Re: [PATCH v4 7/8] Display current tcp memory allocation in kmem cgroup

[Glauber Costa <glommer@parallels.com>]

On 10/03/2011 04:36 PM, Kirill A. Shutemov wrote:
> On Mon, Oct 03, 2011 at 04:26:41PM +0400, Glauber Costa wrote:
>> On 10/03/2011 04:25 PM, Kirill A. Shutemov wrote:
>>> On Mon, Oct 03, 2011 at 04:19:18PM +0400, Glauber Costa wrote:
>>>> On 10/03/2011 04:14 PM, Kirill A. Shutemov wrote:
>>>>> On Mon, Oct 03, 2011 at 02:18:42PM +0400, Glauber Costa wrote:
>>>>>> This patch introduces kmem.tcp_current_memory file, living in the
>>>>>> kmem_cgroup filesystem. It is a simple read-only file that displays the
>>>>>> amount of kernel memory currently consumed by the cgroup.
>>>>>>
>>>>>> Signed-off-by: Glauber Costa<glommer@parallels.com>
>>>>>> CC: David S. Miller<davem@davemloft.net>
>>>>>> CC: Hiroyouki Kamezawa<kamezawa.hiroyu@jp.fujitsu.com>
>>>>>> CC: Eric W. Biederman<ebiederm@xmission.com>
>>>>>> ---
>>>>>>     Documentation/cgroups/memory.txt |    1 +
>>>>>>     mm/memcontrol.c                  |   11 +++++++++++
>>>>>>     2 files changed, 12 insertions(+), 0 deletions(-)
>>>>>>
>>>>>> diff --git a/Documentation/cgroups/memory.txt b/Documentation/cgroups/memory.txt
>>>>>> index 1ffde3e..f5a539d 100644
>>>>>> --- a/Documentation/cgroups/memory.txt
>>>>>> +++ b/Documentation/cgroups/memory.txt
>>>>>> @@ -79,6 +79,7 @@ Brief summary of control files.
>>>>>>      memory.independent_kmem_limit	 # select whether or not kernel memory limits are
>>>>>>     				   independent of user limits
>>>>>>      memory.kmem.tcp.max_memory      # set/show hard limit for tcp buf memory
>>>>>> + memory.kmem.tcp.current_memory  # show current tcp buf memory allocation
>>>>>
>>>>> Both are in pages, right?
>>>>> Shouldn't it be scaled to bytes and named uniform with other memcg file?
>>>>> memory.kmem.tcp.limit_in_bytes/usage_in_bytes.
>>>>>
>>>> You are absolutely correct.
>>>> Since the internal tcp comparison works, I just ended up never noticing
>>>> this.
>>>
>>> Should we have failcnt and max_usage_in_bytes for tcp as well?
>>>
>>
>> Well, we get a fail count from the tracer anyway, so I don't really see
>> a need for that. I see value in having it for the slab allocation
>> itself, but since this only controls the memory pressure framework, I
>> think we can live without it.
>>
>> That said, this is not a strong opinion. I can add it if you'd prefer.
>
> It's good for userspace to have the same set of files for all domains:
>   - memory;
>   - memory.memsw;
>   - memory.kmem;
>   - memory.kmem.tcp;
>   - etc.
> Userspace can reuse code for handling them in this case.
>
Ok. Back on this.

Not all domains have all files anyway.

max_usage seems to be a property of the main memcg, not of its domains.
failcnt is present on memsw, and on that only. The problem here, is that 
this can fail ( and usually will ) in codepaths outside the memory
controller. (see net/core/sock.c:__sk_mem_schedule)

Also, max_usage makes sense for kernel memory as a whole, but I don't 
think it makes sense here as we're only controlling a specific pressure 
condition.

So in a nutshell: I'd like to leave this alone, and add 
kmem.max_usage_in_bytes and kmem.failcnt to the to soon-to-land-here 
slab accounting patches. (Where the actual allocation happens)