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 D81E5A04B6; Mon, 12 Oct 2020 01:02:55 +0200 (CEST) Received: from [92.243.14.124] (localhost [127.0.0.1]) by dpdk.org (Postfix) with ESMTP id 60AC51D598; Mon, 12 Oct 2020 01:02:54 +0200 (CEST) Received: from mga12.intel.com (mga12.intel.com [192.55.52.136]) by dpdk.org (Postfix) with ESMTP id 3437B1BFDD for ; Mon, 12 Oct 2020 01:02:53 +0200 (CEST) IronPort-SDR: oYQ0bRRH+QLnakG6B574BS3bmb44VYbrg26nZyv9lqNW0h33aLREhQeWgIO4kKK0rHe91dFBT4 1xvXEIOaKcDw== X-IronPort-AV: E=McAfee;i="6000,8403,9771"; a="144990987" X-IronPort-AV: E=Sophos;i="5.77,364,1596524400"; d="scan'208";a="144990987" X-Amp-Result: SKIPPED(no attachment in message) X-Amp-File-Uploaded: False Received: from fmsmga005.fm.intel.com ([10.253.24.32]) by fmsmga106.fm.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 11 Oct 2020 16:02:48 -0700 IronPort-SDR: 6Mdrr9GaDWqnCPneHxEI3Ed1yWX9jId5pmEeqrD0YCgdpcfJYCfpZGw5Por/8ppgQ2myKKYK36 Zo6zyzD3LdBw== X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.77,364,1596524400"; d="scan'208";a="520467026" Received: from orsmsx603.amr.corp.intel.com ([10.22.229.16]) by fmsmga005.fm.intel.com with ESMTP; 11 Oct 2020 16:02:47 -0700 Received: from orsmsx607.amr.corp.intel.com (10.22.229.20) by ORSMSX603.amr.corp.intel.com (10.22.229.16) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256) id 15.1.1713.5; Sun, 11 Oct 2020 16:02:47 -0700 Received: from orsmsx601.amr.corp.intel.com (10.22.229.14) by ORSMSX607.amr.corp.intel.com (10.22.229.20) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256) id 15.1.1713.5; Sun, 11 Oct 2020 16:02:47 -0700 Received: from ORSEDG602.ED.cps.intel.com (10.7.248.7) by orsmsx601.amr.corp.intel.com (10.22.229.14) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256) id 15.1.1713.5 via Frontend Transport; Sun, 11 Oct 2020 16:02:47 -0700 Received: from NAM10-DM6-obe.outbound.protection.outlook.com (104.47.58.103) by edgegateway.intel.com (134.134.137.103) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384) id 15.1.1713.5; Sun, 11 Oct 2020 16:02:46 -0700 ARC-Seal: i=1; a=rsa-sha256; s=arcselector9901; d=microsoft.com; cv=none; b=QSoJ+EGHyjl7VTQc1VuSccdjjzYx4/6HBYZJNqZNPfOHb99qaMkQ+yDOuyJXpiR6PY1SaeUxjXJtTVrRWzmW8Q9WUEFOoQzDCz9m8R+bC5S7VWtVXUaIk2fdsGv1IVe22NT+22u+tnF0g3XzPisKaw3p8lfAKLWX81AOpwQ/MWdrTs1lgrDoadPUbCbQ+ocUei3oVNZVB1f4xq7diA2wdLafPpR7gScMMyduU7QYV+pwK6lD3dH6SpH/5A5rjl1RgMIVeWUeX4SYj14s9DYzmCmoDhKmnlAcYUIDu5jdr0/+oVIhuCJojor/kM7Vpel3fICuhPFmwKvHLb6XLcw1Ow== ARC-Message-Signature: i=1; a=rsa-sha256; c=relaxed/relaxed; d=microsoft.com; s=arcselector9901; h=From:Date:Subject:Message-ID:Content-Type:MIME-Version:X-MS-Exchange-SenderADCheck; bh=rgWg4V/qi33Fiww1MN2aw/vVHeJUkENsDKTxmJX206g=; b=AWSH8X1AQuq3YOKUVTwOvXBwHYhamWEK2kqRxULm7W0Z/mr2CMXkE82Ml9zQtnyBwvZCvvBqkAL1a3EdNSkf6vH8oieoxkunEmFwjdWU4OTR1ePFOKXEqMt5pwpugGbsc9t75rz6eS0xGDrgKj+yzjpaAz1p6Nn3+79850EZ9x2OjjhTMeQ0t3kQiJpe7zGMe5DbZoYjmVdkSs/xLqqUiVAZZsvSkBOThGxHkqr0ped/PiI846EwTb7ZXJTq6M5WALpCw5fH8isA30/GDv3l4bkH29fo/plOher4GMIKuSz7Py+breKJlKKa9h3NxcEh7dJt03hSiJMWUw0xUbYYLQ== ARC-Authentication-Results: i=1; mx.microsoft.com 1; spf=pass smtp.mailfrom=intel.com; dmarc=pass action=none header.from=intel.com; dkim=pass header.d=intel.com; arc=none DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=intel.onmicrosoft.com; s=selector2-intel-onmicrosoft-com; h=From:Date:Subject:Message-ID:Content-Type:MIME-Version:X-MS-Exchange-SenderADCheck; bh=rgWg4V/qi33Fiww1MN2aw/vVHeJUkENsDKTxmJX206g=; b=HsoTalPhdoFCLMAgudEil71U/kXGhu3rNAoTfxRnSyrH/BpIM3IKGaSWiC+GgwIia8gZdwjxjuI//E/sY9klbLnvcQpzVubPFmViCUMQxUpLV3EOSlZAjtBvL7DbuWpUnKgbeUyv97bgbaiGe1wK3U0RLcs7fgEz84xGnIoK+EQ= Received: from BYAPR11MB3301.namprd11.prod.outlook.com (2603:10b6:a03:7f::26) by BY5PR11MB4307.namprd11.prod.outlook.com (2603:10b6:a03:1bd::27) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384) id 15.20.3455.25; Sun, 11 Oct 2020 23:02:43 +0000 Received: from BYAPR11MB3301.namprd11.prod.outlook.com ([fe80::f5a4:3f6b:ade3:296b]) by BYAPR11MB3301.namprd11.prod.outlook.com ([fe80::f5a4:3f6b:ade3:296b%3]) with mapi id 15.20.3455.029; Sun, 11 Oct 2020 23:02:43 +0000 From: "Ananyev, Konstantin" To: "Gujjar, Abhinandan S" , Honnappa Nagarahalli , "dev@dpdk.org" , "Doherty, Declan" CC: "jerinj@marvell.com" , "Akhil.goyal@nxp.com" , "Vangati, Narender" , nd , nd Thread-Topic: [dpdk-dev] [v2 1/2] cryptodev: support enqueue callback functions Thread-Index: AQHWhsTbY+pAzzKww0euN2orABmUValrRWzAgAAl9gCAAW0pMIAGJnUAgAAH/lCAAp3uAIAAgn/ggADASgCAASzNQIAHHMUAgAA/FzCADrUggIABX7xwgAAmGoCAA4qBcA== Date: Sun, 11 Oct 2020 23:02:42 +0000 Message-ID: References: <1599549024-195051-1-git-send-email-abhinandan.gujjar@intel.com> In-Reply-To: Accept-Language: en-GB, en-US Content-Language: en-US X-MS-Has-Attach: X-MS-TNEF-Correlator: dlp-product: dlpe-windows dlp-reaction: no-action dlp-version: 11.5.1.3 authentication-results: intel.com; dkim=none (message not signed) header.d=none;intel.com; dmarc=none action=none header.from=intel.com; x-originating-ip: [46.7.39.127] x-ms-publictraffictype: Email x-ms-office365-filtering-correlation-id: a904bf17-00b8-4b85-2390-08d86e39c2ac x-ms-traffictypediagnostic: BY5PR11MB4307: x-ld-processed: 46c98d88-e344-4ed4-8496-4ed7712e255d,ExtAddr x-ms-exchange-transport-forked: True x-microsoft-antispam-prvs: x-ms-oob-tlc-oobclassifiers: OLM:10000; x-ms-exchange-senderadcheck: 1 x-microsoft-antispam: BCL:0; x-microsoft-antispam-message-info: N1AmWsVlYA+pj5s63+dxMSXnbB0l734zlK5DbBUyj4Z8mmfnD1NIY+hbxjKAfKGzx2x7iy2cmVbvxTMpxxAN+RyvFz6Rwstlym6kUQ1vxSAkfo6gEuoDFdj3P6tUePyU70fZIdEZ1Sr+xheZrGp6dPOMd9eEvvlv+wDTBj5tqH4z3lk+A6NbJLFjx1UDRlOt3195nZd0Q/62JGYzVRe7vd4PuixZTj0Ixu9C/fxB/SCnDynCC1Lr6TsRs+e8LF8Y8KVzKND073oKTd4bLm81Zh4sl8pRZDz1n1a5sen6OPs8qBTjkXDmYbY5RB8YcDrYOn6Cf2r49dtbI+NRv38jACx1UzCEJvi68ceO9q7vzWzCVyycVwfo7er7/BiZmANvlSWHxZz0NO4pcodlSZQ2hA== x-forefront-antispam-report: CIP:255.255.255.255; CTRY:; LANG:en; SCL:1; SRV:; IPV:NLI; SFV:NSPM; H:BYAPR11MB3301.namprd11.prod.outlook.com; PTR:; CAT:NONE; SFS:(4636009)(136003)(396003)(39860400002)(346002)(376002)(366004)(83080400001)(478600001)(7696005)(30864003)(83380400001)(54906003)(966005)(6636002)(33656002)(110136005)(316002)(66556008)(64756008)(66446008)(66476007)(86362001)(4326008)(6506007)(53546011)(71200400001)(8936002)(186003)(5660300002)(52536014)(55016002)(2906002)(66946007)(26005)(9686003)(76116006)(8676002)(579004); DIR:OUT; SFP:1102; x-ms-exchange-antispam-messagedata: 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 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: quoted-printable MIME-Version: 1.0 X-MS-Exchange-CrossTenant-AuthAs: Internal X-MS-Exchange-CrossTenant-AuthSource: BYAPR11MB3301.namprd11.prod.outlook.com X-MS-Exchange-CrossTenant-Network-Message-Id: a904bf17-00b8-4b85-2390-08d86e39c2ac X-MS-Exchange-CrossTenant-originalarrivaltime: 11 Oct 2020 23:02:43.0088 (UTC) X-MS-Exchange-CrossTenant-fromentityheader: Hosted X-MS-Exchange-CrossTenant-id: 46c98d88-e344-4ed4-8496-4ed7712e255d X-MS-Exchange-CrossTenant-mailboxtype: HOSTED X-MS-Exchange-CrossTenant-userprincipalname: dU4MRBRGxbRwIqvFlG0fSIpAt5QgGDV1rxlyPFBEAXSTgObVQXA5wrADck1C/bSqYjEaApo7Bff3E1ZAgF7wV9hRM2TdAW6tznKwTj4l118= X-MS-Exchange-Transport-CrossTenantHeadersStamped: BY5PR11MB4307 X-OriginatorOrg: intel.com Subject: Re: [dpdk-dev] [v2 1/2] cryptodev: support enqueue callback functions 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 Abhinandan, >=20 > Hi Konstantin, >=20 >=20 > > -----Original Message----- > > From: Ananyev, Konstantin > > Sent: Friday, October 9, 2020 8:10 PM > > To: Gujjar, Abhinandan S ; Honnappa > > Nagarahalli ; dev@dpdk.org; Doherty, > > Declan > > Cc: jerinj@marvell.com; Akhil.goyal@nxp.com; Vangati, Narender > > ; nd ; nd > > Subject: RE: [dpdk-dev] [v2 1/2] cryptodev: support enqueue callback fu= nctions > > > > Hi Abhinandan, > > > > > > > > Hi Konstantin & Honnappa, > > > > > > Thanks for all the inputs and feedback. > > > > > > @Ananyev, Konstantin, > > > I have measured the perf with and without callback on xeon. Here are = the > > numbers: > > > > > > ./app/dpdk-test-crypto-perf -l 6-7 > > > --vdev=3D"crypto_openssl0,socket_id=3D0,max_nb_sessions=3D128" -- --p= test > > > throughput --devtype crypto_openssl --optype cipher-then-auth > > > --cipher-algo aes-cbc --cipher-op encrypt --cipher-key-sz 16 > > > --auth-algo sha1-hmac --auth-op generate --auth-key-sz 64 --digest-sz > > > 12 --total-ops 10000000 --burst-sz 32 --buffer-sz 64 > > > > > > With callback(+ RCU - totally opaque to data-plane threads) > > > lcore id Buf Size Burst Size Enqueued Dequeued Failed = Enq Failed > > Deq MOps Gbps Cycles/Buf > > > 7 64 32 10000000 10000= 000 0 0 > > 0.8129 0.4162 2694.09 > > > 7 64 32 10000000 10000= 000 0 0 > > 0.8152 0.4174 2686.31 > > > 7 64 32 10000000 10000= 000 0 0 > > 0.8198 0.4197 2671.48 > > > > > > Without callback: > > > lcore id Buf Size Burst Size Enqueued Dequeued Failed = Enq Failed > > Deq MOps Gbps Cycles/Buf > > > > > > 7 64 32 10000000 100000= 00 0 0 > > 0.8234 0.4216 2659.81 > > > 7 64 32 10000000 100000= 00 0 0 > > 0.8247 0.4222 2655.63 > > > 7 64 32 10000000 100000= 00 0 0 > > 0.8123 0.4159 2695.90 > > > > > > Just to cofirm: > > You implemented crypto enqueue callbacks using RCU QSBR online /offline= (as > > suggested below) and numbers above are for: > > 1) callback code in place and some dummy callback installed > That's right. (+ RCU calling online + offline APIs inside rte_cryptodev_e= nqueue_burst()) > > 2) callback code in place but no callbacks installed > No callback code. i.e. Original code. Ok, and if I get things right - difference between mean values is ~15 cycle= s? That's seems like very good result to me. Can I suggest to run one more test for your new callback code in place, but no actual callbacks installed? Thanks Konstantin > > > > Is my understanding correct here? > > Thanks > > Konstantin > > > > > > > > > > Regards > > > Abhinandan > > > > > > > -----Original Message----- > > > > From: Ananyev, Konstantin > > > > Sent: Tuesday, September 29, 2020 2:33 PM > > > > To: Honnappa Nagarahalli ; Gujjar, > > > > Abhinandan S ; dev@dpdk.org; Doherty, > > > > Declan > > > > Cc: jerinj@marvell.com; Akhil.goyal@nxp.com; Vangati, Narender > > > > ; nd ; nd > > > > Subject: RE: [dpdk-dev] [v2 1/2] cryptodev: support enqueue callbac= k > > > > functions > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > +#ifdef RTE_CRYPTODEV_CALLBACKS int > > > > > > > > > > > > > > > +rte_cryptodev_rcu_qsbr_add(uint8_t dev_id, > > > > > > > > > > > > > > > +struct rte_rcu_qsbr > > > > > > > > > > > > > > > +*qsbr) { > > > > > > > > > > > > > > > + > > > > > > > > > > > > > > > + struct rte_cryptodev *dev; > > > > > > > > > > > > > > > + > > > > > > > > > > > > > > > + if (!rte_cryptodev_pmd_is_valid_dev(dev_id)= ) { > > > > > > > > > > > > > > > + CDEV_LOG_ERR("Invalid dev_id=3D%" > > PRIu8, > > > > > > dev_id); > > > > > > > > > > > > > > > + return -EINVAL; > > > > > > > > > > > > > > > + } > > > > > > > > > > > > > > > + > > > > > > > > > > > > > > > + dev =3D &rte_crypto_devices[dev_id]; > > > > > > > > > > > > > > > + dev->qsbr =3D qsbr; > > > > > > > > > > > > > > > + return 0; > > > > > > > > > > > > > > > +} > > > > > > > > > > > > > > > > > > > > > > > > > > > > So if I understand your patch correctly you > > > > > > > > > > > > > > propose a new working model for > > > > > > > > > > > > > > crypto-devs: > > > > > > > > > > > > > > 1. Control-plane has to allocate/setup rcu_qsbr > > > > > > > > > > > > > > and do rte_cryptodev_rcu_qsbr_add(). > > > > > > > > > > > > > > 2. Data-plane has somehow to obtain pointer to > > > > > > > > > > > > > > that rcu_qsbr and wrap > > > > > > > > > > > > > > cryptodev_enqueue() > > > > > > > > > > > > > > with rcu_qsbr_quiescent() or > > > > > > > > rcu_qsbr_online()/rcu_qsbr_offline(). > > > > > > > > > > > > > Yes. I think, it is not a new model. It is same a= s > > > > > > > > > > > > > RCU integration with > > > > > > > > > > LPM. > > > > > > > > > > > > > Please refer: > > > > > > > > > > > > > https://patches.dpdk.org/cover/73673/ > > > > > > > > > > > > > > > > > > > > > > > > I am talking about new working model for crypto-dev > > > > > > > > enqueue/dequeue. > > > > > > > > > > > > As I said above now it becomes data-plane thread > > > > > > > > > > > > responsibility > > > > to: > > > > > > > > > > > > -somehow to obtain pointer to that rcu_qsbr for > > > > > > > > > > > > each cryptodev it is > > > > > > > > > > using. > > > > > > > > > > > > -call rcu sync functions (quiescent/online/offline= ) > > > > > > > > > > > > on a regular > > > > > > basis. > > > > > > > > > > > It is not on regular basis. When data plane comes up, > > > > > > > > > > > they report > > > > > > online. > > > > > > > > > > > They report quiescent when they are done with critica= l > > > > > > > > > > > section or shared > > > > > > > > > > structure. > > > > > > > > > > > > > > > > > > > > I understand that, but it means all existing apps have > > > > > > > > > > to be changed that > > > > > > > > way. > > > > > > > > > > > > > > > > > > > > > All though, there is some dataplane changes involved > > > > > > > > > > > here, I don't think, it > > > > > > > > > > is major. > > > > > > > > > > > > > > > > > > > > I still think our goal here should be to make no visibl= e > > > > > > > > > > changes to the dataplane. > > > > > > > > > > I.E. all necessary data-plane changes need to be hidden > > > > > > > > > > inside CB invocation part. > > > > > > > > > Please note that this is being implemented using the > > > > > > > > > memory reclamation framework documented at > > > > > > > > > https://doc.dpdk.org/guides/prog_guide/rcu_lib.html#resou= r > > > > > > > > > ce-r > > > > > > > > > ecla > > > > > > > > > mati > > > > > > > > > on-framework-for-dpdk > > > > > > > > > > > > > > > > > > While using RCU there are couple of trade-offs that > > > > > > > > > applications have to > > > > > > > > consider: > > > > > > > > > 1) Performance - reporting the quiescent state too often > > > > > > > > > results in performance impact on data plane > > > > > > > > > 2) Amount of outstanding memory to reclaim - reporting > > > > > > > > > less often results in more outstanding memory to reclaim > > > > > > > > > > > > > > > > > > Hence, the quiescent state reporting is left to the appli= cation. > > > > > > > > > The application decides how often it reports the quiescen= t > > > > > > > > > state and has control > > > > > > > > over the data plane performance and the outstanding memory > > > > > > > > to > > > > reclaim. > > > > > > > > > > > > > > > > > > When you say "new working model for crypto-dev > > > > > > > > > enqueue/dequeue", > > > > > > > > > > > > > > > > > > 1) are you comparing these with existing crypto-dev > > > > > > > > > enqueue/dequeue > > > > > > > > APIs? If yes, these are new APIs, it is not breaking anythi= ng. > > > > > > > > > 2) are you comparing these with existing call back > > > > > > > > > functions in ethdev enqueue/dequeue APIs? If yes, agree > > > > > > > > > that this is a new model. But, it is > > > > > > > > possible to support what ethdev supports along with the RCU > > > > > > > > method used in this patch. > > > > > > > > > > > > > > > > What I am talking about: > > > > > > > > > > > > > > > > Existing cryptodev enqueue/dequeue model doesn't require fo= r > > > > > > > > the user to manage any RCU QSBR state manually. > > > > > > > > I believe that addition of ability to add/remove > > > > > > > > enqueue/dequeue callbacks shouldn't change existing working > > model. > > > > > > > > I think that adding/removing such callbacks has to be opaqu= e > > > > > > > > to the user DP code and shouldn't require user to change it= . > > > > > > > > Same as we have now for ethdev callback implementation. > > > > > > > The ethdev callback implementation conveniently leaves the > > > > > > > problem of > > > > > > freeing memory to the user to resolve, it does not handle the i= ssue. > > > > > > > Hence, it "looks" to be opaque to the DP code. However, if th= e > > > > > > > application has to implement a safe way to free the call back > > > > > > > memory, its > > > > > > DP is affected based on call backs are being used or not. > > > > > > > > > > > > Yes, I think that's big drawback in initial ethdev callback > > > > > > implementation - it simply ignores DP/CP sync problem completel= y. > > > > > > Though I think it is possible to have both here: > > > > > > keep callback "opaque" to DP code and provide some sync > > > > > > mechanism between DP/CP. > > > > > > Hopefully one day we can fix ethdev callbacks too. > > > > > The solution we develop can be used in ethdev too. > > > > > > > > > > > > > > > > > > > I think that forcing DP code to be aware that callbacks are > > > > > > > > present or not and to modify its behaviour depending on tha= t > > > > > > > > nearly voids the purpose of having callbacks at all. > > > > > > > > In that case DP can just invoke callback function directly > > > > > > > > from it's > > > > > > codepath . > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Note that now data-plane thread would have to do > > > > > > > > > > > > that always > > > > > > > > > > > > - even if there are now callbacks installed for tha= t > > > > > > > > > > > > cryptodev queue > > > > > > > > right now. > > > > > > > > > > > > All that changes behaviour of existing apps and I > > > > > > > > > > > > presume would reduce adoption of that fature. > > > > > > > > > If I understand this correct, you are talking about a cas= e > > > > > > > > > where in the application might be > > > > > > > > > registering/unregistering multiple times during its > > > > > > > > > lifetime. In this case, yes, the application might be > > > > > > > > > reporting the > > > > > > > > quiescent state even when it has not registered the call ba= cks. > > > > > > > > But, it has the flexibility to not report it if it implemen= ts additional > > logic. > > > > > > > > > Note that we are assuming that the application has to > > > > > > > > > report quiescent state only for using callback functions. > > > > > > > > > Most probably the application has > > > > > > > > other requirements to use RCU. > > > > > > > > > Why not support what is done for ethdev call back > > > > > > > > > functions along with > > > > > > > > providing RCU method? > > > > > > > > > > > > > > > > > > > > There is always trade off involved! > > > > > > > > > > > In the previous patch, you suggested that some lazy > > > > > > > > > > > app may not free up the memory allocated by add cb. > > > > > > > > > > > For such apps, this patch has sync mechanism with som= e > > > > > > > > > > > additional cost of CP & DP > > > > > > changes. > > > > > > > > > > > > > > > > > > > > Sigh, it is not about laziness of the app. > > > > > > > > > > The problem with current ethedev cb mechanism and yours > > > > > > > > > > V1 (which was just a clone of it) - CP doesn't know whe= n > > > > > > > > > > it is safe after CB removal to free related memory. > > > > > > > > > > > > > > > > > > > > > > I still think all this callback mechanism should be > > > > > > > > > > > > totally opaque to data-plane threads - user > > > > > > > > > > > > shouldn't change his app code depending on would > > > > > > > > > > > > some enqueue/dequeue callbacks be > > > > > > > > installed or not. > > > > > > > > > > > I am not sure, how that can be implemented with > > > > > > > > > > > existing RCU > > > > > > design. > > > > > > > > > > > > > > > > > > > > As I said below the simplest way - with calling rcu > > > > > > > > > > onine/offline inside CB invocation block. > > > > > > > > > > That's why I asked you - did you try that approach and > > > > > > > > > > what is the perf numbers? > > > > > > > > > > I presume with no callbacks installed the perf change > > > > > > > > > > should be nearly > > > > > > > > zero. > > > > > > > > > > > > > > > > > > > > > @Honnappa Nagarahalli, Do you have any suggestions? > > > > > > > > > Reporting quiescent state in the call back functions has > > > > > > > > > several > > > > > > > > disadvantages: > > > > > > > > > 1) it will have performance impacts and the impacts will > > > > > > > > > increase as the > > > > > > > > number of data plane threads increase. > > > > > > > > > 2) It will require additional configuration parameters to > > > > > > > > > control how often the quiescent state is reported to > > > > > > > > > control the performance > > > > > > impact. > > > > > > > > > 3) Does not take advantage of the fact that most probably > > > > > > > > > the application is using RCU already > > > > > > > > > 4) There are few difficulties as well, please see below. > > > > > > > > > > > > > > > > I suggested Abhinandan to use RCU library because it is > > > > > > > > already there, and I thought it would be good not to re-imp= lement > > the wheel. > > > > > > > > Though if you feel librte_rcu doesn't match that task - > > > > > > > > fine, let's do it without librte_rcu. > > > > > > > > After all, what we need here - just an atomic ref count per > > > > > > > > queue that we are going to increment at entering and leavin= g > > > > > > > > list of callbacks inside enqueue/dequeue. > > > > > > > Ok, looks like I missed the point that a queue is used by a > > > > > > > single data plane > > > > > > thread. > > > > > > > Along with ref count increment you need the memory orderings > > > > > > > to avoid > > > > > > race conditions. These will be the same ones used in RCU. > > > > > > > On the control plane, you need to read this counter and poll > > > > > > > for the > > > > > > counter updates. All this is same cost as in RCU. > > > > > > > > > > > > Agree. > > > > > > > > > > > > > To control the cost, you > > > > > > > will have to control the rate of quiescent state reporting an= d > > > > > > > might have to > > > > > > expose this as a configuration parameter. > > > > > > > > > > > > > > The other important information you have to consider is if th= e > > > > > > > thread is making any blocking calls, which may be in some > > > > > > > other library. The thread is supposed to call > > > > > > > rcu_qsbr_thread_offline API before calling a > > > > > > blocking call. This allows the RCU to know that this particular > > > > > > thread will not report quiescent state. The cryptodev library > > > > > > might not have > > > > that information. > > > > > > > > > > > > > > If you want to go ahead with this design, you can still use > > > > > > > RCU with single thread configuration (like you have mentioned > > > > > > > below) and hide the > > > > > > details from the application. > > > > > > > > > > > > Yes, same thought here - use rcu_qsbr online/offline for DP > > > > > > part and hide actual sync details inside callback code. > > > > > We can give it a try. If we can have the performance numbers, we > > > > > can decide about how to control how often these APIs are called o= n > > > > > the data > > > > plane. > > > > > > > > To avoid misunderstanding: I am talking about calling online/offlin= e > > > > with every > > > > cryptodev_enqueue() traversal over CB list. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > That seems quite a big change and I don't think > > > > > > > > > > > > > > it is acceptable for most users. > > > > > > > > > > > > > > From my perspective adding/installing call-back= s > > > > > > > > > > > > > > to the dev has to be opaque to the data-plane c= ode. > > > > > > > > > > > > > > Also note that different callbacks can be > > > > > > > > > > > > > > installed by different entities (libs) and migh= t > > > > > > > > > > > > > > have no idea about each > > > > > > other. > > > > > > > > > > > > > > That's why I thought it would be better to make > > > > > > > > > > > > > > all this RCU stuff internal inside cryptodev: > > > > > > > > > > > > > > hide all this rcu_qsbr allocation/setup > > > > > > > > > > > > > > inside cryptod somehow to > > > > > > > > > > > > obtain pointer to that rcu_qsbr ev init/queue setup > > > > > > > > > > > > > > invoke rcu_qsbr_online()/rcu_qsbr_offline() > > > > > > > > > > > > > > inside > > > > > > > > > > > > cryptodev_enqueue(). > > > > > > > > > This will bring in the application related information > > > > > > > > > such as the thread ID > > > > > > > > into the library. > > > > > > > > > > > > > > > > I don't think it would. > > > > > > > > Cryptodev enqueue/dequeue functions are not supposed to be > > > > > > > > thread safe (same as rx/tx burst). > > > > > > > > So we can always use RCU with just one thread(thread_id =3D= 0). > > > > > > > Agree, the memory that needs to be freed is accessed by a > > > > > > > single thread > > > > > > on the data plane. RCU with one thread would suffice. > > > > > > > > > > > > > > > But as I said above - if you feel RCU lib is an overhead > > > > > > > > here, that's fine - I think it would be easy enough to do w= ithout > > librte_rcu. > > > > > > > > > > > > > > > > > If the same API calls are being made from multiple data > > > > > > > > > plane threads, you need a way to configure that > > > > > > > > > information to the library. So, it is better to leave > > > > > > > > > those details for the application to > > > > handle. > > > > > > > > > > > > > > > > > > > > > > I have already tried exploring above stuffs. Ther= e > > > > > > > > > > > > > are too many > > > > > > > > > > constraints. > > > > > > > > > > > > > The changes don't fit in, as per RCU design. > > > > > > > > > > > > > > > > > > > > > > > > Hmm could you be more specific here - what > > > > > > > > > > > > constraints are you referring to? > > > > > > > > > > > > > > > > > > > > > > > > > Moreover, having rcu api under enqueue_burst() > > > > > > > > > > > > > will affect the > > > > > > > > > > > > performance. > > > > > > > > > > > > > > > > > > > > > > > > It most likely will. Though my expectation it will > > > > > > > > > > > > affect performance only when some callbacks are > > > > > > > > > > > > installed. My thought > > > > > > > > here: > > > > > > > > > > > > callback function by itself will affect > > > > > > > > > > > > cryptdev_enqueue performance anyway, > > > > > > > > > > > With existing callback design, I have measured the > > > > > > > > > > > performance(with > > > > > > > > > > crypto perf test) on xeon. > > > > > > > > > > > It was almost negligible and same was shared with Dec= lan. > > > > > > > > > > > > > > > > > > > > I am asking about different thing: did you try alternat= e > > > > > > > > > > approach I described, that wouldn't require changes in > > > > > > > > > > the user data- > > > > > > plane code. > > > > > > > > > > > > > > > > > > > > > That is one of the reasons, I didn't want to add to > > > > > > > > > > > many stuffs in to the > > > > > > > > > > callback. > > > > > > > > > > > The best part of existing design is crypto lib is not= much > > modified. > > > > > > > > > > > The changes are either pushed to CP or DP. > > > > > > > > > > > > > > > > > > > > > > so adding extra overhead for sync is probably ok here= . > > > > > > > > > > > > > > > > > > > > I think that extra overhead when callbacks are present > > > > > > > > > > is expected and probably acceptable. > > > > > > > > > > Changes in the upper-layer data-plane code - probably n= ot. > > > > > > > > > > > > > > > > > > > > > > Though for situation when no callbacks are installe= d > > > > > > > > > > > > - perfomance should be left unaffected (or impact > > > > > > > > > > > > should be as small > > > > > > > > as possible). > > > > > > > > > > > > > > > > > > > > > > > > > The changes are more on control plane side, which > > > > > > > > > > > > > is one > > > > time. > > > > > > > > > > > > > The data plane changes are minimal. > > > > > > > > > > > > > > > > > > > > > > > > I still think upper layer data-plane code should > > > > > > > > > > > > stay unaffected (zero changes). > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >