Re: re-enable IOCB_NOWAIT writes to files

[Jeff Layton <jlayton@kernel.org>]

On Fri, 2025-11-14 at 09:01 -0800, Darrick J. Wong wrote:
> On Fri, Nov 14, 2025 at 09:04:58AM -0500, Jeff Layton wrote:
> > On Fri, 2025-11-14 at 07:26 +0100, Christoph Hellwig wrote:
> > > Hi all,
> > > 
> > > commit 66fa3cedf16a ("fs: Add async write file modification handling.")
> > > effectively disabled IOCB_NOWAIT writes as timestamp updates currently
> > > always require blocking, and the modern timestamp resolution means we
> > > always update timestamps.  This leads to a lot of context switches from
> > > applications using io_uring to submit file writes, making it often worse
> > > than using the legacy aio code that is not using IOCB_NOWAIT.
> > > 
> > > This series allows non-blocking updates for lazytime if the file system
> > > supports it, and adds that support for XFS.
> > > 
> > > It also fixes the layering bypass in btrfs when updating timestamps on
> > > device files for devices removed from btrfs usage, and FMODE_NOCMTIME
> > > handling in the VFS now that nfsd started using it.  Note that I'm still
> > > not sure that nfsd usage is fully correct for all file systems, as only
> > > XFS explicitly supports FMODE_NOCMTIME, but at least the generic code
> > > does the right thing now.
> > > 
> > > Diffstat:
> > >  Documentation/filesystems/locking.rst |    2 
> > >  Documentation/filesystems/vfs.rst     |    6 ++
> > >  fs/btrfs/inode.c                      |    3 +
> > >  fs/btrfs/volumes.c                    |   11 +--
> > >  fs/fat/misc.c                         |    3 +
> > >  fs/fs-writeback.c                     |   53 ++++++++++++++----
> > >  fs/gfs2/inode.c                       |    6 +-
> > >  fs/inode.c                            |  100 +++++++++++-----------------------
> > >  fs/internal.h                         |    3 -
> > >  fs/orangefs/inode.c                   |    7 ++
> > >  fs/overlayfs/inode.c                  |    3 +
> > >  fs/sync.c                             |    4 -
> > >  fs/ubifs/file.c                       |    9 +--
> > >  fs/utimes.c                           |    1 
> > >  fs/xfs/xfs_iops.c                     |   29 ++++++++-
> > >  fs/xfs/xfs_super.c                    |   29 ---------
> > >  include/linux/fs.h                    |   17 +++--
> > >  include/trace/events/writeback.h      |    6 --
> > >  18 files changed, 152 insertions(+), 140 deletions(-)
> > 
> > This all looks pretty reasonable to me. There are a few changelog and
> > subject line typos, but the code changes look fine. You can add:
> > 
> > Reviewed-by: Jeff Layton <jlayton@kernel.org>
> > 
> > As far as nfsd's usage of FMODE_NOCMTIME, it looks OK to me. That's
> > implemented today by the check in file_modified_flags(), which is
> > generic and should work across filesystems.
> > 
> > The main exception is xfs_exchange_range() which has some special
> > handling for it, but nfsd doesn't use that functionality so that
> > shouldn't be an issue.
> > 
> > Am I missing some subtlety?
> 
> In exchangerange specifically?
> 
> The FMODE_NOCMTIME checks in xfs_exchange_range exist to tell the
> exchange-range code to update cmtime, but only if it decides to actually
> go through with the mapping exchange.  Since the mapping exchange
> requires a transaction anyway, it's cheap to bundle in timestamp
> updates.
> 
> Also there's no way that we can do nonblocking exchangerange so a NOWAIT
> flag wouldn't be much help here anyway.
> 
> (I hope that answers your question)
> 
> 

Christoph mentioned nfsd might be doing something wrong, which is my
main interest here. nfsd doesn't have a way to expose exchangerange
functionality right now, but if it did then it seems like that would
just work too.

HCH says:

> Nothing requires file_update_time / file_modified_flags are helpers
> that a file system may or may not call.  I've not done an audit
> if everyone actually uses them.

I'll have to think about how to efficiently audit that. The good news
is that nfsd really only cares about the write() and page_mkwrite()
codepaths. For other activity, the delegation will be broken and
recalled.
-- 
Jeff Layton <jlayton@kernel.org>