[PATCH 1/2] drm/amdkfd: add a new flag to manage where VRAM allocations go

[PATCH 1/2] drm/amdkfd: add a new flag to manage where VRAM allocations go

[Alex Deucher]

On big and small APUs we send KFD VRAM allocations to GTT
since the carve out is either non-existent or relatively
small.  However, if someone sets the carve out size to be
relatively large, we may end up using GTT rather than VRAM.

No change of logic with this patch, but it allows the
driver to determine which logic to use based on the
carve out size in the future.

Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h              |  5 +++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c       |  4 ++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 16 ++++++++--------
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c          |  3 +++
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c         |  2 +-
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c             |  4 ++--
 drivers/gpu/drm/amd/amdkfd/kfd_svm.h             |  2 +-
 7 files changed, 22 insertions(+), 14 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 5e55a44f9eef6..dea0246a23c65 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -1209,6 +1209,11 @@ struct amdgpu_device {
 	struct mutex                    enforce_isolation_mutex;
 
 	struct amdgpu_init_level *init_lvl;
+
+	/* This flag is used to determine how VRAM allocations are handles for APUs
+	 * in KFD; VRAM or GTT.
+	 */
+	bool                            apu_prefer_gtt;
 };
 
 static inline uint32_t amdgpu_ip_version(const struct amdgpu_device *adev,
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
index 2c1b38c5cfc62..0312231b703e1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
@@ -459,7 +459,7 @@ void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev,
 		else
 			mem_info->local_mem_size_private =
 					KFD_XCP_MEMORY_SIZE(adev, xcp->id);
-	} else if (adev->flags & AMD_IS_APU) {
+	} else if (adev->apu_prefer_gtt) {
 		mem_info->local_mem_size_public = (ttm_tt_pages_limit() << PAGE_SHIFT);
 		mem_info->local_mem_size_private = 0;
 	} else {
@@ -818,7 +818,7 @@ u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id)
 		}
 		do_div(tmp, adev->xcp_mgr->num_xcp_per_mem_partition);
 		return ALIGN_DOWN(tmp, PAGE_SIZE);
-	} else if (adev->flags & AMD_IS_APU) {
+	} else if (adev->apu_prefer_gtt) {
 		return (ttm_tt_pages_limit() << PAGE_SHIFT);
 	} else {
 		return adev->gmc.real_vram_size;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 2b4d3840e5bca..60062c10b083f 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -197,7 +197,7 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 			return -EINVAL;
 
 		vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id);
-		if (adev->flags & AMD_IS_APU) {
+		if (adev->apu_prefer_gtt) {
 			system_mem_needed = size;
 			ttm_mem_needed = size;
 		}
@@ -234,7 +234,7 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 	if (adev && xcp_id >= 0) {
 		adev->kfd.vram_used[xcp_id] += vram_needed;
 		adev->kfd.vram_used_aligned[xcp_id] +=
-				(adev->flags & AMD_IS_APU) ?
+				adev->apu_prefer_gtt ?
 				vram_needed :
 				ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
 	}
@@ -262,7 +262,7 @@ void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
 
 		if (adev) {
 			adev->kfd.vram_used[xcp_id] -= size;
-			if (adev->flags & AMD_IS_APU) {
+			if (adev->apu_prefer_gtt) {
 				adev->kfd.vram_used_aligned[xcp_id] -= size;
 				kfd_mem_limit.system_mem_used -= size;
 				kfd_mem_limit.ttm_mem_used -= size;
@@ -890,7 +890,7 @@ static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
 	 * if peer device has large BAR. In contrast, access over xGMI is
 	 * allowed for both small and large BAR configurations of peer device
 	 */
-	if ((adev != bo_adev && !(adev->flags & AMD_IS_APU)) &&
+	if ((adev != bo_adev && !adev->apu_prefer_gtt) &&
 	    ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) ||
 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) ||
 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
@@ -1667,7 +1667,7 @@ size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
 		- reserved_for_pt
 		- reserved_for_ras;
 
-	if (adev->flags & AMD_IS_APU) {
+	if (adev->apu_prefer_gtt) {
 		system_mem_available = no_system_mem_limit ?
 					kfd_mem_limit.max_system_mem_limit :
 					kfd_mem_limit.max_system_mem_limit -
@@ -1715,7 +1715,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
 
-		if (adev->flags & AMD_IS_APU) {
+		if (adev->apu_prefer_gtt) {
 			domain = AMDGPU_GEM_DOMAIN_GTT;
 			alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
 			alloc_flags = 0;
@@ -1966,7 +1966,7 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
 	if (size) {
 		if (!is_imported &&
 		   (mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM ||
-		   ((adev->flags & AMD_IS_APU) &&
+		   (adev->apu_prefer_gtt &&
 		    mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT)))
 			*size = bo_size;
 		else
@@ -2393,7 +2393,7 @@ static int import_obj_create(struct amdgpu_device *adev,
 	(*mem)->bo = bo;
 	(*mem)->va = va;
 	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) &&
-			 !(adev->flags & AMD_IS_APU) ?
+			 !adev->apu_prefer_gtt ?
 			 AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
 
 	(*mem)->mapped_to_gpu_memory = 0;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index 8b6162217b322..9d2a3e2ad9df5 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -1985,6 +1985,9 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
 	DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
 		 (unsigned int)(gtt_size / (1024 * 1024)));
 
+	if (adev->flags & AMD_IS_APU)
+		adev->apu_prefer_gtt = true;
+
 	/* Initialize doorbell pool on PCI BAR */
 	r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_DOORBELL, adev->doorbell.size / PAGE_SIZE);
 	if (r) {
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
index d05d199b5e448..79251f22b7022 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
@@ -1027,7 +1027,7 @@ int kgd2kfd_init_zone_device(struct amdgpu_device *adev)
 	if (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 0, 1))
 		return -EINVAL;
 
-	if (adev->flags & AMD_IS_APU)
+	if (adev->apu_prefer_gtt)
 		return 0;
 
 	pgmap = &kfddev->pgmap;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index e32e19196f6b2..db3034b00dac2 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -2692,7 +2692,7 @@ svm_range_best_restore_location(struct svm_range *prange,
 		return -1;
 	}
 
-	if (node->adev->flags & AMD_IS_APU)
+	if (node->adev->apu_prefer_gtt)
 		return 0;
 
 	if (prange->preferred_loc == gpuid ||
@@ -3441,7 +3441,7 @@ svm_range_best_prefetch_location(struct svm_range *prange)
 		goto out;
 	}
 
-	if (bo_node->adev->flags & AMD_IS_APU) {
+	if (bo_node->adev->apu_prefer_gtt) {
 		best_loc = 0;
 		goto out;
 	}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
index bddd24f04669e..6ea23c78009ce 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
@@ -202,7 +202,7 @@ void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_s
  * is initialized to not 0 when page migration register device memory.
  */
 #define KFD_IS_SVM_API_SUPPORTED(adev) ((adev)->kfd.pgmap.type != 0 ||\
-					((adev)->flags & AMD_IS_APU))
+					((adev)->apu_prefer_gtt))
 
 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo);
 
-- 
2.48.1

[PATCH 2/2] drm/amdkfd: use GTT for VRAM on APUs only if GTT is larger

[Alex Deucher]

If the user has configured a large carveout on a small APU,
only use GTT for VRAM allocations if GTT is larger than
VRAM.

Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index 9d2a3e2ad9df5..96ec26946db60 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -1985,8 +1985,10 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
 	DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
 		 (unsigned int)(gtt_size / (1024 * 1024)));
 
-	if (adev->flags & AMD_IS_APU)
-		adev->apu_prefer_gtt = true;
+	if (adev->flags & AMD_IS_APU) {
+		if (adev->gmc.real_vram_size >= gtt_size)
+			adev->apu_prefer_gtt = true;
+	}
 
 	/* Initialize doorbell pool on PCI BAR */
 	r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_DOORBELL, adev->doorbell.size / PAGE_SIZE);
-- 
2.48.1

Re: [PATCH 2/2] drm/amdkfd: use GTT for VRAM on APUs only if GTT is larger

[Philip Yang]

[-- Attachment #1: Type: text/html, Size: 1888 bytes --]

Re: [PATCH 2/2] drm/amdkfd: use GTT for VRAM on APUs only if GTT is larger

[Alex Deucher]

On Fri, Jan 31, 2025 at 12:13 PM Philip Yang <yangp@amd.com> wrote:
>
>
> On 2025-01-30 15:51, Alex Deucher wrote:
>
> If the user has configured a large carveout on a small APU,
> only use GTT for VRAM allocations if GTT is larger than
> VRAM.
>
> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> index 9d2a3e2ad9df5..96ec26946db60 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> @@ -1985,8 +1985,10 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>   DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
>   (unsigned int)(gtt_size / (1024 * 1024)));
>
> - if (adev->flags & AMD_IS_APU)
> - adev->apu_prefer_gtt = true;
> + if (adev->flags & AMD_IS_APU) {
> + if (adev->gmc.real_vram_size >= gtt_size)
>
> This will not work for MI300 large APU, maybe the condition should be reversed
>
> + if (adev->gmc.real_vram_size < gtt_size)

yes, sorry, I reversed the check.

Alex

>
> Regards,
> Philip
>
> + adev->apu_prefer_gtt = true;
> + }
>
>   /* Initialize doorbell pool on PCI BAR */
>   r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_DOORBELL, adev->doorbell.size / PAGE_SIZE);