Re: [PATCH v5 0/4] NFSD: Implement NFSD_IO_DIRECT for NFS WRITE

[Mike Snitzer <snitzer@kernel.org>]

On Tue, Oct 21, 2025 at 09:35:32AM -0400, Chuck Lever wrote:
> On 10/21/25 7:12 AM, Jeff Layton wrote:
> > On Mon, 2025-10-20 at 12:44 -0400, Chuck Lever wrote:
> >> On 10/20/25 12:33 PM, Mike Snitzer wrote:
> >>> On Mon, Oct 20, 2025 at 12:25:42PM -0400, Chuck Lever wrote:
> >>> Just a bit concerned about removing IOCB_SYNC in that
> >>> we're altering stable_how to be NFS_FILE_SYNC.
> >> Commit 3f3503adb332 ("NFSD: Use vfs_iocb_iter_write()") introduces
> >> the first use of IOCB_ flags in NFSD's write path, and it uses
> >> IOCB_DSYNC. The patch has Reviewed-by's from Christoph, Neil, and
> >> Jeff.
> >>
> >> Should we be concerned that IOCB_DSYNC does not persist time stamp
> >> changes that might be lost during an unplanned server boot?
> >>
> >> As a reminder to the thread, Section 3.3.7 of RFC 1813 says:
> >>
> >>          If stable is FILE_SYNC, the server must commit the data
> >>          written plus all file system metadata to stable storage
> >>          before returning results.
> >>
> >> The text is a bit blurry about whether "file system metadata" means
> >> all of the outstanding metadata changes for every file, or just the
> >> metadata changes for the target file handle.
> >>
> >> NFSD has historically treated DATA_SYNC and FILE_SYNC identically,
> >> as the Linux NFS client does not use DATA_SYNC (IIRC).
> >>
> > 
> > Surely it just meant for the one file.
> 
> Well yes that is the traditional understanding. I'm merely pointing out
> that the actual text is not quite as specific as what we've come to
> understand.
> 
> 
> > FILE_SYNC is only applicable to
> > WRITE/COMMIT operations and those only deal with a single file at a
> > time.
> 
> True but you may recall that NFSD's COMMIT used to ignore the range
> arguments and flush the whole file. Some file systems used to flush
> all dirty data in this case, IIRC.
> 
> There's always been a bit of a mismatch between the spec and what NFSD
> has implemented.
> 
> 
> > If the client gets back FILE_SYNC on a write, it should _not_
> > assume that all outstanding dirty data to all files has been sync'ed.
> 
> Agreed.
> 
> But back to Mike's point.
> 
> - The spec says NFS_DATA_SYNC means persist file data.
> 
> - The spec says NFS_FILE_SYNC means persist file data and file
>   attributes.
> 
> - After consulting with the section describing COMMIT, I think that
>   COMMIT is supposed to persist both file data and attributes.
> 
> And my reading of the code in fs/nfsd/vfs.c is that NFSD does the
> equivalent of NFS_DATA_SYNC in all of these cases, and has done for
> as long as I cared to chase the commit log.
> 
> Moveover, commit 3f3503adb332 did not introduce this behavior.
> 
> Previous to that commit, nfsd_vfs_write() passed RWF_SYNC to
> vfs_iter_write(). This API uses kiocb_set_rw_flags() to convert the RWF
> flag into an IOCB flag. kiocb_set_rw_flags does this:
> 
>         kiocb_flags |= (__force int) (flags & RWF_SUPPORTED);
>         if (flags & RWF_SYNC)
>                 kiocb_flags |= IOCB_DSYNC;
> 
> And that's where I copied IOCB_DSYNC from. The use of RWF_SYNC was
> introduced in 2016 by commit 24368aad47dc ("nfsd: use RWF_SYNC").
> 
> So we've tacitly agreed to let NFSD fall short of the specs in this
> regard for some time. However I don't believe this is documented
> anywhere.
> 
> Based on this reasoning, IOCB_DSYNC is historically correct for the
> DIRECT WRITE path and its fallbacks. I'm guessing that an O_DIRECT WRITE
> is going to persist the written data but won't persist file attribute
> changes either.
> 
> I'm open to making NFSD adhere more strictly to the spec language, but
> I bet there will be a performance impact. Maybe that impact will be
> unnoticeable on modern storage devices.

I was able to get NFSD Direct's use of IOCB_DSYNC|IOCB_SYNC tested
with Jonathan Flynn last week, there was no performance difference on
his modern test cluster (8 "enterprise" NVMe devices in the server).

Was very comforting to learn IOCB_SYNC didn't cause any performance
loss.

Mike