diff options
author | Tejun Heo <tj@kernel.org> | 2014-06-30 23:47:32 +0400 |
---|---|---|
committer | Tejun Heo <tj@kernel.org> | 2014-07-02 00:42:28 +0400 |
commit | 76bb5ab8f6e3e7bebdcefec4146ff305e7d0b465 (patch) | |
tree | 58eeef9dbb6e8f5400407b5000b04e4d75f9dbad /kernel | |
parent | 3a32bd72d77058d768dbb38183ad517f720dd1bc (diff) | |
download | linux-76bb5ab8f6e3e7bebdcefec4146ff305e7d0b465.tar.xz |
cpuset: break kernfs active protection in cpuset_write_resmask()
Writing to either "cpuset.cpus" or "cpuset.mems" file flushes
cpuset_hotplug_work so that cpu or memory hotunplug doesn't end up
migrating tasks off a cpuset after new resources are added to it.
As cpuset_hotplug_work calls into cgroup core via
cgroup_transfer_tasks(), this flushing adds the dependency to cgroup
core locking from cpuset_write_resmak(). This used to be okay because
cgroup interface files were protected by a different mutex; however,
8353da1f91f1 ("cgroup: remove cgroup_tree_mutex") simplified the
cgroup core locking and this dependency became a deadlock hazard -
cgroup file removal performed under cgroup core lock tries to drain
on-going file operation which is trying to flush cpuset_hotplug_work
blocked on the same cgroup core lock.
The locking simplification was done because kernfs added an a lot
easier way to deal with circular dependencies involving kernfs active
protection. Let's use the same strategy in cpuset and break active
protection in cpuset_write_resmask(). While it isn't the prettiest,
this is a very rare, likely unique, situation which also goes away on
the unified hierarchy.
The commands to trigger the deadlock warning without the patch and the
lockdep output follow.
localhost:/ # mount -t cgroup -o cpuset xxx /cpuset
localhost:/ # mkdir /cpuset/tmp
localhost:/ # echo 1 > /cpuset/tmp/cpuset.cpus
localhost:/ # echo 0 > cpuset/tmp/cpuset.mems
localhost:/ # echo $$ > /cpuset/tmp/tasks
localhost:/ # echo 0 > /sys/devices/system/cpu/cpu1/online
======================================================
[ INFO: possible circular locking dependency detected ]
3.16.0-rc1-0.1-default+ #7 Not tainted
-------------------------------------------------------
kworker/1:0/32649 is trying to acquire lock:
(cgroup_mutex){+.+.+.}, at: [<ffffffff8110e3d7>] cgroup_transfer_tasks+0x37/0x150
but task is already holding lock:
(cpuset_hotplug_work){+.+...}, at: [<ffffffff81085412>] process_one_work+0x192/0x520
which lock already depends on the new lock.
the existing dependency chain (in reverse order) is:
-> #2 (cpuset_hotplug_work){+.+...}:
...
-> #1 (s_active#175){++++.+}:
...
-> #0 (cgroup_mutex){+.+.+.}:
...
other info that might help us debug this:
Chain exists of:
cgroup_mutex --> s_active#175 --> cpuset_hotplug_work
Possible unsafe locking scenario:
CPU0 CPU1
---- ----
lock(cpuset_hotplug_work);
lock(s_active#175);
lock(cpuset_hotplug_work);
lock(cgroup_mutex);
*** DEADLOCK ***
2 locks held by kworker/1:0/32649:
#0: ("events"){.+.+.+}, at: [<ffffffff81085412>] process_one_work+0x192/0x520
#1: (cpuset_hotplug_work){+.+...}, at: [<ffffffff81085412>] process_one_work+0x192/0x520
stack backtrace:
CPU: 1 PID: 32649 Comm: kworker/1:0 Not tainted 3.16.0-rc1-0.1-default+ #7
...
Call Trace:
[<ffffffff815a5f78>] dump_stack+0x72/0x8a
[<ffffffff810c263f>] print_circular_bug+0x10f/0x120
[<ffffffff810c481e>] check_prev_add+0x43e/0x4b0
[<ffffffff810c4ee6>] validate_chain+0x656/0x7c0
[<ffffffff810c53d2>] __lock_acquire+0x382/0x660
[<ffffffff810c57a9>] lock_acquire+0xf9/0x170
[<ffffffff815aa13f>] mutex_lock_nested+0x6f/0x380
[<ffffffff8110e3d7>] cgroup_transfer_tasks+0x37/0x150
[<ffffffff811129c0>] hotplug_update_tasks_insane+0x110/0x1d0
[<ffffffff81112bbd>] cpuset_hotplug_update_tasks+0x13d/0x180
[<ffffffff811148ec>] cpuset_hotplug_workfn+0x18c/0x630
[<ffffffff810854d4>] process_one_work+0x254/0x520
[<ffffffff810875dd>] worker_thread+0x13d/0x3d0
[<ffffffff8108e0c8>] kthread+0xf8/0x100
[<ffffffff815acaec>] ret_from_fork+0x7c/0xb0
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Li Zefan <lizefan@huawei.com>
Tested-by: Li Zefan <lizefan@huawei.com>
Diffstat (limited to 'kernel')
-rw-r--r-- | kernel/cpuset.c | 12 |
1 files changed, 12 insertions, 0 deletions
diff --git a/kernel/cpuset.c b/kernel/cpuset.c index d3df02e76643..116a4164720a 100644 --- a/kernel/cpuset.c +++ b/kernel/cpuset.c @@ -1623,7 +1623,17 @@ static ssize_t cpuset_write_resmask(struct kernfs_open_file *of, * resources, wait for the previously scheduled operations before * proceeding, so that we don't end up keep removing tasks added * after execution capability is restored. + * + * cpuset_hotplug_work calls back into cgroup core via + * cgroup_transfer_tasks() and waiting for it from a cgroupfs + * operation like this one can lead to a deadlock through kernfs + * active_ref protection. Let's break the protection. Losing the + * protection is okay as we check whether @cs is online after + * grabbing cpuset_mutex anyway. This only happens on the legacy + * hierarchies. */ + css_get(&cs->css); + kernfs_break_active_protection(of->kn); flush_work(&cpuset_hotplug_work); mutex_lock(&cpuset_mutex); @@ -1651,6 +1661,8 @@ static ssize_t cpuset_write_resmask(struct kernfs_open_file *of, free_trial_cpuset(trialcs); out_unlock: mutex_unlock(&cpuset_mutex); + kernfs_unbreak_active_protection(of->kn); + css_put(&cs->css); return retval ?: nbytes; } |