From mboxrd@z Thu Jan 1 00:00:00 1970 From: "Phil Yang (Arm Technology China)" Subject: Re: [PATCH v2 2/3] kni: fix kni fifo synchronization Date: Fri, 21 Sep 2018 09:00:31 +0000 Message-ID: References: <1537363820-3827-1-git-send-email-phil.yang@arm.com> <1537364560-4124-1-git-send-email-phil.yang@arm.com> <1537364560-4124-2-git-send-email-phil.yang@arm.com> <20180920082846.GB19425@jerin> <20180920153700.GA9459@jerin> <20180921055529.GA15861@jerin> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: quoted-printable Cc: "dev@dpdk.org" , nd , "kkokkilagadda@caviumnetworks.com" , "Gavin Hu (Arm Technology China)" , "ferruh.yigit@intel.com" , Ola Liljedahl To: Honnappa Nagarahalli , Jerin Jacob Return-path: Received: from EUR03-DB5-obe.outbound.protection.outlook.com (mail-eopbgr40085.outbound.protection.outlook.com [40.107.4.85]) by dpdk.org (Postfix) with ESMTP id 4195B1041 for ; Fri, 21 Sep 2018 11:00:33 +0200 (CEST) In-Reply-To: Content-Language: en-US List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org Sender: "dev" + Ola Liljedahl Thanks, Phil Yang > -----Original Message----- > From: Honnappa Nagarahalli > Sent: Friday, September 21, 2018 2:37 PM > To: Jerin Jacob > Cc: Phil Yang (Arm Technology China) ; dev@dpdk.org; n= d > ; kkokkilagadda@caviumnetworks.com; Gavin Hu (Arm > Technology China) ; ferruh.yigit@intel.com > Subject: RE: [PATCH v2 2/3] kni: fix kni fifo synchronization >=20 > > > > > > > > > > > > > > @@ -69,5 +89,13 @@ kni_fifo_get(struct rte_kni_fifo *fifo, > > > > > > > void **data, unsigned num) static inline uint32_t > > > > > > > kni_fifo_count(struct rte_kni_fifo *fifo) { > > > > > > > +#ifdef RTE_USE_C11_MEM_MODEL > > > > > > > + unsigned fifo_write =3D __atomic_load_n(&fifo->write, > > > > > > > + __ATOMIC_AC= QUIRE); > > > > > > > + unsigned fifo_read =3D __atomic_load_n(&fifo->read, > > > > > > > + > > > > > > > +__ATOMIC_ACQUIRE); > > > > > > > > > > > > Isn't too heavy to have two __ATOMIC_ACQUIREs? a simple > > > > > > rte_smp_rmb() would be enough here. Right? > > > > > > or > > > > > > Do we need __ATOMIC_ACQUIRE for fifo_write case? > > > > > > > > > > > We also had some amount of debate internally on this: > > > > > 1) We do not want to use rte_smp_rmb() as we want to keep the > > > > > memory > > > > models separated (for ex: while using C11, use C11 everywhere). It > > > > is also not sufficient, please see 3) below. > > > > > > > > But Nothing technically wrong in using rte_smp_rmb() here in terms > > > > functionally and code generated by the compiler. > > > > > > rte_smp_rmb() generates 'DMB ISHLD'. This works fine, but it is not o= ptimal. > > 'LDAR' is a better option which is generated when C11 atomics are used. > > > > Yes. But which one is optimal 1 x DMB ISHLD vs 2 x LDAR ? >=20 > Good point. I am not sure which one is optimal, it needs to be measured. = 'DMB > ISHLD' orders 'all' earlier loads against 'all' later loads and stores. '= LDAR' orders > the 'specific' load with 'all' later loads and stores. >=20 > > > > > > > > > > > > > > 2) This API can get called from writer or reader, so both the > > > > > loads have to be __ATOMIC_ACQUIRE > > > > > 3) Other option is to use __ATOMIC_RELAXED. That would allow any > > > > loads/stores around of this API to get reordered, especially since > > > > this is an inline function. This would put burden on the > > > > application to manage the ordering depending on its usage. It will > > > > also require the application to understand the implementation of th= is API. > > > > > > > > __ATOMIC_RELAXED may be fine too for _count() case as it may not > > > > very important to get the exact count for the exact very moment, > > > > Application can retry. > > > > > > > > I am in favor of performance effective implementation. > > > > > > The requirement on the correctness of the count depends on the usage > > > of > > this function. I see the following usage: > > > > > > In the file kni_net.c, function: kni_net_tx: > > > > > > if (kni_fifo_free_count(kni->tx_q) =3D=3D 0 || > > > kni_fifo_count(kni->alloc_q) =3D=3D 0) { > > > /** > > > * If no free entry in tx_q or no entry in alloc_q, > > > * drops skb and goes out. > > > */ > > > goto drop; > > > } > > > > > > There is no retry here, the packet is dropped. > > > > OK. Then pick an implementation which is an optimal this case. > > I think, then rte_smp_rmb() makes sense here as > > a) no #ifdef clutter > > b) it is optimal compared to 2 x LDAR > > > As I understand, one of the principals of using C11 model is to match the= store > releases and load acquires. IMO, combining C11 memory model with barrier > based functions makes the code unreadable. > I realized rte_smp_rmb() is required for x86 as well to prevent compiler > reordering. We can add that in the non-C11 case. This way, we will have c= lean > code for both the options (similar to rte_ring). > So, if 'RTE_USE_C11_MEM_MODEL' is set to 'n', then the 'rte_smp_rmb' woul= d > be used. >=20 > We can look at handling the #ifdef clutter based on Ferruh's feedback. >=20 > > > > > > > > > > > > > > > > > > > > > > > > > > Other than that, I prefer to avoid ifdef clutter by > > > > > > introducing two separate file just like ring C11 implementation= . > > > > > > > > > > > > I don't have strong opinion on this this part, I let KNI > > > > > > MAINTAINER to decide on how to accommodate this change. > > > > > > > > > > I prefer to change this as well, I am open for suggestions. > > > > > Introducing two separate files would be too much for this library= . > > > > > A better > > > > way would be to have something similar to 'smp_store_release' > > > > provided by the kernel. i.e. create #defines for loads/stores. > > > > Hide the clutter behind the #defines. > > > > > > > > No Strong opinion on this, leaving to KNI Maintainer. > > > Will wait on this before re-spinning the patch > > > > > > > > > > > This patch needs to split by two, > > > > a) Fixes for non C11 implementation(i.e new addition to > > > > rte_smp_wmb()) > > > > b) add support for C11 implementation. > > > Agree > > > > > > > > > > > > > > > > > > > > > > > > > + return (fifo->len + fifo_write - fifo_read) & > > > > > > > +(fifo->len - 1); #else > > > > > > > return (fifo->len + fifo->write - fifo->read) & > > > > > > > (fifo->len > > > > > > > - 1); > Requires rte_smp_rmb() for x86 to prevent compiler reordering. >=20 > > > > > > > +#endif > > > > > > > } > > > > > > > -- > > > > > > > 2.7.4 > > > > > > >