[Devel] [PATCH 2/7] tcp: Charge socket buffers into cg memory (v2)

[Pavel Emelyanov]

>>>> +void memcg_charge_kmem_nofail(struct mem_cgroup *memcg, u64 size)
>>>>  {
>>>> +	struct res_counter *fail_res;
>>>> +
>>>> +	/*
>>>> +	 * FIXME -- strictly speaking, this value should _also_
>>>> +	 * be charged into kmem counter. But since res_counter_charge
>>>> +	 * is sub-optimal (takes locks) AND we do not care much
>>>> +	 * about kmem limits (at least for now) we can just directly
>>>> +	 * charge into mem counter.
>>>> +	 */
>>>
>>> Please charge kmem too. As I've already told you it should not make any
>>> difference in terms of performance, because we already have a bottleneck
>>> of the same bandwidth.
>>>
>>> Anyway, if we see any performance degradation, I will convert
>>> mem_cgroup->kmem to a percpu counter.
>>
>> No, let's do it vice-versa -- first you fix the locking, then I update this code.
> 
> I don't understand why, because you provide no arguments and keep
> ignoring my reasoning why I think charging kmem along with res is OK,
> which is one paragraph above.

The bandwidth of the bottleneck doesn't look to be the same -- the res counters
in question are not in one cache-line and adding one more (btw, do we have swap
account turned on by default?) will not come unnoticed.

Yet again -- I don't mind changing this and charge TCP into kmem too, I'll do
it, but after this charging becomes fast enough.

-- Pavel