lkml.org 
[lkml]   [2017]   [Dec]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRE: [PATCH v3 2/2] Protected O_CREAT open in sticky directories
Date
From: Solar Designer
> Sent: 30 November 2017 17:52
>
> On Thu, Nov 30, 2017 at 04:53:06PM +0000, David Laight wrote:
> > From: Salvatore Mesoraca
> > > if a program tries to open a file, in a sticky directory,
> > > with the O_CREAT flag and without the O_EXCL, it probably has a bug.
> > > This feature allows to detect and potentially block programs that
> > > act this way, it can be used to find vulnerabilities (like those
> > > prevented by patch #1) and to do policy enforcement.
>
> > I presume the 'vulnerabilities' are related to symlinks being created
> > just before the open?
>
> That's one way to exploit them. We already have a sysctl to try and
> block that specific attack, and we're considering adding more, for other
> attacks. But we'd also like an easy way to learn of the vulnerabilities
> without anyone trying to exploit them yet, hence this patch.
>
> > Trouble is this change breaks a lot of general use of /tmp.
>
> That's general misuse of /tmp. Things like "command > /tmp/file"
> without having pre-created the file with O_EXCL e.g. by mktemp(1).

I'm sorry, I've been using Unix for over 30 years.
/tmp is a place that temporary files were created - nothing special.
Traditionally it was emptied on every boot.
There was never anything that required files be created in any
specific way.

> > I always assumed that code that cared would use O_EXCL and
> > everything else wasn't worth subverting.
>
> Opinions will vary on whether it's worth subverting or not, and someone
> may reasonably want to configure this differently on different systems.
>
> > I found code in vi (and elsewhere) that subverted these checks
> > by opening with O_WRONLY if stat() showed the file existed and
> > O_CREAT|O_EXCL if it didn't.
>
> Yes, such misuses of /tmp and the like by use of those programs - where
> the potential consequences would often be less severe (if your example
> above is correct, it sounds like it'd be data spoofing but not
> overwriting another file over a link) - could remain unnoticed.

It seemed to me that code had been added to avoid issues caused by
this policing of opens.
But the code added is itself racy - so makes the whole thing pointless.

> > I'm pretty sure that traditionally a lot of these opens were done
> > with O_CREAT|O_TRUNC.
> > Implementing that as unlink() followed by a create would stop
> > 'random' (ok all) symlinks being followed.
>
> That's racy. We have O_EXCL, and it must be used along with O_CREAT
> when the directory is untrusted. (If it were only about symlinks, we
> also already have O_NOFOLLOW.)

Not racy if done in the kernel itself.

> > Overall I'm pretty sure this change will break things badly somewhere.
>
> Sure. We want it to visibly break what was subtly broken, so that the
> breakage can be seen and corrected without an attempted attack.

Maybe, but it will break some user system when they do a kernel update.
It won't necessarily break a developer's system first.

And, AFAICT, there is already code that subverts any tests by doing
a check for the file existing and then selecting between two different
types on open.

David

\
 
 \ /
  Last update: 2017-12-01 10:47    [W:0.064 / U:0.076 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site