Re: [RFC PATCH 03/13] selftests/vm/userfaultfd: wake after copy failure

[Peter Xu <peterx@redhat.com>]

On Mon, Dec 21, 2020 at 07:51:52PM +0000, Nadav Amit wrote:
> > On Dec 21, 2020, at 11:28 AM, Peter Xu <peterx@redhat.com> wrote:
> > 
> > On Sat, Nov 28, 2020 at 04:45:38PM -0800, Nadav Amit wrote:
> >> From: Nadav Amit <namit@vmware.com>
> >> 
> >> When userfaultfd copy-ioctl fails since the PTE already exists, an
> >> -EEXIST error is returned and the faulting thread is not woken. The
> >> current userfaultfd test does not wake the faulting thread in such case.
> >> The assumption is presumably that another thread set the PTE through
> >> copy/wp ioctl and would wake the faulting thread or that alternatively
> >> the fault handler would realize there is no need to "must_wait" and
> >> continue. This is not necessarily true.
> >> 
> >> There is an assumption that the "must_wait" tests in handle_userfault()
> >> are sufficient to provide definitive answer whether the offending PTE is
> >> populated or not. However, userfaultfd_must_wait() test is lockless.
> >> Consequently, concurrent calls to ptep_modify_prot_start(), for
> >> instance, can clear the PTE and can cause userfaultfd_must_wait()
> >> to wrongly assume it is not populated and a wait is needed.
> > 
> > Yes userfaultfd_must_wait() is lockless, however my understanding is that we'll
> > enqueue before reading the page table, which seems to me that we'll always get
> > notified even the race happens.  Should apply to either UFFDIO_WRITEPROTECT or
> > UFFDIO_COPY, iiuc, as long as we follow the order of (1) modify pgtable (2)
> > wake sleeping threads.  Then it also means that when must_wait() returned true,
> > it should always get waked up when fault resolved.
> > 
> > Taking UFFDIO_COPY as example, even if UFFDIO_COPY happen right before
> > must_wait() calls:
> > 
> >       worker thread                       uffd thread
> >       -------------                       -----------
> > 
> >   handle_userfault
> >    spin_lock(fault_pending_wqh)
> >    enqueue()
> >    set_current_state(INTERRUPTIBLE)
> >    spin_unlock(fault_pending_wqh)
> >    must_wait()
> >      lockless walk page table
> >                                           UFFDIO_COPY
> >                                             fill in the hole
> >                                             wake up threads
> >                                               (this will wake up worker thread too?)
> >    schedule()
> >      (which may return immediately?)
> > 
> > While here fault_pending_wqh is lock protected. I just feel like there's some
> > other reason to cause the thread to stall.  Or did I miss something?
> 
> But what happens if the copy completed before the enqueuing? Assume
> the page is write-protected during UFFDIO_COPY:
> 
> 
> cpu0					cpu1		
> ----					----			
> handle_userfault
> 					UFFDIO_COPY
> 					[ write-protected ]
> 				 	 fill in the hole
> 				 	 wake up threads
> 				 	 [nothing to wake]
> 							
> 					UFFD_WP (unprotect)
> 					 logically marks as unprotected
> 					 [nothing to wake]
> 
>  spin_lock(fault_pending_wqh)
>   enqueue()
>   set_current_state(INTERRUPTIBLE)
>   spin_unlock(fault_pending_wqh)
>   must_wait()
> 
> 					[ #PF on the same PTE
> 					 due to write-protection ]
> 
> 					...
> 					 wp_page_copy()
> 					  ptep_clear_flush_notify()
> 					  [ PTE is clear ]
> 					
>    lockless walk page table
>     pte_none(*pte) -> must wait
> 
> Note that additional scenarios are possible. For instance, instead of
> wp_page_copy(), we can have other change_pte_range() (due to worker’s
> mprotect() or NUMA balancing), calling ptep_modify_prot_start() and clearing
> the PTE.
> 
> Am I missing something?

Ah I see your point, thanks.  I think you're right:

Reviewed-by: Peter Xu <peterx@redhat.com>

Would you mind adding something like above into the commit message if you're
going to repost?  IMHO it would even be nicer to mention why
UFFDIO_WRITEPROTECT does not need this extra wakeup (I think it's because it'll
do the wakeup unconditionally anyway).

-- 
Peter Xu