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-dpdk.h"
47 #include "netdev-vport.h"
49 #include "odp-execute.h"
51 #include "ofp-print.h"
55 #include "poll-loop.h"
65 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
67 /* By default, choose a priority in the middle. */
68 #define NETDEV_RULE_PRIORITY 0x8000
71 /* Use per thread recirc_depth to prevent recirculation loop. */
72 #define MAX_RECIRC_DEPTH 5
73 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
75 /* Configuration parameters. */
76 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
79 enum { MAX_QUEUE_LEN = 128 }; /* Maximum number of packets per queue. */
80 enum { QUEUE_MASK = MAX_QUEUE_LEN - 1 };
81 BUILD_ASSERT_DECL(IS_POW2(MAX_QUEUE_LEN));
83 /* Protects against changes to 'dp_netdevs'. */
84 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
86 /* Contains all 'struct dp_netdev's. */
87 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
88 = SHASH_INITIALIZER(&dp_netdevs);
90 struct dp_netdev_upcall {
91 struct dpif_upcall upcall; /* Queued upcall information. */
92 struct ofpbuf buf; /* ofpbuf instance for upcall.packet. */
95 /* A queue passing packets from a struct dp_netdev to its clients (handlers).
101 * Any access at all requires the owning 'dp_netdev''s queue_rwlock and
103 struct dp_netdev_queue {
104 struct ovs_mutex mutex;
105 struct seq *seq; /* Incremented whenever a packet is queued. */
106 struct dp_netdev_upcall upcalls[MAX_QUEUE_LEN] OVS_GUARDED;
107 unsigned int head OVS_GUARDED;
108 unsigned int tail OVS_GUARDED;
111 /* Datapath based on the network device interface from netdev.h.
117 * Some members, marked 'const', are immutable. Accessing other members
118 * requires synchronization, as noted in more detail below.
120 * Acquisition order is, from outermost to innermost:
122 * dp_netdev_mutex (global)
129 const struct dpif_class *const class;
130 const char *const name;
131 struct ovs_refcount ref_cnt;
132 atomic_flag destroyed;
136 * Readers of 'cls' and 'flow_table' must take a 'cls->rwlock' read lock.
138 * Writers of 'cls' and 'flow_table' must take the 'flow_mutex' and then
139 * the 'cls->rwlock' write lock. (The outer 'flow_mutex' allows writers to
140 * atomically perform multiple operations on 'cls' and 'flow_table'.)
142 struct ovs_mutex flow_mutex;
143 struct classifier cls; /* Classifier. Protected by cls.rwlock. */
144 struct hmap flow_table OVS_GUARDED; /* Flow table. */
148 * 'queue_rwlock' protects the modification of 'handler_queues' and
149 * 'n_handlers'. The queue elements are protected by its
150 * 'handler_queues''s mutex. */
151 struct fat_rwlock queue_rwlock;
152 struct dp_netdev_queue *handler_queues;
157 * ovsthread_stats is internally synchronized. */
158 struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */
162 * Any lookup into 'ports' or any access to the dp_netdev_ports found
163 * through 'ports' requires taking 'port_rwlock'. */
164 struct ovs_rwlock port_rwlock;
165 struct hmap ports OVS_GUARDED;
166 struct seq *port_seq; /* Incremented whenever a port changes. */
168 /* Forwarding threads. */
169 struct latch exit_latch;
170 struct pmd_thread *pmd_threads;
171 size_t n_pmd_threads;
175 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
177 OVS_REQ_RDLOCK(dp->port_rwlock);
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 hmap_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. */
206 /* A flow in dp_netdev's 'flow_table'.
212 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
213 * its dp_netdev's classifier. The text below calls this classifier 'cls'.
218 * The thread safety rules described here for "struct dp_netdev_flow" are
219 * motivated by two goals:
221 * - Prevent threads that read members of "struct dp_netdev_flow" from
222 * reading bad data due to changes by some thread concurrently modifying
225 * - Prevent two threads making changes to members of a given "struct
226 * dp_netdev_flow" from interfering with each other.
232 * A flow 'flow' may be accessed without a risk of being freed by code that
233 * holds a read-lock or write-lock on 'cls->rwlock' or that owns a reference to
234 * 'flow->ref_cnt' (or both). Code that needs to hold onto a flow for a while
235 * should take 'cls->rwlock', find the flow it needs, increment 'flow->ref_cnt'
236 * with dpif_netdev_flow_ref(), and drop 'cls->rwlock'.
238 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
239 * flow from being deleted from 'cls' (that's 'cls->rwlock') and it doesn't
240 * protect members of 'flow' from modification.
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. */
255 * Reading or writing these members requires 'mutex'. */
256 struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
259 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
262 static void dp_netdev_flow_free(struct dp_netdev_flow *);
264 /* Contained by struct dp_netdev_flow's 'stats' member. */
265 struct dp_netdev_flow_stats {
266 struct ovs_mutex mutex; /* Guards all the other members. */
268 long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
269 long long int packet_count OVS_GUARDED; /* Number of packets matched. */
270 long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
271 uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
274 /* A set of datapath actions within a "struct dp_netdev_flow".
280 * A struct dp_netdev_actions 'actions' is protected with RCU. */
281 struct dp_netdev_actions {
282 /* These members are immutable: they do not change during the struct's
284 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
285 unsigned int size; /* Size of 'actions', in bytes. */
288 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
290 struct dp_netdev_actions *dp_netdev_flow_get_actions(
291 const struct dp_netdev_flow *);
292 static void dp_netdev_actions_free(struct dp_netdev_actions *);
294 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
295 * the performance overhead of interrupt processing. Therefore netdev can
296 * not implement rx-wait for these devices. dpif-netdev needs to poll
297 * these device to check for recv buffer. pmd-thread does polling for
298 * devices assigned to itself thread.
300 * DPDK used PMD for accessing NIC.
302 * A thread that receives packets from PMD ports, looks them up in the flow
303 * table, and executes the actions it finds.
306 struct dp_netdev *dp;
309 atomic_uint change_seq;
312 /* Interface to netdev-based datapath. */
315 struct dp_netdev *dp;
316 uint64_t last_port_seq;
319 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
320 struct dp_netdev_port **portp)
321 OVS_REQ_RDLOCK(dp->port_rwlock);
322 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
323 struct dp_netdev_port **portp)
324 OVS_REQ_RDLOCK(dp->port_rwlock);
325 static void dp_netdev_free(struct dp_netdev *)
326 OVS_REQUIRES(dp_netdev_mutex);
327 static void dp_netdev_flow_flush(struct dp_netdev *);
328 static int do_add_port(struct dp_netdev *dp, const char *devname,
329 const char *type, odp_port_t port_no)
330 OVS_REQ_WRLOCK(dp->port_rwlock);
331 static int do_del_port(struct dp_netdev *dp, odp_port_t port_no)
332 OVS_REQ_WRLOCK(dp->port_rwlock);
333 static void dp_netdev_destroy_all_queues(struct dp_netdev *dp)
334 OVS_REQ_WRLOCK(dp->queue_rwlock);
335 static int dpif_netdev_open(const struct dpif_class *, const char *name,
336 bool create, struct dpif **);
337 static int dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *,
338 int queue_no, int type,
339 const struct miniflow *,
340 const struct nlattr *userdata);
341 static void dp_netdev_execute_actions(struct dp_netdev *dp,
342 const struct miniflow *,
343 struct ofpbuf *, bool may_steal,
344 struct pkt_metadata *,
345 const struct nlattr *actions,
347 static void dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet,
348 struct pkt_metadata *);
350 static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
352 static struct dpif_netdev *
353 dpif_netdev_cast(const struct dpif *dpif)
355 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
356 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
359 static struct dp_netdev *
360 get_dp_netdev(const struct dpif *dpif)
362 return dpif_netdev_cast(dpif)->dp;
366 dpif_netdev_enumerate(struct sset *all_dps)
368 struct shash_node *node;
370 ovs_mutex_lock(&dp_netdev_mutex);
371 SHASH_FOR_EACH(node, &dp_netdevs) {
372 sset_add(all_dps, node->name);
374 ovs_mutex_unlock(&dp_netdev_mutex);
380 dpif_netdev_class_is_dummy(const struct dpif_class *class)
382 return class != &dpif_netdev_class;
386 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
388 return strcmp(type, "internal") ? type
389 : dpif_netdev_class_is_dummy(class) ? "dummy"
394 create_dpif_netdev(struct dp_netdev *dp)
396 uint16_t netflow_id = hash_string(dp->name, 0);
397 struct dpif_netdev *dpif;
399 ovs_refcount_ref(&dp->ref_cnt);
401 dpif = xmalloc(sizeof *dpif);
402 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
404 dpif->last_port_seq = seq_read(dp->port_seq);
409 /* Choose an unused, non-zero port number and return it on success.
410 * Return ODPP_NONE on failure. */
412 choose_port(struct dp_netdev *dp, const char *name)
413 OVS_REQ_RDLOCK(dp->port_rwlock)
417 if (dp->class != &dpif_netdev_class) {
421 /* If the port name begins with "br", start the number search at
422 * 100 to make writing tests easier. */
423 if (!strncmp(name, "br", 2)) {
427 /* If the port name contains a number, try to assign that port number.
428 * This can make writing unit tests easier because port numbers are
430 for (p = name; *p != '\0'; p++) {
431 if (isdigit((unsigned char) *p)) {
432 port_no = start_no + strtol(p, NULL, 10);
433 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
434 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
435 return u32_to_odp(port_no);
442 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
443 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
444 return u32_to_odp(port_no);
452 create_dp_netdev(const char *name, const struct dpif_class *class,
453 struct dp_netdev **dpp)
454 OVS_REQUIRES(dp_netdev_mutex)
456 struct dp_netdev *dp;
459 dp = xzalloc(sizeof *dp);
460 shash_add(&dp_netdevs, name, dp);
462 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
463 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
464 ovs_refcount_init(&dp->ref_cnt);
465 atomic_flag_clear(&dp->destroyed);
467 ovs_mutex_init(&dp->flow_mutex);
468 classifier_init(&dp->cls, NULL);
469 hmap_init(&dp->flow_table);
471 fat_rwlock_init(&dp->queue_rwlock);
473 ovsthread_stats_init(&dp->stats);
475 ovs_rwlock_init(&dp->port_rwlock);
476 hmap_init(&dp->ports);
477 dp->port_seq = seq_create();
478 latch_init(&dp->exit_latch);
480 ovs_rwlock_wrlock(&dp->port_rwlock);
481 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
482 ovs_rwlock_unlock(&dp->port_rwlock);
493 dpif_netdev_open(const struct dpif_class *class, const char *name,
494 bool create, struct dpif **dpifp)
496 struct dp_netdev *dp;
499 ovs_mutex_lock(&dp_netdev_mutex);
500 dp = shash_find_data(&dp_netdevs, name);
502 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
504 error = (dp->class != class ? EINVAL
509 *dpifp = create_dpif_netdev(dp);
511 ovs_mutex_unlock(&dp_netdev_mutex);
517 dp_netdev_purge_queues(struct dp_netdev *dp)
518 OVS_REQ_WRLOCK(dp->queue_rwlock)
522 for (i = 0; i < dp->n_handlers; i++) {
523 struct dp_netdev_queue *q = &dp->handler_queues[i];
525 ovs_mutex_lock(&q->mutex);
526 while (q->tail != q->head) {
527 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
528 ofpbuf_uninit(&u->upcall.packet);
529 ofpbuf_uninit(&u->buf);
531 ovs_mutex_unlock(&q->mutex);
535 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
536 * through the 'dp_netdevs' shash while freeing 'dp'. */
538 dp_netdev_free(struct dp_netdev *dp)
539 OVS_REQUIRES(dp_netdev_mutex)
541 struct dp_netdev_port *port, *next;
542 struct dp_netdev_stats *bucket;
545 shash_find_and_delete(&dp_netdevs, dp->name);
547 dp_netdev_set_pmd_threads(dp, 0);
548 free(dp->pmd_threads);
550 dp_netdev_flow_flush(dp);
551 ovs_rwlock_wrlock(&dp->port_rwlock);
552 HMAP_FOR_EACH_SAFE (port, next, node, &dp->ports) {
553 do_del_port(dp, port->port_no);
555 ovs_rwlock_unlock(&dp->port_rwlock);
557 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
558 ovs_mutex_destroy(&bucket->mutex);
559 free_cacheline(bucket);
561 ovsthread_stats_destroy(&dp->stats);
563 fat_rwlock_wrlock(&dp->queue_rwlock);
564 dp_netdev_destroy_all_queues(dp);
565 fat_rwlock_unlock(&dp->queue_rwlock);
567 fat_rwlock_destroy(&dp->queue_rwlock);
569 classifier_destroy(&dp->cls);
570 hmap_destroy(&dp->flow_table);
571 ovs_mutex_destroy(&dp->flow_mutex);
572 seq_destroy(dp->port_seq);
573 hmap_destroy(&dp->ports);
574 latch_destroy(&dp->exit_latch);
575 free(CONST_CAST(char *, dp->name));
580 dp_netdev_unref(struct dp_netdev *dp)
583 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
584 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
585 ovs_mutex_lock(&dp_netdev_mutex);
586 if (ovs_refcount_unref(&dp->ref_cnt) == 1) {
589 ovs_mutex_unlock(&dp_netdev_mutex);
594 dpif_netdev_close(struct dpif *dpif)
596 struct dp_netdev *dp = get_dp_netdev(dpif);
603 dpif_netdev_destroy(struct dpif *dpif)
605 struct dp_netdev *dp = get_dp_netdev(dpif);
607 if (!atomic_flag_test_and_set(&dp->destroyed)) {
608 if (ovs_refcount_unref(&dp->ref_cnt) == 1) {
609 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
618 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
620 struct dp_netdev *dp = get_dp_netdev(dpif);
621 struct dp_netdev_stats *bucket;
624 fat_rwlock_rdlock(&dp->cls.rwlock);
625 stats->n_flows = hmap_count(&dp->flow_table);
626 fat_rwlock_unlock(&dp->cls.rwlock);
628 stats->n_hit = stats->n_missed = stats->n_lost = 0;
629 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
630 ovs_mutex_lock(&bucket->mutex);
631 stats->n_hit += bucket->n[DP_STAT_HIT];
632 stats->n_missed += bucket->n[DP_STAT_MISS];
633 stats->n_lost += bucket->n[DP_STAT_LOST];
634 ovs_mutex_unlock(&bucket->mutex);
636 stats->n_masks = UINT32_MAX;
637 stats->n_mask_hit = UINT64_MAX;
643 dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
647 for (i = 0; i < dp->n_pmd_threads; i++) {
648 struct pmd_thread *f = &dp->pmd_threads[i];
651 atomic_add(&f->change_seq, 1, &id);
656 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
658 OVS_REQ_WRLOCK(dp->port_rwlock)
660 struct netdev_saved_flags *sf;
661 struct dp_netdev_port *port;
662 struct netdev *netdev;
663 enum netdev_flags flags;
664 const char *open_type;
668 /* XXX reject devices already in some dp_netdev. */
670 /* Open and validate network device. */
671 open_type = dpif_netdev_port_open_type(dp->class, type);
672 error = netdev_open(devname, open_type, &netdev);
676 /* XXX reject non-Ethernet devices */
678 netdev_get_flags(netdev, &flags);
679 if (flags & NETDEV_LOOPBACK) {
680 VLOG_ERR("%s: cannot add a loopback device", devname);
681 netdev_close(netdev);
685 port = xzalloc(sizeof *port);
686 port->port_no = port_no;
687 port->netdev = netdev;
688 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
689 port->type = xstrdup(type);
690 for (i = 0; i < netdev_n_rxq(netdev); i++) {
691 error = netdev_rxq_open(netdev, &port->rxq[i], i);
693 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
694 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
695 devname, ovs_strerror(errno));
696 netdev_close(netdev);
701 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
703 for (i = 0; i < netdev_n_rxq(netdev); i++) {
704 netdev_rxq_close(port->rxq[i]);
706 netdev_close(netdev);
713 if (netdev_is_pmd(netdev)) {
715 dp_netdev_set_pmd_threads(dp, NR_THREADS);
716 dp_netdev_reload_pmd_threads(dp);
718 ovs_refcount_init(&port->ref_cnt);
720 hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0));
721 seq_change(dp->port_seq);
727 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
728 odp_port_t *port_nop)
730 struct dp_netdev *dp = get_dp_netdev(dpif);
731 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
732 const char *dpif_port;
736 ovs_rwlock_wrlock(&dp->port_rwlock);
737 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
738 if (*port_nop != ODPP_NONE) {
740 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
742 port_no = choose_port(dp, dpif_port);
743 error = port_no == ODPP_NONE ? EFBIG : 0;
747 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
749 ovs_rwlock_unlock(&dp->port_rwlock);
755 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
757 struct dp_netdev *dp = get_dp_netdev(dpif);
760 ovs_rwlock_wrlock(&dp->port_rwlock);
761 error = port_no == ODPP_LOCAL ? EINVAL : do_del_port(dp, port_no);
762 ovs_rwlock_unlock(&dp->port_rwlock);
768 is_valid_port_number(odp_port_t port_no)
770 return port_no != ODPP_NONE;
773 static struct dp_netdev_port *
774 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
775 OVS_REQ_RDLOCK(dp->port_rwlock)
777 struct dp_netdev_port *port;
779 HMAP_FOR_EACH_IN_BUCKET (port, node, hash_int(odp_to_u32(port_no), 0),
781 if (port->port_no == port_no) {
789 get_port_by_number(struct dp_netdev *dp,
790 odp_port_t port_no, struct dp_netdev_port **portp)
791 OVS_REQ_RDLOCK(dp->port_rwlock)
793 if (!is_valid_port_number(port_no)) {
797 *portp = dp_netdev_lookup_port(dp, port_no);
798 return *portp ? 0 : ENOENT;
803 port_ref(struct dp_netdev_port *port)
806 ovs_refcount_ref(&port->ref_cnt);
811 port_unref(struct dp_netdev_port *port)
813 if (port && ovs_refcount_unref(&port->ref_cnt) == 1) {
814 int n_rxq = netdev_n_rxq(port->netdev);
817 netdev_close(port->netdev);
818 netdev_restore_flags(port->sf);
820 for (i = 0; i < n_rxq; i++) {
821 netdev_rxq_close(port->rxq[i]);
830 get_port_by_name(struct dp_netdev *dp,
831 const char *devname, struct dp_netdev_port **portp)
832 OVS_REQ_RDLOCK(dp->port_rwlock)
834 struct dp_netdev_port *port;
836 HMAP_FOR_EACH (port, node, &dp->ports) {
837 if (!strcmp(netdev_get_name(port->netdev), devname)) {
846 do_del_port(struct dp_netdev *dp, odp_port_t port_no)
847 OVS_REQ_WRLOCK(dp->port_rwlock)
849 struct dp_netdev_port *port;
852 error = get_port_by_number(dp, port_no, &port);
857 hmap_remove(&dp->ports, &port->node);
858 seq_change(dp->port_seq);
859 if (netdev_is_pmd(port->netdev)) {
860 dp_netdev_reload_pmd_threads(dp);
868 answer_port_query(const struct dp_netdev_port *port,
869 struct dpif_port *dpif_port)
871 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
872 dpif_port->type = xstrdup(port->type);
873 dpif_port->port_no = port->port_no;
877 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
878 struct dpif_port *dpif_port)
880 struct dp_netdev *dp = get_dp_netdev(dpif);
881 struct dp_netdev_port *port;
884 ovs_rwlock_rdlock(&dp->port_rwlock);
885 error = get_port_by_number(dp, port_no, &port);
886 if (!error && dpif_port) {
887 answer_port_query(port, dpif_port);
889 ovs_rwlock_unlock(&dp->port_rwlock);
895 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
896 struct dpif_port *dpif_port)
898 struct dp_netdev *dp = get_dp_netdev(dpif);
899 struct dp_netdev_port *port;
902 ovs_rwlock_rdlock(&dp->port_rwlock);
903 error = get_port_by_name(dp, devname, &port);
904 if (!error && dpif_port) {
905 answer_port_query(port, dpif_port);
907 ovs_rwlock_unlock(&dp->port_rwlock);
913 dp_netdev_flow_free(struct dp_netdev_flow *flow)
915 struct dp_netdev_flow_stats *bucket;
918 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
919 ovs_mutex_destroy(&bucket->mutex);
920 free_cacheline(bucket);
922 ovsthread_stats_destroy(&flow->stats);
924 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
925 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
930 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
931 OVS_REQ_WRLOCK(dp->cls.rwlock)
932 OVS_REQUIRES(dp->flow_mutex)
934 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
935 struct hmap_node *node = CONST_CAST(struct hmap_node *, &flow->node);
937 classifier_remove(&dp->cls, cr);
938 hmap_remove(&dp->flow_table, node);
939 ovsrcu_postpone(dp_netdev_flow_free, flow);
943 dp_netdev_flow_flush(struct dp_netdev *dp)
945 struct dp_netdev_flow *netdev_flow, *next;
947 ovs_mutex_lock(&dp->flow_mutex);
948 fat_rwlock_wrlock(&dp->cls.rwlock);
949 HMAP_FOR_EACH_SAFE (netdev_flow, next, node, &dp->flow_table) {
950 dp_netdev_remove_flow(dp, netdev_flow);
952 fat_rwlock_unlock(&dp->cls.rwlock);
953 ovs_mutex_unlock(&dp->flow_mutex);
957 dpif_netdev_flow_flush(struct dpif *dpif)
959 struct dp_netdev *dp = get_dp_netdev(dpif);
961 dp_netdev_flow_flush(dp);
965 struct dp_netdev_port_state {
972 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
974 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
979 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
980 struct dpif_port *dpif_port)
982 struct dp_netdev_port_state *state = state_;
983 struct dp_netdev *dp = get_dp_netdev(dpif);
984 struct hmap_node *node;
987 ovs_rwlock_rdlock(&dp->port_rwlock);
988 node = hmap_at_position(&dp->ports, &state->bucket, &state->offset);
990 struct dp_netdev_port *port;
992 port = CONTAINER_OF(node, struct dp_netdev_port, node);
995 state->name = xstrdup(netdev_get_name(port->netdev));
996 dpif_port->name = state->name;
997 dpif_port->type = port->type;
998 dpif_port->port_no = port->port_no;
1004 ovs_rwlock_unlock(&dp->port_rwlock);
1010 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1012 struct dp_netdev_port_state *state = state_;
1019 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1021 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1022 uint64_t new_port_seq;
1025 new_port_seq = seq_read(dpif->dp->port_seq);
1026 if (dpif->last_port_seq != new_port_seq) {
1027 dpif->last_port_seq = new_port_seq;
1037 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1039 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1041 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1044 static struct dp_netdev_flow *
1045 dp_netdev_flow_cast(const struct cls_rule *cr)
1047 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1050 static struct dp_netdev_flow *
1051 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
1052 OVS_EXCLUDED(dp->cls.rwlock)
1054 struct dp_netdev_flow *netdev_flow;
1055 struct cls_rule *rule;
1057 fat_rwlock_rdlock(&dp->cls.rwlock);
1058 rule = classifier_lookup_miniflow_first(&dp->cls, key);
1059 netdev_flow = dp_netdev_flow_cast(rule);
1060 fat_rwlock_unlock(&dp->cls.rwlock);
1065 static struct dp_netdev_flow *
1066 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1067 OVS_REQ_RDLOCK(dp->cls.rwlock)
1069 struct dp_netdev_flow *netdev_flow;
1071 HMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1073 if (flow_equal(&netdev_flow->flow, flow)) {
1082 get_dpif_flow_stats(struct dp_netdev_flow *netdev_flow,
1083 struct dpif_flow_stats *stats)
1085 struct dp_netdev_flow_stats *bucket;
1088 memset(stats, 0, sizeof *stats);
1089 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1090 ovs_mutex_lock(&bucket->mutex);
1091 stats->n_packets += bucket->packet_count;
1092 stats->n_bytes += bucket->byte_count;
1093 stats->used = MAX(stats->used, bucket->used);
1094 stats->tcp_flags |= bucket->tcp_flags;
1095 ovs_mutex_unlock(&bucket->mutex);
1100 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1101 const struct nlattr *mask_key,
1102 uint32_t mask_key_len, const struct flow *flow,
1106 enum odp_key_fitness fitness;
1108 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1110 /* This should not happen: it indicates that
1111 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1112 * disagree on the acceptable form of a mask. Log the problem
1113 * as an error, with enough details to enable debugging. */
1114 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1116 if (!VLOG_DROP_ERR(&rl)) {
1120 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1122 VLOG_ERR("internal error parsing flow mask %s (%s)",
1123 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1130 enum mf_field_id id;
1131 /* No mask key, unwildcard everything except fields whose
1132 * prerequisities are not met. */
1133 memset(mask, 0x0, sizeof *mask);
1135 for (id = 0; id < MFF_N_IDS; ++id) {
1136 /* Skip registers and metadata. */
1137 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1138 && id != MFF_METADATA) {
1139 const struct mf_field *mf = mf_from_id(id);
1140 if (mf_are_prereqs_ok(mf, flow)) {
1141 mf_mask_field(mf, mask);
1147 /* Force unwildcard the in_port.
1149 * We need to do this even in the case where we unwildcard "everything"
1150 * above because "everything" only includes the 16-bit OpenFlow port number
1151 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1152 * port number mask->in_port.odp_port. */
1153 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1159 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1164 if (odp_flow_key_to_flow(key, key_len, flow)) {
1165 /* This should not happen: it indicates that odp_flow_key_from_flow()
1166 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1167 * flow. Log the problem as an error, with enough details to enable
1169 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1171 if (!VLOG_DROP_ERR(&rl)) {
1175 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1176 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1183 in_port = flow->in_port.odp_port;
1184 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1192 dpif_netdev_flow_get(const struct dpif *dpif,
1193 const struct nlattr *nl_key, size_t nl_key_len,
1194 struct ofpbuf **bufp,
1195 struct nlattr **maskp, size_t *mask_len,
1196 struct nlattr **actionsp, size_t *actions_len,
1197 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);
1218 if (maskp || actionsp) {
1219 struct dp_netdev_actions *actions;
1222 actions = dp_netdev_flow_get_actions(netdev_flow);
1223 len += maskp ? sizeof(struct odputil_keybuf) : 0;
1224 len += actionsp ? actions->size : 0;
1226 *bufp = ofpbuf_new(len);
1228 struct flow_wildcards wc;
1230 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1231 odp_flow_key_from_mask(*bufp, &wc.masks, &netdev_flow->flow,
1232 odp_to_u32(wc.masks.in_port.odp_port),
1234 *maskp = ofpbuf_data(*bufp);
1235 *mask_len = ofpbuf_size(*bufp);
1238 struct dp_netdev_actions *actions;
1240 actions = dp_netdev_flow_get_actions(netdev_flow);
1241 *actionsp = ofpbuf_put(*bufp, actions->actions, actions->size);
1242 *actions_len = actions->size;
1253 dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow,
1254 const struct flow_wildcards *wc,
1255 const struct nlattr *actions,
1257 OVS_REQUIRES(dp->flow_mutex)
1259 struct dp_netdev_flow *netdev_flow;
1262 netdev_flow = xzalloc(sizeof *netdev_flow);
1263 *CONST_CAST(struct flow *, &netdev_flow->flow) = *flow;
1265 ovsthread_stats_init(&netdev_flow->stats);
1267 ovsrcu_set(&netdev_flow->actions,
1268 dp_netdev_actions_create(actions, actions_len));
1270 match_init(&match, flow, wc);
1271 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1272 &match, NETDEV_RULE_PRIORITY);
1273 fat_rwlock_wrlock(&dp->cls.rwlock);
1274 classifier_insert(&dp->cls,
1275 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1276 hmap_insert(&dp->flow_table,
1277 CONST_CAST(struct hmap_node *, &netdev_flow->node),
1278 flow_hash(flow, 0));
1279 fat_rwlock_unlock(&dp->cls.rwlock);
1285 clear_stats(struct dp_netdev_flow *netdev_flow)
1287 struct dp_netdev_flow_stats *bucket;
1290 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1291 ovs_mutex_lock(&bucket->mutex);
1293 bucket->packet_count = 0;
1294 bucket->byte_count = 0;
1295 bucket->tcp_flags = 0;
1296 ovs_mutex_unlock(&bucket->mutex);
1301 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1303 struct dp_netdev *dp = get_dp_netdev(dpif);
1304 struct dp_netdev_flow *netdev_flow;
1306 struct miniflow miniflow;
1307 struct flow_wildcards wc;
1310 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &flow);
1314 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1315 put->mask, put->mask_len,
1320 miniflow_init(&miniflow, &flow);
1322 ovs_mutex_lock(&dp->flow_mutex);
1323 netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
1325 if (put->flags & DPIF_FP_CREATE) {
1326 if (hmap_count(&dp->flow_table) < MAX_FLOWS) {
1328 memset(put->stats, 0, sizeof *put->stats);
1330 error = dp_netdev_flow_add(dp, &flow, &wc, put->actions,
1339 if (put->flags & DPIF_FP_MODIFY
1340 && flow_equal(&flow, &netdev_flow->flow)) {
1341 struct dp_netdev_actions *new_actions;
1342 struct dp_netdev_actions *old_actions;
1344 new_actions = dp_netdev_actions_create(put->actions,
1347 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1348 ovsrcu_set(&netdev_flow->actions, new_actions);
1351 get_dpif_flow_stats(netdev_flow, put->stats);
1353 if (put->flags & DPIF_FP_ZERO_STATS) {
1354 clear_stats(netdev_flow);
1357 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1358 } else if (put->flags & DPIF_FP_CREATE) {
1361 /* Overlapping flow. */
1365 ovs_mutex_unlock(&dp->flow_mutex);
1366 miniflow_destroy(&miniflow);
1372 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1374 struct dp_netdev *dp = get_dp_netdev(dpif);
1375 struct dp_netdev_flow *netdev_flow;
1379 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1384 ovs_mutex_lock(&dp->flow_mutex);
1385 fat_rwlock_wrlock(&dp->cls.rwlock);
1386 netdev_flow = dp_netdev_find_flow(dp, &key);
1389 get_dpif_flow_stats(netdev_flow, del->stats);
1391 dp_netdev_remove_flow(dp, netdev_flow);
1395 fat_rwlock_unlock(&dp->cls.rwlock);
1396 ovs_mutex_unlock(&dp->flow_mutex);
1401 struct dp_netdev_flow_state {
1402 struct odputil_keybuf keybuf;
1403 struct odputil_keybuf maskbuf;
1404 struct dpif_flow_stats stats;
1407 struct dp_netdev_flow_iter {
1411 struct ovs_mutex mutex;
1415 dpif_netdev_flow_dump_state_init(void **statep)
1417 struct dp_netdev_flow_state *state;
1419 *statep = state = xmalloc(sizeof *state);
1423 dpif_netdev_flow_dump_state_uninit(void *state_)
1425 struct dp_netdev_flow_state *state = state_;
1431 dpif_netdev_flow_dump_start(const struct dpif *dpif OVS_UNUSED, void **iterp)
1433 struct dp_netdev_flow_iter *iter;
1435 *iterp = iter = xmalloc(sizeof *iter);
1439 ovs_mutex_init(&iter->mutex);
1443 /* XXX the caller must use 'actions' without quiescing */
1445 dpif_netdev_flow_dump_next(const struct dpif *dpif, void *iter_, void *state_,
1446 const struct nlattr **key, size_t *key_len,
1447 const struct nlattr **mask, size_t *mask_len,
1448 const struct nlattr **actions, size_t *actions_len,
1449 const struct dpif_flow_stats **stats)
1451 struct dp_netdev_flow_iter *iter = iter_;
1452 struct dp_netdev_flow_state *state = state_;
1453 struct dp_netdev *dp = get_dp_netdev(dpif);
1454 struct dp_netdev_flow *netdev_flow;
1455 struct flow_wildcards wc;
1458 ovs_mutex_lock(&iter->mutex);
1459 error = iter->status;
1461 struct hmap_node *node;
1463 fat_rwlock_rdlock(&dp->cls.rwlock);
1464 node = hmap_at_position(&dp->flow_table, &iter->bucket, &iter->offset);
1466 netdev_flow = CONTAINER_OF(node, struct dp_netdev_flow, node);
1468 fat_rwlock_unlock(&dp->cls.rwlock);
1470 iter->status = error = EOF;
1473 ovs_mutex_unlock(&iter->mutex);
1478 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1483 ofpbuf_use_stack(&buf, &state->keybuf, sizeof state->keybuf);
1484 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1485 netdev_flow->flow.in_port.odp_port);
1487 *key = ofpbuf_data(&buf);
1488 *key_len = ofpbuf_size(&buf);
1494 ofpbuf_use_stack(&buf, &state->maskbuf, sizeof state->maskbuf);
1495 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1496 odp_to_u32(wc.masks.in_port.odp_port),
1499 *mask = ofpbuf_data(&buf);
1500 *mask_len = ofpbuf_size(&buf);
1503 if (actions || stats) {
1505 struct dp_netdev_actions *dp_actions =
1506 dp_netdev_flow_get_actions(netdev_flow);
1508 *actions = dp_actions->actions;
1509 *actions_len = dp_actions->size;
1513 get_dpif_flow_stats(netdev_flow, &state->stats);
1514 *stats = &state->stats;
1522 dpif_netdev_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *iter_)
1524 struct dp_netdev_flow_iter *iter = iter_;
1526 ovs_mutex_destroy(&iter->mutex);
1532 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1534 struct dp_netdev *dp = get_dp_netdev(dpif);
1535 struct pkt_metadata *md = &execute->md;
1537 struct miniflow flow;
1538 uint32_t buf[FLOW_U32S];
1541 if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
1542 ofpbuf_size(execute->packet) > UINT16_MAX) {
1546 /* Extract flow key. */
1547 miniflow_initialize(&key.flow, key.buf);
1548 miniflow_extract(execute->packet, md, &key.flow);
1550 ovs_rwlock_rdlock(&dp->port_rwlock);
1551 dp_netdev_execute_actions(dp, &key.flow, execute->packet, false, md,
1552 execute->actions, execute->actions_len);
1553 ovs_rwlock_unlock(&dp->port_rwlock);
1559 dp_netdev_destroy_all_queues(struct dp_netdev *dp)
1560 OVS_REQ_WRLOCK(dp->queue_rwlock)
1564 dp_netdev_purge_queues(dp);
1566 for (i = 0; i < dp->n_handlers; i++) {
1567 struct dp_netdev_queue *q = &dp->handler_queues[i];
1569 ovs_mutex_destroy(&q->mutex);
1570 seq_destroy(q->seq);
1572 free(dp->handler_queues);
1573 dp->handler_queues = NULL;
1578 dp_netdev_refresh_queues(struct dp_netdev *dp, uint32_t n_handlers)
1579 OVS_REQ_WRLOCK(dp->queue_rwlock)
1581 if (dp->n_handlers != n_handlers) {
1584 dp_netdev_destroy_all_queues(dp);
1586 dp->n_handlers = n_handlers;
1587 dp->handler_queues = xzalloc(n_handlers * sizeof *dp->handler_queues);
1589 for (i = 0; i < n_handlers; i++) {
1590 struct dp_netdev_queue *q = &dp->handler_queues[i];
1592 ovs_mutex_init(&q->mutex);
1593 q->seq = seq_create();
1599 dpif_netdev_recv_set(struct dpif *dpif, bool enable)
1601 struct dp_netdev *dp = get_dp_netdev(dpif);
1603 if ((dp->handler_queues != NULL) == enable) {
1607 fat_rwlock_wrlock(&dp->queue_rwlock);
1609 dp_netdev_destroy_all_queues(dp);
1611 dp_netdev_refresh_queues(dp, 1);
1613 fat_rwlock_unlock(&dp->queue_rwlock);
1619 dpif_netdev_handlers_set(struct dpif *dpif, uint32_t n_handlers)
1621 struct dp_netdev *dp = get_dp_netdev(dpif);
1623 fat_rwlock_wrlock(&dp->queue_rwlock);
1624 if (dp->handler_queues) {
1625 dp_netdev_refresh_queues(dp, n_handlers);
1627 fat_rwlock_unlock(&dp->queue_rwlock);
1633 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1634 uint32_t queue_id, uint32_t *priority)
1636 *priority = queue_id;
1641 dp_netdev_recv_check(const struct dp_netdev *dp, const uint32_t handler_id)
1642 OVS_REQ_RDLOCK(dp->queue_rwlock)
1644 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1646 if (!dp->handler_queues) {
1647 VLOG_WARN_RL(&rl, "receiving upcall disabled");
1651 if (handler_id >= dp->n_handlers) {
1652 VLOG_WARN_RL(&rl, "handler index out of bound");
1660 dpif_netdev_recv(struct dpif *dpif, uint32_t handler_id,
1661 struct dpif_upcall *upcall, struct ofpbuf *buf)
1663 struct dp_netdev *dp = get_dp_netdev(dpif);
1664 struct dp_netdev_queue *q;
1667 fat_rwlock_rdlock(&dp->queue_rwlock);
1669 if (!dp_netdev_recv_check(dp, handler_id)) {
1674 q = &dp->handler_queues[handler_id];
1675 ovs_mutex_lock(&q->mutex);
1676 if (q->head != q->tail) {
1677 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
1679 *upcall = u->upcall;
1686 ovs_mutex_unlock(&q->mutex);
1689 fat_rwlock_unlock(&dp->queue_rwlock);
1695 dpif_netdev_recv_wait(struct dpif *dpif, uint32_t handler_id)
1697 struct dp_netdev *dp = get_dp_netdev(dpif);
1698 struct dp_netdev_queue *q;
1701 fat_rwlock_rdlock(&dp->queue_rwlock);
1703 if (!dp_netdev_recv_check(dp, handler_id)) {
1707 q = &dp->handler_queues[handler_id];
1708 ovs_mutex_lock(&q->mutex);
1709 seq = seq_read(q->seq);
1710 if (q->head != q->tail) {
1711 poll_immediate_wake();
1713 seq_wait(q->seq, seq);
1716 ovs_mutex_unlock(&q->mutex);
1719 fat_rwlock_unlock(&dp->queue_rwlock);
1723 dpif_netdev_recv_purge(struct dpif *dpif)
1725 struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif);
1727 fat_rwlock_wrlock(&dpif_netdev->dp->queue_rwlock);
1728 dp_netdev_purge_queues(dpif_netdev->dp);
1729 fat_rwlock_unlock(&dpif_netdev->dp->queue_rwlock);
1732 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1733 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1735 struct dp_netdev_actions *
1736 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1738 struct dp_netdev_actions *netdev_actions;
1740 netdev_actions = xmalloc(sizeof *netdev_actions);
1741 netdev_actions->actions = xmemdup(actions, size);
1742 netdev_actions->size = size;
1744 return netdev_actions;
1747 struct dp_netdev_actions *
1748 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
1750 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
1754 dp_netdev_actions_free(struct dp_netdev_actions *actions)
1756 free(actions->actions);
1762 dp_netdev_process_rxq_port(struct dp_netdev *dp,
1763 struct dp_netdev_port *port,
1764 struct netdev_rxq *rxq)
1766 struct ofpbuf *packet[NETDEV_MAX_RX_BATCH];
1769 error = netdev_rxq_recv(rxq, packet, &c);
1771 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port->port_no);
1774 for (i = 0; i < c; i++) {
1775 dp_netdev_port_input(dp, packet[i], &md);
1777 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1778 static struct vlog_rate_limit rl
1779 = VLOG_RATE_LIMIT_INIT(1, 5);
1781 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1782 netdev_get_name(port->netdev),
1783 ovs_strerror(error));
1788 dpif_netdev_run(struct dpif *dpif)
1790 struct dp_netdev_port *port;
1791 struct dp_netdev *dp = get_dp_netdev(dpif);
1793 ovs_rwlock_rdlock(&dp->port_rwlock);
1795 HMAP_FOR_EACH (port, node, &dp->ports) {
1796 if (!netdev_is_pmd(port->netdev)) {
1799 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1800 dp_netdev_process_rxq_port(dp, port, port->rxq[i]);
1805 ovs_rwlock_unlock(&dp->port_rwlock);
1809 dpif_netdev_wait(struct dpif *dpif)
1811 struct dp_netdev_port *port;
1812 struct dp_netdev *dp = get_dp_netdev(dpif);
1814 ovs_rwlock_rdlock(&dp->port_rwlock);
1816 HMAP_FOR_EACH (port, node, &dp->ports) {
1817 if (!netdev_is_pmd(port->netdev)) {
1820 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1821 netdev_rxq_wait(port->rxq[i]);
1825 ovs_rwlock_unlock(&dp->port_rwlock);
1829 struct dp_netdev_port *port;
1830 struct netdev_rxq *rx;
1834 pmd_load_queues(struct pmd_thread *f,
1835 struct rxq_poll **ppoll_list, int poll_cnt)
1837 struct dp_netdev *dp = f->dp;
1838 struct rxq_poll *poll_list = *ppoll_list;
1839 struct dp_netdev_port *port;
1844 /* Simple scheduler for netdev rx polling. */
1845 ovs_rwlock_rdlock(&dp->port_rwlock);
1846 for (i = 0; i < poll_cnt; i++) {
1847 port_unref(poll_list[i].port);
1853 HMAP_FOR_EACH (port, node, &f->dp->ports) {
1854 if (netdev_is_pmd(port->netdev)) {
1857 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1858 if ((index % dp->n_pmd_threads) == id) {
1859 poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1));
1862 poll_list[poll_cnt].port = port;
1863 poll_list[poll_cnt].rx = port->rxq[i];
1871 ovs_rwlock_unlock(&dp->port_rwlock);
1872 *ppoll_list = poll_list;
1877 pmd_thread_main(void *f_)
1879 struct pmd_thread *f = f_;
1880 struct dp_netdev *dp = f->dp;
1881 unsigned int lc = 0;
1882 struct rxq_poll *poll_list;
1883 unsigned int port_seq;
1890 pmd_thread_setaffinity_cpu(f->id);
1892 poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
1893 atomic_read(&f->change_seq, &port_seq);
1896 unsigned int c_port_seq;
1899 for (i = 0; i < poll_cnt; i++) {
1900 dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx);
1906 /* TODO: need completely userspace based signaling method.
1907 * to keep this thread entirely in userspace.
1908 * For now using atomic counter. */
1910 atomic_read_explicit(&f->change_seq, &c_port_seq, memory_order_consume);
1911 if (c_port_seq != port_seq) {
1917 if (!latch_is_set(&f->dp->exit_latch)){
1921 for (i = 0; i < poll_cnt; i++) {
1922 port_unref(poll_list[i].port);
1930 dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n)
1934 if (n == dp->n_pmd_threads) {
1938 /* Stop existing threads. */
1939 latch_set(&dp->exit_latch);
1940 dp_netdev_reload_pmd_threads(dp);
1941 for (i = 0; i < dp->n_pmd_threads; i++) {
1942 struct pmd_thread *f = &dp->pmd_threads[i];
1944 xpthread_join(f->thread, NULL);
1946 latch_poll(&dp->exit_latch);
1947 free(dp->pmd_threads);
1949 /* Start new threads. */
1950 dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
1951 dp->n_pmd_threads = n;
1953 for (i = 0; i < n; i++) {
1954 struct pmd_thread *f = &dp->pmd_threads[i];
1958 atomic_store(&f->change_seq, 1);
1960 /* Each thread will distribute all devices rx-queues among
1962 f->thread = ovs_thread_create("pmd", pmd_thread_main, f);
1968 dp_netdev_flow_stats_new_cb(void)
1970 struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1971 ovs_mutex_init(&bucket->mutex);
1976 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
1977 const struct ofpbuf *packet,
1978 const struct miniflow *key)
1980 uint16_t tcp_flags = miniflow_get_tcp_flags(key);
1981 long long int now = time_msec();
1982 struct dp_netdev_flow_stats *bucket;
1984 bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
1985 dp_netdev_flow_stats_new_cb);
1987 ovs_mutex_lock(&bucket->mutex);
1988 bucket->used = MAX(now, bucket->used);
1989 bucket->packet_count++;
1990 bucket->byte_count += ofpbuf_size(packet);
1991 bucket->tcp_flags |= tcp_flags;
1992 ovs_mutex_unlock(&bucket->mutex);
1996 dp_netdev_stats_new_cb(void)
1998 struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1999 ovs_mutex_init(&bucket->mutex);
2004 dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type)
2006 struct dp_netdev_stats *bucket;
2008 bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
2009 ovs_mutex_lock(&bucket->mutex);
2011 ovs_mutex_unlock(&bucket->mutex);
2015 dp_netdev_input(struct dp_netdev *dp, struct ofpbuf *packet,
2016 struct pkt_metadata *md)
2017 OVS_REQ_RDLOCK(dp->port_rwlock)
2019 struct dp_netdev_flow *netdev_flow;
2021 struct miniflow flow;
2022 uint32_t buf[FLOW_U32S];
2025 if (ofpbuf_size(packet) < ETH_HEADER_LEN) {
2026 ofpbuf_delete(packet);
2029 miniflow_initialize(&key.flow, key.buf);
2030 miniflow_extract(packet, md, &key.flow);
2032 netdev_flow = dp_netdev_lookup_flow(dp, &key.flow);
2034 struct dp_netdev_actions *actions;
2036 dp_netdev_flow_used(netdev_flow, packet, &key.flow);
2038 actions = dp_netdev_flow_get_actions(netdev_flow);
2039 dp_netdev_execute_actions(dp, &key.flow, packet, true, md,
2040 actions->actions, actions->size);
2041 dp_netdev_count_packet(dp, DP_STAT_HIT);
2042 } else if (dp->handler_queues) {
2043 dp_netdev_count_packet(dp, DP_STAT_MISS);
2044 dp_netdev_output_userspace(dp, packet,
2045 miniflow_hash_5tuple(&key.flow, 0)
2047 DPIF_UC_MISS, &key.flow, NULL);
2048 ofpbuf_delete(packet);
2053 dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet,
2054 struct pkt_metadata *md)
2055 OVS_REQ_RDLOCK(dp->port_rwlock)
2057 uint32_t *recirc_depth = recirc_depth_get();
2060 dp_netdev_input(dp, packet, md);
2064 dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet,
2065 int queue_no, int type, const struct miniflow *key,
2066 const struct nlattr *userdata)
2068 struct dp_netdev_queue *q;
2071 fat_rwlock_rdlock(&dp->queue_rwlock);
2072 q = &dp->handler_queues[queue_no];
2073 ovs_mutex_lock(&q->mutex);
2074 if (q->head - q->tail < MAX_QUEUE_LEN) {
2075 struct dp_netdev_upcall *u = &q->upcalls[q->head++ & QUEUE_MASK];
2076 struct dpif_upcall *upcall = &u->upcall;
2077 struct ofpbuf *buf = &u->buf;
2082 upcall->type = type;
2084 /* Allocate buffer big enough for everything. */
2085 buf_size = ODPUTIL_FLOW_KEY_BYTES;
2087 buf_size += NLA_ALIGN(userdata->nla_len);
2089 buf_size += ofpbuf_size(packet);
2090 ofpbuf_init(buf, buf_size);
2093 miniflow_expand(key, &flow);
2094 odp_flow_key_from_flow(buf, &flow, NULL, flow.in_port.odp_port);
2095 upcall->key = ofpbuf_data(buf);
2096 upcall->key_len = ofpbuf_size(buf);
2100 upcall->userdata = ofpbuf_put(buf, userdata,
2101 NLA_ALIGN(userdata->nla_len));
2104 data = ofpbuf_put(buf, ofpbuf_data(packet), ofpbuf_size(packet));
2105 ofpbuf_use_stub(&upcall->packet, data, ofpbuf_size(packet));
2106 ofpbuf_set_size(&upcall->packet, ofpbuf_size(packet));
2112 dp_netdev_count_packet(dp, DP_STAT_LOST);
2115 ovs_mutex_unlock(&q->mutex);
2116 fat_rwlock_unlock(&dp->queue_rwlock);
2121 struct dp_netdev_execute_aux {
2122 struct dp_netdev *dp;
2123 const struct miniflow *key;
2127 dp_execute_cb(void *aux_, struct ofpbuf *packet,
2128 struct pkt_metadata *md,
2129 const struct nlattr *a, bool may_steal)
2130 OVS_NO_THREAD_SAFETY_ANALYSIS
2132 struct dp_netdev_execute_aux *aux = aux_;
2133 int type = nl_attr_type(a);
2134 struct dp_netdev_port *p;
2135 uint32_t *depth = recirc_depth_get();
2137 switch ((enum ovs_action_attr)type) {
2138 case OVS_ACTION_ATTR_OUTPUT:
2139 p = dp_netdev_lookup_port(aux->dp, u32_to_odp(nl_attr_get_u32(a)));
2141 netdev_send(p->netdev, packet, may_steal);
2142 } else if (may_steal) {
2143 ofpbuf_delete(packet);
2148 case OVS_ACTION_ATTR_USERSPACE: {
2149 const struct nlattr *userdata;
2151 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2153 if (aux->dp->n_handlers > 0) {
2154 dp_netdev_output_userspace(aux->dp, packet,
2155 miniflow_hash_5tuple(aux->key, 0)
2156 % aux->dp->n_handlers,
2157 DPIF_UC_ACTION, aux->key,
2162 ofpbuf_delete(packet);
2167 case OVS_ACTION_ATTR_HASH: {
2168 const struct ovs_action_hash *hash_act;
2171 hash_act = nl_attr_get(a);
2172 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
2173 /* Hash need not be symmetric, nor does it need to include
2175 hash = miniflow_hash_5tuple(aux->key, hash_act->hash_basis);
2177 hash = 1; /* 0 is not valid */
2181 VLOG_WARN("Unknown hash algorithm specified for the hash action.");
2189 case OVS_ACTION_ATTR_RECIRC:
2190 if (*depth < MAX_RECIRC_DEPTH) {
2191 struct pkt_metadata recirc_md = *md;
2192 struct ofpbuf *recirc_packet;
2194 recirc_packet = may_steal ? packet : ofpbuf_clone(packet);
2195 recirc_md.recirc_id = nl_attr_get_u32(a);
2198 dp_netdev_input(aux->dp, recirc_packet, &recirc_md);
2204 ofpbuf_delete(packet);
2206 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
2210 case OVS_ACTION_ATTR_PUSH_VLAN:
2211 case OVS_ACTION_ATTR_POP_VLAN:
2212 case OVS_ACTION_ATTR_PUSH_MPLS:
2213 case OVS_ACTION_ATTR_POP_MPLS:
2214 case OVS_ACTION_ATTR_SET:
2215 case OVS_ACTION_ATTR_SAMPLE:
2216 case OVS_ACTION_ATTR_UNSPEC:
2217 case __OVS_ACTION_ATTR_MAX:
2223 dp_netdev_execute_actions(struct dp_netdev *dp, const struct miniflow *key,
2224 struct ofpbuf *packet, bool may_steal,
2225 struct pkt_metadata *md,
2226 const struct nlattr *actions, size_t actions_len)
2228 struct dp_netdev_execute_aux aux = {dp, key};
2230 odp_execute_actions(&aux, packet, may_steal, md,
2231 actions, actions_len, dp_execute_cb);
2234 const struct dpif_class dpif_netdev_class = {
2236 dpif_netdev_enumerate,
2237 dpif_netdev_port_open_type,
2240 dpif_netdev_destroy,
2243 dpif_netdev_get_stats,
2244 dpif_netdev_port_add,
2245 dpif_netdev_port_del,
2246 dpif_netdev_port_query_by_number,
2247 dpif_netdev_port_query_by_name,
2248 NULL, /* port_get_pid */
2249 dpif_netdev_port_dump_start,
2250 dpif_netdev_port_dump_next,
2251 dpif_netdev_port_dump_done,
2252 dpif_netdev_port_poll,
2253 dpif_netdev_port_poll_wait,
2254 dpif_netdev_flow_get,
2255 dpif_netdev_flow_put,
2256 dpif_netdev_flow_del,
2257 dpif_netdev_flow_flush,
2258 dpif_netdev_flow_dump_state_init,
2259 dpif_netdev_flow_dump_start,
2260 dpif_netdev_flow_dump_next,
2262 dpif_netdev_flow_dump_done,
2263 dpif_netdev_flow_dump_state_uninit,
2264 dpif_netdev_execute,
2266 dpif_netdev_recv_set,
2267 dpif_netdev_handlers_set,
2268 dpif_netdev_queue_to_priority,
2270 dpif_netdev_recv_wait,
2271 dpif_netdev_recv_purge,
2275 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2276 const char *argv[], void *aux OVS_UNUSED)
2278 struct dp_netdev_port *port;
2279 struct dp_netdev *dp;
2282 ovs_mutex_lock(&dp_netdev_mutex);
2283 dp = shash_find_data(&dp_netdevs, argv[1]);
2284 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2285 ovs_mutex_unlock(&dp_netdev_mutex);
2286 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2289 ovs_refcount_ref(&dp->ref_cnt);
2290 ovs_mutex_unlock(&dp_netdev_mutex);
2292 ovs_rwlock_wrlock(&dp->port_rwlock);
2293 if (get_port_by_name(dp, argv[2], &port)) {
2294 unixctl_command_reply_error(conn, "unknown port");
2298 port_no = u32_to_odp(atoi(argv[3]));
2299 if (!port_no || port_no == ODPP_NONE) {
2300 unixctl_command_reply_error(conn, "bad port number");
2303 if (dp_netdev_lookup_port(dp, port_no)) {
2304 unixctl_command_reply_error(conn, "port number already in use");
2307 hmap_remove(&dp->ports, &port->node);
2308 port->port_no = port_no;
2309 hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0));
2310 seq_change(dp->port_seq);
2311 unixctl_command_reply(conn, NULL);
2314 ovs_rwlock_unlock(&dp->port_rwlock);
2315 dp_netdev_unref(dp);
2319 dpif_dummy_register__(const char *type)
2321 struct dpif_class *class;
2323 class = xmalloc(sizeof *class);
2324 *class = dpif_netdev_class;
2325 class->type = xstrdup(type);
2326 dp_register_provider(class);
2330 dpif_dummy_register(bool override)
2337 dp_enumerate_types(&types);
2338 SSET_FOR_EACH (type, &types) {
2339 if (!dp_unregister_provider(type)) {
2340 dpif_dummy_register__(type);
2343 sset_destroy(&types);
2346 dpif_dummy_register__("dummy");
2348 unixctl_command_register("dpif-dummy/change-port-number",
2349 "DP PORT NEW-NUMBER",
2350 3, 3, dpif_dummy_change_port_number, NULL);