Skip to contentSkip to conversation

drm/amdkfd: Page aligned VRAM reserve size

PATCH In review Philip Yang 3 participants 11 messages Jan 10, 2023

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

PATCH 1/1 drm/amdkfd: Page aligned VRAM reserve size

Use page aligned size to reserve VRAM usage because page aligned TTM BO size is used to unreserve VRAM usage, otherwise this cause vram_used accounting unbalanced.

Change vram_used definition type to int64_t to be able to trigger WARN_ONCE(adev && adev->kfd.vram_used < 0, "..."), to help debug the accouting issue with warning and backtrace.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       | 2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
index fb41869e357a..333780491867 100644
@@ -97,7 +97,7 @@ struct amdgpu_amdkfd_fence {
 
 struct amdgpu_kfd_dev {
 	struct kfd_dev *dev;
-	uint64_t vram_used;
+	int64_t vram_used;
 	uint64_t vram_used_aligned;
 	bool init_complete;
 	struct work_struct reset_work;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -1
index 2a118669d0e3..f23d94e57762 100644
@@ -151,7 +151,7 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 		 * to avoid fragmentation caused by 4K allocations in the tail
 		 * 2M BO chunk.
 		 */
-		vram_needed = size;
+		vram_needed = PAGE_ALIGN(size);
 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
 		system_mem_needed = size;
 	} else if (!(alloc_flag &
Signature · 1 line
2.35.1
maintainer
#1 lore raw

Am 2023-01-09 um 19:01 schrieb Philip Yang:

Quoted 32 lines

Use page aligned size to reserve VRAM usage because page aligned TTM BO size is used to unreserve VRAM usage, otherwise this cause vram_used accounting unbalanced.

Change vram_used definition type to int64_t to be able to trigger WARN_ONCE(adev && adev->kfd.vram_used < 0, "..."), to help debug the accouting issue with warning and backtrace.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       | 2 +-
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 2 +-
  2 files changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
index fb41869e357a..333780491867 100644
@@ -97,7 +97,7 @@ struct amdgpu_amdkfd_fence {
 
  struct amdgpu_kfd_dev {
  	struct kfd_dev *dev;
-	uint64_t vram_used;
+	int64_t vram_used;
  	uint64_t vram_used_aligned;
  	bool init_complete;
  	struct work_struct reset_work;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
index 2a118669d0e3..f23d94e57762 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, · line 154

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -1
@@ -151,7 +151,7 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
  		 * to avoid fragmentation caused by 4K allocations in the tail
  		 * 2M BO chunk.
  		 */
-		vram_needed = size;
+		vram_needed = PAGE_ALIGN(size);

This only solves part of the problem. size is used in other places in this function that should all use the page-aligned size. I think we should do the page-alignment at a much higher level, in kfd_ioctl_alloc_memory_of_gpu. That way all the kernel code can safely assume that buffer sizes are page aligned, and we avoid future surprises.

Regards,
  Felix

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
  	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
  		system_mem_needed = size;
  	} else if (!(alloc_flag &
#2 lore raw
Quoted 32 lines · Felix Kuehling

Am 2023-01-09 um 19:01 schrieb Philip Yang:

Quoted 31 lines

Use page aligned size to reserve VRAM usage because page aligned TTM BO size is used to unreserve VRAM usage, otherwise this cause vram_used accounting unbalanced.

Change vram_used definition type to int64_t to be able to trigger WARN_ONCE(adev && adev->kfd.vram_used < 0, "..."), to help debug the accouting issue with warning and backtrace.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 2 +-

  2 files changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
index fb41869e357a..333780491867 100644
@@ -97,7 +97,7 @@ struct amdgpu_amdkfd_fence {
     struct amdgpu_kfd_dev {
       struct kfd_dev *dev;
-    uint64_t vram_used;
+    int64_t vram_used;
       uint64_t vram_used_aligned;
       bool init_complete;
       struct work_struct reset_work;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
index 2a118669d0e3..f23d94e57762 100644
@@ -151,7 +151,7 @@ int amdgpu_amdkfd_reserve_mem_limit(struct 
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, · line 154

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -1
 amdgpu_device *adev,
            * to avoid fragmentation caused by 4K allocations in the tail
            * 2M BO chunk.
            */
-        vram_needed = size;
+        vram_needed = PAGE_ALIGN(size);

This only solves part of the problem. size is used in other places in this function that should all use the page-aligned size. I think we should do the page-alignment at a much higher level, in kfd_ioctl_alloc_memory_of_gpu. That way all the kernel code can safely assume that buffer sizes are page aligned, and we avoid future surprises.

yes, the error handling unreserve should use aligned_size too, and size is also used as number of pages in amdgpu_bo_create for DOMAIN_GWS etc, we can not pass aligned size at higher level, I will send v2 patch for review.

Regards,

Philip

Quoted 8 lines

Regards,
  Felix

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
       } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
           system_mem_needed = size;

      } else if (!(alloc_flag &

#3 lore raw

PATCH 1/1 drm/amdkfd: Page aligned memory reserve size

In reply to Philip Yang

Use page aligned size to reserve memory usage because page aligned TTM BO size is used to unreserve memory usage, otherwise no page aligned size causes memory usage accounting unbalanced.

Change vram_used definition type to int64_t to be able to trigger WARN_ONCE(adev && adev->kfd.vram_used < 0, "..."), to help debug the accouting issue with warning and backtrace.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       | 2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 9 ++++++---
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         | 7 +++++--
 3 files changed, 12 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
index fb41869e357a..333780491867 100644
@@ -97,7 +97,7 @@ struct amdgpu_amdkfd_fence {
 
 struct amdgpu_kfd_dev {
 	struct kfd_dev *dev;
-	uint64_t vram_used;
+	int64_t vram_used;
 	uint64_t vram_used_aligned;
 	bool init_complete;
 	struct work_struct reset_work;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +6 -3
index 2a118669d0e3..7efee672bc41 100644
@@ -1598,6 +1598,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	struct amdgpu_bo *bo;
 	struct drm_gem_object *gobj = NULL;
 	u32 domain, alloc_domain;
+	uint64_t aligned_size;
 	u64 alloc_flags;
 	int ret;
 
@@ -1653,13 +1654,15 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	 * the memory.
 	 */
 	if ((*mem)->aql_queue)
-		size = size >> 1;
+		size >>= 1;
+
+	aligned_size = PAGE_ALIGN(size);
 
 	(*mem)->alloc_flags = flags;
 
 	amdgpu_sync_create(&(*mem)->sync);
 
-	ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, flags);
+	ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags);
 	if (ret) {
 		pr_debug("Insufficient memory\n");
 		goto err_reserve_limit;
@@ -1725,7 +1728,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	/* Don't unreserve system mem limit twice */
 	goto err_reserve_limit;
 err_bo_create:
-	amdgpu_amdkfd_unreserve_mem_limit(adev, size, flags);
+	amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags);
 err_reserve_limit:
 	mutex_destroy(&(*mem)->lock);
 	if (gobj)
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +5 -2
index 6d291aa6386b..e11451100a20 100644
@@ -1127,8 +1127,11 @@ static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
 	}
 
 	/* Update the VRAM usage count */
-	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
-		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + args->size);
+	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
+		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
+			args->size >>= 1;
+		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(args->size));
+	}
 
 	mutex_unlock(&p->mutex);
 
Signature · 1 line
2.35.1
maintainer
#4 lore raw

Re: [PATCH] drm/amdkfd: Page aligned memory reserve size

Am 2023-01-10 um 12:11 schrieb Philip Yang:

Quoted 41 lines

Use page aligned size to reserve memory usage because page aligned TTM BO size is used to unreserve memory usage, otherwise no page aligned size causes memory usage accounting unbalanced.

Change vram_used definition type to int64_t to be able to trigger WARN_ONCE(adev && adev->kfd.vram_used < 0, "..."), to help debug the accouting issue with warning and backtrace.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       | 2 +-
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 9 ++++++---
  drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         | 7 +++++--
  3 files changed, 12 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
index fb41869e357a..333780491867 100644
@@ -97,7 +97,7 @@ struct amdgpu_amdkfd_fence {
 
  struct amdgpu_kfd_dev {
  	struct kfd_dev *dev;
-	uint64_t vram_used;
+	int64_t vram_used;
  	uint64_t vram_used_aligned;
  	bool init_complete;
  	struct work_struct reset_work;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -0
index 2a118669d0e3..7efee672bc41 100644
@@ -1598,6 +1598,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
  	struct amdgpu_bo *bo;
  	struct drm_gem_object *gobj = NULL;
  	u32 domain, alloc_domain;
+	uint64_t aligned_size;
  	u64 alloc_flags;
  	int ret;
 

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( · line 1659

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +3 -1
@@ -1653,13 +1654,15 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
  	 * the memory.
  	 */
  	if ((*mem)->aql_queue)
-		size = size >> 1;
+		size >>= 1;
+
+	aligned_size = PAGE_ALIGN(size);

Why do you need a new variable for this? Can't you just update size to be page-aligned here? Is the unaligned size still needed anywhere?

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +2 -2
  	(*mem)->alloc_flags = flags;
 
  	amdgpu_sync_create(&(*mem)->sync);
 
-	ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, flags);
+	ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags);
  	if (ret) {
  		pr_debug("Insufficient memory\n");
  		goto err_reserve_limit;
@@ -1725,7 +1728,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
  	/* Don't unreserve system mem limit twice */
  	goto err_reserve_limit;
  err_bo_create:
-	amdgpu_amdkfd_unreserve_mem_limit(adev, size, flags);
+	amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags);
  err_reserve_limit:
  	mutex_destroy(&(*mem)->lock);
  	if (gobj)
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
index 6d291aa6386b..e11451100a20 100644

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep, · line 1132

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +3 -2
@@ -1127,8 +1127,11 @@ static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
  	}
 
  	/* Update the VRAM usage count */
-	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
-		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + args->size);
+	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
+		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
+			args->size >>= 1;

This will return the updated size to user mode. That's probably not what you want. It may be harmless, but technically it breaks the ABI. It would be better to use a local variable for the updated size.

Regards,
  Felix

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2 -0
+		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(args->size));
+	}
 
  	mutex_unlock(&p->mutex);
 
#5 lore raw

Re: [PATCH] drm/amdkfd: Page aligned memory reserve size

Quoted 41 lines · Felix Kuehling

Am 2023-01-10 um 12:11 schrieb Philip Yang:

Quoted 40 lines

Use page aligned size to reserve memory usage because page aligned TTM BO size is used to unreserve memory usage, otherwise no page aligned size causes memory usage accounting unbalanced.

Change vram_used definition type to int64_t to be able to trigger WARN_ONCE(adev && adev->kfd.vram_used < 0, "..."), to help debug the accouting issue with warning and backtrace.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 9 ++++++--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         | 7 +++++--

  3 files changed, 12 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
index fb41869e357a..333780491867 100644
@@ -97,7 +97,7 @@ struct amdgpu_amdkfd_fence {
     struct amdgpu_kfd_dev {
       struct kfd_dev *dev;
-    uint64_t vram_used;
+    int64_t vram_used;
       uint64_t vram_used_aligned;
       bool init_complete;
       struct work_struct reset_work;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -0
index 2a118669d0e3..7efee672bc41 100644
@@ -1598,6 +1598,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
       struct amdgpu_bo *bo;
       struct drm_gem_object *gobj = NULL;
       u32 domain, alloc_domain;
+    uint64_t aligned_size;
       u64 alloc_flags;
       int ret;

  @@ -1653,13 +1654,15 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
       * the memory.

Quoted 8 lines
Quoted 6 lines

       */
      if ((*mem)->aql_queue)
-        size = size >> 1;
+        size >>= 1;
+
+    aligned_size = PAGE_ALIGN(size);

Why do you need a new variable for this? Can't you just update size to be page-aligned here? Is the unaligned size still needed anywhere?

amdgpu_gem_object_create ->...-> amdgpu_bo_create needs the original size for domain GWS etc, as the size is used as number of pages, not bytes.

Quoted 25 lines
Quoted 25 lines

        (*mem)->alloc_flags = flags; amdgpu_sync_create(&(*mem)->sync);
  -    ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, flags); +    ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags); if (ret) {
          pr_debug("Insufficient memory\n");
          goto err_reserve_limit;

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
@@ -1725,7 +1728,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
       /* Don't unreserve system mem limit twice */
       goto err_reserve_limit;

  err_bo_create:
-    amdgpu_amdkfd_unreserve_mem_limit(adev, size, flags); +    amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags); err_reserve_limit:
      mutex_destroy(&(*mem)->lock);
      if (gobj)

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
index 6d291aa6386b..e11451100a20 100644
@@ -1127,8 +1127,11 @@ static int 
 kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
       }
Quoted 8 lines

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep, · line 1132

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +3 -2
         /* Update the VRAM usage count */
-    if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
-        WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + args->size);
+    if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
+        if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
+            args->size >>= 1;

This will return the updated size to user mode. That's probably not what you want. It may be harmless, but technically it breaks the ABI. It would be better to use a local variable for the updated size.

Submit v3 patch to fix this.

Thanks,

Philip

Quoted 9 lines

Regards,
  Felix

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2 -0
+        WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + 
 PAGE_ALIGN(args->size));
+    }
         mutex_unlock(&p->mutex);
maintainer
#6 lore raw

Re: [PATCH] drm/amdkfd: Page aligned memory reserve size

Am 2023-01-10 um 15:44 schrieb Philip Yang:

Quoted 51 lines · Felix Kuehling
Quoted 41 lines

Am 2023-01-10 um 12:11 schrieb Philip Yang:

Quoted 40 lines

Use page aligned size to reserve memory usage because page aligned TTM BO size is used to unreserve memory usage, otherwise no page aligned size causes memory usage accounting unbalanced.

Change vram_used definition type to int64_t to be able to trigger WARN_ONCE(adev && adev->kfd.vram_used < 0, "..."), to help debug the accouting issue with warning and backtrace.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 9 ++++++--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         | 7 +++++--

  3 files changed, 12 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
index fb41869e357a..333780491867 100644
@@ -97,7 +97,7 @@ struct amdgpu_amdkfd_fence {
     struct amdgpu_kfd_dev {
       struct kfd_dev *dev;
-    uint64_t vram_used;
+    int64_t vram_used;
       uint64_t vram_used_aligned;
       bool init_complete;
       struct work_struct reset_work;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -0
index 2a118669d0e3..7efee672bc41 100644
@@ -1598,6 +1598,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
       struct amdgpu_bo *bo;
       struct drm_gem_object *gobj = NULL;
       u32 domain, alloc_domain;
+    uint64_t aligned_size;
       u64 alloc_flags;
       int ret;

  @@ -1653,13 +1654,15 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
       * the memory.

Quoted 8 lines
Quoted 6 lines

       */
      if ((*mem)->aql_queue)
-        size = size >> 1;
+        size >>= 1;
+
+    aligned_size = PAGE_ALIGN(size);

Why do you need a new variable for this? Can't you just update size to be page-aligned here? Is the unaligned size still needed anywhere?

Quoted 1 line

amdgpu_gem_object_create ->...-> amdgpu_bo_create needs the original size for domain GWS etc, as the size is used as number of pages, not bytes.

I don't think GWS is ever allocated through this code path. This type of memory is not exposed in the KFD ioctl API. KFD allocates a GWS BO using amdgpu_amdkfd_alloc_gws.

Regards,
  Felix

Quoted 49 lines
Quoted 25 lines
Quoted 25 lines

        (*mem)->alloc_flags = flags; amdgpu_sync_create(&(*mem)->sync);
  -    ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, flags); +    ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags); if (ret) {
          pr_debug("Insufficient memory\n");
          goto err_reserve_limit;

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
@@ -1725,7 +1728,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
       /* Don't unreserve system mem limit twice */
       goto err_reserve_limit;

  err_bo_create:
-    amdgpu_amdkfd_unreserve_mem_limit(adev, size, flags); +    amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags); err_reserve_limit:
      mutex_destroy(&(*mem)->lock);
      if (gobj)

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
index 6d291aa6386b..e11451100a20 100644
@@ -1127,8 +1127,11 @@ static int 
 kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
       }
Quoted 8 lines

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep, · line 1132

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +3 -2
         /* Update the VRAM usage count */
-    if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
-        WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + args->size);
+    if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
+        if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
+            args->size >>= 1;

This will return the updated size to user mode. That's probably not what you want. It may be harmless, but technically it breaks the ABI. It would be better to use a local variable for the updated size.

Submit v3 patch to fix this.

Thanks,

Philip

Quoted 9 lines

Regards,
  Felix

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2 -0
+        WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + 
 PAGE_ALIGN(args->size));
+    }
         mutex_unlock(&p->mutex);
#7 lore raw

Re: [PATCH] drm/amdkfd: Page aligned memory reserve size

Quoted 52 lines · Felix Kuehling

Am 2023-01-10 um 15:44 schrieb Philip Yang:

Quoted 51 lines · Felix Kuehling
Quoted 49 lines

Am 2023-01-10 um 12:11 schrieb Philip Yang:

Quoted 38 lines

Use page aligned size to reserve memory usage because page aligned TTM BO size is used to unreserve memory usage, otherwise no page aligned size causes memory usage accounting unbalanced.

Change vram_used definition type to int64_t to be able to trigger WARN_ONCE(adev && adev->kfd.vram_used < 0, "..."), to help debug the accouting issue with warning and backtrace.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 9 ++++++--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         | 7 +++++--

  3 files changed, 12 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
index fb41869e357a..333780491867 100644
@@ -97,7 +97,7 @@ struct amdgpu_amdkfd_fence {
     struct amdgpu_kfd_dev {
       struct kfd_dev *dev;
-    uint64_t vram_used;
+    int64_t vram_used;
       uint64_t vram_used_aligned;
       bool init_complete;
       struct work_struct reset_work;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -0
index 2a118669d0e3..7efee672bc41 100644
@@ -1598,6 +1598,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
       struct amdgpu_bo *bo;
       struct drm_gem_object *gobj = NULL;
       u32 domain, alloc_domain;
+    uint64_t aligned_size;
       u64 alloc_flags;
       int ret;
Quoted 8 lines

  @@ -1653,13 +1654,15 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
       * the memory.
       */
      if ((*mem)->aql_queue)
-        size = size >> 1;
+        size >>= 1;
+
+    aligned_size = PAGE_ALIGN(size);

Why do you need a new variable for this? Can't you just update size to be page-aligned here? Is the unaligned size still needed anywhere?

Quoted 3 lines
Quoted 1 line

amdgpu_gem_object_create ->...-> amdgpu_bo_create needs the original size for domain GWS etc, as the size is used as number of pages, not bytes.

I don't think GWS is ever allocated through this code path. This type of memory is not exposed in the KFD ioctl API. KFD allocates a GWS BO using amdgpu_amdkfd_alloc_gws.

yes, as size is used in pr_debug, it is better to show the original size in debug log to match application, also notice the pr_debug size should base on aql_queue flag, will send v4 patch with the fix.

Regards,

Philip

Quoted 54 lines

Regards,
  Felix

Quoted 49 lines
Quoted 25 lines
Quoted 25 lines

        (*mem)->alloc_flags = flags; amdgpu_sync_create(&(*mem)->sync);
  -    ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, flags); +    ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags); if (ret) {
          pr_debug("Insufficient memory\n");
          goto err_reserve_limit;

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
@@ -1725,7 +1728,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
       /* Don't unreserve system mem limit twice */
       goto err_reserve_limit;

  err_bo_create:
-    amdgpu_amdkfd_unreserve_mem_limit(adev, size, flags); +    amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags); err_reserve_limit:
      mutex_destroy(&(*mem)->lock);
      if (gobj)

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
index 6d291aa6386b..e11451100a20 100644
@@ -1127,8 +1127,11 @@ static int 
 kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
       }
Quoted 8 lines

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep, · line 1132

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +3 -2
         /* Update the VRAM usage count */
-    if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
-        WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + args->size);
+    if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
+        if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
+            args->size >>= 1;

This will return the updated size to user mode. That's probably not what you want. It may be harmless, but technically it breaks the ABI. It would be better to use a local variable for the updated size.

Submit v3 patch to fix this.

Thanks,

Philip

Quoted 9 lines

Regards,
  Felix

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2 -0
+ WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + 
 PAGE_ALIGN(args->size));
+    }
         mutex_unlock(&p->mutex);
#8 lore raw

PATCH v3 1/1 drm/amdkfd: Page aligned memory reserve size

In reply to Philip Yang

Use page aligned size to reserve memory usage because page aligned TTM BO size is used to unreserve memory usage, otherwise no page aligned size causes memory usage accounting unbalanced.

Change vram_used definition type to int64_t to be able to trigger WARN_ONCE(adev && adev->kfd.vram_used < 0, "..."), to help debug the accounting issue with warning and backtrace.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       | 2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 9 ++++++---
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         | 9 +++++++--
 3 files changed, 14 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
index fb41869e357a..333780491867 100644
@@ -97,7 +97,7 @@ struct amdgpu_amdkfd_fence {
 
 struct amdgpu_kfd_dev {
 	struct kfd_dev *dev;
-	uint64_t vram_used;
+	int64_t vram_used;
 	uint64_t vram_used_aligned;
 	bool init_complete;
 	struct work_struct reset_work;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +6 -3
index cd5de5e08d2f..08e2c1dc92d5 100644
@@ -1598,6 +1598,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	struct amdgpu_bo *bo;
 	struct drm_gem_object *gobj = NULL;
 	u32 domain, alloc_domain;
+	uint64_t aligned_size;
 	u64 alloc_flags;
 	int ret;
 
@@ -1653,13 +1654,15 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	 * the memory.
 	 */
 	if ((*mem)->aql_queue)
-		size = size >> 1;
+		size >>= 1;
+
+	aligned_size = PAGE_ALIGN(size);
 
 	(*mem)->alloc_flags = flags;
 
 	amdgpu_sync_create(&(*mem)->sync);
 
-	ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, flags);
+	ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags);
 	if (ret) {
 		pr_debug("Insufficient memory\n");
 		goto err_reserve_limit;
@@ -1725,7 +1728,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	/* Don't unreserve system mem limit twice */
 	goto err_reserve_limit;
 err_bo_create:
-	amdgpu_amdkfd_unreserve_mem_limit(adev, size, flags);
+	amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags);
 err_reserve_limit:
 	mutex_destroy(&(*mem)->lock);
 	if (gobj)
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +7 -2
index 6d291aa6386b..f79b8e964140 100644
@@ -1127,8 +1127,13 @@ static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
 	}
 
 	/* Update the VRAM usage count */
-	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
-		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + args->size);
+	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
+		uint64_t size = args->size;
+
+		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
+			size >>= 1;
+		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
+	}
 
 	mutex_unlock(&p->mutex);
 
Signature · 1 line
2.35.1
#9 lore raw

PATCH v4 1/1 drm/amdkfd: Page aligned memory reserve size

In reply to Philip Yang

Use page aligned size to reserve memory usage because page aligned TTM BO size is used to unreserve memory usage, otherwise no page aligned size causes memory usage accounting unbalanced.

Change vram_used definition type to int64_t to be able to trigger WARN_ONCE(adev && adev->kfd.vram_used < 0, "..."), to help debug the accounting issue with warning and backtrace.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 12 +++++++-----
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         |  9 +++++++--
 3 files changed, 15 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
index fb41869e357a..333780491867 100644
@@ -97,7 +97,7 @@ struct amdgpu_amdkfd_fence {
 
 struct amdgpu_kfd_dev {
 	struct kfd_dev *dev;
-	uint64_t vram_used;
+	int64_t vram_used;
 	uint64_t vram_used_aligned;
 	bool init_complete;
 	struct work_struct reset_work;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +7 -5
index cd5de5e08d2f..6f236ded5f12 100644
@@ -1598,6 +1598,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	struct amdgpu_bo *bo;
 	struct drm_gem_object *gobj = NULL;
 	u32 domain, alloc_domain;
+	uint64_t aligned_size;
 	u64 alloc_flags;
 	int ret;
 
@@ -1653,22 +1654,23 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	 * the memory.
 	 */
 	if ((*mem)->aql_queue)
-		size = size >> 1;
+		size >>= 1;
+	aligned_size = PAGE_ALIGN(size);
 
 	(*mem)->alloc_flags = flags;
 
 	amdgpu_sync_create(&(*mem)->sync);
 
-	ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, flags);
+	ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags);
 	if (ret) {
 		pr_debug("Insufficient memory\n");
 		goto err_reserve_limit;
 	}
 
 	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n",
-			va, size, domain_string(alloc_domain));
+			va, (*mem)->aql_queue ? size << 1 : size, domain_string(alloc_domain));
 
-	ret = amdgpu_gem_object_create(adev, size, 1, alloc_domain, alloc_flags,
+	ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags,
 				       bo_type, NULL, &gobj);
 	if (ret) {
 		pr_debug("Failed to create BO on domain %s. ret %d\n",
@@ -1725,7 +1727,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	/* Don't unreserve system mem limit twice */
 	goto err_reserve_limit;
 err_bo_create:
-	amdgpu_amdkfd_unreserve_mem_limit(adev, size, flags);
+	amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags);
 err_reserve_limit:
 	mutex_destroy(&(*mem)->lock);
 	if (gobj)
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +7 -2
index 6d291aa6386b..f79b8e964140 100644
@@ -1127,8 +1127,13 @@ static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
 	}
 
 	/* Update the VRAM usage count */
-	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
-		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + args->size);
+	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
+		uint64_t size = args->size;
+
+		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
+			size >>= 1;
+		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
+	}
 
 	mutex_unlock(&p->mutex);
 
Signature · 1 line
2.35.1
maintainer Reviewed-by on v4 1/1
#10 lore raw

Re: [PATCH v4] drm/amdkfd: Page aligned memory reserve size

Am 2023-01-10 um 17:41 schrieb Philip Yang:

Quoted 9 lines

Use page aligned size to reserve memory usage because page aligned TTM BO size is used to unreserve memory usage, otherwise no page aligned size causes memory usage accounting unbalanced.

Change vram_used definition type to int64_t to be able to trigger WARN_ONCE(adev && adev->kfd.vram_used < 0, "..."), to help debug the accounting issue with warning and backtrace.

Quoted 88 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       |  2 +-
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 12 +++++++-----
  drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         |  9 +++++++--
  3 files changed, 15 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
index fb41869e357a..333780491867 100644
@@ -97,7 +97,7 @@ struct amdgpu_amdkfd_fence {
 
  struct amdgpu_kfd_dev {
  	struct kfd_dev *dev;
-	uint64_t vram_used;
+	int64_t vram_used;
  	uint64_t vram_used_aligned;
  	bool init_complete;
  	struct work_struct reset_work;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +7 -5
index cd5de5e08d2f..6f236ded5f12 100644
@@ -1598,6 +1598,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
  	struct amdgpu_bo *bo;
  	struct drm_gem_object *gobj = NULL;
  	u32 domain, alloc_domain;
+	uint64_t aligned_size;
  	u64 alloc_flags;
  	int ret;
 
@@ -1653,22 +1654,23 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
  	 * the memory.
  	 */
  	if ((*mem)->aql_queue)
-		size = size >> 1;
+		size >>= 1;
+	aligned_size = PAGE_ALIGN(size);
 
  	(*mem)->alloc_flags = flags;
 
  	amdgpu_sync_create(&(*mem)->sync);
 
-	ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, flags);
+	ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags);
  	if (ret) {
  		pr_debug("Insufficient memory\n");
  		goto err_reserve_limit;
  	}
 
  	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n",
-			va, size, domain_string(alloc_domain));
+			va, (*mem)->aql_queue ? size << 1 : size, domain_string(alloc_domain));
 
-	ret = amdgpu_gem_object_create(adev, size, 1, alloc_domain, alloc_flags,
+	ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags,
  				       bo_type, NULL, &gobj);
  	if (ret) {
  		pr_debug("Failed to create BO on domain %s. ret %d\n",
@@ -1725,7 +1727,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
  	/* Don't unreserve system mem limit twice */
  	goto err_reserve_limit;
  err_bo_create:
-	amdgpu_amdkfd_unreserve_mem_limit(adev, size, flags);
+	amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags);
  err_reserve_limit:
  	mutex_destroy(&(*mem)->lock);
  	if (gobj)
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +7 -2
index 6d291aa6386b..f79b8e964140 100644
@@ -1127,8 +1127,13 @@ static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
  	}
 
  	/* Update the VRAM usage count */
-	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
-		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + args->size);
+	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
+		uint64_t size = args->size;
+
+		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
+			size >>= 1;
+		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
+	}
 
  	mutex_unlock(&p->mutex);