Re: [BUG] While changing the cpufreq governor, kernel hits a bug in workqueue.c

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2008-07-10 at 13:11 +0200, Johannes Weiner wrote:
> Hi,
> 
> [added Peter on CC, lockdep confuses me]
> 
> Nageswara R Sastry <rnsastry@linux.vnet.ibm.com> writes:
> 
> > Hi Johannes,
> >
> > Johannes Weiner wrote:
> >
> >>> * I am seeing the circular locking dependency with the above patch
> >>> too.
> >>
> >> Uhm.  Failure or no failure?  A possible dead-lock report _is_ a
> >> failure.  So, do you get one or not?  And if so, could you send me the
> >> dmesg parts?
> >>
> >> Thanks a lot,
> >>
> >> 	Hannes
> >
> > I could see a circular locking dependency and sysfs hang with the new
> > patch named "cpufreq: cancel self-rearming work synchroneously"
> > also. Please find the dmesg output. Please let me know if you need
> > more information. Thank a lot for your coordination.
> >
> > =======================================================
> > [ INFO: possible circular locking dependency detected ]
> > 2.6.25.9.cpufreq #2
> > -------------------------------------------------------
> > S06cpuspeed/3427 is trying to acquire lock:
> >  (&(&dbs_info->work)->work){--..}, at: [<c012f46c>]
> > __cancel_work_timer+0x80/0x177
> >
> > but task is already holding lock:
> >  (dbs_mutex){--..}, at: [<c041e7db>] cpufreq_governor_dbs+0x26a/0x2f7
> 
> Hmmm, it's weird.  This path should be okay.  I wonder where the
> dependency work -> dbs_mutex comes from.  The mutex is nowhere taken
> with the work lock held (I removed this in the new version of the patch,
> can you double-check you applied to correct patch?).
> 
> So the chain should really be dbs_mutex -> work-lock.
> 
> > which lock already depends on the new lock.
> >
> >
> > the existing dependency chain (in reverse order) is:
> >
> > -> #2 (dbs_mutex){--..}:
> >        [<c013aa76>] add_lock_to_list+0x61/0x83
> >        [<c013cfa3>] __lock_acquire+0x953/0xb05
> >        [<c041e5e3>] cpufreq_governor_dbs+0x72/0x2f7
> >        [<c013d1b4>] lock_acquire+0x5f/0x79
> >        [<c041e5e3>] cpufreq_governor_dbs+0x72/0x2f7
> >        [<c04cdaa7>] mutex_lock_nested+0xce/0x222
> >        [<c041e5e3>] cpufreq_governor_dbs+0x72/0x2f7
> >        [<c041e5e3>] cpufreq_governor_dbs+0x72/0x2f7
> >        [<c041e5e3>] cpufreq_governor_dbs+0x72/0x2f7
> >        [<c041c87e>] __cpufreq_governor+0x73/0xa6
> >        [<c041c9ec>] __cpufreq_set_policy+0x13b/0x19e
> >        [<c041d6b9>] cpufreq_add_dev+0x3b4/0x4aa
> >        [<c041d29a>] handle_update+0x0/0x21
> >        [<c02ee310>] sysdev_driver_register+0x48/0x9a
> >        [<c041c75f>] cpufreq_register_driver+0x9b/0x147
> >        [<c06b742c>] kernel_init+0x130/0x26f
> >        [<c06b72fc>] kernel_init+0x0/0x26f
> >        [<c06b72fc>] kernel_init+0x0/0x26f
> >        [<c0105527>] kernel_thread_helper+0x7/0x10
> >        [<ffffffff>] 0xffffffff
> >
> > -> #1 (&per_cpu(cpu_policy_rwsem, cpu)){----}:
> >        [<c013cfa3>] __lock_acquire+0x953/0xb05
> >        [<c041d198>] lock_policy_rwsem_write+0x30/0x56
> >        [<c010a83b>] save_stack_trace+0x1a/0x35
> >        [<c013d1b4>] lock_acquire+0x5f/0x79
> >        [<c041d198>] lock_policy_rwsem_write+0x30/0x56
> >        [<c04cdfd9>] down_write+0x2b/0x44
> >        [<c041d198>] lock_policy_rwsem_write+0x30/0x56
> >        [<c041d198>] lock_policy_rwsem_write+0x30/0x56
> >        [<c041e362>] do_dbs_timer+0x40/0x24f
> >        [<c012ee7f>] run_workqueue+0x81/0x187
> >        [<c012eeba>] run_workqueue+0xbc/0x187
> >        [<c012ee7f>] run_workqueue+0x81/0x187
> >        [<c041e322>] do_dbs_timer+0x0/0x24f
> >        [<c012f6fa>] worker_thread+0x0/0xbd
> >        [<c012f7ad>] worker_thread+0xb3/0xbd
> >        [<c0131acc>] autoremove_wake_function+0x0/0x2d
> >        [<c0131a1b>] kthread+0x38/0x5d
> >        [<c01319e3>] kthread+0x0/0x5d
> >        [<c0105527>] kernel_thread_helper+0x7/0x10
> >        [<ffffffff>] 0xffffffff
> >
> > -> #0 (&(&dbs_info->work)->work){--..}:
> >        [<c013b6a2>] print_circular_bug_tail+0x2a/0x61
> >        [<c013cec8>] __lock_acquire+0x878/0xb05
> >        [<c013d1b4>] lock_acquire+0x5f/0x79
> >        [<c012f46c>] __cancel_work_timer+0x80/0x177
> >        [<c012f497>] __cancel_work_timer+0xab/0x177
> >        [<c012f46c>] __cancel_work_timer+0x80/0x177
> >        [<c013c0ee>] mark_held_locks+0x39/0x53
> >        [<c04cdbe8>] mutex_lock_nested+0x20f/0x222
> >        [<c013c277>] trace_hardirqs_on+0xe7/0x10e
> >        [<c04cdbf3>] mutex_lock_nested+0x21a/0x222
> >        [<c041e7db>] cpufreq_governor_dbs+0x26a/0x2f7
> >        [<c041e7ed>] cpufreq_governor_dbs+0x27c/0x2f7
> >        [<c041c87e>] __cpufreq_governor+0x73/0xa6
> >        [<c041c9da>] __cpufreq_set_policy+0x129/0x19e
> >        [<c041ce0f>] store_scaling_governor+0x112/0x135
> >        [<c041d29a>] handle_update+0x0/0x21
> >        [<c0410065>] atkbd_set_leds+0x5/0xcf
> >        [<c041ccfd>] store_scaling_governor+0x0/0x135
> >        [<c041d7eb>] store+0x3c/0x54
> >        [<c01a09e8>] sysfs_write_file+0xa9/0xdd
> >        [<c01a093f>] sysfs_write_file+0x0/0xdd
> >        [<c016e45a>] vfs_write+0x83/0xf6
> >        [<c016e9a0>] sys_write+0x3c/0x63
> >        [<c0104816>] sysenter_past_esp+0x5f/0xa5
> >        [<ffffffff>] 0xffffffff
> 
> Uhm, this dependency is as new as the actual lockdep detection (the same
> backtrace as the whole event, see below).  What is lockdep doing here?
> Shouldn't this be the callpath where the lock was taken for the first
> time?
> 
> I can not see where the chain is ever work-lock -> dbs_mutex, so how
> does lockdep come to the conclusion this would be the correct order?

Lockdep does not directly observe work-lock -> dbs_mutex, but an
indirect dependency through cpu_policy_rwsem.

So in do_dbs_timer() we get:

 work-lock
   cpu_policy_rwsem

then from:

  cpufreq_add_dev()
    __cpufreq_set_policy()
      __cpufreq_governor()

we obtain the dependency:

  cpu_policy_rwsem
    dbs_mutex

So together we get the chain:

  work-lock
    cpu_policy_rwsem
      dbs_mutex

And the below trace shows us trying to obtain the work-lock while
holding the dbs_mutex, which is in clear violation of the above locking
rules.

> > other info that might help us debug this:
> >
> > 3 locks held by S06cpuspeed/3427:
> >  #0:  (&buffer->mutex){--..}, at: [<c01a0963>] sysfs_write_file+0x24/0xdd
> >  #1:  (&per_cpu(cpu_policy_rwsem, cpu)){----}, at: [<c041d198>]
> > lock_policy_rwsem_write+0x30/0x56
> >  #2:  (dbs_mutex){--..}, at: [<c041e7db>] cpufreq_governor_dbs+0x26a/0x2f7
> >
> > stack backtrace:
> > Pid: 3427, comm: S06cpuspeed Not tainted 2.6.25.9.cpufreq #2
> >  [<c013b6cf>] print_circular_bug_tail+0x57/0x61
> >  [<c013cec8>] __lock_acquire+0x878/0xb05
> >  [<c013d1b4>] lock_acquire+0x5f/0x79
> >  [<c012f46c>] __cancel_work_timer+0x80/0x177
> >  [<c012f497>] __cancel_work_timer+0xab/0x177
> >  [<c012f46c>] __cancel_work_timer+0x80/0x177
> >  [<c013c0ee>] mark_held_locks+0x39/0x53
> >  [<c04cdbe8>] mutex_lock_nested+0x20f/0x222
> >  [<c013c277>] trace_hardirqs_on+0xe7/0x10e
> >  [<c04cdbf3>] mutex_lock_nested+0x21a/0x222
> >  [<c041e7db>] cpufreq_governor_dbs+0x26a/0x2f7
> >  [<c041e7ed>] cpufreq_governor_dbs+0x27c/0x2f7
> >  [<c041c87e>] __cpufreq_governor+0x73/0xa6
> >  [<c041c9da>] __cpufreq_set_policy+0x129/0x19e
> >  [<c041ce0f>] store_scaling_governor+0x112/0x135
> >  [<c041d29a>] handle_update+0x0/0x21
> >  [<c0410065>] atkbd_set_leds+0x5/0xcf
> >  [<c041ccfd>] store_scaling_governor+0x0/0x135
> >  [<c041d7eb>] store+0x3c/0x54
> >  [<c01a09e8>] sysfs_write_file+0xa9/0xdd
> >  [<c01a093f>] sysfs_write_file+0x0/0xdd
> >  [<c016e45a>] vfs_write+0x83/0xf6
> >  [<c016e9a0>] sys_write+0x3c/0x63
> >  [<c0104816>] sysenter_past_esp+0x5f/0xa5
> >  =======================
> >
> > Thanks!!
> > Regards
> > R.Nageswara Sastry
> 
> 	Hannes