2 * Copyright (c) 2014 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
31 #include "dpif-netdev.h"
33 #include "netdev-dpdk.h"
34 #include "netdev-provider.h"
35 #include "netdev-vport.h"
37 #include "ofp-print.h"
40 #include "ovs-thread.h"
42 #include "packet-dpif.h"
46 #include "unaligned.h"
51 VLOG_DEFINE_THIS_MODULE(dpdk);
52 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
54 #define DPDK_PORT_WATCHDOG_INTERVAL 5
56 #define OVS_CACHE_LINE_SIZE CACHE_LINE_SIZE
57 #define OVS_VPORT_DPDK "ovs_dpdk"
60 * need to reserve tons of extra space in the mbufs so we can align the
61 * DMA addresses to 4KB.
64 #define MTU_TO_MAX_LEN(mtu) ((mtu) + ETHER_HDR_LEN + ETHER_CRC_LEN)
65 #define MBUF_SIZE(mtu) (MTU_TO_MAX_LEN(mtu) + (512) + \
66 sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM)
68 /* XXX: mempool size should be based on system resources. */
69 #define NB_MBUF (4096 * 64)
70 #define MP_CACHE_SZ (256 * 2)
73 #define NON_PMD_THREAD_TX_QUEUE 0
75 #define NIC_PORT_RX_Q_SIZE 2048 /* Size of Physical NIC RX Queue, Max (n+32<=4096)*/
76 #define NIC_PORT_TX_Q_SIZE 2048 /* Size of Physical NIC TX Queue, Max (n+32<=4096)*/
78 /* XXX: Needs per NIC value for these constants. */
79 #define RX_PTHRESH 32 /* Default values of RX prefetch threshold reg. */
80 #define RX_HTHRESH 32 /* Default values of RX host threshold reg. */
81 #define RX_WTHRESH 16 /* Default values of RX write-back threshold reg. */
83 #define TX_PTHRESH 36 /* Default values of TX prefetch threshold reg. */
84 #define TX_HTHRESH 0 /* Default values of TX host threshold reg. */
85 #define TX_WTHRESH 0 /* Default values of TX write-back threshold reg. */
87 static const struct rte_eth_conf port_conf = {
89 .mq_mode = ETH_MQ_RX_RSS,
91 .header_split = 0, /* Header Split disabled */
92 .hw_ip_checksum = 0, /* IP checksum offload disabled */
93 .hw_vlan_filter = 0, /* VLAN filtering disabled */
94 .jumbo_frame = 0, /* Jumbo Frame Support disabled */
100 .rss_hf = ETH_RSS_IPV4_TCP | ETH_RSS_IPV4 | ETH_RSS_IPV6
101 | ETH_RSS_IPV4_UDP | ETH_RSS_IPV6_TCP | ETH_RSS_IPV6_UDP,
105 .mq_mode = ETH_MQ_TX_NONE,
109 static const struct rte_eth_rxconf rx_conf = {
111 .pthresh = RX_PTHRESH,
112 .hthresh = RX_HTHRESH,
113 .wthresh = RX_WTHRESH,
117 static const struct rte_eth_txconf tx_conf = {
119 .pthresh = TX_PTHRESH,
120 .hthresh = TX_HTHRESH,
121 .wthresh = TX_WTHRESH,
125 .txq_flags = ETH_TXQ_FLAGS_NOMULTSEGS|ETH_TXQ_FLAGS_NOOFFLOADS,
128 enum { MAX_RX_QUEUE_LEN = 192 };
129 enum { MAX_TX_QUEUE_LEN = 384 };
130 enum { DPDK_RING_SIZE = 256 };
131 BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE));
132 enum { DRAIN_TSC = 200000ULL };
134 static int rte_eal_init_ret = ENODEV;
136 static struct ovs_mutex dpdk_mutex = OVS_MUTEX_INITIALIZER;
138 /* Contains all 'struct dpdk_dev's. */
139 static struct list dpdk_list OVS_GUARDED_BY(dpdk_mutex)
140 = LIST_INITIALIZER(&dpdk_list);
142 static struct list dpdk_mp_list OVS_GUARDED_BY(dpdk_mutex)
143 = LIST_INITIALIZER(&dpdk_mp_list);
145 /* This mutex must be used by non pmd threads when allocating or freeing
146 * mbufs through mempools. Since dpdk_queue_pkts() and dpdk_queue_flush() may
147 * use mempools, a non pmd thread should hold this mutex while calling them */
148 struct ovs_mutex nonpmd_mempool_mutex = OVS_MUTEX_INITIALIZER;
151 struct rte_mempool *mp;
155 struct list list_node OVS_GUARDED_BY(dpdk_mutex);
158 /* There should be one 'struct dpdk_tx_queue' created for
160 struct dpdk_tx_queue {
161 rte_spinlock_t tx_lock;
162 bool flush_tx; /* Set to true to flush queue everytime */
163 /* pkts are queued. */
166 struct rte_mbuf *burst_pkts[MAX_TX_QUEUE_LEN];
169 /* dpdk has no way to remove dpdk ring ethernet devices
170 so we have to keep them around once they've been created
173 static struct list dpdk_ring_list OVS_GUARDED_BY(dpdk_mutex)
174 = LIST_INITIALIZER(&dpdk_ring_list);
177 /* For the client rings */
178 struct rte_ring *cring_tx;
179 struct rte_ring *cring_rx;
180 int user_port_id; /* User given port no, parsed from port name */
181 int eth_port_id; /* ethernet device port id */
182 struct list list_node OVS_GUARDED_BY(dpdk_mutex);
190 struct dpdk_tx_queue *tx_q;
192 struct ovs_mutex mutex OVS_ACQ_AFTER(dpdk_mutex);
194 struct dpdk_mp *dpdk_mp;
198 struct netdev_stats stats;
200 uint8_t hwaddr[ETH_ADDR_LEN];
201 enum netdev_flags flags;
203 struct rte_eth_link link;
207 struct list list_node OVS_GUARDED_BY(dpdk_mutex);
210 struct netdev_rxq_dpdk {
211 struct netdev_rxq up;
215 static bool thread_is_pmd(void);
217 static int netdev_dpdk_construct(struct netdev *);
220 is_dpdk_class(const struct netdev_class *class)
222 return class->construct == netdev_dpdk_construct;
225 /* XXX: use dpdk malloc for entire OVS. infact huge page shld be used
226 * for all other sengments data, bss and text. */
229 dpdk_rte_mzalloc(size_t sz)
233 ptr = rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE);
240 /* XXX this function should be called only by pmd threads (or by non pmd
241 * threads holding the nonpmd_mempool_mutex) */
243 free_dpdk_buf(struct dpif_packet *p)
245 struct rte_mbuf *pkt = (struct rte_mbuf *) p;
247 rte_pktmbuf_free_seg(pkt);
251 __rte_pktmbuf_init(struct rte_mempool *mp,
252 void *opaque_arg OVS_UNUSED,
254 unsigned i OVS_UNUSED)
256 struct rte_mbuf *m = _m;
257 uint32_t buf_len = mp->elt_size - sizeof(struct dpif_packet);
259 RTE_MBUF_ASSERT(mp->elt_size >= sizeof(struct dpif_packet));
261 memset(m, 0, mp->elt_size);
263 /* start of buffer is just after mbuf structure */
264 m->buf_addr = (char *)m + sizeof(struct dpif_packet);
265 m->buf_physaddr = rte_mempool_virt2phy(mp, m) +
266 sizeof(struct dpif_packet);
267 m->buf_len = (uint16_t)buf_len;
269 /* keep some headroom between start of buffer and data */
270 m->pkt.data = (char*) m->buf_addr + RTE_MIN(RTE_PKTMBUF_HEADROOM, m->buf_len);
272 /* init some constant fields */
273 m->type = RTE_MBUF_PKT;
276 m->pkt.in_port = 0xff;
280 ovs_rte_pktmbuf_init(struct rte_mempool *mp,
281 void *opaque_arg OVS_UNUSED,
283 unsigned i OVS_UNUSED)
285 struct rte_mbuf *m = _m;
287 __rte_pktmbuf_init(mp, opaque_arg, _m, i);
289 ofpbuf_init_dpdk((struct ofpbuf *) m, m->buf_len);
292 static struct dpdk_mp *
293 dpdk_mp_get(int socket_id, int mtu) OVS_REQUIRES(dpdk_mutex)
295 struct dpdk_mp *dmp = NULL;
296 char mp_name[RTE_MEMPOOL_NAMESIZE];
298 LIST_FOR_EACH (dmp, list_node, &dpdk_mp_list) {
299 if (dmp->socket_id == socket_id && dmp->mtu == mtu) {
305 dmp = dpdk_rte_mzalloc(sizeof *dmp);
306 dmp->socket_id = socket_id;
310 if (snprintf(mp_name, RTE_MEMPOOL_NAMESIZE, "ovs_mp_%d_%d", dmp->mtu,
311 dmp->socket_id) < 0) {
315 dmp->mp = rte_mempool_create(mp_name, NB_MBUF, MBUF_SIZE(mtu),
317 sizeof(struct rte_pktmbuf_pool_private),
318 rte_pktmbuf_pool_init, NULL,
319 ovs_rte_pktmbuf_init, NULL,
322 if (dmp->mp == NULL) {
326 list_push_back(&dpdk_mp_list, &dmp->list_node);
331 dpdk_mp_put(struct dpdk_mp *dmp)
339 ovs_assert(dmp->refcount >= 0);
342 /* I could not find any API to destroy mp. */
343 if (dmp->refcount == 0) {
344 list_delete(dmp->list_node);
345 /* destroy mp-pool. */
351 check_link_status(struct netdev_dpdk *dev)
353 struct rte_eth_link link;
355 rte_eth_link_get_nowait(dev->port_id, &link);
357 if (dev->link.link_status != link.link_status) {
358 netdev_change_seq_changed(&dev->up);
360 dev->link_reset_cnt++;
362 if (dev->link.link_status) {
363 VLOG_DBG_RL(&rl, "Port %d Link Up - speed %u Mbps - %s",
364 dev->port_id, (unsigned)dev->link.link_speed,
365 (dev->link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
366 ("full-duplex") : ("half-duplex"));
368 VLOG_DBG_RL(&rl, "Port %d Link Down", dev->port_id);
374 dpdk_watchdog(void *dummy OVS_UNUSED)
376 struct netdev_dpdk *dev;
378 pthread_detach(pthread_self());
381 ovs_mutex_lock(&dpdk_mutex);
382 LIST_FOR_EACH (dev, list_node, &dpdk_list) {
383 ovs_mutex_lock(&dev->mutex);
384 check_link_status(dev);
385 ovs_mutex_unlock(&dev->mutex);
387 ovs_mutex_unlock(&dpdk_mutex);
388 xsleep(DPDK_PORT_WATCHDOG_INTERVAL);
395 dpdk_eth_dev_init(struct netdev_dpdk *dev) OVS_REQUIRES(dpdk_mutex)
397 struct rte_pktmbuf_pool_private *mbp_priv;
398 struct ether_addr eth_addr;
402 if (dev->port_id < 0 || dev->port_id >= rte_eth_dev_count()) {
406 diag = rte_eth_dev_configure(dev->port_id, dev->up.n_rxq, dev->up.n_txq,
409 VLOG_ERR("eth dev config error %d",diag);
413 for (i = 0; i < dev->up.n_txq; i++) {
414 diag = rte_eth_tx_queue_setup(dev->port_id, i, NIC_PORT_TX_Q_SIZE,
415 dev->socket_id, &tx_conf);
417 VLOG_ERR("eth dev tx queue setup error %d",diag);
422 for (i = 0; i < dev->up.n_rxq; i++) {
423 diag = rte_eth_rx_queue_setup(dev->port_id, i, NIC_PORT_RX_Q_SIZE,
425 &rx_conf, dev->dpdk_mp->mp);
427 VLOG_ERR("eth dev rx queue setup error %d",diag);
432 diag = rte_eth_dev_start(dev->port_id);
434 VLOG_ERR("eth dev start error %d",diag);
438 rte_eth_promiscuous_enable(dev->port_id);
439 rte_eth_allmulticast_enable(dev->port_id);
441 memset(ð_addr, 0x0, sizeof(eth_addr));
442 rte_eth_macaddr_get(dev->port_id, ð_addr);
443 VLOG_INFO_RL(&rl, "Port %d: "ETH_ADDR_FMT"",
444 dev->port_id, ETH_ADDR_ARGS(eth_addr.addr_bytes));
446 memcpy(dev->hwaddr, eth_addr.addr_bytes, ETH_ADDR_LEN);
447 rte_eth_link_get_nowait(dev->port_id, &dev->link);
449 mbp_priv = rte_mempool_get_priv(dev->dpdk_mp->mp);
450 dev->buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
452 dev->flags = NETDEV_UP | NETDEV_PROMISC;
456 static struct netdev_dpdk *
457 netdev_dpdk_cast(const struct netdev *netdev)
459 return CONTAINER_OF(netdev, struct netdev_dpdk, up);
462 static struct netdev *
463 netdev_dpdk_alloc(void)
465 struct netdev_dpdk *netdev = dpdk_rte_mzalloc(sizeof *netdev);
470 netdev_dpdk_set_txq(struct netdev_dpdk *netdev, unsigned int n_txqs)
474 netdev->tx_q = dpdk_rte_mzalloc(n_txqs * sizeof *netdev->tx_q);
475 /* Each index is considered as a cpu core id, since there should
476 * be one tx queue for each cpu core. */
477 for (i = 0; i < n_txqs; i++) {
478 int core_id = ovs_numa_get_numa_id(i);
480 rte_spinlock_init(&netdev->tx_q[i].tx_lock);
481 /* If the corresponding core is not on the same numa node
482 * as 'netdev', flags the 'flush_tx'. */
483 netdev->tx_q[i].flush_tx = netdev->socket_id == core_id;
488 netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no)
489 OVS_REQUIRES(dpdk_mutex)
491 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
494 ovs_mutex_init(&netdev->mutex);
496 ovs_mutex_lock(&netdev->mutex);
498 /* XXX: need to discover device node at run time. */
499 netdev->socket_id = SOCKET0;
500 netdev_dpdk_set_txq(netdev, NR_QUEUE);
501 netdev->port_id = port_no;
503 netdev->mtu = ETHER_MTU;
504 netdev->max_packet_len = MTU_TO_MAX_LEN(netdev->mtu);
506 netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, netdev->mtu);
507 if (!netdev->dpdk_mp) {
512 netdev_->n_txq = NR_QUEUE;
513 netdev_->n_rxq = NR_QUEUE;
514 err = dpdk_eth_dev_init(netdev);
519 list_push_back(&dpdk_list, &netdev->list_node);
523 rte_free(netdev->tx_q);
525 ovs_mutex_unlock(&netdev->mutex);
530 dpdk_dev_parse_name(const char dev_name[], const char prefix[],
531 unsigned int *port_no)
535 if (strncmp(dev_name, prefix, strlen(prefix))) {
539 cport = dev_name + strlen(prefix);
540 *port_no = strtol(cport, 0, 0); /* string must be null terminated */
545 netdev_dpdk_construct(struct netdev *netdev)
547 unsigned int port_no;
550 if (rte_eal_init_ret) {
551 return rte_eal_init_ret;
554 /* Names always start with "dpdk" */
555 err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no);
560 ovs_mutex_lock(&dpdk_mutex);
561 err = netdev_dpdk_init(netdev, port_no);
562 ovs_mutex_unlock(&dpdk_mutex);
567 netdev_dpdk_destruct(struct netdev *netdev_)
569 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
571 ovs_mutex_lock(&dev->mutex);
572 rte_eth_dev_stop(dev->port_id);
573 ovs_mutex_unlock(&dev->mutex);
575 ovs_mutex_lock(&dpdk_mutex);
577 list_remove(&dev->list_node);
578 dpdk_mp_put(dev->dpdk_mp);
579 ovs_mutex_unlock(&dpdk_mutex);
581 ovs_mutex_destroy(&dev->mutex);
585 netdev_dpdk_dealloc(struct netdev *netdev_)
587 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
593 netdev_dpdk_get_config(const struct netdev *netdev_, struct smap *args)
595 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
597 ovs_mutex_lock(&dev->mutex);
599 /* XXX: Allow to configure number of queues. */
600 smap_add_format(args, "configured_rx_queues", "%u", netdev_->n_rxq);
601 smap_add_format(args, "configured_tx_queues", "%u", netdev_->n_rxq);
602 ovs_mutex_unlock(&dev->mutex);
608 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
610 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
612 return netdev->socket_id;
615 /* Sets the number of tx queues and rx queues for the dpdk interface.
616 * If the configuration fails, do not try restoring its old configuration
617 * and just returns the error. */
619 netdev_dpdk_set_multiq(struct netdev *netdev_, unsigned int n_txq,
622 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
625 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
629 ovs_mutex_lock(&netdev->mutex);
630 rte_eth_dev_stop(netdev->port_id);
631 netdev->up.n_txq = n_txq;
632 netdev->up.n_rxq = n_rxq;
633 err = dpdk_eth_dev_init(netdev);
634 if (!err && netdev->up.n_txq != n_txq) {
635 rte_free(netdev->tx_q);
636 netdev_dpdk_set_txq(netdev, n_txq);
638 ovs_mutex_unlock(&netdev->mutex);
643 static struct netdev_rxq *
644 netdev_dpdk_rxq_alloc(void)
646 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
651 static struct netdev_rxq_dpdk *
652 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
654 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
658 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
660 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
661 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
663 ovs_mutex_lock(&netdev->mutex);
664 rx->port_id = netdev->port_id;
665 ovs_mutex_unlock(&netdev->mutex);
671 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
676 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
678 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
684 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
686 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
689 while (nb_tx != txq->count) {
692 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
701 if (OVS_UNLIKELY(nb_tx != txq->count)) {
702 /* free buffers, which we couldn't transmit, one at a time (each
703 * packet could come from a different mempool) */
706 for (i = nb_tx; i < txq->count; i++) {
707 rte_pktmbuf_free_seg(txq->burst_pkts[i]);
709 ovs_mutex_lock(&dev->mutex);
710 dev->stats.tx_dropped += txq->count-nb_tx;
711 ovs_mutex_unlock(&dev->mutex);
715 txq->tsc = rte_get_timer_cycles();
719 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
721 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
723 if (txq->count == 0) {
726 rte_spinlock_lock(&txq->tx_lock);
727 dpdk_queue_flush__(dev, qid);
728 rte_spinlock_unlock(&txq->tx_lock);
732 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dpif_packet **packets,
735 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
736 struct netdev *netdev = rx->up.netdev;
737 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
740 /* There is only one tx queue for this core. Do not flush other
742 if (rxq_->queue_id == rte_lcore_id()) {
743 dpdk_queue_flush(dev, rxq_->queue_id);
746 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
747 (struct rte_mbuf **) packets,
748 MIN((int)NETDEV_MAX_RX_BATCH,
749 (int)MAX_RX_QUEUE_LEN));
760 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
761 struct rte_mbuf **pkts, int cnt)
763 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
768 rte_spinlock_lock(&txq->tx_lock);
770 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
771 int tocopy = MIN(freeslots, cnt-i);
773 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
774 tocopy * sizeof (struct rte_mbuf *));
776 txq->count += tocopy;
779 if (txq->count == MAX_TX_QUEUE_LEN || txq->flush_tx) {
780 dpdk_queue_flush__(dev, qid);
782 diff_tsc = rte_get_timer_cycles() - txq->tsc;
783 if (diff_tsc >= DRAIN_TSC) {
784 dpdk_queue_flush__(dev, qid);
787 rte_spinlock_unlock(&txq->tx_lock);
790 /* Tx function. Transmit packets indefinitely */
792 dpdk_do_tx_copy(struct netdev *netdev, struct dpif_packet ** pkts, int cnt)
793 OVS_NO_THREAD_SAFETY_ANALYSIS
795 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
796 struct rte_mbuf *mbufs[cnt];
801 /* If we are on a non pmd thread we have to use the mempool mutex, because
802 * every non pmd thread shares the same mempool cache */
804 if (!thread_is_pmd()) {
805 ovs_mutex_lock(&nonpmd_mempool_mutex);
808 for (i = 0; i < cnt; i++) {
809 int size = ofpbuf_size(&pkts[i]->ofpbuf);
811 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
812 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
813 (int)size , dev->max_packet_len);
819 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
821 if (!mbufs[newcnt]) {
826 /* We have to do a copy for now */
827 memcpy(mbufs[newcnt]->pkt.data, ofpbuf_data(&pkts[i]->ofpbuf), size);
829 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
830 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
835 if (OVS_UNLIKELY(dropped)) {
836 ovs_mutex_lock(&dev->mutex);
837 dev->stats.tx_dropped += dropped;
838 ovs_mutex_unlock(&dev->mutex);
841 dpdk_queue_pkts(dev, NON_PMD_THREAD_TX_QUEUE, mbufs, newcnt);
842 dpdk_queue_flush(dev, NON_PMD_THREAD_TX_QUEUE);
844 if (!thread_is_pmd()) {
845 ovs_mutex_unlock(&nonpmd_mempool_mutex);
850 netdev_dpdk_send(struct netdev *netdev, int qid, struct dpif_packet **pkts,
851 int cnt, bool may_steal)
853 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
857 if (!may_steal || pkts[0]->ofpbuf.source != OFPBUF_DPDK) {
858 dpdk_do_tx_copy(netdev, pkts, cnt);
861 for (i = 0; i < cnt; i++) {
862 dpif_packet_delete(pkts[i]);
869 qid = rte_lcore_id();
871 for (i = 0; i < cnt; i++) {
872 int size = ofpbuf_size(&pkts[i]->ofpbuf);
873 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
874 if (next_tx_idx != i) {
875 dpdk_queue_pkts(dev, qid,
876 (struct rte_mbuf **)&pkts[next_tx_idx],
880 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
881 (int)size , dev->max_packet_len);
883 dpif_packet_delete(pkts[i]);
888 if (next_tx_idx != cnt) {
889 dpdk_queue_pkts(dev, qid,
890 (struct rte_mbuf **)&pkts[next_tx_idx],
894 if (OVS_UNLIKELY(dropped)) {
895 ovs_mutex_lock(&dev->mutex);
896 dev->stats.tx_dropped += dropped;
897 ovs_mutex_unlock(&dev->mutex);
906 netdev_dpdk_set_etheraddr(struct netdev *netdev,
907 const uint8_t mac[ETH_ADDR_LEN])
909 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
911 ovs_mutex_lock(&dev->mutex);
912 if (!eth_addr_equals(dev->hwaddr, mac)) {
913 memcpy(dev->hwaddr, mac, ETH_ADDR_LEN);
914 netdev_change_seq_changed(netdev);
916 ovs_mutex_unlock(&dev->mutex);
922 netdev_dpdk_get_etheraddr(const struct netdev *netdev,
923 uint8_t mac[ETH_ADDR_LEN])
925 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
927 ovs_mutex_lock(&dev->mutex);
928 memcpy(mac, dev->hwaddr, ETH_ADDR_LEN);
929 ovs_mutex_unlock(&dev->mutex);
935 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
937 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
939 ovs_mutex_lock(&dev->mutex);
941 ovs_mutex_unlock(&dev->mutex);
947 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
949 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
951 struct dpdk_mp *old_mp;
954 ovs_mutex_lock(&dpdk_mutex);
955 ovs_mutex_lock(&dev->mutex);
956 if (dev->mtu == mtu) {
961 mp = dpdk_mp_get(dev->socket_id, dev->mtu);
967 rte_eth_dev_stop(dev->port_id);
970 old_mp = dev->dpdk_mp;
973 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
975 err = dpdk_eth_dev_init(dev);
979 dev->dpdk_mp = old_mp;
980 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
981 dpdk_eth_dev_init(dev);
986 netdev_change_seq_changed(netdev);
988 ovs_mutex_unlock(&dev->mutex);
989 ovs_mutex_unlock(&dpdk_mutex);
994 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
997 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
999 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1000 struct rte_eth_stats rte_stats;
1003 netdev_dpdk_get_carrier(netdev, &gg);
1004 ovs_mutex_lock(&dev->mutex);
1005 rte_eth_stats_get(dev->port_id, &rte_stats);
1007 memset(stats, 0, sizeof(*stats));
1009 stats->rx_packets = rte_stats.ipackets;
1010 stats->tx_packets = rte_stats.opackets;
1011 stats->rx_bytes = rte_stats.ibytes;
1012 stats->tx_bytes = rte_stats.obytes;
1013 stats->rx_errors = rte_stats.ierrors;
1014 stats->tx_errors = rte_stats.oerrors;
1015 stats->multicast = rte_stats.imcasts;
1017 stats->tx_dropped = dev->stats.tx_dropped;
1018 ovs_mutex_unlock(&dev->mutex);
1024 netdev_dpdk_get_features(const struct netdev *netdev_,
1025 enum netdev_features *current,
1026 enum netdev_features *advertised OVS_UNUSED,
1027 enum netdev_features *supported OVS_UNUSED,
1028 enum netdev_features *peer OVS_UNUSED)
1030 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1031 struct rte_eth_link link;
1033 ovs_mutex_lock(&dev->mutex);
1035 ovs_mutex_unlock(&dev->mutex);
1037 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
1038 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
1039 *current = NETDEV_F_AUTONEG;
1041 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
1042 if (link.link_speed == ETH_LINK_SPEED_10) {
1043 *current = NETDEV_F_10MB_HD;
1045 if (link.link_speed == ETH_LINK_SPEED_100) {
1046 *current = NETDEV_F_100MB_HD;
1048 if (link.link_speed == ETH_LINK_SPEED_1000) {
1049 *current = NETDEV_F_1GB_HD;
1051 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
1052 if (link.link_speed == ETH_LINK_SPEED_10) {
1053 *current = NETDEV_F_10MB_FD;
1055 if (link.link_speed == ETH_LINK_SPEED_100) {
1056 *current = NETDEV_F_100MB_FD;
1058 if (link.link_speed == ETH_LINK_SPEED_1000) {
1059 *current = NETDEV_F_1GB_FD;
1061 if (link.link_speed == ETH_LINK_SPEED_10000) {
1062 *current = NETDEV_F_10GB_FD;
1070 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1072 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1075 ovs_mutex_lock(&dev->mutex);
1076 ifindex = dev->port_id;
1077 ovs_mutex_unlock(&dev->mutex);
1083 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1085 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1087 ovs_mutex_lock(&dev->mutex);
1088 check_link_status(dev);
1089 *carrier = dev->link.link_status;
1090 ovs_mutex_unlock(&dev->mutex);
1095 static long long int
1096 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1098 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1099 long long int carrier_resets;
1101 ovs_mutex_lock(&dev->mutex);
1102 carrier_resets = dev->link_reset_cnt;
1103 ovs_mutex_unlock(&dev->mutex);
1105 return carrier_resets;
1109 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1110 long long int interval OVS_UNUSED)
1116 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1117 enum netdev_flags off, enum netdev_flags on,
1118 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1122 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1126 *old_flagsp = dev->flags;
1130 if (dev->flags == *old_flagsp) {
1134 if (dev->flags & NETDEV_UP) {
1135 err = rte_eth_dev_start(dev->port_id);
1140 if (dev->flags & NETDEV_PROMISC) {
1141 rte_eth_promiscuous_enable(dev->port_id);
1144 if (!(dev->flags & NETDEV_UP)) {
1145 rte_eth_dev_stop(dev->port_id);
1152 netdev_dpdk_update_flags(struct netdev *netdev_,
1153 enum netdev_flags off, enum netdev_flags on,
1154 enum netdev_flags *old_flagsp)
1156 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1159 ovs_mutex_lock(&netdev->mutex);
1160 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1161 ovs_mutex_unlock(&netdev->mutex);
1167 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1169 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1170 struct rte_eth_dev_info dev_info;
1172 if (dev->port_id < 0)
1175 ovs_mutex_lock(&dev->mutex);
1176 rte_eth_dev_info_get(dev->port_id, &dev_info);
1177 ovs_mutex_unlock(&dev->mutex);
1179 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1181 smap_add_format(args, "port_no", "%d", dev->port_id);
1182 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1183 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1184 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1185 smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen);
1186 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1187 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1188 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1189 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1190 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1191 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1193 smap_add_format(args, "pci-vendor_id", "0x%u", dev_info.pci_dev->id.vendor_id);
1194 smap_add_format(args, "pci-device_id", "0x%x", dev_info.pci_dev->id.device_id);
1200 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1201 OVS_REQUIRES(dev->mutex)
1203 enum netdev_flags old_flags;
1206 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1208 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1213 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1214 const char *argv[], void *aux OVS_UNUSED)
1218 if (!strcasecmp(argv[argc - 1], "up")) {
1220 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1223 unixctl_command_reply_error(conn, "Invalid Admin State");
1228 struct netdev *netdev = netdev_from_name(argv[1]);
1229 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1230 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1232 ovs_mutex_lock(&dpdk_dev->mutex);
1233 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1234 ovs_mutex_unlock(&dpdk_dev->mutex);
1236 netdev_close(netdev);
1238 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1239 netdev_close(netdev);
1243 struct netdev_dpdk *netdev;
1245 ovs_mutex_lock(&dpdk_mutex);
1246 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1247 ovs_mutex_lock(&netdev->mutex);
1248 netdev_dpdk_set_admin_state__(netdev, up);
1249 ovs_mutex_unlock(&netdev->mutex);
1251 ovs_mutex_unlock(&dpdk_mutex);
1253 unixctl_command_reply(conn, "OK");
1257 dpdk_common_init(void)
1259 unixctl_command_register("netdev-dpdk/set-admin-state",
1260 "[netdev] up|down", 1, 2,
1261 netdev_dpdk_set_admin_state, NULL);
1263 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
1267 dpdk_class_init(void)
1271 result = rte_eal_pci_probe();
1273 VLOG_ERR("Cannot probe PCI");
1277 VLOG_INFO("Ethernet Device Count: %d", (int)rte_eth_dev_count());
1285 dpdk_ring_create(const char dev_name[], unsigned int port_no,
1286 unsigned int *eth_port_id)
1288 struct dpdk_ring *ivshmem;
1292 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
1293 if (ivshmem == NULL) {
1297 err = snprintf(ring_name, 10, "%s_tx", dev_name);
1302 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0, 0);
1303 if (ivshmem->cring_tx == NULL) {
1308 err = snprintf(ring_name, 10, "%s_rx", dev_name);
1313 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0, 0);
1314 if (ivshmem->cring_rx == NULL) {
1319 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
1320 &ivshmem->cring_tx, 1, SOCKET0);
1327 ivshmem->user_port_id = port_no;
1328 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
1329 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
1331 *eth_port_id = ivshmem->eth_port_id;
1336 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
1338 struct dpdk_ring *ivshmem;
1339 unsigned int port_no;
1342 /* Names always start with "dpdkr" */
1343 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
1348 /* look through our list to find the device */
1349 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
1350 if (ivshmem->user_port_id == port_no) {
1351 VLOG_INFO("Found dpdk ring device %s:\n", dev_name);
1352 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
1356 /* Need to create the device rings */
1357 return dpdk_ring_create(dev_name, port_no, eth_port_id);
1361 netdev_dpdk_ring_construct(struct netdev *netdev)
1363 unsigned int port_no = 0;
1366 if (rte_eal_init_ret) {
1367 return rte_eal_init_ret;
1370 ovs_mutex_lock(&dpdk_mutex);
1372 err = dpdk_ring_open(netdev->name, &port_no);
1377 err = netdev_dpdk_init(netdev, port_no);
1380 ovs_mutex_unlock(&dpdk_mutex);
1384 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, MULTIQ) \
1388 NULL, /* netdev_dpdk_run */ \
1389 NULL, /* netdev_dpdk_wait */ \
1391 netdev_dpdk_alloc, \
1393 netdev_dpdk_destruct, \
1394 netdev_dpdk_dealloc, \
1395 netdev_dpdk_get_config, \
1396 NULL, /* netdev_dpdk_set_config */ \
1397 NULL, /* get_tunnel_config */ \
1398 netdev_dpdk_get_numa_id, /* get_numa_id */ \
1399 MULTIQ, /* set_multiq */ \
1401 netdev_dpdk_send, /* send */ \
1402 NULL, /* send_wait */ \
1404 netdev_dpdk_set_etheraddr, \
1405 netdev_dpdk_get_etheraddr, \
1406 netdev_dpdk_get_mtu, \
1407 netdev_dpdk_set_mtu, \
1408 netdev_dpdk_get_ifindex, \
1409 netdev_dpdk_get_carrier, \
1410 netdev_dpdk_get_carrier_resets, \
1411 netdev_dpdk_set_miimon, \
1412 netdev_dpdk_get_stats, \
1413 netdev_dpdk_get_features, \
1414 NULL, /* set_advertisements */ \
1416 NULL, /* set_policing */ \
1417 NULL, /* get_qos_types */ \
1418 NULL, /* get_qos_capabilities */ \
1419 NULL, /* get_qos */ \
1420 NULL, /* set_qos */ \
1421 NULL, /* get_queue */ \
1422 NULL, /* set_queue */ \
1423 NULL, /* delete_queue */ \
1424 NULL, /* get_queue_stats */ \
1425 NULL, /* queue_dump_start */ \
1426 NULL, /* queue_dump_next */ \
1427 NULL, /* queue_dump_done */ \
1428 NULL, /* dump_queue_stats */ \
1430 NULL, /* get_in4 */ \
1431 NULL, /* set_in4 */ \
1432 NULL, /* get_in6 */ \
1433 NULL, /* add_router */ \
1434 NULL, /* get_next_hop */ \
1435 netdev_dpdk_get_status, \
1436 NULL, /* arp_lookup */ \
1438 netdev_dpdk_update_flags, \
1440 netdev_dpdk_rxq_alloc, \
1441 netdev_dpdk_rxq_construct, \
1442 netdev_dpdk_rxq_destruct, \
1443 netdev_dpdk_rxq_dealloc, \
1444 netdev_dpdk_rxq_recv, \
1445 NULL, /* rx_wait */ \
1446 NULL, /* rxq_drain */ \
1450 dpdk_init(int argc, char **argv)
1454 if (argc < 2 || strcmp(argv[1], "--dpdk"))
1457 /* Make sure program name passed to rte_eal_init() is vswitchd. */
1463 /* Make sure things are initialized ... */
1464 result = rte_eal_init(argc, argv);
1466 ovs_abort(result, "Cannot init EAL\n");
1469 rte_memzone_dump(stdout);
1470 rte_eal_init_ret = 0;
1472 if (argc > result) {
1473 argv[result] = argv[0];
1476 /* We are called from the main thread here */
1477 thread_set_nonpmd();
1482 const struct netdev_class dpdk_class =
1486 netdev_dpdk_construct,
1487 netdev_dpdk_set_multiq);
1489 const struct netdev_class dpdk_ring_class =
1493 netdev_dpdk_ring_construct,
1497 netdev_dpdk_register(void)
1499 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
1501 if (rte_eal_init_ret) {
1505 if (ovsthread_once_start(&once)) {
1507 netdev_register_provider(&dpdk_class);
1508 netdev_register_provider(&dpdk_ring_class);
1509 ovsthread_once_done(&once);
1514 pmd_thread_setaffinity_cpu(int cpu)
1520 CPU_SET(cpu, &cpuset);
1521 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
1523 VLOG_ERR("Thread affinity error %d",err);
1526 /* lcore_id 0 is reseved for use by non pmd threads. */
1527 RTE_PER_LCORE(_lcore_id) = cpu + 1;
1533 thread_set_nonpmd(void)
1535 /* We cannot have RTE_MAX_LCORE pmd threads, because lcore_id 0 is reserved
1536 * for non pmd threads */
1537 BUILD_ASSERT(NR_PMD_THREADS < RTE_MAX_LCORE);
1538 /* We have to use 0 to allow non pmd threads to perform certain DPDK
1539 * operations, like rte_eth_dev_configure(). */
1540 RTE_PER_LCORE(_lcore_id) = 0;
1546 return rte_lcore_id() != 0;