2 * Copyright (c) 2014, 2015, 2016, 2017 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_RX_MULTICAST_PACKETS "rx_multicast_packets"
122 #define XSTAT_TX_MULTICAST_PACKETS "tx_multicast_packets"
123 #define XSTAT_RX_BROADCAST_PACKETS "rx_broadcast_packets"
124 #define XSTAT_TX_BROADCAST_PACKETS "tx_broadcast_packets"
125 #define XSTAT_RX_UNDERSIZED_ERRORS "rx_undersized_errors"
126 #define XSTAT_RX_OVERSIZE_ERRORS "rx_oversize_errors"
127 #define XSTAT_RX_FRAGMENTED_ERRORS "rx_fragmented_errors"
128 #define XSTAT_RX_JABBER_ERRORS "rx_jabber_errors"
132 /* Default size of Physical NIC RXQ */
133 #define NIC_PORT_DEFAULT_RXQ_SIZE 2048
134 /* Default size of Physical NIC TXQ */
135 #define NIC_PORT_DEFAULT_TXQ_SIZE 2048
136 /* Maximum size of Physical NIC Queues */
137 #define NIC_PORT_MAX_Q_SIZE 4096
139 #define OVS_VHOST_MAX_QUEUE_NUM 1024 /* Maximum number of vHost TX queues. */
140 #define OVS_VHOST_QUEUE_MAP_UNKNOWN (-1) /* Mapping not initialized. */
141 #define OVS_VHOST_QUEUE_DISABLED (-2) /* Queue was disabled by guest and not
142 * yet mapped to another queue. */
144 #define DPDK_ETH_PORT_ID_INVALID RTE_MAX_ETHPORTS
146 /* DPDK library uses uint8_t for port_id. */
147 typedef uint8_t dpdk_port_t
;
149 #define VHOST_ENQ_RETRY_NUM 8
150 #define IF_NAME_SZ (PATH_MAX > IFNAMSIZ ? PATH_MAX : IFNAMSIZ)
152 static const struct rte_eth_conf port_conf
= {
154 .mq_mode
= ETH_MQ_RX_RSS
,
156 .header_split
= 0, /* Header Split disabled */
157 .hw_ip_checksum
= 0, /* IP checksum offload disabled */
158 .hw_vlan_filter
= 0, /* VLAN filtering disabled */
159 .jumbo_frame
= 0, /* Jumbo Frame Support disabled */
165 .rss_hf
= ETH_RSS_IP
| ETH_RSS_UDP
| ETH_RSS_TCP
,
169 .mq_mode
= ETH_MQ_TX_NONE
,
173 enum { DPDK_RING_SIZE
= 256 };
174 BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE
));
175 enum { DRAIN_TSC
= 200000ULL };
182 /* Quality of Service */
184 /* An instance of a QoS configuration. Always associated with a particular
187 * Each QoS implementation subclasses this with whatever additional data it
191 const struct dpdk_qos_ops
*ops
;
195 /* A particular implementation of dpdk QoS operations.
197 * The functions below return 0 if successful or a positive errno value on
198 * failure, except where otherwise noted. All of them must be provided, except
199 * where otherwise noted.
201 struct dpdk_qos_ops
{
203 /* Name of the QoS type */
204 const char *qos_name
;
206 /* Called to construct a qos_conf object. The implementation should make
207 * the appropriate calls to configure QoS according to 'details'.
209 * The contents of 'details' should be documented as valid for 'ovs_name'
210 * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml
211 * (which is built as ovs-vswitchd.conf.db(8)).
213 * This function must return 0 if and only if it sets '*conf' to an
214 * initialized 'struct qos_conf'.
216 * For all QoS implementations it should always be non-null.
218 int (*qos_construct
)(const struct smap
*details
, struct qos_conf
**conf
);
220 /* Destroys the data structures allocated by the implementation as part of
223 * For all QoS implementations it should always be non-null.
225 void (*qos_destruct
)(struct qos_conf
*conf
);
227 /* Retrieves details of 'conf' configuration into 'details'.
229 * The contents of 'details' should be documented as valid for 'ovs_name'
230 * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml
231 * (which is built as ovs-vswitchd.conf.db(8)).
233 int (*qos_get
)(const struct qos_conf
*conf
, struct smap
*details
);
235 /* Returns true if 'conf' is already configured according to 'details'.
237 * The contents of 'details' should be documented as valid for 'ovs_name'
238 * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml
239 * (which is built as ovs-vswitchd.conf.db(8)).
241 * For all QoS implementations it should always be non-null.
243 bool (*qos_is_equal
)(const struct qos_conf
*conf
,
244 const struct smap
*details
);
246 /* Modify an array of rte_mbufs. The modification is specific to
247 * each qos implementation.
249 * The function should take and array of mbufs and an int representing
250 * the current number of mbufs present in the array.
252 * After the function has performed a qos modification to the array of
253 * mbufs it returns an int representing the number of mbufs now present in
254 * the array. This value is can then be passed to the port send function
255 * along with the modified array for transmission.
257 * For all QoS implementations it should always be non-null.
259 int (*qos_run
)(struct qos_conf
*qos_conf
, struct rte_mbuf
**pkts
,
263 /* dpdk_qos_ops for each type of user space QoS implementation */
264 static const struct dpdk_qos_ops egress_policer_ops
;
267 * Array of dpdk_qos_ops, contains pointer to all supported QoS
270 static const struct dpdk_qos_ops
*const qos_confs
[] = {
275 static struct ovs_mutex dpdk_mutex
= OVS_MUTEX_INITIALIZER
;
277 /* Contains all 'struct dpdk_dev's. */
278 static struct ovs_list dpdk_list
OVS_GUARDED_BY(dpdk_mutex
)
279 = OVS_LIST_INITIALIZER(&dpdk_list
);
281 static struct ovs_mutex dpdk_mp_mutex
OVS_ACQ_AFTER(dpdk_mutex
)
282 = OVS_MUTEX_INITIALIZER
;
284 static struct ovs_list dpdk_mp_list
OVS_GUARDED_BY(dpdk_mp_mutex
)
285 = OVS_LIST_INITIALIZER(&dpdk_mp_list
);
288 struct rte_mempool
*mp
;
292 struct ovs_list list_node
OVS_GUARDED_BY(dpdk_mp_mutex
);
295 /* There should be one 'struct dpdk_tx_queue' created for
297 struct dpdk_tx_queue
{
298 rte_spinlock_t tx_lock
; /* Protects the members and the NIC queue
299 * from concurrent access. It is used only
300 * if the queue is shared among different
301 * pmd threads (see 'concurrent_txq'). */
302 int map
; /* Mapping of configured vhost-user queues
303 * to enabled by guest. */
306 /* dpdk has no way to remove dpdk ring ethernet devices
307 so we have to keep them around once they've been created
310 static struct ovs_list dpdk_ring_list
OVS_GUARDED_BY(dpdk_mutex
)
311 = OVS_LIST_INITIALIZER(&dpdk_ring_list
);
314 /* For the client rings */
315 struct rte_ring
*cring_tx
;
316 struct rte_ring
*cring_rx
;
317 unsigned int user_port_id
; /* User given port no, parsed from port name */
318 dpdk_port_t eth_port_id
; /* ethernet device port id */
319 struct ovs_list list_node
OVS_GUARDED_BY(dpdk_mutex
);
322 struct ingress_policer
{
323 struct rte_meter_srtcm_params app_srtcm_params
;
324 struct rte_meter_srtcm in_policer
;
325 rte_spinlock_t policer_lock
;
328 enum dpdk_hw_ol_features
{
329 NETDEV_RX_CHECKSUM_OFFLOAD
= 1 << 0,
336 enum dpdk_dev_type type
;
338 struct dpdk_tx_queue
*tx_q
;
340 struct ovs_mutex mutex
OVS_ACQ_AFTER(dpdk_mutex
);
342 struct dpdk_mp
*dpdk_mp
;
346 struct netdev_stats stats
;
348 rte_spinlock_t stats_lock
;
350 struct eth_addr hwaddr
;
351 enum netdev_flags flags
;
353 struct rte_eth_link link
;
356 /* virtio identifier for vhost devices */
359 /* True if vHost device is 'up' and has been reconfigured at least once */
360 bool vhost_reconfigured
;
362 /* Identifier used to distinguish vhost devices from each other. */
363 char vhost_id
[PATH_MAX
];
365 /* Device arguments for dpdk ports */
368 /* If true, device was attached by rte_eth_dev_attach(). */
372 struct ovs_list list_node
OVS_GUARDED_BY(dpdk_mutex
);
374 /* QoS configuration and lock for the device */
375 OVSRCU_TYPE(struct qos_conf
*) qos_conf
;
377 /* The following properties cannot be changed when a device is running,
378 * so we remember the request and update them next time
379 * netdev_dpdk*_reconfigure() is called */
383 int requested_rxq_size
;
384 int requested_txq_size
;
386 /* Number of rx/tx descriptors for physical devices */
390 /* Socket ID detected when vHost device is brought up */
391 int requested_socket_id
;
393 /* Denotes whether vHost port is client/server mode */
394 uint64_t vhost_driver_flags
;
396 /* Ingress Policer */
397 OVSRCU_TYPE(struct ingress_policer
*) ingress_policer
;
398 uint32_t policer_rate
;
399 uint32_t policer_burst
;
401 /* DPDK-ETH Flow control */
402 struct rte_eth_fc_conf fc_conf
;
404 /* DPDK-ETH hardware offload features,
405 * from the enum set 'dpdk_hw_ol_features' */
406 uint32_t hw_ol_features
;
409 struct netdev_rxq_dpdk
{
410 struct netdev_rxq up
;
414 static int netdev_dpdk_class_init(void);
415 static int netdev_dpdk_vhost_class_init(void);
417 int netdev_dpdk_get_vid(const struct netdev_dpdk
*dev
);
419 struct ingress_policer
*
420 netdev_dpdk_get_ingress_policer(const struct netdev_dpdk
*dev
);
423 is_dpdk_class(const struct netdev_class
*class)
425 return class->init
== netdev_dpdk_class_init
426 || class->init
== netdev_dpdk_vhost_class_init
;
429 /* DPDK NIC drivers allocate RX buffers at a particular granularity, typically
430 * aligned at 1k or less. If a declared mbuf size is not a multiple of this
431 * value, insufficient buffers are allocated to accomodate the packet in its
432 * entirety. Furthermore, certain drivers need to ensure that there is also
433 * sufficient space in the Rx buffer to accommodate two VLAN tags (for QinQ
434 * frames). If the RX buffer is too small, then the driver enables scatter RX
435 * behaviour, which reduces performance. To prevent this, use a buffer size
436 * that is closest to 'mtu', but which satisfies the aforementioned criteria.
439 dpdk_buf_size(int mtu
)
441 return ROUND_UP((MTU_TO_MAX_FRAME_LEN(mtu
) + RTE_PKTMBUF_HEADROOM
),
442 NETDEV_DPDK_MBUF_ALIGN
);
445 /* Allocates an area of 'sz' bytes from DPDK. The memory is zero'ed.
447 * Unlike xmalloc(), this function can return NULL on failure. */
449 dpdk_rte_mzalloc(size_t sz
)
451 return rte_zmalloc(OVS_VPORT_DPDK
, sz
, OVS_CACHE_LINE_SIZE
);
455 free_dpdk_buf(struct dp_packet
*p
)
457 struct rte_mbuf
*pkt
= (struct rte_mbuf
*) p
;
459 rte_pktmbuf_free(pkt
);
463 ovs_rte_pktmbuf_init(struct rte_mempool
*mp OVS_UNUSED
,
464 void *opaque_arg OVS_UNUSED
,
466 unsigned i OVS_UNUSED
)
468 struct rte_mbuf
*pkt
= _p
;
470 dp_packet_init_dpdk((struct dp_packet
*) pkt
, pkt
->buf_len
);
473 static struct dpdk_mp
*
474 dpdk_mp_create(int socket_id
, int mtu
)
480 dmp
= dpdk_rte_mzalloc(sizeof *dmp
);
484 dmp
->socket_id
= socket_id
;
487 /* XXX: this is a really rough method of provisioning memory.
488 * It's impossible to determine what the exact memory requirements are
489 * when the number of ports and rxqs that utilize a particular mempool can
490 * change dynamically at runtime. For now, use this rough heurisitic.
492 if (mtu
>= ETHER_MTU
) {
493 mp_size
= MAX_NB_MBUF
;
495 mp_size
= MIN_NB_MBUF
;
499 mp_name
= xasprintf("ovs_mp_%d_%d_%u", dmp
->mtu
, dmp
->socket_id
,
502 dmp
->mp
= rte_pktmbuf_pool_create(mp_name
, mp_size
,
504 sizeof (struct dp_packet
)
505 - sizeof (struct rte_mbuf
),
507 - sizeof(struct dp_packet
),
510 VLOG_DBG("Allocated \"%s\" mempool with %u mbufs",
515 /* rte_pktmbuf_pool_create has done some initialization of the
516 * rte_mbuf part of each dp_packet, while ovs_rte_pktmbuf_init
517 * initializes some OVS specific fields of dp_packet.
519 rte_mempool_obj_iter(dmp
->mp
, ovs_rte_pktmbuf_init
, NULL
);
522 } while (rte_errno
== ENOMEM
&& (mp_size
/= 2) >= MIN_NB_MBUF
);
528 static struct dpdk_mp
*
529 dpdk_mp_get(int socket_id
, int mtu
)
533 ovs_mutex_lock(&dpdk_mp_mutex
);
534 LIST_FOR_EACH (dmp
, list_node
, &dpdk_mp_list
) {
535 if (dmp
->socket_id
== socket_id
&& dmp
->mtu
== mtu
) {
541 dmp
= dpdk_mp_create(socket_id
, mtu
);
543 ovs_list_push_back(&dpdk_mp_list
, &dmp
->list_node
);
547 ovs_mutex_unlock(&dpdk_mp_mutex
);
553 dpdk_mp_put(struct dpdk_mp
*dmp
)
559 ovs_mutex_lock(&dpdk_mp_mutex
);
560 ovs_assert(dmp
->refcount
);
562 if (!--dmp
->refcount
) {
563 ovs_list_remove(&dmp
->list_node
);
564 rte_mempool_free(dmp
->mp
);
567 ovs_mutex_unlock(&dpdk_mp_mutex
);
570 /* Tries to allocate new mempool on requested_socket_id with
571 * mbuf size corresponding to requested_mtu.
572 * On success new configuration will be applied.
573 * On error, device will be left unchanged. */
575 netdev_dpdk_mempool_configure(struct netdev_dpdk
*dev
)
576 OVS_REQUIRES(dev
->mutex
)
578 uint32_t buf_size
= dpdk_buf_size(dev
->requested_mtu
);
581 mp
= dpdk_mp_get(dev
->requested_socket_id
, FRAME_LEN_TO_MTU(buf_size
));
583 VLOG_ERR("Failed to create memory pool for netdev "
584 "%s, with MTU %d on socket %d: %s\n",
585 dev
->up
.name
, dev
->requested_mtu
, dev
->requested_socket_id
,
586 rte_strerror(rte_errno
));
589 dpdk_mp_put(dev
->dpdk_mp
);
591 dev
->mtu
= dev
->requested_mtu
;
592 dev
->socket_id
= dev
->requested_socket_id
;
593 dev
->max_packet_len
= MTU_TO_FRAME_LEN(dev
->mtu
);
600 check_link_status(struct netdev_dpdk
*dev
)
602 struct rte_eth_link link
;
604 rte_eth_link_get_nowait(dev
->port_id
, &link
);
606 if (dev
->link
.link_status
!= link
.link_status
) {
607 netdev_change_seq_changed(&dev
->up
);
609 dev
->link_reset_cnt
++;
611 if (dev
->link
.link_status
) {
612 VLOG_DBG_RL(&rl
, "Port %"PRIu8
" Link Up - speed %u Mbps - %s",
613 dev
->port_id
, (unsigned) dev
->link
.link_speed
,
614 (dev
->link
.link_duplex
== ETH_LINK_FULL_DUPLEX
) ?
615 ("full-duplex") : ("half-duplex"));
617 VLOG_DBG_RL(&rl
, "Port %"PRIu8
" Link Down", dev
->port_id
);
623 dpdk_watchdog(void *dummy OVS_UNUSED
)
625 struct netdev_dpdk
*dev
;
627 pthread_detach(pthread_self());
630 ovs_mutex_lock(&dpdk_mutex
);
631 LIST_FOR_EACH (dev
, list_node
, &dpdk_list
) {
632 ovs_mutex_lock(&dev
->mutex
);
633 if (dev
->type
== DPDK_DEV_ETH
) {
634 check_link_status(dev
);
636 ovs_mutex_unlock(&dev
->mutex
);
638 ovs_mutex_unlock(&dpdk_mutex
);
639 xsleep(DPDK_PORT_WATCHDOG_INTERVAL
);
646 dpdk_eth_dev_queue_setup(struct netdev_dpdk
*dev
, int n_rxq
, int n_txq
)
650 struct rte_eth_conf conf
= port_conf
;
652 if (dev
->mtu
> ETHER_MTU
) {
653 conf
.rxmode
.jumbo_frame
= 1;
654 conf
.rxmode
.max_rx_pkt_len
= dev
->max_packet_len
;
656 conf
.rxmode
.jumbo_frame
= 0;
657 conf
.rxmode
.max_rx_pkt_len
= 0;
659 conf
.rxmode
.hw_ip_checksum
= (dev
->hw_ol_features
&
660 NETDEV_RX_CHECKSUM_OFFLOAD
) != 0;
661 /* A device may report more queues than it makes available (this has
662 * been observed for Intel xl710, which reserves some of them for
663 * SRIOV): rte_eth_*_queue_setup will fail if a queue is not
664 * available. When this happens we can retry the configuration
665 * and request less queues */
666 while (n_rxq
&& n_txq
) {
668 VLOG_INFO("Retrying setup with (rxq:%d txq:%d)", n_rxq
, n_txq
);
671 diag
= rte_eth_dev_configure(dev
->port_id
, n_rxq
, n_txq
, &conf
);
673 VLOG_WARN("Interface %s eth_dev setup error %s\n",
674 dev
->up
.name
, rte_strerror(-diag
));
678 for (i
= 0; i
< n_txq
; i
++) {
679 diag
= rte_eth_tx_queue_setup(dev
->port_id
, i
, dev
->txq_size
,
680 dev
->socket_id
, NULL
);
682 VLOG_INFO("Interface %s txq(%d) setup error: %s",
683 dev
->up
.name
, i
, rte_strerror(-diag
));
689 /* Retry with less tx queues */
694 for (i
= 0; i
< n_rxq
; i
++) {
695 diag
= rte_eth_rx_queue_setup(dev
->port_id
, i
, dev
->rxq_size
,
696 dev
->socket_id
, NULL
,
699 VLOG_INFO("Interface %s rxq(%d) setup error: %s",
700 dev
->up
.name
, i
, rte_strerror(-diag
));
706 /* Retry with less rx queues */
711 dev
->up
.n_rxq
= n_rxq
;
712 dev
->up
.n_txq
= n_txq
;
721 dpdk_eth_checksum_offload_configure(struct netdev_dpdk
*dev
)
722 OVS_REQUIRES(dev
->mutex
)
724 struct rte_eth_dev_info info
;
725 bool rx_csum_ol_flag
= false;
726 uint32_t rx_chksm_offload_capa
= DEV_RX_OFFLOAD_UDP_CKSUM
|
727 DEV_RX_OFFLOAD_TCP_CKSUM
|
728 DEV_RX_OFFLOAD_IPV4_CKSUM
;
729 rte_eth_dev_info_get(dev
->port_id
, &info
);
730 rx_csum_ol_flag
= (dev
->hw_ol_features
& NETDEV_RX_CHECKSUM_OFFLOAD
) != 0;
732 if (rx_csum_ol_flag
&&
733 (info
.rx_offload_capa
& rx_chksm_offload_capa
) !=
734 rx_chksm_offload_capa
) {
735 VLOG_WARN_ONCE("Rx checksum offload is not supported on device %"PRIu8
,
737 dev
->hw_ol_features
&= ~NETDEV_RX_CHECKSUM_OFFLOAD
;
740 netdev_request_reconfigure(&dev
->up
);
744 dpdk_eth_flow_ctrl_setup(struct netdev_dpdk
*dev
) OVS_REQUIRES(dev
->mutex
)
746 if (rte_eth_dev_flow_ctrl_set(dev
->port_id
, &dev
->fc_conf
)) {
747 VLOG_WARN("Failed to enable flow control on device %"PRIu8
,
753 dpdk_eth_dev_init(struct netdev_dpdk
*dev
)
754 OVS_REQUIRES(dev
->mutex
)
756 struct rte_pktmbuf_pool_private
*mbp_priv
;
757 struct rte_eth_dev_info info
;
758 struct ether_addr eth_addr
;
762 rte_eth_dev_info_get(dev
->port_id
, &info
);
764 n_rxq
= MIN(info
.max_rx_queues
, dev
->up
.n_rxq
);
765 n_txq
= MIN(info
.max_tx_queues
, dev
->up
.n_txq
);
767 diag
= dpdk_eth_dev_queue_setup(dev
, n_rxq
, n_txq
);
769 VLOG_ERR("Interface %s(rxq:%d txq:%d) configure error: %s",
770 dev
->up
.name
, n_rxq
, n_txq
, rte_strerror(-diag
));
774 diag
= rte_eth_dev_start(dev
->port_id
);
776 VLOG_ERR("Interface %s start error: %s", dev
->up
.name
,
777 rte_strerror(-diag
));
781 rte_eth_promiscuous_enable(dev
->port_id
);
782 rte_eth_allmulticast_enable(dev
->port_id
);
784 memset(ð_addr
, 0x0, sizeof(eth_addr
));
785 rte_eth_macaddr_get(dev
->port_id
, ð_addr
);
786 VLOG_INFO_RL(&rl
, "Port %"PRIu8
": "ETH_ADDR_FMT
,
787 dev
->port_id
, ETH_ADDR_BYTES_ARGS(eth_addr
.addr_bytes
));
789 memcpy(dev
->hwaddr
.ea
, eth_addr
.addr_bytes
, ETH_ADDR_LEN
);
790 rte_eth_link_get_nowait(dev
->port_id
, &dev
->link
);
792 mbp_priv
= rte_mempool_get_priv(dev
->dpdk_mp
->mp
);
793 dev
->buf_size
= mbp_priv
->mbuf_data_room_size
- RTE_PKTMBUF_HEADROOM
;
795 /* Get the Flow control configuration for DPDK-ETH */
796 diag
= rte_eth_dev_flow_ctrl_get(dev
->port_id
, &dev
->fc_conf
);
798 VLOG_DBG("cannot get flow control parameters on port=%"PRIu8
", err=%d",
805 static struct netdev_dpdk
*
806 netdev_dpdk_cast(const struct netdev
*netdev
)
808 return CONTAINER_OF(netdev
, struct netdev_dpdk
, up
);
811 static struct netdev
*
812 netdev_dpdk_alloc(void)
814 struct netdev_dpdk
*dev
;
816 dev
= dpdk_rte_mzalloc(sizeof *dev
);
824 static struct dpdk_tx_queue
*
825 netdev_dpdk_alloc_txq(unsigned int n_txqs
)
827 struct dpdk_tx_queue
*txqs
;
830 txqs
= dpdk_rte_mzalloc(n_txqs
* sizeof *txqs
);
832 for (i
= 0; i
< n_txqs
; i
++) {
833 /* Initialize map for vhost devices. */
834 txqs
[i
].map
= OVS_VHOST_QUEUE_MAP_UNKNOWN
;
835 rte_spinlock_init(&txqs
[i
].tx_lock
);
843 common_construct(struct netdev
*netdev
, dpdk_port_t port_no
,
844 enum dpdk_dev_type type
, int socket_id
)
845 OVS_REQUIRES(dpdk_mutex
)
847 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
849 ovs_mutex_init(&dev
->mutex
);
851 rte_spinlock_init(&dev
->stats_lock
);
853 /* If the 'sid' is negative, it means that the kernel fails
854 * to obtain the pci numa info. In that situation, always
856 dev
->socket_id
= socket_id
< 0 ? SOCKET0
: socket_id
;
857 dev
->requested_socket_id
= dev
->socket_id
;
858 dev
->port_id
= port_no
;
861 dev
->requested_mtu
= ETHER_MTU
;
862 dev
->max_packet_len
= MTU_TO_FRAME_LEN(dev
->mtu
);
863 ovsrcu_index_init(&dev
->vid
, -1);
864 dev
->vhost_reconfigured
= false;
865 dev
->attached
= false;
867 ovsrcu_init(&dev
->qos_conf
, NULL
);
869 ovsrcu_init(&dev
->ingress_policer
, NULL
);
870 dev
->policer_rate
= 0;
871 dev
->policer_burst
= 0;
875 dev
->requested_n_rxq
= NR_QUEUE
;
876 dev
->requested_n_txq
= NR_QUEUE
;
877 dev
->requested_rxq_size
= NIC_PORT_DEFAULT_RXQ_SIZE
;
878 dev
->requested_txq_size
= NIC_PORT_DEFAULT_TXQ_SIZE
;
880 /* Initialize the flow control to NULL */
881 memset(&dev
->fc_conf
, 0, sizeof dev
->fc_conf
);
883 /* Initilize the hardware offload flags to 0 */
884 dev
->hw_ol_features
= 0;
886 dev
->flags
= NETDEV_UP
| NETDEV_PROMISC
;
888 ovs_list_push_back(&dpdk_list
, &dev
->list_node
);
890 netdev_request_reconfigure(netdev
);
895 /* dev_name must be the prefix followed by a positive decimal number.
896 * (no leading + or - signs are allowed) */
898 dpdk_dev_parse_name(const char dev_name
[], const char prefix
[],
899 unsigned int *port_no
)
903 if (strncmp(dev_name
, prefix
, strlen(prefix
))) {
907 cport
= dev_name
+ strlen(prefix
);
909 if (str_to_uint(cport
, 10, port_no
)) {
917 vhost_common_construct(struct netdev
*netdev
)
918 OVS_REQUIRES(dpdk_mutex
)
920 int socket_id
= rte_lcore_to_socket_id(rte_get_master_lcore());
921 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
923 dev
->tx_q
= netdev_dpdk_alloc_txq(OVS_VHOST_MAX_QUEUE_NUM
);
928 return common_construct(netdev
, DPDK_ETH_PORT_ID_INVALID
,
929 DPDK_DEV_VHOST
, socket_id
);
933 netdev_dpdk_vhost_construct(struct netdev
*netdev
)
935 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
936 const char *name
= netdev
->name
;
939 /* 'name' is appended to 'vhost_sock_dir' and used to create a socket in
940 * the file system. '/' or '\' would traverse directories, so they're not
941 * acceptable in 'name'. */
942 if (strchr(name
, '/') || strchr(name
, '\\')) {
943 VLOG_ERR("\"%s\" is not a valid name for a vhost-user port. "
944 "A valid name must not include '/' or '\\'",
949 ovs_mutex_lock(&dpdk_mutex
);
950 /* Take the name of the vhost-user port and append it to the location where
951 * the socket is to be created, then register the socket.
953 snprintf(dev
->vhost_id
, sizeof dev
->vhost_id
, "%s/%s",
954 dpdk_get_vhost_sock_dir(), name
);
956 dev
->vhost_driver_flags
&= ~RTE_VHOST_USER_CLIENT
;
957 err
= rte_vhost_driver_register(dev
->vhost_id
, dev
->vhost_driver_flags
);
959 VLOG_ERR("vhost-user socket device setup failure for socket %s\n",
962 fatal_signal_add_file_to_unlink(dev
->vhost_id
);
963 VLOG_INFO("Socket %s created for vhost-user port %s\n",
964 dev
->vhost_id
, name
);
966 err
= vhost_common_construct(netdev
);
968 ovs_mutex_unlock(&dpdk_mutex
);
973 netdev_dpdk_vhost_client_construct(struct netdev
*netdev
)
977 ovs_mutex_lock(&dpdk_mutex
);
978 err
= vhost_common_construct(netdev
);
979 ovs_mutex_unlock(&dpdk_mutex
);
984 netdev_dpdk_construct(struct netdev
*netdev
)
988 ovs_mutex_lock(&dpdk_mutex
);
989 err
= common_construct(netdev
, DPDK_ETH_PORT_ID_INVALID
,
990 DPDK_DEV_ETH
, SOCKET0
);
991 ovs_mutex_unlock(&dpdk_mutex
);
996 common_destruct(struct netdev_dpdk
*dev
)
997 OVS_REQUIRES(dpdk_mutex
)
998 OVS_EXCLUDED(dev
->mutex
)
1000 rte_free(dev
->tx_q
);
1001 dpdk_mp_put(dev
->dpdk_mp
);
1003 ovs_list_remove(&dev
->list_node
);
1004 free(ovsrcu_get_protected(struct ingress_policer
*,
1005 &dev
->ingress_policer
));
1006 ovs_mutex_destroy(&dev
->mutex
);
1010 netdev_dpdk_destruct(struct netdev
*netdev
)
1012 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1013 char devname
[RTE_ETH_NAME_MAX_LEN
];
1015 ovs_mutex_lock(&dpdk_mutex
);
1017 rte_eth_dev_stop(dev
->port_id
);
1019 if (dev
->attached
) {
1020 rte_eth_dev_close(dev
->port_id
);
1021 if (rte_eth_dev_detach(dev
->port_id
, devname
) < 0) {
1022 VLOG_ERR("Device '%s' can not be detached", dev
->devargs
);
1024 VLOG_INFO("Device '%s' detached", devname
);
1029 common_destruct(dev
);
1031 ovs_mutex_unlock(&dpdk_mutex
);
1034 /* rte_vhost_driver_unregister() can call back destroy_device(), which will
1035 * try to acquire 'dpdk_mutex' and possibly 'dev->mutex'. To avoid a
1036 * deadlock, none of the mutexes must be held while calling this function. */
1038 dpdk_vhost_driver_unregister(struct netdev_dpdk
*dev OVS_UNUSED
,
1040 OVS_EXCLUDED(dpdk_mutex
)
1041 OVS_EXCLUDED(dev
->mutex
)
1043 return rte_vhost_driver_unregister(vhost_id
);
1047 netdev_dpdk_vhost_destruct(struct netdev
*netdev
)
1049 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1052 ovs_mutex_lock(&dpdk_mutex
);
1054 /* Guest becomes an orphan if still attached. */
1055 if (netdev_dpdk_get_vid(dev
) >= 0
1056 && !(dev
->vhost_driver_flags
& RTE_VHOST_USER_CLIENT
)) {
1057 VLOG_ERR("Removing port '%s' while vhost device still attached.",
1059 VLOG_ERR("To restore connectivity after re-adding of port, VM on "
1060 "socket '%s' must be restarted.", dev
->vhost_id
);
1063 vhost_id
= xstrdup(dev
->vhost_id
);
1065 common_destruct(dev
);
1067 ovs_mutex_unlock(&dpdk_mutex
);
1073 if (dpdk_vhost_driver_unregister(dev
, vhost_id
)) {
1074 VLOG_ERR("%s: Unable to unregister vhost driver for socket '%s'.\n",
1075 netdev
->name
, vhost_id
);
1076 } else if (!(dev
->vhost_driver_flags
& RTE_VHOST_USER_CLIENT
)) {
1077 /* OVS server mode - remove this socket from list for deletion */
1078 fatal_signal_remove_file_to_unlink(vhost_id
);
1085 netdev_dpdk_dealloc(struct netdev
*netdev
)
1087 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1093 netdev_dpdk_get_config(const struct netdev
*netdev
, struct smap
*args
)
1095 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1097 ovs_mutex_lock(&dev
->mutex
);
1099 smap_add_format(args
, "requested_rx_queues", "%d", dev
->requested_n_rxq
);
1100 smap_add_format(args
, "configured_rx_queues", "%d", netdev
->n_rxq
);
1101 smap_add_format(args
, "requested_tx_queues", "%d", dev
->requested_n_txq
);
1102 smap_add_format(args
, "configured_tx_queues", "%d", netdev
->n_txq
);
1103 smap_add_format(args
, "mtu", "%d", dev
->mtu
);
1105 if (dev
->type
== DPDK_DEV_ETH
) {
1106 smap_add_format(args
, "requested_rxq_descriptors", "%d",
1107 dev
->requested_rxq_size
);
1108 smap_add_format(args
, "configured_rxq_descriptors", "%d",
1110 smap_add_format(args
, "requested_txq_descriptors", "%d",
1111 dev
->requested_txq_size
);
1112 smap_add_format(args
, "configured_txq_descriptors", "%d",
1114 if (dev
->hw_ol_features
& NETDEV_RX_CHECKSUM_OFFLOAD
) {
1115 smap_add(args
, "rx_csum_offload", "true");
1118 ovs_mutex_unlock(&dev
->mutex
);
1123 static struct netdev_dpdk
*
1124 netdev_dpdk_lookup_by_port_id(dpdk_port_t port_id
)
1125 OVS_REQUIRES(dpdk_mutex
)
1127 struct netdev_dpdk
*dev
;
1129 LIST_FOR_EACH (dev
, list_node
, &dpdk_list
) {
1130 if (dev
->port_id
== port_id
) {
1139 netdev_dpdk_process_devargs(struct netdev_dpdk
*dev
,
1140 const char *devargs
, char **errp
)
1142 /* Get the name up to the first comma. */
1143 char *name
= xmemdup0(devargs
, strcspn(devargs
, ","));
1144 dpdk_port_t new_port_id
= DPDK_ETH_PORT_ID_INVALID
;
1146 if (!rte_eth_dev_count()
1147 || rte_eth_dev_get_port_by_name(name
, &new_port_id
)
1148 || !rte_eth_dev_is_valid_port(new_port_id
)) {
1149 /* Device not found in DPDK, attempt to attach it */
1150 if (!rte_eth_dev_attach(devargs
, &new_port_id
)) {
1151 /* Attach successful */
1152 dev
->attached
= true;
1153 VLOG_INFO("Device '%s' attached to DPDK", devargs
);
1155 /* Attach unsuccessful */
1156 new_port_id
= DPDK_ETH_PORT_ID_INVALID
;
1157 VLOG_WARN_BUF(errp
, "Error attaching device '%s' to DPDK",
1167 dpdk_set_rxq_config(struct netdev_dpdk
*dev
, const struct smap
*args
)
1168 OVS_REQUIRES(dev
->mutex
)
1172 new_n_rxq
= MAX(smap_get_int(args
, "n_rxq", NR_QUEUE
), 1);
1173 if (new_n_rxq
!= dev
->requested_n_rxq
) {
1174 dev
->requested_n_rxq
= new_n_rxq
;
1175 netdev_request_reconfigure(&dev
->up
);
1180 dpdk_process_queue_size(struct netdev
*netdev
, const struct smap
*args
,
1181 const char *flag
, int default_size
, int *new_size
)
1183 int queue_size
= smap_get_int(args
, flag
, default_size
);
1185 if (queue_size
<= 0 || queue_size
> NIC_PORT_MAX_Q_SIZE
1186 || !is_pow2(queue_size
)) {
1187 queue_size
= default_size
;
1190 if (queue_size
!= *new_size
) {
1191 *new_size
= queue_size
;
1192 netdev_request_reconfigure(netdev
);
1197 netdev_dpdk_set_config(struct netdev
*netdev
, const struct smap
*args
,
1200 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1201 bool rx_fc_en
, tx_fc_en
, autoneg
;
1202 enum rte_eth_fc_mode fc_mode
;
1203 static const enum rte_eth_fc_mode fc_mode_set
[2][2] = {
1204 {RTE_FC_NONE
, RTE_FC_TX_PAUSE
},
1205 {RTE_FC_RX_PAUSE
, RTE_FC_FULL
}
1209 const char *new_devargs
;
1212 ovs_mutex_lock(&dpdk_mutex
);
1213 ovs_mutex_lock(&dev
->mutex
);
1215 dpdk_set_rxq_config(dev
, args
);
1217 dpdk_process_queue_size(netdev
, args
, "n_rxq_desc",
1218 NIC_PORT_DEFAULT_RXQ_SIZE
,
1219 &dev
->requested_rxq_size
);
1220 dpdk_process_queue_size(netdev
, args
, "n_txq_desc",
1221 NIC_PORT_DEFAULT_TXQ_SIZE
,
1222 &dev
->requested_txq_size
);
1224 new_devargs
= smap_get(args
, "dpdk-devargs");
1226 if (dev
->devargs
&& strcmp(new_devargs
, dev
->devargs
)) {
1227 /* The user requested a new device. If we return error, the caller
1228 * will delete this netdev and try to recreate it. */
1233 /* dpdk-devargs is required for device configuration */
1234 if (new_devargs
&& new_devargs
[0]) {
1235 /* Don't process dpdk-devargs if value is unchanged and port id
1237 if (!(dev
->devargs
&& !strcmp(dev
->devargs
, new_devargs
)
1238 && rte_eth_dev_is_valid_port(dev
->port_id
))) {
1239 dpdk_port_t new_port_id
= netdev_dpdk_process_devargs(dev
,
1242 if (!rte_eth_dev_is_valid_port(new_port_id
)) {
1244 } else if (new_port_id
== dev
->port_id
) {
1245 /* Already configured, do not reconfigure again */
1248 struct netdev_dpdk
*dup_dev
;
1250 dup_dev
= netdev_dpdk_lookup_by_port_id(new_port_id
);
1252 VLOG_WARN_BUF(errp
, "'%s' is trying to use device '%s' "
1253 "which is already in use by '%s'",
1254 netdev_get_name(netdev
), new_devargs
,
1255 netdev_get_name(&dup_dev
->up
));
1258 int sid
= rte_eth_dev_socket_id(new_port_id
);
1260 dev
->requested_socket_id
= sid
< 0 ? SOCKET0
: sid
;
1261 dev
->devargs
= xstrdup(new_devargs
);
1262 dev
->port_id
= new_port_id
;
1263 netdev_request_reconfigure(&dev
->up
);
1269 VLOG_WARN_BUF(errp
, "'%s' is missing 'options:dpdk-devargs'. "
1270 "The old 'dpdk<port_id>' names are not supported",
1271 netdev_get_name(netdev
));
1279 rx_fc_en
= smap_get_bool(args
, "rx-flow-ctrl", false);
1280 tx_fc_en
= smap_get_bool(args
, "tx-flow-ctrl", false);
1281 autoneg
= smap_get_bool(args
, "flow-ctrl-autoneg", false);
1283 fc_mode
= fc_mode_set
[tx_fc_en
][rx_fc_en
];
1284 if (dev
->fc_conf
.mode
!= fc_mode
|| autoneg
!= dev
->fc_conf
.autoneg
) {
1285 dev
->fc_conf
.mode
= fc_mode
;
1286 dev
->fc_conf
.autoneg
= autoneg
;
1287 dpdk_eth_flow_ctrl_setup(dev
);
1290 /* Rx checksum offload configuration */
1291 /* By default the Rx checksum offload is ON */
1292 rx_chksm_ofld
= smap_get_bool(args
, "rx-checksum-offload", true);
1293 temp_flag
= (dev
->hw_ol_features
& NETDEV_RX_CHECKSUM_OFFLOAD
)
1295 if (temp_flag
!= rx_chksm_ofld
) {
1296 dev
->hw_ol_features
^= NETDEV_RX_CHECKSUM_OFFLOAD
;
1297 dpdk_eth_checksum_offload_configure(dev
);
1301 ovs_mutex_unlock(&dev
->mutex
);
1302 ovs_mutex_unlock(&dpdk_mutex
);
1308 netdev_dpdk_ring_set_config(struct netdev
*netdev
, const struct smap
*args
,
1309 char **errp OVS_UNUSED
)
1311 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1313 ovs_mutex_lock(&dev
->mutex
);
1314 dpdk_set_rxq_config(dev
, args
);
1315 ovs_mutex_unlock(&dev
->mutex
);
1321 netdev_dpdk_vhost_client_set_config(struct netdev
*netdev
,
1322 const struct smap
*args
,
1323 char **errp OVS_UNUSED
)
1325 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1328 ovs_mutex_lock(&dev
->mutex
);
1329 if (!(dev
->vhost_driver_flags
& RTE_VHOST_USER_CLIENT
)) {
1330 path
= smap_get(args
, "vhost-server-path");
1331 if (path
&& strcmp(path
, dev
->vhost_id
)) {
1332 strcpy(dev
->vhost_id
, path
);
1333 netdev_request_reconfigure(netdev
);
1336 ovs_mutex_unlock(&dev
->mutex
);
1342 netdev_dpdk_get_numa_id(const struct netdev
*netdev
)
1344 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1346 return dev
->socket_id
;
1349 /* Sets the number of tx queues for the dpdk interface. */
1351 netdev_dpdk_set_tx_multiq(struct netdev
*netdev
, unsigned int n_txq
)
1353 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1355 ovs_mutex_lock(&dev
->mutex
);
1357 if (dev
->requested_n_txq
== n_txq
) {
1361 dev
->requested_n_txq
= n_txq
;
1362 netdev_request_reconfigure(netdev
);
1365 ovs_mutex_unlock(&dev
->mutex
);
1369 static struct netdev_rxq
*
1370 netdev_dpdk_rxq_alloc(void)
1372 struct netdev_rxq_dpdk
*rx
= dpdk_rte_mzalloc(sizeof *rx
);
1381 static struct netdev_rxq_dpdk
*
1382 netdev_rxq_dpdk_cast(const struct netdev_rxq
*rxq
)
1384 return CONTAINER_OF(rxq
, struct netdev_rxq_dpdk
, up
);
1388 netdev_dpdk_rxq_construct(struct netdev_rxq
*rxq
)
1390 struct netdev_rxq_dpdk
*rx
= netdev_rxq_dpdk_cast(rxq
);
1391 struct netdev_dpdk
*dev
= netdev_dpdk_cast(rxq
->netdev
);
1393 ovs_mutex_lock(&dev
->mutex
);
1394 rx
->port_id
= dev
->port_id
;
1395 ovs_mutex_unlock(&dev
->mutex
);
1401 netdev_dpdk_rxq_destruct(struct netdev_rxq
*rxq OVS_UNUSED
)
1406 netdev_dpdk_rxq_dealloc(struct netdev_rxq
*rxq
)
1408 struct netdev_rxq_dpdk
*rx
= netdev_rxq_dpdk_cast(rxq
);
1413 /* Tries to transmit 'pkts' to txq 'qid' of device 'dev'. Takes ownership of
1414 * 'pkts', even in case of failure.
1416 * Returns the number of packets that weren't transmitted. */
1418 netdev_dpdk_eth_tx_burst(struct netdev_dpdk
*dev
, int qid
,
1419 struct rte_mbuf
**pkts
, int cnt
)
1423 while (nb_tx
!= cnt
) {
1426 ret
= rte_eth_tx_burst(dev
->port_id
, qid
, pkts
+ nb_tx
, cnt
- nb_tx
);
1434 if (OVS_UNLIKELY(nb_tx
!= cnt
)) {
1435 /* Free buffers, which we couldn't transmit, one at a time (each
1436 * packet could come from a different mempool) */
1439 for (i
= nb_tx
; i
< cnt
; i
++) {
1440 rte_pktmbuf_free(pkts
[i
]);
1448 netdev_dpdk_policer_pkt_handle(struct rte_meter_srtcm
*meter
,
1449 struct rte_mbuf
*pkt
, uint64_t time
)
1451 uint32_t pkt_len
= rte_pktmbuf_pkt_len(pkt
) - sizeof(struct ether_hdr
);
1453 return rte_meter_srtcm_color_blind_check(meter
, time
, pkt_len
) ==
1458 netdev_dpdk_policer_run(struct rte_meter_srtcm
*meter
,
1459 struct rte_mbuf
**pkts
, int pkt_cnt
)
1463 struct rte_mbuf
*pkt
= NULL
;
1464 uint64_t current_time
= rte_rdtsc();
1466 for (i
= 0; i
< pkt_cnt
; i
++) {
1468 /* Handle current packet */
1469 if (netdev_dpdk_policer_pkt_handle(meter
, pkt
, current_time
)) {
1475 rte_pktmbuf_free(pkt
);
1483 ingress_policer_run(struct ingress_policer
*policer
, struct rte_mbuf
**pkts
,
1488 rte_spinlock_lock(&policer
->policer_lock
);
1489 cnt
= netdev_dpdk_policer_run(&policer
->in_policer
, pkts
, pkt_cnt
);
1490 rte_spinlock_unlock(&policer
->policer_lock
);
1496 is_vhost_running(struct netdev_dpdk
*dev
)
1498 return (netdev_dpdk_get_vid(dev
) >= 0 && dev
->vhost_reconfigured
);
1502 netdev_dpdk_vhost_update_rx_size_counters(struct netdev_stats
*stats
,
1503 unsigned int packet_size
)
1505 /* Hard-coded search for the size bucket. */
1506 if (packet_size
< 256) {
1507 if (packet_size
>= 128) {
1508 stats
->rx_128_to_255_packets
++;
1509 } else if (packet_size
<= 64) {
1510 stats
->rx_1_to_64_packets
++;
1512 stats
->rx_65_to_127_packets
++;
1515 if (packet_size
>= 1523) {
1516 stats
->rx_1523_to_max_packets
++;
1517 } else if (packet_size
>= 1024) {
1518 stats
->rx_1024_to_1522_packets
++;
1519 } else if (packet_size
< 512) {
1520 stats
->rx_256_to_511_packets
++;
1522 stats
->rx_512_to_1023_packets
++;
1528 netdev_dpdk_vhost_update_rx_counters(struct netdev_stats
*stats
,
1529 struct dp_packet
**packets
, int count
,
1533 unsigned int packet_size
;
1534 struct dp_packet
*packet
;
1536 stats
->rx_packets
+= count
;
1537 stats
->rx_dropped
+= dropped
;
1538 for (i
= 0; i
< count
; i
++) {
1539 packet
= packets
[i
];
1540 packet_size
= dp_packet_size(packet
);
1542 if (OVS_UNLIKELY(packet_size
< ETH_HEADER_LEN
)) {
1543 /* This only protects the following multicast counting from
1544 * too short packets, but it does not stop the packet from
1545 * further processing. */
1547 stats
->rx_length_errors
++;
1551 netdev_dpdk_vhost_update_rx_size_counters(stats
, packet_size
);
1553 struct eth_header
*eh
= (struct eth_header
*) dp_packet_data(packet
);
1554 if (OVS_UNLIKELY(eth_addr_is_multicast(eh
->eth_dst
))) {
1558 stats
->rx_bytes
+= packet_size
;
1563 * The receive path for the vhost port is the TX path out from guest.
1566 netdev_dpdk_vhost_rxq_recv(struct netdev_rxq
*rxq
,
1567 struct dp_packet_batch
*batch
)
1569 struct netdev_dpdk
*dev
= netdev_dpdk_cast(rxq
->netdev
);
1570 int qid
= rxq
->queue_id
;
1571 struct ingress_policer
*policer
= netdev_dpdk_get_ingress_policer(dev
);
1573 uint16_t dropped
= 0;
1575 if (OVS_UNLIKELY(!is_vhost_running(dev
)
1576 || !(dev
->flags
& NETDEV_UP
))) {
1580 nb_rx
= rte_vhost_dequeue_burst(netdev_dpdk_get_vid(dev
),
1581 qid
* VIRTIO_QNUM
+ VIRTIO_TXQ
,
1583 (struct rte_mbuf
**) batch
->packets
,
1591 nb_rx
= ingress_policer_run(policer
,
1592 (struct rte_mbuf
**) batch
->packets
,
1597 rte_spinlock_lock(&dev
->stats_lock
);
1598 netdev_dpdk_vhost_update_rx_counters(&dev
->stats
, batch
->packets
,
1600 rte_spinlock_unlock(&dev
->stats_lock
);
1602 batch
->count
= (int) nb_rx
;
1607 netdev_dpdk_rxq_recv(struct netdev_rxq
*rxq
, struct dp_packet_batch
*batch
)
1609 struct netdev_rxq_dpdk
*rx
= netdev_rxq_dpdk_cast(rxq
);
1610 struct netdev_dpdk
*dev
= netdev_dpdk_cast(rxq
->netdev
);
1611 struct ingress_policer
*policer
= netdev_dpdk_get_ingress_policer(dev
);
1615 if (OVS_UNLIKELY(!(dev
->flags
& NETDEV_UP
))) {
1619 nb_rx
= rte_eth_rx_burst(rx
->port_id
, rxq
->queue_id
,
1620 (struct rte_mbuf
**) batch
->packets
,
1628 nb_rx
= ingress_policer_run(policer
,
1629 (struct rte_mbuf
**) batch
->packets
,
1634 /* Update stats to reflect dropped packets */
1635 if (OVS_UNLIKELY(dropped
)) {
1636 rte_spinlock_lock(&dev
->stats_lock
);
1637 dev
->stats
.rx_dropped
+= dropped
;
1638 rte_spinlock_unlock(&dev
->stats_lock
);
1641 batch
->count
= nb_rx
;
1647 netdev_dpdk_qos_run(struct netdev_dpdk
*dev
, struct rte_mbuf
**pkts
,
1650 struct qos_conf
*qos_conf
= ovsrcu_get(struct qos_conf
*, &dev
->qos_conf
);
1653 rte_spinlock_lock(&qos_conf
->lock
);
1654 cnt
= qos_conf
->ops
->qos_run(qos_conf
, pkts
, cnt
);
1655 rte_spinlock_unlock(&qos_conf
->lock
);
1662 netdev_dpdk_filter_packet_len(struct netdev_dpdk
*dev
, struct rte_mbuf
**pkts
,
1667 struct rte_mbuf
*pkt
;
1669 for (i
= 0; i
< pkt_cnt
; i
++) {
1671 if (OVS_UNLIKELY(pkt
->pkt_len
> dev
->max_packet_len
)) {
1672 VLOG_WARN_RL(&rl
, "%s: Too big size %" PRIu32
" max_packet_len %d",
1673 dev
->up
.name
, pkt
->pkt_len
, dev
->max_packet_len
);
1674 rte_pktmbuf_free(pkt
);
1678 if (OVS_UNLIKELY(i
!= cnt
)) {
1688 netdev_dpdk_vhost_update_tx_counters(struct netdev_stats
*stats
,
1689 struct dp_packet
**packets
,
1694 int sent
= attempted
- dropped
;
1696 stats
->tx_packets
+= sent
;
1697 stats
->tx_dropped
+= dropped
;
1699 for (i
= 0; i
< sent
; i
++) {
1700 stats
->tx_bytes
+= dp_packet_size(packets
[i
]);
1705 __netdev_dpdk_vhost_send(struct netdev
*netdev
, int qid
,
1706 struct dp_packet
**pkts
, int cnt
)
1708 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1709 struct rte_mbuf
**cur_pkts
= (struct rte_mbuf
**) pkts
;
1710 unsigned int total_pkts
= cnt
;
1711 unsigned int dropped
= 0;
1714 qid
= dev
->tx_q
[qid
% netdev
->n_txq
].map
;
1716 if (OVS_UNLIKELY(!is_vhost_running(dev
) || qid
< 0
1717 || !(dev
->flags
& NETDEV_UP
))) {
1718 rte_spinlock_lock(&dev
->stats_lock
);
1719 dev
->stats
.tx_dropped
+= cnt
;
1720 rte_spinlock_unlock(&dev
->stats_lock
);
1724 rte_spinlock_lock(&dev
->tx_q
[qid
].tx_lock
);
1726 cnt
= netdev_dpdk_filter_packet_len(dev
, cur_pkts
, cnt
);
1727 /* Check has QoS has been configured for the netdev */
1728 cnt
= netdev_dpdk_qos_run(dev
, cur_pkts
, cnt
);
1729 dropped
= total_pkts
- cnt
;
1732 int vhost_qid
= qid
* VIRTIO_QNUM
+ VIRTIO_RXQ
;
1733 unsigned int tx_pkts
;
1735 tx_pkts
= rte_vhost_enqueue_burst(netdev_dpdk_get_vid(dev
),
1736 vhost_qid
, cur_pkts
, cnt
);
1737 if (OVS_LIKELY(tx_pkts
)) {
1738 /* Packets have been sent.*/
1740 /* Prepare for possible retry.*/
1741 cur_pkts
= &cur_pkts
[tx_pkts
];
1743 /* No packets sent - do not retry.*/
1746 } while (cnt
&& (retries
++ <= VHOST_ENQ_RETRY_NUM
));
1748 rte_spinlock_unlock(&dev
->tx_q
[qid
].tx_lock
);
1750 rte_spinlock_lock(&dev
->stats_lock
);
1751 netdev_dpdk_vhost_update_tx_counters(&dev
->stats
, pkts
, total_pkts
,
1753 rte_spinlock_unlock(&dev
->stats_lock
);
1756 for (i
= 0; i
< total_pkts
- dropped
; i
++) {
1757 dp_packet_delete(pkts
[i
]);
1761 /* Tx function. Transmit packets indefinitely */
1763 dpdk_do_tx_copy(struct netdev
*netdev
, int qid
, struct dp_packet_batch
*batch
)
1764 OVS_NO_THREAD_SAFETY_ANALYSIS
1766 #if !defined(__CHECKER__) && !defined(_WIN32)
1767 const size_t PKT_ARRAY_SIZE
= batch
->count
;
1769 /* Sparse or MSVC doesn't like variable length array. */
1770 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
1772 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1773 struct rte_mbuf
*pkts
[PKT_ARRAY_SIZE
];
1778 dp_packet_batch_apply_cutlen(batch
);
1780 for (i
= 0; i
< batch
->count
; i
++) {
1781 int size
= dp_packet_size(batch
->packets
[i
]);
1783 if (OVS_UNLIKELY(size
> dev
->max_packet_len
)) {
1784 VLOG_WARN_RL(&rl
, "Too big size %d max_packet_len %d",
1785 (int) size
, dev
->max_packet_len
);
1791 pkts
[newcnt
] = rte_pktmbuf_alloc(dev
->dpdk_mp
->mp
);
1793 if (!pkts
[newcnt
]) {
1794 dropped
+= batch
->count
- i
;
1798 /* We have to do a copy for now */
1799 memcpy(rte_pktmbuf_mtod(pkts
[newcnt
], void *),
1800 dp_packet_data(batch
->packets
[i
]), size
);
1802 rte_pktmbuf_data_len(pkts
[newcnt
]) = size
;
1803 rte_pktmbuf_pkt_len(pkts
[newcnt
]) = size
;
1808 if (dev
->type
== DPDK_DEV_VHOST
) {
1809 __netdev_dpdk_vhost_send(netdev
, qid
, (struct dp_packet
**) pkts
,
1812 unsigned int qos_pkts
= newcnt
;
1814 /* Check if QoS has been configured for this netdev. */
1815 newcnt
= netdev_dpdk_qos_run(dev
, pkts
, newcnt
);
1817 dropped
+= qos_pkts
- newcnt
;
1818 dropped
+= netdev_dpdk_eth_tx_burst(dev
, qid
, pkts
, newcnt
);
1821 if (OVS_UNLIKELY(dropped
)) {
1822 rte_spinlock_lock(&dev
->stats_lock
);
1823 dev
->stats
.tx_dropped
+= dropped
;
1824 rte_spinlock_unlock(&dev
->stats_lock
);
1829 netdev_dpdk_vhost_send(struct netdev
*netdev
, int qid
,
1830 struct dp_packet_batch
*batch
,
1831 bool may_steal
, bool concurrent_txq OVS_UNUSED
)
1834 if (OVS_UNLIKELY(!may_steal
|| batch
->packets
[0]->source
!= DPBUF_DPDK
)) {
1835 dpdk_do_tx_copy(netdev
, qid
, batch
);
1836 dp_packet_delete_batch(batch
, may_steal
);
1838 dp_packet_batch_apply_cutlen(batch
);
1839 __netdev_dpdk_vhost_send(netdev
, qid
, batch
->packets
, batch
->count
);
1845 netdev_dpdk_send__(struct netdev_dpdk
*dev
, int qid
,
1846 struct dp_packet_batch
*batch
, bool may_steal
,
1847 bool concurrent_txq
)
1849 if (OVS_UNLIKELY(!(dev
->flags
& NETDEV_UP
))) {
1850 dp_packet_delete_batch(batch
, may_steal
);
1854 if (OVS_UNLIKELY(concurrent_txq
)) {
1855 qid
= qid
% dev
->up
.n_txq
;
1856 rte_spinlock_lock(&dev
->tx_q
[qid
].tx_lock
);
1859 if (OVS_UNLIKELY(!may_steal
||
1860 batch
->packets
[0]->source
!= DPBUF_DPDK
)) {
1861 struct netdev
*netdev
= &dev
->up
;
1863 dpdk_do_tx_copy(netdev
, qid
, batch
);
1864 dp_packet_delete_batch(batch
, may_steal
);
1867 int cnt
= batch
->count
;
1868 struct rte_mbuf
**pkts
= (struct rte_mbuf
**) batch
->packets
;
1870 dp_packet_batch_apply_cutlen(batch
);
1872 cnt
= netdev_dpdk_filter_packet_len(dev
, pkts
, cnt
);
1873 cnt
= netdev_dpdk_qos_run(dev
, pkts
, cnt
);
1874 dropped
= batch
->count
- cnt
;
1876 dropped
+= netdev_dpdk_eth_tx_burst(dev
, qid
, pkts
, cnt
);
1878 if (OVS_UNLIKELY(dropped
)) {
1879 rte_spinlock_lock(&dev
->stats_lock
);
1880 dev
->stats
.tx_dropped
+= dropped
;
1881 rte_spinlock_unlock(&dev
->stats_lock
);
1885 if (OVS_UNLIKELY(concurrent_txq
)) {
1886 rte_spinlock_unlock(&dev
->tx_q
[qid
].tx_lock
);
1891 netdev_dpdk_eth_send(struct netdev
*netdev
, int qid
,
1892 struct dp_packet_batch
*batch
, bool may_steal
,
1893 bool concurrent_txq
)
1895 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1897 netdev_dpdk_send__(dev
, qid
, batch
, may_steal
, concurrent_txq
);
1902 netdev_dpdk_set_etheraddr(struct netdev
*netdev
, const struct eth_addr mac
)
1904 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1906 ovs_mutex_lock(&dev
->mutex
);
1907 if (!eth_addr_equals(dev
->hwaddr
, mac
)) {
1909 netdev_change_seq_changed(netdev
);
1911 ovs_mutex_unlock(&dev
->mutex
);
1917 netdev_dpdk_get_etheraddr(const struct netdev
*netdev
, struct eth_addr
*mac
)
1919 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1921 ovs_mutex_lock(&dev
->mutex
);
1923 ovs_mutex_unlock(&dev
->mutex
);
1929 netdev_dpdk_get_mtu(const struct netdev
*netdev
, int *mtup
)
1931 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1933 ovs_mutex_lock(&dev
->mutex
);
1935 ovs_mutex_unlock(&dev
->mutex
);
1941 netdev_dpdk_set_mtu(struct netdev
*netdev
, int mtu
)
1943 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1945 if (MTU_TO_FRAME_LEN(mtu
) > NETDEV_DPDK_MAX_PKT_LEN
1946 || mtu
< ETHER_MIN_MTU
) {
1947 VLOG_WARN("%s: unsupported MTU %d\n", dev
->up
.name
, mtu
);
1951 ovs_mutex_lock(&dev
->mutex
);
1952 if (dev
->requested_mtu
!= mtu
) {
1953 dev
->requested_mtu
= mtu
;
1954 netdev_request_reconfigure(netdev
);
1956 ovs_mutex_unlock(&dev
->mutex
);
1962 netdev_dpdk_get_carrier(const struct netdev
*netdev
, bool *carrier
);
1965 netdev_dpdk_vhost_get_stats(const struct netdev
*netdev
,
1966 struct netdev_stats
*stats
)
1968 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
1970 ovs_mutex_lock(&dev
->mutex
);
1972 rte_spinlock_lock(&dev
->stats_lock
);
1973 /* Supported Stats */
1974 stats
->rx_packets
+= dev
->stats
.rx_packets
;
1975 stats
->tx_packets
+= dev
->stats
.tx_packets
;
1976 stats
->rx_dropped
= dev
->stats
.rx_dropped
;
1977 stats
->tx_dropped
+= dev
->stats
.tx_dropped
;
1978 stats
->multicast
= dev
->stats
.multicast
;
1979 stats
->rx_bytes
= dev
->stats
.rx_bytes
;
1980 stats
->tx_bytes
= dev
->stats
.tx_bytes
;
1981 stats
->rx_errors
= dev
->stats
.rx_errors
;
1982 stats
->rx_length_errors
= dev
->stats
.rx_length_errors
;
1984 stats
->rx_1_to_64_packets
= dev
->stats
.rx_1_to_64_packets
;
1985 stats
->rx_65_to_127_packets
= dev
->stats
.rx_65_to_127_packets
;
1986 stats
->rx_128_to_255_packets
= dev
->stats
.rx_128_to_255_packets
;
1987 stats
->rx_256_to_511_packets
= dev
->stats
.rx_256_to_511_packets
;
1988 stats
->rx_512_to_1023_packets
= dev
->stats
.rx_512_to_1023_packets
;
1989 stats
->rx_1024_to_1522_packets
= dev
->stats
.rx_1024_to_1522_packets
;
1990 stats
->rx_1523_to_max_packets
= dev
->stats
.rx_1523_to_max_packets
;
1992 rte_spinlock_unlock(&dev
->stats_lock
);
1994 ovs_mutex_unlock(&dev
->mutex
);
2000 netdev_dpdk_convert_xstats(struct netdev_stats
*stats
,
2001 const struct rte_eth_xstat
*xstats
,
2002 const struct rte_eth_xstat_name
*names
,
2003 const unsigned int size
)
2005 for (unsigned int i
= 0; i
< size
; i
++) {
2006 if (strcmp(XSTAT_RX_64_PACKETS
, names
[i
].name
) == 0) {
2007 stats
->rx_1_to_64_packets
= xstats
[i
].value
;
2008 } else if (strcmp(XSTAT_RX_65_TO_127_PACKETS
, names
[i
].name
) == 0) {
2009 stats
->rx_65_to_127_packets
= xstats
[i
].value
;
2010 } else if (strcmp(XSTAT_RX_128_TO_255_PACKETS
, names
[i
].name
) == 0) {
2011 stats
->rx_128_to_255_packets
= xstats
[i
].value
;
2012 } else if (strcmp(XSTAT_RX_256_TO_511_PACKETS
, names
[i
].name
) == 0) {
2013 stats
->rx_256_to_511_packets
= xstats
[i
].value
;
2014 } else if (strcmp(XSTAT_RX_512_TO_1023_PACKETS
, names
[i
].name
) == 0) {
2015 stats
->rx_512_to_1023_packets
= xstats
[i
].value
;
2016 } else if (strcmp(XSTAT_RX_1024_TO_1522_PACKETS
, names
[i
].name
) == 0) {
2017 stats
->rx_1024_to_1522_packets
= xstats
[i
].value
;
2018 } else if (strcmp(XSTAT_RX_1523_TO_MAX_PACKETS
, names
[i
].name
) == 0) {
2019 stats
->rx_1523_to_max_packets
= xstats
[i
].value
;
2020 } else if (strcmp(XSTAT_TX_64_PACKETS
, names
[i
].name
) == 0) {
2021 stats
->tx_1_to_64_packets
= xstats
[i
].value
;
2022 } else if (strcmp(XSTAT_TX_65_TO_127_PACKETS
, names
[i
].name
) == 0) {
2023 stats
->tx_65_to_127_packets
= xstats
[i
].value
;
2024 } else if (strcmp(XSTAT_TX_128_TO_255_PACKETS
, names
[i
].name
) == 0) {
2025 stats
->tx_128_to_255_packets
= xstats
[i
].value
;
2026 } else if (strcmp(XSTAT_TX_256_TO_511_PACKETS
, names
[i
].name
) == 0) {
2027 stats
->tx_256_to_511_packets
= xstats
[i
].value
;
2028 } else if (strcmp(XSTAT_TX_512_TO_1023_PACKETS
, names
[i
].name
) == 0) {
2029 stats
->tx_512_to_1023_packets
= xstats
[i
].value
;
2030 } else if (strcmp(XSTAT_TX_1024_TO_1522_PACKETS
, names
[i
].name
) == 0) {
2031 stats
->tx_1024_to_1522_packets
= xstats
[i
].value
;
2032 } else if (strcmp(XSTAT_TX_1523_TO_MAX_PACKETS
, names
[i
].name
) == 0) {
2033 stats
->tx_1523_to_max_packets
= xstats
[i
].value
;
2034 } else if (strcmp(XSTAT_RX_MULTICAST_PACKETS
, names
[i
].name
) == 0) {
2035 stats
->multicast
= xstats
[i
].value
;
2036 } else if (strcmp(XSTAT_TX_MULTICAST_PACKETS
, names
[i
].name
) == 0) {
2037 stats
->tx_multicast_packets
= xstats
[i
].value
;
2038 } else if (strcmp(XSTAT_RX_BROADCAST_PACKETS
, names
[i
].name
) == 0) {
2039 stats
->rx_broadcast_packets
= xstats
[i
].value
;
2040 } else if (strcmp(XSTAT_TX_BROADCAST_PACKETS
, names
[i
].name
) == 0) {
2041 stats
->tx_broadcast_packets
= xstats
[i
].value
;
2042 } else if (strcmp(XSTAT_RX_UNDERSIZED_ERRORS
, names
[i
].name
) == 0) {
2043 stats
->rx_undersized_errors
= xstats
[i
].value
;
2044 } else if (strcmp(XSTAT_RX_FRAGMENTED_ERRORS
, names
[i
].name
) == 0) {
2045 stats
->rx_fragmented_errors
= xstats
[i
].value
;
2046 } else if (strcmp(XSTAT_RX_JABBER_ERRORS
, names
[i
].name
) == 0) {
2047 stats
->rx_jabber_errors
= xstats
[i
].value
;
2053 netdev_dpdk_get_stats(const struct netdev
*netdev
, struct netdev_stats
*stats
)
2055 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2056 struct rte_eth_stats rte_stats
;
2059 netdev_dpdk_get_carrier(netdev
, &gg
);
2060 ovs_mutex_lock(&dev
->mutex
);
2062 struct rte_eth_xstat
*rte_xstats
= NULL
;
2063 struct rte_eth_xstat_name
*rte_xstats_names
= NULL
;
2064 int rte_xstats_len
, rte_xstats_new_len
, rte_xstats_ret
;
2066 if (rte_eth_stats_get(dev
->port_id
, &rte_stats
)) {
2067 VLOG_ERR("Can't get ETH statistics for port: %"PRIu8
, dev
->port_id
);
2068 ovs_mutex_unlock(&dev
->mutex
);
2072 /* Get length of statistics */
2073 rte_xstats_len
= rte_eth_xstats_get_names(dev
->port_id
, NULL
, 0);
2074 if (rte_xstats_len
< 0) {
2075 VLOG_WARN("Cannot get XSTATS values for port: %"PRIu8
, dev
->port_id
);
2078 /* Reserve memory for xstats names and values */
2079 rte_xstats_names
= xcalloc(rte_xstats_len
, sizeof *rte_xstats_names
);
2080 rte_xstats
= xcalloc(rte_xstats_len
, sizeof *rte_xstats
);
2082 /* Retreive xstats names */
2083 rte_xstats_new_len
= rte_eth_xstats_get_names(dev
->port_id
,
2086 if (rte_xstats_new_len
!= rte_xstats_len
) {
2087 VLOG_WARN("Cannot get XSTATS names for port: %"PRIu8
, dev
->port_id
);
2090 /* Retreive xstats values */
2091 memset(rte_xstats
, 0xff, sizeof *rte_xstats
* rte_xstats_len
);
2092 rte_xstats_ret
= rte_eth_xstats_get(dev
->port_id
, rte_xstats
,
2094 if (rte_xstats_ret
> 0 && rte_xstats_ret
<= rte_xstats_len
) {
2095 netdev_dpdk_convert_xstats(stats
, rte_xstats
, rte_xstats_names
,
2098 VLOG_WARN("Cannot get XSTATS values for port: %"PRIu8
, dev
->port_id
);
2103 free(rte_xstats_names
);
2105 stats
->rx_packets
= rte_stats
.ipackets
;
2106 stats
->tx_packets
= rte_stats
.opackets
;
2107 stats
->rx_bytes
= rte_stats
.ibytes
;
2108 stats
->tx_bytes
= rte_stats
.obytes
;
2109 stats
->rx_errors
= rte_stats
.ierrors
;
2110 stats
->tx_errors
= rte_stats
.oerrors
;
2112 rte_spinlock_lock(&dev
->stats_lock
);
2113 stats
->tx_dropped
= dev
->stats
.tx_dropped
;
2114 stats
->rx_dropped
= dev
->stats
.rx_dropped
;
2115 rte_spinlock_unlock(&dev
->stats_lock
);
2117 /* These are the available DPDK counters for packets not received due to
2118 * local resource constraints in DPDK and NIC respectively. */
2119 stats
->rx_dropped
+= rte_stats
.rx_nombuf
+ rte_stats
.imissed
;
2120 stats
->rx_missed_errors
= rte_stats
.imissed
;
2122 ovs_mutex_unlock(&dev
->mutex
);
2128 netdev_dpdk_get_features(const struct netdev
*netdev
,
2129 enum netdev_features
*current
,
2130 enum netdev_features
*advertised
,
2131 enum netdev_features
*supported
,
2132 enum netdev_features
*peer
)
2134 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2135 struct rte_eth_link link
;
2137 ovs_mutex_lock(&dev
->mutex
);
2139 ovs_mutex_unlock(&dev
->mutex
);
2141 if (link
.link_duplex
== ETH_LINK_HALF_DUPLEX
) {
2142 if (link
.link_speed
== ETH_SPEED_NUM_10M
) {
2143 *current
= NETDEV_F_10MB_HD
;
2145 if (link
.link_speed
== ETH_SPEED_NUM_100M
) {
2146 *current
= NETDEV_F_100MB_HD
;
2148 if (link
.link_speed
== ETH_SPEED_NUM_1G
) {
2149 *current
= NETDEV_F_1GB_HD
;
2151 } else if (link
.link_duplex
== ETH_LINK_FULL_DUPLEX
) {
2152 if (link
.link_speed
== ETH_SPEED_NUM_10M
) {
2153 *current
= NETDEV_F_10MB_FD
;
2155 if (link
.link_speed
== ETH_SPEED_NUM_100M
) {
2156 *current
= NETDEV_F_100MB_FD
;
2158 if (link
.link_speed
== ETH_SPEED_NUM_1G
) {
2159 *current
= NETDEV_F_1GB_FD
;
2161 if (link
.link_speed
== ETH_SPEED_NUM_10G
) {
2162 *current
= NETDEV_F_10GB_FD
;
2166 if (link
.link_autoneg
) {
2167 *current
|= NETDEV_F_AUTONEG
;
2170 *advertised
= *supported
= *peer
= 0;
2175 static struct ingress_policer
*
2176 netdev_dpdk_policer_construct(uint32_t rate
, uint32_t burst
)
2178 struct ingress_policer
*policer
= NULL
;
2179 uint64_t rate_bytes
;
2180 uint64_t burst_bytes
;
2183 policer
= xmalloc(sizeof *policer
);
2184 rte_spinlock_init(&policer
->policer_lock
);
2186 /* rte_meter requires bytes so convert kbits rate and burst to bytes. */
2187 rate_bytes
= rate
* 1000/8;
2188 burst_bytes
= burst
* 1000/8;
2190 policer
->app_srtcm_params
.cir
= rate_bytes
;
2191 policer
->app_srtcm_params
.cbs
= burst_bytes
;
2192 policer
->app_srtcm_params
.ebs
= 0;
2193 err
= rte_meter_srtcm_config(&policer
->in_policer
,
2194 &policer
->app_srtcm_params
);
2196 VLOG_ERR("Could not create rte meter for ingress policer");
2204 netdev_dpdk_set_policing(struct netdev
* netdev
, uint32_t policer_rate
,
2205 uint32_t policer_burst
)
2207 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2208 struct ingress_policer
*policer
;
2210 /* Force to 0 if no rate specified,
2211 * default to 8000 kbits if burst is 0,
2212 * else stick with user-specified value.
2214 policer_burst
= (!policer_rate
? 0
2215 : !policer_burst
? 8000
2218 ovs_mutex_lock(&dev
->mutex
);
2220 policer
= ovsrcu_get_protected(struct ingress_policer
*,
2221 &dev
->ingress_policer
);
2223 if (dev
->policer_rate
== policer_rate
&&
2224 dev
->policer_burst
== policer_burst
) {
2225 /* Assume that settings haven't changed since we last set them. */
2226 ovs_mutex_unlock(&dev
->mutex
);
2230 /* Destroy any existing ingress policer for the device if one exists */
2232 ovsrcu_postpone(free
, policer
);
2235 if (policer_rate
!= 0) {
2236 policer
= netdev_dpdk_policer_construct(policer_rate
, policer_burst
);
2240 ovsrcu_set(&dev
->ingress_policer
, policer
);
2241 dev
->policer_rate
= policer_rate
;
2242 dev
->policer_burst
= policer_burst
;
2243 ovs_mutex_unlock(&dev
->mutex
);
2249 netdev_dpdk_get_ifindex(const struct netdev
*netdev
)
2251 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2253 ovs_mutex_lock(&dev
->mutex
);
2254 /* Calculate hash from the netdev name. Ensure that ifindex is a 24-bit
2255 * postive integer to meet RFC 2863 recommendations.
2257 int ifindex
= hash_string(netdev
->name
, 0) % 0xfffffe + 1;
2258 ovs_mutex_unlock(&dev
->mutex
);
2264 netdev_dpdk_get_carrier(const struct netdev
*netdev
, bool *carrier
)
2266 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2268 ovs_mutex_lock(&dev
->mutex
);
2269 check_link_status(dev
);
2270 *carrier
= dev
->link
.link_status
;
2272 ovs_mutex_unlock(&dev
->mutex
);
2278 netdev_dpdk_vhost_get_carrier(const struct netdev
*netdev
, bool *carrier
)
2280 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2282 ovs_mutex_lock(&dev
->mutex
);
2284 if (is_vhost_running(dev
)) {
2290 ovs_mutex_unlock(&dev
->mutex
);
2295 static long long int
2296 netdev_dpdk_get_carrier_resets(const struct netdev
*netdev
)
2298 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2299 long long int carrier_resets
;
2301 ovs_mutex_lock(&dev
->mutex
);
2302 carrier_resets
= dev
->link_reset_cnt
;
2303 ovs_mutex_unlock(&dev
->mutex
);
2305 return carrier_resets
;
2309 netdev_dpdk_set_miimon(struct netdev
*netdev OVS_UNUSED
,
2310 long long int interval OVS_UNUSED
)
2316 netdev_dpdk_update_flags__(struct netdev_dpdk
*dev
,
2317 enum netdev_flags off
, enum netdev_flags on
,
2318 enum netdev_flags
*old_flagsp
)
2319 OVS_REQUIRES(dev
->mutex
)
2321 if ((off
| on
) & ~(NETDEV_UP
| NETDEV_PROMISC
)) {
2325 *old_flagsp
= dev
->flags
;
2329 if (dev
->flags
== *old_flagsp
) {
2333 if (dev
->type
== DPDK_DEV_ETH
) {
2334 if (dev
->flags
& NETDEV_PROMISC
) {
2335 rte_eth_promiscuous_enable(dev
->port_id
);
2338 netdev_change_seq_changed(&dev
->up
);
2340 /* If DPDK_DEV_VHOST device's NETDEV_UP flag was changed and vhost is
2341 * running then change netdev's change_seq to trigger link state
2344 if ((NETDEV_UP
& ((*old_flagsp
^ on
) | (*old_flagsp
^ off
)))
2345 && is_vhost_running(dev
)) {
2346 netdev_change_seq_changed(&dev
->up
);
2348 /* Clear statistics if device is getting up. */
2349 if (NETDEV_UP
& on
) {
2350 rte_spinlock_lock(&dev
->stats_lock
);
2351 memset(&dev
->stats
, 0, sizeof dev
->stats
);
2352 rte_spinlock_unlock(&dev
->stats_lock
);
2361 netdev_dpdk_update_flags(struct netdev
*netdev
,
2362 enum netdev_flags off
, enum netdev_flags on
,
2363 enum netdev_flags
*old_flagsp
)
2365 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2368 ovs_mutex_lock(&dev
->mutex
);
2369 error
= netdev_dpdk_update_flags__(dev
, off
, on
, old_flagsp
);
2370 ovs_mutex_unlock(&dev
->mutex
);
2376 netdev_dpdk_get_status(const struct netdev
*netdev
, struct smap
*args
)
2378 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2379 struct rte_eth_dev_info dev_info
;
2381 if (!rte_eth_dev_is_valid_port(dev
->port_id
)) {
2385 ovs_mutex_lock(&dev
->mutex
);
2386 rte_eth_dev_info_get(dev
->port_id
, &dev_info
);
2387 ovs_mutex_unlock(&dev
->mutex
);
2389 smap_add_format(args
, "port_no", "%d", dev
->port_id
);
2390 smap_add_format(args
, "numa_id", "%d",
2391 rte_eth_dev_socket_id(dev
->port_id
));
2392 smap_add_format(args
, "driver_name", "%s", dev_info
.driver_name
);
2393 smap_add_format(args
, "min_rx_bufsize", "%u", dev_info
.min_rx_bufsize
);
2394 smap_add_format(args
, "max_rx_pktlen", "%u", dev
->max_packet_len
);
2395 smap_add_format(args
, "max_rx_queues", "%u", dev_info
.max_rx_queues
);
2396 smap_add_format(args
, "max_tx_queues", "%u", dev_info
.max_tx_queues
);
2397 smap_add_format(args
, "max_mac_addrs", "%u", dev_info
.max_mac_addrs
);
2398 smap_add_format(args
, "max_hash_mac_addrs", "%u",
2399 dev_info
.max_hash_mac_addrs
);
2400 smap_add_format(args
, "max_vfs", "%u", dev_info
.max_vfs
);
2401 smap_add_format(args
, "max_vmdq_pools", "%u", dev_info
.max_vmdq_pools
);
2403 if (dev_info
.pci_dev
) {
2404 smap_add_format(args
, "pci-vendor_id", "0x%u",
2405 dev_info
.pci_dev
->id
.vendor_id
);
2406 smap_add_format(args
, "pci-device_id", "0x%x",
2407 dev_info
.pci_dev
->id
.device_id
);
2414 netdev_dpdk_set_admin_state__(struct netdev_dpdk
*dev
, bool admin_state
)
2415 OVS_REQUIRES(dev
->mutex
)
2417 enum netdev_flags old_flags
;
2420 netdev_dpdk_update_flags__(dev
, 0, NETDEV_UP
, &old_flags
);
2422 netdev_dpdk_update_flags__(dev
, NETDEV_UP
, 0, &old_flags
);
2427 netdev_dpdk_set_admin_state(struct unixctl_conn
*conn
, int argc
,
2428 const char *argv
[], void *aux OVS_UNUSED
)
2432 if (!strcasecmp(argv
[argc
- 1], "up")) {
2434 } else if ( !strcasecmp(argv
[argc
- 1], "down")) {
2437 unixctl_command_reply_error(conn
, "Invalid Admin State");
2442 struct netdev
*netdev
= netdev_from_name(argv
[1]);
2443 if (netdev
&& is_dpdk_class(netdev
->netdev_class
)) {
2444 struct netdev_dpdk
*dpdk_dev
= netdev_dpdk_cast(netdev
);
2446 ovs_mutex_lock(&dpdk_dev
->mutex
);
2447 netdev_dpdk_set_admin_state__(dpdk_dev
, up
);
2448 ovs_mutex_unlock(&dpdk_dev
->mutex
);
2450 netdev_close(netdev
);
2452 unixctl_command_reply_error(conn
, "Not a DPDK Interface");
2453 netdev_close(netdev
);
2457 struct netdev_dpdk
*netdev
;
2459 ovs_mutex_lock(&dpdk_mutex
);
2460 LIST_FOR_EACH (netdev
, list_node
, &dpdk_list
) {
2461 ovs_mutex_lock(&netdev
->mutex
);
2462 netdev_dpdk_set_admin_state__(netdev
, up
);
2463 ovs_mutex_unlock(&netdev
->mutex
);
2465 ovs_mutex_unlock(&dpdk_mutex
);
2467 unixctl_command_reply(conn
, "OK");
2471 * Set virtqueue flags so that we do not receive interrupts.
2474 set_irq_status(int vid
)
2479 for (i
= 0; i
< rte_vhost_get_queue_num(vid
); i
++) {
2480 idx
= i
* VIRTIO_QNUM
;
2481 rte_vhost_enable_guest_notification(vid
, idx
+ VIRTIO_RXQ
, 0);
2482 rte_vhost_enable_guest_notification(vid
, idx
+ VIRTIO_TXQ
, 0);
2487 * Fixes mapping for vhost-user tx queues. Must be called after each
2488 * enabling/disabling of queues and n_txq modifications.
2491 netdev_dpdk_remap_txqs(struct netdev_dpdk
*dev
)
2492 OVS_REQUIRES(dev
->mutex
)
2494 int *enabled_queues
, n_enabled
= 0;
2495 int i
, k
, total_txqs
= dev
->up
.n_txq
;
2497 enabled_queues
= xcalloc(total_txqs
, sizeof *enabled_queues
);
2499 for (i
= 0; i
< total_txqs
; i
++) {
2500 /* Enabled queues always mapped to themselves. */
2501 if (dev
->tx_q
[i
].map
== i
) {
2502 enabled_queues
[n_enabled
++] = i
;
2506 if (n_enabled
== 0 && total_txqs
!= 0) {
2507 enabled_queues
[0] = OVS_VHOST_QUEUE_DISABLED
;
2512 for (i
= 0; i
< total_txqs
; i
++) {
2513 if (dev
->tx_q
[i
].map
!= i
) {
2514 dev
->tx_q
[i
].map
= enabled_queues
[k
];
2515 k
= (k
+ 1) % n_enabled
;
2519 VLOG_DBG("TX queue mapping for %s\n", dev
->vhost_id
);
2520 for (i
= 0; i
< total_txqs
; i
++) {
2521 VLOG_DBG("%2d --> %2d", i
, dev
->tx_q
[i
].map
);
2524 free(enabled_queues
);
2528 * A new virtio-net device is added to a vhost port.
2533 struct netdev_dpdk
*dev
;
2534 bool exists
= false;
2536 char ifname
[IF_NAME_SZ
];
2538 rte_vhost_get_ifname(vid
, ifname
, sizeof ifname
);
2540 ovs_mutex_lock(&dpdk_mutex
);
2541 /* Add device to the vhost port with the same name as that passed down. */
2542 LIST_FOR_EACH(dev
, list_node
, &dpdk_list
) {
2543 ovs_mutex_lock(&dev
->mutex
);
2544 if (strncmp(ifname
, dev
->vhost_id
, IF_NAME_SZ
) == 0) {
2545 uint32_t qp_num
= rte_vhost_get_queue_num(vid
);
2547 /* Get NUMA information */
2548 newnode
= rte_vhost_get_numa_node(vid
);
2549 if (newnode
== -1) {
2551 VLOG_INFO("Error getting NUMA info for vHost Device '%s'",
2554 newnode
= dev
->socket_id
;
2557 if (dev
->requested_n_txq
!= qp_num
2558 || dev
->requested_n_rxq
!= qp_num
2559 || dev
->requested_socket_id
!= newnode
) {
2560 dev
->requested_socket_id
= newnode
;
2561 dev
->requested_n_rxq
= qp_num
;
2562 dev
->requested_n_txq
= qp_num
;
2563 netdev_request_reconfigure(&dev
->up
);
2565 /* Reconfiguration not required. */
2566 dev
->vhost_reconfigured
= true;
2569 ovsrcu_index_set(&dev
->vid
, vid
);
2572 /* Disable notifications. */
2573 set_irq_status(vid
);
2574 netdev_change_seq_changed(&dev
->up
);
2575 ovs_mutex_unlock(&dev
->mutex
);
2578 ovs_mutex_unlock(&dev
->mutex
);
2580 ovs_mutex_unlock(&dpdk_mutex
);
2583 VLOG_INFO("vHost Device '%s' can't be added - name not found", ifname
);
2588 VLOG_INFO("vHost Device '%s' has been added on numa node %i",
2594 /* Clears mapping for all available queues of vhost interface. */
2596 netdev_dpdk_txq_map_clear(struct netdev_dpdk
*dev
)
2597 OVS_REQUIRES(dev
->mutex
)
2601 for (i
= 0; i
< dev
->up
.n_txq
; i
++) {
2602 dev
->tx_q
[i
].map
= OVS_VHOST_QUEUE_MAP_UNKNOWN
;
2607 * Remove a virtio-net device from the specific vhost port. Use dev->remove
2608 * flag to stop any more packets from being sent or received to/from a VM and
2609 * ensure all currently queued packets have been sent/received before removing
2613 destroy_device(int vid
)
2615 struct netdev_dpdk
*dev
;
2616 bool exists
= false;
2617 char ifname
[IF_NAME_SZ
];
2619 rte_vhost_get_ifname(vid
, ifname
, sizeof ifname
);
2621 ovs_mutex_lock(&dpdk_mutex
);
2622 LIST_FOR_EACH (dev
, list_node
, &dpdk_list
) {
2623 if (netdev_dpdk_get_vid(dev
) == vid
) {
2625 ovs_mutex_lock(&dev
->mutex
);
2626 dev
->vhost_reconfigured
= false;
2627 ovsrcu_index_set(&dev
->vid
, -1);
2628 netdev_dpdk_txq_map_clear(dev
);
2630 netdev_change_seq_changed(&dev
->up
);
2631 ovs_mutex_unlock(&dev
->mutex
);
2637 ovs_mutex_unlock(&dpdk_mutex
);
2641 * Wait for other threads to quiesce after setting the 'virtio_dev'
2642 * to NULL, before returning.
2644 ovsrcu_synchronize();
2646 * As call to ovsrcu_synchronize() will end the quiescent state,
2647 * put thread back into quiescent state before returning.
2649 ovsrcu_quiesce_start();
2650 VLOG_INFO("vHost Device '%s' has been removed", ifname
);
2652 VLOG_INFO("vHost Device '%s' not found", ifname
);
2657 vring_state_changed(int vid
, uint16_t queue_id
, int enable
)
2659 struct netdev_dpdk
*dev
;
2660 bool exists
= false;
2661 int qid
= queue_id
/ VIRTIO_QNUM
;
2662 char ifname
[IF_NAME_SZ
];
2664 rte_vhost_get_ifname(vid
, ifname
, sizeof ifname
);
2666 if (queue_id
% VIRTIO_QNUM
== VIRTIO_TXQ
) {
2670 ovs_mutex_lock(&dpdk_mutex
);
2671 LIST_FOR_EACH (dev
, list_node
, &dpdk_list
) {
2672 ovs_mutex_lock(&dev
->mutex
);
2673 if (strncmp(ifname
, dev
->vhost_id
, IF_NAME_SZ
) == 0) {
2675 dev
->tx_q
[qid
].map
= qid
;
2677 dev
->tx_q
[qid
].map
= OVS_VHOST_QUEUE_DISABLED
;
2679 netdev_dpdk_remap_txqs(dev
);
2681 ovs_mutex_unlock(&dev
->mutex
);
2684 ovs_mutex_unlock(&dev
->mutex
);
2686 ovs_mutex_unlock(&dpdk_mutex
);
2689 VLOG_INFO("State of queue %d ( tx_qid %d ) of vhost device '%s'"
2690 "changed to \'%s\'", queue_id
, qid
, ifname
,
2691 (enable
== 1) ? "enabled" : "disabled");
2693 VLOG_INFO("vHost Device '%s' not found", ifname
);
2701 netdev_dpdk_get_vid(const struct netdev_dpdk
*dev
)
2703 return ovsrcu_index_get(&dev
->vid
);
2706 struct ingress_policer
*
2707 netdev_dpdk_get_ingress_policer(const struct netdev_dpdk
*dev
)
2709 return ovsrcu_get(struct ingress_policer
*, &dev
->ingress_policer
);
2713 * These callbacks allow virtio-net devices to be added to vhost ports when
2714 * configuration has been fully complete.
2716 static const struct virtio_net_device_ops virtio_net_device_ops
=
2718 .new_device
= new_device
,
2719 .destroy_device
= destroy_device
,
2720 .vring_state_changed
= vring_state_changed
2724 start_vhost_loop(void *dummy OVS_UNUSED
)
2726 pthread_detach(pthread_self());
2727 /* Put the vhost thread into quiescent state. */
2728 ovsrcu_quiesce_start();
2729 rte_vhost_driver_session_start();
2734 netdev_dpdk_class_init(void)
2736 static struct ovsthread_once once
= OVSTHREAD_ONCE_INITIALIZER
;
2738 /* This function can be called for different classes. The initialization
2739 * needs to be done only once */
2740 if (ovsthread_once_start(&once
)) {
2741 ovs_thread_create("dpdk_watchdog", dpdk_watchdog
, NULL
);
2742 unixctl_command_register("netdev-dpdk/set-admin-state",
2743 "[netdev] up|down", 1, 2,
2744 netdev_dpdk_set_admin_state
, NULL
);
2746 ovsthread_once_done(&once
);
2753 netdev_dpdk_vhost_class_init(void)
2755 static struct ovsthread_once once
= OVSTHREAD_ONCE_INITIALIZER
;
2757 /* This function can be called for different classes. The initialization
2758 * needs to be done only once */
2759 if (ovsthread_once_start(&once
)) {
2760 rte_vhost_driver_callback_register(&virtio_net_device_ops
);
2761 rte_vhost_feature_disable(1ULL << VIRTIO_NET_F_HOST_TSO4
2762 | 1ULL << VIRTIO_NET_F_HOST_TSO6
2763 | 1ULL << VIRTIO_NET_F_CSUM
);
2764 ovs_thread_create("vhost_thread", start_vhost_loop
, NULL
);
2766 ovsthread_once_done(&once
);
2775 dpdk_ring_create(const char dev_name
[], unsigned int port_no
,
2776 dpdk_port_t
*eth_port_id
)
2778 struct dpdk_ring
*ring_pair
;
2782 ring_pair
= dpdk_rte_mzalloc(sizeof *ring_pair
);
2787 /* XXX: Add support for multiquque ring. */
2788 ring_name
= xasprintf("%s_tx", dev_name
);
2790 /* Create single producer tx ring, netdev does explicit locking. */
2791 ring_pair
->cring_tx
= rte_ring_create(ring_name
, DPDK_RING_SIZE
, SOCKET0
,
2794 if (ring_pair
->cring_tx
== NULL
) {
2795 rte_free(ring_pair
);
2799 ring_name
= xasprintf("%s_rx", dev_name
);
2801 /* Create single consumer rx ring, netdev does explicit locking. */
2802 ring_pair
->cring_rx
= rte_ring_create(ring_name
, DPDK_RING_SIZE
, SOCKET0
,
2805 if (ring_pair
->cring_rx
== NULL
) {
2806 rte_free(ring_pair
);
2810 port_id
= rte_eth_from_rings(dev_name
, &ring_pair
->cring_rx
, 1,
2811 &ring_pair
->cring_tx
, 1, SOCKET0
);
2814 rte_free(ring_pair
);
2818 ring_pair
->user_port_id
= port_no
;
2819 ring_pair
->eth_port_id
= port_id
;
2820 *eth_port_id
= port_id
;
2822 ovs_list_push_back(&dpdk_ring_list
, &ring_pair
->list_node
);
2828 dpdk_ring_open(const char dev_name
[], dpdk_port_t
*eth_port_id
)
2829 OVS_REQUIRES(dpdk_mutex
)
2831 struct dpdk_ring
*ring_pair
;
2832 unsigned int port_no
;
2835 /* Names always start with "dpdkr" */
2836 err
= dpdk_dev_parse_name(dev_name
, "dpdkr", &port_no
);
2841 /* Look through our list to find the device */
2842 LIST_FOR_EACH (ring_pair
, list_node
, &dpdk_ring_list
) {
2843 if (ring_pair
->user_port_id
== port_no
) {
2844 VLOG_INFO("Found dpdk ring device %s:", dev_name
);
2845 /* Really all that is needed */
2846 *eth_port_id
= ring_pair
->eth_port_id
;
2850 /* Need to create the device rings */
2851 return dpdk_ring_create(dev_name
, port_no
, eth_port_id
);
2855 netdev_dpdk_ring_send(struct netdev
*netdev
, int qid
,
2856 struct dp_packet_batch
*batch
, bool may_steal
,
2857 bool concurrent_txq
)
2859 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2862 /* When using 'dpdkr' and sending to a DPDK ring, we want to ensure that
2863 * the rss hash field is clear. This is because the same mbuf may be
2864 * modified by the consumer of the ring and return into the datapath
2865 * without recalculating the RSS hash. */
2866 for (i
= 0; i
< batch
->count
; i
++) {
2867 dp_packet_rss_invalidate(batch
->packets
[i
]);
2870 netdev_dpdk_send__(dev
, qid
, batch
, may_steal
, concurrent_txq
);
2875 netdev_dpdk_ring_construct(struct netdev
*netdev
)
2877 dpdk_port_t port_no
= 0;
2880 ovs_mutex_lock(&dpdk_mutex
);
2882 err
= dpdk_ring_open(netdev
->name
, &port_no
);
2887 err
= common_construct(netdev
, port_no
, DPDK_DEV_ETH
,
2888 rte_eth_dev_socket_id(port_no
));
2890 ovs_mutex_unlock(&dpdk_mutex
);
2897 * Initialize QoS configuration operations.
2900 qos_conf_init(struct qos_conf
*conf
, const struct dpdk_qos_ops
*ops
)
2903 rte_spinlock_init(&conf
->lock
);
2907 * Search existing QoS operations in qos_ops and compare each set of
2908 * operations qos_name to name. Return a dpdk_qos_ops pointer to a match,
2911 static const struct dpdk_qos_ops
*
2912 qos_lookup_name(const char *name
)
2914 const struct dpdk_qos_ops
*const *opsp
;
2916 for (opsp
= qos_confs
; *opsp
!= NULL
; opsp
++) {
2917 const struct dpdk_qos_ops
*ops
= *opsp
;
2918 if (!strcmp(name
, ops
->qos_name
)) {
2926 netdev_dpdk_get_qos_types(const struct netdev
*netdev OVS_UNUSED
,
2929 const struct dpdk_qos_ops
*const *opsp
;
2931 for (opsp
= qos_confs
; *opsp
!= NULL
; opsp
++) {
2932 const struct dpdk_qos_ops
*ops
= *opsp
;
2933 if (ops
->qos_construct
&& ops
->qos_name
[0] != '\0') {
2934 sset_add(types
, ops
->qos_name
);
2941 netdev_dpdk_get_qos(const struct netdev
*netdev
,
2942 const char **typep
, struct smap
*details
)
2944 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2945 struct qos_conf
*qos_conf
;
2948 ovs_mutex_lock(&dev
->mutex
);
2949 qos_conf
= ovsrcu_get_protected(struct qos_conf
*, &dev
->qos_conf
);
2951 *typep
= qos_conf
->ops
->qos_name
;
2952 error
= (qos_conf
->ops
->qos_get
2953 ? qos_conf
->ops
->qos_get(qos_conf
, details
): 0);
2955 /* No QoS configuration set, return an empty string */
2958 ovs_mutex_unlock(&dev
->mutex
);
2964 netdev_dpdk_set_qos(struct netdev
*netdev
, const char *type
,
2965 const struct smap
*details
)
2967 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
2968 const struct dpdk_qos_ops
*new_ops
= NULL
;
2969 struct qos_conf
*qos_conf
, *new_qos_conf
= NULL
;
2972 ovs_mutex_lock(&dev
->mutex
);
2974 qos_conf
= ovsrcu_get_protected(struct qos_conf
*, &dev
->qos_conf
);
2976 new_ops
= qos_lookup_name(type
);
2978 if (!new_ops
|| !new_ops
->qos_construct
) {
2979 new_qos_conf
= NULL
;
2980 if (type
&& type
[0]) {
2983 } else if (qos_conf
&& qos_conf
->ops
== new_ops
2984 && qos_conf
->ops
->qos_is_equal(qos_conf
, details
)) {
2985 new_qos_conf
= qos_conf
;
2987 error
= new_ops
->qos_construct(details
, &new_qos_conf
);
2991 VLOG_ERR("Failed to set QoS type %s on port %s: %s",
2992 type
, netdev
->name
, rte_strerror(error
));
2995 if (new_qos_conf
!= qos_conf
) {
2996 ovsrcu_set(&dev
->qos_conf
, new_qos_conf
);
2998 ovsrcu_postpone(qos_conf
->ops
->qos_destruct
, qos_conf
);
3002 ovs_mutex_unlock(&dev
->mutex
);
3007 /* egress-policer details */
3009 struct egress_policer
{
3010 struct qos_conf qos_conf
;
3011 struct rte_meter_srtcm_params app_srtcm_params
;
3012 struct rte_meter_srtcm egress_meter
;
3016 egress_policer_details_to_param(const struct smap
*details
,
3017 struct rte_meter_srtcm_params
*params
)
3019 memset(params
, 0, sizeof *params
);
3020 params
->cir
= smap_get_ullong(details
, "cir", 0);
3021 params
->cbs
= smap_get_ullong(details
, "cbs", 0);
3026 egress_policer_qos_construct(const struct smap
*details
,
3027 struct qos_conf
**conf
)
3029 struct egress_policer
*policer
;
3032 policer
= xmalloc(sizeof *policer
);
3033 qos_conf_init(&policer
->qos_conf
, &egress_policer_ops
);
3034 egress_policer_details_to_param(details
, &policer
->app_srtcm_params
);
3035 err
= rte_meter_srtcm_config(&policer
->egress_meter
,
3036 &policer
->app_srtcm_params
);
3038 *conf
= &policer
->qos_conf
;
3049 egress_policer_qos_destruct(struct qos_conf
*conf
)
3051 struct egress_policer
*policer
= CONTAINER_OF(conf
, struct egress_policer
,
3057 egress_policer_qos_get(const struct qos_conf
*conf
, struct smap
*details
)
3059 struct egress_policer
*policer
=
3060 CONTAINER_OF(conf
, struct egress_policer
, qos_conf
);
3062 smap_add_format(details
, "cir", "%"PRIu64
, policer
->app_srtcm_params
.cir
);
3063 smap_add_format(details
, "cbs", "%"PRIu64
, policer
->app_srtcm_params
.cbs
);
3069 egress_policer_qos_is_equal(const struct qos_conf
*conf
,
3070 const struct smap
*details
)
3072 struct egress_policer
*policer
=
3073 CONTAINER_OF(conf
, struct egress_policer
, qos_conf
);
3074 struct rte_meter_srtcm_params params
;
3076 egress_policer_details_to_param(details
, ¶ms
);
3078 return !memcmp(¶ms
, &policer
->app_srtcm_params
, sizeof params
);
3082 egress_policer_run(struct qos_conf
*conf
, struct rte_mbuf
**pkts
, int pkt_cnt
)
3085 struct egress_policer
*policer
=
3086 CONTAINER_OF(conf
, struct egress_policer
, qos_conf
);
3088 cnt
= netdev_dpdk_policer_run(&policer
->egress_meter
, pkts
, pkt_cnt
);
3093 static const struct dpdk_qos_ops egress_policer_ops
= {
3094 "egress-policer", /* qos_name */
3095 egress_policer_qos_construct
,
3096 egress_policer_qos_destruct
,
3097 egress_policer_qos_get
,
3098 egress_policer_qos_is_equal
,
3103 netdev_dpdk_reconfigure(struct netdev
*netdev
)
3105 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
3108 ovs_mutex_lock(&dev
->mutex
);
3110 if (netdev
->n_txq
== dev
->requested_n_txq
3111 && netdev
->n_rxq
== dev
->requested_n_rxq
3112 && dev
->mtu
== dev
->requested_mtu
3113 && dev
->rxq_size
== dev
->requested_rxq_size
3114 && dev
->txq_size
== dev
->requested_txq_size
3115 && dev
->socket_id
== dev
->requested_socket_id
) {
3116 /* Reconfiguration is unnecessary */
3121 rte_eth_dev_stop(dev
->port_id
);
3123 if (dev
->mtu
!= dev
->requested_mtu
3124 || dev
->socket_id
!= dev
->requested_socket_id
) {
3125 err
= netdev_dpdk_mempool_configure(dev
);
3131 netdev
->n_txq
= dev
->requested_n_txq
;
3132 netdev
->n_rxq
= dev
->requested_n_rxq
;
3134 dev
->rxq_size
= dev
->requested_rxq_size
;
3135 dev
->txq_size
= dev
->requested_txq_size
;
3137 rte_free(dev
->tx_q
);
3138 err
= dpdk_eth_dev_init(dev
);
3139 dev
->tx_q
= netdev_dpdk_alloc_txq(netdev
->n_txq
);
3144 netdev_change_seq_changed(netdev
);
3147 ovs_mutex_unlock(&dev
->mutex
);
3152 dpdk_vhost_reconfigure_helper(struct netdev_dpdk
*dev
)
3153 OVS_REQUIRES(dev
->mutex
)
3155 dev
->up
.n_txq
= dev
->requested_n_txq
;
3156 dev
->up
.n_rxq
= dev
->requested_n_rxq
;
3159 /* Enable TX queue 0 by default if it wasn't disabled. */
3160 if (dev
->tx_q
[0].map
== OVS_VHOST_QUEUE_MAP_UNKNOWN
) {
3161 dev
->tx_q
[0].map
= 0;
3164 netdev_dpdk_remap_txqs(dev
);
3166 if (dev
->requested_socket_id
!= dev
->socket_id
3167 || dev
->requested_mtu
!= dev
->mtu
) {
3168 err
= netdev_dpdk_mempool_configure(dev
);
3172 netdev_change_seq_changed(&dev
->up
);
3176 if (netdev_dpdk_get_vid(dev
) >= 0) {
3177 dev
->vhost_reconfigured
= true;
3184 netdev_dpdk_vhost_reconfigure(struct netdev
*netdev
)
3186 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
3189 ovs_mutex_lock(&dev
->mutex
);
3190 err
= dpdk_vhost_reconfigure_helper(dev
);
3191 ovs_mutex_unlock(&dev
->mutex
);
3197 netdev_dpdk_vhost_client_reconfigure(struct netdev
*netdev
)
3199 struct netdev_dpdk
*dev
= netdev_dpdk_cast(netdev
);
3202 ovs_mutex_lock(&dev
->mutex
);
3204 /* Configure vHost client mode if requested and if the following criteria
3206 * 1. Device hasn't been registered yet.
3207 * 2. A path has been specified.
3209 if (!(dev
->vhost_driver_flags
& RTE_VHOST_USER_CLIENT
)
3210 && strlen(dev
->vhost_id
)) {
3211 /* Register client-mode device */
3212 err
= rte_vhost_driver_register(dev
->vhost_id
,
3213 RTE_VHOST_USER_CLIENT
);
3215 VLOG_ERR("vhost-user device setup failure for device %s\n",
3219 /* Configuration successful */
3220 dev
->vhost_driver_flags
|= RTE_VHOST_USER_CLIENT
;
3221 VLOG_INFO("vHost User device '%s' created in 'client' mode, "
3222 "using client socket '%s'",
3223 dev
->up
.name
, dev
->vhost_id
);
3227 err
= dpdk_vhost_reconfigure_helper(dev
);
3230 ovs_mutex_unlock(&dev
->mutex
);
3235 #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, DESTRUCT, \
3236 SET_CONFIG, SET_TX_MULTIQ, SEND, \
3237 GET_CARRIER, GET_STATS, \
3238 GET_FEATURES, GET_STATUS, \
3239 RECONFIGURE, RXQ_RECV) \
3242 true, /* is_pmd */ \
3244 NULL, /* netdev_dpdk_run */ \
3245 NULL, /* netdev_dpdk_wait */ \
3247 netdev_dpdk_alloc, \
3250 netdev_dpdk_dealloc, \
3251 netdev_dpdk_get_config, \
3253 NULL, /* get_tunnel_config */ \
3254 NULL, /* build header */ \
3255 NULL, /* push header */ \
3256 NULL, /* pop header */ \
3257 netdev_dpdk_get_numa_id, /* get_numa_id */ \
3261 NULL, /* send_wait */ \
3263 netdev_dpdk_set_etheraddr, \
3264 netdev_dpdk_get_etheraddr, \
3265 netdev_dpdk_get_mtu, \
3266 netdev_dpdk_set_mtu, \
3267 netdev_dpdk_get_ifindex, \
3269 netdev_dpdk_get_carrier_resets, \
3270 netdev_dpdk_set_miimon, \
3273 NULL, /* set_advertisements */ \
3275 netdev_dpdk_set_policing, \
3276 netdev_dpdk_get_qos_types, \
3277 NULL, /* get_qos_capabilities */ \
3278 netdev_dpdk_get_qos, \
3279 netdev_dpdk_set_qos, \
3280 NULL, /* get_queue */ \
3281 NULL, /* set_queue */ \
3282 NULL, /* delete_queue */ \
3283 NULL, /* get_queue_stats */ \
3284 NULL, /* queue_dump_start */ \
3285 NULL, /* queue_dump_next */ \
3286 NULL, /* queue_dump_done */ \
3287 NULL, /* dump_queue_stats */ \
3289 NULL, /* set_in4 */ \
3290 NULL, /* get_addr_list */ \
3291 NULL, /* add_router */ \
3292 NULL, /* get_next_hop */ \
3294 NULL, /* arp_lookup */ \
3296 netdev_dpdk_update_flags, \
3299 netdev_dpdk_rxq_alloc, \
3300 netdev_dpdk_rxq_construct, \
3301 netdev_dpdk_rxq_destruct, \
3302 netdev_dpdk_rxq_dealloc, \
3304 NULL, /* rx_wait */ \
3305 NULL, /* rxq_drain */ \
3308 static const struct netdev_class dpdk_class
=
3311 netdev_dpdk_class_init
,
3312 netdev_dpdk_construct
,
3313 netdev_dpdk_destruct
,
3314 netdev_dpdk_set_config
,
3315 netdev_dpdk_set_tx_multiq
,
3316 netdev_dpdk_eth_send
,
3317 netdev_dpdk_get_carrier
,
3318 netdev_dpdk_get_stats
,
3319 netdev_dpdk_get_features
,
3320 netdev_dpdk_get_status
,
3321 netdev_dpdk_reconfigure
,
3322 netdev_dpdk_rxq_recv
);
3324 static const struct netdev_class dpdk_ring_class
=
3327 netdev_dpdk_class_init
,
3328 netdev_dpdk_ring_construct
,
3329 netdev_dpdk_destruct
,
3330 netdev_dpdk_ring_set_config
,
3331 netdev_dpdk_set_tx_multiq
,
3332 netdev_dpdk_ring_send
,
3333 netdev_dpdk_get_carrier
,
3334 netdev_dpdk_get_stats
,
3335 netdev_dpdk_get_features
,
3336 netdev_dpdk_get_status
,
3337 netdev_dpdk_reconfigure
,
3338 netdev_dpdk_rxq_recv
);
3340 static const struct netdev_class dpdk_vhost_class
=
3343 netdev_dpdk_vhost_class_init
,
3344 netdev_dpdk_vhost_construct
,
3345 netdev_dpdk_vhost_destruct
,
3348 netdev_dpdk_vhost_send
,
3349 netdev_dpdk_vhost_get_carrier
,
3350 netdev_dpdk_vhost_get_stats
,
3353 netdev_dpdk_vhost_reconfigure
,
3354 netdev_dpdk_vhost_rxq_recv
);
3355 static const struct netdev_class dpdk_vhost_client_class
=
3357 "dpdkvhostuserclient",
3358 netdev_dpdk_vhost_class_init
,
3359 netdev_dpdk_vhost_client_construct
,
3360 netdev_dpdk_vhost_destruct
,
3361 netdev_dpdk_vhost_client_set_config
,
3363 netdev_dpdk_vhost_send
,
3364 netdev_dpdk_vhost_get_carrier
,
3365 netdev_dpdk_vhost_get_stats
,
3368 netdev_dpdk_vhost_client_reconfigure
,
3369 netdev_dpdk_vhost_rxq_recv
);
3372 netdev_dpdk_register(void)
3374 netdev_register_provider(&dpdk_class
);
3375 netdev_register_provider(&dpdk_ring_class
);
3376 netdev_register_provider(&dpdk_vhost_class
);
3377 netdev_register_provider(&dpdk_vhost_client_class
);