Message-ID: <202011241347.4AFCBDF62@keescook>
Date:   Tue, 24 Nov 2020 13:53:37 -0800
From:   Kees Cook <keescook@...omium.org>
To:     Petr Mladek <pmladek@...e.com>
Cc:     Steven Rostedt <rostedt@...dmis.org>,
        Alan Stern <stern@...land.harvard.edu>,
        Sergey Senozhatsky <sergey.senozhatsky@...il.com>,
        Kernel development list <linux-kernel@...r.kernel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Linus Torvalds <torvalds@...ux-foundation.org>
Subject: Re: Printk specifiers for __user pointers

On Mon, Nov 23, 2020 at 10:53:24AM +0100, Petr Mladek wrote:
> On Fri 2020-11-20 13:42:42, Steven Rostedt wrote:
> > On Fri, 20 Nov 2020 11:44:12 -0500
> > Alan Stern <stern@...land.harvard.edu> wrote:
> > 
> > > To the VSPRINTF maintainers:
> > > 
> > > Documentation/core-api/printk-formats.rst lists a large number of format 
> > > specifiers for pointers of various sorts.  Yet as far as I can see, 
> > > there is no specifier meant for use with __user pointers.
> > > 
> > > The security implications of printing the true, unmangled value of a 
> > > __user pointer are minimal, since doing so does not leak any kernel 
> > > information.  So %px would work, but tools like checkpatch.pl don't like 
> > > it.
> 
> Just to be sure as I am not a security expert. Is there really that
> big difference in the risk? The following scenarios come to my mind:
> 
> 1. The address would show a well defined location in the userspace
>    application? Could it be used to attack the application?

Yes -- this is the primary risk in my view. Exposing addresses of any
kind can be a risk. While an unprivileged user may not have direct
access to dmesg, there tend to be many indirect ways to see its
contents. As such, exposing a userspace address (when not then also
terminating the process, as seen with the segv reporting) poses a
potential exposure risk. I admit it's not a LARGE risk, but modern
attacks use these kind of building blocks to construct all the steps to
reaching their target.

> 2. The address shows a location that is being accessed by kernel.
>    Could not it be used to pass a value that might be used to attack
>    kernel?

This is also a risk: it provides feedback about where something may be
as a target within a confused-deputy style attack. (i.e. set up one
process to confuse the kernel, and exploit it from another).

> > > Should a new specifier be added?  If not, should we simply use %px?
> > 
> > There's currently no user of '%pu' (although there is a '%pus'. Perhaps we
> > should have a '%pux'?
> > 
> > I would even state that if it is used, that if makes sure that the value is
> > indeed a user space pointer (goes through the same checks as accessing user
> > space), before its printed, otherwise it shows "(fault)" or something.
> 
> I have mixed feelings about this.
> 
> One one hand, it might make sense to mark locations where userspace
> address is printed. We could easily decide how to print them (hash or
> value) and we could check that it is really from a userspace one.
> 
> But I have few concerns:
> 
> 1. The existing "%pus" has a kind of opposite meaning. It says what
>    address space should be used when the kernel and userspace address
>    space is overlapping.
> 
> 2. There is the history with "%pk". It did not work because people did
>    not use it.
> 
> 3. I am not sure about the output when the address is not from
>    userspace. Printing ("fault") is not much helpful. Printing
>    hashed value might be confusing. Well, I am still not sure
>    that it is really safe to print real userspace addresses
>    by default.

I think this should just be %px. Or better yet, not printed at all. See
Linus's prior comments:
https://www.kernel.org/doc/html/latest/process/deprecated.html#p-format-specifier

-- 
Kees Cook

Powered by blists - more mailing lists