drm/amdgpu: optimize VM fencing
authorChristian König <christian.koenig@amd.com>
Tue, 26 Jan 2016 10:40:46 +0000 (11:40 +0100)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 10 Feb 2016 19:17:06 +0000 (14:17 -0500)
No need to fence every page table, just the page directory is enough.

Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

index 7b660db..161652b 100644 (file)
@@ -624,36 +624,25 @@ static void amdgpu_vm_frag_ptes(struct amdgpu_device *adev,
  *
  * Global and local mutex must be locked!
  */
-static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
-                                struct amdgpu_gart *gtt,
-                                uint32_t gtt_flags,
-                                struct amdgpu_vm *vm,
-                                struct amdgpu_ib *ib,
-                                uint64_t start, uint64_t end,
-                                uint64_t dst, uint32_t flags)
+static void amdgpu_vm_update_ptes(struct amdgpu_device *adev,
+                                 struct amdgpu_gart *gtt,
+                                 uint32_t gtt_flags,
+                                 struct amdgpu_vm *vm,
+                                 struct amdgpu_ib *ib,
+                                 uint64_t start, uint64_t end,
+                                 uint64_t dst, uint32_t flags)
 {
        uint64_t mask = AMDGPU_VM_PTE_COUNT - 1;
        uint64_t last_pte = ~0, last_dst = ~0;
-       void *owner = AMDGPU_FENCE_OWNER_VM;
        unsigned count = 0;
        uint64_t addr;
 
-       /* sync to everything on unmapping */
-       if (!(flags & AMDGPU_PTE_VALID))
-               owner = AMDGPU_FENCE_OWNER_UNDEFINED;
-
        /* walk over the address space and update the page tables */
        for (addr = start; addr < end; ) {
                uint64_t pt_idx = addr >> amdgpu_vm_block_size;
                struct amdgpu_bo *pt = vm->page_tables[pt_idx].entry.robj;
                unsigned nptes;
                uint64_t pte;
-               int r;
-
-               amdgpu_sync_resv(adev, &ib->sync, pt->tbo.resv, owner);
-               r = reservation_object_reserve_shared(pt->tbo.resv);
-               if (r)
-                       return r;
 
                if ((addr & ~mask) == (end & ~mask))
                        nptes = end - addr;
@@ -687,8 +676,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
                                    last_pte, last_pte + 8 * count,
                                    last_dst, flags);
        }
-
-       return 0;
 }
 
 /**
@@ -716,11 +703,16 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
                                       struct fence **fence)
 {
        struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
+       void *owner = AMDGPU_FENCE_OWNER_VM;
        unsigned nptes, ncmds, ndw;
        struct amdgpu_ib *ib;
        struct fence *f = NULL;
        int r;
 
+       /* sync to everything on unmapping */
+       if (!(flags & AMDGPU_PTE_VALID))
+               owner = AMDGPU_FENCE_OWNER_UNDEFINED;
+
        nptes = last - start + 1;
 
        /*
@@ -761,15 +753,17 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
                return r;
        }
 
-       ib->length_dw = 0;
+       r = amdgpu_sync_resv(adev, &ib->sync, vm->page_directory->tbo.resv,
+                            owner);
+       if (r)
+               goto error_free;
 
-       r = amdgpu_vm_update_ptes(adev, gtt, gtt_flags, vm, ib, start,
-                                 last + 1, addr, flags);
-       if (r) {
-               amdgpu_ib_free(adev, ib);
-               kfree(ib);
-               return r;
-       }
+       r = reservation_object_reserve_shared(vm->page_directory->tbo.resv);
+       if (r)
+               goto error_free;
+
+       amdgpu_vm_update_ptes(adev, gtt, gtt_flags, vm, ib, start, last + 1,
+                             addr, flags);
 
        amdgpu_vm_pad_ib(adev, ib);
        WARN_ON(ib->length_dw > ndw);