On Tue, Jun 01, 2021 at 02:15:35PM -0700, Hugh Dickins wrote:
> There is a race between THP unmapping and truncation, when truncate sees
> pmd_none() and skips the entry, after munmap's zap_huge_pmd() cleared it,
> but before its page_remove_rmap() gets to decrement compound_mapcount:
> generating false "BUG: Bad page cache" reports that the page is still
> mapped when deleted. This commit fixes that, but not in the way I hoped.
>
> The first attempt used try_to_unmap(page, TTU_SYNC|TTU_IGNORE_MLOCK)
> instead of unmap_mapping_range() in truncate_cleanup_page(): it has often
> been an annoyance that we usually call unmap_mapping_range() with no pages
> locked, but there apply it to a single locked page. try_to_unmap() looks
> more suitable for a single locked page.
>
> However, try_to_unmap_one() contains a VM_BUG_ON_PAGE(!pvmw.pte,page):
> it is used to insert THP migration entries, but not used to unmap THPs.
> Copy zap_huge_pmd() and add THP handling now? Perhaps, but their TLB
> needs are different, I'm too ignorant of the DAX cases, and couldn't
> decide how far to go for anon+swap. Set that aside.
>
> The second attempt took a different tack: make no change in truncate.c,
> but modify zap_huge_pmd() to insert an invalidated huge pmd instead of
> clearing it initially, then pmd_clear() between page_remove_rmap() and
> unlocking at the end. Nice. But powerpc blows that approach out of the
> water, with its serialize_against_pte_lookup(), and interesting pgtable
> usage. It would need serious help to get working on powerpc (with a
> minor optimization issue on s390 too). Set that aside.
>
> Just add an "if (page_mapped(page)) synchronize_rcu();" or other such
> delay, after unmapping in truncate_cleanup_page()? Perhaps, but though
> that's likely to reduce or eliminate the number of incidents, it would
> give less assurance of whether we had identified the problem correctly.
>
> This successful iteration introduces "unmap_mapping_page(page)" instead
> of try_to_unmap(), and goes the usual unmap_mapping_range_tree() route,
> with an addition to details. Then zap_pmd_range() watches for this case,
> and does spin_unlock(pmd_lock) if so - just like page_vma_mapped_walk()
> now does in the PVMW_SYNC case. Not pretty, but safe.
>
> Note that unmap_mapping_page() is doing a VM_BUG_ON(!PageLocked) to
> assert its interface; but currently that's only used to make sure that
> page->mapping is stable, and zap_pmd_range() doesn't care if the page is
> locked or not. Along these lines, in invalidate_inode_pages2_range()
> move the initial unmap_mapping_range() out from under page lock, before
> then calling unmap_mapping_page() under page lock if still mapped.
>
> Fixes: fc127da085c2 ("truncate: handle file thp")
> Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
> Cc: <stable@xxxxxxxxxxxxxxx>
I think adding support for THP in try_to_unmap_one() is the most
future-proof way. We would need it there eventually.
But this works too:
Acked-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
--
Kirill A. Shutemov
