Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Sat, 8 Oct 2022 21:03:28 -0700
From:   Kees Cook <keescook@chromium.org>
To:     Paulo Miguel Almeida <paulo.miguel.almeida.rodenas@gmail.com>
Cc:     Christine Caulfield <ccaulfie@redhat.com>,
        David Teigland <teigland@redhat.com>, cluster-devel@redhat.com,
        linux-kernel@vger.kernel.org, linux-hardening@vger.kernel.org
Subject: Re: [PATCH v2][next] dlm: Replace one-element array with
 flexible-array member
Message-ID: <202210082044.51106145BD@keescook>
References: <Y0IFEUjwXGZFf7bB@mail.google.com>
 <378C6BDE-0A68-4938-86CD-495BD5F35BE6@chromium.org>
 <Y0IsXXYnS4DnWkMW@mail.google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <Y0IsXXYnS4DnWkMW@mail.google.com>
Precedence: bulk

On Sun, Oct 09, 2022 at 03:05:17PM +1300, Paulo Miguel Almeida wrote:
> On Sat, Oct 08, 2022 at 05:18:35PM -0700, Kees Cook wrote:
> > This is allocating 1 more byte than before, since the struct size didn't change. But this has always allocated too much space, due to the struct padding. For a "no binary changes" patch, the above "+ 1" needs to be left off.
> 
> That's true. I agree that leaving "+ 1" would work and produce a
> no-binary-changes patch due to the existing padding that the structure
> has. OTOH, I thought that relying on that space could bite us in the
> future if anyone tweaks the struct again...so my reaction was to ensure 
> that the NUL-terminator space was always guaranteed to be there.
> Hence, the change on c693 (objdump above).
> 
> What do you think? Should we keep or leave the above
> "+ 1" after the rationale above?

I think it depends on what's expected from this allocation. Christine or
David, can you speak to this?

> > I would expect the correct allocation size to be:
> > offsetof(typeof(*ls), ls_name) + namelen
> 
> Fair point, I will make this change.

Well, only do that if we don't depend on the padding nor a trailing
%NUL. :)

> > Question, though: is ls_name _expected_ to be %NUL terminated
> 
> Yes, it is. I tracked down ls_name's utilisations and it is passed down to 
> a bunch of routines that expects it to be NUL-terminated such as
> snprintf and vsnprintf.

Agreed: I see the string functions it gets passed to. So, then the next
question I have is does "namelen" take into account the %NUL, and is
"name" %NUL terminated? Those answers appear to be "no" and "yes",
respectively:

static int new_lockspace(const char *name, ...)
{
	...
        int namelen = strlen(name);


The comparisons for ls->ls_namelen are all done without the %NUL count:

                if (ls->ls_namelen != namelen)
                        continue;
                if (memcmp(ls->ls_name, name, namelen))
                        continue;

> >, and was the prior 3 bytes of extra allocation accidentally required?
> > 
> 
> I am assuming that you are refering to ls_namelen in the struct dlm_ls
> (please correct me if this isn't what you meant).

No, I meant ls_name (the pahole output shows the trailing 3 bytes of
padding before. And with your patch it becomes 4 bytes of trailing
padding.

So I think this is "accidentally correct", since it's so carefully using
memcmp() and not strcmp().

Given the existing padding on the structure, through, it likely needs
to keep a certain amount of minimum padding.

original size was actually this, so you could use this for the new
calculation to get the same values as before:

	offsetof(typeof(*ls), ls_name) + 4 + namelen;

In reality, it may be possible to do this to get exactly what is needed,
but no less than the struct itself:

	max(offsetof(typeof(*ls), ls_name) + 1 + namelen, sizeof(*ls));

-Kees

-- 
Kees Cook