Bug 144412

*** This bug has been split off bug 144403 ***

------- Original comment by Josh Bressers (Security Response Team) on 2005.01.06
14:56 -------

This has been reported to vendor-sec by Paul Starzetz

I have found an exploitable flaw in the page fault handler, however only
in the SMP case.

The problem is this:

[A]     down_read(&mm->mmap_sem);

        vma = find_vma(mm, address);
        if (!vma)
                goto bad_area;
        if (vma->vm_start <= address)
                goto good_area;
        if (!(vma->vm_flags & VM_GROWSDOWN))
                goto bad_area;
        if (error_code & 4) {
                /*
                 * accessing the stack below %esp is always a bug.
                 * The "+ 32" is there due to some instructions (like
                 * pusha) doing post-decrement on the stack and that
                 * doesn't show up until later..
                 */
                if (address + 32 < regs->esp)
                        goto bad_area;
        }
        if (expand_stack(vma, address))
[B]             goto bad_area;


an exploitable race scenario looks as follows:

1) one thread issues down_write on the sem (remap, madvise, ...)

2) two other threads faul below a VM_GROWSDOWN segment (note that they can
fault anywhere below vm_start since esp is arbitrary) and sleep in [A]

3) first thread releases the sem and the two others run again, both find
the same VMA but:

thread1 ----------F1----------[   VMA   ]
thread2 ---------------F2-----[   VMA   ]

where F1/2 faul address.

If timed carefully we get:

thread1 expands stack to F1, installs pte1
thread2 expands stack to F2, installs pte2

resulting in pte1 not covered by the VMA. Techniques like in mremap_pte
can be applied to further exploit this condition.

Please do not argue that the race window is small - I have seen even
smaller windows opening like a barn :-]

The critical section is only from [A] to [B], we do not care about timings
of handle_mm_fault etc, since the VMA is later consulted only for page
flags.

Note that this also races with ptrace/proc etc (everything using
access_process_vm/get_user_pages).