Re: [PATCH 2/2] ntfs: take invalidate_lock in ntfs_filemap_page_mkwrite()

[Hongling Zeng]

在 2026年08月31日 15:24, Namjae Jeon 写道:
> >          if (NInoWofCompressed(NTFS_I(inode)))
> > @@ -684,7 +685,14 @@ static vm_fault_t ntfs_filemap_page_mkwrite(struct vm_fault *vmf)
> >          sb_start_pagefault(inode->i_sb);
> >          file_update_time(vmf->vma->vm_file);
> >
> > +       /*
> > +        * Serialize against truncate/fallocate which hold the lock
> > +        * exclusively while invalidating pagecache and changing extents.
> > +        */
> > +       filemap_invalidate_lock_shared(mapping);
> >          ret = iomap_page_mkwrite(vmf, &ntfs_page_mkwrite_iomap_ops, NULL);
> > +       filemap_invalidate_unlock_shared(mapping);
> > +
> Please move filemap_invalidate_unlock() after truncate_pagecache() and
> pagecache_isize_extended(). Otherwise, ntfs_filemap_page_mkwrite() can
> race with the page-cache cleanup after fallocate extends the file.
>
> diff --git a/fs/ntfs/file.c b/fs/ntfs/file.c
> index 1969e4f444f7..a266105180a9 100644
> --- a/fs/ntfs/file.c
> +++ b/fs/ntfs/file.c
> @@ -1178,13 +1178,15 @@ static long ntfs_fallocate(struct file *file,
> int mode, loff_t offset, loff_t le
>
>          err = file_modified(file);
>   out:
> +       if (!err && mode == 0 && NInoNonResident(ni) &&
> +           offset > old_size) {
> +               truncate_pagecache(vi, old_size);
> +               pagecache_isize_extended(vi, old_size, offset);
> +       }
> +
>          filemap_invalidate_unlock(vi->i_mapping);
> +
>          if (!err) {
> -               if (mode == 0 && NInoNonResident(ni) &&
> -                   offset > old_size) {
> -                       truncate_pagecache(vi, old_size);
> -                       pagecache_isize_extended(vi, old_size, offset);
> -               }
>                  NInoSetFileNameDirty(ni);
>                  inode_set_mtime_to_ts(vi, inode_set_ctime_current(vi));
>                  mark_inode_dirty(vi);
  Thanks for the review!

  You're right. ntfs_fallocate() dropped invalidate_lock before calling
  truncate_pagecache() and pagecache_isize_extended(), so a write fault
  holding the lock shared could interleave with the page-cache cleanup
  and have its folio discarded by truncate_inode_pages(), losing the
  mmap write. Moving both calls inside the lock hold range fixes it.

  I've posted v2 and addresses your comment.