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_IPV4_TCP | ETH_RSS_IPV4 | ETH_RSS_IPV6
121 | ETH_RSS_IPV4_UDP | ETH_RSS_IPV6_TCP | ETH_RSS_IPV6_UDP,
125 .mq_mode = ETH_MQ_TX_NONE,
129 static const struct rte_eth_rxconf rx_conf = {
131 .pthresh = RX_PTHRESH,
132 .hthresh = RX_HTHRESH,
133 .wthresh = RX_WTHRESH,
137 static const struct rte_eth_txconf tx_conf = {
139 .pthresh = TX_PTHRESH,
140 .hthresh = TX_HTHRESH,
141 .wthresh = TX_WTHRESH,
145 .txq_flags = ETH_TXQ_FLAGS_NOMULTSEGS|ETH_TXQ_FLAGS_NOOFFLOADS,
148 enum { MAX_RX_QUEUE_LEN = 192 };
149 enum { MAX_TX_QUEUE_LEN = 384 };
150 enum { DPDK_RING_SIZE = 256 };
151 BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE));
152 enum { DRAIN_TSC = 200000ULL };
159 static int rte_eal_init_ret = ENODEV;
161 static struct ovs_mutex dpdk_mutex = OVS_MUTEX_INITIALIZER;
163 /* Contains all 'struct dpdk_dev's. */
164 static struct ovs_list dpdk_list OVS_GUARDED_BY(dpdk_mutex)
165 = OVS_LIST_INITIALIZER(&dpdk_list);
167 static struct ovs_list dpdk_mp_list OVS_GUARDED_BY(dpdk_mutex)
168 = OVS_LIST_INITIALIZER(&dpdk_mp_list);
170 /* This mutex must be used by non pmd threads when allocating or freeing
171 * mbufs through mempools. Since dpdk_queue_pkts() and dpdk_queue_flush() may
172 * use mempools, a non pmd thread should hold this mutex while calling them */
173 struct ovs_mutex nonpmd_mempool_mutex = OVS_MUTEX_INITIALIZER;
176 struct rte_mempool *mp;
180 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
183 /* There should be one 'struct dpdk_tx_queue' created for
185 struct dpdk_tx_queue {
186 bool flush_tx; /* Set to true to flush queue everytime */
187 /* pkts are queued. */
190 struct rte_mbuf *burst_pkts[MAX_TX_QUEUE_LEN];
193 /* dpdk has no way to remove dpdk ring ethernet devices
194 so we have to keep them around once they've been created
197 static struct ovs_list dpdk_ring_list OVS_GUARDED_BY(dpdk_mutex)
198 = OVS_LIST_INITIALIZER(&dpdk_ring_list);
201 /* For the client rings */
202 struct rte_ring *cring_tx;
203 struct rte_ring *cring_rx;
204 int user_port_id; /* User given port no, parsed from port name */
205 int eth_port_id; /* ethernet device port id */
206 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
213 enum dpdk_dev_type type;
215 struct dpdk_tx_queue *tx_q;
217 struct ovs_mutex mutex OVS_ACQ_AFTER(dpdk_mutex);
219 struct dpdk_mp *dpdk_mp;
223 struct netdev_stats stats;
225 uint8_t hwaddr[ETH_ADDR_LEN];
226 enum netdev_flags flags;
228 struct rte_eth_link link;
231 /* virtio-net structure for vhost device */
232 OVSRCU_TYPE(struct virtio_net *) virtio_dev;
235 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
236 rte_spinlock_t txq_lock;
239 struct netdev_rxq_dpdk {
240 struct netdev_rxq up;
244 static bool thread_is_pmd(void);
246 static int netdev_dpdk_construct(struct netdev *);
248 struct virtio_net * netdev_dpdk_get_virtio(const struct netdev_dpdk *dev);
251 is_dpdk_class(const struct netdev_class *class)
253 return class->construct == netdev_dpdk_construct;
256 /* XXX: use dpdk malloc for entire OVS. in fact huge page should be used
257 * for all other segments data, bss and text. */
260 dpdk_rte_mzalloc(size_t sz)
264 ptr = rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE);
271 /* XXX this function should be called only by pmd threads (or by non pmd
272 * threads holding the nonpmd_mempool_mutex) */
274 free_dpdk_buf(struct dp_packet *p)
276 struct rte_mbuf *pkt = (struct rte_mbuf *) p;
278 rte_pktmbuf_free_seg(pkt);
282 __rte_pktmbuf_init(struct rte_mempool *mp,
283 void *opaque_arg OVS_UNUSED,
285 unsigned i OVS_UNUSED)
287 struct rte_mbuf *m = _m;
288 uint32_t buf_len = mp->elt_size - sizeof(struct dp_packet);
290 RTE_MBUF_ASSERT(mp->elt_size >= sizeof(struct dp_packet));
292 memset(m, 0, mp->elt_size);
294 /* start of buffer is just after mbuf structure */
295 m->buf_addr = (char *)m + sizeof(struct dp_packet);
296 m->buf_physaddr = rte_mempool_virt2phy(mp, m) +
297 sizeof(struct dp_packet);
298 m->buf_len = (uint16_t)buf_len;
300 /* keep some headroom between start of buffer and data */
301 m->data_off = RTE_MIN(RTE_PKTMBUF_HEADROOM, m->buf_len);
303 /* init some constant fields */
310 ovs_rte_pktmbuf_init(struct rte_mempool *mp,
311 void *opaque_arg OVS_UNUSED,
313 unsigned i OVS_UNUSED)
315 struct rte_mbuf *m = _m;
317 __rte_pktmbuf_init(mp, opaque_arg, _m, i);
319 dp_packet_init_dpdk((struct dp_packet *) m, m->buf_len);
322 static struct dpdk_mp *
323 dpdk_mp_get(int socket_id, int mtu) OVS_REQUIRES(dpdk_mutex)
325 struct dpdk_mp *dmp = NULL;
326 char mp_name[RTE_MEMPOOL_NAMESIZE];
329 LIST_FOR_EACH (dmp, list_node, &dpdk_mp_list) {
330 if (dmp->socket_id == socket_id && dmp->mtu == mtu) {
336 dmp = dpdk_rte_mzalloc(sizeof *dmp);
337 dmp->socket_id = socket_id;
341 mp_size = MAX_NB_MBUF;
343 if (snprintf(mp_name, RTE_MEMPOOL_NAMESIZE, "ovs_mp_%d_%d_%u",
344 dmp->mtu, dmp->socket_id, mp_size) < 0) {
348 dmp->mp = rte_mempool_create(mp_name, mp_size, MBUF_SIZE(mtu),
350 sizeof(struct rte_pktmbuf_pool_private),
351 rte_pktmbuf_pool_init, NULL,
352 ovs_rte_pktmbuf_init, NULL,
354 } while (!dmp->mp && rte_errno == ENOMEM && (mp_size /= 2) >= MIN_NB_MBUF);
356 if (dmp->mp == NULL) {
359 VLOG_DBG("Allocated \"%s\" mempool with %u mbufs", mp_name, mp_size );
362 list_push_back(&dpdk_mp_list, &dmp->list_node);
367 dpdk_mp_put(struct dpdk_mp *dmp)
375 ovs_assert(dmp->refcount >= 0);
378 /* I could not find any API to destroy mp. */
379 if (dmp->refcount == 0) {
380 list_delete(dmp->list_node);
381 /* destroy mp-pool. */
387 check_link_status(struct netdev_dpdk *dev)
389 struct rte_eth_link link;
391 rte_eth_link_get_nowait(dev->port_id, &link);
393 if (dev->link.link_status != link.link_status) {
394 netdev_change_seq_changed(&dev->up);
396 dev->link_reset_cnt++;
398 if (dev->link.link_status) {
399 VLOG_DBG_RL(&rl, "Port %d Link Up - speed %u Mbps - %s",
400 dev->port_id, (unsigned)dev->link.link_speed,
401 (dev->link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
402 ("full-duplex") : ("half-duplex"));
404 VLOG_DBG_RL(&rl, "Port %d Link Down", dev->port_id);
410 dpdk_watchdog(void *dummy OVS_UNUSED)
412 struct netdev_dpdk *dev;
414 pthread_detach(pthread_self());
417 ovs_mutex_lock(&dpdk_mutex);
418 LIST_FOR_EACH (dev, list_node, &dpdk_list) {
419 ovs_mutex_lock(&dev->mutex);
420 check_link_status(dev);
421 ovs_mutex_unlock(&dev->mutex);
423 ovs_mutex_unlock(&dpdk_mutex);
424 xsleep(DPDK_PORT_WATCHDOG_INTERVAL);
431 dpdk_eth_dev_init(struct netdev_dpdk *dev) OVS_REQUIRES(dpdk_mutex)
433 struct rte_pktmbuf_pool_private *mbp_priv;
434 struct ether_addr eth_addr;
438 if (dev->port_id < 0 || dev->port_id >= rte_eth_dev_count()) {
442 diag = rte_eth_dev_configure(dev->port_id, dev->up.n_rxq, dev->up.n_txq,
445 VLOG_ERR("eth dev config error %d",diag);
449 for (i = 0; i < dev->up.n_txq; i++) {
450 diag = rte_eth_tx_queue_setup(dev->port_id, i, NIC_PORT_TX_Q_SIZE,
451 dev->socket_id, &tx_conf);
453 VLOG_ERR("eth dev tx queue setup error %d",diag);
458 for (i = 0; i < dev->up.n_rxq; i++) {
459 diag = rte_eth_rx_queue_setup(dev->port_id, i, NIC_PORT_RX_Q_SIZE,
461 &rx_conf, dev->dpdk_mp->mp);
463 VLOG_ERR("eth dev rx queue setup error %d",diag);
468 diag = rte_eth_dev_start(dev->port_id);
470 VLOG_ERR("eth dev start error %d",diag);
474 rte_eth_promiscuous_enable(dev->port_id);
475 rte_eth_allmulticast_enable(dev->port_id);
477 memset(ð_addr, 0x0, sizeof(eth_addr));
478 rte_eth_macaddr_get(dev->port_id, ð_addr);
479 VLOG_INFO_RL(&rl, "Port %d: "ETH_ADDR_FMT"",
480 dev->port_id, ETH_ADDR_ARGS(eth_addr.addr_bytes));
482 memcpy(dev->hwaddr, eth_addr.addr_bytes, ETH_ADDR_LEN);
483 rte_eth_link_get_nowait(dev->port_id, &dev->link);
485 mbp_priv = rte_mempool_get_priv(dev->dpdk_mp->mp);
486 dev->buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
488 dev->flags = NETDEV_UP | NETDEV_PROMISC;
492 static struct netdev_dpdk *
493 netdev_dpdk_cast(const struct netdev *netdev)
495 return CONTAINER_OF(netdev, struct netdev_dpdk, up);
498 static struct netdev *
499 netdev_dpdk_alloc(void)
501 struct netdev_dpdk *netdev = dpdk_rte_mzalloc(sizeof *netdev);
506 netdev_dpdk_alloc_txq(struct netdev_dpdk *netdev, unsigned int n_txqs)
510 netdev->tx_q = dpdk_rte_mzalloc(n_txqs * sizeof *netdev->tx_q);
511 /* Each index is considered as a cpu core id, since there should
512 * be one tx queue for each cpu core. */
513 for (i = 0; i < n_txqs; i++) {
514 int numa_id = ovs_numa_get_numa_id(i);
516 /* If the corresponding core is not on the same numa node
517 * as 'netdev', flags the 'flush_tx'. */
518 netdev->tx_q[i].flush_tx = netdev->socket_id == numa_id;
523 netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no,
524 enum dpdk_dev_type type)
525 OVS_REQUIRES(dpdk_mutex)
527 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
531 ovs_mutex_init(&netdev->mutex);
532 ovs_mutex_lock(&netdev->mutex);
534 /* If the 'sid' is negative, it means that the kernel fails
535 * to obtain the pci numa info. In that situation, always
537 if (type == DPDK_DEV_ETH) {
538 sid = rte_eth_dev_socket_id(port_no);
540 sid = rte_lcore_to_socket_id(rte_get_master_lcore());
543 netdev->socket_id = sid < 0 ? SOCKET0 : sid;
544 netdev->port_id = port_no;
547 netdev->mtu = ETHER_MTU;
548 netdev->max_packet_len = MTU_TO_MAX_LEN(netdev->mtu);
549 rte_spinlock_init(&netdev->txq_lock);
551 netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, netdev->mtu);
552 if (!netdev->dpdk_mp) {
557 netdev_->n_txq = NR_QUEUE;
558 netdev_->n_rxq = NR_QUEUE;
560 if (type == DPDK_DEV_ETH) {
561 netdev_dpdk_alloc_txq(netdev, NR_QUEUE);
562 err = dpdk_eth_dev_init(netdev);
568 list_push_back(&dpdk_list, &netdev->list_node);
572 rte_free(netdev->tx_q);
574 ovs_mutex_unlock(&netdev->mutex);
579 dpdk_dev_parse_name(const char dev_name[], const char prefix[],
580 unsigned int *port_no)
584 if (strncmp(dev_name, prefix, strlen(prefix))) {
588 cport = dev_name + strlen(prefix);
589 *port_no = strtol(cport, 0, 0); /* string must be null terminated */
594 netdev_dpdk_vhost_construct(struct netdev *netdev_)
598 if (rte_eal_init_ret) {
599 return rte_eal_init_ret;
602 ovs_mutex_lock(&dpdk_mutex);
603 err = netdev_dpdk_init(netdev_, -1, DPDK_DEV_VHOST);
604 ovs_mutex_unlock(&dpdk_mutex);
610 netdev_dpdk_construct(struct netdev *netdev)
612 unsigned int port_no;
615 if (rte_eal_init_ret) {
616 return rte_eal_init_ret;
619 /* Names always start with "dpdk" */
620 err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no);
625 ovs_mutex_lock(&dpdk_mutex);
626 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
627 ovs_mutex_unlock(&dpdk_mutex);
632 netdev_dpdk_destruct(struct netdev *netdev_)
634 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
636 ovs_mutex_lock(&dev->mutex);
637 rte_eth_dev_stop(dev->port_id);
638 ovs_mutex_unlock(&dev->mutex);
640 ovs_mutex_lock(&dpdk_mutex);
642 list_remove(&dev->list_node);
643 dpdk_mp_put(dev->dpdk_mp);
644 ovs_mutex_unlock(&dpdk_mutex);
648 netdev_dpdk_vhost_destruct(struct netdev *netdev_)
650 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
652 /* Can't remove a port while a guest is attached to it. */
653 if (netdev_dpdk_get_virtio(dev) != NULL) {
654 VLOG_ERR("Can not remove port, vhost device still attached");
658 ovs_mutex_lock(&dpdk_mutex);
659 list_remove(&dev->list_node);
660 dpdk_mp_put(dev->dpdk_mp);
661 ovs_mutex_unlock(&dpdk_mutex);
665 netdev_dpdk_dealloc(struct netdev *netdev_)
667 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
673 netdev_dpdk_get_config(const struct netdev *netdev_, struct smap *args)
675 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
677 ovs_mutex_lock(&dev->mutex);
679 smap_add_format(args, "configured_rx_queues", "%d", netdev_->n_rxq);
680 smap_add_format(args, "configured_tx_queues", "%d", netdev_->n_txq);
681 ovs_mutex_unlock(&dev->mutex);
687 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
689 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
691 return netdev->socket_id;
694 /* Sets the number of tx queues and rx queues for the dpdk interface.
695 * If the configuration fails, do not try restoring its old configuration
696 * and just returns the error. */
698 netdev_dpdk_set_multiq(struct netdev *netdev_, unsigned int n_txq,
701 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
704 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
708 ovs_mutex_lock(&dpdk_mutex);
709 ovs_mutex_lock(&netdev->mutex);
711 rte_eth_dev_stop(netdev->port_id);
713 netdev->up.n_txq = n_txq;
714 netdev->up.n_rxq = n_rxq;
716 rte_free(netdev->tx_q);
717 netdev_dpdk_alloc_txq(netdev, n_txq);
718 err = dpdk_eth_dev_init(netdev);
720 ovs_mutex_unlock(&netdev->mutex);
721 ovs_mutex_unlock(&dpdk_mutex);
727 netdev_dpdk_vhost_set_multiq(struct netdev *netdev_, unsigned int n_txq,
730 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
733 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
737 ovs_mutex_lock(&dpdk_mutex);
738 ovs_mutex_lock(&netdev->mutex);
740 netdev->up.n_txq = n_txq;
741 netdev->up.n_rxq = n_rxq;
743 ovs_mutex_unlock(&netdev->mutex);
744 ovs_mutex_unlock(&dpdk_mutex);
749 static struct netdev_rxq *
750 netdev_dpdk_rxq_alloc(void)
752 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
757 static struct netdev_rxq_dpdk *
758 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
760 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
764 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
766 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
767 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
769 ovs_mutex_lock(&netdev->mutex);
770 rx->port_id = netdev->port_id;
771 ovs_mutex_unlock(&netdev->mutex);
777 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
782 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
784 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
790 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
792 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
795 while (nb_tx != txq->count) {
798 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
807 if (OVS_UNLIKELY(nb_tx != txq->count)) {
808 /* free buffers, which we couldn't transmit, one at a time (each
809 * packet could come from a different mempool) */
812 for (i = nb_tx; i < txq->count; i++) {
813 rte_pktmbuf_free_seg(txq->burst_pkts[i]);
815 ovs_mutex_lock(&dev->mutex);
816 dev->stats.tx_dropped += txq->count-nb_tx;
817 ovs_mutex_unlock(&dev->mutex);
821 txq->tsc = rte_get_timer_cycles();
825 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
827 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
829 if (txq->count == 0) {
832 dpdk_queue_flush__(dev, qid);
836 is_vhost_running(struct virtio_net *dev)
838 return (dev != NULL && (dev->flags & VIRTIO_DEV_RUNNING));
842 * The receive path for the vhost port is the TX path out from guest.
845 netdev_dpdk_vhost_rxq_recv(struct netdev_rxq *rxq_,
846 struct dp_packet **packets, int *c)
848 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
849 struct netdev *netdev = rx->up.netdev;
850 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
851 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
855 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
859 nb_rx = rte_vhost_dequeue_burst(virtio_dev, qid,
860 vhost_dev->dpdk_mp->mp,
861 (struct rte_mbuf **)packets,
867 vhost_dev->stats.rx_packets += (uint64_t)nb_rx;
873 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dp_packet **packets,
876 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
877 struct netdev *netdev = rx->up.netdev;
878 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
881 /* There is only one tx queue for this core. Do not flush other
883 if (rxq_->queue_id == rte_lcore_id()) {
884 dpdk_queue_flush(dev, rxq_->queue_id);
887 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
888 (struct rte_mbuf **) packets,
889 MIN((int)NETDEV_MAX_RX_BATCH,
890 (int)MAX_RX_QUEUE_LEN));
901 __netdev_dpdk_vhost_send(struct netdev *netdev, struct dp_packet **pkts,
902 int cnt, bool may_steal)
904 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
905 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
908 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
909 ovs_mutex_lock(&vhost_dev->mutex);
910 vhost_dev->stats.tx_dropped+= cnt;
911 ovs_mutex_unlock(&vhost_dev->mutex);
915 /* There is vHost TX single queue, So we need to lock it for TX. */
916 rte_spinlock_lock(&vhost_dev->txq_lock);
917 tx_pkts = rte_vhost_enqueue_burst(virtio_dev, VIRTIO_RXQ,
918 (struct rte_mbuf **)pkts, cnt);
920 vhost_dev->stats.tx_packets += tx_pkts;
921 vhost_dev->stats.tx_dropped += (cnt - tx_pkts);
922 rte_spinlock_unlock(&vhost_dev->txq_lock);
926 for (i = 0; i < cnt; i++) {
927 dp_packet_delete(pkts[i]);
933 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
934 struct rte_mbuf **pkts, int cnt)
936 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
942 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
943 int tocopy = MIN(freeslots, cnt-i);
945 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
946 tocopy * sizeof (struct rte_mbuf *));
948 txq->count += tocopy;
951 if (txq->count == MAX_TX_QUEUE_LEN || txq->flush_tx) {
952 dpdk_queue_flush__(dev, qid);
954 diff_tsc = rte_get_timer_cycles() - txq->tsc;
955 if (diff_tsc >= DRAIN_TSC) {
956 dpdk_queue_flush__(dev, qid);
961 /* Tx function. Transmit packets indefinitely */
963 dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet **pkts,
965 OVS_NO_THREAD_SAFETY_ANALYSIS
967 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
968 struct rte_mbuf *mbufs[cnt];
973 /* If we are on a non pmd thread we have to use the mempool mutex, because
974 * every non pmd thread shares the same mempool cache */
976 if (!thread_is_pmd()) {
977 ovs_mutex_lock(&nonpmd_mempool_mutex);
980 for (i = 0; i < cnt; i++) {
981 int size = dp_packet_size(pkts[i]);
983 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
984 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
985 (int)size , dev->max_packet_len);
991 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
993 if (!mbufs[newcnt]) {
998 /* We have to do a copy for now */
999 memcpy(rte_pktmbuf_mtod(mbufs[newcnt], void *), dp_packet_data(pkts[i]), size);
1001 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
1002 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
1007 if (OVS_UNLIKELY(dropped)) {
1008 ovs_mutex_lock(&dev->mutex);
1009 dev->stats.tx_dropped += dropped;
1010 ovs_mutex_unlock(&dev->mutex);
1013 if (dev->type == DPDK_DEV_VHOST) {
1014 __netdev_dpdk_vhost_send(netdev, (struct dp_packet **) mbufs, newcnt, true);
1016 dpdk_queue_pkts(dev, qid, mbufs, newcnt);
1017 dpdk_queue_flush(dev, qid);
1020 if (!thread_is_pmd()) {
1021 ovs_mutex_unlock(&nonpmd_mempool_mutex);
1026 netdev_dpdk_vhost_send(struct netdev *netdev, int qid OVS_UNUSED, struct dp_packet **pkts,
1027 int cnt, bool may_steal)
1029 if (OVS_UNLIKELY(pkts[0]->source != DPBUF_DPDK)) {
1032 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1034 for (i = 0; i < cnt; i++) {
1035 dp_packet_delete(pkts[i]);
1039 __netdev_dpdk_vhost_send(netdev, pkts, cnt, may_steal);
1045 netdev_dpdk_send__(struct netdev_dpdk *dev, int qid,
1046 struct dp_packet **pkts, int cnt, bool may_steal)
1050 if (OVS_UNLIKELY(!may_steal ||
1051 pkts[0]->source != DPBUF_DPDK)) {
1052 struct netdev *netdev = &dev->up;
1054 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1057 for (i = 0; i < cnt; i++) {
1058 dp_packet_delete(pkts[i]);
1062 int next_tx_idx = 0;
1065 for (i = 0; i < cnt; i++) {
1066 int size = dp_packet_size(pkts[i]);
1067 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1068 if (next_tx_idx != i) {
1069 dpdk_queue_pkts(dev, qid,
1070 (struct rte_mbuf **)&pkts[next_tx_idx],
1074 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1075 (int)size , dev->max_packet_len);
1077 dp_packet_delete(pkts[i]);
1079 next_tx_idx = i + 1;
1082 if (next_tx_idx != cnt) {
1083 dpdk_queue_pkts(dev, qid,
1084 (struct rte_mbuf **)&pkts[next_tx_idx],
1088 if (OVS_UNLIKELY(dropped)) {
1089 ovs_mutex_lock(&dev->mutex);
1090 dev->stats.tx_dropped += dropped;
1091 ovs_mutex_unlock(&dev->mutex);
1097 netdev_dpdk_eth_send(struct netdev *netdev, int qid,
1098 struct dp_packet **pkts, int cnt, bool may_steal)
1100 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1102 netdev_dpdk_send__(dev, qid, pkts, cnt, may_steal);
1107 netdev_dpdk_set_etheraddr(struct netdev *netdev,
1108 const uint8_t mac[ETH_ADDR_LEN])
1110 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1112 ovs_mutex_lock(&dev->mutex);
1113 if (!eth_addr_equals(dev->hwaddr, mac)) {
1114 memcpy(dev->hwaddr, mac, ETH_ADDR_LEN);
1115 netdev_change_seq_changed(netdev);
1117 ovs_mutex_unlock(&dev->mutex);
1123 netdev_dpdk_get_etheraddr(const struct netdev *netdev,
1124 uint8_t mac[ETH_ADDR_LEN])
1126 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1128 ovs_mutex_lock(&dev->mutex);
1129 memcpy(mac, dev->hwaddr, ETH_ADDR_LEN);
1130 ovs_mutex_unlock(&dev->mutex);
1136 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
1138 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1140 ovs_mutex_lock(&dev->mutex);
1142 ovs_mutex_unlock(&dev->mutex);
1148 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
1150 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1152 struct dpdk_mp *old_mp;
1155 ovs_mutex_lock(&dpdk_mutex);
1156 ovs_mutex_lock(&dev->mutex);
1157 if (dev->mtu == mtu) {
1162 mp = dpdk_mp_get(dev->socket_id, dev->mtu);
1168 rte_eth_dev_stop(dev->port_id);
1171 old_mp = dev->dpdk_mp;
1174 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1176 err = dpdk_eth_dev_init(dev);
1180 dev->dpdk_mp = old_mp;
1181 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1182 dpdk_eth_dev_init(dev);
1186 dpdk_mp_put(old_mp);
1187 netdev_change_seq_changed(netdev);
1189 ovs_mutex_unlock(&dev->mutex);
1190 ovs_mutex_unlock(&dpdk_mutex);
1195 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
1198 netdev_dpdk_vhost_get_stats(const struct netdev *netdev,
1199 struct netdev_stats *stats)
1201 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1203 ovs_mutex_lock(&dev->mutex);
1204 memset(stats, 0, sizeof(*stats));
1205 /* Unsupported Stats */
1206 stats->rx_errors = UINT64_MAX;
1207 stats->tx_errors = UINT64_MAX;
1208 stats->multicast = UINT64_MAX;
1209 stats->collisions = UINT64_MAX;
1210 stats->rx_crc_errors = UINT64_MAX;
1211 stats->rx_fifo_errors = UINT64_MAX;
1212 stats->rx_frame_errors = UINT64_MAX;
1213 stats->rx_length_errors = UINT64_MAX;
1214 stats->rx_missed_errors = UINT64_MAX;
1215 stats->rx_over_errors = UINT64_MAX;
1216 stats->tx_aborted_errors = UINT64_MAX;
1217 stats->tx_carrier_errors = UINT64_MAX;
1218 stats->tx_errors = UINT64_MAX;
1219 stats->tx_fifo_errors = UINT64_MAX;
1220 stats->tx_heartbeat_errors = UINT64_MAX;
1221 stats->tx_window_errors = UINT64_MAX;
1222 stats->rx_bytes += UINT64_MAX;
1223 stats->rx_dropped += UINT64_MAX;
1224 stats->tx_bytes += UINT64_MAX;
1226 /* Supported Stats */
1227 stats->rx_packets += dev->stats.rx_packets;
1228 stats->tx_packets += dev->stats.tx_packets;
1229 stats->tx_dropped += dev->stats.tx_dropped;
1230 ovs_mutex_unlock(&dev->mutex);
1236 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
1238 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1239 struct rte_eth_stats rte_stats;
1242 netdev_dpdk_get_carrier(netdev, &gg);
1243 ovs_mutex_lock(&dev->mutex);
1244 rte_eth_stats_get(dev->port_id, &rte_stats);
1246 memset(stats, 0, sizeof(*stats));
1248 stats->rx_packets = rte_stats.ipackets;
1249 stats->tx_packets = rte_stats.opackets;
1250 stats->rx_bytes = rte_stats.ibytes;
1251 stats->tx_bytes = rte_stats.obytes;
1252 stats->rx_errors = rte_stats.ierrors;
1253 stats->tx_errors = rte_stats.oerrors;
1254 stats->multicast = rte_stats.imcasts;
1256 stats->tx_dropped = dev->stats.tx_dropped;
1257 ovs_mutex_unlock(&dev->mutex);
1263 netdev_dpdk_get_features(const struct netdev *netdev_,
1264 enum netdev_features *current,
1265 enum netdev_features *advertised OVS_UNUSED,
1266 enum netdev_features *supported OVS_UNUSED,
1267 enum netdev_features *peer OVS_UNUSED)
1269 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1270 struct rte_eth_link link;
1272 ovs_mutex_lock(&dev->mutex);
1274 ovs_mutex_unlock(&dev->mutex);
1276 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
1277 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
1278 *current = NETDEV_F_AUTONEG;
1280 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
1281 if (link.link_speed == ETH_LINK_SPEED_10) {
1282 *current = NETDEV_F_10MB_HD;
1284 if (link.link_speed == ETH_LINK_SPEED_100) {
1285 *current = NETDEV_F_100MB_HD;
1287 if (link.link_speed == ETH_LINK_SPEED_1000) {
1288 *current = NETDEV_F_1GB_HD;
1290 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
1291 if (link.link_speed == ETH_LINK_SPEED_10) {
1292 *current = NETDEV_F_10MB_FD;
1294 if (link.link_speed == ETH_LINK_SPEED_100) {
1295 *current = NETDEV_F_100MB_FD;
1297 if (link.link_speed == ETH_LINK_SPEED_1000) {
1298 *current = NETDEV_F_1GB_FD;
1300 if (link.link_speed == ETH_LINK_SPEED_10000) {
1301 *current = NETDEV_F_10GB_FD;
1309 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1311 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1314 ovs_mutex_lock(&dev->mutex);
1315 ifindex = dev->port_id;
1316 ovs_mutex_unlock(&dev->mutex);
1322 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1324 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1326 ovs_mutex_lock(&dev->mutex);
1327 check_link_status(dev);
1328 *carrier = dev->link.link_status;
1330 ovs_mutex_unlock(&dev->mutex);
1336 netdev_dpdk_vhost_get_carrier(const struct netdev *netdev_, bool *carrier)
1338 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1339 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(dev);
1341 ovs_mutex_lock(&dev->mutex);
1343 if (is_vhost_running(virtio_dev)) {
1349 ovs_mutex_unlock(&dev->mutex);
1354 static long long int
1355 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1357 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1358 long long int carrier_resets;
1360 ovs_mutex_lock(&dev->mutex);
1361 carrier_resets = dev->link_reset_cnt;
1362 ovs_mutex_unlock(&dev->mutex);
1364 return carrier_resets;
1368 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1369 long long int interval OVS_UNUSED)
1375 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1376 enum netdev_flags off, enum netdev_flags on,
1377 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1381 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1385 *old_flagsp = dev->flags;
1389 if (dev->flags == *old_flagsp) {
1393 if (dev->type == DPDK_DEV_ETH) {
1394 if (dev->flags & NETDEV_UP) {
1395 err = rte_eth_dev_start(dev->port_id);
1400 if (dev->flags & NETDEV_PROMISC) {
1401 rte_eth_promiscuous_enable(dev->port_id);
1404 if (!(dev->flags & NETDEV_UP)) {
1405 rte_eth_dev_stop(dev->port_id);
1413 netdev_dpdk_update_flags(struct netdev *netdev_,
1414 enum netdev_flags off, enum netdev_flags on,
1415 enum netdev_flags *old_flagsp)
1417 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1420 ovs_mutex_lock(&netdev->mutex);
1421 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1422 ovs_mutex_unlock(&netdev->mutex);
1428 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1430 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1431 struct rte_eth_dev_info dev_info;
1433 if (dev->port_id < 0)
1436 ovs_mutex_lock(&dev->mutex);
1437 rte_eth_dev_info_get(dev->port_id, &dev_info);
1438 ovs_mutex_unlock(&dev->mutex);
1440 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1442 smap_add_format(args, "port_no", "%d", dev->port_id);
1443 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1444 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1445 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1446 smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen);
1447 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1448 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1449 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1450 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1451 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1452 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1454 smap_add_format(args, "pci-vendor_id", "0x%u", dev_info.pci_dev->id.vendor_id);
1455 smap_add_format(args, "pci-device_id", "0x%x", dev_info.pci_dev->id.device_id);
1461 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1462 OVS_REQUIRES(dev->mutex)
1464 enum netdev_flags old_flags;
1467 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1469 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1474 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1475 const char *argv[], void *aux OVS_UNUSED)
1479 if (!strcasecmp(argv[argc - 1], "up")) {
1481 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1484 unixctl_command_reply_error(conn, "Invalid Admin State");
1489 struct netdev *netdev = netdev_from_name(argv[1]);
1490 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1491 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1493 ovs_mutex_lock(&dpdk_dev->mutex);
1494 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1495 ovs_mutex_unlock(&dpdk_dev->mutex);
1497 netdev_close(netdev);
1499 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1500 netdev_close(netdev);
1504 struct netdev_dpdk *netdev;
1506 ovs_mutex_lock(&dpdk_mutex);
1507 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1508 ovs_mutex_lock(&netdev->mutex);
1509 netdev_dpdk_set_admin_state__(netdev, up);
1510 ovs_mutex_unlock(&netdev->mutex);
1512 ovs_mutex_unlock(&dpdk_mutex);
1514 unixctl_command_reply(conn, "OK");
1518 * Set virtqueue flags so that we do not receive interrupts.
1521 set_irq_status(struct virtio_net *dev)
1523 dev->virtqueue[VIRTIO_RXQ]->used->flags = VRING_USED_F_NO_NOTIFY;
1524 dev->virtqueue[VIRTIO_TXQ]->used->flags = VRING_USED_F_NO_NOTIFY;
1528 * A new virtio-net device is added to a vhost port.
1531 new_device(struct virtio_net *dev)
1533 struct netdev_dpdk *netdev;
1534 bool exists = false;
1536 ovs_mutex_lock(&dpdk_mutex);
1537 /* Add device to the vhost port with the same name as that passed down. */
1538 LIST_FOR_EACH(netdev, list_node, &dpdk_list) {
1539 if (strncmp(dev->ifname, netdev->up.name, IFNAMSIZ) == 0) {
1540 ovs_mutex_lock(&netdev->mutex);
1541 ovsrcu_set(&netdev->virtio_dev, dev);
1542 ovs_mutex_unlock(&netdev->mutex);
1544 dev->flags |= VIRTIO_DEV_RUNNING;
1545 /* Disable notifications. */
1546 set_irq_status(dev);
1550 ovs_mutex_unlock(&dpdk_mutex);
1553 VLOG_INFO("vHost Device '%s' (%ld) can't be added - name not found",
1554 dev->ifname, dev->device_fh);
1559 VLOG_INFO("vHost Device '%s' (%ld) has been added",
1560 dev->ifname, dev->device_fh);
1565 * Remove a virtio-net device from the specific vhost port. Use dev->remove
1566 * flag to stop any more packets from being sent or received to/from a VM and
1567 * ensure all currently queued packets have been sent/received before removing
1571 destroy_device(volatile struct virtio_net *dev)
1573 struct netdev_dpdk *vhost_dev;
1575 ovs_mutex_lock(&dpdk_mutex);
1576 LIST_FOR_EACH (vhost_dev, list_node, &dpdk_list) {
1577 if (netdev_dpdk_get_virtio(vhost_dev) == dev) {
1579 ovs_mutex_lock(&vhost_dev->mutex);
1580 dev->flags &= ~VIRTIO_DEV_RUNNING;
1581 ovsrcu_set(&vhost_dev->virtio_dev, NULL);
1582 ovs_mutex_unlock(&vhost_dev->mutex);
1585 * Wait for other threads to quiesce before
1586 * setting the virtio_dev to NULL.
1588 ovsrcu_synchronize();
1590 * As call to ovsrcu_synchronize() will end the quiescent state,
1591 * put thread back into quiescent state before returning.
1593 ovsrcu_quiesce_start();
1596 ovs_mutex_unlock(&dpdk_mutex);
1598 VLOG_INFO("vHost Device '%s' (%ld) has been removed",
1599 dev->ifname, dev->device_fh);
1603 netdev_dpdk_get_virtio(const struct netdev_dpdk *dev)
1605 return ovsrcu_get(struct virtio_net *, &dev->virtio_dev);
1609 * These callbacks allow virtio-net devices to be added to vhost ports when
1610 * configuration has been fully complete.
1612 const struct virtio_net_device_ops virtio_net_device_ops =
1614 .new_device = new_device,
1615 .destroy_device = destroy_device,
1619 start_cuse_session_loop(void *dummy OVS_UNUSED)
1621 pthread_detach(pthread_self());
1622 /* Put the cuse thread into quiescent state. */
1623 ovsrcu_quiesce_start();
1624 rte_vhost_driver_session_start();
1629 dpdk_vhost_class_init(void)
1633 rte_vhost_driver_callback_register(&virtio_net_device_ops);
1635 /* Register CUSE device to handle IOCTLs.
1636 * Unless otherwise specified on the vswitchd command line, cuse_dev_name
1637 * is set to vhost-net.
1639 err = rte_vhost_driver_register(cuse_dev_name);
1642 VLOG_ERR("CUSE device setup failure.");
1646 ovs_thread_create("cuse_thread", start_cuse_session_loop, NULL);
1651 dpdk_common_init(void)
1653 unixctl_command_register("netdev-dpdk/set-admin-state",
1654 "[netdev] up|down", 1, 2,
1655 netdev_dpdk_set_admin_state, NULL);
1657 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
1663 dpdk_ring_create(const char dev_name[], unsigned int port_no,
1664 unsigned int *eth_port_id)
1666 struct dpdk_ring *ivshmem;
1670 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
1671 if (ivshmem == NULL) {
1675 /* XXX: Add support for multiquque ring. */
1676 err = snprintf(ring_name, 10, "%s_tx", dev_name);
1681 /* Create single consumer/producer rings, netdev does explicit locking. */
1682 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1683 RING_F_SP_ENQ | RING_F_SC_DEQ);
1684 if (ivshmem->cring_tx == NULL) {
1689 err = snprintf(ring_name, 10, "%s_rx", dev_name);
1694 /* Create single consumer/producer rings, netdev does explicit locking. */
1695 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1696 RING_F_SP_ENQ | RING_F_SC_DEQ);
1697 if (ivshmem->cring_rx == NULL) {
1702 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
1703 &ivshmem->cring_tx, 1, SOCKET0);
1710 ivshmem->user_port_id = port_no;
1711 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
1712 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
1714 *eth_port_id = ivshmem->eth_port_id;
1719 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
1721 struct dpdk_ring *ivshmem;
1722 unsigned int port_no;
1725 /* Names always start with "dpdkr" */
1726 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
1731 /* look through our list to find the device */
1732 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
1733 if (ivshmem->user_port_id == port_no) {
1734 VLOG_INFO("Found dpdk ring device %s:", dev_name);
1735 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
1739 /* Need to create the device rings */
1740 return dpdk_ring_create(dev_name, port_no, eth_port_id);
1744 netdev_dpdk_ring_send(struct netdev *netdev, int qid OVS_UNUSED,
1745 struct dp_packet **pkts, int cnt, bool may_steal)
1747 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1749 /* DPDK Rings have a single TX queue, Therefore needs locking. */
1750 rte_spinlock_lock(&dev->txq_lock);
1751 netdev_dpdk_send__(dev, 0, pkts, cnt, may_steal);
1752 rte_spinlock_unlock(&dev->txq_lock);
1757 netdev_dpdk_ring_construct(struct netdev *netdev)
1759 unsigned int port_no = 0;
1762 if (rte_eal_init_ret) {
1763 return rte_eal_init_ret;
1766 ovs_mutex_lock(&dpdk_mutex);
1768 err = dpdk_ring_open(netdev->name, &port_no);
1773 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
1776 ovs_mutex_unlock(&dpdk_mutex);
1780 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, DESTRUCT, MULTIQ, SEND, \
1781 GET_CARRIER, GET_STATS, GET_FEATURES, GET_STATUS, RXQ_RECV) \
1785 NULL, /* netdev_dpdk_run */ \
1786 NULL, /* netdev_dpdk_wait */ \
1788 netdev_dpdk_alloc, \
1791 netdev_dpdk_dealloc, \
1792 netdev_dpdk_get_config, \
1793 NULL, /* netdev_dpdk_set_config */ \
1794 NULL, /* get_tunnel_config */ \
1795 NULL, /* build header */ \
1796 NULL, /* push header */ \
1797 NULL, /* pop header */ \
1798 netdev_dpdk_get_numa_id, /* get_numa_id */ \
1799 MULTIQ, /* set_multiq */ \
1802 NULL, /* send_wait */ \
1804 netdev_dpdk_set_etheraddr, \
1805 netdev_dpdk_get_etheraddr, \
1806 netdev_dpdk_get_mtu, \
1807 netdev_dpdk_set_mtu, \
1808 netdev_dpdk_get_ifindex, \
1810 netdev_dpdk_get_carrier_resets, \
1811 netdev_dpdk_set_miimon, \
1814 NULL, /* set_advertisements */ \
1816 NULL, /* set_policing */ \
1817 NULL, /* get_qos_types */ \
1818 NULL, /* get_qos_capabilities */ \
1819 NULL, /* get_qos */ \
1820 NULL, /* set_qos */ \
1821 NULL, /* get_queue */ \
1822 NULL, /* set_queue */ \
1823 NULL, /* delete_queue */ \
1824 NULL, /* get_queue_stats */ \
1825 NULL, /* queue_dump_start */ \
1826 NULL, /* queue_dump_next */ \
1827 NULL, /* queue_dump_done */ \
1828 NULL, /* dump_queue_stats */ \
1830 NULL, /* get_in4 */ \
1831 NULL, /* set_in4 */ \
1832 NULL, /* get_in6 */ \
1833 NULL, /* add_router */ \
1834 NULL, /* get_next_hop */ \
1836 NULL, /* arp_lookup */ \
1838 netdev_dpdk_update_flags, \
1840 netdev_dpdk_rxq_alloc, \
1841 netdev_dpdk_rxq_construct, \
1842 netdev_dpdk_rxq_destruct, \
1843 netdev_dpdk_rxq_dealloc, \
1845 NULL, /* rx_wait */ \
1846 NULL, /* rxq_drain */ \
1850 dpdk_init(int argc, char **argv)
1854 char *pragram_name = argv[0];
1856 if (argc < 2 || strcmp(argv[1], "--dpdk"))
1859 /* Remove the --dpdk argument from arg list.*/
1863 /* If the cuse_dev_name parameter has been provided, set 'cuse_dev_name' to
1864 * this string if it meets the correct criteria. Otherwise, set it to the
1865 * default (vhost-net).
1867 if (!strcmp(argv[1], "--cuse_dev_name") &&
1868 (strlen(argv[2]) <= NAME_MAX)) {
1870 cuse_dev_name = strdup(argv[2]);
1872 /* Remove the cuse_dev_name configuration parameters from the argument
1873 * list, so that the correct elements are passed to the DPDK
1874 * initialization function
1877 argv += 2; /* Increment by two to bypass the cuse_dev_name arguments */
1880 VLOG_ERR("User-provided cuse_dev_name in use: /dev/%s", cuse_dev_name);
1882 cuse_dev_name = "vhost-net";
1883 VLOG_INFO("No cuse_dev_name provided - defaulting to /dev/vhost-net");
1886 /* Keep the program name argument as this is needed for call to
1889 argv[0] = pragram_name;
1891 /* Make sure things are initialized ... */
1892 result = rte_eal_init(argc, argv);
1894 ovs_abort(result, "Cannot init EAL");
1897 rte_memzone_dump(stdout);
1898 rte_eal_init_ret = 0;
1900 if (argc > result) {
1901 argv[result] = argv[0];
1904 /* We are called from the main thread here */
1905 thread_set_nonpmd();
1907 return result + 1 + base;
1910 const struct netdev_class dpdk_class =
1914 netdev_dpdk_construct,
1915 netdev_dpdk_destruct,
1916 netdev_dpdk_set_multiq,
1917 netdev_dpdk_eth_send,
1918 netdev_dpdk_get_carrier,
1919 netdev_dpdk_get_stats,
1920 netdev_dpdk_get_features,
1921 netdev_dpdk_get_status,
1922 netdev_dpdk_rxq_recv);
1924 const struct netdev_class dpdk_ring_class =
1928 netdev_dpdk_ring_construct,
1929 netdev_dpdk_destruct,
1931 netdev_dpdk_ring_send,
1932 netdev_dpdk_get_carrier,
1933 netdev_dpdk_get_stats,
1934 netdev_dpdk_get_features,
1935 netdev_dpdk_get_status,
1936 netdev_dpdk_rxq_recv);
1938 const struct netdev_class dpdk_vhost_class =
1941 dpdk_vhost_class_init,
1942 netdev_dpdk_vhost_construct,
1943 netdev_dpdk_vhost_destruct,
1944 netdev_dpdk_vhost_set_multiq,
1945 netdev_dpdk_vhost_send,
1946 netdev_dpdk_vhost_get_carrier,
1947 netdev_dpdk_vhost_get_stats,
1950 netdev_dpdk_vhost_rxq_recv);
1953 netdev_dpdk_register(void)
1955 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
1957 if (rte_eal_init_ret) {
1961 if (ovsthread_once_start(&once)) {
1963 netdev_register_provider(&dpdk_class);
1964 netdev_register_provider(&dpdk_ring_class);
1965 netdev_register_provider(&dpdk_vhost_class);
1966 ovsthread_once_done(&once);
1971 pmd_thread_setaffinity_cpu(int cpu)
1977 CPU_SET(cpu, &cpuset);
1978 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
1980 VLOG_ERR("Thread affinity error %d",err);
1983 /* NON_PMD_CORE_ID is reserved for use by non pmd threads. */
1984 ovs_assert(cpu != NON_PMD_CORE_ID);
1985 RTE_PER_LCORE(_lcore_id) = cpu;
1991 thread_set_nonpmd(void)
1993 /* We have to use NON_PMD_CORE_ID to allow non-pmd threads to perform
1994 * certain DPDK operations, like rte_eth_dev_configure(). */
1995 RTE_PER_LCORE(_lcore_id) = NON_PMD_CORE_ID;
2001 return rte_lcore_id() != NON_PMD_CORE_ID;