Re: mm: deadlock between get_online_cpus/pcpu_alloc

[Vlastimil Babka]

On 02/07/2017 09:48 AM, Michal Hocko wrote:
> On Mon 06-02-17 22:05:30, Mel Gorman wrote:
>>> Unfortunately it does not seem to help.
>>
>> I'm a little stuck on how to best handle this. get_online_cpus() can
>> halt forever if the hotplug operation is holding the mutex when calling
>> pcpu_alloc. One option would be to add a try_get_online_cpus() helper which
>> trylocks the mutex. However, given that drain is so unlikely to actually
>> make that make a difference when racing against parallel allocations,
>> I think this should be acceptable.
>>
>> Any objections?
>>
>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>> index 3b93879990fd..a3192447e906 100644
>> --- a/mm/page_alloc.c
>> +++ b/mm/page_alloc.c
>> @@ -3432,7 +3432,17 @@ __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
>>  	 */
>>  	if (!page && !drained) {
>>  		unreserve_highatomic_pageblock(ac, false);
>> -		drain_all_pages(NULL);
>> +
>> +		/*
>> +		 * Only drain from contexts allocating for user allocations.
>> +		 * Kernel allocations could be holding a CPU hotplug-related
>> +		 * mutex, particularly hot-add allocating per-cpu structures
>> +		 * while hotplug-related mutex's are held which would prevent
>> +		 * get_online_cpus ever returning.
>> +		 */
>> +		if (gfp_mask & __GFP_HARDWALL)
>> +			drain_all_pages(NULL);
>> +
> 
> This wouldn't work AFAICS. If you look at the lockdep splat, the path
> which reverses the locking order (takes pcpu_alloc_mutex prior to
> cpu_hotplug.lock is bpf_array_alloc_percpu which is GFP_USER and thus
> __GFP_HARDWALL.
> 
> I believe we shouldn't pull any dependency on the hotplug locks inside
> the allocator. This is just too fragile! Can we simply drop the
> get_online_cpus()? Why do we need it, anyway? Say we are racing with the

It was added after I noticed in review that queue_work_on() has a
comment that caller must ensure that cpu can't go away, and wondered
about it. Also noted that a similar lru_add_drain_all() does it too.

> cpu offlining. I have to check the code but my impression was that WQ
> code will ignore the cpu requested by the work item when the cpu is
> going offline. If the offline happens while the worker function already
> executes then it has to wait as we run with preemption disabled so we
> should be safe here. Or am I missing something obvious?

Tejun suggested an alternative solution to avoiding get_online_cpus() in
this thread:
https://lkml.kernel.org/r/<20170123170329.GA7820@xxxxxxxxxxxxxxx>