Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
MIME-Version: 1.0
References: <20190329012836.47013-1-shakeelb@google.com> <20190329075206.GA28616@dhcp22.suse.cz>
In-Reply-To: <20190329075206.GA28616@dhcp22.suse.cz>
From:   Shakeel Butt <shakeelb@google.com>
Date:   Fri, 29 Mar 2019 09:00:32 -0700
Message-ID: <CALvZod5foPZvaD5jTBH9a7bdtMys6MdDLV=D-BLcUMU1Q7EZPw@mail.gmail.com>
Subject: Re: [RFC PATCH] mm, kvm: account kvm_vcpu_mmap to kmemcg
To:     Michal Hocko <mhocko@kernel.org>
Cc:     Johannes Weiner <hannes@cmpxchg.org>,
        Vladimir Davydov <vdavydov.dev@gmail.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Matthew Wilcox <willy@infradead.org>,
        Paolo Bonzini <pbonzini@redhat.com>,
        Ben Gardon <bgardon@google.com>,
        =?UTF-8?B?UmFkaW0gS3LEjW3DocWZ?= <rkrcmar@redhat.com>,
        Linux MM <linux-mm@kvack.org>, kvm@vger.kernel.org,
        LKML <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Fri, Mar 29, 2019 at 12:52 AM Michal Hocko <mhocko@kernel.org> wrote:
>
> On Thu 28-03-19 18:28:36, Shakeel Butt wrote:
> > A VCPU of a VM can allocate upto three pages which can be mmap'ed by the
> > user space application. At the moment this memory is not charged. On a
> > large machine running large number of VMs (or small number of VMs having
> > large number of VCPUs), this unaccounted memory can be very significant.
>
> Is this really the case. How many machines are we talking about? Say I
> have a virtual machines with 1K cpus, this will result in 12MB. Is this
> significant to the overal size of the virtual machine to even care?
>

Think of having ~1K VMs having 100s of vcpus and the page size can be
larger than 4k. This is not something happening now but we are moving
in that direction. Also

> > So, this memory should be charged to a kmemcg. However that is not
> > possible as these pages are mmapped to the userspace and PageKmemcg()
> > was designed with the assumption that such pages will never be mmapped
> > to the userspace.
> >
> > One way to solve this problem is by introducing an additional memcg
> > charging API similar to mem_cgroup_[un]charge_skmem(). However skmem
> > charging API usage is contained and shared and no new users are
> > expected but the pages which can be mmapped and should be charged to
> > kmemcg can and will increase. So, requiring the usage for such API will
> > increase the maintenance burden. The simplest solution is to remove the
> > assumption of no mmapping PageKmemcg() pages to user space.
>
> IIRC the only purpose of PageKmemcg is to keep accounting in the legacy
> memcg right. Spending a page flag for that is just no-go.

PgaeKmemcg is used for both v1 and v2.

> If PageKmemcg
> cannot reuse mapping then we have to find a better place for it (e.g.
> bottom bit in the page->memcg pointer or rethink the whole PageKmemcg.
>

Johannes have proposal, I will look more into those.

Shakeel