Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Tue, 28 Jun 2011 15:53:19 -0700
From: Andrew Morton <>
To: Linus Torvalds <>
Cc: Vasiliy Kulikov <>,,
Subject: Re: [Security] CVE request: kernel: taskstats/procfs io infoleak
 (was: taskstats authorized_keys presence infoleak PoC)

On Sun, 26 Jun 2011 19:57:23 -0700 Linus Torvalds <> wrote:

> On Fri, Jun 24, 2011 at 5:34 AM, Vasiliy Kulikov <> wrote:
> >
> > I think it needs 2 CVE, one for /proc/PID/io and another for taskstats.
> Hmm. Should we just round them down to 1kB boundaries or something?
> People *do* want to know about IO accounting, but I agree that giving
> things at a byte granularity ends up giving way too much information.
> When you can see how many bytes something read off a tty, that's a
> problem.
> Returning accounting information at a 1k granularity should make it
> impractical to use that to guess keys etc. It still gives *some*
> information (and enough for rough statistics), but it doesn't give the
> level of detail required for any simple attack.
> Sometimes excessive precision isn't a good thing.
> Andrew - the IO_ACCT stuff went through you (back in 2006), the
> taskstats did too, methinks. Comments?

Random thoughts:

a) I haven't thought very hard about it, but isn't it the case that
   fuzzifying the byte counts in this manner will still permit the
   length of these things to be determined, albeit with a larger data

b) Where does the problem lie?  Is it with the kernel, which exposes
   accurate accounting?  Or is it with userspace, which accidentally
   exposes sensitive information by failing to account for the kernel's
   exposure of accurate accounting information?  

   - Assumes that userspace can be changed to obscure this
     information.  Erroneously, I think ;)

c) Should this information be world-readable?  Perhaps we should add
   more rational privileges here.  Back-compatibility issues.

If rounding the counts to a 1k granularity will indeed defeat the
attack (I'm unsure) then I'd suggest that a fix would be to perform
that fuzzification if the receiving process doesn't have suitable
permissions.  So if the user is reading his own stats or is root, he
still gets byte-resolution results.  This keeps the stats as useful as
we can make them and reduces the back-compatibility damage.

What might be the extent of the back-compatibility damage?  It's hard
to believe that anyone would care about a 1k error in bulk IO stats. 
But if there's someone out there who uses these interfaces to detect
whether the monitored task is doing *anything* then we'll break them. 
eg, "did my data logging task just receive a packet from my
scintillator experiment".  

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.