[LTP] [RFC PATCH 2/5] CGroup API rewrite

[Richard Palethorpe]

Hello Li,

Li Wang <liwang@redhat.com> writes:

> Hi Richard,
>
> Richard Palethorpe <rpalethorpe@suse.de> wrote:
>
>> ...
>> >> > As I was mentioned in 0/5 that maybe we should create test_cgroup_dir
>> >> > for different progress so that the test could use the same controller
>> >> with
>> >> > various configurations in parallel.
>> >> >
>> >> > e.g. child_1 set SIZE to memory.limit_in_bytes
>> >> >        child_2 set SIZE/2 to memory.limit_in_bytes
>> >> >
>> >> > Any possibility to move this to tst_cgroup_move_current?
>> >>
>> >> Yes I suppose we can try this. Is there a test which already requires
>> it?
>> >>
>> >
>> > So far we don't have such a test, I remember that in the previous Lib is
>> > also to keep expandability.
>>
>> I think we have at least two problems:
>>
>> 1) This allows many CGroups to be created for each test and we must
>>    clean them up, adding some complication.
>>
>> 2) It's not clear if a future test will require the CGroup hierarchy to
>>    be the same as the process hierarchy or different. Depending what
>>    behaviour for tst_cgroup_move_current we choose it will make some
>>    configurations impossible.
>>
>> So if we add this then it will add complexity, but I am not sure it will
>> help in the future. If we make it flexible enough to support any
>> hierarchy this will add a lot of complication.
>>
>
> Okay. I do NOT insist to implement a future feature at this early
> moment, but do u think we'd better mention this in documents?
> To let people(avoid abusing it) know that the current CGroup lib
> hasn't supported children using the same controller in parallel?

Yes, this is a good point.

>
> And another new query is, do we really need to export many
> cleanup-levels to users? I guess we can handle it in the library
> with intelligently choose levels no matter in parallel or sequential.

No, I should remove it. I was using it for debugging, but there is
probably a simpler way.

>
> i.e. As now only one directory(test-pid/) created for one test in a
> hierarchy, how about going with TST_CGROUP_CLEANUP_ROOT
> level by default, unless it detects more or equal to two directories
> (that probably means parallel) goes TST_CGROUP_CLEANUP_TEST?

That is likely to result in race conditions. However it won't cleanup
any CGroups which already exist (except the test CGroup). So the user
can pre-create the LTP CGroup before running tests in parallel.

>
>
>>
>> Also if we did need this feature in the future, then we can add some new
>> functions which take a sub-group (or hierarchy) parameter. e.g.
>>
>> void tst_cgroup_move(enum tst_cgroup_ctrl type, pid_t pid,
>>                      struct tst_cgroup_tree *path);
>>
>
> Sounds good to me.
>
>
>>
>> Alternate versions of other functions would also need to be added. Also
>> some changes to the internal data structures may be needed. However it
>> would keep the current API functions simple.
>>
>> So until we have a test which requires this, I think the best option is
>> to do nothing :-)
>>
>> --
>> Thank you,
>> Richard.
>>
>>


-- 
Thank you,
Richard.