[PATCH 19/35] drm/amdkfd: support xgmi same hive mapping

[Felix Kuehling <Felix.Kuehling@amd.com>]

From: Philip Yang <Philip.Yang@amd.com>

amdgpu_gmc_get_vm_pte use bo_va->is_xgmi same hive information to set
pte flags to update GPU mapping. Add local structure variable bo_va, and
update bo_va.is_xgmi, pass it to mapping->bo_va while mapping to GPU.

Assuming xgmi pstate is hi after boot.

Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 27 ++++++++++++++++++++++++---
 1 file changed, 24 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 7d91dc49a5a9..8a4d0a3935b6 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -26,6 +26,8 @@
 #include "amdgpu_object.h"
 #include "amdgpu_vm.h"
 #include "amdgpu_mn.h"
+#include "amdgpu.h"
+#include "amdgpu_xgmi.h"
 #include "kfd_priv.h"
 #include "kfd_svm.h"
 
@@ -923,10 +925,11 @@ static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
 static int
 svm_range_map_to_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		     struct svm_range *prange, bool reserve_vm,
-		     struct dma_fence **fence)
+		     struct amdgpu_device *bo_adev, struct dma_fence **fence)
 {
 	struct ttm_validate_buffer tv[2];
 	struct ww_acquire_ctx ticket;
+	struct amdgpu_bo_va bo_va;
 	struct list_head list;
 	dma_addr_t *pages_addr;
 	uint64_t pte_flags;
@@ -963,6 +966,11 @@ svm_range_map_to_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		}
 	}
 
+	if (prange->svm_bo && prange->mm_nodes) {
+		bo_va.is_xgmi = amdgpu_xgmi_same_hive(adev, bo_adev);
+		prange->mapping.bo_va = &bo_va;
+	}
+
 	prange->mapping.start = prange->it_node.start;
 	prange->mapping.last = prange->it_node.last;
 	prange->mapping.offset = prange->offset;
@@ -970,7 +978,7 @@ svm_range_map_to_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 	prange->mapping.flags = pte_flags;
 	pages_addr = prange->pages_addr;
 
-	r = amdgpu_vm_bo_update_mapping(adev, adev, vm, false, false, NULL,
+	r = amdgpu_vm_bo_update_mapping(adev, bo_adev, vm, false, false, NULL,
 					prange->mapping.start,
 					prange->mapping.last, pte_flags,
 					prange->mapping.offset,
@@ -994,6 +1002,7 @@ svm_range_map_to_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		*fence = dma_fence_get(vm->last_update);
 
 unreserve_out:
+	prange->mapping.bo_va = NULL;
 	if (reserve_vm)
 		ttm_eu_backoff_reservation(&ticket, &list);
 out:
@@ -1004,6 +1013,7 @@ static int svm_range_map_to_gpus(struct svm_range *prange, bool reserve_vm)
 {
 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
 	struct kfd_process_device *pdd;
+	struct amdgpu_device *bo_adev;
 	struct amdgpu_device *adev;
 	struct kfd_process *p;
 	struct kfd_dev *dev;
@@ -1011,6 +1021,11 @@ static int svm_range_map_to_gpus(struct svm_range *prange, bool reserve_vm)
 	uint32_t gpuidx;
 	int r = 0;
 
+	if (prange->svm_bo && prange->mm_nodes)
+		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
+	else
+		bo_adev = NULL;
+
 	bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
 		  MAX_GPU_INSTANCE);
 	p = container_of(prange->svms, struct kfd_process, svms);
@@ -1027,8 +1042,14 @@ static int svm_range_map_to_gpus(struct svm_range *prange, bool reserve_vm)
 			return -EINVAL;
 		adev = (struct amdgpu_device *)dev->kgd;
 
+		if (bo_adev && adev != bo_adev &&
+		    !amdgpu_xgmi_same_hive(adev, bo_adev)) {
+			pr_debug("cannot map to device idx %d\n", gpuidx);
+			continue;
+		}
+
 		r = svm_range_map_to_gpu(adev, pdd->vm, prange, reserve_vm,
-					 &fence);
+					 bo_adev, &fence);
 		if (r)
 			break;
 
-- 
2.29.2

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel