2 * Copyright (c) 2014, 2015 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.
30 #include <sys/types.h>
34 #include "dp-packet.h"
35 #include "dpif-netdev.h"
37 #include "netdev-dpdk.h"
38 #include "netdev-provider.h"
39 #include "netdev-vport.h"
41 #include "ofp-print.h"
43 #include "ovs-thread.h"
48 #include "unaligned.h"
51 #include "openvswitch/vlog.h"
53 #include "rte_config.h"
55 #include "rte_virtio_net.h"
57 VLOG_DEFINE_THIS_MODULE(dpdk);
58 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
60 #define DPDK_PORT_WATCHDOG_INTERVAL 5
62 #define OVS_CACHE_LINE_SIZE CACHE_LINE_SIZE
63 #define OVS_VPORT_DPDK "ovs_dpdk"
66 * need to reserve tons of extra space in the mbufs so we can align the
67 * DMA addresses to 4KB.
68 * The minimum mbuf size is limited to avoid scatter behaviour and drop in
69 * performance for standard Ethernet MTU.
71 #define MTU_TO_MAX_LEN(mtu) ((mtu) + ETHER_HDR_LEN + ETHER_CRC_LEN)
72 #define MBUF_SIZE_MTU(mtu) (MTU_TO_MAX_LEN(mtu) \
73 + sizeof(struct dp_packet) \
74 + RTE_PKTMBUF_HEADROOM)
75 #define MBUF_SIZE_DRIVER (2048 \
76 + sizeof (struct rte_mbuf) \
77 + RTE_PKTMBUF_HEADROOM)
78 #define MBUF_SIZE(mtu) MAX(MBUF_SIZE_MTU(mtu), MBUF_SIZE_DRIVER)
80 /* Max and min number of packets in the mempool. OVS tries to allocate a
81 * mempool with MAX_NB_MBUF: if this fails (because the system doesn't have
82 * enough hugepages) we keep halving the number until the allocation succeeds
83 * or we reach MIN_NB_MBUF */
85 #define MAX_NB_MBUF (4096 * 64)
86 #define MIN_NB_MBUF (4096 * 4)
87 #define MP_CACHE_SZ RTE_MEMPOOL_CACHE_MAX_SIZE
89 /* MAX_NB_MBUF can be divided by 2 many times, until MIN_NB_MBUF */
90 BUILD_ASSERT_DECL(MAX_NB_MBUF % ROUND_DOWN_POW2(MAX_NB_MBUF/MIN_NB_MBUF) == 0);
92 /* The smallest possible NB_MBUF that we're going to try should be a multiple
93 * of MP_CACHE_SZ. This is advised by DPDK documentation. */
94 BUILD_ASSERT_DECL((MAX_NB_MBUF / ROUND_DOWN_POW2(MAX_NB_MBUF/MIN_NB_MBUF))
99 #define NIC_PORT_RX_Q_SIZE 2048 /* Size of Physical NIC RX Queue, Max (n+32<=4096)*/
100 #define NIC_PORT_TX_Q_SIZE 2048 /* Size of Physical NIC TX Queue, Max (n+32<=4096)*/
102 static char *cuse_dev_name = NULL; /* Character device cuse_dev_name. */
103 static char *vhost_sock_dir = NULL; /* Location of vhost-user sockets */
106 * Maximum amount of time in micro seconds to try and enqueue to vhost.
108 #define VHOST_ENQ_RETRY_USECS 100
110 static const struct rte_eth_conf port_conf = {
112 .mq_mode = ETH_MQ_RX_RSS,
114 .header_split = 0, /* Header Split disabled */
115 .hw_ip_checksum = 0, /* IP checksum offload disabled */
116 .hw_vlan_filter = 0, /* VLAN filtering disabled */
117 .jumbo_frame = 0, /* Jumbo Frame Support disabled */
123 .rss_hf = ETH_RSS_IP | ETH_RSS_UDP | ETH_RSS_TCP,
127 .mq_mode = ETH_MQ_TX_NONE,
131 enum { MAX_TX_QUEUE_LEN = 384 };
132 enum { DPDK_RING_SIZE = 256 };
133 BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE));
134 enum { DRAIN_TSC = 200000ULL };
141 static int rte_eal_init_ret = ENODEV;
143 static struct ovs_mutex dpdk_mutex = OVS_MUTEX_INITIALIZER;
145 /* Contains all 'struct dpdk_dev's. */
146 static struct ovs_list dpdk_list OVS_GUARDED_BY(dpdk_mutex)
147 = OVS_LIST_INITIALIZER(&dpdk_list);
149 static struct ovs_list dpdk_mp_list OVS_GUARDED_BY(dpdk_mutex)
150 = OVS_LIST_INITIALIZER(&dpdk_mp_list);
152 /* This mutex must be used by non pmd threads when allocating or freeing
153 * mbufs through mempools. Since dpdk_queue_pkts() and dpdk_queue_flush() may
154 * use mempools, a non pmd thread should hold this mutex while calling them */
155 static struct ovs_mutex nonpmd_mempool_mutex = OVS_MUTEX_INITIALIZER;
158 struct rte_mempool *mp;
162 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
165 /* There should be one 'struct dpdk_tx_queue' created for
167 struct dpdk_tx_queue {
168 bool flush_tx; /* Set to true to flush queue everytime */
169 /* pkts are queued. */
171 rte_spinlock_t tx_lock; /* Protects the members and the NIC queue
172 * from concurrent access. It is used only
173 * if the queue is shared among different
174 * pmd threads (see 'txq_needs_locking'). */
176 struct rte_mbuf *burst_pkts[MAX_TX_QUEUE_LEN];
179 /* dpdk has no way to remove dpdk ring ethernet devices
180 so we have to keep them around once they've been created
183 static struct ovs_list dpdk_ring_list OVS_GUARDED_BY(dpdk_mutex)
184 = OVS_LIST_INITIALIZER(&dpdk_ring_list);
187 /* For the client rings */
188 struct rte_ring *cring_tx;
189 struct rte_ring *cring_rx;
190 int user_port_id; /* User given port no, parsed from port name */
191 int eth_port_id; /* ethernet device port id */
192 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
199 enum dpdk_dev_type type;
201 struct dpdk_tx_queue *tx_q;
203 struct ovs_mutex mutex OVS_ACQ_AFTER(dpdk_mutex);
205 struct dpdk_mp *dpdk_mp;
209 struct netdev_stats stats;
211 rte_spinlock_t stats_lock;
213 struct eth_addr hwaddr;
214 enum netdev_flags flags;
216 struct rte_eth_link link;
219 /* The user might request more txqs than the NIC has. We remap those
220 * ('up.n_txq') on these ('real_n_txq').
221 * If the numbers match, 'txq_needs_locking' is false, otherwise it is
222 * true and we will take a spinlock on transmission */
224 bool txq_needs_locking;
226 /* Spinlock for vhost transmission. Other DPDK devices use spinlocks in
228 rte_spinlock_t vhost_tx_lock;
230 /* virtio-net structure for vhost device */
231 OVSRCU_TYPE(struct virtio_net *) virtio_dev;
233 /* Identifier used to distinguish vhost devices from each other */
234 char vhost_id[PATH_MAX];
237 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
240 struct netdev_rxq_dpdk {
241 struct netdev_rxq up;
245 static bool thread_is_pmd(void);
247 static int netdev_dpdk_construct(struct netdev *);
249 struct virtio_net * netdev_dpdk_get_virtio(const struct netdev_dpdk *dev);
252 is_dpdk_class(const struct netdev_class *class)
254 return class->construct == netdev_dpdk_construct;
257 /* XXX: use dpdk malloc for entire OVS. in fact huge page should be used
258 * for all other segments data, bss and text. */
261 dpdk_rte_mzalloc(size_t sz)
265 ptr = rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE);
272 /* XXX this function should be called only by pmd threads (or by non pmd
273 * threads holding the nonpmd_mempool_mutex) */
275 free_dpdk_buf(struct dp_packet *p)
277 struct rte_mbuf *pkt = (struct rte_mbuf *) p;
279 rte_pktmbuf_free_seg(pkt);
283 __rte_pktmbuf_init(struct rte_mempool *mp,
284 void *opaque_arg OVS_UNUSED,
286 unsigned i OVS_UNUSED)
288 struct rte_mbuf *m = _m;
289 uint32_t buf_len = mp->elt_size - sizeof(struct dp_packet);
291 RTE_MBUF_ASSERT(mp->elt_size >= sizeof(struct dp_packet));
293 memset(m, 0, mp->elt_size);
295 /* start of buffer is just after mbuf structure */
296 m->buf_addr = (char *)m + sizeof(struct dp_packet);
297 m->buf_physaddr = rte_mempool_virt2phy(mp, m) +
298 sizeof(struct dp_packet);
299 m->buf_len = (uint16_t)buf_len;
301 /* keep some headroom between start of buffer and data */
302 m->data_off = RTE_MIN(RTE_PKTMBUF_HEADROOM, m->buf_len);
304 /* init some constant fields */
311 ovs_rte_pktmbuf_init(struct rte_mempool *mp,
312 void *opaque_arg OVS_UNUSED,
314 unsigned i OVS_UNUSED)
316 struct rte_mbuf *m = _m;
318 __rte_pktmbuf_init(mp, opaque_arg, _m, i);
320 dp_packet_init_dpdk((struct dp_packet *) m, m->buf_len);
323 static struct dpdk_mp *
324 dpdk_mp_get(int socket_id, int mtu) OVS_REQUIRES(dpdk_mutex)
326 struct dpdk_mp *dmp = NULL;
327 char mp_name[RTE_MEMPOOL_NAMESIZE];
330 LIST_FOR_EACH (dmp, list_node, &dpdk_mp_list) {
331 if (dmp->socket_id == socket_id && dmp->mtu == mtu) {
337 dmp = dpdk_rte_mzalloc(sizeof *dmp);
338 dmp->socket_id = socket_id;
342 mp_size = MAX_NB_MBUF;
344 if (snprintf(mp_name, RTE_MEMPOOL_NAMESIZE, "ovs_mp_%d_%d_%u",
345 dmp->mtu, dmp->socket_id, mp_size) < 0) {
349 dmp->mp = rte_mempool_create(mp_name, mp_size, MBUF_SIZE(mtu),
351 sizeof(struct rte_pktmbuf_pool_private),
352 rte_pktmbuf_pool_init, NULL,
353 ovs_rte_pktmbuf_init, NULL,
355 } while (!dmp->mp && rte_errno == ENOMEM && (mp_size /= 2) >= MIN_NB_MBUF);
357 if (dmp->mp == NULL) {
360 VLOG_DBG("Allocated \"%s\" mempool with %u mbufs", mp_name, mp_size );
363 list_push_back(&dpdk_mp_list, &dmp->list_node);
368 dpdk_mp_put(struct dpdk_mp *dmp)
376 ovs_assert(dmp->refcount >= 0);
379 /* I could not find any API to destroy mp. */
380 if (dmp->refcount == 0) {
381 list_delete(dmp->list_node);
382 /* destroy mp-pool. */
388 check_link_status(struct netdev_dpdk *dev)
390 struct rte_eth_link link;
392 rte_eth_link_get_nowait(dev->port_id, &link);
394 if (dev->link.link_status != link.link_status) {
395 netdev_change_seq_changed(&dev->up);
397 dev->link_reset_cnt++;
399 if (dev->link.link_status) {
400 VLOG_DBG_RL(&rl, "Port %d Link Up - speed %u Mbps - %s",
401 dev->port_id, (unsigned)dev->link.link_speed,
402 (dev->link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
403 ("full-duplex") : ("half-duplex"));
405 VLOG_DBG_RL(&rl, "Port %d Link Down", dev->port_id);
411 dpdk_watchdog(void *dummy OVS_UNUSED)
413 struct netdev_dpdk *dev;
415 pthread_detach(pthread_self());
418 ovs_mutex_lock(&dpdk_mutex);
419 LIST_FOR_EACH (dev, list_node, &dpdk_list) {
420 ovs_mutex_lock(&dev->mutex);
421 check_link_status(dev);
422 ovs_mutex_unlock(&dev->mutex);
424 ovs_mutex_unlock(&dpdk_mutex);
425 xsleep(DPDK_PORT_WATCHDOG_INTERVAL);
432 dpdk_eth_dev_queue_setup(struct netdev_dpdk *dev, int n_rxq, int n_txq)
437 /* A device may report more queues than it makes available (this has
438 * been observed for Intel xl710, which reserves some of them for
439 * SRIOV): rte_eth_*_queue_setup will fail if a queue is not
440 * available. When this happens we can retry the configuration
441 * and request less queues */
442 while (n_rxq && n_txq) {
444 VLOG_INFO("Retrying setup with (rxq:%d txq:%d)", n_rxq, n_txq);
447 diag = rte_eth_dev_configure(dev->port_id, n_rxq, n_txq, &port_conf);
452 for (i = 0; i < n_txq; i++) {
453 diag = rte_eth_tx_queue_setup(dev->port_id, i, NIC_PORT_TX_Q_SIZE,
454 dev->socket_id, NULL);
456 VLOG_INFO("Interface %s txq(%d) setup error: %s",
457 dev->up.name, i, rte_strerror(-diag));
463 /* Retry with less tx queues */
468 for (i = 0; i < n_rxq; i++) {
469 diag = rte_eth_rx_queue_setup(dev->port_id, i, NIC_PORT_RX_Q_SIZE,
470 dev->socket_id, NULL,
473 VLOG_INFO("Interface %s rxq(%d) setup error: %s",
474 dev->up.name, i, rte_strerror(-diag));
480 /* Retry with less rx queues */
485 dev->up.n_rxq = n_rxq;
486 dev->real_n_txq = n_txq;
496 dpdk_eth_dev_init(struct netdev_dpdk *dev) OVS_REQUIRES(dpdk_mutex)
498 struct rte_pktmbuf_pool_private *mbp_priv;
499 struct rte_eth_dev_info info;
500 struct ether_addr eth_addr;
504 if (dev->port_id < 0 || dev->port_id >= rte_eth_dev_count()) {
508 rte_eth_dev_info_get(dev->port_id, &info);
510 n_rxq = MIN(info.max_rx_queues, dev->up.n_rxq);
511 n_txq = MIN(info.max_tx_queues, dev->up.n_txq);
513 diag = dpdk_eth_dev_queue_setup(dev, n_rxq, n_txq);
515 VLOG_ERR("Interface %s(rxq:%d txq:%d) configure error: %s",
516 dev->up.name, n_rxq, n_txq, rte_strerror(-diag));
520 diag = rte_eth_dev_start(dev->port_id);
522 VLOG_ERR("Interface %s start error: %s", dev->up.name,
523 rte_strerror(-diag));
527 rte_eth_promiscuous_enable(dev->port_id);
528 rte_eth_allmulticast_enable(dev->port_id);
530 memset(ð_addr, 0x0, sizeof(eth_addr));
531 rte_eth_macaddr_get(dev->port_id, ð_addr);
532 VLOG_INFO_RL(&rl, "Port %d: "ETH_ADDR_FMT"",
533 dev->port_id, ETH_ADDR_BYTES_ARGS(eth_addr.addr_bytes));
535 memcpy(dev->hwaddr.ea, eth_addr.addr_bytes, ETH_ADDR_LEN);
536 rte_eth_link_get_nowait(dev->port_id, &dev->link);
538 mbp_priv = rte_mempool_get_priv(dev->dpdk_mp->mp);
539 dev->buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
541 dev->flags = NETDEV_UP | NETDEV_PROMISC;
545 static struct netdev_dpdk *
546 netdev_dpdk_cast(const struct netdev *netdev)
548 return CONTAINER_OF(netdev, struct netdev_dpdk, up);
551 static struct netdev *
552 netdev_dpdk_alloc(void)
554 struct netdev_dpdk *netdev = dpdk_rte_mzalloc(sizeof *netdev);
559 netdev_dpdk_alloc_txq(struct netdev_dpdk *netdev, unsigned int n_txqs)
563 netdev->tx_q = dpdk_rte_mzalloc(n_txqs * sizeof *netdev->tx_q);
564 for (i = 0; i < n_txqs; i++) {
565 int numa_id = ovs_numa_get_numa_id(i);
567 if (!netdev->txq_needs_locking) {
568 /* Each index is considered as a cpu core id, since there should
569 * be one tx queue for each cpu core. If the corresponding core
570 * is not on the same numa node as 'netdev', flags the
572 netdev->tx_q[i].flush_tx = netdev->socket_id == numa_id;
574 /* Queues are shared among CPUs. Always flush */
575 netdev->tx_q[i].flush_tx = true;
577 rte_spinlock_init(&netdev->tx_q[i].tx_lock);
582 netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no,
583 enum dpdk_dev_type type)
584 OVS_REQUIRES(dpdk_mutex)
586 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
590 ovs_mutex_init(&netdev->mutex);
591 ovs_mutex_lock(&netdev->mutex);
593 rte_spinlock_init(&netdev->stats_lock);
595 /* If the 'sid' is negative, it means that the kernel fails
596 * to obtain the pci numa info. In that situation, always
598 if (type == DPDK_DEV_ETH) {
599 sid = rte_eth_dev_socket_id(port_no);
601 sid = rte_lcore_to_socket_id(rte_get_master_lcore());
604 netdev->socket_id = sid < 0 ? SOCKET0 : sid;
605 netdev->port_id = port_no;
608 netdev->mtu = ETHER_MTU;
609 netdev->max_packet_len = MTU_TO_MAX_LEN(netdev->mtu);
611 netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, netdev->mtu);
612 if (!netdev->dpdk_mp) {
617 netdev_->n_txq = NR_QUEUE;
618 netdev_->n_rxq = NR_QUEUE;
619 netdev->real_n_txq = NR_QUEUE;
621 if (type == DPDK_DEV_ETH) {
622 netdev_dpdk_alloc_txq(netdev, NR_QUEUE);
623 err = dpdk_eth_dev_init(netdev);
629 list_push_back(&dpdk_list, &netdev->list_node);
633 rte_free(netdev->tx_q);
635 ovs_mutex_unlock(&netdev->mutex);
640 dpdk_dev_parse_name(const char dev_name[], const char prefix[],
641 unsigned int *port_no)
645 if (strncmp(dev_name, prefix, strlen(prefix))) {
649 cport = dev_name + strlen(prefix);
650 *port_no = strtol(cport, NULL, 0); /* string must be null terminated */
655 vhost_construct_helper(struct netdev *netdev_) OVS_REQUIRES(dpdk_mutex)
657 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
659 if (rte_eal_init_ret) {
660 return rte_eal_init_ret;
663 rte_spinlock_init(&netdev->vhost_tx_lock);
664 return netdev_dpdk_init(netdev_, -1, DPDK_DEV_VHOST);
668 netdev_dpdk_vhost_cuse_construct(struct netdev *netdev_)
670 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
673 ovs_mutex_lock(&dpdk_mutex);
674 strncpy(netdev->vhost_id, netdev->up.name, sizeof(netdev->vhost_id));
675 err = vhost_construct_helper(netdev_);
676 ovs_mutex_unlock(&dpdk_mutex);
681 netdev_dpdk_vhost_user_construct(struct netdev *netdev_)
683 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
686 ovs_mutex_lock(&dpdk_mutex);
687 /* Take the name of the vhost-user port and append it to the location where
688 * the socket is to be created, then register the socket.
690 snprintf(netdev->vhost_id, sizeof(netdev->vhost_id), "%s/%s",
691 vhost_sock_dir, netdev_->name);
692 err = rte_vhost_driver_register(netdev->vhost_id);
694 VLOG_ERR("vhost-user socket device setup failure for socket %s\n",
697 VLOG_INFO("Socket %s created for vhost-user port %s\n", netdev->vhost_id, netdev_->name);
698 err = vhost_construct_helper(netdev_);
699 ovs_mutex_unlock(&dpdk_mutex);
704 netdev_dpdk_construct(struct netdev *netdev)
706 unsigned int port_no;
709 if (rte_eal_init_ret) {
710 return rte_eal_init_ret;
713 /* Names always start with "dpdk" */
714 err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no);
719 ovs_mutex_lock(&dpdk_mutex);
720 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
721 ovs_mutex_unlock(&dpdk_mutex);
726 netdev_dpdk_destruct(struct netdev *netdev_)
728 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
730 ovs_mutex_lock(&dev->mutex);
731 rte_eth_dev_stop(dev->port_id);
732 ovs_mutex_unlock(&dev->mutex);
734 ovs_mutex_lock(&dpdk_mutex);
736 list_remove(&dev->list_node);
737 dpdk_mp_put(dev->dpdk_mp);
738 ovs_mutex_unlock(&dpdk_mutex);
742 netdev_dpdk_vhost_destruct(struct netdev *netdev_)
744 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
746 /* Can't remove a port while a guest is attached to it. */
747 if (netdev_dpdk_get_virtio(dev) != NULL) {
748 VLOG_ERR("Can not remove port, vhost device still attached");
752 if (rte_vhost_driver_unregister(dev->vhost_id)) {
753 VLOG_ERR("Unable to remove vhost-user socket %s", dev->vhost_id);
756 ovs_mutex_lock(&dpdk_mutex);
757 list_remove(&dev->list_node);
758 dpdk_mp_put(dev->dpdk_mp);
759 ovs_mutex_unlock(&dpdk_mutex);
763 netdev_dpdk_dealloc(struct netdev *netdev_)
765 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
771 netdev_dpdk_get_config(const struct netdev *netdev_, struct smap *args)
773 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
775 ovs_mutex_lock(&dev->mutex);
777 smap_add_format(args, "configured_rx_queues", "%d", netdev_->n_rxq);
778 smap_add_format(args, "requested_tx_queues", "%d", netdev_->n_txq);
779 smap_add_format(args, "configured_tx_queues", "%d", dev->real_n_txq);
780 ovs_mutex_unlock(&dev->mutex);
786 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
788 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
790 return netdev->socket_id;
793 /* Sets the number of tx queues and rx queues for the dpdk interface.
794 * If the configuration fails, do not try restoring its old configuration
795 * and just returns the error. */
797 netdev_dpdk_set_multiq(struct netdev *netdev_, unsigned int n_txq,
800 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
802 int old_rxq, old_txq;
804 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
808 ovs_mutex_lock(&dpdk_mutex);
809 ovs_mutex_lock(&netdev->mutex);
811 rte_eth_dev_stop(netdev->port_id);
813 old_txq = netdev->up.n_txq;
814 old_rxq = netdev->up.n_rxq;
815 netdev->up.n_txq = n_txq;
816 netdev->up.n_rxq = n_rxq;
818 rte_free(netdev->tx_q);
819 err = dpdk_eth_dev_init(netdev);
820 netdev_dpdk_alloc_txq(netdev, netdev->real_n_txq);
822 /* If there has been an error, it means that the requested queues
823 * have not been created. Restore the old numbers. */
824 netdev->up.n_txq = old_txq;
825 netdev->up.n_rxq = old_rxq;
828 netdev->txq_needs_locking = netdev->real_n_txq != netdev->up.n_txq;
830 ovs_mutex_unlock(&netdev->mutex);
831 ovs_mutex_unlock(&dpdk_mutex);
837 netdev_dpdk_vhost_set_multiq(struct netdev *netdev_, unsigned int n_txq,
840 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
843 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
847 ovs_mutex_lock(&dpdk_mutex);
848 ovs_mutex_lock(&netdev->mutex);
850 netdev->up.n_txq = n_txq;
851 netdev->real_n_txq = 1;
852 netdev->up.n_rxq = 1;
854 ovs_mutex_unlock(&netdev->mutex);
855 ovs_mutex_unlock(&dpdk_mutex);
860 static struct netdev_rxq *
861 netdev_dpdk_rxq_alloc(void)
863 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
868 static struct netdev_rxq_dpdk *
869 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
871 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
875 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
877 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
878 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
880 ovs_mutex_lock(&netdev->mutex);
881 rx->port_id = netdev->port_id;
882 ovs_mutex_unlock(&netdev->mutex);
888 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
893 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
895 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
901 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
903 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
906 while (nb_tx != txq->count) {
909 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
918 if (OVS_UNLIKELY(nb_tx != txq->count)) {
919 /* free buffers, which we couldn't transmit, one at a time (each
920 * packet could come from a different mempool) */
923 for (i = nb_tx; i < txq->count; i++) {
924 rte_pktmbuf_free_seg(txq->burst_pkts[i]);
926 rte_spinlock_lock(&dev->stats_lock);
927 dev->stats.tx_dropped += txq->count-nb_tx;
928 rte_spinlock_unlock(&dev->stats_lock);
932 txq->tsc = rte_get_timer_cycles();
936 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
938 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
940 if (txq->count == 0) {
943 dpdk_queue_flush__(dev, qid);
947 is_vhost_running(struct virtio_net *dev)
949 return (dev != NULL && (dev->flags & VIRTIO_DEV_RUNNING));
953 netdev_dpdk_vhost_update_rx_counters(struct netdev_stats *stats,
954 struct dp_packet **packets, int count)
957 struct dp_packet *packet;
959 stats->rx_packets += count;
960 for (i = 0; i < count; i++) {
963 if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) {
964 /* This only protects the following multicast counting from
965 * too short packets, but it does not stop the packet from
966 * further processing. */
968 stats->rx_length_errors++;
972 struct eth_header *eh = (struct eth_header *) dp_packet_data(packet);
973 if (OVS_UNLIKELY(eth_addr_is_multicast(eh->eth_dst))) {
977 stats->rx_bytes += dp_packet_size(packet);
982 * The receive path for the vhost port is the TX path out from guest.
985 netdev_dpdk_vhost_rxq_recv(struct netdev_rxq *rxq_,
986 struct dp_packet **packets, int *c)
988 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
989 struct netdev *netdev = rx->up.netdev;
990 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
991 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
995 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
999 nb_rx = rte_vhost_dequeue_burst(virtio_dev, qid,
1000 vhost_dev->dpdk_mp->mp,
1001 (struct rte_mbuf **)packets,
1007 rte_spinlock_lock(&vhost_dev->stats_lock);
1008 netdev_dpdk_vhost_update_rx_counters(&vhost_dev->stats, packets, nb_rx);
1009 rte_spinlock_unlock(&vhost_dev->stats_lock);
1016 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dp_packet **packets,
1019 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
1020 struct netdev *netdev = rx->up.netdev;
1021 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1024 /* There is only one tx queue for this core. Do not flush other
1026 * Do not flush tx queue which is shared among CPUs
1027 * since it is always flushed */
1028 if (rxq_->queue_id == rte_lcore_id() &&
1029 OVS_LIKELY(!dev->txq_needs_locking)) {
1030 dpdk_queue_flush(dev, rxq_->queue_id);
1033 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
1034 (struct rte_mbuf **) packets,
1046 netdev_dpdk_vhost_update_tx_counters(struct netdev_stats *stats,
1047 struct dp_packet **packets,
1052 int sent = attempted - dropped;
1054 stats->tx_packets += sent;
1055 stats->tx_dropped += dropped;
1057 for (i = 0; i < sent; i++) {
1058 stats->tx_bytes += dp_packet_size(packets[i]);
1063 __netdev_dpdk_vhost_send(struct netdev *netdev, struct dp_packet **pkts,
1064 int cnt, bool may_steal)
1066 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
1067 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
1068 struct rte_mbuf **cur_pkts = (struct rte_mbuf **) pkts;
1069 unsigned int total_pkts = cnt;
1072 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
1073 rte_spinlock_lock(&vhost_dev->stats_lock);
1074 vhost_dev->stats.tx_dropped+= cnt;
1075 rte_spinlock_unlock(&vhost_dev->stats_lock);
1079 /* There is vHost TX single queue, So we need to lock it for TX. */
1080 rte_spinlock_lock(&vhost_dev->vhost_tx_lock);
1083 unsigned int tx_pkts;
1085 tx_pkts = rte_vhost_enqueue_burst(virtio_dev, VIRTIO_RXQ,
1087 if (OVS_LIKELY(tx_pkts)) {
1088 /* Packets have been sent.*/
1090 /* Prepare for possible next iteration.*/
1091 cur_pkts = &cur_pkts[tx_pkts];
1093 uint64_t timeout = VHOST_ENQ_RETRY_USECS * rte_get_timer_hz() / 1E6;
1094 unsigned int expired = 0;
1097 start = rte_get_timer_cycles();
1101 * Unable to enqueue packets to vhost interface.
1102 * Check available entries before retrying.
1104 while (!rte_vring_available_entries(virtio_dev, VIRTIO_RXQ)) {
1105 if (OVS_UNLIKELY((rte_get_timer_cycles() - start) > timeout)) {
1111 /* break out of main loop. */
1116 rte_spinlock_unlock(&vhost_dev->vhost_tx_lock);
1118 rte_spinlock_lock(&vhost_dev->stats_lock);
1119 netdev_dpdk_vhost_update_tx_counters(&vhost_dev->stats, pkts, total_pkts,
1121 rte_spinlock_unlock(&vhost_dev->stats_lock);
1127 for (i = 0; i < total_pkts; i++) {
1128 dp_packet_delete(pkts[i]);
1134 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
1135 struct rte_mbuf **pkts, int cnt)
1137 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
1143 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
1144 int tocopy = MIN(freeslots, cnt-i);
1146 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
1147 tocopy * sizeof (struct rte_mbuf *));
1149 txq->count += tocopy;
1152 if (txq->count == MAX_TX_QUEUE_LEN || txq->flush_tx) {
1153 dpdk_queue_flush__(dev, qid);
1155 diff_tsc = rte_get_timer_cycles() - txq->tsc;
1156 if (diff_tsc >= DRAIN_TSC) {
1157 dpdk_queue_flush__(dev, qid);
1162 /* Tx function. Transmit packets indefinitely */
1164 dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet **pkts,
1166 OVS_NO_THREAD_SAFETY_ANALYSIS
1168 #if !defined(__CHECKER__) && !defined(_WIN32)
1169 const size_t PKT_ARRAY_SIZE = cnt;
1171 /* Sparse or MSVC doesn't like variable length array. */
1172 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
1174 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1175 struct rte_mbuf *mbufs[PKT_ARRAY_SIZE];
1180 /* If we are on a non pmd thread we have to use the mempool mutex, because
1181 * every non pmd thread shares the same mempool cache */
1183 if (!thread_is_pmd()) {
1184 ovs_mutex_lock(&nonpmd_mempool_mutex);
1187 for (i = 0; i < cnt; i++) {
1188 int size = dp_packet_size(pkts[i]);
1190 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1191 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1192 (int)size , dev->max_packet_len);
1198 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
1200 if (!mbufs[newcnt]) {
1205 /* We have to do a copy for now */
1206 memcpy(rte_pktmbuf_mtod(mbufs[newcnt], void *), dp_packet_data(pkts[i]), size);
1208 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
1209 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
1214 if (OVS_UNLIKELY(dropped)) {
1215 rte_spinlock_lock(&dev->stats_lock);
1216 dev->stats.tx_dropped += dropped;
1217 rte_spinlock_unlock(&dev->stats_lock);
1220 if (dev->type == DPDK_DEV_VHOST) {
1221 __netdev_dpdk_vhost_send(netdev, (struct dp_packet **) mbufs, newcnt, true);
1223 dpdk_queue_pkts(dev, qid, mbufs, newcnt);
1224 dpdk_queue_flush(dev, qid);
1227 if (!thread_is_pmd()) {
1228 ovs_mutex_unlock(&nonpmd_mempool_mutex);
1233 netdev_dpdk_vhost_send(struct netdev *netdev, int qid OVS_UNUSED, struct dp_packet **pkts,
1234 int cnt, bool may_steal)
1236 if (OVS_UNLIKELY(pkts[0]->source != DPBUF_DPDK)) {
1239 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1241 for (i = 0; i < cnt; i++) {
1242 dp_packet_delete(pkts[i]);
1246 __netdev_dpdk_vhost_send(netdev, pkts, cnt, may_steal);
1252 netdev_dpdk_send__(struct netdev_dpdk *dev, int qid,
1253 struct dp_packet **pkts, int cnt, bool may_steal)
1257 if (OVS_UNLIKELY(dev->txq_needs_locking)) {
1258 qid = qid % dev->real_n_txq;
1259 rte_spinlock_lock(&dev->tx_q[qid].tx_lock);
1262 if (OVS_UNLIKELY(!may_steal ||
1263 pkts[0]->source != DPBUF_DPDK)) {
1264 struct netdev *netdev = &dev->up;
1266 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1269 for (i = 0; i < cnt; i++) {
1270 dp_packet_delete(pkts[i]);
1274 int next_tx_idx = 0;
1277 for (i = 0; i < cnt; i++) {
1278 int size = dp_packet_size(pkts[i]);
1280 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1281 if (next_tx_idx != i) {
1282 dpdk_queue_pkts(dev, qid,
1283 (struct rte_mbuf **)&pkts[next_tx_idx],
1287 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1288 (int)size , dev->max_packet_len);
1290 dp_packet_delete(pkts[i]);
1292 next_tx_idx = i + 1;
1295 if (next_tx_idx != cnt) {
1296 dpdk_queue_pkts(dev, qid,
1297 (struct rte_mbuf **)&pkts[next_tx_idx],
1301 if (OVS_UNLIKELY(dropped)) {
1302 rte_spinlock_lock(&dev->stats_lock);
1303 dev->stats.tx_dropped += dropped;
1304 rte_spinlock_unlock(&dev->stats_lock);
1308 if (OVS_UNLIKELY(dev->txq_needs_locking)) {
1309 rte_spinlock_unlock(&dev->tx_q[qid].tx_lock);
1314 netdev_dpdk_eth_send(struct netdev *netdev, int qid,
1315 struct dp_packet **pkts, int cnt, bool may_steal)
1317 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1319 netdev_dpdk_send__(dev, qid, pkts, cnt, may_steal);
1324 netdev_dpdk_set_etheraddr(struct netdev *netdev, const struct eth_addr mac)
1326 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1328 ovs_mutex_lock(&dev->mutex);
1329 if (!eth_addr_equals(dev->hwaddr, mac)) {
1331 netdev_change_seq_changed(netdev);
1333 ovs_mutex_unlock(&dev->mutex);
1339 netdev_dpdk_get_etheraddr(const struct netdev *netdev, struct eth_addr *mac)
1341 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1343 ovs_mutex_lock(&dev->mutex);
1345 ovs_mutex_unlock(&dev->mutex);
1351 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
1353 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1355 ovs_mutex_lock(&dev->mutex);
1357 ovs_mutex_unlock(&dev->mutex);
1363 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
1365 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1367 struct dpdk_mp *old_mp;
1370 ovs_mutex_lock(&dpdk_mutex);
1371 ovs_mutex_lock(&dev->mutex);
1372 if (dev->mtu == mtu) {
1377 mp = dpdk_mp_get(dev->socket_id, dev->mtu);
1383 rte_eth_dev_stop(dev->port_id);
1386 old_mp = dev->dpdk_mp;
1389 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1391 err = dpdk_eth_dev_init(dev);
1395 dev->dpdk_mp = old_mp;
1396 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1397 dpdk_eth_dev_init(dev);
1401 dpdk_mp_put(old_mp);
1402 netdev_change_seq_changed(netdev);
1404 ovs_mutex_unlock(&dev->mutex);
1405 ovs_mutex_unlock(&dpdk_mutex);
1410 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
1413 netdev_dpdk_vhost_get_stats(const struct netdev *netdev,
1414 struct netdev_stats *stats)
1416 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1418 ovs_mutex_lock(&dev->mutex);
1419 memset(stats, 0, sizeof(*stats));
1420 /* Unsupported Stats */
1421 stats->collisions = UINT64_MAX;
1422 stats->rx_crc_errors = UINT64_MAX;
1423 stats->rx_fifo_errors = UINT64_MAX;
1424 stats->rx_frame_errors = UINT64_MAX;
1425 stats->rx_missed_errors = UINT64_MAX;
1426 stats->rx_over_errors = UINT64_MAX;
1427 stats->tx_aborted_errors = UINT64_MAX;
1428 stats->tx_carrier_errors = UINT64_MAX;
1429 stats->tx_errors = UINT64_MAX;
1430 stats->tx_fifo_errors = UINT64_MAX;
1431 stats->tx_heartbeat_errors = UINT64_MAX;
1432 stats->tx_window_errors = UINT64_MAX;
1433 stats->rx_dropped += UINT64_MAX;
1435 rte_spinlock_lock(&dev->stats_lock);
1436 /* Supported Stats */
1437 stats->rx_packets += dev->stats.rx_packets;
1438 stats->tx_packets += dev->stats.tx_packets;
1439 stats->tx_dropped += dev->stats.tx_dropped;
1440 stats->multicast = dev->stats.multicast;
1441 stats->rx_bytes = dev->stats.rx_bytes;
1442 stats->tx_bytes = dev->stats.tx_bytes;
1443 stats->rx_errors = dev->stats.rx_errors;
1444 stats->rx_length_errors = dev->stats.rx_length_errors;
1445 rte_spinlock_unlock(&dev->stats_lock);
1447 ovs_mutex_unlock(&dev->mutex);
1453 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
1455 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1456 struct rte_eth_stats rte_stats;
1459 netdev_dpdk_get_carrier(netdev, &gg);
1460 ovs_mutex_lock(&dev->mutex);
1461 rte_eth_stats_get(dev->port_id, &rte_stats);
1463 memset(stats, 0, sizeof(*stats));
1465 stats->rx_packets = rte_stats.ipackets;
1466 stats->tx_packets = rte_stats.opackets;
1467 stats->rx_bytes = rte_stats.ibytes;
1468 stats->tx_bytes = rte_stats.obytes;
1469 /* DPDK counts imissed as errors, but count them here as dropped instead */
1470 stats->rx_errors = rte_stats.ierrors - rte_stats.imissed;
1471 stats->tx_errors = rte_stats.oerrors;
1472 stats->multicast = rte_stats.imcasts;
1474 rte_spinlock_lock(&dev->stats_lock);
1475 stats->tx_dropped = dev->stats.tx_dropped;
1476 rte_spinlock_unlock(&dev->stats_lock);
1478 /* These are the available DPDK counters for packets not received due to
1479 * local resource constraints in DPDK and NIC respectively. */
1480 stats->rx_dropped = rte_stats.rx_nombuf + rte_stats.imissed;
1481 stats->collisions = UINT64_MAX;
1483 stats->rx_length_errors = rte_stats.ibadlen;
1484 stats->rx_over_errors = UINT64_MAX;
1485 stats->rx_crc_errors = rte_stats.ibadcrc;
1486 stats->rx_frame_errors = UINT64_MAX;
1487 stats->rx_fifo_errors = UINT64_MAX;
1488 stats->rx_missed_errors = rte_stats.imissed;
1490 stats->tx_aborted_errors = UINT64_MAX;
1491 stats->tx_carrier_errors = UINT64_MAX;
1492 stats->tx_fifo_errors = UINT64_MAX;
1493 stats->tx_heartbeat_errors = UINT64_MAX;
1494 stats->tx_window_errors = UINT64_MAX;
1496 ovs_mutex_unlock(&dev->mutex);
1502 netdev_dpdk_get_features(const struct netdev *netdev_,
1503 enum netdev_features *current,
1504 enum netdev_features *advertised OVS_UNUSED,
1505 enum netdev_features *supported OVS_UNUSED,
1506 enum netdev_features *peer OVS_UNUSED)
1508 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1509 struct rte_eth_link link;
1511 ovs_mutex_lock(&dev->mutex);
1513 ovs_mutex_unlock(&dev->mutex);
1515 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
1516 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
1517 *current = NETDEV_F_AUTONEG;
1519 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
1520 if (link.link_speed == ETH_LINK_SPEED_10) {
1521 *current = NETDEV_F_10MB_HD;
1523 if (link.link_speed == ETH_LINK_SPEED_100) {
1524 *current = NETDEV_F_100MB_HD;
1526 if (link.link_speed == ETH_LINK_SPEED_1000) {
1527 *current = NETDEV_F_1GB_HD;
1529 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
1530 if (link.link_speed == ETH_LINK_SPEED_10) {
1531 *current = NETDEV_F_10MB_FD;
1533 if (link.link_speed == ETH_LINK_SPEED_100) {
1534 *current = NETDEV_F_100MB_FD;
1536 if (link.link_speed == ETH_LINK_SPEED_1000) {
1537 *current = NETDEV_F_1GB_FD;
1539 if (link.link_speed == ETH_LINK_SPEED_10000) {
1540 *current = NETDEV_F_10GB_FD;
1548 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1550 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1553 ovs_mutex_lock(&dev->mutex);
1554 ifindex = dev->port_id;
1555 ovs_mutex_unlock(&dev->mutex);
1561 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1563 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1565 ovs_mutex_lock(&dev->mutex);
1566 check_link_status(dev);
1567 *carrier = dev->link.link_status;
1569 ovs_mutex_unlock(&dev->mutex);
1575 netdev_dpdk_vhost_get_carrier(const struct netdev *netdev_, bool *carrier)
1577 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1578 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(dev);
1580 ovs_mutex_lock(&dev->mutex);
1582 if (is_vhost_running(virtio_dev)) {
1588 ovs_mutex_unlock(&dev->mutex);
1593 static long long int
1594 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1596 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1597 long long int carrier_resets;
1599 ovs_mutex_lock(&dev->mutex);
1600 carrier_resets = dev->link_reset_cnt;
1601 ovs_mutex_unlock(&dev->mutex);
1603 return carrier_resets;
1607 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1608 long long int interval OVS_UNUSED)
1614 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1615 enum netdev_flags off, enum netdev_flags on,
1616 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1620 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1624 *old_flagsp = dev->flags;
1628 if (dev->flags == *old_flagsp) {
1632 if (dev->type == DPDK_DEV_ETH) {
1633 if (dev->flags & NETDEV_UP) {
1634 err = rte_eth_dev_start(dev->port_id);
1639 if (dev->flags & NETDEV_PROMISC) {
1640 rte_eth_promiscuous_enable(dev->port_id);
1643 if (!(dev->flags & NETDEV_UP)) {
1644 rte_eth_dev_stop(dev->port_id);
1652 netdev_dpdk_update_flags(struct netdev *netdev_,
1653 enum netdev_flags off, enum netdev_flags on,
1654 enum netdev_flags *old_flagsp)
1656 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1659 ovs_mutex_lock(&netdev->mutex);
1660 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1661 ovs_mutex_unlock(&netdev->mutex);
1667 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1669 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1670 struct rte_eth_dev_info dev_info;
1672 if (dev->port_id < 0)
1675 ovs_mutex_lock(&dev->mutex);
1676 rte_eth_dev_info_get(dev->port_id, &dev_info);
1677 ovs_mutex_unlock(&dev->mutex);
1679 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1681 smap_add_format(args, "port_no", "%d", dev->port_id);
1682 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1683 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1684 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1685 smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen);
1686 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1687 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1688 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1689 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1690 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1691 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1693 smap_add_format(args, "pci-vendor_id", "0x%u", dev_info.pci_dev->id.vendor_id);
1694 smap_add_format(args, "pci-device_id", "0x%x", dev_info.pci_dev->id.device_id);
1700 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1701 OVS_REQUIRES(dev->mutex)
1703 enum netdev_flags old_flags;
1706 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1708 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1713 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1714 const char *argv[], void *aux OVS_UNUSED)
1718 if (!strcasecmp(argv[argc - 1], "up")) {
1720 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1723 unixctl_command_reply_error(conn, "Invalid Admin State");
1728 struct netdev *netdev = netdev_from_name(argv[1]);
1729 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1730 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1732 ovs_mutex_lock(&dpdk_dev->mutex);
1733 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1734 ovs_mutex_unlock(&dpdk_dev->mutex);
1736 netdev_close(netdev);
1738 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1739 netdev_close(netdev);
1743 struct netdev_dpdk *netdev;
1745 ovs_mutex_lock(&dpdk_mutex);
1746 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1747 ovs_mutex_lock(&netdev->mutex);
1748 netdev_dpdk_set_admin_state__(netdev, up);
1749 ovs_mutex_unlock(&netdev->mutex);
1751 ovs_mutex_unlock(&dpdk_mutex);
1753 unixctl_command_reply(conn, "OK");
1757 * Set virtqueue flags so that we do not receive interrupts.
1760 set_irq_status(struct virtio_net *dev)
1762 dev->virtqueue[VIRTIO_RXQ]->used->flags = VRING_USED_F_NO_NOTIFY;
1763 dev->virtqueue[VIRTIO_TXQ]->used->flags = VRING_USED_F_NO_NOTIFY;
1767 * A new virtio-net device is added to a vhost port.
1770 new_device(struct virtio_net *dev)
1772 struct netdev_dpdk *netdev;
1773 bool exists = false;
1775 ovs_mutex_lock(&dpdk_mutex);
1776 /* Add device to the vhost port with the same name as that passed down. */
1777 LIST_FOR_EACH(netdev, list_node, &dpdk_list) {
1778 if (strncmp(dev->ifname, netdev->vhost_id, IF_NAME_SZ) == 0) {
1779 ovs_mutex_lock(&netdev->mutex);
1780 ovsrcu_set(&netdev->virtio_dev, dev);
1781 ovs_mutex_unlock(&netdev->mutex);
1783 dev->flags |= VIRTIO_DEV_RUNNING;
1784 /* Disable notifications. */
1785 set_irq_status(dev);
1789 ovs_mutex_unlock(&dpdk_mutex);
1792 VLOG_INFO("vHost Device '%s' %"PRIu64" can't be added - name not "
1793 "found", dev->ifname, dev->device_fh);
1798 VLOG_INFO("vHost Device '%s' %"PRIu64" has been added", dev->ifname,
1804 * Remove a virtio-net device from the specific vhost port. Use dev->remove
1805 * flag to stop any more packets from being sent or received to/from a VM and
1806 * ensure all currently queued packets have been sent/received before removing
1810 destroy_device(volatile struct virtio_net *dev)
1812 struct netdev_dpdk *vhost_dev;
1814 ovs_mutex_lock(&dpdk_mutex);
1815 LIST_FOR_EACH (vhost_dev, list_node, &dpdk_list) {
1816 if (netdev_dpdk_get_virtio(vhost_dev) == dev) {
1818 ovs_mutex_lock(&vhost_dev->mutex);
1819 dev->flags &= ~VIRTIO_DEV_RUNNING;
1820 ovsrcu_set(&vhost_dev->virtio_dev, NULL);
1821 ovs_mutex_unlock(&vhost_dev->mutex);
1824 * Wait for other threads to quiesce before
1825 * setting the virtio_dev to NULL.
1827 ovsrcu_synchronize();
1829 * As call to ovsrcu_synchronize() will end the quiescent state,
1830 * put thread back into quiescent state before returning.
1832 ovsrcu_quiesce_start();
1835 ovs_mutex_unlock(&dpdk_mutex);
1837 VLOG_INFO("vHost Device '%s' %"PRIu64" has been removed", dev->ifname,
1842 netdev_dpdk_get_virtio(const struct netdev_dpdk *dev)
1844 return ovsrcu_get(struct virtio_net *, &dev->virtio_dev);
1848 * These callbacks allow virtio-net devices to be added to vhost ports when
1849 * configuration has been fully complete.
1851 static const struct virtio_net_device_ops virtio_net_device_ops =
1853 .new_device = new_device,
1854 .destroy_device = destroy_device,
1858 start_vhost_loop(void *dummy OVS_UNUSED)
1860 pthread_detach(pthread_self());
1861 /* Put the cuse thread into quiescent state. */
1862 ovsrcu_quiesce_start();
1863 rte_vhost_driver_session_start();
1868 dpdk_vhost_class_init(void)
1870 rte_vhost_driver_callback_register(&virtio_net_device_ops);
1871 ovs_thread_create("vhost_thread", start_vhost_loop, NULL);
1876 dpdk_vhost_cuse_class_init(void)
1881 /* Register CUSE device to handle IOCTLs.
1882 * Unless otherwise specified on the vswitchd command line, cuse_dev_name
1883 * is set to vhost-net.
1885 err = rte_vhost_driver_register(cuse_dev_name);
1888 VLOG_ERR("CUSE device setup failure.");
1892 dpdk_vhost_class_init();
1897 dpdk_vhost_user_class_init(void)
1899 dpdk_vhost_class_init();
1904 dpdk_common_init(void)
1906 unixctl_command_register("netdev-dpdk/set-admin-state",
1907 "[netdev] up|down", 1, 2,
1908 netdev_dpdk_set_admin_state, NULL);
1910 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
1916 dpdk_ring_create(const char dev_name[], unsigned int port_no,
1917 unsigned int *eth_port_id)
1919 struct dpdk_ring *ivshmem;
1923 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
1924 if (ivshmem == NULL) {
1928 /* XXX: Add support for multiquque ring. */
1929 err = snprintf(ring_name, 10, "%s_tx", dev_name);
1934 /* Create single producer tx ring, netdev does explicit locking. */
1935 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1937 if (ivshmem->cring_tx == NULL) {
1942 err = snprintf(ring_name, 10, "%s_rx", dev_name);
1947 /* Create single consumer rx ring, netdev does explicit locking. */
1948 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1950 if (ivshmem->cring_rx == NULL) {
1955 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
1956 &ivshmem->cring_tx, 1, SOCKET0);
1963 ivshmem->user_port_id = port_no;
1964 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
1965 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
1967 *eth_port_id = ivshmem->eth_port_id;
1972 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
1974 struct dpdk_ring *ivshmem;
1975 unsigned int port_no;
1978 /* Names always start with "dpdkr" */
1979 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
1984 /* look through our list to find the device */
1985 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
1986 if (ivshmem->user_port_id == port_no) {
1987 VLOG_INFO("Found dpdk ring device %s:", dev_name);
1988 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
1992 /* Need to create the device rings */
1993 return dpdk_ring_create(dev_name, port_no, eth_port_id);
1997 netdev_dpdk_ring_send(struct netdev *netdev_, int qid,
1998 struct dp_packet **pkts, int cnt, bool may_steal)
2000 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
2003 /* When using 'dpdkr' and sending to a DPDK ring, we want to ensure that the
2004 * rss hash field is clear. This is because the same mbuf may be modified by
2005 * the consumer of the ring and return into the datapath without recalculating
2007 for (i = 0; i < cnt; i++) {
2008 dp_packet_rss_invalidate(pkts[i]);
2011 netdev_dpdk_send__(netdev, qid, pkts, cnt, may_steal);
2016 netdev_dpdk_ring_construct(struct netdev *netdev)
2018 unsigned int port_no = 0;
2021 if (rte_eal_init_ret) {
2022 return rte_eal_init_ret;
2025 ovs_mutex_lock(&dpdk_mutex);
2027 err = dpdk_ring_open(netdev->name, &port_no);
2032 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
2035 ovs_mutex_unlock(&dpdk_mutex);
2039 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, DESTRUCT, MULTIQ, SEND, \
2040 GET_CARRIER, GET_STATS, GET_FEATURES, GET_STATUS, RXQ_RECV) \
2044 NULL, /* netdev_dpdk_run */ \
2045 NULL, /* netdev_dpdk_wait */ \
2047 netdev_dpdk_alloc, \
2050 netdev_dpdk_dealloc, \
2051 netdev_dpdk_get_config, \
2052 NULL, /* netdev_dpdk_set_config */ \
2053 NULL, /* get_tunnel_config */ \
2054 NULL, /* build header */ \
2055 NULL, /* push header */ \
2056 NULL, /* pop header */ \
2057 netdev_dpdk_get_numa_id, /* get_numa_id */ \
2058 MULTIQ, /* set_multiq */ \
2061 NULL, /* send_wait */ \
2063 netdev_dpdk_set_etheraddr, \
2064 netdev_dpdk_get_etheraddr, \
2065 netdev_dpdk_get_mtu, \
2066 netdev_dpdk_set_mtu, \
2067 netdev_dpdk_get_ifindex, \
2069 netdev_dpdk_get_carrier_resets, \
2070 netdev_dpdk_set_miimon, \
2073 NULL, /* set_advertisements */ \
2075 NULL, /* set_policing */ \
2076 NULL, /* get_qos_types */ \
2077 NULL, /* get_qos_capabilities */ \
2078 NULL, /* get_qos */ \
2079 NULL, /* set_qos */ \
2080 NULL, /* get_queue */ \
2081 NULL, /* set_queue */ \
2082 NULL, /* delete_queue */ \
2083 NULL, /* get_queue_stats */ \
2084 NULL, /* queue_dump_start */ \
2085 NULL, /* queue_dump_next */ \
2086 NULL, /* queue_dump_done */ \
2087 NULL, /* dump_queue_stats */ \
2089 NULL, /* get_in4 */ \
2090 NULL, /* set_in4 */ \
2091 NULL, /* get_in6 */ \
2092 NULL, /* add_router */ \
2093 NULL, /* get_next_hop */ \
2095 NULL, /* arp_lookup */ \
2097 netdev_dpdk_update_flags, \
2099 netdev_dpdk_rxq_alloc, \
2100 netdev_dpdk_rxq_construct, \
2101 netdev_dpdk_rxq_destruct, \
2102 netdev_dpdk_rxq_dealloc, \
2104 NULL, /* rx_wait */ \
2105 NULL, /* rxq_drain */ \
2109 process_vhost_flags(char *flag, char *default_val, int size,
2110 char **argv, char **new_val)
2114 /* Depending on which version of vhost is in use, process the vhost-specific
2115 * flag if it is provided on the vswitchd command line, otherwise resort to
2118 * For vhost-user: Process "-vhost_sock_dir" to set the custom location of
2119 * the vhost-user socket(s).
2120 * For vhost-cuse: Process "-cuse_dev_name" to set the custom name of the
2121 * vhost-cuse character device.
2123 if (!strcmp(argv[1], flag) && (strlen(argv[2]) <= size)) {
2125 *new_val = strdup(argv[2]);
2126 VLOG_INFO("User-provided %s in use: %s", flag, *new_val);
2128 VLOG_INFO("No %s provided - defaulting to %s", flag, default_val);
2129 *new_val = default_val;
2136 dpdk_init(int argc, char **argv)
2140 char *pragram_name = argv[0];
2142 if (argc < 2 || strcmp(argv[1], "--dpdk"))
2145 /* Remove the --dpdk argument from arg list.*/
2149 /* Reject --user option */
2151 for (i = 0; i < argc; i++) {
2152 if (!strcmp(argv[i], "--user")) {
2153 VLOG_ERR("Can not mix --dpdk and --user options, aborting.");
2158 if (process_vhost_flags("-cuse_dev_name", strdup("vhost-net"),
2159 PATH_MAX, argv, &cuse_dev_name)) {
2161 if (process_vhost_flags("-vhost_sock_dir", strdup(ovs_rundir()),
2162 NAME_MAX, argv, &vhost_sock_dir)) {
2166 err = stat(vhost_sock_dir, &s);
2168 VLOG_ERR("vHostUser socket DIR '%s' does not exist.",
2173 /* Remove the vhost flag configuration parameters from the argument
2174 * list, so that the correct elements are passed to the DPDK
2175 * initialization function
2178 argv += 2; /* Increment by two to bypass the vhost flag arguments */
2182 /* Keep the program name argument as this is needed for call to
2185 argv[0] = pragram_name;
2187 /* Make sure things are initialized ... */
2188 result = rte_eal_init(argc, argv);
2190 ovs_abort(result, "Cannot init EAL");
2193 rte_memzone_dump(stdout);
2194 rte_eal_init_ret = 0;
2196 if (argc > result) {
2197 argv[result] = argv[0];
2200 /* We are called from the main thread here */
2201 RTE_PER_LCORE(_lcore_id) = NON_PMD_CORE_ID;
2203 return result + 1 + base;
2206 static const struct netdev_class dpdk_class =
2210 netdev_dpdk_construct,
2211 netdev_dpdk_destruct,
2212 netdev_dpdk_set_multiq,
2213 netdev_dpdk_eth_send,
2214 netdev_dpdk_get_carrier,
2215 netdev_dpdk_get_stats,
2216 netdev_dpdk_get_features,
2217 netdev_dpdk_get_status,
2218 netdev_dpdk_rxq_recv);
2220 static const struct netdev_class dpdk_ring_class =
2224 netdev_dpdk_ring_construct,
2225 netdev_dpdk_destruct,
2226 netdev_dpdk_set_multiq,
2227 netdev_dpdk_ring_send,
2228 netdev_dpdk_get_carrier,
2229 netdev_dpdk_get_stats,
2230 netdev_dpdk_get_features,
2231 netdev_dpdk_get_status,
2232 netdev_dpdk_rxq_recv);
2234 static const struct netdev_class OVS_UNUSED dpdk_vhost_cuse_class =
2237 dpdk_vhost_cuse_class_init,
2238 netdev_dpdk_vhost_cuse_construct,
2239 netdev_dpdk_vhost_destruct,
2240 netdev_dpdk_vhost_set_multiq,
2241 netdev_dpdk_vhost_send,
2242 netdev_dpdk_vhost_get_carrier,
2243 netdev_dpdk_vhost_get_stats,
2246 netdev_dpdk_vhost_rxq_recv);
2248 static const struct netdev_class OVS_UNUSED dpdk_vhost_user_class =
2251 dpdk_vhost_user_class_init,
2252 netdev_dpdk_vhost_user_construct,
2253 netdev_dpdk_vhost_destruct,
2254 netdev_dpdk_vhost_set_multiq,
2255 netdev_dpdk_vhost_send,
2256 netdev_dpdk_vhost_get_carrier,
2257 netdev_dpdk_vhost_get_stats,
2260 netdev_dpdk_vhost_rxq_recv);
2263 netdev_dpdk_register(void)
2265 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
2267 if (rte_eal_init_ret) {
2271 if (ovsthread_once_start(&once)) {
2273 netdev_register_provider(&dpdk_class);
2274 netdev_register_provider(&dpdk_ring_class);
2276 netdev_register_provider(&dpdk_vhost_cuse_class);
2278 netdev_register_provider(&dpdk_vhost_user_class);
2280 ovsthread_once_done(&once);
2285 pmd_thread_setaffinity_cpu(unsigned cpu)
2291 CPU_SET(cpu, &cpuset);
2292 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
2294 VLOG_ERR("Thread affinity error %d",err);
2297 /* NON_PMD_CORE_ID is reserved for use by non pmd threads. */
2298 ovs_assert(cpu != NON_PMD_CORE_ID);
2299 RTE_PER_LCORE(_lcore_id) = cpu;
2307 return rte_lcore_id() != NON_PMD_CORE_ID;