cpufreq: Fix sysfs deadlock with concurrent hotplug/frequency switch
authorStephen Boyd <sboyd@codeaurora.org>
Fri, 20 Jul 2012 18:14:38 +0000 (18:14 +0000)
committerRafael J. Wysocki <rjw@sisk.pl>
Fri, 20 Jul 2012 19:39:25 +0000 (21:39 +0200)
commita9144436271583115a2230db15d0b6ae2c481d3c
tree89817e96e6e0f863f5eb1972609650753140ef58
parent53df1ad52545854fc34d336b26f3086b2fb2d6f7
cpufreq: Fix sysfs deadlock with concurrent hotplug/frequency switch

Running one program that continuously hotplugs and replugs a cpu
concurrently with another program that continuously writes to the
scaling_setspeed node eventually deadlocks with:

=============================================
[ INFO: possible recursive locking detected ]
3.4.0 #37 Tainted: G        W
---------------------------------------------
filemonkey/122 is trying to acquire lock:
 (s_active#13){++++.+}, at: [<c01a3d28>] sysfs_remove_dir+0x9c/0xb4

but task is already holding lock:
 (s_active#13){++++.+}, at: [<c01a22f0>] sysfs_write_file+0xe8/0x140

other info that might help us debug this:
 Possible unsafe locking scenario:

       CPU0
       ----
  lock(s_active#13);
  lock(s_active#13);

 *** DEADLOCK ***

 May be due to missing lock nesting notation

2 locks held by filemonkey/122:
 #0:  (&buffer->mutex){+.+.+.}, at: [<c01a2230>] sysfs_write_file+0x28/0x140
 #1:  (s_active#13){++++.+}, at: [<c01a22f0>] sysfs_write_file+0xe8/0x140

stack backtrace:
[<c0014fcc>] (unwind_backtrace+0x0/0x120) from [<c00ca600>] (validate_chain+0x6f8/0x1054)
[<c00ca600>] (validate_chain+0x6f8/0x1054) from [<c00cb778>] (__lock_acquire+0x81c/0x8d8)
[<c00cb778>] (__lock_acquire+0x81c/0x8d8) from [<c00cb9c0>] (lock_acquire+0x18c/0x1e8)
[<c00cb9c0>] (lock_acquire+0x18c/0x1e8) from [<c01a3ba8>] (sysfs_addrm_finish+0xd0/0x180)
[<c01a3ba8>] (sysfs_addrm_finish+0xd0/0x180) from [<c01a3d28>] (sysfs_remove_dir+0x9c/0xb4)
[<c01a3d28>] (sysfs_remove_dir+0x9c/0xb4) from [<c02d0e5c>] (kobject_del+0x10/0x38)
[<c02d0e5c>] (kobject_del+0x10/0x38) from [<c02d0f74>] (kobject_release+0xf0/0x194)
[<c02d0f74>] (kobject_release+0xf0/0x194) from [<c0565a98>] (cpufreq_cpu_put+0xc/0x24)
[<c0565a98>] (cpufreq_cpu_put+0xc/0x24) from [<c05683f0>] (store+0x6c/0x74)
[<c05683f0>] (store+0x6c/0x74) from [<c01a2314>] (sysfs_write_file+0x10c/0x140)
[<c01a2314>] (sysfs_write_file+0x10c/0x140) from [<c014af44>] (vfs_write+0xb0/0x128)
[<c014af44>] (vfs_write+0xb0/0x128) from [<c014b06c>] (sys_write+0x3c/0x68)
[<c014b06c>] (sys_write+0x3c/0x68) from [<c000e0e0>] (ret_fast_syscall+0x0/0x3c)

This is because store() in cpufreq.c indirectly calls
kobject_get() via cpufreq_cpu_get() and is the last one to call
kobject_put() via cpufreq_cpu_put(). Sysfs code should not call
kobject_get() or kobject_put() directly (see the comment around
sysfs_schedule_callback() for more information).

Fix this deadlock by introducing two new functions:

struct cpufreq_policy *cpufreq_cpu_get_sysfs(unsigned int cpu)
void cpufreq_cpu_put_sysfs(struct cpufreq_policy *data)

which do the same thing as cpufreq_cpu_{get,put}() but don't call
kobject functions.

To easily trigger this deadlock you can insert an msleep() with a
reasonably large value right after the fail label at the bottom
of the store() function in cpufreq.c and then write
scaling_setspeed in one task and offline the cpu in another. The
first task will hang and be detected by the hung task detector.

Signed-off-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
drivers/cpufreq/cpufreq.c