From: Greg Banks <gnb@sgi.com>
Subject: Re: [RFC,
	PATCH 1/4] svc: Fix skip computation in svc_defer and svc_revisit
Date: Wed, 24 Oct 2007 12:45:09 +1000
Message-ID: <20071024024509.GB17462@sgi.com>
References: <20071019214058.31422.35615.stgit@dell3.ogc.int>
	<20071019214523.31422.28978.stgit@dell3.ogc.int>
	<20071023024530.GB6354@sgi.com>
	<20071023180354.GE30496@fieldses.org>
Mime-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Cc: neilb@suse.de, nfs@lists.sourceforge.net
To: "J. Bruce Fields" <bfields@fieldses.org>
In-Reply-To: <20071023180354.GE30496@fieldses.org>
Sender: nfs-bounces@lists.sourceforge.net
Errors-To: nfs-bounces@lists.sourceforge.net

On Tue, Oct 23, 2007 at 02:03:55PM -0400, J. Bruce Fields wrote:
> On Tue, Oct 23, 2007 at 12:45:30PM +1000, Greg Banks wrote:
> > On Fri, Oct 19, 2007 at 04:45:23PM -0500, Tom Tucker wrote:
> > > 
> > > The rq_arg.len includes the size of the transport header. The computations
> > > assumed that it did not.
> > 
> > ok, assuming that your latest svc_rdma_xdr_decode_req(), which I
> > haven't read yet, doesn't change rq_arg.len as it parses the RDMA
> > chunking header.
> > 
> > The reason I mention that is that one set of patches I was working
> > with some time ago had different semantics for rq_arg.len; when
> > parsing the chunking header, rq_arg.len was adjusted downward as
> > rq_head[0].iov_base was adjusted forward.  This kept rq_arg.len equal
> > to the sum of rq_{head[0],pages[...],tail[0]}.iov_len, which seemed
> > like a good idea.
> > 
> > But your way seems valid too.
> 
> The integrity and privacy code may be a particularly unpleasant source
> of assumptions about the various length fields; see e.g.
> unwrap_integ_data() and unwrap_priv_data().  (But maybe nobody cares
> about krb5i/krb5p over rdma?)

More or less the entire point of NFS/RDMA is that the CPU never
has to see the data that goes on the wire; no memory copies (*),
no message digest calculations, no encryptions.  This makes krb5[ip]
over RDMA rather an unhelpful combination.

But if there's some code there that makes actual assumptions about
the xdr_buf.len field, we need to find it, document it and ensure
we're maintaining that assumption in the new code.

(*) Last time I looked the write path still had an unnecessary memcpy.

Greg.
-- 
Greg Banks, R&D Software Engineer, SGI Australian Software Group.
Apparently, I'm Bedevere.  Which MPHG character are you?
I don't speak for SGI.

-------------------------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc.
Still grepping through log files to find problems?  Stop.
Now Search log events and configuration files using AJAX and a browser.
Download your FREE copy of Splunk now >> http://get.splunk.com/
_______________________________________________
NFS maillist  -  NFS@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/nfs