drm/amdgpu: allow higher level PD invalidations
authorChristian König <christian.koenig@amd.com>
Wed, 22 Jan 2020 12:41:02 +0000 (13:41 +0100)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 5 Feb 2020 04:30:39 +0000 (23:30 -0500)
Allow partial invalidation on unallocated PDs. This is useful when we
need to silence faults to stop interrupt floods on Vega.

Signed-off-by: Christian König <christian.koenig@amd.com>
Tested-by: Tom St Denis <tom.stdenis@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

index ca46622..71fc336 100644 (file)
@@ -1467,9 +1467,8 @@ static int amdgpu_vm_update_ptes(struct amdgpu_vm_update_params *params,
                         * smaller than the address shift. Go to the next
                         * child entry and try again.
                         */
-                       if (!amdgpu_vm_pt_descendant(adev, &cursor))
-                               return -ENOENT;
-                       continue;
+                       if (amdgpu_vm_pt_descendant(adev, &cursor))
+                               continue;
                } else if (frag >= parent_shift) {
                        /* If the fragment size is even larger than the parent
                         * shift we should go up one level and check it again.
@@ -1480,8 +1479,20 @@ static int amdgpu_vm_update_ptes(struct amdgpu_vm_update_params *params,
                }
 
                pt = cursor.entry->base.bo;
-               if (!pt)
-                       return -ENOENT;
+               if (!pt) {
+                       /* We need all PDs and PTs for mapping something, */
+                       if (flags & AMDGPU_PTE_VALID)
+                               return -ENOENT;
+
+                       /* but unmapping something can happen at a higher
+                        * level.
+                        */
+                       if (!amdgpu_vm_pt_ancestor(&cursor))
+                               return -EINVAL;
+
+                       pt = cursor.entry->base.bo;
+                       shift = parent_shift;
+               }
 
                /* Looks good so far, calculate parameters for the update */
                incr = (uint64_t)AMDGPU_GPU_PAGE_SIZE << shift;
@@ -1495,6 +1506,10 @@ static int amdgpu_vm_update_ptes(struct amdgpu_vm_update_params *params,
                        uint64_t upd_end = min(entry_end, frag_end);
                        unsigned nptes = (upd_end - frag_start) >> shift;
 
+                       /* This can happen when we set higher level PDs to
+                        * silent to stop fault floods.
+                        */
+                       nptes = max(nptes, 1u);
                        amdgpu_vm_update_flags(params, pt, cursor.level,
                                               pe_start, dst, nptes, incr,
                                               flags | AMDGPU_PTE_FRAG(frag));