sysfs: don't use global workqueue in sysfs_schedule_callback()
authorAlex Chiang <achiang@hp.com>
Wed, 25 Mar 2009 21:11:36 +0000 (15:11 -0600)
committerGreg Kroah-Hartman <gregkh@suse.de>
Thu, 16 Apr 2009 23:17:08 +0000 (16:17 -0700)
commitd110271e1f4140a9fb06d968b1afe9ca56a6064e
treeeb9ecf34607b5bb40ed26e339c5185cdb341e92f
parentc19f83669a02d4fa047d0d40f518e90f6f19c4c6
sysfs: don't use global workqueue in sysfs_schedule_callback()

A sysfs attribute using sysfs_schedule_callback() to commit suicide
may end up calling device_unregister(), which will eventually call
a driver's ->remove function.

Drivers may call flush_scheduled_work() in their shutdown routines,
in which case lockdep will complain with something like the following:

  =============================================
  [ INFO: possible recursive locking detected ]
  2.6.29-rc8-kk #1
  ---------------------------------------------
  events/4/56 is trying to acquire lock:
  (events){--..}, at: [<ffffffff80257fc0>] flush_workqueue+0x0/0xa0

  but task is already holding lock:
  (events){--..}, at: [<ffffffff80257648>] run_workqueue+0x108/0x230

  other info that might help us debug this:
  3 locks held by events/4/56:
  #0:  (events){--..}, at: [<ffffffff80257648>] run_workqueue+0x108/0x230
  #1:  (&ss->work){--..}, at: [<ffffffff80257648>] run_workqueue+0x108/0x230
  #2:  (pci_remove_rescan_mutex){--..}, at: [<ffffffff803c10d1>] remove_callback+0x21/0x40

  stack backtrace:
  Pid: 56, comm: events/4 Not tainted 2.6.29-rc8-kk #1
  Call Trace:
  [<ffffffff8026dfcd>] validate_chain+0xb7d/0x1260
  [<ffffffff8026eade>] __lock_acquire+0x42e/0xa40
  [<ffffffff8026f148>] lock_acquire+0x58/0x80
  [<ffffffff80257fc0>] ? flush_workqueue+0x0/0xa0
  [<ffffffff8025800d>] flush_workqueue+0x4d/0xa0
  [<ffffffff80257fc0>] ? flush_workqueue+0x0/0xa0
  [<ffffffff80258070>] flush_scheduled_work+0x10/0x20
  [<ffffffffa0144065>] e1000_remove+0x55/0xfe [e1000e]
  [<ffffffff8033ee30>] ? sysfs_schedule_callback_work+0x0/0x50
  [<ffffffff803bfeb2>] pci_device_remove+0x32/0x70
  [<ffffffff80441da9>] __device_release_driver+0x59/0x90
  [<ffffffff80441edb>] device_release_driver+0x2b/0x40
  [<ffffffff804419d6>] bus_remove_device+0xa6/0x120
  [<ffffffff8043e46b>] device_del+0x12b/0x190
  [<ffffffff8043e4f6>] device_unregister+0x26/0x70
  [<ffffffff803ba969>] pci_stop_dev+0x49/0x60
  [<ffffffff803baab0>] pci_remove_bus_device+0x40/0xc0
  [<ffffffff803c10d9>] remove_callback+0x29/0x40
  [<ffffffff8033ee4f>] sysfs_schedule_callback_work+0x1f/0x50
  [<ffffffff8025769a>] run_workqueue+0x15a/0x230
  [<ffffffff80257648>] ? run_workqueue+0x108/0x230
  [<ffffffff8025846f>] worker_thread+0x9f/0x100
  [<ffffffff8025bce0>] ? autoremove_wake_function+0x0/0x40
  [<ffffffff802583d0>] ? worker_thread+0x0/0x100
  [<ffffffff8025b89d>] kthread+0x4d/0x80
  [<ffffffff8020d4ba>] child_rip+0xa/0x20
  [<ffffffff8020cebc>] ? restore_args+0x0/0x30
  [<ffffffff8025b850>] ? kthread+0x0/0x80
  [<ffffffff8020d4b0>] ? child_rip+0x0/0x20

Although we know that the device_unregister path will never acquire
a lock that a driver might try to acquire in its ->remove, in general
we should never attempt to flush a workqueue from within the same
workqueue, and lockdep rightly complains.

So as long as sysfs attributes cannot commit suicide directly and we
are stuck with this callback mechanism, put the sysfs callbacks on
their own workqueue instead of the global one.

This has the side benefit that if a suicidal sysfs attribute kicks
off a long chain of ->remove callbacks, we no longer induce a long
delay on the global queue.

This also fixes a missing module_put in the error path introduced
by sysfs-only-allow-one-scheduled-removal-callback-per-kobj.patch.

We never destroy the workqueue, but I'm not sure that's a
problem.

Reported-by: Kenji Kaneshige <kaneshige.kenji@jp.fujitsu.com>
Tested-by: Kenji Kaneshige <kaneshige.kenji@jp.fujitsu.com>
Signed-off-by: Alex Chiang <achiang@hp.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
fs/sysfs/file.c