Re: [PATCH -v2] cgroup: fix deadlock caused by cgroup_mutex and cpu_hotplug_lock

[chenridong <chenridong@huawei.com>]

On 2024/7/24 19:08, Hillf Danton wrote:
> On Fri, 19 Jul 2024 02:52:32 +0000 Chen Ridong <chenridong@huawei.com>
>> We found a hung_task problem as shown below:
>>
>> INFO: task kworker/0:0:8 blocked for more than 327 seconds.
>> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
>> task:kworker/0:0     state:D stack:13920 pid:8     ppid:2       flags:0x00004000
>> Workqueue: events cgroup_bpf_release
>> Call Trace:
>>   <TASK>
>>   __schedule+0x5a2/0x2050
>>   ? find_held_lock+0x33/0x100
>>   ? wq_worker_sleeping+0x9e/0xe0
>>   schedule+0x9f/0x180
>>   schedule_preempt_disabled+0x25/0x50
>>   __mutex_lock+0x512/0x740
>>   ? cgroup_bpf_release+0x1e/0x4d0
>>   ? cgroup_bpf_release+0xcf/0x4d0
>>   ? process_scheduled_works+0x161/0x8a0
>>   ? cgroup_bpf_release+0x1e/0x4d0
>>   ? mutex_lock_nested+0x2b/0x40
>>   ? __pfx_delay_tsc+0x10/0x10
>>   mutex_lock_nested+0x2b/0x40
>>   cgroup_bpf_release+0xcf/0x4d0
>>   ? process_scheduled_works+0x161/0x8a0
>>   ? trace_event_raw_event_workqueue_execute_start+0x64/0xd0
>>   ? process_scheduled_works+0x161/0x8a0
>>   process_scheduled_works+0x23a/0x8a0
>>   worker_thread+0x231/0x5b0
>>   ? __pfx_worker_thread+0x10/0x10
>>   kthread+0x14d/0x1c0
>>   ? __pfx_kthread+0x10/0x10
>>   ret_from_fork+0x59/0x70
>>   ? __pfx_kthread+0x10/0x10
>>   ret_from_fork_asm+0x1b/0x30
>>   </TASK>
>>
>> This issue can be reproduced by the following methods:
>> 1. A large number of cpuset cgroups are deleted.
>> 2. Set cpu on and off repeatly.
>> 3. Set watchdog_thresh repeatly.
>>
>> The reason for this issue is cgroup_mutex and cpu_hotplug_lock are
>> acquired in different tasks, which may lead to deadlock.
>> It can lead to a deadlock through the following steps:
>> 1. A large number of cgroups are deleted, which will put a large
>>     number of cgroup_bpf_release works into system_wq. The max_active
>>     of system_wq is WQ_DFL_ACTIVE(256). When cgroup_bpf_release can not
>>     get cgroup_metux, it may cram system_wq, and it will block work
>>     enqueued later.
>> 2. Setting watchdog_thresh will hold cpu_hotplug_lock.read and put
>>     smp_call_on_cpu work into system_wq. However it may be blocked by
>>     step 1.
>> 3. Cpu offline requires cpu_hotplug_lock.write, which is blocked by step 2.
>> 4. When a cpuset is deleted, cgroup release work is placed on
>>     cgroup_destroy_wq, it will hold cgroup_metux and acquire
>>     cpu_hotplug_lock.read. Acquiring cpu_hotplug_lock.read is blocked by
>>     cpu_hotplug_lock.write as mentioned by step 3. Finally, it forms a
>>     loop and leads to a deadlock.
>>
>> cgroup_destroy_wq(step4)	cpu offline(step3)		WatchDog(step2)			system_wq(step1)
>> 												......
>> 								__lockup_detector_reconfigure:
>> 								P(cpu_hotplug_lock.read)
>> 								...
>> 				...
>> 				percpu_down_write:
>> 				P(cpu_hotplug_lock.write)
>> 												...256+ works
>> 												cgroup_bpf_release:
>> 												P(cgroup_mutex)
>> 								smp_call_on_cpu:
>> 								Wait system_wq
>> ...
>> css_killed_work_fn:
>> P(cgroup_mutex)
>> ...
>> cpuset_css_offline:
>> P(cpu_hotplug_lock.read)
>>
> 	worker_thread()
> 	manage_workers()
> 	maybe_create_worker()
> 	create_worker() // has nothing to do with WQ_DFL_ACTIVE
> 	process_scheduled_works()
> 
> Given idle worker created independent of WQ_DFL_ACTIVE before handling
> work item, no deadlock could rise in your scenario above.

Hello Hillf, did you mean to say this issue couldn't happen?
I wish it hadn't happen, as it took me a long time to figure out.
However, it did happen. It could be reproduced with the method I 
offered, You can access the scripts using this link: 
https://lore.kernel.org/cgroups/e90c32d2-2a85-4f28-9154-09c7d320cb60@huawei.com/T/#t.

It's not about how the pool's workers were created, but rather the 
limit(system_wq ) of workqueue. If system_wq reaches its max_active 
limit, the work enqueued afterward will be placed on 
pwq->inactive_works. In this scenario described above, the problem is 
all active works(cgroup_bpf_release) are blocked, and the 
inactive_works(smp_call_on_cpu) couldn't be executed when it forms a loop.

We have discussed this before in V1, you can find in the Link.

Thanks
Ridong