drm/radeon: limit how much memory TTM can move per IB according to VRAM usage
authorMarek Olšák <marek.olsak@amd.com>
Sat, 1 Mar 2014 23:56:22 +0000 (00:56 +0100)
committerChristian König <christian.koenig@amd.com>
Mon, 3 Mar 2014 10:00:24 +0000 (11:00 +0100)
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
drivers/gpu/drm/radeon/radeon_cs.c
drivers/gpu/drm/radeon/radeon_object.c
drivers/gpu/drm/radeon/radeon_object.h

index 07e165128dbf67549682ce72f0bffed52249ca74..5abae403ea4fe4b83c7ba2902422c1ed860408ce 100644 (file)
@@ -168,7 +168,7 @@ static int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
 
        radeon_cs_buckets_get_list(&buckets, &p->validated);
 
-       return radeon_bo_list_validate(&p->ticket, &p->validated, p->ring);
+       return radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring);
 }
 
 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
index 8399fe02176978bb1ddfa57a17edab0d8099eb41..ed03f2d158534e286be9c281032d1dd0b9acea32 100644 (file)
@@ -366,29 +366,105 @@ void radeon_bo_fini(struct radeon_device *rdev)
        arch_phys_wc_del(rdev->mc.vram_mtrr);
 }
 
-int radeon_bo_list_validate(struct ww_acquire_ctx *ticket,
+/* Returns how many bytes TTM can move per IB.
+ */
+static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
+{
+       u64 real_vram_size = rdev->mc.real_vram_size;
+       u64 vram_usage = atomic64_read(&rdev->vram_usage);
+
+       /* This function is based on the current VRAM usage.
+        *
+        * - If all of VRAM is free, allow relocating the number of bytes that
+        *   is equal to 1/4 of the size of VRAM for this IB.
+
+        * - If more than one half of VRAM is occupied, only allow relocating
+        *   1 MB of data for this IB.
+        *
+        * - From 0 to one half of used VRAM, the threshold decreases
+        *   linearly.
+        *         __________________
+        * 1/4 of -|\               |
+        * VRAM    | \              |
+        *         |  \             |
+        *         |   \            |
+        *         |    \           |
+        *         |     \          |
+        *         |      \         |
+        *         |       \________|1 MB
+        *         |----------------|
+        *    VRAM 0 %             100 %
+        *         used            used
+        *
+        * Note: It's a threshold, not a limit. The threshold must be crossed
+        * for buffer relocations to stop, so any buffer of an arbitrary size
+        * can be moved as long as the threshold isn't crossed before
+        * the relocation takes place. We don't want to disable buffer
+        * relocations completely.
+        *
+        * The idea is that buffers should be placed in VRAM at creation time
+        * and TTM should only do a minimum number of relocations during
+        * command submission. In practice, you need to submit at least
+        * a dozen IBs to move all buffers to VRAM if they are in GTT.
+        *
+        * Also, things can get pretty crazy under memory pressure and actual
+        * VRAM usage can change a lot, so playing safe even at 50% does
+        * consistently increase performance.
+        */
+
+       u64 half_vram = real_vram_size >> 1;
+       u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
+       u64 bytes_moved_threshold = half_free_vram >> 1;
+       return max(bytes_moved_threshold, 1024*1024ull);
+}
+
+int radeon_bo_list_validate(struct radeon_device *rdev,
+                           struct ww_acquire_ctx *ticket,
                            struct list_head *head, int ring)
 {
        struct radeon_bo_list *lobj;
        struct radeon_bo *bo;
-       u32 domain;
        int r;
+       u64 bytes_moved = 0, initial_bytes_moved;
+       u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
 
        r = ttm_eu_reserve_buffers(ticket, head);
        if (unlikely(r != 0)) {
                return r;
        }
+
        list_for_each_entry(lobj, head, tv.head) {
                bo = lobj->bo;
                if (!bo->pin_count) {
-                       domain = lobj->domain;
-                       
+                       u32 domain = lobj->domain;
+                       u32 current_domain =
+                               radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
+
+                       /* Check if this buffer will be moved and don't move it
+                        * if we have moved too many buffers for this IB already.
+                        *
+                        * Note that this allows moving at least one buffer of
+                        * any size, because it doesn't take the current "bo"
+                        * into account. We don't want to disallow buffer moves
+                        * completely.
+                        */
+                       if (current_domain != RADEON_GEM_DOMAIN_CPU &&
+                           (domain & current_domain) == 0 && /* will be moved */
+                           bytes_moved > bytes_moved_threshold) {
+                               /* don't move it */
+                               domain = current_domain;
+                       }
+
                retry:
                        radeon_ttm_placement_from_domain(bo, domain);
                        if (ring == R600_RING_TYPE_UVD_INDEX)
                                radeon_uvd_force_into_uvd_segment(bo);
-                       r = ttm_bo_validate(&bo->tbo, &bo->placement,
-                                               true, false);
+
+                       initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
+                       r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
+                       bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
+                                      initial_bytes_moved;
+
                        if (unlikely(r)) {
                                if (r != -ERESTARTSYS && domain != lobj->alt_domain) {
                                        domain = lobj->alt_domain;
index 6c3ca9edc2f43f75b9232a3ac189309455dc3ce7..7dff64d1f592719c63bd0e58651a64c5e8ff6291 100644 (file)
@@ -138,7 +138,8 @@ extern int radeon_bo_evict_vram(struct radeon_device *rdev);
 extern void radeon_bo_force_delete(struct radeon_device *rdev);
 extern int radeon_bo_init(struct radeon_device *rdev);
 extern void radeon_bo_fini(struct radeon_device *rdev);
-extern int radeon_bo_list_validate(struct ww_acquire_ctx *ticket,
+extern int radeon_bo_list_validate(struct radeon_device *rdev,
+                                  struct ww_acquire_ctx *ticket,
                                   struct list_head *head, int ring);
 extern int radeon_bo_fbdev_mmap(struct radeon_bo *bo,
                                struct vm_area_struct *vma);