From mboxrd@z Thu Jan  1 00:00:00 1970
From: Glauber Costa <glommer-bzQdu9zFT3WakBO8gow8eQ@public.gmane.org>
Subject: Re: [PATCH v2 02/13] memcg: Kernel memory accounting infrastructure.
Date: Mon, 12 Mar 2012 16:38:26 +0400
Message-ID: <4F5DEE42.6050607@parallels.com>
References: <1331325556-16447-1-git-send-email-ssouhlal@FreeBSD.org> <1331325556-16447-3-git-send-email-ssouhlal@FreeBSD.org>
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit
Return-path: <cgroups-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org>
In-Reply-To: <1331325556-16447-3-git-send-email-ssouhlal-HZy0K5TPuP5AfugRpC6u6w@public.gmane.org>
Sender: cgroups-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
List-ID: <cgroups.vger.kernel.org>
Content-Type: text/plain; charset="us-ascii"; format="flowed"
To: Suleiman Souhlal <ssouhlal-HZy0K5TPuP5AfugRpC6u6w@public.gmane.org>
Cc: cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, suleiman-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org, kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org, penberg-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org, cl-vYTEC60ixJUAvxtiuMwx3w@public.gmane.org, yinghan-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org, hughd-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org, gthelen-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org, peterz-wEGCiKHe2LqWVfeAwA7xHQ@public.gmane.org, dan.magenheimer-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org, hannes-druUgvl0LCNAfugRpC6u6w@public.gmane.org, mgorman-l3A5Bk7waGM@public.gmane.org, James.Bottomley-d9PhHud1JfjCXq6kfMZ53/egYHeGw8Jk@public.gmane.org, linux-mm-Bw31MaZKKs3YtjvyW6yDsg@public.gmane.org, devel-GEFAQzZX7r8dnm+yROfE0A@public.gmane.org, linux-kernel-u79uwXL29TY76Z2rM5mHXA@public.gmane.org

On 03/10/2012 12:39 AM, Suleiman Souhlal wrote:
> +#ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
> +int
> +memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, long long delta)
> +{
> +	struct res_counter *fail_res;
> +	struct mem_cgroup *_memcg;
> +	int may_oom, ret;
> +
> +	may_oom = (gfp&  __GFP_WAIT)&&  (gfp&  __GFP_FS)&&
> +	    !(gfp&  __GFP_NORETRY);
> +
> +	ret = 0;
> +
> +	_memcg = memcg;
> +	if (memcg&&  !mem_cgroup_test_flag(memcg,
> +	    MEMCG_INDEPENDENT_KMEM_LIMIT)) {
> +		ret = __mem_cgroup_try_charge(NULL, gfp, delta / PAGE_SIZE,
> +		&_memcg, may_oom);
> +		if (ret == -ENOMEM)
> +			return ret;
> +	}
> +
> +	if (memcg&&  _memcg == memcg)
> +		ret = res_counter_charge(&memcg->kmem, delta,&fail_res);
> +
> +	return ret;
> +}
> +
> +void
Ok.

So I've spent most of the day today trying to come up with a way not to 
kill the whole performance we gain from consume_stock() by this 
res_counter_charge() to kmem afterwards...

You mentioned you want to still be able to bill to memcg->kmem mostly 
for debugging/display purposes. So we're surely not using all of the 
res_counter infrastructure (limiting, soft limits, etc)

I was thinking: Can't we have a percpu_counter that we use for this 
purpose when !kmem_independent ?

we may not even need to bloat the struct, since we can fold it into a 
union with struct res_counter kmem (which is bigger than a percpu 
counter anyway).

We just need to be a bit more careful not to allow kmem_independent to 
change when we already have charges to any of them (but we need to do it 
anyway)