Re: [PATCH 09/12] x86/process: Pin the target stack in get_wchan()

[Andy Lutomirski]

On Thu, Sep 22, 2016 at 3:44 PM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
> On Fri, Sep 16, 2016 at 7:00 PM, Jann Horn <jann@xxxxxxxxx> wrote:
>> On Tue, Sep 13, 2016 at 02:29:29PM -0700, Andy Lutomirski wrote:
>>> This will prevent a crash if get_wchan() runs after the task stack
>>> is freed.
>>
>> I think I found some more stuff. Have a look at KSTK_EIP() and KSTK_ESP(), I think
>> they read from the saved userspace registers area at the top of the kernel stack?
>>
>> Used on remote processes in:
>>   vma_is_stack_for_task() (via /proc/$pid/maps)
>
> This isn't used in /proc/$pid/maps -- it's only used in
> /proc/$pid/task/$tid/maps.  I wonder if anyone actually cares about it
> -- it certainly won't work reliably.
>
> I could pin the stack in vma_is_stack_for_task, but it seems
> potentially better to me to change it to vma_is_stack_for_current()
> and remove the offending caller in /proc, replacing it with "return
> 0".  Thoughts?

The history here is strange:

Before March 2012, we used to only consider the "stack" of the mm --
that is, the VMA that the mm actually treats as a stack.  Then:

commit b76437579d1344b612cf1851ae610c636cec7db0
Author: Siddhesh Poyarekar <siddhesh.poyarekar@xxxxxxxxx>
Date:   Wed Mar 21 16:34:04 2012 -0700

    procfs: mark thread stack correctly in proc/<pid>/maps

and we did something extra horrible to try to find out whose stack was
where.  This got partially reverted by:

commit 65376df582174ffcec9e6471bf5b0dd79ba05e4a
Author: Johannes Weiner <hannes@xxxxxxxxxxx>
Date:   Tue Feb 2 16:57:29 2016 -0800

    proc: revert /proc/<pid>/maps [stack:TID] annotation

and now we're in the current situation where it's fast but still racy.

Any objection if I finish reverting the patch and restore the pre-2012
behavior?  Frankly, I wouldn't mind trying to excise KSTK_EIP and
KSTK_ESP from the kernel entirely, to be replaced with
current_user_sp() and current_user_ip().  Having those macros around
seems likely to make people think they're safe to use.

>
>>   do_task_stat() (/proc/$pid/stat)
>
> Like this:
>
>         mm = get_task_mm(task);
>         if (mm) {
>                 vsize = task_vsize(mm);
>                 if (permitted) {
>                         eip = KSTK_EIP(task);
>                         esp = KSTK_ESP(task);
>                 }
>         }
>
> Can we just delete this outright?  It seems somewhere between mostly
> and entirely useless, and it also seems dangerous.  Until very
> recently, on x86_64, this would have been a potential info leak, as
> SYSCALL followed closely by a hardware interrupt would cause *kernel*
> values to land in task_pt_regs().  I don't even want to think about
> what this code does if the task is in vm86 mode.  I wouldn't be at all
> surprised if non-x86 architectures have all kinds of interesting
> thinks happen if you do this to a task that isn't running normal
> non-atomic kernel code at the time.
>
> I would advocate for unconditionally returning zeros in these two stat fields.

--Andy