Re: [PATCH v2] mm/page_isolation: fix a deadlock with printk()

[David Hildenbrand <david@redhat.com>]

On 10.10.19 20:06, Michal Hocko wrote:
> On Thu 10-10-19 13:48:06, Qian Cai wrote:
>> On Thu, 2019-10-10 at 19:30 +0200, Michal Hocko wrote:
>>> On Thu 10-10-19 10:47:38, Qian Cai wrote:
>>>> On Thu, 2019-10-10 at 16:18 +0200, Michal Hocko wrote:
>>>>> On Thu 10-10-19 09:11:52, Qian Cai wrote:
>>>>>> On Thu, 2019-10-10 at 12:59 +0200, Michal Hocko wrote:
>>>>>>> On Thu 10-10-19 05:01:44, Qian Cai wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>>> On Oct 9, 2019, at 12:23 PM, Michal Hocko <mhocko@kernel.org> wrote:
>>>>>>>>>
>>>>>>>>> If this was only about the memory offline code then I would agree. But
>>>>>>>>> we are talking about any printk from the zone->lock context and that is
>>>>>>>>> a bigger deal. Besides that it is quite natural that the printk code
>>>>>>>>> should be more universal and allow to be also called from the MM
>>>>>>>>> contexts as much as possible. If there is any really strong reason this
>>>>>>>>> is not possible then it should be documented at least.
>>>>>>>>
>>>>>>>> Where is the best place to document this? I am thinking about under
>>>>>>>> the “struct zone” definition’s lock field in mmzone.h.
>>>>>>>
>>>>>>> I am not sure TBH and I do not think we have reached the state where
>>>>>>> this would be the only way forward.
>>>>>>
>>>>>> How about I revised the changelog to focus on memory offline rather than making
>>>>>> a rule that nobody should call printk() with zone->lock held?
>>>>>
>>>>> If you are to remove the CONFIG_DEBUG_VM printk then I am all for it. I
>>>>> am still not convinced that fiddling with dump_page in the isolation
>>>>> code is justified though.
>>>>
>>>> No, dump_page() there has to be fixed together for memory offline to be useful.
>>>> What's the other options it has here?
>>>
>>> I would really prefer to not repeat myself
>>> http://lkml.kernel.org/r/20191010074049.GD18412@dhcp22.suse.cz
>>
>> Care to elaborate what does that mean? I am confused on if you finally agree on
>> no printk() while held zone->lock or not. You said "If there is absolutely
>> no way around that then we might have to bite a bullet and consider some
>> of MM locks a land of no printk." which makes me think you agreed, but your
>> stance from the last reply seems you were opposite to it.
> 
> I really do mean that the first step is to remove the dependency from
> the printk and remove any allocation from the console callbacks. If that
> turns out to be infeasible then we have to bite the bullet and think of
> a way to drop all printks from all locks that participate in an atomic
> allocation requests.
> 

I second that and dropping the useless printk() as Michal mentioned. I 
would beg to not uglify the offlining/isolation code with __nolock 
variants or dropping locks somewhere down in a function. If everything 
fails, I rather want to see the prinkt's gone or returning details in a 
struct back to the caller, that can print it instead.

e.g.,

struct unmovable_page_info {
	const char *reason;
	struct page *page;
...
};

You should get the idea.

-- 

Thanks,

David / dhildenb