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 ovs_mutex_lock(&dpdk_mutex);
753 list_remove(&dev->list_node);
754 dpdk_mp_put(dev->dpdk_mp);
755 ovs_mutex_unlock(&dpdk_mutex);
759 netdev_dpdk_dealloc(struct netdev *netdev_)
761 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
767 netdev_dpdk_get_config(const struct netdev *netdev_, struct smap *args)
769 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
771 ovs_mutex_lock(&dev->mutex);
773 smap_add_format(args, "configured_rx_queues", "%d", netdev_->n_rxq);
774 smap_add_format(args, "requested_tx_queues", "%d", netdev_->n_txq);
775 smap_add_format(args, "configured_tx_queues", "%d", dev->real_n_txq);
776 ovs_mutex_unlock(&dev->mutex);
782 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
784 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
786 return netdev->socket_id;
789 /* Sets the number of tx queues and rx queues for the dpdk interface.
790 * If the configuration fails, do not try restoring its old configuration
791 * and just returns the error. */
793 netdev_dpdk_set_multiq(struct netdev *netdev_, unsigned int n_txq,
796 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
798 int old_rxq, old_txq;
800 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
804 ovs_mutex_lock(&dpdk_mutex);
805 ovs_mutex_lock(&netdev->mutex);
807 rte_eth_dev_stop(netdev->port_id);
809 old_txq = netdev->up.n_txq;
810 old_rxq = netdev->up.n_rxq;
811 netdev->up.n_txq = n_txq;
812 netdev->up.n_rxq = n_rxq;
814 rte_free(netdev->tx_q);
815 err = dpdk_eth_dev_init(netdev);
816 netdev_dpdk_alloc_txq(netdev, netdev->real_n_txq);
818 /* If there has been an error, it means that the requested queues
819 * have not been created. Restore the old numbers. */
820 netdev->up.n_txq = old_txq;
821 netdev->up.n_rxq = old_rxq;
824 netdev->txq_needs_locking = netdev->real_n_txq != netdev->up.n_txq;
826 ovs_mutex_unlock(&netdev->mutex);
827 ovs_mutex_unlock(&dpdk_mutex);
833 netdev_dpdk_vhost_set_multiq(struct netdev *netdev_, unsigned int n_txq,
836 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
839 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
843 ovs_mutex_lock(&dpdk_mutex);
844 ovs_mutex_lock(&netdev->mutex);
846 netdev->up.n_txq = n_txq;
847 netdev->real_n_txq = 1;
848 netdev->up.n_rxq = 1;
850 ovs_mutex_unlock(&netdev->mutex);
851 ovs_mutex_unlock(&dpdk_mutex);
856 static struct netdev_rxq *
857 netdev_dpdk_rxq_alloc(void)
859 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
864 static struct netdev_rxq_dpdk *
865 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
867 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
871 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
873 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
874 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
876 ovs_mutex_lock(&netdev->mutex);
877 rx->port_id = netdev->port_id;
878 ovs_mutex_unlock(&netdev->mutex);
884 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
889 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
891 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
897 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
899 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
902 while (nb_tx != txq->count) {
905 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
914 if (OVS_UNLIKELY(nb_tx != txq->count)) {
915 /* free buffers, which we couldn't transmit, one at a time (each
916 * packet could come from a different mempool) */
919 for (i = nb_tx; i < txq->count; i++) {
920 rte_pktmbuf_free_seg(txq->burst_pkts[i]);
922 rte_spinlock_lock(&dev->stats_lock);
923 dev->stats.tx_dropped += txq->count-nb_tx;
924 rte_spinlock_unlock(&dev->stats_lock);
928 txq->tsc = rte_get_timer_cycles();
932 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
934 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
936 if (txq->count == 0) {
939 dpdk_queue_flush__(dev, qid);
943 is_vhost_running(struct virtio_net *dev)
945 return (dev != NULL && (dev->flags & VIRTIO_DEV_RUNNING));
949 netdev_dpdk_vhost_update_rx_counters(struct netdev_stats *stats,
950 struct dp_packet **packets, int count)
953 struct dp_packet *packet;
955 stats->rx_packets += count;
956 for (i = 0; i < count; i++) {
959 if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) {
960 /* This only protects the following multicast counting from
961 * too short packets, but it does not stop the packet from
962 * further processing. */
964 stats->rx_length_errors++;
968 struct eth_header *eh = (struct eth_header *) dp_packet_data(packet);
969 if (OVS_UNLIKELY(eth_addr_is_multicast(eh->eth_dst))) {
973 stats->rx_bytes += dp_packet_size(packet);
978 * The receive path for the vhost port is the TX path out from guest.
981 netdev_dpdk_vhost_rxq_recv(struct netdev_rxq *rxq_,
982 struct dp_packet **packets, int *c)
984 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
985 struct netdev *netdev = rx->up.netdev;
986 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
987 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
991 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
995 nb_rx = rte_vhost_dequeue_burst(virtio_dev, qid,
996 vhost_dev->dpdk_mp->mp,
997 (struct rte_mbuf **)packets,
1003 rte_spinlock_lock(&vhost_dev->stats_lock);
1004 netdev_dpdk_vhost_update_rx_counters(&vhost_dev->stats, packets, nb_rx);
1005 rte_spinlock_unlock(&vhost_dev->stats_lock);
1012 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dp_packet **packets,
1015 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
1016 struct netdev *netdev = rx->up.netdev;
1017 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1020 /* There is only one tx queue for this core. Do not flush other
1022 * Do not flush tx queue which is shared among CPUs
1023 * since it is always flushed */
1024 if (rxq_->queue_id == rte_lcore_id() &&
1025 OVS_LIKELY(!dev->txq_needs_locking)) {
1026 dpdk_queue_flush(dev, rxq_->queue_id);
1029 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
1030 (struct rte_mbuf **) packets,
1042 netdev_dpdk_vhost_update_tx_counters(struct netdev_stats *stats,
1043 struct dp_packet **packets,
1048 int sent = attempted - dropped;
1050 stats->tx_packets += sent;
1051 stats->tx_dropped += dropped;
1053 for (i = 0; i < sent; i++) {
1054 stats->tx_bytes += dp_packet_size(packets[i]);
1059 __netdev_dpdk_vhost_send(struct netdev *netdev, struct dp_packet **pkts,
1060 int cnt, bool may_steal)
1062 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
1063 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
1064 struct rte_mbuf **cur_pkts = (struct rte_mbuf **) pkts;
1065 unsigned int total_pkts = cnt;
1068 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
1069 rte_spinlock_lock(&vhost_dev->stats_lock);
1070 vhost_dev->stats.tx_dropped+= cnt;
1071 rte_spinlock_unlock(&vhost_dev->stats_lock);
1075 /* There is vHost TX single queue, So we need to lock it for TX. */
1076 rte_spinlock_lock(&vhost_dev->vhost_tx_lock);
1079 unsigned int tx_pkts;
1081 tx_pkts = rte_vhost_enqueue_burst(virtio_dev, VIRTIO_RXQ,
1083 if (OVS_LIKELY(tx_pkts)) {
1084 /* Packets have been sent.*/
1086 /* Prepare for possible next iteration.*/
1087 cur_pkts = &cur_pkts[tx_pkts];
1089 uint64_t timeout = VHOST_ENQ_RETRY_USECS * rte_get_timer_hz() / 1E6;
1090 unsigned int expired = 0;
1093 start = rte_get_timer_cycles();
1097 * Unable to enqueue packets to vhost interface.
1098 * Check available entries before retrying.
1100 while (!rte_vring_available_entries(virtio_dev, VIRTIO_RXQ)) {
1101 if (OVS_UNLIKELY((rte_get_timer_cycles() - start) > timeout)) {
1107 /* break out of main loop. */
1112 rte_spinlock_unlock(&vhost_dev->vhost_tx_lock);
1114 rte_spinlock_lock(&vhost_dev->stats_lock);
1115 netdev_dpdk_vhost_update_tx_counters(&vhost_dev->stats, pkts, total_pkts,
1117 rte_spinlock_unlock(&vhost_dev->stats_lock);
1123 for (i = 0; i < total_pkts; i++) {
1124 dp_packet_delete(pkts[i]);
1130 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
1131 struct rte_mbuf **pkts, int cnt)
1133 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
1139 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
1140 int tocopy = MIN(freeslots, cnt-i);
1142 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
1143 tocopy * sizeof (struct rte_mbuf *));
1145 txq->count += tocopy;
1148 if (txq->count == MAX_TX_QUEUE_LEN || txq->flush_tx) {
1149 dpdk_queue_flush__(dev, qid);
1151 diff_tsc = rte_get_timer_cycles() - txq->tsc;
1152 if (diff_tsc >= DRAIN_TSC) {
1153 dpdk_queue_flush__(dev, qid);
1158 /* Tx function. Transmit packets indefinitely */
1160 dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet **pkts,
1162 OVS_NO_THREAD_SAFETY_ANALYSIS
1164 #if !defined(__CHECKER__) && !defined(_WIN32)
1165 const size_t PKT_ARRAY_SIZE = cnt;
1167 /* Sparse or MSVC doesn't like variable length array. */
1168 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
1170 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1171 struct rte_mbuf *mbufs[PKT_ARRAY_SIZE];
1176 /* If we are on a non pmd thread we have to use the mempool mutex, because
1177 * every non pmd thread shares the same mempool cache */
1179 if (!thread_is_pmd()) {
1180 ovs_mutex_lock(&nonpmd_mempool_mutex);
1183 for (i = 0; i < cnt; i++) {
1184 int size = dp_packet_size(pkts[i]);
1186 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1187 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1188 (int)size , dev->max_packet_len);
1194 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
1196 if (!mbufs[newcnt]) {
1201 /* We have to do a copy for now */
1202 memcpy(rte_pktmbuf_mtod(mbufs[newcnt], void *), dp_packet_data(pkts[i]), size);
1204 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
1205 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
1210 if (OVS_UNLIKELY(dropped)) {
1211 rte_spinlock_lock(&dev->stats_lock);
1212 dev->stats.tx_dropped += dropped;
1213 rte_spinlock_unlock(&dev->stats_lock);
1216 if (dev->type == DPDK_DEV_VHOST) {
1217 __netdev_dpdk_vhost_send(netdev, (struct dp_packet **) mbufs, newcnt, true);
1219 dpdk_queue_pkts(dev, qid, mbufs, newcnt);
1220 dpdk_queue_flush(dev, qid);
1223 if (!thread_is_pmd()) {
1224 ovs_mutex_unlock(&nonpmd_mempool_mutex);
1229 netdev_dpdk_vhost_send(struct netdev *netdev, int qid OVS_UNUSED, struct dp_packet **pkts,
1230 int cnt, bool may_steal)
1232 if (OVS_UNLIKELY(pkts[0]->source != DPBUF_DPDK)) {
1235 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1237 for (i = 0; i < cnt; i++) {
1238 dp_packet_delete(pkts[i]);
1242 __netdev_dpdk_vhost_send(netdev, pkts, cnt, may_steal);
1248 netdev_dpdk_send__(struct netdev_dpdk *dev, int qid,
1249 struct dp_packet **pkts, int cnt, bool may_steal)
1253 if (OVS_UNLIKELY(dev->txq_needs_locking)) {
1254 qid = qid % dev->real_n_txq;
1255 rte_spinlock_lock(&dev->tx_q[qid].tx_lock);
1258 if (OVS_UNLIKELY(!may_steal ||
1259 pkts[0]->source != DPBUF_DPDK)) {
1260 struct netdev *netdev = &dev->up;
1262 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1265 for (i = 0; i < cnt; i++) {
1266 dp_packet_delete(pkts[i]);
1270 int next_tx_idx = 0;
1273 for (i = 0; i < cnt; i++) {
1274 int size = dp_packet_size(pkts[i]);
1276 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1277 if (next_tx_idx != i) {
1278 dpdk_queue_pkts(dev, qid,
1279 (struct rte_mbuf **)&pkts[next_tx_idx],
1283 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1284 (int)size , dev->max_packet_len);
1286 dp_packet_delete(pkts[i]);
1288 next_tx_idx = i + 1;
1291 if (next_tx_idx != cnt) {
1292 dpdk_queue_pkts(dev, qid,
1293 (struct rte_mbuf **)&pkts[next_tx_idx],
1297 if (OVS_UNLIKELY(dropped)) {
1298 rte_spinlock_lock(&dev->stats_lock);
1299 dev->stats.tx_dropped += dropped;
1300 rte_spinlock_unlock(&dev->stats_lock);
1304 if (OVS_UNLIKELY(dev->txq_needs_locking)) {
1305 rte_spinlock_unlock(&dev->tx_q[qid].tx_lock);
1310 netdev_dpdk_eth_send(struct netdev *netdev, int qid,
1311 struct dp_packet **pkts, int cnt, bool may_steal)
1313 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1315 netdev_dpdk_send__(dev, qid, pkts, cnt, may_steal);
1320 netdev_dpdk_set_etheraddr(struct netdev *netdev, const struct eth_addr mac)
1322 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1324 ovs_mutex_lock(&dev->mutex);
1325 if (!eth_addr_equals(dev->hwaddr, mac)) {
1327 netdev_change_seq_changed(netdev);
1329 ovs_mutex_unlock(&dev->mutex);
1335 netdev_dpdk_get_etheraddr(const struct netdev *netdev, struct eth_addr *mac)
1337 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1339 ovs_mutex_lock(&dev->mutex);
1341 ovs_mutex_unlock(&dev->mutex);
1347 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
1349 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1351 ovs_mutex_lock(&dev->mutex);
1353 ovs_mutex_unlock(&dev->mutex);
1359 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
1361 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1363 struct dpdk_mp *old_mp;
1366 ovs_mutex_lock(&dpdk_mutex);
1367 ovs_mutex_lock(&dev->mutex);
1368 if (dev->mtu == mtu) {
1373 mp = dpdk_mp_get(dev->socket_id, dev->mtu);
1379 rte_eth_dev_stop(dev->port_id);
1382 old_mp = dev->dpdk_mp;
1385 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1387 err = dpdk_eth_dev_init(dev);
1391 dev->dpdk_mp = old_mp;
1392 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1393 dpdk_eth_dev_init(dev);
1397 dpdk_mp_put(old_mp);
1398 netdev_change_seq_changed(netdev);
1400 ovs_mutex_unlock(&dev->mutex);
1401 ovs_mutex_unlock(&dpdk_mutex);
1406 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
1409 netdev_dpdk_vhost_get_stats(const struct netdev *netdev,
1410 struct netdev_stats *stats)
1412 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1414 ovs_mutex_lock(&dev->mutex);
1415 memset(stats, 0, sizeof(*stats));
1416 /* Unsupported Stats */
1417 stats->collisions = UINT64_MAX;
1418 stats->rx_crc_errors = UINT64_MAX;
1419 stats->rx_fifo_errors = UINT64_MAX;
1420 stats->rx_frame_errors = UINT64_MAX;
1421 stats->rx_missed_errors = UINT64_MAX;
1422 stats->rx_over_errors = UINT64_MAX;
1423 stats->tx_aborted_errors = UINT64_MAX;
1424 stats->tx_carrier_errors = UINT64_MAX;
1425 stats->tx_errors = UINT64_MAX;
1426 stats->tx_fifo_errors = UINT64_MAX;
1427 stats->tx_heartbeat_errors = UINT64_MAX;
1428 stats->tx_window_errors = UINT64_MAX;
1429 stats->rx_dropped += UINT64_MAX;
1431 rte_spinlock_lock(&dev->stats_lock);
1432 /* Supported Stats */
1433 stats->rx_packets += dev->stats.rx_packets;
1434 stats->tx_packets += dev->stats.tx_packets;
1435 stats->tx_dropped += dev->stats.tx_dropped;
1436 stats->multicast = dev->stats.multicast;
1437 stats->rx_bytes = dev->stats.rx_bytes;
1438 stats->tx_bytes = dev->stats.tx_bytes;
1439 stats->rx_errors = dev->stats.rx_errors;
1440 stats->rx_length_errors = dev->stats.rx_length_errors;
1441 rte_spinlock_unlock(&dev->stats_lock);
1443 ovs_mutex_unlock(&dev->mutex);
1449 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
1451 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1452 struct rte_eth_stats rte_stats;
1455 netdev_dpdk_get_carrier(netdev, &gg);
1456 ovs_mutex_lock(&dev->mutex);
1457 rte_eth_stats_get(dev->port_id, &rte_stats);
1459 memset(stats, 0, sizeof(*stats));
1461 stats->rx_packets = rte_stats.ipackets;
1462 stats->tx_packets = rte_stats.opackets;
1463 stats->rx_bytes = rte_stats.ibytes;
1464 stats->tx_bytes = rte_stats.obytes;
1465 /* DPDK counts imissed as errors, but count them here as dropped instead */
1466 stats->rx_errors = rte_stats.ierrors - rte_stats.imissed;
1467 stats->tx_errors = rte_stats.oerrors;
1468 stats->multicast = rte_stats.imcasts;
1470 rte_spinlock_lock(&dev->stats_lock);
1471 stats->tx_dropped = dev->stats.tx_dropped;
1472 rte_spinlock_unlock(&dev->stats_lock);
1474 /* These are the available DPDK counters for packets not received due to
1475 * local resource constraints in DPDK and NIC respectively. */
1476 stats->rx_dropped = rte_stats.rx_nombuf + rte_stats.imissed;
1477 stats->collisions = UINT64_MAX;
1479 stats->rx_length_errors = rte_stats.ibadlen;
1480 stats->rx_over_errors = UINT64_MAX;
1481 stats->rx_crc_errors = rte_stats.ibadcrc;
1482 stats->rx_frame_errors = UINT64_MAX;
1483 stats->rx_fifo_errors = UINT64_MAX;
1484 stats->rx_missed_errors = rte_stats.imissed;
1486 stats->tx_aborted_errors = UINT64_MAX;
1487 stats->tx_carrier_errors = UINT64_MAX;
1488 stats->tx_fifo_errors = UINT64_MAX;
1489 stats->tx_heartbeat_errors = UINT64_MAX;
1490 stats->tx_window_errors = UINT64_MAX;
1492 ovs_mutex_unlock(&dev->mutex);
1498 netdev_dpdk_get_features(const struct netdev *netdev_,
1499 enum netdev_features *current,
1500 enum netdev_features *advertised OVS_UNUSED,
1501 enum netdev_features *supported OVS_UNUSED,
1502 enum netdev_features *peer OVS_UNUSED)
1504 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1505 struct rte_eth_link link;
1507 ovs_mutex_lock(&dev->mutex);
1509 ovs_mutex_unlock(&dev->mutex);
1511 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
1512 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
1513 *current = NETDEV_F_AUTONEG;
1515 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
1516 if (link.link_speed == ETH_LINK_SPEED_10) {
1517 *current = NETDEV_F_10MB_HD;
1519 if (link.link_speed == ETH_LINK_SPEED_100) {
1520 *current = NETDEV_F_100MB_HD;
1522 if (link.link_speed == ETH_LINK_SPEED_1000) {
1523 *current = NETDEV_F_1GB_HD;
1525 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
1526 if (link.link_speed == ETH_LINK_SPEED_10) {
1527 *current = NETDEV_F_10MB_FD;
1529 if (link.link_speed == ETH_LINK_SPEED_100) {
1530 *current = NETDEV_F_100MB_FD;
1532 if (link.link_speed == ETH_LINK_SPEED_1000) {
1533 *current = NETDEV_F_1GB_FD;
1535 if (link.link_speed == ETH_LINK_SPEED_10000) {
1536 *current = NETDEV_F_10GB_FD;
1544 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1546 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1549 ovs_mutex_lock(&dev->mutex);
1550 ifindex = dev->port_id;
1551 ovs_mutex_unlock(&dev->mutex);
1557 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1559 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1561 ovs_mutex_lock(&dev->mutex);
1562 check_link_status(dev);
1563 *carrier = dev->link.link_status;
1565 ovs_mutex_unlock(&dev->mutex);
1571 netdev_dpdk_vhost_get_carrier(const struct netdev *netdev_, bool *carrier)
1573 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1574 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(dev);
1576 ovs_mutex_lock(&dev->mutex);
1578 if (is_vhost_running(virtio_dev)) {
1584 ovs_mutex_unlock(&dev->mutex);
1589 static long long int
1590 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1592 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1593 long long int carrier_resets;
1595 ovs_mutex_lock(&dev->mutex);
1596 carrier_resets = dev->link_reset_cnt;
1597 ovs_mutex_unlock(&dev->mutex);
1599 return carrier_resets;
1603 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1604 long long int interval OVS_UNUSED)
1610 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1611 enum netdev_flags off, enum netdev_flags on,
1612 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1616 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1620 *old_flagsp = dev->flags;
1624 if (dev->flags == *old_flagsp) {
1628 if (dev->type == DPDK_DEV_ETH) {
1629 if (dev->flags & NETDEV_UP) {
1630 err = rte_eth_dev_start(dev->port_id);
1635 if (dev->flags & NETDEV_PROMISC) {
1636 rte_eth_promiscuous_enable(dev->port_id);
1639 if (!(dev->flags & NETDEV_UP)) {
1640 rte_eth_dev_stop(dev->port_id);
1648 netdev_dpdk_update_flags(struct netdev *netdev_,
1649 enum netdev_flags off, enum netdev_flags on,
1650 enum netdev_flags *old_flagsp)
1652 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1655 ovs_mutex_lock(&netdev->mutex);
1656 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1657 ovs_mutex_unlock(&netdev->mutex);
1663 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1665 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1666 struct rte_eth_dev_info dev_info;
1668 if (dev->port_id < 0)
1671 ovs_mutex_lock(&dev->mutex);
1672 rte_eth_dev_info_get(dev->port_id, &dev_info);
1673 ovs_mutex_unlock(&dev->mutex);
1675 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1677 smap_add_format(args, "port_no", "%d", dev->port_id);
1678 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1679 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1680 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1681 smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen);
1682 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1683 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1684 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1685 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1686 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1687 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1689 smap_add_format(args, "pci-vendor_id", "0x%u", dev_info.pci_dev->id.vendor_id);
1690 smap_add_format(args, "pci-device_id", "0x%x", dev_info.pci_dev->id.device_id);
1696 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1697 OVS_REQUIRES(dev->mutex)
1699 enum netdev_flags old_flags;
1702 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1704 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1709 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1710 const char *argv[], void *aux OVS_UNUSED)
1714 if (!strcasecmp(argv[argc - 1], "up")) {
1716 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1719 unixctl_command_reply_error(conn, "Invalid Admin State");
1724 struct netdev *netdev = netdev_from_name(argv[1]);
1725 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1726 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1728 ovs_mutex_lock(&dpdk_dev->mutex);
1729 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1730 ovs_mutex_unlock(&dpdk_dev->mutex);
1732 netdev_close(netdev);
1734 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1735 netdev_close(netdev);
1739 struct netdev_dpdk *netdev;
1741 ovs_mutex_lock(&dpdk_mutex);
1742 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1743 ovs_mutex_lock(&netdev->mutex);
1744 netdev_dpdk_set_admin_state__(netdev, up);
1745 ovs_mutex_unlock(&netdev->mutex);
1747 ovs_mutex_unlock(&dpdk_mutex);
1749 unixctl_command_reply(conn, "OK");
1753 * Set virtqueue flags so that we do not receive interrupts.
1756 set_irq_status(struct virtio_net *dev)
1758 dev->virtqueue[VIRTIO_RXQ]->used->flags = VRING_USED_F_NO_NOTIFY;
1759 dev->virtqueue[VIRTIO_TXQ]->used->flags = VRING_USED_F_NO_NOTIFY;
1763 * A new virtio-net device is added to a vhost port.
1766 new_device(struct virtio_net *dev)
1768 struct netdev_dpdk *netdev;
1769 bool exists = false;
1771 ovs_mutex_lock(&dpdk_mutex);
1772 /* Add device to the vhost port with the same name as that passed down. */
1773 LIST_FOR_EACH(netdev, list_node, &dpdk_list) {
1774 if (strncmp(dev->ifname, netdev->vhost_id, IF_NAME_SZ) == 0) {
1775 ovs_mutex_lock(&netdev->mutex);
1776 ovsrcu_set(&netdev->virtio_dev, dev);
1777 ovs_mutex_unlock(&netdev->mutex);
1779 dev->flags |= VIRTIO_DEV_RUNNING;
1780 /* Disable notifications. */
1781 set_irq_status(dev);
1785 ovs_mutex_unlock(&dpdk_mutex);
1788 VLOG_INFO("vHost Device '%s' %"PRIu64" can't be added - name not "
1789 "found", dev->ifname, dev->device_fh);
1794 VLOG_INFO("vHost Device '%s' %"PRIu64" has been added", dev->ifname,
1800 * Remove a virtio-net device from the specific vhost port. Use dev->remove
1801 * flag to stop any more packets from being sent or received to/from a VM and
1802 * ensure all currently queued packets have been sent/received before removing
1806 destroy_device(volatile struct virtio_net *dev)
1808 struct netdev_dpdk *vhost_dev;
1810 ovs_mutex_lock(&dpdk_mutex);
1811 LIST_FOR_EACH (vhost_dev, list_node, &dpdk_list) {
1812 if (netdev_dpdk_get_virtio(vhost_dev) == dev) {
1814 ovs_mutex_lock(&vhost_dev->mutex);
1815 dev->flags &= ~VIRTIO_DEV_RUNNING;
1816 ovsrcu_set(&vhost_dev->virtio_dev, NULL);
1817 ovs_mutex_unlock(&vhost_dev->mutex);
1820 * Wait for other threads to quiesce before
1821 * setting the virtio_dev to NULL.
1823 ovsrcu_synchronize();
1825 * As call to ovsrcu_synchronize() will end the quiescent state,
1826 * put thread back into quiescent state before returning.
1828 ovsrcu_quiesce_start();
1831 ovs_mutex_unlock(&dpdk_mutex);
1833 VLOG_INFO("vHost Device '%s' %"PRIu64" has been removed", dev->ifname,
1838 netdev_dpdk_get_virtio(const struct netdev_dpdk *dev)
1840 return ovsrcu_get(struct virtio_net *, &dev->virtio_dev);
1844 * These callbacks allow virtio-net devices to be added to vhost ports when
1845 * configuration has been fully complete.
1847 static const struct virtio_net_device_ops virtio_net_device_ops =
1849 .new_device = new_device,
1850 .destroy_device = destroy_device,
1854 start_vhost_loop(void *dummy OVS_UNUSED)
1856 pthread_detach(pthread_self());
1857 /* Put the cuse thread into quiescent state. */
1858 ovsrcu_quiesce_start();
1859 rte_vhost_driver_session_start();
1864 dpdk_vhost_class_init(void)
1866 rte_vhost_driver_callback_register(&virtio_net_device_ops);
1867 ovs_thread_create("vhost_thread", start_vhost_loop, NULL);
1872 dpdk_vhost_cuse_class_init(void)
1877 /* Register CUSE device to handle IOCTLs.
1878 * Unless otherwise specified on the vswitchd command line, cuse_dev_name
1879 * is set to vhost-net.
1881 err = rte_vhost_driver_register(cuse_dev_name);
1884 VLOG_ERR("CUSE device setup failure.");
1888 dpdk_vhost_class_init();
1893 dpdk_vhost_user_class_init(void)
1895 dpdk_vhost_class_init();
1900 dpdk_common_init(void)
1902 unixctl_command_register("netdev-dpdk/set-admin-state",
1903 "[netdev] up|down", 1, 2,
1904 netdev_dpdk_set_admin_state, NULL);
1906 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
1912 dpdk_ring_create(const char dev_name[], unsigned int port_no,
1913 unsigned int *eth_port_id)
1915 struct dpdk_ring *ivshmem;
1919 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
1920 if (ivshmem == NULL) {
1924 /* XXX: Add support for multiquque ring. */
1925 err = snprintf(ring_name, 10, "%s_tx", dev_name);
1930 /* Create single consumer/producer rings, netdev does explicit locking. */
1931 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1932 RING_F_SP_ENQ | RING_F_SC_DEQ);
1933 if (ivshmem->cring_tx == NULL) {
1938 err = snprintf(ring_name, 10, "%s_rx", dev_name);
1943 /* Create single consumer/producer rings, netdev does explicit locking. */
1944 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
1945 RING_F_SP_ENQ | RING_F_SC_DEQ);
1946 if (ivshmem->cring_rx == NULL) {
1951 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
1952 &ivshmem->cring_tx, 1, SOCKET0);
1959 ivshmem->user_port_id = port_no;
1960 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
1961 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
1963 *eth_port_id = ivshmem->eth_port_id;
1968 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
1970 struct dpdk_ring *ivshmem;
1971 unsigned int port_no;
1974 /* Names always start with "dpdkr" */
1975 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
1980 /* look through our list to find the device */
1981 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
1982 if (ivshmem->user_port_id == port_no) {
1983 VLOG_INFO("Found dpdk ring device %s:", dev_name);
1984 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
1988 /* Need to create the device rings */
1989 return dpdk_ring_create(dev_name, port_no, eth_port_id);
1993 netdev_dpdk_ring_send(struct netdev *netdev_, int qid,
1994 struct dp_packet **pkts, int cnt, bool may_steal)
1996 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1999 /* When using 'dpdkr' and sending to a DPDK ring, we want to ensure that the
2000 * rss hash field is clear. This is because the same mbuf may be modified by
2001 * the consumer of the ring and return into the datapath without recalculating
2003 for (i = 0; i < cnt; i++) {
2004 dp_packet_rss_invalidate(pkts[i]);
2007 netdev_dpdk_send__(netdev, qid, pkts, cnt, may_steal);
2012 netdev_dpdk_ring_construct(struct netdev *netdev)
2014 unsigned int port_no = 0;
2017 if (rte_eal_init_ret) {
2018 return rte_eal_init_ret;
2021 ovs_mutex_lock(&dpdk_mutex);
2023 err = dpdk_ring_open(netdev->name, &port_no);
2028 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
2031 ovs_mutex_unlock(&dpdk_mutex);
2035 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, DESTRUCT, MULTIQ, SEND, \
2036 GET_CARRIER, GET_STATS, GET_FEATURES, GET_STATUS, RXQ_RECV) \
2040 NULL, /* netdev_dpdk_run */ \
2041 NULL, /* netdev_dpdk_wait */ \
2043 netdev_dpdk_alloc, \
2046 netdev_dpdk_dealloc, \
2047 netdev_dpdk_get_config, \
2048 NULL, /* netdev_dpdk_set_config */ \
2049 NULL, /* get_tunnel_config */ \
2050 NULL, /* build header */ \
2051 NULL, /* push header */ \
2052 NULL, /* pop header */ \
2053 netdev_dpdk_get_numa_id, /* get_numa_id */ \
2054 MULTIQ, /* set_multiq */ \
2057 NULL, /* send_wait */ \
2059 netdev_dpdk_set_etheraddr, \
2060 netdev_dpdk_get_etheraddr, \
2061 netdev_dpdk_get_mtu, \
2062 netdev_dpdk_set_mtu, \
2063 netdev_dpdk_get_ifindex, \
2065 netdev_dpdk_get_carrier_resets, \
2066 netdev_dpdk_set_miimon, \
2069 NULL, /* set_advertisements */ \
2071 NULL, /* set_policing */ \
2072 NULL, /* get_qos_types */ \
2073 NULL, /* get_qos_capabilities */ \
2074 NULL, /* get_qos */ \
2075 NULL, /* set_qos */ \
2076 NULL, /* get_queue */ \
2077 NULL, /* set_queue */ \
2078 NULL, /* delete_queue */ \
2079 NULL, /* get_queue_stats */ \
2080 NULL, /* queue_dump_start */ \
2081 NULL, /* queue_dump_next */ \
2082 NULL, /* queue_dump_done */ \
2083 NULL, /* dump_queue_stats */ \
2085 NULL, /* get_in4 */ \
2086 NULL, /* set_in4 */ \
2087 NULL, /* get_in6 */ \
2088 NULL, /* add_router */ \
2089 NULL, /* get_next_hop */ \
2091 NULL, /* arp_lookup */ \
2093 netdev_dpdk_update_flags, \
2095 netdev_dpdk_rxq_alloc, \
2096 netdev_dpdk_rxq_construct, \
2097 netdev_dpdk_rxq_destruct, \
2098 netdev_dpdk_rxq_dealloc, \
2100 NULL, /* rx_wait */ \
2101 NULL, /* rxq_drain */ \
2105 process_vhost_flags(char *flag, char *default_val, int size,
2106 char **argv, char **new_val)
2110 /* Depending on which version of vhost is in use, process the vhost-specific
2111 * flag if it is provided on the vswitchd command line, otherwise resort to
2114 * For vhost-user: Process "-cuse_dev_name" to set the custom location of
2115 * the vhost-user socket(s).
2116 * For vhost-cuse: Process "-vhost_sock_dir" to set the custom name of the
2117 * vhost-cuse character device.
2119 if (!strcmp(argv[1], flag) && (strlen(argv[2]) <= size)) {
2121 *new_val = strdup(argv[2]);
2122 VLOG_INFO("User-provided %s in use: %s", flag, *new_val);
2124 VLOG_INFO("No %s provided - defaulting to %s", flag, default_val);
2125 *new_val = default_val;
2132 dpdk_init(int argc, char **argv)
2136 char *pragram_name = argv[0];
2138 if (argc < 2 || strcmp(argv[1], "--dpdk"))
2141 /* Remove the --dpdk argument from arg list.*/
2145 /* Reject --user option */
2147 for (i = 0; i < argc; i++) {
2148 if (!strcmp(argv[i], "--user")) {
2149 VLOG_ERR("Can not mix --dpdk and --user options, aborting.");
2154 if (process_vhost_flags("-cuse_dev_name", strdup("vhost-net"),
2155 PATH_MAX, argv, &cuse_dev_name)) {
2157 if (process_vhost_flags("-vhost_sock_dir", strdup(ovs_rundir()),
2158 NAME_MAX, argv, &vhost_sock_dir)) {
2162 err = stat(vhost_sock_dir, &s);
2164 VLOG_ERR("vHostUser socket DIR '%s' does not exist.",
2169 /* Remove the vhost flag configuration parameters from the argument
2170 * list, so that the correct elements are passed to the DPDK
2171 * initialization function
2174 argv += 2; /* Increment by two to bypass the vhost flag arguments */
2178 /* Keep the program name argument as this is needed for call to
2181 argv[0] = pragram_name;
2183 /* Make sure things are initialized ... */
2184 result = rte_eal_init(argc, argv);
2186 ovs_abort(result, "Cannot init EAL");
2189 rte_memzone_dump(stdout);
2190 rte_eal_init_ret = 0;
2192 if (argc > result) {
2193 argv[result] = argv[0];
2196 /* We are called from the main thread here */
2197 RTE_PER_LCORE(_lcore_id) = NON_PMD_CORE_ID;
2199 return result + 1 + base;
2202 static const struct netdev_class dpdk_class =
2206 netdev_dpdk_construct,
2207 netdev_dpdk_destruct,
2208 netdev_dpdk_set_multiq,
2209 netdev_dpdk_eth_send,
2210 netdev_dpdk_get_carrier,
2211 netdev_dpdk_get_stats,
2212 netdev_dpdk_get_features,
2213 netdev_dpdk_get_status,
2214 netdev_dpdk_rxq_recv);
2216 static const struct netdev_class dpdk_ring_class =
2220 netdev_dpdk_ring_construct,
2221 netdev_dpdk_destruct,
2222 netdev_dpdk_set_multiq,
2223 netdev_dpdk_ring_send,
2224 netdev_dpdk_get_carrier,
2225 netdev_dpdk_get_stats,
2226 netdev_dpdk_get_features,
2227 netdev_dpdk_get_status,
2228 netdev_dpdk_rxq_recv);
2230 static const struct netdev_class OVS_UNUSED dpdk_vhost_cuse_class =
2233 dpdk_vhost_cuse_class_init,
2234 netdev_dpdk_vhost_cuse_construct,
2235 netdev_dpdk_vhost_destruct,
2236 netdev_dpdk_vhost_set_multiq,
2237 netdev_dpdk_vhost_send,
2238 netdev_dpdk_vhost_get_carrier,
2239 netdev_dpdk_vhost_get_stats,
2242 netdev_dpdk_vhost_rxq_recv);
2244 static const struct netdev_class OVS_UNUSED dpdk_vhost_user_class =
2247 dpdk_vhost_user_class_init,
2248 netdev_dpdk_vhost_user_construct,
2249 netdev_dpdk_vhost_destruct,
2250 netdev_dpdk_vhost_set_multiq,
2251 netdev_dpdk_vhost_send,
2252 netdev_dpdk_vhost_get_carrier,
2253 netdev_dpdk_vhost_get_stats,
2256 netdev_dpdk_vhost_rxq_recv);
2259 netdev_dpdk_register(void)
2261 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
2263 if (rte_eal_init_ret) {
2267 if (ovsthread_once_start(&once)) {
2269 netdev_register_provider(&dpdk_class);
2270 netdev_register_provider(&dpdk_ring_class);
2272 netdev_register_provider(&dpdk_vhost_cuse_class);
2274 netdev_register_provider(&dpdk_vhost_user_class);
2276 ovsthread_once_done(&once);
2281 pmd_thread_setaffinity_cpu(unsigned cpu)
2287 CPU_SET(cpu, &cpuset);
2288 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
2290 VLOG_ERR("Thread affinity error %d",err);
2293 /* NON_PMD_CORE_ID is reserved for use by non pmd threads. */
2294 ovs_assert(cpu != NON_PMD_CORE_ID);
2295 RTE_PER_LCORE(_lcore_id) = cpu;
2303 return rte_lcore_id() != NON_PMD_CORE_ID;