From patchwork Tue Dec 20 20:24:21 2022 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Dave Taht X-Patchwork-Id: 635580 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by smtp.lore.kernel.org (Postfix) with ESMTP id EF837C4332F for ; Tue, 20 Dec 2022 20:24:40 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S233914AbiLTUYj (ORCPT ); Tue, 20 Dec 2022 15:24:39 -0500 Received: from lindbergh.monkeyblade.net ([23.128.96.19]:58452 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S229906AbiLTUYh (ORCPT ); Tue, 20 Dec 2022 15:24:37 -0500 Received: from mail-wm1-x330.google.com (mail-wm1-x330.google.com [IPv6:2a00:1450:4864:20::330]) by lindbergh.monkeyblade.net (Postfix) with ESMTPS id 7E4123AF for ; Tue, 20 Dec 2022 12:24:36 -0800 (PST) Received: by mail-wm1-x330.google.com with SMTP id v124-20020a1cac82000000b003cf7a4ea2caso12010552wme.5 for ; Tue, 20 Dec 2022 12:24:36 -0800 (PST) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20210112; h=to:subject:message-id:date:from:mime-version:from:to:cc:subject :date:message-id:reply-to; bh=c2bFmxYT3zvM7I3Ys48664PKtgjsbzVbriVicmQMdGw=; b=jSWkCIpHIW4m8uVPwg+nUCL1OBhT77YlzX+W02eY3BqewrtkOLe7Y8cmO1xVMGLGEO oqVG9ZWCwgVjxq22ZJ1A3PuOwmgd/YqnJMgsZx+jAqHL4v3YkKzj1faqXDkHFQk53Ll9 l9m5fH401mTOEp2kbdiOHog/0mAIIoHJ43eyR+YGmIC+GA4I8HoMDlduckAuQmW2bpBw dJMc0VU1YADCD4EqbLcLXJSZed4ebQYmt4KV/c6daIhBZ2vqm+1y4z1B3Nl2FJND1evC m2RV9vSqAhDhNKCoxanLhNXP3KkkwC1GVpDqPp7U39yqtQoOqZVqqqVRIrG3acgwLFl8 yIcw== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=to:subject:message-id:date:from:mime-version:x-gm-message-state :from:to:cc:subject:date:message-id:reply-to; bh=c2bFmxYT3zvM7I3Ys48664PKtgjsbzVbriVicmQMdGw=; b=29tyIqxGKHXnAFRukSqzHwNiXWb158JY9jD35lM1PWJuZzw4hJIQOQ53Izaq7FEVny Kx4kDU33fO23gshypUAw2Fzr4e+pplzm5SoHEqwtHlNExWQchgdBcqUeDDJzl7naMw4q /MXlIBM43sHWywUPCUT5Cpx62c6g4LlfgUesDzvNvkCXgAHdwow+EvfX/qR2GUXw99Nj U4pUeJcfRLtrXbjk1XpF49iE5GiNPa30pgpS97oUP9ItnpreQ6VK0parufezO6i6BGHe ZxEGjOiHBR9daF5Ac4LLMZ7ZW9FiX4au64L7kHWnpSXd7gSBdJnyf+PmNQwJXcjyyz+o tF8Q== X-Gm-Message-State: AFqh2krOmbYEFQlztbTdo8oQLwptq7JTlbe94Eulg/DZXbQKxj8WciTG QrWh+64CGa9UPMr0i7IBZRb/fOcAlH/l/VImDk/K2OZckk0= X-Google-Smtp-Source: AMrXdXuSq2yf8Bgcpq6fUuZvxIgsc4lWZnigfzRkTDiOBwUub5E5UWnIZpQ+V3PWDTcGplUP5UMTnk6O7SL83ywOuFc= X-Received: by 2002:a05:600c:22d4:b0:3cf:a6e8:b59b with SMTP id 20-20020a05600c22d400b003cfa6e8b59bmr1503613wmg.128.1671567874927; Tue, 20 Dec 2022 12:24:34 -0800 (PST) MIME-Version: 1.0 From: Dave Taht Date: Tue, 20 Dec 2022 12:24:21 -0800 Message-ID: Subject: a nuking the mac80211 changing codel parameters patch To: reiser4@gmail.com, Make-Wifi-fast , OpenWrt Development List , linux-wireless Precedence: bulk List-ID: X-Mailing-List: linux-wireless@vger.kernel.org This is the single, most buggy, piece of code in "my" portion of wifi today. It is so wrong, yet thus far I cannot get it out of linux or find an acceptable substitute. It makes it hard to sleep at night knowing this code has been so wrong... and now in millions , maybe even 10s of millions, of devices by now.... Since I've been ranting about the wrongness of this for years, I keep hoping that we can excise it, especially for wifi6 devices and even more especially on 6ghz spectrum... but just about everything, somehow, would benefit hugely if we could somehow do more of the right thing here. I'd tried, last time I got this bee in my bonnet, tried to nuke this call here: https://forum.openwrt.org/t/reducing-multiplexing-latencies-still-further-in-wifi/133605/ As it is, I really encourage folk, especially with mt79 and to some extent mt76 ac or ath10k, to try out the attached patch, measure tcp rtts, and throughput, etc. A slightly less aggressive patch might suit wifi-n.... Maybe there's a reason for keeping this code in linux wifi that I do not understand. But here are my pithy comments as to why this part of mac80211 is so wrong... static void sta_update_codel_params(struct sta_info *sta, u32 thr) { - if (thr && thr < STA_SLOW_THRESHOLD * sta->local->num_sta) { 1) sta->local->num_sta is the number of associated, rather than active, stations. "Active" stations in the last 50ms or so, might have been a better thing to use, but as most people have far more than that associated, we end up with really lousy codel parameters, all the time. Mistake numero uno! 2) The STA_SLOW_THRESHOLD was completely arbitrary in 2016. - sta->cparams.target = MS2TIME(50); This, by itself, was probably not too bad. 30ms might have been better, at the time, when we were battling powersave etc, but 20ms was enough, really, to cover most scenarios, even where we had low rate 2Ghz multicast to cope with. Even then, codel has a hard time finding any sane drop rate at all, with a target this high. - sta->cparams.interval = MS2TIME(300); But this was horrible, a total mistake, that is leading to codel being completely ineffective in almost any scenario on clients or APS. 100ms, even 80ms, here, would be vastly better than this insanity. I'm seeing 5+seconds of delay accumulated in a bunch of otherwise happily fq-ing APs.... 100ms of observed jitter during a flow is enough. Certainly (in 2016) there were interactions with powersave that I did not understand, and still don't, but if you are transmitting in the first place, powersave shouldn't be a problemmmm..... - sta->cparams.ecn = false; At the time we were pretty nervous about ecn, I'm kind of sanguine about it now, and reliably indicating ecn seems better than turning it off for any reason. - } else { - sta->cparams.target = MS2TIME(20); - sta->cparams.interval = MS2TIME(100); - sta->cparams.ecn = true; - } And if we aint gonna fiddle with these, we don't need these either. In production, on p2p wireless, I've had 8ms and 80ms for target and interval for years now, and it works great. It is obviously too low, for those that prize bandwidth over latency (I personally would prefer TXOPs shrink intelligently as well as bandwidth, as you add stations, some of which happens naturally by fq-codels scheduling mechanisms, others don't, I even run with 2ms txops by default on everything myself) + return; Ideally we could kill this entire call off entirely. } A pre-thx for anyone actually trying the attached patch and reporting back on any results. https://forum.openwrt.org/t/reducing-multiplexing-latencies-still-further-in-wifi/133605/ diff --git a/net/mac80211/sta_info.c b/net/mac80211/sta_info.c index 97d24e9..457209a 100644 --- a/net/mac80211/sta_info.c +++ b/net/mac80211/sta_info.c @@ -2766,15 +2766,7 @@ unsigned long ieee80211_sta_last_active(struct sta_info *sta) static void sta_update_codel_params(struct sta_info *sta, u32 thr) { - if (thr && thr < STA_SLOW_THRESHOLD * sta->local->num_sta) { - sta->cparams.target = MS2TIME(50); - sta->cparams.interval = MS2TIME(300); - sta->cparams.ecn = false; - } else { - sta->cparams.target = MS2TIME(20); - sta->cparams.interval = MS2TIME(100); - sta->cparams.ecn = true; - } + return; } void ieee80211_sta_set_expected_throughput(struct ieee80211_sta *pubsta,