mm/page_counter: fix various data races at memsw
authorQian Cai <cai@lca.pw>
Sat, 15 Aug 2020 00:31:34 +0000 (17:31 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 15 Aug 2020 02:56:57 +0000 (19:56 -0700)
Commit 3e32cb2e0a12 ("mm: memcontrol: lockless page counters") could had
memcg->memsw->watermark and memcg->memsw->failcnt been accessed
concurrently as reported by KCSAN,

 BUG: KCSAN: data-race in page_counter_try_charge / page_counter_try_charge

 read to 0xffff8fb18c4cd190 of 8 bytes by task 1081 on cpu 59:
  page_counter_try_charge+0x4d/0x150 mm/page_counter.c:138
  try_charge+0x131/0xd50 mm/memcontrol.c:2405
  __memcg_kmem_charge_memcg+0x58/0x140
  __memcg_kmem_charge+0xcc/0x280
  __alloc_pages_nodemask+0x1e1/0x450
  alloc_pages_current+0xa6/0x120
  pte_alloc_one+0x17/0xd0
  __pte_alloc+0x3a/0x1f0
  copy_p4d_range+0xc36/0x1990
  copy_page_range+0x21d/0x360
  dup_mmap+0x5f5/0x7a0
  dup_mm+0xa2/0x240
  copy_process+0x1b3f/0x3460
  _do_fork+0xaa/0xa20
  __x64_sys_clone+0x13b/0x170
  do_syscall_64+0x91/0xb47
  entry_SYSCALL_64_after_hwframe+0x49/0xbe

 write to 0xffff8fb18c4cd190 of 8 bytes by task 1153 on cpu 120:
  page_counter_try_charge+0x5b/0x150 mm/page_counter.c:139
  try_charge+0x131/0xd50 mm/memcontrol.c:2405
  mem_cgroup_try_charge+0x159/0x460
  mem_cgroup_try_charge_delay+0x3d/0xa0
  wp_page_copy+0x14d/0x930
  do_wp_page+0x107/0x7b0
  __handle_mm_fault+0xce6/0xd40
  handle_mm_fault+0xfc/0x2f0
  do_page_fault+0x263/0x6f9
  page_fault+0x34/0x40

 BUG: KCSAN: data-race in page_counter_try_charge / page_counter_try_charge

 write to 0xffff88809bbf2158 of 8 bytes by task 11782 on cpu 0:
  page_counter_try_charge+0x100/0x170 mm/page_counter.c:129
  try_charge+0x185/0xbf0 mm/memcontrol.c:2405
  __memcg_kmem_charge_memcg+0x4a/0xe0 mm/memcontrol.c:2837
  __memcg_kmem_charge+0xcf/0x1b0 mm/memcontrol.c:2877
  __alloc_pages_nodemask+0x26c/0x310 mm/page_alloc.c:4780

 read to 0xffff88809bbf2158 of 8 bytes by task 11814 on cpu 1:
  page_counter_try_charge+0xef/0x170 mm/page_counter.c:129
  try_charge+0x185/0xbf0 mm/memcontrol.c:2405
  __memcg_kmem_charge_memcg+0x4a/0xe0 mm/memcontrol.c:2837
  __memcg_kmem_charge+0xcf/0x1b0 mm/memcontrol.c:2877
  __alloc_pages_nodemask+0x26c/0x310 mm/page_alloc.c:4780

Since watermark could be compared or set to garbage due to a data race
which would change the code logic, fix it by adding a pair of READ_ONCE()
and WRITE_ONCE() in those places.

The "failcnt" counter is tolerant of some degree of inaccuracy and is only
used to report stats, a data race will not be harmful, thus mark it as an
intentional data race using the data_race() macro.

Fixes: 3e32cb2e0a12 ("mm: memcontrol: lockless page counters")
Reported-by: syzbot+f36cfe60b1006a94f9dc@syzkaller.appspotmail.com
Signed-off-by: Qian Cai <cai@lca.pw>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Marco Elver <elver@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Link: http://lkml.kernel.org/r/1581519682-23594-1-git-send-email-cai@lca.pw
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/page_counter.c

index b466384..afe22ad 100644 (file)
@@ -77,8 +77,8 @@ void page_counter_charge(struct page_counter *counter, unsigned long nr_pages)
                 * This is indeed racy, but we can live with some
                 * inaccuracy in the watermark.
                 */
-               if (new > c->watermark)
-                       c->watermark = new;
+               if (new > READ_ONCE(c->watermark))
+                       WRITE_ONCE(c->watermark, new);
        }
 }
 
@@ -119,9 +119,10 @@ bool page_counter_try_charge(struct page_counter *counter,
                        propagate_protected_usage(c, new);
                        /*
                         * This is racy, but we can live with some
-                        * inaccuracy in the failcnt.
+                        * inaccuracy in the failcnt which is only used
+                        * to report stats.
                         */
-                       c->failcnt++;
+                       data_race(c->failcnt++);
                        *fail = c;
                        goto failed;
                }
@@ -130,8 +131,8 @@ bool page_counter_try_charge(struct page_counter *counter,
                 * Just like with failcnt, we can live with some
                 * inaccuracy in the watermark.
                 */
-               if (new > c->watermark)
-                       c->watermark = new;
+               if (new > READ_ONCE(c->watermark))
+                       WRITE_ONCE(c->watermark, new);
        }
        return true;