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.
18 #include "netdev-dpdk.h"
26 #include <rte_config.h>
27 #include <rte_cycles.h>
28 #include <rte_errno.h>
29 #include <rte_eth_ring.h>
30 #include <rte_ethdev.h>
31 #include <rte_malloc.h>
33 #include <rte_meter.h>
34 #include <rte_virtio_net.h>
37 #include "dp-packet.h"
39 #include "dpif-netdev.h"
40 #include "fatal-signal.h"
41 #include "netdev-provider.h"
42 #include "netdev-vport.h"
44 #include "openvswitch/dynamic-string.h"
45 #include "openvswitch/list.h"
46 #include "openvswitch/ofp-print.h"
47 #include "openvswitch/vlog.h"
49 #include "ovs-thread.h"
52 #include "openvswitch/shash.h"
55 #include "unaligned.h"
59 VLOG_DEFINE_THIS_MODULE(netdev_dpdk
);
60 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(5, 20);
62 #define DPDK_PORT_WATCHDOG_INTERVAL 5
64 #define OVS_CACHE_LINE_SIZE CACHE_LINE_SIZE
65 #define OVS_VPORT_DPDK "ovs_dpdk"
68 * need to reserve tons of extra space in the mbufs so we can align the
69 * DMA addresses to 4KB.
70 * The minimum mbuf size is limited to avoid scatter behaviour and drop in
71 * performance for standard Ethernet MTU.
73 #define ETHER_HDR_MAX_LEN (ETHER_HDR_LEN + ETHER_CRC_LEN \
74 + (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) \
78 - ETHER_HDR_LEN - ETHER_CRC_LEN)
79 #define MBUF_SIZE(mtu) (MTU_TO_MAX_FRAME_LEN(mtu) \
80 + sizeof(struct dp_packet) \
81 + RTE_PKTMBUF_HEADROOM)
82 #define NETDEV_DPDK_MBUF_ALIGN 1024
83 #define NETDEV_DPDK_MAX_PKT_LEN 9728
85 /* Max and min number of packets in the mempool. OVS tries to allocate a
86 * mempool with MAX_NB_MBUF: if this fails (because the system doesn't have
87 * enough hugepages) we keep halving the number until the allocation succeeds
88 * or we reach MIN_NB_MBUF */
90 #define MAX_NB_MBUF (4096 * 64)
91 #define MIN_NB_MBUF (4096 * 4)
92 #define MP_CACHE_SZ RTE_MEMPOOL_CACHE_MAX_SIZE
94 /* MAX_NB_MBUF can be divided by 2 many times, until MIN_NB_MBUF */
95 BUILD_ASSERT_DECL(MAX_NB_MBUF
% ROUND_DOWN_POW2(MAX_NB_MBUF
/MIN_NB_MBUF
) == 0);
97 /* The smallest possible NB_MBUF that we're going to try should be a multiple
98 * of MP_CACHE_SZ. This is advised by DPDK documentation. */
99 BUILD_ASSERT_DECL((MAX_NB_MBUF
/ ROUND_DOWN_POW2(MAX_NB_MBUF
/MIN_NB_MBUF
))
103 * DPDK XSTATS Counter names definition
105 #define XSTAT_RX_64_PACKETS "rx_size_64_packets"
106 #define XSTAT_RX_65_TO_127_PACKETS "rx_size_65_to_127_packets"
107 #define XSTAT_RX_128_TO_255_PACKETS "rx_size_128_to_255_packets"
108 #define XSTAT_RX_256_TO_511_PACKETS "rx_size_256_to_511_packets"
109 #define XSTAT_RX_512_TO_1023_PACKETS "rx_size_512_to_1023_packets"
110 #define XSTAT_RX_1024_TO_1522_PACKETS "rx_size_1024_to_1522_packets"
111 #define XSTAT_RX_1523_TO_MAX_PACKETS "rx_size_1523_to_max_packets"
113 #define XSTAT_TX_64_PACKETS "tx_size_64_packets"
114 #define XSTAT_TX_65_TO_127_PACKETS "tx_size_65_to_127_packets"
115 #define XSTAT_TX_128_TO_255_PACKETS "tx_size_128_to_255_packets"
116 #define XSTAT_TX_256_TO_511_PACKETS "tx_size_256_to_511_packets"
117 #define XSTAT_TX_512_TO_1023_PACKETS "tx_size_512_to_1023_packets"
118 #define XSTAT_TX_1024_TO_1522_PACKETS "tx_size_1024_to_1522_packets"
119 #define XSTAT_TX_1523_TO_MAX_PACKETS "tx_size_1523_to_max_packets"
121 #define XSTAT_TX_MULTICAST_PACKETS "tx_multicast_packets"
122 #define XSTAT_RX_BROADCAST_PACKETS "rx_broadcast_packets"
123 #define XSTAT_TX_BROADCAST_PACKETS "tx_broadcast_packets"
124 #define XSTAT_RX_UNDERSIZED_ERRORS "rx_undersized_errors"
125 #define XSTAT_RX_OVERSIZE_ERRORS "rx_oversize_errors"
126 #define XSTAT_RX_FRAGMENTED_ERRORS "rx_fragmented_errors"
127 #define XSTAT_RX_JABBER_ERRORS "rx_jabber_errors"
131 /* Default size of Physical NIC RXQ */
132 #define NIC_PORT_DEFAULT_RXQ_SIZE 2048
133 /* Default size of Physical NIC TXQ */
134 #define NIC_PORT_DEFAULT_TXQ_SIZE 2048
135 /* Maximum size of Physical NIC Queues */
136 #define NIC_PORT_MAX_Q_SIZE 4096
138 #define OVS_VHOST_MAX_QUEUE_NUM 1024 /* Maximum number of vHost TX queues. */
139 #define OVS_VHOST_QUEUE_MAP_UNKNOWN (-1) /* Mapping not initialized. */
140 #define OVS_VHOST_QUEUE_DISABLED (-2) /* Queue was disabled by guest and not
141 * yet mapped to another queue. */
143 #define VHOST_ENQ_RETRY_NUM 8
144 #define IF_NAME_SZ (PATH_MAX > IFNAMSIZ ? PATH_MAX : IFNAMSIZ)
146 static const struct rte_eth_conf port_conf
= {
148 .mq_mode
= ETH_MQ_RX_RSS
,
150 .header_split
= 0, /* Header Split disabled */
151 .hw_ip_checksum
= 0, /* IP checksum offload disabled */
152 .hw_vlan_filter
= 0, /* VLAN filtering disabled */
153 .jumbo_frame
= 0, /* Jumbo Frame Support disabled */
159 .rss_hf
= ETH_RSS_IP
| ETH_RSS_UDP
| ETH_RSS_TCP
,
163 .mq_mode
= ETH_MQ_TX_NONE
,
167 enum { DPDK_RING_SIZE
= 256 };
168 BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE
));
169 enum { DRAIN_TSC
= 200000ULL };
176 /* Quality of Service */
178 /* An instance of a QoS configuration. Always associated with a particular
181 * Each QoS implementation subclasses this with whatever additional data it
185 const struct dpdk_qos_ops
*ops
;
189 /* A particular implementation of dpdk QoS operations.
191 * The functions below return 0 if successful or a positive errno value on
192 * failure, except where otherwise noted. All of them must be provided, except
193 * where otherwise noted.
195 struct dpdk_qos_ops
{
197 /* Name of the QoS type */
198 const char *qos_name
;
200 /* Called to construct a qos_conf object. The implementation should make
201 * the appropriate calls to configure QoS according to 'details'.
203 * The contents of 'details' should be documented as valid for 'ovs_name'
204 * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml
205 * (which is built as ovs-vswitchd.conf.db(8)).
207 * This function must return 0 if and only if it sets '*conf' to an
208 * initialized 'struct qos_conf'.
210 * For all QoS implementations it should always be non-null.
212 int (*qos_construct
)(const struct smap
*details
, struct qos_conf
**conf
);
214 /* Destroys the data structures allocated by the implementation as part of
217 * For all QoS implementations it should always be non-null.
219 void (*qos_destruct
)(struct qos_conf
*conf
);
221 /* Retrieves details of 'conf' configuration into 'details'.
223 * The contents of 'details' should be documented as valid for 'ovs_name'
224 * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml
225 * (which is built as ovs-vswitchd.conf.db(8)).
227 int (*qos_get
)(const struct qos_conf
*conf
, struct smap
*details
);
229 /* Returns true if 'conf' is already configured according to 'details'.
231 * The contents of 'details' should be documented as valid for 'ovs_name'
232 * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml
233 * (which is built as ovs-vswitchd.conf.db(8)).
235 * For all QoS implementations it should always be non-null.
237 bool (*qos_is_equal
)(const struct qos_conf
*conf
,
238 const struct smap
*details
);
240 /* Modify an array of rte_mbufs. The modification is specific to
241 * each qos implementation.
243 * The function should take and array of mbufs and an int representing
244 * the current number of mbufs present in the array.
246 * After the function has performed a qos modification to the array of
247 * mbufs it returns an int representing the number of mbufs now present in
248 * the array. This value is can then be passed to the port send function
249 * along with the modified array for transmission.
251 * For all QoS implementations it should always be non-null.
253 int (*qos_run
)(struct qos_conf
*qos_conf
, struct rte_mbuf
**pkts
,
257 /* dpdk_qos_ops for each type of user space QoS implementation */
258 static const struct dpdk_qos_ops egress_policer_ops
;
261 * Array of dpdk_qos_ops, contains pointer to all supported QoS
264 static const struct dpdk_qos_ops
*const qos_confs
[] = {
269 static struct ovs_mutex dpdk_mutex
= OVS_MUTEX_INITIALIZER
;
271 /* Contains all 'struct dpdk_dev's. */
272 static struct ovs_list dpdk_list
OVS_GUARDED_BY(dpdk_mutex
)
273 = OVS_LIST_INITIALIZER(&dpdk_list
);
275 static struct ovs_mutex dpdk_mp_mutex
OVS_ACQ_AFTER(dpdk_mutex
)
276 = OVS_MUTEX_INITIALIZER
;
278 static struct ovs_list dpdk_mp_list
OVS_GUARDED_BY(dpdk_mp_mutex
)
279 = OVS_LIST_INITIALIZER(&dpdk_mp_list
);
282 struct rte_mempool
*mp
;
286 struct ovs_list list_node
OVS_GUARDED_BY(dpdk_mp_mutex
);
289 /* There should be one 'struct dpdk_tx_queue' created for
291 struct dpdk_tx_queue
{
292 rte_spinlock_t tx_lock
; /* Protects the members and the NIC queue
293 * from concurrent access. It is used only
294 * if the queue is shared among different
295 * pmd threads (see 'concurrent_txq'). */
296 int map
; /* Mapping of configured vhost-user queues
297 * to enabled by guest. */
300 /* dpdk has no way to remove dpdk ring ethernet devices
301 so we have to keep them around once they've been created
304 static struct ovs_list dpdk_ring_list
OVS_GUARDED_BY(dpdk_mutex
)
305 = OVS_LIST_INITIALIZER(&dpdk_ring_list
);
308 /* For the client rings */
309 struct rte_ring
*cring_tx
;
310 struct rte_ring
*cring_rx
;
311 unsigned int user_port_id
; /* User given port no, parsed from port name */
312 int eth_port_id
; /* ethernet device port id */
313 struct ovs_list list_node
OVS_GUARDED_BY(dpdk_mutex
);
316 struct ingress_policer
{
317 struct rte_meter_srtcm_params app_srtcm_params
;
318 struct rte_meter_srtcm in_policer
;
319 rte_spinlock_t policer_lock
;
322 enum dpdk_hw_ol_features
{
323 NETDEV_RX_CHECKSUM_OFFLOAD
= 1 << 0,
330 enum dpdk_dev_type type
;
332 struct dpdk_tx_queue
*tx_q
;
334 struct ovs_mutex mutex
OVS_ACQ_AFTER(dpdk_mutex
);
336 struct dpdk_mp
*dpdk_mp
;
340 struct netdev_stats stats
;
342 rte_spinlock_t stats_lock
;
344 struct eth_addr hwaddr
;
345 enum netdev_flags flags
;
347 struct rte_eth_link link
;
350 /* virtio identifier for vhost devices */
353 /* True if vHost device is 'up' and has been reconfigured at least once */
354 bool vhost_reconfigured
;
356 /* Identifier used to distinguish vhost devices from each other. */
357 char vhost_id
[PATH_MAX
];
360 struct ovs_list list_node
OVS_GUARDED_BY(dpdk_mutex
);
362 /* QoS configuration and lock for the device */
363 OVSRCU_TYPE(struct qos_conf
*) qos_conf
;
365 /* The following properties cannot be changed when a device is running,
366 * so we remember the request and update them next time
367 * netdev_dpdk*_reconfigure() is called */
371 int requested_rxq_size
;
372 int requested_txq_size
;
374 /* Number of rx/tx descriptors for physical devices */
378 /* Socket ID detected when vHost device is brought up */
379 int requested_socket_id
;
381 /* Denotes whether vHost port is client/server mode */
382 uint64_t vhost_driver_flags
;
384 /* Ingress Policer */
385 OVSRCU_TYPE(struct ingress_policer
*) ingress_policer
;
386 uint32_t policer_rate
;
387 uint32_t policer_burst
;
389 /* DPDK-ETH Flow control */
390 struct rte_eth_fc_conf fc_conf
;
392 /* DPDK-ETH hardware offload features,
393 * from the enum set 'dpdk_hw_ol_features' */
394 uint32_t hw_ol_features
;
397 struct netdev_rxq_dpdk
{
398 struct netdev_rxq up
;
402 static int netdev_dpdk_class_init(void);
403 static int netdev_dpdk_vhost_class_init(void);
405 int netdev_dpdk_get_vid(const struct netdev_dpdk
*dev
);
407 struct ingress_policer
*
408 netdev_dpdk_get_ingress_policer(const struct netdev_dpdk
*dev
);
411 is_dpdk_class(const struct netdev_class
*class)
413 return class->init
== netdev_dpdk_class_init
414 || class->init
== netdev_dpdk_vhost_class_init
;
417 /* DPDK NIC drivers allocate RX buffers at a particular granularity, typically
418 * aligned at 1k or less. If a declared mbuf size is not a multiple of this
419 * value, insufficient buffers are allocated to accomodate the packet in its
420 * entirety. Furthermore, certain drivers need to ensure that there is also
421 * sufficient space in the Rx buffer to accommodate two VLAN tags (for QinQ
422 * frames). If the RX buffer is too small, then the driver enables scatter RX
423 * behaviour, which reduces performance. To prevent this, use a buffer size
424 * that is closest to 'mtu', but which satisfies the aforementioned criteria.
427 dpdk_buf_size(int mtu
)
429 return ROUND_UP((MTU_TO_MAX_FRAME_LEN(mtu
) + RTE_PKTMBUF_HEADROOM
),
430 NETDEV_DPDK_MBUF_ALIGN
);
433 /* Allocates an area of 'sz' bytes from DPDK. The memory is zero'ed.
435 * Unlike xmalloc(), this function can return NULL on failure. */
437 dpdk_rte_mzalloc(size_t sz
)
439 return rte_zmalloc(OVS_VPORT_DPDK
, sz
, OVS_CACHE_LINE_SIZE
);
443 free_dpdk_buf(struct dp_packet
*p
)
445 struct rte_mbuf
*pkt
= (struct rte_mbuf
*) p
;
447 rte_pktmbuf_free(pkt
);
451 ovs_rte_pktmbuf_init(struct rte_mempool
*mp
,
452 void *opaque_arg OVS_UNUSED
,
454 unsigned i OVS_UNUSED
)
456 struct rte_mbuf
*pkt
= _p
;
458 rte_pktmbuf_init(mp
, opaque_arg
, _p
, i
);
460 dp_packet_init_dpdk((struct dp_packet
*) pkt
, pkt
->buf_len
);
463 static struct dpdk_mp
*
464 dpdk_mp_create(int socket_id
, int mtu
)
466 struct rte_pktmbuf_pool_private mbp_priv
;
471 dmp
= dpdk_rte_mzalloc(sizeof *dmp
);
475 dmp
->socket_id
= socket_id
;
478 mbp_priv
.mbuf_data_room_size
= MBUF_SIZE(mtu
) - sizeof(struct dp_packet
);
479 mbp_priv
.mbuf_priv_size
= sizeof(struct dp_packet
)
480 - sizeof(struct rte_mbuf
);
481 /* XXX: this is a really rough method of provisioning memory.
482 * It's impossible to determine what the exact memory requirements are
483 * when the number of ports and rxqs that utilize a particular mempool can
484 * change dynamically at runtime. For now, use this rough heurisitic.
486 if (mtu
>= ETHER_MTU
) {
487 mp_size
= MAX_NB_MBUF
;
489 mp_size
= MIN_NB_MBUF
;
493 mp_name
= xasprintf("ovs_mp_%d_%d_%u", dmp
->mtu
, dmp
->socket_id
,
496 dmp
->mp
= rte_mempool_create(mp_name
, mp_size
, MBUF_SIZE(mtu
),
498 sizeof(struct rte_pktmbuf_pool_private
),
499 rte_pktmbuf_pool_init
, &mbp_priv
,
500 ovs_rte_pktmbuf_init
, NULL
,
503 VLOG_DBG("Allocated \"%s\" mempool with %u mbufs",
510 } while (rte_errno
== ENOMEM
&& (mp_size
/= 2) >= MIN_NB_MBUF
);
516 static struct dpdk_mp
*
517 dpdk_mp_get(int socket_id
, int mtu
)
521 ovs_mutex_lock(&dpdk_mp_mutex
);
522 LIST_FOR_EACH (dmp
, list_node
, &dpdk_mp_list
) {
523 if (dmp
->socket_id
== socket_id
&& dmp
->mtu
== mtu
) {
529 dmp
= dpdk_mp_create(socket_id
, mtu
);
530 ovs_list_push_back(&dpdk_mp_list
, &dmp
->list_node
);
533 ovs_mutex_unlock(&dpdk_mp_mutex
);
539 dpdk_mp_put(struct dpdk_mp
*dmp
)
545 ovs_mutex_lock(&dpdk_mp_mutex
);
546 ovs_assert(dmp
->refcount
);
548 if (!--dmp
->refcount
) {
549 ovs_list_remove(&dmp
->list_node
);
550 rte_mempool_free(dmp
->mp
);
553 ovs_mutex_unlock(&dpdk_mp_mutex
);
556 /* Tries to allocate new mempool on requested_socket_id with
557 * mbuf size corresponding to requested_mtu.
558 * On success new configuration will be applied.
559 * On error, device will be left unchanged. */
561 netdev_dpdk_mempool_configure(struct netdev_dpdk
*dev
)
562 OVS_REQUIRES(dev
->mutex
)
564 uint32_t buf_size
= dpdk_buf_size(dev
->requested_mtu
);
567 mp
= dpdk_mp_get(dev
->requested_socket_id
, FRAME_LEN_TO_MTU(buf_size
));
569 VLOG_ERR("Insufficient memory to create memory pool for netdev "
570 "%s, with MTU %d on socket %d\n",
571 dev
->up
.name
, dev
->requested_mtu
, dev
->requested_socket_id
);
574 dpdk_mp_put(dev
->dpdk_mp
);
576 dev
->mtu
= dev
->requested_mtu
;
577 dev
->socket_id
= dev
->requested_socket_id
;
578 dev
->max_packet_len
= MTU_TO_FRAME_LEN(dev
->mtu
);
585 check_link_status(struct netdev_dpdk
*dev
)
587 struct rte_eth_link link
;
589 rte_eth_link_get_nowait(dev
->port_id
, &link
);
591 if (dev
->link
.link_status
!= link
.link_status
) {
592 netdev_change_seq_changed(&dev
->up
);
594 dev
->link_reset_cnt
++;
596 if (dev
->link
.link_status
) {
597 VLOG_DBG_RL(&rl
, "Port %d Link Up - speed %u Mbps - %s",
598 dev
->port_id
, (unsigned) dev
->link
.link_speed
,
599 (dev
->link
.link_duplex
== ETH_LINK_FULL_DUPLEX
) ?
600 ("full-duplex") : ("half-duplex"));
602 VLOG_DBG_RL(&rl
, "Port %d Link Down", dev
->port_id
);
608 dpdk_watchdog(void *dummy OVS_UNUSED
)
610 struct netdev_dpdk
*dev
;
612 pthread_detach(pthread_self());
615 ovs_mutex_lock(&dpdk_mutex
);
616 LIST_FOR_EACH (dev
, list_node
, &dpdk_list
) {
617 ovs_mutex_lock(&dev
->mutex
);
618 if (dev
->type
== DPDK_DEV_ETH
) {
619 check_link_status(dev
);
621 ovs_mutex_unlock(&dev
->mutex
);
623 ovs_mutex_unlock(&dpdk_mutex
);
624 xsleep(DPDK_PORT_WATCHDOG_INTERVAL
);
631 dpdk_eth_dev_queue_setup(struct netdev_dpdk
*dev
, int n_rxq
, int n_txq
)
635 struct rte_eth_conf conf
= port_conf
;
637 if (dev
->mtu
> ETHER_MTU
) {
638 conf
.rxmode
.jumbo_frame
= 1;
639 conf
.rxmode
.max_rx_pkt_len
= dev
->max_packet_len
;
641 conf
.rxmode
.jumbo_frame
= 0;
642 conf
.rxmode
.max_rx_pkt_len
= 0;
644 conf
.rxmode
.hw_ip_checksum
= (dev
->hw_ol_features
&
645 NETDEV_RX_CHECKSUM_OFFLOAD
) != 0;
646 /* A device may report more queues than it makes available (this has
647 * been observed for Intel xl710, which reserves some of them for
648 * SRIOV): rte_eth_*_queue_setup will fail if a queue is not
649 * available. When this happens we can retry the configuration
650 * and request less queues */
651 while (n_rxq
&& n_txq
) {
653 VLOG_INFO("Retrying setup with (rxq:%d txq:%d)", n_rxq
, n_txq
);
656 diag
= rte_eth_dev_configure(dev
->port_id
, n_rxq
, n_txq
, &conf
);
658 VLOG_WARN("Interface %s eth_dev setup error %s\n",
659 dev
->up
.name
, rte_strerror(-diag
));
663 for (i
= 0; i
< n_txq
; i
++) {
664 diag
= rte_eth_tx_queue_setup(dev
->port_id
, i
, dev
->txq_size
,
665 dev
->socket_id
, NULL
);
667 VLOG_INFO("Interface %s txq(%d) setup error: %s",
668 dev
->up
.name
, i
, rte_strerror(-diag
));
674 /* Retry with less tx queues */
679 for (i
= 0; i
< n_rxq
; i
++) {
680 diag
= rte_eth_rx_queue_setup(dev
->port_id
, i
, dev
->rxq_size
,
681 dev
->socket_id
, NULL
,
684 VLOG_INFO("Interface %s rxq(%d) setup error: %s",
685 dev
->up
.name
, i
, rte_strerror(-diag
));
691 /* Retry with less rx queues */
696 dev
->up
.n_rxq
= n_rxq
;
697 dev
->up
.n_txq
= n_txq
;
706 dpdk_eth_checksum_offload_configure(struct netdev_dpdk
*dev
)
707 OVS_REQUIRES(dev
->mutex
)
709 struct rte_eth_dev_info info
;
710 bool rx_csum_ol_flag
= false;
711 uint32_t rx_chksm_offload_capa
= DEV_RX_OFFLOAD_UDP_CKSUM
|
712 DEV_RX_OFFLOAD_TCP_CKSUM
|
713 DEV_RX_OFFLOAD_IPV4_CKSUM
;
714 rte_eth_dev_info_get(dev
->port_id
, &info
);
715 rx_csum_ol_flag
= (dev
->hw_ol_features
& NETDEV_RX_CHECKSUM_OFFLOAD
) != 0;
717 if (rx_csum_ol_flag
&&
718 (info
.rx_offload_capa
& rx_chksm_offload_capa
) !=
719 rx_chksm_offload_capa
) {
720 VLOG_WARN_ONCE("Rx checksum offload is not supported on device %d",
722 dev
->hw_ol_features
&= ~NETDEV_RX_CHECKSUM_OFFLOAD
;
725 netdev_request_reconfigure(&dev
->up
);
729 dpdk_eth_flow_ctrl_setup(struct netdev_dpdk
*dev
) OVS_REQUIRES(dev
->mutex
)
731 if (rte_eth_dev_flow_ctrl_set(dev
->port_id
, &dev
->fc_conf
)) {
732 VLOG_WARN("Failed to enable flow control on device %d", dev
->port_id
);
737 dpdk_eth_dev_init(struct netdev_dpdk
*dev
)
738 OVS_REQUIRES(dev
->mutex
)
740 struct rte_pktmbuf_pool_private
*mbp_priv
;
741 struct rte_eth_dev_info info
;
742 struct ether_addr eth_addr
;
746 if (!rte_eth_dev_is_valid_port(dev
->port_id
)) {
750 rte_eth_dev_info_get(dev
->port_id
, &info
);
752 n_rxq
= MIN(info
.max_rx_queues
, dev
->up
.n_rxq
);
753 n_txq
= MIN(info
.max_tx_queues
, dev
->up
.n_txq
);
755 diag
= dpdk_eth_dev_queue_setup(dev
, n_rxq
, n_txq
);
757 VLOG_ERR("Interface %s(rxq:%d txq:%d) configure error: %s",
758 dev
->up
.name
, n_rxq
, n_txq
, rte_strerror(-diag
));
762 diag
= rte_eth_dev_start(dev
->port_id
);
764 VLOG_ERR("Interface %s start error: %s", dev
->up
.name
,
765 rte_strerror(-diag
));
769 rte_eth_promiscuous_enable(dev
->port_id
);
770 rte_eth_allmulticast_enable(dev
->port_id
);
772 memset(ð_addr
, 0x0, sizeof(eth_addr
));
773 rte_eth_macaddr_get(dev
->port_id
, ð_addr
);
774 VLOG_INFO_RL(&rl
, "Port %d: "ETH_ADDR_FMT
"",
775 dev
->port_id
, ETH_ADDR_BYTES_ARGS(eth_addr
.addr_bytes
));
777 memcpy(dev
->hwaddr
.ea
, eth_addr
.addr_bytes
, ETH_ADDR_LEN
);
778 rte_eth_link_get_nowait(dev
->port_id
, &dev
->link
);
780 mbp_priv
= rte_mempool_get_priv(dev
->dpdk_mp
->mp
);
781 dev
->buf_size
= mbp_priv
->mbuf_data_room_size
- RTE_PKTMBUF_HEADROOM
;
783 dev
->flags
= NETDEV_UP
| NETDEV_PROMISC
;
785 /* Get the Flow control configuration for DPDK-ETH */
786 diag
= rte_eth_dev_flow_ctrl_get(dev
->port_id
, &dev
->fc_conf
);
788 VLOG_DBG("cannot get flow control parameters on port=%d, err=%d",
795 static struct netdev_dpdk
*
796 netdev_dpdk_cast(const struct netdev
*netdev
)
798 return CONTAINER_OF(netdev
, struct netdev_dpdk
, up
);
801 static struct netdev
*
802 netdev_dpdk_alloc(void)
804 struct netdev_dpdk
*dev
;
806 dev
= dpdk_rte_mzalloc(sizeof *dev
);
814 static struct dpdk_tx_queue
*
815 netdev_dpdk_alloc_txq(unsigned int n_txqs
)
817 struct dpdk_tx_queue
*txqs
;
820 txqs
= dpdk_rte_mzalloc(n_txqs
* sizeof *txqs
);
822 for (i
= 0; i
< n_txqs
; i
++) {
823 /* Initialize map for vhost devices. */
824 txqs
[i
].map
= OVS_VHOST_QUEUE_MAP_UNKNOWN
;
825 rte_spinlock_init(&txqs
[i
].tx_lock
);
833 netdev_dpdk_init(struct netdev
*netdev
, unsigned int port_no
,
834 enum dpdk_dev_type type
)
835 OVS_REQUIRES(dpdk_mutex
)
837 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
841 ovs_mutex_init(&dev
->mutex
);
842 ovs_mutex_lock(&dev
->mutex
);
844 rte_spinlock_init(&dev
->stats_lock
);
846 /* If the 'sid' is negative, it means that the kernel fails
847 * to obtain the pci numa info. In that situation, always
849 if (type
== DPDK_DEV_ETH
) {
850 sid
= rte_eth_dev_socket_id(port_no
);
852 sid
= rte_lcore_to_socket_id(rte_get_master_lcore());
855 dev
->socket_id
= sid
< 0 ? SOCKET0
: sid
;
856 dev
->requested_socket_id
= dev
->socket_id
;
857 dev
->port_id
= port_no
;
860 dev
->requested_mtu
= dev
->mtu
= ETHER_MTU
;
861 dev
->max_packet_len
= MTU_TO_FRAME_LEN(dev
->mtu
);
862 ovsrcu_index_init(&dev
->vid
, -1);
863 dev
->vhost_reconfigured
= false;
865 err
= netdev_dpdk_mempool_configure(dev
);
870 ovsrcu_init(&dev
->qos_conf
, NULL
);
872 ovsrcu_init(&dev
->ingress_policer
, NULL
);
873 dev
->policer_rate
= 0;
874 dev
->policer_burst
= 0;
876 netdev
->n_rxq
= NR_QUEUE
;
877 netdev
->n_txq
= NR_QUEUE
;
878 dev
->requested_n_rxq
= netdev
->n_rxq
;
879 dev
->requested_n_txq
= netdev
->n_txq
;
880 dev
->rxq_size
= NIC_PORT_DEFAULT_RXQ_SIZE
;
881 dev
->txq_size
= NIC_PORT_DEFAULT_TXQ_SIZE
;
882 dev
->requested_rxq_size
= dev
->rxq_size
;
883 dev
->requested_txq_size
= dev
->txq_size
;
885 /* Initialize the flow control to NULL */
886 memset(&dev
->fc_conf
, 0, sizeof dev
->fc_conf
);
888 /* Initilize the hardware offload flags to 0 */
889 dev
->hw_ol_features
= 0;
890 if (type
== DPDK_DEV_ETH
) {
891 err
= dpdk_eth_dev_init(dev
);
895 dev
->tx_q
= netdev_dpdk_alloc_txq(netdev
->n_txq
);
897 dev
->tx_q
= netdev_dpdk_alloc_txq(OVS_VHOST_MAX_QUEUE_NUM
);
898 /* Enable DPDK_DEV_VHOST device and set promiscuous mode flag. */
899 dev
->flags
= NETDEV_UP
| NETDEV_PROMISC
;
907 ovs_list_push_back(&dpdk_list
, &dev
->list_node
);
910 ovs_mutex_unlock(&dev
->mutex
);
914 /* dev_name must be the prefix followed by a positive decimal number.
915 * (no leading + or - signs are allowed) */
917 dpdk_dev_parse_name(const char dev_name
[], const char prefix
[],
918 unsigned int *port_no
)
922 if (strncmp(dev_name
, prefix
, strlen(prefix
))) {
926 cport
= dev_name
+ strlen(prefix
);
928 if (str_to_uint(cport
, 10, port_no
)) {
936 netdev_dpdk_vhost_construct(struct netdev
*netdev
)
938 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
939 const char *name
= netdev
->name
;
942 /* 'name' is appended to 'vhost_sock_dir' and used to create a socket in
943 * the file system. '/' or '\' would traverse directories, so they're not
944 * acceptable in 'name'. */
945 if (strchr(name
, '/') || strchr(name
, '\\')) {
946 VLOG_ERR("\"%s\" is not a valid name for a vhost-user port. "
947 "A valid name must not include '/' or '\\'",
952 ovs_mutex_lock(&dpdk_mutex
);
953 /* Take the name of the vhost-user port and append it to the location where
954 * the socket is to be created, then register the socket.
956 snprintf(dev
->vhost_id
, sizeof dev
->vhost_id
, "%s/%s",
957 dpdk_get_vhost_sock_dir(), name
);
959 dev
->vhost_driver_flags
&= ~RTE_VHOST_USER_CLIENT
;
960 err
= rte_vhost_driver_register(dev
->vhost_id
, dev
->vhost_driver_flags
);
962 VLOG_ERR("vhost-user socket device setup failure for socket %s\n",
965 fatal_signal_add_file_to_unlink(dev
->vhost_id
);
966 VLOG_INFO("Socket %s created for vhost-user port %s\n",
967 dev
->vhost_id
, name
);
969 err
= netdev_dpdk_init(netdev
, -1, DPDK_DEV_VHOST
);
971 ovs_mutex_unlock(&dpdk_mutex
);
976 netdev_dpdk_vhost_client_construct(struct netdev
*netdev
)
980 ovs_mutex_lock(&dpdk_mutex
);
981 err
= netdev_dpdk_init(netdev
, -1, DPDK_DEV_VHOST
);
982 ovs_mutex_unlock(&dpdk_mutex
);
987 netdev_dpdk_construct(struct netdev
*netdev
)
989 unsigned int port_no
;
992 /* Names always start with "dpdk" */
993 err
= dpdk_dev_parse_name(netdev
->name
, "dpdk", &port_no
);
998 ovs_mutex_lock(&dpdk_mutex
);
999 err
= netdev_dpdk_init(netdev
, port_no
, DPDK_DEV_ETH
);
1000 ovs_mutex_unlock(&dpdk_mutex
);
1005 netdev_dpdk_destruct(struct netdev
*netdev
)
1007 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1009 ovs_mutex_lock(&dpdk_mutex
);
1010 ovs_mutex_lock(&dev
->mutex
);
1012 rte_eth_dev_stop(dev
->port_id
);
1013 free(ovsrcu_get_protected(struct ingress_policer
*,
1014 &dev
->ingress_policer
));
1016 rte_free(dev
->tx_q
);
1017 ovs_list_remove(&dev
->list_node
);
1018 dpdk_mp_put(dev
->dpdk_mp
);
1020 ovs_mutex_unlock(&dev
->mutex
);
1021 ovs_mutex_unlock(&dpdk_mutex
);
1024 /* rte_vhost_driver_unregister() can call back destroy_device(), which will
1025 * try to acquire 'dpdk_mutex' and possibly 'dev->mutex'. To avoid a
1026 * deadlock, none of the mutexes must be held while calling this function. */
1028 dpdk_vhost_driver_unregister(struct netdev_dpdk
*dev OVS_UNUSED
,
1030 OVS_EXCLUDED(dpdk_mutex
)
1031 OVS_EXCLUDED(dev
->mutex
)
1033 return rte_vhost_driver_unregister(vhost_id
);
1037 netdev_dpdk_vhost_destruct(struct netdev
*netdev
)
1039 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1042 ovs_mutex_lock(&dpdk_mutex
);
1043 ovs_mutex_lock(&dev
->mutex
);
1045 /* Guest becomes an orphan if still attached. */
1046 if (netdev_dpdk_get_vid(dev
) >= 0
1047 && !(dev
->vhost_driver_flags
& RTE_VHOST_USER_CLIENT
)) {
1048 VLOG_ERR("Removing port '%s' while vhost device still attached.",
1050 VLOG_ERR("To restore connectivity after re-adding of port, VM on "
1051 "socket '%s' must be restarted.", dev
->vhost_id
);
1054 free(ovsrcu_get_protected(struct ingress_policer
*,
1055 &dev
->ingress_policer
));
1057 rte_free(dev
->tx_q
);
1058 ovs_list_remove(&dev
->list_node
);
1059 dpdk_mp_put(dev
->dpdk_mp
);
1061 vhost_id
= xstrdup(dev
->vhost_id
);
1063 ovs_mutex_unlock(&dev
->mutex
);
1064 ovs_mutex_unlock(&dpdk_mutex
);
1070 if (dpdk_vhost_driver_unregister(dev
, vhost_id
)) {
1071 VLOG_ERR("%s: Unable to unregister vhost driver for socket '%s'.\n",
1072 netdev
->name
, vhost_id
);
1073 } else if (!(dev
->vhost_driver_flags
& RTE_VHOST_USER_CLIENT
)) {
1074 /* OVS server mode - remove this socket from list for deletion */
1075 fatal_signal_remove_file_to_unlink(vhost_id
);
1082 netdev_dpdk_dealloc(struct netdev
*netdev
)
1084 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1090 netdev_dpdk_get_config(const struct netdev
*netdev
, struct smap
*args
)
1092 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1094 ovs_mutex_lock(&dev
->mutex
);
1096 smap_add_format(args
, "requested_rx_queues", "%d", dev
->requested_n_rxq
);
1097 smap_add_format(args
, "configured_rx_queues", "%d", netdev
->n_rxq
);
1098 smap_add_format(args
, "requested_tx_queues", "%d", dev
->requested_n_txq
);
1099 smap_add_format(args
, "configured_tx_queues", "%d", netdev
->n_txq
);
1100 smap_add_format(args
, "mtu", "%d", dev
->mtu
);
1102 if (dev
->type
== DPDK_DEV_ETH
) {
1103 smap_add_format(args
, "requested_rxq_descriptors", "%d",
1104 dev
->requested_rxq_size
);
1105 smap_add_format(args
, "configured_rxq_descriptors", "%d",
1107 smap_add_format(args
, "requested_txq_descriptors", "%d",
1108 dev
->requested_txq_size
);
1109 smap_add_format(args
, "configured_txq_descriptors", "%d",
1111 if (dev
->hw_ol_features
& NETDEV_RX_CHECKSUM_OFFLOAD
) {
1112 smap_add(args
, "rx_csum_offload", "true");
1115 ovs_mutex_unlock(&dev
->mutex
);
1121 dpdk_set_rxq_config(struct netdev_dpdk
*dev
, const struct smap
*args
)
1122 OVS_REQUIRES(dev
->mutex
)
1126 new_n_rxq
= MAX(smap_get_int(args
, "n_rxq", NR_QUEUE
), 1);
1127 if (new_n_rxq
!= dev
->requested_n_rxq
) {
1128 dev
->requested_n_rxq
= new_n_rxq
;
1129 netdev_request_reconfigure(&dev
->up
);
1134 dpdk_process_queue_size(struct netdev
*netdev
, const struct smap
*args
,
1135 const char *flag
, int default_size
, int *new_size
)
1137 int queue_size
= smap_get_int(args
, flag
, default_size
);
1139 if (queue_size
<= 0 || queue_size
> NIC_PORT_MAX_Q_SIZE
1140 || !is_pow2(queue_size
)) {
1141 queue_size
= default_size
;
1144 if (queue_size
!= *new_size
) {
1145 *new_size
= queue_size
;
1146 netdev_request_reconfigure(netdev
);
1151 netdev_dpdk_set_config(struct netdev
*netdev
, const struct smap
*args
)
1153 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1154 bool rx_fc_en
, tx_fc_en
, autoneg
;
1155 enum rte_eth_fc_mode fc_mode
;
1156 static const enum rte_eth_fc_mode fc_mode_set
[2][2] = {
1157 {RTE_FC_NONE
, RTE_FC_TX_PAUSE
},
1158 {RTE_FC_RX_PAUSE
, RTE_FC_FULL
}
1163 ovs_mutex_lock(&dev
->mutex
);
1165 dpdk_set_rxq_config(dev
, args
);
1167 dpdk_process_queue_size(netdev
, args
, "n_rxq_desc",
1168 NIC_PORT_DEFAULT_RXQ_SIZE
,
1169 &dev
->requested_rxq_size
);
1170 dpdk_process_queue_size(netdev
, args
, "n_txq_desc",
1171 NIC_PORT_DEFAULT_TXQ_SIZE
,
1172 &dev
->requested_txq_size
);
1174 rx_fc_en
= smap_get_bool(args
, "rx-flow-ctrl", false);
1175 tx_fc_en
= smap_get_bool(args
, "tx-flow-ctrl", false);
1176 autoneg
= smap_get_bool(args
, "flow-ctrl-autoneg", false);
1178 fc_mode
= fc_mode_set
[tx_fc_en
][rx_fc_en
];
1179 if (dev
->fc_conf
.mode
!= fc_mode
|| autoneg
!= dev
->fc_conf
.autoneg
) {
1180 dev
->fc_conf
.mode
= fc_mode
;
1181 dev
->fc_conf
.autoneg
= autoneg
;
1182 dpdk_eth_flow_ctrl_setup(dev
);
1185 /* Rx checksum offload configuration */
1186 /* By default the Rx checksum offload is ON */
1187 rx_chksm_ofld
= smap_get_bool(args
, "rx-checksum-offload", true);
1188 temp_flag
= (dev
->hw_ol_features
& NETDEV_RX_CHECKSUM_OFFLOAD
)
1190 if (temp_flag
!= rx_chksm_ofld
) {
1191 dev
->hw_ol_features
^= NETDEV_RX_CHECKSUM_OFFLOAD
;
1192 dpdk_eth_checksum_offload_configure(dev
);
1194 ovs_mutex_unlock(&dev
->mutex
);
1200 netdev_dpdk_ring_set_config(struct netdev
*netdev
, const struct smap
*args
)
1202 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1204 ovs_mutex_lock(&dev
->mutex
);
1205 dpdk_set_rxq_config(dev
, args
);
1206 ovs_mutex_unlock(&dev
->mutex
);
1212 netdev_dpdk_vhost_client_set_config(struct netdev
*netdev
,
1213 const struct smap
*args
)
1215 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1218 ovs_mutex_lock(&dev
->mutex
);
1219 if (!(dev
->vhost_driver_flags
& RTE_VHOST_USER_CLIENT
)) {
1220 path
= smap_get(args
, "vhost-server-path");
1221 if (path
&& strcmp(path
, dev
->vhost_id
)) {
1222 strcpy(dev
->vhost_id
, path
);
1223 netdev_request_reconfigure(netdev
);
1226 ovs_mutex_unlock(&dev
->mutex
);
1232 netdev_dpdk_get_numa_id(const struct netdev
*netdev
)
1234 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1236 return dev
->socket_id
;
1239 /* Sets the number of tx queues for the dpdk interface. */
1241 netdev_dpdk_set_tx_multiq(struct netdev
*netdev
, unsigned int n_txq
)
1243 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1245 ovs_mutex_lock(&dev
->mutex
);
1247 if (dev
->requested_n_txq
== n_txq
) {
1251 dev
->requested_n_txq
= n_txq
;
1252 netdev_request_reconfigure(netdev
);
1255 ovs_mutex_unlock(&dev
->mutex
);
1259 static struct netdev_rxq
*
1260 netdev_dpdk_rxq_alloc(void)
1262 struct netdev_rxq_dpdk
*rx
= dpdk_rte_mzalloc(sizeof *rx
);
1271 static struct netdev_rxq_dpdk
*
1272 netdev_rxq_dpdk_cast(const struct netdev_rxq
*rxq
)
1274 return CONTAINER_OF(rxq
, struct netdev_rxq_dpdk
, up
);
1278 netdev_dpdk_rxq_construct(struct netdev_rxq
*rxq
)
1280 struct netdev_rxq_dpdk
*rx
= netdev_rxq_dpdk_cast(rxq
);
1281 struct netdev_dpdk
*dev
= netdev_dpdk_cast(rxq
->netdev
);
1283 ovs_mutex_lock(&dev
->mutex
);
1284 rx
->port_id
= dev
->port_id
;
1285 ovs_mutex_unlock(&dev
->mutex
);
1291 netdev_dpdk_rxq_destruct(struct netdev_rxq
*rxq OVS_UNUSED
)
1296 netdev_dpdk_rxq_dealloc(struct netdev_rxq
*rxq
)
1298 struct netdev_rxq_dpdk
*rx
= netdev_rxq_dpdk_cast(rxq
);
1303 /* Tries to transmit 'pkts' to txq 'qid' of device 'dev'. Takes ownership of
1304 * 'pkts', even in case of failure.
1306 * Returns the number of packets that weren't transmitted. */
1308 netdev_dpdk_eth_tx_burst(struct netdev_dpdk
*dev
, int qid
,
1309 struct rte_mbuf
**pkts
, int cnt
)
1313 while (nb_tx
!= cnt
) {
1316 ret
= rte_eth_tx_burst(dev
->port_id
, qid
, pkts
+ nb_tx
, cnt
- nb_tx
);
1324 if (OVS_UNLIKELY(nb_tx
!= cnt
)) {
1325 /* Free buffers, which we couldn't transmit, one at a time (each
1326 * packet could come from a different mempool) */
1329 for (i
= nb_tx
; i
< cnt
; i
++) {
1330 rte_pktmbuf_free(pkts
[i
]);
1338 netdev_dpdk_policer_pkt_handle(struct rte_meter_srtcm
*meter
,
1339 struct rte_mbuf
*pkt
, uint64_t time
)
1341 uint32_t pkt_len
= rte_pktmbuf_pkt_len(pkt
) - sizeof(struct ether_hdr
);
1343 return rte_meter_srtcm_color_blind_check(meter
, time
, pkt_len
) ==
1348 netdev_dpdk_policer_run(struct rte_meter_srtcm
*meter
,
1349 struct rte_mbuf
**pkts
, int pkt_cnt
)
1353 struct rte_mbuf
*pkt
= NULL
;
1354 uint64_t current_time
= rte_rdtsc();
1356 for (i
= 0; i
< pkt_cnt
; i
++) {
1358 /* Handle current packet */
1359 if (netdev_dpdk_policer_pkt_handle(meter
, pkt
, current_time
)) {
1365 rte_pktmbuf_free(pkt
);
1373 ingress_policer_run(struct ingress_policer
*policer
, struct rte_mbuf
**pkts
,
1378 rte_spinlock_lock(&policer
->policer_lock
);
1379 cnt
= netdev_dpdk_policer_run(&policer
->in_policer
, pkts
, pkt_cnt
);
1380 rte_spinlock_unlock(&policer
->policer_lock
);
1386 is_vhost_running(struct netdev_dpdk
*dev
)
1388 return (netdev_dpdk_get_vid(dev
) >= 0 && dev
->vhost_reconfigured
);
1392 netdev_dpdk_vhost_update_rx_size_counters(struct netdev_stats
*stats
,
1393 unsigned int packet_size
)
1395 /* Hard-coded search for the size bucket. */
1396 if (packet_size
< 256) {
1397 if (packet_size
>= 128) {
1398 stats
->rx_128_to_255_packets
++;
1399 } else if (packet_size
<= 64) {
1400 stats
->rx_1_to_64_packets
++;
1402 stats
->rx_65_to_127_packets
++;
1405 if (packet_size
>= 1523) {
1406 stats
->rx_1523_to_max_packets
++;
1407 } else if (packet_size
>= 1024) {
1408 stats
->rx_1024_to_1522_packets
++;
1409 } else if (packet_size
< 512) {
1410 stats
->rx_256_to_511_packets
++;
1412 stats
->rx_512_to_1023_packets
++;
1418 netdev_dpdk_vhost_update_rx_counters(struct netdev_stats
*stats
,
1419 struct dp_packet
**packets
, int count
,
1423 unsigned int packet_size
;
1424 struct dp_packet
*packet
;
1426 stats
->rx_packets
+= count
;
1427 stats
->rx_dropped
+= dropped
;
1428 for (i
= 0; i
< count
; i
++) {
1429 packet
= packets
[i
];
1430 packet_size
= dp_packet_size(packet
);
1432 if (OVS_UNLIKELY(packet_size
< ETH_HEADER_LEN
)) {
1433 /* This only protects the following multicast counting from
1434 * too short packets, but it does not stop the packet from
1435 * further processing. */
1437 stats
->rx_length_errors
++;
1441 netdev_dpdk_vhost_update_rx_size_counters(stats
, packet_size
);
1443 struct eth_header
*eh
= (struct eth_header
*) dp_packet_data(packet
);
1444 if (OVS_UNLIKELY(eth_addr_is_multicast(eh
->eth_dst
))) {
1448 stats
->rx_bytes
+= packet_size
;
1453 * The receive path for the vhost port is the TX path out from guest.
1456 netdev_dpdk_vhost_rxq_recv(struct netdev_rxq
*rxq
,
1457 struct dp_packet_batch
*batch
)
1459 struct netdev_dpdk
*dev
= netdev_dpdk_cast(rxq
->netdev
);
1460 int qid
= rxq
->queue_id
;
1461 struct ingress_policer
*policer
= netdev_dpdk_get_ingress_policer(dev
);
1463 uint16_t dropped
= 0;
1465 if (OVS_UNLIKELY(!is_vhost_running(dev
)
1466 || !(dev
->flags
& NETDEV_UP
))) {
1470 nb_rx
= rte_vhost_dequeue_burst(netdev_dpdk_get_vid(dev
),
1471 qid
* VIRTIO_QNUM
+ VIRTIO_TXQ
,
1473 (struct rte_mbuf
**) batch
->packets
,
1481 nb_rx
= ingress_policer_run(policer
,
1482 (struct rte_mbuf
**) batch
->packets
,
1487 rte_spinlock_lock(&dev
->stats_lock
);
1488 netdev_dpdk_vhost_update_rx_counters(&dev
->stats
, batch
->packets
,
1490 rte_spinlock_unlock(&dev
->stats_lock
);
1492 batch
->count
= (int) nb_rx
;
1497 netdev_dpdk_rxq_recv(struct netdev_rxq
*rxq
, struct dp_packet_batch
*batch
)
1499 struct netdev_rxq_dpdk
*rx
= netdev_rxq_dpdk_cast(rxq
);
1500 struct netdev_dpdk
*dev
= netdev_dpdk_cast(rxq
->netdev
);
1501 struct ingress_policer
*policer
= netdev_dpdk_get_ingress_policer(dev
);
1505 nb_rx
= rte_eth_rx_burst(rx
->port_id
, rxq
->queue_id
,
1506 (struct rte_mbuf
**) batch
->packets
,
1514 nb_rx
= ingress_policer_run(policer
,
1515 (struct rte_mbuf
**) batch
->packets
,
1520 /* Update stats to reflect dropped packets */
1521 if (OVS_UNLIKELY(dropped
)) {
1522 rte_spinlock_lock(&dev
->stats_lock
);
1523 dev
->stats
.rx_dropped
+= dropped
;
1524 rte_spinlock_unlock(&dev
->stats_lock
);
1527 batch
->count
= nb_rx
;
1533 netdev_dpdk_qos_run(struct netdev_dpdk
*dev
, struct rte_mbuf
**pkts
,
1536 struct qos_conf
*qos_conf
= ovsrcu_get(struct qos_conf
*, &dev
->qos_conf
);
1539 rte_spinlock_lock(&qos_conf
->lock
);
1540 cnt
= qos_conf
->ops
->qos_run(qos_conf
, pkts
, cnt
);
1541 rte_spinlock_unlock(&qos_conf
->lock
);
1548 netdev_dpdk_filter_packet_len(struct netdev_dpdk
*dev
, struct rte_mbuf
**pkts
,
1553 struct rte_mbuf
*pkt
;
1555 for (i
= 0; i
< pkt_cnt
; i
++) {
1557 if (OVS_UNLIKELY(pkt
->pkt_len
> dev
->max_packet_len
)) {
1558 VLOG_WARN_RL(&rl
, "%s: Too big size %" PRIu32
" max_packet_len %d",
1559 dev
->up
.name
, pkt
->pkt_len
, dev
->max_packet_len
);
1560 rte_pktmbuf_free(pkt
);
1564 if (OVS_UNLIKELY(i
!= cnt
)) {
1574 netdev_dpdk_vhost_update_tx_counters(struct netdev_stats
*stats
,
1575 struct dp_packet
**packets
,
1580 int sent
= attempted
- dropped
;
1582 stats
->tx_packets
+= sent
;
1583 stats
->tx_dropped
+= dropped
;
1585 for (i
= 0; i
< sent
; i
++) {
1586 stats
->tx_bytes
+= dp_packet_size(packets
[i
]);
1591 __netdev_dpdk_vhost_send(struct netdev
*netdev
, int qid
,
1592 struct dp_packet
**pkts
, int cnt
)
1594 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1595 struct rte_mbuf
**cur_pkts
= (struct rte_mbuf
**) pkts
;
1596 unsigned int total_pkts
= cnt
;
1597 unsigned int dropped
= 0;
1600 qid
= dev
->tx_q
[qid
% netdev
->n_txq
].map
;
1602 if (OVS_UNLIKELY(!is_vhost_running(dev
) || qid
< 0
1603 || !(dev
->flags
& NETDEV_UP
))) {
1604 rte_spinlock_lock(&dev
->stats_lock
);
1605 dev
->stats
.tx_dropped
+= cnt
;
1606 rte_spinlock_unlock(&dev
->stats_lock
);
1610 rte_spinlock_lock(&dev
->tx_q
[qid
].tx_lock
);
1612 cnt
= netdev_dpdk_filter_packet_len(dev
, cur_pkts
, cnt
);
1613 /* Check has QoS has been configured for the netdev */
1614 cnt
= netdev_dpdk_qos_run(dev
, cur_pkts
, cnt
);
1615 dropped
= total_pkts
- cnt
;
1618 int vhost_qid
= qid
* VIRTIO_QNUM
+ VIRTIO_RXQ
;
1619 unsigned int tx_pkts
;
1621 tx_pkts
= rte_vhost_enqueue_burst(netdev_dpdk_get_vid(dev
),
1622 vhost_qid
, cur_pkts
, cnt
);
1623 if (OVS_LIKELY(tx_pkts
)) {
1624 /* Packets have been sent.*/
1626 /* Prepare for possible retry.*/
1627 cur_pkts
= &cur_pkts
[tx_pkts
];
1629 /* No packets sent - do not retry.*/
1632 } while (cnt
&& (retries
++ <= VHOST_ENQ_RETRY_NUM
));
1634 rte_spinlock_unlock(&dev
->tx_q
[qid
].tx_lock
);
1636 rte_spinlock_lock(&dev
->stats_lock
);
1637 netdev_dpdk_vhost_update_tx_counters(&dev
->stats
, pkts
, total_pkts
,
1639 rte_spinlock_unlock(&dev
->stats_lock
);
1642 for (i
= 0; i
< total_pkts
- dropped
; i
++) {
1643 dp_packet_delete(pkts
[i
]);
1647 /* Tx function. Transmit packets indefinitely */
1649 dpdk_do_tx_copy(struct netdev
*netdev
, int qid
, struct dp_packet_batch
*batch
)
1650 OVS_NO_THREAD_SAFETY_ANALYSIS
1652 #if !defined(__CHECKER__) && !defined(_WIN32)
1653 const size_t PKT_ARRAY_SIZE
= batch
->count
;
1655 /* Sparse or MSVC doesn't like variable length array. */
1656 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
1658 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1659 struct rte_mbuf
*pkts
[PKT_ARRAY_SIZE
];
1664 dp_packet_batch_apply_cutlen(batch
);
1666 for (i
= 0; i
< batch
->count
; i
++) {
1667 int size
= dp_packet_size(batch
->packets
[i
]);
1669 if (OVS_UNLIKELY(size
> dev
->max_packet_len
)) {
1670 VLOG_WARN_RL(&rl
, "Too big size %d max_packet_len %d",
1671 (int) size
, dev
->max_packet_len
);
1677 pkts
[newcnt
] = rte_pktmbuf_alloc(dev
->dpdk_mp
->mp
);
1679 if (!pkts
[newcnt
]) {
1680 dropped
+= batch
->count
- i
;
1684 /* We have to do a copy for now */
1685 memcpy(rte_pktmbuf_mtod(pkts
[newcnt
], void *),
1686 dp_packet_data(batch
->packets
[i
]), size
);
1688 rte_pktmbuf_data_len(pkts
[newcnt
]) = size
;
1689 rte_pktmbuf_pkt_len(pkts
[newcnt
]) = size
;
1694 if (dev
->type
== DPDK_DEV_VHOST
) {
1695 __netdev_dpdk_vhost_send(netdev
, qid
, (struct dp_packet
**) pkts
,
1698 unsigned int qos_pkts
= newcnt
;
1700 /* Check if QoS has been configured for this netdev. */
1701 newcnt
= netdev_dpdk_qos_run(dev
, pkts
, newcnt
);
1703 dropped
+= qos_pkts
- newcnt
;
1704 dropped
+= netdev_dpdk_eth_tx_burst(dev
, qid
, pkts
, newcnt
);
1707 if (OVS_UNLIKELY(dropped
)) {
1708 rte_spinlock_lock(&dev
->stats_lock
);
1709 dev
->stats
.tx_dropped
+= dropped
;
1710 rte_spinlock_unlock(&dev
->stats_lock
);
1715 netdev_dpdk_vhost_send(struct netdev
*netdev
, int qid
,
1716 struct dp_packet_batch
*batch
,
1717 bool may_steal
, bool concurrent_txq OVS_UNUSED
)
1720 if (OVS_UNLIKELY(!may_steal
|| batch
->packets
[0]->source
!= DPBUF_DPDK
)) {
1721 dpdk_do_tx_copy(netdev
, qid
, batch
);
1722 dp_packet_delete_batch(batch
, may_steal
);
1724 dp_packet_batch_apply_cutlen(batch
);
1725 __netdev_dpdk_vhost_send(netdev
, qid
, batch
->packets
, batch
->count
);
1731 netdev_dpdk_send__(struct netdev_dpdk
*dev
, int qid
,
1732 struct dp_packet_batch
*batch
, bool may_steal
,
1733 bool concurrent_txq
)
1735 if (OVS_UNLIKELY(concurrent_txq
)) {
1736 qid
= qid
% dev
->up
.n_txq
;
1737 rte_spinlock_lock(&dev
->tx_q
[qid
].tx_lock
);
1740 if (OVS_UNLIKELY(!may_steal
||
1741 batch
->packets
[0]->source
!= DPBUF_DPDK
)) {
1742 struct netdev
*netdev
= &dev
->up
;
1744 dpdk_do_tx_copy(netdev
, qid
, batch
);
1745 dp_packet_delete_batch(batch
, may_steal
);
1748 int cnt
= batch
->count
;
1749 struct rte_mbuf
**pkts
= (struct rte_mbuf
**) batch
->packets
;
1751 dp_packet_batch_apply_cutlen(batch
);
1753 cnt
= netdev_dpdk_filter_packet_len(dev
, pkts
, cnt
);
1754 cnt
= netdev_dpdk_qos_run(dev
, pkts
, cnt
);
1755 dropped
= batch
->count
- cnt
;
1757 dropped
+= netdev_dpdk_eth_tx_burst(dev
, qid
, pkts
, cnt
);
1759 if (OVS_UNLIKELY(dropped
)) {
1760 rte_spinlock_lock(&dev
->stats_lock
);
1761 dev
->stats
.tx_dropped
+= dropped
;
1762 rte_spinlock_unlock(&dev
->stats_lock
);
1766 if (OVS_UNLIKELY(concurrent_txq
)) {
1767 rte_spinlock_unlock(&dev
->tx_q
[qid
].tx_lock
);
1772 netdev_dpdk_eth_send(struct netdev
*netdev
, int qid
,
1773 struct dp_packet_batch
*batch
, bool may_steal
,
1774 bool concurrent_txq
)
1776 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1778 netdev_dpdk_send__(dev
, qid
, batch
, may_steal
, concurrent_txq
);
1783 netdev_dpdk_set_etheraddr(struct netdev
*netdev
, const struct eth_addr mac
)
1785 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1787 ovs_mutex_lock(&dev
->mutex
);
1788 if (!eth_addr_equals(dev
->hwaddr
, mac
)) {
1790 netdev_change_seq_changed(netdev
);
1792 ovs_mutex_unlock(&dev
->mutex
);
1798 netdev_dpdk_get_etheraddr(const struct netdev
*netdev
, struct eth_addr
*mac
)
1800 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1802 ovs_mutex_lock(&dev
->mutex
);
1804 ovs_mutex_unlock(&dev
->mutex
);
1810 netdev_dpdk_get_mtu(const struct netdev
*netdev
, int *mtup
)
1812 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1814 ovs_mutex_lock(&dev
->mutex
);
1816 ovs_mutex_unlock(&dev
->mutex
);
1822 netdev_dpdk_set_mtu(struct netdev
*netdev
, int mtu
)
1824 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1826 if (MTU_TO_FRAME_LEN(mtu
) > NETDEV_DPDK_MAX_PKT_LEN
1827 || mtu
< ETHER_MIN_MTU
) {
1828 VLOG_WARN("%s: unsupported MTU %d\n", dev
->up
.name
, mtu
);
1832 ovs_mutex_lock(&dev
->mutex
);
1833 if (dev
->requested_mtu
!= mtu
) {
1834 dev
->requested_mtu
= mtu
;
1835 netdev_request_reconfigure(netdev
);
1837 ovs_mutex_unlock(&dev
->mutex
);
1843 netdev_dpdk_get_carrier(const struct netdev
*netdev
, bool *carrier
);
1846 netdev_dpdk_vhost_get_stats(const struct netdev
*netdev
,
1847 struct netdev_stats
*stats
)
1849 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1851 ovs_mutex_lock(&dev
->mutex
);
1853 rte_spinlock_lock(&dev
->stats_lock
);
1854 /* Supported Stats */
1855 stats
->rx_packets
+= dev
->stats
.rx_packets
;
1856 stats
->tx_packets
+= dev
->stats
.tx_packets
;
1857 stats
->rx_dropped
= dev
->stats
.rx_dropped
;
1858 stats
->tx_dropped
+= dev
->stats
.tx_dropped
;
1859 stats
->multicast
= dev
->stats
.multicast
;
1860 stats
->rx_bytes
= dev
->stats
.rx_bytes
;
1861 stats
->tx_bytes
= dev
->stats
.tx_bytes
;
1862 stats
->rx_errors
= dev
->stats
.rx_errors
;
1863 stats
->rx_length_errors
= dev
->stats
.rx_length_errors
;
1865 stats
->rx_1_to_64_packets
= dev
->stats
.rx_1_to_64_packets
;
1866 stats
->rx_65_to_127_packets
= dev
->stats
.rx_65_to_127_packets
;
1867 stats
->rx_128_to_255_packets
= dev
->stats
.rx_128_to_255_packets
;
1868 stats
->rx_256_to_511_packets
= dev
->stats
.rx_256_to_511_packets
;
1869 stats
->rx_512_to_1023_packets
= dev
->stats
.rx_512_to_1023_packets
;
1870 stats
->rx_1024_to_1522_packets
= dev
->stats
.rx_1024_to_1522_packets
;
1871 stats
->rx_1523_to_max_packets
= dev
->stats
.rx_1523_to_max_packets
;
1873 rte_spinlock_unlock(&dev
->stats_lock
);
1875 ovs_mutex_unlock(&dev
->mutex
);
1881 netdev_dpdk_convert_xstats(struct netdev_stats
*stats
,
1882 const struct rte_eth_xstat
*xstats
,
1883 const struct rte_eth_xstat_name
*names
,
1884 const unsigned int size
)
1886 for (unsigned int i
= 0; i
< size
; i
++) {
1887 if (strcmp(XSTAT_RX_64_PACKETS
, names
[i
].name
) == 0) {
1888 stats
->rx_1_to_64_packets
= xstats
[i
].value
;
1889 } else if (strcmp(XSTAT_RX_65_TO_127_PACKETS
, names
[i
].name
) == 0) {
1890 stats
->rx_65_to_127_packets
= xstats
[i
].value
;
1891 } else if (strcmp(XSTAT_RX_128_TO_255_PACKETS
, names
[i
].name
) == 0) {
1892 stats
->rx_128_to_255_packets
= xstats
[i
].value
;
1893 } else if (strcmp(XSTAT_RX_256_TO_511_PACKETS
, names
[i
].name
) == 0) {
1894 stats
->rx_256_to_511_packets
= xstats
[i
].value
;
1895 } else if (strcmp(XSTAT_RX_512_TO_1023_PACKETS
, names
[i
].name
) == 0) {
1896 stats
->rx_512_to_1023_packets
= xstats
[i
].value
;
1897 } else if (strcmp(XSTAT_RX_1024_TO_1522_PACKETS
, names
[i
].name
) == 0) {
1898 stats
->rx_1024_to_1522_packets
= xstats
[i
].value
;
1899 } else if (strcmp(XSTAT_RX_1523_TO_MAX_PACKETS
, names
[i
].name
) == 0) {
1900 stats
->rx_1523_to_max_packets
= xstats
[i
].value
;
1901 } else if (strcmp(XSTAT_TX_64_PACKETS
, names
[i
].name
) == 0) {
1902 stats
->tx_1_to_64_packets
= xstats
[i
].value
;
1903 } else if (strcmp(XSTAT_TX_65_TO_127_PACKETS
, names
[i
].name
) == 0) {
1904 stats
->tx_65_to_127_packets
= xstats
[i
].value
;
1905 } else if (strcmp(XSTAT_TX_128_TO_255_PACKETS
, names
[i
].name
) == 0) {
1906 stats
->tx_128_to_255_packets
= xstats
[i
].value
;
1907 } else if (strcmp(XSTAT_TX_256_TO_511_PACKETS
, names
[i
].name
) == 0) {
1908 stats
->tx_256_to_511_packets
= xstats
[i
].value
;
1909 } else if (strcmp(XSTAT_TX_512_TO_1023_PACKETS
, names
[i
].name
) == 0) {
1910 stats
->tx_512_to_1023_packets
= xstats
[i
].value
;
1911 } else if (strcmp(XSTAT_TX_1024_TO_1522_PACKETS
, names
[i
].name
) == 0) {
1912 stats
->tx_1024_to_1522_packets
= xstats
[i
].value
;
1913 } else if (strcmp(XSTAT_TX_1523_TO_MAX_PACKETS
, names
[i
].name
) == 0) {
1914 stats
->tx_1523_to_max_packets
= xstats
[i
].value
;
1915 } else if (strcmp(XSTAT_TX_MULTICAST_PACKETS
, names
[i
].name
) == 0) {
1916 stats
->tx_multicast_packets
= xstats
[i
].value
;
1917 } else if (strcmp(XSTAT_RX_BROADCAST_PACKETS
, names
[i
].name
) == 0) {
1918 stats
->rx_broadcast_packets
= xstats
[i
].value
;
1919 } else if (strcmp(XSTAT_TX_BROADCAST_PACKETS
, names
[i
].name
) == 0) {
1920 stats
->tx_broadcast_packets
= xstats
[i
].value
;
1921 } else if (strcmp(XSTAT_RX_UNDERSIZED_ERRORS
, names
[i
].name
) == 0) {
1922 stats
->rx_undersized_errors
= xstats
[i
].value
;
1923 } else if (strcmp(XSTAT_RX_FRAGMENTED_ERRORS
, names
[i
].name
) == 0) {
1924 stats
->rx_fragmented_errors
= xstats
[i
].value
;
1925 } else if (strcmp(XSTAT_RX_JABBER_ERRORS
, names
[i
].name
) == 0) {
1926 stats
->rx_jabber_errors
= xstats
[i
].value
;
1932 netdev_dpdk_get_stats(const struct netdev
*netdev
, struct netdev_stats
*stats
)
1934 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1935 struct rte_eth_stats rte_stats
;
1938 netdev_dpdk_get_carrier(netdev
, &gg
);
1939 ovs_mutex_lock(&dev
->mutex
);
1941 struct rte_eth_xstat
*rte_xstats
= NULL
;
1942 struct rte_eth_xstat_name
*rte_xstats_names
= NULL
;
1943 int rte_xstats_len
, rte_xstats_new_len
, rte_xstats_ret
;
1945 if (rte_eth_stats_get(dev
->port_id
, &rte_stats
)) {
1946 VLOG_ERR("Can't get ETH statistics for port: %i.", dev
->port_id
);
1947 ovs_mutex_unlock(&dev
->mutex
);
1951 /* Get length of statistics */
1952 rte_xstats_len
= rte_eth_xstats_get_names(dev
->port_id
, NULL
, 0);
1953 if (rte_xstats_len
< 0) {
1954 VLOG_WARN("Cannot get XSTATS values for port: %i", dev
->port_id
);
1957 /* Reserve memory for xstats names and values */
1958 rte_xstats_names
= xcalloc(rte_xstats_len
, sizeof *rte_xstats_names
);
1959 rte_xstats
= xcalloc(rte_xstats_len
, sizeof *rte_xstats
);
1961 /* Retreive xstats names */
1962 rte_xstats_new_len
= rte_eth_xstats_get_names(dev
->port_id
,
1965 if (rte_xstats_new_len
!= rte_xstats_len
) {
1966 VLOG_WARN("Cannot get XSTATS names for port: %i.", dev
->port_id
);
1969 /* Retreive xstats values */
1970 memset(rte_xstats
, 0xff, sizeof *rte_xstats
* rte_xstats_len
);
1971 rte_xstats_ret
= rte_eth_xstats_get(dev
->port_id
, rte_xstats
,
1973 if (rte_xstats_ret
> 0 && rte_xstats_ret
<= rte_xstats_len
) {
1974 netdev_dpdk_convert_xstats(stats
, rte_xstats
, rte_xstats_names
,
1977 VLOG_WARN("Cannot get XSTATS values for port: %i.", dev
->port_id
);
1982 free(rte_xstats_names
);
1984 stats
->rx_packets
= rte_stats
.ipackets
;
1985 stats
->tx_packets
= rte_stats
.opackets
;
1986 stats
->rx_bytes
= rte_stats
.ibytes
;
1987 stats
->tx_bytes
= rte_stats
.obytes
;
1988 /* DPDK counts imissed as errors, but count them here as dropped instead */
1989 stats
->rx_errors
= rte_stats
.ierrors
- rte_stats
.imissed
;
1990 stats
->tx_errors
= rte_stats
.oerrors
;
1992 rte_spinlock_lock(&dev
->stats_lock
);
1993 stats
->tx_dropped
= dev
->stats
.tx_dropped
;
1994 stats
->rx_dropped
= dev
->stats
.rx_dropped
;
1995 rte_spinlock_unlock(&dev
->stats_lock
);
1997 /* These are the available DPDK counters for packets not received due to
1998 * local resource constraints in DPDK and NIC respectively. */
1999 stats
->rx_dropped
+= rte_stats
.rx_nombuf
+ rte_stats
.imissed
;
2000 stats
->rx_missed_errors
= rte_stats
.imissed
;
2002 ovs_mutex_unlock(&dev
->mutex
);
2008 netdev_dpdk_get_features(const struct netdev
*netdev
,
2009 enum netdev_features
*current
,
2010 enum netdev_features
*advertised
,
2011 enum netdev_features
*supported
,
2012 enum netdev_features
*peer
)
2014 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2015 struct rte_eth_link link
;
2017 ovs_mutex_lock(&dev
->mutex
);
2019 ovs_mutex_unlock(&dev
->mutex
);
2021 if (link
.link_duplex
== ETH_LINK_HALF_DUPLEX
) {
2022 if (link
.link_speed
== ETH_SPEED_NUM_10M
) {
2023 *current
= NETDEV_F_10MB_HD
;
2025 if (link
.link_speed
== ETH_SPEED_NUM_100M
) {
2026 *current
= NETDEV_F_100MB_HD
;
2028 if (link
.link_speed
== ETH_SPEED_NUM_1G
) {
2029 *current
= NETDEV_F_1GB_HD
;
2031 } else if (link
.link_duplex
== ETH_LINK_FULL_DUPLEX
) {
2032 if (link
.link_speed
== ETH_SPEED_NUM_10M
) {
2033 *current
= NETDEV_F_10MB_FD
;
2035 if (link
.link_speed
== ETH_SPEED_NUM_100M
) {
2036 *current
= NETDEV_F_100MB_FD
;
2038 if (link
.link_speed
== ETH_SPEED_NUM_1G
) {
2039 *current
= NETDEV_F_1GB_FD
;
2041 if (link
.link_speed
== ETH_SPEED_NUM_10G
) {
2042 *current
= NETDEV_F_10GB_FD
;
2046 if (link
.link_autoneg
) {
2047 *current
|= NETDEV_F_AUTONEG
;
2050 *advertised
= *supported
= *peer
= 0;
2055 static struct ingress_policer
*
2056 netdev_dpdk_policer_construct(uint32_t rate
, uint32_t burst
)
2058 struct ingress_policer
*policer
= NULL
;
2059 uint64_t rate_bytes
;
2060 uint64_t burst_bytes
;
2063 policer
= xmalloc(sizeof *policer
);
2064 rte_spinlock_init(&policer
->policer_lock
);
2066 /* rte_meter requires bytes so convert kbits rate and burst to bytes. */
2067 rate_bytes
= rate
* 1000/8;
2068 burst_bytes
= burst
* 1000/8;
2070 policer
->app_srtcm_params
.cir
= rate_bytes
;
2071 policer
->app_srtcm_params
.cbs
= burst_bytes
;
2072 policer
->app_srtcm_params
.ebs
= 0;
2073 err
= rte_meter_srtcm_config(&policer
->in_policer
,
2074 &policer
->app_srtcm_params
);
2076 VLOG_ERR("Could not create rte meter for ingress policer");
2084 netdev_dpdk_set_policing(struct netdev
* netdev
, uint32_t policer_rate
,
2085 uint32_t policer_burst
)
2087 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2088 struct ingress_policer
*policer
;
2090 /* Force to 0 if no rate specified,
2091 * default to 8000 kbits if burst is 0,
2092 * else stick with user-specified value.
2094 policer_burst
= (!policer_rate
? 0
2095 : !policer_burst
? 8000
2098 ovs_mutex_lock(&dev
->mutex
);
2100 policer
= ovsrcu_get_protected(struct ingress_policer
*,
2101 &dev
->ingress_policer
);
2103 if (dev
->policer_rate
== policer_rate
&&
2104 dev
->policer_burst
== policer_burst
) {
2105 /* Assume that settings haven't changed since we last set them. */
2106 ovs_mutex_unlock(&dev
->mutex
);
2110 /* Destroy any existing ingress policer for the device if one exists */
2112 ovsrcu_postpone(free
, policer
);
2115 if (policer_rate
!= 0) {
2116 policer
= netdev_dpdk_policer_construct(policer_rate
, policer_burst
);
2120 ovsrcu_set(&dev
->ingress_policer
, policer
);
2121 dev
->policer_rate
= policer_rate
;
2122 dev
->policer_burst
= policer_burst
;
2123 ovs_mutex_unlock(&dev
->mutex
);
2129 netdev_dpdk_get_ifindex(const struct netdev
*netdev
)
2131 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2134 ovs_mutex_lock(&dev
->mutex
);
2135 ifindex
= dev
->port_id
;
2136 ovs_mutex_unlock(&dev
->mutex
);
2142 netdev_dpdk_get_carrier(const struct netdev
*netdev
, bool *carrier
)
2144 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2146 ovs_mutex_lock(&dev
->mutex
);
2147 check_link_status(dev
);
2148 *carrier
= dev
->link
.link_status
;
2150 ovs_mutex_unlock(&dev
->mutex
);
2156 netdev_dpdk_vhost_get_carrier(const struct netdev
*netdev
, bool *carrier
)
2158 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2160 ovs_mutex_lock(&dev
->mutex
);
2162 if (is_vhost_running(dev
)) {
2168 ovs_mutex_unlock(&dev
->mutex
);
2173 static long long int
2174 netdev_dpdk_get_carrier_resets(const struct netdev
*netdev
)
2176 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2177 long long int carrier_resets
;
2179 ovs_mutex_lock(&dev
->mutex
);
2180 carrier_resets
= dev
->link_reset_cnt
;
2181 ovs_mutex_unlock(&dev
->mutex
);
2183 return carrier_resets
;
2187 netdev_dpdk_set_miimon(struct netdev
*netdev OVS_UNUSED
,
2188 long long int interval OVS_UNUSED
)
2194 netdev_dpdk_update_flags__(struct netdev_dpdk
*dev
,
2195 enum netdev_flags off
, enum netdev_flags on
,
2196 enum netdev_flags
*old_flagsp
)
2197 OVS_REQUIRES(dev
->mutex
)
2201 if ((off
| on
) & ~(NETDEV_UP
| NETDEV_PROMISC
)) {
2205 *old_flagsp
= dev
->flags
;
2209 if (dev
->flags
== *old_flagsp
) {
2213 if (dev
->type
== DPDK_DEV_ETH
) {
2214 if (dev
->flags
& NETDEV_UP
) {
2215 err
= rte_eth_dev_start(dev
->port_id
);
2220 if (dev
->flags
& NETDEV_PROMISC
) {
2221 rte_eth_promiscuous_enable(dev
->port_id
);
2224 if (!(dev
->flags
& NETDEV_UP
)) {
2225 rte_eth_dev_stop(dev
->port_id
);
2228 netdev_change_seq_changed(&dev
->up
);
2230 /* If DPDK_DEV_VHOST device's NETDEV_UP flag was changed and vhost is
2231 * running then change netdev's change_seq to trigger link state
2234 if ((NETDEV_UP
& ((*old_flagsp
^ on
) | (*old_flagsp
^ off
)))
2235 && is_vhost_running(dev
)) {
2236 netdev_change_seq_changed(&dev
->up
);
2238 /* Clear statistics if device is getting up. */
2239 if (NETDEV_UP
& on
) {
2240 rte_spinlock_lock(&dev
->stats_lock
);
2241 memset(&dev
->stats
, 0, sizeof dev
->stats
);
2242 rte_spinlock_unlock(&dev
->stats_lock
);
2251 netdev_dpdk_update_flags(struct netdev
*netdev
,
2252 enum netdev_flags off
, enum netdev_flags on
,
2253 enum netdev_flags
*old_flagsp
)
2255 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2258 ovs_mutex_lock(&dev
->mutex
);
2259 error
= netdev_dpdk_update_flags__(dev
, off
, on
, old_flagsp
);
2260 ovs_mutex_unlock(&dev
->mutex
);
2266 netdev_dpdk_get_status(const struct netdev
*netdev
, struct smap
*args
)
2268 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2269 struct rte_eth_dev_info dev_info
;
2271 if (!rte_eth_dev_is_valid_port(dev
->port_id
)) {
2275 ovs_mutex_lock(&dev
->mutex
);
2276 rte_eth_dev_info_get(dev
->port_id
, &dev_info
);
2277 ovs_mutex_unlock(&dev
->mutex
);
2279 smap_add_format(args
, "port_no", "%d", dev
->port_id
);
2280 smap_add_format(args
, "numa_id", "%d",
2281 rte_eth_dev_socket_id(dev
->port_id
));
2282 smap_add_format(args
, "driver_name", "%s", dev_info
.driver_name
);
2283 smap_add_format(args
, "min_rx_bufsize", "%u", dev_info
.min_rx_bufsize
);
2284 smap_add_format(args
, "max_rx_pktlen", "%u", dev
->max_packet_len
);
2285 smap_add_format(args
, "max_rx_queues", "%u", dev_info
.max_rx_queues
);
2286 smap_add_format(args
, "max_tx_queues", "%u", dev_info
.max_tx_queues
);
2287 smap_add_format(args
, "max_mac_addrs", "%u", dev_info
.max_mac_addrs
);
2288 smap_add_format(args
, "max_hash_mac_addrs", "%u",
2289 dev_info
.max_hash_mac_addrs
);
2290 smap_add_format(args
, "max_vfs", "%u", dev_info
.max_vfs
);
2291 smap_add_format(args
, "max_vmdq_pools", "%u", dev_info
.max_vmdq_pools
);
2293 if (dev_info
.pci_dev
) {
2294 smap_add_format(args
, "pci-vendor_id", "0x%u",
2295 dev_info
.pci_dev
->id
.vendor_id
);
2296 smap_add_format(args
, "pci-device_id", "0x%x",
2297 dev_info
.pci_dev
->id
.device_id
);
2304 netdev_dpdk_set_admin_state__(struct netdev_dpdk
*dev
, bool admin_state
)
2305 OVS_REQUIRES(dev
->mutex
)
2307 enum netdev_flags old_flags
;
2310 netdev_dpdk_update_flags__(dev
, 0, NETDEV_UP
, &old_flags
);
2312 netdev_dpdk_update_flags__(dev
, NETDEV_UP
, 0, &old_flags
);
2317 netdev_dpdk_set_admin_state(struct unixctl_conn
*conn
, int argc
,
2318 const char *argv
[], void *aux OVS_UNUSED
)
2322 if (!strcasecmp(argv
[argc
- 1], "up")) {
2324 } else if ( !strcasecmp(argv
[argc
- 1], "down")) {
2327 unixctl_command_reply_error(conn
, "Invalid Admin State");
2332 struct netdev
*netdev
= netdev_from_name(argv
[1]);
2333 if (netdev
&& is_dpdk_class(netdev
->netdev_class
)) {
2334 struct netdev_dpdk
*dpdk_dev
= netdev_dpdk_cast(netdev
);
2336 ovs_mutex_lock(&dpdk_dev
->mutex
);
2337 netdev_dpdk_set_admin_state__(dpdk_dev
, up
);
2338 ovs_mutex_unlock(&dpdk_dev
->mutex
);
2340 netdev_close(netdev
);
2342 unixctl_command_reply_error(conn
, "Not a DPDK Interface");
2343 netdev_close(netdev
);
2347 struct netdev_dpdk
*netdev
;
2349 ovs_mutex_lock(&dpdk_mutex
);
2350 LIST_FOR_EACH (netdev
, list_node
, &dpdk_list
) {
2351 ovs_mutex_lock(&netdev
->mutex
);
2352 netdev_dpdk_set_admin_state__(netdev
, up
);
2353 ovs_mutex_unlock(&netdev
->mutex
);
2355 ovs_mutex_unlock(&dpdk_mutex
);
2357 unixctl_command_reply(conn
, "OK");
2361 netdev_dpdk_attach(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2362 const char *argv
[], void *aux OVS_UNUSED
)
2368 ovs_mutex_lock(&dpdk_mutex
);
2370 ret
= rte_eth_dev_attach(argv
[1], &port_id
);
2372 response
= xasprintf("Error attaching device '%s'", argv
[1]);
2373 ovs_mutex_unlock(&dpdk_mutex
);
2374 unixctl_command_reply_error(conn
, response
);
2379 response
= xasprintf("Device '%s' has been attached as 'dpdk%d'",
2382 ovs_mutex_unlock(&dpdk_mutex
);
2383 unixctl_command_reply(conn
, response
);
2388 netdev_dpdk_detach(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2389 const char *argv
[], void *aux OVS_UNUSED
)
2393 unsigned int parsed_port
;
2395 char devname
[RTE_ETH_NAME_MAX_LEN
];
2397 ovs_mutex_lock(&dpdk_mutex
);
2399 ret
= dpdk_dev_parse_name(argv
[1], "dpdk", &parsed_port
);
2401 response
= xasprintf("'%s' is not a valid port", argv
[1]);
2405 port_id
= parsed_port
;
2407 struct netdev
*netdev
= netdev_from_name(argv
[1]);
2409 netdev_close(netdev
);
2410 response
= xasprintf("Port '%s' is being used. Remove it before"
2411 "detaching", argv
[1]);
2415 rte_eth_dev_close(port_id
);
2417 ret
= rte_eth_dev_detach(port_id
, devname
);
2419 response
= xasprintf("Port '%s' can not be detached", argv
[1]);
2423 response
= xasprintf("Port '%s' has been detached", argv
[1]);
2425 ovs_mutex_unlock(&dpdk_mutex
);
2426 unixctl_command_reply(conn
, response
);
2431 ovs_mutex_unlock(&dpdk_mutex
);
2432 unixctl_command_reply_error(conn
, response
);
2437 * Set virtqueue flags so that we do not receive interrupts.
2440 set_irq_status(int vid
)
2445 for (i
= 0; i
< rte_vhost_get_queue_num(vid
); i
++) {
2446 idx
= i
* VIRTIO_QNUM
;
2447 rte_vhost_enable_guest_notification(vid
, idx
+ VIRTIO_RXQ
, 0);
2448 rte_vhost_enable_guest_notification(vid
, idx
+ VIRTIO_TXQ
, 0);
2453 * Fixes mapping for vhost-user tx queues. Must be called after each
2454 * enabling/disabling of queues and n_txq modifications.
2457 netdev_dpdk_remap_txqs(struct netdev_dpdk
*dev
)
2458 OVS_REQUIRES(dev
->mutex
)
2460 int *enabled_queues
, n_enabled
= 0;
2461 int i
, k
, total_txqs
= dev
->up
.n_txq
;
2463 enabled_queues
= xcalloc(total_txqs
, sizeof *enabled_queues
);
2465 for (i
= 0; i
< total_txqs
; i
++) {
2466 /* Enabled queues always mapped to themselves. */
2467 if (dev
->tx_q
[i
].map
== i
) {
2468 enabled_queues
[n_enabled
++] = i
;
2472 if (n_enabled
== 0 && total_txqs
!= 0) {
2473 enabled_queues
[0] = OVS_VHOST_QUEUE_DISABLED
;
2478 for (i
= 0; i
< total_txqs
; i
++) {
2479 if (dev
->tx_q
[i
].map
!= i
) {
2480 dev
->tx_q
[i
].map
= enabled_queues
[k
];
2481 k
= (k
+ 1) % n_enabled
;
2485 VLOG_DBG("TX queue mapping for %s\n", dev
->vhost_id
);
2486 for (i
= 0; i
< total_txqs
; i
++) {
2487 VLOG_DBG("%2d --> %2d", i
, dev
->tx_q
[i
].map
);
2490 free(enabled_queues
);
2494 * A new virtio-net device is added to a vhost port.
2499 struct netdev_dpdk
*dev
;
2500 bool exists
= false;
2502 char ifname
[IF_NAME_SZ
];
2504 rte_vhost_get_ifname(vid
, ifname
, sizeof ifname
);
2506 ovs_mutex_lock(&dpdk_mutex
);
2507 /* Add device to the vhost port with the same name as that passed down. */
2508 LIST_FOR_EACH(dev
, list_node
, &dpdk_list
) {
2509 ovs_mutex_lock(&dev
->mutex
);
2510 if (strncmp(ifname
, dev
->vhost_id
, IF_NAME_SZ
) == 0) {
2511 uint32_t qp_num
= rte_vhost_get_queue_num(vid
);
2513 /* Get NUMA information */
2514 newnode
= rte_vhost_get_numa_node(vid
);
2515 if (newnode
== -1) {
2517 VLOG_INFO("Error getting NUMA info for vHost Device '%s'",
2520 newnode
= dev
->socket_id
;
2523 if (dev
->requested_n_txq
!= qp_num
2524 || dev
->requested_n_rxq
!= qp_num
2525 || dev
->requested_socket_id
!= newnode
) {
2526 dev
->requested_socket_id
= newnode
;
2527 dev
->requested_n_rxq
= qp_num
;
2528 dev
->requested_n_txq
= qp_num
;
2529 netdev_request_reconfigure(&dev
->up
);
2531 /* Reconfiguration not required. */
2532 dev
->vhost_reconfigured
= true;
2535 ovsrcu_index_set(&dev
->vid
, vid
);
2538 /* Disable notifications. */
2539 set_irq_status(vid
);
2540 netdev_change_seq_changed(&dev
->up
);
2541 ovs_mutex_unlock(&dev
->mutex
);
2544 ovs_mutex_unlock(&dev
->mutex
);
2546 ovs_mutex_unlock(&dpdk_mutex
);
2549 VLOG_INFO("vHost Device '%s' can't be added - name not found", ifname
);
2554 VLOG_INFO("vHost Device '%s' has been added on numa node %i",
2560 /* Clears mapping for all available queues of vhost interface. */
2562 netdev_dpdk_txq_map_clear(struct netdev_dpdk
*dev
)
2563 OVS_REQUIRES(dev
->mutex
)
2567 for (i
= 0; i
< dev
->up
.n_txq
; i
++) {
2568 dev
->tx_q
[i
].map
= OVS_VHOST_QUEUE_MAP_UNKNOWN
;
2573 * Remove a virtio-net device from the specific vhost port. Use dev->remove
2574 * flag to stop any more packets from being sent or received to/from a VM and
2575 * ensure all currently queued packets have been sent/received before removing
2579 destroy_device(int vid
)
2581 struct netdev_dpdk
*dev
;
2582 bool exists
= false;
2583 char ifname
[IF_NAME_SZ
];
2585 rte_vhost_get_ifname(vid
, ifname
, sizeof ifname
);
2587 ovs_mutex_lock(&dpdk_mutex
);
2588 LIST_FOR_EACH (dev
, list_node
, &dpdk_list
) {
2589 if (netdev_dpdk_get_vid(dev
) == vid
) {
2591 ovs_mutex_lock(&dev
->mutex
);
2592 dev
->vhost_reconfigured
= false;
2593 ovsrcu_index_set(&dev
->vid
, -1);
2594 netdev_dpdk_txq_map_clear(dev
);
2596 netdev_change_seq_changed(&dev
->up
);
2597 ovs_mutex_unlock(&dev
->mutex
);
2603 ovs_mutex_unlock(&dpdk_mutex
);
2607 * Wait for other threads to quiesce after setting the 'virtio_dev'
2608 * to NULL, before returning.
2610 ovsrcu_synchronize();
2612 * As call to ovsrcu_synchronize() will end the quiescent state,
2613 * put thread back into quiescent state before returning.
2615 ovsrcu_quiesce_start();
2616 VLOG_INFO("vHost Device '%s' has been removed", ifname
);
2618 VLOG_INFO("vHost Device '%s' not found", ifname
);
2623 vring_state_changed(int vid
, uint16_t queue_id
, int enable
)
2625 struct netdev_dpdk
*dev
;
2626 bool exists
= false;
2627 int qid
= queue_id
/ VIRTIO_QNUM
;
2628 char ifname
[IF_NAME_SZ
];
2630 rte_vhost_get_ifname(vid
, ifname
, sizeof ifname
);
2632 if (queue_id
% VIRTIO_QNUM
== VIRTIO_TXQ
) {
2636 ovs_mutex_lock(&dpdk_mutex
);
2637 LIST_FOR_EACH (dev
, list_node
, &dpdk_list
) {
2638 ovs_mutex_lock(&dev
->mutex
);
2639 if (strncmp(ifname
, dev
->vhost_id
, IF_NAME_SZ
) == 0) {
2641 dev
->tx_q
[qid
].map
= qid
;
2643 dev
->tx_q
[qid
].map
= OVS_VHOST_QUEUE_DISABLED
;
2645 netdev_dpdk_remap_txqs(dev
);
2647 ovs_mutex_unlock(&dev
->mutex
);
2650 ovs_mutex_unlock(&dev
->mutex
);
2652 ovs_mutex_unlock(&dpdk_mutex
);
2655 VLOG_INFO("State of queue %d ( tx_qid %d ) of vhost device '%s'"
2656 "changed to \'%s\'", queue_id
, qid
, ifname
,
2657 (enable
== 1) ? "enabled" : "disabled");
2659 VLOG_INFO("vHost Device '%s' not found", ifname
);
2667 netdev_dpdk_get_vid(const struct netdev_dpdk
*dev
)
2669 return ovsrcu_index_get(&dev
->vid
);
2672 struct ingress_policer
*
2673 netdev_dpdk_get_ingress_policer(const struct netdev_dpdk
*dev
)
2675 return ovsrcu_get(struct ingress_policer
*, &dev
->ingress_policer
);
2679 * These callbacks allow virtio-net devices to be added to vhost ports when
2680 * configuration has been fully complete.
2682 static const struct virtio_net_device_ops virtio_net_device_ops
=
2684 .new_device
= new_device
,
2685 .destroy_device
= destroy_device
,
2686 .vring_state_changed
= vring_state_changed
2690 start_vhost_loop(void *dummy OVS_UNUSED
)
2692 pthread_detach(pthread_self());
2693 /* Put the vhost thread into quiescent state. */
2694 ovsrcu_quiesce_start();
2695 rte_vhost_driver_session_start();
2700 netdev_dpdk_class_init(void)
2702 static struct ovsthread_once once
= OVSTHREAD_ONCE_INITIALIZER
;
2704 /* This function can be called for different classes. The initialization
2705 * needs to be done only once */
2706 if (ovsthread_once_start(&once
)) {
2707 ovs_thread_create("dpdk_watchdog", dpdk_watchdog
, NULL
);
2708 unixctl_command_register("netdev-dpdk/set-admin-state",
2709 "[netdev] up|down", 1, 2,
2710 netdev_dpdk_set_admin_state
, NULL
);
2711 unixctl_command_register("netdev-dpdk/attach",
2712 "pci address of device", 1, 1,
2713 netdev_dpdk_attach
, NULL
);
2714 unixctl_command_register("netdev-dpdk/detach",
2716 netdev_dpdk_detach
, NULL
);
2718 ovsthread_once_done(&once
);
2725 netdev_dpdk_vhost_class_init(void)
2727 static struct ovsthread_once once
= OVSTHREAD_ONCE_INITIALIZER
;
2729 /* This function can be called for different classes. The initialization
2730 * needs to be done only once */
2731 if (ovsthread_once_start(&once
)) {
2732 rte_vhost_driver_callback_register(&virtio_net_device_ops
);
2733 rte_vhost_feature_disable(1ULL << VIRTIO_NET_F_HOST_TSO4
2734 | 1ULL << VIRTIO_NET_F_HOST_TSO6
2735 | 1ULL << VIRTIO_NET_F_CSUM
2736 | 1ULL << VIRTIO_RING_F_INDIRECT_DESC
);
2737 ovs_thread_create("vhost_thread", start_vhost_loop
, NULL
);
2739 ovsthread_once_done(&once
);
2748 dpdk_ring_create(const char dev_name
[], unsigned int port_no
,
2749 unsigned int *eth_port_id
)
2751 struct dpdk_ring
*ring_pair
;
2755 ring_pair
= dpdk_rte_mzalloc(sizeof *ring_pair
);
2760 /* XXX: Add support for multiquque ring. */
2761 ring_name
= xasprintf("%s_tx", dev_name
);
2763 /* Create single producer tx ring, netdev does explicit locking. */
2764 ring_pair
->cring_tx
= rte_ring_create(ring_name
, DPDK_RING_SIZE
, SOCKET0
,
2767 if (ring_pair
->cring_tx
== NULL
) {
2768 rte_free(ring_pair
);
2772 ring_name
= xasprintf("%s_rx", dev_name
);
2774 /* Create single consumer rx ring, netdev does explicit locking. */
2775 ring_pair
->cring_rx
= rte_ring_create(ring_name
, DPDK_RING_SIZE
, SOCKET0
,
2778 if (ring_pair
->cring_rx
== NULL
) {
2779 rte_free(ring_pair
);
2783 port_id
= rte_eth_from_rings(dev_name
, &ring_pair
->cring_rx
, 1,
2784 &ring_pair
->cring_tx
, 1, SOCKET0
);
2787 rte_free(ring_pair
);
2791 ring_pair
->user_port_id
= port_no
;
2792 ring_pair
->eth_port_id
= port_id
;
2793 *eth_port_id
= port_id
;
2795 ovs_list_push_back(&dpdk_ring_list
, &ring_pair
->list_node
);
2801 dpdk_ring_open(const char dev_name
[], unsigned int *eth_port_id
)
2802 OVS_REQUIRES(dpdk_mutex
)
2804 struct dpdk_ring
*ring_pair
;
2805 unsigned int port_no
;
2808 /* Names always start with "dpdkr" */
2809 err
= dpdk_dev_parse_name(dev_name
, "dpdkr", &port_no
);
2814 /* Look through our list to find the device */
2815 LIST_FOR_EACH (ring_pair
, list_node
, &dpdk_ring_list
) {
2816 if (ring_pair
->user_port_id
== port_no
) {
2817 VLOG_INFO("Found dpdk ring device %s:", dev_name
);
2818 /* Really all that is needed */
2819 *eth_port_id
= ring_pair
->eth_port_id
;
2823 /* Need to create the device rings */
2824 return dpdk_ring_create(dev_name
, port_no
, eth_port_id
);
2828 netdev_dpdk_ring_send(struct netdev
*netdev
, int qid
,
2829 struct dp_packet_batch
*batch
, bool may_steal
,
2830 bool concurrent_txq
)
2832 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2835 /* When using 'dpdkr' and sending to a DPDK ring, we want to ensure that
2836 * the rss hash field is clear. This is because the same mbuf may be
2837 * modified by the consumer of the ring and return into the datapath
2838 * without recalculating the RSS hash. */
2839 for (i
= 0; i
< batch
->count
; i
++) {
2840 dp_packet_rss_invalidate(batch
->packets
[i
]);
2843 netdev_dpdk_send__(dev
, qid
, batch
, may_steal
, concurrent_txq
);
2848 netdev_dpdk_ring_construct(struct netdev
*netdev
)
2850 unsigned int port_no
= 0;
2853 ovs_mutex_lock(&dpdk_mutex
);
2855 err
= dpdk_ring_open(netdev
->name
, &port_no
);
2860 err
= netdev_dpdk_init(netdev
, port_no
, DPDK_DEV_ETH
);
2863 ovs_mutex_unlock(&dpdk_mutex
);
2870 * Initialize QoS configuration operations.
2873 qos_conf_init(struct qos_conf
*conf
, const struct dpdk_qos_ops
*ops
)
2876 rte_spinlock_init(&conf
->lock
);
2880 * Search existing QoS operations in qos_ops and compare each set of
2881 * operations qos_name to name. Return a dpdk_qos_ops pointer to a match,
2884 static const struct dpdk_qos_ops
*
2885 qos_lookup_name(const char *name
)
2887 const struct dpdk_qos_ops
*const *opsp
;
2889 for (opsp
= qos_confs
; *opsp
!= NULL
; opsp
++) {
2890 const struct dpdk_qos_ops
*ops
= *opsp
;
2891 if (!strcmp(name
, ops
->qos_name
)) {
2899 netdev_dpdk_get_qos_types(const struct netdev
*netdev OVS_UNUSED
,
2902 const struct dpdk_qos_ops
*const *opsp
;
2904 for (opsp
= qos_confs
; *opsp
!= NULL
; opsp
++) {
2905 const struct dpdk_qos_ops
*ops
= *opsp
;
2906 if (ops
->qos_construct
&& ops
->qos_name
[0] != '\0') {
2907 sset_add(types
, ops
->qos_name
);
2914 netdev_dpdk_get_qos(const struct netdev
*netdev
,
2915 const char **typep
, struct smap
*details
)
2917 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2918 struct qos_conf
*qos_conf
;
2921 ovs_mutex_lock(&dev
->mutex
);
2922 qos_conf
= ovsrcu_get_protected(struct qos_conf
*, &dev
->qos_conf
);
2924 *typep
= qos_conf
->ops
->qos_name
;
2925 error
= (qos_conf
->ops
->qos_get
2926 ? qos_conf
->ops
->qos_get(qos_conf
, details
): 0);
2928 /* No QoS configuration set, return an empty string */
2931 ovs_mutex_unlock(&dev
->mutex
);
2937 netdev_dpdk_set_qos(struct netdev
*netdev
, const char *type
,
2938 const struct smap
*details
)
2940 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2941 const struct dpdk_qos_ops
*new_ops
= NULL
;
2942 struct qos_conf
*qos_conf
, *new_qos_conf
= NULL
;
2945 ovs_mutex_lock(&dev
->mutex
);
2947 qos_conf
= ovsrcu_get_protected(struct qos_conf
*, &dev
->qos_conf
);
2949 new_ops
= qos_lookup_name(type
);
2951 if (!new_ops
|| !new_ops
->qos_construct
) {
2952 new_qos_conf
= NULL
;
2953 if (type
&& type
[0]) {
2956 } else if (qos_conf
&& qos_conf
->ops
== new_ops
2957 && qos_conf
->ops
->qos_is_equal(qos_conf
, details
)) {
2958 new_qos_conf
= qos_conf
;
2960 error
= new_ops
->qos_construct(details
, &new_qos_conf
);
2964 VLOG_ERR("Failed to set QoS type %s on port %s: %s",
2965 type
, netdev
->name
, rte_strerror(error
));
2968 if (new_qos_conf
!= qos_conf
) {
2969 ovsrcu_set(&dev
->qos_conf
, new_qos_conf
);
2971 ovsrcu_postpone(qos_conf
->ops
->qos_destruct
, qos_conf
);
2975 ovs_mutex_unlock(&dev
->mutex
);
2980 /* egress-policer details */
2982 struct egress_policer
{
2983 struct qos_conf qos_conf
;
2984 struct rte_meter_srtcm_params app_srtcm_params
;
2985 struct rte_meter_srtcm egress_meter
;
2989 egress_policer_details_to_param(const struct smap
*details
,
2990 struct rte_meter_srtcm_params
*params
)
2992 memset(params
, 0, sizeof *params
);
2993 params
->cir
= smap_get_ullong(details
, "cir", 0);
2994 params
->cbs
= smap_get_ullong(details
, "cbs", 0);
2999 egress_policer_qos_construct(const struct smap
*details
,
3000 struct qos_conf
**conf
)
3002 struct egress_policer
*policer
;
3005 policer
= xmalloc(sizeof *policer
);
3006 qos_conf_init(&policer
->qos_conf
, &egress_policer_ops
);
3007 egress_policer_details_to_param(details
, &policer
->app_srtcm_params
);
3008 err
= rte_meter_srtcm_config(&policer
->egress_meter
,
3009 &policer
->app_srtcm_params
);
3011 *conf
= &policer
->qos_conf
;
3022 egress_policer_qos_destruct(struct qos_conf
*conf
)
3024 struct egress_policer
*policer
= CONTAINER_OF(conf
, struct egress_policer
,
3030 egress_policer_qos_get(const struct qos_conf
*conf
, struct smap
*details
)
3032 struct egress_policer
*policer
=
3033 CONTAINER_OF(conf
, struct egress_policer
, qos_conf
);
3035 smap_add_format(details
, "cir", "%"PRIu64
, policer
->app_srtcm_params
.cir
);
3036 smap_add_format(details
, "cbs", "%"PRIu64
, policer
->app_srtcm_params
.cbs
);
3042 egress_policer_qos_is_equal(const struct qos_conf
*conf
, const struct smap
*details
)
3044 struct egress_policer
*policer
=
3045 CONTAINER_OF(conf
, struct egress_policer
, qos_conf
);
3046 struct rte_meter_srtcm_params params
;
3048 egress_policer_details_to_param(details
, ¶ms
);
3050 return !memcmp(¶ms
, &policer
->app_srtcm_params
, sizeof params
);
3054 egress_policer_run(struct qos_conf
*conf
, struct rte_mbuf
**pkts
, int pkt_cnt
)
3057 struct egress_policer
*policer
=
3058 CONTAINER_OF(conf
, struct egress_policer
, qos_conf
);
3060 cnt
= netdev_dpdk_policer_run(&policer
->egress_meter
, pkts
, pkt_cnt
);
3065 static const struct dpdk_qos_ops egress_policer_ops
= {
3066 "egress-policer", /* qos_name */
3067 egress_policer_qos_construct
,
3068 egress_policer_qos_destruct
,
3069 egress_policer_qos_get
,
3070 egress_policer_qos_is_equal
,
3075 netdev_dpdk_reconfigure(struct netdev
*netdev
)
3077 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
3080 ovs_mutex_lock(&dev
->mutex
);
3082 if (netdev
->n_txq
== dev
->requested_n_txq
3083 && netdev
->n_rxq
== dev
->requested_n_rxq
3084 && dev
->mtu
== dev
->requested_mtu
3085 && dev
->rxq_size
== dev
->requested_rxq_size
3086 && dev
->txq_size
== dev
->requested_txq_size
) {
3087 /* Reconfiguration is unnecessary */
3092 rte_eth_dev_stop(dev
->port_id
);
3094 if (dev
->mtu
!= dev
->requested_mtu
) {
3095 netdev_dpdk_mempool_configure(dev
);
3098 netdev
->n_txq
= dev
->requested_n_txq
;
3099 netdev
->n_rxq
= dev
->requested_n_rxq
;
3101 dev
->rxq_size
= dev
->requested_rxq_size
;
3102 dev
->txq_size
= dev
->requested_txq_size
;
3104 rte_free(dev
->tx_q
);
3105 err
= dpdk_eth_dev_init(dev
);
3106 dev
->tx_q
= netdev_dpdk_alloc_txq(netdev
->n_txq
);
3111 netdev_change_seq_changed(netdev
);
3114 ovs_mutex_unlock(&dev
->mutex
);
3119 dpdk_vhost_reconfigure_helper(struct netdev_dpdk
*dev
)
3120 OVS_REQUIRES(dev
->mutex
)
3122 dev
->up
.n_txq
= dev
->requested_n_txq
;
3123 dev
->up
.n_rxq
= dev
->requested_n_rxq
;
3125 /* Enable TX queue 0 by default if it wasn't disabled. */
3126 if (dev
->tx_q
[0].map
== OVS_VHOST_QUEUE_MAP_UNKNOWN
) {
3127 dev
->tx_q
[0].map
= 0;
3130 netdev_dpdk_remap_txqs(dev
);
3132 if (dev
->requested_socket_id
!= dev
->socket_id
3133 || dev
->requested_mtu
!= dev
->mtu
) {
3134 if (!netdev_dpdk_mempool_configure(dev
)) {
3135 netdev_change_seq_changed(&dev
->up
);
3139 if (netdev_dpdk_get_vid(dev
) >= 0) {
3140 dev
->vhost_reconfigured
= true;
3145 netdev_dpdk_vhost_reconfigure(struct netdev
*netdev
)
3147 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
3149 ovs_mutex_lock(&dev
->mutex
);
3150 dpdk_vhost_reconfigure_helper(dev
);
3151 ovs_mutex_unlock(&dev
->mutex
);
3156 netdev_dpdk_vhost_client_reconfigure(struct netdev
*netdev
)
3158 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
3161 ovs_mutex_lock(&dev
->mutex
);
3163 dpdk_vhost_reconfigure_helper(dev
);
3165 /* Configure vHost client mode if requested and if the following criteria
3167 * 1. Device hasn't been registered yet.
3168 * 2. A path has been specified.
3170 if (!(dev
->vhost_driver_flags
& RTE_VHOST_USER_CLIENT
)
3171 && strlen(dev
->vhost_id
)) {
3172 /* Register client-mode device */
3173 err
= rte_vhost_driver_register(dev
->vhost_id
,
3174 RTE_VHOST_USER_CLIENT
);
3176 VLOG_ERR("vhost-user device setup failure for device %s\n",
3179 /* Configuration successful */
3180 dev
->vhost_driver_flags
|= RTE_VHOST_USER_CLIENT
;
3181 VLOG_INFO("vHost User device '%s' created in 'client' mode, "
3182 "using client socket '%s'",
3183 dev
->up
.name
, dev
->vhost_id
);
3187 ovs_mutex_unlock(&dev
->mutex
);
3192 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, DESTRUCT, \
3193 SET_CONFIG, SET_TX_MULTIQ, SEND, \
3194 GET_CARRIER, GET_STATS, \
3195 GET_FEATURES, GET_STATUS, \
3196 RECONFIGURE, RXQ_RECV) \
3199 true, /* is_pmd */ \
3201 NULL, /* netdev_dpdk_run */ \
3202 NULL, /* netdev_dpdk_wait */ \
3204 netdev_dpdk_alloc, \
3207 netdev_dpdk_dealloc, \
3208 netdev_dpdk_get_config, \
3210 NULL, /* get_tunnel_config */ \
3211 NULL, /* build header */ \
3212 NULL, /* push header */ \
3213 NULL, /* pop header */ \
3214 netdev_dpdk_get_numa_id, /* get_numa_id */ \
3218 NULL, /* send_wait */ \
3220 netdev_dpdk_set_etheraddr, \
3221 netdev_dpdk_get_etheraddr, \
3222 netdev_dpdk_get_mtu, \
3223 netdev_dpdk_set_mtu, \
3224 netdev_dpdk_get_ifindex, \
3226 netdev_dpdk_get_carrier_resets, \
3227 netdev_dpdk_set_miimon, \
3230 NULL, /* set_advertisements */ \
3232 netdev_dpdk_set_policing, \
3233 netdev_dpdk_get_qos_types, \
3234 NULL, /* get_qos_capabilities */ \
3235 netdev_dpdk_get_qos, \
3236 netdev_dpdk_set_qos, \
3237 NULL, /* get_queue */ \
3238 NULL, /* set_queue */ \
3239 NULL, /* delete_queue */ \
3240 NULL, /* get_queue_stats */ \
3241 NULL, /* queue_dump_start */ \
3242 NULL, /* queue_dump_next */ \
3243 NULL, /* queue_dump_done */ \
3244 NULL, /* dump_queue_stats */ \
3246 NULL, /* set_in4 */ \
3247 NULL, /* get_addr_list */ \
3248 NULL, /* add_router */ \
3249 NULL, /* get_next_hop */ \
3251 NULL, /* arp_lookup */ \
3253 netdev_dpdk_update_flags, \
3256 netdev_dpdk_rxq_alloc, \
3257 netdev_dpdk_rxq_construct, \
3258 netdev_dpdk_rxq_destruct, \
3259 netdev_dpdk_rxq_dealloc, \
3261 NULL, /* rx_wait */ \
3262 NULL, /* rxq_drain */ \
3265 static const struct netdev_class dpdk_class
=
3268 netdev_dpdk_class_init
,
3269 netdev_dpdk_construct
,
3270 netdev_dpdk_destruct
,
3271 netdev_dpdk_set_config
,
3272 netdev_dpdk_set_tx_multiq
,
3273 netdev_dpdk_eth_send
,
3274 netdev_dpdk_get_carrier
,
3275 netdev_dpdk_get_stats
,
3276 netdev_dpdk_get_features
,
3277 netdev_dpdk_get_status
,
3278 netdev_dpdk_reconfigure
,
3279 netdev_dpdk_rxq_recv
);
3281 static const struct netdev_class dpdk_ring_class
=
3284 netdev_dpdk_class_init
,
3285 netdev_dpdk_ring_construct
,
3286 netdev_dpdk_destruct
,
3287 netdev_dpdk_ring_set_config
,
3288 netdev_dpdk_set_tx_multiq
,
3289 netdev_dpdk_ring_send
,
3290 netdev_dpdk_get_carrier
,
3291 netdev_dpdk_get_stats
,
3292 netdev_dpdk_get_features
,
3293 netdev_dpdk_get_status
,
3294 netdev_dpdk_reconfigure
,
3295 netdev_dpdk_rxq_recv
);
3297 static const struct netdev_class dpdk_vhost_class
=
3300 netdev_dpdk_vhost_class_init
,
3301 netdev_dpdk_vhost_construct
,
3302 netdev_dpdk_vhost_destruct
,
3305 netdev_dpdk_vhost_send
,
3306 netdev_dpdk_vhost_get_carrier
,
3307 netdev_dpdk_vhost_get_stats
,
3310 netdev_dpdk_vhost_reconfigure
,
3311 netdev_dpdk_vhost_rxq_recv
);
3312 static const struct netdev_class dpdk_vhost_client_class
=
3314 "dpdkvhostuserclient",
3315 netdev_dpdk_vhost_class_init
,
3316 netdev_dpdk_vhost_client_construct
,
3317 netdev_dpdk_vhost_destruct
,
3318 netdev_dpdk_vhost_client_set_config
,
3320 netdev_dpdk_vhost_send
,
3321 netdev_dpdk_vhost_get_carrier
,
3322 netdev_dpdk_vhost_get_stats
,
3325 netdev_dpdk_vhost_client_reconfigure
,
3326 netdev_dpdk_vhost_rxq_recv
);
3329 netdev_dpdk_register(void)
3331 netdev_register_provider(&dpdk_class
);
3332 netdev_register_provider(&dpdk_ring_class
);
3333 netdev_register_provider(&dpdk_vhost_class
);
3334 netdev_register_provider(&dpdk_vhost_client_class
);