sharing protocol defintions between client and server?

sharing protocol defintions between client and server?

[Christoph Hellwig]

>From a lot of the recent work it seems like there's basically no
sharing of protocol definitions between the Linux NFS client and
server, which seems fairly annoying to me.  Is there a good historic
reason for this and did we ever attempt to change it?  Maybe even
use some kernel-specific rpcgen variant to generate them directly
from the spec?

Re: sharing protocol defintions between client and server?

[J. Bruce Fields]

On Wed, Nov 13, 2013 at 08:21:16AM -0800, Christoph Hellwig wrote:
> >From a lot of the recent work it seems like there's basically no
> sharing of protocol definitions between the Linux NFS client and
> server, which seems fairly annoying to me.

What are you thinking of exactly?

I suspect that there could be more sharing.

> Is there a good historic
> reason for this and did we ever attempt to change it?  Maybe even
> use some kernel-specific rpcgen variant to generate them directly
> from the spec?

I don't know if anyone's looked into it.

There's a lot of hand-written xdr code (the encoders/decoders
themselves, the maxsz constants in fs/nfs/nfs*xdr.c, the _rsize junk in
fs/nfsd/nfs4proc.c, ...).  It seems error-prone to me though I can't
recall if we've had a lot of bugs there in practice.

--b.

Re: sharing protocol defintions between client and server?

[Christoph Hellwig]

On Thu, Nov 14, 2013 at 10:05:46AM -0500, J. Bruce Fields wrote:
> On Wed, Nov 13, 2013 at 08:21:16AM -0800, Christoph Hellwig wrote:
> > >From a lot of the recent work it seems like there's basically no
> > sharing of protocol definitions between the Linux NFS client and
> > server, which seems fairly annoying to me.
> 
> What are you thinking of exactly?
> 
> I suspect that there could be more sharing.

To make it easy I'll just quote from the SEEK patches.

client:

+struct nfs42_seek_args {
+	struct nfs4_sequence_args	seq_args;
+
+	struct nfs_fh			*sa_fh;
+	nfs4_stateid			*sa_stateid;
+	u64				sa_offset;
+	u32				sa_what;
+};
+
+struct nfs42_seek_res {
+	struct nfs4_sequence_res	seq_res;
+	unsigned int			status;
+
+	u32	sr_eof;
+	u32	sr_whence;
+	u64	sr_offset;
+	u64	sr_length;
+	u32	sr_allocated;
+};
+#endif

server:

+struct nfsd4_seek {
+	/* request */
+	stateid_t	seek_stateid;
+	loff_t		seek_offset;
+	u32		seek_whence;
+
+	/* response */
+	u64		seek_pos;
+	u32		seek_eof;
+	u64		seek_length;
+	u32		seek_allocated;
+};

note that a lot of server operations also seem to have separate
args and result substrutures.  In general I'd love to have one
structure for the actual on-the wire operation in a header, and then
client and server could build in-memory versions around them.

Of course just generating those from the XDR would be even better.

Maybe I'll play around with doing a krpcgen that we can initially
just use for producing the structures, for which it should be
pretty clear benefit.  If we're lucky we might be to also move
some marshalling/unmarshalling over to it later.

Re: sharing protocol defintions between client and server?

[J. Bruce Fields]

On Thu, Nov 14, 2013 at 08:40:46AM -0800, Christoph Hellwig wrote:
> On Thu, Nov 14, 2013 at 10:05:46AM -0500, J. Bruce Fields wrote:
> > On Wed, Nov 13, 2013 at 08:21:16AM -0800, Christoph Hellwig wrote:
> > > >From a lot of the recent work it seems like there's basically no
> > > sharing of protocol definitions between the Linux NFS client and
> > > server, which seems fairly annoying to me.
> > 
> > What are you thinking of exactly?
> > 
> > I suspect that there could be more sharing.
> 
> To make it easy I'll just quote from the SEEK patches.
> 
> client:
> 
> +struct nfs42_seek_args {
> +	struct nfs4_sequence_args	seq_args;
> +
> +	struct nfs_fh			*sa_fh;
> +	nfs4_stateid			*sa_stateid;
> +	u64				sa_offset;
> +	u32				sa_what;
> +};
> +
> +struct nfs42_seek_res {
> +	struct nfs4_sequence_res	seq_res;
> +	unsigned int			status;
> +
> +	u32	sr_eof;
> +	u32	sr_whence;
> +	u64	sr_offset;
> +	u64	sr_length;
> +	u32	sr_allocated;
> +};
> +#endif
> 
> server:
> 
> +struct nfsd4_seek {
> +	/* request */
> +	stateid_t	seek_stateid;
> +	loff_t		seek_offset;
> +	u32		seek_whence;
> +
> +	/* response */
> +	u64		seek_pos;
> +	u32		seek_eof;
> +	u64		seek_length;
> +	u32		seek_allocated;
> +};
> 
> note that a lot of server operations also seem to have separate
> args and result substrutures.  In general I'd love to have one
> structure for the actual on-the wire operation in a header, and then
> client and server could build in-memory versions around them.
> 
> Of course just generating those from the XDR would be even better.

Sometimes a lot of those fields are unused.  And we can currently do
things like read/write directly from/to the data structures these are
eventually going to end up in.

So in theory the existing code might be more efficient.

In practice I don't know if it is.  And it'd be nice to be able to
autogenerate huge swaths of boring code.

--b.

> Maybe I'll play around with doing a krpcgen that we can initially
> just use for producing the structures, for which it should be
> pretty clear benefit.  If we're lucky we might be to also move
> some marshalling/unmarshalling over to it later.
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html