Re: [PATCH 00/23] userfaultfd-wp: Support shmem and hugetlbfs

[Peter Xu]

On Mon, Mar 22, 2021 at 08:48:49PM -0400, Peter Xu wrote:
> This patchset is based on tag v5.12-rc3-mmots-2021-03-17-22-26.  To run the
> selftest, need to apply the two patches to fix minor mode page leak:
> 
> https://lore.kernel.org/lkml/20210322175132.36659-1-peterx@xxxxxxxxxx/
> https://lore.kernel.org/lkml/20210322204836.1650221-1-axelrasmussen@xxxxxxxxxx/
> 
> Since I didn't get any NACK in the previous RFC series for months, I decided to
> remove the RFC tag starting from this version, so this is v1 of uffd-wp support
> on shmem & hugetlb.
> 
> The whole series can also be found online [1].
> 
> The major comment I'd like to get is on the new idea of swap special pte.  That
> comes from suggestions from both Hugh and Andrea and I appreciated a lot for
> those discussions.
> 
> In short, the so-called "swap special pte" in this patchset is a new type of
> pte that doesn't exist in the past, but it got used initially in this series in
> file-backed memories.  It is used to persist information even if the ptes got
> dropped meanwhile when the page cache still existed.  For example, when
> splitting a file-backed huge pmd, we could be simply dropping the pmd entry
> then wait until another fault coming.  It's okay in the past since all
> information in the pte can be retained from the page cache when the next page
> fault triggers.  However in this case, uffd-wp is per-pte information which
> cannot be kept in page cache, so that information needs to be maintained
> somehow still in the pgtable entry, even if the pgtable entry is going to be
> dropped.  Here instead of replacing with a none entry, we used the "swap
> special pte".  Then when the next page fault triggers, we can observe orig_pte
> to retain this information.
> 
> I'm copy-pasting some commit message from the patch "mm/swap: Introduce the
> idea of special swap ptes", where it tried to explain this pte in another angle:
> 
>     We used to have special swap entries, like migration entries, hw-poison
>     entries, device private entries, etc.
> 
>     Those "special swap entries" reside in the range that they need to be at least
>     swap entries first, and their types are decided by swp_type(entry).
> 
>     This patch introduces another idea called "special swap ptes".
> 
>     It's very easy to get confused against "special swap entries", but a speical
>     swap pte should never contain a swap entry at all.  It means, it's illegal to
>     call pte_to_swp_entry() upon a special swap pte.
> 
>     Make the uffd-wp special pte to be the first special swap pte.
> 
>     Before this patch, is_swap_pte()==true means one of the below:
> 
>        (a.1) The pte has a normal swap entry (non_swap_entry()==false).  For
>              example, when an anonymous page got swapped out.
> 
>        (a.2) The pte has a special swap entry (non_swap_entry()==true).  For
>              example, a migration entry, a hw-poison entry, etc.
> 
>     After this patch, is_swap_pte()==true means one of the below, where case (b) is
>     added:
> 
>      (a) The pte contains a swap entry.
> 
>        (a.1) The pte has a normal swap entry (non_swap_entry()==false).  For
>              example, when an anonymous page got swapped out.
> 
>        (a.2) The pte has a special swap entry (non_swap_entry()==true).  For
>              example, a migration entry, a hw-poison entry, etc.
> 
>      (b) The pte does not contain a swap entry at all (so it cannot be passed
>          into pte_to_swp_entry()).  For example, uffd-wp special swap pte.
> 
> Hugetlbfs needs similar thing because it's also file-backed.  I directly reused
> the same special pte there, though the shmem/hugetlb change on supporting this
> new pte is different since they don't share code path a lot.
> 
> Patch layout
> ============
> 
> Part (1): Shmem support, this is where the special swap pte is introduced.
> Some zap rework is needed within the process:
> 
>   shmem/userfaultfd: Take care of UFFDIO_COPY_MODE_WP
>   mm: Clear vmf->pte after pte_unmap_same() returns
>   mm/userfaultfd: Introduce special pte for unmapped file-backed mem
>   mm/swap: Introduce the idea of special swap ptes
>   shmem/userfaultfd: Handle uffd-wp special pte in page fault handler
>   mm: Drop first_index/last_index in zap_details
>   mm: Introduce zap_details.zap_flags
>   mm: Introduce ZAP_FLAG_SKIP_SWAP
>   mm: Pass zap_flags into unmap_mapping_pages()
>   shmem/userfaultfd: Persist uffd-wp bit across zapping for file-backed
>   shmem/userfaultfd: Allow wr-protect none pte for file-backed mem
>   shmem/userfaultfd: Allows file-back mem to be uffd wr-protected on thps
>   shmem/userfaultfd: Handle the left-overed special swap ptes
>   shmem/userfaultfd: Pass over uffd-wp special swap pte when fork()
> 
> Part (2): Hugetlb support, we need to disable huge pmd sharing for uffd-wp
> because not compatible just like uffd minor mode.  The rest is the changes
> required to teach hugetlbfs understand the special swap pte too that introduced
> with the uffd-wp change:
> 
>   hugetlb/userfaultfd: Hook page faults for uffd write protection
>   hugetlb/userfaultfd: Take care of UFFDIO_COPY_MODE_WP
>   hugetlb/userfaultfd: Handle UFFDIO_WRITEPROTECT
>   hugetlb: Pass vma into huge_pte_alloc()
>   hugetlb/userfaultfd: Forbid huge pmd sharing when uffd enabled
>   mm/hugetlb: Introduce huge version of special swap pte helpers
>   mm/hugetlb: Move flush_hugetlb_tlb_range() into hugetlb.h
>   hugetlb/userfaultfd: Unshare all pmds for hugetlbfs when register wp
>   hugetlb/userfaultfd: Handle uffd-wp special pte in hugetlb pf handler
>   hugetlb/userfaultfd: Allow wr-protect none ptes
>   hugetlb/userfaultfd: Only drop uffd-wp special pte if required
> 
> Part (3): Enable both features in code and test
> 
>   userfaultfd: Enable write protection for shmem & hugetlbfs
>   userfaultfd/selftests: Enable uffd-wp for shmem/hugetlbfs
> 
> Tests
> =========
> 
> I've tested it using either userfaultfd kselftest program, but also with
> umapsort [2] which should be even stricter.  Tested page swapping in/out during
> umapsort.
> 
> If anyone would like to try umapsort, need to use an extremely hacked version
> of umap library [3], because by default umap only supports anonymous.  So to
> test it we need to build [3] then [2].
> 
> Any comment would be greatly welcomed.  Thanks,
> 
> [1] https://github.com/xzpeter/linux/tree/uffd-wp-shmem-hugetlbfs
> [2] https://github.com/LLNL/umap-apps
> [3] https://github.com/xzpeter/umap/tree/peter-shmem-hugetlbfs

Hugh, Mike, Andrew,

Any comment for this series?

Thanks,

-- 
Peter Xu