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"
50 #include "unaligned.h"
53 #include "openvswitch/vlog.h"
55 #include "rte_config.h"
57 #include "rte_meter.h"
58 #include "rte_virtio_net.h"
60 VLOG_DEFINE_THIS_MODULE(dpdk);
61 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
63 #define DPDK_PORT_WATCHDOG_INTERVAL 5
65 #define OVS_CACHE_LINE_SIZE CACHE_LINE_SIZE
66 #define OVS_VPORT_DPDK "ovs_dpdk"
69 * need to reserve tons of extra space in the mbufs so we can align the
70 * DMA addresses to 4KB.
71 * The minimum mbuf size is limited to avoid scatter behaviour and drop in
72 * performance for standard Ethernet MTU.
74 #define ETHER_HDR_MAX_LEN (ETHER_HDR_LEN + ETHER_CRC_LEN + (2 * VLAN_HEADER_LEN))
75 #define MTU_TO_FRAME_LEN(mtu) ((mtu) + ETHER_HDR_LEN + ETHER_CRC_LEN)
76 #define MTU_TO_MAX_FRAME_LEN(mtu) ((mtu) + ETHER_HDR_MAX_LEN)
77 #define FRAME_LEN_TO_MTU(frame_len) ((frame_len)- ETHER_HDR_LEN - ETHER_CRC_LEN)
78 #define MBUF_SIZE(mtu) ( MTU_TO_MAX_FRAME_LEN(mtu) \
79 + sizeof(struct dp_packet) \
80 + RTE_PKTMBUF_HEADROOM)
81 #define NETDEV_DPDK_MBUF_ALIGN 1024
83 /* Max and min number of packets in the mempool. OVS tries to allocate a
84 * mempool with MAX_NB_MBUF: if this fails (because the system doesn't have
85 * enough hugepages) we keep halving the number until the allocation succeeds
86 * or we reach MIN_NB_MBUF */
88 #define MAX_NB_MBUF (4096 * 64)
89 #define MIN_NB_MBUF (4096 * 4)
90 #define MP_CACHE_SZ RTE_MEMPOOL_CACHE_MAX_SIZE
92 /* MAX_NB_MBUF can be divided by 2 many times, until MIN_NB_MBUF */
93 BUILD_ASSERT_DECL(MAX_NB_MBUF % ROUND_DOWN_POW2(MAX_NB_MBUF/MIN_NB_MBUF) == 0);
95 /* The smallest possible NB_MBUF that we're going to try should be a multiple
96 * of MP_CACHE_SZ. This is advised by DPDK documentation. */
97 BUILD_ASSERT_DECL((MAX_NB_MBUF / ROUND_DOWN_POW2(MAX_NB_MBUF/MIN_NB_MBUF))
102 #define NIC_PORT_RX_Q_SIZE 2048 /* Size of Physical NIC RX Queue, Max (n+32<=4096)*/
103 #define NIC_PORT_TX_Q_SIZE 2048 /* Size of Physical NIC TX Queue, Max (n+32<=4096)*/
105 #define OVS_VHOST_MAX_QUEUE_NUM 1024 /* Maximum number of vHost TX queues. */
107 static char *cuse_dev_name = NULL; /* Character device cuse_dev_name. */
108 static char *vhost_sock_dir = NULL; /* Location of vhost-user sockets */
111 * Maximum amount of time in micro seconds to try and enqueue to vhost.
113 #define VHOST_ENQ_RETRY_USECS 100
115 static const struct rte_eth_conf port_conf = {
117 .mq_mode = ETH_MQ_RX_RSS,
119 .header_split = 0, /* Header Split disabled */
120 .hw_ip_checksum = 0, /* IP checksum offload disabled */
121 .hw_vlan_filter = 0, /* VLAN filtering disabled */
122 .jumbo_frame = 0, /* Jumbo Frame Support disabled */
128 .rss_hf = ETH_RSS_IP | ETH_RSS_UDP | ETH_RSS_TCP,
132 .mq_mode = ETH_MQ_TX_NONE,
136 enum { MAX_TX_QUEUE_LEN = 384 };
137 enum { DPDK_RING_SIZE = 256 };
138 BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE));
139 enum { DRAIN_TSC = 200000ULL };
146 static int rte_eal_init_ret = ENODEV;
148 static struct ovs_mutex dpdk_mutex = OVS_MUTEX_INITIALIZER;
150 /* Quality of Service */
152 /* An instance of a QoS configuration. Always associated with a particular
155 * Each QoS implementation subclasses this with whatever additional data it
159 const struct dpdk_qos_ops *ops;
162 /* A particular implementation of dpdk QoS operations.
164 * The functions below return 0 if successful or a positive errno value on
165 * failure, except where otherwise noted. All of them must be provided, except
166 * where otherwise noted.
168 struct dpdk_qos_ops {
170 /* Name of the QoS type */
171 const char *qos_name;
173 /* Called to construct the QoS implementation on 'netdev'. The
174 * implementation should make the appropriate calls to configure QoS
175 * according to 'details'. The implementation may assume that any current
176 * QoS configuration already installed should be destroyed before
177 * constructing the new configuration.
179 * The contents of 'details' should be documented as valid for 'ovs_name'
180 * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml
181 * (which is built as ovs-vswitchd.conf.db(8)).
183 * This function must return 0 if and only if it sets 'netdev->qos_conf'
184 * to an initialized 'struct qos_conf'.
186 * For all QoS implementations it should always be non-null.
188 int (*qos_construct)(struct netdev *netdev, const struct smap *details);
190 /* Destroys the data structures allocated by the implementation as part of
193 * For all QoS implementations it should always be non-null.
195 void (*qos_destruct)(struct netdev *netdev, struct qos_conf *conf);
197 /* Retrieves details of 'netdev->qos_conf' configuration into 'details'.
199 * The contents of 'details' should be documented as valid for 'ovs_name'
200 * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml
201 * (which is built as ovs-vswitchd.conf.db(8)).
203 int (*qos_get)(const struct netdev *netdev, struct smap *details);
205 /* Reconfigures 'netdev->qos_conf' according to 'details', performing any
206 * required calls to complete the reconfiguration.
208 * The contents of 'details' should be documented as valid for 'ovs_name'
209 * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml
210 * (which is built as ovs-vswitchd.conf.db(8)).
212 * This function may be null if 'qos_conf' is not configurable.
214 int (*qos_set)(struct netdev *netdev, const struct smap *details);
216 /* Modify an array of rte_mbufs. The modification is specific to
217 * each qos implementation.
219 * The function should take and array of mbufs and an int representing
220 * the current number of mbufs present in the array.
222 * After the function has performed a qos modification to the array of
223 * mbufs it returns an int representing the number of mbufs now present in
224 * the array. This value is can then be passed to the port send function
225 * along with the modified array for transmission.
227 * For all QoS implementations it should always be non-null.
229 int (*qos_run)(struct netdev *netdev, struct rte_mbuf **pkts,
233 /* dpdk_qos_ops for each type of user space QoS implementation */
234 static const struct dpdk_qos_ops egress_policer_ops;
237 * Array of dpdk_qos_ops, contains pointer to all supported QoS
240 static const struct dpdk_qos_ops *const qos_confs[] = {
245 /* Contains all 'struct dpdk_dev's. */
246 static struct ovs_list dpdk_list OVS_GUARDED_BY(dpdk_mutex)
247 = OVS_LIST_INITIALIZER(&dpdk_list);
249 static struct ovs_list dpdk_mp_list OVS_GUARDED_BY(dpdk_mutex)
250 = OVS_LIST_INITIALIZER(&dpdk_mp_list);
252 /* This mutex must be used by non pmd threads when allocating or freeing
253 * mbufs through mempools. Since dpdk_queue_pkts() and dpdk_queue_flush() may
254 * use mempools, a non pmd thread should hold this mutex while calling them */
255 static struct ovs_mutex nonpmd_mempool_mutex = OVS_MUTEX_INITIALIZER;
258 struct rte_mempool *mp;
262 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
265 /* There should be one 'struct dpdk_tx_queue' created for
267 struct dpdk_tx_queue {
268 bool flush_tx; /* Set to true to flush queue everytime */
269 /* pkts are queued. */
271 rte_spinlock_t tx_lock; /* Protects the members and the NIC queue
272 * from concurrent access. It is used only
273 * if the queue is shared among different
274 * pmd threads (see 'txq_needs_locking'). */
275 int map; /* Mapping of configured vhost-user queues
276 * to enabled by guest. */
278 struct rte_mbuf *burst_pkts[MAX_TX_QUEUE_LEN];
281 /* dpdk has no way to remove dpdk ring ethernet devices
282 so we have to keep them around once they've been created
285 static struct ovs_list dpdk_ring_list OVS_GUARDED_BY(dpdk_mutex)
286 = OVS_LIST_INITIALIZER(&dpdk_ring_list);
289 /* For the client rings */
290 struct rte_ring *cring_tx;
291 struct rte_ring *cring_rx;
292 unsigned int user_port_id; /* User given port no, parsed from port name */
293 int eth_port_id; /* ethernet device port id */
294 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
301 enum dpdk_dev_type type;
303 struct dpdk_tx_queue *tx_q;
305 struct ovs_mutex mutex OVS_ACQ_AFTER(dpdk_mutex);
307 struct dpdk_mp *dpdk_mp;
311 struct netdev_stats stats;
313 rte_spinlock_t stats_lock;
315 struct eth_addr hwaddr;
316 enum netdev_flags flags;
318 struct rte_eth_link link;
321 /* The user might request more txqs than the NIC has. We remap those
322 * ('up.n_txq') on these ('real_n_txq').
323 * If the numbers match, 'txq_needs_locking' is false, otherwise it is
324 * true and we will take a spinlock on transmission */
327 bool txq_needs_locking;
329 /* virtio-net structure for vhost device */
330 OVSRCU_TYPE(struct virtio_net *) virtio_dev;
332 /* Identifier used to distinguish vhost devices from each other */
333 char vhost_id[PATH_MAX];
336 struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex);
338 /* QoS configuration and lock for the device */
339 struct qos_conf *qos_conf;
340 rte_spinlock_t qos_lock;
344 struct netdev_rxq_dpdk {
345 struct netdev_rxq up;
349 static bool dpdk_thread_is_pmd(void);
351 static int netdev_dpdk_construct(struct netdev *);
353 struct virtio_net * netdev_dpdk_get_virtio(const struct netdev_dpdk *dev);
356 is_dpdk_class(const struct netdev_class *class)
358 return class->construct == netdev_dpdk_construct;
361 /* DPDK NIC drivers allocate RX buffers at a particular granularity, typically
362 * aligned at 1k or less. If a declared mbuf size is not a multiple of this
363 * value, insufficient buffers are allocated to accomodate the packet in its
364 * entirety. Furthermore, certain drivers need to ensure that there is also
365 * sufficient space in the Rx buffer to accommodate two VLAN tags (for QinQ
366 * frames). If the RX buffer is too small, then the driver enables scatter RX
367 * behaviour, which reduces performance. To prevent this, use a buffer size that
368 * is closest to 'mtu', but which satisfies the aforementioned criteria.
371 dpdk_buf_size(int mtu)
373 return ROUND_UP((MTU_TO_MAX_FRAME_LEN(mtu) + RTE_PKTMBUF_HEADROOM),
374 NETDEV_DPDK_MBUF_ALIGN);
377 /* XXX: use dpdk malloc for entire OVS. in fact huge page should be used
378 * for all other segments data, bss and text. */
381 dpdk_rte_mzalloc(size_t sz)
385 ptr = rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE);
392 /* XXX this function should be called only by pmd threads (or by non pmd
393 * threads holding the nonpmd_mempool_mutex) */
395 free_dpdk_buf(struct dp_packet *p)
397 struct rte_mbuf *pkt = (struct rte_mbuf *) p;
399 rte_pktmbuf_free(pkt);
403 ovs_rte_pktmbuf_init(struct rte_mempool *mp,
404 void *opaque_arg OVS_UNUSED,
406 unsigned i OVS_UNUSED)
408 struct rte_mbuf *m = _m;
410 rte_pktmbuf_init(mp, opaque_arg, _m, i);
412 dp_packet_init_dpdk((struct dp_packet *) m, m->buf_len);
415 static struct dpdk_mp *
416 dpdk_mp_get(int socket_id, int mtu) OVS_REQUIRES(dpdk_mutex)
418 struct dpdk_mp *dmp = NULL;
419 char mp_name[RTE_MEMPOOL_NAMESIZE];
421 struct rte_pktmbuf_pool_private mbp_priv;
423 LIST_FOR_EACH (dmp, list_node, &dpdk_mp_list) {
424 if (dmp->socket_id == socket_id && dmp->mtu == mtu) {
430 dmp = dpdk_rte_mzalloc(sizeof *dmp);
431 dmp->socket_id = socket_id;
434 mbp_priv.mbuf_data_room_size = MBUF_SIZE(mtu) - sizeof(struct dp_packet);
435 mbp_priv.mbuf_priv_size = sizeof (struct dp_packet) - sizeof (struct rte_mbuf);
437 mp_size = MAX_NB_MBUF;
439 if (snprintf(mp_name, RTE_MEMPOOL_NAMESIZE, "ovs_mp_%d_%d_%u",
440 dmp->mtu, dmp->socket_id, mp_size) < 0) {
444 dmp->mp = rte_mempool_create(mp_name, mp_size, MBUF_SIZE(mtu),
446 sizeof(struct rte_pktmbuf_pool_private),
447 rte_pktmbuf_pool_init, &mbp_priv,
448 ovs_rte_pktmbuf_init, NULL,
450 } while (!dmp->mp && rte_errno == ENOMEM && (mp_size /= 2) >= MIN_NB_MBUF);
452 if (dmp->mp == NULL) {
455 VLOG_DBG("Allocated \"%s\" mempool with %u mbufs", mp_name, mp_size );
458 list_push_back(&dpdk_mp_list, &dmp->list_node);
463 dpdk_mp_put(struct dpdk_mp *dmp)
471 ovs_assert(dmp->refcount >= 0);
474 /* I could not find any API to destroy mp. */
475 if (dmp->refcount == 0) {
476 list_delete(dmp->list_node);
477 /* destroy mp-pool. */
483 check_link_status(struct netdev_dpdk *dev)
485 struct rte_eth_link link;
487 rte_eth_link_get_nowait(dev->port_id, &link);
489 if (dev->link.link_status != link.link_status) {
490 netdev_change_seq_changed(&dev->up);
492 dev->link_reset_cnt++;
494 if (dev->link.link_status) {
495 VLOG_DBG_RL(&rl, "Port %d Link Up - speed %u Mbps - %s",
496 dev->port_id, (unsigned)dev->link.link_speed,
497 (dev->link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
498 ("full-duplex") : ("half-duplex"));
500 VLOG_DBG_RL(&rl, "Port %d Link Down", dev->port_id);
506 dpdk_watchdog(void *dummy OVS_UNUSED)
508 struct netdev_dpdk *dev;
510 pthread_detach(pthread_self());
513 ovs_mutex_lock(&dpdk_mutex);
514 LIST_FOR_EACH (dev, list_node, &dpdk_list) {
515 ovs_mutex_lock(&dev->mutex);
516 check_link_status(dev);
517 ovs_mutex_unlock(&dev->mutex);
519 ovs_mutex_unlock(&dpdk_mutex);
520 xsleep(DPDK_PORT_WATCHDOG_INTERVAL);
527 dpdk_eth_dev_queue_setup(struct netdev_dpdk *dev, int n_rxq, int n_txq)
532 /* A device may report more queues than it makes available (this has
533 * been observed for Intel xl710, which reserves some of them for
534 * SRIOV): rte_eth_*_queue_setup will fail if a queue is not
535 * available. When this happens we can retry the configuration
536 * and request less queues */
537 while (n_rxq && n_txq) {
539 VLOG_INFO("Retrying setup with (rxq:%d txq:%d)", n_rxq, n_txq);
542 diag = rte_eth_dev_configure(dev->port_id, n_rxq, n_txq, &port_conf);
547 for (i = 0; i < n_txq; i++) {
548 diag = rte_eth_tx_queue_setup(dev->port_id, i, NIC_PORT_TX_Q_SIZE,
549 dev->socket_id, NULL);
551 VLOG_INFO("Interface %s txq(%d) setup error: %s",
552 dev->up.name, i, rte_strerror(-diag));
558 /* Retry with less tx queues */
563 for (i = 0; i < n_rxq; i++) {
564 diag = rte_eth_rx_queue_setup(dev->port_id, i, NIC_PORT_RX_Q_SIZE,
565 dev->socket_id, NULL,
568 VLOG_INFO("Interface %s rxq(%d) setup error: %s",
569 dev->up.name, i, rte_strerror(-diag));
575 /* Retry with less rx queues */
580 dev->up.n_rxq = n_rxq;
581 dev->real_n_txq = n_txq;
591 dpdk_eth_dev_init(struct netdev_dpdk *dev) OVS_REQUIRES(dpdk_mutex)
593 struct rte_pktmbuf_pool_private *mbp_priv;
594 struct rte_eth_dev_info info;
595 struct ether_addr eth_addr;
599 if (dev->port_id < 0 || dev->port_id >= rte_eth_dev_count()) {
603 rte_eth_dev_info_get(dev->port_id, &info);
605 n_rxq = MIN(info.max_rx_queues, dev->up.n_rxq);
606 n_txq = MIN(info.max_tx_queues, dev->up.n_txq);
608 diag = dpdk_eth_dev_queue_setup(dev, n_rxq, n_txq);
610 VLOG_ERR("Interface %s(rxq:%d txq:%d) configure error: %s",
611 dev->up.name, n_rxq, n_txq, rte_strerror(-diag));
615 diag = rte_eth_dev_start(dev->port_id);
617 VLOG_ERR("Interface %s start error: %s", dev->up.name,
618 rte_strerror(-diag));
622 rte_eth_promiscuous_enable(dev->port_id);
623 rte_eth_allmulticast_enable(dev->port_id);
625 memset(ð_addr, 0x0, sizeof(eth_addr));
626 rte_eth_macaddr_get(dev->port_id, ð_addr);
627 VLOG_INFO_RL(&rl, "Port %d: "ETH_ADDR_FMT"",
628 dev->port_id, ETH_ADDR_BYTES_ARGS(eth_addr.addr_bytes));
630 memcpy(dev->hwaddr.ea, eth_addr.addr_bytes, ETH_ADDR_LEN);
631 rte_eth_link_get_nowait(dev->port_id, &dev->link);
633 mbp_priv = rte_mempool_get_priv(dev->dpdk_mp->mp);
634 dev->buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
636 dev->flags = NETDEV_UP | NETDEV_PROMISC;
640 static struct netdev_dpdk *
641 netdev_dpdk_cast(const struct netdev *netdev)
643 return CONTAINER_OF(netdev, struct netdev_dpdk, up);
646 static struct netdev *
647 netdev_dpdk_alloc(void)
649 struct netdev_dpdk *netdev = dpdk_rte_mzalloc(sizeof *netdev);
654 netdev_dpdk_alloc_txq(struct netdev_dpdk *netdev, unsigned int n_txqs)
658 netdev->tx_q = dpdk_rte_mzalloc(n_txqs * sizeof *netdev->tx_q);
659 for (i = 0; i < n_txqs; i++) {
660 int numa_id = ovs_numa_get_numa_id(i);
662 if (!netdev->txq_needs_locking) {
663 /* Each index is considered as a cpu core id, since there should
664 * be one tx queue for each cpu core. If the corresponding core
665 * is not on the same numa node as 'netdev', flags the
667 netdev->tx_q[i].flush_tx = netdev->socket_id == numa_id;
669 /* Queues are shared among CPUs. Always flush */
670 netdev->tx_q[i].flush_tx = true;
673 /* Initialize map for vhost devices. */
674 netdev->tx_q[i].map = -1;
675 rte_spinlock_init(&netdev->tx_q[i].tx_lock);
680 netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no,
681 enum dpdk_dev_type type)
682 OVS_REQUIRES(dpdk_mutex)
684 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
689 ovs_mutex_init(&netdev->mutex);
690 ovs_mutex_lock(&netdev->mutex);
692 rte_spinlock_init(&netdev->stats_lock);
694 /* If the 'sid' is negative, it means that the kernel fails
695 * to obtain the pci numa info. In that situation, always
697 if (type == DPDK_DEV_ETH) {
698 sid = rte_eth_dev_socket_id(port_no);
700 sid = rte_lcore_to_socket_id(rte_get_master_lcore());
703 netdev->socket_id = sid < 0 ? SOCKET0 : sid;
704 netdev->port_id = port_no;
707 netdev->mtu = ETHER_MTU;
708 netdev->max_packet_len = MTU_TO_FRAME_LEN(netdev->mtu);
710 buf_size = dpdk_buf_size(netdev->mtu);
711 netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, FRAME_LEN_TO_MTU(buf_size));
712 if (!netdev->dpdk_mp) {
717 /* Initialise QoS configuration to NULL and qos lock to unlocked */
718 netdev->qos_conf = NULL;
719 rte_spinlock_init(&netdev->qos_lock);
721 netdev_->n_txq = NR_QUEUE;
722 netdev_->n_rxq = NR_QUEUE;
723 netdev_->requested_n_rxq = NR_QUEUE;
724 netdev->real_n_txq = NR_QUEUE;
726 if (type == DPDK_DEV_ETH) {
727 netdev_dpdk_alloc_txq(netdev, NR_QUEUE);
728 err = dpdk_eth_dev_init(netdev);
733 netdev_dpdk_alloc_txq(netdev, OVS_VHOST_MAX_QUEUE_NUM);
736 list_push_back(&dpdk_list, &netdev->list_node);
740 rte_free(netdev->tx_q);
742 ovs_mutex_unlock(&netdev->mutex);
746 /* dev_name must be the prefix followed by a positive decimal number.
747 * (no leading + or - signs are allowed) */
749 dpdk_dev_parse_name(const char dev_name[], const char prefix[],
750 unsigned int *port_no)
754 if (strncmp(dev_name, prefix, strlen(prefix))) {
758 cport = dev_name + strlen(prefix);
760 if (str_to_uint(cport, 10, port_no)) {
768 vhost_construct_helper(struct netdev *netdev_) OVS_REQUIRES(dpdk_mutex)
770 if (rte_eal_init_ret) {
771 return rte_eal_init_ret;
774 return netdev_dpdk_init(netdev_, -1, DPDK_DEV_VHOST);
778 netdev_dpdk_vhost_cuse_construct(struct netdev *netdev_)
780 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
783 ovs_mutex_lock(&dpdk_mutex);
784 strncpy(netdev->vhost_id, netdev->up.name, sizeof(netdev->vhost_id));
785 err = vhost_construct_helper(netdev_);
786 ovs_mutex_unlock(&dpdk_mutex);
791 netdev_dpdk_vhost_user_construct(struct netdev *netdev_)
793 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
794 const char *name = netdev_->name;
797 /* 'name' is appended to 'vhost_sock_dir' and used to create a socket in
798 * the file system. '/' or '\' would traverse directories, so they're not
799 * acceptable in 'name'. */
800 if (strchr(name, '/') || strchr(name, '\\')) {
801 VLOG_ERR("\"%s\" is not a valid name for a vhost-user port. "
802 "A valid name must not include '/' or '\\'",
807 ovs_mutex_lock(&dpdk_mutex);
808 /* Take the name of the vhost-user port and append it to the location where
809 * the socket is to be created, then register the socket.
811 snprintf(netdev->vhost_id, sizeof(netdev->vhost_id), "%s/%s",
812 vhost_sock_dir, name);
814 err = rte_vhost_driver_register(netdev->vhost_id);
816 VLOG_ERR("vhost-user socket device setup failure for socket %s\n",
819 fatal_signal_add_file_to_unlink(netdev->vhost_id);
820 VLOG_INFO("Socket %s created for vhost-user port %s\n",
821 netdev->vhost_id, name);
822 err = vhost_construct_helper(netdev_);
825 ovs_mutex_unlock(&dpdk_mutex);
830 netdev_dpdk_construct(struct netdev *netdev)
832 unsigned int port_no;
835 if (rte_eal_init_ret) {
836 return rte_eal_init_ret;
839 /* Names always start with "dpdk" */
840 err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no);
845 ovs_mutex_lock(&dpdk_mutex);
846 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
847 ovs_mutex_unlock(&dpdk_mutex);
852 netdev_dpdk_destruct(struct netdev *netdev_)
854 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
856 ovs_mutex_lock(&dev->mutex);
857 rte_eth_dev_stop(dev->port_id);
858 ovs_mutex_unlock(&dev->mutex);
860 ovs_mutex_lock(&dpdk_mutex);
862 list_remove(&dev->list_node);
863 dpdk_mp_put(dev->dpdk_mp);
864 ovs_mutex_unlock(&dpdk_mutex);
868 netdev_dpdk_vhost_destruct(struct netdev *netdev_)
870 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
872 /* Guest becomes an orphan if still attached. */
873 if (netdev_dpdk_get_virtio(dev) != NULL) {
874 VLOG_ERR("Removing port '%s' while vhost device still attached.",
876 VLOG_ERR("To restore connectivity after re-adding of port, VM on socket"
877 " '%s' must be restarted.",
881 if (rte_vhost_driver_unregister(dev->vhost_id)) {
882 VLOG_ERR("Unable to remove vhost-user socket %s", dev->vhost_id);
884 fatal_signal_remove_file_to_unlink(dev->vhost_id);
887 ovs_mutex_lock(&dpdk_mutex);
889 list_remove(&dev->list_node);
890 dpdk_mp_put(dev->dpdk_mp);
891 ovs_mutex_unlock(&dpdk_mutex);
895 netdev_dpdk_dealloc(struct netdev *netdev_)
897 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
903 netdev_dpdk_get_config(const struct netdev *netdev, struct smap *args)
905 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
907 ovs_mutex_lock(&dev->mutex);
909 smap_add_format(args, "requested_rx_queues", "%d", netdev->requested_n_rxq);
910 smap_add_format(args, "configured_rx_queues", "%d", netdev->n_rxq);
911 smap_add_format(args, "requested_tx_queues", "%d", netdev->n_txq);
912 smap_add_format(args, "configured_tx_queues", "%d", dev->real_n_txq);
913 ovs_mutex_unlock(&dev->mutex);
919 netdev_dpdk_set_config(struct netdev *netdev, const struct smap *args)
921 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
923 ovs_mutex_lock(&dev->mutex);
924 netdev->requested_n_rxq = MAX(smap_get_int(args, "n_rxq",
925 netdev->requested_n_rxq), 1);
926 netdev_change_seq_changed(netdev);
927 ovs_mutex_unlock(&dev->mutex);
933 netdev_dpdk_get_numa_id(const struct netdev *netdev_)
935 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
937 return netdev->socket_id;
940 /* Sets the number of tx queues and rx queues for the dpdk interface.
941 * If the configuration fails, do not try restoring its old configuration
942 * and just returns the error. */
944 netdev_dpdk_set_multiq(struct netdev *netdev_, unsigned int n_txq,
947 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
949 int old_rxq, old_txq;
951 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
955 ovs_mutex_lock(&dpdk_mutex);
956 ovs_mutex_lock(&netdev->mutex);
958 rte_eth_dev_stop(netdev->port_id);
960 old_txq = netdev->up.n_txq;
961 old_rxq = netdev->up.n_rxq;
962 netdev->up.n_txq = n_txq;
963 netdev->up.n_rxq = n_rxq;
965 rte_free(netdev->tx_q);
966 err = dpdk_eth_dev_init(netdev);
967 netdev_dpdk_alloc_txq(netdev, netdev->real_n_txq);
969 /* If there has been an error, it means that the requested queues
970 * have not been created. Restore the old numbers. */
971 netdev->up.n_txq = old_txq;
972 netdev->up.n_rxq = old_rxq;
975 netdev->txq_needs_locking = netdev->real_n_txq != netdev->up.n_txq;
977 ovs_mutex_unlock(&netdev->mutex);
978 ovs_mutex_unlock(&dpdk_mutex);
984 netdev_dpdk_vhost_cuse_set_multiq(struct netdev *netdev_, unsigned int n_txq,
987 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
990 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
994 ovs_mutex_lock(&dpdk_mutex);
995 ovs_mutex_lock(&netdev->mutex);
997 netdev->up.n_txq = n_txq;
998 netdev->real_n_txq = 1;
999 netdev->up.n_rxq = 1;
1000 netdev->txq_needs_locking = netdev->real_n_txq != netdev->up.n_txq;
1002 ovs_mutex_unlock(&netdev->mutex);
1003 ovs_mutex_unlock(&dpdk_mutex);
1009 netdev_dpdk_vhost_set_multiq(struct netdev *netdev_, unsigned int n_txq,
1012 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1015 if (netdev->up.n_txq == n_txq && netdev->up.n_rxq == n_rxq) {
1019 ovs_mutex_lock(&dpdk_mutex);
1020 ovs_mutex_lock(&netdev->mutex);
1022 netdev->up.n_txq = n_txq;
1023 netdev->up.n_rxq = n_rxq;
1025 ovs_mutex_unlock(&netdev->mutex);
1026 ovs_mutex_unlock(&dpdk_mutex);
1031 static struct netdev_rxq *
1032 netdev_dpdk_rxq_alloc(void)
1034 struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx);
1039 static struct netdev_rxq_dpdk *
1040 netdev_rxq_dpdk_cast(const struct netdev_rxq *rx)
1042 return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up);
1046 netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_)
1048 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
1049 struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev);
1051 ovs_mutex_lock(&netdev->mutex);
1052 rx->port_id = netdev->port_id;
1053 ovs_mutex_unlock(&netdev->mutex);
1059 netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED)
1064 netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_)
1066 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
1072 dpdk_queue_flush__(struct netdev_dpdk *dev, int qid)
1074 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
1077 while (nb_tx != txq->count) {
1080 ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx,
1081 txq->count - nb_tx);
1089 if (OVS_UNLIKELY(nb_tx != txq->count)) {
1090 /* free buffers, which we couldn't transmit, one at a time (each
1091 * packet could come from a different mempool) */
1094 for (i = nb_tx; i < txq->count; i++) {
1095 rte_pktmbuf_free(txq->burst_pkts[i]);
1097 rte_spinlock_lock(&dev->stats_lock);
1098 dev->stats.tx_dropped += txq->count-nb_tx;
1099 rte_spinlock_unlock(&dev->stats_lock);
1103 txq->tsc = rte_get_timer_cycles();
1107 dpdk_queue_flush(struct netdev_dpdk *dev, int qid)
1109 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
1111 if (txq->count == 0) {
1114 dpdk_queue_flush__(dev, qid);
1118 is_vhost_running(struct virtio_net *dev)
1120 return (dev != NULL && (dev->flags & VIRTIO_DEV_RUNNING));
1124 netdev_dpdk_vhost_update_rx_counters(struct netdev_stats *stats,
1125 struct dp_packet **packets, int count)
1128 struct dp_packet *packet;
1130 stats->rx_packets += count;
1131 for (i = 0; i < count; i++) {
1132 packet = packets[i];
1134 if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) {
1135 /* This only protects the following multicast counting from
1136 * too short packets, but it does not stop the packet from
1137 * further processing. */
1139 stats->rx_length_errors++;
1143 struct eth_header *eh = (struct eth_header *) dp_packet_data(packet);
1144 if (OVS_UNLIKELY(eth_addr_is_multicast(eh->eth_dst))) {
1148 stats->rx_bytes += dp_packet_size(packet);
1153 * The receive path for the vhost port is the TX path out from guest.
1156 netdev_dpdk_vhost_rxq_recv(struct netdev_rxq *rxq_,
1157 struct dp_packet **packets, int *c)
1159 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
1160 struct netdev *netdev = rx->up.netdev;
1161 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
1162 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
1163 int qid = rxq_->queue_id;
1166 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev))) {
1170 if (rxq_->queue_id >= vhost_dev->real_n_rxq) {
1174 nb_rx = rte_vhost_dequeue_burst(virtio_dev, qid * VIRTIO_QNUM + VIRTIO_TXQ,
1175 vhost_dev->dpdk_mp->mp,
1176 (struct rte_mbuf **)packets,
1182 rte_spinlock_lock(&vhost_dev->stats_lock);
1183 netdev_dpdk_vhost_update_rx_counters(&vhost_dev->stats, packets, nb_rx);
1184 rte_spinlock_unlock(&vhost_dev->stats_lock);
1191 netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dp_packet **packets,
1194 struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_);
1195 struct netdev *netdev = rx->up.netdev;
1196 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1199 /* There is only one tx queue for this core. Do not flush other
1201 * Do not flush tx queue which is shared among CPUs
1202 * since it is always flushed */
1203 if (rxq_->queue_id == rte_lcore_id() &&
1204 OVS_LIKELY(!dev->txq_needs_locking)) {
1205 dpdk_queue_flush(dev, rxq_->queue_id);
1208 nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id,
1209 (struct rte_mbuf **) packets,
1221 netdev_dpdk_qos_run__(struct netdev_dpdk *dev, struct rte_mbuf **pkts,
1224 struct netdev *netdev = &dev->up;
1226 if (dev->qos_conf != NULL) {
1227 rte_spinlock_lock(&dev->qos_lock);
1228 if (dev->qos_conf != NULL) {
1229 cnt = dev->qos_conf->ops->qos_run(netdev, pkts, cnt);
1231 rte_spinlock_unlock(&dev->qos_lock);
1238 netdev_dpdk_vhost_update_tx_counters(struct netdev_stats *stats,
1239 struct dp_packet **packets,
1244 int sent = attempted - dropped;
1246 stats->tx_packets += sent;
1247 stats->tx_dropped += dropped;
1249 for (i = 0; i < sent; i++) {
1250 stats->tx_bytes += dp_packet_size(packets[i]);
1255 __netdev_dpdk_vhost_send(struct netdev *netdev, int qid,
1256 struct dp_packet **pkts, int cnt,
1259 struct netdev_dpdk *vhost_dev = netdev_dpdk_cast(netdev);
1260 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev);
1261 struct rte_mbuf **cur_pkts = (struct rte_mbuf **) pkts;
1262 unsigned int total_pkts = cnt;
1263 unsigned int qos_pkts = cnt;
1266 qid = vhost_dev->tx_q[qid % vhost_dev->real_n_txq].map;
1268 if (OVS_UNLIKELY(!is_vhost_running(virtio_dev) || qid == -1)) {
1269 rte_spinlock_lock(&vhost_dev->stats_lock);
1270 vhost_dev->stats.tx_dropped+= cnt;
1271 rte_spinlock_unlock(&vhost_dev->stats_lock);
1275 rte_spinlock_lock(&vhost_dev->tx_q[qid].tx_lock);
1277 /* Check has QoS has been configured for the netdev */
1278 cnt = netdev_dpdk_qos_run__(vhost_dev, cur_pkts, cnt);
1282 int vhost_qid = qid * VIRTIO_QNUM + VIRTIO_RXQ;
1283 unsigned int tx_pkts;
1285 tx_pkts = rte_vhost_enqueue_burst(virtio_dev, vhost_qid,
1287 if (OVS_LIKELY(tx_pkts)) {
1288 /* Packets have been sent.*/
1290 /* Prepare for possible next iteration.*/
1291 cur_pkts = &cur_pkts[tx_pkts];
1293 uint64_t timeout = VHOST_ENQ_RETRY_USECS * rte_get_timer_hz() / 1E6;
1294 unsigned int expired = 0;
1297 start = rte_get_timer_cycles();
1301 * Unable to enqueue packets to vhost interface.
1302 * Check available entries before retrying.
1304 while (!rte_vring_available_entries(virtio_dev, vhost_qid)) {
1305 if (OVS_UNLIKELY((rte_get_timer_cycles() - start) > timeout)) {
1311 /* break out of main loop. */
1317 rte_spinlock_unlock(&vhost_dev->tx_q[qid].tx_lock);
1319 rte_spinlock_lock(&vhost_dev->stats_lock);
1321 netdev_dpdk_vhost_update_tx_counters(&vhost_dev->stats, pkts, total_pkts,
1323 rte_spinlock_unlock(&vhost_dev->stats_lock);
1329 for (i = 0; i < total_pkts; i++) {
1330 dp_packet_delete(pkts[i]);
1336 dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
1337 struct rte_mbuf **pkts, int cnt)
1339 struct dpdk_tx_queue *txq = &dev->tx_q[qid];
1345 int freeslots = MAX_TX_QUEUE_LEN - txq->count;
1346 int tocopy = MIN(freeslots, cnt-i);
1348 memcpy(&txq->burst_pkts[txq->count], &pkts[i],
1349 tocopy * sizeof (struct rte_mbuf *));
1351 txq->count += tocopy;
1354 if (txq->count == MAX_TX_QUEUE_LEN || txq->flush_tx) {
1355 dpdk_queue_flush__(dev, qid);
1357 diff_tsc = rte_get_timer_cycles() - txq->tsc;
1358 if (diff_tsc >= DRAIN_TSC) {
1359 dpdk_queue_flush__(dev, qid);
1364 /* Tx function. Transmit packets indefinitely */
1366 dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet **pkts,
1368 OVS_NO_THREAD_SAFETY_ANALYSIS
1370 #if !defined(__CHECKER__) && !defined(_WIN32)
1371 const size_t PKT_ARRAY_SIZE = cnt;
1373 /* Sparse or MSVC doesn't like variable length array. */
1374 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
1376 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1377 struct rte_mbuf *mbufs[PKT_ARRAY_SIZE];
1382 /* If we are on a non pmd thread we have to use the mempool mutex, because
1383 * every non pmd thread shares the same mempool cache */
1385 if (!dpdk_thread_is_pmd()) {
1386 ovs_mutex_lock(&nonpmd_mempool_mutex);
1389 for (i = 0; i < cnt; i++) {
1390 int size = dp_packet_size(pkts[i]);
1392 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1393 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1394 (int)size , dev->max_packet_len);
1400 mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
1402 if (!mbufs[newcnt]) {
1407 /* We have to do a copy for now */
1408 memcpy(rte_pktmbuf_mtod(mbufs[newcnt], void *), dp_packet_data(pkts[i]), size);
1410 rte_pktmbuf_data_len(mbufs[newcnt]) = size;
1411 rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
1416 if (dev->type == DPDK_DEV_VHOST) {
1417 __netdev_dpdk_vhost_send(netdev, qid, (struct dp_packet **) mbufs, newcnt, true);
1419 unsigned int qos_pkts = newcnt;
1421 /* Check if QoS has been configured for this netdev. */
1422 newcnt = netdev_dpdk_qos_run__(dev, mbufs, newcnt);
1424 dropped += qos_pkts - newcnt;
1425 dpdk_queue_pkts(dev, qid, mbufs, newcnt);
1426 dpdk_queue_flush(dev, qid);
1429 if (OVS_UNLIKELY(dropped)) {
1430 rte_spinlock_lock(&dev->stats_lock);
1431 dev->stats.tx_dropped += dropped;
1432 rte_spinlock_unlock(&dev->stats_lock);
1435 if (!dpdk_thread_is_pmd()) {
1436 ovs_mutex_unlock(&nonpmd_mempool_mutex);
1441 netdev_dpdk_vhost_send(struct netdev *netdev, int qid, struct dp_packet **pkts,
1442 int cnt, bool may_steal)
1444 if (OVS_UNLIKELY(pkts[0]->source != DPBUF_DPDK)) {
1447 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1449 for (i = 0; i < cnt; i++) {
1450 dp_packet_delete(pkts[i]);
1454 __netdev_dpdk_vhost_send(netdev, qid, pkts, cnt, may_steal);
1460 netdev_dpdk_send__(struct netdev_dpdk *dev, int qid,
1461 struct dp_packet **pkts, int cnt, bool may_steal)
1465 if (OVS_UNLIKELY(dev->txq_needs_locking)) {
1466 qid = qid % dev->real_n_txq;
1467 rte_spinlock_lock(&dev->tx_q[qid].tx_lock);
1470 if (OVS_UNLIKELY(!may_steal ||
1471 pkts[0]->source != DPBUF_DPDK)) {
1472 struct netdev *netdev = &dev->up;
1474 dpdk_do_tx_copy(netdev, qid, pkts, cnt);
1477 for (i = 0; i < cnt; i++) {
1478 dp_packet_delete(pkts[i]);
1482 int next_tx_idx = 0;
1484 unsigned int qos_pkts = 0;
1485 unsigned int temp_cnt = 0;
1487 for (i = 0; i < cnt; i++) {
1488 int size = dp_packet_size(pkts[i]);
1490 if (OVS_UNLIKELY(size > dev->max_packet_len)) {
1491 if (next_tx_idx != i) {
1492 temp_cnt = i - next_tx_idx;
1493 qos_pkts = temp_cnt;
1495 temp_cnt = netdev_dpdk_qos_run__(dev, (struct rte_mbuf**)pkts,
1497 dropped += qos_pkts - temp_cnt;
1498 dpdk_queue_pkts(dev, qid,
1499 (struct rte_mbuf **)&pkts[next_tx_idx],
1504 VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
1505 (int)size , dev->max_packet_len);
1507 dp_packet_delete(pkts[i]);
1509 next_tx_idx = i + 1;
1512 if (next_tx_idx != cnt) {
1516 cnt = netdev_dpdk_qos_run__(dev, (struct rte_mbuf**)pkts, cnt);
1517 dropped += qos_pkts - cnt;
1518 dpdk_queue_pkts(dev, qid, (struct rte_mbuf **)&pkts[next_tx_idx],
1522 if (OVS_UNLIKELY(dropped)) {
1523 rte_spinlock_lock(&dev->stats_lock);
1524 dev->stats.tx_dropped += dropped;
1525 rte_spinlock_unlock(&dev->stats_lock);
1529 if (OVS_UNLIKELY(dev->txq_needs_locking)) {
1530 rte_spinlock_unlock(&dev->tx_q[qid].tx_lock);
1535 netdev_dpdk_eth_send(struct netdev *netdev, int qid,
1536 struct dp_packet **pkts, int cnt, bool may_steal)
1538 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1540 netdev_dpdk_send__(dev, qid, pkts, cnt, may_steal);
1545 netdev_dpdk_set_etheraddr(struct netdev *netdev, const struct eth_addr mac)
1547 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1549 ovs_mutex_lock(&dev->mutex);
1550 if (!eth_addr_equals(dev->hwaddr, mac)) {
1552 netdev_change_seq_changed(netdev);
1554 ovs_mutex_unlock(&dev->mutex);
1560 netdev_dpdk_get_etheraddr(const struct netdev *netdev, struct eth_addr *mac)
1562 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1564 ovs_mutex_lock(&dev->mutex);
1566 ovs_mutex_unlock(&dev->mutex);
1572 netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup)
1574 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1576 ovs_mutex_lock(&dev->mutex);
1578 ovs_mutex_unlock(&dev->mutex);
1584 netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu)
1586 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1587 int old_mtu, err, dpdk_mtu;
1588 struct dpdk_mp *old_mp;
1592 ovs_mutex_lock(&dpdk_mutex);
1593 ovs_mutex_lock(&dev->mutex);
1594 if (dev->mtu == mtu) {
1599 buf_size = dpdk_buf_size(mtu);
1600 dpdk_mtu = FRAME_LEN_TO_MTU(buf_size);
1602 mp = dpdk_mp_get(dev->socket_id, dpdk_mtu);
1608 rte_eth_dev_stop(dev->port_id);
1611 old_mp = dev->dpdk_mp;
1614 dev->max_packet_len = MTU_TO_FRAME_LEN(dev->mtu);
1616 err = dpdk_eth_dev_init(dev);
1620 dev->dpdk_mp = old_mp;
1621 dev->max_packet_len = MTU_TO_FRAME_LEN(dev->mtu);
1622 dpdk_eth_dev_init(dev);
1626 dpdk_mp_put(old_mp);
1627 netdev_change_seq_changed(netdev);
1629 ovs_mutex_unlock(&dev->mutex);
1630 ovs_mutex_unlock(&dpdk_mutex);
1635 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier);
1638 netdev_dpdk_vhost_get_stats(const struct netdev *netdev,
1639 struct netdev_stats *stats)
1641 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1643 ovs_mutex_lock(&dev->mutex);
1644 memset(stats, 0, sizeof(*stats));
1645 /* Unsupported Stats */
1646 stats->collisions = UINT64_MAX;
1647 stats->rx_crc_errors = UINT64_MAX;
1648 stats->rx_fifo_errors = UINT64_MAX;
1649 stats->rx_frame_errors = UINT64_MAX;
1650 stats->rx_missed_errors = UINT64_MAX;
1651 stats->rx_over_errors = UINT64_MAX;
1652 stats->tx_aborted_errors = UINT64_MAX;
1653 stats->tx_carrier_errors = UINT64_MAX;
1654 stats->tx_errors = UINT64_MAX;
1655 stats->tx_fifo_errors = UINT64_MAX;
1656 stats->tx_heartbeat_errors = UINT64_MAX;
1657 stats->tx_window_errors = UINT64_MAX;
1658 stats->rx_dropped += UINT64_MAX;
1660 rte_spinlock_lock(&dev->stats_lock);
1661 /* Supported Stats */
1662 stats->rx_packets += dev->stats.rx_packets;
1663 stats->tx_packets += dev->stats.tx_packets;
1664 stats->tx_dropped += dev->stats.tx_dropped;
1665 stats->multicast = dev->stats.multicast;
1666 stats->rx_bytes = dev->stats.rx_bytes;
1667 stats->tx_bytes = dev->stats.tx_bytes;
1668 stats->rx_errors = dev->stats.rx_errors;
1669 stats->rx_length_errors = dev->stats.rx_length_errors;
1670 rte_spinlock_unlock(&dev->stats_lock);
1672 ovs_mutex_unlock(&dev->mutex);
1678 netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats)
1680 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1681 struct rte_eth_stats rte_stats;
1684 netdev_dpdk_get_carrier(netdev, &gg);
1685 ovs_mutex_lock(&dev->mutex);
1686 rte_eth_stats_get(dev->port_id, &rte_stats);
1688 memset(stats, 0, sizeof(*stats));
1690 stats->rx_packets = rte_stats.ipackets;
1691 stats->tx_packets = rte_stats.opackets;
1692 stats->rx_bytes = rte_stats.ibytes;
1693 stats->tx_bytes = rte_stats.obytes;
1694 /* DPDK counts imissed as errors, but count them here as dropped instead */
1695 stats->rx_errors = rte_stats.ierrors - rte_stats.imissed;
1696 stats->tx_errors = rte_stats.oerrors;
1697 stats->multicast = rte_stats.imcasts;
1699 rte_spinlock_lock(&dev->stats_lock);
1700 stats->tx_dropped = dev->stats.tx_dropped;
1701 rte_spinlock_unlock(&dev->stats_lock);
1703 /* These are the available DPDK counters for packets not received due to
1704 * local resource constraints in DPDK and NIC respectively. */
1705 stats->rx_dropped = rte_stats.rx_nombuf + rte_stats.imissed;
1706 stats->collisions = UINT64_MAX;
1708 stats->rx_length_errors = UINT64_MAX;
1709 stats->rx_over_errors = UINT64_MAX;
1710 stats->rx_crc_errors = UINT64_MAX;
1711 stats->rx_frame_errors = UINT64_MAX;
1712 stats->rx_fifo_errors = UINT64_MAX;
1713 stats->rx_missed_errors = rte_stats.imissed;
1715 stats->tx_aborted_errors = UINT64_MAX;
1716 stats->tx_carrier_errors = UINT64_MAX;
1717 stats->tx_fifo_errors = UINT64_MAX;
1718 stats->tx_heartbeat_errors = UINT64_MAX;
1719 stats->tx_window_errors = UINT64_MAX;
1721 ovs_mutex_unlock(&dev->mutex);
1727 netdev_dpdk_get_features(const struct netdev *netdev_,
1728 enum netdev_features *current,
1729 enum netdev_features *advertised OVS_UNUSED,
1730 enum netdev_features *supported OVS_UNUSED,
1731 enum netdev_features *peer OVS_UNUSED)
1733 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1734 struct rte_eth_link link;
1736 ovs_mutex_lock(&dev->mutex);
1738 ovs_mutex_unlock(&dev->mutex);
1740 if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) {
1741 if (link.link_speed == ETH_LINK_SPEED_AUTONEG) {
1742 *current = NETDEV_F_AUTONEG;
1744 } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) {
1745 if (link.link_speed == ETH_LINK_SPEED_10) {
1746 *current = NETDEV_F_10MB_HD;
1748 if (link.link_speed == ETH_LINK_SPEED_100) {
1749 *current = NETDEV_F_100MB_HD;
1751 if (link.link_speed == ETH_LINK_SPEED_1000) {
1752 *current = NETDEV_F_1GB_HD;
1754 } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) {
1755 if (link.link_speed == ETH_LINK_SPEED_10) {
1756 *current = NETDEV_F_10MB_FD;
1758 if (link.link_speed == ETH_LINK_SPEED_100) {
1759 *current = NETDEV_F_100MB_FD;
1761 if (link.link_speed == ETH_LINK_SPEED_1000) {
1762 *current = NETDEV_F_1GB_FD;
1764 if (link.link_speed == ETH_LINK_SPEED_10000) {
1765 *current = NETDEV_F_10GB_FD;
1773 netdev_dpdk_get_ifindex(const struct netdev *netdev)
1775 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
1778 ovs_mutex_lock(&dev->mutex);
1779 ifindex = dev->port_id;
1780 ovs_mutex_unlock(&dev->mutex);
1786 netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier)
1788 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1790 ovs_mutex_lock(&dev->mutex);
1791 check_link_status(dev);
1792 *carrier = dev->link.link_status;
1794 ovs_mutex_unlock(&dev->mutex);
1800 netdev_dpdk_vhost_get_carrier(const struct netdev *netdev_, bool *carrier)
1802 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1803 struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(dev);
1805 ovs_mutex_lock(&dev->mutex);
1807 if (is_vhost_running(virtio_dev)) {
1813 ovs_mutex_unlock(&dev->mutex);
1818 static long long int
1819 netdev_dpdk_get_carrier_resets(const struct netdev *netdev_)
1821 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1822 long long int carrier_resets;
1824 ovs_mutex_lock(&dev->mutex);
1825 carrier_resets = dev->link_reset_cnt;
1826 ovs_mutex_unlock(&dev->mutex);
1828 return carrier_resets;
1832 netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED,
1833 long long int interval OVS_UNUSED)
1839 netdev_dpdk_update_flags__(struct netdev_dpdk *dev,
1840 enum netdev_flags off, enum netdev_flags on,
1841 enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex)
1845 if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) {
1849 *old_flagsp = dev->flags;
1853 if (dev->flags == *old_flagsp) {
1857 if (dev->type == DPDK_DEV_ETH) {
1858 if (dev->flags & NETDEV_UP) {
1859 err = rte_eth_dev_start(dev->port_id);
1864 if (dev->flags & NETDEV_PROMISC) {
1865 rte_eth_promiscuous_enable(dev->port_id);
1868 if (!(dev->flags & NETDEV_UP)) {
1869 rte_eth_dev_stop(dev->port_id);
1877 netdev_dpdk_update_flags(struct netdev *netdev_,
1878 enum netdev_flags off, enum netdev_flags on,
1879 enum netdev_flags *old_flagsp)
1881 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
1884 ovs_mutex_lock(&netdev->mutex);
1885 error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp);
1886 ovs_mutex_unlock(&netdev->mutex);
1892 netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args)
1894 struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_);
1895 struct rte_eth_dev_info dev_info;
1897 if (dev->port_id < 0)
1900 ovs_mutex_lock(&dev->mutex);
1901 rte_eth_dev_info_get(dev->port_id, &dev_info);
1902 ovs_mutex_unlock(&dev->mutex);
1904 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1906 smap_add_format(args, "port_no", "%d", dev->port_id);
1907 smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id));
1908 smap_add_format(args, "driver_name", "%s", dev_info.driver_name);
1909 smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize);
1910 smap_add_format(args, "max_rx_pktlen", "%u", dev->max_packet_len);
1911 smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues);
1912 smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues);
1913 smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs);
1914 smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs);
1915 smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs);
1916 smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools);
1918 if (dev_info.pci_dev) {
1919 smap_add_format(args, "pci-vendor_id", "0x%u",
1920 dev_info.pci_dev->id.vendor_id);
1921 smap_add_format(args, "pci-device_id", "0x%x",
1922 dev_info.pci_dev->id.device_id);
1929 netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state)
1930 OVS_REQUIRES(dev->mutex)
1932 enum netdev_flags old_flags;
1935 netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags);
1937 netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags);
1942 netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc,
1943 const char *argv[], void *aux OVS_UNUSED)
1947 if (!strcasecmp(argv[argc - 1], "up")) {
1949 } else if ( !strcasecmp(argv[argc - 1], "down")) {
1952 unixctl_command_reply_error(conn, "Invalid Admin State");
1957 struct netdev *netdev = netdev_from_name(argv[1]);
1958 if (netdev && is_dpdk_class(netdev->netdev_class)) {
1959 struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev);
1961 ovs_mutex_lock(&dpdk_dev->mutex);
1962 netdev_dpdk_set_admin_state__(dpdk_dev, up);
1963 ovs_mutex_unlock(&dpdk_dev->mutex);
1965 netdev_close(netdev);
1967 unixctl_command_reply_error(conn, "Not a DPDK Interface");
1968 netdev_close(netdev);
1972 struct netdev_dpdk *netdev;
1974 ovs_mutex_lock(&dpdk_mutex);
1975 LIST_FOR_EACH (netdev, list_node, &dpdk_list) {
1976 ovs_mutex_lock(&netdev->mutex);
1977 netdev_dpdk_set_admin_state__(netdev, up);
1978 ovs_mutex_unlock(&netdev->mutex);
1980 ovs_mutex_unlock(&dpdk_mutex);
1982 unixctl_command_reply(conn, "OK");
1986 * Set virtqueue flags so that we do not receive interrupts.
1989 set_irq_status(struct virtio_net *dev)
1994 for (i = 0; i < dev->virt_qp_nb; i++) {
1995 idx = i * VIRTIO_QNUM;
1996 rte_vhost_enable_guest_notification(dev, idx + VIRTIO_RXQ, 0);
1997 rte_vhost_enable_guest_notification(dev, idx + VIRTIO_TXQ, 0);
2002 * Fixes mapping for vhost-user tx queues. Must be called after each
2003 * enabling/disabling of queues and real_n_txq modifications.
2006 netdev_dpdk_remap_txqs(struct netdev_dpdk *netdev)
2007 OVS_REQUIRES(netdev->mutex)
2009 int *enabled_queues, n_enabled = 0;
2010 int i, k, total_txqs = netdev->real_n_txq;
2012 enabled_queues = dpdk_rte_mzalloc(total_txqs * sizeof *enabled_queues);
2014 for (i = 0; i < total_txqs; i++) {
2015 /* Enabled queues always mapped to themselves. */
2016 if (netdev->tx_q[i].map == i) {
2017 enabled_queues[n_enabled++] = i;
2021 if (n_enabled == 0 && total_txqs != 0) {
2022 enabled_queues[0] = -1;
2027 for (i = 0; i < total_txqs; i++) {
2028 if (netdev->tx_q[i].map != i) {
2029 netdev->tx_q[i].map = enabled_queues[k];
2030 k = (k + 1) % n_enabled;
2034 VLOG_DBG("TX queue mapping for %s\n", netdev->vhost_id);
2035 for (i = 0; i < total_txqs; i++) {
2036 VLOG_DBG("%2d --> %2d", i, netdev->tx_q[i].map);
2039 rte_free(enabled_queues);
2043 netdev_dpdk_vhost_set_queues(struct netdev_dpdk *netdev, struct virtio_net *dev)
2044 OVS_REQUIRES(netdev->mutex)
2048 qp_num = dev->virt_qp_nb;
2049 if (qp_num > netdev->up.n_rxq) {
2050 VLOG_ERR("vHost Device '%s' %"PRIu64" can't be added - "
2051 "too many queues %d > %d", dev->ifname, dev->device_fh,
2052 qp_num, netdev->up.n_rxq);
2056 netdev->real_n_rxq = qp_num;
2057 netdev->real_n_txq = qp_num;
2058 netdev->txq_needs_locking = true;
2060 netdev_dpdk_remap_txqs(netdev);
2066 * A new virtio-net device is added to a vhost port.
2069 new_device(struct virtio_net *dev)
2071 struct netdev_dpdk *netdev;
2072 bool exists = false;
2074 ovs_mutex_lock(&dpdk_mutex);
2075 /* Add device to the vhost port with the same name as that passed down. */
2076 LIST_FOR_EACH(netdev, list_node, &dpdk_list) {
2077 if (strncmp(dev->ifname, netdev->vhost_id, IF_NAME_SZ) == 0) {
2078 ovs_mutex_lock(&netdev->mutex);
2079 if (netdev_dpdk_vhost_set_queues(netdev, dev)) {
2080 ovs_mutex_unlock(&netdev->mutex);
2081 ovs_mutex_unlock(&dpdk_mutex);
2084 ovsrcu_set(&netdev->virtio_dev, dev);
2086 dev->flags |= VIRTIO_DEV_RUNNING;
2087 /* Disable notifications. */
2088 set_irq_status(dev);
2089 ovs_mutex_unlock(&netdev->mutex);
2093 ovs_mutex_unlock(&dpdk_mutex);
2096 VLOG_INFO("vHost Device '%s' %"PRIu64" can't be added - name not "
2097 "found", dev->ifname, dev->device_fh);
2102 VLOG_INFO("vHost Device '%s' %"PRIu64" has been added", dev->ifname,
2108 * Remove a virtio-net device from the specific vhost port. Use dev->remove
2109 * flag to stop any more packets from being sent or received to/from a VM and
2110 * ensure all currently queued packets have been sent/received before removing
2114 destroy_device(volatile struct virtio_net *dev)
2116 struct netdev_dpdk *vhost_dev;
2117 bool exists = false;
2119 ovs_mutex_lock(&dpdk_mutex);
2120 LIST_FOR_EACH (vhost_dev, list_node, &dpdk_list) {
2121 if (netdev_dpdk_get_virtio(vhost_dev) == dev) {
2123 ovs_mutex_lock(&vhost_dev->mutex);
2124 dev->flags &= ~VIRTIO_DEV_RUNNING;
2125 ovsrcu_set(&vhost_dev->virtio_dev, NULL);
2127 ovs_mutex_unlock(&vhost_dev->mutex);
2132 ovs_mutex_unlock(&dpdk_mutex);
2134 if (exists == true) {
2136 * Wait for other threads to quiesce after setting the 'virtio_dev'
2137 * to NULL, before returning.
2139 ovsrcu_synchronize();
2141 * As call to ovsrcu_synchronize() will end the quiescent state,
2142 * put thread back into quiescent state before returning.
2144 ovsrcu_quiesce_start();
2145 VLOG_INFO("vHost Device '%s' %"PRIu64" has been removed", dev->ifname,
2148 VLOG_INFO("vHost Device '%s' %"PRIu64" not found", dev->ifname,
2155 vring_state_changed(struct virtio_net *dev, uint16_t queue_id, int enable)
2157 struct netdev_dpdk *vhost_dev;
2158 bool exists = false;
2159 int qid = queue_id / VIRTIO_QNUM;
2161 if (queue_id % VIRTIO_QNUM == VIRTIO_TXQ) {
2165 ovs_mutex_lock(&dpdk_mutex);
2166 LIST_FOR_EACH (vhost_dev, list_node, &dpdk_list) {
2167 if (strncmp(dev->ifname, vhost_dev->vhost_id, IF_NAME_SZ) == 0) {
2168 ovs_mutex_lock(&vhost_dev->mutex);
2170 vhost_dev->tx_q[qid].map = qid;
2172 vhost_dev->tx_q[qid].map = -1;
2174 netdev_dpdk_remap_txqs(vhost_dev);
2176 ovs_mutex_unlock(&vhost_dev->mutex);
2180 ovs_mutex_unlock(&dpdk_mutex);
2183 VLOG_INFO("State of queue %d ( tx_qid %d ) of vhost device '%s' %"
2184 PRIu64" changed to \'%s\'", queue_id, qid, dev->ifname,
2185 dev->device_fh, (enable == 1) ? "enabled" : "disabled");
2187 VLOG_INFO("vHost Device '%s' %"PRIu64" not found", dev->ifname,
2196 netdev_dpdk_get_virtio(const struct netdev_dpdk *dev)
2198 return ovsrcu_get(struct virtio_net *, &dev->virtio_dev);
2202 * These callbacks allow virtio-net devices to be added to vhost ports when
2203 * configuration has been fully complete.
2205 static const struct virtio_net_device_ops virtio_net_device_ops =
2207 .new_device = new_device,
2208 .destroy_device = destroy_device,
2209 .vring_state_changed = vring_state_changed
2213 start_vhost_loop(void *dummy OVS_UNUSED)
2215 pthread_detach(pthread_self());
2216 /* Put the cuse thread into quiescent state. */
2217 ovsrcu_quiesce_start();
2218 rte_vhost_driver_session_start();
2223 dpdk_vhost_class_init(void)
2225 rte_vhost_driver_callback_register(&virtio_net_device_ops);
2226 ovs_thread_create("vhost_thread", start_vhost_loop, NULL);
2231 dpdk_vhost_cuse_class_init(void)
2236 /* Register CUSE device to handle IOCTLs.
2237 * Unless otherwise specified on the vswitchd command line, cuse_dev_name
2238 * is set to vhost-net.
2240 err = rte_vhost_driver_register(cuse_dev_name);
2243 VLOG_ERR("CUSE device setup failure.");
2247 dpdk_vhost_class_init();
2252 dpdk_vhost_user_class_init(void)
2254 dpdk_vhost_class_init();
2259 dpdk_common_init(void)
2261 unixctl_command_register("netdev-dpdk/set-admin-state",
2262 "[netdev] up|down", 1, 2,
2263 netdev_dpdk_set_admin_state, NULL);
2265 ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL);
2271 dpdk_ring_create(const char dev_name[], unsigned int port_no,
2272 unsigned int *eth_port_id)
2274 struct dpdk_ring *ivshmem;
2275 char ring_name[RTE_RING_NAMESIZE];
2278 ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem);
2279 if (ivshmem == NULL) {
2283 /* XXX: Add support for multiquque ring. */
2284 err = snprintf(ring_name, sizeof(ring_name), "%s_tx", dev_name);
2289 /* Create single producer tx ring, netdev does explicit locking. */
2290 ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
2292 if (ivshmem->cring_tx == NULL) {
2297 err = snprintf(ring_name, sizeof(ring_name), "%s_rx", dev_name);
2302 /* Create single consumer rx ring, netdev does explicit locking. */
2303 ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0,
2305 if (ivshmem->cring_rx == NULL) {
2310 err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1,
2311 &ivshmem->cring_tx, 1, SOCKET0);
2318 ivshmem->user_port_id = port_no;
2319 ivshmem->eth_port_id = rte_eth_dev_count() - 1;
2320 list_push_back(&dpdk_ring_list, &ivshmem->list_node);
2322 *eth_port_id = ivshmem->eth_port_id;
2327 dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex)
2329 struct dpdk_ring *ivshmem;
2330 unsigned int port_no;
2333 /* Names always start with "dpdkr" */
2334 err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no);
2339 /* look through our list to find the device */
2340 LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) {
2341 if (ivshmem->user_port_id == port_no) {
2342 VLOG_INFO("Found dpdk ring device %s:", dev_name);
2343 *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */
2347 /* Need to create the device rings */
2348 return dpdk_ring_create(dev_name, port_no, eth_port_id);
2352 netdev_dpdk_ring_send(struct netdev *netdev_, int qid,
2353 struct dp_packet **pkts, int cnt, bool may_steal)
2355 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
2358 /* When using 'dpdkr' and sending to a DPDK ring, we want to ensure that the
2359 * rss hash field is clear. This is because the same mbuf may be modified by
2360 * the consumer of the ring and return into the datapath without recalculating
2362 for (i = 0; i < cnt; i++) {
2363 dp_packet_rss_invalidate(pkts[i]);
2366 netdev_dpdk_send__(netdev, qid, pkts, cnt, may_steal);
2371 netdev_dpdk_ring_construct(struct netdev *netdev)
2373 unsigned int port_no = 0;
2376 if (rte_eal_init_ret) {
2377 return rte_eal_init_ret;
2380 ovs_mutex_lock(&dpdk_mutex);
2382 err = dpdk_ring_open(netdev->name, &port_no);
2387 err = netdev_dpdk_init(netdev, port_no, DPDK_DEV_ETH);
2390 ovs_mutex_unlock(&dpdk_mutex);
2397 * Initialize QoS configuration operations.
2400 qos_conf_init(struct qos_conf *conf, const struct dpdk_qos_ops *ops)
2406 * Search existing QoS operations in qos_ops and compare each set of
2407 * operations qos_name to name. Return a dpdk_qos_ops pointer to a match,
2410 static const struct dpdk_qos_ops *
2411 qos_lookup_name(const char *name)
2413 const struct dpdk_qos_ops *const *opsp;
2415 for (opsp = qos_confs; *opsp != NULL; opsp++) {
2416 const struct dpdk_qos_ops *ops = *opsp;
2417 if (!strcmp(name, ops->qos_name)) {
2425 * Call qos_destruct to clean up items associated with the netdevs
2426 * qos_conf. Set netdevs qos_conf to NULL.
2429 qos_delete_conf(struct netdev *netdev_)
2431 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
2433 rte_spinlock_lock(&netdev->qos_lock);
2434 if (netdev->qos_conf) {
2435 if (netdev->qos_conf->ops->qos_destruct) {
2436 netdev->qos_conf->ops->qos_destruct(netdev_, netdev->qos_conf);
2438 netdev->qos_conf = NULL;
2440 rte_spinlock_unlock(&netdev->qos_lock);
2444 netdev_dpdk_get_qos_types(const struct netdev *netdev OVS_UNUSED,
2447 const struct dpdk_qos_ops *const *opsp;
2449 for (opsp = qos_confs; *opsp != NULL; opsp++) {
2450 const struct dpdk_qos_ops *ops = *opsp;
2451 if (ops->qos_construct && ops->qos_name[0] != '\0') {
2452 sset_add(types, ops->qos_name);
2459 netdev_dpdk_get_qos(const struct netdev *netdev_,
2460 const char **typep, struct smap *details)
2462 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
2465 ovs_mutex_lock(&netdev->mutex);
2466 if(netdev->qos_conf) {
2467 *typep = netdev->qos_conf->ops->qos_name;
2468 error = (netdev->qos_conf->ops->qos_get
2469 ? netdev->qos_conf->ops->qos_get(netdev_, details): 0);
2471 ovs_mutex_unlock(&netdev->mutex);
2477 netdev_dpdk_set_qos(struct netdev *netdev_,
2478 const char *type, const struct smap *details)
2480 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
2481 const struct dpdk_qos_ops *new_ops = NULL;
2484 /* If type is empty or unsupported then the current QoS configuration
2485 * for the dpdk-netdev can be destroyed */
2486 new_ops = qos_lookup_name(type);
2488 if (type[0] == '\0' || !new_ops || !new_ops->qos_construct) {
2489 qos_delete_conf(netdev_);
2493 ovs_mutex_lock(&netdev->mutex);
2495 if (netdev->qos_conf) {
2496 if (new_ops == netdev->qos_conf->ops) {
2497 error = new_ops->qos_set ? new_ops->qos_set(netdev_, details) : 0;
2499 /* Delete existing QoS configuration. */
2500 qos_delete_conf(netdev_);
2501 ovs_assert(netdev->qos_conf == NULL);
2503 /* Install new QoS configuration. */
2504 error = new_ops->qos_construct(netdev_, details);
2505 ovs_assert((error == 0) == (netdev->qos_conf != NULL));
2508 error = new_ops->qos_construct(netdev_, details);
2509 ovs_assert((error == 0) == (netdev->qos_conf != NULL));
2512 ovs_mutex_unlock(&netdev->mutex);
2516 /* egress-policer details */
2518 struct egress_policer {
2519 struct qos_conf qos_conf;
2520 struct rte_meter_srtcm_params app_srtcm_params;
2521 struct rte_meter_srtcm egress_meter;
2524 static struct egress_policer *
2525 egress_policer_get__(const struct netdev *netdev_)
2527 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
2528 return CONTAINER_OF(netdev->qos_conf, struct egress_policer, qos_conf);
2532 egress_policer_qos_construct(struct netdev *netdev_,
2533 const struct smap *details)
2535 struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_);
2536 struct egress_policer *policer;
2541 rte_spinlock_lock(&netdev->qos_lock);
2542 policer = xmalloc(sizeof *policer);
2543 qos_conf_init(&policer->qos_conf, &egress_policer_ops);
2544 netdev->qos_conf = &policer->qos_conf;
2545 cir_s = smap_get(details, "cir");
2546 cbs_s = smap_get(details, "cbs");
2547 policer->app_srtcm_params.cir = cir_s ? strtoull(cir_s, NULL, 10) : 0;
2548 policer->app_srtcm_params.cbs = cbs_s ? strtoull(cbs_s, NULL, 10) : 0;
2549 policer->app_srtcm_params.ebs = 0;
2550 err = rte_meter_srtcm_config(&policer->egress_meter,
2551 &policer->app_srtcm_params);
2552 rte_spinlock_unlock(&netdev->qos_lock);
2558 egress_policer_qos_destruct(struct netdev *netdev_ OVS_UNUSED,
2559 struct qos_conf *conf)
2561 struct egress_policer *policer = CONTAINER_OF(conf, struct egress_policer,
2567 egress_policer_qos_get(const struct netdev *netdev, struct smap *details)
2569 struct egress_policer *policer = egress_policer_get__(netdev);
2570 smap_add_format(details, "cir", "%llu",
2571 1ULL * policer->app_srtcm_params.cir);
2572 smap_add_format(details, "cbs", "%llu",
2573 1ULL * policer->app_srtcm_params.cbs);
2578 egress_policer_qos_set(struct netdev *netdev_, const struct smap *details)
2580 struct egress_policer *policer;
2585 policer = egress_policer_get__(netdev_);
2586 cir_s = smap_get(details, "cir");
2587 cbs_s = smap_get(details, "cbs");
2588 policer->app_srtcm_params.cir = cir_s ? strtoull(cir_s, NULL, 10) : 0;
2589 policer->app_srtcm_params.cbs = cbs_s ? strtoull(cbs_s, NULL, 10) : 0;
2590 policer->app_srtcm_params.ebs = 0;
2591 err = rte_meter_srtcm_config(&policer->egress_meter,
2592 &policer->app_srtcm_params);
2598 egress_policer_pkt_handle__(struct rte_meter_srtcm *meter,
2599 struct rte_mbuf *pkt, uint64_t time)
2601 uint32_t pkt_len = rte_pktmbuf_pkt_len(pkt) - sizeof(struct ether_hdr);
2603 return rte_meter_srtcm_color_blind_check(meter, time, pkt_len) ==
2608 egress_policer_run(struct netdev *netdev_, struct rte_mbuf **pkts,
2613 struct egress_policer *policer = egress_policer_get__(netdev_);
2614 struct rte_mbuf *pkt = NULL;
2615 uint64_t current_time = rte_rdtsc();
2617 for(i = 0; i < pkt_cnt; i++) {
2619 /* Handle current packet */
2620 if (egress_policer_pkt_handle__(&policer->egress_meter, pkt,
2627 rte_pktmbuf_free(pkt);
2634 static const struct dpdk_qos_ops egress_policer_ops = {
2635 "egress-policer", /* qos_name */
2636 egress_policer_qos_construct,
2637 egress_policer_qos_destruct,
2638 egress_policer_qos_get,
2639 egress_policer_qos_set,
2643 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, DESTRUCT, MULTIQ, SEND, \
2644 GET_CARRIER, GET_STATS, GET_FEATURES, GET_STATUS, RXQ_RECV) \
2647 true, /* is_pmd */ \
2649 NULL, /* netdev_dpdk_run */ \
2650 NULL, /* netdev_dpdk_wait */ \
2652 netdev_dpdk_alloc, \
2655 netdev_dpdk_dealloc, \
2656 netdev_dpdk_get_config, \
2657 netdev_dpdk_set_config, \
2658 NULL, /* get_tunnel_config */ \
2659 NULL, /* build header */ \
2660 NULL, /* push header */ \
2661 NULL, /* pop header */ \
2662 netdev_dpdk_get_numa_id, /* get_numa_id */ \
2663 MULTIQ, /* set_multiq */ \
2666 NULL, /* send_wait */ \
2668 netdev_dpdk_set_etheraddr, \
2669 netdev_dpdk_get_etheraddr, \
2670 netdev_dpdk_get_mtu, \
2671 netdev_dpdk_set_mtu, \
2672 netdev_dpdk_get_ifindex, \
2674 netdev_dpdk_get_carrier_resets, \
2675 netdev_dpdk_set_miimon, \
2678 NULL, /* set_advertisements */ \
2680 NULL, /* set_policing */ \
2681 netdev_dpdk_get_qos_types, \
2682 NULL, /* get_qos_capabilities */ \
2683 netdev_dpdk_get_qos, \
2684 netdev_dpdk_set_qos, \
2685 NULL, /* get_queue */ \
2686 NULL, /* set_queue */ \
2687 NULL, /* delete_queue */ \
2688 NULL, /* get_queue_stats */ \
2689 NULL, /* queue_dump_start */ \
2690 NULL, /* queue_dump_next */ \
2691 NULL, /* queue_dump_done */ \
2692 NULL, /* dump_queue_stats */ \
2694 NULL, /* get_in4 */ \
2695 NULL, /* set_in4 */ \
2696 NULL, /* get_addr_list */ \
2697 NULL, /* add_router */ \
2698 NULL, /* get_next_hop */ \
2700 NULL, /* arp_lookup */ \
2702 netdev_dpdk_update_flags, \
2704 netdev_dpdk_rxq_alloc, \
2705 netdev_dpdk_rxq_construct, \
2706 netdev_dpdk_rxq_destruct, \
2707 netdev_dpdk_rxq_dealloc, \
2709 NULL, /* rx_wait */ \
2710 NULL, /* rxq_drain */ \
2714 process_vhost_flags(char *flag, char *default_val, int size,
2715 char **argv, char **new_val)
2719 /* Depending on which version of vhost is in use, process the vhost-specific
2720 * flag if it is provided on the vswitchd command line, otherwise resort to
2723 * For vhost-user: Process "-vhost_sock_dir" to set the custom location of
2724 * the vhost-user socket(s).
2725 * For vhost-cuse: Process "-cuse_dev_name" to set the custom name of the
2726 * vhost-cuse character device.
2728 if (!strcmp(argv[1], flag) && (strlen(argv[2]) <= size)) {
2730 *new_val = xstrdup(argv[2]);
2731 VLOG_INFO("User-provided %s in use: %s", flag, *new_val);
2733 VLOG_INFO("No %s provided - defaulting to %s", flag, default_val);
2734 *new_val = default_val;
2741 dpdk_init(int argc, char **argv)
2745 char *pragram_name = argv[0];
2747 if (argc < 2 || strcmp(argv[1], "--dpdk"))
2750 /* Remove the --dpdk argument from arg list.*/
2754 /* Reject --user option */
2756 for (i = 0; i < argc; i++) {
2757 if (!strcmp(argv[i], "--user")) {
2758 VLOG_ERR("Can not mix --dpdk and --user options, aborting.");
2763 if (process_vhost_flags("-cuse_dev_name", xstrdup("vhost-net"),
2764 PATH_MAX, argv, &cuse_dev_name)) {
2766 if (process_vhost_flags("-vhost_sock_dir", xstrdup(ovs_rundir()),
2767 NAME_MAX, argv, &vhost_sock_dir)) {
2771 err = stat(vhost_sock_dir, &s);
2773 VLOG_ERR("vHostUser socket DIR '%s' does not exist.",
2778 /* Remove the vhost flag configuration parameters from the argument
2779 * list, so that the correct elements are passed to the DPDK
2780 * initialization function
2783 argv += 2; /* Increment by two to bypass the vhost flag arguments */
2787 /* Keep the program name argument as this is needed for call to
2790 argv[0] = pragram_name;
2792 /* Make sure things are initialized ... */
2793 result = rte_eal_init(argc, argv);
2795 ovs_abort(result, "Cannot init EAL");
2798 rte_memzone_dump(stdout);
2799 rte_eal_init_ret = 0;
2801 if (argc > result) {
2802 argv[result] = argv[0];
2805 /* We are called from the main thread here */
2806 RTE_PER_LCORE(_lcore_id) = NON_PMD_CORE_ID;
2808 return result + 1 + base;
2811 static const struct netdev_class dpdk_class =
2815 netdev_dpdk_construct,
2816 netdev_dpdk_destruct,
2817 netdev_dpdk_set_multiq,
2818 netdev_dpdk_eth_send,
2819 netdev_dpdk_get_carrier,
2820 netdev_dpdk_get_stats,
2821 netdev_dpdk_get_features,
2822 netdev_dpdk_get_status,
2823 netdev_dpdk_rxq_recv);
2825 static const struct netdev_class dpdk_ring_class =
2829 netdev_dpdk_ring_construct,
2830 netdev_dpdk_destruct,
2831 netdev_dpdk_set_multiq,
2832 netdev_dpdk_ring_send,
2833 netdev_dpdk_get_carrier,
2834 netdev_dpdk_get_stats,
2835 netdev_dpdk_get_features,
2836 netdev_dpdk_get_status,
2837 netdev_dpdk_rxq_recv);
2839 static const struct netdev_class OVS_UNUSED dpdk_vhost_cuse_class =
2842 dpdk_vhost_cuse_class_init,
2843 netdev_dpdk_vhost_cuse_construct,
2844 netdev_dpdk_vhost_destruct,
2845 netdev_dpdk_vhost_cuse_set_multiq,
2846 netdev_dpdk_vhost_send,
2847 netdev_dpdk_vhost_get_carrier,
2848 netdev_dpdk_vhost_get_stats,
2851 netdev_dpdk_vhost_rxq_recv);
2853 static const struct netdev_class OVS_UNUSED dpdk_vhost_user_class =
2856 dpdk_vhost_user_class_init,
2857 netdev_dpdk_vhost_user_construct,
2858 netdev_dpdk_vhost_destruct,
2859 netdev_dpdk_vhost_set_multiq,
2860 netdev_dpdk_vhost_send,
2861 netdev_dpdk_vhost_get_carrier,
2862 netdev_dpdk_vhost_get_stats,
2865 netdev_dpdk_vhost_rxq_recv);
2868 netdev_dpdk_register(void)
2870 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
2872 if (rte_eal_init_ret) {
2876 if (ovsthread_once_start(&once)) {
2878 netdev_register_provider(&dpdk_class);
2879 netdev_register_provider(&dpdk_ring_class);
2881 netdev_register_provider(&dpdk_vhost_cuse_class);
2883 netdev_register_provider(&dpdk_vhost_user_class);
2885 ovsthread_once_done(&once);
2890 pmd_thread_setaffinity_cpu(unsigned cpu)
2896 CPU_SET(cpu, &cpuset);
2897 err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
2899 VLOG_ERR("Thread affinity error %d",err);
2902 /* NON_PMD_CORE_ID is reserved for use by non pmd threads. */
2903 ovs_assert(cpu != NON_PMD_CORE_ID);
2904 RTE_PER_LCORE(_lcore_id) = cpu;
2910 dpdk_thread_is_pmd(void)
2912 return rte_lcore_id() != NON_PMD_CORE_ID;