Skip to contentSkip to conversation

drm/amdgpu: Fetch MES scheduler/KIQ versions

PATCH Superseded Graham Sider 3 participants 9 messages Jun 10, 2022 – Jun 13, 2022

Subsystems (3)

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/3 drm/amdgpu: Fetch MES scheduler/KIQ versions

Store MES scheduler and MES KIQ version numbers in amdgpu_mes for GFX11.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h |  3 +++
 drivers/gpu/drm/amd/amdgpu/mes_v11_0.c  | 12 ++++++++++++
 2 files changed, 15 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +3 -0
index 92ddee5e33db..aa06c8396ee0 100644
@@ -64,6 +64,9 @@ struct amdgpu_mes {
 
 	spinlock_t                      queue_id_lock;
 
+	uint32_t			sched_version;
+	uint32_t			kiq_version;
+
 	uint32_t                        total_max_queue;
 	uint32_t                        doorbell_id_offset;
 	uint32_t                        max_doorbell_slices;
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +12 -0
index e4eb87689f7f..2a9ef308e71c 100644
@@ -890,6 +890,18 @@ static int mes_v11_0_queue_init(struct amdgpu_device *adev,
 		mes_v11_0_queue_init_register(ring);
 	}
 
+	/* get MES scheduler/KIQ versions */
+	mutex_lock(&adev->srbm_mutex);
+	soc21_grbm_select(adev, 3, pipe, 0, 0);
+
+	if (pipe == AMDGPU_MES_SCHED_PIPE)
+		adev->mes.sched_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
+	else if (pipe == AMDGPU_MES_KIQ_PIPE && adev->enable_mes_kiq)
+		adev->mes.kiq_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
+
+	soc21_grbm_select(adev, 0, 0, 0, 0);
+	mutex_unlock(&adev->srbm_mutex);
+
 	return 0;
 }
 
Signature · 1 line
2.25.1
#1 lore raw

PATCH v2 2/3 drm/amdkfd: Enable GFX11 usermode queue oversubscription

Starting with GFX11, MES requires wptr BOs to be GTT allocated/mapped to GART for usermode queues in order to support oversubscription. In the case that work is submitted to an unmapped queue, MES must have a GART wptr address to determine whether the queue should be mapped.

This change is accompanied with changes in MES and is applicable for MES_VERSION >= 3.

v2: Update MES_VERSION check from 2 to 3.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 39 ++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 45 ++++++++++++++++++-
 .../drm/amd/amdkfd/kfd_device_queue_manager.c |  9 +++-
 .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c  |  2 +
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |  3 ++
 .../amd/amdkfd/kfd_process_queue_manager.c    | 19 +++++---
 7 files changed, 110 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -0
index 429b16ba10bf..dba26d1e3be9 100644
@@ -301,6 +301,7 @@ int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct amdgpu_device *adev,
 		struct kgd_mem *mem, void **kptr, uint64_t *size);
 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct amdgpu_device *adev,
 		struct kgd_mem *mem);
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo);
 
 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info,
 					    struct dma_fence **ef);
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +39 -0
index efab923056f4..2d452655eb04 100644
@@ -2030,6 +2030,45 @@ int amdgpu_amdkfd_gpuvm_sync_memory(
 	return ret;
 }
 
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo)
+{
+	int ret;
+
+	ret = amdgpu_bo_reserve(bo, true);
+	if (ret) {
+		pr_err("Failed to reserve bo. ret %d\n", ret);
+		goto err_reserve_bo_failed;
+	}
+
+	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
+	if (ret) {
+		pr_err("Failed to pin bo. ret %d\n", ret);
+		goto err_pin_bo_failed;
+	}
+
+	ret = amdgpu_ttm_alloc_gart(&bo->tbo);
+	if (ret) {
+		pr_err("Failed to bind bo to GART. ret %d\n", ret);
+		goto err_map_bo_gart_failed;
+	}
+
+	amdgpu_amdkfd_remove_eviction_fence(
+		bo, bo->kfd_bo->process_info->eviction_fence);
+	list_del_init(&bo->kfd_bo->validate_list.head);
+
+	amdgpu_bo_unreserve(bo);
+
+	return 0;
+
+err_map_bo_gart_failed:
+	amdgpu_bo_unpin(bo);
+err_pin_bo_failed:
+	amdgpu_bo_unreserve(bo);
+err_reserve_bo_failed:
+
+	return ret;
+}
+
 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct amdgpu_device *adev,
 		struct kgd_mem *mem, void **kptr, uint64_t *size)
 {
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +43 -2

68 lines not on this page · raw message

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +8 -1
index b39d89c52887..d8de2fbdfc7d 100644
@@ -208,6 +208,7 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
 	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
 	struct mes_add_queue_input queue_input;
 	int r, queue_type;
+	uint64_t wptr_addr_off;
 
 	if (dqm->is_hws_hang)
 		return -EIO;
@@ -227,7 +228,13 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
 					AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
 	queue_input.doorbell_offset = q->properties.doorbell_off;
 	queue_input.mqd_addr = q->gart_mqd_addr;
-	queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
+
+	if (q->wptr_bo) {
+		wptr_addr_off = (uint64_t)q->properties.write_ptr - (uint64_t)q->wptr_bo->kfd_bo->va;
+		queue_input.wptr_addr = ((uint64_t)q->wptr_bo->tbo.resource->start << PAGE_SHIFT) + wptr_addr_off;
+	} else
+		queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
+
 	queue_input.paging = false;
 	queue_input.tba_addr = qpd->tba_addr;
 	queue_input.tma_addr = qpd->tma_addr;
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +2 -0
index f1654b4da856..35e74bdd81da 100644
@@ -377,6 +377,8 @@ static void update_mqd_sdma(struct mqd_manager *mm, void *mqd,
 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
+	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
+	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
 	m->sdmax_rlcx_doorbell_offset =
 		q->doorbell_off << SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;
 
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +3 -0
index a5d3963537d7..dcddee0d6f06 100644
@@ -639,6 +639,8 @@ struct queue {
 	void *gang_ctx_bo;
 	uint64_t gang_ctx_gpu_addr;
 	void *gang_ctx_cpu_ptr;
+
+	struct amdgpu_bo *wptr_bo;
 };
 
 enum KFD_MQD_TYPE {
@@ -1404,6 +1406,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
 			    struct file *f,
 			    struct queue_properties *properties,
 			    unsigned int *qid,
+			    struct amdgpu_bo *wptr_bo,
 			    const struct kfd_criu_queue_priv_data *q_data,
 			    const void *restore_mqd,
 			    const void *restore_ctl_stack,
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +14 -5
index f99e09dc43ea..ede1462a55c8 100644
@@ -190,7 +190,8 @@ void pqm_uninit(struct process_queue_manager *pqm)
 static int init_user_queue(struct process_queue_manager *pqm,
 				struct kfd_dev *dev, struct queue **q,
 				struct queue_properties *q_properties,
-				struct file *f, unsigned int qid)
+				struct file *f, struct amdgpu_bo *wptr_bo,
+				unsigned int qid)
 {
 	int retval;
 
@@ -221,6 +222,9 @@ static int init_user_queue(struct process_queue_manager *pqm,
 			goto cleanup;
 		}
 		memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
+
+		if (wptr_bo)
+			(*q)->wptr_bo = amdgpu_bo_ref(wptr_bo);
 	}
 
 	pr_debug("PQM After init queue");
@@ -237,6 +241,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
 			    struct file *f,
 			    struct queue_properties *properties,
 			    unsigned int *qid,
+			    struct amdgpu_bo *wptr_bo,
 			    const struct kfd_criu_queue_priv_data *q_data,
 			    const void *restore_mqd,
 			    const void *restore_ctl_stack,
@@ -299,7 +304,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
 		 * allocate_sdma_queue() in create_queue() has the
 		 * corresponding check logic.
 		 */
-		retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
+		retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
 		if (retval != 0)
 			goto err_create_queue;
 		pqn->q = q;
@@ -320,7 +325,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
 			goto err_create_queue;
 		}
 
-		retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
+		retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
 		if (retval != 0)
 			goto err_create_queue;
 		pqn->q = q;
@@ -457,9 +462,13 @@ int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
 			pdd->qpd.num_gws = 0;
 		}
 
-		if (dev->shared_resources.enable_mes)
+		if (dev->shared_resources.enable_mes) {
 			amdgpu_amdkfd_free_gtt_mem(dev->adev,
 						   pqn->q->gang_ctx_bo);
+			if (pqn->q->wptr_bo)
+				amdgpu_amdkfd_free_gtt_mem(dev->adev, pqn->q->wptr_bo);
+
+		}
 		uninit_queue(pqn->q);
 	}
 
@@ -900,7 +909,7 @@ int kfd_criu_restore_queue(struct kfd_process *p,
 
 	print_queue_properties(&qp);
 
-	ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, q_data, mqd, ctl_stack,
+	ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, NULL, q_data, mqd, ctl_stack,
 				NULL);
 	if (ret) {
 		pr_err("Failed to create new queue err:%d\n", ret);
Signature · 1 line
2.25.1
maintainer
#2 lore raw

Re: [PATCH v2 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

Am 2022-06-10 um 13:13 schrieb Graham Sider:

Quoted 38 lines

Starting with GFX11, MES requires wptr BOs to be GTT allocated/mapped to GART for usermode queues in order to support oversubscription. In the case that work is submitted to an unmapped queue, MES must have a GART wptr address to determine whether the queue should be mapped.

This change is accompanied with changes in MES and is applicable for MES_VERSION >= 3.

v2: Update MES_VERSION check from 2 to 3.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
  .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 39 ++++++++++++++++
  drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 45 ++++++++++++++++++-
  .../drm/amd/amdkfd/kfd_device_queue_manager.c |  9 +++-
  .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c  |  2 +
  drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |  3 ++
  .../amd/amdkfd/kfd_process_queue_manager.c    | 19 +++++---
  7 files changed, 110 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -0
index 429b16ba10bf..dba26d1e3be9 100644
@@ -301,6 +301,7 @@ int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct amdgpu_device *adev,
  		struct kgd_mem *mem, void **kptr, uint64_t *size);
  void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct amdgpu_device *adev,
  		struct kgd_mem *mem);
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo);
 
  int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info,
  					    struct dma_fence **ef);
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
index efab923056f4..2d452655eb04 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_sync_memory( · line 2033

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -0
@@ -2030,6 +2030,45 @@ int amdgpu_amdkfd_gpuvm_sync_memory(
  	return ret;
  }
 
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo)

I think this function should take a reference count on the bo to ensure the pinned BO is not freed prematurely (even if broken user mode frees the BO before destroying the queue). Add a comment that the correct way to release the reference and unpin/unmap the BO is to call amdgpu_amdkfd_free_gtt_mem. See another comment below about moving the amdgpu_bo_ref here.

Quoted 14 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +14 -0
+{
+	int ret;
+
+	ret = amdgpu_bo_reserve(bo, true);
+	if (ret) {
+		pr_err("Failed to reserve bo. ret %d\n", ret);
+		goto err_reserve_bo_failed;
+	}
+
+	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
+	if (ret) {
+		pr_err("Failed to pin bo. ret %d\n", ret);
+		goto err_pin_bo_failed;
+	}

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_sync_memory( · line 2057

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +10 -0
+
+	ret = amdgpu_ttm_alloc_gart(&bo->tbo);
+	if (ret) {
+		pr_err("Failed to bind bo to GART. ret %d\n", ret);
+		goto err_map_bo_gart_failed;
+	}
+
+	amdgpu_amdkfd_remove_eviction_fence(
+		bo, bo->kfd_bo->process_info->eviction_fence);
+	list_del_init(&bo->kfd_bo->validate_list.head);

Please see Lang, Yu's patch "drm/amdkfd: add pinned BOs to kfd_bo_list". We realized that pinned BOs still need to be on the validate list. So please remove the list_del_init here.

Quoted 22 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +14 -0
+
+	amdgpu_bo_unreserve(bo);
+
+	return 0;
+
+err_map_bo_gart_failed:
+	amdgpu_bo_unpin(bo);
+err_pin_bo_failed:
+	amdgpu_bo_unreserve(bo);
+err_reserve_bo_failed:
+
+	return ret;
+}
+
  int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct amdgpu_device *adev,
  		struct kgd_mem *mem, void **kptr, uint64_t *size)
  {
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
index e9766e165c38..58d5ebed1b32 100644

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, · line 295

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +3 -0
@@ -290,6 +290,11 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
  	struct queue_properties q_properties;
  	uint32_t doorbell_offset_in_process = 0;
 
+	struct amdgpu_bo_va_mapping *wptr_mapping;
+	struct interval_tree_node *wptr_node;
+	struct amdgpu_vm *wptr_vm;

It's good practice to minimize the scope of local variables. The above three variables are only used inside the "if" below. So they should be declared there.

Quoted 23 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +16 -0
+	struct amdgpu_bo *wptr_bo = NULL;
+
  	memset(&q_properties, 0, sizeof(struct queue_properties));
 
  	pr_debug("Creating queue ioctl\n");
@@ -316,12 +321,44 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
  		goto err_bind_process;
  	}
 
+	/* Starting with GFX11, MES requires wptr BOs to be GTT allocated/mapped to
+	 * GART for usermode queues in order to support oversubscription. In the
+	 * case that work is submitted to an unmapped queue, MES must have a GART
+	 * wptr address to determine whether the queue should be mapped.
+	 */
+	if (dev->shared_resources.enable_mes && (dev->adev->mes.sched_version & 0xff) >= 3) {
+		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
+		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
+		if (err)
+			goto err_wptr_bo_reserve;
+
+		wptr_node = interval_tree_iter_first(&wptr_vm->va,
+				args->write_pointer_address >> PAGE_SHIFT,
+				args->write_pointer_address >> PAGE_SHIFT);

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, · line 347

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +10 -0
+		if (!wptr_node) {
+			pr_err("Failed to lookup wptr bo\n");
+			err = -EINVAL;
+			goto err_wptr_bo_lookup;
+		}
+
+		wptr_mapping = container_of((struct rb_node *)wptr_node, struct amdgpu_bo_va_mapping, rb);
+		wptr_bo = wptr_mapping->bo_va->base.bo;
+
+		amdgpu_bo_unreserve(wptr_vm->root.bo);

You could move this to just after the interval_tree_iter_first call, even before it checks for errors. That way you don't need to deal with the unreserve conditionally on the error handling code path. And you can probably just use one error label err_wptr_map_gart that handles all 3 failure cases.

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +8 -0
+
+		err = amdgpu_amdkfd_map_gtt_bo_to_gart(dev->adev, wptr_bo);
+		if (err) {
+			pr_err("Failed to map wptr bo to GART\n");
+			goto err_wptr_bo_gart_map;
+		}
+	}
+

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, · line 363

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2 -2
  	pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
  			p->pasid,
  			dev->id);
 
-	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, NULL, NULL, NULL,
-			&doorbell_offset_in_process);
+	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, wptr_bo,
+			NULL, NULL, NULL, &doorbell_offset_in_process);
  	if (err != 0)
  		goto err_create_queue;

The error handling for this and all later failures should gart-unmap the wptr_bo again.

Quoted 89 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +4 -0
@@ -353,7 +390,11 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
  	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0);
  	return 0;
 
+err_wptr_bo_lookup:
+	amdgpu_bo_unreserve(wptr_vm->root.bo);
  err_create_queue:
+err_wptr_bo_gart_map:
+err_wptr_bo_reserve:
  err_bind_process:
  err_pdd:
  	mutex_unlock(&p->mutex);
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +8 -1
index b39d89c52887..d8de2fbdfc7d 100644
@@ -208,6 +208,7 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
  	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
  	struct mes_add_queue_input queue_input;
  	int r, queue_type;
+	uint64_t wptr_addr_off;
 
  	if (dqm->is_hws_hang)
  		return -EIO;
@@ -227,7 +228,13 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
  					AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
  	queue_input.doorbell_offset = q->properties.doorbell_off;
  	queue_input.mqd_addr = q->gart_mqd_addr;
-	queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
+
+	if (q->wptr_bo) {
+		wptr_addr_off = (uint64_t)q->properties.write_ptr - (uint64_t)q->wptr_bo->kfd_bo->va;
+		queue_input.wptr_addr = ((uint64_t)q->wptr_bo->tbo.resource->start << PAGE_SHIFT) + wptr_addr_off;
+	} else
+		queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
+
  	queue_input.paging = false;
  	queue_input.tba_addr = qpd->tba_addr;
  	queue_input.tma_addr = qpd->tma_addr;
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +2 -0
index f1654b4da856..35e74bdd81da 100644
@@ -377,6 +377,8 @@ static void update_mqd_sdma(struct mqd_manager *mm, void *mqd,
  	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
  	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
  	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
+	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
+	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
  	m->sdmax_rlcx_doorbell_offset =
  		q->doorbell_off << SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;
 
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +3 -0
index a5d3963537d7..dcddee0d6f06 100644
@@ -639,6 +639,8 @@ struct queue {
  	void *gang_ctx_bo;
  	uint64_t gang_ctx_gpu_addr;
  	void *gang_ctx_cpu_ptr;
+
+	struct amdgpu_bo *wptr_bo;
  };
 
  enum KFD_MQD_TYPE {
@@ -1404,6 +1406,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
  			    struct file *f,
  			    struct queue_properties *properties,
  			    unsigned int *qid,
+			    struct amdgpu_bo *wptr_bo,
  			    const struct kfd_criu_queue_priv_data *q_data,
  			    const void *restore_mqd,
  			    const void *restore_ctl_stack,
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +2 -1
index f99e09dc43ea..ede1462a55c8 100644
@@ -190,7 +190,8 @@ void pqm_uninit(struct process_queue_manager *pqm)
  static int init_user_queue(struct process_queue_manager *pqm,
  				struct kfd_dev *dev, struct queue **q,
  				struct queue_properties *q_properties,
-				struct file *f, unsigned int qid)
+				struct file *f, struct amdgpu_bo *wptr_bo,
+				unsigned int qid)
  {
  	int retval;
 

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c static int init_user_queue(struct process_queue_manager *pqm, · line 227

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +3 -0
@@ -221,6 +222,9 @@ static int init_user_queue(struct process_queue_manager *pqm,
  			goto cleanup;
  		}
  		memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
+
+		if (wptr_bo)
+			(*q)->wptr_bo = amdgpu_bo_ref(wptr_bo);

I think it would be clearer if you took that reference in amdgpu_amdkfd_map_gtt_bo_to_gart. That way it would be a simple assignment here with no "if". And all the BO reference handling would be more neatly contained in amdgpu_amdkfd code.

Regards,
  Felix

Quoted 53 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +9 -4
  	}
 
  	pr_debug("PQM After init queue");
@@ -237,6 +241,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
  			    struct file *f,
  			    struct queue_properties *properties,
  			    unsigned int *qid,
+			    struct amdgpu_bo *wptr_bo,
  			    const struct kfd_criu_queue_priv_data *q_data,
  			    const void *restore_mqd,
  			    const void *restore_ctl_stack,
@@ -299,7 +304,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
  		 * allocate_sdma_queue() in create_queue() has the
  		 * corresponding check logic.
  		 */
-		retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
+		retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
  		if (retval != 0)
  			goto err_create_queue;
  		pqn->q = q;
@@ -320,7 +325,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
  			goto err_create_queue;
  		}
 
-		retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
+		retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
  		if (retval != 0)
  			goto err_create_queue;
  		pqn->q = q;
@@ -457,9 +462,13 @@ int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
  			pdd->qpd.num_gws = 0;
  		}
 
-		if (dev->shared_resources.enable_mes)
+		if (dev->shared_resources.enable_mes) {
  			amdgpu_amdkfd_free_gtt_mem(dev->adev,
  						   pqn->q->gang_ctx_bo);
+			if (pqn->q->wptr_bo)
+				amdgpu_amdkfd_free_gtt_mem(dev->adev, pqn->q->wptr_bo);
+
+		}
  		uninit_queue(pqn->q);
  	}
 
@@ -900,7 +909,7 @@ int kfd_criu_restore_queue(struct kfd_process *p,
 
  	print_queue_properties(&qp);
 
-	ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, q_data, mqd, ctl_stack,
+	ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, NULL, q_data, mqd, ctl_stack,
  				NULL);
  	if (ret) {
  		pr_err("Failed to create new queue err:%d\n", ret);
#3 lore raw

RE: [PATCH v2 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

[AMD Official Use Only - General]

Thanks for the great comments Felix--will apply these.

Best,
Graham

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Friday, June 10, 2022 7:06 PM
To: Sider, Graham <[email protected]>; [email protected]
Cc: Joshi, Mukul <[email protected]>; Yang, Philip <[email protected]>
Subject: Re: [PATCH v2 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

Am 2022-06-10 um 13:13 schrieb Graham Sider:

Quoted 41 lines

Starting with GFX11, MES requires wptr BOs to be GTT allocated/mapped to GART for usermode queues in order to support oversubscription. In the case that work is submitted to an unmapped queue, MES must have a GART wptr address to determine whether the queue should be mapped.

This change is accompanied with changes in MES and is applicable for MES_VERSION >= 3.

v2: Update MES_VERSION check from 2 to 3.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
  .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 39 ++++++++++++++++
  drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 45 ++++++++++++++++++-
  .../drm/amd/amdkfd/kfd_device_queue_manager.c |  9 +++-
  .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c  |  2 +
  drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |  3 ++
  .../amd/amdkfd/kfd_process_queue_manager.c    | 19 +++++---
  7 files changed, 110 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +2 -0
index 429b16ba10bf..dba26d1e3be9 100644
@@ -301,6 +301,7 @@ int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct amdgpu_device *adev,
  		struct kgd_mem *mem, void **kptr, uint64_t *size);
  void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct amdgpu_device *adev,
  		struct kgd_mem *mem);
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, 
+struct amdgpu_bo *bo);
 
  int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info,
  					    struct dma_fence **ef);
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
index efab923056f4..2d452655eb04 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_sync_memory( · line 2034

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +2 -0
@@ -2030,6 +2030,45 @@ int amdgpu_amdkfd_gpuvm_sync_memory(
  	return ret;
  }
 
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, 
+struct amdgpu_bo *bo)

I think this function should take a reference count on the bo to ensure the pinned BO is not freed prematurely (even if broken user mode frees the BO before destroying the queue). Add a comment that the correct way to release the reference and unpin/unmap the BO is to call amdgpu_amdkfd_free_gtt_mem. See another comment below about moving the amdgpu_bo_ref here.

Quoted 14 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +14 -0
+{
+	int ret;
+
+	ret = amdgpu_bo_reserve(bo, true);
+	if (ret) {
+		pr_err("Failed to reserve bo. ret %d\n", ret);
+		goto err_reserve_bo_failed;
+	}
+
+	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
+	if (ret) {
+		pr_err("Failed to pin bo. ret %d\n", ret);
+		goto err_pin_bo_failed;
+	}

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_sync_memory( · line 2057

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +10 -0
+
+	ret = amdgpu_ttm_alloc_gart(&bo->tbo);
+	if (ret) {
+		pr_err("Failed to bind bo to GART. ret %d\n", ret);
+		goto err_map_bo_gart_failed;
+	}
+
+	amdgpu_amdkfd_remove_eviction_fence(
+		bo, bo->kfd_bo->process_info->eviction_fence);
+	list_del_init(&bo->kfd_bo->validate_list.head);

Please see Lang, Yu's patch "drm/amdkfd: add pinned BOs to kfd_bo_list".
We realized that pinned BOs still need to be on the validate list. So please remove the list_del_init here.

Quoted 22 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +14 -0
+
+	amdgpu_bo_unreserve(bo);
+
+	return 0;
+
+err_map_bo_gart_failed:
+	amdgpu_bo_unpin(bo);
+err_pin_bo_failed:
+	amdgpu_bo_unreserve(bo);
+err_reserve_bo_failed:
+
+	return ret;
+}
+
  int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct amdgpu_device *adev,
  		struct kgd_mem *mem, void **kptr, uint64_t *size)
  {
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
index e9766e165c38..58d5ebed1b32 100644

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, · line 295

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +3 -0
@@ -290,6 +290,11 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
  	struct queue_properties q_properties;
  	uint32_t doorbell_offset_in_process = 0;
 
+	struct amdgpu_bo_va_mapping *wptr_mapping;
+	struct interval_tree_node *wptr_node;
+	struct amdgpu_vm *wptr_vm;

It's good practice to minimize the scope of local variables. The above three variables are only used inside the "if" below. So they should be declared there.

Quoted 23 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +16 -0
+	struct amdgpu_bo *wptr_bo = NULL;
+
  	memset(&q_properties, 0, sizeof(struct queue_properties));
 
  	pr_debug("Creating queue ioctl\n");
@@ -316,12 +321,44 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
  		goto err_bind_process;
  	}
 
+	/* Starting with GFX11, MES requires wptr BOs to be GTT allocated/mapped to
+	 * GART for usermode queues in order to support oversubscription. In the
+	 * case that work is submitted to an unmapped queue, MES must have a GART
+	 * wptr address to determine whether the queue should be mapped.
+	 */
+	if (dev->shared_resources.enable_mes && (dev->adev->mes.sched_version & 0xff) >= 3) {
+		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
+		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
+		if (err)
+			goto err_wptr_bo_reserve;
+
+		wptr_node = interval_tree_iter_first(&wptr_vm->va,
+				args->write_pointer_address >> PAGE_SHIFT,
+				args->write_pointer_address >> PAGE_SHIFT);

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, · line 347

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +10 -0
+		if (!wptr_node) {
+			pr_err("Failed to lookup wptr bo\n");
+			err = -EINVAL;
+			goto err_wptr_bo_lookup;
+		}
+
+		wptr_mapping = container_of((struct rb_node *)wptr_node, struct amdgpu_bo_va_mapping, rb);
+		wptr_bo = wptr_mapping->bo_va->base.bo;
+
+		amdgpu_bo_unreserve(wptr_vm->root.bo);

You could move this to just after the interval_tree_iter_first call, even before it checks for errors. That way you don't need to deal with the unreserve conditionally on the error handling code path. And you can probably just use one error label err_wptr_map_gart that handles all 3 failure cases.

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +8 -0
+
+		err = amdgpu_amdkfd_map_gtt_bo_to_gart(dev->adev, wptr_bo);
+		if (err) {
+			pr_err("Failed to map wptr bo to GART\n");
+			goto err_wptr_bo_gart_map;
+		}
+	}
+

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, · line 363

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2 -2
  	pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
  			p->pasid,
  			dev->id);
 
-	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, NULL, NULL, NULL,
-			&doorbell_offset_in_process);
+	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, wptr_bo,
+			NULL, NULL, NULL, &doorbell_offset_in_process);
  	if (err != 0)
  		goto err_create_queue;

The error handling for this and all later failures should gart-unmap the wptr_bo again.

Quoted 89 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +4 -0
@@ -353,7 +390,11 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
  	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0);
  	return 0;
 
+err_wptr_bo_lookup:
+	amdgpu_bo_unreserve(wptr_vm->root.bo);
  err_create_queue:
+err_wptr_bo_gart_map:
+err_wptr_bo_reserve:
  err_bind_process:
  err_pdd:
  	mutex_unlock(&p->mutex);
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +8 -1
index b39d89c52887..d8de2fbdfc7d 100644
@@ -208,6 +208,7 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
  	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
  	struct mes_add_queue_input queue_input;
  	int r, queue_type;
+	uint64_t wptr_addr_off;
 
  	if (dqm->is_hws_hang)
  		return -EIO;
@@ -227,7 +228,13 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
  					AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
  	queue_input.doorbell_offset = q->properties.doorbell_off;
  	queue_input.mqd_addr = q->gart_mqd_addr;
-	queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
+
+	if (q->wptr_bo) {
+		wptr_addr_off = (uint64_t)q->properties.write_ptr - (uint64_t)q->wptr_bo->kfd_bo->va;
+		queue_input.wptr_addr = ((uint64_t)q->wptr_bo->tbo.resource->start << PAGE_SHIFT) + wptr_addr_off;
+	} else
+		queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
+
  	queue_input.paging = false;
  	queue_input.tba_addr = qpd->tba_addr;
  	queue_input.tma_addr = qpd->tma_addr;
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +2 -0
index f1654b4da856..35e74bdd81da 100644
@@ -377,6 +377,8 @@ static void update_mqd_sdma(struct mqd_manager *mm, void *mqd,
  	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
  	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
  	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
+	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
+	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
  	m->sdmax_rlcx_doorbell_offset =
  		q->doorbell_off << SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;
 
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +3 -0
index a5d3963537d7..dcddee0d6f06 100644
@@ -639,6 +639,8 @@ struct queue {
  	void *gang_ctx_bo;
  	uint64_t gang_ctx_gpu_addr;
  	void *gang_ctx_cpu_ptr;
+
+	struct amdgpu_bo *wptr_bo;
  };
 
  enum KFD_MQD_TYPE {
@@ -1404,6 +1406,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
  			    struct file *f,
  			    struct queue_properties *properties,
  			    unsigned int *qid,
+			    struct amdgpu_bo *wptr_bo,
  			    const struct kfd_criu_queue_priv_data *q_data,
  			    const void *restore_mqd,
  			    const void *restore_ctl_stack,
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +2 -1
index f99e09dc43ea..ede1462a55c8 100644
@@ -190,7 +190,8 @@ void pqm_uninit(struct process_queue_manager *pqm)
  static int init_user_queue(struct process_queue_manager *pqm,
  				struct kfd_dev *dev, struct queue **q,
  				struct queue_properties *q_properties,
-				struct file *f, unsigned int qid)
+				struct file *f, struct amdgpu_bo *wptr_bo,
+				unsigned int qid)
  {
  	int retval;
 

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c static int init_user_queue(struct process_queue_manager *pqm, · line 227

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +3 -0
@@ -221,6 +222,9 @@ static int init_user_queue(struct process_queue_manager *pqm,
  			goto cleanup;
  		}
  		memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
+
+		if (wptr_bo)
+			(*q)->wptr_bo = amdgpu_bo_ref(wptr_bo);

I think it would be clearer if you took that reference in amdgpu_amdkfd_map_gtt_bo_to_gart. That way it would be a simple assignment here with no "if". And all the BO reference handling would be more neatly contained in amdgpu_amdkfd code.

Regards,
  Felix

Quoted 53 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +9 -4
  	}
 
  	pr_debug("PQM After init queue");
@@ -237,6 +241,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
  			    struct file *f,
  			    struct queue_properties *properties,
  			    unsigned int *qid,
+			    struct amdgpu_bo *wptr_bo,
  			    const struct kfd_criu_queue_priv_data *q_data,
  			    const void *restore_mqd,
  			    const void *restore_ctl_stack,
@@ -299,7 +304,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
  		 * allocate_sdma_queue() in create_queue() has the
  		 * corresponding check logic.
  		 */
-		retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
+		retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
  		if (retval != 0)
  			goto err_create_queue;
  		pqn->q = q;
@@ -320,7 +325,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
  			goto err_create_queue;
  		}
 
-		retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
+		retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
  		if (retval != 0)
  			goto err_create_queue;
  		pqn->q = q;
@@ -457,9 +462,13 @@ int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
  			pdd->qpd.num_gws = 0;
  		}
 
-		if (dev->shared_resources.enable_mes)
+		if (dev->shared_resources.enable_mes) {
  			amdgpu_amdkfd_free_gtt_mem(dev->adev,
  						   pqn->q->gang_ctx_bo);
+			if (pqn->q->wptr_bo)
+				amdgpu_amdkfd_free_gtt_mem(dev->adev, pqn->q->wptr_bo);
+
+		}
  		uninit_queue(pqn->q);
  	}
 
@@ -900,7 +909,7 @@ int kfd_criu_restore_queue(struct kfd_process *p,
 
  	print_queue_properties(&qp);
 
-	ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, q_data, mqd, ctl_stack,
+	ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, NULL, q_data, mqd, ctl_stack,
  				NULL);
  	if (ret) {
  		pr_err("Failed to create new queue err:%d\n", ret);
#4 lore raw

Re: [PATCH v2 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

In reply to Graham Sider
Quoted 114 lines · Graham Sider

Starting with GFX11, MES requires wptr BOs to be GTT allocated/mapped to GART for usermode queues in order to support oversubscription. In the case that work is submitted to an unmapped queue, MES must have a GART wptr address to determine whether the queue should be mapped.

This change is accompanied with changes in MES and is applicable for MES_VERSION >= 3.

v2: Update MES_VERSION check from 2 to 3.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 39 ++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 45 ++++++++++++++++++-
 .../drm/amd/amdkfd/kfd_device_queue_manager.c |  9 +++-
 .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c  |  2 +
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |  3 ++
 .../amd/amdkfd/kfd_process_queue_manager.c    | 19 +++++---
 7 files changed, 110 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -0
index 429b16ba10bf..dba26d1e3be9 100644
@@ -301,6 +301,7 @@ int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct amdgpu_device *adev,
 		struct kgd_mem *mem, void **kptr, uint64_t *size);
 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct amdgpu_device *adev,
 		struct kgd_mem *mem);
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo);
 
 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info,
 					    struct dma_fence **ef);
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +39 -0
index efab923056f4..2d452655eb04 100644
@@ -2030,6 +2030,45 @@ int amdgpu_amdkfd_gpuvm_sync_memory(
 	return ret;
 }
 
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo)
+{
+	int ret;
+
+	ret = amdgpu_bo_reserve(bo, true);
+	if (ret) {
+		pr_err("Failed to reserve bo. ret %d\n", ret);
+		goto err_reserve_bo_failed;
+	}
+
+	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
+	if (ret) {
+		pr_err("Failed to pin bo. ret %d\n", ret);
+		goto err_pin_bo_failed;
+	}
+
+	ret = amdgpu_ttm_alloc_gart(&bo->tbo);
+	if (ret) {
+		pr_err("Failed to bind bo to GART. ret %d\n", ret);
+		goto err_map_bo_gart_failed;
+	}
+
+	amdgpu_amdkfd_remove_eviction_fence(
+		bo, bo->kfd_bo->process_info->eviction_fence);
+	list_del_init(&bo->kfd_bo->validate_list.head);
+
+	amdgpu_bo_unreserve(bo);
+
+	return 0;
+
+err_map_bo_gart_failed:
+	amdgpu_bo_unpin(bo);
+err_pin_bo_failed:
+	amdgpu_bo_unreserve(bo);
+err_reserve_bo_failed:
+
+	return ret;
+}
+
 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct amdgpu_device *adev,
 		struct kgd_mem *mem, void **kptr, uint64_t *size)
 {
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +16 -0
index e9766e165c38..58d5ebed1b32 100644
@@ -290,6 +290,11 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 	struct queue_properties q_properties;
 	uint32_t doorbell_offset_in_process = 0;
 
+	struct amdgpu_bo_va_mapping *wptr_mapping;
+	struct interval_tree_node *wptr_node;
+	struct amdgpu_vm *wptr_vm;
+	struct amdgpu_bo *wptr_bo = NULL;
+
 	memset(&q_properties, 0, sizeof(struct queue_properties));
 
 	pr_debug("Creating queue ioctl\n");
@@ -316,12 +321,44 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 		goto err_bind_process;
 	}
 
+	/* Starting with GFX11, MES requires wptr BOs to be GTT allocated/mapped to
+	 * GART for usermode queues in order to support oversubscription. In the
+	 * case that work is submitted to an unmapped queue, MES must have a GART
+	 * wptr address to determine whether the queue should be mapped.
+	 */
+	if (dev->shared_resources.enable_mes && (dev->adev->mes.sched_version & 0xff) >= 3) {
+		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
+		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
+		if (err)
+			goto err_wptr_bo_reserve;
+

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, · line 344

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +10 -0
+		wptr_node = interval_tree_iter_first(&wptr_vm->va,
+				args->write_pointer_address >> PAGE_SHIFT,
+				args->write_pointer_address >> PAGE_SHIFT);
+		if (!wptr_node) {
+			pr_err("Failed to lookup wptr bo\n");
+			err = -EINVAL;
+			goto err_wptr_bo_lookup;
+		}
+
+		wptr_mapping = container_of((struct rb_node *)wptr_node, struct amdgpu_bo_va_mapping, rb);

You can use amdgpu_vm_bo_lookup_mapping() to find a mapping by address.

Regards,
Lang

Quoted 172 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +17 -2
+		wptr_bo = wptr_mapping->bo_va->base.bo;
+
+		amdgpu_bo_unreserve(wptr_vm->root.bo);
+
+		err = amdgpu_amdkfd_map_gtt_bo_to_gart(dev->adev, wptr_bo);
+		if (err) {
+			pr_err("Failed to map wptr bo to GART\n");
+			goto err_wptr_bo_gart_map;
+		}
+	}
+
 	pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
 			p->pasid,
 			dev->id);
 
-	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, NULL, NULL, NULL,
-			&doorbell_offset_in_process);
+	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, wptr_bo,
+			NULL, NULL, NULL, &doorbell_offset_in_process);
 	if (err != 0)
 		goto err_create_queue;
 
@@ -353,7 +390,11 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0);
 	return 0;
 
+err_wptr_bo_lookup:
+	amdgpu_bo_unreserve(wptr_vm->root.bo);
 err_create_queue:
+err_wptr_bo_gart_map:
+err_wptr_bo_reserve:
 err_bind_process:
 err_pdd:
 	mutex_unlock(&p->mutex);
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +8 -1
index b39d89c52887..d8de2fbdfc7d 100644
@@ -208,6 +208,7 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
 	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
 	struct mes_add_queue_input queue_input;
 	int r, queue_type;
+	uint64_t wptr_addr_off;
 
 	if (dqm->is_hws_hang)
 		return -EIO;
@@ -227,7 +228,13 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
 					AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
 	queue_input.doorbell_offset = q->properties.doorbell_off;
 	queue_input.mqd_addr = q->gart_mqd_addr;
-	queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
+
+	if (q->wptr_bo) {
+		wptr_addr_off = (uint64_t)q->properties.write_ptr - (uint64_t)q->wptr_bo->kfd_bo->va;
+		queue_input.wptr_addr = ((uint64_t)q->wptr_bo->tbo.resource->start << PAGE_SHIFT) + wptr_addr_off;
+	} else
+		queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
+
 	queue_input.paging = false;
 	queue_input.tba_addr = qpd->tba_addr;
 	queue_input.tma_addr = qpd->tma_addr;
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +2 -0
index f1654b4da856..35e74bdd81da 100644
@@ -377,6 +377,8 @@ static void update_mqd_sdma(struct mqd_manager *mm, void *mqd,
 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
+	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
+	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
 	m->sdmax_rlcx_doorbell_offset =
 		q->doorbell_off << SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;
 
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +3 -0
index a5d3963537d7..dcddee0d6f06 100644
@@ -639,6 +639,8 @@ struct queue {
 	void *gang_ctx_bo;
 	uint64_t gang_ctx_gpu_addr;
 	void *gang_ctx_cpu_ptr;
+
+	struct amdgpu_bo *wptr_bo;
 };
 
 enum KFD_MQD_TYPE {
@@ -1404,6 +1406,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
 			    struct file *f,
 			    struct queue_properties *properties,
 			    unsigned int *qid,
+			    struct amdgpu_bo *wptr_bo,
 			    const struct kfd_criu_queue_priv_data *q_data,
 			    const void *restore_mqd,
 			    const void *restore_ctl_stack,
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +14 -5
index f99e09dc43ea..ede1462a55c8 100644
@@ -190,7 +190,8 @@ void pqm_uninit(struct process_queue_manager *pqm)
 static int init_user_queue(struct process_queue_manager *pqm,
 				struct kfd_dev *dev, struct queue **q,
 				struct queue_properties *q_properties,
-				struct file *f, unsigned int qid)
+				struct file *f, struct amdgpu_bo *wptr_bo,
+				unsigned int qid)
 {
 	int retval;
 
@@ -221,6 +222,9 @@ static int init_user_queue(struct process_queue_manager *pqm,
 			goto cleanup;
 		}
 		memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
+
+		if (wptr_bo)
+			(*q)->wptr_bo = amdgpu_bo_ref(wptr_bo);
 	}
 
 	pr_debug("PQM After init queue");
@@ -237,6 +241,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
 			    struct file *f,
 			    struct queue_properties *properties,
 			    unsigned int *qid,
+			    struct amdgpu_bo *wptr_bo,
 			    const struct kfd_criu_queue_priv_data *q_data,
 			    const void *restore_mqd,
 			    const void *restore_ctl_stack,
@@ -299,7 +304,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
 		 * allocate_sdma_queue() in create_queue() has the
 		 * corresponding check logic.
 		 */
-		retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
+		retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
 		if (retval != 0)
 			goto err_create_queue;
 		pqn->q = q;
@@ -320,7 +325,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
 			goto err_create_queue;
 		}
 
-		retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
+		retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
 		if (retval != 0)
 			goto err_create_queue;
 		pqn->q = q;
@@ -457,9 +462,13 @@ int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
 			pdd->qpd.num_gws = 0;
 		}
 
-		if (dev->shared_resources.enable_mes)
+		if (dev->shared_resources.enable_mes) {
 			amdgpu_amdkfd_free_gtt_mem(dev->adev,
 						   pqn->q->gang_ctx_bo);
+			if (pqn->q->wptr_bo)
+				amdgpu_amdkfd_free_gtt_mem(dev->adev, pqn->q->wptr_bo);
+
+		}
 		uninit_queue(pqn->q);
 	}
 
@@ -900,7 +909,7 @@ int kfd_criu_restore_queue(struct kfd_process *p,
 
 	print_queue_properties(&qp);
 
-	ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, q_data, mqd, ctl_stack,
+	ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, NULL, q_data, mqd, ctl_stack,
 				NULL);
 	if (ret) {
 		pr_err("Failed to create new queue err:%d\n", ret);
Signature · 1 line
2.25.1
#5 lore raw

RE: [PATCH v2 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

[Public]

Quoted 154 lines

-----Original Message-----
From: Yu, Lang <[email protected]>
Sent: Friday, June 10, 2022 10:38 PM
To: Sider, Graham <[email protected]>
Cc: [email protected]; Joshi, Mukul <[email protected]>; Kuehling, Felix <[email protected]>; Yang, Philip <[email protected]>
Subject: Re: [PATCH v2 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

Quoted 8 lines · Graham Sider

Starting with GFX11, MES requires wptr BOs to be GTT allocated/mapped to GART for usermode queues in order to support oversubscription. In the case that work is submitted to an unmapped queue, MES must have a GART wptr address to determine whether the queue should be mapped.

This change is accompanied with changes in MES and is applicable for MES_VERSION >= 3.

Quoted 6 lines

v2: Update MES_VERSION check from 2 to 3.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 39

++++++++++++++++

Quoted 1 line
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 45

++++++++++++++++++-

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int

Quoted 12 lines
 .../drm/amd/amdkfd/kfd_device_queue_manager.c |  9 +++-
.../gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c  |  2 +
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |  3 ++
 .../amd/amdkfd/kfd_process_queue_manager.c    | 19 +++++---
 7 files changed, 110 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0 -0
index 429b16ba10bf..dba26d1e3be9 100644
@@ -301,6 +301,7 @@ int

amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct amdgpu_device *adev,

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0 -0
 		struct kgd_mem *mem, void **kptr, uint64_t *size);  void

amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct

amdgpu_device *adev,

Quoted 45 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +2 -0
 		struct kgd_mem *mem);
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev,
+struct amdgpu_bo *bo);
 
 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info,
 					    struct dma_fence **ef);
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +30 -0
index efab923056f4..2d452655eb04 100644
@@ -2030,6 +2030,45 @@ int amdgpu_amdkfd_gpuvm_sync_memory(
 	return ret;
 }
 
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev,
+struct amdgpu_bo *bo) {
+	int ret;
+
+	ret = amdgpu_bo_reserve(bo, true);
+	if (ret) {
+		pr_err("Failed to reserve bo. ret %d\n", ret);
+		goto err_reserve_bo_failed;
+	}
+
+	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
+	if (ret) {
+		pr_err("Failed to pin bo. ret %d\n", ret);
+		goto err_pin_bo_failed;
+	}
+
+	ret = amdgpu_ttm_alloc_gart(&bo->tbo);
+	if (ret) {
+		pr_err("Failed to bind bo to GART. ret %d\n", ret);
+		goto err_map_bo_gart_failed;
+	}
+
+	amdgpu_amdkfd_remove_eviction_fence(
+		bo, bo->kfd_bo->process_info->eviction_fence);
+	list_del_init(&bo->kfd_bo->validate_list.head);
+
+	amdgpu_bo_unreserve(bo);
+
+	return 0;
+

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_sync_memory( · line 2072

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +9 -0
+err_map_bo_gart_failed:
+	amdgpu_bo_unpin(bo);
+err_pin_bo_failed:
+	amdgpu_bo_unreserve(bo);
+err_reserve_bo_failed:
+
+	return ret;
+}
+
 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct

amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_sync_memory( · line 2073

Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
 		struct kgd_mem *mem, void **kptr, uint64_t *size)  { diff --

git

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep,

Quoted 6 lines

a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index e9766e165c38..58d5ebed1b32 100644

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
@@ -290,6 +290,11 @@ static int kfd_ioctl_create_queue(struct file *filep,

struct kfd_process *p,

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file

Quoted 12 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +5 -0
 	struct queue_properties q_properties;
 	uint32_t doorbell_offset_in_process = 0;
 
+	struct amdgpu_bo_va_mapping *wptr_mapping;
+	struct interval_tree_node *wptr_node;
+	struct amdgpu_vm *wptr_vm;
+	struct amdgpu_bo *wptr_bo = NULL;
+
 	memset(&q_properties, 0, sizeof(struct queue_properties));
 
 	pr_debug("Creating queue ioctl\n");
@@ -316,12 +321,44 @@ static int kfd_ioctl_create_queue(struct file

*filep, struct kfd_process *p,

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
 		goto err_bind_process;
 	}
 
+	/* Starting with GFX11, MES requires wptr BOs to be GTT

allocated/mapped to

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file · line 325

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
+	 * GART for usermode queues in order to support oversubscription.

In the

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file · line 326

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
+	 * case that work is submitted to an unmapped queue, MES must

have a GART

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,

Quoted 11 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +10 -0
+	 * wptr address to determine whether the queue should be mapped.
+	 */
+	if (dev->shared_resources.enable_mes && (dev->adev-
 mes.sched_version & 0xff) >= 3) {
+		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
+		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
+		if (err)
+			goto err_wptr_bo_reserve;
+
+		wptr_node = interval_tree_iter_first(&wptr_vm->va,
+				args->write_pointer_address >>

PAGE_SHIFT,

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file · line 338

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
+				args->write_pointer_address >>

PAGE_SHIFT);

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,

Quoted 7 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +7 -0
+		if (!wptr_node) {
+			pr_err("Failed to lookup wptr bo\n");
+			err = -EINVAL;
+			goto err_wptr_bo_lookup;
+		}
+
+		wptr_mapping = container_of((struct rb_node *)wptr_node,

struct

Quoted 7 lines

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file · line 346

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
+amdgpu_bo_va_mapping, rb);

You can use amdgpu_vm_bo_lookup_mapping() to find a mapping by address.

Regards,
Lang

Good catch, will use this. Thanks!

Best,
Graham

Quoted 211 lines

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +5 -0
+		wptr_bo = wptr_mapping->bo_va->base.bo;
+
+		amdgpu_bo_unreserve(wptr_vm->root.bo);
+
+		err = amdgpu_amdkfd_map_gtt_bo_to_gart(dev->adev,

wptr_bo);

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,

Quoted 11 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +6 -1
+		if (err) {
+			pr_err("Failed to map wptr bo to GART\n");
+			goto err_wptr_bo_gart_map;
+		}
+	}
+
 	pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
 			p->pasid,
 			dev->id);
 
-	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties,

&queue_id, NULL, NULL, NULL,

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file · line 362

Quoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -1
-			&doorbell_offset_in_process);
+	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties,

&queue_id, wptr_bo,

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep,

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
+			NULL, NULL, NULL, &doorbell_offset_in_process);
 	if (err != 0)
 		goto err_create_queue;
@@ -353,7 +390,11 @@ static int kfd_ioctl_create_queue(struct file *filep,

struct kfd_process *p,

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, · line 390

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
 	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev,

queue_id, false, NULL, 0);

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int add_queue_mes(struct

Quoted 16 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +4 -0
 	return 0;
 
+err_wptr_bo_lookup:
+	amdgpu_bo_unreserve(wptr_vm->root.bo);
 err_create_queue:
+err_wptr_bo_gart_map:
+err_wptr_bo_reserve:
 err_bind_process:
 err_pdd:
 	mutex_unlock(&p->mutex);
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
index b39d89c52887..d8de2fbdfc7d 100644
@@ -208,6 +208,7 @@ static int add_queue_mes(struct

device_queue_manager *dqm, struct queue *q,

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int add_queue_mes(struct

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -0
 	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
 	struct mes_add_queue_input queue_input;
 	int r, queue_type;
+	uint64_t wptr_addr_off;
 
 	if (dqm->is_hws_hang)
 		return -EIO;
@@ -227,7 +228,13 @@ static int add_queue_mes(struct

device_queue_manager *dqm, struct queue *q,

Quoted 1 line
	AMDGPU_MES_PRIORITY_LEVEL_NORMAL;

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +3 -1
 	queue_input.doorbell_offset = q->properties.doorbell_off;
 	queue_input.mqd_addr = q->gart_mqd_addr;
-	queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
+
+	if (q->wptr_bo) {
+		wptr_addr_off = (uint64_t)q->properties.write_ptr -

(uint64_t)q->wptr_bo->kfd_bo->va;

drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c static void update_mqd_sdma(struct mqd_manager

Quoted 15 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -0
+		queue_input.wptr_addr = ((uint64_t)q->wptr_bo-

tbo.resource->start << PAGE_SHIFT) + wptr_addr_off;

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -0
+	} else
+		queue_input.wptr_addr = (uint64_t)q-

properties.write_ptr;

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -0
+
 	queue_input.paging = false;
 	queue_input.tba_addr = qpd->tba_addr;
 	queue_input.tma_addr = qpd->tma_addr; diff --git

a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c index f1654b4da856..35e74bdd81da 100644

drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +0 -0
@@ -377,6 +377,8 @@ static void update_mqd_sdma(struct mqd_manager

*mm, void *mqd,

drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c static void update_mqd_sdma(struct mqd_manager · line 377

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +0 -0
 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >>

8);

drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c static void update_mqd_sdma(struct mqd_manager

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +3 -0
 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q-
 read_ptr);
 	m->sdmax_rlcx_rb_rptr_addr_hi =
 upper_32_bits((uint64_t)q->read_ptr);
+	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits((uint64_t)q-
 write_ptr);
+	m->sdmax_rlcx_rb_wptr_poll_addr_hi =
+upper_32_bits((uint64_t)q->write_ptr);
 	m->sdmax_rlcx_doorbell_offset =
 		q->doorbell_off <<

SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;

drivers/gpu/drm/amd/amdkfd/kfd_priv.h int pqm_create_queue(struct

Quoted 16 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +2 -0
index a5d3963537d7..dcddee0d6f06 100644
@@ -639,6 +639,8 @@ struct queue {
 	void *gang_ctx_bo;
 	uint64_t gang_ctx_gpu_addr;
 	void *gang_ctx_cpu_ptr;
+
+	struct amdgpu_bo *wptr_bo;
 };
 
 enum KFD_MQD_TYPE {
@@ -1404,6 +1406,7 @@ int pqm_create_queue(struct

process_queue_manager *pqm,

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c void pqm_uninit(struct process_queue_manager

Quoted 13 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1 -0
 			    struct file *f,
 			    struct queue_properties *properties,
 			    unsigned int *qid,
+			    struct amdgpu_bo *wptr_bo,
 			    const struct kfd_criu_queue_priv_data *q_data,
 			    const void *restore_mqd,
 			    const void *restore_ctl_stack, diff --git

a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c index f99e09dc43ea..ede1462a55c8 100644

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +0 -0
@@ -190,7 +190,8 @@ void pqm_uninit(struct process_queue_manager

*pqm)

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c static int init_user_queue(struct

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +2 -1
 static int init_user_queue(struct process_queue_manager *pqm,
 				struct kfd_dev *dev, struct queue **q,
 				struct queue_properties *q_properties,
-				struct file *f, unsigned int qid)
+				struct file *f, struct amdgpu_bo *wptr_bo,
+				unsigned int qid)
 {
 	int retval;
 
@@ -221,6 +222,9 @@ static int init_user_queue(struct

process_queue_manager *pqm,

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c static int init_user_queue(struct process_queue_manager *pqm,

Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +0 -0
 			goto cleanup;
 		}
 		memset((*q)->gang_ctx_cpu_ptr, 0,

AMDGPU_MES_GANG_CTX_SIZE);

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c int pqm_create_queue(struct

Quoted 7 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +3 -0
+
+		if (wptr_bo)
+			(*q)->wptr_bo = amdgpu_bo_ref(wptr_bo);
 	}
 
 	pr_debug("PQM After init queue");
@@ -237,6 +241,7 @@ int pqm_create_queue(struct

process_queue_manager *pqm,

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c int pqm_create_queue(struct · line 247

Quoted 7 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -0
 			    struct file *f,
 			    struct queue_properties *properties,
 			    unsigned int *qid,
+			    struct amdgpu_bo *wptr_bo,
 			    const struct kfd_criu_queue_priv_data *q_data,
 			    const void *restore_mqd,
 			    const void *restore_ctl_stack, @@ -299,7 +304,7

@@ int

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c int pqm_create_queue(struct process_queue_manager *pqm,

Quoted 6 lines

pqm_create_queue(struct process_queue_manager *pqm,

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -1
 		 * allocate_sdma_queue() in create_queue() has the
 		 * corresponding check logic.
 		 */
-		retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
+		retval = init_user_queue(pqm, dev, &q, properties, f,

wptr_bo,

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c int pqm_create_queue(struct

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -0
+*qid);
 		if (retval != 0)
 			goto err_create_queue;
 		pqn->q = q;
@@ -320,7 +325,7 @@ int pqm_create_queue(struct

process_queue_manager *pqm,

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c int pqm_create_queue(struct process_queue_manager *pqm,

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -1
 			goto err_create_queue;
 		}
 
-		retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
+		retval = init_user_queue(pqm, dev, &q, properties, f,

wptr_bo,

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c int pqm_destroy_queue(struct

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -0
+*qid);
 		if (retval != 0)
 			goto err_create_queue;
 		pqn->q = q;
@@ -457,9 +462,13 @@ int pqm_destroy_queue(struct

process_queue_manager *pqm, unsigned int qid)

Quoted 16 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +5 -1
 			pdd->qpd.num_gws = 0;
 		}
 
-		if (dev->shared_resources.enable_mes)
+		if (dev->shared_resources.enable_mes) {
 			amdgpu_amdkfd_free_gtt_mem(dev->adev,
 						   pqn->q->gang_ctx_bo);
+			if (pqn->q->wptr_bo)
+				amdgpu_amdkfd_free_gtt_mem(dev-
 adev, pqn->q->wptr_bo);
+
+		}
 		uninit_queue(pqn->q);
 	}

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c int kfd_criu_restore_queue(struct kfd_process *p, · line 903

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +0 -1
@@ -900,7 +909,7 @@ int kfd_criu_restore_queue(struct kfd_process *p,
 
 	print_queue_properties(&qp);
 
-	ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp,

&queue_id, q_data, mqd, ctl_stack,

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c int kfd_criu_restore_queue(struct kfd_process *p, · line 912

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -0
+	ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp,

&queue_id,

Quoted 7 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -1
+NULL, q_data, mqd, ctl_stack,
 				NULL);
 	if (ret) {
 		pr_err("Failed to create new queue err:%d\n", ret);
--

2.25.1

#6 lore raw

PATCH v2 3/3 drm/amdgpu: Update mes_v11_api_def.h

In reply to Graham Sider

Update MES API to support oversubscription without aggregated doorbell for usermode queues.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c               | 1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h               | 1 +
 drivers/gpu/drm/amd/amdgpu/mes_v11_0.c                | 3 +++
 drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 2 ++
 drivers/gpu/drm/amd/include/mes_v11_api_def.h         | 4 +++-
 5 files changed, 10 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +1 -0
index 2e86baa32c55..3d9a81a8fa1c 100644
@@ -681,6 +681,7 @@ int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
 	queue_input.wptr_addr = qprops->wptr_gpu_addr;
 	queue_input.queue_type = qprops->queue_type;
 	queue_input.paging = qprops->paging;
+	queue_input.oversubscription_no_aggregated_en = 0;
 
 	r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
 	if (r) {
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +1 -0
index aa06c8396ee0..26765a9946a9 100644
@@ -207,6 +207,7 @@ struct mes_add_queue_input {
 	uint32_t        debug_vmid;
 	uint64_t	tba_addr;
 	uint64_t	tma_addr;
+	uint64_t	oversubscription_no_aggregated_en;
 };
 
 struct mes_remove_queue_input {
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +3 -0
index 2a9ef308e71c..95a1394d3943 100644
@@ -163,6 +163,8 @@ static int mes_v11_0_add_hw_queue(struct amdgpu_mes *mes,
 	mes_add_queue_pkt.gws_size = input->gws_size;
 	mes_add_queue_pkt.trap_handler_addr = input->tba_addr;
 	mes_add_queue_pkt.tma_addr = input->tma_addr;
+	mes_add_queue_pkt.oversubscription_no_aggregated_en =
+		input->oversubscription_no_aggregated_en;
 
 	mes_add_queue_pkt.api_status.api_completion_fence_addr =
 		mes->ring.fence_drv.gpu_addr;
@@ -341,6 +343,7 @@ static int mes_v11_0_set_hw_resources(struct amdgpu_mes *mes)
 	mes_set_hw_res_pkt.disable_reset = 1;
 	mes_set_hw_res_pkt.disable_mes_log = 1;
 	mes_set_hw_res_pkt.use_different_vmid_compute = 1;
+	mes_set_hw_res_pkt.oversubscription_timer = 50;
 
 	mes_set_hw_res_pkt.api_status.api_completion_fence_addr =
 		mes->ring.fence_drv.gpu_addr;
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -0
index d8de2fbdfc7d..762bc6059387 100644
@@ -235,6 +235,8 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
 	} else
 		queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
 
+	queue_input.oversubscription_no_aggregated_en = 1;
+
 	queue_input.paging = false;
 	queue_input.tba_addr = qpd->tba_addr;
 	queue_input.tma_addr = qpd->tma_addr;
drivers/gpu/drm/amd/include/mes_v11_api_def.h +3 -1
index f9d02d7bdf77..95f0246eb045 100644
@@ -226,6 +226,7 @@ union MESAPI_SET_HW_RESOURCES {
 			};
 			uint32_t	uint32_t_all;
 		};
+		uint32_t	oversubscription_timer;
 	};
 
 	uint32_t	max_dwords_in_api[API_FRAME_SIZE_IN_DWORDS];
@@ -265,7 +266,8 @@ union MESAPI__ADD_QUEUE {
 			uint32_t is_gang_suspended	: 1;
 			uint32_t is_tmz_queue		: 1;
 			uint32_t map_kiq_utility_queue  : 1;
-			uint32_t reserved		: 23;
+			uint32_t oversubscription_no_aggregated_en : 1;
+			uint32_t reserved		: 22;
 		};
 		struct MES_API_STATUS		api_status;
 		uint64_t                        tma_addr;
Signature · 1 line
2.25.1
maintainer Acked-by on v2 3/3
#7 lore raw

Re: [PATCH v2 3/3] drm/amdgpu: Update mes_v11_api_def.h

Am 2022-06-10 um 13:13 schrieb Graham Sider:

Quoted 4 lines

Update MES API to support oversubscription without aggregated doorbell for usermode queues.

Quoted 88 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c               | 1 +
  drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h               | 1 +
  drivers/gpu/drm/amd/amdgpu/mes_v11_0.c                | 3 +++
  drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 2 ++
  drivers/gpu/drm/amd/include/mes_v11_api_def.h         | 4 +++-
  5 files changed, 10 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +1 -0
index 2e86baa32c55..3d9a81a8fa1c 100644
@@ -681,6 +681,7 @@ int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
  	queue_input.wptr_addr = qprops->wptr_gpu_addr;
  	queue_input.queue_type = qprops->queue_type;
  	queue_input.paging = qprops->paging;
+	queue_input.oversubscription_no_aggregated_en = 0;
 
  	r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
  	if (r) {
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +1 -0
index aa06c8396ee0..26765a9946a9 100644
@@ -207,6 +207,7 @@ struct mes_add_queue_input {
  	uint32_t        debug_vmid;
  	uint64_t	tba_addr;
  	uint64_t	tma_addr;
+	uint64_t	oversubscription_no_aggregated_en;
  };
 
  struct mes_remove_queue_input {
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +3 -0
index 2a9ef308e71c..95a1394d3943 100644
@@ -163,6 +163,8 @@ static int mes_v11_0_add_hw_queue(struct amdgpu_mes *mes,
  	mes_add_queue_pkt.gws_size = input->gws_size;
  	mes_add_queue_pkt.trap_handler_addr = input->tba_addr;
  	mes_add_queue_pkt.tma_addr = input->tma_addr;
+	mes_add_queue_pkt.oversubscription_no_aggregated_en =
+		input->oversubscription_no_aggregated_en;
 
  	mes_add_queue_pkt.api_status.api_completion_fence_addr =
  		mes->ring.fence_drv.gpu_addr;
@@ -341,6 +343,7 @@ static int mes_v11_0_set_hw_resources(struct amdgpu_mes *mes)
  	mes_set_hw_res_pkt.disable_reset = 1;
  	mes_set_hw_res_pkt.disable_mes_log = 1;
  	mes_set_hw_res_pkt.use_different_vmid_compute = 1;
+	mes_set_hw_res_pkt.oversubscription_timer = 50;
 
  	mes_set_hw_res_pkt.api_status.api_completion_fence_addr =
  		mes->ring.fence_drv.gpu_addr;
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -0
index d8de2fbdfc7d..762bc6059387 100644
@@ -235,6 +235,8 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
  	} else
  		queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
 
+	queue_input.oversubscription_no_aggregated_en = 1;
+
  	queue_input.paging = false;
  	queue_input.tba_addr = qpd->tba_addr;
  	queue_input.tma_addr = qpd->tma_addr;
drivers/gpu/drm/amd/include/mes_v11_api_def.h +3 -1
index f9d02d7bdf77..95f0246eb045 100644
@@ -226,6 +226,7 @@ union MESAPI_SET_HW_RESOURCES {
  			};
  			uint32_t	uint32_t_all;
  		};
+		uint32_t	oversubscription_timer;
  	};
 
  	uint32_t	max_dwords_in_api[API_FRAME_SIZE_IN_DWORDS];
@@ -265,7 +266,8 @@ union MESAPI__ADD_QUEUE {
  			uint32_t is_gang_suspended	: 1;
  			uint32_t is_tmz_queue		: 1;
  			uint32_t map_kiq_utility_queue  : 1;
-			uint32_t reserved		: 23;
+			uint32_t oversubscription_no_aggregated_en : 1;
+			uint32_t reserved		: 22;
  		};
  		struct MES_API_STATUS		api_status;
  		uint64_t                        tma_addr;
maintainer Acked-by on v2 1/3
#8 lore raw
In reply to Graham Sider

Am 2022-06-10 um 13:13 schrieb Graham Sider:

Quoted 3 lines

Store MES scheduler and MES KIQ version numbers in amdgpu_mes for GFX11.

Quoted 42 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h |  3 +++
  drivers/gpu/drm/amd/amdgpu/mes_v11_0.c  | 12 ++++++++++++
  2 files changed, 15 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +3 -0
index 92ddee5e33db..aa06c8396ee0 100644
@@ -64,6 +64,9 @@ struct amdgpu_mes {
 
  	spinlock_t                      queue_id_lock;
 
+	uint32_t			sched_version;
+	uint32_t			kiq_version;
+
  	uint32_t                        total_max_queue;
  	uint32_t                        doorbell_id_offset;
  	uint32_t                        max_doorbell_slices;
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +12 -0
index e4eb87689f7f..2a9ef308e71c 100644
@@ -890,6 +890,18 @@ static int mes_v11_0_queue_init(struct amdgpu_device *adev,
  		mes_v11_0_queue_init_register(ring);
  	}
 
+	/* get MES scheduler/KIQ versions */
+	mutex_lock(&adev->srbm_mutex);
+	soc21_grbm_select(adev, 3, pipe, 0, 0);
+
+	if (pipe == AMDGPU_MES_SCHED_PIPE)
+		adev->mes.sched_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
+	else if (pipe == AMDGPU_MES_KIQ_PIPE && adev->enable_mes_kiq)
+		adev->mes.kiq_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
+
+	soc21_grbm_select(adev, 0, 0, 0, 0);
+	mutex_unlock(&adev->srbm_mutex);
+
  	return 0;
  }