On Thu, 2025-07-31 at 17:28 -0400, Mike Snitzer wrote:
> On Thu, Jul 31, 2025 at 04:58:00PM -0400, Jeff Layton wrote:
> > On Thu, 2025-07-31 at 15:44 -0400, Mike Snitzer wrote:
> > > NFS doesn't have any DIO alignment constraints but it doesn't support
> > > STATX_DIOALIGN, so update NFSD such that it doesn't disable the use of
> > > NFSD_IO_DIRECT if it is reexporting NFS.
> > >
> > > Signed-off-by: Mike Snitzer <snitzer@xxxxxxxxxx>
> > > ---
> > > fs/nfs/export.c | 3 ++-
> > > fs/nfsd/filecache.c | 11 +++++++++++
> > > include/linux/exportfs.h | 13 +++++++++++++
> > > 3 files changed, 26 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/fs/nfs/export.c b/fs/nfs/export.c
> > > index e9c233b6fd209..2cae75ba6b35d 100644
> > > --- a/fs/nfs/export.c
> > > +++ b/fs/nfs/export.c
> > > @@ -155,5 +155,6 @@ const struct export_operations nfs_export_ops = {
> > > EXPORT_OP_REMOTE_FS |
> > > EXPORT_OP_NOATOMIC_ATTR |
> > > EXPORT_OP_FLUSH_ON_CLOSE |
> > > - EXPORT_OP_NOLOCKS,
> > > + EXPORT_OP_NOLOCKS |
> > > + EXPORT_OP_NO_DIOALIGN_NEEDED,
> > > };
> > > diff --git a/fs/nfsd/filecache.c b/fs/nfsd/filecache.c
> > > index 5601e839a72da..ea489dd44fd9a 100644
> > > --- a/fs/nfsd/filecache.c
> > > +++ b/fs/nfsd/filecache.c
> > > @@ -1066,6 +1066,17 @@ nfsd_file_getattr(const struct svc_fh *fhp, struct nfsd_file *nf)
> > > nfsd_io_cache_write != NFSD_IO_DIRECT))
> > > return nfs_ok;
> > >
> > > + if (exportfs_handles_unaligned_dio(nf->nf_file->f_path.mnt->mnt_sb->s_export_op)) {
> > > + /* Underlying filesystem doesn't support STATX_DIOALIGN
> > > + * but it can handle all unaligned DIO, so establish
> > > + * DIO alignment that is accommodating.
> > > + */
> > > + nf->nf_dio_mem_align = 4;
> > > + nf->nf_dio_offset_align = PAGE_SIZE;
> > > + nf->nf_dio_read_offset_align = nf->nf_dio_offset_align;
> > > + return nfs_ok;
> > > + }
> > > +
> > > status = fh_getattr(fhp, &stat);
> > > if (status != nfs_ok)
> > > return status;
> > > diff --git a/include/linux/exportfs.h b/include/linux/exportfs.h
> > > index 9369a607224c1..626b8486dd985 100644
> > > --- a/include/linux/exportfs.h
> > > +++ b/include/linux/exportfs.h
> > > @@ -247,6 +247,7 @@ struct export_operations {
> > > */
> > > #define EXPORT_OP_FLUSH_ON_CLOSE (0x20) /* fs flushes file data on close */
> > > #define EXPORT_OP_NOLOCKS (0x40) /* no file locking support */
> > > +#define EXPORT_OP_NO_DIOALIGN_NEEDED (0x80) /* fs can handle unaligned DIO */
> > > unsigned long flags;
> > > };
> > >
> > > @@ -262,6 +263,18 @@ exportfs_cannot_lock(const struct export_operations *export_ops)
> > > return export_ops->flags & EXPORT_OP_NOLOCKS;
> > > }
> > >
> > > +/**
> > > + * exportfs_handles_unaligned_dio() - check if export can handle unaligned DIO
> > > + * @export_ops: the nfs export operations to check
> > > + *
> > > + * Returns true if the export can handle unaligned DIO.
> > > + */
> > > +static inline bool
> > > +exportfs_handles_unaligned_dio(const struct export_operations *export_ops)
> > > +{
> > > + return export_ops->flags & EXPORT_OP_NO_DIOALIGN_NEEDED;
> > > +}
> > > +
> > > extern int exportfs_encode_inode_fh(struct inode *inode, struct fid *fid,
> > > int *max_len, struct inode *parent,
> > > int flags);
> >
> >
> > Would it not be simpler (better?) to add support for STATX_DIOALIGN to
> > NFS, and just have it report '1' for both values?
>
> I suppose adding NFS support for STATX_DIOALIGN, that doesn't actually
> go over the wire, does make sense.
>
The NFS protocol doesn't have any alignment restrictions. The NFS
client supports DIO, but doesn't enforce any sort of alignment
restriction on userland.
> But I wouldn't think setting them to 1 valid. Pretty sure they need
> to be a power-of-2 (since they are used as masks passed to
> iov_iter_is_aligned).
>
2^0 == 1 :-)
This might be a good thing to bring up to the greater fsdevel
community. What should filesystems that support DIO but don't enforce
any alignment restrictions report for that attribute?
'1' would seem to be the natural thing to return in that case. Maybe we
need to special case that in some of the helpers?
> In addition, we want to make sure NFS's default DIO alignment (which
> isn't informed by actual DIO alignment advertised by NFSD's underlying
> filesystem and hardware, e.g. XFS and NVMe) is able to be compatible
> with both finer (512b) and coarser (4096b) grained DIO alignment.
> Only way to achieve that would be to skew toward the coarse-grained
> end of the spectrum, right?
>
> More conservative but more likely to work with everything.
>
I don't think NFS has ever enforced a particular alignment on userland,
at least not with regular network transport. Does RDMA change this?
In any case, I'm fine with taking this for now as a stopgap fix, but we
should aim to plumb proper support for STATX_DIOALIGN in the client
sometime soon. Applications are going to start using that attribute,
and if they get back that it's unsupported, some may fail or fall back
on buffered I/O on NFS.
--
Jeff Layton <jlayton@xxxxxxxxxx>
