From mboxrd@z Thu Jan 1 00:00:00 1970 From: Matan Azrad Subject: Re: [PATCH v3 6/7] net/mlx4: mitigate Tx path memory barriers Date: Mon, 30 Oct 2017 19:47:20 +0000 Message-ID: References: <1508768520-4810-1-git-send-email-ophirmu@mellanox.com> <1509358049-18854-1-git-send-email-matan@mellanox.com> <1509358049-18854-7-git-send-email-matan@mellanox.com> <20171030142350.GC26782@6wind.com> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: quoted-printable Cc: "dev@dpdk.org" , Ophir Munk To: Adrien Mazarguil Return-path: Received: from EUR01-DB5-obe.outbound.protection.outlook.com (mail-db5eur01on0051.outbound.protection.outlook.com [104.47.2.51]) by dpdk.org (Postfix) with ESMTP id 369A21B28B for ; Mon, 30 Oct 2017 20:47:23 +0100 (CET) In-Reply-To: <20171030142350.GC26782@6wind.com> 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" Hi Adrien > -----Original Message----- > From: Adrien Mazarguil [mailto:adrien.mazarguil@6wind.com] > Sent: Monday, October 30, 2017 4:24 PM > To: Matan Azrad > Cc: dev@dpdk.org; Ophir Munk > Subject: Re: [PATCH v3 6/7] net/mlx4: mitigate Tx path memory barriers >=20 > On Mon, Oct 30, 2017 at 10:07:28AM +0000, Matan Azrad wrote: > > Replace most of the memory barriers by compiler barriers since they > > are all targeted to the DRAM; This improves code efficiency for > > systems which force store order between different addresses. > > > > Only the doorbell record store should be protected by memory barrier > > since it is targeted to the PCI memory domain. > > > > Limit pre byte count store compiler barrier for systems with cache > > line size smaller than 64B (TXBB size). > > > > Signed-off-by: Matan Azrad >=20 > This sounds like an interesting performance improvement, can you share th= e > typical or expected amount (percentage/hard numbers) for a given use case > as part of the commit log? >=20 Yes, it improves performance, I will share numbers. > More comments below. >=20 > > --- > > drivers/net/mlx4/mlx4_rxtx.c | 11 ++++++----- > > 1 file changed, 6 insertions(+), 5 deletions(-) > > > > diff --git a/drivers/net/mlx4/mlx4_rxtx.c > > b/drivers/net/mlx4/mlx4_rxtx.c index 8ea8851..482c399 100644 > > --- a/drivers/net/mlx4/mlx4_rxtx.c > > +++ b/drivers/net/mlx4/mlx4_rxtx.c > > @@ -168,7 +168,7 @@ struct pv { > > /* > > * Make sure we read the CQE after we read the ownership > bit. > > */ > > - rte_rmb(); > > + rte_io_rmb(); >=20 > OK for this one since the rest of the code should not be run due to the > condition (I'm not even sure even a compiler barrier is necessary at all = here). >=20 > > #ifndef NDEBUG > > if (unlikely((cqe->owner_sr_opcode & > MLX4_CQE_OPCODE_MASK) =3D=3D > > MLX4_CQE_OPCODE_ERROR)) { > > @@ -203,7 +203,7 @@ struct pv { > > */ > > cq->cons_index =3D cons_index; > > *cq->set_ci_db =3D rte_cpu_to_be_32(cq->cons_index & > MLX4_CQ_DB_CI_MASK); > > - rte_wmb(); > > + rte_io_wmb(); >=20 > This one could be removed entirely as well, which is more or less what th= e > move to a compiler barrier does. Nothing in subsequent code depends on > this doorbell being written, so this can piggy back on any subsequent > rte_wmb(). Yes, you right, probably this code was taken from multi thread implementati= on. >=20 > On the other hand in my opinion a barrier (compiler or otherwise) might b= e > needed before the doorbell write, to make clear it cannot somehow be done > earlier in case something attempts to optimize it away. >=20 I think we can remove it entirely (compiler can't optimize the ci_db store = since in depends in previous code (cons_index). > > sq->tail =3D sq->tail + nr_txbbs; > > /* Update the list of packets posted for transmission. */ > > elts_comp -=3D pkts; > > @@ -321,6 +321,7 @@ static int handle_multi_segs(struct rte_mbuf *buf, > > * control segment. > > */ > > if ((uintptr_t)dseg & (uintptr_t)(MLX4_TXBB_SIZE - 1)) { > > +#if RTE_CACHE_LINE_SIZE < 64 > > /* > > * Need a barrier here before writing the byte_count > > * fields to make sure that all the data is visible @@ - > 331,6 > > +332,7 @@ static int handle_multi_segs(struct rte_mbuf *buf, > > * data, and end up sending the wrong data. > > */ > > rte_io_wmb(); > > +#endif /* RTE_CACHE_LINE_SIZE */ >=20 > Interesting one. >=20 > > dseg->byte_count =3D byte_count; > > } else { > > /* > > @@ -469,8 +471,7 @@ static int handle_multi_segs(struct rte_mbuf *buf, > > break; > > } > > #endif /* NDEBUG */ > > - /* Need a barrier here before byte count store. */ > > - rte_io_wmb(); > > + /* Never be TXBB aligned, no need compiler barrier. > */ >=20 > The reason there was a barrier here at all was unclear, so if it's really= useless, > you don't even need to describe why. It is because there is a barrier in multi segment similar stage. I think it can help for future review. >=20 > > dseg->byte_count =3D rte_cpu_to_be_32(buf- > >data_len); > > > > /* Fill the control parameters for this packet. */ @@ - > 533,7 > > +534,7 @@ static int handle_multi_segs(struct rte_mbuf *buf, > > * setting ownership bit (because HW can start > > * executing as soon as we do). > > */ > > - rte_wmb(); > > + rte_io_wmb(); >=20 > This one looks dangerous. A compiler barrier is not strong enough to > guarantee the order in which CPU will execute instructions, it only makes > sure what follows the barrier doesn't appear before it in the generated c= ode. >=20 As I investigated, I understood that for CPUs which don't save store order = between different addresses(arm,ppc), the rte_io_wmb is converted to rte_wm= b. So for thus who save it(x86) we just need the right order in compiler code = because all the relevant stores are targeted to same memory domain(DRAM) an= d therefore also the actual store is guaranteed. Unlike doorbell store which directed to different memory domain (PCI). So the only place which need rte_wmb() is before doorbell write. > Unless the comment above this barrier is wrong, this change may cause har= d- > to-debug issues down the road, you should drop it. >=20 > > ctrl->owner_opcode =3D rte_cpu_to_be_32(owner_opcode | > > ((sq->head & sq->txbb_cnt) ? > > MLX4_BIT_WQE_OWN : > 0)); > > -- > > 1.8.3.1 > > >=20 > -- > Adrien Mazarguil > 6WIND Thanks!