[Devel] [PATCH rh7 v4] mm/memcg: fix cache growth above cache.limit_in_bytes

[Andrey Ryabinin]

On 7/30/20 6:52 PM, Evgenii Shatokhin wrote:
> Hi,
> 
> On 30.07.2020 18:02, Andrey Ryabinin wrote:
>> Exceeding cache above cache.limit_in_bytes schedules high_work_func()
>> which tries to reclaim 32 pages. If cache generated fast enough or it allows
>> cgroup to steadily grow above cache.limit_in_bytes because we don't reclaim
>> enough. Try to reclaim exceeded amount of cache instead.
>>
>> https://jira.sw.ru/browse/PSBM-106384
>> Signed-off-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
>> ---
>>
>>   - Changes since v1: add bug link to changelog
>>   - Changes since v2: Fix cache_overused check (We should check if it's positive).
>>      Made this stupid bug during cleanup, patch was tested without bogus cleanup,
>>      so it shoud work.
>>   - Chnages since v3: Compilation fixes, properly tested now.
>>
>>   mm/memcontrol.c | 10 +++++++---
>>   1 file changed, 7 insertions(+), 3 deletions(-)
>>
>> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
>> index 3cf200f506c3..16cbd451a588 100644
>> --- a/mm/memcontrol.c
>> +++ b/mm/memcontrol.c
>> @@ -3080,12 +3080,16 @@ static void reclaim_high(struct mem_cgroup *memcg,
>>   {
>>         do {
>> +        long cache_overused;
>> +
>>           if (page_counter_read(&memcg->memory) > memcg->high)
>>               try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, 0);
>>   -        if (page_counter_read(&memcg->cache) > memcg->cache.limit)
>> -            try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask,
>> -                        MEM_CGROUP_RECLAIM_NOSWAP);
>> +        cache_overused = page_counter_read(&memcg->cache) -
>> +            memcg->cache.limit;
> 
> If cache_overused is less than 32 pages, the kernel would try to reclaim less than before the patch. It it OK, or should it try to reclaim at least 32 pages?

It's ok, try_to_free_mem_cgroup_pages will increase it:

unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
					   unsigned long nr_pages,
					   gfp_t gfp_mask,
					   int flags)

....
		.nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),