Commit b81ea1b (PM / QoS: Fix concurrency issues and memory leaks in
device PM QoS) put calls to pm_qos_sysfs_add_latency(),
pm_qos_sysfs_add_flags(), pm_qos_sysfs_remove_latency(), and
pm_qos_sysfs_remove_flags() under dev_pm_qos_mtx, which was a
mistake, because it may lead to deadlocks in some situations.
For example, if pm_qos_remote_wakeup_store() is run in parallel
with dev_pm_qos_constraints_destroy(), they may deadlock in the
following way:
======================================================
[ INFO: possible circular locking dependency detected ]
3.9.0-rc4-next-20130328-sasha-00014-g91a3267 #319 Tainted: G W
-------------------------------------------------------
trinity-child6/12371 is trying to acquire lock:
(s_active#54){++++.+}, at: [<ffffffff81301631>] sysfs_addrm_finish+0x31/0x60
but task is already holding lock:
(dev_pm_qos_mtx){+.+.+.}, at: [<ffffffff81f07cc3>] dev_pm_qos_constraints_destroy+0x23/0x250
which lock already depends on the new lock.
the existing dependency chain (in reverse order) is:
-> #1 (dev_pm_qos_mtx){+.+.+.}:
[<ffffffff811811da>] lock_acquire+0x1aa/0x240
[<ffffffff83dab809>] __mutex_lock_common+0x59/0x5e0
[<ffffffff83dabebf>] mutex_lock_nested+0x3f/0x50
[<ffffffff81f07f2f>] dev_pm_qos_update_flags+0x3f/0xc0
[<ffffffff81f05f4f>] pm_qos_remote_wakeup_store+0x3f/0x70
[<ffffffff81efbb43>] dev_attr_store+0x13/0x20
[<ffffffff812ffdaa>] sysfs_write_file+0xfa/0x150
[<ffffffff8127f2c1>] __kernel_write+0x81/0x150
[<ffffffff812afc2d>] write_pipe_buf+0x4d/0x80
[<ffffffff812af57c>] splice_from_pipe_feed+0x7c/0x120
[<ffffffff812afa25>] __splice_from_pipe+0x45/0x80
[<ffffffff812b14fc>] splice_from_pipe+0x4c/0x70
[<ffffffff812b1538>] default_file_splice_write+0x18/0x30
[<ffffffff812afae3>] do_splice_from+0x83/0xb0
[<ffffffff812afb2e>] direct_splice_actor+0x1e/0x20
[<ffffffff812b0277>] splice_direct_to_actor+0xe7/0x200
[<ffffffff812b15bc>] do_splice_direct+0x4c/0x70
[<ffffffff8127eda9>] do_sendfile+0x169/0x300
[<ffffffff8127ff94>] SyS_sendfile64+0x64/0xb0
[<ffffffff83db7d18>] tracesys+0xe1/0xe6
-> #0 (s_active#54){++++.+}:
[<ffffffff811800cf>] __lock_acquire+0x15bf/0x1e50
[<ffffffff811811da>] lock_acquire+0x1aa/0x240
[<ffffffff81300aa2>] sysfs_deactivate+0x122/0x1a0
[<ffffffff81301631>] sysfs_addrm_finish+0x31/0x60
[<ffffffff812ff77f>] sysfs_hash_and_remove+0x7f/0xb0
[<ffffffff813035a1>] sysfs_unmerge_group+0x51/0x70
[<ffffffff81f068f4>] pm_qos_sysfs_remove_flags+0x14/0x20
[<ffffffff81f07490>] __dev_pm_qos_hide_flags+0x30/0x70
[<ffffffff81f07cd5>] dev_pm_qos_constraints_destroy+0x35/0x250
[<ffffffff81f06931>] dpm_sysfs_remove+0x11/0x50
[<ffffffff81efcf6f>] device_del+0x3f/0x1b0
[<ffffffff81efd128>] device_unregister+0x48/0x60
[<ffffffff82d4083c>] usb_hub_remove_port_device+0x1c/0x20
[<ffffffff82d2a9cd>] hub_disconnect+0xdd/0x160
[<ffffffff82d36ab7>] usb_unbind_interface+0x67/0x170
[<ffffffff81f001a7>] __device_release_driver+0x87/0xe0
[<ffffffff81f00559>] device_release_driver+0x29/0x40
[<ffffffff81effc58>] bus_remove_device+0x148/0x160
[<ffffffff81efd07f>] device_del+0x14f/0x1b0
[<ffffffff82d344f9>] usb_disable_device+0xf9/0x280
[<ffffffff82d34ff8>] usb_set_configuration+0x268/0x840
[<ffffffff82d3a7fc>] usb_remove_store+0x4c/0x80
[<ffffffff81efbb43>] dev_attr_store+0x13/0x20
[<ffffffff812ffdaa>] sysfs_write_file+0xfa/0x150
[<ffffffff8127f71d>] do_loop_readv_writev+0x4d/0x90
[<ffffffff8127f999>] do_readv_writev+0xf9/0x1e0
[<ffffffff8127faba>] vfs_writev+0x3a/0x60
[<ffffffff8127fc60>] SyS_writev+0x50/0xd0
[<ffffffff83db7d18>] tracesys+0xe1/0xe6
other info that might help us debug this:
Possible unsafe locking scenario:
CPU0 CPU1
---- ----
lock(dev_pm_qos_mtx);
lock(s_active#54);
lock(dev_pm_qos_mtx);
lock(s_active#54);
*** DEADLOCK ***
To avoid that, remove the calls to functions mentioned above from
under dev_pm_qos_mtx and introduce a separate lock to prevent races
between functions that add or remove device PM QoS sysfs attributes
from happening.
Reported-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
0f70306929