1 files changed, 573 insertions, 195 deletions

diff --git a/drivers/accel/ivpu/ivpu_job.c b/drivers/accel/ivpu/ivpu_job.c
index 7149312f16e1..4f8564e2878a 100644
--- a/drivers/accel/ivpu/ivpu_job.c
+++ b/drivers/accel/ivpu/ivpu_job.c
@@ -1,6 +1,6 @@
 // SPDX-License-Identifier: GPL-2.0-only
 /*
- * Copyright (C) 2020-2024 Intel Corporation
+ * Copyright (C) 2020-2025 Intel Corporation
  */
 
 #include <drm/drm_file.h>
@@ -8,6 +8,7 @@
 #include <linux/bitfield.h>
 #include <linux/highmem.h>
 #include <linux/pci.h>
+#include <linux/pm_runtime.h>
 #include <linux/module.h>
 #include <uapi/drm/ivpu_accel.h>
 
@@ -17,6 +18,7 @@
 #include "ivpu_ipc.h"
 #include "ivpu_job.h"
 #include "ivpu_jsm_msg.h"
+#include "ivpu_mmu.h"
 #include "ivpu_pm.h"
 #include "ivpu_trace.h"
 #include "vpu_boot_api.h"
@@ -32,22 +34,20 @@ static void ivpu_cmdq_ring_db(struct ivpu_device *vdev, struct ivpu_cmdq *cmdq)
 static int ivpu_preemption_buffers_create(struct ivpu_device *vdev,
 					  struct ivpu_file_priv *file_priv, struct ivpu_cmdq *cmdq)
 {
-	u64 primary_size = ALIGN(vdev->fw->primary_preempt_buf_size, PAGE_SIZE);
-	u64 secondary_size = ALIGN(vdev->fw->secondary_preempt_buf_size, PAGE_SIZE);
-
-	if (vdev->fw->sched_mode != VPU_SCHEDULING_MODE_HW ||
-	    ivpu_test_mode & IVPU_TEST_MODE_MIP_DISABLE)
+	if (ivpu_fw_preempt_buf_size(vdev) == 0)
 		return 0;
 
 	cmdq->primary_preempt_buf = ivpu_bo_create(vdev, &file_priv->ctx, &vdev->hw->ranges.user,
-						   primary_size, DRM_IVPU_BO_WC);
+						   vdev->fw->primary_preempt_buf_size,
+						   DRM_IVPU_BO_WC);
 	if (!cmdq->primary_preempt_buf) {
 		ivpu_err(vdev, "Failed to create primary preemption buffer\n");
 		return -ENOMEM;
 	}
 
 	cmdq->secondary_preempt_buf = ivpu_bo_create(vdev, &file_priv->ctx, &vdev->hw->ranges.dma,
-						     secondary_size, DRM_IVPU_BO_WC);
+						     vdev->fw->secondary_preempt_buf_size,
+						     DRM_IVPU_BO_WC);
 	if (!cmdq->secondary_preempt_buf) {
 		ivpu_err(vdev, "Failed to create secondary preemption buffer\n");
 		goto err_free_primary;
@@ -64,69 +64,135 @@ err_free_primary:
 static void ivpu_preemption_buffers_free(struct ivpu_device *vdev,
 					 struct ivpu_file_priv *file_priv, struct ivpu_cmdq *cmdq)
 {
-	if (vdev->fw->sched_mode != VPU_SCHEDULING_MODE_HW)
-		return;
-
 	if (cmdq->primary_preempt_buf)
 		ivpu_bo_free(cmdq->primary_preempt_buf);
 	if (cmdq->secondary_preempt_buf)
 		ivpu_bo_free(cmdq->secondary_preempt_buf);
 }
 
+static int ivpu_preemption_job_init(struct ivpu_device *vdev, struct ivpu_file_priv *file_priv,
+				    struct ivpu_cmdq *cmdq, struct ivpu_job *job)
+{
+	int ret;
+
+	/* Use preemption buffer provided by the user space */
+	if (job->primary_preempt_buf)
+		return 0;
+
+	if (!cmdq->primary_preempt_buf) {
+		/* Allocate per command queue preemption buffers */
+		ret = ivpu_preemption_buffers_create(vdev, file_priv, cmdq);
+		if (ret)
+			return ret;
+	}
+
+	/* Use preemption buffers allocated by the kernel */
+	job->primary_preempt_buf = cmdq->primary_preempt_buf;
+	job->secondary_preempt_buf = cmdq->secondary_preempt_buf;
+
+	return 0;
+}
+
 static struct ivpu_cmdq *ivpu_cmdq_alloc(struct ivpu_file_priv *file_priv)
 {
 	struct ivpu_device *vdev = file_priv->vdev;
 	struct ivpu_cmdq *cmdq;
-	int ret;
 
 	cmdq = kzalloc(sizeof(*cmdq), GFP_KERNEL);
 	if (!cmdq)
 		return NULL;
 
-	ret = xa_alloc_cyclic(&vdev->db_xa, &cmdq->db_id, NULL, vdev->db_limit, &vdev->db_next,
-			      GFP_KERNEL);
-	if (ret < 0) {
-		ivpu_err(vdev, "Failed to allocate doorbell id: %d\n", ret);
-		goto err_free_cmdq;
-	}
-
-	ret = xa_alloc_cyclic(&file_priv->cmdq_xa, &cmdq->id, cmdq, file_priv->cmdq_limit,
-			      &file_priv->cmdq_id_next, GFP_KERNEL);
-	if (ret < 0) {
-		ivpu_err(vdev, "Failed to allocate command queue id: %d\n", ret);
-		goto err_erase_db_xa;
-	}
-
 	cmdq->mem = ivpu_bo_create_global(vdev, SZ_4K, DRM_IVPU_BO_WC | DRM_IVPU_BO_MAPPABLE);
 	if (!cmdq->mem)
-		goto err_erase_cmdq_xa;
-
-	ret = ivpu_preemption_buffers_create(vdev, file_priv, cmdq);
-	if (ret)
-		ivpu_warn(vdev, "Failed to allocate preemption buffers, preemption limited\n");
+		goto err_free_cmdq;
 
 	return cmdq;
 
-err_erase_cmdq_xa:
-	xa_erase(&file_priv->cmdq_xa, cmdq->id);
-err_erase_db_xa:
-	xa_erase(&vdev->db_xa, cmdq->db_id);
 err_free_cmdq:
 	kfree(cmdq);
 	return NULL;
 }
 
-static void ivpu_cmdq_free(struct ivpu_file_priv *file_priv, struct ivpu_cmdq *cmdq)
+/**
+ * ivpu_cmdq_get_entry_count - Calculate the number of entries in the command queue.
+ * @cmdq: Pointer to the command queue structure.
+ *
+ * Returns the number of entries that can fit in the command queue memory.
+ */
+static inline u32 ivpu_cmdq_get_entry_count(struct ivpu_cmdq *cmdq)
 {
-	if (!cmdq)
-		return;
+	size_t size = ivpu_bo_size(cmdq->mem) - sizeof(struct vpu_job_queue_header);
+
+	return size / sizeof(struct vpu_job_queue_entry);
+}
+
+/**
+ * ivpu_cmdq_get_flags - Get command queue flags based on input flags and test mode.
+ * @vdev: Pointer to the ivpu device structure.
+ * @flags: Input flags to determine the command queue flags.
+ *
+ * Returns the calculated command queue flags, considering both the input flags
+ * and the current test mode settings.
+ */
+static u32 ivpu_cmdq_get_flags(struct ivpu_device *vdev, u32 flags)
+{
+	u32 cmdq_flags = 0;
+
+	if ((flags & DRM_IVPU_CMDQ_FLAG_TURBO) && (ivpu_hw_ip_gen(vdev) >= IVPU_HW_IP_40XX))
+		cmdq_flags |= VPU_JOB_QUEUE_FLAGS_TURBO_MODE;
 
+	/* Test mode can override the TURBO flag coming from the application */
+	if (ivpu_test_mode & IVPU_TEST_MODE_TURBO_ENABLE)
+		cmdq_flags |= VPU_JOB_QUEUE_FLAGS_TURBO_MODE;
+	if (ivpu_test_mode & IVPU_TEST_MODE_TURBO_DISABLE)
+		cmdq_flags &= ~VPU_JOB_QUEUE_FLAGS_TURBO_MODE;
+
+	return cmdq_flags;
+}
+
+static void ivpu_cmdq_free(struct ivpu_file_priv *file_priv, struct ivpu_cmdq *cmdq)
+{
 	ivpu_preemption_buffers_free(file_priv->vdev, file_priv, cmdq);
 	ivpu_bo_free(cmdq->mem);
-	xa_erase(&file_priv->vdev->db_xa, cmdq->db_id);
 	kfree(cmdq);
 }
 
+static struct ivpu_cmdq *ivpu_cmdq_create(struct ivpu_file_priv *file_priv, u8 priority, u32 flags)
+{
+	struct ivpu_device *vdev = file_priv->vdev;
+	struct ivpu_cmdq *cmdq = NULL;
+	int ret;
+
+	lockdep_assert_held(&file_priv->lock);
+
+	cmdq = ivpu_cmdq_alloc(file_priv);
+	if (!cmdq) {
+		ivpu_err(vdev, "Failed to allocate command queue\n");
+		return NULL;
+	}
+	ret = xa_alloc_cyclic(&file_priv->cmdq_xa, &cmdq->id, cmdq, file_priv->cmdq_limit,
+			      &file_priv->cmdq_id_next, GFP_KERNEL);
+	if (ret < 0) {
+		ivpu_err(vdev, "Failed to allocate command queue ID: %d\n", ret);
+		goto err_free_cmdq;
+	}
+
+	cmdq->entry_count = ivpu_cmdq_get_entry_count(cmdq);
+	cmdq->priority = priority;
+
+	cmdq->jobq = (struct vpu_job_queue *)ivpu_bo_vaddr(cmdq->mem);
+	cmdq->jobq->header.engine_idx = VPU_ENGINE_COMPUTE;
+	cmdq->jobq->header.flags = ivpu_cmdq_get_flags(vdev, flags);
+
+	ivpu_dbg(vdev, JOB, "Command queue %d created, ctx %d, flags 0x%08x\n",
+		 cmdq->id, file_priv->ctx.id, cmdq->jobq->header.flags);
+	return cmdq;
+
+err_free_cmdq:
+	ivpu_cmdq_free(file_priv, cmdq);
+	return NULL;
+}
+
 static int ivpu_hws_cmdq_init(struct ivpu_file_priv *file_priv, struct ivpu_cmdq *cmdq, u16 engine,
 			      u8 priority)
 {
@@ -152,6 +218,13 @@ static int ivpu_register_db(struct ivpu_file_priv *file_priv, struct ivpu_cmdq *
 	struct ivpu_device *vdev = file_priv->vdev;
 	int ret;
 
+	ret = xa_alloc_cyclic(&vdev->db_xa, &cmdq->db_id, NULL, vdev->db_limit, &vdev->db_next,
+			      GFP_KERNEL);
+	if (ret < 0) {
+		ivpu_err(vdev, "Failed to allocate doorbell ID: %d\n", ret);
+		return ret;
+	}
+
 	if (vdev->fw->sched_mode == VPU_SCHEDULING_MODE_HW)
 		ret = ivpu_jsm_hws_register_db(vdev, file_priv->ctx.id, cmdq->id, cmdq->db_id,
 					       cmdq->mem->vpu_addr, ivpu_bo_size(cmdq->mem));
@@ -159,42 +232,39 @@ static int ivpu_register_db(struct ivpu_file_priv *file_priv, struct ivpu_cmdq *
 		ret = ivpu_jsm_register_db(vdev, file_priv->ctx.id, cmdq->db_id,
 					   cmdq->mem->vpu_addr, ivpu_bo_size(cmdq->mem));
 
-	if (!ret)
-		ivpu_dbg(vdev, JOB, "DB %d registered to cmdq %d ctx %d\n",
-			 cmdq->db_id, cmdq->id, file_priv->ctx.id);
+	if (!ret) {
+		ivpu_dbg(vdev, JOB, "DB %d registered to cmdq %d ctx %d priority %d\n",
+			 cmdq->db_id, cmdq->id, file_priv->ctx.id, cmdq->priority);
+	} else {
+		xa_erase(&vdev->db_xa, cmdq->db_id);
+		cmdq->db_id = 0;
+	}
 
 	return ret;
 }
 
-static int
-ivpu_cmdq_init(struct ivpu_file_priv *file_priv, struct ivpu_cmdq *cmdq, u8 priority)
+static void ivpu_cmdq_jobq_reset(struct ivpu_device *vdev, struct vpu_job_queue *jobq)
+{
+	jobq->header.head = 0;
+	jobq->header.tail = 0;
+
+	wmb(); /* Flush WC buffer for jobq->header */
+}
+
+static int ivpu_cmdq_register(struct ivpu_file_priv *file_priv, struct ivpu_cmdq *cmdq)
 {
 	struct ivpu_device *vdev = file_priv->vdev;
-	struct vpu_job_queue_header *jobq_header;
 	int ret;
 
 	lockdep_assert_held(&file_priv->lock);
 
-	if (cmdq->db_registered)
+	if (cmdq->db_id)
 		return 0;
 
-	cmdq->entry_count = (u32)((ivpu_bo_size(cmdq->mem) - sizeof(struct vpu_job_queue_header)) /
-				  sizeof(struct vpu_job_queue_entry));
-
-	cmdq->jobq = (struct vpu_job_queue *)ivpu_bo_vaddr(cmdq->mem);
-	jobq_header = &cmdq->jobq->header;
-	jobq_header->engine_idx = VPU_ENGINE_COMPUTE;
-	jobq_header->head = 0;
-	jobq_header->tail = 0;
-	if (ivpu_test_mode & IVPU_TEST_MODE_TURBO) {
-		ivpu_dbg(vdev, JOB, "Turbo mode enabled");
-		jobq_header->flags = VPU_JOB_QUEUE_FLAGS_TURBO_MODE;
-	}
-
-	wmb(); /* Flush WC buffer for jobq->header */
+	ivpu_cmdq_jobq_reset(vdev, cmdq->jobq);
 
 	if (vdev->fw->sched_mode == VPU_SCHEDULING_MODE_HW) {
-		ret = ivpu_hws_cmdq_init(file_priv, cmdq, VPU_ENGINE_COMPUTE, priority);
+		ret = ivpu_hws_cmdq_init(file_priv, cmdq, VPU_ENGINE_COMPUTE, cmdq->priority);
 		if (ret)
 			return ret;
 	}
@@ -203,58 +273,84 @@ ivpu_cmdq_init(struct ivpu_file_priv *file_priv, struct ivpu_cmdq *cmdq, u8 prio
 	if (ret)
 		return ret;
 
-	cmdq->db_registered = true;
-
 	return 0;
 }
 
-static int ivpu_cmdq_fini(struct ivpu_file_priv *file_priv, struct ivpu_cmdq *cmdq)
+static int ivpu_cmdq_unregister(struct ivpu_file_priv *file_priv, struct ivpu_cmdq *cmdq)
 {
 	struct ivpu_device *vdev = file_priv->vdev;
 	int ret;
 
 	lockdep_assert_held(&file_priv->lock);
 
-	if (!cmdq->db_registered)
+	if (!cmdq->db_id)
 		return 0;
 
-	cmdq->db_registered = false;
+	ret = ivpu_jsm_unregister_db(vdev, cmdq->db_id);
+	if (!ret)
+		ivpu_dbg(vdev, JOB, "DB %d unregistered\n", cmdq->db_id);
 
 	if (vdev->fw->sched_mode == VPU_SCHEDULING_MODE_HW) {
 		ret = ivpu_jsm_hws_destroy_cmdq(vdev, file_priv->ctx.id, cmdq->id);
 		if (!ret)
-			ivpu_dbg(vdev, JOB, "Command queue %d destroyed\n", cmdq->id);
+			ivpu_dbg(vdev, JOB, "Command queue %d destroyed, ctx %d\n",
+				 cmdq->id, file_priv->ctx.id);
 	}
 
-	ret = ivpu_jsm_unregister_db(vdev, cmdq->db_id);
-	if (!ret)
-		ivpu_dbg(vdev, JOB, "DB %d unregistered\n", cmdq->db_id);
+	xa_erase(&file_priv->vdev->db_xa, cmdq->db_id);
+	cmdq->db_id = 0;
 
 	return 0;
 }
 
-static struct ivpu_cmdq *ivpu_cmdq_acquire(struct ivpu_file_priv *file_priv, u8 priority)
+static inline u8 ivpu_job_to_jsm_priority(u8 priority)
+{
+	if (priority == DRM_IVPU_JOB_PRIORITY_DEFAULT)
+		return VPU_JOB_SCHEDULING_PRIORITY_BAND_NORMAL;
+
+	return priority - 1;
+}
+
+static void ivpu_cmdq_destroy(struct ivpu_file_priv *file_priv, struct ivpu_cmdq *cmdq)
+{
+	ivpu_cmdq_unregister(file_priv, cmdq);
+	xa_erase(&file_priv->cmdq_xa, cmdq->id);
+	ivpu_cmdq_free(file_priv, cmdq);
+}
+
+static struct ivpu_cmdq *ivpu_cmdq_acquire_legacy(struct ivpu_file_priv *file_priv, u8 priority)
 {
 	struct ivpu_cmdq *cmdq;
-	unsigned long cmdq_id;
-	int ret;
+	unsigned long id;
 
 	lockdep_assert_held(&file_priv->lock);
 
-	xa_for_each(&file_priv->cmdq_xa, cmdq_id, cmdq)
-		if (cmdq->priority == priority)
+	xa_for_each(&file_priv->cmdq_xa, id, cmdq)
+		if (cmdq->is_legacy && cmdq->priority == priority)
 			break;
 
 	if (!cmdq) {
-		cmdq = ivpu_cmdq_alloc(file_priv);
+		cmdq = ivpu_cmdq_create(file_priv, priority, 0);
 		if (!cmdq)
 			return NULL;
-		cmdq->priority = priority;
+		cmdq->is_legacy = true;
 	}
 
-	ret = ivpu_cmdq_init(file_priv, cmdq, priority);
-	if (ret)
+	return cmdq;
+}
+
+static struct ivpu_cmdq *ivpu_cmdq_acquire(struct ivpu_file_priv *file_priv, u32 cmdq_id)
+{
+	struct ivpu_device *vdev = file_priv->vdev;
+	struct ivpu_cmdq *cmdq;
+
+	lockdep_assert_held(&file_priv->lock);
+
+	cmdq = xa_load(&file_priv->cmdq_xa, cmdq_id);
+	if (!cmdq) {
+		ivpu_dbg(vdev, IOCTL, "Failed to find command queue with ID: %u\n", cmdq_id);
 		return NULL;
+	}
 
 	return cmdq;
 }
@@ -266,11 +362,8 @@ void ivpu_cmdq_release_all_locked(struct ivpu_file_priv *file_priv)
 
 	lockdep_assert_held(&file_priv->lock);
 
-	xa_for_each(&file_priv->cmdq_xa, cmdq_id, cmdq) {
-		xa_erase(&file_priv->cmdq_xa, cmdq_id);
-		ivpu_cmdq_fini(file_priv, cmdq);
-		ivpu_cmdq_free(file_priv, cmdq);
-	}
+	xa_for_each(&file_priv->cmdq_xa, cmdq_id, cmdq)
+		ivpu_cmdq_destroy(file_priv, cmdq);
 }
 
 /*
@@ -286,8 +379,10 @@ static void ivpu_cmdq_reset(struct ivpu_file_priv *file_priv)
 
 	mutex_lock(&file_priv->lock);
 
-	xa_for_each(&file_priv->cmdq_xa, cmdq_id, cmdq)
-		cmdq->db_registered = false;
+	xa_for_each(&file_priv->cmdq_xa, cmdq_id, cmdq) {
+		xa_erase(&file_priv->vdev->db_xa, cmdq->db_id);
+		cmdq->db_id = 0;
+	}
 
 	mutex_unlock(&file_priv->lock);
 }
@@ -305,25 +400,24 @@ void ivpu_cmdq_reset_all_contexts(struct ivpu_device *vdev)
 	mutex_unlock(&vdev->context_list_lock);
 }
 
-static void ivpu_cmdq_fini_all(struct ivpu_file_priv *file_priv)
-{
-	struct ivpu_cmdq *cmdq;
-	unsigned long cmdq_id;
-
-	xa_for_each(&file_priv->cmdq_xa, cmdq_id, cmdq)
-		ivpu_cmdq_fini(file_priv, cmdq);
-}
-
 void ivpu_context_abort_locked(struct ivpu_file_priv *file_priv)
 {
 	struct ivpu_device *vdev = file_priv->vdev;
+	struct ivpu_cmdq *cmdq;
+	unsigned long cmdq_id;
 
 	lockdep_assert_held(&file_priv->lock);
+	ivpu_dbg(vdev, JOB, "Context ID: %u abort\n", file_priv->ctx.id);
 
-	ivpu_cmdq_fini_all(file_priv);
+	xa_for_each(&file_priv->cmdq_xa, cmdq_id, cmdq)
+		ivpu_cmdq_unregister(file_priv, cmdq);
 
 	if (vdev->fw->sched_mode == VPU_SCHEDULING_MODE_OS)
 		ivpu_jsm_context_release(vdev, file_priv->ctx.id);
+
+	ivpu_mmu_disable_ssid_events(vdev, file_priv->ctx.id);
+
+	file_priv->aborted = true;
 }
 
 static int ivpu_cmdq_push_job(struct ivpu_cmdq *cmdq, struct ivpu_job *job)
@@ -348,17 +442,14 @@ static int ivpu_cmdq_push_job(struct ivpu_cmdq *cmdq, struct ivpu_job *job)
 	if (unlikely(ivpu_test_mode & IVPU_TEST_MODE_NULL_SUBMISSION))
 		entry->flags = VPU_JOB_FLAGS_NULL_SUBMISSION_MASK;
 
-	if (vdev->fw->sched_mode == VPU_SCHEDULING_MODE_HW) {
-		if (cmdq->primary_preempt_buf) {
-			entry->primary_preempt_buf_addr = cmdq->primary_preempt_buf->vpu_addr;
-			entry->primary_preempt_buf_size = ivpu_bo_size(cmdq->primary_preempt_buf);
-		}
+	if (job->primary_preempt_buf) {
+		entry->primary_preempt_buf_addr = job->primary_preempt_buf->vpu_addr;
+		entry->primary_preempt_buf_size = ivpu_bo_size(job->primary_preempt_buf);
+	}
 
-		if (cmdq->secondary_preempt_buf) {
-			entry->secondary_preempt_buf_addr = cmdq->secondary_preempt_buf->vpu_addr;
-			entry->secondary_preempt_buf_size =
-				ivpu_bo_size(cmdq->secondary_preempt_buf);
-		}
+	if (job->secondary_preempt_buf) {
+		entry->secondary_preempt_buf_addr = job->secondary_preempt_buf->vpu_addr;
+		entry->secondary_preempt_buf_size = ivpu_bo_size(job->secondary_preempt_buf);
 	}
 
 	wmb(); /* Ensure that tail is updated after filling entry */
@@ -416,8 +507,8 @@ static void ivpu_job_destroy(struct ivpu_job *job)
 	struct ivpu_device *vdev = job->vdev;
 	u32 i;
 
-	ivpu_dbg(vdev, JOB, "Job destroyed: id %3u ctx %2d engine %d",
-		 job->job_id, job->file_priv->ctx.id, job->engine_idx);
+	ivpu_dbg(vdev, JOB, "Job destroyed: id %3u ctx %2d cmdq_id %u engine %d",
+		 job->job_id, job->file_priv->ctx.id, job->cmdq_id, job->engine_idx);
 
 	for (i = 0; i < job->bo_count; i++)
 		if (job->bos[i])
@@ -443,7 +534,7 @@ ivpu_job_create(struct ivpu_file_priv *file_priv, u32 engine_idx, u32 bo_count)
 	job->bo_count = bo_count;
 	job->done_fence = ivpu_fence_create(vdev);
 	if (!job->done_fence) {
-		ivpu_warn_ratelimited(vdev, "Failed to create a fence\n");
+		ivpu_err(vdev, "Failed to create a fence\n");
 		goto err_free_job;
 	}
 
@@ -462,41 +553,90 @@ static struct ivpu_job *ivpu_job_remove_from_submitted_jobs(struct ivpu_device *
 {
 	struct ivpu_job *job;
 
-	xa_lock(&vdev->submitted_jobs_xa);
-	job = __xa_erase(&vdev->submitted_jobs_xa, job_id);
+	lockdep_assert_held(&vdev->submitted_jobs_lock);
 
+	job = xa_erase(&vdev->submitted_jobs_xa, job_id);
 	if (xa_empty(&vdev->submitted_jobs_xa) && job) {
 		vdev->busy_time = ktime_add(ktime_sub(ktime_get(), vdev->busy_start_ts),
 					    vdev->busy_time);
 	}
 
-	xa_unlock(&vdev->submitted_jobs_xa);
-
 	return job;
 }
 
+bool ivpu_job_handle_engine_error(struct ivpu_device *vdev, u32 job_id, u32 job_status)
+{
+	lockdep_assert_held(&vdev->submitted_jobs_lock);
+
+	switch (job_status) {
+	case VPU_JSM_STATUS_PROCESSING_ERR:
+	case VPU_JSM_STATUS_ENGINE_RESET_REQUIRED_MIN ... VPU_JSM_STATUS_ENGINE_RESET_REQUIRED_MAX:
+	{
+		struct ivpu_job *job = xa_load(&vdev->submitted_jobs_xa, job_id);
+
+		if (!job)
+			return false;
+
+		/* Trigger an engine reset */
+		guard(mutex)(&job->file_priv->lock);
+
+		job->job_status = job_status;
+
+		if (job->file_priv->has_mmu_faults)
+			return false;
+
+		/*
+		 * Mark context as faulty and defer destruction of the job to jobs abort thread
+		 * handler to synchronize between both faults and jobs returning context violation
+		 * status and ensure both are handled in the same way
+		 */
+		job->file_priv->has_mmu_faults = true;
+		queue_work(system_percpu_wq, &vdev->context_abort_work);
+		return true;
+	}
+	default:
+		/* Complete job with error status, engine reset not required */
+		break;
+	}
+
+	return false;
+}
+
 static int ivpu_job_signal_and_destroy(struct ivpu_device *vdev, u32 job_id, u32 job_status)
 {
 	struct ivpu_job *job;
 
-	job = ivpu_job_remove_from_submitted_jobs(vdev, job_id);
+	lockdep_assert_held(&vdev->submitted_jobs_lock);
+
+	job = xa_load(&vdev->submitted_jobs_xa, job_id);
 	if (!job)
 		return -ENOENT;
 
-	if (job->file_priv->has_mmu_faults)
-		job_status = DRM_IVPU_JOB_STATUS_ABORTED;
+	ivpu_job_remove_from_submitted_jobs(vdev, job_id);
+
+	if (job->job_status == VPU_JSM_STATUS_SUCCESS) {
+		if (job->file_priv->has_mmu_faults)
+			job->job_status = DRM_IVPU_JOB_STATUS_ABORTED;
+		else
+			job->job_status = job_status;
+	}
 
-	job->bos[CMD_BUF_IDX]->job_status = job_status;
+	job->bos[CMD_BUF_IDX]->job_status = job->job_status;
 	dma_fence_signal(job->done_fence);
 
 	trace_job("done", job);
-	ivpu_dbg(vdev, JOB, "Job complete:  id %3u ctx %2d engine %d status 0x%x\n",
-		 job->job_id, job->file_priv->ctx.id, job->engine_idx, job_status);
+	ivpu_dbg(vdev, JOB, "Job complete:  id %3u ctx %2d cmdq_id %u engine %d status 0x%x\n",
+		 job->job_id, job->file_priv->ctx.id, job->cmdq_id, job->engine_idx,
+		 job->job_status);
 
 	ivpu_job_destroy(job);
 	ivpu_stop_job_timeout_detection(vdev);
 
 	ivpu_rpm_put(vdev);
+
+	if (!xa_empty(&vdev->submitted_jobs_xa))
+		ivpu_start_job_timeout_detection(vdev);
+
 	return 0;
 }
 
@@ -505,11 +645,29 @@ void ivpu_jobs_abort_all(struct ivpu_device *vdev)
 	struct ivpu_job *job;
 	unsigned long id;
 
+	mutex_lock(&vdev->submitted_jobs_lock);
+
 	xa_for_each(&vdev->submitted_jobs_xa, id, job)
 		ivpu_job_signal_and_destroy(vdev, id, DRM_IVPU_JOB_STATUS_ABORTED);
+
+	mutex_unlock(&vdev->submitted_jobs_lock);
+}
+
+void ivpu_cmdq_abort_all_jobs(struct ivpu_device *vdev, u32 ctx_id, u32 cmdq_id)
+{
+	struct ivpu_job *job;
+	unsigned long id;
+
+	mutex_lock(&vdev->submitted_jobs_lock);
+
+	xa_for_each(&vdev->submitted_jobs_xa, id, job)
+		if (job->file_priv->ctx.id == ctx_id && job->cmdq_id == cmdq_id)
+			ivpu_job_signal_and_destroy(vdev, id, DRM_IVPU_JOB_STATUS_ABORTED);
+
+	mutex_unlock(&vdev->submitted_jobs_lock);
 }
 
-static int ivpu_job_submit(struct ivpu_job *job, u8 priority)
+static int ivpu_job_submit(struct ivpu_job *job, u8 priority, u32 cmdq_id)
 {
 	struct ivpu_file_priv *file_priv = job->file_priv;
 	struct ivpu_device *vdev = job->vdev;
@@ -521,25 +679,41 @@ static int ivpu_job_submit(struct ivpu_job *job, u8 priority)
 	if (ret < 0)
 		return ret;
 
+	mutex_lock(&vdev->submitted_jobs_lock);
 	mutex_lock(&file_priv->lock);
 
-	cmdq = ivpu_cmdq_acquire(file_priv, priority);
+	if (cmdq_id == 0)
+		cmdq = ivpu_cmdq_acquire_legacy(file_priv, priority);
+	else
+		cmdq = ivpu_cmdq_acquire(file_priv, cmdq_id);
 	if (!cmdq) {
-		ivpu_warn_ratelimited(vdev, "Failed to get job queue, ctx %d engine %d prio %d\n",
-				      file_priv->ctx.id, job->engine_idx, priority);
 		ret = -EINVAL;
-		goto err_unlock_file_priv;
+		goto err_unlock;
+	}
+
+	ret = ivpu_cmdq_register(file_priv, cmdq);
+	if (ret) {
+		ivpu_err(vdev, "Failed to register command queue: %d\n", ret);
+		goto err_unlock;
+	}
+
+	ret = ivpu_preemption_job_init(vdev, file_priv, cmdq, job);
+	if (ret) {
+		ivpu_err(vdev, "Failed to initialize preemption buffers for job %d: %d\n",
+			 job->job_id, ret);
+		goto err_unlock;
 	}
 
-	xa_lock(&vdev->submitted_jobs_xa);
+	job->cmdq_id = cmdq->id;
+
 	is_first_job = xa_empty(&vdev->submitted_jobs_xa);
-	ret = __xa_alloc_cyclic(&vdev->submitted_jobs_xa, &job->job_id, job, file_priv->job_limit,
-				&file_priv->job_id_next, GFP_KERNEL);
+	ret = xa_alloc_cyclic(&vdev->submitted_jobs_xa, &job->job_id, job, file_priv->job_limit,
+			      &file_priv->job_id_next, GFP_KERNEL);
 	if (ret < 0) {
 		ivpu_dbg(vdev, JOB, "Too many active jobs in ctx %d\n",
 			 file_priv->ctx.id);
 		ret = -EBUSY;
-		goto err_unlock_submitted_jobs_xa;
+		goto err_unlock;
 	}
 
 	ret = ivpu_cmdq_push_job(cmdq, job);
@@ -558,34 +732,34 @@ static int ivpu_job_submit(struct ivpu_job *job, u8 priority)
 	}
 
 	trace_job("submit", job);
-	ivpu_dbg(vdev, JOB, "Job submitted: id %3u ctx %2d engine %d prio %d addr 0x%llx next %d\n",
-		 job->job_id, file_priv->ctx.id, job->engine_idx, priority,
+	ivpu_dbg(vdev, JOB, "Job submitted: id %3u ctx %2d cmdq_id %u engine %d prio %d addr 0x%llx next %d\n",
+		 job->job_id, file_priv->ctx.id, cmdq->id, job->engine_idx, cmdq->priority,
 		 job->cmd_buf_vpu_addr, cmdq->jobq->header.tail);
 
-	xa_unlock(&vdev->submitted_jobs_xa);
-
 	mutex_unlock(&file_priv->lock);
 
-	if (unlikely(ivpu_test_mode & IVPU_TEST_MODE_NULL_HW))
+	if (unlikely(ivpu_test_mode & IVPU_TEST_MODE_NULL_HW)) {
 		ivpu_job_signal_and_destroy(vdev, job->job_id, VPU_JSM_STATUS_SUCCESS);
+	}
+
+	mutex_unlock(&vdev->submitted_jobs_lock);
 
 	return 0;
 
 err_erase_xa:
-	__xa_erase(&vdev->submitted_jobs_xa, job->job_id);
-err_unlock_submitted_jobs_xa:
-	xa_unlock(&vdev->submitted_jobs_xa);
-err_unlock_file_priv:
+	xa_erase(&vdev->submitted_jobs_xa, job->job_id);
+err_unlock:
 	mutex_unlock(&file_priv->lock);
+	mutex_unlock(&vdev->submitted_jobs_lock);
 	ivpu_rpm_put(vdev);
 	return ret;
 }
 
 static int
 ivpu_job_prepare_bos_for_submit(struct drm_file *file, struct ivpu_job *job, u32 *buf_handles,
-				u32 buf_count, u32 commands_offset)
+				u32 buf_count, u32 commands_offset, u32 preempt_buffer_index)
 {
-	struct ivpu_file_priv *file_priv = file->driver_priv;
+	struct ivpu_file_priv *file_priv = job->file_priv;
 	struct ivpu_device *vdev = file_priv->vdev;
 	struct ww_acquire_ctx acquire_ctx;
 	enum dma_resv_usage usage;
@@ -596,40 +770,58 @@ ivpu_job_prepare_bos_for_submit(struct drm_file *file, struct ivpu_job *job, u32
 	for (i = 0; i < buf_count; i++) {
 		struct drm_gem_object *obj = drm_gem_object_lookup(file, buf_handles[i]);
 
-		if (!obj)
+		if (!obj) {
+			ivpu_dbg(vdev, IOCTL, "Failed to lookup GEM object with handle %u\n",
+				 buf_handles[i]);
 			return -ENOENT;
+		}
 
 		job->bos[i] = to_ivpu_bo(obj);
 
-		ret = ivpu_bo_pin(job->bos[i]);
+		ret = ivpu_bo_bind(job->bos[i]);
 		if (ret)
 			return ret;
 	}
 
 	bo = job->bos[CMD_BUF_IDX];
 	if (!dma_resv_test_signaled(bo->base.base.resv, DMA_RESV_USAGE_READ)) {
-		ivpu_warn(vdev, "Buffer is already in use\n");
+		ivpu_dbg(vdev, IOCTL, "Buffer is already in use by another job\n");
 		return -EBUSY;
 	}
 
 	if (commands_offset >= ivpu_bo_size(bo)) {
-		ivpu_warn(vdev, "Invalid command buffer offset %u\n", commands_offset);
+		ivpu_dbg(vdev, IOCTL, "Invalid commands offset %u for buffer size %zu\n",
+			 commands_offset, ivpu_bo_size(bo));
 		return -EINVAL;
 	}
 
 	job->cmd_buf_vpu_addr = bo->vpu_addr + commands_offset;
 
+	if (preempt_buffer_index) {
+		struct ivpu_bo *preempt_bo = job->bos[preempt_buffer_index];
+
+		if (ivpu_bo_size(preempt_bo) < ivpu_fw_preempt_buf_size(vdev)) {
+			ivpu_dbg(vdev, IOCTL, "Preemption buffer is too small\n");
+			return -EINVAL;
+		}
+		if (ivpu_bo_is_mappable(preempt_bo)) {
+			ivpu_dbg(vdev, IOCTL, "Preemption buffer cannot be mappable\n");
+			return -EINVAL;
+		}
+		job->primary_preempt_buf = preempt_bo;
+	}
+
 	ret = drm_gem_lock_reservations((struct drm_gem_object **)job->bos, buf_count,
 					&acquire_ctx);
 	if (ret) {
-		ivpu_warn(vdev, "Failed to lock reservations: %d\n", ret);
+		ivpu_warn_ratelimited(vdev, "Failed to lock reservations: %d\n", ret);
 		return ret;
 	}
 
 	for (i = 0; i < buf_count; i++) {
 		ret = dma_resv_reserve_fences(job->bos[i]->base.base.resv, 1);
 		if (ret) {
-			ivpu_warn(vdev, "Failed to reserve fences: %d\n", ret);
+			ivpu_warn_ratelimited(vdev, "Failed to reserve fences: %d\n", ret);
 			goto unlock_reservations;
 		}
 	}
@@ -647,49 +839,20 @@ unlock_reservations:
 	return ret;
 }
 
-static inline u8 ivpu_job_to_hws_priority(struct ivpu_file_priv *file_priv, u8 priority)
+static int ivpu_submit(struct drm_file *file, struct ivpu_file_priv *file_priv, u32 cmdq_id,
+		       u32 buffer_count, u32 engine, void __user *buffers_ptr, u32 cmds_offset,
+		       u32 preempt_buffer_index, u8 priority)
 {
-	if (priority == DRM_IVPU_JOB_PRIORITY_DEFAULT)
-		return DRM_IVPU_JOB_PRIORITY_NORMAL;
-
-	return priority - 1;
-}
-
-int ivpu_submit_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
-{
-	struct ivpu_file_priv *file_priv = file->driver_priv;
 	struct ivpu_device *vdev = file_priv->vdev;
-	struct drm_ivpu_submit *params = data;
 	struct ivpu_job *job;
 	u32 *buf_handles;
 	int idx, ret;
-	u8 priority;
-
-	if (params->engine != DRM_IVPU_ENGINE_COMPUTE)
-		return -EINVAL;
-
-	if (params->priority > DRM_IVPU_JOB_PRIORITY_REALTIME)
-		return -EINVAL;
-
-	if (params->buffer_count == 0 || params->buffer_count > JOB_MAX_BUFFER_COUNT)
-		return -EINVAL;
-
-	if (!IS_ALIGNED(params->commands_offset, 8))
-		return -EINVAL;
-
-	if (!file_priv->ctx.id)
-		return -EINVAL;
-
-	if (file_priv->has_mmu_faults)
-		return -EBADFD;
 
-	buf_handles = kcalloc(params->buffer_count, sizeof(u32), GFP_KERNEL);
+	buf_handles = kcalloc(buffer_count, sizeof(u32), GFP_KERNEL);
 	if (!buf_handles)
 		return -ENOMEM;
 
-	ret = copy_from_user(buf_handles,
-			     (void __user *)params->buffers_ptr,
-			     params->buffer_count * sizeof(u32));
+	ret = copy_from_user(buf_handles, buffers_ptr, buffer_count * sizeof(u32));
 	if (ret) {
 		ret = -EFAULT;
 		goto err_free_handles;
@@ -700,27 +863,22 @@ int ivpu_submit_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 		goto err_free_handles;
 	}
 
-	ivpu_dbg(vdev, JOB, "Submit ioctl: ctx %u buf_count %u\n",
-		 file_priv->ctx.id, params->buffer_count);
+	ivpu_dbg(vdev, JOB, "Submit ioctl: ctx %u cmdq_id %u buf_count %u\n",
+		 file_priv->ctx.id, cmdq_id, buffer_count);
 
-	job = ivpu_job_create(file_priv, params->engine, params->buffer_count);
+	job = ivpu_job_create(file_priv, engine, buffer_count);
 	if (!job) {
-		ivpu_err(vdev, "Failed to create job\n");
 		ret = -ENOMEM;
 		goto err_exit_dev;
 	}
 
-	ret = ivpu_job_prepare_bos_for_submit(file, job, buf_handles, params->buffer_count,
-					      params->commands_offset);
-	if (ret) {
-		ivpu_err(vdev, "Failed to prepare job: %d\n", ret);
+	ret = ivpu_job_prepare_bos_for_submit(file, job, buf_handles, buffer_count, cmds_offset,
+					      preempt_buffer_index);
+	if (ret)
 		goto err_destroy_job;
-	}
-
-	priority = ivpu_job_to_hws_priority(file_priv, params->priority);
 
 	down_read(&vdev->pm->reset_lock);
-	ret = ivpu_job_submit(job, priority);
+	ret = ivpu_job_submit(job, priority, cmdq_id);
 	up_read(&vdev->pm->reset_lock);
 	if (ret)
 		goto err_signal_fence;
@@ -740,12 +898,176 @@ err_free_handles:
 	return ret;
 }
 
+int ivpu_submit_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
+{
+	struct ivpu_file_priv *file_priv = file->driver_priv;
+	struct ivpu_device *vdev = file_priv->vdev;
+	struct drm_ivpu_submit *args = data;
+	u8 priority;
+
+	if (args->engine != DRM_IVPU_ENGINE_COMPUTE) {
+		ivpu_dbg(vdev, IOCTL, "Invalid engine %d\n", args->engine);
+		return -EINVAL;
+	}
+
+	if (args->priority > DRM_IVPU_JOB_PRIORITY_REALTIME) {
+		ivpu_dbg(vdev, IOCTL, "Invalid priority %d\n", args->priority);
+		return -EINVAL;
+	}
+
+	if (args->buffer_count == 0 || args->buffer_count > JOB_MAX_BUFFER_COUNT) {
+		ivpu_dbg(vdev, IOCTL, "Invalid buffer count %u\n", args->buffer_count);
+		return -EINVAL;
+	}
+
+	if (!IS_ALIGNED(args->commands_offset, 8)) {
+		ivpu_dbg(vdev, IOCTL, "Invalid commands offset %u\n", args->commands_offset);
+		return -EINVAL;
+	}
+
+	if (!file_priv->ctx.id) {
+		ivpu_dbg(vdev, IOCTL, "Context not initialized\n");
+		return -EINVAL;
+	}
+
+	if (file_priv->has_mmu_faults) {
+		ivpu_dbg(vdev, IOCTL, "Context %u has MMU faults\n", file_priv->ctx.id);
+		return -EBADFD;
+	}
+
+	priority = ivpu_job_to_jsm_priority(args->priority);
+
+	return ivpu_submit(file, file_priv, 0, args->buffer_count, args->engine,
+			   (void __user *)args->buffers_ptr, args->commands_offset, 0, priority);
+}
+
+int ivpu_cmdq_submit_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
+{
+	struct ivpu_file_priv *file_priv = file->driver_priv;
+	struct ivpu_device *vdev = file_priv->vdev;
+	struct drm_ivpu_cmdq_submit *args = data;
+
+	if (!ivpu_is_capable(file_priv->vdev, DRM_IVPU_CAP_MANAGE_CMDQ)) {
+		ivpu_dbg(vdev, IOCTL, "Command queue management not supported\n");
+		return -ENODEV;
+	}
+
+	if (args->cmdq_id < IVPU_CMDQ_MIN_ID || args->cmdq_id > IVPU_CMDQ_MAX_ID) {
+		ivpu_dbg(vdev, IOCTL, "Invalid command queue ID %u\n", args->cmdq_id);
+		return -EINVAL;
+	}
+
+	if (args->buffer_count == 0 || args->buffer_count > JOB_MAX_BUFFER_COUNT) {
+		ivpu_dbg(vdev, IOCTL, "Invalid buffer count %u\n", args->buffer_count);
+		return -EINVAL;
+	}
+
+	if (args->preempt_buffer_index >= args->buffer_count) {
+		ivpu_dbg(vdev, IOCTL, "Invalid preemption buffer index %u\n",
+			 args->preempt_buffer_index);
+		return -EINVAL;
+	}
+
+	if (!IS_ALIGNED(args->commands_offset, 8)) {
+		ivpu_dbg(vdev, IOCTL, "Invalid commands offset %u\n", args->commands_offset);
+		return -EINVAL;
+	}
+
+	if (!file_priv->ctx.id) {
+		ivpu_dbg(vdev, IOCTL, "Context not initialized\n");
+		return -EINVAL;
+	}
+
+	if (file_priv->has_mmu_faults) {
+		ivpu_dbg(vdev, IOCTL, "Context %u has MMU faults\n", file_priv->ctx.id);
+		return -EBADFD;
+	}
+
+	return ivpu_submit(file, file_priv, args->cmdq_id, args->buffer_count, VPU_ENGINE_COMPUTE,
+			   (void __user *)args->buffers_ptr, args->commands_offset,
+			   args->preempt_buffer_index, 0);
+}
+
+int ivpu_cmdq_create_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
+{
+	struct ivpu_file_priv *file_priv = file->driver_priv;
+	struct ivpu_device *vdev = file_priv->vdev;
+	struct drm_ivpu_cmdq_create *args = data;
+	struct ivpu_cmdq *cmdq;
+	int ret;
+
+	if (!ivpu_is_capable(vdev, DRM_IVPU_CAP_MANAGE_CMDQ)) {
+		ivpu_dbg(vdev, IOCTL, "Command queue management not supported\n");
+		return -ENODEV;
+	}
+
+	if (args->priority > DRM_IVPU_JOB_PRIORITY_REALTIME) {
+		ivpu_dbg(vdev, IOCTL, "Invalid priority %d\n", args->priority);
+		return -EINVAL;
+	}
+
+	ret = ivpu_rpm_get(vdev);
+	if (ret < 0)
+		return ret;
+
+	mutex_lock(&file_priv->lock);
+
+	cmdq = ivpu_cmdq_create(file_priv, ivpu_job_to_jsm_priority(args->priority), args->flags);
+	if (cmdq)
+		args->cmdq_id = cmdq->id;
+
+	mutex_unlock(&file_priv->lock);
+
+	ivpu_rpm_put(vdev);
+
+	return cmdq ? 0 : -ENOMEM;
+}
+
+int ivpu_cmdq_destroy_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
+{
+	struct ivpu_file_priv *file_priv = file->driver_priv;
+	struct ivpu_device *vdev = file_priv->vdev;
+	struct drm_ivpu_cmdq_destroy *args = data;
+	struct ivpu_cmdq *cmdq;
+	u32 cmdq_id = 0;
+	int ret;
+
+	if (!ivpu_is_capable(vdev, DRM_IVPU_CAP_MANAGE_CMDQ)) {
+		ivpu_dbg(vdev, IOCTL, "Command queue management not supported\n");
+		return -ENODEV;
+	}
+
+	ret = ivpu_rpm_get(vdev);
+	if (ret < 0)
+		return ret;
+
+	mutex_lock(&file_priv->lock);
+
+	cmdq = xa_load(&file_priv->cmdq_xa, args->cmdq_id);
+	if (!cmdq || cmdq->is_legacy) {
+		ret = -ENOENT;
+	} else {
+		cmdq_id = cmdq->id;
+		ivpu_cmdq_destroy(file_priv, cmdq);
+		ret = 0;
+	}
+
+	mutex_unlock(&file_priv->lock);
+
+	/* Abort any pending jobs only if cmdq was destroyed */
+	if (!ret)
+		ivpu_cmdq_abort_all_jobs(vdev, file_priv->ctx.id, cmdq_id);
+
+	ivpu_rpm_put(vdev);
+
+	return ret;
+}
+
 static void
 ivpu_job_done_callback(struct ivpu_device *vdev, struct ivpu_ipc_hdr *ipc_hdr,
 		       struct vpu_jsm_msg *jsm_msg)
 {
 	struct vpu_ipc_msg_payload_job_done *payload;
-	int ret;
 
 	if (!jsm_msg) {
 		ivpu_err(vdev, "IPC message has no JSM payload\n");
@@ -758,9 +1080,12 @@ ivpu_job_done_callback(struct ivpu_device *vdev, struct ivpu_ipc_hdr *ipc_hdr,
 	}
 
 	payload = (struct vpu_ipc_msg_payload_job_done *)&jsm_msg->payload;
-	ret = ivpu_job_signal_and_destroy(vdev, payload->job_id, payload->job_status);
-	if (!ret && !xa_empty(&vdev->submitted_jobs_xa))
-		ivpu_start_job_timeout_detection(vdev);
+
+	mutex_lock(&vdev->submitted_jobs_lock);
+	if (!ivpu_job_handle_engine_error(vdev, payload->job_id, payload->job_status))
+		/* No engine error, complete the job normally */
+		ivpu_job_signal_and_destroy(vdev, payload->job_id, payload->job_status);
+	mutex_unlock(&vdev->submitted_jobs_lock);
 }
 
 void ivpu_job_done_consumer_init(struct ivpu_device *vdev)
@@ -773,3 +1098,56 @@ void ivpu_job_done_consumer_fini(struct ivpu_device *vdev)
 {
 	ivpu_ipc_consumer_del(vdev, &vdev->job_done_consumer);
 }
+
+void ivpu_context_abort_work_fn(struct work_struct *work)
+{
+	struct ivpu_device *vdev = container_of(work, struct ivpu_device, context_abort_work);
+	struct ivpu_file_priv *file_priv;
+	struct ivpu_job *job;
+	unsigned long ctx_id;
+	unsigned long id;
+
+	if (drm_WARN_ON(&vdev->drm, pm_runtime_get_if_active(vdev->drm.dev) <= 0))
+		return;
+
+	if (vdev->fw->sched_mode == VPU_SCHEDULING_MODE_HW)
+		if (ivpu_jsm_reset_engine(vdev, 0))
+			goto runtime_put;
+
+	mutex_lock(&vdev->context_list_lock);
+	xa_for_each(&vdev->context_xa, ctx_id, file_priv) {
+		if (!file_priv->has_mmu_faults || file_priv->aborted)
+			continue;
+
+		mutex_lock(&file_priv->lock);
+		ivpu_context_abort_locked(file_priv);
+		mutex_unlock(&file_priv->lock);
+	}
+	mutex_unlock(&vdev->context_list_lock);
+
+	/*
+	 * We will not receive new MMU event interrupts until existing events are discarded
+	 * however, we want to discard these events only after aborting the faulty context
+	 * to avoid generating new faults from that context
+	 */
+	ivpu_mmu_discard_events(vdev);
+
+	if (vdev->fw->sched_mode != VPU_SCHEDULING_MODE_HW)
+		goto runtime_put;
+
+	if (ivpu_jsm_hws_resume_engine(vdev, 0))
+		goto runtime_put;
+	/*
+	 * In hardware scheduling mode NPU already has stopped processing jobs
+	 * and won't send us any further notifications, thus we have to free job related resources
+	 * and notify userspace
+	 */
+	mutex_lock(&vdev->submitted_jobs_lock);
+	xa_for_each(&vdev->submitted_jobs_xa, id, job)
+		if (job->file_priv->aborted)
+			ivpu_job_signal_and_destroy(vdev, job->job_id, DRM_IVPU_JOB_STATUS_ABORTED);
+	mutex_unlock(&vdev->submitted_jobs_lock);
+
+runtime_put:
+	pm_runtime_put_autosuspend(vdev->drm.dev);
+}