Re: [fs/exec] 80bd5afdd8: xfstests.generic.633.fail

From: Christian Brauner <brauner@kernel.org>
To: Matthew Wilcox <willy@infradead.org>
Cc: kernel test robot <oliver.sang@intel.com>,
	Kees Cook <keescook@chromium.org>,
	Ariadne Conill <ariadne@dereferenced.org>,
	0day robot <lkp@intel.com>,
	Michael Kerrisk <mtk.manpages@gmail.com>,
	Rich Felker <dalias@libc.org>,
	Eric Biederman <ebiederm@xmission.com>,
	Alexander Viro <viro@zeniv.linux.org.uk>,
	LKML <linux-kernel@vger.kernel.org>,
	lkp@lists.01.org, Andrew Morton <akpm@linux-foundation.org>,
	linux-fsdevel@vger.kernel.org, linux-mm@kvack.org,
	stable@vger.kernel.org
Subject: Re: [fs/exec]  80bd5afdd8: xfstests.generic.633.fail
Date: Mon, 31 Jan 2022 18:13:44 +0100	[thread overview]
Message-ID: <20220131171344.77iifun5wdilbqdz@wittgenstein> (raw)
In-Reply-To: <20220131161415.wlvtsd4ecehyg3x5@wittgenstein>

On Mon, Jan 31, 2022 at 05:14:15PM +0100, Christian Brauner wrote:
> On Mon, Jan 31, 2022 at 03:51:21PM +0000, Matthew Wilcox wrote:
> > On Mon, Jan 31, 2022 at 04:37:07PM +0100, Christian Brauner wrote:
> > > On Mon, Jan 31, 2022 at 03:19:22PM +0000, Matthew Wilcox wrote:
> > > > On Mon, Jan 31, 2022 at 04:08:19PM +0100, Christian Brauner wrote:
> > > > > On Mon, Jan 31, 2022 at 10:43:52PM +0800, kernel test robot wrote:
> > > > > I can fix this rather simply in our upstream fstests with:
> > > > > 
> > > > > static char *argv[] = {
> > > > > 	"",
> > > > > };
> > > > > 
> > > > > I guess.
> > > > > 
> > > > > But doesn't
> > > > > 
> > > > > static char *argv[] = {
> > > > > 	NULL,
> > > > > };
> > > > > 
> > > > > seem something that should work especially with execveat()?
> > > > 
> > > > The problem is that the exec'ed program sees an argc of 0, which is the
> > > > problem we're trying to work around in the kernel (instead of leaving
> > > > it to ld.so to fix for suid programs).
> > > 
> > > Ok, just seems a bit more intuitive for path-based exec than for
> > > fd-based execveat().
> > > 
> > > What's argv[0] supposed to contain in these cases?
> > > 
> > > 1. execveat(fd, NULL, ..., AT_EMPTY_PATH)
> > > 2. execveat(fd, "my-file", ..., )
> > > 
> > > "" in both 1. and 2.?
> > > "" in 1. and "my-file" in 2.?
> > 
> > You didn't specify argv for either of those, so I have no idea.
> > Programs shouldn't be assuming anything about argv[0]; it's purely
> > advisory.  Unfortunately, some of them do.  And some of them are suid.
> 
> Yes, programs shouldn't assume anything about argv[0]. But a lot of
> programs are used to setting argv[0] to the name of the executed binary.
> The exec* manpages examples do this. Just looking at a random selftest, e.g.
> 
> bpf/prog_tests/test_lsm.c
> 
> where we find:
> 
> 	char *CMD_ARGS[] = {"true", NULL};
> 	execvp(CMD_ARGS[0], CMD_ARGS);
> 
> I'm just wondering how common this is for execveat() because it is not
> as clear what the actual name of the binary is in these two examples
> 
> 	1.
> 	fd = open("/bin/true", );
> 	char *CMD_ARGS[] = {"", NULL};
> 	execveat(fd, NULL, ..., AT_EMPTY_PATH)
> 	
> 	2.
> 	fd = open("/bin", );
> 	char *CMD_ARGS[] = {"true", NULL};
> 	execveat(fd, CMD_ARGS[0], CMD_ARGS 0)
> 
> in other words, the changes that you see CMD_ARGS[0] == NULL for
> execveat() seem higher than for path-based exec.
> 
> To counter that we should probably at least update the execveat()
> manpage with a recommendation what CMD_ARGS[0] should be set to if it
> isn't allowed to be set to NULL anymore. This is why was asking what
> argv[0] is supposed to be if the binary doesn't take any arguments.

Sent a fix to our fstests now replacing the argv[0] as NULL with "".