From mboxrd@z Thu Jan 1 00:00:00 1970 From: Sowmini Varadhan Subject: Re: [PATCH net-next 2/2] sunvnet: Re-check for a VIO_DESC_READY data descriptor after short udelay() Date: Tue, 02 Sep 2014 06:27:54 -0400 Message-ID: <54059BAA.1080307@oracle.com> References: <20140829201809.GH14753@oracle.com> <20140901.204715.2121922580015092507.davem@davemloft.net> Reply-To: sowmini.varadhan@oracle.com Mime-Version: 1.0 Content-Type: text/plain; charset=windows-1252; format=flowed Content-Transfer-Encoding: 7bit Cc: raghuram.kothakota@oracle.com, netdev@vger.kernel.org To: David Miller Return-path: Received: from aserp1040.oracle.com ([141.146.126.69]:45805 "EHLO aserp1040.oracle.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752355AbaIBK16 (ORCPT ); Tue, 2 Sep 2014 06:27:58 -0400 In-Reply-To: <20140901.204715.2121922580015092507.davem@davemloft.net> Sender: netdev-owner@vger.kernel.org List-ID: On 09/01/2014 11:47 PM, David Miller wrote: > > If there were no more packets coming, this is wasted useless polling > time in atomic context. turns out that this gives a 20-30% perf improvement for tests like iperf. when there are no more packets coming, the extra 12 microsecond delay is not that big of a deal anyway. The point was that the extra 12 micro-second tax in the quiescent network state is less expensive than exiting interrupt context, taking another interrupt and doing another ldc_read, when there is actually a burst of packets. notice that there are many other such udelay() loops elsewhere in the code. I can remove the retries and submit patch 1/1 again later today. > > Everything should be event based, and we should not be compensating > and making sacrifices for a producer slower than we are as a consumer. > -- > To unsubscribe from this list: send the line "unsubscribe netdev" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html >