Re: [PATCH] mm/userfaultfd: fix memory corruption due to writeprotect

[Yu Zhao <yuzhao@google.com>]

On Sat, Dec 19, 2020 at 01:34:29PM -0800, Nadav Amit wrote:
> [ cc’ing some more people who have experience with similar problems ]
> 
> > On Dec 19, 2020, at 11:15 AM, Andrea Arcangeli <aarcange@redhat.com> wrote:
> > 
> > Hello,
> > 
> > On Fri, Dec 18, 2020 at 08:30:06PM -0800, Nadav Amit wrote:
> >> Analyzing this problem indicates that there is a real bug since
> >> mmap_lock is only taken for read in mwriteprotect_range(). This might
> > 
> > Never having to take the mmap_sem for writing, and in turn never
> > blocking, in order to modify the pagetables is quite an important
> > feature in uffd that justifies uffd instead of mprotect. It's not the
> > most important reason to use uffd, but it'd be nice if that guarantee
> > would remain also for the UFFDIO_WRITEPROTECT API, not only for the
> > other pgtable manipulations.
> > 
> >> Consider the following scenario with 3 CPUs (cpu2 is not shown):
> >> 
> >> cpu0				cpu1
> >> ----				----
> >> userfaultfd_writeprotect()
> >> [ write-protecting ]
> >> mwriteprotect_range()
> >> mmap_read_lock()
> >> change_protection()
> >>  change_protection_range()
> >>   ...
> >>   change_pte_range()
> >>   [ defer TLB flushes]
> >> 				userfaultfd_writeprotect()
> >> 				 mmap_read_lock()
> >> 				 change_protection()
> >> 				 [ write-unprotect ]
> >> 				 ...
> >> 				  [ unprotect PTE logically ]
> >> 				...
> >> 				[ page-fault]
> >> 				...
> >> 				wp_page_copy()
> >> 				[ set new writable page in PTE]

I don't see any problem in this example -- wp_page_copy() calls
ptep_clear_flush_notify(), which should take care of the stale entry
left by cpu0.

That being said, I suspect the memory corruption you observed is
related this example, with cpu1 running something else that flushes
conditionally depending on pte_write().

Do you know which type of pages were corrupted? file, anon, etc.

> > Can't we check mm_tlb_flush_pending(vma->vm_mm) if MM_CP_UFFD_WP_ALL
> > is set and do an explicit (potentially spurious) tlb flush before
> > write-unprotect?
> 
> There is a concrete scenario that I actually encountered and then there is a
> general problem.
> 
> In general, the kernel code assumes that PTEs that are read from the
> page-tables are coherent across all the TLBs, excluding permission promotion
> (i.e., the PTE may have higher permissions in the page-tables than those
> that are cached in the TLBs).
> 
> We therefore need to both: (a) protect change_protection_range() from the
> changes of others who might defer TLB flushes without taking mmap_sem for
> write (e.g., try_to_unmap_one()); and (b) to protect others (e.g.,
> page-fault handlers) from concurrent changes of change_protection().
> 
> We have already encountered several similar bugs, and debugging such issues
> s time consuming and these bugs impact is substantial (memory corruption,
> security). So I think we should only stick to general solutions.
> 
> So perhaps your the approach of your proposed solution is feasible, but it
> would have to be applied all over the place: we will need to add a check for
> mm_tlb_flush_pending() and conditionally flush the TLB in every case in
> which PTEs are read and there might be an assumption that the
> access-permission reflect what the TLBs hold. This includes page-fault
> handlers, but also NUMA migration code in change_protection(), softdirty
> cleanup in clear_refs_write() and maybe others.
> 
> [ I have in mind another solution, such as keeping in each page-table a 
> “table-generation” which is the mm-generation at the time of the change,
> and only flush if “table-generation”==“mm-generation”, but it requires
> some thought on how to avoid adding new memory barriers. ]
> 
> IOW: I think the change that you suggest is insufficient, and a proper
> solution is too intrusive for “stable".
> 
> As for performance, I can add another patch later to remove the TLB flush
> that is unnecessarily performed during change_protection_range() that does
> permission promotion. I know that your concern is about the “protect” case
> but I cannot think of a good immediate solution that avoids taking mmap_lock
> for write.
> 
> Thoughts?
>