2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 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 "dpif-netdev.h"
24 #include <netinet/in.h>
25 #include <sys/socket.h>
30 #include <sys/ioctl.h>
34 #include "classifier.h"
38 #include "dpif-provider.h"
40 #include "dynamic-string.h"
41 #include "fat-rwlock.h"
46 #include "meta-flow.h"
48 #include "netdev-dpdk.h"
49 #include "netdev-vport.h"
51 #include "odp-execute.h"
53 #include "ofp-print.h"
56 #include "packet-dpif.h"
58 #include "poll-loop.h"
68 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
70 /* By default, choose a priority in the middle. */
71 #define NETDEV_RULE_PRIORITY 0x8000
73 #define FLOW_DUMP_MAX_BATCH 50
74 /* Use per thread recirc_depth to prevent recirculation loop. */
75 #define MAX_RECIRC_DEPTH 5
76 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
78 /* Configuration parameters. */
79 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
82 enum { MAX_QUEUE_LEN = 128 }; /* Maximum number of packets per queue. */
83 enum { QUEUE_MASK = MAX_QUEUE_LEN - 1 };
84 BUILD_ASSERT_DECL(IS_POW2(MAX_QUEUE_LEN));
86 /* Protects against changes to 'dp_netdevs'. */
87 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
89 /* Contains all 'struct dp_netdev's. */
90 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
91 = SHASH_INITIALIZER(&dp_netdevs);
93 struct dp_netdev_upcall {
94 struct dpif_upcall upcall; /* Queued upcall information. */
95 struct ofpbuf buf; /* ofpbuf instance for upcall.packet. */
98 /* A queue passing packets from a struct dp_netdev to its clients (handlers).
104 * Any access at all requires the owning 'dp_netdev''s queue_rwlock and
106 struct dp_netdev_queue {
107 struct ovs_mutex mutex;
108 struct seq *seq; /* Incremented whenever a packet is queued. */
109 struct dp_netdev_upcall upcalls[MAX_QUEUE_LEN] OVS_GUARDED;
110 unsigned int head OVS_GUARDED;
111 unsigned int tail OVS_GUARDED;
114 /* Datapath based on the network device interface from netdev.h.
120 * Some members, marked 'const', are immutable. Accessing other members
121 * requires synchronization, as noted in more detail below.
123 * Acquisition order is, from outermost to innermost:
125 * dp_netdev_mutex (global)
131 const struct dpif_class *const class;
132 const char *const name;
133 struct ovs_refcount ref_cnt;
134 atomic_flag destroyed;
138 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
139 * changes to 'cls' must be made while still holding the 'flow_mutex'.
141 struct ovs_mutex flow_mutex;
142 struct classifier cls;
143 struct cmap flow_table OVS_GUARDED; /* Flow table. */
147 * 'queue_rwlock' protects the modification of 'handler_queues' and
148 * 'n_handlers'. The queue elements are protected by its
149 * 'handler_queues''s mutex. */
150 struct fat_rwlock queue_rwlock;
151 struct dp_netdev_queue *handler_queues;
156 * ovsthread_stats is internally synchronized. */
157 struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */
161 * Protected by RCU. Take the mutex to add or remove ports. */
162 struct ovs_mutex port_mutex;
164 struct seq *port_seq; /* Incremented whenever a port changes. */
166 /* Forwarding threads. */
167 struct latch exit_latch;
168 struct pmd_thread *pmd_threads;
169 size_t n_pmd_threads;
173 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
177 DP_STAT_HIT, /* Packets that matched in the flow table. */
178 DP_STAT_MISS, /* Packets that did not match. */
179 DP_STAT_LOST, /* Packets not passed up to the client. */
183 /* Contained by struct dp_netdev's 'stats' member. */
184 struct dp_netdev_stats {
185 struct ovs_mutex mutex; /* Protects 'n'. */
187 /* Indexed by DP_STAT_*, protected by 'mutex'. */
188 unsigned long long int n[DP_N_STATS] OVS_GUARDED;
192 /* A port in a netdev-based datapath. */
193 struct dp_netdev_port {
194 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
196 struct netdev *netdev;
197 struct netdev_saved_flags *sf;
198 struct netdev_rxq **rxq;
199 struct ovs_refcount ref_cnt;
200 char *type; /* Port type as requested by user. */
204 /* Stores a miniflow */
206 /* There are fields in the flow structure that we never use. Therefore we can
207 * save a few words of memory */
208 #define NETDEV_KEY_BUF_SIZE_U32 (FLOW_U32S - MINI_N_INLINE \
209 - FLOW_U32_SIZE(regs) \
210 - FLOW_U32_SIZE(metadata) \
212 struct netdev_flow_key {
213 struct miniflow flow;
214 uint32_t buf[NETDEV_KEY_BUF_SIZE_U32];
217 /* A flow in dp_netdev's 'flow_table'.
223 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
224 * its dp_netdev's classifier. The text below calls this classifier 'cls'.
229 * The thread safety rules described here for "struct dp_netdev_flow" are
230 * motivated by two goals:
232 * - Prevent threads that read members of "struct dp_netdev_flow" from
233 * reading bad data due to changes by some thread concurrently modifying
236 * - Prevent two threads making changes to members of a given "struct
237 * dp_netdev_flow" from interfering with each other.
243 * A flow 'flow' may be accessed without a risk of being freed by code that
244 * holds a read-lock or write-lock on 'cls->rwlock' or that owns a reference to
245 * 'flow->ref_cnt' (or both). Code that needs to hold onto a flow for a while
246 * should take 'cls->rwlock', find the flow it needs, increment 'flow->ref_cnt'
247 * with dpif_netdev_flow_ref(), and drop 'cls->rwlock'.
249 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
250 * flow from being deleted from 'cls' (that's 'cls->rwlock') and it doesn't
251 * protect members of 'flow' from modification.
253 * Some members, marked 'const', are immutable. Accessing other members
254 * requires synchronization, as noted in more detail below.
256 struct dp_netdev_flow {
257 /* Packet classification. */
258 const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
260 /* Hash table index by unmasked flow. */
261 const struct cmap_node node; /* In owning dp_netdev's 'flow_table'. */
262 const struct flow flow; /* The flow that created this entry. */
266 * Reading or writing these members requires 'mutex'. */
267 struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
270 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
273 static void dp_netdev_flow_free(struct dp_netdev_flow *);
275 /* Contained by struct dp_netdev_flow's 'stats' member. */
276 struct dp_netdev_flow_stats {
277 struct ovs_mutex mutex; /* Guards all the other members. */
279 long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
280 long long int packet_count OVS_GUARDED; /* Number of packets matched. */
281 long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
282 uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
285 /* A set of datapath actions within a "struct dp_netdev_flow".
291 * A struct dp_netdev_actions 'actions' is protected with RCU. */
292 struct dp_netdev_actions {
293 /* These members are immutable: they do not change during the struct's
295 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
296 unsigned int size; /* Size of 'actions', in bytes. */
299 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
301 struct dp_netdev_actions *dp_netdev_flow_get_actions(
302 const struct dp_netdev_flow *);
303 static void dp_netdev_actions_free(struct dp_netdev_actions *);
305 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
306 * the performance overhead of interrupt processing. Therefore netdev can
307 * not implement rx-wait for these devices. dpif-netdev needs to poll
308 * these device to check for recv buffer. pmd-thread does polling for
309 * devices assigned to itself thread.
311 * DPDK used PMD for accessing NIC.
313 * A thread that receives packets from PMD ports, looks them up in the flow
314 * table, and executes the actions it finds.
317 struct dp_netdev *dp;
320 atomic_uint change_seq;
323 /* Interface to netdev-based datapath. */
326 struct dp_netdev *dp;
327 uint64_t last_port_seq;
330 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
331 struct dp_netdev_port **portp);
332 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
333 struct dp_netdev_port **portp);
334 static void dp_netdev_free(struct dp_netdev *)
335 OVS_REQUIRES(dp_netdev_mutex);
336 static void dp_netdev_flow_flush(struct dp_netdev *);
337 static int do_add_port(struct dp_netdev *dp, const char *devname,
338 const char *type, odp_port_t port_no)
339 OVS_REQUIRES(dp->port_mutex);
340 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
341 OVS_REQUIRES(dp->port_mutex);
342 static void dp_netdev_destroy_all_queues(struct dp_netdev *dp)
343 OVS_REQ_WRLOCK(dp->queue_rwlock);
344 static int dpif_netdev_open(const struct dpif_class *, const char *name,
345 bool create, struct dpif **);
346 static int dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *,
347 int queue_no, int type,
348 const struct miniflow *,
349 const struct nlattr *userdata);
350 static void dp_netdev_execute_actions(struct dp_netdev *dp,
351 struct dpif_packet **, int c,
352 bool may_steal, struct pkt_metadata *,
353 const struct nlattr *actions,
355 static void dp_netdev_port_input(struct dp_netdev *dp,
356 struct dpif_packet **packets, int cnt,
359 static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
361 static struct dpif_netdev *
362 dpif_netdev_cast(const struct dpif *dpif)
364 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
365 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
368 static struct dp_netdev *
369 get_dp_netdev(const struct dpif *dpif)
371 return dpif_netdev_cast(dpif)->dp;
375 dpif_netdev_enumerate(struct sset *all_dps,
376 const struct dpif_class *dpif_class)
378 struct shash_node *node;
380 ovs_mutex_lock(&dp_netdev_mutex);
381 SHASH_FOR_EACH(node, &dp_netdevs) {
382 struct dp_netdev *dp = node->data;
383 if (dpif_class != dp->class) {
384 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
385 * If the class doesn't match, skip this dpif. */
388 sset_add(all_dps, node->name);
390 ovs_mutex_unlock(&dp_netdev_mutex);
396 dpif_netdev_class_is_dummy(const struct dpif_class *class)
398 return class != &dpif_netdev_class;
402 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
404 return strcmp(type, "internal") ? type
405 : dpif_netdev_class_is_dummy(class) ? "dummy"
410 create_dpif_netdev(struct dp_netdev *dp)
412 uint16_t netflow_id = hash_string(dp->name, 0);
413 struct dpif_netdev *dpif;
415 ovs_refcount_ref(&dp->ref_cnt);
417 dpif = xmalloc(sizeof *dpif);
418 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
420 dpif->last_port_seq = seq_read(dp->port_seq);
425 /* Choose an unused, non-zero port number and return it on success.
426 * Return ODPP_NONE on failure. */
428 choose_port(struct dp_netdev *dp, const char *name)
429 OVS_REQUIRES(dp->port_mutex)
433 if (dp->class != &dpif_netdev_class) {
437 /* If the port name begins with "br", start the number search at
438 * 100 to make writing tests easier. */
439 if (!strncmp(name, "br", 2)) {
443 /* If the port name contains a number, try to assign that port number.
444 * This can make writing unit tests easier because port numbers are
446 for (p = name; *p != '\0'; p++) {
447 if (isdigit((unsigned char) *p)) {
448 port_no = start_no + strtol(p, NULL, 10);
449 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
450 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
451 return u32_to_odp(port_no);
458 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
459 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
460 return u32_to_odp(port_no);
468 create_dp_netdev(const char *name, const struct dpif_class *class,
469 struct dp_netdev **dpp)
470 OVS_REQUIRES(dp_netdev_mutex)
472 struct dp_netdev *dp;
475 dp = xzalloc(sizeof *dp);
476 shash_add(&dp_netdevs, name, dp);
478 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
479 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
480 ovs_refcount_init(&dp->ref_cnt);
481 atomic_flag_clear(&dp->destroyed);
483 ovs_mutex_init(&dp->flow_mutex);
484 classifier_init(&dp->cls, NULL);
485 cmap_init(&dp->flow_table);
487 fat_rwlock_init(&dp->queue_rwlock);
489 ovsthread_stats_init(&dp->stats);
491 ovs_mutex_init(&dp->port_mutex);
492 cmap_init(&dp->ports);
493 dp->port_seq = seq_create();
494 latch_init(&dp->exit_latch);
496 ovs_mutex_lock(&dp->port_mutex);
497 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
498 ovs_mutex_unlock(&dp->port_mutex);
509 dpif_netdev_open(const struct dpif_class *class, const char *name,
510 bool create, struct dpif **dpifp)
512 struct dp_netdev *dp;
515 ovs_mutex_lock(&dp_netdev_mutex);
516 dp = shash_find_data(&dp_netdevs, name);
518 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
520 error = (dp->class != class ? EINVAL
525 *dpifp = create_dpif_netdev(dp);
527 ovs_mutex_unlock(&dp_netdev_mutex);
533 dp_netdev_purge_queues(struct dp_netdev *dp)
534 OVS_REQ_WRLOCK(dp->queue_rwlock)
538 for (i = 0; i < dp->n_handlers; i++) {
539 struct dp_netdev_queue *q = &dp->handler_queues[i];
541 ovs_mutex_lock(&q->mutex);
542 while (q->tail != q->head) {
543 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
544 ofpbuf_uninit(&u->upcall.packet);
545 ofpbuf_uninit(&u->buf);
547 ovs_mutex_unlock(&q->mutex);
551 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
552 * through the 'dp_netdevs' shash while freeing 'dp'. */
554 dp_netdev_free(struct dp_netdev *dp)
555 OVS_REQUIRES(dp_netdev_mutex)
557 struct dp_netdev_port *port;
558 struct dp_netdev_stats *bucket;
561 shash_find_and_delete(&dp_netdevs, dp->name);
563 dp_netdev_set_pmd_threads(dp, 0);
564 free(dp->pmd_threads);
566 dp_netdev_flow_flush(dp);
567 ovs_mutex_lock(&dp->port_mutex);
568 CMAP_FOR_EACH (port, node, &dp->ports) {
569 do_del_port(dp, port);
571 ovs_mutex_unlock(&dp->port_mutex);
573 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
574 ovs_mutex_destroy(&bucket->mutex);
575 free_cacheline(bucket);
577 ovsthread_stats_destroy(&dp->stats);
579 fat_rwlock_wrlock(&dp->queue_rwlock);
580 dp_netdev_destroy_all_queues(dp);
581 fat_rwlock_unlock(&dp->queue_rwlock);
583 fat_rwlock_destroy(&dp->queue_rwlock);
585 classifier_destroy(&dp->cls);
586 cmap_destroy(&dp->flow_table);
587 ovs_mutex_destroy(&dp->flow_mutex);
588 seq_destroy(dp->port_seq);
589 cmap_destroy(&dp->ports);
590 latch_destroy(&dp->exit_latch);
591 free(CONST_CAST(char *, dp->name));
596 dp_netdev_unref(struct dp_netdev *dp)
599 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
600 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
601 ovs_mutex_lock(&dp_netdev_mutex);
602 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
605 ovs_mutex_unlock(&dp_netdev_mutex);
610 dpif_netdev_close(struct dpif *dpif)
612 struct dp_netdev *dp = get_dp_netdev(dpif);
619 dpif_netdev_destroy(struct dpif *dpif)
621 struct dp_netdev *dp = get_dp_netdev(dpif);
623 if (!atomic_flag_test_and_set(&dp->destroyed)) {
624 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
625 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
634 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
636 struct dp_netdev *dp = get_dp_netdev(dpif);
637 struct dp_netdev_stats *bucket;
640 stats->n_flows = cmap_count(&dp->flow_table);
642 stats->n_hit = stats->n_missed = stats->n_lost = 0;
643 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
644 ovs_mutex_lock(&bucket->mutex);
645 stats->n_hit += bucket->n[DP_STAT_HIT];
646 stats->n_missed += bucket->n[DP_STAT_MISS];
647 stats->n_lost += bucket->n[DP_STAT_LOST];
648 ovs_mutex_unlock(&bucket->mutex);
650 stats->n_masks = UINT32_MAX;
651 stats->n_mask_hit = UINT64_MAX;
657 dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
661 for (i = 0; i < dp->n_pmd_threads; i++) {
662 struct pmd_thread *f = &dp->pmd_threads[i];
665 atomic_add(&f->change_seq, 1, &id);
670 hash_port_no(odp_port_t port_no)
672 return hash_int(odp_to_u32(port_no), 0);
676 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
678 OVS_REQUIRES(dp->port_mutex)
680 struct netdev_saved_flags *sf;
681 struct dp_netdev_port *port;
682 struct netdev *netdev;
683 enum netdev_flags flags;
684 const char *open_type;
688 /* XXX reject devices already in some dp_netdev. */
690 /* Open and validate network device. */
691 open_type = dpif_netdev_port_open_type(dp->class, type);
692 error = netdev_open(devname, open_type, &netdev);
696 /* XXX reject non-Ethernet devices */
698 netdev_get_flags(netdev, &flags);
699 if (flags & NETDEV_LOOPBACK) {
700 VLOG_ERR("%s: cannot add a loopback device", devname);
701 netdev_close(netdev);
705 port = xzalloc(sizeof *port);
706 port->port_no = port_no;
707 port->netdev = netdev;
708 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
709 port->type = xstrdup(type);
710 for (i = 0; i < netdev_n_rxq(netdev); i++) {
711 error = netdev_rxq_open(netdev, &port->rxq[i], i);
713 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
714 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
715 devname, ovs_strerror(errno));
716 netdev_close(netdev);
721 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
723 for (i = 0; i < netdev_n_rxq(netdev); i++) {
724 netdev_rxq_close(port->rxq[i]);
726 netdev_close(netdev);
733 if (netdev_is_pmd(netdev)) {
735 dp_netdev_set_pmd_threads(dp, NR_PMD_THREADS);
736 dp_netdev_reload_pmd_threads(dp);
738 ovs_refcount_init(&port->ref_cnt);
740 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
741 seq_change(dp->port_seq);
747 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
748 odp_port_t *port_nop)
750 struct dp_netdev *dp = get_dp_netdev(dpif);
751 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
752 const char *dpif_port;
756 ovs_mutex_lock(&dp->port_mutex);
757 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
758 if (*port_nop != ODPP_NONE) {
760 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
762 port_no = choose_port(dp, dpif_port);
763 error = port_no == ODPP_NONE ? EFBIG : 0;
767 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
769 ovs_mutex_unlock(&dp->port_mutex);
775 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
777 struct dp_netdev *dp = get_dp_netdev(dpif);
780 ovs_mutex_lock(&dp->port_mutex);
781 if (port_no == ODPP_LOCAL) {
784 struct dp_netdev_port *port;
786 error = get_port_by_number(dp, port_no, &port);
788 do_del_port(dp, port);
791 ovs_mutex_unlock(&dp->port_mutex);
797 is_valid_port_number(odp_port_t port_no)
799 return port_no != ODPP_NONE;
802 static struct dp_netdev_port *
803 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
805 struct dp_netdev_port *port;
807 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
808 if (port->port_no == port_no) {
816 get_port_by_number(struct dp_netdev *dp,
817 odp_port_t port_no, struct dp_netdev_port **portp)
819 if (!is_valid_port_number(port_no)) {
823 *portp = dp_netdev_lookup_port(dp, port_no);
824 return *portp ? 0 : ENOENT;
829 port_ref(struct dp_netdev_port *port)
832 ovs_refcount_ref(&port->ref_cnt);
837 port_destroy__(struct dp_netdev_port *port)
839 int n_rxq = netdev_n_rxq(port->netdev);
842 netdev_close(port->netdev);
843 netdev_restore_flags(port->sf);
845 for (i = 0; i < n_rxq; i++) {
846 netdev_rxq_close(port->rxq[i]);
854 port_unref(struct dp_netdev_port *port)
856 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
857 ovsrcu_postpone(port_destroy__, port);
862 get_port_by_name(struct dp_netdev *dp,
863 const char *devname, struct dp_netdev_port **portp)
864 OVS_REQUIRES(dp->port_mutex)
866 struct dp_netdev_port *port;
868 CMAP_FOR_EACH (port, node, &dp->ports) {
869 if (!strcmp(netdev_get_name(port->netdev), devname)) {
878 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
879 OVS_REQUIRES(dp->port_mutex)
881 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
882 seq_change(dp->port_seq);
883 if (netdev_is_pmd(port->netdev)) {
884 dp_netdev_reload_pmd_threads(dp);
891 answer_port_query(const struct dp_netdev_port *port,
892 struct dpif_port *dpif_port)
894 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
895 dpif_port->type = xstrdup(port->type);
896 dpif_port->port_no = port->port_no;
900 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
901 struct dpif_port *dpif_port)
903 struct dp_netdev *dp = get_dp_netdev(dpif);
904 struct dp_netdev_port *port;
907 error = get_port_by_number(dp, port_no, &port);
908 if (!error && dpif_port) {
909 answer_port_query(port, dpif_port);
916 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
917 struct dpif_port *dpif_port)
919 struct dp_netdev *dp = get_dp_netdev(dpif);
920 struct dp_netdev_port *port;
923 ovs_mutex_lock(&dp->port_mutex);
924 error = get_port_by_name(dp, devname, &port);
925 if (!error && dpif_port) {
926 answer_port_query(port, dpif_port);
928 ovs_mutex_unlock(&dp->port_mutex);
934 dp_netdev_flow_free(struct dp_netdev_flow *flow)
936 struct dp_netdev_flow_stats *bucket;
939 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
940 ovs_mutex_destroy(&bucket->mutex);
941 free_cacheline(bucket);
943 ovsthread_stats_destroy(&flow->stats);
945 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
946 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
951 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
952 OVS_REQUIRES(dp->flow_mutex)
954 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
955 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
957 classifier_remove(&dp->cls, cr);
958 cmap_remove(&dp->flow_table, node, flow_hash(&flow->flow, 0));
959 ovsrcu_postpone(dp_netdev_flow_free, flow);
963 dp_netdev_flow_flush(struct dp_netdev *dp)
965 struct dp_netdev_flow *netdev_flow;
967 ovs_mutex_lock(&dp->flow_mutex);
968 CMAP_FOR_EACH_SAFE (netdev_flow, node, &dp->flow_table) {
969 dp_netdev_remove_flow(dp, netdev_flow);
971 ovs_mutex_unlock(&dp->flow_mutex);
975 dpif_netdev_flow_flush(struct dpif *dpif)
977 struct dp_netdev *dp = get_dp_netdev(dpif);
979 dp_netdev_flow_flush(dp);
983 struct dp_netdev_port_state {
984 struct cmap_position position;
989 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
991 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
996 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
997 struct dpif_port *dpif_port)
999 struct dp_netdev_port_state *state = state_;
1000 struct dp_netdev *dp = get_dp_netdev(dpif);
1001 struct cmap_node *node;
1004 node = cmap_next_position(&dp->ports, &state->position);
1006 struct dp_netdev_port *port;
1008 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1011 state->name = xstrdup(netdev_get_name(port->netdev));
1012 dpif_port->name = state->name;
1013 dpif_port->type = port->type;
1014 dpif_port->port_no = port->port_no;
1025 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1027 struct dp_netdev_port_state *state = state_;
1034 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1036 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1037 uint64_t new_port_seq;
1040 new_port_seq = seq_read(dpif->dp->port_seq);
1041 if (dpif->last_port_seq != new_port_seq) {
1042 dpif->last_port_seq = new_port_seq;
1052 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1054 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1056 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1059 static struct dp_netdev_flow *
1060 dp_netdev_flow_cast(const struct cls_rule *cr)
1062 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1065 static struct dp_netdev_flow *
1066 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
1068 struct dp_netdev_flow *netdev_flow;
1069 struct cls_rule *rule;
1071 classifier_lookup_miniflow_batch(&dp->cls, &key, &rule, 1);
1072 netdev_flow = dp_netdev_flow_cast(rule);
1077 static struct dp_netdev_flow *
1078 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1080 struct dp_netdev_flow *netdev_flow;
1082 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1084 if (flow_equal(&netdev_flow->flow, flow)) {
1093 get_dpif_flow_stats(struct dp_netdev_flow *netdev_flow,
1094 struct dpif_flow_stats *stats)
1096 struct dp_netdev_flow_stats *bucket;
1099 memset(stats, 0, sizeof *stats);
1100 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1101 ovs_mutex_lock(&bucket->mutex);
1102 stats->n_packets += bucket->packet_count;
1103 stats->n_bytes += bucket->byte_count;
1104 stats->used = MAX(stats->used, bucket->used);
1105 stats->tcp_flags |= bucket->tcp_flags;
1106 ovs_mutex_unlock(&bucket->mutex);
1111 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1112 const struct nlattr *mask_key,
1113 uint32_t mask_key_len, const struct flow *flow,
1117 enum odp_key_fitness fitness;
1119 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1121 /* This should not happen: it indicates that
1122 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1123 * disagree on the acceptable form of a mask. Log the problem
1124 * as an error, with enough details to enable debugging. */
1125 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1127 if (!VLOG_DROP_ERR(&rl)) {
1131 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1133 VLOG_ERR("internal error parsing flow mask %s (%s)",
1134 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1141 enum mf_field_id id;
1142 /* No mask key, unwildcard everything except fields whose
1143 * prerequisities are not met. */
1144 memset(mask, 0x0, sizeof *mask);
1146 for (id = 0; id < MFF_N_IDS; ++id) {
1147 /* Skip registers and metadata. */
1148 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1149 && id != MFF_METADATA) {
1150 const struct mf_field *mf = mf_from_id(id);
1151 if (mf_are_prereqs_ok(mf, flow)) {
1152 mf_mask_field(mf, mask);
1158 /* Force unwildcard the in_port.
1160 * We need to do this even in the case where we unwildcard "everything"
1161 * above because "everything" only includes the 16-bit OpenFlow port number
1162 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1163 * port number mask->in_port.odp_port. */
1164 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1170 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1175 if (odp_flow_key_to_flow(key, key_len, flow)) {
1176 /* This should not happen: it indicates that odp_flow_key_from_flow()
1177 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1178 * flow. Log the problem as an error, with enough details to enable
1180 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1182 if (!VLOG_DROP_ERR(&rl)) {
1186 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1187 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1194 in_port = flow->in_port.odp_port;
1195 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1203 dpif_netdev_flow_get(const struct dpif *dpif,
1204 const struct nlattr *nl_key, size_t nl_key_len,
1205 struct ofpbuf **bufp,
1206 struct nlattr **maskp, size_t *mask_len,
1207 struct nlattr **actionsp, size_t *actions_len,
1208 struct dpif_flow_stats *stats)
1210 struct dp_netdev *dp = get_dp_netdev(dpif);
1211 struct dp_netdev_flow *netdev_flow;
1215 error = dpif_netdev_flow_from_nlattrs(nl_key, nl_key_len, &key);
1220 netdev_flow = dp_netdev_find_flow(dp, &key);
1224 get_dpif_flow_stats(netdev_flow, stats);
1228 struct flow_wildcards wc;
1230 *bufp = ofpbuf_new(sizeof(struct odputil_keybuf));
1231 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1232 odp_flow_key_from_mask(*bufp, &wc.masks, &netdev_flow->flow,
1233 odp_to_u32(wc.masks.in_port.odp_port),
1235 *maskp = ofpbuf_data(*bufp);
1236 *mask_len = ofpbuf_size(*bufp);
1239 struct dp_netdev_actions *actions;
1241 actions = dp_netdev_flow_get_actions(netdev_flow);
1242 *actionsp = actions->actions;
1243 *actions_len = actions->size;
1253 dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow,
1254 const struct flow_wildcards *wc,
1255 const struct nlattr *actions,
1257 OVS_REQUIRES(dp->flow_mutex)
1259 struct dp_netdev_flow *netdev_flow;
1262 netdev_flow = xzalloc(sizeof *netdev_flow);
1263 *CONST_CAST(struct flow *, &netdev_flow->flow) = *flow;
1265 ovsthread_stats_init(&netdev_flow->stats);
1267 ovsrcu_set(&netdev_flow->actions,
1268 dp_netdev_actions_create(actions, actions_len));
1270 match_init(&match, flow, wc);
1271 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1272 &match, NETDEV_RULE_PRIORITY);
1273 cmap_insert(&dp->flow_table,
1274 CONST_CAST(struct cmap_node *, &netdev_flow->node),
1275 flow_hash(flow, 0));
1276 classifier_insert(&dp->cls,
1277 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1283 clear_stats(struct dp_netdev_flow *netdev_flow)
1285 struct dp_netdev_flow_stats *bucket;
1288 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1289 ovs_mutex_lock(&bucket->mutex);
1291 bucket->packet_count = 0;
1292 bucket->byte_count = 0;
1293 bucket->tcp_flags = 0;
1294 ovs_mutex_unlock(&bucket->mutex);
1299 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1301 struct dp_netdev *dp = get_dp_netdev(dpif);
1302 struct dp_netdev_flow *netdev_flow;
1304 struct miniflow miniflow;
1305 struct flow_wildcards wc;
1308 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &flow);
1312 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1313 put->mask, put->mask_len,
1318 miniflow_init(&miniflow, &flow);
1320 ovs_mutex_lock(&dp->flow_mutex);
1321 netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
1323 if (put->flags & DPIF_FP_CREATE) {
1324 if (cmap_count(&dp->flow_table) < MAX_FLOWS) {
1326 memset(put->stats, 0, sizeof *put->stats);
1328 error = dp_netdev_flow_add(dp, &flow, &wc, put->actions,
1337 if (put->flags & DPIF_FP_MODIFY
1338 && flow_equal(&flow, &netdev_flow->flow)) {
1339 struct dp_netdev_actions *new_actions;
1340 struct dp_netdev_actions *old_actions;
1342 new_actions = dp_netdev_actions_create(put->actions,
1345 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1346 ovsrcu_set(&netdev_flow->actions, new_actions);
1349 get_dpif_flow_stats(netdev_flow, put->stats);
1351 if (put->flags & DPIF_FP_ZERO_STATS) {
1352 clear_stats(netdev_flow);
1355 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1356 } else if (put->flags & DPIF_FP_CREATE) {
1359 /* Overlapping flow. */
1363 ovs_mutex_unlock(&dp->flow_mutex);
1364 miniflow_destroy(&miniflow);
1370 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1372 struct dp_netdev *dp = get_dp_netdev(dpif);
1373 struct dp_netdev_flow *netdev_flow;
1377 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1382 ovs_mutex_lock(&dp->flow_mutex);
1383 netdev_flow = dp_netdev_find_flow(dp, &key);
1386 get_dpif_flow_stats(netdev_flow, del->stats);
1388 dp_netdev_remove_flow(dp, netdev_flow);
1392 ovs_mutex_unlock(&dp->flow_mutex);
1397 struct dpif_netdev_flow_dump {
1398 struct dpif_flow_dump up;
1399 struct cmap_position pos;
1401 struct ovs_mutex mutex;
1404 static struct dpif_netdev_flow_dump *
1405 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1407 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1410 static struct dpif_flow_dump *
1411 dpif_netdev_flow_dump_create(const struct dpif *dpif_)
1413 struct dpif_netdev_flow_dump *dump;
1415 dump = xmalloc(sizeof *dump);
1416 dpif_flow_dump_init(&dump->up, dpif_);
1417 memset(&dump->pos, 0, sizeof dump->pos);
1419 ovs_mutex_init(&dump->mutex);
1425 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1427 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1429 ovs_mutex_destroy(&dump->mutex);
1434 struct dpif_netdev_flow_dump_thread {
1435 struct dpif_flow_dump_thread up;
1436 struct dpif_netdev_flow_dump *dump;
1437 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1438 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1441 static struct dpif_netdev_flow_dump_thread *
1442 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1444 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1447 static struct dpif_flow_dump_thread *
1448 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1450 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1451 struct dpif_netdev_flow_dump_thread *thread;
1453 thread = xmalloc(sizeof *thread);
1454 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1455 thread->dump = dump;
1460 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1462 struct dpif_netdev_flow_dump_thread *thread
1463 = dpif_netdev_flow_dump_thread_cast(thread_);
1469 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
1470 struct dpif_flow *flows, int max_flows)
1472 struct dpif_netdev_flow_dump_thread *thread
1473 = dpif_netdev_flow_dump_thread_cast(thread_);
1474 struct dpif_netdev_flow_dump *dump = thread->dump;
1475 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
1476 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
1477 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
1481 ovs_mutex_lock(&dump->mutex);
1482 if (!dump->status) {
1483 for (n_flows = 0; n_flows < MIN(max_flows, FLOW_DUMP_MAX_BATCH);
1485 struct cmap_node *node;
1487 node = cmap_next_position(&dp->flow_table, &dump->pos);
1492 netdev_flows[n_flows] = CONTAINER_OF(node, struct dp_netdev_flow,
1496 ovs_mutex_unlock(&dump->mutex);
1498 for (i = 0; i < n_flows; i++) {
1499 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
1500 struct odputil_keybuf *keybuf = &thread->keybuf[i];
1501 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
1502 struct dpif_flow *f = &flows[i];
1503 struct dp_netdev_actions *dp_actions;
1504 struct flow_wildcards wc;
1507 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1510 ofpbuf_use_stack(&buf, keybuf, sizeof *keybuf);
1511 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1512 netdev_flow->flow.in_port.odp_port, true);
1513 f->key = ofpbuf_data(&buf);
1514 f->key_len = ofpbuf_size(&buf);
1517 ofpbuf_use_stack(&buf, maskbuf, sizeof *maskbuf);
1518 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1519 odp_to_u32(wc.masks.in_port.odp_port),
1521 f->mask = ofpbuf_data(&buf);
1522 f->mask_len = ofpbuf_size(&buf);
1525 dp_actions = dp_netdev_flow_get_actions(netdev_flow);
1526 f->actions = dp_actions->actions;
1527 f->actions_len = dp_actions->size;
1530 get_dpif_flow_stats(netdev_flow, &f->stats);
1537 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1539 struct dp_netdev *dp = get_dp_netdev(dpif);
1540 struct dpif_packet packet, *pp;
1541 struct pkt_metadata *md = &execute->md;
1543 if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
1544 ofpbuf_size(execute->packet) > UINT16_MAX) {
1548 packet.ofpbuf = *execute->packet;
1551 dp_netdev_execute_actions(dp, &pp, 1, false, md,
1552 execute->actions, execute->actions_len);
1554 /* Even though may_steal is set to false, some actions could modify or
1555 * reallocate the ofpbuf memory. We need to pass those changes to the
1557 *execute->packet = packet.ofpbuf;
1563 dp_netdev_destroy_all_queues(struct dp_netdev *dp)
1564 OVS_REQ_WRLOCK(dp->queue_rwlock)
1568 dp_netdev_purge_queues(dp);
1570 for (i = 0; i < dp->n_handlers; i++) {
1571 struct dp_netdev_queue *q = &dp->handler_queues[i];
1573 ovs_mutex_destroy(&q->mutex);
1574 seq_destroy(q->seq);
1576 free(dp->handler_queues);
1577 dp->handler_queues = NULL;
1582 dp_netdev_refresh_queues(struct dp_netdev *dp, uint32_t n_handlers)
1583 OVS_REQ_WRLOCK(dp->queue_rwlock)
1585 if (dp->n_handlers != n_handlers) {
1588 dp_netdev_destroy_all_queues(dp);
1590 dp->n_handlers = n_handlers;
1591 dp->handler_queues = xzalloc(n_handlers * sizeof *dp->handler_queues);
1593 for (i = 0; i < n_handlers; i++) {
1594 struct dp_netdev_queue *q = &dp->handler_queues[i];
1596 ovs_mutex_init(&q->mutex);
1597 q->seq = seq_create();
1603 dpif_netdev_recv_set(struct dpif *dpif, bool enable)
1605 struct dp_netdev *dp = get_dp_netdev(dpif);
1607 if ((dp->handler_queues != NULL) == enable) {
1611 fat_rwlock_wrlock(&dp->queue_rwlock);
1613 dp_netdev_destroy_all_queues(dp);
1615 dp_netdev_refresh_queues(dp, 1);
1617 fat_rwlock_unlock(&dp->queue_rwlock);
1623 dpif_netdev_handlers_set(struct dpif *dpif, uint32_t n_handlers)
1625 struct dp_netdev *dp = get_dp_netdev(dpif);
1627 fat_rwlock_wrlock(&dp->queue_rwlock);
1628 if (dp->handler_queues) {
1629 dp_netdev_refresh_queues(dp, n_handlers);
1631 fat_rwlock_unlock(&dp->queue_rwlock);
1637 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1638 uint32_t queue_id, uint32_t *priority)
1640 *priority = queue_id;
1645 dp_netdev_recv_check(const struct dp_netdev *dp, const uint32_t handler_id)
1646 OVS_REQ_RDLOCK(dp->queue_rwlock)
1648 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1650 if (!dp->handler_queues) {
1651 VLOG_WARN_RL(&rl, "receiving upcall disabled");
1655 if (handler_id >= dp->n_handlers) {
1656 VLOG_WARN_RL(&rl, "handler index out of bound");
1664 dpif_netdev_recv(struct dpif *dpif, uint32_t handler_id,
1665 struct dpif_upcall *upcall, struct ofpbuf *buf)
1667 struct dp_netdev *dp = get_dp_netdev(dpif);
1668 struct dp_netdev_queue *q;
1671 fat_rwlock_rdlock(&dp->queue_rwlock);
1673 if (!dp_netdev_recv_check(dp, handler_id)) {
1678 q = &dp->handler_queues[handler_id];
1679 ovs_mutex_lock(&q->mutex);
1680 if (q->head != q->tail) {
1681 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
1683 *upcall = u->upcall;
1690 ovs_mutex_unlock(&q->mutex);
1693 fat_rwlock_unlock(&dp->queue_rwlock);
1699 dpif_netdev_recv_wait(struct dpif *dpif, uint32_t handler_id)
1701 struct dp_netdev *dp = get_dp_netdev(dpif);
1702 struct dp_netdev_queue *q;
1705 fat_rwlock_rdlock(&dp->queue_rwlock);
1707 if (!dp_netdev_recv_check(dp, handler_id)) {
1711 q = &dp->handler_queues[handler_id];
1712 ovs_mutex_lock(&q->mutex);
1713 seq = seq_read(q->seq);
1714 if (q->head != q->tail) {
1715 poll_immediate_wake();
1717 seq_wait(q->seq, seq);
1720 ovs_mutex_unlock(&q->mutex);
1723 fat_rwlock_unlock(&dp->queue_rwlock);
1727 dpif_netdev_recv_purge(struct dpif *dpif)
1729 struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif);
1731 fat_rwlock_wrlock(&dpif_netdev->dp->queue_rwlock);
1732 dp_netdev_purge_queues(dpif_netdev->dp);
1733 fat_rwlock_unlock(&dpif_netdev->dp->queue_rwlock);
1736 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1737 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1739 struct dp_netdev_actions *
1740 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1742 struct dp_netdev_actions *netdev_actions;
1744 netdev_actions = xmalloc(sizeof *netdev_actions);
1745 netdev_actions->actions = xmemdup(actions, size);
1746 netdev_actions->size = size;
1748 return netdev_actions;
1751 struct dp_netdev_actions *
1752 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
1754 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
1758 dp_netdev_actions_free(struct dp_netdev_actions *actions)
1760 free(actions->actions);
1766 dp_netdev_process_rxq_port(struct dp_netdev *dp,
1767 struct dp_netdev_port *port,
1768 struct netdev_rxq *rxq)
1770 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1773 error = netdev_rxq_recv(rxq, packets, &cnt);
1775 dp_netdev_port_input(dp, packets, cnt, port->port_no);
1776 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1777 static struct vlog_rate_limit rl
1778 = VLOG_RATE_LIMIT_INIT(1, 5);
1780 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1781 netdev_get_name(port->netdev),
1782 ovs_strerror(error));
1787 dpif_netdev_run(struct dpif *dpif)
1789 struct dp_netdev_port *port;
1790 struct dp_netdev *dp = get_dp_netdev(dpif);
1792 CMAP_FOR_EACH (port, node, &dp->ports) {
1793 if (!netdev_is_pmd(port->netdev)) {
1796 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1797 dp_netdev_process_rxq_port(dp, port, port->rxq[i]);
1804 dpif_netdev_wait(struct dpif *dpif)
1806 struct dp_netdev_port *port;
1807 struct dp_netdev *dp = get_dp_netdev(dpif);
1809 ovs_mutex_lock(&dp_netdev_mutex);
1810 CMAP_FOR_EACH (port, node, &dp->ports) {
1811 if (!netdev_is_pmd(port->netdev)) {
1814 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1815 netdev_rxq_wait(port->rxq[i]);
1819 ovs_mutex_unlock(&dp_netdev_mutex);
1823 struct dp_netdev_port *port;
1824 struct netdev_rxq *rx;
1828 pmd_load_queues(struct pmd_thread *f,
1829 struct rxq_poll **ppoll_list, int poll_cnt)
1831 struct dp_netdev *dp = f->dp;
1832 struct rxq_poll *poll_list = *ppoll_list;
1833 struct dp_netdev_port *port;
1838 /* Simple scheduler for netdev rx polling. */
1839 for (i = 0; i < poll_cnt; i++) {
1840 port_unref(poll_list[i].port);
1846 CMAP_FOR_EACH (port, node, &f->dp->ports) {
1847 if (netdev_is_pmd(port->netdev)) {
1850 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1851 if ((index % dp->n_pmd_threads) == id) {
1852 poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1));
1855 poll_list[poll_cnt].port = port;
1856 poll_list[poll_cnt].rx = port->rxq[i];
1864 *ppoll_list = poll_list;
1869 pmd_thread_main(void *f_)
1871 struct pmd_thread *f = f_;
1872 struct dp_netdev *dp = f->dp;
1873 unsigned int lc = 0;
1874 struct rxq_poll *poll_list;
1875 unsigned int port_seq;
1882 pmd_thread_setaffinity_cpu(f->id);
1884 poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
1885 atomic_read(&f->change_seq, &port_seq);
1888 unsigned int c_port_seq;
1891 for (i = 0; i < poll_cnt; i++) {
1892 dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx);
1898 /* TODO: need completely userspace based signaling method.
1899 * to keep this thread entirely in userspace.
1900 * For now using atomic counter. */
1902 atomic_read_explicit(&f->change_seq, &c_port_seq, memory_order_consume);
1903 if (c_port_seq != port_seq) {
1909 if (!latch_is_set(&f->dp->exit_latch)){
1913 for (i = 0; i < poll_cnt; i++) {
1914 port_unref(poll_list[i].port);
1922 dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n)
1926 if (n == dp->n_pmd_threads) {
1930 /* Stop existing threads. */
1931 latch_set(&dp->exit_latch);
1932 dp_netdev_reload_pmd_threads(dp);
1933 for (i = 0; i < dp->n_pmd_threads; i++) {
1934 struct pmd_thread *f = &dp->pmd_threads[i];
1936 xpthread_join(f->thread, NULL);
1938 latch_poll(&dp->exit_latch);
1939 free(dp->pmd_threads);
1941 /* Start new threads. */
1942 dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
1943 dp->n_pmd_threads = n;
1945 for (i = 0; i < n; i++) {
1946 struct pmd_thread *f = &dp->pmd_threads[i];
1950 atomic_store(&f->change_seq, 1);
1952 /* Each thread will distribute all devices rx-queues among
1954 f->thread = ovs_thread_create("pmd", pmd_thread_main, f);
1960 dp_netdev_flow_stats_new_cb(void)
1962 struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1963 ovs_mutex_init(&bucket->mutex);
1968 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
1972 long long int now = time_msec();
1973 struct dp_netdev_flow_stats *bucket;
1975 bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
1976 dp_netdev_flow_stats_new_cb);
1978 ovs_mutex_lock(&bucket->mutex);
1979 bucket->used = MAX(now, bucket->used);
1980 bucket->packet_count += cnt;
1981 bucket->byte_count += size;
1982 bucket->tcp_flags |= tcp_flags;
1983 ovs_mutex_unlock(&bucket->mutex);
1987 dp_netdev_stats_new_cb(void)
1989 struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1990 ovs_mutex_init(&bucket->mutex);
1995 dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type, int cnt)
1997 struct dp_netdev_stats *bucket;
1999 bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
2000 ovs_mutex_lock(&bucket->mutex);
2001 bucket->n[type] += cnt;
2002 ovs_mutex_unlock(&bucket->mutex);
2005 struct packet_batch {
2006 unsigned int packet_count;
2007 unsigned int byte_count;
2010 struct dp_netdev_flow *flow;
2012 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
2013 struct pkt_metadata md;
2017 packet_batch_update(struct packet_batch *batch,
2018 struct dpif_packet *packet, const struct miniflow *mf)
2020 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
2021 batch->packets[batch->packet_count++] = packet;
2022 batch->byte_count += ofpbuf_size(&packet->ofpbuf);
2026 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow,
2027 struct pkt_metadata *md)
2032 batch->packet_count = 0;
2033 batch->byte_count = 0;
2034 batch->tcp_flags = 0;
2038 packet_batch_execute(struct packet_batch *batch, struct dp_netdev *dp)
2040 struct dp_netdev_actions *actions;
2041 struct dp_netdev_flow *flow = batch->flow;
2043 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
2046 actions = dp_netdev_flow_get_actions(flow);
2048 dp_netdev_execute_actions(dp, batch->packets,
2049 batch->packet_count, true, &batch->md,
2050 actions->actions, actions->size);
2052 dp_netdev_count_packet(dp, DP_STAT_HIT, batch->packet_count);
2056 dp_netdev_input(struct dp_netdev *dp, struct dpif_packet **packets, int cnt,
2057 struct pkt_metadata *md)
2059 struct packet_batch batches[NETDEV_MAX_RX_BATCH];
2060 struct netdev_flow_key keys[NETDEV_MAX_RX_BATCH];
2061 const struct miniflow *mfs[NETDEV_MAX_RX_BATCH]; /* NULL at bad packets. */
2062 struct cls_rule *rules[NETDEV_MAX_RX_BATCH];
2063 size_t n_batches, i;
2065 for (i = 0; i < cnt; i++) {
2066 if (OVS_UNLIKELY(ofpbuf_size(&packets[i]->ofpbuf) < ETH_HEADER_LEN)) {
2067 dpif_packet_delete(packets[i]);
2072 miniflow_initialize(&keys[i].flow, keys[i].buf);
2073 miniflow_extract(&packets[i]->ofpbuf, md, &keys[i].flow);
2074 mfs[i] = &keys[i].flow;
2077 classifier_lookup_miniflow_batch(&dp->cls, mfs, rules, cnt);
2080 for (i = 0; i < cnt; i++) {
2081 struct dp_netdev_flow *flow;
2082 struct packet_batch *batch;
2085 if (OVS_UNLIKELY(!mfs[i])) {
2089 if (OVS_UNLIKELY(!rules[i])) {
2091 dp_netdev_count_packet(dp, DP_STAT_MISS, 1);
2093 if (OVS_LIKELY(dp->handler_queues)) {
2094 uint32_t hash = miniflow_hash_5tuple(mfs[i], 0);
2095 struct ofpbuf *buf = &packets[i]->ofpbuf;
2097 dp_netdev_output_userspace(dp, buf, hash % dp->n_handlers,
2098 DPIF_UC_MISS, mfs[i], NULL);
2101 dpif_packet_delete(packets[i]);
2105 /* XXX: This O(n^2) algortihm makes sense if we're operating under the
2106 * assumption that the number of distinct flows (and therefore the
2107 * number of distinct batches) is quite small. If this turns out not
2108 * to be the case, it may make sense to pre sort based on the
2109 * netdev_flow pointer. That done we can get the appropriate batching
2110 * in O(n * log(n)) instead. */
2112 flow = dp_netdev_flow_cast(rules[i]);
2113 for (j = 0; j < n_batches; j++) {
2114 if (batches[j].flow == flow) {
2115 batch = &batches[j];
2121 batch = &batches[n_batches++];
2122 packet_batch_init(batch, flow, md);
2124 packet_batch_update(batch, packets[i], mfs[i]);
2127 for (i = 0; i < n_batches; i++) {
2128 packet_batch_execute(&batches[i], dp);
2133 dp_netdev_port_input(struct dp_netdev *dp, struct dpif_packet **packets,
2134 int cnt, odp_port_t port_no)
2136 uint32_t *recirc_depth = recirc_depth_get();
2137 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port_no);
2140 dp_netdev_input(dp, packets, cnt, &md);
2144 dp_netdev_queue_userspace_packet(struct dp_netdev_queue *q,
2145 struct ofpbuf *packet, int type,
2146 const struct miniflow *key,
2147 const struct nlattr *userdata)
2148 OVS_REQUIRES(q->mutex)
2150 if (q->head - q->tail < MAX_QUEUE_LEN) {
2151 struct dp_netdev_upcall *u = &q->upcalls[q->head++ & QUEUE_MASK];
2152 struct dpif_upcall *upcall = &u->upcall;
2153 struct ofpbuf *buf = &u->buf;
2158 upcall->type = type;
2160 /* Allocate buffer big enough for everything. */
2161 buf_size = ODPUTIL_FLOW_KEY_BYTES;
2163 buf_size += NLA_ALIGN(userdata->nla_len);
2165 buf_size += ofpbuf_size(packet);
2166 ofpbuf_init(buf, buf_size);
2169 miniflow_expand(key, &flow);
2170 odp_flow_key_from_flow(buf, &flow, NULL, flow.in_port.odp_port, true);
2171 upcall->key = ofpbuf_data(buf);
2172 upcall->key_len = ofpbuf_size(buf);
2176 upcall->userdata = ofpbuf_put(buf, userdata,
2177 NLA_ALIGN(userdata->nla_len));
2180 /* We have to perform a copy of the packet, because we cannot send DPDK
2181 * mbufs to a non pmd thread. When the upcall processing will be done
2182 * in the pmd thread, this copy can be avoided */
2183 data = ofpbuf_put(buf, ofpbuf_data(packet), ofpbuf_size(packet));
2184 ofpbuf_use_stub(&upcall->packet, data, ofpbuf_size(packet));
2185 ofpbuf_set_size(&upcall->packet, ofpbuf_size(packet));
2196 dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet,
2197 int queue_no, int type,
2198 const struct miniflow *key,
2199 const struct nlattr *userdata)
2201 struct dp_netdev_queue *q;
2204 fat_rwlock_rdlock(&dp->queue_rwlock);
2205 q = &dp->handler_queues[queue_no];
2206 ovs_mutex_lock(&q->mutex);
2207 error = dp_netdev_queue_userspace_packet(q, packet, type, key,
2209 if (error == ENOBUFS) {
2210 dp_netdev_count_packet(dp, DP_STAT_LOST, 1);
2212 ovs_mutex_unlock(&q->mutex);
2213 fat_rwlock_unlock(&dp->queue_rwlock);
2218 struct dp_netdev_execute_aux {
2219 struct dp_netdev *dp;
2223 dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
2224 struct pkt_metadata *md,
2225 const struct nlattr *a, bool may_steal)
2226 OVS_NO_THREAD_SAFETY_ANALYSIS
2228 struct dp_netdev_execute_aux *aux = aux_;
2229 int type = nl_attr_type(a);
2230 struct dp_netdev_port *p;
2231 uint32_t *depth = recirc_depth_get();
2234 switch ((enum ovs_action_attr)type) {
2235 case OVS_ACTION_ATTR_OUTPUT:
2236 p = dp_netdev_lookup_port(aux->dp, u32_to_odp(nl_attr_get_u32(a)));
2237 if (OVS_LIKELY(p)) {
2238 netdev_send(p->netdev, packets, cnt, may_steal);
2239 } else if (may_steal) {
2240 for (i = 0; i < cnt; i++) {
2241 dpif_packet_delete(packets[i]);
2246 case OVS_ACTION_ATTR_USERSPACE: {
2247 const struct nlattr *userdata;
2248 struct netdev_flow_key key;
2250 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2252 miniflow_initialize(&key.flow, key.buf);
2254 for (i = 0; i < cnt; i++) {
2255 struct ofpbuf *packet;
2257 packet = &packets[i]->ofpbuf;
2259 miniflow_extract(packet, md, &key.flow);
2261 dp_netdev_output_userspace(aux->dp, packet,
2262 miniflow_hash_5tuple(&key.flow, 0)
2263 % aux->dp->n_handlers,
2264 DPIF_UC_ACTION, &key.flow,
2267 dpif_packet_delete(packets[i]);
2273 case OVS_ACTION_ATTR_HASH: {
2274 const struct ovs_action_hash *hash_act;
2275 struct netdev_flow_key key;
2278 hash_act = nl_attr_get(a);
2280 miniflow_initialize(&key.flow, key.buf);
2282 for (i = 0; i < cnt; i++) {
2284 /* TODO: this is slow. Use RSS hash in the future */
2285 miniflow_extract(&packets[i]->ofpbuf, md, &key.flow);
2287 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
2288 /* Hash need not be symmetric, nor does it need to include
2290 hash = miniflow_hash_5tuple(&key.flow, hash_act->hash_basis);
2292 VLOG_WARN("Unknown hash algorithm specified "
2293 "for the hash action.");
2298 hash = 1; /* 0 is not valid */
2304 packets[i]->dp_hash = hash;
2309 case OVS_ACTION_ATTR_RECIRC:
2310 if (*depth < MAX_RECIRC_DEPTH) {
2313 for (i = 0; i < cnt; i++) {
2314 struct dpif_packet *recirc_pkt;
2315 struct pkt_metadata recirc_md = *md;
2317 recirc_pkt = (may_steal) ? packets[i]
2318 : dpif_packet_clone(packets[i]);
2320 recirc_md.recirc_id = nl_attr_get_u32(a);
2322 /* Hash is private to each packet */
2323 recirc_md.dp_hash = packets[i]->dp_hash;
2325 dp_netdev_input(aux->dp, &recirc_pkt, 1, &recirc_md);
2331 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
2333 for (i = 0; i < cnt; i++) {
2334 dpif_packet_delete(packets[i]);
2340 case OVS_ACTION_ATTR_PUSH_VLAN:
2341 case OVS_ACTION_ATTR_POP_VLAN:
2342 case OVS_ACTION_ATTR_PUSH_MPLS:
2343 case OVS_ACTION_ATTR_POP_MPLS:
2344 case OVS_ACTION_ATTR_SET:
2345 case OVS_ACTION_ATTR_SAMPLE:
2346 case OVS_ACTION_ATTR_UNSPEC:
2347 case __OVS_ACTION_ATTR_MAX:
2353 dp_netdev_execute_actions(struct dp_netdev *dp,
2354 struct dpif_packet **packets, int cnt,
2355 bool may_steal, struct pkt_metadata *md,
2356 const struct nlattr *actions, size_t actions_len)
2358 struct dp_netdev_execute_aux aux = {dp};
2360 odp_execute_actions(&aux, packets, cnt, may_steal, md, actions,
2361 actions_len, dp_execute_cb);
2364 const struct dpif_class dpif_netdev_class = {
2366 dpif_netdev_enumerate,
2367 dpif_netdev_port_open_type,
2370 dpif_netdev_destroy,
2373 dpif_netdev_get_stats,
2374 dpif_netdev_port_add,
2375 dpif_netdev_port_del,
2376 dpif_netdev_port_query_by_number,
2377 dpif_netdev_port_query_by_name,
2378 NULL, /* port_get_pid */
2379 dpif_netdev_port_dump_start,
2380 dpif_netdev_port_dump_next,
2381 dpif_netdev_port_dump_done,
2382 dpif_netdev_port_poll,
2383 dpif_netdev_port_poll_wait,
2384 dpif_netdev_flow_get,
2385 dpif_netdev_flow_put,
2386 dpif_netdev_flow_del,
2387 dpif_netdev_flow_flush,
2388 dpif_netdev_flow_dump_create,
2389 dpif_netdev_flow_dump_destroy,
2390 dpif_netdev_flow_dump_thread_create,
2391 dpif_netdev_flow_dump_thread_destroy,
2392 dpif_netdev_flow_dump_next,
2393 dpif_netdev_execute,
2395 dpif_netdev_recv_set,
2396 dpif_netdev_handlers_set,
2397 dpif_netdev_queue_to_priority,
2399 dpif_netdev_recv_wait,
2400 dpif_netdev_recv_purge,
2404 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2405 const char *argv[], void *aux OVS_UNUSED)
2407 struct dp_netdev_port *old_port;
2408 struct dp_netdev_port *new_port;
2409 struct dp_netdev *dp;
2412 ovs_mutex_lock(&dp_netdev_mutex);
2413 dp = shash_find_data(&dp_netdevs, argv[1]);
2414 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2415 ovs_mutex_unlock(&dp_netdev_mutex);
2416 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2419 ovs_refcount_ref(&dp->ref_cnt);
2420 ovs_mutex_unlock(&dp_netdev_mutex);
2422 ovs_mutex_lock(&dp->port_mutex);
2423 if (get_port_by_name(dp, argv[2], &old_port)) {
2424 unixctl_command_reply_error(conn, "unknown port");
2428 port_no = u32_to_odp(atoi(argv[3]));
2429 if (!port_no || port_no == ODPP_NONE) {
2430 unixctl_command_reply_error(conn, "bad port number");
2433 if (dp_netdev_lookup_port(dp, port_no)) {
2434 unixctl_command_reply_error(conn, "port number already in use");
2438 /* Remove old port. */
2439 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
2440 ovsrcu_postpone(free, old_port);
2442 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
2443 new_port = xmemdup(old_port, sizeof *old_port);
2444 new_port->port_no = port_no;
2445 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
2447 seq_change(dp->port_seq);
2448 unixctl_command_reply(conn, NULL);
2451 ovs_mutex_unlock(&dp->port_mutex);
2452 dp_netdev_unref(dp);
2456 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
2457 const char *argv[], void *aux OVS_UNUSED)
2459 struct dp_netdev_port *port;
2460 struct dp_netdev *dp;
2462 ovs_mutex_lock(&dp_netdev_mutex);
2463 dp = shash_find_data(&dp_netdevs, argv[1]);
2464 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2465 ovs_mutex_unlock(&dp_netdev_mutex);
2466 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2469 ovs_refcount_ref(&dp->ref_cnt);
2470 ovs_mutex_unlock(&dp_netdev_mutex);
2472 ovs_mutex_lock(&dp->port_mutex);
2473 if (get_port_by_name(dp, argv[2], &port)) {
2474 unixctl_command_reply_error(conn, "unknown port");
2475 } else if (port->port_no == ODPP_LOCAL) {
2476 unixctl_command_reply_error(conn, "can't delete local port");
2478 do_del_port(dp, port);
2479 unixctl_command_reply(conn, NULL);
2481 ovs_mutex_unlock(&dp->port_mutex);
2483 dp_netdev_unref(dp);
2487 dpif_dummy_register__(const char *type)
2489 struct dpif_class *class;
2491 class = xmalloc(sizeof *class);
2492 *class = dpif_netdev_class;
2493 class->type = xstrdup(type);
2494 dp_register_provider(class);
2498 dpif_dummy_register(bool override)
2505 dp_enumerate_types(&types);
2506 SSET_FOR_EACH (type, &types) {
2507 if (!dp_unregister_provider(type)) {
2508 dpif_dummy_register__(type);
2511 sset_destroy(&types);
2514 dpif_dummy_register__("dummy");
2516 unixctl_command_register("dpif-dummy/change-port-number",
2517 "DP PORT NEW-NUMBER",
2518 3, 3, dpif_dummy_change_port_number, NULL);
2519 unixctl_command_register("dpif-dummy/delete-port", "DP PORT",
2520 2, 2, dpif_dummy_delete_port, NULL);