Re: [PATCH memcg v4] net: set proper memcg for net_init hooks allocations

From: Roman Gushchin <roman.gushchin@linux.dev>
To: Vasily Averin <vvs@openvz.org>
Cc: "Shakeel Butt" <shakeelb@google.com>,
	"Michal Koutný" <mkoutny@suse.com>,
	"Vlastimil Babka" <vbabka@suse.cz>,
	kernel@openvz.org, "Florian Westphal" <fw@strlen.de>,
	LKML <linux-kernel@vger.kernel.org>,
	"Michal Hocko" <mhocko@suse.com>,
	Cgroups <cgroups@vger.kernel.org>,
	netdev <netdev@vger.kernel.org>,
	"David S. Miller" <davem@davemloft.net>,
	"Jakub Kicinski" <kuba@kernel.org>,
	"Paolo Abeni" <pabeni@redhat.com>
Subject: Re: [PATCH memcg v4] net: set proper memcg for net_init hooks allocations
Date: Wed, 27 Apr 2022 15:36:03 -0700	[thread overview]
Message-ID: <YmnFUwhmqJwYGQ5j@carbon> (raw)
In-Reply-To: <6b18f82d-1950-b38e-f3f5-94f6c23f0edb@openvz.org>

On Thu, Apr 28, 2022 at 01:16:53AM +0300, Vasily Averin wrote:
> On 4/27/22 18:06, Shakeel Butt wrote:
> > On Wed, Apr 27, 2022 at 5:22 AM Michal Koutný <mkoutny@suse.com> wrote:
> >>
> >> On Tue, Apr 26, 2022 at 10:23:32PM -0700, Shakeel Butt <shakeelb@google.com> wrote:
> >>> [...]
> >>>>
> >>>> +static inline struct mem_cgroup *get_mem_cgroup_from_obj(void *p)
> >>>> +{
> >>>> +       struct mem_cgroup *memcg;
> >>>> +
> >>>
> >>> Do we need memcg_kmem_enabled() check here or maybe
> >>> mem_cgroup_from_obj() should be doing memcg_kmem_enabled() instead of
> >>> mem_cgroup_disabled() as we can have "cgroup.memory=nokmem" boot
> >>> param.
> 
> Shakeel, unfortunately I'm not ready to answer this question right now.
> I even did not noticed that memcg_kmem_enabled() and mem_cgroup_disabled()
> have a different nature.
> If you have no objections I'm going to keep this place as is and investigate
> this question later. 
> 
> >> I reckon such a guard is on the charge side and readers should treat
> >> NULL and root_mem_group equally. Or is there a case when these two are
> >> different?
> >>
> >> (I can see it's different semantics when stored in current->active_memcg
> >> (and active_memcg() getter) but for such "outer" callers like here it
> >> seems equal.)
> 
> Dear Michal,
> I may have misunderstood your point of view, so let me explain my vision
> in more detail.
> I do not think that NULL and root_mem_cgroup are equal here:
> - we have enabled cgroups and well-defined root_mem_cgroup,
> - this function is called from inside memcg-limited container,
> - we tried to get memcg from net, but without success,
>   and as result got NULL from  mem_cgroup_from_obj()
>   (frankly speaking I do not think this situation is really possible)
> If we keep memcg = NULL, then current's memcg will not be masked and
> net_init's allocations will be accounted to current's memcg. 
> So we need to set active_memcg to root_mem_cgroup, it helps to avoid
> incorrect accounting.

It's way out of scope of this patch, but I think we need to stop
using NULL as root_mem_cgroup/system scope indicator. Remaining use cases
will be like end of cgroup iteration, active memcg not set, parent of the
root memcg, etc.
We can point root_mem_cgroup at a statically allocated structure
on both CONFIG_MEMCG and !CONFIG_MEMCG.
Does it sound reasonable or I'm missing some important points?

Thanks!