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"
37 #include "dpif-provider.h"
39 #include "dynamic-string.h"
44 #include "meta-flow.h"
46 #include "netdev-vport.h"
48 #include "odp-execute.h"
50 #include "ofp-print.h"
54 #include "poll-loop.h"
64 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
66 /* By default, choose a priority in the middle. */
67 #define NETDEV_RULE_PRIORITY 0x8000
71 /* Configuration parameters. */
72 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
75 enum { MAX_QUEUE_LEN = 128 }; /* Maximum number of packets per queue. */
76 enum { QUEUE_MASK = MAX_QUEUE_LEN - 1 };
77 BUILD_ASSERT_DECL(IS_POW2(MAX_QUEUE_LEN));
79 /* Protects against changes to 'dp_netdevs'. */
80 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
82 /* Contains all 'struct dp_netdev's. */
83 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
84 = SHASH_INITIALIZER(&dp_netdevs);
86 struct dp_netdev_upcall {
87 struct dpif_upcall upcall; /* Queued upcall information. */
88 struct ofpbuf buf; /* ofpbuf instance for upcall.packet. */
91 /* A queue passing packets from a struct dp_netdev to its clients (handlers).
97 * Any access at all requires the owning 'dp_netdev''s queue_rwlock and
99 struct dp_netdev_queue {
100 struct ovs_mutex mutex;
101 struct seq *seq; /* Incremented whenever a packet is queued. */
102 struct dp_netdev_upcall upcalls[MAX_QUEUE_LEN] OVS_GUARDED;
103 unsigned int head OVS_GUARDED;
104 unsigned int tail OVS_GUARDED;
107 /* Datapath based on the network device interface from netdev.h.
113 * Some members, marked 'const', are immutable. Accessing other members
114 * requires synchronization, as noted in more detail below.
116 * Acquisition order is, from outermost to innermost:
118 * dp_netdev_mutex (global)
125 const struct dpif_class *const class;
126 const char *const name;
127 struct ovs_refcount ref_cnt;
128 atomic_flag destroyed;
132 * Readers of 'cls' and 'flow_table' must take a 'cls->rwlock' read lock.
134 * Writers of 'cls' and 'flow_table' must take the 'flow_mutex' and then
135 * the 'cls->rwlock' write lock. (The outer 'flow_mutex' allows writers to
136 * atomically perform multiple operations on 'cls' and 'flow_table'.)
138 struct ovs_mutex flow_mutex;
139 struct classifier cls; /* Classifier. Protected by cls.rwlock. */
140 struct hmap flow_table OVS_GUARDED; /* Flow table. */
144 * 'queue_rwlock' protects the modification of 'handler_queues' and
145 * 'n_handlers'. The queue elements are protected by its
146 * 'handler_queues''s mutex. */
147 struct fat_rwlock queue_rwlock;
148 struct dp_netdev_queue *handler_queues;
153 * ovsthread_stats is internally synchronized. */
154 struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */
158 * Any lookup into 'ports' or any access to the dp_netdev_ports found
159 * through 'ports' requires taking 'port_rwlock'. */
160 struct ovs_rwlock port_rwlock;
161 struct hmap ports OVS_GUARDED;
162 struct seq *port_seq; /* Incremented whenever a port changes. */
164 /* Forwarding threads. */
165 struct latch exit_latch;
166 struct pmd_thread *pmd_threads;
167 size_t n_pmd_threads;
171 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
173 OVS_REQ_RDLOCK(dp->port_rwlock);
176 DP_STAT_HIT, /* Packets that matched in the flow table. */
177 DP_STAT_MISS, /* Packets that did not match. */
178 DP_STAT_LOST, /* Packets not passed up to the client. */
182 /* Contained by struct dp_netdev's 'stats' member. */
183 struct dp_netdev_stats {
184 struct ovs_mutex mutex; /* Protects 'n'. */
186 /* Indexed by DP_STAT_*, protected by 'mutex'. */
187 unsigned long long int n[DP_N_STATS] OVS_GUARDED;
191 /* A port in a netdev-based datapath. */
192 struct dp_netdev_port {
193 struct hmap_node node; /* Node in dp_netdev's 'ports'. */
195 struct netdev *netdev;
196 struct netdev_saved_flags *sf;
197 struct netdev_rxq **rxq;
198 struct ovs_refcount ref_cnt;
199 char *type; /* Port type as requested by user. */
202 /* A flow in dp_netdev's 'flow_table'.
208 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
209 * its dp_netdev's classifier. The text below calls this classifier 'cls'.
214 * The thread safety rules described here for "struct dp_netdev_flow" are
215 * motivated by two goals:
217 * - Prevent threads that read members of "struct dp_netdev_flow" from
218 * reading bad data due to changes by some thread concurrently modifying
221 * - Prevent two threads making changes to members of a given "struct
222 * dp_netdev_flow" from interfering with each other.
228 * A flow 'flow' may be accessed without a risk of being freed by code that
229 * holds a read-lock or write-lock on 'cls->rwlock' or that owns a reference to
230 * 'flow->ref_cnt' (or both). Code that needs to hold onto a flow for a while
231 * should take 'cls->rwlock', find the flow it needs, increment 'flow->ref_cnt'
232 * with dpif_netdev_flow_ref(), and drop 'cls->rwlock'.
234 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
235 * flow from being deleted from 'cls' (that's 'cls->rwlock') and it doesn't
236 * protect members of 'flow' from modification (that's 'flow->mutex').
238 * 'flow->mutex' protects the members of 'flow' from modification. It doesn't
239 * protect the flow from being deleted from 'cls' (that's 'cls->rwlock') and it
240 * doesn't prevent the flow from being freed (that's 'flow->ref_cnt').
242 * Some members, marked 'const', are immutable. Accessing other members
243 * requires synchronization, as noted in more detail below.
245 struct dp_netdev_flow {
246 /* Packet classification. */
247 const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
249 /* Hash table index by unmasked flow. */
250 const struct hmap_node node; /* In owning dp_netdev's 'flow_table'. */
251 const struct flow flow; /* The flow that created this entry. */
253 /* Protects members marked OVS_GUARDED.
255 * Acquire after datapath's flow_mutex. */
256 struct ovs_mutex mutex OVS_ACQ_AFTER(dp_netdev_mutex);
260 * Reading or writing these members requires 'mutex'. */
261 struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
265 * Reading 'actions' requires 'mutex'.
266 * Writing 'actions' requires 'mutex' and (to allow for transactions) the
267 * datapath's flow_mutex. */
268 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
271 static void dp_netdev_flow_free(struct dp_netdev_flow *);
273 /* Contained by struct dp_netdev_flow's 'stats' member. */
274 struct dp_netdev_flow_stats {
275 struct ovs_mutex mutex; /* Guards all the other members. */
277 long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
278 long long int packet_count OVS_GUARDED; /* Number of packets matched. */
279 long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
280 uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
283 /* A set of datapath actions within a "struct dp_netdev_flow".
289 * A struct dp_netdev_actions 'actions' may be accessed without a risk of being
290 * freed by code that holds a read-lock or write-lock on 'flow->mutex' (where
291 * 'flow' is the dp_netdev_flow for which 'flow->actions == actions') or that
292 * owns a reference to 'actions->ref_cnt' (or both). */
293 struct dp_netdev_actions {
294 /* These members are immutable: they do not change during the struct's
296 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
297 unsigned int size; /* Size of 'actions', in bytes. */
300 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
302 struct dp_netdev_actions *dp_netdev_flow_get_actions(
303 const struct dp_netdev_flow *);
304 static void dp_netdev_actions_free(struct dp_netdev_actions *);
306 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
307 * the performance overhead of interrupt processing. Therefore netdev can
308 * not implement rx-wait for these devices. dpif-netdev needs to poll
309 * these device to check for recv buffer. pmd-thread does polling for
310 * devices assigned to itself thread.
312 * DPDK used PMD for accessing NIC.
314 * A thread that receives packets from PMD ports, looks them up in the flow
315 * table, and executes the actions it finds.
318 struct dp_netdev *dp;
321 atomic_uint change_seq;
325 /* Interface to netdev-based datapath. */
328 struct dp_netdev *dp;
329 uint64_t last_port_seq;
332 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
333 struct dp_netdev_port **portp)
334 OVS_REQ_RDLOCK(dp->port_rwlock);
335 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
336 struct dp_netdev_port **portp)
337 OVS_REQ_RDLOCK(dp->port_rwlock);
338 static void dp_netdev_free(struct dp_netdev *)
339 OVS_REQUIRES(dp_netdev_mutex);
340 static void dp_netdev_flow_flush(struct dp_netdev *);
341 static int do_add_port(struct dp_netdev *dp, const char *devname,
342 const char *type, odp_port_t port_no)
343 OVS_REQ_WRLOCK(dp->port_rwlock);
344 static int do_del_port(struct dp_netdev *dp, odp_port_t port_no)
345 OVS_REQ_WRLOCK(dp->port_rwlock);
346 static void dp_netdev_destroy_all_queues(struct dp_netdev *dp)
347 OVS_REQ_WRLOCK(dp->queue_rwlock);
348 static int dpif_netdev_open(const struct dpif_class *, const char *name,
349 bool create, struct dpif **);
350 static int dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *,
351 int queue_no, int type,
353 const struct nlattr *userdata);
354 static void dp_netdev_execute_actions(struct dp_netdev *dp,
355 const struct flow *, struct ofpbuf *, bool may_steal,
356 struct pkt_metadata *,
357 const struct nlattr *actions,
359 static void dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet,
360 struct pkt_metadata *);
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)
380 struct shash_node *node;
382 ovs_mutex_lock(&dp_netdev_mutex);
383 SHASH_FOR_EACH(node, &dp_netdevs) {
384 sset_add(all_dps, node->name);
386 ovs_mutex_unlock(&dp_netdev_mutex);
392 dpif_netdev_class_is_dummy(const struct dpif_class *class)
394 return class != &dpif_netdev_class;
398 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
400 return strcmp(type, "internal") ? type
401 : dpif_netdev_class_is_dummy(class) ? "dummy"
406 create_dpif_netdev(struct dp_netdev *dp)
408 uint16_t netflow_id = hash_string(dp->name, 0);
409 struct dpif_netdev *dpif;
411 ovs_refcount_ref(&dp->ref_cnt);
413 dpif = xmalloc(sizeof *dpif);
414 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
416 dpif->last_port_seq = seq_read(dp->port_seq);
421 /* Choose an unused, non-zero port number and return it on success.
422 * Return ODPP_NONE on failure. */
424 choose_port(struct dp_netdev *dp, const char *name)
425 OVS_REQ_RDLOCK(dp->port_rwlock)
429 if (dp->class != &dpif_netdev_class) {
433 /* If the port name begins with "br", start the number search at
434 * 100 to make writing tests easier. */
435 if (!strncmp(name, "br", 2)) {
439 /* If the port name contains a number, try to assign that port number.
440 * This can make writing unit tests easier because port numbers are
442 for (p = name; *p != '\0'; p++) {
443 if (isdigit((unsigned char) *p)) {
444 port_no = start_no + strtol(p, NULL, 10);
445 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
446 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
447 return u32_to_odp(port_no);
454 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
455 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
456 return u32_to_odp(port_no);
464 create_dp_netdev(const char *name, const struct dpif_class *class,
465 struct dp_netdev **dpp)
466 OVS_REQUIRES(dp_netdev_mutex)
468 struct dp_netdev *dp;
471 dp = xzalloc(sizeof *dp);
472 shash_add(&dp_netdevs, name, dp);
474 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
475 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
476 ovs_refcount_init(&dp->ref_cnt);
477 atomic_flag_clear(&dp->destroyed);
479 ovs_mutex_init(&dp->flow_mutex);
480 classifier_init(&dp->cls, NULL);
481 hmap_init(&dp->flow_table);
483 fat_rwlock_init(&dp->queue_rwlock);
485 ovsthread_stats_init(&dp->stats);
487 ovs_rwlock_init(&dp->port_rwlock);
488 hmap_init(&dp->ports);
489 dp->port_seq = seq_create();
490 latch_init(&dp->exit_latch);
492 ovs_rwlock_wrlock(&dp->port_rwlock);
493 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
494 ovs_rwlock_unlock(&dp->port_rwlock);
505 dpif_netdev_open(const struct dpif_class *class, const char *name,
506 bool create, struct dpif **dpifp)
508 struct dp_netdev *dp;
511 ovs_mutex_lock(&dp_netdev_mutex);
512 dp = shash_find_data(&dp_netdevs, name);
514 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
516 error = (dp->class != class ? EINVAL
521 *dpifp = create_dpif_netdev(dp);
523 ovs_mutex_unlock(&dp_netdev_mutex);
529 dp_netdev_purge_queues(struct dp_netdev *dp)
530 OVS_REQ_WRLOCK(dp->queue_rwlock)
534 for (i = 0; i < dp->n_handlers; i++) {
535 struct dp_netdev_queue *q = &dp->handler_queues[i];
537 ovs_mutex_lock(&q->mutex);
538 while (q->tail != q->head) {
539 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
540 ofpbuf_uninit(&u->upcall.packet);
541 ofpbuf_uninit(&u->buf);
543 ovs_mutex_unlock(&q->mutex);
547 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
548 * through the 'dp_netdevs' shash while freeing 'dp'. */
550 dp_netdev_free(struct dp_netdev *dp)
551 OVS_REQUIRES(dp_netdev_mutex)
553 struct dp_netdev_port *port, *next;
554 struct dp_netdev_stats *bucket;
557 shash_find_and_delete(&dp_netdevs, dp->name);
559 dp_netdev_set_pmd_threads(dp, 0);
560 free(dp->pmd_threads);
562 dp_netdev_flow_flush(dp);
563 ovs_rwlock_wrlock(&dp->port_rwlock);
564 HMAP_FOR_EACH_SAFE (port, next, node, &dp->ports) {
565 do_del_port(dp, port->port_no);
567 ovs_rwlock_unlock(&dp->port_rwlock);
569 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
570 ovs_mutex_destroy(&bucket->mutex);
571 free_cacheline(bucket);
573 ovsthread_stats_destroy(&dp->stats);
575 fat_rwlock_wrlock(&dp->queue_rwlock);
576 dp_netdev_destroy_all_queues(dp);
577 fat_rwlock_unlock(&dp->queue_rwlock);
579 fat_rwlock_destroy(&dp->queue_rwlock);
581 classifier_destroy(&dp->cls);
582 hmap_destroy(&dp->flow_table);
583 ovs_mutex_destroy(&dp->flow_mutex);
584 seq_destroy(dp->port_seq);
585 hmap_destroy(&dp->ports);
586 latch_destroy(&dp->exit_latch);
587 free(CONST_CAST(char *, dp->name));
592 dp_netdev_unref(struct dp_netdev *dp)
595 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
596 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
597 ovs_mutex_lock(&dp_netdev_mutex);
598 if (ovs_refcount_unref(&dp->ref_cnt) == 1) {
601 ovs_mutex_unlock(&dp_netdev_mutex);
606 dpif_netdev_close(struct dpif *dpif)
608 struct dp_netdev *dp = get_dp_netdev(dpif);
615 dpif_netdev_destroy(struct dpif *dpif)
617 struct dp_netdev *dp = get_dp_netdev(dpif);
619 if (!atomic_flag_test_and_set(&dp->destroyed)) {
620 if (ovs_refcount_unref(&dp->ref_cnt) == 1) {
621 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
630 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
632 struct dp_netdev *dp = get_dp_netdev(dpif);
633 struct dp_netdev_stats *bucket;
636 fat_rwlock_rdlock(&dp->cls.rwlock);
637 stats->n_flows = hmap_count(&dp->flow_table);
638 fat_rwlock_unlock(&dp->cls.rwlock);
640 stats->n_hit = stats->n_missed = stats->n_lost = 0;
641 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
642 ovs_mutex_lock(&bucket->mutex);
643 stats->n_hit += bucket->n[DP_STAT_HIT];
644 stats->n_missed += bucket->n[DP_STAT_MISS];
645 stats->n_lost += bucket->n[DP_STAT_LOST];
646 ovs_mutex_unlock(&bucket->mutex);
648 stats->n_masks = UINT32_MAX;
649 stats->n_mask_hit = UINT64_MAX;
655 dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
659 for (i = 0; i < dp->n_pmd_threads; i++) {
660 struct pmd_thread *f = &dp->pmd_threads[i];
663 atomic_add(&f->change_seq, 1, &id);
668 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
670 OVS_REQ_WRLOCK(dp->port_rwlock)
672 struct netdev_saved_flags *sf;
673 struct dp_netdev_port *port;
674 struct netdev *netdev;
675 enum netdev_flags flags;
676 const char *open_type;
680 /* XXX reject devices already in some dp_netdev. */
682 /* Open and validate network device. */
683 open_type = dpif_netdev_port_open_type(dp->class, type);
684 error = netdev_open(devname, open_type, &netdev);
688 /* XXX reject non-Ethernet devices */
690 netdev_get_flags(netdev, &flags);
691 if (flags & NETDEV_LOOPBACK) {
692 VLOG_ERR("%s: cannot add a loopback device", devname);
693 netdev_close(netdev);
697 port = xzalloc(sizeof *port);
698 port->port_no = port_no;
699 port->netdev = netdev;
700 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
701 port->type = xstrdup(type);
702 for (i = 0; i < netdev_n_rxq(netdev); i++) {
703 error = netdev_rxq_open(netdev, &port->rxq[i], i);
705 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
706 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
707 devname, ovs_strerror(errno));
708 netdev_close(netdev);
713 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
715 for (i = 0; i < netdev_n_rxq(netdev); i++) {
716 netdev_rxq_close(port->rxq[i]);
718 netdev_close(netdev);
725 if (netdev_is_pmd(netdev)) {
727 dp_netdev_set_pmd_threads(dp, NR_THREADS);
728 dp_netdev_reload_pmd_threads(dp);
730 ovs_refcount_init(&port->ref_cnt);
732 hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0));
733 seq_change(dp->port_seq);
739 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
740 odp_port_t *port_nop)
742 struct dp_netdev *dp = get_dp_netdev(dpif);
743 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
744 const char *dpif_port;
748 ovs_rwlock_wrlock(&dp->port_rwlock);
749 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
750 if (*port_nop != ODPP_NONE) {
752 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
754 port_no = choose_port(dp, dpif_port);
755 error = port_no == ODPP_NONE ? EFBIG : 0;
759 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
761 ovs_rwlock_unlock(&dp->port_rwlock);
767 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
769 struct dp_netdev *dp = get_dp_netdev(dpif);
772 ovs_rwlock_wrlock(&dp->port_rwlock);
773 error = port_no == ODPP_LOCAL ? EINVAL : do_del_port(dp, port_no);
774 ovs_rwlock_unlock(&dp->port_rwlock);
780 is_valid_port_number(odp_port_t port_no)
782 return port_no != ODPP_NONE;
785 static struct dp_netdev_port *
786 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
787 OVS_REQ_RDLOCK(dp->port_rwlock)
789 struct dp_netdev_port *port;
791 HMAP_FOR_EACH_IN_BUCKET (port, node, hash_int(odp_to_u32(port_no), 0),
793 if (port->port_no == port_no) {
801 get_port_by_number(struct dp_netdev *dp,
802 odp_port_t port_no, struct dp_netdev_port **portp)
803 OVS_REQ_RDLOCK(dp->port_rwlock)
805 if (!is_valid_port_number(port_no)) {
809 *portp = dp_netdev_lookup_port(dp, port_no);
810 return *portp ? 0 : ENOENT;
815 port_ref(struct dp_netdev_port *port)
818 ovs_refcount_ref(&port->ref_cnt);
823 port_unref(struct dp_netdev_port *port)
825 if (port && ovs_refcount_unref(&port->ref_cnt) == 1) {
828 netdev_close(port->netdev);
829 netdev_restore_flags(port->sf);
831 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
832 netdev_rxq_close(port->rxq[i]);
840 get_port_by_name(struct dp_netdev *dp,
841 const char *devname, struct dp_netdev_port **portp)
842 OVS_REQ_RDLOCK(dp->port_rwlock)
844 struct dp_netdev_port *port;
846 HMAP_FOR_EACH (port, node, &dp->ports) {
847 if (!strcmp(netdev_get_name(port->netdev), devname)) {
856 do_del_port(struct dp_netdev *dp, odp_port_t port_no)
857 OVS_REQ_WRLOCK(dp->port_rwlock)
859 struct dp_netdev_port *port;
862 error = get_port_by_number(dp, port_no, &port);
867 hmap_remove(&dp->ports, &port->node);
868 seq_change(dp->port_seq);
869 if (netdev_is_pmd(port->netdev)) {
870 dp_netdev_reload_pmd_threads(dp);
878 answer_port_query(const struct dp_netdev_port *port,
879 struct dpif_port *dpif_port)
881 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
882 dpif_port->type = xstrdup(port->type);
883 dpif_port->port_no = port->port_no;
887 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
888 struct dpif_port *dpif_port)
890 struct dp_netdev *dp = get_dp_netdev(dpif);
891 struct dp_netdev_port *port;
894 ovs_rwlock_rdlock(&dp->port_rwlock);
895 error = get_port_by_number(dp, port_no, &port);
896 if (!error && dpif_port) {
897 answer_port_query(port, dpif_port);
899 ovs_rwlock_unlock(&dp->port_rwlock);
905 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
906 struct dpif_port *dpif_port)
908 struct dp_netdev *dp = get_dp_netdev(dpif);
909 struct dp_netdev_port *port;
912 ovs_rwlock_rdlock(&dp->port_rwlock);
913 error = get_port_by_name(dp, devname, &port);
914 if (!error && dpif_port) {
915 answer_port_query(port, dpif_port);
917 ovs_rwlock_unlock(&dp->port_rwlock);
923 dp_netdev_flow_free(struct dp_netdev_flow *flow)
925 struct dp_netdev_flow_stats *bucket;
928 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
929 ovs_mutex_destroy(&bucket->mutex);
930 free_cacheline(bucket);
932 ovsthread_stats_destroy(&flow->stats);
934 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
935 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
936 ovs_mutex_destroy(&flow->mutex);
941 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
942 OVS_REQ_WRLOCK(dp->cls.rwlock)
943 OVS_REQUIRES(dp->flow_mutex)
945 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
946 struct hmap_node *node = CONST_CAST(struct hmap_node *, &flow->node);
948 classifier_remove(&dp->cls, cr);
949 hmap_remove(&dp->flow_table, node);
950 ovsrcu_postpone(dp_netdev_flow_free, flow);
954 dp_netdev_flow_flush(struct dp_netdev *dp)
956 struct dp_netdev_flow *netdev_flow, *next;
958 ovs_mutex_lock(&dp->flow_mutex);
959 fat_rwlock_wrlock(&dp->cls.rwlock);
960 HMAP_FOR_EACH_SAFE (netdev_flow, next, node, &dp->flow_table) {
961 dp_netdev_remove_flow(dp, netdev_flow);
963 fat_rwlock_unlock(&dp->cls.rwlock);
964 ovs_mutex_unlock(&dp->flow_mutex);
968 dpif_netdev_flow_flush(struct dpif *dpif)
970 struct dp_netdev *dp = get_dp_netdev(dpif);
972 dp_netdev_flow_flush(dp);
976 struct dp_netdev_port_state {
983 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
985 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
990 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
991 struct dpif_port *dpif_port)
993 struct dp_netdev_port_state *state = state_;
994 struct dp_netdev *dp = get_dp_netdev(dpif);
995 struct hmap_node *node;
998 ovs_rwlock_rdlock(&dp->port_rwlock);
999 node = hmap_at_position(&dp->ports, &state->bucket, &state->offset);
1001 struct dp_netdev_port *port;
1003 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1006 state->name = xstrdup(netdev_get_name(port->netdev));
1007 dpif_port->name = state->name;
1008 dpif_port->type = port->type;
1009 dpif_port->port_no = port->port_no;
1015 ovs_rwlock_unlock(&dp->port_rwlock);
1021 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1023 struct dp_netdev_port_state *state = state_;
1030 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1032 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1033 uint64_t new_port_seq;
1036 new_port_seq = seq_read(dpif->dp->port_seq);
1037 if (dpif->last_port_seq != new_port_seq) {
1038 dpif->last_port_seq = new_port_seq;
1048 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1050 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1052 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1055 static struct dp_netdev_flow *
1056 dp_netdev_flow_cast(const struct cls_rule *cr)
1058 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1061 static struct dp_netdev_flow *
1062 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct flow *flow)
1063 OVS_EXCLUDED(dp->cls.rwlock)
1065 struct dp_netdev_flow *netdev_flow;
1067 fat_rwlock_rdlock(&dp->cls.rwlock);
1068 netdev_flow = dp_netdev_flow_cast(classifier_lookup(&dp->cls, flow, NULL));
1069 fat_rwlock_unlock(&dp->cls.rwlock);
1074 static struct dp_netdev_flow *
1075 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1076 OVS_REQ_RDLOCK(dp->cls.rwlock)
1078 struct dp_netdev_flow *netdev_flow;
1080 HMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1082 if (flow_equal(&netdev_flow->flow, flow)) {
1091 get_dpif_flow_stats(struct dp_netdev_flow *netdev_flow,
1092 struct dpif_flow_stats *stats)
1094 struct dp_netdev_flow_stats *bucket;
1097 memset(stats, 0, sizeof *stats);
1098 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1099 ovs_mutex_lock(&bucket->mutex);
1100 stats->n_packets += bucket->packet_count;
1101 stats->n_bytes += bucket->byte_count;
1102 stats->used = MAX(stats->used, bucket->used);
1103 stats->tcp_flags |= bucket->tcp_flags;
1104 ovs_mutex_unlock(&bucket->mutex);
1109 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1110 const struct nlattr *mask_key,
1111 uint32_t mask_key_len, const struct flow *flow,
1115 enum odp_key_fitness fitness;
1117 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1119 /* This should not happen: it indicates that
1120 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1121 * disagree on the acceptable form of a mask. Log the problem
1122 * as an error, with enough details to enable debugging. */
1123 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1125 if (!VLOG_DROP_ERR(&rl)) {
1129 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1131 VLOG_ERR("internal error parsing flow mask %s (%s)",
1132 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1138 /* Force unwildcard the in_port. */
1139 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
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);
1162 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1167 if (odp_flow_key_to_flow(key, key_len, flow)) {
1168 /* This should not happen: it indicates that odp_flow_key_from_flow()
1169 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1170 * flow. Log the problem as an error, with enough details to enable
1172 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1174 if (!VLOG_DROP_ERR(&rl)) {
1178 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1179 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1186 in_port = flow->in_port.odp_port;
1187 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1195 dpif_netdev_flow_get(const struct dpif *dpif,
1196 const struct nlattr *nl_key, size_t nl_key_len,
1197 struct ofpbuf **actionsp, struct dpif_flow_stats *stats)
1199 struct dp_netdev *dp = get_dp_netdev(dpif);
1200 struct dp_netdev_flow *netdev_flow;
1204 error = dpif_netdev_flow_from_nlattrs(nl_key, nl_key_len, &key);
1209 fat_rwlock_rdlock(&dp->cls.rwlock);
1210 netdev_flow = dp_netdev_find_flow(dp, &key);
1211 fat_rwlock_unlock(&dp->cls.rwlock);
1215 get_dpif_flow_stats(netdev_flow, stats);
1219 struct dp_netdev_actions *actions;
1221 actions = dp_netdev_flow_get_actions(netdev_flow);
1222 *actionsp = ofpbuf_clone_data(actions->actions, actions->size);
1232 dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow,
1233 const struct flow_wildcards *wc,
1234 const struct nlattr *actions,
1236 OVS_REQUIRES(dp->flow_mutex)
1238 struct dp_netdev_flow *netdev_flow;
1241 netdev_flow = xzalloc(sizeof *netdev_flow);
1242 *CONST_CAST(struct flow *, &netdev_flow->flow) = *flow;
1244 ovs_mutex_init(&netdev_flow->mutex);
1246 ovsthread_stats_init(&netdev_flow->stats);
1248 ovsrcu_set(&netdev_flow->actions,
1249 dp_netdev_actions_create(actions, actions_len));
1251 match_init(&match, flow, wc);
1252 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1253 &match, NETDEV_RULE_PRIORITY);
1254 fat_rwlock_wrlock(&dp->cls.rwlock);
1255 classifier_insert(&dp->cls,
1256 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1257 hmap_insert(&dp->flow_table,
1258 CONST_CAST(struct hmap_node *, &netdev_flow->node),
1259 flow_hash(flow, 0));
1260 fat_rwlock_unlock(&dp->cls.rwlock);
1266 clear_stats(struct dp_netdev_flow *netdev_flow)
1268 struct dp_netdev_flow_stats *bucket;
1271 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1272 ovs_mutex_lock(&bucket->mutex);
1274 bucket->packet_count = 0;
1275 bucket->byte_count = 0;
1276 bucket->tcp_flags = 0;
1277 ovs_mutex_unlock(&bucket->mutex);
1282 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1284 struct dp_netdev *dp = get_dp_netdev(dpif);
1285 struct dp_netdev_flow *netdev_flow;
1287 struct flow_wildcards wc;
1290 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &flow);
1294 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1295 put->mask, put->mask_len,
1301 ovs_mutex_lock(&dp->flow_mutex);
1302 netdev_flow = dp_netdev_lookup_flow(dp, &flow);
1304 if (put->flags & DPIF_FP_CREATE) {
1305 if (hmap_count(&dp->flow_table) < MAX_FLOWS) {
1307 memset(put->stats, 0, sizeof *put->stats);
1309 error = dp_netdev_flow_add(dp, &flow, &wc, put->actions,
1318 if (put->flags & DPIF_FP_MODIFY
1319 && flow_equal(&flow, &netdev_flow->flow)) {
1320 struct dp_netdev_actions *new_actions;
1321 struct dp_netdev_actions *old_actions;
1323 new_actions = dp_netdev_actions_create(put->actions,
1326 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1327 ovsrcu_set(&netdev_flow->actions, new_actions);
1330 get_dpif_flow_stats(netdev_flow, put->stats);
1332 if (put->flags & DPIF_FP_ZERO_STATS) {
1333 clear_stats(netdev_flow);
1336 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1337 } else if (put->flags & DPIF_FP_CREATE) {
1340 /* Overlapping flow. */
1344 ovs_mutex_unlock(&dp->flow_mutex);
1350 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1352 struct dp_netdev *dp = get_dp_netdev(dpif);
1353 struct dp_netdev_flow *netdev_flow;
1357 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1362 ovs_mutex_lock(&dp->flow_mutex);
1363 fat_rwlock_wrlock(&dp->cls.rwlock);
1364 netdev_flow = dp_netdev_find_flow(dp, &key);
1367 get_dpif_flow_stats(netdev_flow, del->stats);
1369 dp_netdev_remove_flow(dp, netdev_flow);
1373 fat_rwlock_unlock(&dp->cls.rwlock);
1374 ovs_mutex_unlock(&dp->flow_mutex);
1379 struct dp_netdev_flow_state {
1380 struct dp_netdev_actions *actions;
1381 struct odputil_keybuf keybuf;
1382 struct odputil_keybuf maskbuf;
1383 struct dpif_flow_stats stats;
1386 struct dp_netdev_flow_iter {
1390 struct ovs_mutex mutex;
1394 dpif_netdev_flow_dump_state_init(void **statep)
1396 struct dp_netdev_flow_state *state;
1398 *statep = state = xmalloc(sizeof *state);
1399 state->actions = NULL;
1403 dpif_netdev_flow_dump_state_uninit(void *state_)
1405 struct dp_netdev_flow_state *state = state_;
1411 dpif_netdev_flow_dump_start(const struct dpif *dpif OVS_UNUSED, void **iterp)
1413 struct dp_netdev_flow_iter *iter;
1415 *iterp = iter = xmalloc(sizeof *iter);
1419 ovs_mutex_init(&iter->mutex);
1423 /* XXX the caller must use 'actions' without quiescing */
1425 dpif_netdev_flow_dump_next(const struct dpif *dpif, void *iter_, void *state_,
1426 const struct nlattr **key, size_t *key_len,
1427 const struct nlattr **mask, size_t *mask_len,
1428 const struct nlattr **actions, size_t *actions_len,
1429 const struct dpif_flow_stats **stats)
1431 struct dp_netdev_flow_iter *iter = iter_;
1432 struct dp_netdev_flow_state *state = state_;
1433 struct dp_netdev *dp = get_dp_netdev(dpif);
1434 struct dp_netdev_flow *netdev_flow;
1437 ovs_mutex_lock(&iter->mutex);
1438 error = iter->status;
1440 struct hmap_node *node;
1442 fat_rwlock_rdlock(&dp->cls.rwlock);
1443 node = hmap_at_position(&dp->flow_table, &iter->bucket, &iter->offset);
1445 netdev_flow = CONTAINER_OF(node, struct dp_netdev_flow, node);
1447 fat_rwlock_unlock(&dp->cls.rwlock);
1449 iter->status = error = EOF;
1452 ovs_mutex_unlock(&iter->mutex);
1460 ofpbuf_use_stack(&buf, &state->keybuf, sizeof state->keybuf);
1461 odp_flow_key_from_flow(&buf, &netdev_flow->flow,
1462 netdev_flow->flow.in_port.odp_port);
1465 *key_len = buf.size;
1470 struct flow_wildcards wc;
1472 ofpbuf_use_stack(&buf, &state->maskbuf, sizeof state->maskbuf);
1473 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1474 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1475 odp_to_u32(wc.masks.in_port.odp_port),
1479 *mask_len = buf.size;
1482 if (actions || stats) {
1483 state->actions = NULL;
1486 state->actions = dp_netdev_flow_get_actions(netdev_flow);
1487 *actions = state->actions->actions;
1488 *actions_len = state->actions->size;
1492 get_dpif_flow_stats(netdev_flow, &state->stats);
1493 *stats = &state->stats;
1501 dpif_netdev_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *iter_)
1503 struct dp_netdev_flow_iter *iter = iter_;
1505 ovs_mutex_destroy(&iter->mutex);
1511 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1513 struct dp_netdev *dp = get_dp_netdev(dpif);
1514 struct pkt_metadata *md = &execute->md;
1517 if (execute->packet->size < ETH_HEADER_LEN ||
1518 execute->packet->size > UINT16_MAX) {
1522 /* Extract flow key. */
1523 flow_extract(execute->packet, md, &key);
1525 ovs_rwlock_rdlock(&dp->port_rwlock);
1526 dp_netdev_execute_actions(dp, &key, execute->packet, false, md,
1527 execute->actions, execute->actions_len);
1528 ovs_rwlock_unlock(&dp->port_rwlock);
1534 dp_netdev_destroy_all_queues(struct dp_netdev *dp)
1535 OVS_REQ_WRLOCK(dp->queue_rwlock)
1539 dp_netdev_purge_queues(dp);
1541 for (i = 0; i < dp->n_handlers; i++) {
1542 struct dp_netdev_queue *q = &dp->handler_queues[i];
1544 ovs_mutex_destroy(&q->mutex);
1545 seq_destroy(q->seq);
1547 free(dp->handler_queues);
1548 dp->handler_queues = NULL;
1553 dp_netdev_refresh_queues(struct dp_netdev *dp, uint32_t n_handlers)
1554 OVS_REQ_WRLOCK(dp->queue_rwlock)
1556 if (dp->n_handlers != n_handlers) {
1559 dp_netdev_destroy_all_queues(dp);
1561 dp->n_handlers = n_handlers;
1562 dp->handler_queues = xzalloc(n_handlers * sizeof *dp->handler_queues);
1564 for (i = 0; i < n_handlers; i++) {
1565 struct dp_netdev_queue *q = &dp->handler_queues[i];
1567 ovs_mutex_init(&q->mutex);
1568 q->seq = seq_create();
1574 dpif_netdev_recv_set(struct dpif *dpif, bool enable)
1576 struct dp_netdev *dp = get_dp_netdev(dpif);
1578 if ((dp->handler_queues != NULL) == enable) {
1582 fat_rwlock_wrlock(&dp->queue_rwlock);
1584 dp_netdev_destroy_all_queues(dp);
1586 dp_netdev_refresh_queues(dp, 1);
1588 fat_rwlock_unlock(&dp->queue_rwlock);
1594 dpif_netdev_handlers_set(struct dpif *dpif, uint32_t n_handlers)
1596 struct dp_netdev *dp = get_dp_netdev(dpif);
1598 fat_rwlock_wrlock(&dp->queue_rwlock);
1599 if (dp->handler_queues) {
1600 dp_netdev_refresh_queues(dp, n_handlers);
1602 fat_rwlock_unlock(&dp->queue_rwlock);
1608 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1609 uint32_t queue_id, uint32_t *priority)
1611 *priority = queue_id;
1616 dp_netdev_recv_check(const struct dp_netdev *dp, const uint32_t handler_id)
1617 OVS_REQ_RDLOCK(dp->queue_rwlock)
1619 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1621 if (!dp->handler_queues) {
1622 VLOG_WARN_RL(&rl, "receiving upcall disabled");
1626 if (handler_id >= dp->n_handlers) {
1627 VLOG_WARN_RL(&rl, "handler index out of bound");
1635 dpif_netdev_recv(struct dpif *dpif, uint32_t handler_id,
1636 struct dpif_upcall *upcall, struct ofpbuf *buf)
1638 struct dp_netdev *dp = get_dp_netdev(dpif);
1639 struct dp_netdev_queue *q;
1642 fat_rwlock_rdlock(&dp->queue_rwlock);
1644 if (!dp_netdev_recv_check(dp, handler_id)) {
1649 q = &dp->handler_queues[handler_id];
1650 ovs_mutex_lock(&q->mutex);
1651 if (q->head != q->tail) {
1652 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
1654 *upcall = u->upcall;
1661 ovs_mutex_unlock(&q->mutex);
1664 fat_rwlock_unlock(&dp->queue_rwlock);
1670 dpif_netdev_recv_wait(struct dpif *dpif, uint32_t handler_id)
1672 struct dp_netdev *dp = get_dp_netdev(dpif);
1673 struct dp_netdev_queue *q;
1676 fat_rwlock_rdlock(&dp->queue_rwlock);
1678 if (!dp_netdev_recv_check(dp, handler_id)) {
1682 q = &dp->handler_queues[handler_id];
1683 ovs_mutex_lock(&q->mutex);
1684 seq = seq_read(q->seq);
1685 if (q->head != q->tail) {
1686 poll_immediate_wake();
1688 seq_wait(q->seq, seq);
1691 ovs_mutex_unlock(&q->mutex);
1694 fat_rwlock_unlock(&dp->queue_rwlock);
1698 dpif_netdev_recv_purge(struct dpif *dpif)
1700 struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif);
1702 fat_rwlock_wrlock(&dpif_netdev->dp->queue_rwlock);
1703 dp_netdev_purge_queues(dpif_netdev->dp);
1704 fat_rwlock_unlock(&dpif_netdev->dp->queue_rwlock);
1707 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1708 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1710 struct dp_netdev_actions *
1711 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1713 struct dp_netdev_actions *netdev_actions;
1715 netdev_actions = xmalloc(sizeof *netdev_actions);
1716 netdev_actions->actions = xmemdup(actions, size);
1717 netdev_actions->size = size;
1719 return netdev_actions;
1722 struct dp_netdev_actions *
1723 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
1725 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
1729 dp_netdev_actions_free(struct dp_netdev_actions *actions)
1731 free(actions->actions);
1737 dp_netdev_process_rxq_port(struct dp_netdev *dp,
1738 struct dp_netdev_port *port,
1739 struct netdev_rxq *rxq)
1741 struct ofpbuf *packet[NETDEV_MAX_RX_BATCH];
1744 error = netdev_rxq_recv(rxq, packet, &c);
1746 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port->port_no);
1749 for (i = 0; i < c; i++) {
1750 dp_netdev_port_input(dp, packet[i], &md);
1752 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1753 static struct vlog_rate_limit rl
1754 = VLOG_RATE_LIMIT_INIT(1, 5);
1756 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1757 netdev_get_name(port->netdev),
1758 ovs_strerror(error));
1763 dpif_netdev_run(struct dpif *dpif)
1765 struct dp_netdev_port *port;
1766 struct dp_netdev *dp = get_dp_netdev(dpif);
1768 ovs_rwlock_rdlock(&dp->port_rwlock);
1770 HMAP_FOR_EACH (port, node, &dp->ports) {
1771 if (!netdev_is_pmd(port->netdev)) {
1774 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1775 dp_netdev_process_rxq_port(dp, port, port->rxq[i]);
1780 ovs_rwlock_unlock(&dp->port_rwlock);
1784 dpif_netdev_wait(struct dpif *dpif)
1786 struct dp_netdev_port *port;
1787 struct dp_netdev *dp = get_dp_netdev(dpif);
1789 ovs_rwlock_rdlock(&dp->port_rwlock);
1791 HMAP_FOR_EACH (port, node, &dp->ports) {
1792 if (!netdev_is_pmd(port->netdev)) {
1795 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1796 netdev_rxq_wait(port->rxq[i]);
1800 ovs_rwlock_unlock(&dp->port_rwlock);
1804 struct dp_netdev_port *port;
1805 struct netdev_rxq *rx;
1809 pmd_load_queues(struct pmd_thread *f,
1810 struct rxq_poll **ppoll_list, int poll_cnt)
1812 struct dp_netdev *dp = f->dp;
1813 struct rxq_poll *poll_list = *ppoll_list;
1814 struct dp_netdev_port *port;
1819 /* Simple scheduler for netdev rx polling. */
1820 ovs_rwlock_rdlock(&dp->port_rwlock);
1821 for (i = 0; i < poll_cnt; i++) {
1822 port_unref(poll_list[i].port);
1828 HMAP_FOR_EACH (port, node, &f->dp->ports) {
1829 if (netdev_is_pmd(port->netdev)) {
1832 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1833 if ((index % dp->n_pmd_threads) == id) {
1834 poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1));
1837 poll_list[poll_cnt].port = port;
1838 poll_list[poll_cnt].rx = port->rxq[i];
1846 ovs_rwlock_unlock(&dp->port_rwlock);
1847 *ppoll_list = poll_list;
1852 pmd_thread_main(void *f_)
1854 struct pmd_thread *f = f_;
1855 struct dp_netdev *dp = f->dp;
1856 unsigned int lc = 0;
1857 struct rxq_poll *poll_list;
1858 unsigned int port_seq;
1862 f->name = xasprintf("pmd_%u", ovsthread_id_self());
1863 set_subprogram_name("%s", f->name);
1868 poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
1869 atomic_read(&f->change_seq, &port_seq);
1872 unsigned int c_port_seq;
1875 for (i = 0; i < poll_cnt; i++) {
1876 dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx);
1882 /* TODO: need completely userspace based signaling method.
1883 * to keep this thread entirely in userspace.
1884 * For now using atomic counter. */
1886 atomic_read_explicit(&f->change_seq, &c_port_seq, memory_order_consume);
1887 if (c_port_seq != port_seq) {
1893 if (!latch_is_set(&f->dp->exit_latch)){
1897 for (i = 0; i < poll_cnt; i++) {
1898 port_unref(poll_list[i].port);
1907 dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n)
1911 if (n == dp->n_pmd_threads) {
1915 /* Stop existing threads. */
1916 latch_set(&dp->exit_latch);
1917 dp_netdev_reload_pmd_threads(dp);
1918 for (i = 0; i < dp->n_pmd_threads; i++) {
1919 struct pmd_thread *f = &dp->pmd_threads[i];
1921 xpthread_join(f->thread, NULL);
1923 latch_poll(&dp->exit_latch);
1924 free(dp->pmd_threads);
1926 /* Start new threads. */
1927 dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
1928 dp->n_pmd_threads = n;
1930 for (i = 0; i < n; i++) {
1931 struct pmd_thread *f = &dp->pmd_threads[i];
1935 atomic_store(&f->change_seq, 1);
1937 /* Each thread will distribute all devices rx-queues among
1939 xpthread_create(&f->thread, NULL, pmd_thread_main, f);
1945 dp_netdev_flow_stats_new_cb(void)
1947 struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1948 ovs_mutex_init(&bucket->mutex);
1953 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
1954 const struct ofpbuf *packet,
1955 const struct flow *key)
1957 uint16_t tcp_flags = ntohs(key->tcp_flags);
1958 long long int now = time_msec();
1959 struct dp_netdev_flow_stats *bucket;
1961 bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
1962 dp_netdev_flow_stats_new_cb);
1964 ovs_mutex_lock(&bucket->mutex);
1965 bucket->used = MAX(now, bucket->used);
1966 bucket->packet_count++;
1967 bucket->byte_count += packet->size;
1968 bucket->tcp_flags |= tcp_flags;
1969 ovs_mutex_unlock(&bucket->mutex);
1973 dp_netdev_stats_new_cb(void)
1975 struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1976 ovs_mutex_init(&bucket->mutex);
1981 dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type)
1983 struct dp_netdev_stats *bucket;
1985 bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
1986 ovs_mutex_lock(&bucket->mutex);
1988 ovs_mutex_unlock(&bucket->mutex);
1992 dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet,
1993 struct pkt_metadata *md)
1995 struct dp_netdev_flow *netdev_flow;
1998 if (packet->size < ETH_HEADER_LEN) {
1999 ofpbuf_delete(packet);
2002 flow_extract(packet, md, &key);
2003 netdev_flow = dp_netdev_lookup_flow(dp, &key);
2005 struct dp_netdev_actions *actions;
2007 dp_netdev_flow_used(netdev_flow, packet, &key);
2009 actions = dp_netdev_flow_get_actions(netdev_flow);
2010 dp_netdev_execute_actions(dp, &key, packet, true, md,
2011 actions->actions, actions->size);
2012 dp_netdev_count_packet(dp, DP_STAT_HIT);
2013 } else if (dp->handler_queues) {
2014 dp_netdev_count_packet(dp, DP_STAT_MISS);
2015 dp_netdev_output_userspace(dp, packet,
2016 flow_hash_5tuple(&key, 0) % dp->n_handlers,
2017 DPIF_UC_MISS, &key, NULL);
2018 ofpbuf_delete(packet);
2023 dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet,
2024 int queue_no, int type, const struct flow *flow,
2025 const struct nlattr *userdata)
2027 struct dp_netdev_queue *q;
2030 fat_rwlock_rdlock(&dp->queue_rwlock);
2031 q = &dp->handler_queues[queue_no];
2032 ovs_mutex_lock(&q->mutex);
2033 if (q->head - q->tail < MAX_QUEUE_LEN) {
2034 struct dp_netdev_upcall *u = &q->upcalls[q->head++ & QUEUE_MASK];
2035 struct dpif_upcall *upcall = &u->upcall;
2036 struct ofpbuf *buf = &u->buf;
2039 upcall->type = type;
2041 /* Allocate buffer big enough for everything. */
2042 buf_size = ODPUTIL_FLOW_KEY_BYTES;
2044 buf_size += NLA_ALIGN(userdata->nla_len);
2046 buf_size += packet->size;
2047 ofpbuf_init(buf, buf_size);
2050 odp_flow_key_from_flow(buf, flow, flow->in_port.odp_port);
2051 upcall->key = buf->data;
2052 upcall->key_len = buf->size;
2056 upcall->userdata = ofpbuf_put(buf, userdata,
2057 NLA_ALIGN(userdata->nla_len));
2060 upcall->packet.data = ofpbuf_put(buf, packet->data, packet->size);
2061 upcall->packet.size = packet->size;
2067 dp_netdev_count_packet(dp, DP_STAT_LOST);
2070 ovs_mutex_unlock(&q->mutex);
2071 fat_rwlock_unlock(&dp->queue_rwlock);
2076 struct dp_netdev_execute_aux {
2077 struct dp_netdev *dp;
2078 const struct flow *key;
2082 dp_execute_cb(void *aux_, struct ofpbuf *packet,
2083 const struct pkt_metadata *md OVS_UNUSED,
2084 const struct nlattr *a, bool may_steal)
2085 OVS_NO_THREAD_SAFETY_ANALYSIS
2087 struct dp_netdev_execute_aux *aux = aux_;
2088 int type = nl_attr_type(a);
2089 struct dp_netdev_port *p;
2091 switch ((enum ovs_action_attr)type) {
2092 case OVS_ACTION_ATTR_OUTPUT:
2093 p = dp_netdev_lookup_port(aux->dp, u32_to_odp(nl_attr_get_u32(a)));
2095 netdev_send(p->netdev, packet, may_steal);
2099 case OVS_ACTION_ATTR_USERSPACE: {
2100 const struct nlattr *userdata;
2102 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2104 dp_netdev_output_userspace(aux->dp, packet,
2105 flow_hash_5tuple(aux->key, 0)
2106 % aux->dp->n_handlers,
2107 DPIF_UC_ACTION, aux->key,
2111 ofpbuf_delete(packet);
2115 case OVS_ACTION_ATTR_PUSH_VLAN:
2116 case OVS_ACTION_ATTR_POP_VLAN:
2117 case OVS_ACTION_ATTR_PUSH_MPLS:
2118 case OVS_ACTION_ATTR_POP_MPLS:
2119 case OVS_ACTION_ATTR_SET:
2120 case OVS_ACTION_ATTR_SAMPLE:
2121 case OVS_ACTION_ATTR_UNSPEC:
2122 case __OVS_ACTION_ATTR_MAX:
2129 dp_netdev_execute_actions(struct dp_netdev *dp, const struct flow *key,
2130 struct ofpbuf *packet, bool may_steal,
2131 struct pkt_metadata *md,
2132 const struct nlattr *actions, size_t actions_len)
2134 struct dp_netdev_execute_aux aux = {dp, key};
2136 odp_execute_actions(&aux, packet, may_steal, md,
2137 actions, actions_len, dp_execute_cb);
2140 const struct dpif_class dpif_netdev_class = {
2142 dpif_netdev_enumerate,
2143 dpif_netdev_port_open_type,
2146 dpif_netdev_destroy,
2149 dpif_netdev_get_stats,
2150 dpif_netdev_port_add,
2151 dpif_netdev_port_del,
2152 dpif_netdev_port_query_by_number,
2153 dpif_netdev_port_query_by_name,
2154 NULL, /* port_get_pid */
2155 dpif_netdev_port_dump_start,
2156 dpif_netdev_port_dump_next,
2157 dpif_netdev_port_dump_done,
2158 dpif_netdev_port_poll,
2159 dpif_netdev_port_poll_wait,
2160 dpif_netdev_flow_get,
2161 dpif_netdev_flow_put,
2162 dpif_netdev_flow_del,
2163 dpif_netdev_flow_flush,
2164 dpif_netdev_flow_dump_state_init,
2165 dpif_netdev_flow_dump_start,
2166 dpif_netdev_flow_dump_next,
2168 dpif_netdev_flow_dump_done,
2169 dpif_netdev_flow_dump_state_uninit,
2170 dpif_netdev_execute,
2172 dpif_netdev_recv_set,
2173 dpif_netdev_handlers_set,
2174 dpif_netdev_queue_to_priority,
2176 dpif_netdev_recv_wait,
2177 dpif_netdev_recv_purge,
2181 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2182 const char *argv[], void *aux OVS_UNUSED)
2184 struct dp_netdev_port *port;
2185 struct dp_netdev *dp;
2188 ovs_mutex_lock(&dp_netdev_mutex);
2189 dp = shash_find_data(&dp_netdevs, argv[1]);
2190 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2191 ovs_mutex_unlock(&dp_netdev_mutex);
2192 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2195 ovs_refcount_ref(&dp->ref_cnt);
2196 ovs_mutex_unlock(&dp_netdev_mutex);
2198 ovs_rwlock_wrlock(&dp->port_rwlock);
2199 if (get_port_by_name(dp, argv[2], &port)) {
2200 unixctl_command_reply_error(conn, "unknown port");
2204 port_no = u32_to_odp(atoi(argv[3]));
2205 if (!port_no || port_no == ODPP_NONE) {
2206 unixctl_command_reply_error(conn, "bad port number");
2209 if (dp_netdev_lookup_port(dp, port_no)) {
2210 unixctl_command_reply_error(conn, "port number already in use");
2213 hmap_remove(&dp->ports, &port->node);
2214 port->port_no = port_no;
2215 hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0));
2216 seq_change(dp->port_seq);
2217 unixctl_command_reply(conn, NULL);
2220 ovs_rwlock_unlock(&dp->port_rwlock);
2221 dp_netdev_unref(dp);
2225 dpif_dummy_register__(const char *type)
2227 struct dpif_class *class;
2229 class = xmalloc(sizeof *class);
2230 *class = dpif_netdev_class;
2231 class->type = xstrdup(type);
2232 dp_register_provider(class);
2236 dpif_dummy_register(bool override)
2243 dp_enumerate_types(&types);
2244 SSET_FOR_EACH (type, &types) {
2245 if (!dp_unregister_provider(type)) {
2246 dpif_dummy_register__(type);
2249 sset_destroy(&types);
2252 dpif_dummy_register__("dummy");
2254 unixctl_command_register("dpif-dummy/change-port-number",
2255 "DP PORT NEW-NUMBER",
2256 3, 3, dpif_dummy_change_port_number, NULL);