mm/slub: optimize alloc/free fastpath by removing preemption on/off

We had to insert a preempt enable/disable in the fastpath a while ago in
order to guarantee that tid and kmem_cache_cpu are retrieved on the same
cpu.  It is the problem only for CONFIG_PREEMPT in which scheduler can
move the process to other cpu during retrieving data.

Now, I reach the solution to remove preempt enable/disable in the
fastpath.  If tid is matched with kmem_cache_cpu's tid after tid and
kmem_cache_cpu are retrieved by separate this_cpu operation, it means
that they are retrieved on the same cpu.  If not matched, we just have
to retry it.

With this guarantee, preemption enable/disable isn't need at all even if
CONFIG_PREEMPT, so this patch removes it.

I saw roughly 5% win in a fast-path loop over kmem_cache_alloc/free in
CONFIG_PREEMPT.  (14.821 ns -> 14.049 ns)

Below is the result of Christoph's slab_test reported by Jesper Dangaard
Brouer.

* Before

 Single thread testing
 =====================
 1. Kmalloc: Repeatedly allocate then free test
 10000 times kmalloc(8) -> 49 cycles kfree -> 62 cycles
 10000 times kmalloc(16) -> 48 cycles kfree -> 64 cycles
 10000 times kmalloc(32) -> 53 cycles kfree -> 70 cycles
 10000 times kmalloc(64) -> 64 cycles kfree -> 77 cycles
 10000 times kmalloc(128) -> 74 cycles kfree -> 84 cycles
 10000 times kmalloc(256) -> 84 cycles kfree -> 114 cycles
 10000 times kmalloc(512) -> 83 cycles kfree -> 116 cycles
 10000 times kmalloc(1024) -> 81 cycles kfree -> 120 cycles
 10000 times kmalloc(2048) -> 104 cycles kfree -> 136 cycles
 10000 times kmalloc(4096) -> 142 cycles kfree -> 165 cycles
 10000 times kmalloc(8192) -> 238 cycles kfree -> 226 cycles
 10000 times kmalloc(16384) -> 403 cycles kfree -> 264 cycles
 2. Kmalloc: alloc/free test
 10000 times kmalloc(8)/kfree -> 68 cycles
 10000 times kmalloc(16)/kfree -> 68 cycles
 10000 times kmalloc(32)/kfree -> 69 cycles
 10000 times kmalloc(64)/kfree -> 68 cycles
 10000 times kmalloc(128)/kfree -> 68 cycles
 10000 times kmalloc(256)/kfree -> 68 cycles
 10000 times kmalloc(512)/kfree -> 74 cycles
 10000 times kmalloc(1024)/kfree -> 75 cycles
 10000 times kmalloc(2048)/kfree -> 74 cycles
 10000 times kmalloc(4096)/kfree -> 74 cycles
 10000 times kmalloc(8192)/kfree -> 75 cycles
 10000 times kmalloc(16384)/kfree -> 510 cycles

* After

 Single thread testing
 =====================
 1. Kmalloc: Repeatedly allocate then free test
 10000 times kmalloc(8) -> 46 cycles kfree -> 61 cycles
 10000 times kmalloc(16) -> 46 cycles kfree -> 63 cycles
 10000 times kmalloc(32) -> 49 cycles kfree -> 69 cycles
 10000 times kmalloc(64) -> 57 cycles kfree -> 76 cycles
 10000 times kmalloc(128) -> 66 cycles kfree -> 83 cycles
 10000 times kmalloc(256) -> 84 cycles kfree -> 110 cycles
 10000 times kmalloc(512) -> 77 cycles kfree -> 114 cycles
 10000 times kmalloc(1024) -> 80 cycles kfree -> 116 cycles
 10000 times kmalloc(2048) -> 102 cycles kfree -> 131 cycles
 10000 times kmalloc(4096) -> 135 cycles kfree -> 163 cycles
 10000 times kmalloc(8192) -> 238 cycles kfree -> 218 cycles
 10000 times kmalloc(16384) -> 399 cycles kfree -> 262 cycles
 2. Kmalloc: alloc/free test
 10000 times kmalloc(8)/kfree -> 65 cycles
 10000 times kmalloc(16)/kfree -> 66 cycles
 10000 times kmalloc(32)/kfree -> 65 cycles
 10000 times kmalloc(64)/kfree -> 66 cycles
 10000 times kmalloc(128)/kfree -> 66 cycles
 10000 times kmalloc(256)/kfree -> 71 cycles
 10000 times kmalloc(512)/kfree -> 72 cycles
 10000 times kmalloc(1024)/kfree -> 71 cycles
 10000 times kmalloc(2048)/kfree -> 71 cycles
 10000 times kmalloc(4096)/kfree -> 71 cycles
 10000 times kmalloc(8192)/kfree -> 65 cycles
 10000 times kmalloc(16384)/kfree -> 511 cycles

Most of the results are better than before.

Note that this change slightly worses performance in !CONFIG_PREEMPT,
roughly 0.3%.  Implementing each case separately would help performance,
but, since it's so marginal, I didn't do that.  This would help
maintanance since we have same code for all cases.

Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Christoph Lameter <cl@linux.com>
Tested-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This commit is contained in:
Joonsoo Kim 2015-02-10 14:09:32 -08:00 коммит произвёл Linus Torvalds
Родитель 913e027ca1
Коммит 9aabf810a6
1 изменённых файлов: 23 добавлений и 12 удалений

Просмотреть файл

@ -2398,13 +2398,24 @@ redo:
* reading from one cpu area. That does not matter as long
* as we end up on the original cpu again when doing the cmpxchg.
*
* Preemption is disabled for the retrieval of the tid because that
* must occur from the current processor. We cannot allow rescheduling
* on a different processor between the determination of the pointer
* and the retrieval of the tid.
* We should guarantee that tid and kmem_cache are retrieved on
* the same cpu. It could be different if CONFIG_PREEMPT so we need
* to check if it is matched or not.
*/
preempt_disable();
c = this_cpu_ptr(s->cpu_slab);
do {
tid = this_cpu_read(s->cpu_slab->tid);
c = raw_cpu_ptr(s->cpu_slab);
} while (IS_ENABLED(CONFIG_PREEMPT) && unlikely(tid != c->tid));
/*
* Irqless object alloc/free algorithm used here depends on sequence
* of fetching cpu_slab's data. tid should be fetched before anything
* on c to guarantee that object and page associated with previous tid
* won't be used with current tid. If we fetch tid first, object and
* page could be one associated with next tid and our alloc/free
* request will be failed. In this case, we will retry. So, no problem.
*/
barrier();
/*
* The transaction ids are globally unique per cpu and per operation on
@ -2412,8 +2423,6 @@ redo:
* occurs on the right processor and that there was no operation on the
* linked list in between.
*/
tid = c->tid;
preempt_enable();
object = c->freelist;
page = c->page;
@ -2659,11 +2668,13 @@ redo:
* data is retrieved via this pointer. If we are on the same cpu
* during the cmpxchg then the free will succedd.
*/
preempt_disable();
c = this_cpu_ptr(s->cpu_slab);
do {
tid = this_cpu_read(s->cpu_slab->tid);
c = raw_cpu_ptr(s->cpu_slab);
} while (IS_ENABLED(CONFIG_PREEMPT) && unlikely(tid != c->tid));
tid = c->tid;
preempt_enable();
/* Same with comment on barrier() in slab_alloc_node() */
barrier();
if (likely(page == c->page)) {
set_freepointer(s, object, c->freelist);