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 #include "rte_config.h"
52 #include "rte_virtio_net.h"
54 VLOG_DEFINE_THIS_MODULE(dpdk);
55 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
57 #define DPDK_PORT_WATCHDOG_INTERVAL 5
59 #define OVS_CACHE_LINE_SIZE CACHE_LINE_SIZE
60 #define OVS_VPORT_DPDK "ovs_dpdk"
63 * need to reserve tons of extra space in the mbufs so we can align the
64 * DMA addresses to 4KB.
67 #define MTU_TO_MAX_LEN(mtu) ((mtu) + ETHER_HDR_LEN + ETHER_CRC_LEN)
68 #define MBUF_SIZE(mtu) (MTU_TO_MAX_LEN(mtu) + (512) + \
69 sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM)
71 /* Max and min number of packets in the mempool. OVS tries to allocate a
72 * mempool with MAX_NB_MBUF: if this fails (because the system doesn't have
73 * enough hugepages) we keep halving the number until the allocation succeeds
74 * or we reach MIN_NB_MBUF */
76 #define MAX_NB_MBUF (4096 * 64)
77 #define MIN_NB_MBUF (4096 * 4)
78 #define MP_CACHE_SZ RTE_MEMPOOL_CACHE_MAX_SIZE
80 /* MAX_NB_MBUF can be divided by 2 many times, until MIN_NB_MBUF */
81 BUILD_ASSERT_DECL(MAX_NB_MBUF % ROUND_DOWN_POW2(MAX_NB_MBUF/MIN_NB_MBUF) == 0);
83 /* The smallest possible NB_MBUF that we're going to try should be a multiple
84 * of MP_CACHE_SZ. This is advised by DPDK documentation. */
85 BUILD_ASSERT_DECL((MAX_NB_MBUF / ROUND_DOWN_POW2(MAX_NB_MBUF/MIN_NB_MBUF))
90 #define NIC_PORT_RX_Q_SIZE 2048 /* Size of Physical NIC RX Queue, Max (n+32<=4096)*/
91 #define NIC_PORT_TX_Q_SIZE 2048 /* Size of Physical NIC TX Queue, Max (n+32<=4096)*/
93 /* Character device cuse_dev_name. */
94 static char *cuse_dev_name = NULL;
97 * Maximum amount of time in micro seconds to try and enqueue to vhost.
99 #define VHOST_ENQ_RETRY_USECS 100
101 static const struct rte_eth_conf port_conf = {
103 .mq_mode = ETH_MQ_RX_RSS,
105 .header_split = 0, /* Header Split disabled */
106 .hw_ip_checksum = 0, /* IP checksum offload disabled */
107 .hw_vlan_filter = 0, /* VLAN filtering disabled */
108 .jumbo_frame = 0, /* Jumbo Frame Support disabled */
114 .rss_hf = ETH_RSS_IP | ETH_RSS_UDP | ETH_RSS_TCP,
118 .mq_mode = ETH_MQ_TX_NONE,
122 enum { MAX_TX_QUEUE_LEN = 384 };
123 enum { DPDK_RING_SIZE = 256 };
124 BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE));
125 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 ovs_list dpdk_list OVS_GUARDED_BY(dpdk_mutex)
138 = OVS_LIST_INITIALIZER(&dpdk_list);
140 static struct ovs_list dpdk_mp_list OVS_GUARDED_BY(dpdk_mutex)
141 = OVS_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 static struct ovs_mutex nonpmd_mempool_mutex = OVS_MUTEX_INITIALIZER;
149 struct rte_mempool *mp;
153 struct ovs_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 ovs_list dpdk_ring_list OVS_GUARDED_BY(dpdk_mutex)
171 = OVS_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 ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
186 enum dpdk_dev_type type;
188 struct dpdk_tx_queue *tx_q;
190 struct ovs_mutex mutex OVS_ACQ_AFTER(dpdk_mutex);
192 struct dpdk_mp *dpdk_mp;
196 struct netdev_stats stats;
198 rte_spinlock_t stats_lock;
200 uint8_t hwaddr[ETH_ADDR_LEN];
201 enum netdev_flags flags;
203 struct rte_eth_link link;
206 /* virtio-net structure for vhost device */
207 OVSRCU_TYPE(struct virtio_net *) virtio_dev;
210 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
211 rte_spinlock_t txq_lock;
214 struct netdev_rxq_dpdk {
215 struct netdev_rxq up;
219 static bool thread_is_pmd(void);
221 static int netdev_dpdk_construct(struct netdev *);
223 struct virtio_net * netdev_dpdk_get_virtio(const struct netdev_dpdk *dev);
226 is_dpdk_class(const struct netdev_class *class)
228 return class->construct == netdev_dpdk_construct;
231 /* XXX: use dpdk malloc for entire OVS. in fact huge page should be used
232 * for all other segments data, bss and text. */
235 dpdk_rte_mzalloc(size_t sz)
239 ptr = rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE);
246 /* XXX this function should be called only by pmd threads (or by non pmd
247 * threads holding the nonpmd_mempool_mutex) */
249 free_dpdk_buf(struct dp_packet *p)
251 struct rte_mbuf *pkt = (struct rte_mbuf *) p;
253 rte_pktmbuf_free_seg(pkt);
257 __rte_pktmbuf_init(struct rte_mempool *mp,
258 void *opaque_arg OVS_UNUSED,
260 unsigned i OVS_UNUSED)
262 struct rte_mbuf *m = _m;
263 uint32_t buf_len = mp->elt_size - sizeof(struct dp_packet);
265 RTE_MBUF_ASSERT(mp->elt_size >= sizeof(struct dp_packet));
267 memset(m, 0, mp->elt_size);
269 /* start of buffer is just after mbuf structure */
270 m->buf_addr = (char *)m + sizeof(struct dp_packet);
271 m->buf_physaddr = rte_mempool_virt2phy(mp, m) +
272 sizeof(struct dp_packet);
273 m->buf_len = (uint16_t)buf_len;
275 /* keep some headroom between start of buffer and data */
276 m->data_off = RTE_MIN(RTE_PKTMBUF_HEADROOM, m->buf_len);
278 /* init some constant fields */
285 ovs_rte_pktmbuf_init(struct rte_mempool *mp,
286 void *opaque_arg OVS_UNUSED,
288 unsigned i OVS_UNUSED)
290 struct rte_mbuf *m = _m;
292 __rte_pktmbuf_init(mp, opaque_arg, _m, i);
294 dp_packet_init_dpdk((struct dp_packet *) m, m->buf_len);
297 static struct dpdk_mp *
298 dpdk_mp_get(int socket_id, int mtu) OVS_REQUIRES(dpdk_mutex)
300 struct dpdk_mp *dmp = NULL;
301 char mp_name[RTE_MEMPOOL_NAMESIZE];
304 LIST_FOR_EACH (dmp, list_node, &dpdk_mp_list) {
305 if (dmp->socket_id == socket_id && dmp->mtu == mtu) {
311 dmp = dpdk_rte_mzalloc(sizeof *dmp);
312 dmp->socket_id = socket_id;
316 mp_size = MAX_NB_MBUF;
318 if (snprintf(mp_name, RTE_MEMPOOL_NAMESIZE, "ovs_mp_%d_%d_%u",
319 dmp->mtu, dmp->socket_id, mp_size) < 0) {
323 dmp->mp = rte_mempool_create(mp_name, mp_size, MBUF_SIZE(mtu),
325 sizeof(struct rte_pktmbuf_pool_private),
326 rte_pktmbuf_pool_init, NULL,
327 ovs_rte_pktmbuf_init, NULL,
329 } while (!dmp->mp && rte_errno == ENOMEM && (mp_size /= 2) >= MIN_NB_MBUF);
331 if (dmp->mp == NULL) {
334 VLOG_DBG("Allocated \"%s\" mempool with %u mbufs", mp_name, mp_size );
337 list_push_back(&dpdk_mp_list, &dmp->list_node);
342 dpdk_mp_put(struct dpdk_mp *dmp)
350 ovs_assert(dmp->refcount >= 0);
353 /* I could not find any API to destroy mp. */
354 if (dmp->refcount == 0) {
355 list_delete(dmp->list_node);
356 /* destroy mp-pool. */
362 check_link_status(struct netdev_dpdk *dev)
364 struct rte_eth_link link;
366 rte_eth_link_get_nowait(dev->port_id, &link);
368 if (dev->link.link_status != link.link_status) {
369 netdev_change_seq_changed(&dev->up);
371 dev->link_reset_cnt++;
373 if (dev->link.link_status) {
374 VLOG_DBG_RL(&rl, "Port %d Link Up - speed %u Mbps - %s",
375 dev->port_id, (unsigned)dev->link.link_speed,
376 (dev->link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
377 ("full-duplex") : ("half-duplex"));
379 VLOG_DBG_RL(&rl, "Port %d Link Down", dev->port_id);
385 dpdk_watchdog(void *dummy OVS_UNUSED)
387 struct netdev_dpdk *dev;
389 pthread_detach(pthread_self());
392 ovs_mutex_lock(&dpdk_mutex);
393 LIST_FOR_EACH (dev, list_node, &dpdk_list) {
394 ovs_mutex_lock(&dev->mutex);
395 check_link_status(dev);
396 ovs_mutex_unlock(&dev->mutex);
398 ovs_mutex_unlock(&dpdk_mutex);
399 xsleep(DPDK_PORT_WATCHDOG_INTERVAL);
406 dpdk_eth_dev_init(struct netdev_dpdk *dev) OVS_REQUIRES(dpdk_mutex)
408 struct rte_pktmbuf_pool_private *mbp_priv;
409 struct ether_addr eth_addr;
413 if (dev->port_id < 0 || dev->port_id >= rte_eth_dev_count()) {
417 diag = rte_eth_dev_configure(dev->port_id, dev->up.n_rxq, dev->up.n_txq,
420 VLOG_ERR("eth dev config error %d",diag);
424 for (i = 0; i < dev->up.n_txq; i++) {
425 diag = rte_eth_tx_queue_setup(dev->port_id, i, NIC_PORT_TX_Q_SIZE,
426 dev->socket_id, NULL);
428 VLOG_ERR("eth dev tx queue setup error %d",diag);
433 for (i = 0; i < dev->up.n_rxq; i++) {
434 diag = rte_eth_rx_queue_setup(dev->port_id, i, NIC_PORT_RX_Q_SIZE,
436 NULL, dev->dpdk_mp->mp);
438 VLOG_ERR("eth dev rx queue setup error %d",diag);
443 diag = rte_eth_dev_start(dev->port_id);
445 VLOG_ERR("eth dev start error %d",diag);
449 rte_eth_promiscuous_enable(dev->port_id);
450 rte_eth_allmulticast_enable(dev->port_id);
452 memset(ð_addr, 0x0, sizeof(eth_addr));
453 rte_eth_macaddr_get(dev->port_id, ð_addr);
454 VLOG_INFO_RL(&rl, "Port %d: "ETH_ADDR_FMT"",
455 dev->port_id, ETH_ADDR_ARGS(eth_addr.addr_bytes));
457 memcpy(dev->hwaddr, eth_addr.addr_bytes, ETH_ADDR_LEN);
458 rte_eth_link_get_nowait(dev->port_id, &dev->link);
460 mbp_priv = rte_mempool_get_priv(dev->dpdk_mp->mp);
461 dev->buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
463 dev->flags = NETDEV_UP | NETDEV_PROMISC;
467 static struct netdev_dpdk *
468 netdev_dpdk_cast(const struct netdev *netdev)
470 return CONTAINER_OF(netdev, struct netdev_dpdk, up);
473 static struct netdev *
474 netdev_dpdk_alloc(void)
476 struct netdev_dpdk *netdev = dpdk_rte_mzalloc(sizeof *netdev);
481 netdev_dpdk_alloc_txq(struct netdev_dpdk *netdev, unsigned int n_txqs)
485 netdev->tx_q = dpdk_rte_mzalloc(n_txqs * sizeof *netdev->tx_q);
486 /* Each index is considered as a cpu core id, since there should
487 * be one tx queue for each cpu core. */
488 for (i = 0; i < n_txqs; i++) {
489 int numa_id = ovs_numa_get_numa_id(i);
491 /* If the corresponding core is not on the same numa node
492 * as 'netdev', flags the 'flush_tx'. */
493 netdev->tx_q[i].flush_tx = netdev->socket_id == numa_id;
498 netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no,
499 enum dpdk_dev_type type)
500 OVS_REQUIRES(dpdk_mutex)
502 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
506 ovs_mutex_init(&netdev->mutex);
507 ovs_mutex_lock(&netdev->mutex);
509 rte_spinlock_init(&netdev->stats_lock);
511 /* If the 'sid' is negative, it means that the kernel fails
512 * to obtain the pci numa info. In that situation, always
514 if (type == DPDK_DEV_ETH) {
515 sid = rte_eth_dev_socket_id(port_no);
517 sid = rte_lcore_to_socket_id(rte_get_master_lcore());
520 netdev->socket_id = sid < 0 ? SOCKET0 : sid;
521 netdev->port_id = port_no;
524 netdev->mtu = ETHER_MTU;
525 netdev->max_packet_len = MTU_TO_MAX_LEN(netdev->mtu);
526 rte_spinlock_init(&netdev->txq_lock);
528 netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, netdev->mtu);
529 if (!netdev->dpdk_mp) {
534 netdev_->n_txq = NR_QUEUE;
535 netdev_->n_rxq = NR_QUEUE;
537 if (type == DPDK_DEV_ETH) {
538 netdev_dpdk_alloc_txq(netdev, NR_QUEUE);
539 err = dpdk_eth_dev_init(netdev);
545 list_push_back(&dpdk_list, &netdev->list_node);
549 rte_free(netdev->tx_q);
551 ovs_mutex_unlock(&netdev->mutex);
556 dpdk_dev_parse_name(const char dev_name[], const char prefix[],
557 unsigned int *port_no)
561 if (strncmp(dev_name, prefix, strlen(prefix))) {
565 cport = dev_name + strlen(prefix);
566 *port_no = strtol(cport, NULL, 0); /* string must be null terminated */
571 netdev_dpdk_vhost_construct(struct netdev *netdev_)
575 if (rte_eal_init_ret) {
576 return rte_eal_init_ret;
579 ovs_mutex_lock(&dpdk_mutex);
580 err = netdev_dpdk_init(netdev_, -1, DPDK_DEV_VHOST);
581 ovs_mutex_unlock(&dpdk_mutex);
587 netdev_dpdk_construct(struct netdev *netdev)
589 unsigned int port_no;
592 if (rte_eal_init_ret) {
593 return rte_eal_init_ret;
596 /* Names always start with "dpdk" */
597 err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no);
602 ovs_mutex_lock(&dpdk_mutex);
603 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
604 ovs_mutex_unlock(&dpdk_mutex);
609 netdev_dpdk_destruct(struct netdev *netdev_)
611 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
613 ovs_mutex_lock(&dev->mutex);
614 rte_eth_dev_stop(dev->port_id);
615 ovs_mutex_unlock(&dev->mutex);
617 ovs_mutex_lock(&dpdk_mutex);
619 list_remove(&dev->list_node);
620 dpdk_mp_put(dev->dpdk_mp);
621 ovs_mutex_unlock(&dpdk_mutex);
625 netdev_dpdk_vhost_destruct(struct netdev *netdev_)
627 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
629 /* Can't remove a port while a guest is attached to it. */
630 if (netdev_dpdk_get_virtio(dev) != NULL) {
631 VLOG_ERR("Can not remove port, vhost device still attached");
635 ovs_mutex_lock(&dpdk_mutex);
636 list_remove(&dev->list_node);
637 dpdk_mp_put(dev->dpdk_mp);
638 ovs_mutex_unlock(&dpdk_mutex);
642 netdev_dpdk_dealloc(struct netdev *netdev_)
644 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
650 netdev_dpdk_get_config(const struct netdev *netdev_, struct smap *args)
652 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
654 ovs_mutex_lock(&dev->mutex);
656 smap_add_format(args, "configured_rx_queues", "%d", netdev_->n_rxq);
657 smap_add_format(args, "configured_tx_queues", "%d", netdev_->n_txq);
658 ovs_mutex_unlock(&dev->mutex);
664 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
666 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
668 return netdev->socket_id;
671 /* Sets the number of tx queues and rx queues for the dpdk interface.
672 * If the configuration fails, do not try restoring its old configuration
673 * and just returns the error. */
675 netdev_dpdk_set_multiq(struct netdev *netdev_, unsigned int n_txq,
678 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
681 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
685 ovs_mutex_lock(&dpdk_mutex);
686 ovs_mutex_lock(&netdev->mutex);
688 rte_eth_dev_stop(netdev->port_id);
690 netdev->up.n_txq = n_txq;
691 netdev->up.n_rxq = n_rxq;
693 rte_free(netdev->tx_q);
694 netdev_dpdk_alloc_txq(netdev, n_txq);
695 err = dpdk_eth_dev_init(netdev);
697 ovs_mutex_unlock(&netdev->mutex);
698 ovs_mutex_unlock(&dpdk_mutex);
704 netdev_dpdk_vhost_set_multiq(struct netdev *netdev_, unsigned int n_txq,
707 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
710 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
714 ovs_mutex_lock(&dpdk_mutex);
715 ovs_mutex_lock(&netdev->mutex);
717 netdev->up.n_txq = n_txq;
718 netdev->up.n_rxq = n_rxq;
720 ovs_mutex_unlock(&netdev->mutex);
721 ovs_mutex_unlock(&dpdk_mutex);
726 static struct netdev_rxq *
727 netdev_dpdk_rxq_alloc(void)
729 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
734 static struct netdev_rxq_dpdk *
735 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
737 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
741 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
743 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
744 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
746 ovs_mutex_lock(&netdev->mutex);
747 rx->port_id = netdev->port_id;
748 ovs_mutex_unlock(&netdev->mutex);
754 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
759 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
761 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
767 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
769 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
772 while (nb_tx != txq->count) {
775 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
784 if (OVS_UNLIKELY(nb_tx != txq->count)) {
785 /* free buffers, which we couldn't transmit, one at a time (each
786 * packet could come from a different mempool) */
789 for (i = nb_tx; i < txq->count; i++) {
790 rte_pktmbuf_free_seg(txq->burst_pkts[i]);
792 rte_spinlock_lock(&dev->stats_lock);
793 dev->stats.tx_dropped += txq->count-nb_tx;
794 rte_spinlock_unlock(&dev->stats_lock);
798 txq->tsc = rte_get_timer_cycles();
802 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
804 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
806 if (txq->count == 0) {
809 dpdk_queue_flush__(dev, qid);
813 is_vhost_running(struct virtio_net *dev)
815 return (dev != NULL && (dev->flags & VIRTIO_DEV_RUNNING));
819 * The receive path for the vhost port is the TX path out from guest.
822 netdev_dpdk_vhost_rxq_recv(struct netdev_rxq *rxq_,
823 struct dp_packet **packets, int *c)
825 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
826 struct netdev *netdev = rx->up.netdev;
827 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
828 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
832 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
836 nb_rx = rte_vhost_dequeue_burst(virtio_dev, qid,
837 vhost_dev->dpdk_mp->mp,
838 (struct rte_mbuf **)packets,
844 rte_spinlock_lock(&vhost_dev->stats_lock);
845 vhost_dev->stats.rx_packets += (uint64_t)nb_rx;
846 rte_spinlock_unlock(&vhost_dev->stats_lock);
853 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dp_packet **packets,
856 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
857 struct netdev *netdev = rx->up.netdev;
858 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
861 /* There is only one tx queue for this core. Do not flush other
863 if (rxq_->queue_id == rte_lcore_id()) {
864 dpdk_queue_flush(dev, rxq_->queue_id);
867 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
868 (struct rte_mbuf **) packets,
880 __netdev_dpdk_vhost_send(struct netdev *netdev, struct dp_packet **pkts,
881 int cnt, bool may_steal)
883 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
884 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
885 struct rte_mbuf **cur_pkts = (struct rte_mbuf **) pkts;
886 unsigned int total_pkts = cnt;
889 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
890 rte_spinlock_lock(&vhost_dev->stats_lock);
891 vhost_dev->stats.tx_dropped+= cnt;
892 rte_spinlock_unlock(&vhost_dev->stats_lock);
896 /* There is vHost TX single queue, So we need to lock it for TX. */
897 rte_spinlock_lock(&vhost_dev->txq_lock);
900 unsigned int tx_pkts;
902 tx_pkts = rte_vhost_enqueue_burst(virtio_dev, VIRTIO_RXQ,
904 if (OVS_LIKELY(tx_pkts)) {
905 /* Packets have been sent.*/
907 /* Prepare for possible next iteration.*/
908 cur_pkts = &cur_pkts[tx_pkts];
910 uint64_t timeout = VHOST_ENQ_RETRY_USECS * rte_get_timer_hz() / 1E6;
911 unsigned int expired = 0;
914 start = rte_get_timer_cycles();
918 * Unable to enqueue packets to vhost interface.
919 * Check available entries before retrying.
921 while (!rte_vring_available_entries(virtio_dev, VIRTIO_RXQ)) {
922 if (OVS_UNLIKELY((rte_get_timer_cycles() - start) > timeout)) {
928 /* break out of main loop. */
934 rte_spinlock_lock(&vhost_dev->stats_lock);
935 vhost_dev->stats.tx_packets += (total_pkts - cnt);
936 vhost_dev->stats.tx_dropped += cnt;
937 rte_spinlock_unlock(&vhost_dev->stats_lock);
938 rte_spinlock_unlock(&vhost_dev->txq_lock);
944 for (i = 0; i < total_pkts; i++) {
945 dp_packet_delete(pkts[i]);
951 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
952 struct rte_mbuf **pkts, int cnt)
954 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
960 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
961 int tocopy = MIN(freeslots, cnt-i);
963 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
964 tocopy * sizeof (struct rte_mbuf *));
966 txq->count += tocopy;
969 if (txq->count == MAX_TX_QUEUE_LEN || txq->flush_tx) {
970 dpdk_queue_flush__(dev, qid);
972 diff_tsc = rte_get_timer_cycles() - txq->tsc;
973 if (diff_tsc >= DRAIN_TSC) {
974 dpdk_queue_flush__(dev, qid);
979 /* Tx function. Transmit packets indefinitely */
981 dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet **pkts,
983 OVS_NO_THREAD_SAFETY_ANALYSIS
985 #if !defined(__CHECKER__) && !defined(_WIN32)
986 const size_t PKT_ARRAY_SIZE = cnt;
988 /* Sparse or MSVC doesn't like variable length array. */
989 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
991 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
992 struct rte_mbuf *mbufs[PKT_ARRAY_SIZE];
997 /* If we are on a non pmd thread we have to use the mempool mutex, because
998 * every non pmd thread shares the same mempool cache */
1000 if (!thread_is_pmd()) {
1001 ovs_mutex_lock(&nonpmd_mempool_mutex);
1004 for (i = 0; i < cnt; i++) {
1005 int size = dp_packet_size(pkts[i]);
1007 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1008 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1009 (int)size , dev->max_packet_len);
1015 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
1017 if (!mbufs[newcnt]) {
1022 /* We have to do a copy for now */
1023 memcpy(rte_pktmbuf_mtod(mbufs[newcnt], void *), dp_packet_data(pkts[i]), size);
1025 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
1026 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
1031 if (OVS_UNLIKELY(dropped)) {
1032 rte_spinlock_lock(&dev->stats_lock);
1033 dev->stats.tx_dropped += dropped;
1034 rte_spinlock_unlock(&dev->stats_lock);
1037 if (dev->type == DPDK_DEV_VHOST) {
1038 __netdev_dpdk_vhost_send(netdev, (struct dp_packet **) mbufs, newcnt, true);
1040 dpdk_queue_pkts(dev, qid, mbufs, newcnt);
1041 dpdk_queue_flush(dev, qid);
1044 if (!thread_is_pmd()) {
1045 ovs_mutex_unlock(&nonpmd_mempool_mutex);
1050 netdev_dpdk_vhost_send(struct netdev *netdev, int qid OVS_UNUSED, struct dp_packet **pkts,
1051 int cnt, bool may_steal)
1053 if (OVS_UNLIKELY(pkts[0]->source != DPBUF_DPDK)) {
1056 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1058 for (i = 0; i < cnt; i++) {
1059 dp_packet_delete(pkts[i]);
1063 __netdev_dpdk_vhost_send(netdev, pkts, cnt, may_steal);
1069 netdev_dpdk_send__(struct netdev_dpdk *dev, int qid,
1070 struct dp_packet **pkts, int cnt, bool may_steal)
1074 if (OVS_UNLIKELY(!may_steal ||
1075 pkts[0]->source != DPBUF_DPDK)) {
1076 struct netdev *netdev = &dev->up;
1078 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1081 for (i = 0; i < cnt; i++) {
1082 dp_packet_delete(pkts[i]);
1086 int next_tx_idx = 0;
1089 for (i = 0; i < cnt; i++) {
1090 int size = dp_packet_size(pkts[i]);
1092 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1093 if (next_tx_idx != i) {
1094 dpdk_queue_pkts(dev, qid,
1095 (struct rte_mbuf **)&pkts[next_tx_idx],
1099 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1100 (int)size , dev->max_packet_len);
1102 dp_packet_delete(pkts[i]);
1104 next_tx_idx = i + 1;
1107 if (next_tx_idx != cnt) {
1108 dpdk_queue_pkts(dev, qid,
1109 (struct rte_mbuf **)&pkts[next_tx_idx],
1113 if (OVS_UNLIKELY(dropped)) {
1114 rte_spinlock_lock(&dev->stats_lock);
1115 dev->stats.tx_dropped += dropped;
1116 rte_spinlock_unlock(&dev->stats_lock);
1122 netdev_dpdk_eth_send(struct netdev *netdev, int qid,
1123 struct dp_packet **pkts, int cnt, bool may_steal)
1125 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1127 netdev_dpdk_send__(dev, qid, pkts, cnt, may_steal);
1132 netdev_dpdk_set_etheraddr(struct netdev *netdev,
1133 const uint8_t mac[ETH_ADDR_LEN])
1135 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1137 ovs_mutex_lock(&dev->mutex);
1138 if (!eth_addr_equals(dev->hwaddr, mac)) {
1139 memcpy(dev->hwaddr, mac, ETH_ADDR_LEN);
1140 netdev_change_seq_changed(netdev);
1142 ovs_mutex_unlock(&dev->mutex);
1148 netdev_dpdk_get_etheraddr(const struct netdev *netdev,
1149 uint8_t mac[ETH_ADDR_LEN])
1151 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1153 ovs_mutex_lock(&dev->mutex);
1154 memcpy(mac, dev->hwaddr, ETH_ADDR_LEN);
1155 ovs_mutex_unlock(&dev->mutex);
1161 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
1163 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1165 ovs_mutex_lock(&dev->mutex);
1167 ovs_mutex_unlock(&dev->mutex);
1173 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
1175 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1177 struct dpdk_mp *old_mp;
1180 ovs_mutex_lock(&dpdk_mutex);
1181 ovs_mutex_lock(&dev->mutex);
1182 if (dev->mtu == mtu) {
1187 mp = dpdk_mp_get(dev->socket_id, dev->mtu);
1193 rte_eth_dev_stop(dev->port_id);
1196 old_mp = dev->dpdk_mp;
1199 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1201 err = dpdk_eth_dev_init(dev);
1205 dev->dpdk_mp = old_mp;
1206 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1207 dpdk_eth_dev_init(dev);
1211 dpdk_mp_put(old_mp);
1212 netdev_change_seq_changed(netdev);
1214 ovs_mutex_unlock(&dev->mutex);
1215 ovs_mutex_unlock(&dpdk_mutex);
1220 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
1223 netdev_dpdk_vhost_get_stats(const struct netdev *netdev,
1224 struct netdev_stats *stats)
1226 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1228 ovs_mutex_lock(&dev->mutex);
1229 memset(stats, 0, sizeof(*stats));
1230 /* Unsupported Stats */
1231 stats->rx_errors = UINT64_MAX;
1232 stats->tx_errors = UINT64_MAX;
1233 stats->multicast = UINT64_MAX;
1234 stats->collisions = UINT64_MAX;
1235 stats->rx_crc_errors = UINT64_MAX;
1236 stats->rx_fifo_errors = UINT64_MAX;
1237 stats->rx_frame_errors = UINT64_MAX;
1238 stats->rx_length_errors = UINT64_MAX;
1239 stats->rx_missed_errors = UINT64_MAX;
1240 stats->rx_over_errors = UINT64_MAX;
1241 stats->tx_aborted_errors = UINT64_MAX;
1242 stats->tx_carrier_errors = UINT64_MAX;
1243 stats->tx_errors = UINT64_MAX;
1244 stats->tx_fifo_errors = UINT64_MAX;
1245 stats->tx_heartbeat_errors = UINT64_MAX;
1246 stats->tx_window_errors = UINT64_MAX;
1247 stats->rx_bytes += UINT64_MAX;
1248 stats->rx_dropped += UINT64_MAX;
1249 stats->tx_bytes += UINT64_MAX;
1251 rte_spinlock_lock(&dev->stats_lock);
1252 /* Supported Stats */
1253 stats->rx_packets += dev->stats.rx_packets;
1254 stats->tx_packets += dev->stats.tx_packets;
1255 stats->tx_dropped += dev->stats.tx_dropped;
1256 rte_spinlock_unlock(&dev->stats_lock);
1257 ovs_mutex_unlock(&dev->mutex);
1263 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
1265 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1266 struct rte_eth_stats rte_stats;
1269 netdev_dpdk_get_carrier(netdev, &gg);
1270 ovs_mutex_lock(&dev->mutex);
1271 rte_eth_stats_get(dev->port_id, &rte_stats);
1273 memset(stats, 0, sizeof(*stats));
1275 stats->rx_packets = rte_stats.ipackets;
1276 stats->tx_packets = rte_stats.opackets;
1277 stats->rx_bytes = rte_stats.ibytes;
1278 stats->tx_bytes = rte_stats.obytes;
1279 stats->rx_errors = rte_stats.ierrors;
1280 stats->tx_errors = rte_stats.oerrors;
1281 stats->multicast = rte_stats.imcasts;
1283 rte_spinlock_lock(&dev->stats_lock);
1284 stats->tx_dropped = dev->stats.tx_dropped;
1285 rte_spinlock_unlock(&dev->stats_lock);
1286 ovs_mutex_unlock(&dev->mutex);
1292 netdev_dpdk_get_features(const struct netdev *netdev_,
1293 enum netdev_features *current,
1294 enum netdev_features *advertised OVS_UNUSED,
1295 enum netdev_features *supported OVS_UNUSED,
1296 enum netdev_features *peer OVS_UNUSED)
1298 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1299 struct rte_eth_link link;
1301 ovs_mutex_lock(&dev->mutex);
1303 ovs_mutex_unlock(&dev->mutex);
1305 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
1306 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
1307 *current = NETDEV_F_AUTONEG;
1309 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
1310 if (link.link_speed == ETH_LINK_SPEED_10) {
1311 *current = NETDEV_F_10MB_HD;
1313 if (link.link_speed == ETH_LINK_SPEED_100) {
1314 *current = NETDEV_F_100MB_HD;
1316 if (link.link_speed == ETH_LINK_SPEED_1000) {
1317 *current = NETDEV_F_1GB_HD;
1319 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
1320 if (link.link_speed == ETH_LINK_SPEED_10) {
1321 *current = NETDEV_F_10MB_FD;
1323 if (link.link_speed == ETH_LINK_SPEED_100) {
1324 *current = NETDEV_F_100MB_FD;
1326 if (link.link_speed == ETH_LINK_SPEED_1000) {
1327 *current = NETDEV_F_1GB_FD;
1329 if (link.link_speed == ETH_LINK_SPEED_10000) {
1330 *current = NETDEV_F_10GB_FD;
1338 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1340 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1343 ovs_mutex_lock(&dev->mutex);
1344 ifindex = dev->port_id;
1345 ovs_mutex_unlock(&dev->mutex);
1351 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1353 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1355 ovs_mutex_lock(&dev->mutex);
1356 check_link_status(dev);
1357 *carrier = dev->link.link_status;
1359 ovs_mutex_unlock(&dev->mutex);
1365 netdev_dpdk_vhost_get_carrier(const struct netdev *netdev_, bool *carrier)
1367 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1368 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(dev);
1370 ovs_mutex_lock(&dev->mutex);
1372 if (is_vhost_running(virtio_dev)) {
1378 ovs_mutex_unlock(&dev->mutex);
1383 static long long int
1384 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1386 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1387 long long int carrier_resets;
1389 ovs_mutex_lock(&dev->mutex);
1390 carrier_resets = dev->link_reset_cnt;
1391 ovs_mutex_unlock(&dev->mutex);
1393 return carrier_resets;
1397 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1398 long long int interval OVS_UNUSED)
1404 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1405 enum netdev_flags off, enum netdev_flags on,
1406 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1410 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1414 *old_flagsp = dev->flags;
1418 if (dev->flags == *old_flagsp) {
1422 if (dev->type == DPDK_DEV_ETH) {
1423 if (dev->flags & NETDEV_UP) {
1424 err = rte_eth_dev_start(dev->port_id);
1429 if (dev->flags & NETDEV_PROMISC) {
1430 rte_eth_promiscuous_enable(dev->port_id);
1433 if (!(dev->flags & NETDEV_UP)) {
1434 rte_eth_dev_stop(dev->port_id);
1442 netdev_dpdk_update_flags(struct netdev *netdev_,
1443 enum netdev_flags off, enum netdev_flags on,
1444 enum netdev_flags *old_flagsp)
1446 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1449 ovs_mutex_lock(&netdev->mutex);
1450 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1451 ovs_mutex_unlock(&netdev->mutex);
1457 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1459 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1460 struct rte_eth_dev_info dev_info;
1462 if (dev->port_id < 0)
1465 ovs_mutex_lock(&dev->mutex);
1466 rte_eth_dev_info_get(dev->port_id, &dev_info);
1467 ovs_mutex_unlock(&dev->mutex);
1469 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1471 smap_add_format(args, "port_no", "%d", dev->port_id);
1472 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1473 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1474 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1475 smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen);
1476 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1477 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1478 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1479 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1480 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1481 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1483 smap_add_format(args, "pci-vendor_id", "0x%u", dev_info.pci_dev->id.vendor_id);
1484 smap_add_format(args, "pci-device_id", "0x%x", dev_info.pci_dev->id.device_id);
1490 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1491 OVS_REQUIRES(dev->mutex)
1493 enum netdev_flags old_flags;
1496 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1498 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1503 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1504 const char *argv[], void *aux OVS_UNUSED)
1508 if (!strcasecmp(argv[argc - 1], "up")) {
1510 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1513 unixctl_command_reply_error(conn, "Invalid Admin State");
1518 struct netdev *netdev = netdev_from_name(argv[1]);
1519 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1520 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1522 ovs_mutex_lock(&dpdk_dev->mutex);
1523 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1524 ovs_mutex_unlock(&dpdk_dev->mutex);
1526 netdev_close(netdev);
1528 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1529 netdev_close(netdev);
1533 struct netdev_dpdk *netdev;
1535 ovs_mutex_lock(&dpdk_mutex);
1536 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1537 ovs_mutex_lock(&netdev->mutex);
1538 netdev_dpdk_set_admin_state__(netdev, up);
1539 ovs_mutex_unlock(&netdev->mutex);
1541 ovs_mutex_unlock(&dpdk_mutex);
1543 unixctl_command_reply(conn, "OK");
1547 * Set virtqueue flags so that we do not receive interrupts.
1550 set_irq_status(struct virtio_net *dev)
1552 dev->virtqueue[VIRTIO_RXQ]->used->flags = VRING_USED_F_NO_NOTIFY;
1553 dev->virtqueue[VIRTIO_TXQ]->used->flags = VRING_USED_F_NO_NOTIFY;
1557 * A new virtio-net device is added to a vhost port.
1560 new_device(struct virtio_net *dev)
1562 struct netdev_dpdk *netdev;
1563 bool exists = false;
1565 ovs_mutex_lock(&dpdk_mutex);
1566 /* Add device to the vhost port with the same name as that passed down. */
1567 LIST_FOR_EACH(netdev, list_node, &dpdk_list) {
1568 if (strncmp(dev->ifname, netdev->up.name, IFNAMSIZ) == 0) {
1569 ovs_mutex_lock(&netdev->mutex);
1570 ovsrcu_set(&netdev->virtio_dev, dev);
1571 ovs_mutex_unlock(&netdev->mutex);
1573 dev->flags |= VIRTIO_DEV_RUNNING;
1574 /* Disable notifications. */
1575 set_irq_status(dev);
1579 ovs_mutex_unlock(&dpdk_mutex);
1582 VLOG_INFO("vHost Device '%s' (%ld) can't be added - name not found",
1583 dev->ifname, dev->device_fh);
1588 VLOG_INFO("vHost Device '%s' (%ld) has been added",
1589 dev->ifname, dev->device_fh);
1594 * Remove a virtio-net device from the specific vhost port. Use dev->remove
1595 * flag to stop any more packets from being sent or received to/from a VM and
1596 * ensure all currently queued packets have been sent/received before removing
1600 destroy_device(volatile struct virtio_net *dev)
1602 struct netdev_dpdk *vhost_dev;
1604 ovs_mutex_lock(&dpdk_mutex);
1605 LIST_FOR_EACH (vhost_dev, list_node, &dpdk_list) {
1606 if (netdev_dpdk_get_virtio(vhost_dev) == dev) {
1608 ovs_mutex_lock(&vhost_dev->mutex);
1609 dev->flags &= ~VIRTIO_DEV_RUNNING;
1610 ovsrcu_set(&vhost_dev->virtio_dev, NULL);
1611 ovs_mutex_unlock(&vhost_dev->mutex);
1614 * Wait for other threads to quiesce before
1615 * setting the virtio_dev to NULL.
1617 ovsrcu_synchronize();
1619 * As call to ovsrcu_synchronize() will end the quiescent state,
1620 * put thread back into quiescent state before returning.
1622 ovsrcu_quiesce_start();
1625 ovs_mutex_unlock(&dpdk_mutex);
1627 VLOG_INFO("vHost Device '%s' (%ld) has been removed",
1628 dev->ifname, dev->device_fh);
1632 netdev_dpdk_get_virtio(const struct netdev_dpdk *dev)
1634 return ovsrcu_get(struct virtio_net *, &dev->virtio_dev);
1638 * These callbacks allow virtio-net devices to be added to vhost ports when
1639 * configuration has been fully complete.
1641 static const struct virtio_net_device_ops virtio_net_device_ops =
1643 .new_device = new_device,
1644 .destroy_device = destroy_device,
1648 start_cuse_session_loop(void *dummy OVS_UNUSED)
1650 pthread_detach(pthread_self());
1651 /* Put the cuse thread into quiescent state. */
1652 ovsrcu_quiesce_start();
1653 rte_vhost_driver_session_start();
1658 dpdk_vhost_class_init(void)
1662 rte_vhost_driver_callback_register(&virtio_net_device_ops);
1664 /* Register CUSE device to handle IOCTLs.
1665 * Unless otherwise specified on the vswitchd command line, cuse_dev_name
1666 * is set to vhost-net.
1668 err = rte_vhost_driver_register(cuse_dev_name);
1671 VLOG_ERR("CUSE device setup failure.");
1675 ovs_thread_create("cuse_thread", start_cuse_session_loop, NULL);
1680 dpdk_common_init(void)
1682 unixctl_command_register("netdev-dpdk/set-admin-state",
1683 "[netdev] up|down", 1, 2,
1684 netdev_dpdk_set_admin_state, NULL);
1686 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
1692 dpdk_ring_create(const char dev_name[], unsigned int port_no,
1693 unsigned int *eth_port_id)
1695 struct dpdk_ring *ivshmem;
1699 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
1700 if (ivshmem == NULL) {
1704 /* XXX: Add support for multiquque ring. */
1705 err = snprintf(ring_name, 10, "%s_tx", dev_name);
1710 /* Create single consumer/producer rings, netdev does explicit locking. */
1711 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1712 RING_F_SP_ENQ | RING_F_SC_DEQ);
1713 if (ivshmem->cring_tx == NULL) {
1718 err = snprintf(ring_name, 10, "%s_rx", dev_name);
1723 /* Create single consumer/producer rings, netdev does explicit locking. */
1724 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1725 RING_F_SP_ENQ | RING_F_SC_DEQ);
1726 if (ivshmem->cring_rx == NULL) {
1731 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
1732 &ivshmem->cring_tx, 1, SOCKET0);
1739 ivshmem->user_port_id = port_no;
1740 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
1741 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
1743 *eth_port_id = ivshmem->eth_port_id;
1748 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
1750 struct dpdk_ring *ivshmem;
1751 unsigned int port_no;
1754 /* Names always start with "dpdkr" */
1755 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
1760 /* look through our list to find the device */
1761 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
1762 if (ivshmem->user_port_id == port_no) {
1763 VLOG_INFO("Found dpdk ring device %s:", dev_name);
1764 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
1768 /* Need to create the device rings */
1769 return dpdk_ring_create(dev_name, port_no, eth_port_id);
1773 netdev_dpdk_ring_send(struct netdev *netdev, int qid OVS_UNUSED,
1774 struct dp_packet **pkts, int cnt, bool may_steal)
1776 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1779 /* When using 'dpdkr' and sending to a DPDK ring, we want to ensure that the
1780 * rss hash field is clear. This is because the same mbuf may be modified by
1781 * the consumer of the ring and return into the datapath without recalculating
1783 for (i = 0; i < cnt; i++) {
1784 dp_packet_set_rss_hash(pkts[i], 0);
1787 /* DPDK Rings have a single TX queue, Therefore needs locking. */
1788 rte_spinlock_lock(&dev->txq_lock);
1789 netdev_dpdk_send__(dev, 0, pkts, cnt, may_steal);
1790 rte_spinlock_unlock(&dev->txq_lock);
1795 netdev_dpdk_ring_construct(struct netdev *netdev)
1797 unsigned int port_no = 0;
1800 if (rte_eal_init_ret) {
1801 return rte_eal_init_ret;
1804 ovs_mutex_lock(&dpdk_mutex);
1806 err = dpdk_ring_open(netdev->name, &port_no);
1811 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
1814 ovs_mutex_unlock(&dpdk_mutex);
1818 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, DESTRUCT, MULTIQ, SEND, \
1819 GET_CARRIER, GET_STATS, GET_FEATURES, GET_STATUS, RXQ_RECV) \
1823 NULL, /* netdev_dpdk_run */ \
1824 NULL, /* netdev_dpdk_wait */ \
1826 netdev_dpdk_alloc, \
1829 netdev_dpdk_dealloc, \
1830 netdev_dpdk_get_config, \
1831 NULL, /* netdev_dpdk_set_config */ \
1832 NULL, /* get_tunnel_config */ \
1833 NULL, /* build header */ \
1834 NULL, /* push header */ \
1835 NULL, /* pop header */ \
1836 netdev_dpdk_get_numa_id, /* get_numa_id */ \
1837 MULTIQ, /* set_multiq */ \
1840 NULL, /* send_wait */ \
1842 netdev_dpdk_set_etheraddr, \
1843 netdev_dpdk_get_etheraddr, \
1844 netdev_dpdk_get_mtu, \
1845 netdev_dpdk_set_mtu, \
1846 netdev_dpdk_get_ifindex, \
1848 netdev_dpdk_get_carrier_resets, \
1849 netdev_dpdk_set_miimon, \
1852 NULL, /* set_advertisements */ \
1854 NULL, /* set_policing */ \
1855 NULL, /* get_qos_types */ \
1856 NULL, /* get_qos_capabilities */ \
1857 NULL, /* get_qos */ \
1858 NULL, /* set_qos */ \
1859 NULL, /* get_queue */ \
1860 NULL, /* set_queue */ \
1861 NULL, /* delete_queue */ \
1862 NULL, /* get_queue_stats */ \
1863 NULL, /* queue_dump_start */ \
1864 NULL, /* queue_dump_next */ \
1865 NULL, /* queue_dump_done */ \
1866 NULL, /* dump_queue_stats */ \
1868 NULL, /* get_in4 */ \
1869 NULL, /* set_in4 */ \
1870 NULL, /* get_in6 */ \
1871 NULL, /* add_router */ \
1872 NULL, /* get_next_hop */ \
1874 NULL, /* arp_lookup */ \
1876 netdev_dpdk_update_flags, \
1878 netdev_dpdk_rxq_alloc, \
1879 netdev_dpdk_rxq_construct, \
1880 netdev_dpdk_rxq_destruct, \
1881 netdev_dpdk_rxq_dealloc, \
1883 NULL, /* rx_wait */ \
1884 NULL, /* rxq_drain */ \
1888 dpdk_init(int argc, char **argv)
1892 char *pragram_name = argv[0];
1894 if (argc < 2 || strcmp(argv[1], "--dpdk"))
1897 /* Remove the --dpdk argument from arg list.*/
1901 /* If the cuse_dev_name parameter has been provided, set 'cuse_dev_name' to
1902 * this string if it meets the correct criteria. Otherwise, set it to the
1903 * default (vhost-net).
1905 if (!strcmp(argv[1], "--cuse_dev_name") &&
1906 (strlen(argv[2]) <= NAME_MAX)) {
1908 cuse_dev_name = strdup(argv[2]);
1910 /* Remove the cuse_dev_name configuration parameters from the argument
1911 * list, so that the correct elements are passed to the DPDK
1912 * initialization function
1915 argv += 2; /* Increment by two to bypass the cuse_dev_name arguments */
1918 VLOG_ERR("User-provided cuse_dev_name in use: /dev/%s", cuse_dev_name);
1920 cuse_dev_name = "vhost-net";
1921 VLOG_INFO("No cuse_dev_name provided - defaulting to /dev/vhost-net");
1924 /* Keep the program name argument as this is needed for call to
1927 argv[0] = pragram_name;
1929 /* Make sure things are initialized ... */
1930 result = rte_eal_init(argc, argv);
1932 ovs_abort(result, "Cannot init EAL");
1935 rte_memzone_dump(stdout);
1936 rte_eal_init_ret = 0;
1938 if (argc > result) {
1939 argv[result] = argv[0];
1942 /* We are called from the main thread here */
1943 RTE_PER_LCORE(_lcore_id) = NON_PMD_CORE_ID;
1945 return result + 1 + base;
1948 static const struct netdev_class dpdk_class =
1952 netdev_dpdk_construct,
1953 netdev_dpdk_destruct,
1954 netdev_dpdk_set_multiq,
1955 netdev_dpdk_eth_send,
1956 netdev_dpdk_get_carrier,
1957 netdev_dpdk_get_stats,
1958 netdev_dpdk_get_features,
1959 netdev_dpdk_get_status,
1960 netdev_dpdk_rxq_recv);
1962 static const struct netdev_class dpdk_ring_class =
1966 netdev_dpdk_ring_construct,
1967 netdev_dpdk_destruct,
1969 netdev_dpdk_ring_send,
1970 netdev_dpdk_get_carrier,
1971 netdev_dpdk_get_stats,
1972 netdev_dpdk_get_features,
1973 netdev_dpdk_get_status,
1974 netdev_dpdk_rxq_recv);
1976 static const struct netdev_class dpdk_vhost_class =
1979 dpdk_vhost_class_init,
1980 netdev_dpdk_vhost_construct,
1981 netdev_dpdk_vhost_destruct,
1982 netdev_dpdk_vhost_set_multiq,
1983 netdev_dpdk_vhost_send,
1984 netdev_dpdk_vhost_get_carrier,
1985 netdev_dpdk_vhost_get_stats,
1988 netdev_dpdk_vhost_rxq_recv);
1991 netdev_dpdk_register(void)
1993 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
1995 if (rte_eal_init_ret) {
1999 if (ovsthread_once_start(&once)) {
2001 netdev_register_provider(&dpdk_class);
2002 netdev_register_provider(&dpdk_ring_class);
2003 netdev_register_provider(&dpdk_vhost_class);
2004 ovsthread_once_done(&once);
2009 pmd_thread_setaffinity_cpu(unsigned cpu)
2015 CPU_SET(cpu, &cpuset);
2016 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
2018 VLOG_ERR("Thread affinity error %d",err);
2021 /* NON_PMD_CORE_ID is reserved for use by non pmd threads. */
2022 ovs_assert(cpu != NON_PMD_CORE_ID);
2023 RTE_PER_LCORE(_lcore_id) = cpu;
2031 return rte_lcore_id() != NON_PMD_CORE_ID;