drm/gem: replace misleading comment
authorDavid Herrmann <dh.herrmann@gmail.com>
Sun, 25 May 2014 12:34:08 +0000 (14:34 +0200)
committerDave Airlie <airlied@redhat.com>
Tue, 27 May 2014 05:50:57 +0000 (15:50 +1000)
shmem supports page-relocations during swapin since quite some time. It
was implemented in:

    commit bde05d1ccd512696b09db9dd2e5f33ad19152605
    Author: Hugh Dickins <hughd@google.com>
    Date:   Tue May 29 15:06:38 2012 -0700

        shmem: replace page if mapping excludes its zone

The gem-comment about wrongly placed DMA32 pages is no longer valid.
Replace it with a proper comment but keep the BUG_ON() to verify correct
shmem behavior.

Signed-off-by: David Herrmann <dh.herrmann@gmail.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
drivers/gpu/drm/drm_gem.c

index 9909bef..f7d7119 100644 (file)
@@ -474,21 +474,10 @@ struct page **drm_gem_get_pages(struct drm_gem_object *obj, gfp_t gfpmask)
                        goto fail;
                pages[i] = p;
 
-               /* There is a hypothetical issue w/ drivers that require
-                * buffer memory in the low 4GB.. if the pages are un-
-                * pinned, and swapped out, they can end up swapped back
-                * in above 4GB.  If pages are already in memory, then
-                * shmem_read_mapping_page_gfp will ignore the gfpmask,
-                * even if the already in-memory page disobeys the mask.
-                *
-                * It is only a theoretical issue today, because none of
-                * the devices with this limitation can be populated with
-                * enough memory to trigger the issue.  But this BUG_ON()
-                * is here as a reminder in case the problem with
-                * shmem_read_mapping_page_gfp() isn't solved by the time
-                * it does become a real issue.
-                *
-                * See this thread: http://lkml.org/lkml/2011/7/11/238
+               /* Make sure shmem keeps __GFP_DMA32 allocated pages in the
+                * correct region during swapin. Note that this requires
+                * __GFP_DMA32 to be set in mapping_gfp_mask(inode->i_mapping)
+                * so shmem can relocate pages during swapin if required.
                 */
                BUG_ON((gfpmask & __GFP_DMA32) &&
                                (page_to_pfn(p) >= 0x00100000UL));