From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from dpdk.org (dpdk.org [92.243.14.124]) by inbox.dpdk.org (Postfix) with ESMTP id BFA7BA0535; Tue, 4 Feb 2020 17:02:34 +0100 (CET) Received: from [92.243.14.124] (localhost [127.0.0.1]) by dpdk.org (Postfix) with ESMTP id F1A151C206; Tue, 4 Feb 2020 17:02:33 +0100 (CET) Received: from relay5-d.mail.gandi.net (relay5-d.mail.gandi.net [217.70.183.197]) by dpdk.org (Postfix) with ESMTP id 7C9E11C204 for ; Tue, 4 Feb 2020 17:02:32 +0100 (CET) X-Originating-IP: 37.58.153.206 Received: from [10.0.3.185] (bny206.haproxy.com [37.58.153.206]) (Authenticated sender: grive@u256.net) by relay5-d.mail.gandi.net (Postfix) with ESMTPSA id 8BC4D1C0012; Tue, 4 Feb 2020 16:02:31 +0000 (UTC) To: Thomas Monjalon , Pavan Nikhilesh Bhagavatula , Vamsi Krishna Attunuru Cc: David Marchand , dev@dpdk.org, Jerin Jacob Kollanukkaran References: <3503128.aV6nBDHxoP@xps> <4205291.LvFx2qVVIh@xps> From: Gaetan Rivet Message-ID: <790778b3-8d6c-0ec2-79dd-7fb254f021cd@u256.net> Date: Tue, 4 Feb 2020 17:02:31 +0100 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101 Thunderbird/68.4.2 MIME-Version: 1.0 In-Reply-To: <4205291.LvFx2qVVIh@xps> Content-Type: text/plain; charset=utf-8; format=flowed Content-Language: en-US Content-Transfer-Encoding: 7bit Subject: Re: [dpdk-dev] [PATCH v7] eal: add manual probing option X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.15 Precedence: list List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org Sender: "dev" On 04/02/2020 16:06, Thomas Monjalon wrote: > 04/02/2020 13:43, Gaetan Rivet: >> On 04/02/2020 12:07, Thomas Monjalon wrote: >>> 04/02/2020 11:03, Gaetan Rivet: >>>> On 03/02/2020 23:21, Thomas Monjalon wrote: >>>>> 03/02/2020 06:16, Pavan Nikhilesh Bhagavatula: >>>>>> @David Marchand @thomas@monjalon.net >>>>>> >>>>>> Ping? >>>>>> >>>>>> Are there any more changes required for this patch? It's been in queue since last October. >>>>> >>>>> Sorry we have not decided whether it is a good idea or not. >>>>> >>>>> All changes related to probing are very sensitive, >>>>> and we know a big refactoring would be better than stacking >>>>> more and more options and corner cases. >>>>> >>>>> As we are busy with ABI stability stuff, we did not allocate >>>>> enough time to properly think about this feature. >>>>> Please accept our apologies, and let's consider it as >>>>> a high priority for 20.05 cycle. >>>>> >>>> >>>> Hello Thomas, >>>> >>>> This is unfortunate. I pushed Pavan to accept an alternative implementation of this functionality that was less obtrusive, to make the integration smoother. I took care to alleviate those risks from the common path. >>>> >>>> The big refactoring is needed yes, but considering the current path I'm not seeing it happen in 20.05. If that means taking this patch as-is in 20.05 for Marvell users, I'm not sure much is gained from waiting 3 months, except minimal risk avoidance. >>> >>> >>> Yes, life is full of bad decisions and consequences. >> >> >> Ah, yes, but I stand by my initial opinion, the first implementation [1] was riskier and less useful. >> >>> >>> I still think there is a risk in adding new user expectations, >>> and maintaining some code to workaround unknown issues. >>> >>> The real question here is to know why this patch? >>> Is it to workaround a broken driver? >>> Or to workaround a broken design in EAL and bus drivers? >> >> Two birds - one stone here: OVS needed a way to disable automatic probing cleanly (current workaround seen in multiple deployment is to add a dummy whitelisted device, which will be ignored by the PCI bus --> it sets the bus in whitelist mode but avoid probing anything), and as a bonus this option allows using devices that depends on other devices being probed already (LAG, representors, failsafe, etc). >> >> I'm not sure having a dependent-probe by default is good, and that would be a big change. >> >> If we are doing the genesis of this patch, the initial motivation should be asked for more details from Marvell people and David for the OVS side. >> >> [1]: First proposal: >> http://mails.dpdk.org/archives/dev/2019-September/144166.html >> My arguments: >> http://mails.dpdk.org/archives/dev/2019-September/144564.html > > > OK so there are two needs: > > 1/ Better control whitelist/blacklist mode. > We already know that a rework is needed here. > Unfortunately neither you or me had time to work on it, > and others who were interested disappeared. > > 2/ Associate ports with equivalent properties in applications. > This must be done in applications. > Tweaking the probe order is a hack. > > An application that want to tightly control the port init order, currently (by doing exactly like me here, hotpluging one by one the ports), would still need the bigger hack that consist in inserting a whitelist PCI devargs with a dummy address, depending on a undocumented PCI bus feature consisting in ignoring matching errors but keeping probing policy from failed devargs processing. Instead, with this patch this app can do rte_eal_manual_probe_set(1); rte_eal_init(); to have the same behavior and be able to hotplug ports as it sees fit. You are worried about creating user expectations about this behavior (being forced to replicate in some way the functionality during the rewrite, as I understand it?), but then you are currently forcing users to expect this workaround to exist in the PCI bus, blocking devs from touching it as it will thus break current app configurations. I've seen systemd unit file using this -w dummy flag, as well as the programmatic equivalent. Which is better, to have to rework it cutting short these configs, or to propose beforehand a deprecation path?. This rework won't happen in 20.05, nor in the medium term unless you decide to drive this change. This workaround serves three needs (PCI normalization, port congruence and port dependency) in a low-risk implementation.