From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mga14.intel.com (mga14.intel.com [192.55.52.115]) by dpdk.org (Postfix) with ESMTP id 620D1924E for ; Fri, 23 Oct 2015 04:21:20 +0200 (CEST) Received: from orsmga001.jf.intel.com ([10.7.209.18]) by fmsmga103.fm.intel.com with ESMTP; 22 Oct 2015 19:20:55 -0700 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.20,185,1444719600"; d="scan'208";a="800745205" Received: from fmsmsx103.amr.corp.intel.com ([10.18.124.201]) by orsmga001.jf.intel.com with ESMTP; 22 Oct 2015 19:20:55 -0700 Received: from fmsmsx157.amr.corp.intel.com (10.18.116.73) by FMSMSX103.amr.corp.intel.com (10.18.124.201) with Microsoft SMTP Server (TLS) id 14.3.248.2; Thu, 22 Oct 2015 19:20:55 -0700 Received: from shsmsx152.ccr.corp.intel.com (10.239.6.52) by FMSMSX157.amr.corp.intel.com (10.18.116.73) with Microsoft SMTP Server (TLS) id 14.3.248.2; Thu, 22 Oct 2015 19:20:54 -0700 Received: from shsmsx101.ccr.corp.intel.com ([169.254.1.96]) by SHSMSX152.ccr.corp.intel.com ([169.254.6.106]) with mapi id 14.03.0248.002; Fri, 23 Oct 2015 10:20:52 +0800 From: "Xie, Huawei" To: Stephen Hemminger Thread-Topic: [dpdk-dev] [PATCH v4 6/7] virtio: simple tx routine Thread-Index: AdENOOrj9g4A3zrLRF2F8uH+Pvbylg== Date: Fri, 23 Oct 2015 02:20:52 +0000 Message-ID: References: <1443537953-23917-1-git-send-email-huawei.xie@intel.com> <1445515791-25909-1-git-send-email-huawei.xie@intel.com> <1445515791-25909-7-git-send-email-huawei.xie@intel.com> <20151022095730.5a3c742b@xeon-e3> Accept-Language: en-US Content-Language: en-US X-MS-Has-Attach: X-MS-TNEF-Correlator: x-originating-ip: [10.239.127.40] Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: quoted-printable MIME-Version: 1.0 Cc: "dev@dpdk.org" Subject: Re: [dpdk-dev] [PATCH v4 6/7] virtio: simple tx routine X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.15 Precedence: list List-Id: patches and discussions about DPDK List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Fri, 23 Oct 2015 02:21:22 -0000 On 10/23/2015 10:17 AM, Xie, Huawei wrote:=0A= > On 10/23/2015 12:57 AM, Stephen Hemminger wrote:=0A= >> On Thu, 22 Oct 2015 20:09:50 +0800=0A= >> Huawei Xie wrote:=0A= >>=0A= >>> Changes in v4:=0A= >>> - move virtio_xmit_cleanup ahead to free descriptors earlier=0A= >>>=0A= >>> Changes in v3:=0A= >>> - Remove return at the end of void function=0A= >>> - Remove always_inline attribute for virtio_xmit_cleanup=0A= >>> bulk free of mbufs when clean used ring.=0A= >>> shift operation of idx could be saved if vq_free_cnt means=0A= >>> free slots rather than free descriptors.=0A= >>>=0A= >>> TODO: rearrange vq data structure, pack the stats var together so that = we=0A= >>> could use one vec instruction to update all of them.=0A= >>>=0A= >>> Signed-off-by: Huawei Xie =0A= >>> ---=0A= >>> drivers/net/virtio/virtio_ethdev.h | 3 ++=0A= >>> drivers/net/virtio/virtio_rxtx_simple.c | 93 +++++++++++++++++++++++++= ++++++++=0A= >>> 2 files changed, 96 insertions(+)=0A= >>>=0A= >>> diff --git a/drivers/net/virtio/virtio_ethdev.h b/drivers/net/virtio/vi= rtio_ethdev.h=0A= >>> index d7797ab..ae2d47d 100644=0A= >>> --- a/drivers/net/virtio/virtio_ethdev.h=0A= >>> +++ b/drivers/net/virtio/virtio_ethdev.h=0A= >>> @@ -111,6 +111,9 @@ uint16_t virtio_xmit_pkts(void *tx_queue, struct rt= e_mbuf **tx_pkts,=0A= >>> uint16_t virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkt= s,=0A= >>> uint16_t nb_pkts);=0A= >>> =0A= >>> +uint16_t virtio_xmit_pkts_simple(void *tx_queue, struct rte_mbuf **tx_= pkts,=0A= >>> + uint16_t nb_pkts);=0A= >>> +=0A= >>> /*=0A= >>> * The VIRTIO_NET_F_GUEST_TSO[46] features permit the host to send us= =0A= >>> * frames larger than 1514 bytes. We do not yet support software LRO= =0A= >>> diff --git a/drivers/net/virtio/virtio_rxtx_simple.c b/drivers/net/virt= io/virtio_rxtx_simple.c=0A= >>> index ef17562..79b4f7f 100644=0A= >>> --- a/drivers/net/virtio/virtio_rxtx_simple.c=0A= >>> +++ b/drivers/net/virtio/virtio_rxtx_simple.c=0A= >>> @@ -288,6 +288,99 @@ virtio_recv_pkts_vec(void *rx_queue, struct rte_mb= uf **rx_pkts,=0A= >>> return nb_pkts_received;=0A= >>> }=0A= >>> =0A= >>> +#define VIRTIO_TX_FREE_THRESH 32=0A= >>> +#define VIRTIO_TX_MAX_FREE_BUF_SZ 32=0A= >>> +#define VIRTIO_TX_FREE_NR 32=0A= >>> +/* TODO: vq->tx_free_cnt could mean num of free slots so we could avoi= d shift */=0A= >>> +static inline void=0A= >>> +virtio_xmit_cleanup(struct virtqueue *vq)=0A= >>> +{=0A= >>> + uint16_t i, desc_idx;=0A= >>> + int nb_free =3D 0;=0A= >>> + struct rte_mbuf *m, *free[VIRTIO_TX_MAX_FREE_BUF_SZ];=0A= >>> +=0A= >>> + desc_idx =3D (uint16_t)(vq->vq_used_cons_idx &=0A= >>> + ((vq->vq_nentries >> 1) - 1));=0A= >>> + free[0] =3D (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;=0A= >>> + nb_free =3D 1;=0A= >>> +=0A= >>> + for (i =3D 1; i < VIRTIO_TX_FREE_NR; i++) {=0A= >>> + m =3D (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;=0A= >>> + if (likely(m->pool =3D=3D free[0]->pool))=0A= >>> + free[nb_free++] =3D m;=0A= >>> + else {=0A= >>> + rte_mempool_put_bulk(free[0]->pool, (void **)free,=0A= >>> + nb_free);=0A= >>> + free[0] =3D m;=0A= >>> + nb_free =3D 1;=0A= >>> + }=0A= >>> + }=0A= >>> +=0A= >>> + rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);=0A= >>> + vq->vq_used_cons_idx +=3D VIRTIO_TX_FREE_NR;=0A= >>> + vq->vq_free_cnt +=3D (VIRTIO_TX_FREE_NR << 1);=0A= >>> +}=0A= >> I think you need to handle refcount, here is a similar patch=0A= >> for ixgbe.=0A= > ok, like this:=0A= >=0A= > m =3D (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;=0A= Missed a line=0A= m =3D __rte_pktmbuf_prefree_seg(m)=0A= > if (likely(m !=3D NULL)) {=0A= > ...=0A= >=0A= >> Subject: ixgbe: speed up transmit=0A= >>=0A= >> Coalesce transmit buffers and put them back into the pool=0A= >> in one burst.=0A= >>=0A= >> Signed-off-by: Stephen Hemminger =0A= >>=0A= >> --- a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c=0A= >> +++ b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c=0A= >> @@ -120,12 +120,16 @@ rte_rxmbuf_alloc(struct rte_mempool *mp)=0A= >> * Check for descriptors with their DD bit set and free mbufs.=0A= >> * Return the total number of buffers freed.=0A= >> */=0A= >> +#define TX_FREE_BULK 32=0A= >> +=0A= >> static inline int __attribute__((always_inline))=0A= >> ixgbe_tx_free_bufs(struct ixgbe_tx_queue *txq)=0A= >> {=0A= >> struct ixgbe_tx_entry *txep;=0A= >> uint32_t status;=0A= >> - int i;=0A= >> + int i, n =3D 0;=0A= >> + struct rte_mempool *txpool =3D NULL;=0A= >> + struct rte_mbuf *free_list[TX_FREE_BULK];=0A= >> =0A= >> /* check DD bit on threshold descriptor */=0A= >> status =3D txq->tx_ring[txq->tx_next_dd].wb.status;=0A= >> @@ -138,20 +142,26 @@ ixgbe_tx_free_bufs(struct ixgbe_tx_queue=0A= >> */=0A= >> txep =3D &(txq->sw_ring[txq->tx_next_dd - (txq->tx_rs_thresh - 1)]);= =0A= >> =0A= >> - /* free buffers one at a time */=0A= >> - if ((txq->txq_flags & (uint32_t)ETH_TXQ_FLAGS_NOREFCOUNT) !=3D 0) {=0A= >> - for (i =3D 0; i < txq->tx_rs_thresh; ++i, ++txep) {=0A= >> - txep->mbuf->next =3D NULL;=0A= >> - rte_mempool_put(txep->mbuf->pool, txep->mbuf);=0A= >> - txep->mbuf =3D NULL;=0A= >> - }=0A= >> - } else {=0A= >> - for (i =3D 0; i < txq->tx_rs_thresh; ++i, ++txep) {=0A= >> - rte_pktmbuf_free_seg(txep->mbuf);=0A= >> - txep->mbuf =3D NULL;=0A= >> + for (i =3D 0; i < txq->tx_rs_thresh; ++i, ++txep) {=0A= >> + struct rte_mbuf *m;=0A= >> +=0A= >> + /* free buffers one at a time */=0A= >> + m =3D __rte_pktmbuf_prefree_seg(txep->mbuf);=0A= >> + txep->mbuf =3D NULL;=0A= >> +=0A= >> + if (n >=3D TX_FREE_BULK ||=0A= > check whether m is NULL here.=0A= >> + (n > 0 && m->pool !=3D txpool)) {=0A= >> + rte_mempool_put_bulk(txpool, (void **)free_list, n);=0A= >> + n =3D 0;=0A= >> }=0A= >> +=0A= >> + txpool =3D m->pool;=0A= >> + free_list[n++] =3D m;=0A= >> }=0A= >> =0A= >> + if (n > 0)=0A= >> + rte_mempool_put_bulk(txpool, (void **)free_list, n);=0A= >> +=0A= >> /* buffers were freed, update counters */=0A= >> txq->nb_tx_free =3D (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);= =0A= >> txq->tx_next_dd =3D (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);= =0A= >>=0A= >>=0A= >=0A= =0A=