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.
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"
45 #include "meta-flow.h"
47 #include "netdev-dpdk.h"
48 #include "netdev-vport.h"
50 #include "odp-execute.h"
52 #include "ofp-print.h"
55 #include "packet-dpif.h"
57 #include "poll-loop.h"
67 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
69 /* By default, choose a priority in the middle. */
70 #define NETDEV_RULE_PRIORITY 0x8000
73 /* Use per thread recirc_depth to prevent recirculation loop. */
74 #define MAX_RECIRC_DEPTH 5
75 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
77 /* Configuration parameters. */
78 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
81 enum { MAX_QUEUE_LEN = 128 }; /* Maximum number of packets per queue. */
82 enum { QUEUE_MASK = MAX_QUEUE_LEN - 1 };
83 BUILD_ASSERT_DECL(IS_POW2(MAX_QUEUE_LEN));
85 /* Protects against changes to 'dp_netdevs'. */
86 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
88 /* Contains all 'struct dp_netdev's. */
89 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
90 = SHASH_INITIALIZER(&dp_netdevs);
92 struct dp_netdev_upcall {
93 struct dpif_upcall upcall; /* Queued upcall information. */
94 struct ofpbuf buf; /* ofpbuf instance for upcall.packet. */
97 /* A queue passing packets from a struct dp_netdev to its clients (handlers).
103 * Any access at all requires the owning 'dp_netdev''s queue_rwlock and
105 struct dp_netdev_queue {
106 struct ovs_mutex mutex;
107 struct seq *seq; /* Incremented whenever a packet is queued. */
108 struct dp_netdev_upcall upcalls[MAX_QUEUE_LEN] OVS_GUARDED;
109 unsigned int head OVS_GUARDED;
110 unsigned int tail OVS_GUARDED;
113 /* Datapath based on the network device interface from netdev.h.
119 * Some members, marked 'const', are immutable. Accessing other members
120 * requires synchronization, as noted in more detail below.
122 * Acquisition order is, from outermost to innermost:
124 * 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 * Readers of 'cls' and 'flow_table' must take a 'cls->rwlock' read lock.
140 * Writers of 'cls' and 'flow_table' must take the 'flow_mutex' and then
141 * the 'cls->rwlock' write lock. (The outer 'flow_mutex' allows writers to
142 * atomically perform multiple operations on 'cls' and 'flow_table'.)
144 struct ovs_mutex flow_mutex;
145 struct classifier cls; /* Classifier. Protected by cls.rwlock. */
146 struct hmap flow_table OVS_GUARDED; /* Flow table. */
150 * 'queue_rwlock' protects the modification of 'handler_queues' and
151 * 'n_handlers'. The queue elements are protected by its
152 * 'handler_queues''s mutex. */
153 struct fat_rwlock queue_rwlock;
154 struct dp_netdev_queue *handler_queues;
159 * ovsthread_stats is internally synchronized. */
160 struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */
164 * Protected by RCU. Take the mutex to add or remove ports. */
165 struct ovs_mutex port_mutex;
167 struct seq *port_seq; /* Incremented whenever a port changes. */
169 /* Forwarding threads. */
170 struct latch exit_latch;
171 struct pmd_thread *pmd_threads;
172 size_t n_pmd_threads;
176 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
180 DP_STAT_HIT, /* Packets that matched in the flow table. */
181 DP_STAT_MISS, /* Packets that did not match. */
182 DP_STAT_LOST, /* Packets not passed up to the client. */
186 /* Contained by struct dp_netdev's 'stats' member. */
187 struct dp_netdev_stats {
188 struct ovs_mutex mutex; /* Protects 'n'. */
190 /* Indexed by DP_STAT_*, protected by 'mutex'. */
191 unsigned long long int n[DP_N_STATS] OVS_GUARDED;
195 /* A port in a netdev-based datapath. */
196 struct dp_netdev_port {
197 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
199 struct netdev *netdev;
200 struct netdev_saved_flags *sf;
201 struct netdev_rxq **rxq;
202 struct ovs_refcount ref_cnt;
203 char *type; /* Port type as requested by user. */
207 /* Stores a miniflow */
209 /* There are fields in the flow structure that we never use. Therefore we can
210 * save a few words of memory */
211 #define NETDEV_KEY_BUF_SIZE_U32 (FLOW_U32S \
212 - FLOW_U32_SIZE(regs) \
213 - FLOW_U32_SIZE(metadata) \
215 struct netdev_flow_key {
216 struct miniflow flow;
217 uint32_t buf[NETDEV_KEY_BUF_SIZE_U32];
220 /* A flow in dp_netdev's 'flow_table'.
226 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
227 * its dp_netdev's classifier. The text below calls this classifier 'cls'.
232 * The thread safety rules described here for "struct dp_netdev_flow" are
233 * motivated by two goals:
235 * - Prevent threads that read members of "struct dp_netdev_flow" from
236 * reading bad data due to changes by some thread concurrently modifying
239 * - Prevent two threads making changes to members of a given "struct
240 * dp_netdev_flow" from interfering with each other.
246 * A flow 'flow' may be accessed without a risk of being freed by code that
247 * holds a read-lock or write-lock on 'cls->rwlock' or that owns a reference to
248 * 'flow->ref_cnt' (or both). Code that needs to hold onto a flow for a while
249 * should take 'cls->rwlock', find the flow it needs, increment 'flow->ref_cnt'
250 * with dpif_netdev_flow_ref(), and drop 'cls->rwlock'.
252 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
253 * flow from being deleted from 'cls' (that's 'cls->rwlock') and it doesn't
254 * protect members of 'flow' from modification.
256 * Some members, marked 'const', are immutable. Accessing other members
257 * requires synchronization, as noted in more detail below.
259 struct dp_netdev_flow {
260 /* Packet classification. */
261 const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
263 /* Hash table index by unmasked flow. */
264 const struct hmap_node node; /* In owning dp_netdev's 'flow_table'. */
265 const struct flow flow; /* The flow that created this entry. */
269 * Reading or writing these members requires 'mutex'. */
270 struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
273 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
276 static void dp_netdev_flow_free(struct dp_netdev_flow *);
278 /* Contained by struct dp_netdev_flow's 'stats' member. */
279 struct dp_netdev_flow_stats {
280 struct ovs_mutex mutex; /* Guards all the other members. */
282 long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
283 long long int packet_count OVS_GUARDED; /* Number of packets matched. */
284 long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
285 uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
288 /* A set of datapath actions within a "struct dp_netdev_flow".
294 * A struct dp_netdev_actions 'actions' is protected with RCU. */
295 struct dp_netdev_actions {
296 /* These members are immutable: they do not change during the struct's
298 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
299 unsigned int size; /* Size of 'actions', in bytes. */
302 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
304 struct dp_netdev_actions *dp_netdev_flow_get_actions(
305 const struct dp_netdev_flow *);
306 static void dp_netdev_actions_free(struct dp_netdev_actions *);
308 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
309 * the performance overhead of interrupt processing. Therefore netdev can
310 * not implement rx-wait for these devices. dpif-netdev needs to poll
311 * these device to check for recv buffer. pmd-thread does polling for
312 * devices assigned to itself thread.
314 * DPDK used PMD for accessing NIC.
316 * A thread that receives packets from PMD ports, looks them up in the flow
317 * table, and executes the actions it finds.
320 struct dp_netdev *dp;
323 atomic_uint change_seq;
326 /* Interface to netdev-based datapath. */
329 struct dp_netdev *dp;
330 uint64_t last_port_seq;
333 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
334 struct dp_netdev_port **portp);
335 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
336 struct dp_netdev_port **portp);
337 static void dp_netdev_free(struct dp_netdev *)
338 OVS_REQUIRES(dp_netdev_mutex);
339 static void dp_netdev_flow_flush(struct dp_netdev *);
340 static int do_add_port(struct dp_netdev *dp, const char *devname,
341 const char *type, odp_port_t port_no)
342 OVS_REQUIRES(dp->port_mutex);
343 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
344 OVS_REQUIRES(dp->port_mutex);
345 static void dp_netdev_destroy_all_queues(struct dp_netdev *dp)
346 OVS_REQ_WRLOCK(dp->queue_rwlock);
347 static int dpif_netdev_open(const struct dpif_class *, const char *name,
348 bool create, struct dpif **);
349 static int dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf **,
350 int cnt, int queue_no, int type,
351 const struct miniflow *,
352 const struct nlattr *userdata);
353 static void dp_netdev_execute_actions(struct dp_netdev *dp,
354 struct dpif_packet **, int c,
355 bool may_steal, struct pkt_metadata *,
356 const struct nlattr *actions,
358 static void dp_netdev_port_input(struct dp_netdev *dp,
359 struct dpif_packet **packets, int cnt,
362 static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
364 static struct dpif_netdev *
365 dpif_netdev_cast(const struct dpif *dpif)
367 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
368 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
371 static struct dp_netdev *
372 get_dp_netdev(const struct dpif *dpif)
374 return dpif_netdev_cast(dpif)->dp;
378 dpif_netdev_enumerate(struct sset *all_dps,
379 const struct dpif_class *dpif_class)
381 struct shash_node *node;
383 ovs_mutex_lock(&dp_netdev_mutex);
384 SHASH_FOR_EACH(node, &dp_netdevs) {
385 struct dp_netdev *dp = node->data;
386 if (dpif_class != dp->class) {
387 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
388 * If the class doesn't match, skip this dpif. */
391 sset_add(all_dps, node->name);
393 ovs_mutex_unlock(&dp_netdev_mutex);
399 dpif_netdev_class_is_dummy(const struct dpif_class *class)
401 return class != &dpif_netdev_class;
405 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
407 return strcmp(type, "internal") ? type
408 : dpif_netdev_class_is_dummy(class) ? "dummy"
413 create_dpif_netdev(struct dp_netdev *dp)
415 uint16_t netflow_id = hash_string(dp->name, 0);
416 struct dpif_netdev *dpif;
418 ovs_refcount_ref(&dp->ref_cnt);
420 dpif = xmalloc(sizeof *dpif);
421 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
423 dpif->last_port_seq = seq_read(dp->port_seq);
428 /* Choose an unused, non-zero port number and return it on success.
429 * Return ODPP_NONE on failure. */
431 choose_port(struct dp_netdev *dp, const char *name)
432 OVS_REQUIRES(dp->port_mutex)
436 if (dp->class != &dpif_netdev_class) {
440 /* If the port name begins with "br", start the number search at
441 * 100 to make writing tests easier. */
442 if (!strncmp(name, "br", 2)) {
446 /* If the port name contains a number, try to assign that port number.
447 * This can make writing unit tests easier because port numbers are
449 for (p = name; *p != '\0'; p++) {
450 if (isdigit((unsigned char) *p)) {
451 port_no = start_no + strtol(p, NULL, 10);
452 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
453 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
454 return u32_to_odp(port_no);
461 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
462 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
463 return u32_to_odp(port_no);
471 create_dp_netdev(const char *name, const struct dpif_class *class,
472 struct dp_netdev **dpp)
473 OVS_REQUIRES(dp_netdev_mutex)
475 struct dp_netdev *dp;
478 dp = xzalloc(sizeof *dp);
479 shash_add(&dp_netdevs, name, dp);
481 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
482 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
483 ovs_refcount_init(&dp->ref_cnt);
484 atomic_flag_clear(&dp->destroyed);
486 ovs_mutex_init(&dp->flow_mutex);
487 classifier_init(&dp->cls, NULL);
488 hmap_init(&dp->flow_table);
490 fat_rwlock_init(&dp->queue_rwlock);
492 ovsthread_stats_init(&dp->stats);
494 ovs_mutex_init(&dp->port_mutex);
495 cmap_init(&dp->ports);
496 dp->port_seq = seq_create();
497 latch_init(&dp->exit_latch);
499 ovs_mutex_lock(&dp->port_mutex);
500 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
501 ovs_mutex_unlock(&dp->port_mutex);
512 dpif_netdev_open(const struct dpif_class *class, const char *name,
513 bool create, struct dpif **dpifp)
515 struct dp_netdev *dp;
518 ovs_mutex_lock(&dp_netdev_mutex);
519 dp = shash_find_data(&dp_netdevs, name);
521 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
523 error = (dp->class != class ? EINVAL
528 *dpifp = create_dpif_netdev(dp);
530 ovs_mutex_unlock(&dp_netdev_mutex);
536 dp_netdev_purge_queues(struct dp_netdev *dp)
537 OVS_REQ_WRLOCK(dp->queue_rwlock)
541 for (i = 0; i < dp->n_handlers; i++) {
542 struct dp_netdev_queue *q = &dp->handler_queues[i];
544 ovs_mutex_lock(&q->mutex);
545 while (q->tail != q->head) {
546 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
547 ofpbuf_uninit(&u->upcall.packet);
548 ofpbuf_uninit(&u->buf);
550 ovs_mutex_unlock(&q->mutex);
554 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
555 * through the 'dp_netdevs' shash while freeing 'dp'. */
557 dp_netdev_free(struct dp_netdev *dp)
558 OVS_REQUIRES(dp_netdev_mutex)
560 struct dp_netdev_port *port;
561 struct dp_netdev_stats *bucket;
564 shash_find_and_delete(&dp_netdevs, dp->name);
566 dp_netdev_set_pmd_threads(dp, 0);
567 free(dp->pmd_threads);
569 dp_netdev_flow_flush(dp);
570 ovs_mutex_lock(&dp->port_mutex);
571 CMAP_FOR_EACH (port, node, &dp->ports) {
572 do_del_port(dp, port);
574 ovs_mutex_unlock(&dp->port_mutex);
576 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
577 ovs_mutex_destroy(&bucket->mutex);
578 free_cacheline(bucket);
580 ovsthread_stats_destroy(&dp->stats);
582 fat_rwlock_wrlock(&dp->queue_rwlock);
583 dp_netdev_destroy_all_queues(dp);
584 fat_rwlock_unlock(&dp->queue_rwlock);
586 fat_rwlock_destroy(&dp->queue_rwlock);
588 classifier_destroy(&dp->cls);
589 hmap_destroy(&dp->flow_table);
590 ovs_mutex_destroy(&dp->flow_mutex);
591 seq_destroy(dp->port_seq);
592 cmap_destroy(&dp->ports);
593 latch_destroy(&dp->exit_latch);
594 free(CONST_CAST(char *, dp->name));
599 dp_netdev_unref(struct dp_netdev *dp)
602 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
603 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
604 ovs_mutex_lock(&dp_netdev_mutex);
605 if (ovs_refcount_unref(&dp->ref_cnt) == 1) {
608 ovs_mutex_unlock(&dp_netdev_mutex);
613 dpif_netdev_close(struct dpif *dpif)
615 struct dp_netdev *dp = get_dp_netdev(dpif);
622 dpif_netdev_destroy(struct dpif *dpif)
624 struct dp_netdev *dp = get_dp_netdev(dpif);
626 if (!atomic_flag_test_and_set(&dp->destroyed)) {
627 if (ovs_refcount_unref(&dp->ref_cnt) == 1) {
628 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
637 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
639 struct dp_netdev *dp = get_dp_netdev(dpif);
640 struct dp_netdev_stats *bucket;
643 fat_rwlock_rdlock(&dp->cls.rwlock);
644 stats->n_flows = hmap_count(&dp->flow_table);
645 fat_rwlock_unlock(&dp->cls.rwlock);
647 stats->n_hit = stats->n_missed = stats->n_lost = 0;
648 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
649 ovs_mutex_lock(&bucket->mutex);
650 stats->n_hit += bucket->n[DP_STAT_HIT];
651 stats->n_missed += bucket->n[DP_STAT_MISS];
652 stats->n_lost += bucket->n[DP_STAT_LOST];
653 ovs_mutex_unlock(&bucket->mutex);
655 stats->n_masks = UINT32_MAX;
656 stats->n_mask_hit = UINT64_MAX;
662 dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
666 for (i = 0; i < dp->n_pmd_threads; i++) {
667 struct pmd_thread *f = &dp->pmd_threads[i];
670 atomic_add(&f->change_seq, 1, &id);
675 hash_port_no(odp_port_t port_no)
677 return hash_int(odp_to_u32(port_no), 0);
681 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
683 OVS_REQUIRES(dp->port_mutex)
685 struct netdev_saved_flags *sf;
686 struct dp_netdev_port *port;
687 struct netdev *netdev;
688 enum netdev_flags flags;
689 const char *open_type;
693 /* XXX reject devices already in some dp_netdev. */
695 /* Open and validate network device. */
696 open_type = dpif_netdev_port_open_type(dp->class, type);
697 error = netdev_open(devname, open_type, &netdev);
701 /* XXX reject non-Ethernet devices */
703 netdev_get_flags(netdev, &flags);
704 if (flags & NETDEV_LOOPBACK) {
705 VLOG_ERR("%s: cannot add a loopback device", devname);
706 netdev_close(netdev);
710 port = xzalloc(sizeof *port);
711 port->port_no = port_no;
712 port->netdev = netdev;
713 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
714 port->type = xstrdup(type);
715 for (i = 0; i < netdev_n_rxq(netdev); i++) {
716 error = netdev_rxq_open(netdev, &port->rxq[i], i);
718 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
719 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
720 devname, ovs_strerror(errno));
721 netdev_close(netdev);
726 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
728 for (i = 0; i < netdev_n_rxq(netdev); i++) {
729 netdev_rxq_close(port->rxq[i]);
731 netdev_close(netdev);
738 if (netdev_is_pmd(netdev)) {
740 dp_netdev_set_pmd_threads(dp, NR_THREADS);
741 dp_netdev_reload_pmd_threads(dp);
743 ovs_refcount_init(&port->ref_cnt);
745 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
746 seq_change(dp->port_seq);
752 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
753 odp_port_t *port_nop)
755 struct dp_netdev *dp = get_dp_netdev(dpif);
756 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
757 const char *dpif_port;
761 ovs_mutex_lock(&dp->port_mutex);
762 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
763 if (*port_nop != ODPP_NONE) {
765 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
767 port_no = choose_port(dp, dpif_port);
768 error = port_no == ODPP_NONE ? EFBIG : 0;
772 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
774 ovs_mutex_unlock(&dp->port_mutex);
780 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
782 struct dp_netdev *dp = get_dp_netdev(dpif);
785 ovs_mutex_lock(&dp->port_mutex);
786 if (port_no == ODPP_LOCAL) {
789 struct dp_netdev_port *port;
791 error = get_port_by_number(dp, port_no, &port);
793 do_del_port(dp, port);
796 ovs_mutex_unlock(&dp->port_mutex);
802 is_valid_port_number(odp_port_t port_no)
804 return port_no != ODPP_NONE;
807 static struct dp_netdev_port *
808 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
810 struct dp_netdev_port *port;
812 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
813 if (port->port_no == port_no) {
821 get_port_by_number(struct dp_netdev *dp,
822 odp_port_t port_no, struct dp_netdev_port **portp)
824 if (!is_valid_port_number(port_no)) {
828 *portp = dp_netdev_lookup_port(dp, port_no);
829 return *portp ? 0 : ENOENT;
834 port_ref(struct dp_netdev_port *port)
837 ovs_refcount_ref(&port->ref_cnt);
842 port_destroy__(struct dp_netdev_port *port)
844 int n_rxq = netdev_n_rxq(port->netdev);
847 netdev_close(port->netdev);
848 netdev_restore_flags(port->sf);
850 for (i = 0; i < n_rxq; i++) {
851 netdev_rxq_close(port->rxq[i]);
859 port_unref(struct dp_netdev_port *port)
861 if (port && ovs_refcount_unref(&port->ref_cnt) == 1) {
862 ovsrcu_postpone(port_destroy__, port);
867 get_port_by_name(struct dp_netdev *dp,
868 const char *devname, struct dp_netdev_port **portp)
869 OVS_REQUIRES(dp->port_mutex)
871 struct dp_netdev_port *port;
873 CMAP_FOR_EACH (port, node, &dp->ports) {
874 if (!strcmp(netdev_get_name(port->netdev), devname)) {
883 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
884 OVS_REQUIRES(dp->port_mutex)
886 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
887 seq_change(dp->port_seq);
888 if (netdev_is_pmd(port->netdev)) {
889 dp_netdev_reload_pmd_threads(dp);
896 answer_port_query(const struct dp_netdev_port *port,
897 struct dpif_port *dpif_port)
899 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
900 dpif_port->type = xstrdup(port->type);
901 dpif_port->port_no = port->port_no;
905 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
906 struct dpif_port *dpif_port)
908 struct dp_netdev *dp = get_dp_netdev(dpif);
909 struct dp_netdev_port *port;
912 error = get_port_by_number(dp, port_no, &port);
913 if (!error && dpif_port) {
914 answer_port_query(port, dpif_port);
921 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
922 struct dpif_port *dpif_port)
924 struct dp_netdev *dp = get_dp_netdev(dpif);
925 struct dp_netdev_port *port;
928 ovs_mutex_lock(&dp->port_mutex);
929 error = get_port_by_name(dp, devname, &port);
930 if (!error && dpif_port) {
931 answer_port_query(port, dpif_port);
933 ovs_mutex_unlock(&dp->port_mutex);
939 dp_netdev_flow_free(struct dp_netdev_flow *flow)
941 struct dp_netdev_flow_stats *bucket;
944 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
945 ovs_mutex_destroy(&bucket->mutex);
946 free_cacheline(bucket);
948 ovsthread_stats_destroy(&flow->stats);
950 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
951 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
956 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
957 OVS_REQ_WRLOCK(dp->cls.rwlock)
958 OVS_REQUIRES(dp->flow_mutex)
960 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
961 struct hmap_node *node = CONST_CAST(struct hmap_node *, &flow->node);
963 classifier_remove(&dp->cls, cr);
964 hmap_remove(&dp->flow_table, node);
965 ovsrcu_postpone(dp_netdev_flow_free, flow);
969 dp_netdev_flow_flush(struct dp_netdev *dp)
971 struct dp_netdev_flow *netdev_flow, *next;
973 ovs_mutex_lock(&dp->flow_mutex);
974 fat_rwlock_wrlock(&dp->cls.rwlock);
975 HMAP_FOR_EACH_SAFE (netdev_flow, next, node, &dp->flow_table) {
976 dp_netdev_remove_flow(dp, netdev_flow);
978 fat_rwlock_unlock(&dp->cls.rwlock);
979 ovs_mutex_unlock(&dp->flow_mutex);
983 dpif_netdev_flow_flush(struct dpif *dpif)
985 struct dp_netdev *dp = get_dp_netdev(dpif);
987 dp_netdev_flow_flush(dp);
991 struct dp_netdev_port_state {
992 struct cmap_position position;
997 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
999 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1004 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1005 struct dpif_port *dpif_port)
1007 struct dp_netdev_port_state *state = state_;
1008 struct dp_netdev *dp = get_dp_netdev(dpif);
1009 struct cmap_node *node;
1012 node = cmap_next_position(&dp->ports, &state->position);
1014 struct dp_netdev_port *port;
1016 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1019 state->name = xstrdup(netdev_get_name(port->netdev));
1020 dpif_port->name = state->name;
1021 dpif_port->type = port->type;
1022 dpif_port->port_no = port->port_no;
1033 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1035 struct dp_netdev_port_state *state = state_;
1042 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1044 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1045 uint64_t new_port_seq;
1048 new_port_seq = seq_read(dpif->dp->port_seq);
1049 if (dpif->last_port_seq != new_port_seq) {
1050 dpif->last_port_seq = new_port_seq;
1060 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1062 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1064 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1067 static struct dp_netdev_flow *
1068 dp_netdev_flow_cast(const struct cls_rule *cr)
1070 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1073 static struct dp_netdev_flow *
1074 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
1075 OVS_EXCLUDED(dp->cls.rwlock)
1077 struct dp_netdev_flow *netdev_flow;
1078 struct cls_rule *rule;
1080 fat_rwlock_rdlock(&dp->cls.rwlock);
1081 rule = classifier_lookup_miniflow_first(&dp->cls, key);
1082 netdev_flow = dp_netdev_flow_cast(rule);
1083 fat_rwlock_unlock(&dp->cls.rwlock);
1088 static struct dp_netdev_flow *
1089 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1090 OVS_REQ_RDLOCK(dp->cls.rwlock)
1092 struct dp_netdev_flow *netdev_flow;
1094 HMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1096 if (flow_equal(&netdev_flow->flow, flow)) {
1105 get_dpif_flow_stats(struct dp_netdev_flow *netdev_flow,
1106 struct dpif_flow_stats *stats)
1108 struct dp_netdev_flow_stats *bucket;
1111 memset(stats, 0, sizeof *stats);
1112 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1113 ovs_mutex_lock(&bucket->mutex);
1114 stats->n_packets += bucket->packet_count;
1115 stats->n_bytes += bucket->byte_count;
1116 stats->used = MAX(stats->used, bucket->used);
1117 stats->tcp_flags |= bucket->tcp_flags;
1118 ovs_mutex_unlock(&bucket->mutex);
1123 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1124 const struct nlattr *mask_key,
1125 uint32_t mask_key_len, const struct flow *flow,
1129 enum odp_key_fitness fitness;
1131 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1133 /* This should not happen: it indicates that
1134 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1135 * disagree on the acceptable form of a mask. Log the problem
1136 * as an error, with enough details to enable debugging. */
1137 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1139 if (!VLOG_DROP_ERR(&rl)) {
1143 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1145 VLOG_ERR("internal error parsing flow mask %s (%s)",
1146 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1153 enum mf_field_id id;
1154 /* No mask key, unwildcard everything except fields whose
1155 * prerequisities are not met. */
1156 memset(mask, 0x0, sizeof *mask);
1158 for (id = 0; id < MFF_N_IDS; ++id) {
1159 /* Skip registers and metadata. */
1160 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1161 && id != MFF_METADATA) {
1162 const struct mf_field *mf = mf_from_id(id);
1163 if (mf_are_prereqs_ok(mf, flow)) {
1164 mf_mask_field(mf, mask);
1170 /* Force unwildcard the in_port.
1172 * We need to do this even in the case where we unwildcard "everything"
1173 * above because "everything" only includes the 16-bit OpenFlow port number
1174 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1175 * port number mask->in_port.odp_port. */
1176 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1182 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1187 if (odp_flow_key_to_flow(key, key_len, flow)) {
1188 /* This should not happen: it indicates that odp_flow_key_from_flow()
1189 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1190 * flow. Log the problem as an error, with enough details to enable
1192 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1194 if (!VLOG_DROP_ERR(&rl)) {
1198 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1199 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1206 in_port = flow->in_port.odp_port;
1207 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1215 dpif_netdev_flow_get(const struct dpif *dpif,
1216 const struct nlattr *nl_key, size_t nl_key_len,
1217 struct ofpbuf **actionsp, struct dpif_flow_stats *stats)
1219 struct dp_netdev *dp = get_dp_netdev(dpif);
1220 struct dp_netdev_flow *netdev_flow;
1224 error = dpif_netdev_flow_from_nlattrs(nl_key, nl_key_len, &key);
1229 fat_rwlock_rdlock(&dp->cls.rwlock);
1230 netdev_flow = dp_netdev_find_flow(dp, &key);
1231 fat_rwlock_unlock(&dp->cls.rwlock);
1235 get_dpif_flow_stats(netdev_flow, stats);
1239 struct dp_netdev_actions *actions;
1241 actions = dp_netdev_flow_get_actions(netdev_flow);
1242 *actionsp = ofpbuf_clone_data(actions->actions, actions->size);
1252 dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow,
1253 const struct flow_wildcards *wc,
1254 const struct nlattr *actions,
1256 OVS_REQUIRES(dp->flow_mutex)
1258 struct dp_netdev_flow *netdev_flow;
1261 netdev_flow = xzalloc(sizeof *netdev_flow);
1262 *CONST_CAST(struct flow *, &netdev_flow->flow) = *flow;
1264 ovsthread_stats_init(&netdev_flow->stats);
1266 ovsrcu_set(&netdev_flow->actions,
1267 dp_netdev_actions_create(actions, actions_len));
1269 match_init(&match, flow, wc);
1270 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1271 &match, NETDEV_RULE_PRIORITY);
1272 fat_rwlock_wrlock(&dp->cls.rwlock);
1273 classifier_insert(&dp->cls,
1274 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1275 hmap_insert(&dp->flow_table,
1276 CONST_CAST(struct hmap_node *, &netdev_flow->node),
1277 flow_hash(flow, 0));
1278 fat_rwlock_unlock(&dp->cls.rwlock);
1284 clear_stats(struct dp_netdev_flow *netdev_flow)
1286 struct dp_netdev_flow_stats *bucket;
1289 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1290 ovs_mutex_lock(&bucket->mutex);
1292 bucket->packet_count = 0;
1293 bucket->byte_count = 0;
1294 bucket->tcp_flags = 0;
1295 ovs_mutex_unlock(&bucket->mutex);
1300 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1302 struct dp_netdev *dp = get_dp_netdev(dpif);
1303 struct dp_netdev_flow *netdev_flow;
1305 struct miniflow miniflow;
1306 struct flow_wildcards wc;
1309 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &flow);
1313 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1314 put->mask, put->mask_len,
1319 miniflow_init(&miniflow, &flow);
1321 ovs_mutex_lock(&dp->flow_mutex);
1322 netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
1324 if (put->flags & DPIF_FP_CREATE) {
1325 if (hmap_count(&dp->flow_table) < MAX_FLOWS) {
1327 memset(put->stats, 0, sizeof *put->stats);
1329 error = dp_netdev_flow_add(dp, &flow, &wc, put->actions,
1338 if (put->flags & DPIF_FP_MODIFY
1339 && flow_equal(&flow, &netdev_flow->flow)) {
1340 struct dp_netdev_actions *new_actions;
1341 struct dp_netdev_actions *old_actions;
1343 new_actions = dp_netdev_actions_create(put->actions,
1346 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1347 ovsrcu_set(&netdev_flow->actions, new_actions);
1350 get_dpif_flow_stats(netdev_flow, put->stats);
1352 if (put->flags & DPIF_FP_ZERO_STATS) {
1353 clear_stats(netdev_flow);
1356 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1357 } else if (put->flags & DPIF_FP_CREATE) {
1360 /* Overlapping flow. */
1364 ovs_mutex_unlock(&dp->flow_mutex);
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 fat_rwlock_wrlock(&dp->cls.rwlock);
1384 netdev_flow = dp_netdev_find_flow(dp, &key);
1387 get_dpif_flow_stats(netdev_flow, del->stats);
1389 dp_netdev_remove_flow(dp, netdev_flow);
1393 fat_rwlock_unlock(&dp->cls.rwlock);
1394 ovs_mutex_unlock(&dp->flow_mutex);
1399 struct dpif_netdev_flow_dump {
1400 struct dpif_flow_dump up;
1404 struct ovs_mutex mutex;
1407 static struct dpif_netdev_flow_dump *
1408 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1410 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1413 static struct dpif_flow_dump *
1414 dpif_netdev_flow_dump_create(const struct dpif *dpif_)
1416 struct dpif_netdev_flow_dump *dump;
1418 dump = xmalloc(sizeof *dump);
1419 dpif_flow_dump_init(&dump->up, dpif_);
1423 ovs_mutex_init(&dump->mutex);
1429 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1431 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1433 ovs_mutex_destroy(&dump->mutex);
1438 struct dpif_netdev_flow_dump_thread {
1439 struct dpif_flow_dump_thread up;
1440 struct dpif_netdev_flow_dump *dump;
1441 struct odputil_keybuf keybuf;
1442 struct odputil_keybuf maskbuf;
1445 static struct dpif_netdev_flow_dump_thread *
1446 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1448 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1451 static struct dpif_flow_dump_thread *
1452 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1454 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1455 struct dpif_netdev_flow_dump_thread *thread;
1457 thread = xmalloc(sizeof *thread);
1458 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1459 thread->dump = dump;
1464 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1466 struct dpif_netdev_flow_dump_thread *thread
1467 = dpif_netdev_flow_dump_thread_cast(thread_);
1472 /* XXX the caller must use 'actions' without quiescing */
1474 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
1475 struct dpif_flow *f, int max_flows OVS_UNUSED)
1477 struct dpif_netdev_flow_dump_thread *thread
1478 = dpif_netdev_flow_dump_thread_cast(thread_);
1479 struct dpif_netdev_flow_dump *dump = thread->dump;
1480 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
1481 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
1482 struct dp_netdev_flow *netdev_flow;
1483 struct flow_wildcards wc;
1484 struct dp_netdev_actions *dp_actions;
1488 ovs_mutex_lock(&dump->mutex);
1489 error = dump->status;
1491 struct hmap_node *node;
1493 fat_rwlock_rdlock(&dp->cls.rwlock);
1494 node = hmap_at_position(&dp->flow_table, &dump->bucket, &dump->offset);
1496 netdev_flow = CONTAINER_OF(node, struct dp_netdev_flow, node);
1498 fat_rwlock_unlock(&dp->cls.rwlock);
1500 dump->status = error = EOF;
1503 ovs_mutex_unlock(&dump->mutex);
1508 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1511 ofpbuf_use_stack(&buf, &thread->keybuf, sizeof thread->keybuf);
1512 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1513 netdev_flow->flow.in_port.odp_port, true);
1514 f->key = ofpbuf_data(&buf);
1515 f->key_len = ofpbuf_size(&buf);
1518 ofpbuf_use_stack(&buf, &thread->maskbuf, sizeof thread->maskbuf);
1519 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1520 odp_to_u32(wc.masks.in_port.odp_port),
1522 f->mask = ofpbuf_data(&buf);
1523 f->mask_len = ofpbuf_size(&buf);
1526 dp_actions = dp_netdev_flow_get_actions(netdev_flow);
1527 f->actions = dp_actions->actions;
1528 f->actions_len = dp_actions->size;
1531 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 batch_pkt_execute {
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 batch_pkt_execute *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 batch_pkt_execute *batch, struct dp_netdev_flow *flow,
2027 struct dpif_packet *packet, struct pkt_metadata *md,
2028 const struct miniflow *mf)
2032 batch->packets[0] = packet;
2034 batch->packet_count = 0;
2035 batch->byte_count = 0;
2036 batch->tcp_flags = 0;
2038 packet_batch_update(batch, packet, mf);
2042 packet_batch_execute(struct batch_pkt_execute *batch, struct dp_netdev *dp)
2044 struct dp_netdev_actions *actions;
2045 struct dp_netdev_flow *flow = batch->flow;
2047 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
2050 actions = dp_netdev_flow_get_actions(flow);
2052 dp_netdev_execute_actions(dp, batch->packets,
2053 batch->packet_count, true, &batch->md,
2054 actions->actions, actions->size);
2056 dp_netdev_count_packet(dp, DP_STAT_HIT, batch->packet_count);
2060 dp_netdev_input(struct dp_netdev *dp, struct dpif_packet **packets, int cnt,
2061 struct pkt_metadata *md)
2063 struct batch_pkt_execute batch;
2065 struct netdev_flow_key key;
2071 miniflow_initialize(&key.flow, key.buf);
2073 for (i = 0; i < cnt; i++) {
2074 struct dp_netdev_flow *netdev_flow;
2075 struct ofpbuf *buf = &packets[i]->ofpbuf;
2077 if (ofpbuf_size(buf) < ETH_HEADER_LEN) {
2078 dpif_packet_delete(packets[i]);
2082 miniflow_extract(buf, md, &key.flow);
2084 netdev_flow = dp_netdev_lookup_flow(dp, &key.flow);
2088 packet_batch_init(&batch, netdev_flow, packets[i], md,
2090 } else if (batch.flow == netdev_flow) {
2091 packet_batch_update(&batch, packets[i], &key.flow);
2093 packet_batch_execute(&batch, dp);
2094 packet_batch_init(&batch, netdev_flow, packets[i], md,
2098 /* Packet's flow not in datapath */
2099 dp_netdev_count_packet(dp, DP_STAT_MISS, 1);
2101 if (dp->handler_queues) {
2103 dp_netdev_output_userspace(dp, &buf, 1,
2104 miniflow_hash_5tuple(&key.flow, 0)
2106 DPIF_UC_MISS, &key.flow, NULL);
2108 /* No upcall queue. Freeing the packet */
2109 dpif_packet_delete(packets[i]);
2115 packet_batch_execute(&batch, dp);
2120 dp_netdev_port_input(struct dp_netdev *dp, struct dpif_packet **packets,
2121 int cnt, odp_port_t port_no)
2123 uint32_t *recirc_depth = recirc_depth_get();
2124 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port_no);
2127 dp_netdev_input(dp, packets, cnt, &md);
2131 dp_netdev_queue_userspace_packet(struct dp_netdev_queue *q,
2132 struct ofpbuf *packet, int type,
2133 const struct miniflow *key,
2134 const struct nlattr *userdata)
2135 OVS_REQUIRES(q->mutex)
2137 if (q->head - q->tail < MAX_QUEUE_LEN) {
2138 struct dp_netdev_upcall *u = &q->upcalls[q->head++ & QUEUE_MASK];
2139 struct dpif_upcall *upcall = &u->upcall;
2140 struct ofpbuf *buf = &u->buf;
2144 upcall->type = type;
2146 /* Allocate buffer big enough for everything. */
2147 buf_size = ODPUTIL_FLOW_KEY_BYTES;
2149 buf_size += NLA_ALIGN(userdata->nla_len);
2151 ofpbuf_init(buf, buf_size);
2154 miniflow_expand(key, &flow);
2155 odp_flow_key_from_flow(buf, &flow, NULL, flow.in_port.odp_port, true);
2156 upcall->key = ofpbuf_data(buf);
2157 upcall->key_len = ofpbuf_size(buf);
2161 upcall->userdata = ofpbuf_put(buf, userdata,
2162 NLA_ALIGN(userdata->nla_len));
2165 upcall->packet = *packet;
2171 ofpbuf_delete(packet);
2178 dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf **packets,
2179 int cnt, int queue_no, int type,
2180 const struct miniflow *key,
2181 const struct nlattr *userdata)
2183 struct dp_netdev_queue *q;
2187 fat_rwlock_rdlock(&dp->queue_rwlock);
2188 q = &dp->handler_queues[queue_no];
2189 ovs_mutex_lock(&q->mutex);
2190 for (i = 0; i < cnt; i++) {
2191 struct ofpbuf *packet = packets[i];
2193 error = dp_netdev_queue_userspace_packet(q, packet, type, key,
2195 if (error == ENOBUFS) {
2196 dp_netdev_count_packet(dp, DP_STAT_LOST, 1);
2199 ovs_mutex_unlock(&q->mutex);
2200 fat_rwlock_unlock(&dp->queue_rwlock);
2205 struct dp_netdev_execute_aux {
2206 struct dp_netdev *dp;
2210 dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
2211 struct pkt_metadata *md,
2212 const struct nlattr *a, bool may_steal)
2213 OVS_NO_THREAD_SAFETY_ANALYSIS
2215 struct dp_netdev_execute_aux *aux = aux_;
2216 int type = nl_attr_type(a);
2217 struct dp_netdev_port *p;
2218 uint32_t *depth = recirc_depth_get();
2221 switch ((enum ovs_action_attr)type) {
2222 case OVS_ACTION_ATTR_OUTPUT:
2223 p = dp_netdev_lookup_port(aux->dp, u32_to_odp(nl_attr_get_u32(a)));
2225 netdev_send(p->netdev, packets, cnt, may_steal);
2229 case OVS_ACTION_ATTR_USERSPACE: {
2230 const struct nlattr *userdata;
2231 struct netdev_flow_key key;
2233 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2235 miniflow_initialize(&key.flow, key.buf);
2237 for (i = 0; i < cnt; i++) {
2238 struct ofpbuf *packet, *userspace_packet;
2240 packet = &packets[i]->ofpbuf;
2242 miniflow_extract(packet, md, &key.flow);
2244 userspace_packet = may_steal ? packet : ofpbuf_clone(packet);
2246 dp_netdev_output_userspace(aux->dp, &userspace_packet, 1,
2247 miniflow_hash_5tuple(&key.flow, 0)
2248 % aux->dp->n_handlers,
2249 DPIF_UC_ACTION, &key.flow,
2255 case OVS_ACTION_ATTR_HASH: {
2256 const struct ovs_action_hash *hash_act;
2257 struct netdev_flow_key key;
2260 hash_act = nl_attr_get(a);
2262 miniflow_initialize(&key.flow, key.buf);
2264 for (i = 0; i < cnt; i++) {
2266 /* TODO: this is slow. Use RSS hash in the future */
2267 miniflow_extract(&packets[i]->ofpbuf, md, &key.flow);
2269 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
2270 /* Hash need not be symmetric, nor does it need to include
2272 hash = miniflow_hash_5tuple(&key.flow, hash_act->hash_basis);
2274 VLOG_WARN("Unknown hash algorithm specified "
2275 "for the hash action.");
2280 hash = 1; /* 0 is not valid */
2286 packets[i]->dp_hash = hash;
2291 case OVS_ACTION_ATTR_RECIRC:
2292 if (*depth < MAX_RECIRC_DEPTH) {
2295 for (i = 0; i < cnt; i++) {
2296 struct dpif_packet *recirc_pkt;
2297 struct pkt_metadata recirc_md = *md;
2299 recirc_pkt = (may_steal) ? packets[i]
2300 : dpif_packet_clone(packets[i]);
2302 recirc_md.recirc_id = nl_attr_get_u32(a);
2304 /* Hash is private to each packet */
2305 recirc_md.dp_hash = packets[i]->dp_hash;
2307 dp_netdev_input(aux->dp, &recirc_pkt, 1, &recirc_md);
2313 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
2317 case OVS_ACTION_ATTR_PUSH_VLAN:
2318 case OVS_ACTION_ATTR_POP_VLAN:
2319 case OVS_ACTION_ATTR_PUSH_MPLS:
2320 case OVS_ACTION_ATTR_POP_MPLS:
2321 case OVS_ACTION_ATTR_SET:
2322 case OVS_ACTION_ATTR_SAMPLE:
2323 case OVS_ACTION_ATTR_UNSPEC:
2324 case __OVS_ACTION_ATTR_MAX:
2330 dp_netdev_execute_actions(struct dp_netdev *dp,
2331 struct dpif_packet **packets, int cnt,
2332 bool may_steal, struct pkt_metadata *md,
2333 const struct nlattr *actions, size_t actions_len)
2335 struct dp_netdev_execute_aux aux = {dp};
2337 odp_execute_actions(&aux, packets, cnt, may_steal, md, actions,
2338 actions_len, dp_execute_cb);
2341 const struct dpif_class dpif_netdev_class = {
2343 dpif_netdev_enumerate,
2344 dpif_netdev_port_open_type,
2347 dpif_netdev_destroy,
2350 dpif_netdev_get_stats,
2351 dpif_netdev_port_add,
2352 dpif_netdev_port_del,
2353 dpif_netdev_port_query_by_number,
2354 dpif_netdev_port_query_by_name,
2355 NULL, /* port_get_pid */
2356 dpif_netdev_port_dump_start,
2357 dpif_netdev_port_dump_next,
2358 dpif_netdev_port_dump_done,
2359 dpif_netdev_port_poll,
2360 dpif_netdev_port_poll_wait,
2361 dpif_netdev_flow_get,
2362 dpif_netdev_flow_put,
2363 dpif_netdev_flow_del,
2364 dpif_netdev_flow_flush,
2365 dpif_netdev_flow_dump_create,
2366 dpif_netdev_flow_dump_destroy,
2367 dpif_netdev_flow_dump_thread_create,
2368 dpif_netdev_flow_dump_thread_destroy,
2369 dpif_netdev_flow_dump_next,
2370 dpif_netdev_execute,
2372 dpif_netdev_recv_set,
2373 dpif_netdev_handlers_set,
2374 dpif_netdev_queue_to_priority,
2376 dpif_netdev_recv_wait,
2377 dpif_netdev_recv_purge,
2381 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2382 const char *argv[], void *aux OVS_UNUSED)
2384 struct dp_netdev_port *old_port;
2385 struct dp_netdev_port *new_port;
2386 struct dp_netdev *dp;
2389 ovs_mutex_lock(&dp_netdev_mutex);
2390 dp = shash_find_data(&dp_netdevs, argv[1]);
2391 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2392 ovs_mutex_unlock(&dp_netdev_mutex);
2393 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2396 ovs_refcount_ref(&dp->ref_cnt);
2397 ovs_mutex_unlock(&dp_netdev_mutex);
2399 ovs_mutex_lock(&dp->port_mutex);
2400 if (get_port_by_name(dp, argv[2], &old_port)) {
2401 unixctl_command_reply_error(conn, "unknown port");
2405 port_no = u32_to_odp(atoi(argv[3]));
2406 if (!port_no || port_no == ODPP_NONE) {
2407 unixctl_command_reply_error(conn, "bad port number");
2410 if (dp_netdev_lookup_port(dp, port_no)) {
2411 unixctl_command_reply_error(conn, "port number already in use");
2415 /* Remove old port. */
2416 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
2417 ovsrcu_postpone(free, old_port);
2419 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
2420 new_port = xmemdup(old_port, sizeof *old_port);
2421 new_port->port_no = port_no;
2422 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
2424 seq_change(dp->port_seq);
2425 unixctl_command_reply(conn, NULL);
2428 ovs_mutex_unlock(&dp->port_mutex);
2429 dp_netdev_unref(dp);
2433 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
2434 const char *argv[], void *aux OVS_UNUSED)
2436 struct dp_netdev_port *port;
2437 struct dp_netdev *dp;
2439 ovs_mutex_lock(&dp_netdev_mutex);
2440 dp = shash_find_data(&dp_netdevs, argv[1]);
2441 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2442 ovs_mutex_unlock(&dp_netdev_mutex);
2443 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2446 ovs_refcount_ref(&dp->ref_cnt);
2447 ovs_mutex_unlock(&dp_netdev_mutex);
2449 ovs_mutex_lock(&dp->port_mutex);
2450 if (get_port_by_name(dp, argv[2], &port)) {
2451 unixctl_command_reply_error(conn, "unknown port");
2452 } else if (port->port_no == ODPP_LOCAL) {
2453 unixctl_command_reply_error(conn, "can't delete local port");
2455 do_del_port(dp, port);
2456 unixctl_command_reply(conn, NULL);
2458 ovs_mutex_unlock(&dp->port_mutex);
2460 dp_netdev_unref(dp);
2464 dpif_dummy_register__(const char *type)
2466 struct dpif_class *class;
2468 class = xmalloc(sizeof *class);
2469 *class = dpif_netdev_class;
2470 class->type = xstrdup(type);
2471 dp_register_provider(class);
2475 dpif_dummy_register(bool override)
2482 dp_enumerate_types(&types);
2483 SSET_FOR_EACH (type, &types) {
2484 if (!dp_unregister_provider(type)) {
2485 dpif_dummy_register__(type);
2488 sset_destroy(&types);
2491 dpif_dummy_register__("dummy");
2493 unixctl_command_register("dpif-dummy/change-port-number",
2494 "DP PORT NEW-NUMBER",
2495 3, 3, dpif_dummy_change_port_number, NULL);
2496 unixctl_command_register("dpif-dummy/delete-port", "DP PORT",
2497 2, 2, dpif_dummy_delete_port, NULL);