On 2024/9/9 22:19, Michal Koutný wrote:

On Sat, Aug 17, 2024 at 09:33:34AM GMT, Chen Ridong [off-list ref] wrote:

quoted

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 cpusets are deleted asynchronously, which puts a
    large number of cgroup_bpf_release works into system_wq. The max_active
    of system_wq is WQ_DFL_ACTIVE(256). Consequently, all active works are
    cgroup_bpf_release works, and many cgroup_bpf_release works will be put
    into inactive queue. As illustrated in the diagram, there are 256 (in
    the acvtive queue) + n (in the inactive queue) works.
2. Setting watchdog_thresh will hold cpu_hotplug_lock.read and put
    smp_call_on_cpu work into system_wq. However step 1 has already filled
    system_wq, 'sscs.work' is put into inactive queue. 'sscs.work' has
    to wait until the works that were put into the inacvtive queue earlier
    have executed (n cgroup_bpf_release), so it will be blocked for a while.
3. Cpu offline requires cpu_hotplug_lock.write, which is blocked by step 2.
4. Cpusets that were deleted at step 1 put cgroup_release works into
    cgroup_destroy_wq. They are competing to get cgroup_mutex all the time.
    When cgroup_metux is acqured by work at css_killed_work_fn, it will
    call cpuset_css_offline, which needs to acqure cpu_hotplug_lock.read.
    However, cpuset_css_offline will be blocked for step 3.
5. At this moment, there are 256 works in active queue that are
    cgroup_bpf_release, they are attempting to acquire cgroup_mutex, and as
    a result, all of them are blocked. Consequently, sscs.work can not be
    executed. Ultimately, this situation leads to four processes being
    blocked, forming a deadlock.

system_wq(step1)		WatchDog(step2)			cpu offline(step3)	cgroup_destroy_wq(step4)
...
2000+ cgroups deleted asyn
256 actives + n inactives
				__lockup_detector_reconfigure
				P(cpu_hotplug_lock.read)
				put sscs.work into system_wq
256 + n + 1(sscs.work)
sscs.work wait to be executed
				warting sscs.work finish
								percpu_down_write
								P(cpu_hotplug_lock.write)
								...blocking...
											css_killed_work_fn
											P(cgroup_mutex)
											cpuset_css_offline
											P(cpu_hotplug_lock.read)
											...blocking...
256 cgroup_bpf_release
mutex_lock(&cgroup_mutex);
..blocking...

Thanks, Ridong, for laying this out.
Let me try to extract the core of the deps above.

The correct lock ordering is: cgroup_mutex then cpu_hotplug_lock.
However, the smp_call_on_cpu() under cpus_read_lock may lead to
a deadlock (ABBA over those two locks).

That's right.

This is OK
	thread T					system_wq worker
	
	  						lock(cgroup_mutex) (II)
							...
							unlock(cgroup_mutex)
	down(cpu_hotplug_lock.read)
	smp_call_on_cpu
	  queue_work_on(cpu, system_wq, scss) (I)
							scss.func
	  wait_for_completion(scss)
	up(cpu_hotplug_lock.read)

However, there is no ordering between (I) and (II) so they can also happen
in opposite

	thread T					system_wq worker
	
	down(cpu_hotplug_lock.read)
	smp_call_on_cpu
	  queue_work_on(cpu, system_wq, scss) (I)
	  						lock(cgroup_mutex)  (II)
							...
							unlock(cgroup_mutex)
							scss.func
	  wait_for_completion(scss)
	up(cpu_hotplug_lock.read)

And here the thread T + system_wq worker effectively call
cpu_hotplug_lock and cgroup_mutex in the wrong order. (And since they're
two threads, it won't be caught by lockdep.)

By that reasoning any holder of cgroup_mutex on system_wq makes system
susceptible to a deadlock (in presence of cpu_hotplug_lock waiting
writers + cpuset operations). And the two work items must meet in same
worker's processing hence probability is low (zero?) with less than
WQ_DFL_ACTIVE items.

(And more generally, any lock that is ordered before cpu_hotplug_lock
should not be taken in system_wq work functions. Or at least such works
items should not saturate WQ_DFL_ACTIVE workers.)

Wrt other uses of cgroup_mutex, I only see
   bpf_map_free_in_work
     queue_work(system_unbound_wq)
       bpf_map_free_deferred
         ops->map_free == cgroup_storage_map_free
           cgroup_lock()
which is safe since it uses a different workqueue than system_wq.

quoted

To fix the problem, place cgroup_bpf_release works on cgroup_destroy_wq,
which can break the loop and solve the problem.

Yes, it moves the problematic cgroup_mutex holder away from system_wq
and cgroup_destroy_wq could not cause similar problems because there are
no explicit waiter for particular work items or its flushing.

quoted

System wqs are for misc things which shouldn't create a large number
of concurrent work items.  If something is going to generate

quoted

WQ_DFL_ACTIVE(256) concurrent work

items, it should use its own dedicated workqueue.

Actually, I'm not sure (because I lack workqueue knowledge) if producing
less than WQ_DFL_ACTIVE work items completely eliminates the chance of
two offending work items producing the wrong lock ordering.

If producing less than WQ_DFL_ACTIVE work items, it won't lead to a 
deadlock. Because scss.func can be executed and doesn't have to wait for 
work that holds cgroup_mutex to be completed. Therefore, the probability 
is low and this issue can only be reproduced under pressure test.

quoted

Fixes: 4bfc0bb2c60e ("bpf: decouple the lifetime of cgroup_bpf from cgroup itself")

I'm now indifferent whether this is needed (perhaps in the sense it is
the _latest_ of multiple changes that contributed to possibility of this
deadlock scenario).

quoted

Link: https://lore.kernel.org/cgroups/e90c32d2-2a85-4f28-9154-09c7d320cb60@huawei.com/T/#t (local)
Signed-off-by: Chen Ridong <redacted>
---
  kernel/bpf/cgroup.c             | 2 +-
  kernel/cgroup/cgroup-internal.h | 1 +
  kernel/cgroup/cgroup.c          | 2 +-
  3 files changed, 3 insertions(+), 2 deletions(-)

I have convinved myself now that you can put

Reviewed-by: Michal Koutný <mkoutny@suse.com>

Regards,
Michal

Thank you very much.

Best Regards,
Ridong