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;
Date:   Sun, 9 Dec 2018 22:39:52 +0100
From:   Luc Van Oostenryck <luc.vanoostenryck@gmail.com>
To:     Tycho Andersen <tycho@tycho.ws>
Cc:     Al Viro <viro@zeniv.linux.org.uk>, linux-sparse@vger.kernel.org,
        kernel-hardening@lists.openwall.com, linux-kernel@vger.kernel.org
Subject: Re: [RFC v1] copy_{to,from}_user(): only inline when !__CHECKER__
Message-ID: <20181209213951.kumz33u6prb2seqz@ltop.local>
References: <20181209204449.18906-1-tycho@tycho.ws>
 <20181209210220.GB2217@ZenIV.linux.org.uk>
 <20181209212523.GE30796@cisco>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20181209212523.GE30796@cisco>
User-Agent: NeoMutt/20180622
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Sun, Dec 09, 2018 at 02:25:23PM -0700, Tycho Andersen wrote:
> Hi Al,
> 
> On Sun, Dec 09, 2018 at 09:02:21PM +0000, Al Viro wrote:
> > On Sun, Dec 09, 2018 at 01:44:49PM -0700, Tycho Andersen wrote:
> > > While working on some additional copy_to_user() checks for sparse, I
> > > noticed that sparse's current copy_to_user() checks are not triggered. This
> > > is because copy_to_user() is declared as __always_inline, and sparse
> > > specifically looks for a call instruction to copy_to_user() when it tries
> > > to apply the checks.
> > > 
> > > A quick fix is to explicitly not inline when __CHECKER__ is defined, so
> > > that sparse will be able to analyze all the copy_{to,from}_user calls.
> > > There may be some refactoring in sparse that we can do to fix this,
> > > although it's not immediately obvious to me how, hence the RFC-ness of this
> > > patch.
> > 
> > Which sparse checks do not trigger?  Explain, please - as it is, I had been
> > unable to guess what could "specifically looks for a call instruction" refer
> > to.
> 
> In sparse.c there's check_call_instruction(), which is triggered when
> there's an instruction of OP_CALL type in the basic block. This simply
> compares against the name of the call target to determine whether or
> not to call check_ctu().
> 
> I think what's happening here is that the call is getting inlined, and
> so the OP_CALL goes away, and check_call_instruction() never gets
> called.

Yes, it's what's happening.

There are several more or less bad/good solutions, like:
* add raw_copy_{to,from}_user() in the list of checked function
  (not inlined in most archs).
* add a new annotation to force sparse to check the byte count
  (I'm thinking about __range__/OP_RANGE or something similar).
* do these checks before functions are inlined (but then some
  constant count could not yet be seen as constant).
* ...

Wasn't there some plan to remove all these __always_inline
because of the future 'asm inline'?

-- Luc