Re: [PATCH v3 1/4] ntfs: add named stream ioctls support

[Namjae Jeon]

On Mon, Oct 12, 2026 at 8:47 AM CharSyam <charsyam@xxxxxxxxx> wrote:
>
> Hi, Namjae.
>
>   I have two comments on the named stream write path. I tested both cases
>   with the NTFS module in a QEMU guest.
>
>   1. In __ntfs_inode_non_resident_attr_pwrite(), count == PAGE_SIZE does
>      not imply that pos is page aligned. Both full-page branches need to
>      check the alignment of pos:
>
>          while (count) {
>                  bool full_page = count == PAGE_SIZE &&
>                                   IS_ALIGNED(pos, PAGE_SIZE);
>
>                  if (full_page) {
>                          /* __filemap_get_folio() path */
>                  } else {
>                          /* read_mapping_folio() path */
>                  }
>
>                  if (full_page) {
>                          offset = 0;
>                          attr_len = count;
>                  } else {
>                          offset = offset_in_folio(folio, pos);
>                          attr_len = min_t(size_t, (size_t)count,
>                                           folio_size(folio) - offset);
>                  }
>                  ...
>          }
>
>      To reproduce this, I filled an 8 KiB stream with 'A' and wrote
>      4096 bytes of 'B' at offset 1. The ioctl reported success, but byte 0
>      became 'B' and byte 4096 remained 'A'. With the condition above,
>      byte 0 remained 'A' and byte 4096 became 'B', as expected.
This has already been fixed in ntfs-next by commit 7de32a72b75b7
(“ntfs: require page-aligned full-page attribute writes”).

>
>   2. An error on a later page returns the error even if earlier pages
>      were written. The ioctl then returns the error without copying
>      bytes_returned to userspace, and ntfs_inode_attr_pwrite() skips the
>      base file's mtime/ctime update.
>
>      I reproduced this in QEMU by injecting a test-only -EIO after the
>      first page of an 8 KiB overwrite. The ioctl returned EIO with no
>      transferred byte count, while the first 4 KiB changed from 'A' to
>      'B' and the second 4 KiB remained 'A'. The base file's mtime did not
>      change. The changed data remained after sync and remount. With a
>      test fix that updates the base timestamps when an earlier page was
>      written, mtime changed in the same test; the ioctl still returned EIO.
>
>      What partial-write behavior is intended for NTFS_IOC_STREAM_WRITE?
>      Returning the completed byte count as a short successful transfer
>      would follow write(2) conventions, but the stream is currently
>      extended to pos + count before copying, so EOF handling would also
>      need attention. If the ioctl should return the error, its UAPI should
>      explain that partial changes may remain and define whether their
>      length is reported. Either way, timestamps should reflect data that
>      changed.
Okay, I will fix it.
Thanks!