drm/nouveau: Never evict VRAM buffers to system.

VRAM->system is a synchronous operation: it involves scheduling a
VRAM->TT DMA transfer and stalling the CPU until it's finished so that
we can unbind the new memory from the translation tables. VRAM->TT can
always be performed asynchronously, even if TT is already full and we
have to move something out of it.

Additionally, allowing VRAM->system behaves badly under heavy memory
pressure because once we run out of TT, stuff starts to be moved back
and forth between VRAM and system, and the TT contents are hardly
renewed.

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index 028719fddf761e94b1f1ae44ef50f5fc860e2813..026612471c92de827d52f30b2e183e0253843c42 100644 (file)
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -439,8 +439,7 @@ nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
 
        switch (bo->mem.mem_type) {
        case TTM_PL_VRAM:
-               nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT |
-                                        TTM_PL_FLAG_SYSTEM);
+               nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT);
                break;
        default:
                nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM);