[Intel-gfx] [PATCH 2/2] drm/i915: Serialise updates to GGTT with access through GGTT on Braswell

Daniel Vetter daniel at ffwll.ch
Fri Oct 30 09:14:21 PDT 2015

On Fri, Oct 23, 2015 at 06:43:32PM +0100, Chris Wilson wrote:
> When accessing through the GTT from one CPU whilst concurrently updating
> the GGTT PTEs in another thread, the hardware likes to return random
> data. As we have strong serialisation prevent us from modifying the PTE
> of an active GTT mmapping, we have to conclude that it whilst modifying
> other PTE's that error occurs. (I have not looked for any pattern such
> as modifying PTE within the same page or cacheline as active PTE -
> though checking whether revoking neighbouring objects should be enough
> to test that theory.) The corruption also seems restricted to Braswell
> and disappears with maxcpus=0. This patch stops all access through the
> GTT by other CPUs when we update any PTE by stopping the machine around
> the GGTT update.
> Testcase: igt/gem_concurrent_blit
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89079
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>

Wild guess, since it wouldn't be the first time hw engineers screwed this

Cheers, Daniel

diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c
index d1c5cf89fe77..de983c8e6e54 100644
--- a/drivers/gpu/drm/i915/i915_gem_gtt.c
+++ b/drivers/gpu/drm/i915/i915_gem_gtt.c
@@ -2337,12 +2337,8 @@ int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
-#ifdef writeq
-	writeq(pte, addr);
 	iowrite32((u32)pte, addr);
 	iowrite32(pte >> 32, addr + 4);
 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
Daniel Vetter
Software Engineer, Intel Corporation

More information about the Intel-gfx mailing list