Re: [PATCH v3 0/5] ntfs: support O_TMPFILE

[liubaolin <liubaolin12138@163.com>]

在 2026/11/6 19:37, Namjae Jeon 写道:
> On Fri, Nov 6, 2026 at 3:31 PM Baolin Liu <liubaolin12138@163.com> wrote:
>>
>> From: Baolin Liu <liubaolin@kylinos.cn>
> Hi Baolin,
>>
>> This series adds O_TMPFILE support to fs/ntfs. Applications can prepare a
>> regular file's contents and metadata before publishing it with linkat().
>> Unpublished files are reclaimed when their last reference is released.
>> O_EXCL prevents publication.
>>
>> The series keeps the same five changes and ordering as v1 and v2:
>>
>>    1. Initialize inherited ACLs after allocating the new MFT record.
>>       Anonymous creation needs its inherited ACL stored before returning.
>>    2. Preserve prepared atime and mtime when publishing the inode through
>>       the hard-link path.
>>    3. Initialize the new FILE_NAME's packed EA size, since a temporary
>>       file may already have EAs when it is published.
>>    4. Complete unlinked-file attribute writes before dropping the cache,
>>       including EAs and data copied during resident-to-nonresident
>>       conversion before publication.
>>    5. Add O_TMPFILE creation and publication support.
>>
>> As requested in review, patch 2 gives the new FILE_NAME and directory
>> index entry the same ctime as the inode. The inode ctime is updated only
>> after the link succeeds. Patch 1 also initializes a newly created
>> FILE_NAME's ctime after inherited ACL setup, fixing a timestamp mismatch
>> caused by moving ACL initialization later in v1.
>>
>> In v3, patch 4 also preserves data copied by
>> ntfs_attr_make_non_resident(). As reported in review, growing DATA in a
>> nearly full MFT record can convert SECURITY_DESCRIPTOR and leave its
>> contents only in a dirty folio. Evicting the zero-link attribute inode
>> then discards those contents.
>>
>> Write the resident value to the newly allocated clusters before
>> committing the conversion. If that write fails, release the new clusters
>> and retain the original resident value. Propagate the error to the caller
>> instead of continuing with another attribute. Patches 1-3 and 5 are
>> unchanged in code from v2.
>>
>> While implementing and testing O_TMPFILE, we also found pre-existing
>> NTFS bugs in attribute rollback and ACL lookup. These include stale
>> attribute inodes left cached after failed creation, which can cause
>> subsequent creation to return EIO even after space is freed, incorrect
>> ACL inheritance when the default ACL is not cached, and a fixed-size
>> read buffer that cannot handle larger ACL values. The underlying bugs
>> predate this series.
>>
>> To keep this series focused on O_TMPFILE support and review
>> feedback, we are leaving these existing bugs to a separate bug-fix
>> series. We plan to submit those fixes after this series is accepted and
>> merged. These issues remain present with this series applied.
>>
>> Tested this five-patch series in QEMU with four CPUs, NTFS debug and
>> lockdep enabled. The requested unchanged xfstests generic/004 and
>> generic/389 passed. Tests covered creation, publication, last-reference
>> cleanup, O_EXCL, normal idmapped ownership, named streams, nonresident
>> EAs, direct I/O, mmap, allocation and concurrent publication. Published
>> data and metadata were checked after cache eviction, remount and reboot.
>> Offline inspection compared hard-link ctime in STANDARD_INFORMATION,
>> the new FILE_NAME and the directory index.
>>
>> The conversion regression uses 1 KiB MFT records and varies EA sizes
>> before writing 4 KiB to anonymous and ordinary files. Offline inspection
>> compares SECURITY_DESCRIPTOR contents with the original resident value.
>> The v2 kernel loses converted security descriptors and some EAs; v3
>> preserves them with 512-byte, 4 KiB and 8 KiB clusters, including after a
>> fresh boot. An injected block write error returns EIO and leaves the
>> resident security descriptor and EA intact. Free space is unchanged after
>> that failure, and retrying the write succeeds.
>>
>> Changes in v3:
>> - Extend patch 4 to write converted attribute data before switching the
>>    record to nonresident form, addressing the v2 review.
>> - Preserve the resident value on conversion write failure and propagate
>>    conversion errors.
>> - Add conversion and I/O-error regression tests and rerun generic/004
>>    and generic/389.
>> - No code changes to patches 1-3 and 5; add Hyunchul Lee's Reviewed-by.
>>
>> Changes in v2:
>> - Update FILE_NAME and index ctime as requested in review.
>> - Set the source inode ctime only after successful hard-link creation.
>> - Initialize creation FILE_NAME ctime after inherited ACL setup.
>> - Run generic/004 and generic/389 on the standalone five-patch series.
>> - Rebase onto ntfs-next commit
>>    564c7753647731d19170914c1c637d3b33d88d3b, including the const mnt_idmap
>>    tmpfile callback signature and documentation context.
>> - Resolve patch 1 against the updated early ACL error handling in
>>    __ntfs_create(); the relocated ACL failure still uses MFT/EA rollback.
>> - Keep the original five-patch scope; submit fixes for the pre-existing
>>    attribute rollback and ACL lookup bugs separately after this series
>>    is merged.
>> - No functional changes to patches 3 and 4.
>>
>> Baolin Liu (2):
>>    ntfs: initialize inherited ACLs after allocating the MFT record
>>    ntfs: support O_TMPFILE
>>
>> Xuewen Wang (3):
>>    ntfs: preserve timestamps when creating hard links
>>    ntfs: initialize packed EA size when creating hard links
>>    ntfs: write unlinked file attributes before dropping their cache
> 
> Can you check Sashiko's review comments? I think some of the points
> raised in Sashiko’s review are worth looking into.
> 
> https://sashiko.dev/#/patchset/20261106063059.2904988-1-liubaolin12138%40163.com
> 
> Thanks.

Hi Namjae,
    Thanks for pointing this out. I'll address the review comments and 
send v4 as soon as possible.

Thanks,
Baolin.