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 "dp-packet.h"
32 #include "dpif-netdev.h"
34 #include "netdev-dpdk.h"
35 #include "netdev-provider.h"
36 #include "netdev-vport.h"
38 #include "ofp-print.h"
40 #include "ovs-thread.h"
45 #include "unaligned.h"
48 #include "openvswitch/vlog.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 NIC_PORT_RX_Q_SIZE 2048 /* Size of Physical NIC RX Queue, Max (n+32<=4096)*/
73 #define NIC_PORT_TX_Q_SIZE 2048 /* Size of Physical NIC TX Queue, Max (n+32<=4096)*/
75 /* XXX: Needs per NIC value for these constants. */
76 #define RX_PTHRESH 32 /* Default values of RX prefetch threshold reg. */
77 #define RX_HTHRESH 32 /* Default values of RX host threshold reg. */
78 #define RX_WTHRESH 16 /* Default values of RX write-back threshold reg. */
80 #define TX_PTHRESH 36 /* Default values of TX prefetch threshold reg. */
81 #define TX_HTHRESH 0 /* Default values of TX host threshold reg. */
82 #define TX_WTHRESH 0 /* Default values of TX write-back threshold reg. */
84 static const struct rte_eth_conf port_conf = {
86 .mq_mode = ETH_MQ_RX_RSS,
88 .header_split = 0, /* Header Split disabled */
89 .hw_ip_checksum = 0, /* IP checksum offload disabled */
90 .hw_vlan_filter = 0, /* VLAN filtering disabled */
91 .jumbo_frame = 0, /* Jumbo Frame Support disabled */
97 .rss_hf = ETH_RSS_IPV4_TCP | ETH_RSS_IPV4 | ETH_RSS_IPV6
98 | ETH_RSS_IPV4_UDP | ETH_RSS_IPV6_TCP | ETH_RSS_IPV6_UDP,
102 .mq_mode = ETH_MQ_TX_NONE,
106 static const struct rte_eth_rxconf rx_conf = {
108 .pthresh = RX_PTHRESH,
109 .hthresh = RX_HTHRESH,
110 .wthresh = RX_WTHRESH,
114 static const struct rte_eth_txconf tx_conf = {
116 .pthresh = TX_PTHRESH,
117 .hthresh = TX_HTHRESH,
118 .wthresh = TX_WTHRESH,
122 .txq_flags = ETH_TXQ_FLAGS_NOMULTSEGS|ETH_TXQ_FLAGS_NOOFFLOADS,
125 enum { MAX_RX_QUEUE_LEN = 192 };
126 enum { MAX_TX_QUEUE_LEN = 384 };
127 enum { DPDK_RING_SIZE = 256 };
128 BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE));
129 enum { DRAIN_TSC = 200000ULL };
131 static int rte_eal_init_ret = ENODEV;
133 static struct ovs_mutex dpdk_mutex = OVS_MUTEX_INITIALIZER;
135 /* Contains all 'struct dpdk_dev's. */
136 static struct ovs_list dpdk_list OVS_GUARDED_BY(dpdk_mutex)
137 = OVS_LIST_INITIALIZER(&dpdk_list);
139 static struct ovs_list dpdk_mp_list OVS_GUARDED_BY(dpdk_mutex)
140 = OVS_LIST_INITIALIZER(&dpdk_mp_list);
142 /* This mutex must be used by non pmd threads when allocating or freeing
143 * mbufs through mempools. Since dpdk_queue_pkts() and dpdk_queue_flush() may
144 * use mempools, a non pmd thread should hold this mutex while calling them */
145 struct ovs_mutex nonpmd_mempool_mutex = OVS_MUTEX_INITIALIZER;
148 struct rte_mempool *mp;
152 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
155 /* There should be one 'struct dpdk_tx_queue' created for
157 struct dpdk_tx_queue {
158 bool flush_tx; /* Set to true to flush queue everytime */
159 /* pkts are queued. */
162 struct rte_mbuf *burst_pkts[MAX_TX_QUEUE_LEN];
165 /* dpdk has no way to remove dpdk ring ethernet devices
166 so we have to keep them around once they've been created
169 static struct ovs_list dpdk_ring_list OVS_GUARDED_BY(dpdk_mutex)
170 = OVS_LIST_INITIALIZER(&dpdk_ring_list);
173 /* For the client rings */
174 struct rte_ring *cring_tx;
175 struct rte_ring *cring_rx;
176 int user_port_id; /* User given port no, parsed from port name */
177 int eth_port_id; /* ethernet device port id */
178 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
186 struct dpdk_tx_queue *tx_q;
188 struct ovs_mutex mutex OVS_ACQ_AFTER(dpdk_mutex);
190 struct dpdk_mp *dpdk_mp;
194 struct netdev_stats stats;
196 uint8_t hwaddr[ETH_ADDR_LEN];
197 enum netdev_flags flags;
199 struct rte_eth_link link;
203 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
204 rte_spinlock_t dpdkr_tx_lock;
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 dp_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 dp_packet);
256 RTE_MBUF_ASSERT(mp->elt_size >= sizeof(struct dp_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 dp_packet);
262 m->buf_physaddr = rte_mempool_virt2phy(mp, m) +
263 sizeof(struct dp_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 dp_packet_init_dpdk((struct dp_packet *) 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_alloc_txq(struct netdev_dpdk *netdev, unsigned int n_txqs)
471 netdev->tx_q = dpdk_rte_mzalloc(n_txqs * sizeof *netdev->tx_q);
472 /* Each index is considered as a cpu core id, since there should
473 * be one tx queue for each cpu core. */
474 for (i = 0; i < n_txqs; i++) {
475 int numa_id = ovs_numa_get_numa_id(i);
477 /* If the corresponding core is not on the same numa node
478 * as 'netdev', flags the 'flush_tx'. */
479 netdev->tx_q[i].flush_tx = netdev->socket_id == numa_id;
484 netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no)
485 OVS_REQUIRES(dpdk_mutex)
487 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
491 ovs_mutex_init(&netdev->mutex);
493 ovs_mutex_lock(&netdev->mutex);
495 /* If the 'sid' is negative, it means that the kernel fails
496 * to obtain the pci numa info. In that situation, always
498 sid = rte_eth_dev_socket_id(port_no);
499 netdev->socket_id = sid < 0 ? SOCKET0 : sid;
500 netdev_dpdk_alloc_txq(netdev, NR_QUEUE);
501 netdev->port_id = port_no;
503 netdev->mtu = ETHER_MTU;
504 netdev->max_packet_len = MTU_TO_MAX_LEN(netdev->mtu);
505 rte_spinlock_init(&netdev->dpdkr_tx_lock);
507 netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, netdev->mtu);
508 if (!netdev->dpdk_mp) {
513 netdev_->n_txq = NR_QUEUE;
514 netdev_->n_rxq = NR_QUEUE;
515 err = dpdk_eth_dev_init(netdev);
520 list_push_back(&dpdk_list, &netdev->list_node);
524 rte_free(netdev->tx_q);
526 ovs_mutex_unlock(&netdev->mutex);
531 dpdk_dev_parse_name(const char dev_name[], const char prefix[],
532 unsigned int *port_no)
536 if (strncmp(dev_name, prefix, strlen(prefix))) {
540 cport = dev_name + strlen(prefix);
541 *port_no = strtol(cport, 0, 0); /* string must be null terminated */
546 netdev_dpdk_construct(struct netdev *netdev)
548 unsigned int port_no;
551 if (rte_eal_init_ret) {
552 return rte_eal_init_ret;
555 /* Names always start with "dpdk" */
556 err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no);
561 ovs_mutex_lock(&dpdk_mutex);
562 err = netdev_dpdk_init(netdev, port_no);
563 ovs_mutex_unlock(&dpdk_mutex);
568 netdev_dpdk_destruct(struct netdev *netdev_)
570 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
572 ovs_mutex_lock(&dev->mutex);
573 rte_eth_dev_stop(dev->port_id);
574 ovs_mutex_unlock(&dev->mutex);
576 ovs_mutex_lock(&dpdk_mutex);
578 list_remove(&dev->list_node);
579 dpdk_mp_put(dev->dpdk_mp);
580 ovs_mutex_unlock(&dpdk_mutex);
582 ovs_mutex_destroy(&dev->mutex);
586 netdev_dpdk_dealloc(struct netdev *netdev_)
588 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
594 netdev_dpdk_get_config(const struct netdev *netdev_, struct smap *args)
596 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
598 ovs_mutex_lock(&dev->mutex);
600 smap_add_format(args, "configured_rx_queues", "%d", netdev_->n_rxq);
601 smap_add_format(args, "configured_tx_queues", "%d", netdev_->n_txq);
602 ovs_mutex_unlock(&dev->mutex);
608 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
610 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
612 return netdev->socket_id;
615 /* Sets the number of tx queues and rx queues for the dpdk interface.
616 * If the configuration fails, do not try restoring its old configuration
617 * and just returns the error. */
619 netdev_dpdk_set_multiq(struct netdev *netdev_, unsigned int n_txq,
622 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
625 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
629 ovs_mutex_lock(&dpdk_mutex);
630 ovs_mutex_lock(&netdev->mutex);
632 rte_eth_dev_stop(netdev->port_id);
634 netdev->up.n_txq = n_txq;
635 netdev->up.n_rxq = n_rxq;
636 rte_free(netdev->tx_q);
637 netdev_dpdk_alloc_txq(netdev, n_txq);
638 err = dpdk_eth_dev_init(netdev);
640 ovs_mutex_unlock(&netdev->mutex);
641 ovs_mutex_unlock(&dpdk_mutex);
646 static struct netdev_rxq *
647 netdev_dpdk_rxq_alloc(void)
649 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
654 static struct netdev_rxq_dpdk *
655 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
657 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
661 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
663 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
664 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
666 ovs_mutex_lock(&netdev->mutex);
667 rx->port_id = netdev->port_id;
668 ovs_mutex_unlock(&netdev->mutex);
674 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
679 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
681 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
687 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
689 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
692 while (nb_tx != txq->count) {
695 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
704 if (OVS_UNLIKELY(nb_tx != txq->count)) {
705 /* free buffers, which we couldn't transmit, one at a time (each
706 * packet could come from a different mempool) */
709 for (i = nb_tx; i < txq->count; i++) {
710 rte_pktmbuf_free_seg(txq->burst_pkts[i]);
712 ovs_mutex_lock(&dev->mutex);
713 dev->stats.tx_dropped += txq->count-nb_tx;
714 ovs_mutex_unlock(&dev->mutex);
718 txq->tsc = rte_get_timer_cycles();
722 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
724 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
726 if (txq->count == 0) {
729 dpdk_queue_flush__(dev, qid);
733 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dp_packet **packets,
736 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
737 struct netdev *netdev = rx->up.netdev;
738 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
741 /* There is only one tx queue for this core. Do not flush other
743 if (rxq_->queue_id == rte_lcore_id()) {
744 dpdk_queue_flush(dev, rxq_->queue_id);
747 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
748 (struct rte_mbuf **) packets,
749 MIN((int)NETDEV_MAX_RX_BATCH,
750 (int)MAX_RX_QUEUE_LEN));
761 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
762 struct rte_mbuf **pkts, int cnt)
764 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
770 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
771 int tocopy = MIN(freeslots, cnt-i);
773 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
774 tocopy * sizeof (struct rte_mbuf *));
776 txq->count += tocopy;
779 if (txq->count == MAX_TX_QUEUE_LEN || txq->flush_tx) {
780 dpdk_queue_flush__(dev, qid);
782 diff_tsc = rte_get_timer_cycles() - txq->tsc;
783 if (diff_tsc >= DRAIN_TSC) {
784 dpdk_queue_flush__(dev, qid);
789 /* Tx function. Transmit packets indefinitely */
791 dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet ** pkts,
793 OVS_NO_THREAD_SAFETY_ANALYSIS
795 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
796 struct rte_mbuf *mbufs[cnt];
801 /* If we are on a non pmd thread we have to use the mempool mutex, because
802 * every non pmd thread shares the same mempool cache */
804 if (!thread_is_pmd()) {
805 ovs_mutex_lock(&nonpmd_mempool_mutex);
808 for (i = 0; i < cnt; i++) {
809 int size = dp_packet_size(pkts[i]);
811 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
812 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
813 (int)size , dev->max_packet_len);
819 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
821 if (!mbufs[newcnt]) {
826 /* We have to do a copy for now */
827 memcpy(mbufs[newcnt]->pkt.data, dp_packet_data(pkts[i]), size);
829 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
830 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
835 if (OVS_UNLIKELY(dropped)) {
836 ovs_mutex_lock(&dev->mutex);
837 dev->stats.tx_dropped += dropped;
838 ovs_mutex_unlock(&dev->mutex);
841 dpdk_queue_pkts(dev, qid, mbufs, newcnt);
842 dpdk_queue_flush(dev, qid);
844 if (!thread_is_pmd()) {
845 ovs_mutex_unlock(&nonpmd_mempool_mutex);
850 netdev_dpdk_send__(struct netdev_dpdk *dev, int qid,
851 struct dp_packet **pkts, int cnt, bool may_steal)
855 if (OVS_UNLIKELY(!may_steal ||
856 pkts[0]->source != DPBUF_DPDK)) {
857 struct netdev *netdev = &dev->up;
859 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
862 for (i = 0; i < cnt; i++) {
863 dp_packet_delete(pkts[i]);
870 for (i = 0; i < cnt; i++) {
871 int size = dp_packet_size(pkts[i]);
872 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
873 if (next_tx_idx != i) {
874 dpdk_queue_pkts(dev, qid,
875 (struct rte_mbuf **)&pkts[next_tx_idx],
879 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
880 (int)size , dev->max_packet_len);
882 dp_packet_delete(pkts[i]);
887 if (next_tx_idx != cnt) {
888 dpdk_queue_pkts(dev, qid,
889 (struct rte_mbuf **)&pkts[next_tx_idx],
893 if (OVS_UNLIKELY(dropped)) {
894 ovs_mutex_lock(&dev->mutex);
895 dev->stats.tx_dropped += dropped;
896 ovs_mutex_unlock(&dev->mutex);
902 netdev_dpdk_eth_send(struct netdev *netdev, int qid,
903 struct dp_packet **pkts, int cnt, bool may_steal)
905 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
907 netdev_dpdk_send__(dev, qid, pkts, cnt, may_steal);
912 netdev_dpdk_set_etheraddr(struct netdev *netdev,
913 const uint8_t mac[ETH_ADDR_LEN])
915 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
917 ovs_mutex_lock(&dev->mutex);
918 if (!eth_addr_equals(dev->hwaddr, mac)) {
919 memcpy(dev->hwaddr, mac, ETH_ADDR_LEN);
920 netdev_change_seq_changed(netdev);
922 ovs_mutex_unlock(&dev->mutex);
928 netdev_dpdk_get_etheraddr(const struct netdev *netdev,
929 uint8_t mac[ETH_ADDR_LEN])
931 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
933 ovs_mutex_lock(&dev->mutex);
934 memcpy(mac, dev->hwaddr, ETH_ADDR_LEN);
935 ovs_mutex_unlock(&dev->mutex);
941 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
943 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
945 ovs_mutex_lock(&dev->mutex);
947 ovs_mutex_unlock(&dev->mutex);
953 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
955 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
957 struct dpdk_mp *old_mp;
960 ovs_mutex_lock(&dpdk_mutex);
961 ovs_mutex_lock(&dev->mutex);
962 if (dev->mtu == mtu) {
967 mp = dpdk_mp_get(dev->socket_id, dev->mtu);
973 rte_eth_dev_stop(dev->port_id);
976 old_mp = dev->dpdk_mp;
979 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
981 err = dpdk_eth_dev_init(dev);
985 dev->dpdk_mp = old_mp;
986 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
987 dpdk_eth_dev_init(dev);
992 netdev_change_seq_changed(netdev);
994 ovs_mutex_unlock(&dev->mutex);
995 ovs_mutex_unlock(&dpdk_mutex);
1000 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
1003 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
1005 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1006 struct rte_eth_stats rte_stats;
1009 netdev_dpdk_get_carrier(netdev, &gg);
1010 ovs_mutex_lock(&dev->mutex);
1011 rte_eth_stats_get(dev->port_id, &rte_stats);
1013 memset(stats, 0, sizeof(*stats));
1015 stats->rx_packets = rte_stats.ipackets;
1016 stats->tx_packets = rte_stats.opackets;
1017 stats->rx_bytes = rte_stats.ibytes;
1018 stats->tx_bytes = rte_stats.obytes;
1019 stats->rx_errors = rte_stats.ierrors;
1020 stats->tx_errors = rte_stats.oerrors;
1021 stats->multicast = rte_stats.imcasts;
1023 stats->tx_dropped = dev->stats.tx_dropped;
1024 ovs_mutex_unlock(&dev->mutex);
1030 netdev_dpdk_get_features(const struct netdev *netdev_,
1031 enum netdev_features *current,
1032 enum netdev_features *advertised OVS_UNUSED,
1033 enum netdev_features *supported OVS_UNUSED,
1034 enum netdev_features *peer OVS_UNUSED)
1036 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1037 struct rte_eth_link link;
1039 ovs_mutex_lock(&dev->mutex);
1041 ovs_mutex_unlock(&dev->mutex);
1043 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
1044 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
1045 *current = NETDEV_F_AUTONEG;
1047 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
1048 if (link.link_speed == ETH_LINK_SPEED_10) {
1049 *current = NETDEV_F_10MB_HD;
1051 if (link.link_speed == ETH_LINK_SPEED_100) {
1052 *current = NETDEV_F_100MB_HD;
1054 if (link.link_speed == ETH_LINK_SPEED_1000) {
1055 *current = NETDEV_F_1GB_HD;
1057 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
1058 if (link.link_speed == ETH_LINK_SPEED_10) {
1059 *current = NETDEV_F_10MB_FD;
1061 if (link.link_speed == ETH_LINK_SPEED_100) {
1062 *current = NETDEV_F_100MB_FD;
1064 if (link.link_speed == ETH_LINK_SPEED_1000) {
1065 *current = NETDEV_F_1GB_FD;
1067 if (link.link_speed == ETH_LINK_SPEED_10000) {
1068 *current = NETDEV_F_10GB_FD;
1076 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1078 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1081 ovs_mutex_lock(&dev->mutex);
1082 ifindex = dev->port_id;
1083 ovs_mutex_unlock(&dev->mutex);
1089 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1091 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1093 ovs_mutex_lock(&dev->mutex);
1094 check_link_status(dev);
1095 *carrier = dev->link.link_status;
1096 ovs_mutex_unlock(&dev->mutex);
1101 static long long int
1102 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1104 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1105 long long int carrier_resets;
1107 ovs_mutex_lock(&dev->mutex);
1108 carrier_resets = dev->link_reset_cnt;
1109 ovs_mutex_unlock(&dev->mutex);
1111 return carrier_resets;
1115 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1116 long long int interval OVS_UNUSED)
1122 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1123 enum netdev_flags off, enum netdev_flags on,
1124 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1128 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1132 *old_flagsp = dev->flags;
1136 if (dev->flags == *old_flagsp) {
1140 if (dev->flags & NETDEV_UP) {
1141 err = rte_eth_dev_start(dev->port_id);
1146 if (dev->flags & NETDEV_PROMISC) {
1147 rte_eth_promiscuous_enable(dev->port_id);
1150 if (!(dev->flags & NETDEV_UP)) {
1151 rte_eth_dev_stop(dev->port_id);
1158 netdev_dpdk_update_flags(struct netdev *netdev_,
1159 enum netdev_flags off, enum netdev_flags on,
1160 enum netdev_flags *old_flagsp)
1162 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1165 ovs_mutex_lock(&netdev->mutex);
1166 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1167 ovs_mutex_unlock(&netdev->mutex);
1173 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1175 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1176 struct rte_eth_dev_info dev_info;
1178 if (dev->port_id < 0)
1181 ovs_mutex_lock(&dev->mutex);
1182 rte_eth_dev_info_get(dev->port_id, &dev_info);
1183 ovs_mutex_unlock(&dev->mutex);
1185 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1187 smap_add_format(args, "port_no", "%d", dev->port_id);
1188 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1189 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1190 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1191 smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen);
1192 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1193 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1194 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1195 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1196 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1197 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1199 smap_add_format(args, "pci-vendor_id", "0x%u", dev_info.pci_dev->id.vendor_id);
1200 smap_add_format(args, "pci-device_id", "0x%x", dev_info.pci_dev->id.device_id);
1206 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1207 OVS_REQUIRES(dev->mutex)
1209 enum netdev_flags old_flags;
1212 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1214 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1219 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1220 const char *argv[], void *aux OVS_UNUSED)
1224 if (!strcasecmp(argv[argc - 1], "up")) {
1226 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1229 unixctl_command_reply_error(conn, "Invalid Admin State");
1234 struct netdev *netdev = netdev_from_name(argv[1]);
1235 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1236 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1238 ovs_mutex_lock(&dpdk_dev->mutex);
1239 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1240 ovs_mutex_unlock(&dpdk_dev->mutex);
1242 netdev_close(netdev);
1244 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1245 netdev_close(netdev);
1249 struct netdev_dpdk *netdev;
1251 ovs_mutex_lock(&dpdk_mutex);
1252 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1253 ovs_mutex_lock(&netdev->mutex);
1254 netdev_dpdk_set_admin_state__(netdev, up);
1255 ovs_mutex_unlock(&netdev->mutex);
1257 ovs_mutex_unlock(&dpdk_mutex);
1259 unixctl_command_reply(conn, "OK");
1263 dpdk_common_init(void)
1265 unixctl_command_register("netdev-dpdk/set-admin-state",
1266 "[netdev] up|down", 1, 2,
1267 netdev_dpdk_set_admin_state, NULL);
1269 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
1273 dpdk_class_init(void)
1277 result = rte_eal_pci_probe();
1279 VLOG_ERR("Cannot probe PCI");
1283 VLOG_INFO("Ethernet Device Count: %d", (int)rte_eth_dev_count());
1291 dpdk_ring_create(const char dev_name[], unsigned int port_no,
1292 unsigned int *eth_port_id)
1294 struct dpdk_ring *ivshmem;
1298 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
1299 if (ivshmem == NULL) {
1303 /* XXX: Add support for multiquque ring. */
1304 err = snprintf(ring_name, 10, "%s_tx", dev_name);
1309 /* Create single consumer/producer rings, netdev does explicit locking. */
1310 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1311 RING_F_SP_ENQ | RING_F_SC_DEQ);
1312 if (ivshmem->cring_tx == NULL) {
1317 err = snprintf(ring_name, 10, "%s_rx", dev_name);
1322 /* Create single consumer/producer rings, netdev does explicit locking. */
1323 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1324 RING_F_SP_ENQ | RING_F_SC_DEQ);
1325 if (ivshmem->cring_rx == NULL) {
1330 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
1331 &ivshmem->cring_tx, 1, SOCKET0);
1338 ivshmem->user_port_id = port_no;
1339 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
1340 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
1342 *eth_port_id = ivshmem->eth_port_id;
1347 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
1349 struct dpdk_ring *ivshmem;
1350 unsigned int port_no;
1353 /* Names always start with "dpdkr" */
1354 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
1359 /* look through our list to find the device */
1360 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
1361 if (ivshmem->user_port_id == port_no) {
1362 VLOG_INFO("Found dpdk ring device %s:\n", dev_name);
1363 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
1367 /* Need to create the device rings */
1368 return dpdk_ring_create(dev_name, port_no, eth_port_id);
1372 netdev_dpdk_ring_send(struct netdev *netdev, int qid OVS_UNUSED,
1373 struct dp_packet **pkts, int cnt, bool may_steal)
1375 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1377 /* DPDK Rings have a single TX queue, Therefore needs locking. */
1378 rte_spinlock_lock(&dev->dpdkr_tx_lock);
1379 netdev_dpdk_send__(dev, 0, pkts, cnt, may_steal);
1380 rte_spinlock_unlock(&dev->dpdkr_tx_lock);
1385 netdev_dpdk_ring_construct(struct netdev *netdev)
1387 unsigned int port_no = 0;
1390 if (rte_eal_init_ret) {
1391 return rte_eal_init_ret;
1394 ovs_mutex_lock(&dpdk_mutex);
1396 err = dpdk_ring_open(netdev->name, &port_no);
1401 err = netdev_dpdk_init(netdev, port_no);
1404 ovs_mutex_unlock(&dpdk_mutex);
1408 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, MULTIQ, SEND) \
1412 NULL, /* netdev_dpdk_run */ \
1413 NULL, /* netdev_dpdk_wait */ \
1415 netdev_dpdk_alloc, \
1417 netdev_dpdk_destruct, \
1418 netdev_dpdk_dealloc, \
1419 netdev_dpdk_get_config, \
1420 NULL, /* netdev_dpdk_set_config */ \
1421 NULL, /* get_tunnel_config */ \
1422 NULL, /* build header */ \
1423 NULL, /* push header */ \
1424 NULL, /* pop header */ \
1425 netdev_dpdk_get_numa_id, /* get_numa_id */ \
1426 MULTIQ, /* set_multiq */ \
1429 NULL, /* send_wait */ \
1431 netdev_dpdk_set_etheraddr, \
1432 netdev_dpdk_get_etheraddr, \
1433 netdev_dpdk_get_mtu, \
1434 netdev_dpdk_set_mtu, \
1435 netdev_dpdk_get_ifindex, \
1436 netdev_dpdk_get_carrier, \
1437 netdev_dpdk_get_carrier_resets, \
1438 netdev_dpdk_set_miimon, \
1439 netdev_dpdk_get_stats, \
1440 netdev_dpdk_get_features, \
1441 NULL, /* set_advertisements */ \
1443 NULL, /* set_policing */ \
1444 NULL, /* get_qos_types */ \
1445 NULL, /* get_qos_capabilities */ \
1446 NULL, /* get_qos */ \
1447 NULL, /* set_qos */ \
1448 NULL, /* get_queue */ \
1449 NULL, /* set_queue */ \
1450 NULL, /* delete_queue */ \
1451 NULL, /* get_queue_stats */ \
1452 NULL, /* queue_dump_start */ \
1453 NULL, /* queue_dump_next */ \
1454 NULL, /* queue_dump_done */ \
1455 NULL, /* dump_queue_stats */ \
1457 NULL, /* get_in4 */ \
1458 NULL, /* set_in4 */ \
1459 NULL, /* get_in6 */ \
1460 NULL, /* add_router */ \
1461 NULL, /* get_next_hop */ \
1462 netdev_dpdk_get_status, \
1463 NULL, /* arp_lookup */ \
1465 netdev_dpdk_update_flags, \
1467 netdev_dpdk_rxq_alloc, \
1468 netdev_dpdk_rxq_construct, \
1469 netdev_dpdk_rxq_destruct, \
1470 netdev_dpdk_rxq_dealloc, \
1471 netdev_dpdk_rxq_recv, \
1472 NULL, /* rx_wait */ \
1473 NULL, /* rxq_drain */ \
1477 dpdk_init(int argc, char **argv)
1481 if (argc < 2 || strcmp(argv[1], "--dpdk"))
1484 /* Make sure program name passed to rte_eal_init() is vswitchd. */
1490 /* Make sure things are initialized ... */
1491 result = rte_eal_init(argc, argv);
1493 ovs_abort(result, "Cannot init EAL\n");
1496 rte_memzone_dump(stdout);
1497 rte_eal_init_ret = 0;
1499 if (argc > result) {
1500 argv[result] = argv[0];
1503 /* We are called from the main thread here */
1504 thread_set_nonpmd();
1509 const struct netdev_class dpdk_class =
1513 netdev_dpdk_construct,
1514 netdev_dpdk_set_multiq,
1515 netdev_dpdk_eth_send);
1517 const struct netdev_class dpdk_ring_class =
1521 netdev_dpdk_ring_construct,
1523 netdev_dpdk_ring_send);
1526 netdev_dpdk_register(void)
1528 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
1530 if (rte_eal_init_ret) {
1534 if (ovsthread_once_start(&once)) {
1536 netdev_register_provider(&dpdk_class);
1537 netdev_register_provider(&dpdk_ring_class);
1538 ovsthread_once_done(&once);
1543 pmd_thread_setaffinity_cpu(int cpu)
1549 CPU_SET(cpu, &cpuset);
1550 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
1552 VLOG_ERR("Thread affinity error %d",err);
1555 /* NON_PMD_CORE_ID is reserved for use by non pmd threads. */
1556 ovs_assert(cpu != NON_PMD_CORE_ID);
1557 RTE_PER_LCORE(_lcore_id) = cpu;
1563 thread_set_nonpmd(void)
1565 /* We have to use NON_PMD_CORE_ID to allow non-pmd threads to perform
1566 * certain DPDK operations, like rte_eth_dev_configure(). */
1567 RTE_PER_LCORE(_lcore_id) = NON_PMD_CORE_ID;
1573 return rte_lcore_id() != NON_PMD_CORE_ID;