Currently SLUB is using a strict L1_CACHE_BYTES alignment if
SLAB_HWCACHE_ALIGN is specified. SLAB does not align to a cacheline if the
object is smaller than half of a cacheline. Small objects are then aligned
by SLAB to a fraction of a cacheline.

Make SLUB just forget about the alignment requirement if the object size
is less than L1_CACHE_BYTES. It seems that fractional alignments are no
good because they grow the object and reduce the object density in a cache
line needlessly causing additional cache line fetches.

If we are already throwing the user suggestion of a cache line alignment
away then lets do the best we can. Maybe SLAB_HWCACHE_ALIGN also needs
to be tossed given its wishy-washy handling but doing so would require
an audit of all kmem_cache_allocs throughout the kernel source.

In any case one needs to explictly specify an alignment during
kmem_cache_create to either slab allocator in order to ensure that the
objects are cacheline aligned.

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.21-rc7-mm1/mm/slub.c
===================================================================
--- linux-2.6.21-rc7-mm1.orig/mm/slub.c	2007-04-25 21:23:56.000000000 -0700
+++ linux-2.6.21-rc7-mm1/mm/slub.c	2007-04-25 21:23:59.000000000 -0700
@@ -1482,9 +1482,19 @@ static int calculate_order(int size)
  * various ways of specifying it.
  */
 static unsigned long calculate_alignment(unsigned long flags,
-		unsigned long align)
+		unsigned long align, unsigned long size)
 {
-	if (flags & SLAB_HWCACHE_ALIGN)
+	/*
+	 * If the user wants hardware cache aligned objects then
+	 * follow that suggestion if the object is sufficiently
+	 * large.
+	 *
+	 * The hardware cache alignment cannot override the
+	 * specified alignment though. If that is greater
+	 * then use it.
+	 */
+	if ((flags & SLAB_HWCACHE_ALIGN) &&
+			size > L1_CACHE_BYTES / 2)
 		return max_t(unsigned long, align, L1_CACHE_BYTES);
 
 	if (align < ARCH_SLAB_MINALIGN)
@@ -1673,7 +1683,7 @@ static int calculate_sizes(struct kmem_c
 	 * user specified (this is unecessarily complex due to the attempt
 	 * to be compatible with SLAB. Should be cleaned up some day).
 	 */
-	align = calculate_alignment(flags, align);
+	align = calculate_alignment(flags, align, s->objsize);
 
 	/*
 	 * SLUB stores one object immediately after another beginning from
@@ -2250,7 +2260,7 @@ static struct kmem_cache *find_mergeable
 		return NULL;
 
 	size = ALIGN(size, sizeof(void *));
-	align = calculate_alignment(flags, align);
+	align = calculate_alignment(flags, align, size);
 	size = ALIGN(size, align);
 
 	list_for_each(h, &slab_caches) {

--
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/