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) {
816 netdev_close(port->netdev);
817 netdev_restore_flags(port->sf);
819 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
820 netdev_rxq_close(port->rxq[i]);
828 get_port_by_name(struct dp_netdev *dp,
829 const char *devname, struct dp_netdev_port **portp)
830 OVS_REQ_RDLOCK(dp->port_rwlock)
832 struct dp_netdev_port *port;
834 HMAP_FOR_EACH (port, node, &dp->ports) {
835 if (!strcmp(netdev_get_name(port->netdev), devname)) {
844 do_del_port(struct dp_netdev *dp, odp_port_t port_no)
845 OVS_REQ_WRLOCK(dp->port_rwlock)
847 struct dp_netdev_port *port;
850 error = get_port_by_number(dp, port_no, &port);
855 hmap_remove(&dp->ports, &port->node);
856 seq_change(dp->port_seq);
857 if (netdev_is_pmd(port->netdev)) {
858 dp_netdev_reload_pmd_threads(dp);
866 answer_port_query(const struct dp_netdev_port *port,
867 struct dpif_port *dpif_port)
869 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
870 dpif_port->type = xstrdup(port->type);
871 dpif_port->port_no = port->port_no;
875 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
876 struct dpif_port *dpif_port)
878 struct dp_netdev *dp = get_dp_netdev(dpif);
879 struct dp_netdev_port *port;
882 ovs_rwlock_rdlock(&dp->port_rwlock);
883 error = get_port_by_number(dp, port_no, &port);
884 if (!error && dpif_port) {
885 answer_port_query(port, dpif_port);
887 ovs_rwlock_unlock(&dp->port_rwlock);
893 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
894 struct dpif_port *dpif_port)
896 struct dp_netdev *dp = get_dp_netdev(dpif);
897 struct dp_netdev_port *port;
900 ovs_rwlock_rdlock(&dp->port_rwlock);
901 error = get_port_by_name(dp, devname, &port);
902 if (!error && dpif_port) {
903 answer_port_query(port, dpif_port);
905 ovs_rwlock_unlock(&dp->port_rwlock);
911 dp_netdev_flow_free(struct dp_netdev_flow *flow)
913 struct dp_netdev_flow_stats *bucket;
916 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
917 ovs_mutex_destroy(&bucket->mutex);
918 free_cacheline(bucket);
920 ovsthread_stats_destroy(&flow->stats);
922 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
923 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
928 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
929 OVS_REQ_WRLOCK(dp->cls.rwlock)
930 OVS_REQUIRES(dp->flow_mutex)
932 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
933 struct hmap_node *node = CONST_CAST(struct hmap_node *, &flow->node);
935 classifier_remove(&dp->cls, cr);
936 hmap_remove(&dp->flow_table, node);
937 ovsrcu_postpone(dp_netdev_flow_free, flow);
941 dp_netdev_flow_flush(struct dp_netdev *dp)
943 struct dp_netdev_flow *netdev_flow, *next;
945 ovs_mutex_lock(&dp->flow_mutex);
946 fat_rwlock_wrlock(&dp->cls.rwlock);
947 HMAP_FOR_EACH_SAFE (netdev_flow, next, node, &dp->flow_table) {
948 dp_netdev_remove_flow(dp, netdev_flow);
950 fat_rwlock_unlock(&dp->cls.rwlock);
951 ovs_mutex_unlock(&dp->flow_mutex);
955 dpif_netdev_flow_flush(struct dpif *dpif)
957 struct dp_netdev *dp = get_dp_netdev(dpif);
959 dp_netdev_flow_flush(dp);
963 struct dp_netdev_port_state {
970 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
972 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
977 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
978 struct dpif_port *dpif_port)
980 struct dp_netdev_port_state *state = state_;
981 struct dp_netdev *dp = get_dp_netdev(dpif);
982 struct hmap_node *node;
985 ovs_rwlock_rdlock(&dp->port_rwlock);
986 node = hmap_at_position(&dp->ports, &state->bucket, &state->offset);
988 struct dp_netdev_port *port;
990 port = CONTAINER_OF(node, struct dp_netdev_port, node);
993 state->name = xstrdup(netdev_get_name(port->netdev));
994 dpif_port->name = state->name;
995 dpif_port->type = port->type;
996 dpif_port->port_no = port->port_no;
1002 ovs_rwlock_unlock(&dp->port_rwlock);
1008 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1010 struct dp_netdev_port_state *state = state_;
1017 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1019 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1020 uint64_t new_port_seq;
1023 new_port_seq = seq_read(dpif->dp->port_seq);
1024 if (dpif->last_port_seq != new_port_seq) {
1025 dpif->last_port_seq = new_port_seq;
1035 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1037 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1039 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1042 static struct dp_netdev_flow *
1043 dp_netdev_flow_cast(const struct cls_rule *cr)
1045 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1048 static struct dp_netdev_flow *
1049 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
1050 OVS_EXCLUDED(dp->cls.rwlock)
1052 struct dp_netdev_flow *netdev_flow;
1053 struct cls_rule *rule;
1055 fat_rwlock_rdlock(&dp->cls.rwlock);
1056 rule = classifier_lookup_miniflow_first(&dp->cls, key);
1057 netdev_flow = dp_netdev_flow_cast(rule);
1058 fat_rwlock_unlock(&dp->cls.rwlock);
1063 static struct dp_netdev_flow *
1064 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1065 OVS_REQ_RDLOCK(dp->cls.rwlock)
1067 struct dp_netdev_flow *netdev_flow;
1069 HMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1071 if (flow_equal(&netdev_flow->flow, flow)) {
1080 get_dpif_flow_stats(struct dp_netdev_flow *netdev_flow,
1081 struct dpif_flow_stats *stats)
1083 struct dp_netdev_flow_stats *bucket;
1086 memset(stats, 0, sizeof *stats);
1087 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1088 ovs_mutex_lock(&bucket->mutex);
1089 stats->n_packets += bucket->packet_count;
1090 stats->n_bytes += bucket->byte_count;
1091 stats->used = MAX(stats->used, bucket->used);
1092 stats->tcp_flags |= bucket->tcp_flags;
1093 ovs_mutex_unlock(&bucket->mutex);
1098 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1099 const struct nlattr *mask_key,
1100 uint32_t mask_key_len, const struct flow *flow,
1104 enum odp_key_fitness fitness;
1106 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1108 /* This should not happen: it indicates that
1109 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1110 * disagree on the acceptable form of a mask. Log the problem
1111 * as an error, with enough details to enable debugging. */
1112 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1114 if (!VLOG_DROP_ERR(&rl)) {
1118 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1120 VLOG_ERR("internal error parsing flow mask %s (%s)",
1121 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1128 enum mf_field_id id;
1129 /* No mask key, unwildcard everything except fields whose
1130 * prerequisities are not met. */
1131 memset(mask, 0x0, sizeof *mask);
1133 for (id = 0; id < MFF_N_IDS; ++id) {
1134 /* Skip registers and metadata. */
1135 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1136 && id != MFF_METADATA) {
1137 const struct mf_field *mf = mf_from_id(id);
1138 if (mf_are_prereqs_ok(mf, flow)) {
1139 mf_mask_field(mf, mask);
1145 /* Force unwildcard the in_port.
1147 * We need to do this even in the case where we unwildcard "everything"
1148 * above because "everything" only includes the 16-bit OpenFlow port number
1149 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1150 * port number mask->in_port.odp_port. */
1151 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1157 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1162 if (odp_flow_key_to_flow(key, key_len, flow)) {
1163 /* This should not happen: it indicates that odp_flow_key_from_flow()
1164 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1165 * flow. Log the problem as an error, with enough details to enable
1167 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1169 if (!VLOG_DROP_ERR(&rl)) {
1173 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1174 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1181 in_port = flow->in_port.odp_port;
1182 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1190 dpif_netdev_flow_get(const struct dpif *dpif,
1191 const struct nlattr *nl_key, size_t nl_key_len,
1192 struct ofpbuf **actionsp, struct dpif_flow_stats *stats)
1194 struct dp_netdev *dp = get_dp_netdev(dpif);
1195 struct dp_netdev_flow *netdev_flow;
1199 error = dpif_netdev_flow_from_nlattrs(nl_key, nl_key_len, &key);
1204 fat_rwlock_rdlock(&dp->cls.rwlock);
1205 netdev_flow = dp_netdev_find_flow(dp, &key);
1206 fat_rwlock_unlock(&dp->cls.rwlock);
1210 get_dpif_flow_stats(netdev_flow, stats);
1214 struct dp_netdev_actions *actions;
1216 actions = dp_netdev_flow_get_actions(netdev_flow);
1217 *actionsp = ofpbuf_clone_data(actions->actions, actions->size);
1227 dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow,
1228 const struct flow_wildcards *wc,
1229 const struct nlattr *actions,
1231 OVS_REQUIRES(dp->flow_mutex)
1233 struct dp_netdev_flow *netdev_flow;
1236 netdev_flow = xzalloc(sizeof *netdev_flow);
1237 *CONST_CAST(struct flow *, &netdev_flow->flow) = *flow;
1239 ovsthread_stats_init(&netdev_flow->stats);
1241 ovsrcu_set(&netdev_flow->actions,
1242 dp_netdev_actions_create(actions, actions_len));
1244 match_init(&match, flow, wc);
1245 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1246 &match, NETDEV_RULE_PRIORITY);
1247 fat_rwlock_wrlock(&dp->cls.rwlock);
1248 classifier_insert(&dp->cls,
1249 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1250 hmap_insert(&dp->flow_table,
1251 CONST_CAST(struct hmap_node *, &netdev_flow->node),
1252 flow_hash(flow, 0));
1253 fat_rwlock_unlock(&dp->cls.rwlock);
1259 clear_stats(struct dp_netdev_flow *netdev_flow)
1261 struct dp_netdev_flow_stats *bucket;
1264 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1265 ovs_mutex_lock(&bucket->mutex);
1267 bucket->packet_count = 0;
1268 bucket->byte_count = 0;
1269 bucket->tcp_flags = 0;
1270 ovs_mutex_unlock(&bucket->mutex);
1275 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1277 struct dp_netdev *dp = get_dp_netdev(dpif);
1278 struct dp_netdev_flow *netdev_flow;
1280 struct miniflow miniflow;
1281 struct flow_wildcards wc;
1284 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &flow);
1288 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1289 put->mask, put->mask_len,
1294 miniflow_init(&miniflow, &flow);
1296 ovs_mutex_lock(&dp->flow_mutex);
1297 netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
1299 if (put->flags & DPIF_FP_CREATE) {
1300 if (hmap_count(&dp->flow_table) < MAX_FLOWS) {
1302 memset(put->stats, 0, sizeof *put->stats);
1304 error = dp_netdev_flow_add(dp, &flow, &wc, put->actions,
1313 if (put->flags & DPIF_FP_MODIFY
1314 && flow_equal(&flow, &netdev_flow->flow)) {
1315 struct dp_netdev_actions *new_actions;
1316 struct dp_netdev_actions *old_actions;
1318 new_actions = dp_netdev_actions_create(put->actions,
1321 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1322 ovsrcu_set(&netdev_flow->actions, new_actions);
1325 get_dpif_flow_stats(netdev_flow, put->stats);
1327 if (put->flags & DPIF_FP_ZERO_STATS) {
1328 clear_stats(netdev_flow);
1331 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1332 } else if (put->flags & DPIF_FP_CREATE) {
1335 /* Overlapping flow. */
1339 ovs_mutex_unlock(&dp->flow_mutex);
1345 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1347 struct dp_netdev *dp = get_dp_netdev(dpif);
1348 struct dp_netdev_flow *netdev_flow;
1352 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1357 ovs_mutex_lock(&dp->flow_mutex);
1358 fat_rwlock_wrlock(&dp->cls.rwlock);
1359 netdev_flow = dp_netdev_find_flow(dp, &key);
1362 get_dpif_flow_stats(netdev_flow, del->stats);
1364 dp_netdev_remove_flow(dp, netdev_flow);
1368 fat_rwlock_unlock(&dp->cls.rwlock);
1369 ovs_mutex_unlock(&dp->flow_mutex);
1374 struct dp_netdev_flow_state {
1375 struct odputil_keybuf keybuf;
1376 struct odputil_keybuf maskbuf;
1377 struct dpif_flow_stats stats;
1380 struct dp_netdev_flow_iter {
1384 struct ovs_mutex mutex;
1388 dpif_netdev_flow_dump_state_init(void **statep)
1390 struct dp_netdev_flow_state *state;
1392 *statep = state = xmalloc(sizeof *state);
1396 dpif_netdev_flow_dump_state_uninit(void *state_)
1398 struct dp_netdev_flow_state *state = state_;
1404 dpif_netdev_flow_dump_start(const struct dpif *dpif OVS_UNUSED, void **iterp)
1406 struct dp_netdev_flow_iter *iter;
1408 *iterp = iter = xmalloc(sizeof *iter);
1412 ovs_mutex_init(&iter->mutex);
1416 /* XXX the caller must use 'actions' without quiescing */
1418 dpif_netdev_flow_dump_next(const struct dpif *dpif, void *iter_, void *state_,
1419 const struct nlattr **key, size_t *key_len,
1420 const struct nlattr **mask, size_t *mask_len,
1421 const struct nlattr **actions, size_t *actions_len,
1422 const struct dpif_flow_stats **stats)
1424 struct dp_netdev_flow_iter *iter = iter_;
1425 struct dp_netdev_flow_state *state = state_;
1426 struct dp_netdev *dp = get_dp_netdev(dpif);
1427 struct dp_netdev_flow *netdev_flow;
1428 struct flow_wildcards wc;
1431 ovs_mutex_lock(&iter->mutex);
1432 error = iter->status;
1434 struct hmap_node *node;
1436 fat_rwlock_rdlock(&dp->cls.rwlock);
1437 node = hmap_at_position(&dp->flow_table, &iter->bucket, &iter->offset);
1439 netdev_flow = CONTAINER_OF(node, struct dp_netdev_flow, node);
1441 fat_rwlock_unlock(&dp->cls.rwlock);
1443 iter->status = error = EOF;
1446 ovs_mutex_unlock(&iter->mutex);
1451 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1456 ofpbuf_use_stack(&buf, &state->keybuf, sizeof state->keybuf);
1457 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1458 netdev_flow->flow.in_port.odp_port);
1460 *key = ofpbuf_data(&buf);
1461 *key_len = ofpbuf_size(&buf);
1467 ofpbuf_use_stack(&buf, &state->maskbuf, sizeof state->maskbuf);
1468 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1469 odp_to_u32(wc.masks.in_port.odp_port),
1472 *mask = ofpbuf_data(&buf);
1473 *mask_len = ofpbuf_size(&buf);
1476 if (actions || stats) {
1478 struct dp_netdev_actions *dp_actions =
1479 dp_netdev_flow_get_actions(netdev_flow);
1481 *actions = dp_actions->actions;
1482 *actions_len = dp_actions->size;
1486 get_dpif_flow_stats(netdev_flow, &state->stats);
1487 *stats = &state->stats;
1495 dpif_netdev_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *iter_)
1497 struct dp_netdev_flow_iter *iter = iter_;
1499 ovs_mutex_destroy(&iter->mutex);
1505 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1507 struct dp_netdev *dp = get_dp_netdev(dpif);
1508 struct pkt_metadata *md = &execute->md;
1510 struct miniflow flow;
1511 uint32_t buf[FLOW_U32S];
1514 if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
1515 ofpbuf_size(execute->packet) > UINT16_MAX) {
1519 /* Extract flow key. */
1520 miniflow_initialize(&key.flow, key.buf);
1521 miniflow_extract(execute->packet, md, &key.flow);
1523 ovs_rwlock_rdlock(&dp->port_rwlock);
1524 dp_netdev_execute_actions(dp, &key.flow, execute->packet, false, md,
1525 execute->actions, execute->actions_len);
1526 ovs_rwlock_unlock(&dp->port_rwlock);
1532 dp_netdev_destroy_all_queues(struct dp_netdev *dp)
1533 OVS_REQ_WRLOCK(dp->queue_rwlock)
1537 dp_netdev_purge_queues(dp);
1539 for (i = 0; i < dp->n_handlers; i++) {
1540 struct dp_netdev_queue *q = &dp->handler_queues[i];
1542 ovs_mutex_destroy(&q->mutex);
1543 seq_destroy(q->seq);
1545 free(dp->handler_queues);
1546 dp->handler_queues = NULL;
1551 dp_netdev_refresh_queues(struct dp_netdev *dp, uint32_t n_handlers)
1552 OVS_REQ_WRLOCK(dp->queue_rwlock)
1554 if (dp->n_handlers != n_handlers) {
1557 dp_netdev_destroy_all_queues(dp);
1559 dp->n_handlers = n_handlers;
1560 dp->handler_queues = xzalloc(n_handlers * sizeof *dp->handler_queues);
1562 for (i = 0; i < n_handlers; i++) {
1563 struct dp_netdev_queue *q = &dp->handler_queues[i];
1565 ovs_mutex_init(&q->mutex);
1566 q->seq = seq_create();
1572 dpif_netdev_recv_set(struct dpif *dpif, bool enable)
1574 struct dp_netdev *dp = get_dp_netdev(dpif);
1576 if ((dp->handler_queues != NULL) == enable) {
1580 fat_rwlock_wrlock(&dp->queue_rwlock);
1582 dp_netdev_destroy_all_queues(dp);
1584 dp_netdev_refresh_queues(dp, 1);
1586 fat_rwlock_unlock(&dp->queue_rwlock);
1592 dpif_netdev_handlers_set(struct dpif *dpif, uint32_t n_handlers)
1594 struct dp_netdev *dp = get_dp_netdev(dpif);
1596 fat_rwlock_wrlock(&dp->queue_rwlock);
1597 if (dp->handler_queues) {
1598 dp_netdev_refresh_queues(dp, n_handlers);
1600 fat_rwlock_unlock(&dp->queue_rwlock);
1606 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1607 uint32_t queue_id, uint32_t *priority)
1609 *priority = queue_id;
1614 dp_netdev_recv_check(const struct dp_netdev *dp, const uint32_t handler_id)
1615 OVS_REQ_RDLOCK(dp->queue_rwlock)
1617 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1619 if (!dp->handler_queues) {
1620 VLOG_WARN_RL(&rl, "receiving upcall disabled");
1624 if (handler_id >= dp->n_handlers) {
1625 VLOG_WARN_RL(&rl, "handler index out of bound");
1633 dpif_netdev_recv(struct dpif *dpif, uint32_t handler_id,
1634 struct dpif_upcall *upcall, struct ofpbuf *buf)
1636 struct dp_netdev *dp = get_dp_netdev(dpif);
1637 struct dp_netdev_queue *q;
1640 fat_rwlock_rdlock(&dp->queue_rwlock);
1642 if (!dp_netdev_recv_check(dp, handler_id)) {
1647 q = &dp->handler_queues[handler_id];
1648 ovs_mutex_lock(&q->mutex);
1649 if (q->head != q->tail) {
1650 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
1652 *upcall = u->upcall;
1659 ovs_mutex_unlock(&q->mutex);
1662 fat_rwlock_unlock(&dp->queue_rwlock);
1668 dpif_netdev_recv_wait(struct dpif *dpif, uint32_t handler_id)
1670 struct dp_netdev *dp = get_dp_netdev(dpif);
1671 struct dp_netdev_queue *q;
1674 fat_rwlock_rdlock(&dp->queue_rwlock);
1676 if (!dp_netdev_recv_check(dp, handler_id)) {
1680 q = &dp->handler_queues[handler_id];
1681 ovs_mutex_lock(&q->mutex);
1682 seq = seq_read(q->seq);
1683 if (q->head != q->tail) {
1684 poll_immediate_wake();
1686 seq_wait(q->seq, seq);
1689 ovs_mutex_unlock(&q->mutex);
1692 fat_rwlock_unlock(&dp->queue_rwlock);
1696 dpif_netdev_recv_purge(struct dpif *dpif)
1698 struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif);
1700 fat_rwlock_wrlock(&dpif_netdev->dp->queue_rwlock);
1701 dp_netdev_purge_queues(dpif_netdev->dp);
1702 fat_rwlock_unlock(&dpif_netdev->dp->queue_rwlock);
1705 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1706 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1708 struct dp_netdev_actions *
1709 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1711 struct dp_netdev_actions *netdev_actions;
1713 netdev_actions = xmalloc(sizeof *netdev_actions);
1714 netdev_actions->actions = xmemdup(actions, size);
1715 netdev_actions->size = size;
1717 return netdev_actions;
1720 struct dp_netdev_actions *
1721 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
1723 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
1727 dp_netdev_actions_free(struct dp_netdev_actions *actions)
1729 free(actions->actions);
1735 dp_netdev_process_rxq_port(struct dp_netdev *dp,
1736 struct dp_netdev_port *port,
1737 struct netdev_rxq *rxq)
1739 struct ofpbuf *packet[NETDEV_MAX_RX_BATCH];
1742 error = netdev_rxq_recv(rxq, packet, &c);
1744 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port->port_no);
1747 for (i = 0; i < c; i++) {
1748 dp_netdev_port_input(dp, packet[i], &md);
1750 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1751 static struct vlog_rate_limit rl
1752 = VLOG_RATE_LIMIT_INIT(1, 5);
1754 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1755 netdev_get_name(port->netdev),
1756 ovs_strerror(error));
1761 dpif_netdev_run(struct dpif *dpif)
1763 struct dp_netdev_port *port;
1764 struct dp_netdev *dp = get_dp_netdev(dpif);
1766 ovs_rwlock_rdlock(&dp->port_rwlock);
1768 HMAP_FOR_EACH (port, node, &dp->ports) {
1769 if (!netdev_is_pmd(port->netdev)) {
1772 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1773 dp_netdev_process_rxq_port(dp, port, port->rxq[i]);
1778 ovs_rwlock_unlock(&dp->port_rwlock);
1782 dpif_netdev_wait(struct dpif *dpif)
1784 struct dp_netdev_port *port;
1785 struct dp_netdev *dp = get_dp_netdev(dpif);
1787 ovs_rwlock_rdlock(&dp->port_rwlock);
1789 HMAP_FOR_EACH (port, node, &dp->ports) {
1790 if (!netdev_is_pmd(port->netdev)) {
1793 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1794 netdev_rxq_wait(port->rxq[i]);
1798 ovs_rwlock_unlock(&dp->port_rwlock);
1802 struct dp_netdev_port *port;
1803 struct netdev_rxq *rx;
1807 pmd_load_queues(struct pmd_thread *f,
1808 struct rxq_poll **ppoll_list, int poll_cnt)
1810 struct dp_netdev *dp = f->dp;
1811 struct rxq_poll *poll_list = *ppoll_list;
1812 struct dp_netdev_port *port;
1817 /* Simple scheduler for netdev rx polling. */
1818 ovs_rwlock_rdlock(&dp->port_rwlock);
1819 for (i = 0; i < poll_cnt; i++) {
1820 port_unref(poll_list[i].port);
1826 HMAP_FOR_EACH (port, node, &f->dp->ports) {
1827 if (netdev_is_pmd(port->netdev)) {
1830 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1831 if ((index % dp->n_pmd_threads) == id) {
1832 poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1));
1835 poll_list[poll_cnt].port = port;
1836 poll_list[poll_cnt].rx = port->rxq[i];
1844 ovs_rwlock_unlock(&dp->port_rwlock);
1845 *ppoll_list = poll_list;
1850 pmd_thread_main(void *f_)
1852 struct pmd_thread *f = f_;
1853 struct dp_netdev *dp = f->dp;
1854 unsigned int lc = 0;
1855 struct rxq_poll *poll_list;
1856 unsigned int port_seq;
1863 pmd_thread_setaffinity_cpu(f->id);
1865 poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
1866 atomic_read(&f->change_seq, &port_seq);
1869 unsigned int c_port_seq;
1872 for (i = 0; i < poll_cnt; i++) {
1873 dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx);
1879 /* TODO: need completely userspace based signaling method.
1880 * to keep this thread entirely in userspace.
1881 * For now using atomic counter. */
1883 atomic_read_explicit(&f->change_seq, &c_port_seq, memory_order_consume);
1884 if (c_port_seq != port_seq) {
1890 if (!latch_is_set(&f->dp->exit_latch)){
1894 for (i = 0; i < poll_cnt; i++) {
1895 port_unref(poll_list[i].port);
1903 dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n)
1907 if (n == dp->n_pmd_threads) {
1911 /* Stop existing threads. */
1912 latch_set(&dp->exit_latch);
1913 dp_netdev_reload_pmd_threads(dp);
1914 for (i = 0; i < dp->n_pmd_threads; i++) {
1915 struct pmd_thread *f = &dp->pmd_threads[i];
1917 xpthread_join(f->thread, NULL);
1919 latch_poll(&dp->exit_latch);
1920 free(dp->pmd_threads);
1922 /* Start new threads. */
1923 dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
1924 dp->n_pmd_threads = n;
1926 for (i = 0; i < n; i++) {
1927 struct pmd_thread *f = &dp->pmd_threads[i];
1931 atomic_store(&f->change_seq, 1);
1933 /* Each thread will distribute all devices rx-queues among
1935 f->thread = ovs_thread_create("pmd", pmd_thread_main, f);
1941 dp_netdev_flow_stats_new_cb(void)
1943 struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1944 ovs_mutex_init(&bucket->mutex);
1949 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
1950 const struct ofpbuf *packet,
1951 const struct miniflow *key)
1953 uint16_t tcp_flags = miniflow_get_tcp_flags(key);
1954 long long int now = time_msec();
1955 struct dp_netdev_flow_stats *bucket;
1957 bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
1958 dp_netdev_flow_stats_new_cb);
1960 ovs_mutex_lock(&bucket->mutex);
1961 bucket->used = MAX(now, bucket->used);
1962 bucket->packet_count++;
1963 bucket->byte_count += ofpbuf_size(packet);
1964 bucket->tcp_flags |= tcp_flags;
1965 ovs_mutex_unlock(&bucket->mutex);
1969 dp_netdev_stats_new_cb(void)
1971 struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1972 ovs_mutex_init(&bucket->mutex);
1977 dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type)
1979 struct dp_netdev_stats *bucket;
1981 bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
1982 ovs_mutex_lock(&bucket->mutex);
1984 ovs_mutex_unlock(&bucket->mutex);
1988 dp_netdev_input(struct dp_netdev *dp, struct ofpbuf *packet,
1989 struct pkt_metadata *md)
1990 OVS_REQ_RDLOCK(dp->port_rwlock)
1992 struct dp_netdev_flow *netdev_flow;
1994 struct miniflow flow;
1995 uint32_t buf[FLOW_U32S];
1998 if (ofpbuf_size(packet) < ETH_HEADER_LEN) {
1999 ofpbuf_delete(packet);
2002 miniflow_initialize(&key.flow, key.buf);
2003 miniflow_extract(packet, md, &key.flow);
2005 netdev_flow = dp_netdev_lookup_flow(dp, &key.flow);
2007 struct dp_netdev_actions *actions;
2009 dp_netdev_flow_used(netdev_flow, packet, &key.flow);
2011 actions = dp_netdev_flow_get_actions(netdev_flow);
2012 dp_netdev_execute_actions(dp, &key.flow, packet, true, md,
2013 actions->actions, actions->size);
2014 dp_netdev_count_packet(dp, DP_STAT_HIT);
2015 } else if (dp->handler_queues) {
2016 dp_netdev_count_packet(dp, DP_STAT_MISS);
2017 dp_netdev_output_userspace(dp, packet,
2018 miniflow_hash_5tuple(&key.flow, 0)
2020 DPIF_UC_MISS, &key.flow, NULL);
2021 ofpbuf_delete(packet);
2026 dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet,
2027 struct pkt_metadata *md)
2028 OVS_REQ_RDLOCK(dp->port_rwlock)
2030 uint32_t *recirc_depth = recirc_depth_get();
2033 dp_netdev_input(dp, packet, md);
2037 dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet,
2038 int queue_no, int type, const struct miniflow *key,
2039 const struct nlattr *userdata)
2041 struct dp_netdev_queue *q;
2044 fat_rwlock_rdlock(&dp->queue_rwlock);
2045 q = &dp->handler_queues[queue_no];
2046 ovs_mutex_lock(&q->mutex);
2047 if (q->head - q->tail < MAX_QUEUE_LEN) {
2048 struct dp_netdev_upcall *u = &q->upcalls[q->head++ & QUEUE_MASK];
2049 struct dpif_upcall *upcall = &u->upcall;
2050 struct ofpbuf *buf = &u->buf;
2054 upcall->type = type;
2056 /* Allocate buffer big enough for everything. */
2057 buf_size = ODPUTIL_FLOW_KEY_BYTES;
2059 buf_size += NLA_ALIGN(userdata->nla_len);
2061 buf_size += ofpbuf_size(packet);
2062 ofpbuf_init(buf, buf_size);
2065 miniflow_expand(key, &flow);
2066 odp_flow_key_from_flow(buf, &flow, NULL, flow.in_port.odp_port);
2067 upcall->key = ofpbuf_data(buf);
2068 upcall->key_len = ofpbuf_size(buf);
2072 upcall->userdata = ofpbuf_put(buf, userdata,
2073 NLA_ALIGN(userdata->nla_len));
2076 ofpbuf_set_data(&upcall->packet,
2077 ofpbuf_put(buf, ofpbuf_data(packet), ofpbuf_size(packet)));
2078 ofpbuf_set_size(&upcall->packet, ofpbuf_size(packet));
2084 dp_netdev_count_packet(dp, DP_STAT_LOST);
2087 ovs_mutex_unlock(&q->mutex);
2088 fat_rwlock_unlock(&dp->queue_rwlock);
2093 struct dp_netdev_execute_aux {
2094 struct dp_netdev *dp;
2095 const struct miniflow *key;
2099 dp_execute_cb(void *aux_, struct ofpbuf *packet,
2100 struct pkt_metadata *md,
2101 const struct nlattr *a, bool may_steal)
2102 OVS_NO_THREAD_SAFETY_ANALYSIS
2104 struct dp_netdev_execute_aux *aux = aux_;
2105 int type = nl_attr_type(a);
2106 struct dp_netdev_port *p;
2107 uint32_t *depth = recirc_depth_get();
2109 switch ((enum ovs_action_attr)type) {
2110 case OVS_ACTION_ATTR_OUTPUT:
2111 p = dp_netdev_lookup_port(aux->dp, u32_to_odp(nl_attr_get_u32(a)));
2113 netdev_send(p->netdev, packet, may_steal);
2117 case OVS_ACTION_ATTR_USERSPACE: {
2118 const struct nlattr *userdata;
2120 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2122 dp_netdev_output_userspace(aux->dp, packet,
2123 miniflow_hash_5tuple(aux->key, 0)
2124 % aux->dp->n_handlers,
2125 DPIF_UC_ACTION, aux->key,
2129 ofpbuf_delete(packet);
2134 case OVS_ACTION_ATTR_HASH: {
2135 const struct ovs_action_hash *hash_act;
2138 hash_act = nl_attr_get(a);
2139 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
2140 /* Hash need not be symmetric, nor does it need to include
2142 hash = miniflow_hash_5tuple(aux->key, hash_act->hash_basis);
2144 hash = 1; /* 0 is not valid */
2148 VLOG_WARN("Unknown hash algorithm specified for the hash action.");
2156 case OVS_ACTION_ATTR_RECIRC:
2157 if (*depth < MAX_RECIRC_DEPTH) {
2158 struct pkt_metadata recirc_md = *md;
2159 struct ofpbuf *recirc_packet;
2161 recirc_packet = may_steal ? packet : ofpbuf_clone(packet);
2162 recirc_md.recirc_id = nl_attr_get_u32(a);
2165 dp_netdev_input(aux->dp, recirc_packet, &recirc_md);
2170 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
2174 case OVS_ACTION_ATTR_PUSH_VLAN:
2175 case OVS_ACTION_ATTR_POP_VLAN:
2176 case OVS_ACTION_ATTR_PUSH_MPLS:
2177 case OVS_ACTION_ATTR_POP_MPLS:
2178 case OVS_ACTION_ATTR_SET:
2179 case OVS_ACTION_ATTR_SAMPLE:
2180 case OVS_ACTION_ATTR_UNSPEC:
2181 case __OVS_ACTION_ATTR_MAX:
2187 dp_netdev_execute_actions(struct dp_netdev *dp, const struct miniflow *key,
2188 struct ofpbuf *packet, bool may_steal,
2189 struct pkt_metadata *md,
2190 const struct nlattr *actions, size_t actions_len)
2192 struct dp_netdev_execute_aux aux = {dp, key};
2194 odp_execute_actions(&aux, packet, may_steal, md,
2195 actions, actions_len, dp_execute_cb);
2198 const struct dpif_class dpif_netdev_class = {
2200 dpif_netdev_enumerate,
2201 dpif_netdev_port_open_type,
2204 dpif_netdev_destroy,
2207 dpif_netdev_get_stats,
2208 dpif_netdev_port_add,
2209 dpif_netdev_port_del,
2210 dpif_netdev_port_query_by_number,
2211 dpif_netdev_port_query_by_name,
2212 NULL, /* port_get_pid */
2213 dpif_netdev_port_dump_start,
2214 dpif_netdev_port_dump_next,
2215 dpif_netdev_port_dump_done,
2216 dpif_netdev_port_poll,
2217 dpif_netdev_port_poll_wait,
2218 dpif_netdev_flow_get,
2219 dpif_netdev_flow_put,
2220 dpif_netdev_flow_del,
2221 dpif_netdev_flow_flush,
2222 dpif_netdev_flow_dump_state_init,
2223 dpif_netdev_flow_dump_start,
2224 dpif_netdev_flow_dump_next,
2226 dpif_netdev_flow_dump_done,
2227 dpif_netdev_flow_dump_state_uninit,
2228 dpif_netdev_execute,
2230 dpif_netdev_recv_set,
2231 dpif_netdev_handlers_set,
2232 dpif_netdev_queue_to_priority,
2234 dpif_netdev_recv_wait,
2235 dpif_netdev_recv_purge,
2239 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2240 const char *argv[], void *aux OVS_UNUSED)
2242 struct dp_netdev_port *port;
2243 struct dp_netdev *dp;
2246 ovs_mutex_lock(&dp_netdev_mutex);
2247 dp = shash_find_data(&dp_netdevs, argv[1]);
2248 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2249 ovs_mutex_unlock(&dp_netdev_mutex);
2250 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2253 ovs_refcount_ref(&dp->ref_cnt);
2254 ovs_mutex_unlock(&dp_netdev_mutex);
2256 ovs_rwlock_wrlock(&dp->port_rwlock);
2257 if (get_port_by_name(dp, argv[2], &port)) {
2258 unixctl_command_reply_error(conn, "unknown port");
2262 port_no = u32_to_odp(atoi(argv[3]));
2263 if (!port_no || port_no == ODPP_NONE) {
2264 unixctl_command_reply_error(conn, "bad port number");
2267 if (dp_netdev_lookup_port(dp, port_no)) {
2268 unixctl_command_reply_error(conn, "port number already in use");
2271 hmap_remove(&dp->ports, &port->node);
2272 port->port_no = port_no;
2273 hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0));
2274 seq_change(dp->port_seq);
2275 unixctl_command_reply(conn, NULL);
2278 ovs_rwlock_unlock(&dp->port_rwlock);
2279 dp_netdev_unref(dp);
2283 dpif_dummy_register__(const char *type)
2285 struct dpif_class *class;
2287 class = xmalloc(sizeof *class);
2288 *class = dpif_netdev_class;
2289 class->type = xstrdup(type);
2290 dp_register_provider(class);
2294 dpif_dummy_register(bool override)
2301 dp_enumerate_types(&types);
2302 SSET_FOR_EACH (type, &types) {
2303 if (!dp_unregister_provider(type)) {
2304 dpif_dummy_register__(type);
2307 sset_destroy(&types);
2310 dpif_dummy_register__("dummy");
2312 unixctl_command_register("dpif-dummy/change-port-number",
2313 "DP PORT NEW-NUMBER",
2314 3, 3, dpif_dummy_change_port_number, NULL);