Re: [PATCH RFC] mm/madvise: implement MADV_STOCKPILE (kswapd from user space)

[Konstantin Khlebnikov <khlebnikov@yandex-team.ru>]

On 28.05.2019 10:38, Michal Hocko wrote:
> On Tue 28-05-19 10:30:12, Konstantin Khlebnikov wrote:
>> On 28.05.2019 9:51, Michal Hocko wrote:
>>> On Tue 28-05-19 09:25:13, Konstantin Khlebnikov wrote:
>>>> On 27.05.2019 17:39, Michal Hocko wrote:
>>>>> On Mon 27-05-19 16:21:56, Michal Hocko wrote:
>>>>>> On Mon 27-05-19 16:12:23, Michal Hocko wrote:
>>>>>>> [Cc linux-api. Please always cc this list when proposing a new user
>>>>>>>     visible api. Keeping the rest of the email intact for reference]
>>>>>>>
>>>>>>> On Mon 27-05-19 13:05:58, Konstantin Khlebnikov wrote:
>>>>>> [...]
>>>>>>>> This implements manual kswapd-style memory reclaim initiated by userspace.
>>>>>>>> It reclaims both physical memory and cgroup pages. It works in context of
>>>>>>>> task who calls syscall madvise thus cpu time is accounted correctly.
>>>>>>
>>>>>> I do not follow. Does this mean that the madvise always reclaims from
>>>>>> the memcg the process is member of?
>>>>>
>>>>> OK, I've had a quick look at the implementation (the semantic should be
>>>>> clear from the patch descrition btw.) and it goes all the way up the
>>>>> hierarchy and finally try to impose the same limit to the global state.
>>>>> This doesn't really make much sense to me. For few reasons.
>>>>>
>>>>> First of all it breaks isolation where one subgroup can influence a
>>>>> different hierarchy via parent reclaim.
>>>>
>>>> madvise(NULL, size, MADV_STOCKPILE) is the same as memory allocation and
>>>> freeing immediately, but without pinning memory and provoking oom.
>>>>
>>>> So, there is shouldn't be any isolation or security issues.
>>>>
>>>> At least probably it should be limited with portion of limit (like half)
>>>> instead of whole limit as it does now.
>>>
>>> I do not think so. If a process is running inside a memcg then it is
>>> a subject of a limit and that implies an isolation. What you are
>>> proposing here is to allow escaping that restriction unless I am missing
>>> something. Just consider the following setup
>>>
>>> 		root (total memory = 2G)
>>> 		 / \
>>>              (1G) A   B (1G)
>>>                      / \
>>>              (500M) C   D (500M)
>>>
>>> all of them used up close to the limit and a process inside D requests
>>> shrinking to 250M. Unless I am misunderstanding this implementation
>>> will shrink D, B root to 250M (which means reclaiming C and A as well)
>>> and then globally if that was not sufficient. So you have allowed D to
>>> "allocate" 1,75G of memory effectively, right?
>>
>> It shrinks not 'size' memory - only while usage + size > limit.
>> So, after reclaiming 250M in D all other levels will have 250M free.
> 
> Could you define the exact semantic? Ideally something for the manual
> page please?
> 

Like kswapd which works with thresholds of free memory this one reclaims
until 'free' (i.e. memory which could be allocated without invoking
direct recliam of any kind) is lower than passed 'size' argument.

Thus right after madvise(NULL, size, MADV_STOCKPILE) 'size' bytes
could be allocated in this memory cgroup without extra latency from
reclaimer if there is no other memory consumers.

Reclaimed memory is simply put into free lists in common buddy allocator,
there is no reserves for particular task or cgroup.

If overall memory allocation rate is smooth without rough spikes then
calling MADV_STOCKPILE in loop periodically provides enough room for
allocations and eliminates direct reclaim from all other tasks.
As a result this eliminates unpredictable delays caused by
direct reclaim in random places.