Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:1 as permitted sender) client-ip=2620:137:e000::3:1;
Date:   Tue, 26 Sep 2023 08:38:57 +0200
From:   Mateusz Guzik <mjguzik@gmail.com>
To:     John Johansen <john.johansen@canonical.com>
Cc:     Vinicius Costa Gomes <vinicius.gomes@intel.com>,
        linux-security-module@vger.kernel.org, apparmor@lists.ubuntu.com,
        linux-kernel@vger.kernel.org
Subject: Re: [apparmor] use per-cpu refcounts for apparmor labels?
Message-ID: <20230926063857.h3afce5hagnlkoob@f>
References: <CAGudoHFfG7mARwSqcoLNwV81-KX4Bici5FQHjoNG4f9m83oLyg@mail.gmail.com>
 <87a5t9bypm.fsf@intel.com>
 <c6379a39-42f2-b3f9-c835-bbebe516ba3a@canonical.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <c6379a39-42f2-b3f9-c835-bbebe516ba3a@canonical.com>
Precedence: bulk

On Mon, Sep 25, 2023 at 11:21:26PM -0700, John Johansen wrote:
> On 9/25/23 16:49, Vinicius Costa Gomes wrote:
> > Hi Mateusz,
> > 
> > Mateusz Guzik <mjguzik@gmail.com> writes:
> > 
> > > I'm sanity-checking perf in various microbenchmarks and I found
> > > apparmor to be the main bottleneck in some of them.
> > > 
> > > For example: will-it-scale open1_processes -t 16, top of the profile:
> > >    20.17%  [kernel]                   [k] apparmor_file_alloc_security
> > >    20.08%  [kernel]                   [k] apparmor_file_open
> > >    20.05%  [kernel]                   [k] apparmor_file_free_security
> > >    18.39%  [kernel]                   [k] apparmor_current_getsecid_subj
> > > [snip]
> > > 
> > > This serializes on refing/unrefing apparmor objs, sounds like a great
> > > candidate for per-cpu refcounting instead (I'm assuming they are
> > > expected to be long-lived).
> > > 
> > > I would hack it up myself, but I failed to find a clear spot to switch
> > > back from per-cpu to centalized operation and don't want to put
> > > serious effort into it.
> > > 
> > > Can you sort this out?
> > 
> 
> I will add looking into it on the todo list. Its going to have to come
> after some other major cleanups land, and I am not sure we can make
> the semantic work well for some of these. For other we might get away
> with switching to a critical section like Vinicius's patch has done
> for apparmor_current_getsecid_subj.
> 

Is there an eta?

I looked at dodging ref round trips myself, but then found that ref
manipulation in apparmor_file_alloc_security and the free counterpart
cannot be avoided. Thus per-cpu refs instead.

Perhaps making the label as stale would be a good enough switching
point? Is it *guaranteed* to get labelled as stale before it gets freed?

btw, __aa_proxy_redirect open-codes setting the flag.

> > I was looking at this same workload, and proposed a patch[1] some time
> > ago, see if it helps:
> > 
> > https://lists.ubuntu.com/archives/apparmor/2023-August/012914.html
> > 
> > But my idea was different, in many cases, we are looking at the label
> > associated with the current task, and there's no need to take the
> > refcount.
> > 
> 
> yes, and thanks for that.
>