[PATCH 16/34] drm/amdkfd: support xgmi same hive mapping

[Felix Kuehling]

From: Philip Yang <Philip.Yang at 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 at amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling at 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 f01bc3b4010d..f557f67b9d2d 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"
 
@@ -953,21 +955,27 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,
 static int
 svm_range_map_to_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		     struct svm_range *prange, dma_addr_t *dma_addr,
-		     struct dma_fence **fence)
+		     struct amdgpu_device *bo_adev, struct dma_fence **fence)
 {
+	struct amdgpu_bo_va bo_va;
 	uint64_t pte_flags;
 	int r = 0;
 
 	pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start,
 		 prange->last);
 
+	if (prange->svm_bo && prange->ttm_res) {
+		bo_va.is_xgmi = amdgpu_xgmi_same_hive(adev, bo_adev);
+		prange->mapping.bo_va = &bo_va;
+	}
+
 	prange->mapping.start = prange->start;
 	prange->mapping.last = prange->last;
 	prange->mapping.offset = prange->offset;
 	pte_flags = svm_range_get_pte_flags(adev, prange);
 	prange->mapping.flags = pte_flags;
 
-	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,
@@ -990,6 +998,7 @@ svm_range_map_to_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		*fence = dma_fence_get(vm->last_update);
 
 out:
+	prange->mapping.bo_va = NULL;
 	return r;
 }
 
@@ -997,12 +1006,18 @@ static int svm_range_map_to_gpus(struct svm_range *prange,
 				 unsigned long *bitmap, bool wait)
 {
 	struct kfd_process_device *pdd;
+	struct amdgpu_device *bo_adev;
 	struct amdgpu_device *adev;
 	struct kfd_process *p;
 	struct dma_fence *fence = NULL;
 	uint32_t gpuidx;
 	int r = 0;
 
+	if (prange->svm_bo && prange->ttm_res)
+		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
+	else
+		bo_adev = NULL;
+
 	p = container_of(prange->svms, struct kfd_process, svms);
 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
@@ -1016,8 +1031,14 @@ static int svm_range_map_to_gpus(struct svm_range *prange,
 		if (IS_ERR(pdd))
 			return -EINVAL;
 
+		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,
-					 prange->dma_addr[gpuidx],
+					 prange->dma_addr[gpuidx], bo_adev,
 					 wait ? &fence : NULL);
 		if (r)
 			break;
-- 
2.31.1