Re: [RFC PATCH 1/2] perf_events: add support for per-cpu per-cgroupmonitoring (v4)

[stephane eranian]

On Wed, Oct 13, 2010 at 4:26 AM, Li Zefan <lizf@xxxxxxxxxxxxxx> wrote:
> (Sorry for the late reply. I've been keeping busy..)
>
> Stephane Eranian wrote:
>> On Fri, Oct 8, 2010 at 2:46 AM, Li Zefan <lizf@xxxxxxxxxxxxxx> wrote:
>>>>>>> +#ifdef CONFIG_CGROUPS
>>>>>>> +struct perf_cgroup_time {
>>>>>>> + Â Â u64 time;
>>>>>>> + Â Â u64 timestamp;
>>>>>>> +};
>>>>>>> +
>>>>>>> +struct perf_cgroup {
>>>>>>> + Â Â struct cgroup_subsys_state css;
>>>>>>> + Â Â struct perf_cgroup_time *time;
>>>>>>> +};
>>>>>> Can we avoid adding this perf cgroup subsystem? It has 2 disavantages:
>>>>>>
>>>>> Well, I need to maintain some timing information for each cgroup. This has
>>>>> to be stored somewhere.
>>>>>
>>> Seems you can simply store it in struct perf_event?
>>>
>> No, timing has to be shared by events monitoring the same cgroup at
>> the same time.
>> Works like a timestamp. It needs to be centralized for all events
>> attached to the same cgroup.
>>
>
> I no little about internel perf code, so I don't know if we can store
> this somewhere in perf. The last resort could be store it in struct cgroup.
>
>>>>>> - If one mounted cgroup fs without perf cgroup subsys, he can't monitor it.
>>>>> That's unfortunately true ;-)
>>>>>
>>>>>> - If there are several different cgroup mount points, only one can be
>>>>>> Âmonitored.
>>>>>>
>>>>>> To choose which cgroup hierarchy to monitor, hierarchy id can be passed
>>>>>> from userspace, which is the 2nd column below:
>>>>>>
>>>>> Ok, I will investigate this. As long as the hierarchy id is unique AND it can be
>>>>> searched, then we can use it. Using /proc is fine with me.
>>>>>
>>>>>> $ cat /proc/cgroups
>>>>>> #subsys_name  Âhierarchy    num_cgroups   enabled
>>>>>> debug  0    1    1
>>>>>> net_cls 0 Â Â Â 1 Â Â Â 1
>>>>>>
>>>> If I mount all subsystems:
>>>> mount -t cgroup none /dev/cgroup
>>>> Then, I get:
>>>> #subsys_name Âhierarchy    num_cgroups   enabled
>>>> cpuset    Â1    1    1
>>>> cpu      1    1    1
>>>> perf_event  Â1    1    1
>>>>
>>>> In other words, the hierarchy id is not unique.
>>>> If the perf_event is not mounted, then hierarchy id = 0.
>>>>
>>> Yes, it's unique. ;)
>>>
>>> You mounted them together, and that's a cgroup hierarchy, so
>>> they have the same hierarchy id.
>>>
>>> If you mount them seperately:
>>>
>>> # mount -t cgroup -o debug xxx /cgroup1
>>> # mount -t cgroup -o net_cls xxx /cgroup2/
>>> # cat /proc/cgroups
>>> #subsys_name  Âhierarchy    num_cgroups   enabled
>>> debug  1    1    1
>>> net_cls 2 Â Â Â 1 Â Â Â 1
>>>
>> Ok, but if you mount perf_event twice, you get the
>> same hierarchy id for it:
>>
>> # mount -t cgroup -operf_event none /cgroup
>> # cat /proc/cgroups
>> #subsys_name Âhierarchy    num_cgroups   enabled
>> cpuset    Â0    1    1
>> cpu      0    1    1
>> perf_event  Â1    1    1
>>
>> # mount -t cgroup -operf_event none /cgroup2
>> # cat /proc/cgroups
>> #subsys_name Âhierarchy    num_cgroups   enabled
>> cpuset    Â0    1    1
>> cpu      0    1    1
>> perf_event  Â1    1    1
>>
>> It does not seem like I can mount the same subsystem
>> twice with difference hierarchies:
>>
>> # umount /cgroup2
>> # mount -t cgroup -operf_event,cpuset none /cgroup2
>> mount: none already mounted or /cgroup2 busy
>> # mount -t cgroup Ânone /cgroup2
>> mount: none already mounted or /cgroup2 busy
>>
>>> They now have different hierarchy id, because they belong
>>> to different cgroup hierarchy.
>>>
>>> So pid + hierarchy_id locates the cgroup.
>>>
>>
>> I cannot do task's pid + cgroup hierarchy_id. It's one or the
>> other.
>>
>
> I've looked into the patch again, and I see you pass the fd from
> userspace, so you don't need hierarchy_id.
>
True.

> And to get rid of perf_cgroup subsys, seems you just need to find
> another place to store the time info, somewhere inside perf code
> or in struct cgroup.
>
Something I may have missed since the beginning of our conversation
is why do you think definition perf_cgroup subsys is wrong or useless.
What kind of problem does it introduce. I think it is fine to reject cgroup
mode if the perf cgroup is not mounted.

The other key point is that from a task (on context switch), I need to identify
the perf_cgroup subsys that it corresponds to. How would I have such link
if I don't leverage the existing cgroup infrastructure?
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/