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. */
162 rte_spinlock_t tx_lock; /* Protects the members and the NIC queue
163 * from concurrent access. It is used only
164 * if the queue is shared among different
165 * pmd threads (see 'txq_needs_locking'). */
167 struct rte_mbuf *burst_pkts[MAX_TX_QUEUE_LEN];
170 /* dpdk has no way to remove dpdk ring ethernet devices
171 so we have to keep them around once they've been created
174 static struct ovs_list dpdk_ring_list OVS_GUARDED_BY(dpdk_mutex)
175 = OVS_LIST_INITIALIZER(&dpdk_ring_list);
178 /* For the client rings */
179 struct rte_ring *cring_tx;
180 struct rte_ring *cring_rx;
181 int user_port_id; /* User given port no, parsed from port name */
182 int eth_port_id; /* ethernet device port id */
183 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
190 enum dpdk_dev_type type;
192 struct dpdk_tx_queue *tx_q;
194 struct ovs_mutex mutex OVS_ACQ_AFTER(dpdk_mutex);
196 struct dpdk_mp *dpdk_mp;
200 struct netdev_stats stats;
202 rte_spinlock_t stats_lock;
204 uint8_t hwaddr[ETH_ADDR_LEN];
205 enum netdev_flags flags;
207 struct rte_eth_link link;
210 /* The user might request more txqs than the NIC has. We remap those
211 * ('up.n_txq') on these ('real_n_txq').
212 * If the numbers match, 'txq_needs_locking' is false, otherwise it is
213 * true and we will take a spinlock on transmission */
215 bool txq_needs_locking;
217 /* Spinlock for vhost transmission. Other DPDK devices use spinlocks in
219 rte_spinlock_t vhost_tx_lock;
221 /* virtio-net structure for vhost device */
222 OVSRCU_TYPE(struct virtio_net *) virtio_dev;
225 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
228 struct netdev_rxq_dpdk {
229 struct netdev_rxq up;
233 static bool thread_is_pmd(void);
235 static int netdev_dpdk_construct(struct netdev *);
237 struct virtio_net * netdev_dpdk_get_virtio(const struct netdev_dpdk *dev);
240 is_dpdk_class(const struct netdev_class *class)
242 return class->construct == netdev_dpdk_construct;
245 /* XXX: use dpdk malloc for entire OVS. in fact huge page should be used
246 * for all other segments data, bss and text. */
249 dpdk_rte_mzalloc(size_t sz)
253 ptr = rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE);
260 /* XXX this function should be called only by pmd threads (or by non pmd
261 * threads holding the nonpmd_mempool_mutex) */
263 free_dpdk_buf(struct dp_packet *p)
265 struct rte_mbuf *pkt = (struct rte_mbuf *) p;
267 rte_pktmbuf_free_seg(pkt);
271 __rte_pktmbuf_init(struct rte_mempool *mp,
272 void *opaque_arg OVS_UNUSED,
274 unsigned i OVS_UNUSED)
276 struct rte_mbuf *m = _m;
277 uint32_t buf_len = mp->elt_size - sizeof(struct dp_packet);
279 RTE_MBUF_ASSERT(mp->elt_size >= sizeof(struct dp_packet));
281 memset(m, 0, mp->elt_size);
283 /* start of buffer is just after mbuf structure */
284 m->buf_addr = (char *)m + sizeof(struct dp_packet);
285 m->buf_physaddr = rte_mempool_virt2phy(mp, m) +
286 sizeof(struct dp_packet);
287 m->buf_len = (uint16_t)buf_len;
289 /* keep some headroom between start of buffer and data */
290 m->data_off = RTE_MIN(RTE_PKTMBUF_HEADROOM, m->buf_len);
292 /* init some constant fields */
299 ovs_rte_pktmbuf_init(struct rte_mempool *mp,
300 void *opaque_arg OVS_UNUSED,
302 unsigned i OVS_UNUSED)
304 struct rte_mbuf *m = _m;
306 __rte_pktmbuf_init(mp, opaque_arg, _m, i);
308 dp_packet_init_dpdk((struct dp_packet *) m, m->buf_len);
311 static struct dpdk_mp *
312 dpdk_mp_get(int socket_id, int mtu) OVS_REQUIRES(dpdk_mutex)
314 struct dpdk_mp *dmp = NULL;
315 char mp_name[RTE_MEMPOOL_NAMESIZE];
318 LIST_FOR_EACH (dmp, list_node, &dpdk_mp_list) {
319 if (dmp->socket_id == socket_id && dmp->mtu == mtu) {
325 dmp = dpdk_rte_mzalloc(sizeof *dmp);
326 dmp->socket_id = socket_id;
330 mp_size = MAX_NB_MBUF;
332 if (snprintf(mp_name, RTE_MEMPOOL_NAMESIZE, "ovs_mp_%d_%d_%u",
333 dmp->mtu, dmp->socket_id, mp_size) < 0) {
337 dmp->mp = rte_mempool_create(mp_name, mp_size, MBUF_SIZE(mtu),
339 sizeof(struct rte_pktmbuf_pool_private),
340 rte_pktmbuf_pool_init, NULL,
341 ovs_rte_pktmbuf_init, NULL,
343 } while (!dmp->mp && rte_errno == ENOMEM && (mp_size /= 2) >= MIN_NB_MBUF);
345 if (dmp->mp == NULL) {
348 VLOG_DBG("Allocated \"%s\" mempool with %u mbufs", mp_name, mp_size );
351 list_push_back(&dpdk_mp_list, &dmp->list_node);
356 dpdk_mp_put(struct dpdk_mp *dmp)
364 ovs_assert(dmp->refcount >= 0);
367 /* I could not find any API to destroy mp. */
368 if (dmp->refcount == 0) {
369 list_delete(dmp->list_node);
370 /* destroy mp-pool. */
376 check_link_status(struct netdev_dpdk *dev)
378 struct rte_eth_link link;
380 rte_eth_link_get_nowait(dev->port_id, &link);
382 if (dev->link.link_status != link.link_status) {
383 netdev_change_seq_changed(&dev->up);
385 dev->link_reset_cnt++;
387 if (dev->link.link_status) {
388 VLOG_DBG_RL(&rl, "Port %d Link Up - speed %u Mbps - %s",
389 dev->port_id, (unsigned)dev->link.link_speed,
390 (dev->link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
391 ("full-duplex") : ("half-duplex"));
393 VLOG_DBG_RL(&rl, "Port %d Link Down", dev->port_id);
399 dpdk_watchdog(void *dummy OVS_UNUSED)
401 struct netdev_dpdk *dev;
403 pthread_detach(pthread_self());
406 ovs_mutex_lock(&dpdk_mutex);
407 LIST_FOR_EACH (dev, list_node, &dpdk_list) {
408 ovs_mutex_lock(&dev->mutex);
409 check_link_status(dev);
410 ovs_mutex_unlock(&dev->mutex);
412 ovs_mutex_unlock(&dpdk_mutex);
413 xsleep(DPDK_PORT_WATCHDOG_INTERVAL);
420 dpdk_eth_dev_init(struct netdev_dpdk *dev) OVS_REQUIRES(dpdk_mutex)
422 struct rte_pktmbuf_pool_private *mbp_priv;
423 struct rte_eth_dev_info info;
424 struct ether_addr eth_addr;
428 if (dev->port_id < 0 || dev->port_id >= rte_eth_dev_count()) {
432 rte_eth_dev_info_get(dev->port_id, &info);
433 dev->up.n_rxq = MIN(info.max_rx_queues, dev->up.n_rxq);
434 dev->real_n_txq = MIN(info.max_tx_queues, dev->up.n_txq);
436 diag = rte_eth_dev_configure(dev->port_id, dev->up.n_rxq, dev->real_n_txq,
439 VLOG_ERR("eth dev config error %d. rxq:%d txq:%d", diag, dev->up.n_rxq,
444 for (i = 0; i < dev->real_n_txq; i++) {
445 diag = rte_eth_tx_queue_setup(dev->port_id, i, NIC_PORT_TX_Q_SIZE,
446 dev->socket_id, NULL);
448 VLOG_ERR("eth dev tx queue setup error %d",diag);
453 for (i = 0; i < dev->up.n_rxq; i++) {
454 diag = rte_eth_rx_queue_setup(dev->port_id, i, NIC_PORT_RX_Q_SIZE,
456 NULL, dev->dpdk_mp->mp);
458 VLOG_ERR("eth dev rx queue setup error %d",diag);
463 diag = rte_eth_dev_start(dev->port_id);
465 VLOG_ERR("eth dev start error %d",diag);
469 rte_eth_promiscuous_enable(dev->port_id);
470 rte_eth_allmulticast_enable(dev->port_id);
472 memset(ð_addr, 0x0, sizeof(eth_addr));
473 rte_eth_macaddr_get(dev->port_id, ð_addr);
474 VLOG_INFO_RL(&rl, "Port %d: "ETH_ADDR_FMT"",
475 dev->port_id, ETH_ADDR_ARGS(eth_addr.addr_bytes));
477 memcpy(dev->hwaddr, eth_addr.addr_bytes, ETH_ADDR_LEN);
478 rte_eth_link_get_nowait(dev->port_id, &dev->link);
480 mbp_priv = rte_mempool_get_priv(dev->dpdk_mp->mp);
481 dev->buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
483 dev->flags = NETDEV_UP | NETDEV_PROMISC;
487 static struct netdev_dpdk *
488 netdev_dpdk_cast(const struct netdev *netdev)
490 return CONTAINER_OF(netdev, struct netdev_dpdk, up);
493 static struct netdev *
494 netdev_dpdk_alloc(void)
496 struct netdev_dpdk *netdev = dpdk_rte_mzalloc(sizeof *netdev);
501 netdev_dpdk_alloc_txq(struct netdev_dpdk *netdev, unsigned int n_txqs)
505 netdev->tx_q = dpdk_rte_mzalloc(n_txqs * sizeof *netdev->tx_q);
506 for (i = 0; i < n_txqs; i++) {
507 int numa_id = ovs_numa_get_numa_id(i);
509 if (!netdev->txq_needs_locking) {
510 /* Each index is considered as a cpu core id, since there should
511 * be one tx queue for each cpu core. If the corresponding core
512 * is not on the same numa node as 'netdev', flags the
514 netdev->tx_q[i].flush_tx = netdev->socket_id == numa_id;
516 /* Queues are shared among CPUs. Always flush */
517 netdev->tx_q[i].flush_tx = true;
519 rte_spinlock_init(&netdev->tx_q[i].tx_lock);
524 netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no,
525 enum dpdk_dev_type type)
526 OVS_REQUIRES(dpdk_mutex)
528 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
532 ovs_mutex_init(&netdev->mutex);
533 ovs_mutex_lock(&netdev->mutex);
535 rte_spinlock_init(&netdev->stats_lock);
537 /* If the 'sid' is negative, it means that the kernel fails
538 * to obtain the pci numa info. In that situation, always
540 if (type == DPDK_DEV_ETH) {
541 sid = rte_eth_dev_socket_id(port_no);
543 sid = rte_lcore_to_socket_id(rte_get_master_lcore());
546 netdev->socket_id = sid < 0 ? SOCKET0 : sid;
547 netdev->port_id = port_no;
550 netdev->mtu = ETHER_MTU;
551 netdev->max_packet_len = MTU_TO_MAX_LEN(netdev->mtu);
553 netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, netdev->mtu);
554 if (!netdev->dpdk_mp) {
559 netdev_->n_txq = NR_QUEUE;
560 netdev_->n_rxq = NR_QUEUE;
561 netdev->real_n_txq = NR_QUEUE;
563 if (type == DPDK_DEV_ETH) {
564 netdev_dpdk_alloc_txq(netdev, NR_QUEUE);
565 err = dpdk_eth_dev_init(netdev);
571 list_push_back(&dpdk_list, &netdev->list_node);
575 rte_free(netdev->tx_q);
577 ovs_mutex_unlock(&netdev->mutex);
582 dpdk_dev_parse_name(const char dev_name[], const char prefix[],
583 unsigned int *port_no)
587 if (strncmp(dev_name, prefix, strlen(prefix))) {
591 cport = dev_name + strlen(prefix);
592 *port_no = strtol(cport, NULL, 0); /* string must be null terminated */
597 netdev_dpdk_vhost_construct(struct netdev *netdev_)
599 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
602 if (rte_eal_init_ret) {
603 return rte_eal_init_ret;
606 ovs_mutex_lock(&dpdk_mutex);
607 err = netdev_dpdk_init(netdev_, -1, DPDK_DEV_VHOST);
608 ovs_mutex_unlock(&dpdk_mutex);
610 rte_spinlock_init(&netdev->vhost_tx_lock);
616 netdev_dpdk_construct(struct netdev *netdev)
618 unsigned int port_no;
621 if (rte_eal_init_ret) {
622 return rte_eal_init_ret;
625 /* Names always start with "dpdk" */
626 err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no);
631 ovs_mutex_lock(&dpdk_mutex);
632 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
633 ovs_mutex_unlock(&dpdk_mutex);
638 netdev_dpdk_destruct(struct netdev *netdev_)
640 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
642 ovs_mutex_lock(&dev->mutex);
643 rte_eth_dev_stop(dev->port_id);
644 ovs_mutex_unlock(&dev->mutex);
646 ovs_mutex_lock(&dpdk_mutex);
648 list_remove(&dev->list_node);
649 dpdk_mp_put(dev->dpdk_mp);
650 ovs_mutex_unlock(&dpdk_mutex);
654 netdev_dpdk_vhost_destruct(struct netdev *netdev_)
656 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
658 /* Can't remove a port while a guest is attached to it. */
659 if (netdev_dpdk_get_virtio(dev) != NULL) {
660 VLOG_ERR("Can not remove port, vhost device still attached");
664 ovs_mutex_lock(&dpdk_mutex);
665 list_remove(&dev->list_node);
666 dpdk_mp_put(dev->dpdk_mp);
667 ovs_mutex_unlock(&dpdk_mutex);
671 netdev_dpdk_dealloc(struct netdev *netdev_)
673 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
679 netdev_dpdk_get_config(const struct netdev *netdev_, struct smap *args)
681 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
683 ovs_mutex_lock(&dev->mutex);
685 smap_add_format(args, "configured_rx_queues", "%d", netdev_->n_rxq);
686 smap_add_format(args, "requested_tx_queues", "%d", netdev_->n_txq);
687 smap_add_format(args, "configured_tx_queues", "%d", dev->real_n_txq);
688 ovs_mutex_unlock(&dev->mutex);
694 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
696 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
698 return netdev->socket_id;
701 /* Sets the number of tx queues and rx queues for the dpdk interface.
702 * If the configuration fails, do not try restoring its old configuration
703 * and just returns the error. */
705 netdev_dpdk_set_multiq(struct netdev *netdev_, unsigned int n_txq,
708 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
711 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
715 ovs_mutex_lock(&dpdk_mutex);
716 ovs_mutex_lock(&netdev->mutex);
718 rte_eth_dev_stop(netdev->port_id);
720 netdev->up.n_txq = n_txq;
721 netdev->up.n_rxq = n_rxq;
723 rte_free(netdev->tx_q);
724 err = dpdk_eth_dev_init(netdev);
725 netdev_dpdk_alloc_txq(netdev, netdev->real_n_txq);
727 netdev->txq_needs_locking = netdev->real_n_txq != netdev->up.n_txq;
729 ovs_mutex_unlock(&netdev->mutex);
730 ovs_mutex_unlock(&dpdk_mutex);
736 netdev_dpdk_vhost_set_multiq(struct netdev *netdev_, unsigned int n_txq,
739 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
742 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
746 ovs_mutex_lock(&dpdk_mutex);
747 ovs_mutex_lock(&netdev->mutex);
749 netdev->up.n_txq = n_txq;
750 netdev->real_n_txq = 1;
751 netdev->up.n_rxq = 1;
753 ovs_mutex_unlock(&netdev->mutex);
754 ovs_mutex_unlock(&dpdk_mutex);
759 static struct netdev_rxq *
760 netdev_dpdk_rxq_alloc(void)
762 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
767 static struct netdev_rxq_dpdk *
768 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
770 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
774 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
776 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
777 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
779 ovs_mutex_lock(&netdev->mutex);
780 rx->port_id = netdev->port_id;
781 ovs_mutex_unlock(&netdev->mutex);
787 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
792 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
794 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
800 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
802 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
805 while (nb_tx != txq->count) {
808 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
817 if (OVS_UNLIKELY(nb_tx != txq->count)) {
818 /* free buffers, which we couldn't transmit, one at a time (each
819 * packet could come from a different mempool) */
822 for (i = nb_tx; i < txq->count; i++) {
823 rte_pktmbuf_free_seg(txq->burst_pkts[i]);
825 rte_spinlock_lock(&dev->stats_lock);
826 dev->stats.tx_dropped += txq->count-nb_tx;
827 rte_spinlock_unlock(&dev->stats_lock);
831 txq->tsc = rte_get_timer_cycles();
835 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
837 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
839 if (txq->count == 0) {
842 dpdk_queue_flush__(dev, qid);
846 is_vhost_running(struct virtio_net *dev)
848 return (dev != NULL && (dev->flags & VIRTIO_DEV_RUNNING));
852 * The receive path for the vhost port is the TX path out from guest.
855 netdev_dpdk_vhost_rxq_recv(struct netdev_rxq *rxq_,
856 struct dp_packet **packets, int *c)
858 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
859 struct netdev *netdev = rx->up.netdev;
860 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
861 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
865 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
869 nb_rx = rte_vhost_dequeue_burst(virtio_dev, qid,
870 vhost_dev->dpdk_mp->mp,
871 (struct rte_mbuf **)packets,
877 rte_spinlock_lock(&vhost_dev->stats_lock);
878 vhost_dev->stats.rx_packets += (uint64_t)nb_rx;
879 rte_spinlock_unlock(&vhost_dev->stats_lock);
886 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dp_packet **packets,
889 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
890 struct netdev *netdev = rx->up.netdev;
891 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
894 /* There is only one tx queue for this core. Do not flush other
896 if (rxq_->queue_id == rte_lcore_id()) {
897 dpdk_queue_flush(dev, rxq_->queue_id);
900 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
901 (struct rte_mbuf **) packets,
913 __netdev_dpdk_vhost_send(struct netdev *netdev, struct dp_packet **pkts,
914 int cnt, bool may_steal)
916 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
917 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
918 struct rte_mbuf **cur_pkts = (struct rte_mbuf **) pkts;
919 unsigned int total_pkts = cnt;
922 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
923 rte_spinlock_lock(&vhost_dev->stats_lock);
924 vhost_dev->stats.tx_dropped+= cnt;
925 rte_spinlock_unlock(&vhost_dev->stats_lock);
929 /* There is vHost TX single queue, So we need to lock it for TX. */
930 rte_spinlock_lock(&vhost_dev->vhost_tx_lock);
933 unsigned int tx_pkts;
935 tx_pkts = rte_vhost_enqueue_burst(virtio_dev, VIRTIO_RXQ,
937 if (OVS_LIKELY(tx_pkts)) {
938 /* Packets have been sent.*/
940 /* Prepare for possible next iteration.*/
941 cur_pkts = &cur_pkts[tx_pkts];
943 uint64_t timeout = VHOST_ENQ_RETRY_USECS * rte_get_timer_hz() / 1E6;
944 unsigned int expired = 0;
947 start = rte_get_timer_cycles();
951 * Unable to enqueue packets to vhost interface.
952 * Check available entries before retrying.
954 while (!rte_vring_available_entries(virtio_dev, VIRTIO_RXQ)) {
955 if (OVS_UNLIKELY((rte_get_timer_cycles() - start) > timeout)) {
961 /* break out of main loop. */
966 rte_spinlock_unlock(&vhost_dev->vhost_tx_lock);
968 rte_spinlock_lock(&vhost_dev->stats_lock);
969 vhost_dev->stats.tx_packets += (total_pkts - cnt);
970 vhost_dev->stats.tx_dropped += cnt;
971 rte_spinlock_unlock(&vhost_dev->stats_lock);
977 for (i = 0; i < total_pkts; i++) {
978 dp_packet_delete(pkts[i]);
984 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
985 struct rte_mbuf **pkts, int cnt)
987 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
993 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
994 int tocopy = MIN(freeslots, cnt-i);
996 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
997 tocopy * sizeof (struct rte_mbuf *));
999 txq->count += tocopy;
1002 if (txq->count == MAX_TX_QUEUE_LEN || txq->flush_tx) {
1003 dpdk_queue_flush__(dev, qid);
1005 diff_tsc = rte_get_timer_cycles() - txq->tsc;
1006 if (diff_tsc >= DRAIN_TSC) {
1007 dpdk_queue_flush__(dev, qid);
1012 /* Tx function. Transmit packets indefinitely */
1014 dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet **pkts,
1016 OVS_NO_THREAD_SAFETY_ANALYSIS
1018 #if !defined(__CHECKER__) && !defined(_WIN32)
1019 const size_t PKT_ARRAY_SIZE = cnt;
1021 /* Sparse or MSVC doesn't like variable length array. */
1022 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
1024 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1025 struct rte_mbuf *mbufs[PKT_ARRAY_SIZE];
1030 /* If we are on a non pmd thread we have to use the mempool mutex, because
1031 * every non pmd thread shares the same mempool cache */
1033 if (!thread_is_pmd()) {
1034 ovs_mutex_lock(&nonpmd_mempool_mutex);
1037 for (i = 0; i < cnt; i++) {
1038 int size = dp_packet_size(pkts[i]);
1040 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1041 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1042 (int)size , dev->max_packet_len);
1048 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
1050 if (!mbufs[newcnt]) {
1055 /* We have to do a copy for now */
1056 memcpy(rte_pktmbuf_mtod(mbufs[newcnt], void *), dp_packet_data(pkts[i]), size);
1058 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
1059 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
1064 if (OVS_UNLIKELY(dropped)) {
1065 rte_spinlock_lock(&dev->stats_lock);
1066 dev->stats.tx_dropped += dropped;
1067 rte_spinlock_unlock(&dev->stats_lock);
1070 if (dev->type == DPDK_DEV_VHOST) {
1071 __netdev_dpdk_vhost_send(netdev, (struct dp_packet **) mbufs, newcnt, true);
1073 dpdk_queue_pkts(dev, qid, mbufs, newcnt);
1074 dpdk_queue_flush(dev, qid);
1077 if (!thread_is_pmd()) {
1078 ovs_mutex_unlock(&nonpmd_mempool_mutex);
1083 netdev_dpdk_vhost_send(struct netdev *netdev, int qid OVS_UNUSED, struct dp_packet **pkts,
1084 int cnt, bool may_steal)
1086 if (OVS_UNLIKELY(pkts[0]->source != DPBUF_DPDK)) {
1089 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1091 for (i = 0; i < cnt; i++) {
1092 dp_packet_delete(pkts[i]);
1096 __netdev_dpdk_vhost_send(netdev, pkts, cnt, may_steal);
1102 netdev_dpdk_send__(struct netdev_dpdk *dev, int qid,
1103 struct dp_packet **pkts, int cnt, bool may_steal)
1107 if (OVS_UNLIKELY(dev->txq_needs_locking)) {
1108 qid = qid % dev->real_n_txq;
1109 rte_spinlock_lock(&dev->tx_q[qid].tx_lock);
1112 if (OVS_UNLIKELY(!may_steal ||
1113 pkts[0]->source != DPBUF_DPDK)) {
1114 struct netdev *netdev = &dev->up;
1116 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1119 for (i = 0; i < cnt; i++) {
1120 dp_packet_delete(pkts[i]);
1124 int next_tx_idx = 0;
1127 for (i = 0; i < cnt; i++) {
1128 int size = dp_packet_size(pkts[i]);
1130 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1131 if (next_tx_idx != i) {
1132 dpdk_queue_pkts(dev, qid,
1133 (struct rte_mbuf **)&pkts[next_tx_idx],
1137 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1138 (int)size , dev->max_packet_len);
1140 dp_packet_delete(pkts[i]);
1142 next_tx_idx = i + 1;
1145 if (next_tx_idx != cnt) {
1146 dpdk_queue_pkts(dev, qid,
1147 (struct rte_mbuf **)&pkts[next_tx_idx],
1151 if (OVS_UNLIKELY(dropped)) {
1152 rte_spinlock_lock(&dev->stats_lock);
1153 dev->stats.tx_dropped += dropped;
1154 rte_spinlock_unlock(&dev->stats_lock);
1158 if (OVS_UNLIKELY(dev->txq_needs_locking)) {
1159 rte_spinlock_unlock(&dev->tx_q[qid].tx_lock);
1164 netdev_dpdk_eth_send(struct netdev *netdev, int qid,
1165 struct dp_packet **pkts, int cnt, bool may_steal)
1167 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1169 netdev_dpdk_send__(dev, qid, pkts, cnt, may_steal);
1174 netdev_dpdk_set_etheraddr(struct netdev *netdev,
1175 const uint8_t mac[ETH_ADDR_LEN])
1177 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1179 ovs_mutex_lock(&dev->mutex);
1180 if (!eth_addr_equals(dev->hwaddr, mac)) {
1181 memcpy(dev->hwaddr, mac, ETH_ADDR_LEN);
1182 netdev_change_seq_changed(netdev);
1184 ovs_mutex_unlock(&dev->mutex);
1190 netdev_dpdk_get_etheraddr(const struct netdev *netdev,
1191 uint8_t mac[ETH_ADDR_LEN])
1193 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1195 ovs_mutex_lock(&dev->mutex);
1196 memcpy(mac, dev->hwaddr, ETH_ADDR_LEN);
1197 ovs_mutex_unlock(&dev->mutex);
1203 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
1205 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1207 ovs_mutex_lock(&dev->mutex);
1209 ovs_mutex_unlock(&dev->mutex);
1215 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
1217 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1219 struct dpdk_mp *old_mp;
1222 ovs_mutex_lock(&dpdk_mutex);
1223 ovs_mutex_lock(&dev->mutex);
1224 if (dev->mtu == mtu) {
1229 mp = dpdk_mp_get(dev->socket_id, dev->mtu);
1235 rte_eth_dev_stop(dev->port_id);
1238 old_mp = dev->dpdk_mp;
1241 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1243 err = dpdk_eth_dev_init(dev);
1247 dev->dpdk_mp = old_mp;
1248 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1249 dpdk_eth_dev_init(dev);
1253 dpdk_mp_put(old_mp);
1254 netdev_change_seq_changed(netdev);
1256 ovs_mutex_unlock(&dev->mutex);
1257 ovs_mutex_unlock(&dpdk_mutex);
1262 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
1265 netdev_dpdk_vhost_get_stats(const struct netdev *netdev,
1266 struct netdev_stats *stats)
1268 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1270 ovs_mutex_lock(&dev->mutex);
1271 memset(stats, 0, sizeof(*stats));
1272 /* Unsupported Stats */
1273 stats->rx_errors = UINT64_MAX;
1274 stats->tx_errors = UINT64_MAX;
1275 stats->multicast = UINT64_MAX;
1276 stats->collisions = UINT64_MAX;
1277 stats->rx_crc_errors = UINT64_MAX;
1278 stats->rx_fifo_errors = UINT64_MAX;
1279 stats->rx_frame_errors = UINT64_MAX;
1280 stats->rx_length_errors = UINT64_MAX;
1281 stats->rx_missed_errors = UINT64_MAX;
1282 stats->rx_over_errors = UINT64_MAX;
1283 stats->tx_aborted_errors = UINT64_MAX;
1284 stats->tx_carrier_errors = UINT64_MAX;
1285 stats->tx_errors = UINT64_MAX;
1286 stats->tx_fifo_errors = UINT64_MAX;
1287 stats->tx_heartbeat_errors = UINT64_MAX;
1288 stats->tx_window_errors = UINT64_MAX;
1289 stats->rx_bytes += UINT64_MAX;
1290 stats->rx_dropped += UINT64_MAX;
1291 stats->tx_bytes += UINT64_MAX;
1293 rte_spinlock_lock(&dev->stats_lock);
1294 /* Supported Stats */
1295 stats->rx_packets += dev->stats.rx_packets;
1296 stats->tx_packets += dev->stats.tx_packets;
1297 stats->tx_dropped += dev->stats.tx_dropped;
1298 rte_spinlock_unlock(&dev->stats_lock);
1299 ovs_mutex_unlock(&dev->mutex);
1305 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
1307 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1308 struct rte_eth_stats rte_stats;
1311 netdev_dpdk_get_carrier(netdev, &gg);
1312 ovs_mutex_lock(&dev->mutex);
1313 rte_eth_stats_get(dev->port_id, &rte_stats);
1315 memset(stats, 0, sizeof(*stats));
1317 stats->rx_packets = rte_stats.ipackets;
1318 stats->tx_packets = rte_stats.opackets;
1319 stats->rx_bytes = rte_stats.ibytes;
1320 stats->tx_bytes = rte_stats.obytes;
1321 stats->rx_errors = rte_stats.ierrors;
1322 stats->tx_errors = rte_stats.oerrors;
1323 stats->multicast = rte_stats.imcasts;
1325 rte_spinlock_lock(&dev->stats_lock);
1326 stats->tx_dropped = dev->stats.tx_dropped;
1327 rte_spinlock_unlock(&dev->stats_lock);
1328 ovs_mutex_unlock(&dev->mutex);
1334 netdev_dpdk_get_features(const struct netdev *netdev_,
1335 enum netdev_features *current,
1336 enum netdev_features *advertised OVS_UNUSED,
1337 enum netdev_features *supported OVS_UNUSED,
1338 enum netdev_features *peer OVS_UNUSED)
1340 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1341 struct rte_eth_link link;
1343 ovs_mutex_lock(&dev->mutex);
1345 ovs_mutex_unlock(&dev->mutex);
1347 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
1348 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
1349 *current = NETDEV_F_AUTONEG;
1351 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
1352 if (link.link_speed == ETH_LINK_SPEED_10) {
1353 *current = NETDEV_F_10MB_HD;
1355 if (link.link_speed == ETH_LINK_SPEED_100) {
1356 *current = NETDEV_F_100MB_HD;
1358 if (link.link_speed == ETH_LINK_SPEED_1000) {
1359 *current = NETDEV_F_1GB_HD;
1361 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
1362 if (link.link_speed == ETH_LINK_SPEED_10) {
1363 *current = NETDEV_F_10MB_FD;
1365 if (link.link_speed == ETH_LINK_SPEED_100) {
1366 *current = NETDEV_F_100MB_FD;
1368 if (link.link_speed == ETH_LINK_SPEED_1000) {
1369 *current = NETDEV_F_1GB_FD;
1371 if (link.link_speed == ETH_LINK_SPEED_10000) {
1372 *current = NETDEV_F_10GB_FD;
1380 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1382 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1385 ovs_mutex_lock(&dev->mutex);
1386 ifindex = dev->port_id;
1387 ovs_mutex_unlock(&dev->mutex);
1393 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1395 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1397 ovs_mutex_lock(&dev->mutex);
1398 check_link_status(dev);
1399 *carrier = dev->link.link_status;
1401 ovs_mutex_unlock(&dev->mutex);
1407 netdev_dpdk_vhost_get_carrier(const struct netdev *netdev_, bool *carrier)
1409 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1410 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(dev);
1412 ovs_mutex_lock(&dev->mutex);
1414 if (is_vhost_running(virtio_dev)) {
1420 ovs_mutex_unlock(&dev->mutex);
1425 static long long int
1426 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1428 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1429 long long int carrier_resets;
1431 ovs_mutex_lock(&dev->mutex);
1432 carrier_resets = dev->link_reset_cnt;
1433 ovs_mutex_unlock(&dev->mutex);
1435 return carrier_resets;
1439 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1440 long long int interval OVS_UNUSED)
1446 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1447 enum netdev_flags off, enum netdev_flags on,
1448 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1452 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1456 *old_flagsp = dev->flags;
1460 if (dev->flags == *old_flagsp) {
1464 if (dev->type == DPDK_DEV_ETH) {
1465 if (dev->flags & NETDEV_UP) {
1466 err = rte_eth_dev_start(dev->port_id);
1471 if (dev->flags & NETDEV_PROMISC) {
1472 rte_eth_promiscuous_enable(dev->port_id);
1475 if (!(dev->flags & NETDEV_UP)) {
1476 rte_eth_dev_stop(dev->port_id);
1484 netdev_dpdk_update_flags(struct netdev *netdev_,
1485 enum netdev_flags off, enum netdev_flags on,
1486 enum netdev_flags *old_flagsp)
1488 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1491 ovs_mutex_lock(&netdev->mutex);
1492 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1493 ovs_mutex_unlock(&netdev->mutex);
1499 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1501 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1502 struct rte_eth_dev_info dev_info;
1504 if (dev->port_id < 0)
1507 ovs_mutex_lock(&dev->mutex);
1508 rte_eth_dev_info_get(dev->port_id, &dev_info);
1509 ovs_mutex_unlock(&dev->mutex);
1511 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1513 smap_add_format(args, "port_no", "%d", dev->port_id);
1514 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1515 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1516 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1517 smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen);
1518 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1519 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1520 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1521 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1522 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1523 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1525 smap_add_format(args, "pci-vendor_id", "0x%u", dev_info.pci_dev->id.vendor_id);
1526 smap_add_format(args, "pci-device_id", "0x%x", dev_info.pci_dev->id.device_id);
1532 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1533 OVS_REQUIRES(dev->mutex)
1535 enum netdev_flags old_flags;
1538 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1540 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1545 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1546 const char *argv[], void *aux OVS_UNUSED)
1550 if (!strcasecmp(argv[argc - 1], "up")) {
1552 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1555 unixctl_command_reply_error(conn, "Invalid Admin State");
1560 struct netdev *netdev = netdev_from_name(argv[1]);
1561 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1562 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1564 ovs_mutex_lock(&dpdk_dev->mutex);
1565 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1566 ovs_mutex_unlock(&dpdk_dev->mutex);
1568 netdev_close(netdev);
1570 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1571 netdev_close(netdev);
1575 struct netdev_dpdk *netdev;
1577 ovs_mutex_lock(&dpdk_mutex);
1578 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1579 ovs_mutex_lock(&netdev->mutex);
1580 netdev_dpdk_set_admin_state__(netdev, up);
1581 ovs_mutex_unlock(&netdev->mutex);
1583 ovs_mutex_unlock(&dpdk_mutex);
1585 unixctl_command_reply(conn, "OK");
1589 * Set virtqueue flags so that we do not receive interrupts.
1592 set_irq_status(struct virtio_net *dev)
1594 dev->virtqueue[VIRTIO_RXQ]->used->flags = VRING_USED_F_NO_NOTIFY;
1595 dev->virtqueue[VIRTIO_TXQ]->used->flags = VRING_USED_F_NO_NOTIFY;
1599 * A new virtio-net device is added to a vhost port.
1602 new_device(struct virtio_net *dev)
1604 struct netdev_dpdk *netdev;
1605 bool exists = false;
1607 ovs_mutex_lock(&dpdk_mutex);
1608 /* Add device to the vhost port with the same name as that passed down. */
1609 LIST_FOR_EACH(netdev, list_node, &dpdk_list) {
1610 if (strncmp(dev->ifname, netdev->up.name, IFNAMSIZ) == 0) {
1611 ovs_mutex_lock(&netdev->mutex);
1612 ovsrcu_set(&netdev->virtio_dev, dev);
1613 ovs_mutex_unlock(&netdev->mutex);
1615 dev->flags |= VIRTIO_DEV_RUNNING;
1616 /* Disable notifications. */
1617 set_irq_status(dev);
1621 ovs_mutex_unlock(&dpdk_mutex);
1624 VLOG_INFO("vHost Device '%s' (%ld) can't be added - name not found",
1625 dev->ifname, dev->device_fh);
1630 VLOG_INFO("vHost Device '%s' (%ld) has been added",
1631 dev->ifname, dev->device_fh);
1636 * Remove a virtio-net device from the specific vhost port. Use dev->remove
1637 * flag to stop any more packets from being sent or received to/from a VM and
1638 * ensure all currently queued packets have been sent/received before removing
1642 destroy_device(volatile struct virtio_net *dev)
1644 struct netdev_dpdk *vhost_dev;
1646 ovs_mutex_lock(&dpdk_mutex);
1647 LIST_FOR_EACH (vhost_dev, list_node, &dpdk_list) {
1648 if (netdev_dpdk_get_virtio(vhost_dev) == dev) {
1650 ovs_mutex_lock(&vhost_dev->mutex);
1651 dev->flags &= ~VIRTIO_DEV_RUNNING;
1652 ovsrcu_set(&vhost_dev->virtio_dev, NULL);
1653 ovs_mutex_unlock(&vhost_dev->mutex);
1656 * Wait for other threads to quiesce before
1657 * setting the virtio_dev to NULL.
1659 ovsrcu_synchronize();
1661 * As call to ovsrcu_synchronize() will end the quiescent state,
1662 * put thread back into quiescent state before returning.
1664 ovsrcu_quiesce_start();
1667 ovs_mutex_unlock(&dpdk_mutex);
1669 VLOG_INFO("vHost Device '%s' (%ld) has been removed",
1670 dev->ifname, dev->device_fh);
1674 netdev_dpdk_get_virtio(const struct netdev_dpdk *dev)
1676 return ovsrcu_get(struct virtio_net *, &dev->virtio_dev);
1680 * These callbacks allow virtio-net devices to be added to vhost ports when
1681 * configuration has been fully complete.
1683 static const struct virtio_net_device_ops virtio_net_device_ops =
1685 .new_device = new_device,
1686 .destroy_device = destroy_device,
1690 start_cuse_session_loop(void *dummy OVS_UNUSED)
1692 pthread_detach(pthread_self());
1693 /* Put the cuse thread into quiescent state. */
1694 ovsrcu_quiesce_start();
1695 rte_vhost_driver_session_start();
1700 dpdk_vhost_class_init(void)
1704 rte_vhost_driver_callback_register(&virtio_net_device_ops);
1706 /* Register CUSE device to handle IOCTLs.
1707 * Unless otherwise specified on the vswitchd command line, cuse_dev_name
1708 * is set to vhost-net.
1710 err = rte_vhost_driver_register(cuse_dev_name);
1713 VLOG_ERR("CUSE device setup failure.");
1717 ovs_thread_create("cuse_thread", start_cuse_session_loop, NULL);
1722 dpdk_common_init(void)
1724 unixctl_command_register("netdev-dpdk/set-admin-state",
1725 "[netdev] up|down", 1, 2,
1726 netdev_dpdk_set_admin_state, NULL);
1728 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
1734 dpdk_ring_create(const char dev_name[], unsigned int port_no,
1735 unsigned int *eth_port_id)
1737 struct dpdk_ring *ivshmem;
1741 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
1742 if (ivshmem == NULL) {
1746 /* XXX: Add support for multiquque ring. */
1747 err = snprintf(ring_name, 10, "%s_tx", dev_name);
1752 /* Create single consumer/producer rings, netdev does explicit locking. */
1753 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1754 RING_F_SP_ENQ | RING_F_SC_DEQ);
1755 if (ivshmem->cring_tx == NULL) {
1760 err = snprintf(ring_name, 10, "%s_rx", dev_name);
1765 /* Create single consumer/producer rings, netdev does explicit locking. */
1766 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1767 RING_F_SP_ENQ | RING_F_SC_DEQ);
1768 if (ivshmem->cring_rx == NULL) {
1773 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
1774 &ivshmem->cring_tx, 1, SOCKET0);
1781 ivshmem->user_port_id = port_no;
1782 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
1783 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
1785 *eth_port_id = ivshmem->eth_port_id;
1790 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
1792 struct dpdk_ring *ivshmem;
1793 unsigned int port_no;
1796 /* Names always start with "dpdkr" */
1797 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
1802 /* look through our list to find the device */
1803 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
1804 if (ivshmem->user_port_id == port_no) {
1805 VLOG_INFO("Found dpdk ring device %s:", dev_name);
1806 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
1810 /* Need to create the device rings */
1811 return dpdk_ring_create(dev_name, port_no, eth_port_id);
1815 netdev_dpdk_ring_send(struct netdev *netdev_, int qid,
1816 struct dp_packet **pkts, int cnt, bool may_steal)
1818 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1821 /* When using 'dpdkr' and sending to a DPDK ring, we want to ensure that the
1822 * rss hash field is clear. This is because the same mbuf may be modified by
1823 * the consumer of the ring and return into the datapath without recalculating
1825 for (i = 0; i < cnt; i++) {
1826 dp_packet_set_rss_hash(pkts[i], 0);
1829 netdev_dpdk_send__(netdev, qid, pkts, cnt, may_steal);
1834 netdev_dpdk_ring_construct(struct netdev *netdev)
1836 unsigned int port_no = 0;
1839 if (rte_eal_init_ret) {
1840 return rte_eal_init_ret;
1843 ovs_mutex_lock(&dpdk_mutex);
1845 err = dpdk_ring_open(netdev->name, &port_no);
1850 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
1853 ovs_mutex_unlock(&dpdk_mutex);
1857 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, DESTRUCT, MULTIQ, SEND, \
1858 GET_CARRIER, GET_STATS, GET_FEATURES, GET_STATUS, RXQ_RECV) \
1862 NULL, /* netdev_dpdk_run */ \
1863 NULL, /* netdev_dpdk_wait */ \
1865 netdev_dpdk_alloc, \
1868 netdev_dpdk_dealloc, \
1869 netdev_dpdk_get_config, \
1870 NULL, /* netdev_dpdk_set_config */ \
1871 NULL, /* get_tunnel_config */ \
1872 NULL, /* build header */ \
1873 NULL, /* push header */ \
1874 NULL, /* pop header */ \
1875 netdev_dpdk_get_numa_id, /* get_numa_id */ \
1876 MULTIQ, /* set_multiq */ \
1879 NULL, /* send_wait */ \
1881 netdev_dpdk_set_etheraddr, \
1882 netdev_dpdk_get_etheraddr, \
1883 netdev_dpdk_get_mtu, \
1884 netdev_dpdk_set_mtu, \
1885 netdev_dpdk_get_ifindex, \
1887 netdev_dpdk_get_carrier_resets, \
1888 netdev_dpdk_set_miimon, \
1891 NULL, /* set_advertisements */ \
1893 NULL, /* set_policing */ \
1894 NULL, /* get_qos_types */ \
1895 NULL, /* get_qos_capabilities */ \
1896 NULL, /* get_qos */ \
1897 NULL, /* set_qos */ \
1898 NULL, /* get_queue */ \
1899 NULL, /* set_queue */ \
1900 NULL, /* delete_queue */ \
1901 NULL, /* get_queue_stats */ \
1902 NULL, /* queue_dump_start */ \
1903 NULL, /* queue_dump_next */ \
1904 NULL, /* queue_dump_done */ \
1905 NULL, /* dump_queue_stats */ \
1907 NULL, /* get_in4 */ \
1908 NULL, /* set_in4 */ \
1909 NULL, /* get_in6 */ \
1910 NULL, /* add_router */ \
1911 NULL, /* get_next_hop */ \
1913 NULL, /* arp_lookup */ \
1915 netdev_dpdk_update_flags, \
1917 netdev_dpdk_rxq_alloc, \
1918 netdev_dpdk_rxq_construct, \
1919 netdev_dpdk_rxq_destruct, \
1920 netdev_dpdk_rxq_dealloc, \
1922 NULL, /* rx_wait */ \
1923 NULL, /* rxq_drain */ \
1927 dpdk_init(int argc, char **argv)
1931 char *pragram_name = argv[0];
1933 if (argc < 2 || strcmp(argv[1], "--dpdk"))
1936 /* Remove the --dpdk argument from arg list.*/
1940 /* If the cuse_dev_name parameter has been provided, set 'cuse_dev_name' to
1941 * this string if it meets the correct criteria. Otherwise, set it to the
1942 * default (vhost-net).
1944 if (!strcmp(argv[1], "--cuse_dev_name") &&
1945 (strlen(argv[2]) <= NAME_MAX)) {
1947 cuse_dev_name = strdup(argv[2]);
1949 /* Remove the cuse_dev_name configuration parameters from the argument
1950 * list, so that the correct elements are passed to the DPDK
1951 * initialization function
1954 argv += 2; /* Increment by two to bypass the cuse_dev_name arguments */
1957 VLOG_ERR("User-provided cuse_dev_name in use: /dev/%s", cuse_dev_name);
1959 cuse_dev_name = "vhost-net";
1960 VLOG_INFO("No cuse_dev_name provided - defaulting to /dev/vhost-net");
1963 /* Keep the program name argument as this is needed for call to
1966 argv[0] = pragram_name;
1968 /* Make sure things are initialized ... */
1969 result = rte_eal_init(argc, argv);
1971 ovs_abort(result, "Cannot init EAL");
1974 rte_memzone_dump(stdout);
1975 rte_eal_init_ret = 0;
1977 if (argc > result) {
1978 argv[result] = argv[0];
1981 /* We are called from the main thread here */
1982 RTE_PER_LCORE(_lcore_id) = NON_PMD_CORE_ID;
1984 return result + 1 + base;
1987 static const struct netdev_class dpdk_class =
1991 netdev_dpdk_construct,
1992 netdev_dpdk_destruct,
1993 netdev_dpdk_set_multiq,
1994 netdev_dpdk_eth_send,
1995 netdev_dpdk_get_carrier,
1996 netdev_dpdk_get_stats,
1997 netdev_dpdk_get_features,
1998 netdev_dpdk_get_status,
1999 netdev_dpdk_rxq_recv);
2001 static const struct netdev_class dpdk_ring_class =
2005 netdev_dpdk_ring_construct,
2006 netdev_dpdk_destruct,
2007 netdev_dpdk_set_multiq,
2008 netdev_dpdk_ring_send,
2009 netdev_dpdk_get_carrier,
2010 netdev_dpdk_get_stats,
2011 netdev_dpdk_get_features,
2012 netdev_dpdk_get_status,
2013 netdev_dpdk_rxq_recv);
2015 static const struct netdev_class dpdk_vhost_class =
2018 dpdk_vhost_class_init,
2019 netdev_dpdk_vhost_construct,
2020 netdev_dpdk_vhost_destruct,
2021 netdev_dpdk_vhost_set_multiq,
2022 netdev_dpdk_vhost_send,
2023 netdev_dpdk_vhost_get_carrier,
2024 netdev_dpdk_vhost_get_stats,
2027 netdev_dpdk_vhost_rxq_recv);
2030 netdev_dpdk_register(void)
2032 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
2034 if (rte_eal_init_ret) {
2038 if (ovsthread_once_start(&once)) {
2040 netdev_register_provider(&dpdk_class);
2041 netdev_register_provider(&dpdk_ring_class);
2042 netdev_register_provider(&dpdk_vhost_class);
2043 ovsthread_once_done(&once);
2048 pmd_thread_setaffinity_cpu(unsigned cpu)
2054 CPU_SET(cpu, &cpuset);
2055 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
2057 VLOG_ERR("Thread affinity error %d",err);
2060 /* NON_PMD_CORE_ID is reserved for use by non pmd threads. */
2061 ovs_assert(cpu != NON_PMD_CORE_ID);
2062 RTE_PER_LCORE(_lcore_id) = cpu;
2070 return rte_lcore_id() != NON_PMD_CORE_ID;