[PATCH] drm/i915: Mark the object as dirty when setting to the CPU write domain.

[PATCH] drm/i915: Mark the object as dirty when setting to the CPU write domain.

[Chris Wilson]

Or else we may not write back the written pages upon unbind. For
example the contents of a batch buffer written using a simple mmap or
using shmmem pwrite may be discarded if we are forced to evict
everything whilst pinning the objects for execbuffer.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: stable@kernel.org
---
 drivers/gpu/drm/i915/i915_gem.c |    1 +
 1 files changed, 1 insertions(+), 0 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index f04612f..5d60c3b 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -3076,6 +3076,7 @@ i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj, int write)
 	if (write) {
 		obj->read_domains &= I915_GEM_DOMAIN_CPU;
 		obj->write_domain = I915_GEM_DOMAIN_CPU;
+		obj_priv->dirty = 1;
 	}
 
 	trace_i915_gem_object_change_domain(obj,
-- 
1.7.1

Re: [PATCH] drm/i915: Mark the object as dirty when setting to the CPU write domain.

[Chris Wilson]

On Sun,  9 May 2010 21:42:00 +0100, Chris Wilson <chris@chris-wilson.co.uk> wrote:
> Or else we may not write back the written pages upon unbind. For
> example the contents of a batch buffer written using a simple mmap or
> using shmmem pwrite may be discarded if we are forced to evict
> everything whilst pinning the objects for execbuffer.

Daniel Vetter pointed out that all callers of set_to_cpu_domain(write=TRUE)
manually mark the object as dirty, so this cannot be the cause of the
failures that I see. Well I enjoyed my eureka moment, short though it was.

I think that it would be wise to make set_to_cpu_domain() and its callers
behave consistently with set_to_gtt_domain(), whereby the
obj_priv->dirty=1 is done in the common function and not the callers.  A
task for tomorrow.
-ickle

-- 
Chris Wilson, Intel Open Source Technology Centre