]> git.baikalelectronics.ru Git - kernel.git/commit
mm/mempool: fix a data race in mempool_free()
authorQian Cai <cai@lca.pw>
Sat, 15 Aug 2020 00:31:44 +0000 (17:31 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 15 Aug 2020 02:56:57 +0000 (19:56 -0700)
commit32c23ea377ab74bcc7a576b483a9ea68423dede6
tree2c2eb4a9298aa010c5f674a632ad648f267723f2
parent640902b91aedf6b27073d2b7257667f144f2453f
mm/mempool: fix a data race in mempool_free()

mempool_t pool.curr_nr could be accessed concurrently as noticed by
KCSAN,

 BUG: KCSAN: data-race in mempool_free / remove_element

 write to 0xffffffffa937638c of 4 bytes by task 6359 on cpu 113:
  remove_element+0x4a/0x1c0
  remove_element at mm/mempool.c:132
  mempool_alloc+0x102/0x210
  (inlined by) mempool_alloc at mm/mempool.c:399
  bio_alloc_bioset+0x106/0x2c0
  get_swap_bio+0x49/0x230
  __swap_writepage+0x680/0xc30
  swap_writepage+0x9c/0xf0
  pageout+0x33e/0xae0
  shrink_page_list+0x1f57/0x2870
  shrink_inactive_list+0x316/0x880
  shrink_lruvec+0x8dc/0x1380
  shrink_node+0x317/0xd80
  do_try_to_free_pages+0x1f7/0xa10
  try_to_free_pages+0x26c/0x5e0
  __alloc_pages_slowpath+0x458/0x1290
  <snip>

 read to 0xffffffffa937638c of 4 bytes by interrupt on cpu 64:
  mempool_free+0x3e/0x150
  mempool_free at mm/mempool.c:492
  bio_free+0x192/0x280
  bio_put+0x91/0xd0
  end_swap_bio_write+0x1d8/0x280
  bio_endio+0x2c2/0x5b0
  dec_pending+0x22b/0x440 [dm_mod]
  clone_endio+0xe4/0x2c0 [dm_mod]
  bio_endio+0x2c2/0x5b0
  blk_update_request+0x217/0x940
  scsi_end_request+0x6b/0x4d0
  scsi_io_completion+0xb7/0x7e0
  scsi_finish_command+0x223/0x310
  scsi_softirq_done+0x1d5/0x210
  blk_mq_complete_request+0x224/0x250
  scsi_mq_done+0xc2/0x250
  pqi_raid_io_complete+0x5a/0x70 [smartpqi]
  pqi_irq_handler+0x150/0x1410 [smartpqi]
  __handle_irq_event_percpu+0x90/0x540
  handle_irq_event_percpu+0x49/0xd0
  handle_irq_event+0x85/0xca
  handle_edge_irq+0x13f/0x3e0
  do_IRQ+0x86/0x190
  <snip>

Since the write is under pool->lock but the read is done as lockless.
Even though the commit 2d8a8530c537 ("mempool: fix and document
synchronization and memory barrier usage") introduced the smp_wmb() and
smp_rmb() pair to improve the situation, it is adequate to protect it
from data races which could lead to a logic bug, so fix it by adding
READ_ONCE() for the read.

Signed-off-by: Qian Cai <cai@lca.pw>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Marco Elver <elver@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Link: http://lkml.kernel.org/r/1581446384-2131-1-git-send-email-cai@lca.pw
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/mempool.c