2008-05-06 21:59:42

by [email protected]

[permalink] [raw]
Subject: Re: [PATCH 1/3] svcrdma: Remove redunant call to xpo_release_rqst method.

On Fri, May 02, 2008 at 11:18:04AM -0500, Tom Tucker wrote:
> The xpo_release_rqst method is called from svc_xprt_release and therefore
> this direct call of the provider method is redundant. There is no bug
> here since the provider protects against multiple calls.
>
> Originally, this code was here because the xpo_release_rqst function was
> being used by the RDMA transport to return credits to the client, i.e.
> posting a receive buffer. This had to be done before the send in order to
> avoid a race wherein the client responds immediately with a new request
> but the buffer has not yet been posted. This is a poor design point and
> the recv buffer posting has been modified in the rdma transport.

The comment there refers to the socket code, not the rdma code, and
there was a svc_release_skb() there before the rdma code came along.
I'd like to understand why.

--b.

>
> Signed-off-by: Tom Tucker <[email protected]>
>
> ---
> net/sunrpc/svc_xprt.c | 3 ---
> 1 files changed, 0 insertions(+), 3 deletions(-)
>
> diff --git a/net/sunrpc/svc_xprt.c b/net/sunrpc/svc_xprt.c
> index d8e8d79..74e52d4 100644
> --- a/net/sunrpc/svc_xprt.c
> +++ b/net/sunrpc/svc_xprt.c
> @@ -751,9 +751,6 @@ int svc_send(struct svc_rqst *rqstp)
> if (!xprt)
> return -EFAULT;
>
> - /* release the receive skb before sending the reply */
> - rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
> -
> /* calculate over-all length */
> xb = &rqstp->rq_res;
> xb->len = xb->head[0].iov_len +


2008-05-07 15:05:37

by Tom Tucker

[permalink] [raw]
Subject: Re: [PATCH 1/3] svcrdma: Remove redunant call to xpo_release_rqst method.


On Tue, 2008-05-06 at 17:59 -0400, J. Bruce Fields wrote:
> On Fri, May 02, 2008 at 11:18:04AM -0500, Tom Tucker wrote:
> > The xpo_release_rqst method is called from svc_xprt_release and therefore
> > this direct call of the provider method is redundant. There is no bug
> > here since the provider protects against multiple calls.
> >
> > Originally, this code was here because the xpo_release_rqst function was
> > being used by the RDMA transport to return credits to the client, i.e.
> > posting a receive buffer. This had to be done before the send in order to
> > avoid a race wherein the client responds immediately with a new request
> > but the buffer has not yet been posted. This is a poor design point and
> > the recv buffer posting has been modified in the rdma transport.
>
> The comment there refers to the socket code, not the rdma code, and

Yes, you are correct.

> there was a svc_release_skb() there before the rdma code came along.
> I'd like to understand why.

I believe it's because the socket code has buffer quotas and the skb
you're holding consumes credits in the recv sock buf. So it's
conceivable that you could send a message, another client message
arrives prior to releasing the skb your holding, and this new message
gets dropped unnecessarily. I don't believe it affects the sending at
all because the message you're holding came from sock_recv_datagram and
only consumes recv side credits.

Since this patch is only for "clean up" purposes, I'm inclined to just
ditch this patch and let us noodle on it a little more. Agreed?

>
> --b.
>
> >
> > Signed-off-by: Tom Tucker <[email protected]>
> >
> > ---
> > net/sunrpc/svc_xprt.c | 3 ---
> > 1 files changed, 0 insertions(+), 3 deletions(-)
> >
> > diff --git a/net/sunrpc/svc_xprt.c b/net/sunrpc/svc_xprt.c
> > index d8e8d79..74e52d4 100644
> > --- a/net/sunrpc/svc_xprt.c
> > +++ b/net/sunrpc/svc_xprt.c
> > @@ -751,9 +751,6 @@ int svc_send(struct svc_rqst *rqstp)
> > if (!xprt)
> > return -EFAULT;
> >
> > - /* release the receive skb before sending the reply */
> > - rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
> > -
> > /* calculate over-all length */
> > xb = &rqstp->rq_res;
> > xb->len = xb->head[0].iov_len +
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html