[PATCH] [RFC] drm/radeon: clear WC flag when moving bo from vram to gtt

[Julien Isorce]

But re-add the flag is the bo is moved back to vram.

This fixes "ring 0/3 stalled" issue which happens when the driver
evicts bo from vram to gtt, at least on TAHITI and CAPVERDE.

I do not know the exact reason among the following:
  - si_copy_dma from vram to gtt is slow if WC
	(only for the non-visible part ? specific cases ?)
  - Allow snooping (SNOOPED flag from radeon_vm_bo_update).
  - WC should not be set at all for bo in the GTT
        (same reason why WC is only set for vram domain,
         see mesa::r600_init_resource_fields since mesa commit 5b6a0b7
             "gallium/radeon: set GTT WC on tiled textures")
  - Bug in WC
  - Same reason as why radeon_sa_bo_manager_init is not passing
    WC flags if older than CHIP_BONAIRE (see 810b73d1
        drm/radeon: Use write-combined CPU mappings of IBs on >= CIK)
  - Same as b738ca5d
        Revert "drm/radeon: Use write-combined CPU mappings of ring ..."
  - Same as 96ea47c0
        drm/radeon: Disable uncacheable CPU mappings of GTT with RV6xx
        see https://bugs.freedesktop.org/show_bug.cgi?id=91268#c2

https://bugs.freedesktop.org/show_bug.cgi?id=100712
---
 drivers/gpu/drm/radeon/radeon.h        |  1 +
 drivers/gpu/drm/radeon/radeon_object.c |  1 +
 drivers/gpu/drm/radeon/radeon_ttm.c    | 13 +++++++++++++
 3 files changed, 15 insertions(+)

diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
index 7a39a35..9847f4e 100644
--- a/drivers/gpu/drm/radeon/radeon.h
+++ b/drivers/gpu/drm/radeon/radeon.h
@@ -518,6 +518,7 @@ struct radeon_bo {
 
 	struct radeon_mn		*mn;
 	struct list_head		mn_list;
+	u32				vram_flags;
 };
 #define gem_to_radeon_bo(gobj) container_of((gobj), struct radeon_bo, gem_base)
 
diff --git a/drivers/gpu/drm/radeon/radeon_object.c b/drivers/gpu/drm/radeon/radeon_object.c
index a557869..870f6b0 100644
--- a/drivers/gpu/drm/radeon/radeon_object.c
+++ b/drivers/gpu/drm/radeon/radeon_object.c
@@ -228,6 +228,7 @@ int radeon_bo_create(struct radeon_device *rdev,
 				       RADEON_GEM_DOMAIN_CPU);
 
 	bo->flags = flags;
+	bo->vram_flags = 0;
 	/* PCI GART is always snooped */
 	if (!(rdev->flags & RADEON_IS_PCIE))
 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
diff --git a/drivers/gpu/drm/radeon/radeon_ttm.c b/drivers/gpu/drm/radeon/radeon_ttm.c
index d07ff84..a8743bd 100644
--- a/drivers/gpu/drm/radeon/radeon_ttm.c
+++ b/drivers/gpu/drm/radeon/radeon_ttm.c
@@ -440,6 +440,19 @@ static int radeon_bo_move(struct ttm_buffer_object *bo,
 		r = radeon_move_ram_vram(bo, evict, interruptible,
 					    no_wait_gpu, new_mem);
 	} else {
+		/* Clear WC flag when moving bo from vram to gtt. */
+		if (old_mem->mem_type == TTM_PL_VRAM && new_mem->mem_type == TTM_PL_TT) {
+			if (rbo->flags & RADEON_GEM_GTT_WC) {
+				rbo->vram_flags |= RADEON_GEM_GTT_WC;
+				rbo->flags &= ~RADEON_GEM_GTT_WC;
+			}
+		/* Re-add WC flag when moving back from gtt to vram. */
+		} else if (old_mem->mem_type == TTM_PL_TT && new_mem->mem_type == TTM_PL_VRAM) {
+			if (rbo->vram_flags & RADEON_GEM_GTT_WC) {
+				rbo->flags |= RADEON_GEM_GTT_WC;
+				rbo->vram_flags &= ~RADEON_GEM_GTT_WC;
+			}
+		}
 		r = radeon_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
 	}
 
-- 
2.7.4