From mboxrd@z Thu Jan 1 00:00:00 1970 From: Eric Dumazet Subject: Re: Problems: network rx out-of-order issue Date: Wed, 28 Mar 2018 07:19:16 -0700 Message-ID: <600ea635-9399-e889-6c41-44df47d6f8eb@gmail.com> References: <1522229201.8705.44.camel@mbjswglx818> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 7bit Cc: zhen.jiang@mediatek.com To: Anny Hu , netdev@vger.kernel.org Return-path: Received: from mail-pl0-f51.google.com ([209.85.160.51]:41887 "EHLO mail-pl0-f51.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752899AbeC1OTS (ORCPT ); Wed, 28 Mar 2018 10:19:18 -0400 Received: by mail-pl0-f51.google.com with SMTP id b7-v6so1653517plr.8 for ; Wed, 28 Mar 2018 07:19:18 -0700 (PDT) In-Reply-To: <1522229201.8705.44.camel@mbjswglx818> Content-Language: en-US Sender: netdev-owner@vger.kernel.org List-ID: On 03/28/2018 02:26 AM, Anny Hu wrote: > Dears, > > Recently, we find the following patch will impact multi-core network throughput performance on kernel-4.9, > for it will cause rx packet out-of-order. > commit id: 4cd13c21b207e80ddb1144c576500098f2d5f882 > > [kernel version]: > kernel-4.9 > > [repeat steps] > 1. two our phones with load based on kernel-4.9, enable wifi and set up P2P connection; > 2. two reference phones with load based on kernel-4.4, enable wifi and set up P2P connection; > 3. do iperf UDP downlink test with the following commands; > Client: adb shell iperf -c 192.168.49.xx -u -i 1 -t 60 -b 15M > Server: adb shell iperf -s -u > 4. out-of-order rate of our phones are obviously higher than reference phone: > DUT Lost rate out of order > Our phone 0.011% 164 > Reference phone 0.0091% 1 > > [Code flow] > Wi-Fi driver trigger rx_thread to receive rx packets and call netif_rx_ni() to push packets to tcp stack in thread context. > > [netlog]: > No Time Source ip Dest ip Protocol Len Ip id Info > 3811 2010-01-01 08:11:35.075522 192.168.43.1 192.168.43.90 UDP 144 0x25ee (9710) 60834 -> 5201 Len=100 > 3812 2010-01-01 08:11:35.075805 192.168.43.1 192.168.43.90 UDP 144 0x25f5 (9717) 60834 -> 5201 Len=100 > 3813 2010-01-01 08:11:35.075549 192.168.43.1 192.168.43.90 UDP 144 0x25ef (9711) 60834 -> 5201 Len=100 > 3814 2010-01-01 08:11:35.075944 192.168.43.1 192.168.43.90 UDP 144 0x25f6 (9718) 60834 -> 5201 Len=100 > 3815 2010-01-01 08:11:35.075567 192.168.43.1 192.168.43.90 UDP 144 0x25f0 (9712) 60834 -> 5201 Len=100 > > out-of-order issue happens on: > 1) ip_id=0x25f5 is received before ip_id=0x25ef, which are belong to the same downlink connection; > 2) ip_id=0x25f6 is received before ip_id=0x25f0, which are belong to the same downlink connection. > > > [kernel log]: > We add debug log in rx kernel flow, such as netif_rx_ni() -> enqueue_to_backlog -> do_softirq() -> net_rx_action() -> process_backlog() > > kernel time > | cpu id which process running on > | | process name function name > | | | | > [ 598.837358] (7)[3282:rx_thread]: [RX]enqueue_to_backlog: ip_id=0x25ee, pending=0x8a, backlog_state=1 > > // wifi rx_thread is running on cpu7 to receive rx packets of ip_id 0x25ee ~ 0x25f4, which are only enqueued. > [ 598.837366] (7)[3282:rx_thread]: [RX]enqueue_to_backlog: ip_id=0x25ee, input_qlen=1, proc_qlen=0, pending=0x8a, backlog_state=1 > > // because ksoftirqd is in running queue, do_softirq() exit and net_rx_action isn't executed actually. > > [ 598.837373] (7)[3282:rx_thread]: [RX]do_softirq: preempt=0x1, in_interrupt()=0x0, pending=0x8a, ksoftirqd_running=1 > > [ 598.837383] (7)[3282:rx_thread]: [RX]enqueue_to_backlog: ip_id=0x25ef,input_qlen=2, proc_qlen=0, pending=0x8a, backlog_state=1 > [ 598.837391] (7)[3282:rx_thread]: [RX]do_softirq: preempt=0x1, in_interrupt()=0x0 pending=0x8a, ksoftirqd_running=1 > [ 598.837400] (7)[3282:rx_thread]: [RX]enqueue_to_backlog: ip_id=0x25f0,input_qlen=3, proc_qlen=0, pending=0x8a, backlog_state=1 > [ 598.837407] (7)[3282:rx_thread]: [RX]do_softirq: preempt=0x1, in_interrupt()=0x0, pending=0x8a, ksoftirqd_running=1 > [ 598.837417] (7)[3282:rx_thread]: [RX]enqueue_to_backlog: ip_id=0x25f1,input_qlen=4, proc_qlen=0, pending=0x8a, backlog_state=1 > [ 598.837424] (7)[3282:rx_thread]: [RX]do_softirq: preempt=0x1, in_interrupt()=0x0, pending=0x8a, ksoftirqd_running=1 > [ 598.837434] (7)[3282:rx_thread]: [RX]enqueue_to_backlog: ip_id=0x25f2,input_qlen=5, proc_qlen=0, pending=0x8a, backlog_state=1 > [ 598.837441] (7)[3282:rx_thread]: [RX]do_softirq: preempt=0x1, in_interrupt()=0x0, pending=0x8a, ksoftirqd_running=1 > [ 598.837450] (7)[3282:rx_thread]: [RX]enqueue_to_backlog: ip_id=0x25f3,input_qlen=6, proc_qlen=0, pending=0x8a, backlog_state=1 > [ 598.837458] (7)[3282:rx_thread]: [RX]do_softirq: preempt=0x1, in_interrupt()=0x0, pending=0x8a, ksoftirqd_running=1 > > // after ip_id(0x25f4), there is no rx packet in this wifi interrupt, so rx_thread exit. > [ 598.837470] (7)[3282:rx_thread]: [RX]enqueue_to_backlog: ip_id=0x25f4,input_qlen=7, proc_qlen=0, pending=0x8a, backlog_state=1 > [ 598.837479] (7)[3282:rx_thread]: [RX]do_softirq: preempt=0x1, in_interrupt()=0x0, pending=0x8a, ksoftirqd_running=1 > > // ksoftirqd is scheduled to run and call do_softirq(), so net_rx_action() is executed. > [ 598.837511] (7)[45:ksoftirqd/7]: [RX]net_rx_action: budget=300, time=4295042005 > // net_rx_action() call process_backlog() to handle rx packet of ip_id=0x25ee > > [ 598.837646] (7)[45:ksoftirqd/7]: [RX]process_backlog: ip_id=0x25ee, delay=126308, work=1, input_qlen=0, proc_qlen=6 > > // wifi interrupt happens and trigger rx_thread to receive the next rx packet, which is running on cpu1 this time. > [ 598.837660] (1)[3282:rx_thread]: [RX]enqueue_to_backlog: ip_id=0x25f5, input_qlen=1, proc_qlen=0, pending=0x8, backlog_state=1 > // because ksoftirqd is not running, net_rx softirq is executed immediately. > [ 598.837666] (1)[3282:rx_thread]: [RX]do_softirq: preempt=0x1, in_interrupt()=0x0, pending=0x8, ksoftirqd_running =0 > > // so ksoftirqd on cpu7 and rx_thread on cpu1 are running at the same time. > [ 598.837679] (1)[3282:rx_thread]: [RX]net_rx_action: budget=300, time=4295042005 > > // ip_id=0x25ef is handled on cpu7 by net_rx softirq in ksoftirqd() thread context. > [ 598.837762] (7)[45:ksoftirqd/7]: [RX]process_backlog: ip_id=0x25ef, delay=70385, work=2, input_qlen=0, proc_qlen=5 > > // ip_id=0x25f5 is handled on cpu1 by net_rx softirq in rx_thread context, out-of-order issue happens. > > [ 598.837768] (1)[3282:rx_thread]: [RX]process_backlog: ip_id=0x25f5, delay=82000, work=1, quota=64, input_qlen=0, proc_qlen=0 > > // ip_id=0x25f0 is handled on cpu7 by net_rx softirq in ksoftirqd() thread context. > [ 598.837878] (7)[45:ksoftirqd/7]: [RX]process_backlog:ip_id=0x25f0, delay=47539, work=3, input_qlen=0, proc_qlen=4 > > > Could you please help to check if this out-of-order issue is caused by kernel-4.9 patch above? > > Thanks very much! > A driver calling netif_rx_ni() can have out of order packets, regardless of commit 4cd13c21b207e80ddb1144c57650 This patch has changed scheduling, and might have pointed out ancient bugs/issues. 'thread context' means cpu can change under us (preempt kernels) Which driver are you using exactly ?