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 1FF3E43A51; Fri, 2 Feb 2024 16:09:37 +0100 (CET) Received: from mails.dpdk.org (localhost [127.0.0.1]) by mails.dpdk.org (Postfix) with ESMTP id 0674F42E3D; Fri, 2 Feb 2024 16:09:37 +0100 (CET) Received: from mx0b-0016f401.pphosted.com (mx0b-0016f401.pphosted.com [67.231.156.173]) by mails.dpdk.org (Postfix) with ESMTP id 270F6402AB for ; Fri, 2 Feb 2024 16:09:36 +0100 (CET) Received: from pps.filterd (m0045851.ppops.net [127.0.0.1]) by mx0b-0016f401.pphosted.com (8.17.1.24/8.17.1.24) with ESMTP id 412Ctqwe013876; Fri, 2 Feb 2024 07:09:35 -0800 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=marvell.com; h= from:to:cc:subject:date:message-id:in-reply-to:references :mime-version:content-transfer-encoding:content-type; s= pfpt0220; bh=LnCBdDhArgl7O0n0dAuOXrVWI8B9xm+Hp4JWFyrDuw0=; b=iuL 9meQti0xEzUC2LPlxHf/w536E4cPOF+UYqkyuZxjfkakdzRs83cdkCFa8Zi6uWYz q6kEmvOmsf9A69obgsQ2vAlpIeLtK+VLQiWgTxZXAX4MoDhjUXYsN96A1Ek13AAC I5ZqZrtav3zb0DTlrCpU7RiyHTHk6r1G6i1j9RvkUEP/QJwaxoL7R86sClpHnfUu Q28yYV4TmSWQSjpb9UMHt1g2Kwmx92Bg92Ofy7LMj0+pRRMOeBdzeJsb0zMX6+ui uI5vglX7jaV9Fl0Cl6eMJmaamP5vnL1n/QOJkeyndS7EiikRSOsBi6hn/cdYOqjE urStpgeqej63MPegw6w== Received: from dc5-exch02.marvell.com ([199.233.59.182]) by mx0b-0016f401.pphosted.com (PPS) with ESMTPS id 3w10turc0m-1 (version=TLSv1.2 cipher=ECDHE-RSA-AES256-SHA384 bits=256 verify=NOT); Fri, 02 Feb 2024 07:09:34 -0800 (PST) Received: from DC5-EXCH01.marvell.com (10.69.176.38) by DC5-EXCH02.marvell.com (10.69.176.39) with Microsoft SMTP Server (TLS) id 15.0.1497.48; Fri, 2 Feb 2024 07:09:32 -0800 Received: from maili.marvell.com (10.69.176.80) by DC5-EXCH01.marvell.com (10.69.176.38) with Microsoft SMTP Server id 15.0.1497.48 via Frontend Transport; Fri, 2 Feb 2024 07:09:32 -0800 Received: from MININT-80QBFE8.corp.innovium.com (MININT-80QBFE8.marvell.com [10.28.164.106]) by maili.marvell.com (Postfix) with ESMTP id 5F14E3F7085; Fri, 2 Feb 2024 07:09:30 -0800 (PST) From: To: , Vamsi Attunuru , "Bruce Richardson" , Konstantin Ananyev CC: , Pavan Nikhilesh Subject: [PATCH v6 1/2] net/octeon_ep: improve Rx performance Date: Fri, 2 Feb 2024 20:39:25 +0530 Message-ID: <20240202150927.15846-1-pbhagavatula@marvell.com> X-Mailer: git-send-email 2.25.1 In-Reply-To: <20240202150638.15222-1-pbhagavatula@marvell.com> References: <20240202150638.15222-1-pbhagavatula@marvell.com> MIME-Version: 1.0 Content-Transfer-Encoding: 8bit Content-Type: text/plain X-Proofpoint-GUID: ZMbZercotuYEJnLoXiuyAo27zrLczTXZ X-Proofpoint-ORIG-GUID: ZMbZercotuYEJnLoXiuyAo27zrLczTXZ X-Proofpoint-Virus-Version: vendor=baseguard engine=ICAP:2.0.272,Aquarius:18.0.1011,Hydra:6.0.619,FMLib:17.11.176.26 definitions=2024-02-02_08,2024-01-31_01,2023-05-22_02 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 From: Pavan Nikhilesh Use mempool API instead of pktmbuf alloc to avoid mbuf reset as it will be done by rearm on receive. Reorder refill to avoid unnecessary write commits on mbuf data. Signed-off-by: Pavan Nikhilesh --- v2 Changes: - Fix compilation with distro gcc. v3 Changes: - Fix aarch32 compilation. v4 Changes: - Fix checkpatch. v5 Changes: - Update release notes. v6 Changes: - Fix checkpatch again. doc/guides/rel_notes/release_24_03.rst | 2 ++ drivers/net/octeon_ep/cnxk_ep_rx.c | 4 +-- drivers/net/octeon_ep/cnxk_ep_rx.h | 13 ++++++--- drivers/net/octeon_ep/cnxk_ep_rx_avx.c | 20 +++++++------- drivers/net/octeon_ep/cnxk_ep_rx_sse.c | 38 ++++++++++++++------------ drivers/net/octeon_ep/otx_ep_rxtx.h | 2 +- 6 files changed, 44 insertions(+), 35 deletions(-) diff --git a/doc/guides/rel_notes/release_24_03.rst b/doc/guides/rel_notes/release_24_03.rst index 282a3f9c8c..c8fcaaad6d 100644 --- a/doc/guides/rel_notes/release_24_03.rst +++ b/doc/guides/rel_notes/release_24_03.rst @@ -59,6 +59,8 @@ New Features * Optimize mbuf rearm sequence. * Updated Tx queue mbuf free thresholds from 128 to 256 for better performance. + * Updated Rx queue mbuf refill routine to use mempool alloc and reorder it + to avoid mbuf write commits. * Added optimized SSE Rx routines. * Added optimized AVX2 Rx routines. diff --git a/drivers/net/octeon_ep/cnxk_ep_rx.c b/drivers/net/octeon_ep/cnxk_ep_rx.c index f3e4fb27d1..7465e0a017 100644 --- a/drivers/net/octeon_ep/cnxk_ep_rx.c +++ b/drivers/net/octeon_ep/cnxk_ep_rx.c @@ -76,12 +76,12 @@ cnxk_ep_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts) uint16_t new_pkts; new_pkts = cnxk_ep_rx_pkts_to_process(droq, nb_pkts); - cnxk_ep_process_pkts_scalar(rx_pkts, droq, new_pkts); - /* Refill RX buffers */ if (droq->refill_count >= DROQ_REFILL_THRESHOLD) cnxk_ep_rx_refill(droq); + cnxk_ep_process_pkts_scalar(rx_pkts, droq, new_pkts); + return new_pkts; } diff --git a/drivers/net/octeon_ep/cnxk_ep_rx.h b/drivers/net/octeon_ep/cnxk_ep_rx.h index e71fc0de5c..61263e651e 100644 --- a/drivers/net/octeon_ep/cnxk_ep_rx.h +++ b/drivers/net/octeon_ep/cnxk_ep_rx.h @@ -21,13 +21,16 @@ cnxk_ep_rx_refill_mbuf(struct otx_ep_droq *droq, uint32_t count) uint32_t i; int rc; - rc = rte_pktmbuf_alloc_bulk(droq->mpool, &recv_buf_list[refill_idx], count); + rc = rte_mempool_get_bulk(droq->mpool, (void **)&recv_buf_list[refill_idx], count); if (unlikely(rc)) { droq->stats.rx_alloc_failure++; return rc; } for (i = 0; i < count; i++) { + rte_prefetch_non_temporal(&desc_ring[(refill_idx + 1) & 3]); + if (i < count - 1) + rte_prefetch_non_temporal(recv_buf_list[refill_idx + 1]); buf = recv_buf_list[refill_idx]; desc_ring[refill_idx].buffer_ptr = rte_mbuf_data_iova_default(buf); refill_idx++; @@ -42,9 +45,9 @@ cnxk_ep_rx_refill_mbuf(struct otx_ep_droq *droq, uint32_t count) static inline void cnxk_ep_rx_refill(struct otx_ep_droq *droq) { - uint32_t desc_refilled = 0, count; - uint32_t nb_desc = droq->nb_desc; + const uint32_t nb_desc = droq->nb_desc; uint32_t refill_idx = droq->refill_idx; + uint32_t desc_refilled = 0, count; int rc; if (unlikely(droq->read_idx == refill_idx)) @@ -128,6 +131,8 @@ cnxk_ep_rx_pkts_to_process(struct otx_ep_droq *droq, uint16_t nb_pkts) return RTE_MIN(nb_pkts, droq->pkts_pending); } +#define cnxk_pktmbuf_mtod(m, t) ((t)(void *)((char *)(m)->buf_addr + RTE_PKTMBUF_HEADROOM)) + static __rte_always_inline void cnxk_ep_process_pkts_scalar(struct rte_mbuf **rx_pkts, struct otx_ep_droq *droq, uint16_t new_pkts) { @@ -147,7 +152,7 @@ cnxk_ep_process_pkts_scalar(struct rte_mbuf **rx_pkts, struct otx_ep_droq *droq, void *)); mbuf = recv_buf_list[read_idx]; - info = rte_pktmbuf_mtod(mbuf, struct otx_ep_droq_info *); + info = cnxk_pktmbuf_mtod(mbuf, struct otx_ep_droq_info *); read_idx = otx_ep_incr_index(read_idx, 1, nb_desc); pkt_len = rte_bswap16(info->length >> 48); mbuf->pkt_len = pkt_len; diff --git a/drivers/net/octeon_ep/cnxk_ep_rx_avx.c b/drivers/net/octeon_ep/cnxk_ep_rx_avx.c index ae4615e6da..47eb1d2ef7 100644 --- a/drivers/net/octeon_ep/cnxk_ep_rx_avx.c +++ b/drivers/net/octeon_ep/cnxk_ep_rx_avx.c @@ -49,7 +49,7 @@ cnxk_ep_process_pkts_vec_avx(struct rte_mbuf **rx_pkts, struct otx_ep_droq *droq /* Load rearm data and packet length for shuffle. */ for (i = 0; i < CNXK_EP_OQ_DESC_PER_LOOP_AVX; i++) data[i] = _mm256_set_epi64x(0, - rte_pktmbuf_mtod(m[i], struct otx_ep_droq_info *)->length >> 16, + cnxk_pktmbuf_mtod(m[i], struct otx_ep_droq_info *)->length >> 16, 0, rearm_data); /* Shuffle data to its place and sum the packet length. */ @@ -81,15 +81,15 @@ cnxk_ep_recv_pkts_avx(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkt struct otx_ep_droq *droq = (struct otx_ep_droq *)rx_queue; uint16_t new_pkts, vpkts; + /* Refill RX buffers */ + if (droq->refill_count >= DROQ_REFILL_THRESHOLD) + cnxk_ep_rx_refill(droq); + new_pkts = cnxk_ep_rx_pkts_to_process(droq, nb_pkts); vpkts = RTE_ALIGN_FLOOR(new_pkts, CNXK_EP_OQ_DESC_PER_LOOP_AVX); cnxk_ep_process_pkts_vec_avx(rx_pkts, droq, vpkts); cnxk_ep_process_pkts_scalar(&rx_pkts[vpkts], droq, new_pkts - vpkts); - /* Refill RX buffers */ - if (droq->refill_count >= DROQ_REFILL_THRESHOLD) - cnxk_ep_rx_refill(droq); - return new_pkts; } @@ -99,11 +99,6 @@ cn9k_ep_recv_pkts_avx(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkt struct otx_ep_droq *droq = (struct otx_ep_droq *)rx_queue; uint16_t new_pkts, vpkts; - new_pkts = cnxk_ep_rx_pkts_to_process(droq, nb_pkts); - vpkts = RTE_ALIGN_FLOOR(new_pkts, CNXK_EP_OQ_DESC_PER_LOOP_AVX); - cnxk_ep_process_pkts_vec_avx(rx_pkts, droq, vpkts); - cnxk_ep_process_pkts_scalar(&rx_pkts[vpkts], droq, new_pkts - vpkts); - /* Refill RX buffers */ if (droq->refill_count >= DROQ_REFILL_THRESHOLD) { cnxk_ep_rx_refill(droq); @@ -119,5 +114,10 @@ cn9k_ep_recv_pkts_avx(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkt rte_write32(0, droq->pkts_credit_reg); } + new_pkts = cnxk_ep_rx_pkts_to_process(droq, nb_pkts); + vpkts = RTE_ALIGN_FLOOR(new_pkts, CNXK_EP_OQ_DESC_PER_LOOP_AVX); + cnxk_ep_process_pkts_vec_avx(rx_pkts, droq, vpkts); + cnxk_ep_process_pkts_scalar(&rx_pkts[vpkts], droq, new_pkts - vpkts); + return new_pkts; } diff --git a/drivers/net/octeon_ep/cnxk_ep_rx_sse.c b/drivers/net/octeon_ep/cnxk_ep_rx_sse.c index 67c0c1c862..308c8b2288 100644 --- a/drivers/net/octeon_ep/cnxk_ep_rx_sse.c +++ b/drivers/net/octeon_ep/cnxk_ep_rx_sse.c @@ -18,13 +18,15 @@ static __rte_always_inline void cnxk_ep_process_pkts_vec_sse(struct rte_mbuf **rx_pkts, struct otx_ep_droq *droq, uint16_t new_pkts) { struct rte_mbuf **recv_buf_list = droq->recv_buf_list; - uint32_t bytes_rsvd = 0, read_idx = droq->read_idx; - uint32_t idx0, idx1, idx2, idx3; + uint32_t read_idx = droq->read_idx; struct rte_mbuf *m0, *m1, *m2, *m3; uint16_t nb_desc = droq->nb_desc; + uint32_t idx0, idx1, idx2, idx3; uint16_t pkts = 0; + __m128i bytes; idx0 = read_idx; + bytes = _mm_setzero_si128(); while (pkts < new_pkts) { const __m128i bswap_mask = _mm_set_epi8(0xFF, 0xFF, 12, 13, 0xFF, 0xFF, 8, 9, 0xFF, 0xFF, 4, 5, 0xFF, 0xFF, 0, 1); @@ -42,14 +44,14 @@ cnxk_ep_process_pkts_vec_sse(struct rte_mbuf **rx_pkts, struct otx_ep_droq *droq m3 = recv_buf_list[idx3]; /* Load packet size big-endian. */ - s01 = _mm_set_epi32(rte_pktmbuf_mtod(m3, struct otx_ep_droq_info *)->length >> 48, - rte_pktmbuf_mtod(m1, struct otx_ep_droq_info *)->length >> 48, - rte_pktmbuf_mtod(m2, struct otx_ep_droq_info *)->length >> 48, - rte_pktmbuf_mtod(m0, struct otx_ep_droq_info *)->length >> 48); + s01 = _mm_set_epi32(cnxk_pktmbuf_mtod(m3, struct otx_ep_droq_info *)->length >> 48, + cnxk_pktmbuf_mtod(m1, struct otx_ep_droq_info *)->length >> 48, + cnxk_pktmbuf_mtod(m2, struct otx_ep_droq_info *)->length >> 48, + cnxk_pktmbuf_mtod(m0, struct otx_ep_droq_info *)->length >> 48); /* Convert to little-endian. */ s01 = _mm_shuffle_epi8(s01, bswap_mask); - /* Horizontal add. */ - bytes_rsvd += hadd(s01); + /* Vertical add, consolidate outside loop */ + bytes = _mm_add_epi32(bytes, s01); /* Segregate to packet length and data length. */ s23 = _mm_shuffle_epi32(s01, _MM_SHUFFLE(3, 3, 1, 1)); s01 = _mm_shuffle_epi8(s01, cpy_mask); @@ -79,7 +81,7 @@ cnxk_ep_process_pkts_vec_sse(struct rte_mbuf **rx_pkts, struct otx_ep_droq *droq droq->pkts_pending -= new_pkts; /* Stats */ droq->stats.pkts_received += new_pkts; - droq->stats.bytes_received += bytes_rsvd; + droq->stats.bytes_received += hadd(bytes); } uint16_t __rte_noinline __rte_hot @@ -88,15 +90,15 @@ cnxk_ep_recv_pkts_sse(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkt struct otx_ep_droq *droq = (struct otx_ep_droq *)rx_queue; uint16_t new_pkts, vpkts; + /* Refill RX buffers */ + if (droq->refill_count >= DROQ_REFILL_THRESHOLD) + cnxk_ep_rx_refill(droq); + new_pkts = cnxk_ep_rx_pkts_to_process(droq, nb_pkts); vpkts = RTE_ALIGN_FLOOR(new_pkts, CNXK_EP_OQ_DESC_PER_LOOP_SSE); cnxk_ep_process_pkts_vec_sse(rx_pkts, droq, vpkts); cnxk_ep_process_pkts_scalar(&rx_pkts[vpkts], droq, new_pkts - vpkts); - /* Refill RX buffers */ - if (droq->refill_count >= DROQ_REFILL_THRESHOLD) - cnxk_ep_rx_refill(droq); - return new_pkts; } @@ -106,11 +108,6 @@ cn9k_ep_recv_pkts_sse(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkt struct otx_ep_droq *droq = (struct otx_ep_droq *)rx_queue; uint16_t new_pkts, vpkts; - new_pkts = cnxk_ep_rx_pkts_to_process(droq, nb_pkts); - vpkts = RTE_ALIGN_FLOOR(new_pkts, CNXK_EP_OQ_DESC_PER_LOOP_SSE); - cnxk_ep_process_pkts_vec_sse(rx_pkts, droq, vpkts); - cnxk_ep_process_pkts_scalar(&rx_pkts[vpkts], droq, new_pkts - vpkts); - /* Refill RX buffers */ if (droq->refill_count >= DROQ_REFILL_THRESHOLD) { cnxk_ep_rx_refill(droq); @@ -126,5 +123,10 @@ cn9k_ep_recv_pkts_sse(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkt rte_write32(0, droq->pkts_credit_reg); } + new_pkts = cnxk_ep_rx_pkts_to_process(droq, nb_pkts); + vpkts = RTE_ALIGN_FLOOR(new_pkts, CNXK_EP_OQ_DESC_PER_LOOP_SSE); + cnxk_ep_process_pkts_vec_sse(rx_pkts, droq, vpkts); + cnxk_ep_process_pkts_scalar(&rx_pkts[vpkts], droq, new_pkts - vpkts); + return new_pkts; } diff --git a/drivers/net/octeon_ep/otx_ep_rxtx.h b/drivers/net/octeon_ep/otx_ep_rxtx.h index 0adcbc7814..8f306bd94e 100644 --- a/drivers/net/octeon_ep/otx_ep_rxtx.h +++ b/drivers/net/octeon_ep/otx_ep_rxtx.h @@ -21,7 +21,7 @@ /* SDP_LENGTH_S specifies packet length and is of 8-byte size */ #define OTX_EP_INFO_SIZE 8 -#define DROQ_REFILL_THRESHOLD 16 +#define DROQ_REFILL_THRESHOLD 64 #define OTX2_SDP_REQUEST_ISM (0x1ULL << 63) static inline uint32_t -- 2.25.1