From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mails.dpdk.org (mails.dpdk.org [217.70.189.124]) by inbox.dpdk.org (Postfix) with ESMTP id 7F814A00C2; Mon, 22 Aug 2022 05:48:23 +0200 (CEST) Received: from [217.70.189.124] (localhost [127.0.0.1]) by mails.dpdk.org (Postfix) with ESMTP id 7F671427FF; Mon, 22 Aug 2022 05:48:15 +0200 (CEST) Received: from foss.arm.com (foss.arm.com [217.140.110.172]) by mails.dpdk.org (Postfix) with ESMTP id AF1EC42B6E for ; Mon, 22 Aug 2022 05:48:13 +0200 (CEST) Received: from usa-sjc-imap-foss1.foss.arm.com (unknown [10.121.207.14]) by usa-sjc-mx-foss1.foss.arm.com (Postfix) with ESMTP id 33F9B1E2F; Sun, 21 Aug 2022 20:48:16 -0700 (PDT) Received: from net-arm-n1amp-02.shanghai.arm.com (net-arm-n1amp-02.shanghai.arm.com [10.169.210.108]) by usa-sjc-imap-foss1.foss.arm.com (Postfix) with ESMTPA id 9D0283F718; Sun, 21 Aug 2022 20:48:10 -0700 (PDT) From: Joyce Kong To: jgrajcia@cisco.com, stephen@networkplumber.org, huzaifa.rahman@emumba.com Cc: dev@dpdk.org, nd@arm.com, mb@smartsharesystems.com, ruifeng.wang@arm.com, Joyce Kong Subject: [PATCH v3 2/2] net/memif: add a Tx fast path Date: Mon, 22 Aug 2022 03:47:31 +0000 Message-Id: <20220822034731.528424-3-joyce.kong@arm.com> X-Mailer: git-send-email 2.25.1 In-Reply-To: <20220822034731.528424-1-joyce.kong@arm.com> References: <20220701102815.1444223-2-joyce.kong@arm.com> <20220822034731.528424-1-joyce.kong@arm.com> MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org For memif non-zero-copy mode, there is a branch to compare the mbuf and memif buffer size during memory copying. If all mbufs come from the same mempool, and memif buf size >= mbuf size, add a fast Tx memory copy path without the comparing branch and with mbuf bulk free, otherwise still run the original Tx path. The removal of the branch and bulk free lead to considerable performance uplift. Test with 1p1q on N1SDP AArch64 server, -------------------------------------------- buf size | memif >= mbuf | memif < mbuf | -------------------------------------------- non-zc gain | 10.82% | 0.04% | -------------------------------------------- zc gain | 8.86% | 3.18% | -------------------------------------------- Test with 1p1q on Cascade Lake Xeon X86 server, -------------------------------------------- buf size | memif >= mbuf | memif < mbuf | -------------------------------------------- non-zc gain | 7.32% | -0.85% | -------------------------------------------- zc gain | 12.75% | -0.16% | -------------------------------------------- Signed-off-by: Joyce Kong Reviewed-by: Ruifeng Wang Acked-by: Morten Brørup --- drivers/net/memif/rte_eth_memif.c | 134 ++++++++++++++++++++---------- 1 file changed, 92 insertions(+), 42 deletions(-) diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c index 2ea2a8e266..b0b5abef5a 100644 --- a/drivers/net/memif/rte_eth_memif.c +++ b/drivers/net/memif/rte_eth_memif.c @@ -671,62 +671,112 @@ eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts) n_free = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE) - slot; } - while (n_tx_pkts < nb_pkts && n_free) { - mbuf_head = *bufs++; - nb_segs = mbuf_head->nb_segs; - mbuf = mbuf_head; + uint8_t i; + struct rte_mbuf **buf_tmp = bufs; + mbuf_head = *buf_tmp++; + struct rte_mempool *mp = mbuf_head->pool; + + for (i = 1; i < nb_pkts; i++) { + mbuf_head = *buf_tmp++; + if (mbuf_head->pool != mp) + break; + } + + uint16_t mbuf_size = rte_pktmbuf_data_room_size(mp) - RTE_PKTMBUF_HEADROOM; + if (i == nb_pkts && pmd->cfg.pkt_buffer_size >= mbuf_size) { + buf_tmp = bufs; + while (n_tx_pkts < nb_pkts && n_free) { + mbuf_head = *bufs++; + nb_segs = mbuf_head->nb_segs; + mbuf = mbuf_head; - saved_slot = slot; - d0 = &ring->desc[slot & mask]; - dst_off = 0; - dst_len = (type == MEMIF_RING_C2S) ? - pmd->run.pkt_buffer_size : d0->length; + saved_slot = slot; -next_in_chain: - src_off = 0; - src_len = rte_pktmbuf_data_len(mbuf); +next_in_chain1: + d0 = &ring->desc[slot & mask]; + cp_len = rte_pktmbuf_data_len(mbuf); - while (src_len) { - if (dst_len == 0) { + rte_memcpy((uint8_t *)memif_get_buffer(proc_private, d0), + rte_pktmbuf_mtod(mbuf, void *), cp_len); + + d0->length = cp_len; + mq->n_bytes += cp_len; + slot++; + n_free--; + + if (--nb_segs > 0) { if (n_free) { - slot++; - n_free--; d0->flags |= MEMIF_DESC_FLAG_NEXT; - d0 = &ring->desc[slot & mask]; - dst_off = 0; - dst_len = (type == MEMIF_RING_C2S) ? - pmd->run.pkt_buffer_size : d0->length; - d0->flags = 0; + mbuf = mbuf->next; + goto next_in_chain1; } else { slot = saved_slot; - goto no_free_slots; + goto free_mbufs; } } - cp_len = RTE_MIN(dst_len, src_len); - rte_memcpy((uint8_t *)memif_get_buffer(proc_private, - d0) + dst_off, - rte_pktmbuf_mtod_offset(mbuf, void *, src_off), - cp_len); + n_tx_pkts++; + } +free_mbufs: + rte_pktmbuf_free_bulk(buf_tmp, n_tx_pkts); + } else { + while (n_tx_pkts < nb_pkts && n_free) { + mbuf_head = *bufs++; + nb_segs = mbuf_head->nb_segs; + mbuf = mbuf_head; - mq->n_bytes += cp_len; - src_off += cp_len; - dst_off += cp_len; - src_len -= cp_len; - dst_len -= cp_len; + saved_slot = slot; + d0 = &ring->desc[slot & mask]; + dst_off = 0; + dst_len = (type == MEMIF_RING_C2S) ? + pmd->run.pkt_buffer_size : d0->length; - d0->length = dst_off; - } +next_in_chain2: + src_off = 0; + src_len = rte_pktmbuf_data_len(mbuf); - if (--nb_segs > 0) { - mbuf = mbuf->next; - goto next_in_chain; - } + while (src_len) { + if (dst_len == 0) { + if (n_free) { + slot++; + n_free--; + d0->flags |= MEMIF_DESC_FLAG_NEXT; + d0 = &ring->desc[slot & mask]; + dst_off = 0; + dst_len = (type == MEMIF_RING_C2S) ? + pmd->run.pkt_buffer_size : d0->length; + d0->flags = 0; + } else { + slot = saved_slot; + goto no_free_slots; + } + } + cp_len = RTE_MIN(dst_len, src_len); - n_tx_pkts++; - slot++; - n_free--; - rte_pktmbuf_free(mbuf_head); + rte_memcpy((uint8_t *)memif_get_buffer(proc_private, + d0) + dst_off, + rte_pktmbuf_mtod_offset(mbuf, void *, src_off), + cp_len); + + mq->n_bytes += cp_len; + src_off += cp_len; + dst_off += cp_len; + src_len -= cp_len; + dst_len -= cp_len; + + d0->length = dst_off; + } + + if (--nb_segs > 0) { + mbuf = mbuf->next; + goto next_in_chain2; + } + + n_tx_pkts++; + slot++; + n_free--; + rte_pktmbuf_free(mbuf_head); + } } no_free_slots: -- 2.25.1