]> git.baikalelectronics.ru Git - kernel.git/commit
cpufreq: Fix timer/workqueue corruption by protecting reading governor_enabled
authorJane Li <jiel@marvell.com>
Fri, 3 Jan 2014 09:17:41 +0000 (17:17 +0800)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Mon, 6 Jan 2014 00:22:02 +0000 (01:22 +0100)
commitbb5eb603b6d5629edf021622d78c6ae811060878
tree0c3420a9cbd3ac70235a34a82165879a78aaa4a7
parent5c1798dee5e5199a1932b6a690e4cea9564783b5
cpufreq: Fix timer/workqueue corruption by protecting reading governor_enabled

When a CPU is hot removed we'll cancel all the delayed work items via
gov_cancel_work(). Sometimes the delayed work function determines that
it should adjust the delay for all other CPUs that the policy is
managing. If this scenario occurs, the canceling CPU will cancel its own
work but queue up the other CPUs works to run.

Commit 3617f2 (cpufreq: Fix timer/workqueue corruption due to double
queueing) has tried to fix this, but reading governor_enabled is not
protected by cpufreq_governor_lock. Even though od_dbs_timer() checks
governor_enabled before gov_queue_work(), this scenario may occur. For
example:

 CPU0                                        CPU1
 ----                                        ----
 cpu_down()
  ...                                        <work runs>
  __cpufreq_remove_dev()                     od_dbs_timer()
   __cpufreq_governor()                       policy->governor_enabled
    policy->governor_enabled = false;
    cpufreq_governor_dbs()
     case CPUFREQ_GOV_STOP:
      gov_cancel_work(dbs_data, policy);
       cpu0 work is canceled
        timer is canceled
        cpu1 work is canceled
        <waits for cpu1>
                                              gov_queue_work(*, *, true);
                                               cpu0 work queued
                                               cpu1 work queued
                                               cpu2 work queued
                                               ...
        cpu1 work is canceled
        cpu2 work is canceled
        ...

At the end of the GOV_STOP case cpu0 still has a work queued to
run although the code is expecting all of the works to be
canceled. __cpufreq_remove_dev() will then proceed to
re-initialize all the other CPUs works except for the CPU that is
going down. The CPUFREQ_GOV_START case in cpufreq_governor_dbs()
will trample over the queued work and debugobjects will spit out
a warning:

WARNING: at lib/debugobjects.c:260 debug_print_object+0x94/0xbc()
ODEBUG: init active (active state 0) object type: timer_list hint: delayed_work_timer_fn+0x0/0x14
Modules linked in:
CPU: 1 PID: 1205 Comm: sh Tainted: G        W    3.10.0 #200
[<c01144f0>] (unwind_backtrace+0x0/0xf8) from [<c0111d98>] (show_stack+0x10/0x14)
[<c0111d98>] (show_stack+0x10/0x14) from [<c01272cc>] (warn_slowpath_common+0x4c/0x68)
[<c01272cc>] (warn_slowpath_common+0x4c/0x68) from [<c012737c>] (warn_slowpath_fmt+0x30/0x40)
[<c012737c>] (warn_slowpath_fmt+0x30/0x40) from [<c034c640>] (debug_print_object+0x94/0xbc)
[<c034c640>] (debug_print_object+0x94/0xbc) from [<c034c7f8>] (__debug_object_init+0xc8/0x3c0)
[<c034c7f8>] (__debug_object_init+0xc8/0x3c0) from [<c01360e0>] (init_timer_key+0x20/0x104)
[<c01360e0>] (init_timer_key+0x20/0x104) from [<c04872ac>] (cpufreq_governor_dbs+0x1dc/0x68c)
[<c04872ac>] (cpufreq_governor_dbs+0x1dc/0x68c) from [<c04833a8>] (__cpufreq_governor+0x80/0x1b0)
[<c04833a8>] (__cpufreq_governor+0x80/0x1b0) from [<c0483704>] (__cpufreq_remove_dev.isra.12+0x22c/0x380)
[<c0483704>] (__cpufreq_remove_dev.isra.12+0x22c/0x380) from [<c0692f38>] (cpufreq_cpu_callback+0x48/0x5c)
[<c0692f38>] (cpufreq_cpu_callback+0x48/0x5c) from [<c014fb40>] (notifier_call_chain+0x44/0x84)
[<c014fb40>] (notifier_call_chain+0x44/0x84) from [<c012ae44>] (__cpu_notify+0x2c/0x48)
[<c012ae44>] (__cpu_notify+0x2c/0x48) from [<c068dd40>] (_cpu_down+0x80/0x258)
[<c068dd40>] (_cpu_down+0x80/0x258) from [<c068df40>] (cpu_down+0x28/0x3c)
[<c068df40>] (cpu_down+0x28/0x3c) from [<c068e4c0>] (store_online+0x30/0x74)
[<c068e4c0>] (store_online+0x30/0x74) from [<c03a7308>] (dev_attr_store+0x18/0x24)
[<c03a7308>] (dev_attr_store+0x18/0x24) from [<c0256fe0>] (sysfs_write_file+0x100/0x180)
[<c0256fe0>] (sysfs_write_file+0x100/0x180) from [<c01fec9c>] (vfs_write+0xbc/0x184)
[<c01fec9c>] (vfs_write+0xbc/0x184) from [<c01ff034>] (SyS_write+0x40/0x68)
[<c01ff034>] (SyS_write+0x40/0x68) from [<c010e200>] (ret_fast_syscall+0x0/0x48)

In gov_queue_work(), lock cpufreq_governor_lock before gov_queue_work,
and unlock it after __gov_queue_work(). In this way, governor_enabled
is guaranteed not changed in gov_queue_work().

Signed-off-by: Jane Li <jiel@marvell.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Dmitry Torokhov <dmitry.torokhov@gmail.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
drivers/cpufreq/cpufreq.c
drivers/cpufreq/cpufreq_governor.c
drivers/cpufreq/cpufreq_governor.h