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 NIC_PORT_RX_Q_SIZE 2048 /* Size of Physical NIC RX Queue, Max (n+32<=4096)*/
74 #define NIC_PORT_TX_Q_SIZE 2048 /* Size of Physical NIC TX Queue, Max (n+32<=4096)*/
76 /* XXX: Needs per NIC value for these constants. */
77 #define RX_PTHRESH 32 /* Default values of RX prefetch threshold reg. */
78 #define RX_HTHRESH 32 /* Default values of RX host threshold reg. */
79 #define RX_WTHRESH 16 /* Default values of RX write-back threshold reg. */
81 #define TX_PTHRESH 36 /* Default values of TX prefetch threshold reg. */
82 #define TX_HTHRESH 0 /* Default values of TX host threshold reg. */
83 #define TX_WTHRESH 0 /* Default values of TX write-back threshold reg. */
85 static const struct rte_eth_conf port_conf = {
87 .mq_mode = ETH_MQ_RX_RSS,
89 .header_split = 0, /* Header Split disabled */
90 .hw_ip_checksum = 0, /* IP checksum offload disabled */
91 .hw_vlan_filter = 0, /* VLAN filtering disabled */
92 .jumbo_frame = 0, /* Jumbo Frame Support disabled */
98 .rss_hf = ETH_RSS_IPV4_TCP | ETH_RSS_IPV4 | ETH_RSS_IPV6
99 | ETH_RSS_IPV4_UDP | ETH_RSS_IPV6_TCP | ETH_RSS_IPV6_UDP,
103 .mq_mode = ETH_MQ_TX_NONE,
107 static const struct rte_eth_rxconf rx_conf = {
109 .pthresh = RX_PTHRESH,
110 .hthresh = RX_HTHRESH,
111 .wthresh = RX_WTHRESH,
115 static const struct rte_eth_txconf tx_conf = {
117 .pthresh = TX_PTHRESH,
118 .hthresh = TX_HTHRESH,
119 .wthresh = TX_WTHRESH,
123 .txq_flags = ETH_TXQ_FLAGS_NOMULTSEGS|ETH_TXQ_FLAGS_NOOFFLOADS,
126 enum { MAX_RX_QUEUE_LEN = 192 };
127 enum { MAX_TX_QUEUE_LEN = 384 };
128 enum { DPDK_RING_SIZE = 256 };
129 BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE));
130 enum { DRAIN_TSC = 200000ULL };
132 static int rte_eal_init_ret = ENODEV;
134 static struct ovs_mutex dpdk_mutex = OVS_MUTEX_INITIALIZER;
136 /* Contains all 'struct dpdk_dev's. */
137 static struct list dpdk_list OVS_GUARDED_BY(dpdk_mutex)
138 = LIST_INITIALIZER(&dpdk_list);
140 static struct list dpdk_mp_list OVS_GUARDED_BY(dpdk_mutex)
141 = LIST_INITIALIZER(&dpdk_mp_list);
143 /* This mutex must be used by non pmd threads when allocating or freeing
144 * mbufs through mempools. Since dpdk_queue_pkts() and dpdk_queue_flush() may
145 * use mempools, a non pmd thread should hold this mutex while calling them */
146 struct ovs_mutex nonpmd_mempool_mutex = OVS_MUTEX_INITIALIZER;
149 struct rte_mempool *mp;
153 struct list list_node OVS_GUARDED_BY(dpdk_mutex);
156 /* There should be one 'struct dpdk_tx_queue' created for
158 struct dpdk_tx_queue {
159 bool flush_tx; /* Set to true to flush queue everytime */
160 /* pkts are queued. */
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);
205 rte_spinlock_t dpdkr_tx_lock;
208 struct netdev_rxq_dpdk {
209 struct netdev_rxq up;
213 static bool thread_is_pmd(void);
215 static int netdev_dpdk_construct(struct netdev *);
218 is_dpdk_class(const struct netdev_class *class)
220 return class->construct == netdev_dpdk_construct;
223 /* XXX: use dpdk malloc for entire OVS. infact huge page shld be used
224 * for all other sengments data, bss and text. */
227 dpdk_rte_mzalloc(size_t sz)
231 ptr = rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE);
238 /* XXX this function should be called only by pmd threads (or by non pmd
239 * threads holding the nonpmd_mempool_mutex) */
241 free_dpdk_buf(struct dpif_packet *p)
243 struct rte_mbuf *pkt = (struct rte_mbuf *) p;
245 rte_pktmbuf_free_seg(pkt);
249 __rte_pktmbuf_init(struct rte_mempool *mp,
250 void *opaque_arg OVS_UNUSED,
252 unsigned i OVS_UNUSED)
254 struct rte_mbuf *m = _m;
255 uint32_t buf_len = mp->elt_size - sizeof(struct dpif_packet);
257 RTE_MBUF_ASSERT(mp->elt_size >= sizeof(struct dpif_packet));
259 memset(m, 0, mp->elt_size);
261 /* start of buffer is just after mbuf structure */
262 m->buf_addr = (char *)m + sizeof(struct dpif_packet);
263 m->buf_physaddr = rte_mempool_virt2phy(mp, m) +
264 sizeof(struct dpif_packet);
265 m->buf_len = (uint16_t)buf_len;
267 /* keep some headroom between start of buffer and data */
268 m->pkt.data = (char*) m->buf_addr + RTE_MIN(RTE_PKTMBUF_HEADROOM, m->buf_len);
270 /* init some constant fields */
271 m->type = RTE_MBUF_PKT;
274 m->pkt.in_port = 0xff;
278 ovs_rte_pktmbuf_init(struct rte_mempool *mp,
279 void *opaque_arg OVS_UNUSED,
281 unsigned i OVS_UNUSED)
283 struct rte_mbuf *m = _m;
285 __rte_pktmbuf_init(mp, opaque_arg, _m, i);
287 ofpbuf_init_dpdk((struct ofpbuf *) m, m->buf_len);
290 static struct dpdk_mp *
291 dpdk_mp_get(int socket_id, int mtu) OVS_REQUIRES(dpdk_mutex)
293 struct dpdk_mp *dmp = NULL;
294 char mp_name[RTE_MEMPOOL_NAMESIZE];
296 LIST_FOR_EACH (dmp, list_node, &dpdk_mp_list) {
297 if (dmp->socket_id == socket_id && dmp->mtu == mtu) {
303 dmp = dpdk_rte_mzalloc(sizeof *dmp);
304 dmp->socket_id = socket_id;
308 if (snprintf(mp_name, RTE_MEMPOOL_NAMESIZE, "ovs_mp_%d_%d", dmp->mtu,
309 dmp->socket_id) < 0) {
313 dmp->mp = rte_mempool_create(mp_name, NB_MBUF, MBUF_SIZE(mtu),
315 sizeof(struct rte_pktmbuf_pool_private),
316 rte_pktmbuf_pool_init, NULL,
317 ovs_rte_pktmbuf_init, NULL,
320 if (dmp->mp == NULL) {
324 list_push_back(&dpdk_mp_list, &dmp->list_node);
329 dpdk_mp_put(struct dpdk_mp *dmp)
337 ovs_assert(dmp->refcount >= 0);
340 /* I could not find any API to destroy mp. */
341 if (dmp->refcount == 0) {
342 list_delete(dmp->list_node);
343 /* destroy mp-pool. */
349 check_link_status(struct netdev_dpdk *dev)
351 struct rte_eth_link link;
353 rte_eth_link_get_nowait(dev->port_id, &link);
355 if (dev->link.link_status != link.link_status) {
356 netdev_change_seq_changed(&dev->up);
358 dev->link_reset_cnt++;
360 if (dev->link.link_status) {
361 VLOG_DBG_RL(&rl, "Port %d Link Up - speed %u Mbps - %s",
362 dev->port_id, (unsigned)dev->link.link_speed,
363 (dev->link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
364 ("full-duplex") : ("half-duplex"));
366 VLOG_DBG_RL(&rl, "Port %d Link Down", dev->port_id);
372 dpdk_watchdog(void *dummy OVS_UNUSED)
374 struct netdev_dpdk *dev;
376 pthread_detach(pthread_self());
379 ovs_mutex_lock(&dpdk_mutex);
380 LIST_FOR_EACH (dev, list_node, &dpdk_list) {
381 ovs_mutex_lock(&dev->mutex);
382 check_link_status(dev);
383 ovs_mutex_unlock(&dev->mutex);
385 ovs_mutex_unlock(&dpdk_mutex);
386 xsleep(DPDK_PORT_WATCHDOG_INTERVAL);
393 dpdk_eth_dev_init(struct netdev_dpdk *dev) OVS_REQUIRES(dpdk_mutex)
395 struct rte_pktmbuf_pool_private *mbp_priv;
396 struct ether_addr eth_addr;
400 if (dev->port_id < 0 || dev->port_id >= rte_eth_dev_count()) {
404 diag = rte_eth_dev_configure(dev->port_id, dev->up.n_rxq, dev->up.n_txq,
407 VLOG_ERR("eth dev config error %d",diag);
411 for (i = 0; i < dev->up.n_txq; i++) {
412 diag = rte_eth_tx_queue_setup(dev->port_id, i, NIC_PORT_TX_Q_SIZE,
413 dev->socket_id, &tx_conf);
415 VLOG_ERR("eth dev tx queue setup error %d",diag);
420 for (i = 0; i < dev->up.n_rxq; i++) {
421 diag = rte_eth_rx_queue_setup(dev->port_id, i, NIC_PORT_RX_Q_SIZE,
423 &rx_conf, dev->dpdk_mp->mp);
425 VLOG_ERR("eth dev rx queue setup error %d",diag);
430 diag = rte_eth_dev_start(dev->port_id);
432 VLOG_ERR("eth dev start error %d",diag);
436 rte_eth_promiscuous_enable(dev->port_id);
437 rte_eth_allmulticast_enable(dev->port_id);
439 memset(ð_addr, 0x0, sizeof(eth_addr));
440 rte_eth_macaddr_get(dev->port_id, ð_addr);
441 VLOG_INFO_RL(&rl, "Port %d: "ETH_ADDR_FMT"",
442 dev->port_id, ETH_ADDR_ARGS(eth_addr.addr_bytes));
444 memcpy(dev->hwaddr, eth_addr.addr_bytes, ETH_ADDR_LEN);
445 rte_eth_link_get_nowait(dev->port_id, &dev->link);
447 mbp_priv = rte_mempool_get_priv(dev->dpdk_mp->mp);
448 dev->buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
450 dev->flags = NETDEV_UP | NETDEV_PROMISC;
454 static struct netdev_dpdk *
455 netdev_dpdk_cast(const struct netdev *netdev)
457 return CONTAINER_OF(netdev, struct netdev_dpdk, up);
460 static struct netdev *
461 netdev_dpdk_alloc(void)
463 struct netdev_dpdk *netdev = dpdk_rte_mzalloc(sizeof *netdev);
468 netdev_dpdk_alloc_txq(struct netdev_dpdk *netdev, unsigned int n_txqs)
472 netdev->tx_q = dpdk_rte_mzalloc(n_txqs * sizeof *netdev->tx_q);
473 /* Each index is considered as a cpu core id, since there should
474 * be one tx queue for each cpu core. */
475 for (i = 0; i < n_txqs; i++) {
476 int numa_id = ovs_numa_get_numa_id(i);
478 /* If the corresponding core is not on the same numa node
479 * as 'netdev', flags the 'flush_tx'. */
480 netdev->tx_q[i].flush_tx = netdev->socket_id == numa_id;
485 netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no)
486 OVS_REQUIRES(dpdk_mutex)
488 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
492 ovs_mutex_init(&netdev->mutex);
494 ovs_mutex_lock(&netdev->mutex);
496 /* If the 'sid' is negative, it means that the kernel fails
497 * to obtain the pci numa info. In that situation, always
499 sid = rte_eth_dev_socket_id(port_no);
500 netdev->socket_id = sid < 0 ? SOCKET0 : sid;
501 netdev_dpdk_alloc_txq(netdev, NR_QUEUE);
502 netdev->port_id = port_no;
504 netdev->mtu = ETHER_MTU;
505 netdev->max_packet_len = MTU_TO_MAX_LEN(netdev->mtu);
506 rte_spinlock_init(&netdev->dpdkr_tx_lock);
508 netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, netdev->mtu);
509 if (!netdev->dpdk_mp) {
514 netdev_->n_txq = NR_QUEUE;
515 netdev_->n_rxq = NR_QUEUE;
516 err = dpdk_eth_dev_init(netdev);
521 list_push_back(&dpdk_list, &netdev->list_node);
525 rte_free(netdev->tx_q);
527 ovs_mutex_unlock(&netdev->mutex);
532 dpdk_dev_parse_name(const char dev_name[], const char prefix[],
533 unsigned int *port_no)
537 if (strncmp(dev_name, prefix, strlen(prefix))) {
541 cport = dev_name + strlen(prefix);
542 *port_no = strtol(cport, 0, 0); /* string must be null terminated */
547 netdev_dpdk_construct(struct netdev *netdev)
549 unsigned int port_no;
552 if (rte_eal_init_ret) {
553 return rte_eal_init_ret;
556 /* Names always start with "dpdk" */
557 err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no);
562 ovs_mutex_lock(&dpdk_mutex);
563 err = netdev_dpdk_init(netdev, port_no);
564 ovs_mutex_unlock(&dpdk_mutex);
569 netdev_dpdk_destruct(struct netdev *netdev_)
571 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
573 ovs_mutex_lock(&dev->mutex);
574 rte_eth_dev_stop(dev->port_id);
575 ovs_mutex_unlock(&dev->mutex);
577 ovs_mutex_lock(&dpdk_mutex);
579 list_remove(&dev->list_node);
580 dpdk_mp_put(dev->dpdk_mp);
581 ovs_mutex_unlock(&dpdk_mutex);
583 ovs_mutex_destroy(&dev->mutex);
587 netdev_dpdk_dealloc(struct netdev *netdev_)
589 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
595 netdev_dpdk_get_config(const struct netdev *netdev_, struct smap *args)
597 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
599 ovs_mutex_lock(&dev->mutex);
601 smap_add_format(args, "configured_rx_queues", "%d", netdev_->n_rxq);
602 smap_add_format(args, "configured_tx_queues", "%d", netdev_->n_txq);
603 ovs_mutex_unlock(&dev->mutex);
609 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
611 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
613 return netdev->socket_id;
616 /* Sets the number of tx queues and rx queues for the dpdk interface.
617 * If the configuration fails, do not try restoring its old configuration
618 * and just returns the error. */
620 netdev_dpdk_set_multiq(struct netdev *netdev_, unsigned int n_txq,
623 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
626 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
630 ovs_mutex_lock(&dpdk_mutex);
631 ovs_mutex_lock(&netdev->mutex);
633 rte_eth_dev_stop(netdev->port_id);
635 netdev->up.n_txq = n_txq;
636 netdev->up.n_rxq = n_rxq;
637 rte_free(netdev->tx_q);
638 netdev_dpdk_alloc_txq(netdev, n_txq);
639 err = dpdk_eth_dev_init(netdev);
641 ovs_mutex_unlock(&netdev->mutex);
642 ovs_mutex_unlock(&dpdk_mutex);
647 static struct netdev_rxq *
648 netdev_dpdk_rxq_alloc(void)
650 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
655 static struct netdev_rxq_dpdk *
656 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
658 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
662 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
664 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
665 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
667 ovs_mutex_lock(&netdev->mutex);
668 rx->port_id = netdev->port_id;
669 ovs_mutex_unlock(&netdev->mutex);
675 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
680 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
682 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
688 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
690 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
693 while (nb_tx != txq->count) {
696 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
705 if (OVS_UNLIKELY(nb_tx != txq->count)) {
706 /* free buffers, which we couldn't transmit, one at a time (each
707 * packet could come from a different mempool) */
710 for (i = nb_tx; i < txq->count; i++) {
711 rte_pktmbuf_free_seg(txq->burst_pkts[i]);
713 ovs_mutex_lock(&dev->mutex);
714 dev->stats.tx_dropped += txq->count-nb_tx;
715 ovs_mutex_unlock(&dev->mutex);
719 txq->tsc = rte_get_timer_cycles();
723 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
725 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
727 if (txq->count == 0) {
730 dpdk_queue_flush__(dev, qid);
734 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dpif_packet **packets,
737 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
738 struct netdev *netdev = rx->up.netdev;
739 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
742 /* There is only one tx queue for this core. Do not flush other
744 if (rxq_->queue_id == rte_lcore_id()) {
745 dpdk_queue_flush(dev, rxq_->queue_id);
748 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
749 (struct rte_mbuf **) packets,
750 MIN((int)NETDEV_MAX_RX_BATCH,
751 (int)MAX_RX_QUEUE_LEN));
762 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
763 struct rte_mbuf **pkts, int cnt)
765 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
771 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
772 int tocopy = MIN(freeslots, cnt-i);
774 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
775 tocopy * sizeof (struct rte_mbuf *));
777 txq->count += tocopy;
780 if (txq->count == MAX_TX_QUEUE_LEN || txq->flush_tx) {
781 dpdk_queue_flush__(dev, qid);
783 diff_tsc = rte_get_timer_cycles() - txq->tsc;
784 if (diff_tsc >= DRAIN_TSC) {
785 dpdk_queue_flush__(dev, qid);
790 /* Tx function. Transmit packets indefinitely */
792 dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dpif_packet ** pkts,
794 OVS_NO_THREAD_SAFETY_ANALYSIS
796 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
797 struct rte_mbuf *mbufs[cnt];
802 /* If we are on a non pmd thread we have to use the mempool mutex, because
803 * every non pmd thread shares the same mempool cache */
805 if (!thread_is_pmd()) {
806 ovs_mutex_lock(&nonpmd_mempool_mutex);
809 for (i = 0; i < cnt; i++) {
810 int size = ofpbuf_size(&pkts[i]->ofpbuf);
812 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
813 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
814 (int)size , dev->max_packet_len);
820 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
822 if (!mbufs[newcnt]) {
827 /* We have to do a copy for now */
828 memcpy(mbufs[newcnt]->pkt.data, ofpbuf_data(&pkts[i]->ofpbuf), size);
830 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
831 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
836 if (OVS_UNLIKELY(dropped)) {
837 ovs_mutex_lock(&dev->mutex);
838 dev->stats.tx_dropped += dropped;
839 ovs_mutex_unlock(&dev->mutex);
842 dpdk_queue_pkts(dev, qid, mbufs, newcnt);
843 dpdk_queue_flush(dev, qid);
845 if (!thread_is_pmd()) {
846 ovs_mutex_unlock(&nonpmd_mempool_mutex);
851 netdev_dpdk_send__(struct netdev_dpdk *dev, int qid,
852 struct dpif_packet **pkts, int cnt, bool may_steal)
856 if (OVS_UNLIKELY(!may_steal ||
857 pkts[0]->ofpbuf.source != OFPBUF_DPDK)) {
858 struct netdev *netdev = &dev->up;
860 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
863 for (i = 0; i < cnt; i++) {
864 dpif_packet_delete(pkts[i]);
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);
903 netdev_dpdk_eth_send(struct netdev *netdev, int qid,
904 struct dpif_packet **pkts, int cnt, bool may_steal)
906 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
908 netdev_dpdk_send__(dev, qid, pkts, cnt, may_steal);
913 netdev_dpdk_set_etheraddr(struct netdev *netdev,
914 const uint8_t mac[ETH_ADDR_LEN])
916 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
918 ovs_mutex_lock(&dev->mutex);
919 if (!eth_addr_equals(dev->hwaddr, mac)) {
920 memcpy(dev->hwaddr, mac, ETH_ADDR_LEN);
921 netdev_change_seq_changed(netdev);
923 ovs_mutex_unlock(&dev->mutex);
929 netdev_dpdk_get_etheraddr(const struct netdev *netdev,
930 uint8_t mac[ETH_ADDR_LEN])
932 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
934 ovs_mutex_lock(&dev->mutex);
935 memcpy(mac, dev->hwaddr, ETH_ADDR_LEN);
936 ovs_mutex_unlock(&dev->mutex);
942 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
944 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
946 ovs_mutex_lock(&dev->mutex);
948 ovs_mutex_unlock(&dev->mutex);
954 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
956 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
958 struct dpdk_mp *old_mp;
961 ovs_mutex_lock(&dpdk_mutex);
962 ovs_mutex_lock(&dev->mutex);
963 if (dev->mtu == mtu) {
968 mp = dpdk_mp_get(dev->socket_id, dev->mtu);
974 rte_eth_dev_stop(dev->port_id);
977 old_mp = dev->dpdk_mp;
980 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
982 err = dpdk_eth_dev_init(dev);
986 dev->dpdk_mp = old_mp;
987 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
988 dpdk_eth_dev_init(dev);
993 netdev_change_seq_changed(netdev);
995 ovs_mutex_unlock(&dev->mutex);
996 ovs_mutex_unlock(&dpdk_mutex);
1001 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
1004 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
1006 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1007 struct rte_eth_stats rte_stats;
1010 netdev_dpdk_get_carrier(netdev, &gg);
1011 ovs_mutex_lock(&dev->mutex);
1012 rte_eth_stats_get(dev->port_id, &rte_stats);
1014 memset(stats, 0, sizeof(*stats));
1016 stats->rx_packets = rte_stats.ipackets;
1017 stats->tx_packets = rte_stats.opackets;
1018 stats->rx_bytes = rte_stats.ibytes;
1019 stats->tx_bytes = rte_stats.obytes;
1020 stats->rx_errors = rte_stats.ierrors;
1021 stats->tx_errors = rte_stats.oerrors;
1022 stats->multicast = rte_stats.imcasts;
1024 stats->tx_dropped = dev->stats.tx_dropped;
1025 ovs_mutex_unlock(&dev->mutex);
1031 netdev_dpdk_get_features(const struct netdev *netdev_,
1032 enum netdev_features *current,
1033 enum netdev_features *advertised OVS_UNUSED,
1034 enum netdev_features *supported OVS_UNUSED,
1035 enum netdev_features *peer OVS_UNUSED)
1037 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1038 struct rte_eth_link link;
1040 ovs_mutex_lock(&dev->mutex);
1042 ovs_mutex_unlock(&dev->mutex);
1044 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
1045 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
1046 *current = NETDEV_F_AUTONEG;
1048 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
1049 if (link.link_speed == ETH_LINK_SPEED_10) {
1050 *current = NETDEV_F_10MB_HD;
1052 if (link.link_speed == ETH_LINK_SPEED_100) {
1053 *current = NETDEV_F_100MB_HD;
1055 if (link.link_speed == ETH_LINK_SPEED_1000) {
1056 *current = NETDEV_F_1GB_HD;
1058 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
1059 if (link.link_speed == ETH_LINK_SPEED_10) {
1060 *current = NETDEV_F_10MB_FD;
1062 if (link.link_speed == ETH_LINK_SPEED_100) {
1063 *current = NETDEV_F_100MB_FD;
1065 if (link.link_speed == ETH_LINK_SPEED_1000) {
1066 *current = NETDEV_F_1GB_FD;
1068 if (link.link_speed == ETH_LINK_SPEED_10000) {
1069 *current = NETDEV_F_10GB_FD;
1077 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1079 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1082 ovs_mutex_lock(&dev->mutex);
1083 ifindex = dev->port_id;
1084 ovs_mutex_unlock(&dev->mutex);
1090 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1092 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1094 ovs_mutex_lock(&dev->mutex);
1095 check_link_status(dev);
1096 *carrier = dev->link.link_status;
1097 ovs_mutex_unlock(&dev->mutex);
1102 static long long int
1103 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1105 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1106 long long int carrier_resets;
1108 ovs_mutex_lock(&dev->mutex);
1109 carrier_resets = dev->link_reset_cnt;
1110 ovs_mutex_unlock(&dev->mutex);
1112 return carrier_resets;
1116 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1117 long long int interval OVS_UNUSED)
1123 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1124 enum netdev_flags off, enum netdev_flags on,
1125 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1129 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1133 *old_flagsp = dev->flags;
1137 if (dev->flags == *old_flagsp) {
1141 if (dev->flags & NETDEV_UP) {
1142 err = rte_eth_dev_start(dev->port_id);
1147 if (dev->flags & NETDEV_PROMISC) {
1148 rte_eth_promiscuous_enable(dev->port_id);
1151 if (!(dev->flags & NETDEV_UP)) {
1152 rte_eth_dev_stop(dev->port_id);
1159 netdev_dpdk_update_flags(struct netdev *netdev_,
1160 enum netdev_flags off, enum netdev_flags on,
1161 enum netdev_flags *old_flagsp)
1163 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1166 ovs_mutex_lock(&netdev->mutex);
1167 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1168 ovs_mutex_unlock(&netdev->mutex);
1174 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1176 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1177 struct rte_eth_dev_info dev_info;
1179 if (dev->port_id < 0)
1182 ovs_mutex_lock(&dev->mutex);
1183 rte_eth_dev_info_get(dev->port_id, &dev_info);
1184 ovs_mutex_unlock(&dev->mutex);
1186 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1188 smap_add_format(args, "port_no", "%d", dev->port_id);
1189 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1190 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1191 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1192 smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen);
1193 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1194 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1195 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1196 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1197 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1198 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1200 smap_add_format(args, "pci-vendor_id", "0x%u", dev_info.pci_dev->id.vendor_id);
1201 smap_add_format(args, "pci-device_id", "0x%x", dev_info.pci_dev->id.device_id);
1207 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1208 OVS_REQUIRES(dev->mutex)
1210 enum netdev_flags old_flags;
1213 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1215 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1220 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1221 const char *argv[], void *aux OVS_UNUSED)
1225 if (!strcasecmp(argv[argc - 1], "up")) {
1227 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1230 unixctl_command_reply_error(conn, "Invalid Admin State");
1235 struct netdev *netdev = netdev_from_name(argv[1]);
1236 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1237 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1239 ovs_mutex_lock(&dpdk_dev->mutex);
1240 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1241 ovs_mutex_unlock(&dpdk_dev->mutex);
1243 netdev_close(netdev);
1245 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1246 netdev_close(netdev);
1250 struct netdev_dpdk *netdev;
1252 ovs_mutex_lock(&dpdk_mutex);
1253 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1254 ovs_mutex_lock(&netdev->mutex);
1255 netdev_dpdk_set_admin_state__(netdev, up);
1256 ovs_mutex_unlock(&netdev->mutex);
1258 ovs_mutex_unlock(&dpdk_mutex);
1260 unixctl_command_reply(conn, "OK");
1264 dpdk_common_init(void)
1266 unixctl_command_register("netdev-dpdk/set-admin-state",
1267 "[netdev] up|down", 1, 2,
1268 netdev_dpdk_set_admin_state, NULL);
1270 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
1274 dpdk_class_init(void)
1278 result = rte_eal_pci_probe();
1280 VLOG_ERR("Cannot probe PCI");
1284 VLOG_INFO("Ethernet Device Count: %d", (int)rte_eth_dev_count());
1292 dpdk_ring_create(const char dev_name[], unsigned int port_no,
1293 unsigned int *eth_port_id)
1295 struct dpdk_ring *ivshmem;
1299 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
1300 if (ivshmem == NULL) {
1304 /* XXX: Add support for multiquque ring. */
1305 err = snprintf(ring_name, 10, "%s_tx", dev_name);
1310 /* Create single consumer/producer rings, netdev does explicit locking. */
1311 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1312 RING_F_SP_ENQ | RING_F_SC_DEQ);
1313 if (ivshmem->cring_tx == NULL) {
1318 err = snprintf(ring_name, 10, "%s_rx", dev_name);
1323 /* Create single consumer/producer rings, netdev does explicit locking. */
1324 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1325 RING_F_SP_ENQ | RING_F_SC_DEQ);
1326 if (ivshmem->cring_rx == NULL) {
1331 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
1332 &ivshmem->cring_tx, 1, SOCKET0);
1339 ivshmem->user_port_id = port_no;
1340 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
1341 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
1343 *eth_port_id = ivshmem->eth_port_id;
1348 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
1350 struct dpdk_ring *ivshmem;
1351 unsigned int port_no;
1354 /* Names always start with "dpdkr" */
1355 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
1360 /* look through our list to find the device */
1361 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
1362 if (ivshmem->user_port_id == port_no) {
1363 VLOG_INFO("Found dpdk ring device %s:\n", dev_name);
1364 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
1368 /* Need to create the device rings */
1369 return dpdk_ring_create(dev_name, port_no, eth_port_id);
1373 netdev_dpdk_ring_send(struct netdev *netdev, int qid OVS_UNUSED,
1374 struct dpif_packet **pkts, int cnt, bool may_steal)
1376 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1378 /* DPDK Rings have a single TX queue, Therefore needs locking. */
1379 rte_spinlock_lock(&dev->dpdkr_tx_lock);
1380 netdev_dpdk_send__(dev, 0, pkts, cnt, may_steal);
1381 rte_spinlock_unlock(&dev->dpdkr_tx_lock);
1386 netdev_dpdk_ring_construct(struct netdev *netdev)
1388 unsigned int port_no = 0;
1391 if (rte_eal_init_ret) {
1392 return rte_eal_init_ret;
1395 ovs_mutex_lock(&dpdk_mutex);
1397 err = dpdk_ring_open(netdev->name, &port_no);
1402 err = netdev_dpdk_init(netdev, port_no);
1405 ovs_mutex_unlock(&dpdk_mutex);
1409 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, MULTIQ, SEND) \
1413 NULL, /* netdev_dpdk_run */ \
1414 NULL, /* netdev_dpdk_wait */ \
1416 netdev_dpdk_alloc, \
1418 netdev_dpdk_destruct, \
1419 netdev_dpdk_dealloc, \
1420 netdev_dpdk_get_config, \
1421 NULL, /* netdev_dpdk_set_config */ \
1422 NULL, /* get_tunnel_config */ \
1423 netdev_dpdk_get_numa_id, /* get_numa_id */ \
1424 MULTIQ, /* set_multiq */ \
1427 NULL, /* send_wait */ \
1429 netdev_dpdk_set_etheraddr, \
1430 netdev_dpdk_get_etheraddr, \
1431 netdev_dpdk_get_mtu, \
1432 netdev_dpdk_set_mtu, \
1433 netdev_dpdk_get_ifindex, \
1434 netdev_dpdk_get_carrier, \
1435 netdev_dpdk_get_carrier_resets, \
1436 netdev_dpdk_set_miimon, \
1437 netdev_dpdk_get_stats, \
1438 netdev_dpdk_get_features, \
1439 NULL, /* set_advertisements */ \
1441 NULL, /* set_policing */ \
1442 NULL, /* get_qos_types */ \
1443 NULL, /* get_qos_capabilities */ \
1444 NULL, /* get_qos */ \
1445 NULL, /* set_qos */ \
1446 NULL, /* get_queue */ \
1447 NULL, /* set_queue */ \
1448 NULL, /* delete_queue */ \
1449 NULL, /* get_queue_stats */ \
1450 NULL, /* queue_dump_start */ \
1451 NULL, /* queue_dump_next */ \
1452 NULL, /* queue_dump_done */ \
1453 NULL, /* dump_queue_stats */ \
1455 NULL, /* get_in4 */ \
1456 NULL, /* set_in4 */ \
1457 NULL, /* get_in6 */ \
1458 NULL, /* add_router */ \
1459 NULL, /* get_next_hop */ \
1460 netdev_dpdk_get_status, \
1461 NULL, /* arp_lookup */ \
1463 netdev_dpdk_update_flags, \
1465 netdev_dpdk_rxq_alloc, \
1466 netdev_dpdk_rxq_construct, \
1467 netdev_dpdk_rxq_destruct, \
1468 netdev_dpdk_rxq_dealloc, \
1469 netdev_dpdk_rxq_recv, \
1470 NULL, /* rx_wait */ \
1471 NULL, /* rxq_drain */ \
1475 dpdk_init(int argc, char **argv)
1479 if (argc < 2 || strcmp(argv[1], "--dpdk"))
1482 /* Make sure program name passed to rte_eal_init() is vswitchd. */
1488 /* Make sure things are initialized ... */
1489 result = rte_eal_init(argc, argv);
1491 ovs_abort(result, "Cannot init EAL\n");
1494 rte_memzone_dump(stdout);
1495 rte_eal_init_ret = 0;
1497 if (argc > result) {
1498 argv[result] = argv[0];
1501 /* We are called from the main thread here */
1502 thread_set_nonpmd();
1507 const struct netdev_class dpdk_class =
1511 netdev_dpdk_construct,
1512 netdev_dpdk_set_multiq,
1513 netdev_dpdk_eth_send);
1515 const struct netdev_class dpdk_ring_class =
1519 netdev_dpdk_ring_construct,
1521 netdev_dpdk_ring_send);
1524 netdev_dpdk_register(void)
1526 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
1528 if (rte_eal_init_ret) {
1532 if (ovsthread_once_start(&once)) {
1534 netdev_register_provider(&dpdk_class);
1535 netdev_register_provider(&dpdk_ring_class);
1536 ovsthread_once_done(&once);
1541 pmd_thread_setaffinity_cpu(int cpu)
1547 CPU_SET(cpu, &cpuset);
1548 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
1550 VLOG_ERR("Thread affinity error %d",err);
1553 /* lcore_id 0 is reseved for use by non pmd threads. */
1555 RTE_PER_LCORE(_lcore_id) = cpu;
1561 thread_set_nonpmd(void)
1563 /* We have to use 0 to allow non pmd threads to perform certain DPDK
1564 * operations, like rte_eth_dev_configure(). */
1565 RTE_PER_LCORE(_lcore_id) = 0;
1571 return rte_lcore_id() != 0;