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: <20180919122751.12439-1-tvrtko.ursulin@linux.intel.com>
 <alpine.DEB.2.21.1809281151190.2004@nanos.tec.linutronix.de>
 <20180928164111.i6nba2j6mnegwslw@lakrids.cambridge.arm.com>
 <20180928172340.GA32651@tassilo.jf.intel.com> <20180928174016.i7d24puv7y3jwzf6@lakrids.cambridge.arm.com>
 <20180928204930.GC32651@tassilo.jf.intel.com>
In-Reply-To: <20180928204930.GC32651@tassilo.jf.intel.com>
From:   Jann Horn <jannh@google.com>
Date:   Fri, 28 Sep 2018 22:54:05 +0200
Message-ID: <CAG48ez16ECnLMNggTZQQJOuPK5ZJ8dLSkGDzvxwK1zVh1BFExw@mail.gmail.com>
Subject: Re: [RFC 0/5] perf: Per PMU access controls (paranoid setting)
To:     Andi Kleen <ak@linux.intel.com>
Cc:     Mark Rutland <mark.rutland@arm.com>,
        Thomas Gleixner <tglx@linutronix.de>, tursulin@ursulin.net,
        kernel list <linux-kernel@vger.kernel.org>,
        tvrtko.ursulin@linux.intel.com,
        Peter Zijlstra <peterz@infradead.org>,
        "the arch/x86 maintainers" <x86@kernel.org>,
        "H . Peter Anvin" <hpa@zytor.com>, acme@kernel.org,
        alexander.shishkin@linux.intel.com, jolsa@redhat.com,
        namhyung@kernel.org, maddy@linux.vnet.ibm.com,
        alexey.budankov@linux.intel.com, Kees Cook <keescook@chromium.org>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Fri, Sep 28, 2018 at 10:49 PM Andi Kleen <ak@linux.intel.com> wrote:
>
> On Fri, Sep 28, 2018 at 06:40:17PM +0100, Mark Rutland wrote:
> > On Fri, Sep 28, 2018 at 10:23:40AM -0700, Andi Kleen wrote:
> > > > There's also been prior discussion on these feature in other contexts
> > > > (e.g. android expoits resulting from out-of-tree drivers). It would be
> > > > nice to see those considered.
> > > >
> > > > IIRC The conclusion from prior discussions (e.g. [1]) was that we wanted
> > > > finer granularity of control such that we could limit PMU access to
> > > > specific users -- e.g. disallow arbitrary android apps from poking *any*
> > > > PMU, while allowing some more trusted apps/users to uses *some* specific
> > > > PMUs.
> > > >
> > > > e.g. we could add /sys/bus/event_source/devices/${PMU}/device, protect
> > > > this via the usual fs ACLs, and pass the fd to perf_event_open()
> > > > somehow. A valid fd would act as a capability, taking precedence over
> > > > perf_event_paranoid.
> > >
> > > That sounds like an orthogonal feature. I don't think the original
> > > patchkit would need to be hold up for this. It would be something
> > > in addition.
> >
> > I have to say that I disagree -- these controls will have to interact
> > somehow, and the fewer of them we have, the less complexity we'll have
> > to deal with longer-term.
>
> You're proposing to completely redesign perf_event_open.

And I think it would be a very good redesign. :) I love things that
use file descriptors to represent capabilities.

> This new file descriptor argument doesn't exist today so it would
> need to create a new system call with more arguments

Is that true? The first argument is a pointer to a struct that
contains its own size, so it can be expanded without an ABI break. I
don't see any reason why you couldn't cram more stuff in there.

> (and BTW it would be more than the normal 6 argument limit
> we have, so actually it couldn't even be a standard sycall)
>
> Obviously we would need to keep the old system call around
> for compability, so you would need to worry about this
> interaction in any case!
>
> So tying it together doesn't make any sense, because
> the problem has to be solved separately anyways.
>
> >
> > > BTW can't you already do that with the syscall filter? I assume
> > > the Android sandboxes already use that. Just forbid perf_event_open
> > > for the apps.
> >
> > Note that this was about providing access to *some* PMUs in some cases.
> >
> > IIUC, if that can be done today via a syscall filter, the same is true
> > of per-pmu paranoid settings.
>
> The difference is that the Android sandboxes likely already doing this
> and have all the infrastructure, and it's just another rule.
>
> Requiring syscall filters just to use the PMU on xn system
> that otherwise doesn't need them would be very odd.
>
> -Andi