Re: [PATCH 4/9] userfaultfd/shmem: support UFFDIO_CONTINUE for shmem

[Axel Rasmussen]

On Mon, Apr 12, 2021 at 4:17 PM Peter Xu <peterx@xxxxxxxxxx> wrote:
>
> On Thu, Apr 08, 2021 at 04:43:22PM -0700, Axel Rasmussen wrote:
> > +/*
> > + * Install PTEs, to map dst_addr (within dst_vma) to page.
> > + *
> > + * This function handles MCOPY_ATOMIC_CONTINUE (which is always file-backed),
> > + * whether or not dst_vma is VM_SHARED. It also handles the more general
> > + * MCOPY_ATOMIC_NORMAL case, when dst_vma is *not* VM_SHARED (it may be file
> > + * backed, or not).
> > + *
> > + * Note that MCOPY_ATOMIC_NORMAL for a VM_SHARED dst_vma is handled by
> > + * shmem_mcopy_atomic_pte instead.
> > + */
> > +static int mcopy_atomic_install_ptes(struct mm_struct *dst_mm, pmd_t *dst_pmd,
> > +                                  struct vm_area_struct *dst_vma,
> > +                                  unsigned long dst_addr, struct page *page,
> > +                                  bool newly_allocated, bool wp_copy)
> > +{
> > +     int ret;
> > +     pte_t _dst_pte, *dst_pte;
> > +     int writable;
> > +     bool vm_shared = dst_vma->vm_flags & VM_SHARED;
> > +     spinlock_t *ptl;
> > +     struct inode *inode;
> > +     pgoff_t offset, max_off;
> > +
> > +     _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
> > +     writable = dst_vma->vm_flags & VM_WRITE;
> > +     /* For private, non-anon we need CoW (don't write to page cache!) */
> > +     if (!vma_is_anonymous(dst_vma) && !vm_shared)
> > +             writable = 0;
> > +
> > +     if (writable || vma_is_anonymous(dst_vma))
> > +             _dst_pte = pte_mkdirty(_dst_pte);
> > +     if (writable) {
> > +             if (wp_copy)
> > +                     _dst_pte = pte_mkuffd_wp(_dst_pte);
> > +             else
> > +                     _dst_pte = pte_mkwrite(_dst_pte);
> > +     } else if (vm_shared) {
> > +             /*
> > +              * Since we didn't pte_mkdirty(), mark the page dirty or it
> > +              * could be freed from under us. We could do this
> > +              * unconditionally, but doing it only if !writable is faster.
> > +              */
> > +             set_page_dirty(page);
> > +     }
> > +
> > +     dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
> > +
> > +     if (vma_is_shmem(dst_vma)) {
> > +             /* The shmem MAP_PRIVATE case requires checking the i_size */
>
> When you start to use this function in the last patch it'll be needed too even
> if MAP_SHARED?
>
> How about directly state the reason of doing this ("serialize against truncate
> with the PT lock") instead of commenting about "who will need it"?
>
> > +             inode = dst_vma->vm_file->f_inode;
> > +             offset = linear_page_index(dst_vma, dst_addr);
> > +             max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
> > +             ret = -EFAULT;
> > +             if (unlikely(offset >= max_off))
> > +                     goto out_unlock;
> > +     }
>
> [...]
>
> > +/* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
> > +static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
> > +                             pmd_t *dst_pmd,
> > +                             struct vm_area_struct *dst_vma,
> > +                             unsigned long dst_addr,
> > +                             bool wp_copy)
> > +{
> > +     struct inode *inode = file_inode(dst_vma->vm_file);
> > +     pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
> > +     struct page *page;
> > +     int ret;
> > +
> > +     ret = shmem_getpage(inode, pgoff, &page, SGP_READ);
>
> SGP_READ looks right, as we don't want page allocation.  However I noticed
> there's very slight difference when the page was just fallocated:
>
>         /* fallocated page? */
>         if (page && !PageUptodate(page)) {
>                 if (sgp != SGP_READ)
>                         goto clear;
>                 unlock_page(page);
>                 put_page(page);
>                 page = NULL;
>                 hindex = index;
>         }
>
> I think it won't happen for your case since the page should be uptodate already
> (the other thread should check and modify the page before CONTINUE), but still
> raise this up, since if the page was allocated it smells better to still
> install the fallocated page (do we need to clear the page and SetUptodate)?

Sorry for the somewhat rambling thought process:

My first thought is, I don't really know what PageUptodate means for
shmem pages. If I understand correctly, normally we say PageUptodate()
if the in memory data is more recent or equivalent to the on-disk
data. But, shmem pages are entirely in memory - they are file backed
in name only, in some sense.

fallocate() does all sorts of things so the comment to me seems a bit
ambiguous, but it seems the implication is that we're worried
specifically about the case where the shmem page was recently
allocated with fallocate(mode=0)? In that case, do we use
!PageUptodate() to denote that the page has been allocated, but its
contents are undefined?

I suppose that would make sense, as the action "goto clear;" generally
memset()-s the page to zero it, and then calls SetPageUptodate().

Okay so let's say the following sequence of events happens:

1. Userspace calls fallocate(mode=0) to allocate some shmem pages.
2. Another thread, via a UFFD-registered mapping, manages to trigger a
minor fault on one such page, while we still have !PageUptodate().
(I'm not 100% sure this can happen, but let's say it can.)
3. UFFD handler thread gets the minor fault event, and for whatever
(buggy?) reason does nothing - it doesn't modify the page, it just
calls CONTINUE.

I think if we get to this point, zeroing the page, returning it, and
setting up the PTEs seems somewhat reasonable to me. I suppose
alternatively we could notice that this happened and return an error
to the caller? I'm hesitant to mess with the behavior of
shmem_getpage_gfp() to make such a thing happen though. I do think if
we're going to set up the PTEs instead of returning an error, we
definitely do need to clear and SetPageUptodate() the page first.

In conclusion, I think this behavior is correct.

>
> --
> Peter Xu
>