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"
39 #include "ovs-thread.h"
41 #include "packet-dpif.h"
45 #include "unaligned.h"
50 VLOG_DEFINE_THIS_MODULE(dpdk);
51 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
53 #define DPDK_PORT_WATCHDOG_INTERVAL 5
55 #define OVS_CACHE_LINE_SIZE CACHE_LINE_SIZE
56 #define OVS_VPORT_DPDK "ovs_dpdk"
59 * need to reserve tons of extra space in the mbufs so we can align the
60 * DMA addresses to 4KB.
63 #define MTU_TO_MAX_LEN(mtu) ((mtu) + ETHER_HDR_LEN + ETHER_CRC_LEN)
64 #define MBUF_SIZE(mtu) (MTU_TO_MAX_LEN(mtu) + (512) + \
65 sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM)
67 /* XXX: mempool size should be based on system resources. */
68 #define NB_MBUF (4096 * 64)
69 #define MP_CACHE_SZ (256 * 2)
72 #define NON_PMD_THREAD_TX_QUEUE 0
74 #define NIC_PORT_RX_Q_SIZE 2048 /* Size of Physical NIC RX Queue, Max (n+32<=4096)*/
75 #define NIC_PORT_TX_Q_SIZE 2048 /* Size of Physical NIC TX Queue, Max (n+32<=4096)*/
77 /* XXX: Needs per NIC value for these constants. */
78 #define RX_PTHRESH 32 /* Default values of RX prefetch threshold reg. */
79 #define RX_HTHRESH 32 /* Default values of RX host threshold reg. */
80 #define RX_WTHRESH 16 /* Default values of RX write-back threshold reg. */
82 #define TX_PTHRESH 36 /* Default values of TX prefetch threshold reg. */
83 #define TX_HTHRESH 0 /* Default values of TX host threshold reg. */
84 #define TX_WTHRESH 0 /* Default values of TX write-back threshold reg. */
86 static const struct rte_eth_conf port_conf = {
88 .mq_mode = ETH_MQ_RX_RSS,
90 .header_split = 0, /* Header Split disabled */
91 .hw_ip_checksum = 0, /* IP checksum offload disabled */
92 .hw_vlan_filter = 0, /* VLAN filtering disabled */
93 .jumbo_frame = 0, /* Jumbo Frame Support disabled */
99 .rss_hf = ETH_RSS_IPV4_TCP | ETH_RSS_IPV4 | ETH_RSS_IPV6
100 | ETH_RSS_IPV4_UDP | ETH_RSS_IPV6_TCP | ETH_RSS_IPV6_UDP,
104 .mq_mode = ETH_MQ_TX_NONE,
108 static const struct rte_eth_rxconf rx_conf = {
110 .pthresh = RX_PTHRESH,
111 .hthresh = RX_HTHRESH,
112 .wthresh = RX_WTHRESH,
116 static const struct rte_eth_txconf tx_conf = {
118 .pthresh = TX_PTHRESH,
119 .hthresh = TX_HTHRESH,
120 .wthresh = TX_WTHRESH,
124 .txq_flags = ETH_TXQ_FLAGS_NOMULTSEGS|ETH_TXQ_FLAGS_NOOFFLOADS,
127 enum { MAX_RX_QUEUE_LEN = 192 };
128 enum { MAX_TX_QUEUE_LEN = 384 };
129 enum { DPDK_RING_SIZE = 256 };
130 BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE));
131 enum { DRAIN_TSC = 200000ULL };
133 static int rte_eal_init_ret = ENODEV;
135 static struct ovs_mutex dpdk_mutex = OVS_MUTEX_INITIALIZER;
137 /* Contains all 'struct dpdk_dev's. */
138 static struct list dpdk_list OVS_GUARDED_BY(dpdk_mutex)
139 = LIST_INITIALIZER(&dpdk_list);
141 static struct list dpdk_mp_list OVS_GUARDED_BY(dpdk_mutex)
142 = LIST_INITIALIZER(&dpdk_mp_list);
144 /* This mutex must be used by non pmd threads when allocating or freeing
145 * mbufs through mempools. Since dpdk_queue_pkts() and dpdk_queue_flush() may
146 * use mempools, a non pmd thread should hold this mutex while calling them */
147 struct ovs_mutex nonpmd_mempool_mutex = OVS_MUTEX_INITIALIZER;
150 struct rte_mempool *mp;
154 struct list list_node OVS_GUARDED_BY(dpdk_mutex);
157 /* There should be one 'struct dpdk_tx_queue' created for
159 struct dpdk_tx_queue {
160 rte_spinlock_t tx_lock;
163 struct rte_mbuf *burst_pkts[MAX_TX_QUEUE_LEN];
166 /* dpdk has no way to remove dpdk ring ethernet devices
167 so we have to keep them around once they've been created
170 static struct list dpdk_ring_list OVS_GUARDED_BY(dpdk_mutex)
171 = LIST_INITIALIZER(&dpdk_ring_list);
174 /* For the client rings */
175 struct rte_ring *cring_tx;
176 struct rte_ring *cring_rx;
177 int user_port_id; /* User given port no, parsed from port name */
178 int eth_port_id; /* ethernet device port id */
179 struct list list_node OVS_GUARDED_BY(dpdk_mutex);
187 struct dpdk_tx_queue *tx_q;
189 struct ovs_mutex mutex OVS_ACQ_AFTER(dpdk_mutex);
191 struct dpdk_mp *dpdk_mp;
195 struct netdev_stats stats;
197 uint8_t hwaddr[ETH_ADDR_LEN];
198 enum netdev_flags flags;
200 struct rte_eth_link link;
204 struct list list_node OVS_GUARDED_BY(dpdk_mutex);
207 struct netdev_rxq_dpdk {
208 struct netdev_rxq up;
212 static bool thread_is_pmd(void);
214 static int netdev_dpdk_construct(struct netdev *);
217 is_dpdk_class(const struct netdev_class *class)
219 return class->construct == netdev_dpdk_construct;
222 /* XXX: use dpdk malloc for entire OVS. infact huge page shld be used
223 * for all other sengments data, bss and text. */
226 dpdk_rte_mzalloc(size_t sz)
230 ptr = rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE);
237 /* XXX this function should be called only by pmd threads (or by non pmd
238 * threads holding the nonpmd_mempool_mutex) */
240 free_dpdk_buf(struct dpif_packet *p)
242 struct rte_mbuf *pkt = (struct rte_mbuf *) p;
244 rte_pktmbuf_free_seg(pkt);
248 __rte_pktmbuf_init(struct rte_mempool *mp,
249 void *opaque_arg OVS_UNUSED,
251 unsigned i OVS_UNUSED)
253 struct rte_mbuf *m = _m;
254 uint32_t buf_len = mp->elt_size - sizeof(struct dpif_packet);
256 RTE_MBUF_ASSERT(mp->elt_size >= sizeof(struct dpif_packet));
258 memset(m, 0, mp->elt_size);
260 /* start of buffer is just after mbuf structure */
261 m->buf_addr = (char *)m + sizeof(struct dpif_packet);
262 m->buf_physaddr = rte_mempool_virt2phy(mp, m) +
263 sizeof(struct dpif_packet);
264 m->buf_len = (uint16_t)buf_len;
266 /* keep some headroom between start of buffer and data */
267 m->pkt.data = (char*) m->buf_addr + RTE_MIN(RTE_PKTMBUF_HEADROOM, m->buf_len);
269 /* init some constant fields */
270 m->type = RTE_MBUF_PKT;
273 m->pkt.in_port = 0xff;
277 ovs_rte_pktmbuf_init(struct rte_mempool *mp,
278 void *opaque_arg OVS_UNUSED,
280 unsigned i OVS_UNUSED)
282 struct rte_mbuf *m = _m;
284 __rte_pktmbuf_init(mp, opaque_arg, _m, i);
286 ofpbuf_init_dpdk((struct ofpbuf *) m, m->buf_len);
289 static struct dpdk_mp *
290 dpdk_mp_get(int socket_id, int mtu) OVS_REQUIRES(dpdk_mutex)
292 struct dpdk_mp *dmp = NULL;
293 char mp_name[RTE_MEMPOOL_NAMESIZE];
295 LIST_FOR_EACH (dmp, list_node, &dpdk_mp_list) {
296 if (dmp->socket_id == socket_id && dmp->mtu == mtu) {
302 dmp = dpdk_rte_mzalloc(sizeof *dmp);
303 dmp->socket_id = socket_id;
307 if (snprintf(mp_name, RTE_MEMPOOL_NAMESIZE, "ovs_mp_%d_%d", dmp->mtu,
308 dmp->socket_id) < 0) {
312 dmp->mp = rte_mempool_create(mp_name, NB_MBUF, MBUF_SIZE(mtu),
314 sizeof(struct rte_pktmbuf_pool_private),
315 rte_pktmbuf_pool_init, NULL,
316 ovs_rte_pktmbuf_init, NULL,
319 if (dmp->mp == NULL) {
323 list_push_back(&dpdk_mp_list, &dmp->list_node);
328 dpdk_mp_put(struct dpdk_mp *dmp)
336 ovs_assert(dmp->refcount >= 0);
339 /* I could not find any API to destroy mp. */
340 if (dmp->refcount == 0) {
341 list_delete(dmp->list_node);
342 /* destroy mp-pool. */
348 check_link_status(struct netdev_dpdk *dev)
350 struct rte_eth_link link;
352 rte_eth_link_get_nowait(dev->port_id, &link);
354 if (dev->link.link_status != link.link_status) {
355 netdev_change_seq_changed(&dev->up);
357 dev->link_reset_cnt++;
359 if (dev->link.link_status) {
360 VLOG_DBG_RL(&rl, "Port %d Link Up - speed %u Mbps - %s",
361 dev->port_id, (unsigned)dev->link.link_speed,
362 (dev->link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
363 ("full-duplex") : ("half-duplex"));
365 VLOG_DBG_RL(&rl, "Port %d Link Down", dev->port_id);
371 dpdk_watchdog(void *dummy OVS_UNUSED)
373 struct netdev_dpdk *dev;
375 pthread_detach(pthread_self());
378 ovs_mutex_lock(&dpdk_mutex);
379 LIST_FOR_EACH (dev, list_node, &dpdk_list) {
380 ovs_mutex_lock(&dev->mutex);
381 check_link_status(dev);
382 ovs_mutex_unlock(&dev->mutex);
384 ovs_mutex_unlock(&dpdk_mutex);
385 xsleep(DPDK_PORT_WATCHDOG_INTERVAL);
392 dpdk_eth_dev_init(struct netdev_dpdk *dev) OVS_REQUIRES(dpdk_mutex)
394 struct rte_pktmbuf_pool_private *mbp_priv;
395 struct ether_addr eth_addr;
399 if (dev->port_id < 0 || dev->port_id >= rte_eth_dev_count()) {
403 diag = rte_eth_dev_configure(dev->port_id, dev->up.n_rxq, dev->up.n_txq,
406 VLOG_ERR("eth dev config error %d",diag);
410 for (i = 0; i < dev->up.n_txq; i++) {
411 diag = rte_eth_tx_queue_setup(dev->port_id, i, NIC_PORT_TX_Q_SIZE,
412 dev->socket_id, &tx_conf);
414 VLOG_ERR("eth dev tx queue setup error %d",diag);
419 for (i = 0; i < dev->up.n_rxq; i++) {
420 diag = rte_eth_rx_queue_setup(dev->port_id, i, NIC_PORT_RX_Q_SIZE,
422 &rx_conf, dev->dpdk_mp->mp);
424 VLOG_ERR("eth dev rx queue setup error %d",diag);
429 diag = rte_eth_dev_start(dev->port_id);
431 VLOG_ERR("eth dev start error %d",diag);
435 rte_eth_promiscuous_enable(dev->port_id);
436 rte_eth_allmulticast_enable(dev->port_id);
438 memset(ð_addr, 0x0, sizeof(eth_addr));
439 rte_eth_macaddr_get(dev->port_id, ð_addr);
440 VLOG_INFO_RL(&rl, "Port %d: "ETH_ADDR_FMT"",
441 dev->port_id, ETH_ADDR_ARGS(eth_addr.addr_bytes));
443 memcpy(dev->hwaddr, eth_addr.addr_bytes, ETH_ADDR_LEN);
444 rte_eth_link_get_nowait(dev->port_id, &dev->link);
446 mbp_priv = rte_mempool_get_priv(dev->dpdk_mp->mp);
447 dev->buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
449 dev->flags = NETDEV_UP | NETDEV_PROMISC;
453 static struct netdev_dpdk *
454 netdev_dpdk_cast(const struct netdev *netdev)
456 return CONTAINER_OF(netdev, struct netdev_dpdk, up);
459 static struct netdev *
460 netdev_dpdk_alloc(void)
462 struct netdev_dpdk *netdev = dpdk_rte_mzalloc(sizeof *netdev);
467 netdev_dpdk_set_txq(struct netdev_dpdk *netdev, unsigned int n_txqs)
471 netdev->tx_q = dpdk_rte_mzalloc(n_txqs * sizeof *netdev->tx_q);
472 for (i = 0; i < n_txqs; i++) {
473 rte_spinlock_init(&netdev->tx_q[i].tx_lock);
478 netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no)
479 OVS_REQUIRES(dpdk_mutex)
481 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
484 ovs_mutex_init(&netdev->mutex);
486 ovs_mutex_lock(&netdev->mutex);
488 netdev_dpdk_set_txq(netdev, NR_QUEUE);
489 netdev->port_id = port_no;
491 netdev->mtu = ETHER_MTU;
492 netdev->max_packet_len = MTU_TO_MAX_LEN(netdev->mtu);
494 /* XXX: need to discover device node at run time. */
495 netdev->socket_id = SOCKET0;
497 netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, netdev->mtu);
498 if (!netdev->dpdk_mp) {
503 netdev_->n_txq = NR_QUEUE;
504 netdev_->n_rxq = NR_QUEUE;
505 err = dpdk_eth_dev_init(netdev);
510 list_push_back(&dpdk_list, &netdev->list_node);
514 rte_free(netdev->tx_q);
516 ovs_mutex_unlock(&netdev->mutex);
521 dpdk_dev_parse_name(const char dev_name[], const char prefix[],
522 unsigned int *port_no)
526 if (strncmp(dev_name, prefix, strlen(prefix))) {
530 cport = dev_name + strlen(prefix);
531 *port_no = strtol(cport, 0, 0); /* string must be null terminated */
536 netdev_dpdk_construct(struct netdev *netdev)
538 unsigned int port_no;
541 if (rte_eal_init_ret) {
542 return rte_eal_init_ret;
545 /* Names always start with "dpdk" */
546 err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no);
551 ovs_mutex_lock(&dpdk_mutex);
552 err = netdev_dpdk_init(netdev, port_no);
553 ovs_mutex_unlock(&dpdk_mutex);
558 netdev_dpdk_destruct(struct netdev *netdev_)
560 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
562 ovs_mutex_lock(&dev->mutex);
563 rte_eth_dev_stop(dev->port_id);
564 ovs_mutex_unlock(&dev->mutex);
566 ovs_mutex_lock(&dpdk_mutex);
568 list_remove(&dev->list_node);
569 dpdk_mp_put(dev->dpdk_mp);
570 ovs_mutex_unlock(&dpdk_mutex);
572 ovs_mutex_destroy(&dev->mutex);
576 netdev_dpdk_dealloc(struct netdev *netdev_)
578 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
584 netdev_dpdk_get_config(const struct netdev *netdev_, struct smap *args)
586 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
588 ovs_mutex_lock(&dev->mutex);
590 /* XXX: Allow to configure number of queues. */
591 smap_add_format(args, "configured_rx_queues", "%u", netdev_->n_rxq);
592 smap_add_format(args, "configured_tx_queues", "%u", netdev_->n_rxq);
593 ovs_mutex_unlock(&dev->mutex);
599 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
601 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
603 return netdev->socket_id;
606 /* Sets the number of tx queues and rx queues for the dpdk interface.
607 * If the configuration fails, do not try restoring its old configuration
608 * and just returns the error. */
610 netdev_dpdk_set_multiq(struct netdev *netdev_, unsigned int n_txq,
613 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
616 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
620 ovs_mutex_lock(&netdev->mutex);
621 rte_eth_dev_stop(netdev->port_id);
622 netdev->up.n_txq = n_txq;
623 netdev->up.n_rxq = n_rxq;
624 err = dpdk_eth_dev_init(netdev);
625 if (!err && netdev->up.n_txq != n_txq) {
626 rte_free(netdev->tx_q);
627 netdev_dpdk_set_txq(netdev, n_txq);
629 ovs_mutex_unlock(&netdev->mutex);
634 static struct netdev_rxq *
635 netdev_dpdk_rxq_alloc(void)
637 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
642 static struct netdev_rxq_dpdk *
643 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
645 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
649 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
651 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
652 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
654 ovs_mutex_lock(&netdev->mutex);
655 rx->port_id = netdev->port_id;
656 ovs_mutex_unlock(&netdev->mutex);
662 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
667 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
669 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
675 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
677 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
680 while (nb_tx != txq->count) {
683 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
692 if (OVS_UNLIKELY(nb_tx != txq->count)) {
693 /* free buffers, which we couldn't transmit, one at a time (each
694 * packet could come from a different mempool) */
697 for (i = nb_tx; i < txq->count; i++) {
698 rte_pktmbuf_free_seg(txq->burst_pkts[i]);
700 ovs_mutex_lock(&dev->mutex);
701 dev->stats.tx_dropped += txq->count-nb_tx;
702 ovs_mutex_unlock(&dev->mutex);
706 txq->tsc = rte_get_timer_cycles();
710 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
712 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
714 if (txq->count == 0) {
717 rte_spinlock_lock(&txq->tx_lock);
718 dpdk_queue_flush__(dev, qid);
719 rte_spinlock_unlock(&txq->tx_lock);
723 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dpif_packet **packets,
726 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
727 struct netdev *netdev = rx->up.netdev;
728 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
731 /* There is only one tx queue for this core. Do not flush other
733 if (rxq_->queue_id == rte_lcore_id()) {
734 dpdk_queue_flush(dev, rxq_->queue_id);
737 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
738 (struct rte_mbuf **) packets,
739 MIN((int)NETDEV_MAX_RX_BATCH,
740 (int)MAX_RX_QUEUE_LEN));
751 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
752 struct rte_mbuf **pkts, int cnt)
754 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
759 rte_spinlock_lock(&txq->tx_lock);
761 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
762 int tocopy = MIN(freeslots, cnt-i);
764 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
765 tocopy * sizeof (struct rte_mbuf *));
767 txq->count += tocopy;
770 if (txq->count == MAX_TX_QUEUE_LEN) {
771 dpdk_queue_flush__(dev, qid);
773 diff_tsc = rte_get_timer_cycles() - txq->tsc;
774 if (diff_tsc >= DRAIN_TSC) {
775 dpdk_queue_flush__(dev, qid);
778 rte_spinlock_unlock(&txq->tx_lock);
781 /* Tx function. Transmit packets indefinitely */
783 dpdk_do_tx_copy(struct netdev *netdev, struct dpif_packet ** pkts, int cnt)
784 OVS_NO_THREAD_SAFETY_ANALYSIS
786 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
787 struct rte_mbuf *mbufs[cnt];
792 /* If we are on a non pmd thread we have to use the mempool mutex, because
793 * every non pmd thread shares the same mempool cache */
795 if (!thread_is_pmd()) {
796 ovs_mutex_lock(&nonpmd_mempool_mutex);
799 for (i = 0; i < cnt; i++) {
800 int size = ofpbuf_size(&pkts[i]->ofpbuf);
802 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
803 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
804 (int)size , dev->max_packet_len);
810 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
812 if (!mbufs[newcnt]) {
817 /* We have to do a copy for now */
818 memcpy(mbufs[newcnt]->pkt.data, ofpbuf_data(&pkts[i]->ofpbuf), size);
820 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
821 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
826 if (OVS_UNLIKELY(dropped)) {
827 ovs_mutex_lock(&dev->mutex);
828 dev->stats.tx_dropped += dropped;
829 ovs_mutex_unlock(&dev->mutex);
832 dpdk_queue_pkts(dev, NON_PMD_THREAD_TX_QUEUE, mbufs, newcnt);
833 dpdk_queue_flush(dev, NON_PMD_THREAD_TX_QUEUE);
835 if (!thread_is_pmd()) {
836 ovs_mutex_unlock(&nonpmd_mempool_mutex);
841 netdev_dpdk_send(struct netdev *netdev, int qid, struct dpif_packet **pkts,
842 int cnt, bool may_steal)
844 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
848 if (!may_steal || pkts[0]->ofpbuf.source != OFPBUF_DPDK) {
849 dpdk_do_tx_copy(netdev, pkts, cnt);
852 for (i = 0; i < cnt; i++) {
853 dpif_packet_delete(pkts[i]);
860 qid = rte_lcore_id();
862 for (i = 0; i < cnt; i++) {
863 int size = ofpbuf_size(&pkts[i]->ofpbuf);
864 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
865 if (next_tx_idx != i) {
866 dpdk_queue_pkts(dev, qid,
867 (struct rte_mbuf **)&pkts[next_tx_idx],
871 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
872 (int)size , dev->max_packet_len);
874 dpif_packet_delete(pkts[i]);
879 if (next_tx_idx != cnt) {
880 dpdk_queue_pkts(dev, qid,
881 (struct rte_mbuf **)&pkts[next_tx_idx],
885 if (OVS_UNLIKELY(dropped)) {
886 ovs_mutex_lock(&dev->mutex);
887 dev->stats.tx_dropped += dropped;
888 ovs_mutex_unlock(&dev->mutex);
897 netdev_dpdk_set_etheraddr(struct netdev *netdev,
898 const uint8_t mac[ETH_ADDR_LEN])
900 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
902 ovs_mutex_lock(&dev->mutex);
903 if (!eth_addr_equals(dev->hwaddr, mac)) {
904 memcpy(dev->hwaddr, mac, ETH_ADDR_LEN);
905 netdev_change_seq_changed(netdev);
907 ovs_mutex_unlock(&dev->mutex);
913 netdev_dpdk_get_etheraddr(const struct netdev *netdev,
914 uint8_t mac[ETH_ADDR_LEN])
916 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
918 ovs_mutex_lock(&dev->mutex);
919 memcpy(mac, dev->hwaddr, ETH_ADDR_LEN);
920 ovs_mutex_unlock(&dev->mutex);
926 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
928 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
930 ovs_mutex_lock(&dev->mutex);
932 ovs_mutex_unlock(&dev->mutex);
938 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
940 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
942 struct dpdk_mp *old_mp;
945 ovs_mutex_lock(&dpdk_mutex);
946 ovs_mutex_lock(&dev->mutex);
947 if (dev->mtu == mtu) {
952 mp = dpdk_mp_get(dev->socket_id, dev->mtu);
958 rte_eth_dev_stop(dev->port_id);
961 old_mp = dev->dpdk_mp;
964 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
966 err = dpdk_eth_dev_init(dev);
970 dev->dpdk_mp = old_mp;
971 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
972 dpdk_eth_dev_init(dev);
977 netdev_change_seq_changed(netdev);
979 ovs_mutex_unlock(&dev->mutex);
980 ovs_mutex_unlock(&dpdk_mutex);
985 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
988 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
990 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
991 struct rte_eth_stats rte_stats;
994 netdev_dpdk_get_carrier(netdev, &gg);
995 ovs_mutex_lock(&dev->mutex);
996 rte_eth_stats_get(dev->port_id, &rte_stats);
998 memset(stats, 0, sizeof(*stats));
1000 stats->rx_packets = rte_stats.ipackets;
1001 stats->tx_packets = rte_stats.opackets;
1002 stats->rx_bytes = rte_stats.ibytes;
1003 stats->tx_bytes = rte_stats.obytes;
1004 stats->rx_errors = rte_stats.ierrors;
1005 stats->tx_errors = rte_stats.oerrors;
1006 stats->multicast = rte_stats.imcasts;
1008 stats->tx_dropped = dev->stats.tx_dropped;
1009 ovs_mutex_unlock(&dev->mutex);
1015 netdev_dpdk_get_features(const struct netdev *netdev_,
1016 enum netdev_features *current,
1017 enum netdev_features *advertised OVS_UNUSED,
1018 enum netdev_features *supported OVS_UNUSED,
1019 enum netdev_features *peer OVS_UNUSED)
1021 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1022 struct rte_eth_link link;
1024 ovs_mutex_lock(&dev->mutex);
1026 ovs_mutex_unlock(&dev->mutex);
1028 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
1029 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
1030 *current = NETDEV_F_AUTONEG;
1032 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
1033 if (link.link_speed == ETH_LINK_SPEED_10) {
1034 *current = NETDEV_F_10MB_HD;
1036 if (link.link_speed == ETH_LINK_SPEED_100) {
1037 *current = NETDEV_F_100MB_HD;
1039 if (link.link_speed == ETH_LINK_SPEED_1000) {
1040 *current = NETDEV_F_1GB_HD;
1042 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
1043 if (link.link_speed == ETH_LINK_SPEED_10) {
1044 *current = NETDEV_F_10MB_FD;
1046 if (link.link_speed == ETH_LINK_SPEED_100) {
1047 *current = NETDEV_F_100MB_FD;
1049 if (link.link_speed == ETH_LINK_SPEED_1000) {
1050 *current = NETDEV_F_1GB_FD;
1052 if (link.link_speed == ETH_LINK_SPEED_10000) {
1053 *current = NETDEV_F_10GB_FD;
1061 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1063 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1066 ovs_mutex_lock(&dev->mutex);
1067 ifindex = dev->port_id;
1068 ovs_mutex_unlock(&dev->mutex);
1074 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1076 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1078 ovs_mutex_lock(&dev->mutex);
1079 check_link_status(dev);
1080 *carrier = dev->link.link_status;
1081 ovs_mutex_unlock(&dev->mutex);
1086 static long long int
1087 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1089 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1090 long long int carrier_resets;
1092 ovs_mutex_lock(&dev->mutex);
1093 carrier_resets = dev->link_reset_cnt;
1094 ovs_mutex_unlock(&dev->mutex);
1096 return carrier_resets;
1100 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1101 long long int interval OVS_UNUSED)
1107 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1108 enum netdev_flags off, enum netdev_flags on,
1109 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1113 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1117 *old_flagsp = dev->flags;
1121 if (dev->flags == *old_flagsp) {
1125 if (dev->flags & NETDEV_UP) {
1126 err = rte_eth_dev_start(dev->port_id);
1131 if (dev->flags & NETDEV_PROMISC) {
1132 rte_eth_promiscuous_enable(dev->port_id);
1135 if (!(dev->flags & NETDEV_UP)) {
1136 rte_eth_dev_stop(dev->port_id);
1143 netdev_dpdk_update_flags(struct netdev *netdev_,
1144 enum netdev_flags off, enum netdev_flags on,
1145 enum netdev_flags *old_flagsp)
1147 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1150 ovs_mutex_lock(&netdev->mutex);
1151 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1152 ovs_mutex_unlock(&netdev->mutex);
1158 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1160 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1161 struct rte_eth_dev_info dev_info;
1163 if (dev->port_id < 0)
1166 ovs_mutex_lock(&dev->mutex);
1167 rte_eth_dev_info_get(dev->port_id, &dev_info);
1168 ovs_mutex_unlock(&dev->mutex);
1170 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1172 smap_add_format(args, "port_no", "%d", dev->port_id);
1173 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1174 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1175 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1176 smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen);
1177 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1178 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1179 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1180 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1181 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1182 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1184 smap_add_format(args, "pci-vendor_id", "0x%u", dev_info.pci_dev->id.vendor_id);
1185 smap_add_format(args, "pci-device_id", "0x%x", dev_info.pci_dev->id.device_id);
1191 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1192 OVS_REQUIRES(dev->mutex)
1194 enum netdev_flags old_flags;
1197 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1199 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1204 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1205 const char *argv[], void *aux OVS_UNUSED)
1209 if (!strcasecmp(argv[argc - 1], "up")) {
1211 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1214 unixctl_command_reply_error(conn, "Invalid Admin State");
1219 struct netdev *netdev = netdev_from_name(argv[1]);
1220 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1221 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1223 ovs_mutex_lock(&dpdk_dev->mutex);
1224 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1225 ovs_mutex_unlock(&dpdk_dev->mutex);
1227 netdev_close(netdev);
1229 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1230 netdev_close(netdev);
1234 struct netdev_dpdk *netdev;
1236 ovs_mutex_lock(&dpdk_mutex);
1237 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1238 ovs_mutex_lock(&netdev->mutex);
1239 netdev_dpdk_set_admin_state__(netdev, up);
1240 ovs_mutex_unlock(&netdev->mutex);
1242 ovs_mutex_unlock(&dpdk_mutex);
1244 unixctl_command_reply(conn, "OK");
1248 dpdk_common_init(void)
1250 unixctl_command_register("netdev-dpdk/set-admin-state",
1251 "[netdev] up|down", 1, 2,
1252 netdev_dpdk_set_admin_state, NULL);
1254 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
1258 dpdk_class_init(void)
1262 result = rte_eal_pci_probe();
1264 VLOG_ERR("Cannot probe PCI");
1268 VLOG_INFO("Ethernet Device Count: %d", (int)rte_eth_dev_count());
1276 dpdk_ring_create(const char dev_name[], unsigned int port_no,
1277 unsigned int *eth_port_id)
1279 struct dpdk_ring *ivshmem;
1283 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
1284 if (ivshmem == NULL) {
1288 err = snprintf(ring_name, 10, "%s_tx", dev_name);
1293 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0, 0);
1294 if (ivshmem->cring_tx == NULL) {
1299 err = snprintf(ring_name, 10, "%s_rx", dev_name);
1304 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0, 0);
1305 if (ivshmem->cring_rx == NULL) {
1310 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
1311 &ivshmem->cring_tx, 1, SOCKET0);
1318 ivshmem->user_port_id = port_no;
1319 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
1320 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
1322 *eth_port_id = ivshmem->eth_port_id;
1327 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
1329 struct dpdk_ring *ivshmem;
1330 unsigned int port_no;
1333 /* Names always start with "dpdkr" */
1334 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
1339 /* look through our list to find the device */
1340 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
1341 if (ivshmem->user_port_id == port_no) {
1342 VLOG_INFO("Found dpdk ring device %s:\n", dev_name);
1343 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
1347 /* Need to create the device rings */
1348 return dpdk_ring_create(dev_name, port_no, eth_port_id);
1352 netdev_dpdk_ring_construct(struct netdev *netdev)
1354 unsigned int port_no = 0;
1357 if (rte_eal_init_ret) {
1358 return rte_eal_init_ret;
1361 ovs_mutex_lock(&dpdk_mutex);
1363 err = dpdk_ring_open(netdev->name, &port_no);
1368 err = netdev_dpdk_init(netdev, port_no);
1371 ovs_mutex_unlock(&dpdk_mutex);
1375 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, MULTIQ) \
1379 NULL, /* netdev_dpdk_run */ \
1380 NULL, /* netdev_dpdk_wait */ \
1382 netdev_dpdk_alloc, \
1384 netdev_dpdk_destruct, \
1385 netdev_dpdk_dealloc, \
1386 netdev_dpdk_get_config, \
1387 NULL, /* netdev_dpdk_set_config */ \
1388 NULL, /* get_tunnel_config */ \
1389 netdev_dpdk_get_numa_id, /* get_numa_id */ \
1390 MULTIQ, /* set_multiq */ \
1392 netdev_dpdk_send, /* send */ \
1393 NULL, /* send_wait */ \
1395 netdev_dpdk_set_etheraddr, \
1396 netdev_dpdk_get_etheraddr, \
1397 netdev_dpdk_get_mtu, \
1398 netdev_dpdk_set_mtu, \
1399 netdev_dpdk_get_ifindex, \
1400 netdev_dpdk_get_carrier, \
1401 netdev_dpdk_get_carrier_resets, \
1402 netdev_dpdk_set_miimon, \
1403 netdev_dpdk_get_stats, \
1404 netdev_dpdk_get_features, \
1405 NULL, /* set_advertisements */ \
1407 NULL, /* set_policing */ \
1408 NULL, /* get_qos_types */ \
1409 NULL, /* get_qos_capabilities */ \
1410 NULL, /* get_qos */ \
1411 NULL, /* set_qos */ \
1412 NULL, /* get_queue */ \
1413 NULL, /* set_queue */ \
1414 NULL, /* delete_queue */ \
1415 NULL, /* get_queue_stats */ \
1416 NULL, /* queue_dump_start */ \
1417 NULL, /* queue_dump_next */ \
1418 NULL, /* queue_dump_done */ \
1419 NULL, /* dump_queue_stats */ \
1421 NULL, /* get_in4 */ \
1422 NULL, /* set_in4 */ \
1423 NULL, /* get_in6 */ \
1424 NULL, /* add_router */ \
1425 NULL, /* get_next_hop */ \
1426 netdev_dpdk_get_status, \
1427 NULL, /* arp_lookup */ \
1429 netdev_dpdk_update_flags, \
1431 netdev_dpdk_rxq_alloc, \
1432 netdev_dpdk_rxq_construct, \
1433 netdev_dpdk_rxq_destruct, \
1434 netdev_dpdk_rxq_dealloc, \
1435 netdev_dpdk_rxq_recv, \
1436 NULL, /* rx_wait */ \
1437 NULL, /* rxq_drain */ \
1441 dpdk_init(int argc, char **argv)
1445 if (argc < 2 || strcmp(argv[1], "--dpdk"))
1448 /* Make sure program name passed to rte_eal_init() is vswitchd. */
1454 /* Make sure things are initialized ... */
1455 result = rte_eal_init(argc, argv);
1457 ovs_abort(result, "Cannot init EAL\n");
1460 rte_memzone_dump(stdout);
1461 rte_eal_init_ret = 0;
1463 if (argc > result) {
1464 argv[result] = argv[0];
1467 /* We are called from the main thread here */
1468 thread_set_nonpmd();
1473 const struct netdev_class dpdk_class =
1477 netdev_dpdk_construct,
1478 netdev_dpdk_set_multiq);
1480 const struct netdev_class dpdk_ring_class =
1484 netdev_dpdk_ring_construct,
1488 netdev_dpdk_register(void)
1490 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
1492 if (rte_eal_init_ret) {
1496 if (ovsthread_once_start(&once)) {
1498 netdev_register_provider(&dpdk_class);
1499 netdev_register_provider(&dpdk_ring_class);
1500 ovsthread_once_done(&once);
1505 pmd_thread_setaffinity_cpu(int cpu)
1511 CPU_SET(cpu, &cpuset);
1512 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
1514 VLOG_ERR("Thread affinity error %d",err);
1517 /* lcore_id 0 is reseved for use by non pmd threads. */
1518 RTE_PER_LCORE(_lcore_id) = cpu + 1;
1524 thread_set_nonpmd(void)
1526 /* We cannot have RTE_MAX_LCORE pmd threads, because lcore_id 0 is reserved
1527 * for non pmd threads */
1528 BUILD_ASSERT(NR_PMD_THREADS < RTE_MAX_LCORE);
1529 /* We have to use 0 to allow non pmd threads to perform certain DPDK
1530 * operations, like rte_eth_dev_configure(). */
1531 RTE_PER_LCORE(_lcore_id) = 0;
1537 return rte_lcore_id() != 0;