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:   Mon, 19 Sep 2022 13:02:51 -0700
From:   Kees Cook <keescook@chromium.org>
To:     Josh Triplett <josh@joshtriplett.org>
Cc:     Peter Zijlstra <peterz@infradead.org>,
        Eric Biederman <ebiederm@xmission.com>,
        Alexander Viro <viro@zeniv.linux.org.uk>, linux-mm@kvack.org,
        linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] fs/exec.c: Add fast path for ENOENT on PATH search
 before allocating mm
Message-ID: <202209191256.893576D4@keescook>
References: <5c7333ea4bec2fad1b47a8fa2db7c31e4ffc4f14.1663334978.git.josh@joshtriplett.org>
 <202209160727.5FC78B735@keescook>
 <YyTY+OaClK+JHCOw@localhost>
 <202209161637.9EDAF6B18@keescook>
 <YyUZ0NHfFF+eVe24@localhost>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <YyUZ0NHfFF+eVe24@localhost>
Precedence: bulk

On Sat, Sep 17, 2022 at 01:50:24AM +0100, Josh Triplett wrote:
> On Fri, Sep 16, 2022 at 05:11:18PM -0700, Kees Cook wrote:
> > I don't like the idea of penalizing the _succeeding_ case, though, which
> > happens if we do the path walk twice. So, I went and refactoring the setup
> > order, moving the do_open_execat() up into alloc_bprm() instead of where
> > it was in bprm_exec(). The result makes it so it is, as you observed,
> > before the mm creation and generally expensive argument copying. The
> > difference to your patch seems to only be the allocation of the file
> > table entry, but avoids the double lookup, so I'm hoping the result is
> > actually even faster.
> 
> Thanks for giving this a try; I'd wondered how feasible it would be to
> just do one lookup.
> 
> However, on the same test system with the same test setup, with your
> refactor it seems to go slower:
> fork/execvpe: 38087ns
> fork/execve:  33758ns
> 
> For comparison, the previous numbers (which I re-confirmed):
> 
> Without fast-path:
> fork/execvpe: 49876ns
> fork/execve:  32773ns
> 
> With my original separate-lookup fast-path:
> fork/execvpe: 36890ns
> fork/execve:  31551ns

Hmm, this shows as slower in the *normal* case, which I find rather
surprising -- it's the same work, just reordered.

Can you post a URL to your tests? I'd like to reproduce this and maybe
throw perf at it as well.

> I tried several runs of each, and I seem to get reasonably consistent
> results.
> 
> My test program just creates a pipe once, then loops on
> clock_gettime/fork/execvpe/read, with the spawned child process doing
> clock_gettime/write/exit (in asm to minimize overhead). The test PATH is
> PATH=/usr/local/bin:/usr/local/sbin:/usr/bin:/usr/sbin:/bin:/sbin:. with
> the test program in the current directory.

I'm also curious about less synthetic testing. It'd be nice to see real
workloads with these changes, etc.

-- 
Kees Cook