thermal: cpu_cooling: fix lockdep problems in cpu_cooling
authorRussell King <rmk+kernel@arm.linux.org.uk>
Wed, 12 Aug 2015 09:52:16 +0000 (15:22 +0530)
committerEduardo Valentin <edubezval@gmail.com>
Sat, 15 Aug 2015 01:21:41 +0000 (18:21 -0700)
commit02373d7c69b4270bbab930f8a81b0721be794347
tree4cfe2d1f1621c42bb2ec89eee199fdd1326681eb
parentcf736ea6f902c26e03895dc7f5ccbc55cdc68e6e
thermal: cpu_cooling: fix lockdep problems in cpu_cooling

A recent change to the cpu_cooling code introduced a AB-BA deadlock
scenario between the cpufreq_policy_notifier_list rwsem and the
cooling_cpufreq_lock.  This is caused by cooling_cpufreq_lock being held
before the registration/removal of the notifier block (an operation
which takes the rwsem), and the notifier code itself which takes the
locks in the reverse order:

======================================================
[ INFO: possible circular locking dependency detected ]
3.18.0+ #1453 Not tainted
-------------------------------------------------------
rc.local/770 is trying to acquire lock:
 (cooling_cpufreq_lock){+.+.+.}, at: [<c04abfc4>] cpufreq_thermal_notifier+0x34/0xfc

but task is already holding lock:
 ((cpufreq_policy_notifier_list).rwsem){++++.+}, at: [<c0042f04>]  __blocking_notifier_call_chain+0x34/0x68

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #1 ((cpufreq_policy_notifier_list).rwsem){++++.+}:
       [<c06bc3b0>] down_write+0x44/0x9c
       [<c0043444>] blocking_notifier_chain_register+0x28/0xd8
       [<c04ad610>] cpufreq_register_notifier+0x68/0x90
       [<c04abe4c>] __cpufreq_cooling_register.part.1+0x120/0x180
       [<c04abf44>] __cpufreq_cooling_register+0x98/0xa4
       [<c04abf8c>] cpufreq_cooling_register+0x18/0x1c
       [<bf0046f8>] imx_thermal_probe+0x1c0/0x470 [imx_thermal]
       [<c037cef8>] platform_drv_probe+0x50/0xac
       [<c037b710>] driver_probe_device+0x114/0x234
       [<c037b8cc>] __driver_attach+0x9c/0xa0
       [<c0379d68>] bus_for_each_dev+0x5c/0x90
       [<c037b204>] driver_attach+0x24/0x28
       [<c037ae7c>] bus_add_driver+0xe0/0x1d8
       [<c037c0cc>] driver_register+0x80/0xfc
       [<c037cd80>] __platform_driver_register+0x50/0x64
       [<bf007018>] 0xbf007018
       [<c0008a5c>] do_one_initcall+0x88/0x1d8
       [<c0095da4>] load_module+0x1768/0x1ef8
       [<c0096614>] SyS_init_module+0xe0/0xf4
       [<c000ec00>] ret_fast_syscall+0x0/0x48

-> #0 (cooling_cpufreq_lock){+.+.+.}:
       [<c00619f8>] lock_acquire+0xb0/0x124
       [<c06ba3b4>] mutex_lock_nested+0x5c/0x3d8
       [<c04abfc4>] cpufreq_thermal_notifier+0x34/0xfc
       [<c0042bf4>] notifier_call_chain+0x4c/0x8c
       [<c0042f20>] __blocking_notifier_call_chain+0x50/0x68
       [<c0042f58>] blocking_notifier_call_chain+0x20/0x28
       [<c04ae62c>] cpufreq_set_policy+0x7c/0x1d0
       [<c04af3cc>] store_scaling_governor+0x74/0x9c
       [<c04ad418>] store+0x90/0xc0
       [<c0175384>] sysfs_kf_write+0x54/0x58
       [<c01746b4>] kernfs_fop_write+0xdc/0x190
       [<c010dcc0>] vfs_write+0xac/0x1b4
       [<c010dfec>] SyS_write+0x44/0x90
       [<c000ec00>] ret_fast_syscall+0x0/0x48

other info that might help us debug this:

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock((cpufreq_policy_notifier_list).rwsem);
                               lock(cooling_cpufreq_lock);
                               lock((cpufreq_policy_notifier_list).rwsem);
  lock(cooling_cpufreq_lock);

 *** DEADLOCK ***

7 locks held by rc.local/770:
 #0:  (sb_writers#6){.+.+.+}, at: [<c010dda0>] vfs_write+0x18c/0x1b4
 #1:  (&of->mutex){+.+.+.}, at: [<c0174678>] kernfs_fop_write+0xa0/0x190
 #2:  (s_active#52){.+.+.+}, at: [<c0174680>] kernfs_fop_write+0xa8/0x190
 #3:  (cpu_hotplug.lock){++++++}, at: [<c0026a60>] get_online_cpus+0x34/0x90
 #4:  (cpufreq_rwsem){.+.+.+}, at: [<c04ad3e0>] store+0x58/0xc0
 #5:  (&policy->rwsem){+.+.+.}, at: [<c04ad3f8>] store+0x70/0xc0
 #6:  ((cpufreq_policy_notifier_list).rwsem){++++.+}, at: [<c0042f04>] __blocking_notifier_call_chain+0x34/0x68

stack backtrace:
CPU: 0 PID: 770 Comm: rc.local Not tainted 3.18.0+ #1453
Hardware name: Freescale i.MX6 Quad/DualLite (Device Tree)
Backtrace:
[<c00121c8>] (dump_backtrace) from [<c0012360>] (show_stack+0x18/0x1c)
 r6:c0b85a80 r5:c0b75630 r4:00000000 r3:00000000
[<c0012348>] (show_stack) from [<c06b6c48>] (dump_stack+0x7c/0x98)
[<c06b6bcc>] (dump_stack) from [<c06b42a4>] (print_circular_bug+0x28c/0x2d8)
 r4:c0b85a80 r3:d0071d40
[<c06b4018>] (print_circular_bug) from [<c00613b0>] (__lock_acquire+0x1acc/0x1bb0)
 r10:c0b50660 r8:c09e6d80 r7:d0071d40 r6:c11d0f0c r5:00000007 r4:d0072240
[<c005f8e4>] (__lock_acquire) from [<c00619f8>] (lock_acquire+0xb0/0x124)
 r10:00000000 r9:c04abfc4 r8:00000000 r7:00000000 r6:00000000 r5:c0a06f0c
 r4:00000000
[<c0061948>] (lock_acquire) from [<c06ba3b4>] (mutex_lock_nested+0x5c/0x3d8)
 r10:ec853800 r9:c0a06ed4 r8:d0071d40 r7:c0a06ed4 r6:c11d0f0c r5:00000000
 r4:c04abfc4
[<c06ba358>] (mutex_lock_nested) from [<c04abfc4>] (cpufreq_thermal_notifier+0x34/0xfc)
 r10:ec853800 r9:ec85380c r8:d00d7d3c r7:c0a06ed4 r6:d00d7d3c r5:00000000
 r4:fffffffe
[<c04abf90>] (cpufreq_thermal_notifier) from [<c0042bf4>] (notifier_call_chain+0x4c/0x8c)
 r7:00000000 r6:00000000 r5:00000000 r4:fffffffe
[<c0042ba8>] (notifier_call_chain) from [<c0042f20>] (__blocking_notifier_call_chain+0x50/0x68)
 r8:c0a072a4 r7:00000000 r6:d00d7d3c r5:ffffffff r4:c0a06fc8 r3:ffffffff
[<c0042ed0>] (__blocking_notifier_call_chain) from [<c0042f58>] (blocking_notifier_call_chain+0x20/0x28)
 r7:ec98b540 r6:c13ebc80 r5:ed76e600 r4:d00d7d3c
[<c0042f38>] (blocking_notifier_call_chain) from [<c04ae62c>] (cpufreq_set_policy+0x7c/0x1d0)
[<c04ae5b0>] (cpufreq_set_policy) from [<c04af3cc>] (store_scaling_governor+0x74/0x9c)
 r7:ec98b540 r6:0000000c r5:ec98b540 r4:ed76e600
[<c04af358>] (store_scaling_governor) from [<c04ad418>] (store+0x90/0xc0)
 r6:0000000c r5:ed76e6d4 r4:ed76e600
[<c04ad388>] (store) from [<c0175384>] (sysfs_kf_write+0x54/0x58)
 r8:0000000c r7:d00d7f78 r6:ec98b540 r5:0000000c r4:ec853800 r3:0000000c
[<c0175330>] (sysfs_kf_write) from [<c01746b4>] (kernfs_fop_write+0xdc/0x190)
 r6:ec98b540 r5:00000000 r4:00000000 r3:c0175330
[<c01745d8>] (kernfs_fop_write) from [<c010dcc0>] (vfs_write+0xac/0x1b4)
 r10:0162aa70 r9:d00d6000 r8:0000000c r7:d00d7f78 r6:0162aa70 r5:0000000c
 r4:eccde500
[<c010dc14>] (vfs_write) from [<c010dfec>] (SyS_write+0x44/0x90)
 r10:0162aa70 r8:0000000c r7:eccde500 r6:eccde500 r5:00000000 r4:00000000
[<c010dfa8>] (SyS_write) from [<c000ec00>] (ret_fast_syscall+0x0/0x48)
 r10:00000000 r8:c000edc4 r7:00000004 r6:000216cc r5:0000000c r4:0162aa70

Solve this by moving to finer grained locking - use one mutex to protect
the cpufreq_dev_list as a whole, and a separate lock to ensure correct
ordering of cpufreq notifier registration and removal.

cooling_list_lock is taken within cooling_cpufreq_lock on
(un)registration to preserve the behavior of the code, i.e. to
atomically add/remove to the list and (un)register the notifier.

Fixes: 2dcd851fe4b4 ("thermal: cpu_cooling: Update always cpufreq policy with
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Eduardo Valentin <edubezval@gmail.com>
drivers/thermal/cpu_cooling.c