[LTP] [PATCH] memcg: Account for pages in the per-cpu cache

[Richard Palethorpe]

Hello,

Merged, thanks!

As discussed on bz, it may be better to drain the stocks
instead. However there is no direct way to do it.

My idea would be to find the batch size emperically, but it requires a
rewrite of the test IMO. So let's see if this works for now.

Petr Vorel <pvorel@suse.cz> writes:

> Hi Richie,
>
> [ Cc Li ]
>
>> PAGESIZES is one greater than the batch size for charging pages from
>> the per CPU cache. So with MEM_TO_ALLOC=2*PAGESIZES we have two pages
>> that are below the threshold for being charged.
>
>> Sometimes something triggers a flush and the pages get charged to the
>> global counter anyway and the test passes. We have seen cases where
>> the test times out waiting for this to happen.
>
>> So this patch sets a lower bound to allow those cases to pass. It'll
>> probably speed the test up as well.
>
> LGTM.
> Reviewed-by: Petr Vorel <pvorel@suse.cz>
>
> Kind regards,
> Petr
>
>> Signed-off-by: Richard Palethorpe <rpalethorpe@suse.com>
>> ---
>>  .../controllers/memcg/functional/memcg_subgroup_charge.sh    | 5 +++--
>>  1 file changed, 3 insertions(+), 2 deletions(-)
>
>> diff --git a/testcases/kernel/controllers/memcg/functional/memcg_subgroup_charge.sh b/testcases/kernel/controllers/memcg/functional/memcg_subgroup_charge.sh
>> index 9bcc01258..3b7311422 100755
>> --- a/testcases/kernel/controllers/memcg/functional/memcg_subgroup_charge.sh
>> +++ b/testcases/kernel/controllers/memcg/functional/memcg_subgroup_charge.sh
>> @@ -33,8 +33,8 @@ test_subgroup()
>>  	fi
>
>>  	echo $MEMCG_PROCESS_PID > tasks
>> -	signal_memcg_process $MEM_TO_ALLOC
>> -	check_mem_stat "rss" $MEM_TO_ALLOC
>> +	signal_memcg_process $MIN_CHARGED
>> +	check_mem_stat "rss" $MIN_CHARGED $MEM_TO_ALLOC
>
>>  	cd subgroup
>>  	echo $MEMCG_PROCESS_PID > tasks
>> @@ -66,5 +66,6 @@ test3()
>
>>  # Allocate memory bigger than per-cpu kernel memory
>>  MEM_TO_ALLOC=$((PAGESIZES * 2))
>> +MIN_CHARGED=$((2 * (PAGESIZES - 1)))
>
>>  tst_run


-- 
Thank you,
Richard.