Linux NFS development
 help / color / mirror / Atom feed
From: Peter Leckie <pleckie@sgi.com>
To: Trond Myklebust <trond.myklebust@fys.uio.no>
Cc: talpey@netapp.com, linux-nfs@vger.kernel.org
Subject: Re: [PATCH 01/04]  NFS/RDMA client stall patches
Date: Wed, 11 Jun 2008 18:03:03 +1000	[thread overview]
Message-ID: <484F86B7.7040109@sgi.com> (raw)
In-Reply-To: <1213125899.20459.34.camel@localhost>

Trond Myklebust wrote:
> On Mon, 2008-05-19 at 13:50 +1000, Peter Leckie wrote:
>   
>> Don't call __xprt_get_cong() if this is a retransmit.
>> This prevents __xprt_get_cong() from recursively
>> incrementing the congestion avoidance window for
>> retransmitted data.
>>
>> Signed-off-by: Peter Leckie <pleckie-cP1dWloDopni96+mSzHFpQC/G2K4zDHf@public.gmane.org>
>> Reviewed-by: Greg Banks <gnb-cP1dWloDopni96+mSzHFpQC/G2K4zDHf@public.gmane.org>
>> X-Sgi-Pv: 971446
>>  <http://bugworks/query.cgi/971446>---
>> Index: linux-2.6.25.3/net/sunrpc/xprt.c
>> ===================================================================
>> --- linux-2.6.25.3.orig/net/sunrpc/xprt.c
>> +++ linux-2.6.25.3/net/sunrpc/xprt.c
>> @@ -224,7 +224,8 @@ int xprt_reserve_xprt_cong(struct rpc_ta
>>                         return 1;
>>                 goto out_sleep;
>>         }
>> -       if (__xprt_get_cong(xprt, task)) {
>> +     /*If this is a retransmit don't increment cong*/
>> +       if ((req && req->rq_ntrans) ||__xprt_get_cong(xprt, task)) {
>>                 xprt->snd_task = task;
>>                 if (req) {
>>                         req->rq_bytes_sent = 0;
>>
>>     
>
> Why would we not want to increment the congestion avoidance window on
> retransmitted data? On timeout, xprt_adjust_cwnd will call
> __xprt_put_cong() prior to the retransmission, so I can't see how this
> is a 'recursive increment'.
>
> Trond

That's a good point you raise there I was looking to closely at the tcp 
equivalent, the correct fix for this issue would be to implement a timer 
function for NFS/RDMA pretty much identical to xs_udp_timer(), as follows:


Implement xprt_rdma_timer() to be called when an RPC times out.
This is needed to decrement the cong after an rpc times out preventing
the congestion aviodance from tripping under retransmitts.

Signed-off-by: Peter Leckie <pleckie-cP1dWloDopni96+mSzHFpQC/G2K4zDHf@public.gmane.org>
Reviewed-by: Greg Banks <gnb-cP1dWloDopni96+mSzHFpQC/G2K4zDHf@public.gmane.org>
---

Index: linux-2.6.25.3/net/sunrpc/xprtrdma/transport.c
===================================================================
--- linux-2.6.25.3.orig/net/sunrpc/xprtrdma/transport.c
+++ linux-2.6.25.3/net/sunrpc/xprtrdma/transport.c
@@ -450,6 +450,18 @@ out1:
 }

 /*
+ * xprt_rdma_timer - called when a retransmit timeout occurs on a RDMA 
transport
+ * @task: task that timed out
+ *
+ * Adjust the congestion window after a retransmit timeout has occurred.
+ */
+static void
+xprt_rdma_timer(struct rpc_task *task)
+{
+       xprt_adjust_cwnd(task, -ETIMEDOUT);
+}
+
+/*
  * Close a connection, during shutdown or timeout/reconnect
  */
 static void
@@ -755,7 +767,8 @@ static struct rpc_xprt_ops xprt_rdma_pro
        .send_request           = xprt_rdma_send_request,
        .close                  = xprt_rdma_close,
        .destroy                = xprt_rdma_destroy,
-       .print_stats            = xprt_rdma_print_stats
+       .print_stats            = xprt_rdma_print_stats,
+       .timer                  = xprt_rdma_timer
 };

 static struct xprt_class xprt_rdma = {


Thanks,
Pete


  reply	other threads:[~2008-06-11  8:04 UTC|newest]

Thread overview: 5+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2008-05-19  3:50 [PATCH 01/04] NFS/RDMA client stall patches Peter Leckie
2008-06-10 19:24 ` Trond Myklebust
2008-06-11  8:03   ` Peter Leckie [this message]
2008-06-11 13:53     ` Talpey, Thomas
     [not found]       ` <RTPCLUEXC1-PRDaogxL000001eb-rtwIt2gI0FxT+ZUat5FNkAK/GNPrWCqfQQ4Iyu8u01E@public.gmane.org>
2008-06-12  8:45         ` Peter Leckie

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=484F86B7.7040109@sgi.com \
    --to=pleckie@sgi.com \
    --cc=linux-nfs@vger.kernel.org \
    --cc=talpey@netapp.com \
    --cc=trond.myklebust@fys.uio.no \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox