2 * Copyright (c) 2014, 2015, 2016 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"
36 #include "fatal-signal.h"
38 #include "netdev-dpdk.h"
39 #include "netdev-provider.h"
40 #include "netdev-vport.h"
42 #include "ofp-print.h"
44 #include "ovs-thread.h"
49 #include "unaligned.h"
52 #include "openvswitch/vlog.h"
54 #include "rte_config.h"
56 #include "rte_virtio_net.h"
58 VLOG_DEFINE_THIS_MODULE(dpdk);
59 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
61 #define DPDK_PORT_WATCHDOG_INTERVAL 5
63 #define OVS_CACHE_LINE_SIZE CACHE_LINE_SIZE
64 #define OVS_VPORT_DPDK "ovs_dpdk"
67 * need to reserve tons of extra space in the mbufs so we can align the
68 * DMA addresses to 4KB.
69 * The minimum mbuf size is limited to avoid scatter behaviour and drop in
70 * performance for standard Ethernet MTU.
72 #define MTU_TO_MAX_LEN(mtu) ((mtu) + ETHER_HDR_LEN + ETHER_CRC_LEN)
73 #define MBUF_SIZE_MTU(mtu) (MTU_TO_MAX_LEN(mtu) \
74 + sizeof(struct dp_packet) \
75 + RTE_PKTMBUF_HEADROOM)
76 #define MBUF_SIZE_DRIVER (2048 \
77 + sizeof (struct rte_mbuf) \
78 + RTE_PKTMBUF_HEADROOM)
79 #define MBUF_SIZE(mtu) MAX(MBUF_SIZE_MTU(mtu), MBUF_SIZE_DRIVER)
81 /* Max and min number of packets in the mempool. OVS tries to allocate a
82 * mempool with MAX_NB_MBUF: if this fails (because the system doesn't have
83 * enough hugepages) we keep halving the number until the allocation succeeds
84 * or we reach MIN_NB_MBUF */
86 #define MAX_NB_MBUF (4096 * 64)
87 #define MIN_NB_MBUF (4096 * 4)
88 #define MP_CACHE_SZ RTE_MEMPOOL_CACHE_MAX_SIZE
90 /* MAX_NB_MBUF can be divided by 2 many times, until MIN_NB_MBUF */
91 BUILD_ASSERT_DECL(MAX_NB_MBUF % ROUND_DOWN_POW2(MAX_NB_MBUF/MIN_NB_MBUF) == 0);
93 /* The smallest possible NB_MBUF that we're going to try should be a multiple
94 * of MP_CACHE_SZ. This is advised by DPDK documentation. */
95 BUILD_ASSERT_DECL((MAX_NB_MBUF / ROUND_DOWN_POW2(MAX_NB_MBUF/MIN_NB_MBUF))
100 #define NIC_PORT_RX_Q_SIZE 2048 /* Size of Physical NIC RX Queue, Max (n+32<=4096)*/
101 #define NIC_PORT_TX_Q_SIZE 2048 /* Size of Physical NIC TX Queue, Max (n+32<=4096)*/
103 static char *cuse_dev_name = NULL; /* Character device cuse_dev_name. */
104 static char *vhost_sock_dir = NULL; /* Location of vhost-user sockets */
107 * Maximum amount of time in micro seconds to try and enqueue to vhost.
109 #define VHOST_ENQ_RETRY_USECS 100
111 static const struct rte_eth_conf port_conf = {
113 .mq_mode = ETH_MQ_RX_RSS,
115 .header_split = 0, /* Header Split disabled */
116 .hw_ip_checksum = 0, /* IP checksum offload disabled */
117 .hw_vlan_filter = 0, /* VLAN filtering disabled */
118 .jumbo_frame = 0, /* Jumbo Frame Support disabled */
124 .rss_hf = ETH_RSS_IP | ETH_RSS_UDP | ETH_RSS_TCP,
128 .mq_mode = ETH_MQ_TX_NONE,
132 enum { MAX_TX_QUEUE_LEN = 384 };
133 enum { DPDK_RING_SIZE = 256 };
134 BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE));
135 enum { DRAIN_TSC = 200000ULL };
142 static int rte_eal_init_ret = ENODEV;
144 static struct ovs_mutex dpdk_mutex = OVS_MUTEX_INITIALIZER;
146 /* Contains all 'struct dpdk_dev's. */
147 static struct ovs_list dpdk_list OVS_GUARDED_BY(dpdk_mutex)
148 = OVS_LIST_INITIALIZER(&dpdk_list);
150 static struct ovs_list dpdk_mp_list OVS_GUARDED_BY(dpdk_mutex)
151 = OVS_LIST_INITIALIZER(&dpdk_mp_list);
153 /* This mutex must be used by non pmd threads when allocating or freeing
154 * mbufs through mempools. Since dpdk_queue_pkts() and dpdk_queue_flush() may
155 * use mempools, a non pmd thread should hold this mutex while calling them */
156 static struct ovs_mutex nonpmd_mempool_mutex = OVS_MUTEX_INITIALIZER;
159 struct rte_mempool *mp;
163 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
166 /* There should be one 'struct dpdk_tx_queue' created for
168 struct dpdk_tx_queue {
169 bool flush_tx; /* Set to true to flush queue everytime */
170 /* pkts are queued. */
172 rte_spinlock_t tx_lock; /* Protects the members and the NIC queue
173 * from concurrent access. It is used only
174 * if the queue is shared among different
175 * pmd threads (see 'txq_needs_locking'). */
177 struct rte_mbuf *burst_pkts[MAX_TX_QUEUE_LEN];
180 /* dpdk has no way to remove dpdk ring ethernet devices
181 so we have to keep them around once they've been created
184 static struct ovs_list dpdk_ring_list OVS_GUARDED_BY(dpdk_mutex)
185 = OVS_LIST_INITIALIZER(&dpdk_ring_list);
188 /* For the client rings */
189 struct rte_ring *cring_tx;
190 struct rte_ring *cring_rx;
191 unsigned int user_port_id; /* User given port no, parsed from port name */
192 int eth_port_id; /* ethernet device port id */
193 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
200 enum dpdk_dev_type type;
202 struct dpdk_tx_queue *tx_q;
204 struct ovs_mutex mutex OVS_ACQ_AFTER(dpdk_mutex);
206 struct dpdk_mp *dpdk_mp;
210 struct netdev_stats stats;
212 rte_spinlock_t stats_lock;
214 struct eth_addr hwaddr;
215 enum netdev_flags flags;
217 struct rte_eth_link link;
220 /* The user might request more txqs than the NIC has. We remap those
221 * ('up.n_txq') on these ('real_n_txq').
222 * If the numbers match, 'txq_needs_locking' is false, otherwise it is
223 * true and we will take a spinlock on transmission */
226 bool txq_needs_locking;
228 /* virtio-net structure for vhost device */
229 OVSRCU_TYPE(struct virtio_net *) virtio_dev;
231 /* Identifier used to distinguish vhost devices from each other */
232 char vhost_id[PATH_MAX];
235 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
238 struct netdev_rxq_dpdk {
239 struct netdev_rxq up;
243 static bool dpdk_thread_is_pmd(void);
245 static int netdev_dpdk_construct(struct netdev *);
247 struct virtio_net * netdev_dpdk_get_virtio(const struct netdev_dpdk *dev);
250 is_dpdk_class(const struct netdev_class *class)
252 return class->construct == netdev_dpdk_construct;
255 /* XXX: use dpdk malloc for entire OVS. in fact huge page should be used
256 * for all other segments data, bss and text. */
259 dpdk_rte_mzalloc(size_t sz)
263 ptr = rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE);
270 /* XXX this function should be called only by pmd threads (or by non pmd
271 * threads holding the nonpmd_mempool_mutex) */
273 free_dpdk_buf(struct dp_packet *p)
275 struct rte_mbuf *pkt = (struct rte_mbuf *) p;
277 rte_pktmbuf_free_seg(pkt);
281 __rte_pktmbuf_init(struct rte_mempool *mp,
282 void *opaque_arg OVS_UNUSED,
284 unsigned i OVS_UNUSED)
286 struct rte_mbuf *m = _m;
287 uint32_t buf_len = mp->elt_size - sizeof(struct dp_packet);
289 RTE_MBUF_ASSERT(mp->elt_size >= sizeof(struct dp_packet));
291 memset(m, 0, mp->elt_size);
293 /* start of buffer is just after mbuf structure */
294 m->buf_addr = (char *)m + sizeof(struct dp_packet);
295 m->buf_physaddr = rte_mempool_virt2phy(mp, m) +
296 sizeof(struct dp_packet);
297 m->buf_len = (uint16_t)buf_len;
299 /* keep some headroom between start of buffer and data */
300 m->data_off = RTE_MIN(RTE_PKTMBUF_HEADROOM, m->buf_len);
302 /* init some constant fields */
309 ovs_rte_pktmbuf_init(struct rte_mempool *mp,
310 void *opaque_arg OVS_UNUSED,
312 unsigned i OVS_UNUSED)
314 struct rte_mbuf *m = _m;
316 __rte_pktmbuf_init(mp, opaque_arg, _m, i);
318 dp_packet_init_dpdk((struct dp_packet *) m, m->buf_len);
321 static struct dpdk_mp *
322 dpdk_mp_get(int socket_id, int mtu) OVS_REQUIRES(dpdk_mutex)
324 struct dpdk_mp *dmp = NULL;
325 char mp_name[RTE_MEMPOOL_NAMESIZE];
328 LIST_FOR_EACH (dmp, list_node, &dpdk_mp_list) {
329 if (dmp->socket_id == socket_id && dmp->mtu == mtu) {
335 dmp = dpdk_rte_mzalloc(sizeof *dmp);
336 dmp->socket_id = socket_id;
340 mp_size = MAX_NB_MBUF;
342 if (snprintf(mp_name, RTE_MEMPOOL_NAMESIZE, "ovs_mp_%d_%d_%u",
343 dmp->mtu, dmp->socket_id, mp_size) < 0) {
347 dmp->mp = rte_mempool_create(mp_name, mp_size, MBUF_SIZE(mtu),
349 sizeof(struct rte_pktmbuf_pool_private),
350 rte_pktmbuf_pool_init, NULL,
351 ovs_rte_pktmbuf_init, NULL,
353 } while (!dmp->mp && rte_errno == ENOMEM && (mp_size /= 2) >= MIN_NB_MBUF);
355 if (dmp->mp == NULL) {
358 VLOG_DBG("Allocated \"%s\" mempool with %u mbufs", mp_name, mp_size );
361 list_push_back(&dpdk_mp_list, &dmp->list_node);
366 dpdk_mp_put(struct dpdk_mp *dmp)
374 ovs_assert(dmp->refcount >= 0);
377 /* I could not find any API to destroy mp. */
378 if (dmp->refcount == 0) {
379 list_delete(dmp->list_node);
380 /* destroy mp-pool. */
386 check_link_status(struct netdev_dpdk *dev)
388 struct rte_eth_link link;
390 rte_eth_link_get_nowait(dev->port_id, &link);
392 if (dev->link.link_status != link.link_status) {
393 netdev_change_seq_changed(&dev->up);
395 dev->link_reset_cnt++;
397 if (dev->link.link_status) {
398 VLOG_DBG_RL(&rl, "Port %d Link Up - speed %u Mbps - %s",
399 dev->port_id, (unsigned)dev->link.link_speed,
400 (dev->link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
401 ("full-duplex") : ("half-duplex"));
403 VLOG_DBG_RL(&rl, "Port %d Link Down", dev->port_id);
409 dpdk_watchdog(void *dummy OVS_UNUSED)
411 struct netdev_dpdk *dev;
413 pthread_detach(pthread_self());
416 ovs_mutex_lock(&dpdk_mutex);
417 LIST_FOR_EACH (dev, list_node, &dpdk_list) {
418 ovs_mutex_lock(&dev->mutex);
419 check_link_status(dev);
420 ovs_mutex_unlock(&dev->mutex);
422 ovs_mutex_unlock(&dpdk_mutex);
423 xsleep(DPDK_PORT_WATCHDOG_INTERVAL);
430 dpdk_eth_dev_queue_setup(struct netdev_dpdk *dev, int n_rxq, int n_txq)
435 /* A device may report more queues than it makes available (this has
436 * been observed for Intel xl710, which reserves some of them for
437 * SRIOV): rte_eth_*_queue_setup will fail if a queue is not
438 * available. When this happens we can retry the configuration
439 * and request less queues */
440 while (n_rxq && n_txq) {
442 VLOG_INFO("Retrying setup with (rxq:%d txq:%d)", n_rxq, n_txq);
445 diag = rte_eth_dev_configure(dev->port_id, n_rxq, n_txq, &port_conf);
450 for (i = 0; i < n_txq; i++) {
451 diag = rte_eth_tx_queue_setup(dev->port_id, i, NIC_PORT_TX_Q_SIZE,
452 dev->socket_id, NULL);
454 VLOG_INFO("Interface %s txq(%d) setup error: %s",
455 dev->up.name, i, rte_strerror(-diag));
461 /* Retry with less tx queues */
466 for (i = 0; i < n_rxq; i++) {
467 diag = rte_eth_rx_queue_setup(dev->port_id, i, NIC_PORT_RX_Q_SIZE,
468 dev->socket_id, NULL,
471 VLOG_INFO("Interface %s rxq(%d) setup error: %s",
472 dev->up.name, i, rte_strerror(-diag));
478 /* Retry with less rx queues */
483 dev->up.n_rxq = n_rxq;
484 dev->real_n_txq = n_txq;
494 dpdk_eth_dev_init(struct netdev_dpdk *dev) OVS_REQUIRES(dpdk_mutex)
496 struct rte_pktmbuf_pool_private *mbp_priv;
497 struct rte_eth_dev_info info;
498 struct ether_addr eth_addr;
502 if (dev->port_id < 0 || dev->port_id >= rte_eth_dev_count()) {
506 rte_eth_dev_info_get(dev->port_id, &info);
508 n_rxq = MIN(info.max_rx_queues, dev->up.n_rxq);
509 n_txq = MIN(info.max_tx_queues, dev->up.n_txq);
511 diag = dpdk_eth_dev_queue_setup(dev, n_rxq, n_txq);
513 VLOG_ERR("Interface %s(rxq:%d txq:%d) configure error: %s",
514 dev->up.name, n_rxq, n_txq, rte_strerror(-diag));
518 diag = rte_eth_dev_start(dev->port_id);
520 VLOG_ERR("Interface %s start error: %s", dev->up.name,
521 rte_strerror(-diag));
525 rte_eth_promiscuous_enable(dev->port_id);
526 rte_eth_allmulticast_enable(dev->port_id);
528 memset(ð_addr, 0x0, sizeof(eth_addr));
529 rte_eth_macaddr_get(dev->port_id, ð_addr);
530 VLOG_INFO_RL(&rl, "Port %d: "ETH_ADDR_FMT"",
531 dev->port_id, ETH_ADDR_BYTES_ARGS(eth_addr.addr_bytes));
533 memcpy(dev->hwaddr.ea, eth_addr.addr_bytes, ETH_ADDR_LEN);
534 rte_eth_link_get_nowait(dev->port_id, &dev->link);
536 mbp_priv = rte_mempool_get_priv(dev->dpdk_mp->mp);
537 dev->buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
539 dev->flags = NETDEV_UP | NETDEV_PROMISC;
543 static struct netdev_dpdk *
544 netdev_dpdk_cast(const struct netdev *netdev)
546 return CONTAINER_OF(netdev, struct netdev_dpdk, up);
549 static struct netdev *
550 netdev_dpdk_alloc(void)
552 struct netdev_dpdk *netdev = dpdk_rte_mzalloc(sizeof *netdev);
557 netdev_dpdk_alloc_txq(struct netdev_dpdk *netdev, unsigned int n_txqs)
561 netdev->tx_q = dpdk_rte_mzalloc(n_txqs * sizeof *netdev->tx_q);
562 for (i = 0; i < n_txqs; i++) {
563 int numa_id = ovs_numa_get_numa_id(i);
565 if (!netdev->txq_needs_locking) {
566 /* Each index is considered as a cpu core id, since there should
567 * be one tx queue for each cpu core. If the corresponding core
568 * is not on the same numa node as 'netdev', flags the
570 netdev->tx_q[i].flush_tx = netdev->socket_id == numa_id;
572 /* Queues are shared among CPUs. Always flush */
573 netdev->tx_q[i].flush_tx = true;
575 rte_spinlock_init(&netdev->tx_q[i].tx_lock);
580 netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no,
581 enum dpdk_dev_type type)
582 OVS_REQUIRES(dpdk_mutex)
584 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
588 ovs_mutex_init(&netdev->mutex);
589 ovs_mutex_lock(&netdev->mutex);
591 rte_spinlock_init(&netdev->stats_lock);
593 /* If the 'sid' is negative, it means that the kernel fails
594 * to obtain the pci numa info. In that situation, always
596 if (type == DPDK_DEV_ETH) {
597 sid = rte_eth_dev_socket_id(port_no);
599 sid = rte_lcore_to_socket_id(rte_get_master_lcore());
602 netdev->socket_id = sid < 0 ? SOCKET0 : sid;
603 netdev->port_id = port_no;
606 netdev->mtu = ETHER_MTU;
607 netdev->max_packet_len = MTU_TO_MAX_LEN(netdev->mtu);
609 netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, netdev->mtu);
610 if (!netdev->dpdk_mp) {
615 netdev_->n_txq = NR_QUEUE;
616 netdev_->n_rxq = NR_QUEUE;
617 netdev_->requested_n_rxq = NR_QUEUE;
618 netdev->real_n_txq = NR_QUEUE;
620 if (type == DPDK_DEV_ETH) {
621 netdev_dpdk_alloc_txq(netdev, NR_QUEUE);
622 err = dpdk_eth_dev_init(netdev);
628 list_push_back(&dpdk_list, &netdev->list_node);
632 rte_free(netdev->tx_q);
634 ovs_mutex_unlock(&netdev->mutex);
638 /* dev_name must be the prefix followed by a positive decimal number.
639 * (no leading + or - signs are allowed) */
641 dpdk_dev_parse_name(const char dev_name[], const char prefix[],
642 unsigned int *port_no)
646 if (strncmp(dev_name, prefix, strlen(prefix))) {
650 cport = dev_name + strlen(prefix);
652 if (str_to_uint(cport, 10, port_no)) {
660 vhost_construct_helper(struct netdev *netdev_) OVS_REQUIRES(dpdk_mutex)
662 if (rte_eal_init_ret) {
663 return rte_eal_init_ret;
666 return netdev_dpdk_init(netdev_, -1, DPDK_DEV_VHOST);
670 netdev_dpdk_vhost_cuse_construct(struct netdev *netdev_)
672 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
675 ovs_mutex_lock(&dpdk_mutex);
676 strncpy(netdev->vhost_id, netdev->up.name, sizeof(netdev->vhost_id));
677 err = vhost_construct_helper(netdev_);
678 ovs_mutex_unlock(&dpdk_mutex);
683 netdev_dpdk_vhost_user_construct(struct netdev *netdev_)
685 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
686 const char *name = netdev_->name;
689 /* 'name' is appended to 'vhost_sock_dir' and used to create a socket in
690 * the file system. '/' or '\' would traverse directories, so they're not
691 * acceptable in 'name'. */
692 if (strchr(name, '/') || strchr(name, '\\')) {
693 VLOG_ERR("\"%s\" is not a valid name for a vhost-user port. "
694 "A valid name must not include '/' or '\\'",
699 ovs_mutex_lock(&dpdk_mutex);
700 /* Take the name of the vhost-user port and append it to the location where
701 * the socket is to be created, then register the socket.
703 snprintf(netdev->vhost_id, sizeof(netdev->vhost_id), "%s/%s",
704 vhost_sock_dir, name);
706 err = rte_vhost_driver_register(netdev->vhost_id);
708 VLOG_ERR("vhost-user socket device setup failure for socket %s\n",
711 fatal_signal_add_file_to_unlink(netdev->vhost_id);
712 VLOG_INFO("Socket %s created for vhost-user port %s\n",
713 netdev->vhost_id, name);
714 err = vhost_construct_helper(netdev_);
717 ovs_mutex_unlock(&dpdk_mutex);
722 netdev_dpdk_construct(struct netdev *netdev)
724 unsigned int port_no;
727 if (rte_eal_init_ret) {
728 return rte_eal_init_ret;
731 /* Names always start with "dpdk" */
732 err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no);
737 ovs_mutex_lock(&dpdk_mutex);
738 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
739 ovs_mutex_unlock(&dpdk_mutex);
744 netdev_dpdk_destruct(struct netdev *netdev_)
746 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
748 ovs_mutex_lock(&dev->mutex);
749 rte_eth_dev_stop(dev->port_id);
750 ovs_mutex_unlock(&dev->mutex);
752 ovs_mutex_lock(&dpdk_mutex);
754 list_remove(&dev->list_node);
755 dpdk_mp_put(dev->dpdk_mp);
756 ovs_mutex_unlock(&dpdk_mutex);
760 netdev_dpdk_vhost_destruct(struct netdev *netdev_)
762 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
764 /* Can't remove a port while a guest is attached to it. */
765 if (netdev_dpdk_get_virtio(dev) != NULL) {
766 VLOG_ERR("Can not remove port, vhost device still attached");
770 if (rte_vhost_driver_unregister(dev->vhost_id)) {
771 VLOG_ERR("Unable to remove vhost-user socket %s", dev->vhost_id);
773 fatal_signal_remove_file_to_unlink(dev->vhost_id);
776 ovs_mutex_lock(&dpdk_mutex);
777 list_remove(&dev->list_node);
778 dpdk_mp_put(dev->dpdk_mp);
779 ovs_mutex_unlock(&dpdk_mutex);
783 netdev_dpdk_dealloc(struct netdev *netdev_)
785 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
791 netdev_dpdk_get_config(const struct netdev *netdev, struct smap *args)
793 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
795 ovs_mutex_lock(&dev->mutex);
797 smap_add_format(args, "requested_rx_queues", "%d", netdev->requested_n_rxq);
798 smap_add_format(args, "configured_rx_queues", "%d", netdev->n_rxq);
799 smap_add_format(args, "requested_tx_queues", "%d", netdev->n_txq);
800 smap_add_format(args, "configured_tx_queues", "%d", dev->real_n_txq);
801 ovs_mutex_unlock(&dev->mutex);
807 netdev_dpdk_set_config(struct netdev *netdev, const struct smap *args)
809 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
811 ovs_mutex_lock(&dev->mutex);
812 netdev->requested_n_rxq = MAX(smap_get_int(args, "n_rxq",
813 netdev->requested_n_rxq), 1);
814 netdev_change_seq_changed(netdev);
815 ovs_mutex_unlock(&dev->mutex);
821 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
823 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
825 return netdev->socket_id;
828 /* Sets the number of tx queues and rx queues for the dpdk interface.
829 * If the configuration fails, do not try restoring its old configuration
830 * and just returns the error. */
832 netdev_dpdk_set_multiq(struct netdev *netdev_, unsigned int n_txq,
835 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
837 int old_rxq, old_txq;
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 rte_eth_dev_stop(netdev->port_id);
848 old_txq = netdev->up.n_txq;
849 old_rxq = netdev->up.n_rxq;
850 netdev->up.n_txq = n_txq;
851 netdev->up.n_rxq = n_rxq;
853 rte_free(netdev->tx_q);
854 err = dpdk_eth_dev_init(netdev);
855 netdev_dpdk_alloc_txq(netdev, netdev->real_n_txq);
857 /* If there has been an error, it means that the requested queues
858 * have not been created. Restore the old numbers. */
859 netdev->up.n_txq = old_txq;
860 netdev->up.n_rxq = old_rxq;
863 netdev->txq_needs_locking = netdev->real_n_txq != netdev->up.n_txq;
865 ovs_mutex_unlock(&netdev->mutex);
866 ovs_mutex_unlock(&dpdk_mutex);
872 netdev_dpdk_vhost_cuse_set_multiq(struct netdev *netdev_, unsigned int n_txq,
875 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
878 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
882 ovs_mutex_lock(&dpdk_mutex);
883 ovs_mutex_lock(&netdev->mutex);
885 netdev->up.n_txq = n_txq;
886 netdev->real_n_txq = 1;
887 netdev->up.n_rxq = 1;
888 netdev->txq_needs_locking = netdev->real_n_txq != netdev->up.n_txq;
890 ovs_mutex_unlock(&netdev->mutex);
891 ovs_mutex_unlock(&dpdk_mutex);
897 netdev_dpdk_vhost_set_multiq(struct netdev *netdev_, unsigned int n_txq,
900 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
903 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
907 ovs_mutex_lock(&dpdk_mutex);
908 ovs_mutex_lock(&netdev->mutex);
910 rte_free(netdev->tx_q);
911 netdev->up.n_txq = n_txq;
912 netdev->up.n_rxq = n_rxq;
913 netdev_dpdk_alloc_txq(netdev, netdev->up.n_txq);
915 ovs_mutex_unlock(&netdev->mutex);
916 ovs_mutex_unlock(&dpdk_mutex);
921 static struct netdev_rxq *
922 netdev_dpdk_rxq_alloc(void)
924 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
929 static struct netdev_rxq_dpdk *
930 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
932 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
936 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
938 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
939 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
941 ovs_mutex_lock(&netdev->mutex);
942 rx->port_id = netdev->port_id;
943 ovs_mutex_unlock(&netdev->mutex);
949 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
954 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
956 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
962 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
964 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
967 while (nb_tx != txq->count) {
970 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
979 if (OVS_UNLIKELY(nb_tx != txq->count)) {
980 /* free buffers, which we couldn't transmit, one at a time (each
981 * packet could come from a different mempool) */
984 for (i = nb_tx; i < txq->count; i++) {
985 rte_pktmbuf_free_seg(txq->burst_pkts[i]);
987 rte_spinlock_lock(&dev->stats_lock);
988 dev->stats.tx_dropped += txq->count-nb_tx;
989 rte_spinlock_unlock(&dev->stats_lock);
993 txq->tsc = rte_get_timer_cycles();
997 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
999 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
1001 if (txq->count == 0) {
1004 dpdk_queue_flush__(dev, qid);
1008 is_vhost_running(struct virtio_net *dev)
1010 return (dev != NULL && (dev->flags & VIRTIO_DEV_RUNNING));
1014 netdev_dpdk_vhost_update_rx_counters(struct netdev_stats *stats,
1015 struct dp_packet **packets, int count)
1018 struct dp_packet *packet;
1020 stats->rx_packets += count;
1021 for (i = 0; i < count; i++) {
1022 packet = packets[i];
1024 if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) {
1025 /* This only protects the following multicast counting from
1026 * too short packets, but it does not stop the packet from
1027 * further processing. */
1029 stats->rx_length_errors++;
1033 struct eth_header *eh = (struct eth_header *) dp_packet_data(packet);
1034 if (OVS_UNLIKELY(eth_addr_is_multicast(eh->eth_dst))) {
1038 stats->rx_bytes += dp_packet_size(packet);
1043 * The receive path for the vhost port is the TX path out from guest.
1046 netdev_dpdk_vhost_rxq_recv(struct netdev_rxq *rxq_,
1047 struct dp_packet **packets, int *c)
1049 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
1050 struct netdev *netdev = rx->up.netdev;
1051 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
1052 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
1053 int qid = rxq_->queue_id;
1056 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
1060 if (rxq_->queue_id >= vhost_dev->real_n_rxq) {
1064 nb_rx = rte_vhost_dequeue_burst(virtio_dev, qid * VIRTIO_QNUM + VIRTIO_TXQ,
1065 vhost_dev->dpdk_mp->mp,
1066 (struct rte_mbuf **)packets,
1072 rte_spinlock_lock(&vhost_dev->stats_lock);
1073 netdev_dpdk_vhost_update_rx_counters(&vhost_dev->stats, packets, nb_rx);
1074 rte_spinlock_unlock(&vhost_dev->stats_lock);
1081 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dp_packet **packets,
1084 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
1085 struct netdev *netdev = rx->up.netdev;
1086 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1089 /* There is only one tx queue for this core. Do not flush other
1091 * Do not flush tx queue which is shared among CPUs
1092 * since it is always flushed */
1093 if (rxq_->queue_id == rte_lcore_id() &&
1094 OVS_LIKELY(!dev->txq_needs_locking)) {
1095 dpdk_queue_flush(dev, rxq_->queue_id);
1098 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
1099 (struct rte_mbuf **) packets,
1111 netdev_dpdk_vhost_update_tx_counters(struct netdev_stats *stats,
1112 struct dp_packet **packets,
1117 int sent = attempted - dropped;
1119 stats->tx_packets += sent;
1120 stats->tx_dropped += dropped;
1122 for (i = 0; i < sent; i++) {
1123 stats->tx_bytes += dp_packet_size(packets[i]);
1128 __netdev_dpdk_vhost_send(struct netdev *netdev, int qid,
1129 struct dp_packet **pkts, int cnt,
1132 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
1133 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
1134 struct rte_mbuf **cur_pkts = (struct rte_mbuf **) pkts;
1135 unsigned int total_pkts = cnt;
1138 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
1139 rte_spinlock_lock(&vhost_dev->stats_lock);
1140 vhost_dev->stats.tx_dropped+= cnt;
1141 rte_spinlock_unlock(&vhost_dev->stats_lock);
1145 if (vhost_dev->txq_needs_locking) {
1146 qid = qid % vhost_dev->real_n_txq;
1147 rte_spinlock_lock(&vhost_dev->tx_q[qid].tx_lock);
1151 int vhost_qid = qid * VIRTIO_QNUM + VIRTIO_RXQ;
1152 unsigned int tx_pkts;
1154 tx_pkts = rte_vhost_enqueue_burst(virtio_dev, vhost_qid,
1156 if (OVS_LIKELY(tx_pkts)) {
1157 /* Packets have been sent.*/
1159 /* Prepare for possible next iteration.*/
1160 cur_pkts = &cur_pkts[tx_pkts];
1162 uint64_t timeout = VHOST_ENQ_RETRY_USECS * rte_get_timer_hz() / 1E6;
1163 unsigned int expired = 0;
1166 start = rte_get_timer_cycles();
1170 * Unable to enqueue packets to vhost interface.
1171 * Check available entries before retrying.
1173 while (!rte_vring_available_entries(virtio_dev, vhost_qid)) {
1174 if (OVS_UNLIKELY((rte_get_timer_cycles() - start) > timeout)) {
1180 /* break out of main loop. */
1186 if (vhost_dev->txq_needs_locking) {
1187 rte_spinlock_unlock(&vhost_dev->tx_q[qid].tx_lock);
1190 rte_spinlock_lock(&vhost_dev->stats_lock);
1191 netdev_dpdk_vhost_update_tx_counters(&vhost_dev->stats, pkts, total_pkts,
1193 rte_spinlock_unlock(&vhost_dev->stats_lock);
1199 for (i = 0; i < total_pkts; i++) {
1200 dp_packet_delete(pkts[i]);
1206 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
1207 struct rte_mbuf **pkts, int cnt)
1209 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
1215 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
1216 int tocopy = MIN(freeslots, cnt-i);
1218 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
1219 tocopy * sizeof (struct rte_mbuf *));
1221 txq->count += tocopy;
1224 if (txq->count == MAX_TX_QUEUE_LEN || txq->flush_tx) {
1225 dpdk_queue_flush__(dev, qid);
1227 diff_tsc = rte_get_timer_cycles() - txq->tsc;
1228 if (diff_tsc >= DRAIN_TSC) {
1229 dpdk_queue_flush__(dev, qid);
1234 /* Tx function. Transmit packets indefinitely */
1236 dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet **pkts,
1238 OVS_NO_THREAD_SAFETY_ANALYSIS
1240 #if !defined(__CHECKER__) && !defined(_WIN32)
1241 const size_t PKT_ARRAY_SIZE = cnt;
1243 /* Sparse or MSVC doesn't like variable length array. */
1244 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
1246 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1247 struct rte_mbuf *mbufs[PKT_ARRAY_SIZE];
1252 /* If we are on a non pmd thread we have to use the mempool mutex, because
1253 * every non pmd thread shares the same mempool cache */
1255 if (!dpdk_thread_is_pmd()) {
1256 ovs_mutex_lock(&nonpmd_mempool_mutex);
1259 for (i = 0; i < cnt; i++) {
1260 int size = dp_packet_size(pkts[i]);
1262 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1263 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1264 (int)size , dev->max_packet_len);
1270 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
1272 if (!mbufs[newcnt]) {
1277 /* We have to do a copy for now */
1278 memcpy(rte_pktmbuf_mtod(mbufs[newcnt], void *), dp_packet_data(pkts[i]), size);
1280 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
1281 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
1286 if (OVS_UNLIKELY(dropped)) {
1287 rte_spinlock_lock(&dev->stats_lock);
1288 dev->stats.tx_dropped += dropped;
1289 rte_spinlock_unlock(&dev->stats_lock);
1292 if (dev->type == DPDK_DEV_VHOST) {
1293 __netdev_dpdk_vhost_send(netdev, qid, (struct dp_packet **) mbufs, newcnt, true);
1295 dpdk_queue_pkts(dev, qid, mbufs, newcnt);
1296 dpdk_queue_flush(dev, qid);
1299 if (!dpdk_thread_is_pmd()) {
1300 ovs_mutex_unlock(&nonpmd_mempool_mutex);
1305 netdev_dpdk_vhost_send(struct netdev *netdev, int qid, struct dp_packet **pkts,
1306 int cnt, bool may_steal)
1308 if (OVS_UNLIKELY(pkts[0]->source != DPBUF_DPDK)) {
1311 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1313 for (i = 0; i < cnt; i++) {
1314 dp_packet_delete(pkts[i]);
1318 __netdev_dpdk_vhost_send(netdev, qid, pkts, cnt, may_steal);
1324 netdev_dpdk_send__(struct netdev_dpdk *dev, int qid,
1325 struct dp_packet **pkts, int cnt, bool may_steal)
1329 if (OVS_UNLIKELY(dev->txq_needs_locking)) {
1330 qid = qid % dev->real_n_txq;
1331 rte_spinlock_lock(&dev->tx_q[qid].tx_lock);
1334 if (OVS_UNLIKELY(!may_steal ||
1335 pkts[0]->source != DPBUF_DPDK)) {
1336 struct netdev *netdev = &dev->up;
1338 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1341 for (i = 0; i < cnt; i++) {
1342 dp_packet_delete(pkts[i]);
1346 int next_tx_idx = 0;
1349 for (i = 0; i < cnt; i++) {
1350 int size = dp_packet_size(pkts[i]);
1352 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1353 if (next_tx_idx != i) {
1354 dpdk_queue_pkts(dev, qid,
1355 (struct rte_mbuf **)&pkts[next_tx_idx],
1359 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1360 (int)size , dev->max_packet_len);
1362 dp_packet_delete(pkts[i]);
1364 next_tx_idx = i + 1;
1367 if (next_tx_idx != cnt) {
1368 dpdk_queue_pkts(dev, qid,
1369 (struct rte_mbuf **)&pkts[next_tx_idx],
1373 if (OVS_UNLIKELY(dropped)) {
1374 rte_spinlock_lock(&dev->stats_lock);
1375 dev->stats.tx_dropped += dropped;
1376 rte_spinlock_unlock(&dev->stats_lock);
1380 if (OVS_UNLIKELY(dev->txq_needs_locking)) {
1381 rte_spinlock_unlock(&dev->tx_q[qid].tx_lock);
1386 netdev_dpdk_eth_send(struct netdev *netdev, int qid,
1387 struct dp_packet **pkts, int cnt, bool may_steal)
1389 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1391 netdev_dpdk_send__(dev, qid, pkts, cnt, may_steal);
1396 netdev_dpdk_set_etheraddr(struct netdev *netdev, const struct eth_addr mac)
1398 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1400 ovs_mutex_lock(&dev->mutex);
1401 if (!eth_addr_equals(dev->hwaddr, mac)) {
1403 netdev_change_seq_changed(netdev);
1405 ovs_mutex_unlock(&dev->mutex);
1411 netdev_dpdk_get_etheraddr(const struct netdev *netdev, struct eth_addr *mac)
1413 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1415 ovs_mutex_lock(&dev->mutex);
1417 ovs_mutex_unlock(&dev->mutex);
1423 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
1425 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1427 ovs_mutex_lock(&dev->mutex);
1429 ovs_mutex_unlock(&dev->mutex);
1435 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
1437 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1439 struct dpdk_mp *old_mp;
1442 ovs_mutex_lock(&dpdk_mutex);
1443 ovs_mutex_lock(&dev->mutex);
1444 if (dev->mtu == mtu) {
1449 mp = dpdk_mp_get(dev->socket_id, dev->mtu);
1455 rte_eth_dev_stop(dev->port_id);
1458 old_mp = dev->dpdk_mp;
1461 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1463 err = dpdk_eth_dev_init(dev);
1467 dev->dpdk_mp = old_mp;
1468 dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu);
1469 dpdk_eth_dev_init(dev);
1473 dpdk_mp_put(old_mp);
1474 netdev_change_seq_changed(netdev);
1476 ovs_mutex_unlock(&dev->mutex);
1477 ovs_mutex_unlock(&dpdk_mutex);
1482 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
1485 netdev_dpdk_vhost_get_stats(const struct netdev *netdev,
1486 struct netdev_stats *stats)
1488 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1490 ovs_mutex_lock(&dev->mutex);
1491 memset(stats, 0, sizeof(*stats));
1492 /* Unsupported Stats */
1493 stats->collisions = UINT64_MAX;
1494 stats->rx_crc_errors = UINT64_MAX;
1495 stats->rx_fifo_errors = UINT64_MAX;
1496 stats->rx_frame_errors = UINT64_MAX;
1497 stats->rx_missed_errors = UINT64_MAX;
1498 stats->rx_over_errors = UINT64_MAX;
1499 stats->tx_aborted_errors = UINT64_MAX;
1500 stats->tx_carrier_errors = UINT64_MAX;
1501 stats->tx_errors = UINT64_MAX;
1502 stats->tx_fifo_errors = UINT64_MAX;
1503 stats->tx_heartbeat_errors = UINT64_MAX;
1504 stats->tx_window_errors = UINT64_MAX;
1505 stats->rx_dropped += UINT64_MAX;
1507 rte_spinlock_lock(&dev->stats_lock);
1508 /* Supported Stats */
1509 stats->rx_packets += dev->stats.rx_packets;
1510 stats->tx_packets += dev->stats.tx_packets;
1511 stats->tx_dropped += dev->stats.tx_dropped;
1512 stats->multicast = dev->stats.multicast;
1513 stats->rx_bytes = dev->stats.rx_bytes;
1514 stats->tx_bytes = dev->stats.tx_bytes;
1515 stats->rx_errors = dev->stats.rx_errors;
1516 stats->rx_length_errors = dev->stats.rx_length_errors;
1517 rte_spinlock_unlock(&dev->stats_lock);
1519 ovs_mutex_unlock(&dev->mutex);
1525 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
1527 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1528 struct rte_eth_stats rte_stats;
1531 netdev_dpdk_get_carrier(netdev, &gg);
1532 ovs_mutex_lock(&dev->mutex);
1533 rte_eth_stats_get(dev->port_id, &rte_stats);
1535 memset(stats, 0, sizeof(*stats));
1537 stats->rx_packets = rte_stats.ipackets;
1538 stats->tx_packets = rte_stats.opackets;
1539 stats->rx_bytes = rte_stats.ibytes;
1540 stats->tx_bytes = rte_stats.obytes;
1541 /* DPDK counts imissed as errors, but count them here as dropped instead */
1542 stats->rx_errors = rte_stats.ierrors - rte_stats.imissed;
1543 stats->tx_errors = rte_stats.oerrors;
1544 stats->multicast = rte_stats.imcasts;
1546 rte_spinlock_lock(&dev->stats_lock);
1547 stats->tx_dropped = dev->stats.tx_dropped;
1548 rte_spinlock_unlock(&dev->stats_lock);
1550 /* These are the available DPDK counters for packets not received due to
1551 * local resource constraints in DPDK and NIC respectively. */
1552 stats->rx_dropped = rte_stats.rx_nombuf + rte_stats.imissed;
1553 stats->collisions = UINT64_MAX;
1555 stats->rx_length_errors = UINT64_MAX;
1556 stats->rx_over_errors = UINT64_MAX;
1557 stats->rx_crc_errors = UINT64_MAX;
1558 stats->rx_frame_errors = UINT64_MAX;
1559 stats->rx_fifo_errors = UINT64_MAX;
1560 stats->rx_missed_errors = rte_stats.imissed;
1562 stats->tx_aborted_errors = UINT64_MAX;
1563 stats->tx_carrier_errors = UINT64_MAX;
1564 stats->tx_fifo_errors = UINT64_MAX;
1565 stats->tx_heartbeat_errors = UINT64_MAX;
1566 stats->tx_window_errors = UINT64_MAX;
1568 ovs_mutex_unlock(&dev->mutex);
1574 netdev_dpdk_get_features(const struct netdev *netdev_,
1575 enum netdev_features *current,
1576 enum netdev_features *advertised OVS_UNUSED,
1577 enum netdev_features *supported OVS_UNUSED,
1578 enum netdev_features *peer OVS_UNUSED)
1580 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1581 struct rte_eth_link link;
1583 ovs_mutex_lock(&dev->mutex);
1585 ovs_mutex_unlock(&dev->mutex);
1587 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
1588 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
1589 *current = NETDEV_F_AUTONEG;
1591 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
1592 if (link.link_speed == ETH_LINK_SPEED_10) {
1593 *current = NETDEV_F_10MB_HD;
1595 if (link.link_speed == ETH_LINK_SPEED_100) {
1596 *current = NETDEV_F_100MB_HD;
1598 if (link.link_speed == ETH_LINK_SPEED_1000) {
1599 *current = NETDEV_F_1GB_HD;
1601 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
1602 if (link.link_speed == ETH_LINK_SPEED_10) {
1603 *current = NETDEV_F_10MB_FD;
1605 if (link.link_speed == ETH_LINK_SPEED_100) {
1606 *current = NETDEV_F_100MB_FD;
1608 if (link.link_speed == ETH_LINK_SPEED_1000) {
1609 *current = NETDEV_F_1GB_FD;
1611 if (link.link_speed == ETH_LINK_SPEED_10000) {
1612 *current = NETDEV_F_10GB_FD;
1620 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1622 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1625 ovs_mutex_lock(&dev->mutex);
1626 ifindex = dev->port_id;
1627 ovs_mutex_unlock(&dev->mutex);
1633 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1635 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1637 ovs_mutex_lock(&dev->mutex);
1638 check_link_status(dev);
1639 *carrier = dev->link.link_status;
1641 ovs_mutex_unlock(&dev->mutex);
1647 netdev_dpdk_vhost_get_carrier(const struct netdev *netdev_, bool *carrier)
1649 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1650 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(dev);
1652 ovs_mutex_lock(&dev->mutex);
1654 if (is_vhost_running(virtio_dev)) {
1660 ovs_mutex_unlock(&dev->mutex);
1665 static long long int
1666 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1668 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1669 long long int carrier_resets;
1671 ovs_mutex_lock(&dev->mutex);
1672 carrier_resets = dev->link_reset_cnt;
1673 ovs_mutex_unlock(&dev->mutex);
1675 return carrier_resets;
1679 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1680 long long int interval OVS_UNUSED)
1686 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1687 enum netdev_flags off, enum netdev_flags on,
1688 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1692 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1696 *old_flagsp = dev->flags;
1700 if (dev->flags == *old_flagsp) {
1704 if (dev->type == DPDK_DEV_ETH) {
1705 if (dev->flags & NETDEV_UP) {
1706 err = rte_eth_dev_start(dev->port_id);
1711 if (dev->flags & NETDEV_PROMISC) {
1712 rte_eth_promiscuous_enable(dev->port_id);
1715 if (!(dev->flags & NETDEV_UP)) {
1716 rte_eth_dev_stop(dev->port_id);
1724 netdev_dpdk_update_flags(struct netdev *netdev_,
1725 enum netdev_flags off, enum netdev_flags on,
1726 enum netdev_flags *old_flagsp)
1728 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1731 ovs_mutex_lock(&netdev->mutex);
1732 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1733 ovs_mutex_unlock(&netdev->mutex);
1739 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1741 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1742 struct rte_eth_dev_info dev_info;
1744 if (dev->port_id < 0)
1747 ovs_mutex_lock(&dev->mutex);
1748 rte_eth_dev_info_get(dev->port_id, &dev_info);
1749 ovs_mutex_unlock(&dev->mutex);
1751 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1753 smap_add_format(args, "port_no", "%d", dev->port_id);
1754 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1755 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1756 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1757 smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen);
1758 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1759 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1760 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1761 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1762 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1763 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1765 if (dev_info.pci_dev) {
1766 smap_add_format(args, "pci-vendor_id", "0x%u",
1767 dev_info.pci_dev->id.vendor_id);
1768 smap_add_format(args, "pci-device_id", "0x%x",
1769 dev_info.pci_dev->id.device_id);
1776 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1777 OVS_REQUIRES(dev->mutex)
1779 enum netdev_flags old_flags;
1782 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1784 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1789 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1790 const char *argv[], void *aux OVS_UNUSED)
1794 if (!strcasecmp(argv[argc - 1], "up")) {
1796 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1799 unixctl_command_reply_error(conn, "Invalid Admin State");
1804 struct netdev *netdev = netdev_from_name(argv[1]);
1805 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1806 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1808 ovs_mutex_lock(&dpdk_dev->mutex);
1809 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1810 ovs_mutex_unlock(&dpdk_dev->mutex);
1812 netdev_close(netdev);
1814 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1815 netdev_close(netdev);
1819 struct netdev_dpdk *netdev;
1821 ovs_mutex_lock(&dpdk_mutex);
1822 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1823 ovs_mutex_lock(&netdev->mutex);
1824 netdev_dpdk_set_admin_state__(netdev, up);
1825 ovs_mutex_unlock(&netdev->mutex);
1827 ovs_mutex_unlock(&dpdk_mutex);
1829 unixctl_command_reply(conn, "OK");
1833 * Set virtqueue flags so that we do not receive interrupts.
1836 set_irq_status(struct virtio_net *dev)
1841 for (i = 0; i < dev->virt_qp_nb; i++) {
1842 idx = i * VIRTIO_QNUM;
1843 rte_vhost_enable_guest_notification(dev, idx + VIRTIO_RXQ, 0);
1844 rte_vhost_enable_guest_notification(dev, idx + VIRTIO_TXQ, 0);
1850 netdev_dpdk_vhost_set_queues(struct netdev_dpdk *netdev, struct virtio_net *dev)
1854 qp_num = dev->virt_qp_nb;
1855 if (qp_num > netdev->up.n_rxq) {
1856 VLOG_ERR("vHost Device '%s' %"PRIu64" can't be added - "
1857 "too many queues %d > %d", dev->ifname, dev->device_fh,
1858 qp_num, netdev->up.n_rxq);
1862 netdev->real_n_rxq = qp_num;
1863 netdev->real_n_txq = qp_num;
1864 if (netdev->up.n_txq > netdev->real_n_txq) {
1865 netdev->txq_needs_locking = true;
1867 netdev->txq_needs_locking = false;
1874 * A new virtio-net device is added to a vhost port.
1877 new_device(struct virtio_net *dev)
1879 struct netdev_dpdk *netdev;
1880 bool exists = false;
1882 ovs_mutex_lock(&dpdk_mutex);
1883 /* Add device to the vhost port with the same name as that passed down. */
1884 LIST_FOR_EACH(netdev, list_node, &dpdk_list) {
1885 if (strncmp(dev->ifname, netdev->vhost_id, IF_NAME_SZ) == 0) {
1886 ovs_mutex_lock(&netdev->mutex);
1887 if (netdev_dpdk_vhost_set_queues(netdev, dev)) {
1888 ovs_mutex_unlock(&netdev->mutex);
1889 ovs_mutex_unlock(&dpdk_mutex);
1892 ovsrcu_set(&netdev->virtio_dev, dev);
1894 dev->flags |= VIRTIO_DEV_RUNNING;
1895 /* Disable notifications. */
1896 set_irq_status(dev);
1897 ovs_mutex_unlock(&netdev->mutex);
1901 ovs_mutex_unlock(&dpdk_mutex);
1904 VLOG_INFO("vHost Device '%s' %"PRIu64" can't be added - name not "
1905 "found", dev->ifname, dev->device_fh);
1910 VLOG_INFO("vHost Device '%s' %"PRIu64" has been added", dev->ifname,
1916 * Remove a virtio-net device from the specific vhost port. Use dev->remove
1917 * flag to stop any more packets from being sent or received to/from a VM and
1918 * ensure all currently queued packets have been sent/received before removing
1922 destroy_device(volatile struct virtio_net *dev)
1924 struct netdev_dpdk *vhost_dev;
1925 bool exists = false;
1927 ovs_mutex_lock(&dpdk_mutex);
1928 LIST_FOR_EACH (vhost_dev, list_node, &dpdk_list) {
1929 if (netdev_dpdk_get_virtio(vhost_dev) == dev) {
1931 ovs_mutex_lock(&vhost_dev->mutex);
1932 dev->flags &= ~VIRTIO_DEV_RUNNING;
1933 ovsrcu_set(&vhost_dev->virtio_dev, NULL);
1935 ovs_mutex_unlock(&vhost_dev->mutex);
1940 ovs_mutex_unlock(&dpdk_mutex);
1942 if (exists == true) {
1944 * Wait for other threads to quiesce after setting the 'virtio_dev'
1945 * to NULL, before returning.
1947 ovsrcu_synchronize();
1949 * As call to ovsrcu_synchronize() will end the quiescent state,
1950 * put thread back into quiescent state before returning.
1952 ovsrcu_quiesce_start();
1953 VLOG_INFO("vHost Device '%s' %"PRIu64" has been removed", dev->ifname,
1956 VLOG_INFO("vHost Device '%s' %"PRIu64" not found", dev->ifname,
1963 netdev_dpdk_get_virtio(const struct netdev_dpdk *dev)
1965 return ovsrcu_get(struct virtio_net *, &dev->virtio_dev);
1969 * These callbacks allow virtio-net devices to be added to vhost ports when
1970 * configuration has been fully complete.
1972 static const struct virtio_net_device_ops virtio_net_device_ops =
1974 .new_device = new_device,
1975 .destroy_device = destroy_device,
1979 start_vhost_loop(void *dummy OVS_UNUSED)
1981 pthread_detach(pthread_self());
1982 /* Put the cuse thread into quiescent state. */
1983 ovsrcu_quiesce_start();
1984 rte_vhost_driver_session_start();
1989 dpdk_vhost_class_init(void)
1991 rte_vhost_driver_callback_register(&virtio_net_device_ops);
1992 ovs_thread_create("vhost_thread", start_vhost_loop, NULL);
1997 dpdk_vhost_cuse_class_init(void)
2002 /* Register CUSE device to handle IOCTLs.
2003 * Unless otherwise specified on the vswitchd command line, cuse_dev_name
2004 * is set to vhost-net.
2006 err = rte_vhost_driver_register(cuse_dev_name);
2009 VLOG_ERR("CUSE device setup failure.");
2013 dpdk_vhost_class_init();
2018 dpdk_vhost_user_class_init(void)
2020 dpdk_vhost_class_init();
2025 dpdk_common_init(void)
2027 unixctl_command_register("netdev-dpdk/set-admin-state",
2028 "[netdev] up|down", 1, 2,
2029 netdev_dpdk_set_admin_state, NULL);
2031 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
2037 dpdk_ring_create(const char dev_name[], unsigned int port_no,
2038 unsigned int *eth_port_id)
2040 struct dpdk_ring *ivshmem;
2041 char ring_name[RTE_RING_NAMESIZE];
2044 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
2045 if (ivshmem == NULL) {
2049 /* XXX: Add support for multiquque ring. */
2050 err = snprintf(ring_name, sizeof(ring_name), "%s_tx", dev_name);
2055 /* Create single producer tx ring, netdev does explicit locking. */
2056 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
2058 if (ivshmem->cring_tx == NULL) {
2063 err = snprintf(ring_name, sizeof(ring_name), "%s_rx", dev_name);
2068 /* Create single consumer rx ring, netdev does explicit locking. */
2069 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
2071 if (ivshmem->cring_rx == NULL) {
2076 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
2077 &ivshmem->cring_tx, 1, SOCKET0);
2084 ivshmem->user_port_id = port_no;
2085 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
2086 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
2088 *eth_port_id = ivshmem->eth_port_id;
2093 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
2095 struct dpdk_ring *ivshmem;
2096 unsigned int port_no;
2099 /* Names always start with "dpdkr" */
2100 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
2105 /* look through our list to find the device */
2106 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
2107 if (ivshmem->user_port_id == port_no) {
2108 VLOG_INFO("Found dpdk ring device %s:", dev_name);
2109 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
2113 /* Need to create the device rings */
2114 return dpdk_ring_create(dev_name, port_no, eth_port_id);
2118 netdev_dpdk_ring_send(struct netdev *netdev_, int qid,
2119 struct dp_packet **pkts, int cnt, bool may_steal)
2121 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
2124 /* When using 'dpdkr' and sending to a DPDK ring, we want to ensure that the
2125 * rss hash field is clear. This is because the same mbuf may be modified by
2126 * the consumer of the ring and return into the datapath without recalculating
2128 for (i = 0; i < cnt; i++) {
2129 dp_packet_rss_invalidate(pkts[i]);
2132 netdev_dpdk_send__(netdev, qid, pkts, cnt, may_steal);
2137 netdev_dpdk_ring_construct(struct netdev *netdev)
2139 unsigned int port_no = 0;
2142 if (rte_eal_init_ret) {
2143 return rte_eal_init_ret;
2146 ovs_mutex_lock(&dpdk_mutex);
2148 err = dpdk_ring_open(netdev->name, &port_no);
2153 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
2156 ovs_mutex_unlock(&dpdk_mutex);
2160 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, DESTRUCT, MULTIQ, SEND, \
2161 GET_CARRIER, GET_STATS, GET_FEATURES, GET_STATUS, RXQ_RECV) \
2165 NULL, /* netdev_dpdk_run */ \
2166 NULL, /* netdev_dpdk_wait */ \
2168 netdev_dpdk_alloc, \
2171 netdev_dpdk_dealloc, \
2172 netdev_dpdk_get_config, \
2173 netdev_dpdk_set_config, \
2174 NULL, /* get_tunnel_config */ \
2175 NULL, /* build header */ \
2176 NULL, /* push header */ \
2177 NULL, /* pop header */ \
2178 netdev_dpdk_get_numa_id, /* get_numa_id */ \
2179 MULTIQ, /* set_multiq */ \
2182 NULL, /* send_wait */ \
2184 netdev_dpdk_set_etheraddr, \
2185 netdev_dpdk_get_etheraddr, \
2186 netdev_dpdk_get_mtu, \
2187 netdev_dpdk_set_mtu, \
2188 netdev_dpdk_get_ifindex, \
2190 netdev_dpdk_get_carrier_resets, \
2191 netdev_dpdk_set_miimon, \
2194 NULL, /* set_advertisements */ \
2196 NULL, /* set_policing */ \
2197 NULL, /* get_qos_types */ \
2198 NULL, /* get_qos_capabilities */ \
2199 NULL, /* get_qos */ \
2200 NULL, /* set_qos */ \
2201 NULL, /* get_queue */ \
2202 NULL, /* set_queue */ \
2203 NULL, /* delete_queue */ \
2204 NULL, /* get_queue_stats */ \
2205 NULL, /* queue_dump_start */ \
2206 NULL, /* queue_dump_next */ \
2207 NULL, /* queue_dump_done */ \
2208 NULL, /* dump_queue_stats */ \
2210 NULL, /* get_in4 */ \
2211 NULL, /* set_in4 */ \
2212 NULL, /* get_in6 */ \
2213 NULL, /* add_router */ \
2214 NULL, /* get_next_hop */ \
2216 NULL, /* arp_lookup */ \
2218 netdev_dpdk_update_flags, \
2220 netdev_dpdk_rxq_alloc, \
2221 netdev_dpdk_rxq_construct, \
2222 netdev_dpdk_rxq_destruct, \
2223 netdev_dpdk_rxq_dealloc, \
2225 NULL, /* rx_wait */ \
2226 NULL, /* rxq_drain */ \
2230 process_vhost_flags(char *flag, char *default_val, int size,
2231 char **argv, char **new_val)
2235 /* Depending on which version of vhost is in use, process the vhost-specific
2236 * flag if it is provided on the vswitchd command line, otherwise resort to
2239 * For vhost-user: Process "-vhost_sock_dir" to set the custom location of
2240 * the vhost-user socket(s).
2241 * For vhost-cuse: Process "-cuse_dev_name" to set the custom name of the
2242 * vhost-cuse character device.
2244 if (!strcmp(argv[1], flag) && (strlen(argv[2]) <= size)) {
2246 *new_val = xstrdup(argv[2]);
2247 VLOG_INFO("User-provided %s in use: %s", flag, *new_val);
2249 VLOG_INFO("No %s provided - defaulting to %s", flag, default_val);
2250 *new_val = default_val;
2257 dpdk_init(int argc, char **argv)
2261 char *pragram_name = argv[0];
2263 if (argc < 2 || strcmp(argv[1], "--dpdk"))
2266 /* Remove the --dpdk argument from arg list.*/
2270 /* Reject --user option */
2272 for (i = 0; i < argc; i++) {
2273 if (!strcmp(argv[i], "--user")) {
2274 VLOG_ERR("Can not mix --dpdk and --user options, aborting.");
2279 if (process_vhost_flags("-cuse_dev_name", xstrdup("vhost-net"),
2280 PATH_MAX, argv, &cuse_dev_name)) {
2282 if (process_vhost_flags("-vhost_sock_dir", xstrdup(ovs_rundir()),
2283 NAME_MAX, argv, &vhost_sock_dir)) {
2287 err = stat(vhost_sock_dir, &s);
2289 VLOG_ERR("vHostUser socket DIR '%s' does not exist.",
2294 /* Remove the vhost flag configuration parameters from the argument
2295 * list, so that the correct elements are passed to the DPDK
2296 * initialization function
2299 argv += 2; /* Increment by two to bypass the vhost flag arguments */
2303 /* Keep the program name argument as this is needed for call to
2306 argv[0] = pragram_name;
2308 /* Make sure things are initialized ... */
2309 result = rte_eal_init(argc, argv);
2311 ovs_abort(result, "Cannot init EAL");
2314 rte_memzone_dump(stdout);
2315 rte_eal_init_ret = 0;
2317 if (argc > result) {
2318 argv[result] = argv[0];
2321 /* We are called from the main thread here */
2322 RTE_PER_LCORE(_lcore_id) = NON_PMD_CORE_ID;
2324 return result + 1 + base;
2327 static const struct netdev_class dpdk_class =
2331 netdev_dpdk_construct,
2332 netdev_dpdk_destruct,
2333 netdev_dpdk_set_multiq,
2334 netdev_dpdk_eth_send,
2335 netdev_dpdk_get_carrier,
2336 netdev_dpdk_get_stats,
2337 netdev_dpdk_get_features,
2338 netdev_dpdk_get_status,
2339 netdev_dpdk_rxq_recv);
2341 static const struct netdev_class dpdk_ring_class =
2345 netdev_dpdk_ring_construct,
2346 netdev_dpdk_destruct,
2347 netdev_dpdk_set_multiq,
2348 netdev_dpdk_ring_send,
2349 netdev_dpdk_get_carrier,
2350 netdev_dpdk_get_stats,
2351 netdev_dpdk_get_features,
2352 netdev_dpdk_get_status,
2353 netdev_dpdk_rxq_recv);
2355 static const struct netdev_class OVS_UNUSED dpdk_vhost_cuse_class =
2358 dpdk_vhost_cuse_class_init,
2359 netdev_dpdk_vhost_cuse_construct,
2360 netdev_dpdk_vhost_destruct,
2361 netdev_dpdk_vhost_cuse_set_multiq,
2362 netdev_dpdk_vhost_send,
2363 netdev_dpdk_vhost_get_carrier,
2364 netdev_dpdk_vhost_get_stats,
2367 netdev_dpdk_vhost_rxq_recv);
2369 static const struct netdev_class OVS_UNUSED dpdk_vhost_user_class =
2372 dpdk_vhost_user_class_init,
2373 netdev_dpdk_vhost_user_construct,
2374 netdev_dpdk_vhost_destruct,
2375 netdev_dpdk_vhost_set_multiq,
2376 netdev_dpdk_vhost_send,
2377 netdev_dpdk_vhost_get_carrier,
2378 netdev_dpdk_vhost_get_stats,
2381 netdev_dpdk_vhost_rxq_recv);
2384 netdev_dpdk_register(void)
2386 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
2388 if (rte_eal_init_ret) {
2392 if (ovsthread_once_start(&once)) {
2394 netdev_register_provider(&dpdk_class);
2395 netdev_register_provider(&dpdk_ring_class);
2397 netdev_register_provider(&dpdk_vhost_cuse_class);
2399 netdev_register_provider(&dpdk_vhost_user_class);
2401 ovsthread_once_done(&once);
2406 pmd_thread_setaffinity_cpu(unsigned cpu)
2412 CPU_SET(cpu, &cpuset);
2413 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
2415 VLOG_ERR("Thread affinity error %d",err);
2418 /* NON_PMD_CORE_ID is reserved for use by non pmd threads. */
2419 ovs_assert(cpu != NON_PMD_CORE_ID);
2420 RTE_PER_LCORE(_lcore_id) = cpu;
2426 dpdk_thread_is_pmd(void)
2428 return rte_lcore_id() != NON_PMD_CORE_ID;