From mboxrd@z Thu Jan 1 00:00:00 1970 From: Nix Subject: Re: [PATCH RFT net-next #2 0/6] via-rhine receive buffers rework Date: Wed, 08 Apr 2015 23:43:55 +0100 Message-ID: <87mw2iw89g.fsf@spindle.srvr.nix> References: <87vbh6364u.fsf@spindle.srvr.nix> <20150408215051.GA25326@electric-eye.fr.zoreil.com> Mime-Version: 1.0 Content-Type: text/plain Cc: netdev@vger.kernel.org, "David S. Miller" , rl@hellgate.ch, Bjarke Istrup Pedersen To: Francois Romieu Return-path: Received: from icebox.esperi.org.uk ([81.187.191.129]:41952 "EHLO mail.esperi.org.uk" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753288AbbDHWoD (ORCPT ); Wed, 8 Apr 2015 18:44:03 -0400 In-Reply-To: <20150408215051.GA25326@electric-eye.fr.zoreil.com> (Francois Romieu's message of "Wed, 8 Apr 2015 23:50:51 +0200") Sender: netdev-owner@vger.kernel.org List-ID: On 8 Apr 2015, Francois Romieu stated: > Nix : > [...] >> I am sorry to report that I just had a watchdog-triggered autoreboot >> during testing of this patch series :( with no log messages of any kind. >> looks like the underlying bug is still there, or another bug with the >> same symptoms (i.e. some way to crash inside the rx handler). I qwish I >> could get some debugging output when this happens! > > You may add the patch below on top of the current stack. I don't expect > much difference. Increasing RX_RING_SIZE could be a different story. Will try it tomorrow. > Did you keep netconsole disabled and did you increse via-rhine verbosity > level ? netconsole is active. The verbosity level is default because I didn't notice the driver had one to set! I'll push it up and see what happens. I don't expect too much: this is, after all, a uniprocessor, and if you're stuck in an interrupt handler there's not much it can do... > No shared IRQ ? None relevant that I can see: CPU0 0: 692019 XT-PIC timer 2: 0 XT-PIC cascade 4: 314 XT-PIC serial 5: 114062 XT-PIC adsl 7: 329914 XT-PIC cs5535-clockevt 9: 137080 XT-PIC bdsl 10: 2006 XT-PIC voip 11: 1546540 XT-PIC gordianet 12: 724187 XT-PIC wireless 14: 12313 XT-PIC pata_cs5536 15: 219702 XT-PIC ehci_hcd:usb1, ohci_hcd:usb2 The only shared interrupt is usb2, and since that's an internal-to-the-case thing that's not plugged in to anything, in effect there are none shared at all. (The unplugged interfaces *are* shared -- the last four of the eight interfaces on the box run 10, 7, 10, 7 -- but since only one of those has anything plugged into it, the net effect is no sharing.) My load tests are being performed between the gordianet and wireless interfaces, IRQs 11 and 12, which are entirely unshared. >> However, to give some good news, CPU usage is much lower than before the >> patch: si ~10%, rather than ~80% with spikes of full CPU usage: >> ksoftirqd's CPU usage is steady at about 3% rather than 40--60% with >> spikes to 100%, and some of that will be USB interrupts from the >> continuous USB traffic from my entropy key. > > Huuuuh ? Which entropy key ? It's a USB device that gives you about 4KiB/s over USB: . No longer manufactured, alas :( but as it provides entropy continuously it is a continuous source of interrupts. (But from the USB port, obviously.) -- NULL && (void)