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 05543A328D for ; Tue, 22 Oct 2019 12:16:45 +0200 (CEST) Received: from [92.243.14.124] (localhost [127.0.0.1]) by dpdk.org (Postfix) with ESMTP id 8AE565B3E; Tue, 22 Oct 2019 12:16:44 +0200 (CEST) Received: from dispatch1-us1.ppe-hosted.com (dispatch1-us1.ppe-hosted.com [67.231.154.164]) by dpdk.org (Postfix) with ESMTP id C06DC4C74 for ; Tue, 22 Oct 2019 12:16:43 +0200 (CEST) X-Virus-Scanned: Proofpoint Essentials engine Received: from webmail.solarflare.com (uk.solarflare.com [193.34.186.16]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-SHA384 (256/256 bits)) (No client certificate requested) by mx1-us1.ppe-hosted.com (PPE Hosted ESMTP Server) with ESMTPS id 4A3F1B00064; Tue, 22 Oct 2019 10:16:41 +0000 (UTC) Received: from [192.168.38.17] (91.220.146.112) by ukex01.SolarFlarecom.com (10.17.10.4) with Microsoft SMTP Server (TLS) id 15.0.1395.4; Tue, 22 Oct 2019 11:16:35 +0100 To: Ferruh Yigit , Pavan Nikhilesh Bhagavatula , Jerin Jacob Kollanukkaran CC: "dev@dpdk.org" , Adrien Mazarguil , Thomas Monjalon , "Xiaolong Ye" , Bruce Richardson References: <79136309-b027-1031-5356-97f19dab7e58@intel.com> From: Andrew Rybchenko Message-ID: <3aade004-4150-d5d5-148d-7811a3f8d155@solarflare.com> Date: Tue, 22 Oct 2019 13:16:30 +0300 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101 Thunderbird/60.9.0 MIME-Version: 1.0 In-Reply-To: <79136309-b027-1031-5356-97f19dab7e58@intel.com> Content-Type: text/plain; charset="utf-8"; format=flowed Content-Transfer-Encoding: 8bit Content-Language: en-GB X-Originating-IP: [91.220.146.112] X-ClientProxiedBy: ocex03.SolarFlarecom.com (10.20.40.36) To ukex01.SolarFlarecom.com (10.17.10.4) X-TM-AS-Product-Ver: SMEX-12.5.0.1300-8.5.1010-24994.003 X-TM-AS-Result: No-15.730400-8.000000-10 X-TMASE-MatchedRID: yebcs53SkkD4ECMHJTM/ufZvT2zYoYOwt3aeg7g/usAutoY2UtFqGJWi xnDxF7NQr4VH42r0016YZfjORODtZTM9BBRuZZ1vboe6sMfg+k+hhKooH+/fy9EsTITobgNEZpV YVTTLTSb6DxBeqeXlOvxqRdyqeSp7ajCF67OIrSp0+657dxGJGA/o5bNHEsCTSMg2Oe/b8Ewr6D +mFr1pQDRFsYpZYpn/2mT5DR7XEdAkjSX/fu4YCGXaK3KHx/xpeouvej40T4iHv8otQeUIQuc9u TWZVtE9r2FpZtrxLRbzaflrNdESz0GpYFFdx/uzwS8qUbQKOMg5iooXtStiHiNGK7UC7ElMmzO5 24bhivFCK/zl5wNg7h/AeVs+z6k90ijRgGspNZV9Na8belcvhGfQgc3VUSnOERgtK1vKtV0Nih9 C4BdvuLnl71T889exkjVVIg7nRduf2/wegNnQUdbgzPjrV+wcDvc/j9oMIgVzE9D5svL/QBwEEI VsM/kpja3rP9/YOUv7zOAXcy31DIpAMGiqcHEM20204SCJw/qpvf+jmz45w7xgMf9QE2ebyC1Mv YqDoSoa0+wgjtulISgGY1pnOHdH1Y4IIoI+uMITF1LtYW9la4nAh89Hsnb001c7LSleBSEpXIOs 5jngRilMEKo38KD7qNscV9AN9zVdPa2QGGaXzub3p4cnIXGN6/ovm5YGTGmwzl+IrmkihbNuV9b qcyE32OjKLnafEv3KUguNaXKQsAaChs2LAIZjTauf2PrRb1s6En2bnefhoMLRRNKa+dDoW2g2r2 LI8m19AWQ5+s2aP024rDL/P7srrgOXLTAm6OCeAiCmPx4NwFkMvWAuahr8+gD2vYtOFhgqtq5d3 cxkNQg0+/2gzcJh5PyyMgRbRjIeY4btcjE3pxKzIdaq/OAMEPBgTeh/d58= X-TM-AS-User-Approved-Sender: Yes X-TM-AS-User-Blocked-Sender: No X-TMASE-Result: 10--15.730400-8.000000 X-TMASE-Version: SMEX-12.5.0.1300-8.5.1010-24994.003 X-MDID: 1571739402-bUqyzHk8qMQv Subject: Re: [dpdk-dev] [PATCH v12 0/7] ethdev: add new Rx offload flags 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" Hi, @Pavan, see question below. On 10/21/19 6:34 PM, Ferruh Yigit wrote: > On 10/21/2019 4:19 PM, Pavan Nikhilesh Bhagavatula wrote: >> Hi Ferruh, >> >>> -----Original Message----- >>> From: Ferruh Yigit >>> Sent: Monday, October 21, 2019 8:37 PM >>> To: Andrew Rybchenko ; Pavan Nikhilesh >>> Bhagavatula ; Jerin Jacob Kollanukkaran >>> >>> Cc: dev@dpdk.org; Adrien Mazarguil ; >>> Thomas Monjalon ; Xiaolong Ye >>> ; Bruce Richardson >>> >>> Subject: Re: [dpdk-dev] [PATCH v12 0/7] ethdev: add new Rx >>> offload flags >>> On 10/18/2019 11:31 AM, Andrew Rybchenko wrote: >>>> On 10/18/19 12:42 PM, Ferruh Yigit wrote: >>>>> On 10/18/2019 8:32 AM, Andrew Rybchenko wrote: >>>>>> Hi Ferruh, >>>>>> >>>>>> since I've reviewed I'll reply as I understand it. >>>>>> >>>>>> On 10/17/19 8:43 PM, Ferruh Yigit wrote: >>>>>>> On 10/17/2019 1:02 PM, pbhagavatula@marvell.com wrote: >>>>>>>> From: Pavan Nikhilesh >>>>>>>> >>>>>>>> Add new Rx offload flags `DEV_RX_OFFLOAD_RSS_HASH` and >>>>>>>> `DEV_RX_OFFLOAD_FLOW_MARK`. These flags can be used to >>>>>>>> enable/disable PMD writes to rte_mbuf fields `hash.rss` and >>> `hash.fdir.hi` >>>>>>>> and also `ol_flags:PKT_RX_RSS` and `ol_flags:PKT_RX_FDIR`. >>>>>>> Hi Pavan, >>>>>>> >>>>>>> Initially sorry for involving late, >>>>>>> >>>>>>> When we expose an interface to the applications, they will expect >>> those will be >>>>>>> respected by underlying PMDs. >>>>>>> As far as I can see drivers are updated to report new added Rx >>> offload flags as >>>>>>> supported capabilities but drivers are not using those flags at all, >>> so >>>>>>> application providing that flag won't really enable/disable >>> anything, I think >>>>>>> this is a problem and it is wrong to lie even for the PMDs J >>>>>> It is required to let applications know that the offload is supported. >>>>>> There are a number of cases when an offload cannot be disabled, >>>>>> but it does not mean that the offload must not be advertised. >>>>> Can't disable is something else, although I believe that is rare case, in >>> this >>>>> case driver can enable/disable the RSS and representing this as an >>> offload >>>>> capability. >>>> It is not enabling/disabling the RSS. It is enabling/disabling RSS hash >>>> delivery >>>> together with an mbuf. >>> >>> Got it, it is related to the RSS hash delivery. >>> >>>>> But when user want to configure this offload by setting or unsetting >>> in offload >>>>> config, driver just ignores it. >>>> When application enables offload, it says that it needs it and going to >>> use >>>> (required). When the offload is not enabled, application simply don't >>> care. >>>> So, if the information is still provided it does not harm. >>> >>> Not sure if there is no harm, a config option not respected by >>> underlying PMDs >>> silently is a problem I think. Some time ago the idea of offloads which cannot be disabled by PMD was discussed and, if I remember decision correctly, it was decided that it will overcomplicate it. It was discussed when new offload API is introduced. Logging is helpful sometimes, but it is not a panacea. >>>>>> If driver see benefits from disabling the offload (e.g. avoid delivery >>>>>> of RSS hash from NIC to host), it can do it after the patchset. >>>>> Yes but I think this patchset shouldn't ignore that disabling the >>> feature is not >>>>> implemented yet. If those PMDs that has been updated to report >>> the HASH >>>>> capability has RSS enabled by default, I suggest adding a check for >>> this offload >>>>> in PMD, >>>>> if it is requested to disable (which means not requested for enable), >>> print a >>>>> log saying disabling HASH is not supported and set this flag in the >>> offload >>>>> configuration to say PMD is configured to calculate the HASH. >>>>> Later PMD maintainers may prefer to replace that error log with >>> actual disable code. >>>> It is possible to do. Of course, it is better to provide real offload >>>> values on get, but >>>> eth_conf is const in rte_eth_dev_configure(), so, we can't change it >>> and >>>> it is good. >>>> So, the only way is rte_eth_rx_queue_info_get(). >>>> I guess there is a lot of space for the same improvement for other Rx >>>> offloads >>>> in various PMDs. >>> >>> We don't need the update 'eth_conf' parameter of the >>> 'rte_eth_dev_configure()', >>> that is what user requested, but config stored in 'dev->data->dev_conf' >>> which >>> can be updated. What for? As I understand data->dev_conf should not be used outside librte_ethdev and drivers. Basically it means that all patched drivers should be updated to log information message if the offload is not requested, but will be enabled anyway and updated data->dev_conf.rxmode.offloads. Please, confirm. >>>> Also I worry that it could be not that trivial to do in all effected PMDs. >>> >>> Yes it can be some work, and if this patchset doesn't do it, who will do >>> the work? Pavan, will you do or should I care about it if confirmed? >>>>>>> Specific to `DEV_RX_OFFLOAD_RSS_HASH`, we have already >>> some RSS config >>>>>>> structures and it is part of the 'rte_eth_dev_configure()' API, >>> won't it create >>>>>>> multiple way to do same thing? >>>>>> No, a new offload is responsible for RSS hash delivery from NIC to >>> host >>>>>> and fill in in mbuf returned to application on Rx. >>>>> What you have described is already happening without the new >>> offload flag and >>>>> this is my concern that we are duplicating it. >>>>> >>>>> >>>>> There is a 'struct rte_eth_rxmode' (under 'struct rte_eth_conf') >>>>> which has 'enum rte_eth_rx_mq_mode mq_mode;' >>>>> >>>>> If "mq_mode == ETH_MQ_RX_NONE" hash calculation is disabled, >>> and >>>>> 'mbuf::hash::rss' is not updated. >>>> No-no. It binds RSS distribution and hash delivery. What the new >>>> offload allows to achieve: I want Rx to spread traffic over many Rx >>>> queues, but I don't need RSS hash. >>> >>> I see, so RSS configuration will stay same, but driver needs to take care >>> the >>> new flags to decide to update or not the mbuf::rss::hash field. >>> >>> I don't know if disabling RSS but calculating hash is supported, if not >>> supported that case also should be checked by driver. Yes, it is interesting question what should happen if ETH_MQ_RX_NONE with DEV_RX_OFFLOAD_RSS_HASH. It sounds like rte_ethdev should check dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG is set when DEV_RX_OFFLOAD_RSS_HASH is requested. Otherwise the request is invalid. May be it could be relaxed in the future, but not now. >>>>> (Thanks Bruce to helping finding it out) >>>>> >>>>> >>>>>>> And for the `ol_flags:PKT_RX_RSS` flag, it was already used to >>> mark that >>>>>>> 'mbuf::hash::rss' is valid, right? Is there anything new related that >>> in the set? >>>>>> As I understand you mean, ol_flags::PKT_RX_RSS_HASH. >>>>>> Yes, the new offload allows say if application needs it or now. >>>>>> Basically it decouples RSS distribution and hash delivery. >>>>> Setting 'ol_flags::PKT_RX_RSS_HASH' and 'mbuf::hash::rss' already >>> there and not >>>>> changing. I just want to clarify since this is not clear in the commit log. >>>>> >>>>> Only addition is to add a new flag to control PMD to enable/disable >>> hash >>>>> calculation (which PMDs ignore in the patch ???) >>>> It is not calculation, but delivery of the value from HW to applications. >>> >>> OK >>> >>>> Yes, commit log may/should be improved.> >>>>>>> Specific to the `DEV_RX_OFFLOAD_FLOW_MARK` and >>> `RTE_FLOW_ACTION_FLAG`, they are >>>>>>> rte_flow actions, application can verify and later request these >>> actions via >>>>>>> rte_flow APIs. Why we are adding an additional RX_OFFLOAD flag >>> for them? >>>>>> The reason is basically the same as above. HW needs to know in >>> advance, >>>>>> if application is going to use flow marks and configure Rx queue to >>> enable >>>>>> the information delivery. >>>>> What you described is done via 'rte_flow_create()' API, application >>> will request >>>>> those actions via API and Rx queue will be configured accordingly, >>> this is more >>>>> dynamic approach. Why application need to set this additional >>> configuration flag? >>>> More dynamic approach is definitely better, but it is not always >>> possible. >>>> Some PMDs can't even change MTU dynamically or MTU changing >>> requires >>>> restart which is hardly really a dynamic change. Of course, it is >>>> unlikely that >>>> MTU is changed when traffic is running etc, but still possible. >>>> The information about necessity to support flow marks delivery may >>>> be required on Rx queue setup and cannot be changed dynamically >>> when >>>> Rx queue is running and application would like to add flow rule with >>> mark >>>> action. >>> It doesn't need to be changed dynamically, application can call >>> 'rte_flow_validate()' and learn if it can set this action or not. Perhaps I >>> am >>> missing something, when it is required to have this as configuration >>> option? >>> >>>>> And as above the new RX offload flags ignored by PMDs, hard to >>> understand what >>>>> is the intention here. >>>>> >>>>> >>>>> Above usage of flags feels like the intention is adding some capability >>>>> information for the PMDs more that adding new offload >>> configuration. >>>>> If so this is bigger/older problem, and instead of abusing the offload >>> flags we >>>>> can think of an API to present device capabilities, and move >>> features.ini >>>>> content to the API in long term. >>>> What I really like with these new offload flags for Rx hash and flow >>> mark is >>>> that it makes features which provide information in mbuf on Rx >>> consistent: >>>>  - want timestamp? => DEV_RX_OFFLOAD_TIMESTAMP >>>>  - want Rx checksum flags => DEV_RX_OFFLOAD_CHECKSUM >>>>  - want to strip VLAN? => DEV_RX_OFFLOAD_VLAN_STRIP >>>>  - want RSS hash? => DEV_RX_OFFLOAD_RSS_HASH >>>>  - want flow mark support? => DEV_RX_OFFLOAD_FLOW_MARK >>>> >>>> Also it perfectly fits dynamic mbuf fields and allows to make RSS hash >>>> and flow mark fields dynamic with the new offloads as controls >>> Agree RSS_HASH fits well, my main concern with the patchset is driver >>> implementations are missing and just ignored. >>> >> Ignoring driver implementation is intentional as it involves adding a branch >> in Rx fastpath function for all drivers and might have -ve effects on performance. > Yes it may affect performance. Also it may be too much driver specific > implementation. > > That is why I suggest, following: > For the drivers that claim this capability, > - For the case driver updates the mbuf::rss:hash > Check if this offload requested or not, if not print an error and set internal > config as this offload enabled I would not say it is an error. At maximum it is warning, but I would use just info log level. I think there is nothing critical there. > - For the case driver not updates the mbuf::rss:hash > Check if this offload requested or not, if requested print an error and set > internal config as this offload disabled If so, the offload is not advertised and generic checks in rte_ethdev catch it and return error. > Later PMD maintainers may prefer to replace those errors with actual > implementation if they want. [snip]