Skip to contentSkip to conversation

drm/amdkfd: add helper function for kfd sysfs create

PATCH In review Philip Yang 3 participants 8 messages Jun 22, 2021 – Jun 23, 2021

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/4 drm/amdkfd: add helper function for kfd sysfs create

No functionality change. Modify kfd_sysfs_create_file to use kobject as parameter, so it becomes common helper function to remove duplicate code and will simplify new kfd sysfs file create in future.

Move pr_warn to helper function if sysfs file create failed. Set helper function as void return because caller doesn't use the helper function return value.

---
 drivers/gpu/drm/amd/amdkfd/kfd_process.c | 119 ++++++++---------------
 1 file changed, 39 insertions(+), 80 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_process.c +39 -80
index 09b98a83f670..3147dc8bb051 100644
@@ -484,34 +484,31 @@ int kfd_procfs_add_queue(struct queue *q)
 	return 0;
 }
 
-static int kfd_sysfs_create_file(struct kfd_process *p, struct attribute *attr,
+static void kfd_sysfs_create_file(struct kobject *kobj, struct attribute *attr,
 				 char *name)
 {
-	int ret = 0;
+	int ret;
 
-	if (!p || !attr || !name)
-		return -EINVAL;
+	if (!kobj || !attr || !name)
+		return;
 
 	attr->name = name;
 	attr->mode = KFD_SYSFS_FILE_MODE;
 	sysfs_attr_init(attr);
 
-	ret = sysfs_create_file(p->kobj, attr);
-
-	return ret;
+	ret = sysfs_create_file(kobj, attr);
+	if (ret)
+		pr_warn("Create sysfs %s/%s failed %d", kobj->name, name, ret);
 }
 
-static int kfd_procfs_add_sysfs_stats(struct kfd_process *p)
+static void kfd_procfs_add_sysfs_stats(struct kfd_process *p)
 {
-	int ret = 0;
+	int ret;
 	int i;
 	char stats_dir_filename[MAX_SYSFS_FILENAME_LEN];
 
-	if (!p)
-		return -EINVAL;
-
-	if (!p->kobj)
-		return -EFAULT;
+	if (!p || !p->kobj)
+		return;
 
 	/*
 	 * Create sysfs files for each GPU:
@@ -521,63 +518,43 @@ static int kfd_procfs_add_sysfs_stats(struct kfd_process *p)
 	 */
 	for (i = 0; i < p->n_pdds; i++) {
 		struct kfd_process_device *pdd = p->pdds[i];
-		struct kobject *kobj_stats;
 
 		snprintf(stats_dir_filename, MAX_SYSFS_FILENAME_LEN,
 				"stats_%u", pdd->dev->id);
-		kobj_stats = kfd_alloc_struct(kobj_stats);
-		if (!kobj_stats)
-			return -ENOMEM;
+		pdd->kobj_stats = kfd_alloc_struct(pdd->kobj_stats);
+		if (!pdd->kobj_stats)
+			return;
 
-		ret = kobject_init_and_add(kobj_stats,
-						&procfs_stats_type,
-						p->kobj,
-						stats_dir_filename);
+		ret = kobject_init_and_add(pdd->kobj_stats,
+					   &procfs_stats_type,
+					   p->kobj,
+					   stats_dir_filename);
 
 		if (ret) {
 			pr_warn("Creating KFD proc/stats_%s folder failed",
-					stats_dir_filename);
-			kobject_put(kobj_stats);
-			goto err;
+				stats_dir_filename);
+			kobject_put(pdd->kobj_stats);
+			pdd->kobj_stats = NULL;
+			return;
 		}
 
-		pdd->kobj_stats = kobj_stats;
-		pdd->attr_evict.name = "evicted_ms";
-		pdd->attr_evict.mode = KFD_SYSFS_FILE_MODE;
-		sysfs_attr_init(&pdd->attr_evict);
-		ret = sysfs_create_file(kobj_stats, &pdd->attr_evict);
-		if (ret)
-			pr_warn("Creating eviction stats for gpuid %d failed",
-					(int)pdd->dev->id);
-
+		kfd_sysfs_create_file(pdd->kobj_stats, &pdd->attr_evict,
+				      "evicted_ms");
 		/* Add sysfs file to report compute unit occupancy */
-		if (pdd->dev->kfd2kgd->get_cu_occupancy != NULL) {
-			pdd->attr_cu_occupancy.name = "cu_occupancy";
-			pdd->attr_cu_occupancy.mode = KFD_SYSFS_FILE_MODE;
-			sysfs_attr_init(&pdd->attr_cu_occupancy);
-			ret = sysfs_create_file(kobj_stats,
-						&pdd->attr_cu_occupancy);
-			if (ret)
-				pr_warn("Creating %s failed for gpuid: %d",
-					pdd->attr_cu_occupancy.name,
-					(int)pdd->dev->id);
-		}
+		if (pdd->dev->kfd2kgd->get_cu_occupancy)
+			kfd_sysfs_create_file(pdd->kobj_stats,
+					      &pdd->attr_cu_occupancy,
+					      "cu_occupancy");
 	}
-err:
-	return ret;
 }
 
 
-static int kfd_procfs_add_sysfs_files(struct kfd_process *p)
+static void kfd_procfs_add_sysfs_files(struct kfd_process *p)
 {
-	int ret = 0;
 	int i;
 
-	if (!p)
-		return -EINVAL;
-
-	if (!p->kobj)
-		return -EFAULT;
+	if (!p || !p->kobj)
+		return;
 
 	/*
 	 * Create sysfs files for each GPU:
@@ -589,20 +566,14 @@ static int kfd_procfs_add_sysfs_files(struct kfd_process *p)
 
 		snprintf(pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, "vram_%u",
 			 pdd->dev->id);
-		ret = kfd_sysfs_create_file(p, &pdd->attr_vram, pdd->vram_filename);
-		if (ret)
-			pr_warn("Creating vram usage for gpu id %d failed",
-				(int)pdd->dev->id);
+		kfd_sysfs_create_file(p->kobj, &pdd->attr_vram,
+				      pdd->vram_filename);
 
 		snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, "sdma_%u",
 			 pdd->dev->id);
-		ret = kfd_sysfs_create_file(p, &pdd->attr_sdma, pdd->sdma_filename);
-		if (ret)
-			pr_warn("Creating sdma usage for gpu id %d failed",
-				(int)pdd->dev->id);
+		kfd_sysfs_create_file(p->kobj, &pdd->attr_sdma,
+					    pdd->sdma_filename);
 	}
-
-	return ret;
 }
 
 void kfd_procfs_del_queue(struct queue *q)
@@ -800,28 +771,16 @@ struct kfd_process *kfd_create_process(struct file *filep)
 			goto out;
 		}
 
-		process->attr_pasid.name = "pasid";
-		process->attr_pasid.mode = KFD_SYSFS_FILE_MODE;
-		sysfs_attr_init(&process->attr_pasid);
-		ret = sysfs_create_file(process->kobj, &process->attr_pasid);
-		if (ret)
-			pr_warn("Creating pasid for pid %d failed",
-					(int)process->lead_thread->pid);
+		kfd_sysfs_create_file(process->kobj, &process->attr_pasid,
+				      "pasid");
 
 		process->kobj_queues = kobject_create_and_add("queues",
 							process->kobj);
 		if (!process->kobj_queues)
 			pr_warn("Creating KFD proc/queues folder failed");
 
-		ret = kfd_procfs_add_sysfs_stats(process);
-		if (ret)
-			pr_warn("Creating sysfs stats dir for pid %d failed",
-				(int)process->lead_thread->pid);
-
-		ret = kfd_procfs_add_sysfs_files(process);
-		if (ret)
-			pr_warn("Creating sysfs usage file for pid %d failed",
-				(int)process->lead_thread->pid);
+		kfd_procfs_add_sysfs_stats(process);
+		kfd_procfs_add_sysfs_files(process);
 	}
 out:
 	if (!IS_ERR(process))
Signature · 6 lines
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#1 lore raw

PATCH 2/4 drm/amdkfd: fix sysfs kobj leak

3 cases of kobj leak, which causes memory leak:

kobj_type must have release() method to free memory from release callback. Don't need NULL default_attrs to init kobj.

sysfs files created under kobj_status should be removed with kobj_status as parent kobject.

Remove queue sysfs files when releasing queue from process MMU notifier release callback.

---
 drivers/gpu/drm/amd/amdkfd/kfd_process.c           | 14 ++++++--------
 .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c |  1 +
 2 files changed, 7 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_process.c +6 -8
index 3147dc8bb051..cfc36fceac8a 100644
@@ -451,13 +451,9 @@ static const struct sysfs_ops procfs_stats_ops = {
 	.show = kfd_procfs_stats_show,
 };
 
-static struct attribute *procfs_stats_attrs[] = {
-	NULL
-};
-
 static struct kobj_type procfs_stats_type = {
 	.sysfs_ops = &procfs_stats_ops,
-	.default_attrs = procfs_stats_attrs,
+	.release = kfd_procfs_kobj_release,
 };
 
 int kfd_procfs_add_queue(struct queue *q)
@@ -946,9 +942,11 @@ static void kfd_process_wq_release(struct work_struct *work)
 
 			sysfs_remove_file(p->kobj, &pdd->attr_vram);
 			sysfs_remove_file(p->kobj, &pdd->attr_sdma);
-			sysfs_remove_file(p->kobj, &pdd->attr_evict);
-			if (pdd->dev->kfd2kgd->get_cu_occupancy != NULL)
-				sysfs_remove_file(p->kobj, &pdd->attr_cu_occupancy);
+
+			sysfs_remove_file(pdd->kobj_stats, &pdd->attr_evict);
+			if (pdd->dev->kfd2kgd->get_cu_occupancy)
+				sysfs_remove_file(pdd->kobj_stats,
+						  &pdd->attr_cu_occupancy);
 			kobject_del(pdd->kobj_stats);
 			kobject_put(pdd->kobj_stats);
 			pdd->kobj_stats = NULL;
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -0
index 95a6c36cea4c..243dd1efcdbf 100644
@@ -153,6 +153,7 @@ void pqm_uninit(struct process_queue_manager *pqm)
 		if (pqn->q && pqn->q->gws)
 			amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info,
 				pqn->q->gws);
+		kfd_procfs_del_queue(pqn->q);
 		uninit_queue(pqn->q);
 		list_del(&pqn->process_queue_list);
 		kfree(pqn);
Signature · 6 lines
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#2 lore raw

PATCH 3/4 drm/amdkfd: add sysfs counters for vm fault and migration

In reply to Philip Yang

This is part of SVM profiling API, export sysfs counters for per-process, per-GPU vm retry fault, pages migrated in and out of GPU vram.

counters will not be updated in parallel in GPU retry fault handler and migration to vram/ram path, use READ_ONCE to avoid compiler optimization.

---
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h    |   9 ++
 drivers/gpu/drm/amd/amdkfd/kfd_process.c | 151 ++++++++++++++++++-----
 2 files changed, 131 insertions(+), 29 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +9 -0
index 6dc22fa1e555..3426743ed228 100644
@@ -730,6 +730,15 @@ struct kfd_process_device {
 	 *  number of CU's a device has along with number of other competing processes
 	 */
 	struct attribute attr_cu_occupancy;
+
+	/* sysfs counters for GPU retry fault and page migration tracking */
+	struct kobject *kobj_counters;
+	struct attribute attr_faults;
+	struct attribute attr_page_in;
+	struct attribute attr_page_out;
+	uint64_t faults;
+	uint64_t page_in;
+	uint64_t page_out;
 };
 
 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
drivers/gpu/drm/amd/amdkfd/kfd_process.c +122 -29

190 lines not on this page · raw message

Signature · 6 lines
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#3 lore raw

PATCH 4/4 drm/amdkfd: implement counters for vm fault and migration

In reply to Philip Yang

Add helper function to get process device data structure from adev to update counters.

Update vm faults, page_in, page_out counters will no be executed in parallel, use WRITE_ONCE to avoid any form of compiler optimizations.

---
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 14 ++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c     | 24 ++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_svm.h     |  2 ++
 3 files changed, 40 insertions(+)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +14 -0
index fd8f544f0de2..45b5349283af 100644
@@ -413,6 +413,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 			uint64_t end)
 {
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
+	struct kfd_process_device *pdd;
 	struct dma_fence *mfence = NULL;
 	struct migrate_vma migrate;
 	dma_addr_t *scratch;
@@ -473,6 +474,12 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 out_free:
 	kvfree(buf);
 out:
+	if (!r) {
+		pdd = svm_range_get_pdd_by_adev(prange, adev);
+		if (pdd)
+			WRITE_ONCE(pdd->page_in, pdd->page_in + migrate.cpages);
+	}
+
 	return r;
 }
 
@@ -629,6 +636,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 		       struct vm_area_struct *vma, uint64_t start, uint64_t end)
 {
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
+	struct kfd_process_device *pdd;
 	struct dma_fence *mfence = NULL;
 	struct migrate_vma migrate;
 	dma_addr_t *scratch;
@@ -678,6 +686,12 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 out_free:
 	kvfree(buf);
 out:
+	if (!r) {
+		pdd = svm_range_get_pdd_by_adev(prange, adev);
+		if (pdd)
+			WRITE_ONCE(pdd->page_out,
+				   pdd->page_out + migrate.cpages);
+	}
 	return r;
 }
 
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +24 -0
index 5468ea4264c6..f3323328f01f 100644
@@ -564,6 +564,24 @@ svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id)
 	return (struct amdgpu_device *)pdd->dev->kgd;
 }
 
+struct kfd_process_device *
+svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev)
+{
+	struct kfd_process *p;
+	int32_t gpu_idx, gpuid;
+	int r;
+
+	p = container_of(prange->svms, struct kfd_process, svms);
+
+	r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpu_idx);
+	if (r) {
+		pr_debug("failed to get device id by adev %p\n", adev);
+		return NULL;
+	}
+
+	return kfd_process_device_from_gpuidx(p, gpu_idx);
+}
+
 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
 {
 	struct ttm_operation_ctx ctx = { false, false };
@@ -2315,6 +2333,7 @@ int
 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
 			uint64_t addr)
 {
+	struct kfd_process_device *pdd;
 	struct mm_struct *mm = NULL;
 	struct svm_range_list *svms;
 	struct svm_range *prange;
@@ -2440,6 +2459,11 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
 out_unlock_svms:
 	mutex_unlock(&svms->lock);
 	mmap_read_unlock(mm);
+
+	pdd = svm_range_get_pdd_by_adev(prange, adev);
+	if (pdd)
+		WRITE_ONCE(pdd->faults, pdd->faults + 1);
+
 	mmput(mm);
 out:
 	kfd_unref_process(p);
drivers/gpu/drm/amd/amdkfd/kfd_svm.h +2 -0
index 0c0fc399395e..a9af03994d1a 100644
@@ -174,6 +174,8 @@ void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
 			 unsigned long offset, unsigned long npages);
 void svm_range_free_dma_mappings(struct svm_range *prange);
 void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm);
+struct kfd_process_device *
+svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev);
 
 /* SVM API and HMM page migration work together, device memory type
  * is initialized to not 0 when page migration register device memory.
Signature · 6 lines
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
maintainer Reviewed-by on v1 4/4
#4 lore raw

Re: [PATCH 4/4] drm/amdkfd: implement counters for vm fault and migration

Am 2021-06-22 um 9:32 a.m. schrieb Philip Yang:

Quoted 97 lines

Add helper function to get process device data structure from adev to update counters.

Update vm faults, page_in, page_out counters will no be executed in parallel, use WRITE_ONCE to avoid any form of compiler optimizations.

---
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 14 ++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c     | 24 ++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_svm.h     |  2 ++
 3 files changed, 40 insertions(+)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +14 -0
index fd8f544f0de2..45b5349283af 100644
@@ -413,6 +413,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 			uint64_t end)
 {
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
+	struct kfd_process_device *pdd;
 	struct dma_fence *mfence = NULL;
 	struct migrate_vma migrate;
 	dma_addr_t *scratch;
@@ -473,6 +474,12 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 out_free:
 	kvfree(buf);
 out:
+	if (!r) {
+		pdd = svm_range_get_pdd_by_adev(prange, adev);
+		if (pdd)
+			WRITE_ONCE(pdd->page_in, pdd->page_in + migrate.cpages);
+	}
+
 	return r;
 }
 
@@ -629,6 +636,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 		       struct vm_area_struct *vma, uint64_t start, uint64_t end)
 {
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
+	struct kfd_process_device *pdd;
 	struct dma_fence *mfence = NULL;
 	struct migrate_vma migrate;
 	dma_addr_t *scratch;
@@ -678,6 +686,12 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 out_free:
 	kvfree(buf);
 out:
+	if (!r) {
+		pdd = svm_range_get_pdd_by_adev(prange, adev);
+		if (pdd)
+			WRITE_ONCE(pdd->page_out,
+				   pdd->page_out + migrate.cpages);
+	}
 	return r;
 }
 
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +19 -0
index 5468ea4264c6..f3323328f01f 100644
@@ -564,6 +564,24 @@ svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id)
 	return (struct amdgpu_device *)pdd->dev->kgd;
 }
 
+struct kfd_process_device *
+svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev)
+{
+	struct kfd_process *p;
+	int32_t gpu_idx, gpuid;
+	int r;
+
+	p = container_of(prange->svms, struct kfd_process, svms);
+
+	r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpu_idx);
+	if (r) {
+		pr_debug("failed to get device id by adev %p\n", adev);
+		return NULL;
+	}
+
+	return kfd_process_device_from_gpuidx(p, gpu_idx);
+}
+
 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
 {
 	struct ttm_operation_ctx ctx = { false, false };
@@ -2315,6 +2333,7 @@ int
 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
 			uint64_t addr)
 {
+	struct kfd_process_device *pdd;
 	struct mm_struct *mm = NULL;
 	struct svm_range_list *svms;
 	struct svm_range *prange;

drivers/gpu/drm/amd/amdkfd/kfd_svm.c svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, · line 2463

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +2 -0
@@ -2440,6 +2459,11 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
 out_unlock_svms:
 	mutex_unlock(&svms->lock);
 	mmap_read_unlock(mm);
+
+	pdd = svm_range_get_pdd_by_adev(prange, adev);

svm_range_get_pdd_by_adev needs to do a linear search. You don't need this here because you already know the gpuidx. I think you can just call kfd_process_device_from_gpuidx(p, gpu_idx) here.

With that fixed, the series is

P.S.: Thanks for catching and fixing those memory leaks in patch 2.

Quoted 11 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3 -0
+	if (pdd)
+		WRITE_ONCE(pdd->faults, pdd->faults + 1);
+
 	mmput(mm);
 out:
 	kfd_unref_process(p);
drivers/gpu/drm/amd/amdkfd/kfd_svm.h +0 -0
index 0c0fc399395e..a9af03994d1a 100644

drivers/gpu/drm/amd/amdkfd/kfd_svm.h void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr, · line 181

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.h +2 -0
@@ -174,6 +174,8 @@ void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
 			 unsigned long offset, unsigned long npages);
 void svm_range_free_dma_mappings(struct svm_range *prange);
 void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm);
+struct kfd_process_device *
+svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev);
 
 /* SVM API and HMM page migration work together, device memory type
  * is initialized to not 0 when page migration register device memory.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#6 lore raw

PATCH v2 4/4 drm/amdkfd: implement counters for vm fault and migration

In reply to Philip Yang

Add helper function to get process device data structure from adev to update counters.

Update vm faults, page_in, page_out counters will no be executed in parallel, use WRITE_ONCE to avoid any form of compiler optimizations.

---
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 14 ++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c     | 45 +++++++++++++++++++++++-
 drivers/gpu/drm/amd/amdkfd/kfd_svm.h     |  2 ++
 3 files changed, 60 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +14 -0
index fd8f544f0de2..45b5349283af 100644
@@ -413,6 +413,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 			uint64_t end)
 {
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
+	struct kfd_process_device *pdd;
 	struct dma_fence *mfence = NULL;
 	struct migrate_vma migrate;
 	dma_addr_t *scratch;
@@ -473,6 +474,12 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 out_free:
 	kvfree(buf);
 out:
+	if (!r) {
+		pdd = svm_range_get_pdd_by_adev(prange, adev);
+		if (pdd)
+			WRITE_ONCE(pdd->page_in, pdd->page_in + migrate.cpages);
+	}
+
 	return r;
 }
 
@@ -629,6 +636,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 		       struct vm_area_struct *vma, uint64_t start, uint64_t end)
 {
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
+	struct kfd_process_device *pdd;
 	struct dma_fence *mfence = NULL;
 	struct migrate_vma migrate;
 	dma_addr_t *scratch;
@@ -678,6 +686,12 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 out_free:
 	kvfree(buf);
 out:
+	if (!r) {
+		pdd = svm_range_get_pdd_by_adev(prange, adev);
+		if (pdd)
+			WRITE_ONCE(pdd->page_out,
+				   pdd->page_out + migrate.cpages);
+	}
 	return r;
 }
 
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +44 -1

69 lines not on this page · raw message

drivers/gpu/drm/amd/amdkfd/kfd_svm.h +2 -0
index 0c0fc399395e..a9af03994d1a 100644
@@ -174,6 +174,8 @@ void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
 			 unsigned long offset, unsigned long npages);
 void svm_range_free_dma_mappings(struct svm_range *prange);
 void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm);
+struct kfd_process_device *
+svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev);
 
 /* SVM API and HMM page migration work together, device memory type
  * is initialized to not 0 when page migration register device memory.
Signature · 6 lines
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
maintainer Reviewed-by on v2 4/4
#7 lore raw

Re: [PATCH v2 4/4] drm/amdkfd: implement counters for vm fault and migration

Quoted 7 lines · Philip Yang

Add helper function to get process device data structure from adev to update counters.

Update vm faults, page_in, page_out counters will no be executed in parallel, use WRITE_ONCE to avoid any form of compiler optimizations.

Quoted 134 lines
---
  drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 14 ++++++++
  drivers/gpu/drm/amd/amdkfd/kfd_svm.c     | 45 +++++++++++++++++++++++-
  drivers/gpu/drm/amd/amdkfd/kfd_svm.h     |  2 ++
  3 files changed, 60 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +14 -0
index fd8f544f0de2..45b5349283af 100644
@@ -413,6 +413,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
  			uint64_t end)
  {
  	uint64_t npages = (end - start) >> PAGE_SHIFT;
+	struct kfd_process_device *pdd;
  	struct dma_fence *mfence = NULL;
  	struct migrate_vma migrate;
  	dma_addr_t *scratch;
@@ -473,6 +474,12 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
  out_free:
  	kvfree(buf);
  out:
+	if (!r) {
+		pdd = svm_range_get_pdd_by_adev(prange, adev);
+		if (pdd)
+			WRITE_ONCE(pdd->page_in, pdd->page_in + migrate.cpages);
+	}
+
  	return r;
  }
 
@@ -629,6 +636,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  		       struct vm_area_struct *vma, uint64_t start, uint64_t end)
  {
  	uint64_t npages = (end - start) >> PAGE_SHIFT;
+	struct kfd_process_device *pdd;
  	struct dma_fence *mfence = NULL;
  	struct migrate_vma migrate;
  	dma_addr_t *scratch;
@@ -678,6 +686,12 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  out_free:
  	kvfree(buf);
  out:
+	if (!r) {
+		pdd = svm_range_get_pdd_by_adev(prange, adev);
+		if (pdd)
+			WRITE_ONCE(pdd->page_out,
+				   pdd->page_out + migrate.cpages);
+	}
  	return r;
  }
 
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +44 -1
index 7f5ebc00b349..00b46846a5e0 100644
@@ -564,6 +564,24 @@ svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id)
  	return (struct amdgpu_device *)pdd->dev->kgd;
  }
 
+struct kfd_process_device *
+svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev)
+{
+	struct kfd_process *p;
+	int32_t gpu_idx, gpuid;
+	int r;
+
+	p = container_of(prange->svms, struct kfd_process, svms);
+
+	r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpu_idx);
+	if (r) {
+		pr_debug("failed to get device id by adev %p\n", adev);
+		return NULL;
+	}
+
+	return kfd_process_device_from_gpuidx(p, gpu_idx);
+}
+
  static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
  {
  	struct ttm_operation_ctx ctx = { false, false };
@@ -2311,6 +2329,27 @@ static bool svm_range_skip_recover(struct svm_range *prange)
  	return false;
  }
 
+static void
+svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p,
+		      struct svm_range *prange, int32_t gpuidx)
+{
+	struct kfd_process_device *pdd;
+
+	if (gpuidx == MAX_GPU_INSTANCE)
+		/* fault is on different page of same range
+		 * or fault is skipped to recover later
+		 */
+		pdd = svm_range_get_pdd_by_adev(prange, adev);
+	else
+		/* fault recovered
+		 * or fault cannot recover because GPU no access on the range
+		 */
+		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
+
+	if (pdd)
+		WRITE_ONCE(pdd->faults, pdd->faults + 1);
+}
+
  int
  svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
  			uint64_t addr)
@@ -2320,7 +2359,8 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
  	struct svm_range *prange;
  	struct kfd_process *p;
  	uint64_t timestamp;
-	int32_t best_loc, gpuidx;
+	int32_t best_loc;
+	int32_t gpuidx = MAX_GPU_INSTANCE;
  	bool write_locked = false;
  	int r = 0;
 
@@ -2440,6 +2480,9 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
  out_unlock_svms:
  	mutex_unlock(&svms->lock);
  	mmap_read_unlock(mm);
+
+	svm_range_count_fault(adev, p, prange, gpuidx);
+
  	mmput(mm);
  out:
  	kfd_unref_process(p);
drivers/gpu/drm/amd/amdkfd/kfd_svm.h +0 -0
index 0c0fc399395e..a9af03994d1a 100644

drivers/gpu/drm/amd/amdkfd/kfd_svm.h void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr, · line 181

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.h +2 -0
@@ -174,6 +174,8 @@ void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
  			 unsigned long offset, unsigned long npages);
  void svm_range_free_dma_mappings(struct svm_range *prange);
  void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm);
+struct kfd_process_device *
+svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev);
 
  /* SVM API and HMM page migration work together, device memory type
   * is initialized to not 0 when page migration register device memory.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx