slub: refactoring unfreeze_partials()
Joonsoo Kim [Fri, 8 Jun 2012 17:23:16 +0000 (02:23 +0900)]
Current implementation of unfreeze_partials() is so complicated,
but benefit from it is insignificant. In addition many code in
do {} while loop have a bad influence to a fail rate of cmpxchg_double_slab.
Under current implementation which test status of cpu partial slab
and acquire list_lock in do {} while loop,
we don't need to acquire a list_lock and gain a little benefit
when front of the cpu partial slab is to be discarded, but this is a rare case.
In case that add_partial is performed and cmpxchg_double_slab is failed,
remove_partial should be called case by case.

I think that these are disadvantages of current implementation,
so I do refactoring unfreeze_partials().

Minimizing code in do {} while loop introduce a reduced fail rate
of cmpxchg_double_slab. Below is output of 'slabinfo -r kmalloc-256'
when './perf stat -r 33 hackbench 50 process 4000 > /dev/null' is done.

** before **
Cmpxchg_double Looping
------------------------
Locked Cmpxchg Double redos   182685
Unlocked Cmpxchg Double redos 0

** after **
Cmpxchg_double Looping
------------------------
Locked Cmpxchg Double redos   177995
Unlocked Cmpxchg Double redos 1

We can see cmpxchg_double_slab fail rate is improved slightly.

Bolow is output of './perf stat -r 30 hackbench 50 process 4000 > /dev/null'.

** before **
 Performance counter stats for './hackbench 50 process 4000' (30 runs):

     108517.190463 task-clock                #    7.926 CPUs utilized            ( +-  0.24% )
         2,919,550 context-switches          #    0.027 M/sec                    ( +-  3.07% )
           100,774 CPU-migrations            #    0.929 K/sec                    ( +-  4.72% )
           124,201 page-faults               #    0.001 M/sec                    ( +-  0.15% )
   401,500,234,387 cycles                    #    3.700 GHz                      ( +-  0.24% )
   <not supported> stalled-cycles-frontend
   <not supported> stalled-cycles-backend
   250,576,913,354 instructions              #    0.62  insns per cycle          ( +-  0.13% )
    45,934,956,860 branches                  #  423.297 M/sec                    ( +-  0.14% )
       188,219,787 branch-misses             #    0.41% of all branches          ( +-  0.56% )

      13.691837307 seconds time elapsed                                          ( +-  0.24% )

** after **
 Performance counter stats for './hackbench 50 process 4000' (30 runs):

     107784.479767 task-clock                #    7.928 CPUs utilized            ( +-  0.22% )
         2,834,781 context-switches          #    0.026 M/sec                    ( +-  2.33% )
            93,083 CPU-migrations            #    0.864 K/sec                    ( +-  3.45% )
           123,967 page-faults               #    0.001 M/sec                    ( +-  0.15% )
   398,781,421,836 cycles                    #    3.700 GHz                      ( +-  0.22% )
   <not supported> stalled-cycles-frontend
   <not supported> stalled-cycles-backend
   250,189,160,419 instructions              #    0.63  insns per cycle          ( +-  0.09% )
    45,855,370,128 branches                  #  425.436 M/sec                    ( +-  0.10% )
       169,881,248 branch-misses             #    0.37% of all branches          ( +-  0.43% )

      13.596272341 seconds time elapsed                                          ( +-  0.22% )

No regression is found, but rather we can see slightly better result.

Acked-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Pekka Enberg <penberg@kernel.org>

mm/slub.c

index 4f406cd..f96d8bc 100644 (file)
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -1879,18 +1879,24 @@ redo:
  */
 static void unfreeze_partials(struct kmem_cache *s)
 {
-       struct kmem_cache_node *n = NULL;
+       struct kmem_cache_node *n = NULL, *n2 = NULL;
        struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
        struct page *page, *discard_page = NULL;
 
        while ((page = c->partial)) {
-               enum slab_modes { M_PARTIAL, M_FREE };
-               enum slab_modes l, m;
                struct page new;
                struct page old;
 
                c->partial = page->next;
-               l = M_FREE;
+
+               n2 = get_node(s, page_to_nid(page));
+               if (n != n2) {
+                       if (n)
+                               spin_unlock(&n->list_lock);
+
+                       n = n2;
+                       spin_lock(&n->list_lock);
+               }
 
                do {
 
@@ -1903,43 +1909,17 @@ static void unfreeze_partials(struct kmem_cache *s)
 
                        new.frozen = 0;
 
-                       if (!new.inuse && (!n || n->nr_partial > s->min_partial))
-                               m = M_FREE;
-                       else {
-                               struct kmem_cache_node *n2 = get_node(s,
-                                                       page_to_nid(page));
-
-                               m = M_PARTIAL;
-                               if (n != n2) {
-                                       if (n)
-                                               spin_unlock(&n->list_lock);
-
-                                       n = n2;
-                                       spin_lock(&n->list_lock);
-                               }
-                       }
-
-                       if (l != m) {
-                               if (l == M_PARTIAL) {
-                                       remove_partial(n, page);
-                                       stat(s, FREE_REMOVE_PARTIAL);
-                               } else {
-                                       add_partial(n, page,
-                                               DEACTIVATE_TO_TAIL);
-                                       stat(s, FREE_ADD_PARTIAL);
-                               }
-
-                               l = m;
-                       }
-
                } while (!__cmpxchg_double_slab(s, page,
                                old.freelist, old.counters,
                                new.freelist, new.counters,
                                "unfreezing slab"));
 
-               if (m == M_FREE) {
+               if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
                        page->next = discard_page;
                        discard_page = page;
+               } else {
+                       add_partial(n, page, DEACTIVATE_TO_TAIL);
+                       stat(s, FREE_ADD_PARTIAL);
                }
        }