Skip to contentSkip to conversation

drm/amdgpu: Fetch MES scheduler/KIQ versions

PATCH Superseded Graham Sider 2 participants 6 messages Jun 7, 2022 – Jun 8, 2022

Subsystems (3)

v1 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/4 drm/amdgpu: Fetch MES scheduler/KIQ versions

Store MES scheduler and MES KIQ version numbers in amdgpu_mes.

---
 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 2/4 drm/amdkfd: Pass MES/RS64 information to sysfs

Make MES/RS64 CP enablement and MES scheduler/MES KIQ versions available through sysfs.

---
 drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 8 ++++++++
 1 file changed, 8 insertions(+)
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +8 -0
index 02efae4f5549..51c8a285baaf 100644
@@ -571,6 +571,14 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
 				      dev->gpu->sdma_fw_version);
 		sysfs_show_64bit_prop(buffer, offs, "unique_id",
 				      dev->gpu->adev->unique_id);
+		sysfs_show_32bit_prop(buffer, offs, "mes_enabled",
+				      (uint32_t)dev->gpu->adev->enable_mes);
+		sysfs_show_32bit_prop(buffer, offs, "rs64_cp_enabled",
+				      (uint32_t)dev->gpu->adev->gfx.rs64_enable);
+		sysfs_show_32bit_prop(buffer, offs, "mes_sched_version",
+				      dev->gpu->adev->mes.sched_version);
+		sysfs_show_32bit_prop(buffer, offs, "mes_kiq_version",
+				      dev->gpu->adev->mes.kiq_version);
 
 	}
 
Signature · 1 line
2.25.1
maintainer
#2 lore raw

Re: [PATCH 2/4] drm/amdkfd: Pass MES/RS64 information to sysfs

Quoted 4 lines · Graham Sider

Make MES/RS64 CP enablement and MES scheduler/MES KIQ versions available through sysfs.

Why do we need to expose this to user mode applications? They don't interact directly with the scheduler or the KIQ. I cannot think of any reasonable conditions in usermode that would need the scheduler or KIQ version number. It may make sense in debugfs, but not here.

Regards,
  Felix

Quoted 23 lines
---
  drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 8 ++++++++
  1 file changed, 8 insertions(+)
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +8 -0
index 02efae4f5549..51c8a285baaf 100644
@@ -571,6 +571,14 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
  				      dev->gpu->sdma_fw_version);
  		sysfs_show_64bit_prop(buffer, offs, "unique_id",
  				      dev->gpu->adev->unique_id);
+		sysfs_show_32bit_prop(buffer, offs, "mes_enabled",
+				      (uint32_t)dev->gpu->adev->enable_mes);
+		sysfs_show_32bit_prop(buffer, offs, "rs64_cp_enabled",
+				      (uint32_t)dev->gpu->adev->gfx.rs64_enable);
+		sysfs_show_32bit_prop(buffer, offs, "mes_sched_version",
+				      dev->gpu->adev->mes.sched_version);
+		sysfs_show_32bit_prop(buffer, offs, "mes_kiq_version",
+				      dev->gpu->adev->mes.kiq_version);
 
  	}
 
#3 lore raw

RE: [PATCH 2/4] drm/amdkfd: Pass MES/RS64 information to sysfs

[AMD Official Use Only - General]

I was originally using the MES scheduler version in Thunk to check whether to use userptr or non-paged memory for the queue allocation. Just today though I've changed this to always use non-paged so this isn't needed anymore. Maybe instead we can think about adding this to debugfs, I agree.

Thanks,
Graham

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Wednesday, June 8, 2022 3:31 PM
To: Sider, Graham <[email protected]>; [email protected]
Cc: Yang, Philip <[email protected]>; Joshi, Mukul <[email protected]>
Subject: Re: [PATCH 2/4] drm/amdkfd: Pass MES/RS64 information to sysfs

Quoted 4 lines · Graham Sider

Make MES/RS64 CP enablement and MES scheduler/MES KIQ versions available through sysfs.

Why do we need to expose this to user mode applications? They don't interact directly with the scheduler or the KIQ. I cannot think of any reasonable conditions in usermode that would need the scheduler or KIQ version number. It may make sense in debugfs, but not here.

Regards,
  Felix

Quoted 24 lines
---
  drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 8 ++++++++
  1 file changed, 8 insertions(+)
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +8 -0
index 02efae4f5549..51c8a285baaf 100644
@@ -571,6 +571,14 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
  				      dev->gpu->sdma_fw_version);
  		sysfs_show_64bit_prop(buffer, offs, "unique_id",
  				      dev->gpu->adev->unique_id);
+		sysfs_show_32bit_prop(buffer, offs, "mes_enabled",
+				      (uint32_t)dev->gpu->adev->enable_mes);
+		sysfs_show_32bit_prop(buffer, offs, "rs64_cp_enabled",
+				      (uint32_t)dev->gpu->adev->gfx.rs64_enable);
+		sysfs_show_32bit_prop(buffer, offs, "mes_sched_version",
+				      dev->gpu->adev->mes.sched_version);
+		sysfs_show_32bit_prop(buffer, offs, "mes_kiq_version",
+				      dev->gpu->adev->mes.kiq_version);
 
  	}
 
#4 lore raw

PATCH 3/4 drm/amdkfd: Enable GFX11 usermode queue oversubscription

In reply to 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 >= 2.

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

68 lines not on this page · raw message

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

PATCH 4/4 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