Messages in this thread | | | From | Andy Lutomirski <> | Date | Thu, 22 Sep 2016 15:50:47 -0700 | Subject | Re: [PATCH 09/12] x86/process: Pin the target stack in get_wchan() |
| |
On Thu, Sep 22, 2016 at 3:44 PM, Andy Lutomirski <luto@amacapital.net> wrote: > On Fri, Sep 16, 2016 at 7:00 PM, Jann Horn <jann@thejh.net> 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@gmail.com> 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@cmpxchg.org> 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
| |