From 44a9fa8cc6c7d598163d1885bf69e4bf747a004b Mon Sep 17 00:00:00 2001 From: Eric Anholt Date: Mon, 7 Jan 2008 13:10:50 -0800 Subject: Add additional explanation of DRM_BO_FLAG_CACHED_MAPPED before I forget again. --- shared-core/drm.h | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) (limited to 'shared-core/drm.h') diff --git a/shared-core/drm.h b/shared-core/drm.h index 6b4125f2..a551e4c2 100644 --- a/shared-core/drm.h +++ b/shared-core/drm.h @@ -705,7 +705,14 @@ struct drm_fence_arg { */ #define DRM_BO_FLAG_NO_MOVE (1ULL << 8) -/* Mask: Make sure the buffer is in cached memory when mapped +/* Mask: Make sure the buffer is in cached memory when mapped. In conjunction + * with DRM_BO_FLAG_CACHED it also allows the buffer to be bound into the GART + * with unsnooped PTEs instead of snooped, by using chipset-specific cache + * flushing at bind time. A better name might be DRM_BO_FLAG_TT_UNSNOOPED, + * as the eviction to local memory (TTM unbind) on map is just a side effect + * to prevent aggressive cache prefetch from the GPU disturbing the cache + * management that the DRM is doing. + * * Flags: Acknowledge. * Buffers allocated with this flag should not be used for suballocators * This type may have issues on CPUs with over-aggressive caching -- cgit v1.2.3