From patchwork Fri Jan 22 17:12:08 2021 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Slava Ovsiienko X-Patchwork-Id: 87102 X-Patchwork-Delegate: rasland@nvidia.com Return-Path: X-Original-To: patchwork@inbox.dpdk.org Delivered-To: patchwork@inbox.dpdk.org Received: from mails.dpdk.org (mails.dpdk.org [217.70.189.124]) by inbox.dpdk.org (Postfix) with ESMTP id 3EC74A0A0A; Fri, 22 Jan 2021 18:12:22 +0100 (CET) Received: from [217.70.189.124] (localhost [127.0.0.1]) by mails.dpdk.org (Postfix) with ESMTP id 1024C141040; Fri, 22 Jan 2021 18:12:20 +0100 (CET) Received: from mellanox.co.il (mail-il-dmz.mellanox.com [193.47.165.129]) by mails.dpdk.org (Postfix) with ESMTP id 53F3E14103C for ; Fri, 22 Jan 2021 18:12:18 +0100 (CET) Received: from Internal Mail-Server by MTLPINE1 (envelope-from viacheslavo@nvidia.com) with SMTP; 22 Jan 2021 19:12:13 +0200 Received: from nvidia.com (pegasus12.mtr.labs.mlnx [10.210.17.40]) by labmailer.mlnx (8.13.8/8.13.8) with ESMTP id 10MHCCue025286; Fri, 22 Jan 2021 19:12:13 +0200 From: Viacheslav Ovsiienko To: dev@dpdk.org Cc: rasland@nvidia.com, matan@nvidia.com, orika@nvidia.com, thomas@monjalon.net, akozyrev@nvidia.com, stable@dpdk.org Date: Fri, 22 Jan 2021 17:12:08 +0000 Message-Id: <1611335529-26503-2-git-send-email-viacheslavo@nvidia.com> X-Mailer: git-send-email 1.8.3.1 In-Reply-To: <1611335529-26503-1-git-send-email-viacheslavo@nvidia.com> References: <1608311697-31529-1-git-send-email-viacheslavo@nvidia.com> <1611335529-26503-1-git-send-email-viacheslavo@nvidia.com> Subject: [dpdk-dev] [PATCH v3 1/2] net/mlx5: optimize inline mbuf freeing 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 Sender: "dev" The mlx5 PMD supports packet data inlining by pushing data to the transmit descriptor. If packet is short enough and all data are inline, the mbuf is not needed for data send anymore and can be freed. The mbuf free was performed in the most inner loop building the transmit descriptors. This patch postpones the mbuf free transaction to the tx_burst routine exit, optimizing the loop and allowing the bulk freeing for the multiple mbufs in single pool API call. Cc: stable@dpdk.org Signed-off-by: Viacheslav Ovsiienko --- drivers/net/mlx5/mlx5_rxtx.c | 38 ++++++++++++++++++++++++++++++++++---- drivers/net/mlx5/mlx5_rxtx.h | 1 + 2 files changed, 35 insertions(+), 4 deletions(-) diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c index 3497765..97912dd 100644 --- a/drivers/net/mlx5/mlx5_rxtx.c +++ b/drivers/net/mlx5/mlx5_rxtx.c @@ -1990,6 +1990,17 @@ enum mlx5_txcmp_code { } } } +/* + * No inline version to free buffers for optimal call + * on the tx_burst completion. + */ +static __rte_noinline void +__mlx5_tx_free_mbuf(struct rte_mbuf **__rte_restrict pkts, + unsigned int pkts_n, + unsigned int olx __rte_unused) +{ + mlx5_tx_free_mbuf(pkts, pkts_n, olx); +} /** * Free the mbuf from the elts ring buffer till new tail. @@ -4408,10 +4419,25 @@ enum mlx5_txcmp_code { MLX5_ASSERT(room >= tlen); room -= tlen; /* - * Packet data are completely inlined, - * free the packet immediately. + * Packet data are completely inline, + * we can try to free the packet. + */ + if (likely(loc->pkts_sent == loc->mbuf_free)) { + /* + * All the packets from the burst beginning + * are inline, we can free mbufs directly + * from the origin array on tx_burst exit(). + */ + loc->mbuf_free++; + goto next_mbuf; + } + /* + * In order no to call rte_pktmbuf_free_seg() here, + * in the most inner loop (that might be very + * expensive) we just save the mbuf in elts. */ - rte_pktmbuf_free_seg(loc->mbuf); + txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf; + loc->elts_free--; goto next_mbuf; pointer_empw: /* @@ -4433,6 +4459,7 @@ enum mlx5_txcmp_code { mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx); /* We have to store mbuf in elts.*/ txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf; + loc->elts_free--; room -= MLX5_WQE_DSEG_SIZE; /* Ring buffer wraparound is checked at the loop end.*/ ++dseg; @@ -4442,7 +4469,6 @@ enum mlx5_txcmp_code { slen += dlen; #endif loc->pkts_sent++; - loc->elts_free--; pkts_n--; if (unlikely(!pkts_n || !loc->elts_free)) { /* @@ -4892,6 +4918,8 @@ enum mlx5_txcmp_code { MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi)); if (unlikely(!pkts_n)) return 0; + if (MLX5_TXOFF_CONFIG(INLINE)) + loc.mbuf_free = 0; loc.pkts_sent = 0; loc.pkts_copy = 0; loc.wqe_last = NULL; @@ -5155,6 +5183,8 @@ enum mlx5_txcmp_code { /* Increment sent packets counter. */ txq->stats.opackets += loc.pkts_sent; #endif + if (MLX5_TXOFF_CONFIG(INLINE) && loc.mbuf_free) + __mlx5_tx_free_mbuf(pkts, loc.mbuf_free, olx); return loc.pkts_sent; } diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h index 7756ed3..9dac408 100644 --- a/drivers/net/mlx5/mlx5_rxtx.h +++ b/drivers/net/mlx5/mlx5_rxtx.h @@ -209,6 +209,7 @@ struct mlx5_txq_local { uint16_t wqe_free; /* available wqe remain. */ uint16_t mbuf_off; /* data offset in current mbuf. */ uint16_t mbuf_nseg; /* number of remaining mbuf. */ + uint16_t mbuf_free; /* number of inline mbufs to free. */ }; /* TX queue descriptor. */ From patchwork Fri Jan 22 17:12:09 2021 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Slava Ovsiienko X-Patchwork-Id: 87103 X-Patchwork-Delegate: rasland@nvidia.com Return-Path: X-Original-To: patchwork@inbox.dpdk.org Delivered-To: patchwork@inbox.dpdk.org Received: from mails.dpdk.org (mails.dpdk.org [217.70.189.124]) by inbox.dpdk.org (Postfix) with ESMTP id A809CA0A0A; Fri, 22 Jan 2021 18:12:30 +0100 (CET) Received: from [217.70.189.124] (localhost [127.0.0.1]) by mails.dpdk.org (Postfix) with ESMTP id 45B0B141047; Fri, 22 Jan 2021 18:12:21 +0100 (CET) Received: from mellanox.co.il (mail-il-dmz.mellanox.com [193.47.165.129]) by mails.dpdk.org (Postfix) with ESMTP id 5893314103D for ; Fri, 22 Jan 2021 18:12:18 +0100 (CET) Received: from Internal Mail-Server by MTLPINE1 (envelope-from viacheslavo@nvidia.com) with SMTP; 22 Jan 2021 19:12:14 +0200 Received: from nvidia.com (pegasus12.mtr.labs.mlnx [10.210.17.40]) by labmailer.mlnx (8.13.8/8.13.8) with ESMTP id 10MHCCuf025286; Fri, 22 Jan 2021 19:12:14 +0200 From: Viacheslav Ovsiienko To: dev@dpdk.org Cc: rasland@nvidia.com, matan@nvidia.com, orika@nvidia.com, thomas@monjalon.net, akozyrev@nvidia.com Date: Fri, 22 Jan 2021 17:12:09 +0000 Message-Id: <1611335529-26503-3-git-send-email-viacheslavo@nvidia.com> X-Mailer: git-send-email 1.8.3.1 In-Reply-To: <1611335529-26503-1-git-send-email-viacheslavo@nvidia.com> References: <1608311697-31529-1-git-send-email-viacheslavo@nvidia.com> <1611335529-26503-1-git-send-email-viacheslavo@nvidia.com> Subject: [dpdk-dev] [PATCH v3 2/2] net/mlx5: add mbuf fast free offload support 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 Sender: "dev" This patch adds support of the mbuf fast free offload to the transmit datapath. This offload allows freeing the mbufs on transmit completion in the most efficient way. It requires the all mbufs were allocated from the same pool, have the reference counter value as 1, and have no any externally attached buffers. Signed-off-by: Viacheslav Ovsiienko --- doc/guides/nics/features/mlx5.ini | 1 + doc/guides/nics/mlx5.rst | 12 ++++++++++++ doc/guides/rel_notes/release_21_02.rst | 2 +- drivers/net/mlx5/mlx5_rxtx.c | 35 ++++++++++++++++++++++++++-------- drivers/net/mlx5/mlx5_rxtx.h | 1 + drivers/net/mlx5/mlx5_txq.c | 6 ++++++ 6 files changed, 48 insertions(+), 9 deletions(-) diff --git a/doc/guides/nics/features/mlx5.ini b/doc/guides/nics/features/mlx5.ini index 9d14678..a997543 100644 --- a/doc/guides/nics/features/mlx5.ini +++ b/doc/guides/nics/features/mlx5.ini @@ -9,6 +9,7 @@ Link status = Y Link status event = Y Removal event = Y Rx interrupt = Y +Fast mbuf free = Y Queue start/stop = Y Burst mode info = Y MTU update = Y diff --git a/doc/guides/nics/mlx5.rst b/doc/guides/nics/mlx5.rst index 1765952..e5f8f3f 100644 --- a/doc/guides/nics/mlx5.rst +++ b/doc/guides/nics/mlx5.rst @@ -74,6 +74,7 @@ Features - RX VLAN stripping. - TX VLAN insertion. - RX CRC stripping configuration. +- TX mbuf fast free offload. - Promiscuous mode on PF and VF. - Multicast promiscuous mode on PF and VF. - Hardware checksum offloads. @@ -353,6 +354,17 @@ Limitations for some NICs (such as ConnectX-6 Dx, ConnectX-6 Lx, and BlueField-2). The capability bit ``scatter_fcs_w_decap_disable`` shows NIC support. +- TX mbuf fast free: + + - fast free offload assumes the all mbufs being sent are originated from the + same memory pool and there is no any extra references to the mbufs (the + reference counter for each mbuf is equal 1 on tx_burst call). The latter + means there should be no any externally attached buffers in mbufs. It is + an application responsibility to provide the correct mbufs if the fast + free offload is engaged. The mlx5 PMD implicitly produces the mbufs with + externally attached buffers if MPRQ option is enabled, hence, the fast + free offload is neither supported nor advertised if there is MPRQ enabled. + - Sample flow: - Supports ``RTE_FLOW_ACTION_TYPE_SAMPLE`` action only within NIC Rx and E-Switch steering domain. diff --git a/doc/guides/rel_notes/release_21_02.rst b/doc/guides/rel_notes/release_21_02.rst index f52468c..0410337 100644 --- a/doc/guides/rel_notes/release_21_02.rst +++ b/doc/guides/rel_notes/release_21_02.rst @@ -98,6 +98,7 @@ New Features mirroring action on E-Switch. * Enlarge the number of flow priorities to 21844(0 - 21843) for ingress or egress flow groups greater than 0 and for any transfer flow group. + * Add support for the Tx mbuf fast free offload. * **Updated Wangxun txgbe driver.** @@ -127,7 +128,6 @@ New Features * Added support for aes-cbc sha256-128-hmac cipher combination in OCTEON TX2 crypto PMD lookaside protocol offload for IPsec. - Removed Items ------------- diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c index 97912dd..2e4b87c 100644 --- a/drivers/net/mlx5/mlx5_rxtx.c +++ b/drivers/net/mlx5/mlx5_rxtx.c @@ -1889,6 +1889,8 @@ enum mlx5_txcmp_code { /** * Free the mbufs from the linear array of pointers. * + * @param txq + * Pointer to Tx queue structure. * @param pkts * Pointer to array of packets to be free. * @param pkts_n @@ -1898,7 +1900,8 @@ enum mlx5_txcmp_code { * compile time and may be used for optimization. */ static __rte_always_inline void -mlx5_tx_free_mbuf(struct rte_mbuf **__rte_restrict pkts, +mlx5_tx_free_mbuf(struct mlx5_txq_data *__rte_restrict txq, + struct rte_mbuf **__rte_restrict pkts, unsigned int pkts_n, unsigned int olx __rte_unused) { @@ -1914,6 +1917,16 @@ enum mlx5_txcmp_code { */ MLX5_ASSERT(pkts); MLX5_ASSERT(pkts_n); + /* + * Free mbufs directly to the pool in bulk + * if fast free offload is engaged + */ + if (!MLX5_TXOFF_CONFIG(MULTI) && txq->fast_free) { + mbuf = *pkts; + pool = mbuf->pool; + rte_mempool_put_bulk(pool, (void *)pkts, pkts_n); + return; + } for (;;) { for (;;) { /* @@ -1995,11 +2008,12 @@ enum mlx5_txcmp_code { * on the tx_burst completion. */ static __rte_noinline void -__mlx5_tx_free_mbuf(struct rte_mbuf **__rte_restrict pkts, +__mlx5_tx_free_mbuf(struct mlx5_txq_data *__rte_restrict txq, + struct rte_mbuf **__rte_restrict pkts, unsigned int pkts_n, unsigned int olx __rte_unused) { - mlx5_tx_free_mbuf(pkts, pkts_n, olx); + mlx5_tx_free_mbuf(txq, pkts, pkts_n, olx); } /** @@ -2033,7 +2047,8 @@ enum mlx5_txcmp_code { part = RTE_MIN(part, n_elts); MLX5_ASSERT(part); MLX5_ASSERT(part <= txq->elts_s); - mlx5_tx_free_mbuf(&txq->elts[txq->elts_tail & txq->elts_m], + mlx5_tx_free_mbuf(txq, + &txq->elts[txq->elts_tail & txq->elts_m], part, olx); txq->elts_tail += part; n_elts -= part; @@ -5184,7 +5199,7 @@ enum mlx5_txcmp_code { txq->stats.opackets += loc.pkts_sent; #endif if (MLX5_TXOFF_CONFIG(INLINE) && loc.mbuf_free) - __mlx5_tx_free_mbuf(pkts, loc.mbuf_free, olx); + __mlx5_tx_free_mbuf(txq, pkts, loc.mbuf_free, olx); return loc.pkts_sent; } @@ -5829,17 +5844,19 @@ enum mlx5_txcmp_code { int mlx5_tx_burst_mode_get(struct rte_eth_dev *dev, - uint16_t tx_queue_id __rte_unused, + uint16_t tx_queue_id, struct rte_eth_burst_mode *mode) { eth_tx_burst_t pkt_burst = dev->tx_pkt_burst; + struct mlx5_priv *priv = dev->data->dev_private; + struct mlx5_txq_data *txq = (*priv->txqs)[tx_queue_id]; unsigned int i, olx; for (i = 0; i < RTE_DIM(txoff_func); i++) { if (pkt_burst == txoff_func[i].func) { olx = txoff_func[i].olx; snprintf(mode->info, sizeof(mode->info), - "%s%s%s%s%s%s%s%s%s", + "%s%s%s%s%s%s%s%s%s%s", (olx & MLX5_TXOFF_CONFIG_EMPW) ? ((olx & MLX5_TXOFF_CONFIG_MPW) ? "Legacy MPW" : "Enhanced MPW") : "No MPW", @@ -5858,7 +5875,9 @@ enum mlx5_txcmp_code { (olx & MLX5_TXOFF_CONFIG_METADATA) ? " + METADATA" : "", (olx & MLX5_TXOFF_CONFIG_TXPP) ? - " + TXPP" : ""); + " + TXPP" : "", + (txq && txq->fast_free) ? + " + Fast Free" : ""); return 0; } } diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h index 9dac408..6432030 100644 --- a/drivers/net/mlx5/mlx5_rxtx.h +++ b/drivers/net/mlx5/mlx5_rxtx.h @@ -243,6 +243,7 @@ struct mlx5_txq_data { uint16_t vlan_en:1; /* VLAN insertion in WQE is supported. */ uint16_t db_nc:1; /* Doorbell mapped to non-cached region. */ uint16_t db_heu:1; /* Doorbell heuristic write barrier. */ + uint16_t fast_free:1; /* mbuf fast free on Tx is enabled. */ uint16_t inlen_send; /* Ordinary send data inline size. */ uint16_t inlen_empw; /* eMPW max packet size to inline. */ uint16_t inlen_mode; /* Minimal data length to inline. */ diff --git a/drivers/net/mlx5/mlx5_txq.c b/drivers/net/mlx5/mlx5_txq.c index b81bb4a..bf5e9a9 100644 --- a/drivers/net/mlx5/mlx5_txq.c +++ b/drivers/net/mlx5/mlx5_txq.c @@ -123,6 +123,8 @@ DEV_TX_OFFLOAD_GRE_TNL_TSO | DEV_TX_OFFLOAD_GENEVE_TNL_TSO); } + if (!config->mprq.enabled) + offloads |= DEV_TX_OFFLOAD_MBUF_FAST_FREE; return offloads; } @@ -800,6 +802,10 @@ bool vlan_inline; unsigned int temp; + txq_ctrl->txq.fast_free = + !!((txq_ctrl->txq.offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE) && + !(txq_ctrl->txq.offloads & DEV_TX_OFFLOAD_MULTI_SEGS) && + !config->mprq.enabled); if (config->txqs_inline == MLX5_ARG_UNSET) txqs_inline = #if defined(RTE_ARCH_ARM64)