mbox series

[v4,00/10] fuse direct write consolidation and parallel IO

Message ID 20230918150313.3845114-1-bschubert@ddn.com (mailing list archive)
Headers show
Series fuse direct write consolidation and parallel IO | expand

Message

Bernd Schubert Sept. 18, 2023, 3:03 p.m. UTC 
This series consolidates DIO writes into a single code path via
fuse_cache_write_iter/generic_file_direct_write. Before
it was only used for O_DIRECT and when writeback cache was not enabled.
For server/daemon dio enforcement (FOPEN_DIRECT_IO) another code
path was used before, but I _think_ that is not needed and
just IOCB_DIRECT needs to be set/enforced.
When writeback-cache was enabled another code path was used, with
a fallback to write-through - for direct IO that should not be
needed either.

So far O_DIRECT through fuse_cache_write_iter also took an exclusive
lock, this should not be needed either, at least when server side
sets FOPEN_PARALLEL_DIRECT_WRITES.

Regarding xfstests, without FOPEN_DIRECT_IO there are no differences
between patched and unpatched. With FOPEN_DIRECT_IO the series
fixes generic/647, which was failing unpatched ("pwrite is broken").

v4:
checkpatch cleanup

Addresses review comments
  - Fix shared lock in fuse_cache_write_iter, file_remove_privs()
    needs the exclusive lock.  In order to know if the exclusive
    lock is needed the new exported vfs function file_remove_privs()
    is introduced.
  - Also fixed for the shared lock is the fallback to buffered writes,
    which also needs the exclusive lock.
  - Fixed is the accidental turn around of false/true in
    fuse_dio_wr_exclusive_lock().
  - Added documentation in fuse_write_flags() that the IOCB_DIRECT flag
    cannot be trusted.

New optional patches
- "fuse: Use the existing inode fuse_cache_write_iter"
- "[RFC] fuse: No privilege removal when FOPEN_DIRECT_IO is set"

v3:
Addresses review comments
  - Rename fuse_direct_write_extending_i_size to io_past_eof
  - Change to single line conditions in fuse_dio_wr_exclusive_lock
    (also fixes accidental parenthesis).
  - Add another patch to rename fuse_direct_io to fuse_send_dio
  - Add detailed information into the commit message of the patch
    that consolidates IO paths (5/6, previously 4/5) and also
    update the subject.

v2:
The entire v1 approach to route DIO writes through fuse_direct_write_iter
was turned around and fuse_direct_write_iter is removed instead and all
DIO writes are now routed through fuse_cache_write_iter

Cc: Hao Xu <howeyxu@tencent.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Dharmendra Singh <dsingh@ddn.com>
Cc: linux-fsdevel@vger.kernel.org

Bernd Schubert (10):
  fuse: direct IO can use the write-through code path
  fuse: Create helper function if DIO write needs exclusive lock
  fuse: prepare support for shared lock for DIO writes
  fs: Add and export file_needs_remove_privs
  fuse: enable shared lock for O_DIRECT / handle privilege drop
  fuse: Rename fuse_direct_io
  fuse: Remove fuse_direct_write_iter code path / use IOCB_DIRECT
  fuse: Remove page flush/invaliation in fuse_direct_io
  fuse: Use the existing inode fuse_cache_write_iter
  [RFC] fuse: No privilege removal when FOPEN_DIRECT_IO is set

 fs/fuse/cuse.c     |   5 +-
 fs/fuse/dax.c      |   2 +-
 fs/fuse/file.c     | 189 +++++++++++++++++++++++----------------------
 fs/fuse/fuse_i.h   |   8 +-
 fs/inode.c         |   8 ++
 include/linux/fs.h |   1 +
 6 files changed, 113 insertions(+), 100 deletions(-)