zsmalloc: fix zs_can_compact() integer overflow
authorSergey Senozhatsky <sergey.senozhatsky@gmail.com>
Mon, 9 May 2016 23:28:49 +0000 (16:28 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 10 May 2016 00:40:59 +0000 (17:40 -0700)
commit44f43e99fe70833058482d183e99fdfd11220996
treef682bfc09f7c223790914e0f6b8778ae107516b9
parent1e92a61c4c7ed85c1bec037c046e92d6dc762f32
zsmalloc: fix zs_can_compact() integer overflow

zs_can_compact() has two race conditions in its core calculation:

unsigned long obj_wasted = zs_stat_get(class, OBJ_ALLOCATED) -
zs_stat_get(class, OBJ_USED);

1) classes are not locked, so the numbers of allocated and used
   objects can change by the concurrent ops happening on other CPUs
2) shrinker invokes it from preemptible context

Depending on the circumstances, thus, OBJ_ALLOCATED can become
less than OBJ_USED, which can result in either very high or
negative `total_scan' value calculated later in do_shrink_slab().

do_shrink_slab() has some logic to prevent those cases:

 vmscan: shrink_slab: zs_shrinker_scan+0x0/0x28 [zsmalloc] negative objects to delete nr=-62
 vmscan: shrink_slab: zs_shrinker_scan+0x0/0x28 [zsmalloc] negative objects to delete nr=-62
 vmscan: shrink_slab: zs_shrinker_scan+0x0/0x28 [zsmalloc] negative objects to delete nr=-64
 vmscan: shrink_slab: zs_shrinker_scan+0x0/0x28 [zsmalloc] negative objects to delete nr=-62
 vmscan: shrink_slab: zs_shrinker_scan+0x0/0x28 [zsmalloc] negative objects to delete nr=-62
 vmscan: shrink_slab: zs_shrinker_scan+0x0/0x28 [zsmalloc] negative objects to delete nr=-62

However, due to the way `total_scan' is calculated, not every
shrinker->count_objects() overflow can be spotted and handled.
To demonstrate the latter, I added some debugging code to do_shrink_slab()
(x86_64) and the results were:

 vmscan: OVERFLOW: shrinker->count_objects() == -1 [18446744073709551615]
 vmscan: but total_scan > 0: 92679974445502
 vmscan: resulting total_scan: 92679974445502
[..]
 vmscan: OVERFLOW: shrinker->count_objects() == -1 [18446744073709551615]
 vmscan: but total_scan > 0: 22634041808232578
 vmscan: resulting total_scan: 22634041808232578

Even though shrinker->count_objects() has returned an overflowed value,
the resulting `total_scan' is positive, and, what is more worrisome, it
is insanely huge. This value is getting used later on in
shrinker->scan_objects() loop:

        while (total_scan >= batch_size ||
               total_scan >= freeable) {
                unsigned long ret;
                unsigned long nr_to_scan = min(batch_size, total_scan);

                shrinkctl->nr_to_scan = nr_to_scan;
                ret = shrinker->scan_objects(shrinker, shrinkctl);
                if (ret == SHRINK_STOP)
                        break;
                freed += ret;

                count_vm_events(SLABS_SCANNED, nr_to_scan);
                total_scan -= nr_to_scan;

                cond_resched();
        }

`total_scan >= batch_size' is true for a very-very long time and
'total_scan >= freeable' is also true for quite some time, because
`freeable < 0' and `total_scan' is large enough, for example,
22634041808232578. The only break condition, in the given scheme of
things, is shrinker->scan_objects() == SHRINK_STOP test, which is a
bit too weak to rely on, especially in heavy zsmalloc-usage scenarios.

To fix the issue, take a pool stat snapshot and use it instead of
racy zs_stat_get() calls.

Link: http://lkml.kernel.org/r/20160509140052.3389-1-sergey.senozhatsky@gmail.com
Signed-off-by: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: <stable@vger.kernel.org> [4.3+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/zsmalloc.c