lkml.org 
[lkml]   [2013]   [Oct]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH v2 0/9] procfs: protect /proc/<pid>/* files with file->f_cred
On Thu, Oct 3, 2013 at 1:29 PM, Djalal Harouni <tixxdz@opendz.org> wrote:
> On Thu, Oct 03, 2013 at 08:12:44AM +0200, Ingo Molnar wrote:
>>
>> * Djalal Harouni <tixxdz@opendz.org> wrote:
>>
>> > > Regardless, glibc uses /proc/self/maps, which would be fine here, right?
>> >
>> > I did not touch /proc/self/maps and others, but I'm planning to fix them
>> > if this solution is accepted.
>> >
>> > I'll do the same thing as in /proc/*/stat for maps, let it be 0444, and
>> > try to delay the check to ->read(). So during ->read() perform
>> > ptrace_may_access() on currenct's cred and process_allow_access() on
>> > file's opener cred. This should work.
>>
>> Aren't read() time checks fundamentally unrobust? We generally don't do
>> locking on read()s, so such checks - in the general case - are pretty
>> racy.
> For procfs yes, read() time checks are unrobust, the general agreement on
> procfs is checks should be performed during each syscall.
>
> For the locking on read()/write() IMHO there should be locking by design
> for /proc/pid/* entries. Here we are speaking about content that varies,
> data attached to other processes, so there is already some locking
> mechanism, and for sensitive stuff, we must hold the cred mutex. This
> is the standard from the old days of procfs.
>
>
> And yes some of them are racy, but we can improve it, delay the checks.
>
> From old Linux git history, before the initial git repository build, I
> found that some important checks were done right after gathering the info.
>
>
>> Now procfs might be special, as by its nature of a pseudofilesystem it's
>> far more atomic than other filesystems, but still IMHO it's a lot more
>
>
>> robust security design wise to revoke access to objects you should not
>> have a reference to when a security boundary is crossed, than letting
>> stuff leak through and sprinking all the potential cases that may leak
>> information with permission checks ...
> I agree, but those access should also be checked at the beginning, IOW
> during ->open(). revoke will not help if revoke is not involved at all,
> the task /proc entries may still be valide (no execve).
>
> Currently security boundary is crossed for example arbitrary /proc/*/stack
> (and others).
> 1) The target task does not do an execve (no revoke).
> 2) current task will open these files and *want* and *will* pass the fd to a
> more privileged process to pass the ptrace check which is done only during
> ->read().

What does this? Or are you saying that this is a bad thing?

(And *please* don't write software that *depends* on different
processes having different read()/write() permissions on the *same*
struct file. I've already found multiple privilege escalations based
on that, and I'm pretty sure I can find some more.)

>
>
>> It's also probably faster: security boundary crossings are relatively rare
>> slowpaths, while read()s are much more common...
> Hmm, These two are related? can't get rid of permission checks
> especially on this pseudofilesystem!
>
>
>> So please first get consensus on this fundamental design question before
>> spreading your solution to more areas.
> Of course, I did clean the patchset to prove that it will work, and I
> only implemented full protection for /proc/*/{stack,syscall,stat} other
> files will wait.
>
> But Ingo you can't ignore the fact that:
> /proc/*/{stack,syscall} are 0444 mode
> /proc/*/{stack,syscall} do not have ptrace_may_access() during open()
> /proc/*/{stack,syscall} have the ptrace_may_access() during read()

I think everyone agrees that this is broken. We don't agree on the
fix check. (Also, as described in my other email, your approach may
be really hard to get right.)

--Andy


\
 
 \ /
  Last update: 2013-10-03 19:21    [W:0.106 / U:9.144 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site