use 32 bit division in slab_put_obj()

PATCH In review Benjamin LaHaise 2 participants 2 messages Jan 21, 2006 – Jan 23, 2006 Memory management

Share & tools

Benjamin LaHaise

author

2006-01-21 01:12 UTC #0 lore raw

PATCH 1/1 use 32 bit division in slab_put_obj()

The patch below improves the performance of slab_put_obj(). Without the cast, gcc considers ptrdiff_t a 64 bit signed integer and ends up emitting code to use a full signed 128 bit divide on EM64T, which is substantially slower than a 32 bit unsigned divide. I noticed this when looking at the profile of a case where the slab balance is just on edge and thrashes back and forth freeing a block.

Signed-off-by Benjamin LaHaise <[email protected]>

mm/slab.c renamed from home/bcrl/kernels/v2.6/linux-2.6.16-rc1-mm2/mm/slab.c +5 -1

diff -X work-2.6.16-rc1-mm2/Documentation/dontdiff -urP /home/bcrl/kernels/v2.6/linux-2.6.16-rc1-mm2/mm/slab.c work-2.6.16-rc1-mm2/mm/slab.c
@@ -2267,8 +2267,12 @@
 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, void *objp,
 			  int nodeid)
 {
-	unsigned int objnr = (objp - slabp->s_mem) / cachep->buffer_size;
+	/* Slabs are always <4GB in size, so use a less expensive division. */
+	unsigned objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
 
+#if DEBUG
+	WARN_ON((unsigned long)(objp - slabp->s_mem) > ~0U);
+#endif
 #if 0
 	/* Verify that the slab belongs to the intended node */
 	WARN_ON(slabp->nodeid != nodeid);

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Andrew Morton

2006-01-23 06:39 UTC #1 lore raw

Quoted 16 lines · Benjamin LaHaise

The patch below improves the performance of slab_put_obj(). Without the cast, gcc considers ptrdiff_t a 64 bit signed integer and ends up emitting code to use a full signed 128 bit divide on EM64T, which is substantially slower than a 32 bit unsigned divide. I noticed this when looking at the profile of a case where the slab balance is just on edge and thrashes back and forth freeing a block.
Signed-off-by Benjamin LaHaise <[email protected]>
diff -X work-2.6.16-rc1-mm2/Documentation/dontdiff -urP /home/bcrl/kernels/v2.6/linux-2.6.16-rc1-mm2/mm/slab.c work-2.6.16-rc1-mm2/mm/slab.c
mm/slab.c renamed from home/bcrl/kernels/v2.6/linux-2.6.16-rc1-mm2/mm/slab.c +0 -0
@@ -2267,8 +2267,12 @@
 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, void *objp,
 			  int nodeid)
 {

mm/slab.c line 2278

Quoted 10 lines

mm/slab.c renamed from home/bcrl/kernels/v2.6/linux-2.6.16-rc1-mm2/mm/slab.c +5 -1

-	unsigned int objnr = (objp - slabp->s_mem) / cachep->buffer_size;
+	/* Slabs are always <4GB in size, so use a less expensive division. */
+	unsigned objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
 
+#if DEBUG
+	WARN_ON((unsigned long)(objp - slabp->s_mem) > ~0U);
+#endif
 #if 0
 	/* Verify that the slab belongs to the intended node */
 	WARN_ON(slabp->nodeid != nodeid);

There were actually three places in slab.c which were doing this, so I updated all of them and then redid all the other slab patches in the-mm queue.

mm/slab.c renamed from mm/slab.c~use-32-bit-division-in-slab_put_obj +3 -3

diff -puN mm/slab.c~use-32-bit-division-in-slab_put_obj mm/slab.c
@@ -1398,7 +1398,7 @@ static void check_poison_obj(kmem_cache_
 		struct slab *slabp = page_get_slab(virt_to_page(objp));
 		int objnr;
 
-		objnr = (objp - slabp->s_mem) / cachep->objsize;
+		objnr = (unsigned)(objp - slabp->s_mem) / cachep->objsize;
 		if (objnr) {
 			objp = slabp->s_mem + (objnr - 1) * cachep->objsize;
 			realobj = (char *)objp + obj_dbghead(cachep);

@@ -2341,7 +2341,7 @@ static void *cache_free_debugcheck(kmem_
 	if (cachep->flags & SLAB_STORE_USER)
 		*dbg_userword(cachep, objp) = caller;
 
-	objnr = (objp - slabp->s_mem) / cachep->objsize;
+	objnr = (unsigned)(objp - slabp->s_mem) / cachep->objsize;
 
 	BUG_ON(objnr >= cachep->num);
 	BUG_ON(objp != slabp->s_mem + objnr * cachep->objsize);

@@ -2699,7 +2699,7 @@ static void free_block(kmem_cache_t *cac
 		slabp = page_get_slab(virt_to_page(objp));
 		l3 = cachep->nodelists[node];
 		list_del(&slabp->list);
-		objnr = (objp - slabp->s_mem) / cachep->objsize;
+		objnr = (unsigned)(objp - slabp->s_mem) / cachep->objsize;
 		check_spinlock_acquired_node(cachep, node);
 		check_slabp(cachep, slabp);