Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210223214346.GB6000@sequoia> <20210223215054.GC6000@sequoia>
 <20210223223652.GD6000@sequoia> <CAFqZXNvfux46_f8gnvVvRYMKoes24nwm2n3sPbMjrB8vKTW00g@mail.gmail.com>
In-Reply-To: <CAFqZXNvfux46_f8gnvVvRYMKoes24nwm2n3sPbMjrB8vKTW00g@mail.gmail.com>
From:   Paul Moore <paul@paul-moore.com>
Date:   Thu, 25 Feb 2021 20:06:45 -0500
Message-ID: <CAHC9VhSaU-3_fs83kEA5bxBf9xMsE29B_O5nXFpROk4=y9kgXw@mail.gmail.com>
Subject: Re: [BUG] Race between policy reload sidtab conversion and live conversion
To:     Ondrej Mosnacek <omosnace@redhat.com>
Cc:     Tyler Hicks <tyhicks@linux.microsoft.com>,
        Stephen Smalley <stephen.smalley.work@gmail.com>,
        SElinux list <selinux@vger.kernel.org>,
        Linux kernel mailing list <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, Feb 24, 2021 at 4:35 AM Ondrej Mosnacek <omosnace@redhat.com> wrote:
> After the switch to RCU, we now have:
> 1. Start live conversion of new entries.
> 2. Convert existing entries.
> 3. RCU-assign the new policy pointer to selinux_state.
> [!!! Now actually both old and new sidtab may be referenced by
> readers, since there is no synchronization barrier previously provided
> by the write lock.]
> 4. Wait for synchronize_rcu() to return.
> 5. Now only the new sidtab is visible to readers, so the old one can
> be destroyed.
>
> So the race can happen between 3. and 5., if one thread already sees
> the new sidtab and adds a new entry there, and a second thread still
> has the reference to the old sidtab and also tires to add a new entry;
> live-converting to the new sidtab, which it doesn't expect to change
> by itself. Unfortunately I failed to realize this when reviewing the
> patch :/

It is possible I'm not fully understanding the problem and/or missing
an important detail - it is rather tricky code, and RCU can be very
hard to reason at times - but I think we may be able to solve this
with some lock fixes inside sidtab_context_to_sid().  Let me try to
explain to see if we are on the same page here ...

The problem is when we have two (or more) threads trying to
add/convert the same context into a sid; the task with new_sidtab is
looking to add a new sidtab entry, while the task with old_sidtab is
looking to convert an entry in old_sidtab into a new entry in
new_sidtab.  Boom.

Looking at the code in sidtab_context_to_sid(), when we have two
sidtabs that are currently active (old_sidtab->convert pointer is
valid) and a task with old_sidtab attempts to add a new entry to both
sidtabs it first adds it to the old sidtab then it also adds it to the
new sidtab.  I believe the problem is that in this case while the task
grabs the old_sidtab->lock, it never grabs the new_sidtab->lock which
allows it to race with tasks that already see only new_sidtab.  I
think adding code to sidtab_context_to_sid() which grabs the
new_sidtab->lock when adding entries to the new_sidtab *should* solve
the problem.

Did I miss something important? ;)

-- 
paul moore
www.paul-moore.com