Am 04.03.20 um 20:40 schrieb Yong Zhao:
Because too many things are involved in this workaround, we need more
comments to avoid pitfalls.
Change-Id: I5d7917296dd5f5edb45921118cf8e7d778d40de1
Signed-off-by: Yong Zhao <Yong.Zhao@xxxxxxx>
Acked-by: Christian König <christian.koenig@xxxxxxx>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 5 ++++-
.../gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 18 +++++++++++++++---
2 files changed, 19 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index 1947a326de57..10f6f4b21b44 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -1041,7 +1041,10 @@ int amdgpu_ttm_gart_bind(struct amdgpu_device *adev,
if (r)
goto gart_bind_fail;
- /* Patch mtype of the second part BO */
+ /* The memory type of the first page defaults to UC. Now
+ * modify the memory type to NC from the second page of
+ * the BO onward.
+ */
flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
flags |= AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
index 436b7f518979..5b11190ff6e6 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
@@ -87,9 +87,21 @@ static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd,
int retval;
struct kfd_mem_obj *mqd_mem_obj = NULL;
- /* From V9, for CWSR, the control stack is located on the next page
- * boundary after the mqd, we will use the gtt allocation function
- * instead of sub-allocation function.
+ /* For V9 only, due to a HW bug, the control stack of a user mode
+ * compute queue needs to be allocated just behind the page boundary
+ * of its regular MQD buffer. So we allocate an enlarged MQD buffer:
+ * the first page of the buffer serves as the regular MQD buffer
+ * purpose and the remaining is for control stack. Although the two
+ * parts are in the same buffer object, they need different memory
+ * types: MQD part needs UC (uncached) as usual, while control stack
+ * needs NC (non coherent), which is different from the UC type which
+ * is used when control stack is allocated in user space.
+ *
+ * Because of all those, we use the gtt allocation function instead
+ * of sub-allocation function for this enlarged MQD buffer. Moreover,
+ * in order to achieve two memory types in a single buffer object, we
+ * pass a special bo flag AMDGPU_GEM_CREATE_MQD_GFX9 to instruct
+ * amdgpu memory functions to do so.
*/
if (kfd->cwsr_enabled && (q->type == KFD_QUEUE_TYPE_COMPUTE)) {
mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
