Re: [PATCH v4 29/35] mm: slub: Move flush_cpu_slab() invocations __free_slab() invocations out of IRQ context

[Qian Cai]

On 8/9/2021 4:08 PM, Vlastimil Babka wrote:
> On 8/9/2021 8:44 PM, Mike Galbraith wrote:
>> On Mon, 2021-08-09 at 09:41 -0400, Qian Cai wrote:
>>>
>>>
>>> On 8/5/2021 11:19 AM, Vlastimil Babka wrote:
>>>>
>>>>  
>>>> +static DEFINE_MUTEX(flush_lock);
>>>> +static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
>>>> +
>>>>  static void flush_all(struct kmem_cache *s)
>>>>  {
>>>> -       on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
>>>> +       struct slub_flush_work *sfw;
>>>> +       unsigned int cpu;
>>>> +
>>>> +       mutex_lock(&flush_lock);
>>>
>>> Vlastimil, taking the lock here could trigger a warning during memory
>>> offline/online due to the locking order:
>>>
>>> slab_mutex -> flush_lock
>>
>> Bugger.  That chain ending with cpu_hotplug_lock makes slub_cpu_dead()
>> taking slab_mutex a non-starter for cpu hotplug as well.  It's
>> established early by kernel_init_freeable()..kmem_cache_destroy() as
>> well as by slab_mem_going_offline_callback().
> 
> I suck at reading the lockdep splats, so I don't see yet how the "existing
> reverse order" occurs - I do understand the order in the "lsbug".
> What I also wonder is why didn't this occur also in the older RT trees with this
> patch. I did change the order of locks in flush_all() to take flush_lock first
> and cpus_read_lock() second, as Cyrill Gorcunov suggested. Would the original
> order prevent this? Or we would fail anyway because we already took
> cpus_read_lock() in offline_pages() and now are taking it again - do these nest
> or not?

"lsbug" is just an user-space tool running workloads like memory offline/online
via sysfs. The splat indicated that the existing locking orders on the running
system saw so far are:

flush_lock -> cpu_hotplug_lock (in #1)
  cpu_hotplug_lock -> pck_batch_high_lock (in #2)
    pcp_batch_high_lock -> (memory_chain).rwsem (in #3)
      (memory_chain).rwsem -> slab_mutex (in #4)

Thus, lockdep inferences that taking flush_lock first could later reaching
slab_mutex. Then, in the commit, memory offline (in #0) started to take the locking
order slab_mutex -> flush_lock. Thus, the potential deadlock warning.