Re: [PATCH v2 0/9] procfs: protect /proc/<pid>/* files with file->f_cred

[Julien Tinnes <jln@google.com>]

On Wed, Oct 2, 2013 at 11:07 AM, Kees Cook <keescook@chromium.org> wrote:
>
> On Wed, Oct 2, 2013 at 11:00 AM, Andy Lutomirski <luto@amacapital.net> wrote:
> > On Wed, Oct 2, 2013 at 10:48 AM, Kees Cook <keescook@chromium.org> wrote:
> >> On Wed, Oct 2, 2013 at 9:51 AM, Andy Lutomirski <luto@amacapital.net> wrote:
> >>> On Wed, Oct 2, 2013 at 3:37 PM, Djalal Harouni <tixxdz@opendz.org> wrote:
> >>>> On Tue, Oct 01, 2013 at 06:40:41PM -0700, Andy Lutomirski wrote:
> >>>>> On 10/01/2013 01:26 PM, Djalal Harouni wrote:
> >>>>> > /proc/<pid>/* entries varies at runtime, appropriate permission checks
> >>>>> > need to happen during each system call.
> >>>>> >
> >>>>> > Currently some of these sensitive entries are protected by performing
> >>>>> > the ptrace_may_access() check. However even with that the /proc file
> >>>>> > descriptors can be passed to a more privileged process
> >>>>> > (e.g. a suid-exec) which will pass the classic ptrace_may_access()
> >>>>> > check. In general the ->open() call will be issued by an unprivileged
> >>>>> > process while the ->read(),->write() calls by a more privileged one.
> >>>>> >
> >>>>> > Example of these files are:
> >>>>> > /proc/*/syscall, /proc/*/stack etc.
> >>>>> >
> >>>>> > And any open(/proc/self/*) then suid-exec to read()/write() /proc/self/*
> >>>>> >
> >>>>> >
> >>>>> > These files are protected during read() by the ptrace_may_access(),
> >>>>> > however the file descriptor can be passed to a suid-exec which can be
> >>>>> > used to read data and bypass ASLR. Of course this was discussed several
> >>>>> > times on LKML.
> >>>>>
> >>>>> Can you elaborate on what it is that you're fixing?  That is, can you
> >>>>> give a concrete example of what process opens what file and passes the
> >>>>> fd to what process?
> >>>> Yes, the references were already given in this email:
> >>>> https://lkml.org/lkml/2013/8/31/209
> >>>>
> >>>> This has been discussed several times on lkml:
> >>>> https://lkml.org/lkml/2013/8/28/544
> >>>>
> >>>> https://lkml.org/lkml/2013/8/28/564 (check Kees's references)
> >>>>
> >>>>
> >>>>> I'm having trouble following your description.
> >>>> Process open a /proc file and pass the fd to a more privilaged process
> >>>> that will pass the ptrace_may_access() check, while the original process
> >>>> that opened that file should fail at the ptrace_may_access()
> >>>
> >>> So we're talking about two kinds of attacks, right?
> >>
> >> Correct.
> >>
> >>> Type 1: Unprivileged process does something like open("/proc/1/maps",
> >>> O_RDONLY) and then passes the resulting fd to something privileged.
> >>
> >> ... and then leaks contents back to unprivileged process.
> >>
> >>> Type 2: Unprivileged process does something like
> >>> open("/proc/self/maps", O_RDONLY) and then forks.  The parent calls
> >>> execve on something privileged.
> >>
> >> ... and then parent snoops on file contents for the privileged child.
> >>
> >> (Type 2 is solved currently, IIUC. Type 1 could be reduced in scope by
> >> changing these file modes back to 0400.)
> >>
> >>> Can we really not get away with fixing type 1 by preventing these
> >>> files from being opened in the first place and type 2 by revoking all
> >>> of these fds when a privilege-changing exec happens?
> >>
> >> Type 1 can be done via exec as well. Instead of using a priv exec to
> >> read an arbitrary process, read it could read its own.
> >
> > Right.
> >
> >>
> >> I think revoking the fd would be great. Does that mechanism exist?
> >
> > There's this thing that never got merged.
> >
> > http://thread.gmane.org/gmane.linux.kernel/1523331
> >
> > But doing it more directly should be reasonably straightforward.  Either:
> >
> > (a) when a process execs and privileges change, find all the old proc
> > inodes, mark them dead, and unlink them, or
> >
> > (b) add self_exec_id to all the proc file private_data entries (or
> > somewhere else).  Then just make sure that they're unchanged.  I think
> > the bug last time around was because the self_exec_id and struct pid
> > weren't being compared together.
> >
> > (a) is probably nicer.  I don't know if it'll break things.  Linus
> > seemed to think that the Chrome sandbox was sensitive to this stuff,
> > but I don't know why.
>
> I agree, (a) seems much cleaner. Hm, I don't think Chrome does
> anything with these sensitive files (maps, stack, syscall, etc). But
> let's ask Julien. :)
>
> Julien, do you see any problem with Chrome's sandbox behavior if these
> proc files would be unavailable across privilege changes?

There is nothing that currently jumps to mind in Chromium. However,
anything that breaks "file descriptors are capabilities" inevitably
ends-up breaking something.

For instance, I could easily imagine breakage because a process uses
PR_SET_DUMPABLE (more so than, say, transitions to uid 0) while its
/proc entries are being monitored by another part of the same
application.

Please cc:me on patches.

Julien