Received-SPF: pass (google.com: domain of linux-kernel+bounces-202427-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45d1:ec00::1 as permitted sender) client-ip=2604:1380:45d1:ec00::1;
From: Sasha Levin <sashal@kernel.org>
To: linux-kernel@vger.kernel.org,
	stable@vger.kernel.org
Cc: Lang Yu <Lang.Yu@amd.com>,
	Felix Kuehling <felix.kuehling@amd.com>,
	Alex Deucher <alexander.deucher@amd.com>,
	Sasha Levin <sashal@kernel.org>,
	Felix.Kuehling@amd.com,
	christian.koenig@amd.com,
	Xinhui.Pan@amd.com,
	airlied@gmail.com,
	daniel@ffwll.ch,
	amd-gfx@lists.freedesktop.org,
	dri-devel@lists.freedesktop.org
Subject: [PATCH AUTOSEL 6.8 4/6] drm/amdkfd: Let VRAM allocations go to GTT domain on small APUs
Date: Wed,  5 Jun 2024 07:55:56 -0400
Message-ID: <20240605115602.2964993-4-sashal@kernel.org>
In-Reply-To: <20240605115602.2964993-1-sashal@kernel.org>
References: <20240605115602.2964993-1-sashal@kernel.org>
Precedence: bulk
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit

From: Lang Yu <Lang.Yu@amd.com>

[ Upstream commit eb853413d02c8d9b27942429b261a9eef228f005 ]

Small APUs(i.e., consumer, embedded products) usually have a small
carveout device memory which can't satisfy most compute workloads
memory allocation requirements.

We can't even run a Basic MNIST Example with a default 512MB carveout.
https://github.com/pytorch/examples/tree/main/mnist. Error Log:

"torch.cuda.OutOfMemoryError: HIP out of memory. Tried to allocate
84.00 MiB. GPU 0 has a total capacity of 512.00 MiB of which 0 bytes
is free. Of the allocated memory 103.83 MiB is allocated by PyTorch,
and 22.17 MiB is reserved by PyTorch but unallocated"

Though we can change BIOS settings to enlarge carveout size,
which is inflexible and may bring complaint. On the other hand,
the memory resource can't be effectively used between host and device.

The solution is MI300A approach, i.e., let VRAM allocations go to GTT.
Then device and host can flexibly and effectively share memory resource.

v2: Report local_mem_size_private as 0. (Felix)

Signed-off-by: Lang Yu <Lang.Yu@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c    |  5 +++++
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 20 ++++++++++---------
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c      |  2 +-
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c          |  6 ++++--
 drivers/gpu/drm/amd/amdkfd/kfd_svm.h          |  3 ++-
 5 files changed, 23 insertions(+), 13 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
index 131983ed43465..7bd14a293ee52 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
@@ -455,6 +455,9 @@ 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) {
+		mem_info->local_mem_size_public = (ttm_tt_pages_limit() << PAGE_SHIFT);
+		mem_info->local_mem_size_private = 0;
 	} else {
 		mem_info->local_mem_size_public = adev->gmc.visible_vram_size;
 		mem_info->local_mem_size_private = adev->gmc.real_vram_size -
@@ -803,6 +806,8 @@ 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) {
+		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 36796fbc00315..fd0aef17ab36c 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -196,7 +196,7 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 			return -EINVAL;
 
 		vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id);
-		if (adev->gmc.is_app_apu) {
+		if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
 			system_mem_needed = size;
 			ttm_mem_needed = size;
 		}
@@ -232,7 +232,8 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 		  "adev reference can't be null when vram is used");
 	if (adev && xcp_id >= 0) {
 		adev->kfd.vram_used[xcp_id] += vram_needed;
-		adev->kfd.vram_used_aligned[xcp_id] += adev->gmc.is_app_apu ?
+		adev->kfd.vram_used_aligned[xcp_id] +=
+				(adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) ?
 				vram_needed :
 				ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
 	}
@@ -260,7 +261,7 @@ void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
 
 		if (adev) {
 			adev->kfd.vram_used[xcp_id] -= size;
-			if (adev->gmc.is_app_apu) {
+			if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
 				adev->kfd.vram_used_aligned[xcp_id] -= size;
 				kfd_mem_limit.system_mem_used -= size;
 				kfd_mem_limit.ttm_mem_used -= size;
@@ -887,7 +888,7 @@ 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->gmc.is_app_apu) &&
+	if ((adev != bo_adev && !(adev->gmc.is_app_apu || adev->flags & AMD_IS_APU)) &&
 	    ((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))) {
@@ -1654,7 +1655,7 @@ size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
 		- atomic64_read(&adev->vram_pin_size)
 		- reserved_for_pt;
 
-	if (adev->gmc.is_app_apu) {
+	if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
 		system_mem_available = no_system_mem_limit ?
 					kfd_mem_limit.max_system_mem_limit :
 					kfd_mem_limit.max_system_mem_limit -
@@ -1702,7 +1703,7 @@ 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->gmc.is_app_apu) {
+		if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
 			domain = AMDGPU_GEM_DOMAIN_GTT;
 			alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
 			alloc_flags = 0;
@@ -1949,7 +1950,7 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
 	if (size) {
 		if (!is_imported &&
 		   (mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM ||
-		   (adev->gmc.is_app_apu &&
+		   ((adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) &&
 		    mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT)))
 			*size = bo_size;
 		else
@@ -2372,8 +2373,9 @@ static int import_obj_create(struct amdgpu_device *adev,
 	(*mem)->dmabuf = dma_buf;
 	(*mem)->bo = bo;
 	(*mem)->va = va;
-	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) && !adev->gmc.is_app_apu ?
-		AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
+	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) &&
+			 !(adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) ?
+			 AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
 
 	(*mem)->mapped_to_gpu_memory = 0;
 	(*mem)->process_info = avm->process_info;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
index bdc01ca9609a7..00e4032976231 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
@@ -1009,7 +1009,7 @@ 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->gmc.is_app_apu)
+	if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU)
 		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 c50a0dc9c9c07..94f3614dd6c26 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -2617,7 +2617,8 @@ svm_range_best_restore_location(struct svm_range *prange,
 		return -1;
 	}
 
-	if (node->adev->gmc.is_app_apu)
+	if (node->adev->gmc.is_app_apu ||
+	    node->adev->flags & AMD_IS_APU)
 		return 0;
 
 	if (prange->preferred_loc == gpuid ||
@@ -3335,7 +3336,8 @@ svm_range_best_prefetch_location(struct svm_range *prange)
 		goto out;
 	}
 
-	if (bo_node->adev->gmc.is_app_apu) {
+	if (bo_node->adev->gmc.is_app_apu ||
+	    bo_node->adev->flags & AMD_IS_APU) {
 		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 026863a0abcd3..9c37bd0567efa 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
@@ -201,7 +201,8 @@ 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)->gmc.is_app_apu)
+					(adev)->gmc.is_app_apu ||\
+					((adev)->flags & AMD_IS_APU))
 
 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo);
 
-- 
2.43.0