From patchwork Wed Sep 2 11:44:25 2015 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Ilya Maximets X-Patchwork-Id: 513501 Return-Path: X-Original-To: incoming@patchwork.ozlabs.org Delivered-To: patchwork-incoming@bilbo.ozlabs.org Received: from archives.nicira.com (unknown [IPv6:2600:3c00::f03c:91ff:fe6e:bdf7]) by ozlabs.org (Postfix) with ESMTP id 865DF1401CD for ; Wed, 2 Sep 2015 21:44:41 +1000 (AEST) Received: from archives.nicira.com (localhost [127.0.0.1]) by archives.nicira.com (Postfix) with ESMTP id 9CD4810A5A; Wed, 2 Sep 2015 04:44:40 -0700 (PDT) X-Original-To: dev@openvswitch.org Delivered-To: dev@openvswitch.org Received: from mx1e4.cudamail.com (mx1.cudamail.com [69.90.118.67]) by archives.nicira.com (Postfix) with ESMTPS id 2FFFF1021D for ; Wed, 2 Sep 2015 04:44:39 -0700 (PDT) Received: from bar2.cudamail.com (unknown [192.168.21.12]) by mx1e4.cudamail.com (Postfix) with ESMTPS id 84C111E0578 for ; Wed, 2 Sep 2015 05:44:38 -0600 (MDT) X-ASG-Debug-ID: 1441194277-03dc535341337a0001-byXFYA Received: from mx1-pf1.cudamail.com ([192.168.24.1]) by bar2.cudamail.com with ESMTP id RuBlPxPynQdBdc3E (version=TLSv1 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO) for ; Wed, 02 Sep 2015 05:44:37 -0600 (MDT) X-Barracuda-Envelope-From: i.maximets@samsung.com X-Barracuda-RBL-Trusted-Forwarder: 192.168.24.1 Received: from unknown (HELO mailout2.w1.samsung.com) (210.118.77.12) by mx1-pf1.cudamail.com with ESMTPS (DHE-RSA-AES128-SHA encrypted); 2 Sep 2015 11:44:36 -0000 Received-SPF: none (mx1-pf1.cudamail.com: domain at samsung.com does not designate permitted sender hosts) X-Barracuda-Apparent-Source-IP: 210.118.77.12 X-Barracuda-RBL-IP: 210.118.77.12 Received: from eucpsbgm1.samsung.com (unknown [203.254.199.244]) by mailout2.w1.samsung.com (Oracle Communications Messaging Server 7.0.5.31.0 64bit (built May 5 2014)) with ESMTP id <0NU100LGQSM94Y10@mailout2.w1.samsung.com> for dev@openvswitch.org; Wed, 02 Sep 2015 12:44:33 +0100 (BST) X-AuditID: cbfec7f4-f79c56d0000012ee-7d-55e6e121d16e Received: from eusync4.samsung.com ( [203.254.199.214]) by eucpsbgm1.samsung.com (EUCPMTA) with SMTP id D0.47.04846.121E6E55; Wed, 2 Sep 2015 12:44:33 +0100 (BST) Received: from imaximets.rnd.samsung.ru ([106.109.129.180]) by eusync4.samsung.com (Oracle Communications Messaging Server 7.0.5.31.0 64bit (built May 5 2014)) with ESMTPA id <0NU100KWYSM4OO80@eusync4.samsung.com>; Wed, 02 Sep 2015 12:44:33 +0100 (BST) X-CudaMail-Envelope-Sender: i.maximets@samsung.com From: Ilya Maximets To: dev@openvswitch.org, Daniele Di Proietto X-CudaMail-MID: CM-E1-901010875 X-CudaMail-DTE: 090215 X-CudaMail-Originating-IP: 210.118.77.12 Date: Wed, 02 Sep 2015 14:44:25 +0300 X-ASG-Orig-Subj: [##CM-E1-901010875##][PATCH v4] dpif-netdev: proper tx queue id Message-id: <1441194265-21074-1-git-send-email-i.maximets@samsung.com> X-Mailer: git-send-email 2.1.4 X-Brightmail-Tracker: H4sIAAAAAAAAA+NgFvrEJMWRmVeSWpSXmKPExsVy+t/xa7qKD5+FGqzcrWjxY30nm8XR03uY LV5232O3uNL+k93i781nzBaTZ0s5sHkcOLqdxePZzf+MHn1bVjF6vJv/li2AJYrLJiU1J7Ms tUjfLoEr4/F7j4KXGRU7dr5nb2A87NPFyMkhIWAisX35JyYIW0ziwr31bF2MXBxCAksZJfpn vmGFcFqZJJ5+7WYGqWIT0JE4tfoII4gtIuAqMf32G3aQImaBaYwSdy7NBBslLGAs8eH2bBYQ m0VAVWLVq+9gzbwCbhIzuj4wQqyTkzh//CfzBEbuBYwMqxhFU0uTC4qT0nMN9YoTc4tL89L1 kvNzNzFCQuLLDsbFx6wOMQpwMCrx8DYkPgsVYk0sK67MPcQowcGsJMLbcBMoxJuSWFmVWpQf X1Sak1p8iFGag0VJnHfurvchQgLpiSWp2ampBalFMFkmDk6pBkbTXV3rvb+HcXzrjeJl9p/Q 1ntn/aeZEoXWd2d9OJS349VOq6PyC0PiHkZ/k5q61c1DMPtjMtt+0be6HczRrt6SJyfNaYwR a32nPluv11xqsdZj13dByrPUzzm7PGdfYXLn79L+yMI3iR8kbSf7vNLOsLZb1rBywd+o3ROn fzD0f7VgdVp04XIlluKMREMt5qLiRABn2p8pBQIAAA== X-GBUdb-Analysis: 0, 210.118.77.12, Ugly c=0 p=0 Source New X-MessageSniffer-Rules: 0-0-0-26005-c X-Barracuda-Connect: UNKNOWN[192.168.24.1] X-Barracuda-Start-Time: 1441194277 X-Barracuda-Encrypted: DHE-RSA-AES256-SHA X-Barracuda-URL: https://web.cudamail.com:443/cgi-mod/mark.cgi X-Virus-Scanned: by bsmtpd at cudamail.com X-Barracuda-BRTS-Status: 1 X-ASG-Whitelist: EmailCat (corporate) Cc: Ilya Maximets , Dyasly Sergey Subject: [ovs-dev] [PATCH v4] dpif-netdev: proper tx queue id X-BeenThere: dev@openvswitch.org X-Mailman-Version: 2.1.16 Precedence: list List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , MIME-Version: 1.0 Errors-To: dev-bounces@openvswitch.org Sender: "dev" Currently tx_qid is equal to pmd->core_id. This leads to unexpected behavior if pmd-cpu-mask different from '/(0*)(1|3|7)?(f*)/', e.g. if core_ids are not sequential, or doesn't start from 0, or both. Example: starting 2 pmd threads with 1 port, 2 rxqs per port, pmd-cpu-mask = 00000014 and let dev->real_n_txq = 2 It that case pmd_1->tx_qid = 2, pmd_2->tx_qid = 4 and txq_needs_locking = true (if device hasn't ovs_numa_get_n_cores()+1 queues). In that case, after truncating in netdev_dpdk_send__(): 'qid = qid % dev->real_n_txq;' pmd_1: qid = 2 % 2 = 0 pmd_2: qid = 4 % 2 = 0 So, both threads will call dpdk_queue_pkts() with same qid = 0. This is unexpected behavior if there is 2 tx queues in device. Queue #1 will not be used and both threads will lock queue #0 on each send. Fix that by introducing per pmd thread hash map 'tx_queues', where will be stored all available tx queues for that pmd thread with port_no as a key(hash). All tx_qid-s will be unique per port and sequential to prevent described unexpected mapping after truncating. Implemented infrastructure can be used in the future to choose between all tx queues available for that pmd thread. Signed-off-by: Ilya Maximets --- version 4: * fixed distribution of tx queues if multiqueue is not supported version 3: * fixed failing of unit tests by adding tx queues of non pmd devices to non pmd thread. (they haven't been used by any thread) * pmd_flush_tx_queues --> dp_netdev_pmd_detach_tx_queues * function names changed to dp_netdev_* * dp_netdev_pmd_lookup_txq now looks by port_no. * removed unnecessary dp_netdev_lookup_port in dp_execute_cb for OVS_ACTION_ATTR_OUTPUT. * refactoring lib/dpif-netdev.c | 160 +++++++++++++++++++++++++++++++++++++++++++++++------- 1 file changed, 139 insertions(+), 21 deletions(-) diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c index cf5b064..d430bc9 100644 --- a/lib/dpif-netdev.c +++ b/lib/dpif-netdev.c @@ -371,6 +371,13 @@ struct dp_netdev_pmd_cycles { atomic_ullong n[PMD_N_CYCLES]; }; +struct dp_netdev_pmd_txq { + struct cmap_node node; /* In owning dp_netdev_pmd_thread's */ + /* 'tx_queues'. */ + struct dp_netdev_port *port; + int tx_qid; +}; + /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate * the performance overhead of interrupt processing. Therefore netdev can * not implement rx-wait for these devices. dpif-netdev needs to poll @@ -426,8 +433,8 @@ struct dp_netdev_pmd_thread { /* threads on same numa node. */ unsigned core_id; /* CPU core id of this pmd thread. */ int numa_id; /* numa node id of this pmd thread. */ - int tx_qid; /* Queue id used by this pmd thread to - * send packets on all netdevs */ + struct cmap tx_queues; /* Queue ids used by this pmd thread to + * send packets to ports */ /* Only a pmd thread can write on its own 'cycles' and 'stats'. * The main thread keeps 'stats_zero' and 'cycles_zero' as base @@ -469,6 +476,15 @@ static void dp_netdev_input(struct dp_netdev_pmd_thread *, static void dp_netdev_disable_upcall(struct dp_netdev *); void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd); +static void dp_netdev_configure_non_pmd_txqs(struct dp_netdev_pmd_thread *pmd); +static void dp_netdev_pmd_add_txq(struct dp_netdev_pmd_thread *pmd, + struct dp_netdev_port *port, int queue_id); +static void dp_netdev_pmd_del_txq(struct dp_netdev_pmd_thread *pmd, + struct dp_netdev_pmd_txq *txq); +static void dp_netdev_pmd_detach_tx_queues(struct dp_netdev_pmd_thread *pmd); +static struct dp_netdev_pmd_txq * +dp_netdev_pmd_lookup_txq(const struct dp_netdev_pmd_thread *pmd, + odp_port_t port_no); static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp, int index, unsigned core_id, int numa_id); @@ -1050,6 +1066,7 @@ do_add_port(struct dp_netdev *dp, const char *devname, const char *type, struct netdev_saved_flags *sf; struct dp_netdev_port *port; struct netdev *netdev; + struct dp_netdev_pmd_thread *non_pmd; enum netdev_flags flags; const char *open_type; int error; @@ -1126,6 +1143,11 @@ do_add_port(struct dp_netdev *dp, const char *devname, const char *type, ovs_refcount_init(&port->ref_cnt); cmap_insert(&dp->ports, &port->node, hash_port_no(port_no)); + non_pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID); + if (non_pmd) { + dp_netdev_pmd_add_txq(non_pmd, port, ovs_numa_get_n_cores()); + dp_netdev_pmd_unref(non_pmd); + } if (netdev_is_pmd(netdev)) { dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev)); dp_netdev_reload_pmds(dp); @@ -1307,8 +1329,21 @@ static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port) OVS_REQUIRES(dp->port_mutex) { + struct dp_netdev_pmd_thread *non_pmd; + cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no)); seq_change(dp->port_seq); + + non_pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID); + if (non_pmd) { + /* There is only one txq for each port for non pmd thread */ + struct dp_netdev_pmd_txq *txq; + txq = dp_netdev_pmd_lookup_txq(non_pmd, port->port_no); + if (OVS_LIKELY(txq)) + dp_netdev_pmd_del_txq(non_pmd, txq); + dp_netdev_pmd_unref(non_pmd); + } + if (netdev_is_pmd(port->netdev)) { int numa_id = netdev_get_numa_id(port->netdev); @@ -2578,6 +2613,80 @@ dpif_netdev_wait(struct dpif *dpif) seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq); } +static void +dp_netdev_pmd_add_txq(struct dp_netdev_pmd_thread *pmd, + struct dp_netdev_port *port, int queue_id) +{ + if (port_try_ref(port)) { + struct dp_netdev_pmd_txq * txq = xmalloc(sizeof *txq); + txq->port = port; + txq->tx_qid = queue_id; + cmap_insert(&pmd->tx_queues, &txq->node, + hash_port_no(port->port_no)); + } +} + +/* Configures tx_queues for non pmd thread. */ +static void +dp_netdev_configure_non_pmd_txqs(struct dp_netdev_pmd_thread *pmd) +{ + if (!cmap_is_empty(&pmd->tx_queues)) + dp_netdev_pmd_detach_tx_queues(pmd); + + struct dp_netdev_port *port; + CMAP_FOR_EACH (port, node, &pmd->dp->ports) { + dp_netdev_pmd_add_txq(pmd, port, ovs_numa_get_n_cores()); + } +} + +static void +dp_netdev_pmd_del_txq(struct dp_netdev_pmd_thread *pmd, + struct dp_netdev_pmd_txq *txq) +{ + cmap_remove(&pmd->tx_queues, &txq->node, + hash_port_no(txq->port->port_no)); + port_unref(txq->port); + free(txq); +} + +/* Removes all queues from 'tx_queues' of pmd thread. */ +static void +dp_netdev_pmd_detach_tx_queues(struct dp_netdev_pmd_thread *pmd) +{ + struct dp_netdev_pmd_txq *txq; + + CMAP_FOR_EACH (txq, node, &pmd->tx_queues) { + dp_netdev_pmd_del_txq(pmd, txq); + } +} + +static void OVS_UNUSED +dp_netdev_pmd_tx_queues_print(struct dp_netdev_pmd_thread *pmd) +{ + struct dp_netdev_pmd_txq *txq; + + CMAP_FOR_EACH (txq, node, &pmd->tx_queues) { + VLOG_INFO("Core_id: %d, Port: %s, tx_qid: %d\n", + pmd->core_id, netdev_get_name(txq->port->netdev), + txq->tx_qid); + } +} + +static struct dp_netdev_pmd_txq * +dp_netdev_pmd_lookup_txq(const struct dp_netdev_pmd_thread *pmd, + odp_port_t port_no) +{ + struct dp_netdev_pmd_txq *txq; + + CMAP_FOR_EACH_WITH_HASH (txq, node, hash_port_no(port_no), + &pmd->tx_queues) { + if (txq->port->port_no == port_no) { + return txq; + } + } + return NULL; +} + struct rxq_poll { struct dp_netdev_port *port; struct netdev_rxq *rx; @@ -2589,16 +2698,19 @@ pmd_load_queues(struct dp_netdev_pmd_thread *pmd, { struct rxq_poll *poll_list = *ppoll_list; struct dp_netdev_port *port; - int n_pmds_on_numa, index, i; + int n_pmds_on_numa, rx_index, tx_index, i, n_txq; /* Simple scheduler for netdev rx polling. */ + dp_netdev_pmd_detach_tx_queues(pmd); + for (i = 0; i < poll_cnt; i++) { port_unref(poll_list[i].port); } poll_cnt = 0; n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id); - index = 0; + rx_index = 0; + tx_index = 0; CMAP_FOR_EACH (port, node, &pmd->dp->ports) { /* Calls port_try_ref() to prevent the main thread @@ -2609,7 +2721,7 @@ pmd_load_queues(struct dp_netdev_pmd_thread *pmd, int i; for (i = 0; i < netdev_n_rxq(port->netdev); i++) { - if ((index % n_pmds_on_numa) == pmd->index) { + if ((rx_index % n_pmds_on_numa) == pmd->index) { poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1)); @@ -2618,7 +2730,16 @@ pmd_load_queues(struct dp_netdev_pmd_thread *pmd, poll_list[poll_cnt].rx = port->rxq[i]; poll_cnt++; } - index++; + rx_index++; + } + + n_txq = netdev_n_txq(port->netdev); + /* Last queue reserved for non pmd threads */ + n_txq = n_txq == 1 ? 1 : n_txq - 1; + for (i = 0; i < n_txq; i++) { + if ((tx_index % n_pmds_on_numa) == pmd->index || n_txq == 1) + dp_netdev_pmd_add_txq(pmd, port, i); + tx_index++; } } /* Unrefs the port_try_ref(). */ @@ -2689,6 +2810,8 @@ reload: goto reload; } + dp_netdev_pmd_detach_tx_queues(pmd); + for (i = 0; i < poll_cnt; i++) { port_unref(poll_list[i].port); } @@ -2802,16 +2925,6 @@ dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos) return next; } -static int -core_id_to_qid(unsigned core_id) -{ - if (core_id != NON_PMD_CORE_ID) { - return core_id; - } else { - return ovs_numa_get_n_cores(); - } -} - /* Configures the 'pmd' based on the input argument. */ static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp, @@ -2820,7 +2933,6 @@ dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp, pmd->dp = dp; pmd->index = index; pmd->core_id = core_id; - pmd->tx_qid = core_id_to_qid(core_id); pmd->numa_id = numa_id; ovs_refcount_init(&pmd->ref_cnt); @@ -2831,9 +2943,11 @@ dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp, ovs_mutex_init(&pmd->flow_mutex); dpcls_init(&pmd->cls); cmap_init(&pmd->flow_table); - /* init the 'flow_cache' since there is no + cmap_init(&pmd->tx_queues); + /* init the 'flow_cache' and 'tx_queues' since there is no * actual thread created for NON_PMD_CORE_ID. */ if (core_id == NON_PMD_CORE_ID) { + dp_netdev_configure_non_pmd_txqs(pmd); emc_cache_init(&pmd->flow_cache); } cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node), @@ -2846,6 +2960,7 @@ dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd) dp_netdev_pmd_flow_flush(pmd); dpcls_destroy(&pmd->cls); cmap_destroy(&pmd->flow_table); + cmap_destroy(&pmd->tx_queues); ovs_mutex_destroy(&pmd->flow_mutex); latch_destroy(&pmd->exit_latch); xpthread_cond_destroy(&pmd->cond); @@ -2862,6 +2977,7 @@ dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd) * no actual thread uninit it for NON_PMD_CORE_ID. */ if (pmd->core_id == NON_PMD_CORE_ID) { emc_cache_uninit(&pmd->flow_cache); + dp_netdev_pmd_detach_tx_queues(pmd); } else { latch_set(&pmd->exit_latch); dp_netdev_reload_pmd__(pmd); @@ -3471,13 +3587,15 @@ dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt, struct dp_netdev *dp = pmd->dp; int type = nl_attr_type(a); struct dp_netdev_port *p; + struct dp_netdev_pmd_txq *txq; int i; switch ((enum ovs_action_attr)type) { case OVS_ACTION_ATTR_OUTPUT: - p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a))); - if (OVS_LIKELY(p)) { - netdev_send(p->netdev, pmd->tx_qid, packets, cnt, may_steal); + txq = dp_netdev_pmd_lookup_txq(pmd, u32_to_odp(nl_attr_get_u32(a))); + if (OVS_LIKELY(txq)) { + netdev_send(txq->port->netdev, txq->tx_qid, + packets, cnt, may_steal); return; } break;