From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-6.8 required=3.0 tests=HEADER_FROM_DIFFERENT_DOMAINS, INCLUDES_PATCH,MAILING_LIST_MULTI,SIGNED_OFF_BY,SPF_HELO_NONE,SPF_PASS, URIBL_BLOCKED autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id CE246C4360C for ; Thu, 10 Oct 2019 15:10:03 +0000 (UTC) Received: from dpdk.org (dpdk.org [92.243.14.124]) by mail.kernel.org (Postfix) with ESMTP id 6699B206A1 for ; Thu, 10 Oct 2019 15:10:03 +0000 (UTC) DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org 6699B206A1 Authentication-Results: mail.kernel.org; dmarc=fail (p=none dis=none) header.from=intel.com Authentication-Results: mail.kernel.org; spf=pass smtp.mailfrom=dev-bounces@dpdk.org Received: from [92.243.14.124] (localhost [127.0.0.1]) by dpdk.org (Postfix) with ESMTP id 2FFFF1E4E2; Thu, 10 Oct 2019 17:10:02 +0200 (CEST) Received: from mga18.intel.com (mga18.intel.com [134.134.136.126]) by dpdk.org (Postfix) with ESMTP id 2D7BC1DFEC for ; Thu, 10 Oct 2019 17:09:59 +0200 (CEST) X-Amp-Result: SKIPPED(no attachment in message) X-Amp-File-Uploaded: False Received: from fmsmga001.fm.intel.com ([10.253.24.23]) by orsmga106.jf.intel.com with ESMTP/TLS/DHE-RSA-AES256-GCM-SHA384; 10 Oct 2019 08:09:58 -0700 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.67,280,1566889200"; d="scan'208";a="207207810" Received: from irsmsx104.ger.corp.intel.com ([163.33.3.159]) by fmsmga001.fm.intel.com with ESMTP; 10 Oct 2019 08:09:56 -0700 Received: from irsmsx105.ger.corp.intel.com ([169.254.7.164]) by IRSMSX104.ger.corp.intel.com ([169.254.5.103]) with mapi id 14.03.0439.000; Thu, 10 Oct 2019 16:09:56 +0100 From: "Ananyev, Konstantin" To: Honnappa Nagarahalli , "stephen@networkplumber.org" , "paulmck@linux.ibm.com" CC: "Wang, Yipeng1" , "Medvedkin, Vladimir" , "Ruifeng Wang (Arm Technology China)" , Dharmik Thakkar , "dev@dpdk.org" , nd , nd Thread-Topic: [PATCH v3 1/3] lib/ring: add peek API Thread-Index: AQHVeCGUIgBwOBQCI0OY47a7OfyurKdHoInwgAGlroCABaC3MIACyz6AgAJU8SA= Date: Thu, 10 Oct 2019 15:09:55 +0000 Message-ID: <2601191342CEEE43887BDE71AB9772580191975145@irsmsx105.ger.corp.intel.com> References: <20190906094534.36060-1-ruifeng.wang@arm.com> <20191001062917.35578-1-honnappa.nagarahalli@arm.com> <20191001062917.35578-2-honnappa.nagarahalli@arm.com> <2601191342CEEE43887BDE71AB9772580191970014@irsmsx105.ger.corp.intel.com> <2601191342CEEE43887BDE71AB9772580191971EBE@irsmsx105.ger.corp.intel.com> In-Reply-To: Accept-Language: en-IE, en-US Content-Language: en-US X-MS-Has-Attach: X-MS-TNEF-Correlator: x-titus-metadata-40: eyJDYXRlZ29yeUxhYmVscyI6IiIsIk1ldGFkYXRhIjp7Im5zIjoiaHR0cDpcL1wvd3d3LnRpdHVzLmNvbVwvbnNcL0ludGVsMyIsImlkIjoiNjU0NmI0NGUtZDFmOS00YTQ2LWEwMWEtOWRjNWY0OTBkYmY3IiwicHJvcHMiOlt7Im4iOiJDVFBDbGFzc2lmaWNhdGlvbiIsInZhbHMiOlt7InZhbHVlIjoiQ1RQX05UIn1dfV19LCJTdWJqZWN0TGFiZWxzIjpbXSwiVE1DVmVyc2lvbiI6IjE3LjEwLjE4MDQuNDkiLCJUcnVzdGVkTGFiZWxIYXNoIjoiSjd2NlEzbHN6cEluNHdOZmhQSFFDZWJlbFR1T0RGQitwR0ZRM200czBqZ2pqc05vSmF2VWFlQTBRV3hNUFwvZ1oifQ== x-ctpclassification: CTP_NT dlp-product: dlpe-windows dlp-version: 11.2.0.6 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/ring: add peek API 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" > >=20 > > > > > > > > > > Subject: [PATCH v3 1/3] lib/ring: add peek API > > > > > > > > > > From: Ruifeng Wang > > > > > > > > > > The peek API allows fetching the next available object in the rin= g > > > > > without dequeuing it. This helps in scenarios where dequeuing of > > > > > objects depend on their value. > > > > > > > > > > Signed-off-by: Dharmik Thakkar > > > > > Signed-off-by: Ruifeng Wang > > > > > Reviewed-by: Honnappa Nagarahalli > > > > > Reviewed-by: Gavin Hu > > > > > --- > > > > > lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ > > > > > 1 file changed, 30 insertions(+) > > > > > > > > > > diff --git a/lib/librte_ring/rte_ring.h > > > > > b/lib/librte_ring/rte_ring.h index 2a9f768a1..d3d0d5e18 100644 > > > > > --- a/lib/librte_ring/rte_ring.h > > > > > +++ b/lib/librte_ring/rte_ring.h > > > > > @@ -953,6 +953,36 @@ rte_ring_dequeue_burst(struct rte_ring *r, > > > > > void > > > > **obj_table, > > > > > r->cons.single, available); > > > > > } > > > > > > > > > > +/** > > > > > + * Peek one object from a ring. > > > > > + * > > > > > + * The peek API allows fetching the next available object in the > > > > > +ring > > > > > + * without dequeuing it. This API is not multi-thread safe with > > > > > +respect > > > > > + * to other consumer threads. > > > > > + * > > > > > + * @param r > > > > > + * A pointer to the ring structure. > > > > > + * @param obj_p > > > > > + * A pointer to a void * pointer (object) that will be filled. > > > > > + * @return > > > > > + * - 0: Success, object available > > > > > + * - -ENOENT: Not enough entries in the ring. > > > > > + */ > > > > > +__rte_experimental > > > > > +static __rte_always_inline int > > > > > +rte_ring_peek(struct rte_ring *r, void **obj_p) > > > > > > > > As it is not MT safe, then I think we need _sc_ in the name, to > > > > follow other rte_ring functions naming conventions > > > > (rte_ring_sc_peek() or so). > > > Agree > > > > > > > > > > > As a better alternative what do you think about introducing a > > > > serialized versions of DPDK rte_ring dequeue functions? > > > > Something like that: > > > > > > > > /* same as original ring dequeue, but: > > > > * 1) move cons.head only if cons.head =3D=3D const.tail > > > > * 2) don't update cons.tail > > > > */ > > > > unsigned int > > > > rte_ring_serial_dequeue_bulk(struct rte_ring *r, void **obj_table, > > > > unsigned int n, > > > > unsigned int *available); > > > > > > > > /* sets both cons.head and cons.tail to cons.head + num */ void > > > > rte_ring_serial_dequeue_finish(struct rte_ring *r, uint32_t num); > > > > > > > > /* resets cons.head to const.tail value */ void > > > > rte_ring_serial_dequeue_abort(struct rte_ring *r); > > > > > > > > Then your dq_reclaim cycle function will look like that: > > > > > > > > const uint32_t nb_elt =3D dq->elt_size/8 + 1; uint32_t avl, n; > > > > uintptr_t elt[nb_elt]; ... > > > > > > > > do { > > > > > > > > /* read next elem from the queue */ > > > > n =3D rte_ring_serial_dequeue_bulk(dq->r, elt, nb_elt, &avl); > > > > if (n =3D=3D 0) > > > > break; > > > > > > > > /* wrong period, keep elem in the queue */ if > > > > (rte_rcu_qsbr_check(dr->v, > > > > elt[0]) !=3D 1) { > > > > rte_ring_serial_dequeue_abort(dq->r); > > > > break; > > > > } > > > > > > > > /* can reclaim, remove elem from the queue */ > > > > rte_ring_serial_dequeue_finish(dr->q, nb_elt); > > > > > > > > /*call reclaim function */ > > > > dr->f(dr->p, elt); > > > > > > > > } while (avl >=3D nb_elt); > > > > > > > > That way, I think even rte_rcu_qsbr_dq_reclaim() can be MT safe. > > > > As long as actual reclamation callback itself is MT safe of course. > > > > > > I think it is a great idea. The other writers would still be polling > > > for the current writer to update the tail or update the head. This ma= kes it a > > blocking solution. > > > > Yep, it is a blocking one. > > > > > We can make the other threads not poll i.e. they will quit reclaiming= if they > > see that other writers are dequeuing from the queue. > > > > Actually didn't think about that possibility, but yes should be possibl= e to have > > _try_ semantics too. > > > > >The other way is to use per thread queues. > > > > > > The other requirement I see is to support unbounded-size data > > > structures where in the data structures do not have a pre-determined > > > number of entries. Also, currently the defer queue size is equal to t= he total > > number of entries in a given data structure. There are plans to support > > dynamically resizable defer queue. This means, memory allocation which = will > > affect the lock-free-ness of the solution. > > > > > > So, IMO: > > > 1) The API should provide the capability to support different algorit= hms - > > may be through some flags? > > > 2) The requirements for the ring are pretty unique to the problem we > > > have here (for ex: move the cons-head only if cons-tail is also the s= ame, skip > > polling). So, we should probably implement a ring with-in the RCU libra= ry? > > > > Personally, I think such serialization ring API would be useful for oth= er cases > > too. > > There are few cases when user need to read contents of the queue withou= t > > removing elements from it. > > Let say we do use similar approach inside TLDK to implement TCP transmi= t > > queue. > > If such API would exist in DPDK we can just use it straightway, without > > maintaining a separate one. > ok >=20 > > > > > > > > From the timeline perspective, adding all these capabilities would be > > > difficult to get done with in 19.11 timeline. What I have here > > > satisfies my current needs. I suggest that we make provisions in APIs= now to > > support all these features, but do the implementation in the coming rel= eases. > > Does this sound ok for you? > > > > Not sure I understand your suggestion here... > > Could you explain it a bit more - how new API will look like and what w= ould > > be left for the future. > For this patch, I suggest we do not add any more complexity. If someone w= ants a lock-free/block-free mechanism, it is available by creating > per thread defer queues. >=20 > We push the following to the future: > 1) Dynamically size adjustable defer queue. IMO, with this, the lock-free= /block-free reclamation will not be available (memory allocation > requires locking). The memory for the defer queue will be allocated/freed= in chunks of 'size' elements as the queue grows/shrinks. That one is fine by me. In fact I don't know would be there a real use-case for dynamic defer queue= for rcu var... But I suppose that's subject for another discussion. >=20 > 2) Constant size defer queue with lock-free and block-free reclamation (s= ingle option). The defer queue will be of fixed length 'size'. If the > queue gets full an error is returned. The user could provide a 'size' equ= al to the number of elements in a data structure to ensure queue > never gets full. Ok so for 19.11 what enqueue/dequeue model do you plan to support? - MP/MC - MP/SC - SP/SC - non MT at all (only same single thread can do enqueue and dequeue) And related question: What additional rte_ring API you plan to introduce in that case? - None - rte_ring_sc_peek() - rte_ring_serial_dequeue() >=20 > I would add a 'flags' field in rte_rcu_qsbr_dq_parameters and provide 2 #= defines, one for dynamically variable size defer queue and the > other for constant size defer queue. >=20 > However, IMO, using per thread defer queue is a much simpler way to achie= ve 2. It does not add any significant burden to the user either. >=20 > > > > > > > > > > > > > > +{ > > > > > + uint32_t prod_tail =3D r->prod.tail; > > > > > + uint32_t cons_head =3D r->cons.head; > > > > > + uint32_t count =3D (prod_tail - cons_head) & r->mask; > > > > > + unsigned int n =3D 1; > > > > > + if (count) { > > > > > + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); > > > > > + return 0; > > > > > + } > > > > > + return -ENOENT; > > > > > +} > > > > > + > > > > > #ifdef __cplusplus > > > > > } > > > > > #endif > > > > > -- > > > > > 2.17.1