On Mon, Apr 02, 2018 at 07:10:58AM -0700, Matthew Wilcox wrote:
> 
> Souptick and I have been auditing the various page fault handler routines
> and we've noticed that graphics drivers assume that a signal should be
> able to interrupt a page fault.  In contrast, the page cache takes great
> care to allow only fatal signals to interrupt a page fault.
> 
> I believe (but have not verified) that a non-fatal signal being delivered
> to a task which is in the middle of a page fault may well end up in an
> infinite loop, attempting to handle the page fault and failing forever.
> 
> Here's one of the simpler ones:
> 
>         ret = mutex_lock_interruptible(&etnaviv_obj->lock);
>         if (ret)
>                 return VM_FAULT_NOPAGE;
> 
> (many other drivers do essentially the same thing including i915)
> 
> On seeing NOPAGE, the fault handler believes the PTE is in the page
> table, so does nothing before it returns to arch code at which point
> I get lost in the magic assembler macros.  I believe it will end up
> returning to userspace if the signal is non-fatal, at which point it'll
> go right back into the page fault handler, and mutex_lock_interruptible()
> will immediately fail.  So we've converted a sleeping lock into the most
> expensive spinlock.
> 
> I don't think the graphics drivers really want to be interrupted by
> any signal.  I think they want to be interruptible by fatal signals
> and should use the mutex_lock_killable / fatal_signal_pending family of
> functions.  That's going to be a bit of churn, funnelling TASK_KILLABLE
> / TASK_INTERRUPTIBLE all the way down into the dma-fence code.  Before
> anyone gets started on that, I want to be sure that my analysis is
> correct, and the drivers are doing the wrong thing by using interruptible
> waits in a page fault handler.

So we've done some experiments for the case where the fault originated
from kernel context (copy_to|from_user and friends). The fixup code seems
to retry the copy once after the fault (in copy_user_handle_tail), if that
fails again we get a short read/write. This might result in an -EFAULT,
short read()/write() or anything else really, depending upon the syscall
api.

Except in some code paths in gpu drivers where we convert anything into
-ERESTARTSYS/EINTR if there's a signal pending it won't ever result in the
syscall getting restarted (well except maybe short read/writes if
userspace bothers with that).

So I guess gpu fault handlers indeed break the kernel's expectations, but
then I think we're getting away with that because the inner workings of
gpu memory objects is all heavily abstracted away by opengl/vulkan and
friends.

I guess what we could do is try to only do killable sleeps if it's a
kernel fault, but that means wiring a flag through all the callchains. Not
pretty. Except when there's a magic set of functions that would convert
all interruptible sleeps to killable ones only for us.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Reply via email to