From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-0.8 required=3.0 tests=HEADER_FROM_DIFFERENT_DOMAINS, MAILING_LIST_MULTI,SPF_PASS autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 1FC6DC64EBC for ; Thu, 4 Oct 2018 09:08:17 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.kernel.org (Postfix) with ESMTP id C8B41213A2 for ; Thu, 4 Oct 2018 09:08:16 +0000 (UTC) DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org C8B41213A2 Authentication-Results: mail.kernel.org; dmarc=fail (p=none dis=none) header.from=redhat.com Authentication-Results: mail.kernel.org; spf=none smtp.mailfrom=linux-kernel-owner@vger.kernel.org Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1727420AbeJDQAc (ORCPT ); Thu, 4 Oct 2018 12:00:32 -0400 Received: from mx1.redhat.com ([209.132.183.28]:57606 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1726808AbeJDQAc (ORCPT ); Thu, 4 Oct 2018 12:00:32 -0400 Received: from smtp.corp.redhat.com (int-mx05.intmail.prod.int.phx2.redhat.com [10.5.11.15]) (using TLSv1.2 with cipher AECDH-AES256-SHA (256/256 bits)) (No client certificate requested) by mx1.redhat.com (Postfix) with ESMTPS id 580843081256; Thu, 4 Oct 2018 09:08:14 +0000 (UTC) Received: from ovpn-117-12.ams2.redhat.com (ovpn-117-12.ams2.redhat.com [10.36.117.12]) by smtp.corp.redhat.com (Postfix) with ESMTP id 040BF5D6A6; Thu, 4 Oct 2018 09:08:12 +0000 (UTC) Message-ID: Subject: Re: [PATCH net] udp: Allow kernel service to avoid udp socket rx queue From: Paolo Abeni To: David Howells , netdev@vger.kernel.org Cc: linux-afs@lists.infradead.org, linux-kernel@vger.kernel.org Date: Thu, 04 Oct 2018 11:08:11 +0200 In-Reply-To: <153859250219.15389.11970533498295122206.stgit@warthog.procyon.org.uk> References: <153859250219.15389.11970533498295122206.stgit@warthog.procyon.org.uk> Content-Type: text/plain; charset="UTF-8" Mime-Version: 1.0 Content-Transfer-Encoding: 7bit X-Scanned-By: MIMEDefang 2.79 on 10.5.11.15 X-Greylist: Sender IP whitelisted, not delayed by milter-greylist-4.5.16 (mx1.redhat.com [10.5.110.49]); Thu, 04 Oct 2018 09:08:14 +0000 (UTC) Sender: linux-kernel-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Hi, On Wed, 2018-10-03 at 19:48 +0100, David Howells wrote: > There's a problem somewhere skb_recv_udp() that doesn't decrease the > sk_rmem_alloc counter when a packet is extracted from the receive queue by > a kernel service. If this is the case, it's really bad and need an explicit fix. However it looks like sk_rmem_alloc is reclaimed by skb_recv_udp(), as it ends- up calling udp_rmem_release() on succesfull dequeue. udp_rmem_release() can delay sk_rmem_alloc reclaiming is the rx queue is almost empty, due to commit 6b229cf77d683f634f0edd876c6d1015402303ad. Anyhow I don't see either as that may affect this scenario: if only 1/4 of the maxium receive buffer size is used, the next packet should always be able to land into the rx queue, with the default sk_rcvbuf and every reasonable truesize. > Further, there doesn't seem any point in having the socket buffer being > added to the UDP socket's rx queue since the rxrpc's data_ready handler > takes it straight back out again (more or less, there seem to be occasional > hold ups there). I really would really try to avoid adding another indirect call in the data-path, unless strictly needed (to avoid more RETPOLINE overhead for all other use-case). If skipping altogether the enqueuing makes sense (I guess so, mostily for performance reasons), I *think* you can use the already existing encap_rcv hook, initializing it to the rxrpc input function, and updating such function to pull the udp header and ev. initializing the pktinfo, if needed. Please see e.g. l2tp usage. > Putting in some tracepoints show a significant delay occurring between packets > coming in and thence being delivered to rxrpc: > > -0 [001] ..s2 67.631844: net_rtl8169_napi_rx: enp3s0 skb=07db0a32 > ... > -0 [001] ..s4 68.292778: rxrpc_rx_packet: d5ce8d37:bdb93c60:00000002:09c7 00000006 00000000 02 20 ACK 660967981 skb=07db0a32 > > The "660967981" is the time difference in nanoseconds between the sk_buff > timestamp and the current time. It seems to match the time elapsed between > the two trace lines reasonably well. I've seen anything up to about 4s. Can you please provide more data? specifically can you please add: * a perf probe in rxrpc_data_ready() just after skb_recv_udp() reporting the sk->sk_rmem_alloc and skb->truesize * a perf probe in __udp_enqueue_schedule_skb() just before the 'if (rmem > sk->sk_rcvbuf)' test reporting again sk->sk_rmem_alloc, skb- >truesize and sk->sk_rcvbuf And then provide the perf record -g -e ... /perf script output? Thanks, Paolo