Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Wed, 29 Jun 2011 15:11:07 +0400
From: Vasiliy Kulikov <>
To: Linus Torvalds <>
Cc: Andrew Morton <>,,
Subject: Re: [Security] CVE request: kernel: taskstats/procfs io infoleak
 (was: taskstats authorized_keys presence infoleak PoC)

On Tue, Jun 28, 2011 at 17:49 -0700, Linus Torvalds wrote:
> Actually, due to the whole netlink thing, it's not obvious who the
> data goes to,

In send_cpu_listeners() there is a loop over all listeners,
genlmsg_unicast() is called for exclusively for each listening socket.
It's possible to make 2 taskstats structs, one with precise information,
one with rouned information.

> If you want the exact thing, you can use /proc/<pid>/io, which now
> does the security checking as per Vasiliy.

The patch lacks proper locking against a race with exec (noticed by
KOSAKI).  task->signal->cred_guard_mutex should be fine, but I hesitate
whether it's fine to mix it with lock_task_sighand() and if mix then in
what order.  If keeping ->cred_guard_mutex prevents theads from exiting
then sighand is redundant.

> So some patch like the appended?

1) The filtering on exit looks OK, but fill_stats_for_tgid() is not filtered:

	if (first->signal->stats)
		memcpy(stats, first->signal->stats, sizeof(*stats));
		memset(stats, 0, sizeof(*stats));

2) syscalls counts is probably needs another rounding constant, it is
not measures in kbs.  However, 1024 might be OK if round char number by

> Vasiliy, this is different from your
> 2/2, but it's simpler and I think sufficient. And shouldn't break
> iotop. What do you think? I agree that it's not perfect, but it seems
> to be sufficient at least for the particular passwd attack, no?

Indeed, such rounding does break this specific exploit.

> Or is
> there some way you can fool sshd to read some other user-supplied data
> so that you can trick it into giving multiple values that you control,
> and thus see exactly when the IO counts overflow..

I'm trying to find a way to bypass 1k rounding.  I see 2 abstract ways:

1) a program generates X bytes io traffic for every 1 byte of sensitive
information.  X should be as close to kb boundary as possible.

2) as you say here:


If CONST is known and CONTROLLABLE is controlled by an attacker then he
may find C1 and C1+1 generating X kb - 1 and (X+1) kb traffic,
respectively, revealing len(SENSITIVE).

I cannot find vulnerable programs now, but I believe there are some of them
among widespread programs.

The core problem here is that by giving *some part* of information about
internal task activity the kernel violating the task privacy, strictly
speaking.  A program doing IO expects this activity to be kept private.
This revealted part may or may not reveal sensible information, depends
on the specific program.


Vasiliy Kulikov - bringing security into open computing environments

Powered by blists - more mailing lists

Please check out the Open Source Software Security Wiki, which is counterpart to this mailing list.

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.