_kmd_complete_resume(struct pva_kmd_device *pva)
 				"PVA: Resume priv queue for context %d\n",
 				ctx->ccq_id);
 			syncpt_info = pva_kmd_queue_get_rw_syncpt_info(
-				PVA_PRIV_CCQ_ID, ctx->ccq_id);
+				pva, PVA_PRIV_CCQ_ID, ctx->ccq_id);
 			pva_kmd_set_cmd_init_queue(
 				queue_cmd, PVA_PRIV_CCQ_ID,
 				ctx->ccq_id, /* For privileged queues, queue ID == user CCQ ID*/
@@ -168,7 +168,8 @@ enum pva_error pva_kmd_complete_resume(struct pva_kmd_device *pva)
 
 					syncpt_info =
 						pva_kmd_queue_get_rw_syncpt_info(
-							ctx, queue->queue_id);
+							pva, ctx->ccq_id,
+							queue->queue_id);
 					pva_kmd_set_cmd_init_queue(
 						queue_cmd, queue->ccq_id,
 						queue->queue_id,

drivers/video/tegra/host/pva/src/kmd/common/pva_kmd_queue.c

@@ -135,7 +135,8 @@ enum pva_error pva_kmd_queue_create(struct pva_kmd_context *ctx,
 		goto err_free_kmd_memory;
 	}
 
-	syncpt_info = pva_kmd_queue_get_rw_syncpt_info(ctx, queue->queue_id);
+	syncpt_info = pva_kmd_queue_get_rw_syncpt_info(ctx->pva, ctx->ccq_id,
+						       queue->queue_id);
 	pva_kmd_set_cmd_init_queue(&cmd, queue->ccq_id, queue->queue_id,
 				   queue->queue_memory->iova,
 				   queue->max_num_submit,
@@ -197,12 +198,13 @@ unlock:
 }
 
 const struct pva_syncpt_rw_info *
-pva_kmd_queue_get_rw_syncpt_info(struct pva_kmd_context *ctx, uint8_t queue_id)
+pva_kmd_queue_get_rw_syncpt_info(struct pva_kmd_device *pva, uint8_t ccq_id,
+				 uint8_t queue_id)
 {
 	uint8_t ctx_offset =
-		safe_mulu32(ctx->ccq_id, PVA_NUM_RW_SYNCPTS_PER_CONTEXT);
+		safe_mulu32(ccq_id, PVA_NUM_RW_SYNCPTS_PER_CONTEXT);
 	uint32_t syncpt_index = safe_addu32(ctx_offset, queue_id);
 
 	ASSERT(syncpt_index < PVA_NUM_RW_SYNCPTS);
-	return &ctx->pva->rw_syncpts[syncpt_index];
+	return &pva->rw_syncpts[syncpt_index];
 }

drivers/video/tegra/host/pva/src/kmd/common/pva_kmd_queue.h

@@ -35,6 +35,7 @@ pva_kmd_queue_submit(struct pva_kmd_queue *queue,
 uint32_t pva_kmd_queue_space(struct pva_kmd_queue *queue);
 
 const struct pva_syncpt_rw_info *
-pva_kmd_queue_get_rw_syncpt_info(struct pva_kmd_context *ctx, uint8_t queue_id);
+pva_kmd_queue_get_rw_syncpt_info(struct pva_kmd_device *pva, uint8_t ccq_id,
+				 uint8_t queue_id);
 
 #endif // PVA_KMD_QUEUE_H

drivers/video/tegra/host/pva/src/kmd/common/shim/pva_kmd_shim_debugfs.h

@@ -19,4 +19,7 @@ unsigned long pva_kmd_copy_data_to_user(void *to, const void *from,
 					unsigned long size);
 unsigned long pva_kmd_strtol(const char *str, int base);
 
+enum pva_error pva_kmd_simulate_enter_sc7(struct pva_kmd_device *pva);
+enum pva_error pva_kmd_simulate_exit_sc7(struct pva_kmd_device *pva);
+
 #endif //PVA_KMD_SHIM_DEBUGFS_H

drivers/video/tegra/host/pva/src/kmd/linux/pva_kmd_linux_driver.c

@@ -667,6 +667,59 @@ done:
 	return;
 }
 
+enum pva_error pva_kmd_simulate_enter_sc7(struct pva_kmd_device *pva)
+{
+	struct pva_kmd_linux_device_data *device_data;
+	struct nvpva_device_data *device_props;
+	struct device *dev;
+	int ret;
+
+	device_data = pva_kmd_linux_device_get_data(pva);
+	device_props = device_data->pva_device_properties;
+	dev = &device_props->pdev->dev;
+
+	// The PM core increases the device usage count before calling prepare, so
+	// we need to emulate this behavior as well.
+	pm_runtime_get_noresume(dev);
+
+	ret = pva_kmd_linux_device_prepare_suspend(dev);
+	if (ret != 0) {
+		pva_kmd_log_err("SC7 simulation: prepare suspend failed");
+		return PVA_INTERNAL;
+	}
+	ret = pva_kmd_linux_device_suspend(dev);
+	if (ret != 0) {
+		pva_kmd_log_err("SC7 simulation: suspend failed");
+		return PVA_INTERNAL;
+	}
+
+	return PVA_SUCCESS;
+}
+
+enum pva_error pva_kmd_simulate_exit_sc7(struct pva_kmd_device *pva)
+{
+	struct pva_kmd_linux_device_data *device_data =
+		pva_kmd_linux_device_get_data(pva);
+	struct nvpva_device_data *device_props =
+		device_data->pva_device_properties;
+	struct device *dev = &device_props->pdev->dev;
+	int ret;
+
+	ret = pva_kmd_linux_device_resume(dev);
+	if (ret != 0) {
+		pva_kmd_log_err("SC7 simulation: resume failed");
+		return PVA_INTERNAL;
+	}
+
+	pva_kmd_linux_device_complete_resume(dev);
+
+	// The PM core decreases the device usage count after calling complete, so
+	// we need to emulate this behavior as well.
+	pm_runtime_put(dev);
+
+	return PVA_SUCCESS;
+}
+
 static const struct dev_pm_ops pva_kmd_linux_pm_ops = {
 	SET_SYSTEM_SLEEP_PM_OPS(pva_kmd_linux_device_suspend,
 				pva_kmd_linux_device_resume)

drivers/video/tegra/host/pva/src/kmd/linux/pva_kmd_linux_event_trace.c

@@ -4,7 +4,6 @@
 #include "pva_kmd_linux_device.h"
 #define CREATE_TRACE_POINTS
 #include "trace/events/nvpva_ftrace.h"
-#include <linux/nvhost.h>
 
 static uint32_t get_job_id(uint32_t queue_id, uint64_t submit_id)
 {