drm/i915: optimize the shmem_pwrite slowpath handling
authorDaniel Vetter <daniel.vetter@ffwll.ch>
Thu, 15 Nov 2012 15:53:58 +0000 (16:53 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Thu, 29 Nov 2012 12:49:08 +0000 (13:49 +0100)
commit8dcf015eb967c718962c0690330d9a94d56f2c5d
tree3650113cc42e2f2380b5a5f3123a913061e98489
parenta39a68054f63da0ea3b4806e1bfad79670a93d9f
drm/i915: optimize the shmem_pwrite slowpath handling

Since we drop dev->struct_mutex when going through the slowpath, the
object might have been moved out of the cpu domain. Hence we need to
clflush the entire object to ensure that after the ioctl returns,
everything is coherent again (interwoven writes are ill-defined
anyway).

But we only need to do this if we start in the cpu domain and the
object requires flushing for coherency. So don't do the flushing if
the object is coherent anyway or if we've done in-line clfushing
already.

v2: i915_gem_clflush_object already checks whether the object is
coherent and if so, drops the flushing. Hence we don't need to check
that ourselves, simplifying the condition.

v3: Reorder the checks for better clarity (and adjust the comment
accordingly), suggested by Chris Wilson.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_gem.c