drm/radeon: fence virtual address and free it once idle v4
authorJerome Glisse <jglisse@redhat.com>
Mon, 6 Aug 2012 16:32:21 +0000 (12:32 -0400)
committerAlex Deucher <alexander.deucher@amd.com>
Mon, 13 Aug 2012 14:50:55 +0000 (10:50 -0400)
Virtual address need to be fenced to know when we can safely remove it.
This patch also properly clear the pagetable. Previously it was
serouisly broken.

Kernel 3.5/3.4 need a similar patch but adapted for difference in mutex locking.

v2: For to update pagetable when unbinding bo (don't bailout if
    bo_va->valid is true).
v3: Add kernel 3.5/3.4 comment.
v4: Fix compilation warnings.

Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/radeon/radeon.h
drivers/gpu/drm/radeon/radeon_cs.c
drivers/gpu/drm/radeon/radeon_gart.c
drivers/gpu/drm/radeon/radeon_gem.c
drivers/gpu/drm/radeon/radeon_object.c

index 583e0f19e1c7e32bf82e919239e5777c1b21f2f7..b237a29142d160d62f9a098fb0241f8e6d1f2779 100644 (file)
@@ -300,6 +300,7 @@ struct radeon_bo_va {
        uint64_t                        soffset;
        uint64_t                        eoffset;
        uint32_t                        flags;
+       struct radeon_fence             *fence;
        bool                            valid;
 };
 
index 8a4c49ef0cc4e40170b3300aa7ad9cffeea2bc72..b4a0db24f4ddc4611028b671499d5062bd713649 100644 (file)
@@ -278,6 +278,30 @@ int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
        return 0;
 }
 
+static void radeon_bo_vm_fence_va(struct radeon_cs_parser *parser,
+                                 struct radeon_fence *fence)
+{
+       struct radeon_fpriv *fpriv = parser->filp->driver_priv;
+       struct radeon_vm *vm = &fpriv->vm;
+       struct radeon_bo_list *lobj;
+
+       if (parser->chunk_ib_idx == -1) {
+               return;
+       }
+       if ((parser->cs_flags & RADEON_CS_USE_VM) == 0) {
+               return;
+       }
+
+       list_for_each_entry(lobj, &parser->validated, tv.head) {
+               struct radeon_bo_va *bo_va;
+               struct radeon_bo *rbo = lobj->bo;
+
+               bo_va = radeon_bo_va(rbo, vm);
+               radeon_fence_unref(&bo_va->fence);
+               bo_va->fence = radeon_fence_ref(fence);
+       }
+}
+
 /**
  * cs_parser_fini() - clean parser states
  * @parser:    parser structure holding parsing context.
@@ -290,11 +314,14 @@ static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error)
 {
        unsigned i;
 
-       if (!error)
+       if (!error) {
+               /* fence all bo va before ttm_eu_fence_buffer_objects so bo are still reserved */
+               radeon_bo_vm_fence_va(parser, parser->ib.fence);
                ttm_eu_fence_buffer_objects(&parser->validated,
                                            parser->ib.fence);
-       else
+       } else {
                ttm_eu_backoff_reservation(&parser->validated);
+       }
 
        if (parser->relocs != NULL) {
                for (i = 0; i < parser->nrelocs; i++) {
@@ -388,7 +415,6 @@ static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
 
        if (parser->chunk_ib_idx == -1)
                return 0;
-
        if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
                return 0;
 
index b3720054614db5199684ad1d86924a1fd5c8f34d..9912182682ec8c2cd38d053d3e627d44aabb3969 100644 (file)
@@ -814,7 +814,7 @@ int radeon_vm_bo_update_pte(struct radeon_device *rdev,
                return -EINVAL;
        }
 
-       if (bo_va->valid)
+       if (bo_va->valid && mem)
                return 0;
 
        ngpu_pages = radeon_bo_ngpu_pages(bo);
@@ -859,11 +859,27 @@ int radeon_vm_bo_rmv(struct radeon_device *rdev,
                     struct radeon_bo *bo)
 {
        struct radeon_bo_va *bo_va;
+       int r;
 
        bo_va = radeon_bo_va(bo, vm);
        if (bo_va == NULL)
                return 0;
 
+       /* wait for va use to end */
+       while (bo_va->fence) {
+               r = radeon_fence_wait(bo_va->fence, false);
+               if (r) {
+                       DRM_ERROR("error while waiting for fence: %d\n", r);
+               }
+               if (r == -EDEADLK) {
+                       r = radeon_gpu_reset(rdev);
+                       if (!r)
+                               continue;
+               }
+               break;
+       }
+       radeon_fence_unref(&bo_va->fence);
+
        mutex_lock(&rdev->vm_manager.lock);
        mutex_lock(&vm->mutex);
        radeon_vm_bo_update_pte(rdev, vm, bo, NULL);
@@ -952,12 +968,15 @@ void radeon_vm_fini(struct radeon_device *rdev, struct radeon_vm *vm)
        radeon_vm_unbind_locked(rdev, vm);
        mutex_unlock(&rdev->vm_manager.lock);
 
-       /* remove all bo */
+       /* remove all bo at this point non are busy any more because unbind
+        * waited for the last vm fence to signal
+        */
        r = radeon_bo_reserve(rdev->ring_tmp_bo.bo, false);
        if (!r) {
                bo_va = radeon_bo_va(rdev->ring_tmp_bo.bo, vm);
                list_del_init(&bo_va->bo_list);
                list_del_init(&bo_va->vm_list);
+               radeon_fence_unref(&bo_va->fence);
                radeon_bo_unreserve(rdev->ring_tmp_bo.bo);
                kfree(bo_va);
        }
@@ -969,6 +988,7 @@ void radeon_vm_fini(struct radeon_device *rdev, struct radeon_vm *vm)
                r = radeon_bo_reserve(bo_va->bo, false);
                if (!r) {
                        list_del_init(&bo_va->bo_list);
+                       radeon_fence_unref(&bo_va->fence);
                        radeon_bo_unreserve(bo_va->bo);
                        kfree(bo_va);
                }
index 84d045245739eb3b049e78f88693d6dc1dd229da..1b57b0058ad642e6d0d610dadbf7af235f517c71 100644 (file)
@@ -134,25 +134,16 @@ void radeon_gem_object_close(struct drm_gem_object *obj,
        struct radeon_device *rdev = rbo->rdev;
        struct radeon_fpriv *fpriv = file_priv->driver_priv;
        struct radeon_vm *vm = &fpriv->vm;
-       struct radeon_bo_va *bo_va, *tmp;
 
        if (rdev->family < CHIP_CAYMAN) {
                return;
        }
 
        if (radeon_bo_reserve(rbo, false)) {
+               dev_err(rdev->dev, "leaking bo va because we fail to reserve bo\n");
                return;
        }
-       list_for_each_entry_safe(bo_va, tmp, &rbo->va, bo_list) {
-               if (bo_va->vm == vm) {
-                       /* remove from this vm address space */
-                       mutex_lock(&vm->mutex);
-                       list_del(&bo_va->vm_list);
-                       mutex_unlock(&vm->mutex);
-                       list_del(&bo_va->bo_list);
-                       kfree(bo_va);
-               }
-       }
+       radeon_vm_bo_rmv(rdev, vm, rbo);
        radeon_bo_unreserve(rbo);
 }
 
index 1f1a4c803c1dd267158dece882e8fdf5b2d6b871..1cb014b571ab93d2479f4456de93e6c4c672f1c4 100644 (file)
@@ -52,11 +52,7 @@ void radeon_bo_clear_va(struct radeon_bo *bo)
 
        list_for_each_entry_safe(bo_va, tmp, &bo->va, bo_list) {
                /* remove from all vm address space */
-               mutex_lock(&bo_va->vm->mutex);
-               list_del(&bo_va->vm_list);
-               mutex_unlock(&bo_va->vm->mutex);
-               list_del(&bo_va->bo_list);
-               kfree(bo_va);
+               radeon_vm_bo_rmv(bo->rdev, bo_va->vm, bo);
        }
 }