drm/amdgpu: Set MTYPE in PTE based on BO flags

The same BO may need different MTYPEs and SNOOP flags in PTEs depending
on its current location relative to the mapping GPU. Setting MTYPEs from
clients ahead of time is not practical for coherent memory sharing.
Instead determine the correct MTYPE for the desired coherence model and
current BO location when updating the page tables.

To maintain backwards compatibility with MTYPE-selection in
AMDGPU_VA_OP_MAP, the coherence-model-based MTYPE selection is only
applied if it chooses an MTYPE other than MTYPE_NC (the default).

Add two AMDGPU_GEM_CREATE_... flags to indicate the coherence model. The
default if no flag is specified is non-coherent (i.e. coarse-grained
coherent at dispatch boundaries).

Update amdgpu_amdkfd_gpuvm.c to use this new method to choose the
correct MTYPE depending on the current memory location.

v2:
* check that bo is not NULL (e.g. PRT mappings)
* Fix missing ~ bitmask in gmc_v11_0.c
v3:
* squash in "drm/amdgpu: Inherit coherence flags on dmabuf import"
Suggested-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c

@@ -405,63 +405,15 @@ static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
 
 static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem)
 {
-	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
-	bool coherent = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT;
-	bool uncached = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED;
-	uint32_t mapping_flags;
-	uint64_t pte_flags;
-	bool snoop = false;
+	uint32_t mapping_flags = AMDGPU_VM_PAGE_READABLE |
+				 AMDGPU_VM_MTYPE_DEFAULT;
 
-	mapping_flags = AMDGPU_VM_PAGE_READABLE;
 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE)
 		mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE)
 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
 
-	switch (adev->ip_versions[GC_HWIP][0]) {
-	case IP_VERSION(9, 4, 1):
-	case IP_VERSION(9, 4, 2):
-		if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
-			if (bo_adev == adev) {
-				if (uncached)
-					mapping_flags |= AMDGPU_VM_MTYPE_UC;
-				else if (coherent)
-					mapping_flags |= AMDGPU_VM_MTYPE_CC;
-				else
-					mapping_flags |= AMDGPU_VM_MTYPE_RW;
-				if ((adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2)) &&
-				    adev->gmc.xgmi.connected_to_cpu)
-					snoop = true;
-			} else {
-				if (uncached || coherent)
-					mapping_flags |= AMDGPU_VM_MTYPE_UC;
-				else
-					mapping_flags |= AMDGPU_VM_MTYPE_NC;
-				if (amdgpu_xgmi_same_hive(adev, bo_adev))
-					snoop = true;
-			}
-		} else {
-			if (uncached || coherent)
-				mapping_flags |= AMDGPU_VM_MTYPE_UC;
-			else
-				mapping_flags |= AMDGPU_VM_MTYPE_NC;
-			snoop = true;
-		}
-		break;
-	default:
-		if (uncached || coherent)
-			mapping_flags |= AMDGPU_VM_MTYPE_UC;
-		else
-			mapping_flags |= AMDGPU_VM_MTYPE_NC;
-
-		if (!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM))
-			snoop = true;
-	}
-
-	pte_flags = amdgpu_gem_va_map_flags(adev, mapping_flags);
-	pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
-
-	return pte_flags;
+	return amdgpu_gem_va_map_flags(adev, mapping_flags);
 }
 
 /**
@@ -1673,6 +1625,11 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 		}
 	}
 
+	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT)
+		alloc_flags |= AMDGPU_GEM_CREATE_COHERENT;
+	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED)
+		alloc_flags |= AMDGPU_GEM_CREATE_UNCACHED;
+
 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
 	if (!*mem) {
 		ret = -ENOMEM;

drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c

@@ -328,7 +328,9 @@ amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf)
 	if (dma_buf->ops == &amdgpu_dmabuf_ops) {
 		struct amdgpu_bo *other = gem_to_amdgpu_bo(dma_buf->priv);
 
-		flags |= other->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC;
+		flags |= other->flags & (AMDGPU_GEM_CREATE_CPU_GTT_USWC |
+					 AMDGPU_GEM_CREATE_COHERENT |
+					 AMDGPU_GEM_CREATE_UNCACHED);
 	}
 
 	ret = amdgpu_gem_object_create(adev, dma_buf->size, PAGE_SIZE,

drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c

@@ -612,6 +612,8 @@ static void gmc_v10_0_get_vm_pte(struct amdgpu_device *adev,
 				 struct amdgpu_bo_va_mapping *mapping,
 				 uint64_t *flags)
 {
+	struct amdgpu_bo *bo = mapping->bo_va->base.bo;
+
 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
 
@@ -628,6 +630,11 @@ static void gmc_v10_0_get_vm_pte(struct amdgpu_device *adev,
 		*flags |= AMDGPU_PTE_SYSTEM;
 		*flags &= ~AMDGPU_PTE_VALID;
 	}
+
+	if (bo && bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
+			       AMDGPU_GEM_CREATE_UNCACHED))
+		*flags = (*flags & ~AMDGPU_PTE_MTYPE_NV10_MASK) |
+			 AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
 }
 
 static unsigned gmc_v10_0_get_vbios_fb_size(struct amdgpu_device *adev)

drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c

@@ -503,6 +503,8 @@ static void gmc_v11_0_get_vm_pte(struct amdgpu_device *adev,
 				 struct amdgpu_bo_va_mapping *mapping,
 				 uint64_t *flags)
 {
+	struct amdgpu_bo *bo = mapping->bo_va->base.bo;
+
 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
 
@@ -519,6 +521,11 @@ static void gmc_v11_0_get_vm_pte(struct amdgpu_device *adev,
 		*flags |= AMDGPU_PTE_SYSTEM;
 		*flags &= ~AMDGPU_PTE_VALID;
 	}
+
+	if (bo && bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
+			       AMDGPU_GEM_CREATE_UNCACHED))
+		*flags = (*flags & ~AMDGPU_PTE_MTYPE_NV10_MASK) |
+			 AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
 }
 
 static unsigned gmc_v11_0_get_vbios_fb_size(struct amdgpu_device *adev)

drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c

@@ -1113,6 +1113,74 @@ static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level,
 	}
 }
 
+static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev,
+					 struct amdgpu_bo *bo,
+					 struct amdgpu_bo_va_mapping *mapping,
+					 uint64_t *flags)
+{
+	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
+	bool is_vram = bo->tbo.resource->mem_type == TTM_PL_VRAM;
+	bool coherent = bo->flags & AMDGPU_GEM_CREATE_COHERENT;
+	bool uncached = bo->flags & AMDGPU_GEM_CREATE_UNCACHED;
+	unsigned int mtype;
+	bool snoop = false;
+
+	switch (adev->ip_versions[GC_HWIP][0]) {
+	case IP_VERSION(9, 4, 1):
+	case IP_VERSION(9, 4, 2):
+		if (is_vram) {
+			if (bo_adev == adev) {
+				if (uncached)
+					mtype = MTYPE_UC;
+				else if (coherent)
+					mtype = MTYPE_CC;
+				else
+					mtype = MTYPE_RW;
+				/* FIXME: is this still needed? Or does
+				 * amdgpu_ttm_tt_pde_flags already handle this?
+				 */
+				if (adev->ip_versions[GC_HWIP][0] ==
+					IP_VERSION(9, 4, 2) &&
+				    adev->gmc.xgmi.connected_to_cpu)
+					snoop = true;
+			} else {
+				if (uncached || coherent)
+					mtype = MTYPE_UC;
+				else
+					mtype = MTYPE_NC;
+				if (mapping->bo_va->is_xgmi)
+					snoop = true;
+			}
+		} else {
+			if (uncached || coherent)
+				mtype = MTYPE_UC;
+			else
+				mtype = MTYPE_NC;
+			/* FIXME: is this still needed? Or does
+			 * amdgpu_ttm_tt_pde_flags already handle this?
+			 */
+			snoop = true;
+		}
+		break;
+	default:
+		if (uncached || coherent)
+			mtype = MTYPE_UC;
+		else
+			mtype = MTYPE_NC;
+
+		/* FIXME: is this still needed? Or does
+		 * amdgpu_ttm_tt_pde_flags already handle this?
+		 */
+		if (!is_vram)
+			snoop = true;
+	}
+
+	if (mtype != MTYPE_NC)
+		*flags = (*flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) |
+			 AMDGPU_PTE_MTYPE_VG10(mtype);
+	*flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
+}
+
 static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev,
 				struct amdgpu_bo_va_mapping *mapping,
 				uint64_t *flags)
@@ -1128,14 +1196,9 @@ static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev,
 		*flags &= ~AMDGPU_PTE_VALID;
 	}
 
-	if ((adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 1) ||
-	     adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2)) &&
-	    !(*flags & AMDGPU_PTE_SYSTEM) &&
-	    mapping->bo_va->is_xgmi)
-		*flags |= AMDGPU_PTE_SNOOPED;
-
-	if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2))
-		*flags |= mapping->flags & AMDGPU_PTE_SNOOPED;
+	if (mapping->bo_va->base.bo)
+		gmc_v9_0_get_coherence_flags(adev, mapping->bo_va->base.bo,
+					     mapping, flags);
 }
 
 static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)

include/uapi/drm/amdgpu_drm.h

@@ -144,6 +144,20 @@ extern "C" {
  * content.
  */
 #define AMDGPU_GEM_CREATE_DISCARDABLE		(1 << 12)
+/* Flag that BO is shared coherently between multiple devices or CPU threads.
+ * May depend on GPU instructions to flush caches explicitly
+ *
+ * This influences the choice of MTYPE in the PTEs on GFXv9 and later GPUs and
+ * may override the MTYPE selected in AMDGPU_VA_OP_MAP.
+ */
+#define AMDGPU_GEM_CREATE_COHERENT		(1 << 13)
+/* Flag that BO should not be cached by GPU. Coherent without having to flush
+ * GPU caches explicitly
+ *
+ * This influences the choice of MTYPE in the PTEs on GFXv9 and later GPUs and
+ * may override the MTYPE selected in AMDGPU_VA_OP_MAP.
+ */
+#define AMDGPU_GEM_CREATE_UNCACHED		(1 << 14)
 
 struct drm_amdgpu_gem_create_in  {
 	/** the requested memory size */