[PATCH 7/7] exec: Implement kernel_execve

David Laight David.Laight at ACULAB.COM
Wed Jul 15 16:46:44 UTC 2020


From: Kees Cook <keescook at chromium.org>
> Sent: 15 July 2020 16:09
> 
> On Wed, Jul 15, 2020 at 02:55:50PM +0000, David Laight wrote:
> > From: Christoph Hellwig
> > > Sent: 15 July 2020 07:43
> > > Subject: Re: [PATCH 7/7] exec: Implement kernel_execve
> > >
> > > On Tue, Jul 14, 2020 at 02:49:23PM -0700, Kees Cook wrote:
> > > > On Tue, Jul 14, 2020 at 08:31:40AM -0500, Eric W. Biederman wrote:
> > > > > +static int count_strings_kernel(const char *const *argv)
> > > > > +{
> > > > > +	int i;
> > > > > +
> > > > > +	if (!argv)
> > > > > +		return 0;
> > > > > +
> > > > > +	for (i = 0; argv[i]; ++i) {
> > > > > +		if (i >= MAX_ARG_STRINGS)
> > > > > +			return -E2BIG;
> > > > > +		if (fatal_signal_pending(current))
> > > > > +			return -ERESTARTNOHAND;
> > > > > +		cond_resched();
> > > > > +	}
> > > > > +	return i;
> > > > > +}
> > > >
> > > > I notice count() is only ever called with MAX_ARG_STRINGS. Perhaps
> > > > refactor that too? (And maybe rename it to count_strings_user()?)
> >
> > Thinks....
> > If you setup env[] and argv[] on the new user stack early in exec processing
> > then you may not need any limits at all - except the size of the user stack.
> > Even the get_user() loop will hit an invalid address before the counter
> > wraps (provided it is unsigned long).
> 
> *grumpy noises* Yes, but not in practice (if I'm understanding what you
> mean). The expectations of a number of execution environments can be
> really odd-ball. I've tried to collect the notes from over the years in
> prepare_arg_pages()'s comments, and it mostly boils down to "there has
> to be enough room for the exec to start" otherwise the exec ends up in a
> hard-to-debug failure state (i.e. past the "point of no return", where you
> get no useful information about the cause of the SEGV). So the point has
> been to move as many of the setup checks as early as possible and report
> about them if they fail. The argv processing is already very early, but
> it needs to do the limit checks otherwise it'll just break after the exec
> is underway and the process will just SEGV. (And ... some environments
> will attempt to dynamically check the size of the argv space by growing
> until it sees E2BIG, so we can't just remove it and let those hit SEGV.)

Yes - I bet the code is horrid.
I guess the real problem is that you'd need access to the old process's
user addresses and the new process's stack area at the same time.
Unless there is a suitable hole in the old process's address map
any attempted trick will fall foul of cache aliasing on some
architectures - like anything else that does page-loaning.

I'm sure there are hair-brained schemes that might work.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)



More information about the Linux-security-module-archive mailing list