On Sat, 3 Nov 2007, Hugh Dickins wrote:
> On Sat, 3 Nov 2007, Olivér Pintér wrote:
> > Q: It's needed auch to 2.6.22-stable?
>
> I guess so: though SLUB wasn't on by default in 2.6.22; and it being
> only a slow leak rather than a corruption, I was less inclined to
> agitate about it for releases further back.
>
> But your question makes me realize I never even looked at 2.6.23 or
> 2.6.22 hereabouts, just assumed they were the same; let alone patch
> or build or test them. The patches reject as such because quite a
> lot has changed around (there was no struct kmem_cache_cpu in either).
>
> A hurried look suggests that the leakage problem was there in both,
> but let's wait to hear Christoph's expert opinion.
Okay, here's a version for 2.6.23 and 2.6.22...
Christoph, you've now Acked the 2.6.24 one, thanks:
do you agree this patch below should go to -stable?
Slub has been quite leaky under load. Taking mm_struct as an example, in
a loop of swapping kernel builds, after the first iteration slabinfo shows:
Name Objects Objsize Space Slabs/Part/Cpu O/S O %Fr %Ef Flg
mm_struct 55 840 73.7K 18/7/4 4 0 38 62 A
but Objects and Partials steadily creep up - after the 340th iteration:
mm_struct 110 840 188.4K 46/36/4 4 0 78 49 A
(example taken from 2.6.24-rc1: YMMV).
The culprit turns out to be __slab_alloc(), where it copes with the race
that another task has assigned the cpu slab while we were allocating one.
Don't rush off to load_freelist there: that assumes page->lockless_freelist
is empty, and will lose all its free slots when page->freelist is not empty.
Instead just do a local allocation from lockless_freelist when it has one.
Which fixes the leakage: Objects and Partials then remain stable.
Signed-off-by: Hugh Dickins <[EMAIL PROTECTED]>
---
Version of patch suitable and recommended for both 2.6.23-stable and
2.6.22-stable. I've not run tests on either to observe the mounting
leakage; but a version of the patch below with a printk announcing
when non-empty freelist would overwrite non-empty lockless_freelist
does indeed show up in both (though notably less frequently than in
2.6.24-rc1 - something else seems to be making it more likely now).
But please wait for Christoph's Ack before committing to -stable.
mm/slub.c | 6 ++++++
1 file changed, 6 insertions(+)
--- 2.6.23/mm/slub.c 2007-10-09 21:31:38.000000000 +0100
+++ linux/mm/slub.c 2007-11-03 18:23:07.000000000 +0000
@@ -1517,6 +1517,12 @@ new_slab:
*/
discard_slab(s, page);
page = s->cpu_slab[cpu];
+ if (page->lockless_freelist) {
+ object = page->lockless_freelist;
+ page->lockless_freelist =
+ object[page->offset];
+ return object;
+ }
slab_lock(page);
goto load_freelist;
}
