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 /* XXX: Needs per NIC value for these constants. */
94 #define RX_PTHRESH 32 /* Default values of RX prefetch threshold reg. */
95 #define RX_HTHRESH 32 /* Default values of RX host threshold reg. */
96 #define RX_WTHRESH 16 /* Default values of RX write-back threshold reg. */
98 #define TX_PTHRESH 36 /* Default values of TX prefetch threshold reg. */
99 #define TX_HTHRESH 0 /* Default values of TX host threshold reg. */
100 #define TX_WTHRESH 0 /* Default values of TX write-back threshold reg. */
102 #define MAX_PKT_BURST 32 /* Max burst size for RX/TX */
104 /* Character device cuse_dev_name. */
105 char *cuse_dev_name = NULL;
107 static const struct rte_eth_conf port_conf = {
109 .mq_mode = ETH_MQ_RX_RSS,
111 .header_split = 0, /* Header Split disabled */
112 .hw_ip_checksum = 0, /* IP checksum offload disabled */
113 .hw_vlan_filter = 0, /* VLAN filtering disabled */
114 .jumbo_frame = 0, /* Jumbo Frame Support disabled */
120 .rss_hf = ETH_RSS_IP | ETH_RSS_UDP | ETH_RSS_TCP,
124 .mq_mode = ETH_MQ_TX_NONE,
128 static const struct rte_eth_rxconf rx_conf = {
130 .pthresh = RX_PTHRESH,
131 .hthresh = RX_HTHRESH,
132 .wthresh = RX_WTHRESH,
136 static const struct rte_eth_txconf tx_conf = {
138 .pthresh = TX_PTHRESH,
139 .hthresh = TX_HTHRESH,
140 .wthresh = TX_WTHRESH,
144 .txq_flags = ETH_TXQ_FLAGS_NOMULTSEGS|ETH_TXQ_FLAGS_NOOFFLOADS,
147 enum { MAX_TX_QUEUE_LEN = 384 };
148 enum { DPDK_RING_SIZE = 256 };
149 BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE));
150 enum { DRAIN_TSC = 200000ULL };
157 static int rte_eal_init_ret = ENODEV;
159 static struct ovs_mutex dpdk_mutex = OVS_MUTEX_INITIALIZER;
161 /* Contains all 'struct dpdk_dev's. */
162 static struct ovs_list dpdk_list OVS_GUARDED_BY(dpdk_mutex)
163 = OVS_LIST_INITIALIZER(&dpdk_list);
165 static struct ovs_list dpdk_mp_list OVS_GUARDED_BY(dpdk_mutex)
166 = OVS_LIST_INITIALIZER(&dpdk_mp_list);
168 /* This mutex must be used by non pmd threads when allocating or freeing
169 * mbufs through mempools. Since dpdk_queue_pkts() and dpdk_queue_flush() may
170 * use mempools, a non pmd thread should hold this mutex while calling them */
171 struct ovs_mutex nonpmd_mempool_mutex = OVS_MUTEX_INITIALIZER;
174 struct rte_mempool *mp;
178 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
181 /* There should be one 'struct dpdk_tx_queue' created for
183 struct dpdk_tx_queue {
184 bool flush_tx; /* Set to true to flush queue everytime */
185 /* pkts are queued. */
188 struct rte_mbuf *burst_pkts[MAX_TX_QUEUE_LEN];
191 /* dpdk has no way to remove dpdk ring ethernet devices
192 so we have to keep them around once they've been created
195 static struct ovs_list dpdk_ring_list OVS_GUARDED_BY(dpdk_mutex)
196 = OVS_LIST_INITIALIZER(&dpdk_ring_list);
199 /* For the client rings */
200 struct rte_ring *cring_tx;
201 struct rte_ring *cring_rx;
202 int user_port_id; /* User given port no, parsed from port name */
203 int eth_port_id; /* ethernet device port id */
204 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
211 enum dpdk_dev_type type;
213 struct dpdk_tx_queue *tx_q;
215 struct ovs_mutex mutex OVS_ACQ_AFTER(dpdk_mutex);
217 struct dpdk_mp *dpdk_mp;
221 struct netdev_stats stats;
223 uint8_t hwaddr[ETH_ADDR_LEN];
224 enum netdev_flags flags;
226 struct rte_eth_link link;
229 /* virtio-net structure for vhost device */
230 OVSRCU_TYPE(struct virtio_net *) virtio_dev;
233 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
234 rte_spinlock_t txq_lock;
237 struct netdev_rxq_dpdk {
238 struct netdev_rxq up;
242 static bool thread_is_pmd(void);
244 static int netdev_dpdk_construct(struct netdev *);
246 struct virtio_net * netdev_dpdk_get_virtio(const struct netdev_dpdk *dev);
249 is_dpdk_class(const struct netdev_class *class)
251 return class->construct == netdev_dpdk_construct;
254 /* XXX: use dpdk malloc for entire OVS. in fact huge page should be used
255 * for all other segments data, bss and text. */
258 dpdk_rte_mzalloc(size_t sz)
262 ptr = rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE);
269 /* XXX this function should be called only by pmd threads (or by non pmd
270 * threads holding the nonpmd_mempool_mutex) */
272 free_dpdk_buf(struct dp_packet *p)
274 struct rte_mbuf *pkt = (struct rte_mbuf *) p;
276 rte_pktmbuf_free_seg(pkt);
280 __rte_pktmbuf_init(struct rte_mempool *mp,
281 void *opaque_arg OVS_UNUSED,
283 unsigned i OVS_UNUSED)
285 struct rte_mbuf *m = _m;
286 uint32_t buf_len = mp->elt_size - sizeof(struct dp_packet);
288 RTE_MBUF_ASSERT(mp->elt_size >= sizeof(struct dp_packet));
290 memset(m, 0, mp->elt_size);
292 /* start of buffer is just after mbuf structure */
293 m->buf_addr = (char *)m + sizeof(struct dp_packet);
294 m->buf_physaddr = rte_mempool_virt2phy(mp, m) +
295 sizeof(struct dp_packet);
296 m->buf_len = (uint16_t)buf_len;
298 /* keep some headroom between start of buffer and data */
299 m->data_off = RTE_MIN(RTE_PKTMBUF_HEADROOM, m->buf_len);
301 /* init some constant fields */
308 ovs_rte_pktmbuf_init(struct rte_mempool *mp,
309 void *opaque_arg OVS_UNUSED,
311 unsigned i OVS_UNUSED)
313 struct rte_mbuf *m = _m;
315 __rte_pktmbuf_init(mp, opaque_arg, _m, i);
317 dp_packet_init_dpdk((struct dp_packet *) m, m->buf_len);
320 static struct dpdk_mp *
321 dpdk_mp_get(int socket_id, int mtu) OVS_REQUIRES(dpdk_mutex)
323 struct dpdk_mp *dmp = NULL;
324 char mp_name[RTE_MEMPOOL_NAMESIZE];
327 LIST_FOR_EACH (dmp, list_node, &dpdk_mp_list) {
328 if (dmp->socket_id == socket_id && dmp->mtu == mtu) {
334 dmp = dpdk_rte_mzalloc(sizeof *dmp);
335 dmp->socket_id = socket_id;
339 mp_size = MAX_NB_MBUF;
341 if (snprintf(mp_name, RTE_MEMPOOL_NAMESIZE, "ovs_mp_%d_%d_%u",
342 dmp->mtu, dmp->socket_id, mp_size) < 0) {
346 dmp->mp = rte_mempool_create(mp_name, mp_size, MBUF_SIZE(mtu),
348 sizeof(struct rte_pktmbuf_pool_private),
349 rte_pktmbuf_pool_init, NULL,
350 ovs_rte_pktmbuf_init, NULL,
352 } while (!dmp->mp && rte_errno == ENOMEM && (mp_size /= 2) >= MIN_NB_MBUF);
354 if (dmp->mp == NULL) {
357 VLOG_DBG("Allocated \"%s\" mempool with %u mbufs", mp_name, mp_size );
360 list_push_back(&dpdk_mp_list, &dmp->list_node);
365 dpdk_mp_put(struct dpdk_mp *dmp)
373 ovs_assert(dmp->refcount >= 0);
376 /* I could not find any API to destroy mp. */
377 if (dmp->refcount == 0) {
378 list_delete(dmp->list_node);
379 /* destroy mp-pool. */
385 check_link_status(struct netdev_dpdk *dev)
387 struct rte_eth_link link;
389 rte_eth_link_get_nowait(dev->port_id, &link);
391 if (dev->link.link_status != link.link_status) {
392 netdev_change_seq_changed(&dev->up);
394 dev->link_reset_cnt++;
396 if (dev->link.link_status) {
397 VLOG_DBG_RL(&rl, "Port %d Link Up - speed %u Mbps - %s",
398 dev->port_id, (unsigned)dev->link.link_speed,
399 (dev->link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
400 ("full-duplex") : ("half-duplex"));
402 VLOG_DBG_RL(&rl, "Port %d Link Down", dev->port_id);
408 dpdk_watchdog(void *dummy OVS_UNUSED)
410 struct netdev_dpdk *dev;
412 pthread_detach(pthread_self());
415 ovs_mutex_lock(&dpdk_mutex);
416 LIST_FOR_EACH (dev, list_node, &dpdk_list) {
417 ovs_mutex_lock(&dev->mutex);
418 check_link_status(dev);
419 ovs_mutex_unlock(&dev->mutex);
421 ovs_mutex_unlock(&dpdk_mutex);
422 xsleep(DPDK_PORT_WATCHDOG_INTERVAL);
429 dpdk_eth_dev_init(struct netdev_dpdk *dev) OVS_REQUIRES(dpdk_mutex)
431 struct rte_pktmbuf_pool_private *mbp_priv;
432 struct ether_addr eth_addr;
436 if (dev->port_id < 0 || dev->port_id >= rte_eth_dev_count()) {
440 diag = rte_eth_dev_configure(dev->port_id, dev->up.n_rxq, dev->up.n_txq,
443 VLOG_ERR("eth dev config error %d",diag);
447 for (i = 0; i < dev->up.n_txq; i++) {
448 diag = rte_eth_tx_queue_setup(dev->port_id, i, NIC_PORT_TX_Q_SIZE,
449 dev->socket_id, &tx_conf);
451 VLOG_ERR("eth dev tx queue setup error %d",diag);
456 for (i = 0; i < dev->up.n_rxq; i++) {
457 diag = rte_eth_rx_queue_setup(dev->port_id, i, NIC_PORT_RX_Q_SIZE,
459 &rx_conf, dev->dpdk_mp->mp);
461 VLOG_ERR("eth dev rx queue setup error %d",diag);
466 diag = rte_eth_dev_start(dev->port_id);
468 VLOG_ERR("eth dev start error %d",diag);
472 rte_eth_promiscuous_enable(dev->port_id);
473 rte_eth_allmulticast_enable(dev->port_id);
475 memset(ð_addr, 0x0, sizeof(eth_addr));
476 rte_eth_macaddr_get(dev->port_id, ð_addr);
477 VLOG_INFO_RL(&rl, "Port %d: "ETH_ADDR_FMT"",
478 dev->port_id, ETH_ADDR_ARGS(eth_addr.addr_bytes));
480 memcpy(dev->hwaddr, eth_addr.addr_bytes, ETH_ADDR_LEN);
481 rte_eth_link_get_nowait(dev->port_id, &dev->link);
483 mbp_priv = rte_mempool_get_priv(dev->dpdk_mp->mp);
484 dev->buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
486 dev->flags = NETDEV_UP | NETDEV_PROMISC;
490 static struct netdev_dpdk *
491 netdev_dpdk_cast(const struct netdev *netdev)
493 return CONTAINER_OF(netdev, struct netdev_dpdk, up);
496 static struct netdev *
497 netdev_dpdk_alloc(void)
499 struct netdev_dpdk *netdev = dpdk_rte_mzalloc(sizeof *netdev);
504 netdev_dpdk_alloc_txq(struct netdev_dpdk *netdev, unsigned int n_txqs)
508 netdev->tx_q = dpdk_rte_mzalloc(n_txqs * sizeof *netdev->tx_q);
509 /* Each index is considered as a cpu core id, since there should
510 * be one tx queue for each cpu core. */
511 for (i = 0; i < n_txqs; i++) {
512 int numa_id = ovs_numa_get_numa_id(i);
514 /* If the corresponding core is not on the same numa node
515 * as 'netdev', flags the 'flush_tx'. */
516 netdev->tx_q[i].flush_tx = netdev->socket_id == numa_id;
521 netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no,
522 enum dpdk_dev_type type)
523 OVS_REQUIRES(dpdk_mutex)
525 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
529 ovs_mutex_init(&netdev->mutex);
530 ovs_mutex_lock(&netdev->mutex);
532 /* If the 'sid' is negative, it means that the kernel fails
533 * to obtain the pci numa info. In that situation, always
535 if (type == DPDK_DEV_ETH) {
536 sid = rte_eth_dev_socket_id(port_no);
538 sid = rte_lcore_to_socket_id(rte_get_master_lcore());
541 netdev->socket_id = sid < 0 ? SOCKET0 : sid;
542 netdev->port_id = port_no;
545 netdev->mtu = ETHER_MTU;
546 netdev->max_packet_len = MTU_TO_MAX_LEN(netdev->mtu);
547 rte_spinlock_init(&netdev->txq_lock);
549 netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, netdev->mtu);
550 if (!netdev->dpdk_mp) {
555 netdev_->n_txq = NR_QUEUE;
556 netdev_->n_rxq = NR_QUEUE;
558 if (type == DPDK_DEV_ETH) {
559 netdev_dpdk_alloc_txq(netdev, NR_QUEUE);
560 err = dpdk_eth_dev_init(netdev);
566 list_push_back(&dpdk_list, &netdev->list_node);
570 rte_free(netdev->tx_q);
572 ovs_mutex_unlock(&netdev->mutex);
577 dpdk_dev_parse_name(const char dev_name[], const char prefix[],
578 unsigned int *port_no)
582 if (strncmp(dev_name, prefix, strlen(prefix))) {
586 cport = dev_name + strlen(prefix);
587 *port_no = strtol(cport, 0, 0); /* string must be null terminated */
592 netdev_dpdk_vhost_construct(struct netdev *netdev_)
596 if (rte_eal_init_ret) {
597 return rte_eal_init_ret;
600 ovs_mutex_lock(&dpdk_mutex);
601 err = netdev_dpdk_init(netdev_, -1, DPDK_DEV_VHOST);
602 ovs_mutex_unlock(&dpdk_mutex);
608 netdev_dpdk_construct(struct netdev *netdev)
610 unsigned int port_no;
613 if (rte_eal_init_ret) {
614 return rte_eal_init_ret;
617 /* Names always start with "dpdk" */
618 err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no);
623 ovs_mutex_lock(&dpdk_mutex);
624 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
625 ovs_mutex_unlock(&dpdk_mutex);
630 netdev_dpdk_destruct(struct netdev *netdev_)
632 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
634 ovs_mutex_lock(&dev->mutex);
635 rte_eth_dev_stop(dev->port_id);
636 ovs_mutex_unlock(&dev->mutex);
638 ovs_mutex_lock(&dpdk_mutex);
640 list_remove(&dev->list_node);
641 dpdk_mp_put(dev->dpdk_mp);
642 ovs_mutex_unlock(&dpdk_mutex);
646 netdev_dpdk_vhost_destruct(struct netdev *netdev_)
648 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
650 /* Can't remove a port while a guest is attached to it. */
651 if (netdev_dpdk_get_virtio(dev) != NULL) {
652 VLOG_ERR("Can not remove port, vhost device still attached");
656 ovs_mutex_lock(&dpdk_mutex);
657 list_remove(&dev->list_node);
658 dpdk_mp_put(dev->dpdk_mp);
659 ovs_mutex_unlock(&dpdk_mutex);
663 netdev_dpdk_dealloc(struct netdev *netdev_)
665 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
671 netdev_dpdk_get_config(const struct netdev *netdev_, struct smap *args)
673 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
675 ovs_mutex_lock(&dev->mutex);
677 smap_add_format(args, "configured_rx_queues", "%d", netdev_->n_rxq);
678 smap_add_format(args, "configured_tx_queues", "%d", netdev_->n_txq);
679 ovs_mutex_unlock(&dev->mutex);
685 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
687 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
689 return netdev->socket_id;
692 /* Sets the number of tx queues and rx queues for the dpdk interface.
693 * If the configuration fails, do not try restoring its old configuration
694 * and just returns the error. */
696 netdev_dpdk_set_multiq(struct netdev *netdev_, unsigned int n_txq,
699 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
702 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
706 ovs_mutex_lock(&dpdk_mutex);
707 ovs_mutex_lock(&netdev->mutex);
709 rte_eth_dev_stop(netdev->port_id);
711 netdev->up.n_txq = n_txq;
712 netdev->up.n_rxq = n_rxq;
714 rte_free(netdev->tx_q);
715 netdev_dpdk_alloc_txq(netdev, n_txq);
716 err = dpdk_eth_dev_init(netdev);
718 ovs_mutex_unlock(&netdev->mutex);
719 ovs_mutex_unlock(&dpdk_mutex);
725 netdev_dpdk_vhost_set_multiq(struct netdev *netdev_, unsigned int n_txq,
728 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
731 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
735 ovs_mutex_lock(&dpdk_mutex);
736 ovs_mutex_lock(&netdev->mutex);
738 netdev->up.n_txq = n_txq;
739 netdev->up.n_rxq = n_rxq;
741 ovs_mutex_unlock(&netdev->mutex);
742 ovs_mutex_unlock(&dpdk_mutex);
747 static struct netdev_rxq *
748 netdev_dpdk_rxq_alloc(void)
750 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
755 static struct netdev_rxq_dpdk *
756 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
758 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
762 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
764 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
765 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
767 ovs_mutex_lock(&netdev->mutex);
768 rx->port_id = netdev->port_id;
769 ovs_mutex_unlock(&netdev->mutex);
775 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
780 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
782 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
788 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
790 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
793 while (nb_tx != txq->count) {
796 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
805 if (OVS_UNLIKELY(nb_tx != txq->count)) {
806 /* free buffers, which we couldn't transmit, one at a time (each
807 * packet could come from a different mempool) */
810 for (i = nb_tx; i < txq->count; i++) {
811 rte_pktmbuf_free_seg(txq->burst_pkts[i]);
813 ovs_mutex_lock(&dev->mutex);
814 dev->stats.tx_dropped += txq->count-nb_tx;
815 ovs_mutex_unlock(&dev->mutex);
819 txq->tsc = rte_get_timer_cycles();
823 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
825 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
827 if (txq->count == 0) {
830 dpdk_queue_flush__(dev, qid);
834 is_vhost_running(struct virtio_net *dev)
836 return (dev != NULL && (dev->flags & VIRTIO_DEV_RUNNING));
840 * The receive path for the vhost port is the TX path out from guest.
843 netdev_dpdk_vhost_rxq_recv(struct netdev_rxq *rxq_,
844 struct dp_packet **packets, int *c)
846 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
847 struct netdev *netdev = rx->up.netdev;
848 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
849 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
853 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
857 nb_rx = rte_vhost_dequeue_burst(virtio_dev, qid,
858 vhost_dev->dpdk_mp->mp,
859 (struct rte_mbuf **)packets,
865 vhost_dev->stats.rx_packets += (uint64_t)nb_rx;
871 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dp_packet **packets,
874 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
875 struct netdev *netdev = rx->up.netdev;
876 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
879 /* There is only one tx queue for this core. Do not flush other
881 if (rxq_->queue_id == rte_lcore_id()) {
882 dpdk_queue_flush(dev, rxq_->queue_id);
885 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
886 (struct rte_mbuf **) packets,
887 MIN((int) NETDEV_MAX_RX_BATCH,
888 (int) MAX_PKT_BURST));
899 __netdev_dpdk_vhost_send(struct netdev *netdev, struct dp_packet **pkts,
900 int cnt, bool may_steal)
902 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
903 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
906 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
907 ovs_mutex_lock(&vhost_dev->mutex);
908 vhost_dev->stats.tx_dropped+= cnt;
909 ovs_mutex_unlock(&vhost_dev->mutex);
913 /* There is vHost TX single queue, So we need to lock it for TX. */
914 rte_spinlock_lock(&vhost_dev->txq_lock);
915 tx_pkts = rte_vhost_enqueue_burst(virtio_dev, VIRTIO_RXQ,
916 (struct rte_mbuf **)pkts, cnt);
918 vhost_dev->stats.tx_packets += tx_pkts;
919 vhost_dev->stats.tx_dropped += (cnt - tx_pkts);
920 rte_spinlock_unlock(&vhost_dev->txq_lock);
924 for (i = 0; i < cnt; i++) {
925 dp_packet_delete(pkts[i]);
931 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
932 struct rte_mbuf **pkts, int cnt)
934 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
940 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
941 int tocopy = MIN(freeslots, cnt-i);
943 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
944 tocopy * sizeof (struct rte_mbuf *));
946 txq->count += tocopy;
949 if (txq->count == MAX_TX_QUEUE_LEN || txq->flush_tx) {
950 dpdk_queue_flush__(dev, qid);
952 diff_tsc = rte_get_timer_cycles() - txq->tsc;
953 if (diff_tsc >= DRAIN_TSC) {
954 dpdk_queue_flush__(dev, qid);
959 /* Tx function. Transmit packets indefinitely */
961 dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet **pkts,
963 OVS_NO_THREAD_SAFETY_ANALYSIS
965 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
966 struct rte_mbuf *mbufs[cnt];
971 /* If we are on a non pmd thread we have to use the mempool mutex, because
972 * every non pmd thread shares the same mempool cache */
974 if (!thread_is_pmd()) {
975 ovs_mutex_lock(&nonpmd_mempool_mutex);
978 for (i = 0; i < cnt; i++) {
979 int size = dp_packet_size(pkts[i]);
981 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
982 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
983 (int)size , dev->max_packet_len);
989 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
991 if (!mbufs[newcnt]) {
996 /* We have to do a copy for now */
997 memcpy(rte_pktmbuf_mtod(mbufs[newcnt], void *), dp_packet_data(pkts[i]), size);
999 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
1000 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
1005 if (OVS_UNLIKELY(dropped)) {
1006 ovs_mutex_lock(&dev->mutex);
1007 dev->stats.tx_dropped += dropped;
1008 ovs_mutex_unlock(&dev->mutex);
1011 if (dev->type == DPDK_DEV_VHOST) {
1012 __netdev_dpdk_vhost_send(netdev, (struct dp_packet **) mbufs, newcnt, true);
1014 dpdk_queue_pkts(dev, qid, mbufs, newcnt);
1015 dpdk_queue_flush(dev, qid);
1018 if (!thread_is_pmd()) {
1019 ovs_mutex_unlock(&nonpmd_mempool_mutex);
1024 netdev_dpdk_vhost_send(struct netdev *netdev, int qid OVS_UNUSED, struct dp_packet **pkts,
1025 int cnt, bool may_steal)
1027 if (OVS_UNLIKELY(pkts[0]->source != DPBUF_DPDK)) {
1030 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1032 for (i = 0; i < cnt; i++) {
1033 dp_packet_delete(pkts[i]);
1037 __netdev_dpdk_vhost_send(netdev, pkts, cnt, may_steal);
1043 netdev_dpdk_send__(struct netdev_dpdk *dev, int qid,
1044 struct dp_packet **pkts, int cnt, bool may_steal)
1048 if (OVS_UNLIKELY(!may_steal ||
1049 pkts[0]->source != DPBUF_DPDK)) {
1050 struct netdev *netdev = &dev->up;
1052 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1055 for (i = 0; i < cnt; i++) {
1056 dp_packet_delete(pkts[i]);
1060 int next_tx_idx = 0;
1063 for (i = 0; i < cnt; i++) {
1064 int size = dp_packet_size(pkts[i]);
1066 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1067 if (next_tx_idx != i) {
1068 dpdk_queue_pkts(dev, qid,
1069 (struct rte_mbuf **)&pkts[next_tx_idx],
1073 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1074 (int)size , dev->max_packet_len);
1076 dp_packet_delete(pkts[i]);
1078 next_tx_idx = i + 1;
1081 if (next_tx_idx != cnt) {
1082 dpdk_queue_pkts(dev, qid,
1083 (struct rte_mbuf **)&pkts[next_tx_idx],
1087 if (OVS_UNLIKELY(dropped)) {
1088 ovs_mutex_lock(&dev->mutex);
1089 dev->stats.tx_dropped += dropped;
1090 ovs_mutex_unlock(&dev->mutex);
1096 netdev_dpdk_eth_send(struct netdev *netdev, int qid,
1097 struct dp_packet **pkts, int cnt, bool may_steal)
1099 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1101 netdev_dpdk_send__(dev, qid, pkts, cnt, may_steal);
1106 netdev_dpdk_set_etheraddr(struct netdev *netdev,
1107 const uint8_t mac[ETH_ADDR_LEN])
1109 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1111 ovs_mutex_lock(&dev->mutex);
1112 if (!eth_addr_equals(dev->hwaddr, mac)) {
1113 memcpy(dev->hwaddr, mac, ETH_ADDR_LEN);
1114 netdev_change_seq_changed(netdev);
1116 ovs_mutex_unlock(&dev->mutex);
1122 netdev_dpdk_get_etheraddr(const struct netdev *netdev,
1123 uint8_t mac[ETH_ADDR_LEN])
1125 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1127 ovs_mutex_lock(&dev->mutex);
1128 memcpy(mac, dev->hwaddr, ETH_ADDR_LEN);
1129 ovs_mutex_unlock(&dev->mutex);
1135 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
1137 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1139 ovs_mutex_lock(&dev->mutex);
1141 ovs_mutex_unlock(&dev->mutex);
1147 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
1149 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1151 struct dpdk_mp *old_mp;
1154 ovs_mutex_lock(&dpdk_mutex);
1155 ovs_mutex_lock(&dev->mutex);
1156 if (dev->mtu == mtu) {
1161 mp = dpdk_mp_get(dev->socket_id, dev->mtu);
1167 rte_eth_dev_stop(dev->port_id);
1170 old_mp = dev->dpdk_mp;
1173 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1175 err = dpdk_eth_dev_init(dev);
1179 dev->dpdk_mp = old_mp;
1180 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1181 dpdk_eth_dev_init(dev);
1185 dpdk_mp_put(old_mp);
1186 netdev_change_seq_changed(netdev);
1188 ovs_mutex_unlock(&dev->mutex);
1189 ovs_mutex_unlock(&dpdk_mutex);
1194 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
1197 netdev_dpdk_vhost_get_stats(const struct netdev *netdev,
1198 struct netdev_stats *stats)
1200 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1202 ovs_mutex_lock(&dev->mutex);
1203 memset(stats, 0, sizeof(*stats));
1204 /* Unsupported Stats */
1205 stats->rx_errors = UINT64_MAX;
1206 stats->tx_errors = UINT64_MAX;
1207 stats->multicast = UINT64_MAX;
1208 stats->collisions = UINT64_MAX;
1209 stats->rx_crc_errors = UINT64_MAX;
1210 stats->rx_fifo_errors = UINT64_MAX;
1211 stats->rx_frame_errors = UINT64_MAX;
1212 stats->rx_length_errors = UINT64_MAX;
1213 stats->rx_missed_errors = UINT64_MAX;
1214 stats->rx_over_errors = UINT64_MAX;
1215 stats->tx_aborted_errors = UINT64_MAX;
1216 stats->tx_carrier_errors = UINT64_MAX;
1217 stats->tx_errors = UINT64_MAX;
1218 stats->tx_fifo_errors = UINT64_MAX;
1219 stats->tx_heartbeat_errors = UINT64_MAX;
1220 stats->tx_window_errors = UINT64_MAX;
1221 stats->rx_bytes += UINT64_MAX;
1222 stats->rx_dropped += UINT64_MAX;
1223 stats->tx_bytes += UINT64_MAX;
1225 /* Supported Stats */
1226 stats->rx_packets += dev->stats.rx_packets;
1227 stats->tx_packets += dev->stats.tx_packets;
1228 stats->tx_dropped += dev->stats.tx_dropped;
1229 ovs_mutex_unlock(&dev->mutex);
1235 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
1237 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1238 struct rte_eth_stats rte_stats;
1241 netdev_dpdk_get_carrier(netdev, &gg);
1242 ovs_mutex_lock(&dev->mutex);
1243 rte_eth_stats_get(dev->port_id, &rte_stats);
1245 memset(stats, 0, sizeof(*stats));
1247 stats->rx_packets = rte_stats.ipackets;
1248 stats->tx_packets = rte_stats.opackets;
1249 stats->rx_bytes = rte_stats.ibytes;
1250 stats->tx_bytes = rte_stats.obytes;
1251 stats->rx_errors = rte_stats.ierrors;
1252 stats->tx_errors = rte_stats.oerrors;
1253 stats->multicast = rte_stats.imcasts;
1255 stats->tx_dropped = dev->stats.tx_dropped;
1256 ovs_mutex_unlock(&dev->mutex);
1262 netdev_dpdk_get_features(const struct netdev *netdev_,
1263 enum netdev_features *current,
1264 enum netdev_features *advertised OVS_UNUSED,
1265 enum netdev_features *supported OVS_UNUSED,
1266 enum netdev_features *peer OVS_UNUSED)
1268 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1269 struct rte_eth_link link;
1271 ovs_mutex_lock(&dev->mutex);
1273 ovs_mutex_unlock(&dev->mutex);
1275 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
1276 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
1277 *current = NETDEV_F_AUTONEG;
1279 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
1280 if (link.link_speed == ETH_LINK_SPEED_10) {
1281 *current = NETDEV_F_10MB_HD;
1283 if (link.link_speed == ETH_LINK_SPEED_100) {
1284 *current = NETDEV_F_100MB_HD;
1286 if (link.link_speed == ETH_LINK_SPEED_1000) {
1287 *current = NETDEV_F_1GB_HD;
1289 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
1290 if (link.link_speed == ETH_LINK_SPEED_10) {
1291 *current = NETDEV_F_10MB_FD;
1293 if (link.link_speed == ETH_LINK_SPEED_100) {
1294 *current = NETDEV_F_100MB_FD;
1296 if (link.link_speed == ETH_LINK_SPEED_1000) {
1297 *current = NETDEV_F_1GB_FD;
1299 if (link.link_speed == ETH_LINK_SPEED_10000) {
1300 *current = NETDEV_F_10GB_FD;
1308 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1310 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1313 ovs_mutex_lock(&dev->mutex);
1314 ifindex = dev->port_id;
1315 ovs_mutex_unlock(&dev->mutex);
1321 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1323 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1325 ovs_mutex_lock(&dev->mutex);
1326 check_link_status(dev);
1327 *carrier = dev->link.link_status;
1329 ovs_mutex_unlock(&dev->mutex);
1335 netdev_dpdk_vhost_get_carrier(const struct netdev *netdev_, bool *carrier)
1337 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1338 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(dev);
1340 ovs_mutex_lock(&dev->mutex);
1342 if (is_vhost_running(virtio_dev)) {
1348 ovs_mutex_unlock(&dev->mutex);
1353 static long long int
1354 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1356 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1357 long long int carrier_resets;
1359 ovs_mutex_lock(&dev->mutex);
1360 carrier_resets = dev->link_reset_cnt;
1361 ovs_mutex_unlock(&dev->mutex);
1363 return carrier_resets;
1367 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1368 long long int interval OVS_UNUSED)
1374 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1375 enum netdev_flags off, enum netdev_flags on,
1376 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1380 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1384 *old_flagsp = dev->flags;
1388 if (dev->flags == *old_flagsp) {
1392 if (dev->type == DPDK_DEV_ETH) {
1393 if (dev->flags & NETDEV_UP) {
1394 err = rte_eth_dev_start(dev->port_id);
1399 if (dev->flags & NETDEV_PROMISC) {
1400 rte_eth_promiscuous_enable(dev->port_id);
1403 if (!(dev->flags & NETDEV_UP)) {
1404 rte_eth_dev_stop(dev->port_id);
1412 netdev_dpdk_update_flags(struct netdev *netdev_,
1413 enum netdev_flags off, enum netdev_flags on,
1414 enum netdev_flags *old_flagsp)
1416 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1419 ovs_mutex_lock(&netdev->mutex);
1420 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1421 ovs_mutex_unlock(&netdev->mutex);
1427 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1429 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1430 struct rte_eth_dev_info dev_info;
1432 if (dev->port_id < 0)
1435 ovs_mutex_lock(&dev->mutex);
1436 rte_eth_dev_info_get(dev->port_id, &dev_info);
1437 ovs_mutex_unlock(&dev->mutex);
1439 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1441 smap_add_format(args, "port_no", "%d", dev->port_id);
1442 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1443 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1444 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1445 smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen);
1446 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1447 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1448 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1449 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1450 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1451 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1453 smap_add_format(args, "pci-vendor_id", "0x%u", dev_info.pci_dev->id.vendor_id);
1454 smap_add_format(args, "pci-device_id", "0x%x", dev_info.pci_dev->id.device_id);
1460 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1461 OVS_REQUIRES(dev->mutex)
1463 enum netdev_flags old_flags;
1466 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1468 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1473 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1474 const char *argv[], void *aux OVS_UNUSED)
1478 if (!strcasecmp(argv[argc - 1], "up")) {
1480 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1483 unixctl_command_reply_error(conn, "Invalid Admin State");
1488 struct netdev *netdev = netdev_from_name(argv[1]);
1489 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1490 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1492 ovs_mutex_lock(&dpdk_dev->mutex);
1493 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1494 ovs_mutex_unlock(&dpdk_dev->mutex);
1496 netdev_close(netdev);
1498 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1499 netdev_close(netdev);
1503 struct netdev_dpdk *netdev;
1505 ovs_mutex_lock(&dpdk_mutex);
1506 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1507 ovs_mutex_lock(&netdev->mutex);
1508 netdev_dpdk_set_admin_state__(netdev, up);
1509 ovs_mutex_unlock(&netdev->mutex);
1511 ovs_mutex_unlock(&dpdk_mutex);
1513 unixctl_command_reply(conn, "OK");
1517 * Set virtqueue flags so that we do not receive interrupts.
1520 set_irq_status(struct virtio_net *dev)
1522 dev->virtqueue[VIRTIO_RXQ]->used->flags = VRING_USED_F_NO_NOTIFY;
1523 dev->virtqueue[VIRTIO_TXQ]->used->flags = VRING_USED_F_NO_NOTIFY;
1527 * A new virtio-net device is added to a vhost port.
1530 new_device(struct virtio_net *dev)
1532 struct netdev_dpdk *netdev;
1533 bool exists = false;
1535 ovs_mutex_lock(&dpdk_mutex);
1536 /* Add device to the vhost port with the same name as that passed down. */
1537 LIST_FOR_EACH(netdev, list_node, &dpdk_list) {
1538 if (strncmp(dev->ifname, netdev->up.name, IFNAMSIZ) == 0) {
1539 ovs_mutex_lock(&netdev->mutex);
1540 ovsrcu_set(&netdev->virtio_dev, dev);
1541 ovs_mutex_unlock(&netdev->mutex);
1543 dev->flags |= VIRTIO_DEV_RUNNING;
1544 /* Disable notifications. */
1545 set_irq_status(dev);
1549 ovs_mutex_unlock(&dpdk_mutex);
1552 VLOG_INFO("vHost Device '%s' (%ld) can't be added - name not found",
1553 dev->ifname, dev->device_fh);
1558 VLOG_INFO("vHost Device '%s' (%ld) has been added",
1559 dev->ifname, dev->device_fh);
1564 * Remove a virtio-net device from the specific vhost port. Use dev->remove
1565 * flag to stop any more packets from being sent or received to/from a VM and
1566 * ensure all currently queued packets have been sent/received before removing
1570 destroy_device(volatile struct virtio_net *dev)
1572 struct netdev_dpdk *vhost_dev;
1574 ovs_mutex_lock(&dpdk_mutex);
1575 LIST_FOR_EACH (vhost_dev, list_node, &dpdk_list) {
1576 if (netdev_dpdk_get_virtio(vhost_dev) == dev) {
1578 ovs_mutex_lock(&vhost_dev->mutex);
1579 dev->flags &= ~VIRTIO_DEV_RUNNING;
1580 ovsrcu_set(&vhost_dev->virtio_dev, NULL);
1581 ovs_mutex_unlock(&vhost_dev->mutex);
1584 * Wait for other threads to quiesce before
1585 * setting the virtio_dev to NULL.
1587 ovsrcu_synchronize();
1589 * As call to ovsrcu_synchronize() will end the quiescent state,
1590 * put thread back into quiescent state before returning.
1592 ovsrcu_quiesce_start();
1595 ovs_mutex_unlock(&dpdk_mutex);
1597 VLOG_INFO("vHost Device '%s' (%ld) has been removed",
1598 dev->ifname, dev->device_fh);
1602 netdev_dpdk_get_virtio(const struct netdev_dpdk *dev)
1604 return ovsrcu_get(struct virtio_net *, &dev->virtio_dev);
1608 * These callbacks allow virtio-net devices to be added to vhost ports when
1609 * configuration has been fully complete.
1611 const struct virtio_net_device_ops virtio_net_device_ops =
1613 .new_device = new_device,
1614 .destroy_device = destroy_device,
1618 start_cuse_session_loop(void *dummy OVS_UNUSED)
1620 pthread_detach(pthread_self());
1621 /* Put the cuse thread into quiescent state. */
1622 ovsrcu_quiesce_start();
1623 rte_vhost_driver_session_start();
1628 dpdk_vhost_class_init(void)
1632 rte_vhost_driver_callback_register(&virtio_net_device_ops);
1634 /* Register CUSE device to handle IOCTLs.
1635 * Unless otherwise specified on the vswitchd command line, cuse_dev_name
1636 * is set to vhost-net.
1638 err = rte_vhost_driver_register(cuse_dev_name);
1641 VLOG_ERR("CUSE device setup failure.");
1645 ovs_thread_create("cuse_thread", start_cuse_session_loop, NULL);
1650 dpdk_common_init(void)
1652 unixctl_command_register("netdev-dpdk/set-admin-state",
1653 "[netdev] up|down", 1, 2,
1654 netdev_dpdk_set_admin_state, NULL);
1656 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
1662 dpdk_ring_create(const char dev_name[], unsigned int port_no,
1663 unsigned int *eth_port_id)
1665 struct dpdk_ring *ivshmem;
1669 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
1670 if (ivshmem == NULL) {
1674 /* XXX: Add support for multiquque ring. */
1675 err = snprintf(ring_name, 10, "%s_tx", dev_name);
1680 /* Create single consumer/producer rings, netdev does explicit locking. */
1681 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1682 RING_F_SP_ENQ | RING_F_SC_DEQ);
1683 if (ivshmem->cring_tx == NULL) {
1688 err = snprintf(ring_name, 10, "%s_rx", dev_name);
1693 /* Create single consumer/producer rings, netdev does explicit locking. */
1694 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1695 RING_F_SP_ENQ | RING_F_SC_DEQ);
1696 if (ivshmem->cring_rx == NULL) {
1701 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
1702 &ivshmem->cring_tx, 1, SOCKET0);
1709 ivshmem->user_port_id = port_no;
1710 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
1711 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
1713 *eth_port_id = ivshmem->eth_port_id;
1718 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
1720 struct dpdk_ring *ivshmem;
1721 unsigned int port_no;
1724 /* Names always start with "dpdkr" */
1725 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
1730 /* look through our list to find the device */
1731 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
1732 if (ivshmem->user_port_id == port_no) {
1733 VLOG_INFO("Found dpdk ring device %s:", dev_name);
1734 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
1738 /* Need to create the device rings */
1739 return dpdk_ring_create(dev_name, port_no, eth_port_id);
1743 netdev_dpdk_ring_send(struct netdev *netdev, int qid OVS_UNUSED,
1744 struct dp_packet **pkts, int cnt, bool may_steal)
1746 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1749 /* When using 'dpdkr' and sending to a DPDK ring, we want to ensure that the
1750 * rss hash field is clear. This is because the same mbuf may be modified by
1751 * the consumer of the ring and return into the datapath without recalculating
1753 for (i = 0; i < cnt; i++) {
1754 dp_packet_set_rss_hash(pkts[i], 0);
1757 /* DPDK Rings have a single TX queue, Therefore needs locking. */
1758 rte_spinlock_lock(&dev->txq_lock);
1759 netdev_dpdk_send__(dev, 0, pkts, cnt, may_steal);
1760 rte_spinlock_unlock(&dev->txq_lock);
1765 netdev_dpdk_ring_construct(struct netdev *netdev)
1767 unsigned int port_no = 0;
1770 if (rte_eal_init_ret) {
1771 return rte_eal_init_ret;
1774 ovs_mutex_lock(&dpdk_mutex);
1776 err = dpdk_ring_open(netdev->name, &port_no);
1781 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
1784 ovs_mutex_unlock(&dpdk_mutex);
1788 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, DESTRUCT, MULTIQ, SEND, \
1789 GET_CARRIER, GET_STATS, GET_FEATURES, GET_STATUS, RXQ_RECV) \
1793 NULL, /* netdev_dpdk_run */ \
1794 NULL, /* netdev_dpdk_wait */ \
1796 netdev_dpdk_alloc, \
1799 netdev_dpdk_dealloc, \
1800 netdev_dpdk_get_config, \
1801 NULL, /* netdev_dpdk_set_config */ \
1802 NULL, /* get_tunnel_config */ \
1803 NULL, /* build header */ \
1804 NULL, /* push header */ \
1805 NULL, /* pop header */ \
1806 netdev_dpdk_get_numa_id, /* get_numa_id */ \
1807 MULTIQ, /* set_multiq */ \
1810 NULL, /* send_wait */ \
1812 netdev_dpdk_set_etheraddr, \
1813 netdev_dpdk_get_etheraddr, \
1814 netdev_dpdk_get_mtu, \
1815 netdev_dpdk_set_mtu, \
1816 netdev_dpdk_get_ifindex, \
1818 netdev_dpdk_get_carrier_resets, \
1819 netdev_dpdk_set_miimon, \
1822 NULL, /* set_advertisements */ \
1824 NULL, /* set_policing */ \
1825 NULL, /* get_qos_types */ \
1826 NULL, /* get_qos_capabilities */ \
1827 NULL, /* get_qos */ \
1828 NULL, /* set_qos */ \
1829 NULL, /* get_queue */ \
1830 NULL, /* set_queue */ \
1831 NULL, /* delete_queue */ \
1832 NULL, /* get_queue_stats */ \
1833 NULL, /* queue_dump_start */ \
1834 NULL, /* queue_dump_next */ \
1835 NULL, /* queue_dump_done */ \
1836 NULL, /* dump_queue_stats */ \
1838 NULL, /* get_in4 */ \
1839 NULL, /* set_in4 */ \
1840 NULL, /* get_in6 */ \
1841 NULL, /* add_router */ \
1842 NULL, /* get_next_hop */ \
1844 NULL, /* arp_lookup */ \
1846 netdev_dpdk_update_flags, \
1848 netdev_dpdk_rxq_alloc, \
1849 netdev_dpdk_rxq_construct, \
1850 netdev_dpdk_rxq_destruct, \
1851 netdev_dpdk_rxq_dealloc, \
1853 NULL, /* rx_wait */ \
1854 NULL, /* rxq_drain */ \
1858 dpdk_init(int argc, char **argv)
1862 char *pragram_name = argv[0];
1864 if (argc < 2 || strcmp(argv[1], "--dpdk"))
1867 /* Remove the --dpdk argument from arg list.*/
1871 /* If the cuse_dev_name parameter has been provided, set 'cuse_dev_name' to
1872 * this string if it meets the correct criteria. Otherwise, set it to the
1873 * default (vhost-net).
1875 if (!strcmp(argv[1], "--cuse_dev_name") &&
1876 (strlen(argv[2]) <= NAME_MAX)) {
1878 cuse_dev_name = strdup(argv[2]);
1880 /* Remove the cuse_dev_name configuration parameters from the argument
1881 * list, so that the correct elements are passed to the DPDK
1882 * initialization function
1885 argv += 2; /* Increment by two to bypass the cuse_dev_name arguments */
1888 VLOG_ERR("User-provided cuse_dev_name in use: /dev/%s", cuse_dev_name);
1890 cuse_dev_name = "vhost-net";
1891 VLOG_INFO("No cuse_dev_name provided - defaulting to /dev/vhost-net");
1894 /* Keep the program name argument as this is needed for call to
1897 argv[0] = pragram_name;
1899 /* Make sure things are initialized ... */
1900 result = rte_eal_init(argc, argv);
1902 ovs_abort(result, "Cannot init EAL");
1905 rte_memzone_dump(stdout);
1906 rte_eal_init_ret = 0;
1908 if (argc > result) {
1909 argv[result] = argv[0];
1912 /* We are called from the main thread here */
1913 thread_set_nonpmd();
1915 return result + 1 + base;
1918 const struct netdev_class dpdk_class =
1922 netdev_dpdk_construct,
1923 netdev_dpdk_destruct,
1924 netdev_dpdk_set_multiq,
1925 netdev_dpdk_eth_send,
1926 netdev_dpdk_get_carrier,
1927 netdev_dpdk_get_stats,
1928 netdev_dpdk_get_features,
1929 netdev_dpdk_get_status,
1930 netdev_dpdk_rxq_recv);
1932 const struct netdev_class dpdk_ring_class =
1936 netdev_dpdk_ring_construct,
1937 netdev_dpdk_destruct,
1939 netdev_dpdk_ring_send,
1940 netdev_dpdk_get_carrier,
1941 netdev_dpdk_get_stats,
1942 netdev_dpdk_get_features,
1943 netdev_dpdk_get_status,
1944 netdev_dpdk_rxq_recv);
1946 const struct netdev_class dpdk_vhost_class =
1949 dpdk_vhost_class_init,
1950 netdev_dpdk_vhost_construct,
1951 netdev_dpdk_vhost_destruct,
1952 netdev_dpdk_vhost_set_multiq,
1953 netdev_dpdk_vhost_send,
1954 netdev_dpdk_vhost_get_carrier,
1955 netdev_dpdk_vhost_get_stats,
1958 netdev_dpdk_vhost_rxq_recv);
1961 netdev_dpdk_register(void)
1963 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
1965 if (rte_eal_init_ret) {
1969 if (ovsthread_once_start(&once)) {
1971 netdev_register_provider(&dpdk_class);
1972 netdev_register_provider(&dpdk_ring_class);
1973 netdev_register_provider(&dpdk_vhost_class);
1974 ovsthread_once_done(&once);
1979 pmd_thread_setaffinity_cpu(int cpu)
1985 CPU_SET(cpu, &cpuset);
1986 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
1988 VLOG_ERR("Thread affinity error %d",err);
1991 /* NON_PMD_CORE_ID is reserved for use by non pmd threads. */
1992 ovs_assert(cpu != NON_PMD_CORE_ID);
1993 RTE_PER_LCORE(_lcore_id) = cpu;
1999 thread_set_nonpmd(void)
2001 /* We have to use NON_PMD_CORE_ID to allow non-pmd threads to perform
2002 * certain DPDK operations, like rte_eth_dev_configure(). */
2003 RTE_PER_LCORE(_lcore_id) = NON_PMD_CORE_ID;
2009 return rte_lcore_id() != NON_PMD_CORE_ID;