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 struct dpdk_tx_queue {
158 rte_spinlock_t tx_lock;
161 struct rte_mbuf *burst_pkts[MAX_TX_QUEUE_LEN];
164 /* dpdk has no way to remove dpdk ring ethernet devices
165 so we have to keep them around once they've been created
168 static struct list dpdk_ring_list OVS_GUARDED_BY(dpdk_mutex)
169 = LIST_INITIALIZER(&dpdk_ring_list);
172 /* For the client rings */
173 struct rte_ring *cring_tx;
174 struct rte_ring *cring_rx;
175 int user_port_id; /* User given port no, parsed from port name */
176 int eth_port_id; /* ethernet device port id */
177 struct list list_node OVS_GUARDED_BY(dpdk_mutex);
185 struct dpdk_tx_queue tx_q[NR_QUEUE];
187 struct ovs_mutex mutex OVS_ACQ_AFTER(dpdk_mutex);
189 struct dpdk_mp *dpdk_mp;
193 struct netdev_stats stats;
195 uint8_t hwaddr[ETH_ADDR_LEN];
196 enum netdev_flags flags;
198 struct rte_eth_link link;
202 struct list list_node OVS_GUARDED_BY(dpdk_mutex);
205 struct netdev_rxq_dpdk {
206 struct netdev_rxq up;
210 static bool thread_is_pmd(void);
212 static int netdev_dpdk_construct(struct netdev *);
215 is_dpdk_class(const struct netdev_class *class)
217 return class->construct == netdev_dpdk_construct;
220 /* XXX: use dpdk malloc for entire OVS. infact huge page shld be used
221 * for all other sengments data, bss and text. */
224 dpdk_rte_mzalloc(size_t sz)
228 ptr = rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE);
235 /* XXX this function should be called only by pmd threads (or by non pmd
236 * threads holding the nonpmd_mempool_mutex) */
238 free_dpdk_buf(struct dpif_packet *p)
240 struct rte_mbuf *pkt = (struct rte_mbuf *) p;
242 rte_pktmbuf_free_seg(pkt);
246 __rte_pktmbuf_init(struct rte_mempool *mp,
247 void *opaque_arg OVS_UNUSED,
249 unsigned i OVS_UNUSED)
251 struct rte_mbuf *m = _m;
252 uint32_t buf_len = mp->elt_size - sizeof(struct dpif_packet);
254 RTE_MBUF_ASSERT(mp->elt_size >= sizeof(struct dpif_packet));
256 memset(m, 0, mp->elt_size);
258 /* start of buffer is just after mbuf structure */
259 m->buf_addr = (char *)m + sizeof(struct dpif_packet);
260 m->buf_physaddr = rte_mempool_virt2phy(mp, m) +
261 sizeof(struct dpif_packet);
262 m->buf_len = (uint16_t)buf_len;
264 /* keep some headroom between start of buffer and data */
265 m->pkt.data = (char*) m->buf_addr + RTE_MIN(RTE_PKTMBUF_HEADROOM, m->buf_len);
267 /* init some constant fields */
268 m->type = RTE_MBUF_PKT;
271 m->pkt.in_port = 0xff;
275 ovs_rte_pktmbuf_init(struct rte_mempool *mp,
276 void *opaque_arg OVS_UNUSED,
278 unsigned i OVS_UNUSED)
280 struct rte_mbuf *m = _m;
282 __rte_pktmbuf_init(mp, opaque_arg, _m, i);
284 ofpbuf_init_dpdk((struct ofpbuf *) m, m->buf_len);
287 static struct dpdk_mp *
288 dpdk_mp_get(int socket_id, int mtu) OVS_REQUIRES(dpdk_mutex)
290 struct dpdk_mp *dmp = NULL;
291 char mp_name[RTE_MEMPOOL_NAMESIZE];
293 LIST_FOR_EACH (dmp, list_node, &dpdk_mp_list) {
294 if (dmp->socket_id == socket_id && dmp->mtu == mtu) {
300 dmp = dpdk_rte_mzalloc(sizeof *dmp);
301 dmp->socket_id = socket_id;
305 if (snprintf(mp_name, RTE_MEMPOOL_NAMESIZE, "ovs_mp_%d_%d", dmp->mtu,
306 dmp->socket_id) < 0) {
310 dmp->mp = rte_mempool_create(mp_name, NB_MBUF, MBUF_SIZE(mtu),
312 sizeof(struct rte_pktmbuf_pool_private),
313 rte_pktmbuf_pool_init, NULL,
314 ovs_rte_pktmbuf_init, NULL,
317 if (dmp->mp == NULL) {
321 list_push_back(&dpdk_mp_list, &dmp->list_node);
326 dpdk_mp_put(struct dpdk_mp *dmp)
334 ovs_assert(dmp->refcount >= 0);
337 /* I could not find any API to destroy mp. */
338 if (dmp->refcount == 0) {
339 list_delete(dmp->list_node);
340 /* destroy mp-pool. */
346 check_link_status(struct netdev_dpdk *dev)
348 struct rte_eth_link link;
350 rte_eth_link_get_nowait(dev->port_id, &link);
352 if (dev->link.link_status != link.link_status) {
353 netdev_change_seq_changed(&dev->up);
355 dev->link_reset_cnt++;
357 if (dev->link.link_status) {
358 VLOG_DBG_RL(&rl, "Port %d Link Up - speed %u Mbps - %s",
359 dev->port_id, (unsigned)dev->link.link_speed,
360 (dev->link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
361 ("full-duplex") : ("half-duplex"));
363 VLOG_DBG_RL(&rl, "Port %d Link Down", dev->port_id);
369 dpdk_watchdog(void *dummy OVS_UNUSED)
371 struct netdev_dpdk *dev;
373 pthread_detach(pthread_self());
376 ovs_mutex_lock(&dpdk_mutex);
377 LIST_FOR_EACH (dev, list_node, &dpdk_list) {
378 ovs_mutex_lock(&dev->mutex);
379 check_link_status(dev);
380 ovs_mutex_unlock(&dev->mutex);
382 ovs_mutex_unlock(&dpdk_mutex);
383 xsleep(DPDK_PORT_WATCHDOG_INTERVAL);
390 dpdk_eth_dev_init(struct netdev_dpdk *dev) OVS_REQUIRES(dpdk_mutex)
392 struct rte_pktmbuf_pool_private *mbp_priv;
393 struct ether_addr eth_addr;
397 if (dev->port_id < 0 || dev->port_id >= rte_eth_dev_count()) {
401 diag = rte_eth_dev_configure(dev->port_id, NR_QUEUE, NR_QUEUE, &port_conf);
403 VLOG_ERR("eth dev config error %d",diag);
407 for (i = 0; i < NR_QUEUE; i++) {
408 diag = rte_eth_tx_queue_setup(dev->port_id, i, NIC_PORT_TX_Q_SIZE,
409 dev->socket_id, &tx_conf);
411 VLOG_ERR("eth dev tx queue setup error %d",diag);
416 for (i = 0; i < NR_QUEUE; i++) {
417 diag = rte_eth_rx_queue_setup(dev->port_id, i, NIC_PORT_RX_Q_SIZE,
419 &rx_conf, dev->dpdk_mp->mp);
421 VLOG_ERR("eth dev rx queue setup error %d",diag);
426 diag = rte_eth_dev_start(dev->port_id);
428 VLOG_ERR("eth dev start error %d",diag);
432 rte_eth_promiscuous_enable(dev->port_id);
433 rte_eth_allmulticast_enable(dev->port_id);
435 memset(ð_addr, 0x0, sizeof(eth_addr));
436 rte_eth_macaddr_get(dev->port_id, ð_addr);
437 VLOG_INFO_RL(&rl, "Port %d: "ETH_ADDR_FMT"",
438 dev->port_id, ETH_ADDR_ARGS(eth_addr.addr_bytes));
440 memcpy(dev->hwaddr, eth_addr.addr_bytes, ETH_ADDR_LEN);
441 rte_eth_link_get_nowait(dev->port_id, &dev->link);
443 mbp_priv = rte_mempool_get_priv(dev->dpdk_mp->mp);
444 dev->buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
446 dev->flags = NETDEV_UP | NETDEV_PROMISC;
450 static struct netdev_dpdk *
451 netdev_dpdk_cast(const struct netdev *netdev)
453 return CONTAINER_OF(netdev, struct netdev_dpdk, up);
456 static struct netdev *
457 netdev_dpdk_alloc(void)
459 struct netdev_dpdk *netdev = dpdk_rte_mzalloc(sizeof *netdev);
464 netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no) OVS_REQUIRES(dpdk_mutex)
466 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
470 ovs_mutex_init(&netdev->mutex);
472 ovs_mutex_lock(&netdev->mutex);
474 for (i = 0; i < NR_QUEUE; i++) {
475 rte_spinlock_init(&netdev->tx_q[i].tx_lock);
478 netdev->port_id = port_no;
481 netdev->mtu = ETHER_MTU;
482 netdev->max_packet_len = MTU_TO_MAX_LEN(netdev->mtu);
484 /* XXX: need to discover device node at run time. */
485 netdev->socket_id = SOCKET0;
487 netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, netdev->mtu);
488 if (!netdev->dpdk_mp) {
493 err = dpdk_eth_dev_init(netdev);
497 netdev_->n_txq = NR_QUEUE;
498 netdev_->n_rxq = NR_QUEUE;
500 list_push_back(&dpdk_list, &netdev->list_node);
503 ovs_mutex_unlock(&netdev->mutex);
508 dpdk_dev_parse_name(const char dev_name[], const char prefix[],
509 unsigned int *port_no)
513 if (strncmp(dev_name, prefix, strlen(prefix))) {
517 cport = dev_name + strlen(prefix);
518 *port_no = strtol(cport, 0, 0); /* string must be null terminated */
523 netdev_dpdk_construct(struct netdev *netdev)
525 unsigned int port_no;
528 if (rte_eal_init_ret) {
529 return rte_eal_init_ret;
532 /* Names always start with "dpdk" */
533 err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no);
538 ovs_mutex_lock(&dpdk_mutex);
539 err = netdev_dpdk_init(netdev, port_no);
540 ovs_mutex_unlock(&dpdk_mutex);
545 netdev_dpdk_destruct(struct netdev *netdev_)
547 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
549 ovs_mutex_lock(&dev->mutex);
550 rte_eth_dev_stop(dev->port_id);
551 ovs_mutex_unlock(&dev->mutex);
553 ovs_mutex_lock(&dpdk_mutex);
554 list_remove(&dev->list_node);
555 dpdk_mp_put(dev->dpdk_mp);
556 ovs_mutex_unlock(&dpdk_mutex);
558 ovs_mutex_destroy(&dev->mutex);
562 netdev_dpdk_dealloc(struct netdev *netdev_)
564 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
570 netdev_dpdk_get_config(const struct netdev *netdev_, struct smap *args)
572 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
574 ovs_mutex_lock(&dev->mutex);
576 /* XXX: Allow to configure number of queues. */
577 smap_add_format(args, "configured_rx_queues", "%u", netdev_->n_rxq);
578 smap_add_format(args, "configured_tx_queues", "%u", netdev_->n_rxq);
579 ovs_mutex_unlock(&dev->mutex);
585 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
587 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
589 return netdev->socket_id;
592 static struct netdev_rxq *
593 netdev_dpdk_rxq_alloc(void)
595 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
600 static struct netdev_rxq_dpdk *
601 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
603 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
607 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
609 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
610 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
612 ovs_mutex_lock(&netdev->mutex);
613 rx->port_id = netdev->port_id;
614 ovs_mutex_unlock(&netdev->mutex);
620 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
625 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
627 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
633 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
635 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
638 while (nb_tx != txq->count) {
641 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
650 if (OVS_UNLIKELY(nb_tx != txq->count)) {
651 /* free buffers, which we couldn't transmit, one at a time (each
652 * packet could come from a different mempool) */
655 for (i = nb_tx; i < txq->count; i++) {
656 rte_pktmbuf_free_seg(txq->burst_pkts[i]);
658 ovs_mutex_lock(&dev->mutex);
659 dev->stats.tx_dropped += txq->count-nb_tx;
660 ovs_mutex_unlock(&dev->mutex);
664 txq->tsc = rte_get_timer_cycles();
668 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
670 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
672 if (txq->count == 0) {
675 rte_spinlock_lock(&txq->tx_lock);
676 dpdk_queue_flush__(dev, qid);
677 rte_spinlock_unlock(&txq->tx_lock);
681 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dpif_packet **packets,
684 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
685 struct netdev *netdev = rx->up.netdev;
686 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
689 dpdk_queue_flush(dev, rxq_->queue_id);
691 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
692 (struct rte_mbuf **) packets,
693 MIN((int)NETDEV_MAX_RX_BATCH,
694 (int)MAX_RX_QUEUE_LEN));
705 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
706 struct rte_mbuf **pkts, int cnt)
708 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
713 rte_spinlock_lock(&txq->tx_lock);
715 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
716 int tocopy = MIN(freeslots, cnt-i);
718 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
719 tocopy * sizeof (struct rte_mbuf *));
721 txq->count += tocopy;
724 if (txq->count == MAX_TX_QUEUE_LEN) {
725 dpdk_queue_flush__(dev, qid);
727 diff_tsc = rte_get_timer_cycles() - txq->tsc;
728 if (diff_tsc >= DRAIN_TSC) {
729 dpdk_queue_flush__(dev, qid);
732 rte_spinlock_unlock(&txq->tx_lock);
735 /* Tx function. Transmit packets indefinitely */
737 dpdk_do_tx_copy(struct netdev *netdev, struct dpif_packet ** pkts, int cnt)
738 OVS_NO_THREAD_SAFETY_ANALYSIS
740 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
741 struct rte_mbuf *mbufs[cnt];
746 /* If we are on a non pmd thread we have to use the mempool mutex, because
747 * every non pmd thread shares the same mempool cache */
749 if (!thread_is_pmd()) {
750 ovs_mutex_lock(&nonpmd_mempool_mutex);
753 for (i = 0; i < cnt; i++) {
754 int size = ofpbuf_size(&pkts[i]->ofpbuf);
756 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
757 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
758 (int)size , dev->max_packet_len);
764 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
766 if (!mbufs[newcnt]) {
771 /* We have to do a copy for now */
772 memcpy(mbufs[newcnt]->pkt.data, ofpbuf_data(&pkts[i]->ofpbuf), size);
774 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
775 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
780 if (OVS_UNLIKELY(dropped)) {
781 ovs_mutex_lock(&dev->mutex);
782 dev->stats.tx_dropped += dropped;
783 ovs_mutex_unlock(&dev->mutex);
786 dpdk_queue_pkts(dev, NON_PMD_THREAD_TX_QUEUE, mbufs, newcnt);
787 dpdk_queue_flush(dev, NON_PMD_THREAD_TX_QUEUE);
789 if (!thread_is_pmd()) {
790 ovs_mutex_unlock(&nonpmd_mempool_mutex);
795 netdev_dpdk_send(struct netdev *netdev, int qid, struct dpif_packet **pkts,
796 int cnt, bool may_steal)
798 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
802 if (!may_steal || pkts[0]->ofpbuf.source != OFPBUF_DPDK) {
803 dpdk_do_tx_copy(netdev, pkts, cnt);
806 for (i = 0; i < cnt; i++) {
807 dpif_packet_delete(pkts[i]);
814 qid = rte_lcore_id() % NR_QUEUE;
816 for (i = 0; i < cnt; i++) {
817 int size = ofpbuf_size(&pkts[i]->ofpbuf);
818 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
819 if (next_tx_idx != i) {
820 dpdk_queue_pkts(dev, qid,
821 (struct rte_mbuf **)&pkts[next_tx_idx],
825 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
826 (int)size , dev->max_packet_len);
828 dpif_packet_delete(pkts[i]);
833 if (next_tx_idx != cnt) {
834 dpdk_queue_pkts(dev, qid,
835 (struct rte_mbuf **)&pkts[next_tx_idx],
839 if (OVS_UNLIKELY(dropped)) {
840 ovs_mutex_lock(&dev->mutex);
841 dev->stats.tx_dropped += dropped;
842 ovs_mutex_unlock(&dev->mutex);
851 netdev_dpdk_set_etheraddr(struct netdev *netdev,
852 const uint8_t mac[ETH_ADDR_LEN])
854 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
856 ovs_mutex_lock(&dev->mutex);
857 if (!eth_addr_equals(dev->hwaddr, mac)) {
858 memcpy(dev->hwaddr, mac, ETH_ADDR_LEN);
859 netdev_change_seq_changed(netdev);
861 ovs_mutex_unlock(&dev->mutex);
867 netdev_dpdk_get_etheraddr(const struct netdev *netdev,
868 uint8_t mac[ETH_ADDR_LEN])
870 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
872 ovs_mutex_lock(&dev->mutex);
873 memcpy(mac, dev->hwaddr, ETH_ADDR_LEN);
874 ovs_mutex_unlock(&dev->mutex);
880 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
882 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
884 ovs_mutex_lock(&dev->mutex);
886 ovs_mutex_unlock(&dev->mutex);
892 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
894 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
896 struct dpdk_mp *old_mp;
899 ovs_mutex_lock(&dpdk_mutex);
900 ovs_mutex_lock(&dev->mutex);
901 if (dev->mtu == mtu) {
906 mp = dpdk_mp_get(dev->socket_id, dev->mtu);
912 rte_eth_dev_stop(dev->port_id);
915 old_mp = dev->dpdk_mp;
918 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
920 err = dpdk_eth_dev_init(dev);
924 dev->dpdk_mp = old_mp;
925 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
926 dpdk_eth_dev_init(dev);
931 netdev_change_seq_changed(netdev);
933 ovs_mutex_unlock(&dev->mutex);
934 ovs_mutex_unlock(&dpdk_mutex);
939 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
942 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
944 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
945 struct rte_eth_stats rte_stats;
948 netdev_dpdk_get_carrier(netdev, &gg);
949 ovs_mutex_lock(&dev->mutex);
950 rte_eth_stats_get(dev->port_id, &rte_stats);
952 memset(stats, 0, sizeof(*stats));
954 stats->rx_packets = rte_stats.ipackets;
955 stats->tx_packets = rte_stats.opackets;
956 stats->rx_bytes = rte_stats.ibytes;
957 stats->tx_bytes = rte_stats.obytes;
958 stats->rx_errors = rte_stats.ierrors;
959 stats->tx_errors = rte_stats.oerrors;
960 stats->multicast = rte_stats.imcasts;
962 stats->tx_dropped = dev->stats.tx_dropped;
963 ovs_mutex_unlock(&dev->mutex);
969 netdev_dpdk_get_features(const struct netdev *netdev_,
970 enum netdev_features *current,
971 enum netdev_features *advertised OVS_UNUSED,
972 enum netdev_features *supported OVS_UNUSED,
973 enum netdev_features *peer OVS_UNUSED)
975 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
976 struct rte_eth_link link;
978 ovs_mutex_lock(&dev->mutex);
980 ovs_mutex_unlock(&dev->mutex);
982 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
983 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
984 *current = NETDEV_F_AUTONEG;
986 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
987 if (link.link_speed == ETH_LINK_SPEED_10) {
988 *current = NETDEV_F_10MB_HD;
990 if (link.link_speed == ETH_LINK_SPEED_100) {
991 *current = NETDEV_F_100MB_HD;
993 if (link.link_speed == ETH_LINK_SPEED_1000) {
994 *current = NETDEV_F_1GB_HD;
996 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
997 if (link.link_speed == ETH_LINK_SPEED_10) {
998 *current = NETDEV_F_10MB_FD;
1000 if (link.link_speed == ETH_LINK_SPEED_100) {
1001 *current = NETDEV_F_100MB_FD;
1003 if (link.link_speed == ETH_LINK_SPEED_1000) {
1004 *current = NETDEV_F_1GB_FD;
1006 if (link.link_speed == ETH_LINK_SPEED_10000) {
1007 *current = NETDEV_F_10GB_FD;
1015 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1017 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1020 ovs_mutex_lock(&dev->mutex);
1021 ifindex = dev->port_id;
1022 ovs_mutex_unlock(&dev->mutex);
1028 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1030 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1032 ovs_mutex_lock(&dev->mutex);
1033 check_link_status(dev);
1034 *carrier = dev->link.link_status;
1035 ovs_mutex_unlock(&dev->mutex);
1040 static long long int
1041 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1043 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1044 long long int carrier_resets;
1046 ovs_mutex_lock(&dev->mutex);
1047 carrier_resets = dev->link_reset_cnt;
1048 ovs_mutex_unlock(&dev->mutex);
1050 return carrier_resets;
1054 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1055 long long int interval OVS_UNUSED)
1061 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1062 enum netdev_flags off, enum netdev_flags on,
1063 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1067 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1071 *old_flagsp = dev->flags;
1075 if (dev->flags == *old_flagsp) {
1079 if (dev->flags & NETDEV_UP) {
1080 err = rte_eth_dev_start(dev->port_id);
1085 if (dev->flags & NETDEV_PROMISC) {
1086 rte_eth_promiscuous_enable(dev->port_id);
1089 if (!(dev->flags & NETDEV_UP)) {
1090 rte_eth_dev_stop(dev->port_id);
1097 netdev_dpdk_update_flags(struct netdev *netdev_,
1098 enum netdev_flags off, enum netdev_flags on,
1099 enum netdev_flags *old_flagsp)
1101 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1104 ovs_mutex_lock(&netdev->mutex);
1105 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1106 ovs_mutex_unlock(&netdev->mutex);
1112 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1114 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1115 struct rte_eth_dev_info dev_info;
1117 if (dev->port_id < 0)
1120 ovs_mutex_lock(&dev->mutex);
1121 rte_eth_dev_info_get(dev->port_id, &dev_info);
1122 ovs_mutex_unlock(&dev->mutex);
1124 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1126 smap_add_format(args, "port_no", "%d", dev->port_id);
1127 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1128 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1129 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1130 smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen);
1131 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1132 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1133 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1134 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1135 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1136 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1138 smap_add_format(args, "pci-vendor_id", "0x%u", dev_info.pci_dev->id.vendor_id);
1139 smap_add_format(args, "pci-device_id", "0x%x", dev_info.pci_dev->id.device_id);
1145 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1146 OVS_REQUIRES(dev->mutex)
1148 enum netdev_flags old_flags;
1151 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1153 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1158 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1159 const char *argv[], void *aux OVS_UNUSED)
1163 if (!strcasecmp(argv[argc - 1], "up")) {
1165 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1168 unixctl_command_reply_error(conn, "Invalid Admin State");
1173 struct netdev *netdev = netdev_from_name(argv[1]);
1174 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1175 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1177 ovs_mutex_lock(&dpdk_dev->mutex);
1178 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1179 ovs_mutex_unlock(&dpdk_dev->mutex);
1181 netdev_close(netdev);
1183 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1184 netdev_close(netdev);
1188 struct netdev_dpdk *netdev;
1190 ovs_mutex_lock(&dpdk_mutex);
1191 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1192 ovs_mutex_lock(&netdev->mutex);
1193 netdev_dpdk_set_admin_state__(netdev, up);
1194 ovs_mutex_unlock(&netdev->mutex);
1196 ovs_mutex_unlock(&dpdk_mutex);
1198 unixctl_command_reply(conn, "OK");
1202 dpdk_common_init(void)
1204 unixctl_command_register("netdev-dpdk/set-admin-state",
1205 "[netdev] up|down", 1, 2,
1206 netdev_dpdk_set_admin_state, NULL);
1208 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
1212 dpdk_class_init(void)
1216 result = rte_eal_pci_probe();
1218 VLOG_ERR("Cannot probe PCI");
1222 VLOG_INFO("Ethernet Device Count: %d", (int)rte_eth_dev_count());
1230 dpdk_ring_create(const char dev_name[], unsigned int port_no,
1231 unsigned int *eth_port_id)
1233 struct dpdk_ring *ivshmem;
1237 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
1238 if (ivshmem == NULL) {
1242 err = snprintf(ring_name, 10, "%s_tx", dev_name);
1247 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0, 0);
1248 if (ivshmem->cring_tx == NULL) {
1253 err = snprintf(ring_name, 10, "%s_rx", dev_name);
1258 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0, 0);
1259 if (ivshmem->cring_rx == NULL) {
1264 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
1265 &ivshmem->cring_tx, 1, SOCKET0);
1272 ivshmem->user_port_id = port_no;
1273 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
1274 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
1276 *eth_port_id = ivshmem->eth_port_id;
1281 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
1283 struct dpdk_ring *ivshmem;
1284 unsigned int port_no;
1287 /* Names always start with "dpdkr" */
1288 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
1293 /* look through our list to find the device */
1294 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
1295 if (ivshmem->user_port_id == port_no) {
1296 VLOG_INFO("Found dpdk ring device %s:\n", dev_name);
1297 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
1301 /* Need to create the device rings */
1302 return dpdk_ring_create(dev_name, port_no, eth_port_id);
1306 netdev_dpdk_ring_construct(struct netdev *netdev)
1308 unsigned int port_no = 0;
1311 if (rte_eal_init_ret) {
1312 return rte_eal_init_ret;
1315 ovs_mutex_lock(&dpdk_mutex);
1317 err = dpdk_ring_open(netdev->name, &port_no);
1322 err = netdev_dpdk_init(netdev, port_no);
1325 ovs_mutex_unlock(&dpdk_mutex);
1329 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT) \
1333 NULL, /* netdev_dpdk_run */ \
1334 NULL, /* netdev_dpdk_wait */ \
1336 netdev_dpdk_alloc, \
1338 netdev_dpdk_destruct, \
1339 netdev_dpdk_dealloc, \
1340 netdev_dpdk_get_config, \
1341 NULL, /* netdev_dpdk_set_config */ \
1342 NULL, /* get_tunnel_config */ \
1343 netdev_dpdk_get_numa_id, /* get_numa_id */ \
1345 netdev_dpdk_send, /* send */ \
1346 NULL, /* send_wait */ \
1348 netdev_dpdk_set_etheraddr, \
1349 netdev_dpdk_get_etheraddr, \
1350 netdev_dpdk_get_mtu, \
1351 netdev_dpdk_set_mtu, \
1352 netdev_dpdk_get_ifindex, \
1353 netdev_dpdk_get_carrier, \
1354 netdev_dpdk_get_carrier_resets, \
1355 netdev_dpdk_set_miimon, \
1356 netdev_dpdk_get_stats, \
1357 netdev_dpdk_get_features, \
1358 NULL, /* set_advertisements */ \
1360 NULL, /* set_policing */ \
1361 NULL, /* get_qos_types */ \
1362 NULL, /* get_qos_capabilities */ \
1363 NULL, /* get_qos */ \
1364 NULL, /* set_qos */ \
1365 NULL, /* get_queue */ \
1366 NULL, /* set_queue */ \
1367 NULL, /* delete_queue */ \
1368 NULL, /* get_queue_stats */ \
1369 NULL, /* queue_dump_start */ \
1370 NULL, /* queue_dump_next */ \
1371 NULL, /* queue_dump_done */ \
1372 NULL, /* dump_queue_stats */ \
1374 NULL, /* get_in4 */ \
1375 NULL, /* set_in4 */ \
1376 NULL, /* get_in6 */ \
1377 NULL, /* add_router */ \
1378 NULL, /* get_next_hop */ \
1379 netdev_dpdk_get_status, \
1380 NULL, /* arp_lookup */ \
1382 netdev_dpdk_update_flags, \
1384 netdev_dpdk_rxq_alloc, \
1385 netdev_dpdk_rxq_construct, \
1386 netdev_dpdk_rxq_destruct, \
1387 netdev_dpdk_rxq_dealloc, \
1388 netdev_dpdk_rxq_recv, \
1389 NULL, /* rx_wait */ \
1390 NULL, /* rxq_drain */ \
1394 dpdk_init(int argc, char **argv)
1398 if (argc < 2 || strcmp(argv[1], "--dpdk"))
1401 /* Make sure program name passed to rte_eal_init() is vswitchd. */
1407 /* Make sure things are initialized ... */
1408 result = rte_eal_init(argc, argv);
1410 ovs_abort(result, "Cannot init EAL\n");
1413 rte_memzone_dump(stdout);
1414 rte_eal_init_ret = 0;
1416 if (argc > result) {
1417 argv[result] = argv[0];
1420 /* We are called from the main thread here */
1421 thread_set_nonpmd();
1426 const struct netdev_class dpdk_class =
1430 netdev_dpdk_construct);
1432 const struct netdev_class dpdk_ring_class =
1436 netdev_dpdk_ring_construct);
1439 netdev_dpdk_register(void)
1441 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
1443 if (rte_eal_init_ret) {
1447 if (ovsthread_once_start(&once)) {
1449 netdev_register_provider(&dpdk_class);
1450 netdev_register_provider(&dpdk_ring_class);
1451 ovsthread_once_done(&once);
1456 pmd_thread_setaffinity_cpu(int cpu)
1462 CPU_SET(cpu, &cpuset);
1463 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
1465 VLOG_ERR("Thread affinity error %d",err);
1468 /* lcore_id 0 is reseved for use by non pmd threads. */
1469 RTE_PER_LCORE(_lcore_id) = cpu + 1;
1475 thread_set_nonpmd(void)
1477 /* We cannot have RTE_MAX_LCORE pmd threads, because lcore_id 0 is reserved
1478 * for non pmd threads */
1479 BUILD_ASSERT(NR_PMD_THREADS < RTE_MAX_LCORE);
1480 /* We have to use 0 to allow non pmd threads to perform certain DPDK
1481 * operations, like rte_eth_dev_configure(). */
1482 RTE_PER_LCORE(_lcore_id) = 0;
1488 return rte_lcore_id() != 0;