From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mga07.intel.com (mga07.intel.com [134.134.136.100]) by dpdk.org (Postfix) with ESMTP id D34522BA2 for ; Sat, 27 May 2017 18:51:08 +0200 (CEST) Received: from fmsmga005.fm.intel.com ([10.253.24.32]) by orsmga105.jf.intel.com with ESMTP; 27 May 2017 09:51:07 -0700 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.38,404,1491289200"; d="scan'208";a="107364588" Received: from irsmsx106.ger.corp.intel.com ([163.33.3.31]) by fmsmga005.fm.intel.com with ESMTP; 27 May 2017 09:51:06 -0700 Received: from irsmsx109.ger.corp.intel.com ([169.254.13.250]) by IRSMSX106.ger.corp.intel.com ([169.254.8.236]) with mapi id 14.03.0319.002; Sat, 27 May 2017 17:51:05 +0100 From: "Ananyev, Konstantin" To: "Hu, Jiayu" CC: "dev@dpdk.org" , "Wiles, Keith" , "yuanhan.liu@linux.intel.com" Thread-Topic: [PATCH v3 1/3] lib: add Generic Receive Offload API framework Thread-Index: AQHSvNIioTj7qilZNEunAvoVQTYKUKIAL6+AgAGlgACAAN/38IADoyeAgAEVmeCAAD37gIAAjg7QgAAhXoCAADsncA== Date: Sat, 27 May 2017 16:51:04 +0000 Message-ID: <2601191342CEEE43887BDE71AB9772583FB02195@IRSMSX109.ger.corp.intel.com> References: <1491309106-94264-1-git-send-email-jiayu.hu@intel.com> <1493021398-115955-1-git-send-email-jiayu.hu@intel.com> <1493021398-115955-2-git-send-email-jiayu.hu@intel.com> <2601191342CEEE43887BDE71AB9772583FAF8C27@IRSMSX109.ger.corp.intel.com> <20170523103154.GA2033@localhost.localdomain> <2601191342CEEE43887BDE71AB9772583FAF9BAE@IRSMSX109.ger.corp.intel.com> <20170526072613.GA85810@localhost.localdomain> <2601191342CEEE43887BDE71AB9772583FAFB22A@IRSMSX109.ger.corp.intel.com> <20170527034137.GA22570@localhost.localdomain> <2601191342CEEE43887BDE71AB9772583FAFE69E@IRSMSX109.ger.corp.intel.com> <20170527140929.GA27976@localhost.localdomain> In-Reply-To: <20170527140929.GA27976@localhost.localdomain> Accept-Language: en-IE, en-US Content-Language: en-US X-MS-Has-Attach: X-MS-TNEF-Correlator: dlp-product: dlpe-windows dlp-version: 10.0.102.7 dlp-reaction: no-action x-originating-ip: [163.33.239.182] Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: quoted-printable MIME-Version: 1.0 Subject: Re: [dpdk-dev] [PATCH v3 1/3] lib: add Generic Receive Offload API framework 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: , X-List-Received-Date: Sat, 27 May 2017 16:51:09 -0000 Hi Jiayu, >=20 > Hi Konstantin, >=20 > On Sat, May 27, 2017 at 07:12:16PM +0800, Ananyev, Konstantin wrote: > > > > > > > -----Original Message----- > > > From: Hu, Jiayu > > > Sent: Saturday, May 27, 2017 4:42 AM > > > To: Ananyev, Konstantin > > > Cc: dev@dpdk.org; Wiles, Keith ; yuanhan.liu@l= inux.intel.com > > > Subject: Re: [PATCH v3 1/3] lib: add Generic Receive Offload API fram= ework > > > > > > On Sat, May 27, 2017 at 07:10:21AM +0800, Ananyev, Konstantin wrote: > > > > Hi Jiayu, > > > > > > > > > -----Original Message----- > > > > > From: Hu, Jiayu > > > > > Sent: Friday, May 26, 2017 8:26 AM > > > > > To: Ananyev, Konstantin > > > > > Cc: dev@dpdk.org; Wiles, Keith ; yuanhan.l= iu@linux.intel.com > > > > > Subject: Re: [PATCH v3 1/3] lib: add Generic Receive Offload API = framework > > > > > > > > > > Hi Konstantin, > > > > > > > > > > On Wed, May 24, 2017 at 08:38:25PM +0800, Ananyev, Konstantin wro= te: > > > > > > > > > > > > Hi Jiayu, > > > > > > > > > > > > > > > > > > > > Hi Konstantin, > > > > > > > > > > > > > > Thanks for your comments. My replies/questions are below. > > > > > > > > > > > > > > BRs, > > > > > > > Jiayu > > > > > > > > > > > > > > On Mon, May 22, 2017 at 05:19:19PM +0800, Ananyev, Konstantin= wrote: > > > > > > > > Hi Jiayu, > > > > > > > > My comments/questions below. > > > > > > > > Konstantin > > > > > > > > > > > > > > > > > > > > > > > > > > For applications, DPDK GRO provides three external functi= ons to > > > > > > > > > enable/disable GRO: > > > > > > > > > - rte_gro_init: initialize GRO environment; > > > > > > > > > - rte_gro_enable: enable GRO for a given port; > > > > > > > > > - rte_gro_disable: disable GRO for a given port. > > > > > > > > > Before using GRO, applications should explicitly call rte= _gro_init to > > > > > > > > > initizalize GRO environment. After that, applications can= call > > > > > > > > > rte_gro_enable to enable GRO and call rte_gro_disable to = disable GRO for > > > > > > > > > specific ports. > > > > > > > > > > > > > > > > I think this is too restrictive and wouldn't meet various u= ser's needs. > > > > > > > > User might want to: > > > > > > > > - enable/disable GRO for particular RX queue > > > > > > > > - or even setup different GRO types for different RX queues= , > > > > > > > > i.e, - GRO over IPV4/TCP for queue 0, and GRO over IPV6= /TCP for queue 1, etc. > > > > > > > > > > > > > > The reason for enabling/disabling GRO per-port instead of per= -queue is that LINUX > > > > > > > controls GRO per-port. To control GRO per-queue indeed can pr= ovide more flexibility > > > > > > > to applications. But are there any scenarios that different q= ueues of a port may > > > > > > > require different GRO control (i.e. GRO types and enable/disa= ble GRO)? > > > > > > > > > > > > I think yes. > > > > > > > > > > > > > > > > > > > > > - For various reasons, user might prefer not to use RX call= backs for various reasons, > > > > > > > > But invoke gro() manually at somepoint in his code. > > > > > > > > > > > > > > An application-used GRO library can enable more flexibility t= o applications. Besides, > > > > > > > when perform GRO in ethdev layer or inside PMD drivers, it is= an issue that > > > > > > > rte_eth_rx_burst returns actually received packet number or G= ROed packet number. And > > > > > > > the same issue happens in GSO, and even more seriously. This = is because applications > > > > > > > , like VPP, always rely on the return value of rte_eth_tx_bur= st to decide further > > > > > > > operations. If applications can direcly call GRO/GSO librarie= s, this issue won't exist. > > > > > > > And DPDK is a library, which is not a holistic system like LI= NUX. We don't need to do > > > > > > > the same as LINUX. Therefore, maybe it's a better idea to dir= ectly provide SW > > > > > > > segmentation/reassembling libraries to applications. > > > > > > > > > > > > > > > - Many users would like to control size (number of flows/it= ems per flow), > > > > > > > > max allowed packet size, max timeout, etc., for different= GRO tables. > > > > > > > > - User would need a way to flush all or only timeout packet= s from particular GRO tables. > > > > > > > > > > > > > > > > So I think that API needs to extended to become be much mor= e fine-grained. > > > > > > > > Something like that: > > > > > > > > > > > > > > > > struct rte_gro_tbl_param { > > > > > > > > int32_t socket_id; > > > > > > > > size_t max_flows; > > > > > > > > size_t max_items_per_flow; > > > > > > > > size_t max_pkt_size; > > > > > > > > uint64_t packet_timeout_cycles; > > > > > > > > > > > > > > > > > > > > > > > > ... > > > > > > > > }; > > > > > > > > > > > > > > > > struct rte_gro_tbl; > > > > > > > > strct rte_gro_tbl *rte_gro_tbl_create(const struct rte_gro_= tbl_param *param); > > > > > > > > > > > > > > > > void rte_gro_tbl_destroy(struct rte_gro_tbl *tbl); > > > > > > > > > > > > > > Yes, I agree with you. It's necessary to provide more fine-gr= ained control APIs to > > > > > > > applications. But what's 'packet_timeout_cycles' used for? Is= it for TCP packets? > > > > > > > > > > > > For any packets that sits in the gro_table for too long. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > /* > > > > > > > > * process packets, might store some packets inside the GRO= table, > > > > > > > > * returns number of filled entries in pkt[] > > > > > > > > */ > > > > > > > > uint32_t rte_gro_tbl_process(struct rte_gro_tbl *tbl, struc= t rte_mbuf *pkt[], uint32_t num); > > > > > > > > > > > > > > > > /* > > > > > > > > * retirieves up to num timeouted packets from the table. > > > > > > > > */ > > > > > > > > uint32_t rtre_gro_tbl_timeout(struct rte_gro_tbl *tbl, uint= 64_t tmt, struct rte_mbuf *pkt[], uint32_t num); > > > > > > > > > > > > > > Currently, we implement GRO as RX callback, whose processing = logic is simple: > > > > > > > receive burst packets -> perform GRO -> return to application= . GRO stops after > > > > > > > finishing processing received packets. If we provide rte_gro_= tbl_timeout, when > > > > > > > and who will call it? > > > > > > > > > > > > I mean the following scenario: > > > > > > We receive a packet, find it is eligible for GRO and put it int= o gro_table > > > > > > in expectation - there would be more packets for the same flow. > > > > > > But it could happen that we would never (or for some long time)= receive > > > > > > any new packets for that flow. > > > > > > So the first packet would never be delivered to the upper layer= , > > > > > > or delivered too late. > > > > > > So we need a mechanism to extract such not merged packets > > > > > > and push them to the upper layer. > > > > > > My thought it would be application responsibility to call it fr= om time to time. > > > > > > That's actually another reason why I think we shouldn't use app= lication > > > > > > to always use RX callbacks for GRO. > > > > > > > > > > Currently, we only provide one reassembly function, rte_gro_reass= emble_burst, > > > > > which merges N inputted packets at a time. After finishing proces= sing these > > > > > packets, it returns all of them and reset hashing tables. Therefo= re, there > > > > > are no packets in hashing tables after rte_gro_reassemble_burst r= eturns. > > > > > > > > Ok, sorry I missed that part with rte_hash_reset(). > > > > So gro_ressemble_burst() performs only inline processing on current= input packets > > > > and doesn't try to save packets for future merging, correct? > > > > > > Yes. > > > > > > > Such approach indeed is much lightweight and doesn't require any ex= tra timeouts and flush(). > > > > So my opinion let's keep it like that - nice and simple. > > > > BTW, I think in that case we don't need any hashtables (or any othe= r persistent strucures)at all. > > > > What we need is just a set of GROs (tcp4, tpc6, etc.) we want to pe= rform on given array of packets. > > > > > > Beside GRO types that are desired to perform, maybe it also needs max= _pkt_size and > > > some GRO type specific information? > > > > Yes, but we don't need the actual hash-tables, etc. inside. > > Passing something like struct gro_param seems enough. >=20 > Yes, we can just pass gro_param and allocate hashing tables > inside rte_gro_reassemble_burst. If so, hashing tables of > desired GRO types are created and freed in each invocation > of rte_gro_reassemble_burst. In GRO library, hashing tables > are created by GRO type specific gro_tbl_create_fn. These > gro_tbl_create_fn may allocate hashing table space via malloc > (or rte_malloc). Therefore, we may frequently call malloc/free > when using rte_gro_reassemble_burst. In my opinion, it will > degrade GRO performance greatly. I don't' understand why do we need to put/extract each packet into/from has= h table at all. We have N input packets that need to be grouped/sorted by some criteria. Surely that can be done without any hash-table involved. What is the need for hash table and all the overhead it brings here? Konstantin >=20 > But if we ask applications to input hashing tables, what we > need to do is to reset them after finishing using in > rte_gro_reassemble_burst, rather than to malloc and free each > time. Therefore, I think this way is more efficient. How do you > think? >=20 > > > > > > > > > > > > > > > > > > > If we provide rte_gro_tbl_timeout, we also need to provide anothe= r reassmebly > > > > > function, like rte_gro_reassemble, which processes one given pack= et at a > > > > > time and won't reset hashing tables. Applications decide when to = flush packets > > > > > in hashing tables. And rte_gro_tbl_timeout is one of the ways tha= t can be used > > > > > to flush the packets. Do you mean that? > > > > > > > > Yes, that's what I meant, but as I said above - I think your approa= ch is probably > > > > more preferable - it is much simpler and lightweight. > > > > Konstantin > > > >