Skip to contentSkip to conversation

drm/amdgpu: Fetch MES scheduler/KIQ versions

PATCH Superseded Graham Sider 6 participants 13 messages Jun 13, 2022 – Jun 17, 2022

Subsystems (3)

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#10 lore raw

PATCH v3 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
#0 lore raw

PATCH v3 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 v3 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.
v3:
- Use amdgpu_vm_bo_lookup_mapping for wptr_bo mapping lookup
- Move wptr_bo refcount increment to amdgpu_amdkfd_map_gtt_bo_to_gart
- Remove list_del_init from amdgpu_amdkfd_map_gtt_bo_to_gart
- Cleanup/fix create_queue wptr_bo error handling

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 49 +++++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 37 +++++++++++++-
 .../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    | 17 +++++--
 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 +49 -0

55 lines not on this page · raw message

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +35 -2
index e9766e165c38..1789ed8b79f5 100644
@@ -289,6 +289,7 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 	struct kfd_process_device *pdd;
 	struct queue_properties q_properties;
 	uint32_t doorbell_offset_in_process = 0;
+	struct amdgpu_bo *wptr_bo = NULL;
 
 	memset(&q_properties, 0, sizeof(struct queue_properties));
 
@@ -316,12 +317,41 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 		goto err_bind_process;
 	}
 
+	/* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
+	 * on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
+	 */
+	if (dev->shared_resources.enable_mes && (dev->adev->mes.sched_version & 0xff) >= 3) {
+		struct amdgpu_bo_va_mapping *wptr_mapping;
+		struct amdgpu_vm *wptr_vm;
+
+		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
+		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
+		if (err)
+			goto err_wptr_map_gart;
+
+		wptr_mapping = amdgpu_vm_bo_lookup_mapping(
+				wptr_vm, args->write_pointer_address >> PAGE_SHIFT);
+		amdgpu_bo_unreserve(wptr_vm->root.bo);
+		if (!wptr_mapping) {
+			pr_err("Failed to lookup wptr bo\n");
+			err = -EINVAL;
+			goto err_wptr_map_gart;
+		}
+
+		wptr_bo = wptr_mapping->bo_va->base.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_map_gart;
+		}
+	}
+
 	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;
 
@@ -354,6 +384,9 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 	return 0;
 
 err_create_queue:
+	if (wptr_bo)
+		amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
+err_wptr_map_gart:
 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 +12 -5
index f99e09dc43ea..3a17c1ebc527 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 +222,7 @@ static int init_user_queue(struct process_queue_manager *pqm,
 			goto cleanup;
 		}
 		memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
+		(*q)->wptr_bo = wptr_bo;
 	}
 
 	pr_debug("PQM After init queue");
@@ -237,6 +239,7 @@ 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 +302,7 @@ 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 +323,7 @@ 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 +460,13 @@ 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 +907,7 @@ 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
Reviewed-by on v3 1/3
#12 lore raw
In reply to Graham Sider

[AMD Official Use Only - General]

Regards,
Jack
________________________________
From: amd-gfx <[email protected]> on behalf of Graham Sider <[email protected]>
Sent: Monday, 13 June 2022 23:19
To: [email protected] <[email protected]>
Cc: Joshi, Mukul <[email protected]>; Kuehling, Felix <[email protected]>; Sider, Graham <[email protected]>; Yang, Philip <[email protected]>
Subject: [PATCH v3 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;
 }
 

--
2.25.1

#11 lore raw

Re: [PATCH v3 3/3] drm/amdgpu: Update mes_v11_api_def.h

In reply to Graham Sider

[AMD Official Use Only - General]

Quoted 1 line
Quoted 1 line

+ uint64_t oversubscription_no_aggregated_en;

uint64_t is unnecessary here, 32bit or bool type here better.

With fixed, the patch is Reviewed-by: Jack Xiao <[email protected]>

Regards,
Jack
________________________________
From: amd-gfx <[email protected]> on behalf of Graham Sider <[email protected]>
Sent: Monday, 13 June 2022 23:20
To: [email protected] <[email protected]>
Cc: Joshi, Mukul <[email protected]>; Kuehling, Felix <[email protected]>; Sider, Graham <[email protected]>; Yang, Philip <[email protected]>
Subject: [PATCH v3 3/3] drm/amdgpu: Update mes_v11_api_def.h

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;

--
2.25.1

Reviewed-by on v3 2/3
#2 lore raw

Re: [PATCH v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

In reply to Graham Sider

On 2022-06-13 11:20, Graham Sider wrote:
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. v3: - Use amdgpu_vm_bo_lookup_mapping for wptr_bo mapping lookup - Move wptr_bo refcount increment to amdgpu_amdkfd_map_gtt_bo_to_gart - Remove list_del_init from amdgpu_amdkfd_map_gtt_bo_to_gart - Cleanup/fix create_queue wptr_bo error handling Two nit-pick below, with those fixed, this patch is

Signed-off-by Graham Sider <[email protected]> --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 1 + .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 49 +++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 37 +++++++++++++- .../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 | 17 +++++-- 7 files changed, 110 insertions(+), 8 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index 429b16ba10bf..dba26d1e3be9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -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); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index efab923056f4..888d08128a94 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -2030,6 +2030,55 @@ int amdgpu_amdkfd_gpuvm_sync_memory( return ret; } +/** + * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count + * @adev: Device to which allocated BO belongs + * @bo: Buffer object to be mapped + * + * Before return, bo reference count is incremented. To release the reference and unpin/ + * unmap the BO, call amdgpu_amdkfd_free_gtt_mem. + */ +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); pinned bo should keep in validate_list as PDB/PTB may move and update, please remove list_del_init here.

+ + amdgpu_bo_unreserve(bo); + + bo = amdgpu_bo_ref(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) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index e9766e165c38..1789ed8b79f5 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -289,6 +289,7 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, struct kfd_process_device *pdd; struct queue_properties q_properties; uint32_t doorbell_offset_in_process = 0; + struct amdgpu_bo *wptr_bo = NULL; memset(&q_properties, 0, sizeof(struct queue_properties)); @@ -316,12 +317,41 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, goto err_bind_process; } + /* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work + * on unmapped queues for usermode queue oversubscription (no aggregated doorbell) + */ + if (dev->shared_resources.enable_mes && (dev->adev->mes.sched_version & 0xff) >= 3) { Should we check ip version for GFX11 only? Because GFX10 set enable_mes, and may set adev->mes.sched_version later as well.

Regards,
Philip

+ struct amdgpu_bo_va_mapping *wptr_mapping; + struct amdgpu_vm *wptr_vm; + + wptr_vm = drm_priv_to_vm(pdd->drm_priv); + err = amdgpu_bo_reserve(wptr_vm->root.bo, false); + if (err) + goto err_wptr_map_gart; + + wptr_mapping = amdgpu_vm_bo_lookup_mapping( + wptr_vm, args->write_pointer_address >> PAGE_SHIFT); + amdgpu_bo_unreserve(wptr_vm->root.bo); + if (!wptr_mapping) { + pr_err("Failed to lookup wptr bo\n"); + err = -EINVAL; + goto err_wptr_map_gart; + } + + wptr_bo = wptr_mapping->bo_va->base.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_map_gart; + } + } + 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; @@ -354,6 +384,9 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, return 0; err_create_queue: + if (wptr_bo) + amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo); +err_wptr_map_gart: err_bind_process: err_pdd: mutex_unlock(&p->mutex); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index b39d89c52887..d8de2fbdfc7d 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -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; 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 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c @@ -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; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index a5d3963537d7..dcddee0d6f06 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -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, 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..3a17c1ebc527 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c @@ -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 +222,7 @@ static int init_user_queue(struct process_queue_manager *pqm, goto cleanup; } memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE); + (*q)->wptr_bo = wptr_bo; } pr_debug("PQM After init queue"); @@ -237,6 +239,7 @@ 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 +302,7 @@ 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 +323,7 @@ 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 +460,13 @@ 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 +907,7 @@ 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 v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

[AMD Official Use Only - General]

Quoted 84 lines
Quoted 84 lines

From: Yang, Philip <[email protected]>
Sent: Tuesday, June 14, 2022 2:22 PM
To: Sider, Graham <[email protected]>; [email protected]
Cc: Kuehling, Felix <[email protected]>; Joshi, Mukul <[email protected]>; Yang, Philip <[email protected]>
Subject: Re: [PATCH v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

On 2022-06-13 11:20, Graham Sider wrote:
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.
v3:
- Use amdgpu_vm_bo_lookup_mapping for wptr_bo mapping lookup
- Move wptr_bo refcount increment to amdgpu_amdkfd_map_gtt_bo_to_gart
- Remove list_del_init from amdgpu_amdkfd_map_gtt_bo_to_gart
- Cleanup/fix create_queue wptr_bo error handling
Two nit-pick below, with those fixed, this patch is

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
.../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 49 +++++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 37 +++++++++++++-
.../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    | 17 +++++--
 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 +27 -0
index efab923056f4..888d08128a94 100644
@@ -2030,6 +2030,55 @@ int amdgpu_amdkfd_gpuvm_sync_memory(
 	return ret;
 }
 
+/**
+ * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count
+ * @adev: Device to which allocated BO belongs
+ * @bo: Buffer object to be mapped
+ *
+ * Before return, bo reference count is incremented. To release the reference and unpin/
+ * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
+ */
+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);
Quoted 9 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_sync_memory( · line 2065

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +6 -0
+		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);

pinned bo should keep in validate_list as PDB/PTB may move and update, please remove list_del_init here.

Thought I deleted this line - good catch.

Quoted 35 lines
Quoted 35 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +16 -0
+
+	amdgpu_bo_unreserve(bo);
+
+	bo = amdgpu_bo_ref(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 +1 -0
index e9766e165c38..1789ed8b79f5 100644
@@ -289,6 +289,7 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 	struct kfd_process_device *pdd;
 	struct queue_properties q_properties;
 	uint32_t doorbell_offset_in_process = 0;
+	struct amdgpu_bo *wptr_bo = NULL;
 
 	memset(&q_properties, 0, sizeof(struct queue_properties));
 
@@ -316,12 +317,41 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 		goto err_bind_process;
 	}
 
Quoted 9 lines

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, · line 323

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +4 -0
+	/* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
+	 * on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
+	 */
+	if (dev->shared_resources.enable_mes && (dev->adev->mes.sched_version & 0xff) >= 3) {

Should we check ip version for GFX11 only? Because GFX10 set enable_mes, and may set adev->mes.sched_version later as well. Regards,
Philip

I purposefully omitted setting sched/kiq_version for GFX10 in part for this reason. The MES version numbers for GFX11 don't build on top of GFX10, so it could get very confusing if we go back to set this for GFX10 now. With that in mind I think it's okay to base this check on the MES scheduler version without explicitly checking for GFX11+, but if it's preferred the other way, I'm also fine with that.

Best,
Graham

Quoted 181 lines
Quoted 181 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +30 -2
+		struct amdgpu_bo_va_mapping *wptr_mapping;
+		struct amdgpu_vm *wptr_vm;
+
+		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
+		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
+		if (err)
+			goto err_wptr_map_gart;
+
+		wptr_mapping = amdgpu_vm_bo_lookup_mapping(
+				wptr_vm, args->write_pointer_address >> PAGE_SHIFT);
+		amdgpu_bo_unreserve(wptr_vm->root.bo);
+		if (!wptr_mapping) {
+			pr_err("Failed to lookup wptr bo\n");
+			err = -EINVAL;
+			goto err_wptr_map_gart;
+		}
+
+		wptr_bo = wptr_mapping->bo_va->base.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_map_gart;
+		}
+	}
+
 	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;
 
@@ -354,6 +384,9 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 	return 0;
 
 err_create_queue:
+	if (wptr_bo)
+		amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
+err_wptr_map_gart:
 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 +13 -6
index f99e09dc43ea..3a17c1ebc527 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 +222,7 @@ static int init_user_queue(struct process_queue_manager *pqm,
 			goto cleanup;
 		}
 		memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
+		(*q)->wptr_bo = wptr_bo;
 	}
 
 	pr_debug("PQM After init queue");
@@ -237,6 +239,7 @@ 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 +302,7 @@ 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 +323,7 @@ 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 +460,13 @@ 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 +907,7 @@ 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);
--

2.25.1

maintainer
#4 lore raw

Re: [PATCH v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

In reply to Graham Sider

Am 13.06.22 um 17:20 schrieb Graham Sider:

Quoted 60 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.
v3:
- Use amdgpu_vm_bo_lookup_mapping for wptr_bo mapping lookup
- Move wptr_bo refcount increment to amdgpu_amdkfd_map_gtt_bo_to_gart
- Remove list_del_init from amdgpu_amdkfd_map_gtt_bo_to_gart
- Cleanup/fix create_queue wptr_bo error handling

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
  .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 49 +++++++++++++++++++
  drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 37 +++++++++++++-
  .../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    | 17 +++++--
  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 +13 -0
index efab923056f4..888d08128a94 100644
@@ -2030,6 +2030,55 @@ int amdgpu_amdkfd_gpuvm_sync_memory(
  	return ret;
  }
 
+/**
+ * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count
+ * @adev: Device to which allocated BO belongs
+ * @bo: Buffer object to be mapped
+ *
+ * Before return, bo reference count is incremented. To release the reference and unpin/
+ * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
+ */
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo)
+{
+	int ret;
+
+	ret = amdgpu_bo_reserve(bo, true);

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_sync_memory( · line 2055

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +10 -0
+	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;
+	}

Oh! Is that something we do for every MQD? When yes that here is pretty much a NAK.

We can't do this or create a trivial deny of service attack against the kernel driver.

Regards,
Christian.

Quoted 228 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +26 -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);
+
+	amdgpu_bo_unreserve(bo);
+
+	bo = amdgpu_bo_ref(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 +35 -2
index e9766e165c38..1789ed8b79f5 100644
@@ -289,6 +289,7 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
  	struct kfd_process_device *pdd;
  	struct queue_properties q_properties;
  	uint32_t doorbell_offset_in_process = 0;
+	struct amdgpu_bo *wptr_bo = NULL;
 
  	memset(&q_properties, 0, sizeof(struct queue_properties));
 
@@ -316,12 +317,41 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
  		goto err_bind_process;
  	}
 
+	/* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
+	 * on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
+	 */
+	if (dev->shared_resources.enable_mes && (dev->adev->mes.sched_version & 0xff) >= 3) {
+		struct amdgpu_bo_va_mapping *wptr_mapping;
+		struct amdgpu_vm *wptr_vm;
+
+		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
+		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
+		if (err)
+			goto err_wptr_map_gart;
+
+		wptr_mapping = amdgpu_vm_bo_lookup_mapping(
+				wptr_vm, args->write_pointer_address >> PAGE_SHIFT);
+		amdgpu_bo_unreserve(wptr_vm->root.bo);
+		if (!wptr_mapping) {
+			pr_err("Failed to lookup wptr bo\n");
+			err = -EINVAL;
+			goto err_wptr_map_gart;
+		}
+
+		wptr_bo = wptr_mapping->bo_va->base.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_map_gart;
+		}
+	}
+
  	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;
 
@@ -354,6 +384,9 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
  	return 0;
 
  err_create_queue:
+	if (wptr_bo)
+		amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
+err_wptr_map_gart:
  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 +12 -5
index f99e09dc43ea..3a17c1ebc527 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 +222,7 @@ static int init_user_queue(struct process_queue_manager *pqm,
  			goto cleanup;
  		}
  		memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
+		(*q)->wptr_bo = wptr_bo;
  	}
 
  	pr_debug("PQM After init queue");
@@ -237,6 +239,7 @@ 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 +302,7 @@ 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 +323,7 @@ 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 +460,13 @@ 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 +907,7 @@ 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);
#5 lore raw

RE: [PATCH v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

[AMD Official Use Only - General]

Quoted 99 lines

-----Original Message-----
From: Koenig, Christian <[email protected]>
Sent: Wednesday, June 15, 2022 3:29 AM
To: Sider, Graham <[email protected]>; amd- [email protected]
Cc: Joshi, Mukul <[email protected]>; Kuehling, Felix <[email protected]>; Yang, Philip <[email protected]>
Subject: Re: [PATCH v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

Am 13.06.22 um 17:20 schrieb Graham Sider:

Quoted 5 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.

Quoted 8 lines

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

v2:
- Update MES_VERSION check from 2 to 3.
v3:
- Use amdgpu_vm_bo_lookup_mapping for wptr_bo mapping lookup
- Move wptr_bo refcount increment to

amdgpu_amdkfd_map_gtt_bo_to_gart

Quoted 7 lines

- Remove list_del_init from amdgpu_amdkfd_map_gtt_bo_to_gart
- Cleanup/fix create_queue wptr_bo error handling

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
  .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 49

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

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

Quoted 13 lines
  drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 37 +++++++++++++-
  .../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    | 17 +++++--
  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,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int · line 302

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 12 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 +0 -0
index efab923056f4..888d08128a94 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 +2030,55 @@ int amdgpu_amdkfd_gpuvm_sync_memory(
  	return ret;
  }
 
+/**
+ * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and

increment

Quoted 13 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +13 -0
+reference count
+ * @adev: Device to which allocated BO belongs
+ * @bo: Buffer object to be mapped
+ *
+ * Before return, bo reference count is incremented. To release the
+reference and unpin/
+ * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
+ */
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev,
+struct amdgpu_bo *bo) {
+	int ret;
+
+	ret = amdgpu_bo_reserve(bo, true);

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_sync_memory( · line 2055

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +10 -0
+	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;
+	}

Oh! Is that something we do for every MQD? When yes that here is pretty much a NAK.

Quoted 6 lines

We can't do this or create a trivial deny of service attack against the kernel driver.

Regards,
Christian.

Hi Christian, could you elaborate on this? Right now this is only being used to pin the queue wptr BO.

Best,
Graham

Quoted 273 lines
Quoted 17 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +17 -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);
+
+	amdgpu_bo_unreserve(bo);
+
+	bo = amdgpu_bo_ref(bo);
+
+	return 0;
+

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

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/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep,

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
  		struct kgd_mem *mem, void **kptr, uint64_t *size)
  {
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
index e9766e165c38..1789ed8b79f5 100644
@@ -289,6 +289,7 @@ 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 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
  	struct kfd_process_device *pdd;
  	struct queue_properties q_properties;
  	uint32_t doorbell_offset_in_process = 0;
+	struct amdgpu_bo *wptr_bo = NULL;
 
  	memset(&q_properties, 0, sizeof(struct queue_properties));
 
@@ -316,12 +317,41 @@ 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, wptr BOs must be mapped to GART for MES

to determine work

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file · line 321

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
+	 * on unmapped queues for usermode queue oversubscription (no

aggregated doorbell)

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

Quoted 13 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +12 -0
+	 */
+	if (dev->shared_resources.enable_mes && (dev->adev-
 mes.sched_version & 0xff) >= 3) {
+		struct amdgpu_bo_va_mapping *wptr_mapping;
+		struct amdgpu_vm *wptr_vm;
+
+		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
+		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
+		if (err)
+			goto err_wptr_map_gart;
+
+		wptr_mapping = amdgpu_vm_bo_lookup_mapping(
+				wptr_vm, 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 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +9 -0
+		amdgpu_bo_unreserve(wptr_vm->root.bo);
+		if (!wptr_mapping) {
+			pr_err("Failed to lookup wptr bo\n");
+			err = -EINVAL;
+			goto err_wptr_map_gart;
+		}
+
+		wptr_bo = wptr_mapping->bo_va->base.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_map_gart;
+		}
+	}
+
  	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 354

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;
@@ -354,6 +384,9 @@ static int kfd_ioctl_create_queue(struct file *filep,

struct kfd_process *p,

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

Quoted 15 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +3 -0
  	return 0;
 
  err_create_queue:
+	if (wptr_bo)
+		amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
+err_wptr_map_gart:
  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
  	queue_input.doorbell_offset = q->properties.doorbell_off;
  	queue_input.mqd_addr = q->gart_mqd_addr;
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +3 -1
-	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;
+
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..3a17c1ebc527 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 +222,7 @@ 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 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -0
+		(*q)->wptr_bo = wptr_bo;
  	}
 
  	pr_debug("PQM After init queue");
@@ -237,6 +239,7 @@ 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 245

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 +302,7

@@ int

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c

Quoted 6 lines

pqm_create_queue(struct process_queue_manager *pqm,
* allocate_sdma_queue() in create_queue() has the
* corresponding check logic.
*/

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -1
-		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 +323,7 @@ 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 +460,13 @@ 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 +907,7 @@ 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 910

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 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -0
+NULL, q_data, mqd, ctl_stack,
  				NULL);
  	if (ret) {
  		pr_err("Failed to create new queue err:%d\n", ret);
maintainer
#6 lore raw

Re: [PATCH v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

Am 15.06.22 um 15:17 schrieb Sider, Graham:

Quoted 100 lines

[AMD Official Use Only - General]

Quoted 98 lines

-----Original Message-----
From: Koenig, Christian <[email protected]>
Sent: Wednesday, June 15, 2022 3:29 AM
To: Sider, Graham <[email protected]>; amd- [email protected]
Cc: Joshi, Mukul <[email protected]>; Kuehling, Felix <[email protected]>; Yang, Philip <[email protected]>
Subject: Re: [PATCH v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

Am 13.06.22 um 17:20 schrieb Graham Sider:

Quoted 5 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.

Quoted 8 lines

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

v2:
- Update MES_VERSION check from 2 to 3.
v3:
- Use amdgpu_vm_bo_lookup_mapping for wptr_bo mapping lookup
- Move wptr_bo refcount increment to

amdgpu_amdkfd_map_gtt_bo_to_gart

Quoted 7 lines

- Remove list_del_init from amdgpu_amdkfd_map_gtt_bo_to_gart
- Cleanup/fix create_queue wptr_bo error handling

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
   .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 49

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

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

Quoted 13 lines
   drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 37 +++++++++++++-
   .../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    | 17 +++++--
   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,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int · line 302

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 12 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 +0 -0
index efab923056f4..888d08128a94 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 +2030,55 @@ int amdgpu_amdkfd_gpuvm_sync_memory(
   	return ret;
   }
 
+/**
+ * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and

increment

Quoted 13 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +13 -0
+reference count
+ * @adev: Device to which allocated BO belongs
+ * @bo: Buffer object to be mapped
+ *
+ * Before return, bo reference count is incremented. To release the
+reference and unpin/
+ * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
+ */
+int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev,
+struct amdgpu_bo *bo) {
+	int ret;
+
+	ret = amdgpu_bo_reserve(bo, true);

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_sync_memory( · line 2055

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +10 -0
+	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;
+	}

Oh! Is that something we do for every MQD? When yes that here is pretty much a NAK.

Quoted 7 lines
Quoted 6 lines

We can't do this or create a trivial deny of service attack against the kernel driver.

Regards,
Christian.

Hi Christian, could you elaborate on this? Right now this is only being used to pin the queue wptr BO.

Well is this wptr BO per process, per queue or global?

amdgpu_bo_pin() is only allowed if we pin global resources, otherwise I have to reject that.

Regards,
Christian.

Quoted 275 lines

Best,
Graham

Quoted 271 lines
Quoted 17 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +17 -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);
+
+	amdgpu_bo_unreserve(bo);
+
+	bo = amdgpu_bo_ref(bo);
+
+	return 0;
+

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

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/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file *filep,

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
   		struct kgd_mem *mem, void **kptr, uint64_t *size)
   {
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
index e9766e165c38..1789ed8b79f5 100644
@@ -289,6 +289,7 @@ 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 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
   	struct kfd_process_device *pdd;
   	struct queue_properties q_properties;
   	uint32_t doorbell_offset_in_process = 0;
+	struct amdgpu_bo *wptr_bo = NULL;
 
   	memset(&q_properties, 0, sizeof(struct queue_properties));
 
@@ -316,12 +317,41 @@ 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, wptr BOs must be mapped to GART for MES

to determine work

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file · line 321

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
+	 * on unmapped queues for usermode queue oversubscription (no

aggregated doorbell)

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

Quoted 13 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +12 -0
+	 */
+	if (dev->shared_resources.enable_mes && (dev->adev-
 mes.sched_version & 0xff) >= 3) {
+		struct amdgpu_bo_va_mapping *wptr_mapping;
+		struct amdgpu_vm *wptr_vm;
+
+		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
+		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
+		if (err)
+			goto err_wptr_map_gart;
+
+		wptr_mapping = amdgpu_vm_bo_lookup_mapping(
+				wptr_vm, 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 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +9 -0
+		amdgpu_bo_unreserve(wptr_vm->root.bo);
+		if (!wptr_mapping) {
+			pr_err("Failed to lookup wptr bo\n");
+			err = -EINVAL;
+			goto err_wptr_map_gart;
+		}
+
+		wptr_bo = wptr_mapping->bo_va->base.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_map_gart;
+		}
+	}
+
   	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 354

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;
@@ -354,6 +384,9 @@ static int kfd_ioctl_create_queue(struct file *filep,

struct kfd_process *p,

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

Quoted 15 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +3 -0
   	return 0;
 
   err_create_queue:
+	if (wptr_bo)
+		amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
+err_wptr_map_gart:
   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,

	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 +4 -0
+		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; 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 15 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..3a17c1ebc527 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 +222,7 @@ 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 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -0
+		(*q)->wptr_bo = wptr_bo;
   	}
 
   	pr_debug("PQM After init queue");
@@ -237,6 +239,7 @@ 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 245

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 +302,7

@@ int

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c

Quoted 6 lines

pqm_create_queue(struct process_queue_manager *pqm,
* allocate_sdma_queue() in create_queue() has the
* corresponding check logic.
*/

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -1
-		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 +323,7 @@ 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 +460,13 @@ 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 +907,7 @@ 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 910

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 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -0
+NULL, q_data, mqd, ctl_stack,
   				NULL);
   	if (ret) {
   		pr_err("Failed to create new queue err:%d\n", ret);
#7 lore raw

Re: [PATCH v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

On 2022-06-15 10:06, Christian König wrote:
Am 15.06.22 um 15:17 schrieb Sider, Graham:
[AMD Official Use Only - General]

-----Original Message-----
From: Koenig, Christian <[email protected]>
Sent: Wednesday, June 15, 2022 3:29 AM
To: Sider, Graham <[email protected]>; amd-
[email protected]
Cc: Joshi, Mukul <[email protected]>; Kuehling, Felix <[email protected]>; Yang, Philip <[email protected]>
Subject: Re: [PATCH v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

Am 13.06.22 um 17:20 schrieb 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.
v3:
- Use amdgpu_vm_bo_lookup_mapping for wptr_bo mapping lookup
- Move wptr_bo refcount increment to amdgpu_amdkfd_map_gtt_bo_to_gart
- Remove list_del_init from amdgpu_amdkfd_map_gtt_bo_to_gart
- Cleanup/fix create_queue wptr_bo error handling

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 + .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 49
+++++++++++++++++++
   drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 37 +++++++++++++- .../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    | 17 +++++--

   7 files changed, 110 insertions(+), 8 deletions(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index 429b16ba10bf..dba26d1e3be9 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0 -0
@@ -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);

(quoted diff) +0 -0

b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index efab923056f4..888d08128a94 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
@@ -2030,6 +2030,55 @@ int amdgpu_amdkfd_gpuvm_sync_memory(

       return ret;
   }

+/**
+ * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment
+reference count
+ * @adev: Device to which allocated BO belongs
+ * @bo: Buffer object to be mapped
+ *
+ * Before return, bo reference count is incremented. To release the +reference and unpin/
+ * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
+ */
+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;
+    }

Oh! Is that something we do for every MQD? When yes that here is pretty much a NAK.

We can't do this or create a trivial deny of service attack against the kernel driver.

Regards,
Christian.

Hi Christian, could you elaborate on this? Right now this is only being used to pin the queue wptr BO.

Well is this wptr BO per process, per queue or global?

amdgpu_bo_pin() is only allowed if we pin global resources, otherwise I have to reject that.
wptr BO is per queue, allocated as queue structure, 1 page size on system memory.
KFD limit number of queues globally, max_queues = 127; /* HWS limit */, so this will pin max 508KB and take max 127 GART page mapping.
wptr is updated by app and read by HWS, if we don't pin wptr, we have to evict queue when wptr bo is moved on system memory, then update GART mapping and restore queue.
Regards,
Philip

Regards,
Christian.

Best,
Graham

+
+    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);
+
+    bo = amdgpu_bo_ref(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) {

(quoted diff) +0 -0

b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index e9766e165c38..1789ed8b79f5 100644

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
@@ -289,6 +289,7 @@ static int kfd_ioctl_create_queue(struct file *filep,

struct kfd_process *p,
       struct kfd_process_device *pdd;
       struct queue_properties q_properties;
       uint32_t doorbell_offset_in_process = 0;
+    struct amdgpu_bo *wptr_bo = NULL;

       memset(&q_properties, 0, sizeof(struct queue_properties));

@@ -316,12 +317,41 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
           goto err_bind_process;
       }

+    /* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
+     * on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
+     */
+    if (dev->shared_resources.enable_mes && (dev->adev- mes.sched_version & 0xff) >= 3) {
+        struct amdgpu_bo_va_mapping *wptr_mapping;
+        struct amdgpu_vm *wptr_vm;
+
+        wptr_vm = drm_priv_to_vm(pdd->drm_priv);
+        err = amdgpu_bo_reserve(wptr_vm->root.bo, false); +        if (err)
+            goto err_wptr_map_gart;
+
+        wptr_mapping = amdgpu_vm_bo_lookup_mapping(
+                wptr_vm, args->write_pointer_address >> PAGE_SHIFT);
+        amdgpu_bo_unreserve(wptr_vm->root.bo);
+        if (!wptr_mapping) {
+            pr_err("Failed to lookup wptr bo\n");
+            err = -EINVAL;
+            goto err_wptr_map_gart;
+        }
+
+        wptr_bo = wptr_mapping->bo_va->base.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_map_gart;
+        }
+    }
+
       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;

@@ -354,6 +384,9 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
       return 0;

   err_create_queue:
+    if (wptr_bo)
+        amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo); +err_wptr_map_gart:
   err_bind_process:
   err_pdd:
       mutex_unlock(&p->mutex);

(quoted diff) +0 -0

b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index b39d89c52887..d8de2fbdfc7d 100644

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
@@ -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; 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,
       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;

(quoted diff) +0 -0

b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index a5d3963537d7..dcddee0d6f06 100644

drivers/gpu/drm/amd/amdkfd/kfd_priv.h +0 -0
@@ -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, 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..3a17c1ebc527 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)
   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 +222,7 @@ static int init_user_queue(struct process_queue_manager *pqm,
               goto cleanup;
           }
           memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
+        (*q)->wptr_bo = wptr_bo;
       }

       pr_debug("PQM After init queue");
@@ -237,6 +239,7 @@ 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 +302,7 @@ 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 +323,7 @@ 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 +460,13 @@ 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 +907,7 @@ 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);

Acked-by
#8 lore raw

Re: [PATCH v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

Quoted 129 lines · philip yang

On 2022-06-15 10:06, Christian König wrote:

Am 15.06.22 um 15:17 schrieb Sider, Graham:

[AMD Official Use Only - General]

-----Original Message-----
From: Koenig, Christian <[email protected]>
Sent: Wednesday, June 15, 2022 3:29 AM
To: Sider, Graham <[email protected]>; amd- [email protected]
Cc: Joshi, Mukul <[email protected]>; Kuehling, Felix <[email protected]>; Yang, Philip <[email protected]>
Subject: Re: [PATCH v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

Am 13.06.22 um 17:20 schrieb 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.
v3:
- Use amdgpu_vm_bo_lookup_mapping for wptr_bo mapping lookup
- Move wptr_bo refcount increment to

amdgpu_amdkfd_map_gtt_bo_to_gart

- Remove list_del_init from amdgpu_amdkfd_map_gtt_bo_to_gart
- Cleanup/fix create_queue wptr_bo error handling

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
   .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 49

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

   drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 37 +++++++++++++-
   .../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    | 17 +++++--
   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,

           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);
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +2 -0
+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 +2 -0
index efab923056f4..888d08128a94 100644
@@ -2030,6 +2030,55 @@ int amdgpu_amdkfd_gpuvm_sync_memory(
       return ret;
   }
 
+/**
+ * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and
 

increment

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +23 -0
+reference count
+ * @adev: Device to which allocated BO belongs
+ * @bo: Buffer object to be mapped
+ *
+ * Before return, bo reference count is incremented. To release the
+reference and unpin/
+ * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
+ */
+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;
+    }

Oh! Is that something we do for every MQD? When yes that here is pretty much a NAK.

We can't do this or create a trivial deny of service attack against the kernel driver.

Regards,
Christian.

Hi Christian, could you elaborate on this? Right now this is only being used to pin the queue wptr BO.

Well is this wptr BO per process, per queue or global?

Quoted 8 lines

amdgpu_bo_pin() is only allowed if we pin global resources, otherwise I have to reject that.

wptr BO is per queue, allocated as queue structure, 1 page size on system memory.

KFD limit number of queues globally, max_queues = 127; /* HWS limit */, so this will pin max 508KB and take max 127 GART page mapping.

wptr is updated by app and read by HWS, if we don't pin wptr, we have to evict queue when wptr bo is moved on system memory, then update GART mapping and restore queue.

I talked to Graham about this a bit off line. This is a requirement due to changes in the hardware and the way KFD doorbells work. The fact that we are limited by the max queues supported by the hardware scheduler prevents this from getting out of hand. We would need this space whether the queues were kernel managed or user managed.

Quoted 341 lines

Regards,

Philip

Regards,
Christian.

Best,
Graham

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +26 -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);
+
+    amdgpu_bo_unreserve(bo);
+
+    bo = amdgpu_bo_ref(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 +1 -0
index e9766e165c38..1789ed8b79f5 100644
@@ -289,6 +289,7 @@ static int kfd_ioctl_create_queue(struct file *filep,
 
 struct kfd_process *p,
 
       struct kfd_process_device *pdd;
       struct queue_properties q_properties;
       uint32_t doorbell_offset_in_process = 0;
+    struct amdgpu_bo *wptr_bo = NULL;
       memset(&q_properties, 0, sizeof(struct queue_properties));
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
@@ -316,12 +317,41 @@ static int kfd_ioctl_create_queue(struct file
 
 *filep, struct kfd_process *p,
 
           goto err_bind_process;
       }
 
+    /* Starting with GFX11, wptr BOs must be mapped to GART for MES
 

to determine work

+ * on unmapped queues for usermode queue oversubscription (no

aggregated doorbell)

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2 -0
+     */
+    if (dev->shared_resources.enable_mes && (dev->adev-

mes.sched_version & 0xff) >= 3) {

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +10 -0
+        struct amdgpu_bo_va_mapping *wptr_mapping;
+        struct amdgpu_vm *wptr_vm;
+
+        wptr_vm = drm_priv_to_vm(pdd->drm_priv);
+        err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
+        if (err)
+            goto err_wptr_map_gart;
+
+        wptr_mapping = amdgpu_vm_bo_lookup_mapping(
+                wptr_vm, args->write_pointer_address >>

PAGE_SHIFT);

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +9 -0
+        amdgpu_bo_unreserve(wptr_vm->root.bo);
+        if (!wptr_mapping) {
+            pr_err("Failed to lookup wptr bo\n");
+            err = -EINVAL;
+            goto err_wptr_map_gart;
+        }
+
+        wptr_bo = wptr_mapping->bo_va->base.bo;
+        err = amdgpu_amdkfd_map_gtt_bo_to_gart(dev->adev,

wptr_bo);

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_map_gart;
+        }
+    }
+
       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 +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 +4 -0
+            NULL, NULL, NULL, &doorbell_offset_in_process);
       if (err != 0)
           goto err_create_queue;
@@ -354,6 +384,9 @@ static int kfd_ioctl_create_queue(struct file *filep,
 
 struct kfd_process *p,
 
       return 0;
 
   err_create_queue:
+    if (wptr_bo)
+        amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
+err_wptr_map_gart:
   err_bind_process:
   err_pdd:
       mutex_unlock(&p->mutex);
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -0
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;
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +3 -1
@@ -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;

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -0
+    } 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; 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,
 
       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);

drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +2 -0
+    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, 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..3a17c1ebc527 100644

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +2 -1
@@ -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 +0 -0
@@ -221,6 +222,7 @@ static int init_user_queue(struct
 
 process_queue_manager *pqm,
 
               goto cleanup;
           }
           memset((*q)->gang_ctx_cpu_ptr, 0,

AMDGPU_MES_GANG_CTX_SIZE);

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +2 -0
+        (*q)->wptr_bo = wptr_bo;
       }
 
       pr_debug("PQM After init queue");
@@ -237,6 +239,7 @@ 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 +302,7

@@ int

pqm_create_queue(struct process_queue_manager *pqm,
* allocate_sdma_queue() in create_queue() has the
* corresponding check logic.
*/

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -1
-        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 +2 -1
+*qid);
           if (retval != 0)
               goto err_create_queue;
           pqn->q = q;
@@ -320,7 +323,7 @@ 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,

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +4 -1
+*qid);
           if (retval != 0)
               goto err_create_queue;
           pqn->q = q;
@@ -457,9 +460,13 @@ 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);

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +3 -1
+
+        }
           uninit_queue(pqn->q);
       }
@@ -900,7 +907,7 @@ 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);
#9 lore raw

Re: [PATCH v3 2/3] drm/amdkfd: Enable GFX11 usermode queue oversubscription

In reply to philip yang

Hi Philip,

Am 16.06.22 um 00:41 schrieb philip yang:

Quoted 16 lines

[SNIP]

Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Quoted 5 lines

+    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;
+    }

Oh! Is that something we do for every MQD? When yes that here is pretty much a NAK.

We can't do this or create a trivial deny of service attack against the kernel driver.

Regards,
Christian.

Quoted 8 lines
Quoted 5 lines
Quoted 1 line

Hi Christian, could you elaborate on this? Right now this is only being used to pin the queue wptr BO.

Well is this wptr BO per process, per queue or global?

amdgpu_bo_pin() is only allowed if we pin global resources, otherwise I have to reject that.

wptr BO is per queue, allocated as queue structure, 1 page size on system memory.

Yeah, I was hoping for this explanation as well. My status was still that the WPTR and RPTR are part of the ring buffer.

We should add a check that we really only pin a buffer with 1 page size here, then that should be ok.

Regards,
Christian.

Quoted 290 lines

KFD limit number of queues globally, max_queues = 127; /* HWS limit */, so this will pin max 508KB and take max 127 GART page mapping.

wptr is updated by app and read by HWS, if we don't pin wptr, we have to evict queue when wptr bo is moved on system memory, then update GART mapping and restore queue.

Regards,

Philip

Quoted 282 lines

Regards,
Christian.

Quoted 277 lines

Best,
Graham

Quoted 273 lines
Quoted 26 lines
(quoted diff) +26 -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);
+
+    amdgpu_bo_unreserve(bo);
+
+    bo = amdgpu_bo_ref(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;
+}
+
Quoted 1 line

   int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct

amdgpu_device *adev,

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file · line 289

Quoted 9 lines

           struct kgd_mem *mem, void **kptr, uint64_t *size) {

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
index e9766e165c38..1789ed8b79f5 100644
@@ -289,6 +289,7 @@ 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 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
        struct kfd_process_device *pdd;
        struct queue_properties q_properties;
        uint32_t doorbell_offset_in_process = 0;
+    struct amdgpu_bo *wptr_bo = NULL;
 
        memset(&q_properties, 0, sizeof(struct queue_properties));
@@ -316,12 +317,41 @@ 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 · line 320

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

to determine work

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file · line 321

Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
+     * on unmapped queues for usermode queue oversubscription (no

aggregated doorbell)

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file · line 334

Quoted 13 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +12 -0
+     */
+    if (dev->shared_resources.enable_mes && (dev->adev-
 mes.sched_version & 0xff) >= 3) {
+        struct amdgpu_bo_va_mapping *wptr_mapping;
+        struct amdgpu_vm *wptr_vm;
+
+        wptr_vm = drm_priv_to_vm(pdd->drm_priv);
+        err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
+        if (err)
+            goto err_wptr_map_gart;
+
+        wptr_mapping = amdgpu_vm_bo_lookup_mapping(
+                wptr_vm, args->write_pointer_address >>

PAGE_SHIFT);

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_create_queue(struct file · line 343

Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +9 -0
+ amdgpu_bo_unreserve(wptr_vm->root.bo);
+        if (!wptr_mapping) {
+            pr_err("Failed to lookup wptr bo\n");
+            err = -EINVAL;
+            goto err_wptr_map_gart;
+        }
+
+        wptr_bo = wptr_mapping->bo_va->base.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 · line 324

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_map_gart;
+        }
+    }
+
        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 354

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 · line 384

Quoted 6 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;
@@ -354,6 +384,9 @@ static int kfd_ioctl_create_queue(struct file 
 *filep,

struct kfd_process *p,

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

Quoted 15 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
        return 0;
 

   err_create_queue:
+    if (wptr_bo)
+        amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo); +err_wptr_map_gart:
   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,
    AMDGPU_MES_PRIORITY_LEVEL_NORMAL;

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int add_queue_mes(struct · line 232

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 +4 -0
+        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; 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,

Quoted 1 line

       m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >>

8);

Quoted 10 lines

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 15 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..3a17c1ebc527 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 +222,7 @@ 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 · line 224

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 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -0
+        (*q)->wptr_bo = wptr_bo;
        }
 
        pr_debug("PQM After init queue");
@@ -237,6 +239,7 @@ int pqm_create_queue(struct

process_queue_manager *pqm,

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 +302,7

@@ int

Quoted 6 lines

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,

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

Quoted 5 lines

+*qid);
           if (retval != 0)
               goto err_create_queue;
           pqn->q = q;

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +0 -0
@@ -320,7 +323,7 @@ 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 326

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;

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +0 -0
@@ -457,9 +460,13 @@ 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 +907,7 @@ 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 910

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 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -0
+NULL, q_data, mqd, ctl_stack,
                    NULL);
        if (ret) {

           pr_err("Failed to create new queue err:%d\n", ret);