From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-wi0-f171.google.com (mail-wi0-f171.google.com [209.85.212.171]) by dpdk.org (Postfix) with ESMTP id 73E9C9208 for ; Tue, 15 Sep 2015 12:04:59 +0200 (CEST) Received: by wiclk2 with SMTP id lk2so19820724wic.1 for ; Tue, 15 Sep 2015 03:04:59 -0700 (PDT) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20130820; h=x-gm-message-state:date:from:to:cc:subject:message-id :mail-followup-to:references:mime-version:content-type :content-disposition:content-transfer-encoding:in-reply-to; bh=9BPSVuqZ640klkgH8rifL5OpIpyqDFqR1jf9vYFX+9k=; b=Q/Oywp+C4pxk0Fur0IZBSb0nEoKor94Sw9vwVpWcbV9riBtSdY1ggPr7hGir4kYBK9 odBHA99CTL/fq6hOjzkFT8dSyC7WtrYKG8FG1FHcd4sL/mfC+iDqBUUZJSHkZ+3s7i/2 gWR3/BcdXlU0rRMX3XaN8k++4ELmwg6q7nHbYGqUDL+BF6STyFztgoYPGzREqNP/iF6S WeJugXd0rneeG3mBbsqi6y7NIbPhRdAKcWa3pNs2AjpmtrU92ltfl+KEgPELJCzlqJjF o2MTgbN06owyfdOEjOYFThDwDG8Hz+3OyLg23vXVBDiPZ+AVcTyflQAfskCzNbM023La Wa9w== X-Gm-Message-State: ALoCoQmz/Gvt8qxFSHX59/6N33f0w3pDyi0mtAZi00A04L19HCC8oWbeupsB9SeS3VkMfqvL1f93 X-Received: by 10.194.236.200 with SMTP id uw8mr7679673wjc.126.1442311499218; Tue, 15 Sep 2015 03:04:59 -0700 (PDT) Received: from 6wind.com (guy78-3-82-239-227-177.fbx.proxad.net. [82.239.227.177]) by smtp.gmail.com with ESMTPSA id oh5sm7772915wic.11.2015.09.15.03.04.57 (version=TLSv1 cipher=RC4-SHA bits=128/128); Tue, 15 Sep 2015 03:04:58 -0700 (PDT) Date: Tue, 15 Sep 2015 12:04:43 +0200 From: Adrien Mazarguil To: Marc Sune Message-ID: <20150915100443.GA4924@6wind.com> Mail-Followup-To: Marc Sune , =?utf-8?B?TsOpbGlv?= Laranjeiro , Morten =?utf-8?Q?Br=C3=B8rup?= , Thomas Monjalon , dev@dpdk.org, Olga Shern References: <20150909131037.GA25122@autoinstall.dev.6wind.com> <2699193.9riTyGPe1z@xps13> <2046894.c3eJ0QZGuc@xps13> <98CBD80474FA8B44BF855DF32C47DC358AF3B7@smartserver.smartshare.dk> <98CBD80474FA8B44BF855DF32C47DC358AF3BD@smartserver.smartshare.dk> <20150915082544.GG25122@autoinstall.dev.6wind.com> MIME-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Disposition: inline Content-Transfer-Encoding: 8bit In-Reply-To: Cc: dev@dpdk.org, Morten =?utf-8?Q?Br=C3=B8rup?= Subject: Re: [dpdk-dev] [PATCH v4 0/2] ethdev: add port speed capability bitmap X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.15 Precedence: list List-Id: patches and discussions about DPDK List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Tue, 15 Sep 2015 10:04:59 -0000 Hi Marc, Adding my thoughts to the discussion, see below. On Tue, Sep 15, 2015 at 10:48:03AM +0200, Marc Sune wrote: > I will answer Morten in another mail, because I got his point on the > AUTONEG as a separate bit, and it _makes_ sense to me. > > But Neilo, > > 2015-09-15 10:25 GMT+02:00 Nélio Laranjeiro : > > > On Tue, Sep 15, 2015 at 12:50:11AM +0200, Morten Brørup wrote: > > > Comments inline, marked MB>. > > > > > > Med venlig hilsen / kind regards > > > - Morten Brørup > > > > > > Marc Sune on 14. september 2015 23:34 wrote: > > > > > > 2015-09-14 12:52 GMT+02:00 Morten Brørup : > > > > It is important to consider that a multipath link (bonding etc.) is > > not a physical link, but a logical link (built on top of multiple physical > > links). Regardless whether it is a Layer2 link aggregate (IEEE 802.1ad, > > Ethernet bonding, EtherChannel, DSL pair bonding, etc.) or a Layer3 > > multipath link (e.g. simultaneously using Wi-Fi and mobile networks). So it > > doesn't make sense trying to impose physical link properties on a purely > > logical link. Likewise, it doesn't make sense to impose logical link > > properties on physical links. In other words: Don't consider bonding or any > > other logical link types when designing the PHY API. > > > > > > +1 > > > > +1. I agree with the fact that physical link properties do not make sense for logical links, however in the case of the bonding PMD, the aggregated link speed can be actually useful for applications (assuming it is kept up to date, I think it's the case). The current API certainly allows this. > > > > I think there is consensus that 1/ (PHY capabilities) and 2/ (PHY > > advertisements) should use the same definitions, specifically a bitmap > > field. And when you disregard bonding, I don't see any reason to use > > different definitions for 3/ (PHY negotiation result). This makes it one > > unified API for all three purposes. > > > > > > Agree. > > > > I don't agree with this one, some PMDs don't use the advertise of > > autoneg result to get the speed or the duplex. You make a > > generality from your case above all PMDs. > > > > can you please explain how a particular PMD is recovering the actual link > speed and the duplex has to do with the design of the (general) API? It's not so much about the way PMDs recover link information, rather about the amount of changes required to switch to a bit-field API for the current link speed with no clear advantage. All PMDs must be modified, the initial set of patches isn't complete in this regard. > > Mellanox get the speed, duplex and status information from IOCTLs > > which are not related to your bitmap. So at least for this PMD, there > > is already a conversion from 3 fields to a bitmap, knowing that it will > > use the speed as an integer after. What is the benefit of your solution? > > > > I said already I don't have a strong preference for 3/. But steering the > design of an API through a "minimum common denominator" principle is not a > good idea, specially since we are talking about a super simple mapping > issue for this specific PMD. I think Nelio was using mlx4 as an example, all PMDs have their own particular method to recover it and several must perform calculations to get the final value. Using integers for this task is certainly easier than going through bit-field conversions. > > > > Nelio suggested adding a support function to convert the bitmap field > > to a speed value as an integer. I strongly support this, because you cannot > > expect the bitmap to be ordered by speed. > > > > > > Agree with Nelio&you. This is useful. > > > > It was exactly the extreme opposite, a function which takes a > > rte_eth_link to a bitmap i.e. speed_to_bm (rte_eth_link link) because, > > the speed is mostly used as an integer and not some kind of bitmap. > > > > > > This support function will be able to determine which speed is higher > > when exotic speeds are added to the bitmap. Please extend this conversion > > function to give three output parameters: speed, full/half duplex, auto > > negotiation/non-auto negotiation, or add two separate functions to get the > > duplex and auto-negotiation. > > > > > > Since, Full/Half duplex is for legacy 10/100Mbps only (afaik), I have my > > doubts on using a bit for all speeds. I would suggest to define (unroll) > > 100M (or 100M_FD) and 100M_HD, and the same 10Mbps/1gbps, as Thomas was > > suggesting some mails ago. > > > > > > This was done in v4 (implicitely 100M == 100M_FD). See below. > > > > > > MB> I didn't intend two bits to be allocated in the bitmap for all > > speeds to support full/half duplex, only for the relevant speeds. Since > > full duplex is dominant, I agree with the previous decision (originally > > suggested by Thomas, I think) to make full duplex implicit unless half > > duplex is explicitly specified. E.g. 10M_HD, 10M (alias 10M_FD), 100M_HD, > > 100M (alias 100M_FD), 1000M (or 1G), 2500M, 10G, 40G, 100G, etc. > > > > > > > > > > I haven't read the suggested code, but there should be some means in > > 2/ (advertisements) to disable auto negotiation, e.g. a single bit in the > > bitmap to indicate if the speed/duplex-indicating bits in the bitmap means > > forced speed/duplex (in which case only a single speed/duplex-bit should be > > set) or auto negotiation advertised speed/duplex (in which case multiple > > speed/duplex-bits can be set). > > > > > > Agree. > > > > > > v3/4 of this patch adds the bitmap in the advertised, as per discussed, > > to select a group of speeds This is not implemented by drivers yet (!). > > > > > > So, as of v4 of this patch, there could be: a) autoneg any supported > > speed (=> bitmap == 0) b) autoneg over group of speeds (=> more than one > > bit set in the bitmap) c) forced speed (one and only one set in the bitmap). > > > > > > I think this is precisely what you meant + b) as a bonus > > > > > > MB> This was not what I meant, but it wasn't very clearly written, so > > I'll try again: Add an additional single bit "NO_AUTONEG" (or whatever you > > want to name it) to the 2/ (advertisements) bitmap that explicitly turns > > off auto negotiation and tries to force the selected speed/duplex (i.e. > > only one other bit can be set in the bitmap when the NO_AUTONEG bit is > > set). Your c) makes it impossible to use auto negotiation to advertise a > > specific speed/duplex, e.g. 100M_FD. My suggested NO_AUTONEG bit can also > > be used in 3/ (result) to indicate that the speed was a result of Parallel > > Detection, i.e. that auto negotiation failed or was disabled in either end > > of the link. > > > > > > MB> However, I like your suggestion a). > > > > > > > > > > And some means in 3/ (result) and maybe 2/ (advertisements) to select > > and/or indicate physical interface in dual-personality ports (e.g. ports > > where the PHY has both an SFP and a RJ45 connector, but only one of the two > > can be used at any time). > > > > > > For rte_eth_link_get() I don't have such a strong opinion. You either > > > > > > * encode the link speed and duplex as of now, separating duplex and > > numeric speed. I would suggest to add the encoded speed+duplex bitmap flag > > for consistency (although redundant). > > > * or you return a single value, the bitmap with a single flag set of the > > unrolled speeds, and then have the helpers int rte_eth_speed_from_bm(int > > val_bm) and bool rte_eth_duplex_from_bm(int val_bm). > > > > > > MB> I prefer the latter of the two, only because it makes 3/ (result) > > consistent with 1/ (capabilities) and 2/ (advertisements). > > > > So I agree for 1/ capabilities and 2/ advertisements. > > > > But, I don't agree to modify rte_eth_link_get API > > (and rte_eth_link structure) thus 3/ result. > > We don't need a "consistent" result, we need something usable. This is > > not the case of the bitmap and using some conversion functions. > > Remember that the speed and duplex will not change until the next link > > down and there is a lot of code using speeds as integers. > > Your solution will just increase the number of instruction to get the > > same result, is that a benefit? > > > > So do you think we should care about roughly 10cycles (at very most) which > is what a unique switch case mapping will take? We are not talking about > the dataplane here, Neilo. He wasn't arguing about the number of CPU cycles but the amount of code required to go back and forth between actual link status speed and its bit-field counterpart for no apparent benefit. > The benefit that Morten is arguing here is to have a unified API, > consistent for the user. Once more, I don't have a preference, though I see > what he means. I am not sure if the bitmap for retrieving the link is > really more usable than the current API, which is IMHO what should steer > the discussion, not performance. Everyone agrees that a link speed bit-field is useful as an input value to advertise, request and allow a set of speeds. We do not agree with its usage as the current link speed which is often the result of a computation. We aren't talking about performance. A given link cannot be simultaneously at 10 Gbps and 1 Gbps right? Using a bit-field for the current link speed is confusing at best. Output values do not need to be included in the unified API, they are never converted back into enum values. I'm stressing again the fact that doing so would require a changes in all applications that use the current speed and in PMDs for no good reason. Regards, -- Adrien Mazarguil 6WIND