lkml.org 
[lkml]   [2012]   [Mar]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 0/9] proc: protect /proc/<pid>/* files across execve
On Mon, Mar 12, 2012 at 02:47:43PM -0700, Eric W. Biederman wrote:
> Djalal Harouni <tixxdz@opendz.org> writes:
>
> > On Mon, Mar 12, 2012 at 12:13:15PM -0700, Eric W. Biederman wrote:
> >> Djalal Harouni <tixxdz@opendz.org> writes:
> >>
> >> > Procfs files and other important objects may contain sensitive information
> >> > which must not be seen, inherited or processed across execve.
> >>
> >> So I am dense. /proc/<pid>/mem was special in that it uses a different
> >> set of checks than other files, and to do those access checks
> >> /proc/<pid>/mem needed to look at exec_id.
> > If you are referring to the old protection, yes it was against an ID, but
> > not uniq IDs, so you can execve a suid do some tricks to have a match on IDs
> > and bypass the protection, how: by opening your /proc/self/mem and pass
> > the fd to the exec suid who at read/write time will process its own
> > /proc/self/mem
>
> Yes that case is silly and I don't care. I care that you seem to be
> stomping all over the non-silly cases using the excuse that there
> was one silly case.
I'm not sure I follow you here, these proc files suffer from the same
problems, can you please point what are these non-silly case ?


> >> For all of the access checks that are not written in that silly way.
> >> What is wrong with ptrace_may_access run at every read/write of a file?
> > As it was noted, these files change during runtime, so even if you do the
> > ptrace check at each syscall (which is of course a good thing), you must be
> > sure that you are doing the check on the right target and the processed
> > file belongs really to the appropriate process image of the target.
>
> The right target is by definition the current value of the process in
> question.
>
> /proc/<pid>/<attr> files are supposed to work after an exec.
Yes.

> Adding an exec_id to additional files simply breaks existing
> applications for no good reason.
Can you please point these applications that this patch will break ?
So we do not do it.

I've tested 'ps' but perhaps I've missed something ?

We just return 0 at read() which is a correct thing to do.


> What is needed is for safety is to guard against the race of exec
> happening during a read or a write, so that we don't get access
before and during. I say before since this is what we are tying to
emulate.

> to something we shouldn't have permissions to.
>
> In general that means reference counting or locks. All exec_id can
Locks ? counting yes this perhaps can work as Alan suggested, but a simple
check will catch all the things without any node list nor count inc/dec.

Linus's /proc/pid/mem patch do not even do that, he just keep the old mm
at open time.

> meaningfully be used for in the general case is a trigger to try again.
>
> > This is not news. Alan's historical links:
> > http://lkml.org/lkml/2012/1/29/35
>
> Alan's case refers to how to handle /proc/<pid>/maps the right way.
Alan's case refers to how to handle all the /proc/<pid>/* files and any
other object that should be attached to its process image.

> > The previous discussion on kernel-hardening:
> > (includes some variants described by Vasiliy and other problems which I'll
> > try to discuss here)
> > http://www.openwall.com/lists/kernel-hardening/2012/02/10/1
>
> In your post on openwall I just see you arguing that the current
> and deliberate semantics of the permission checks on the proc files are
> wrong. Because the permission checks happen at access time rather than
> at open time.
Yes if you are speaking about {maps,smaps...} then it should also be at
open time and at each syscall (it depends on files) and for the
/proc/pid/{maps,smaps,numa_maps} yes it is not safe to give non-privileged
processes an fd to an objects attached to a privileged process.

> Well I am sorry. The permission checks have happened at access time for
> ages and that is deliberate.
Yes (perhaps even before I use Linux :) ), but IMHO they are not perfect,
if they were, then we'll not see all the /proc/<pid>/ problems.

The current patches protect /proc/pid/{maps,smaps,numa_maps} without breaking
things, there are no checks at open which is not safe, but I did not add
it since I was not sure and I don't want to break things (glibc
FORITFY_SOURCE ...).

My patches add the ptrace checks at open only for
/proc/<pid>/{environ,pagemap,mem} which is safe, I did not include
/proc/pid/{auxv,maps,smaps,...} so please if you have _real_ cases of
problems just post them.

> Furthermore one of your confused arguments seems to imply that there is
> a such a thing as a /proc/self file distinct from a /proc/<pid> file.
> There not. /proc/self is just a symlink into /proc/<pid> and as such
> does not open new security holes.
No it's not confused, we just use the /proc/self as it is easy to explain.
Sorry if you feel that.

> If you expect /proc/ not to let you find out things about yourself
> if you are a suid executable that just seems silly.
Ah yes this one about suid/privileged, if you are still privileged then
there is no harm, but if you drop your privileges IMHO that means that you
do not want to do privileged operations, but that current == task in
ptrace which is before the creds check will just allow you to do so.

Well, this is another subject.

> So in short you seem to be arguing for changing the semantics of access
> of every file under /proc/<pid> because there is cognitive dissonance
> between your understanding of those files and what is implemented. I
> see no acknowledgement that you are arguing for a semantic change or
> any arguments in favor of that change. At best I see is an argument
> that says you the files don't work the way you expect which is most
> definitely not sufficient for a change.
I really don't understant what you are trying to say/prove here, this is
the same issue of the last /proc/<pid>/mem one that got fixed by Linus and
Olge patches. These are dynamic files.

For arguments I'll re-try.

> Eric
Thanks Eric.

> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/

--
tixxdz
http://opendz.org


\
 
 \ /
  Last update: 2012-03-12 23:41    [W:0.091 / U:0.976 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site