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"
43 #include "meta-flow.h"
45 #include "netdev-vport.h"
47 #include "odp-execute.h"
49 #include "ofp-print.h"
52 #include "poll-loop.h"
62 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
64 /* By default, choose a priority in the middle. */
65 #define NETDEV_RULE_PRIORITY 0x8000
67 /* Configuration parameters. */
68 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
70 /* Enough headroom to add a vlan tag, plus an extra 2 bytes to allow IP
71 * headers to be aligned on a 4-byte boundary. */
72 enum { DP_NETDEV_HEADROOM = 2 + VLAN_HEADER_LEN };
75 enum { N_QUEUES = 2 }; /* Number of queues for dpif_recv(). */
76 enum { MAX_QUEUE_LEN = 128 }; /* Maximum number of packets per queue. */
77 enum { QUEUE_MASK = MAX_QUEUE_LEN - 1 };
78 BUILD_ASSERT_DECL(IS_POW2(MAX_QUEUE_LEN));
80 struct dp_netdev_upcall {
81 struct dpif_upcall upcall; /* Queued upcall information. */
82 struct ofpbuf buf; /* ofpbuf instance for upcall.packet. */
85 struct dp_netdev_queue {
86 struct dp_netdev_upcall upcalls[MAX_QUEUE_LEN];
87 unsigned int head, tail;
90 /* Datapath based on the network device interface from netdev.h. */
92 const struct dpif_class *class;
94 struct ovs_refcount ref_cnt;
95 atomic_flag destroyed;
97 struct dp_netdev_queue queues[N_QUEUES];
98 struct classifier cls; /* Classifier. */
99 struct hmap flow_table; /* Flow table. */
100 struct seq *queue_seq; /* Incremented whenever a packet is queued. */
103 struct ovsthread_counter *n_hit; /* Number of flow table matches. */
104 struct ovsthread_counter *n_missed; /* Number of flow table misses. */
105 struct ovsthread_counter *n_lost; /* Number of misses not passed up. */
109 struct seq *port_seq; /* Incremented whenever a port changes. */
112 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *,
115 /* A port in a netdev-based datapath. */
116 struct dp_netdev_port {
117 struct hmap_node node; /* Node in dp_netdev's 'ports'. */
119 struct netdev *netdev;
120 struct netdev_saved_flags *sf;
121 struct netdev_rx *rx;
122 char *type; /* Port type as requested by user. */
125 /* A flow in dp_netdev's 'flow_table'. */
126 struct dp_netdev_flow {
127 /* Packet classification. */
128 struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
130 /* Hash table index by unmasked flow.*/
131 struct hmap_node node; /* In owning dp_netdev's 'flow_table'. */
132 struct flow flow; /* The flow that created this entry. */
135 long long int used; /* Last used time, in monotonic msecs. */
136 long long int packet_count; /* Number of packets matched. */
137 long long int byte_count; /* Number of bytes matched. */
138 uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
141 struct dp_netdev_actions *actions;
144 /* A set of datapath actions within a "struct dp_netdev_flow".
150 * A struct dp_netdev_actions 'actions' may be accessed without a risk of being
151 * freed by code that holds a read-lock or write-lock on 'flow->mutex' (where
152 * 'flow' is the dp_netdev_flow for which 'flow->actions == actions') or that
153 * owns a reference to 'actions->ref_cnt' (or both). */
154 struct dp_netdev_actions {
155 struct ovs_refcount ref_cnt;
157 /* These members are immutable: they do not change during the struct's
159 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
160 unsigned int size; /* Size of 'actions', in bytes. */
163 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
165 struct dp_netdev_actions *dp_netdev_actions_ref(
166 const struct dp_netdev_actions *);
167 void dp_netdev_actions_unref(struct dp_netdev_actions *);
169 /* Interface to netdev-based datapath. */
172 struct dp_netdev *dp;
173 uint64_t last_port_seq;
176 /* All netdev-based datapaths. */
177 static struct shash dp_netdevs = SHASH_INITIALIZER(&dp_netdevs);
179 /* Global lock for all data. */
180 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
182 static int get_port_by_number(struct dp_netdev *, odp_port_t port_no,
183 struct dp_netdev_port **portp);
184 static int get_port_by_name(struct dp_netdev *, const char *devname,
185 struct dp_netdev_port **portp);
186 static void dp_netdev_free(struct dp_netdev *);
187 static void dp_netdev_flow_flush(struct dp_netdev *);
188 static int do_add_port(struct dp_netdev *, const char *devname,
189 const char *type, odp_port_t port_no);
190 static int do_del_port(struct dp_netdev *, odp_port_t port_no);
191 static int dpif_netdev_open(const struct dpif_class *, const char *name,
192 bool create, struct dpif **);
193 static int dp_netdev_output_userspace(struct dp_netdev *, struct ofpbuf *,
194 int queue_no, const struct flow *,
195 const struct nlattr *userdata);
196 static void dp_netdev_execute_actions(struct dp_netdev *, const struct flow *,
197 struct ofpbuf *, struct pkt_metadata *,
198 const struct nlattr *actions,
200 static void dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet,
201 struct pkt_metadata *md);
203 static struct dpif_netdev *
204 dpif_netdev_cast(const struct dpif *dpif)
206 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
207 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
210 static struct dp_netdev *
211 get_dp_netdev(const struct dpif *dpif)
213 return dpif_netdev_cast(dpif)->dp;
217 dpif_netdev_enumerate(struct sset *all_dps)
219 struct shash_node *node;
221 ovs_mutex_lock(&dp_netdev_mutex);
222 SHASH_FOR_EACH(node, &dp_netdevs) {
223 sset_add(all_dps, node->name);
225 ovs_mutex_unlock(&dp_netdev_mutex);
231 dpif_netdev_class_is_dummy(const struct dpif_class *class)
233 return class != &dpif_netdev_class;
237 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
239 return strcmp(type, "internal") ? type
240 : dpif_netdev_class_is_dummy(class) ? "dummy"
245 create_dpif_netdev(struct dp_netdev *dp)
247 uint16_t netflow_id = hash_string(dp->name, 0);
248 struct dpif_netdev *dpif;
250 ovs_refcount_ref(&dp->ref_cnt);
252 dpif = xmalloc(sizeof *dpif);
253 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
255 dpif->last_port_seq = seq_read(dp->port_seq);
260 /* Choose an unused, non-zero port number and return it on success.
261 * Return ODPP_NONE on failure. */
263 choose_port(struct dp_netdev *dp, const char *name)
267 if (dp->class != &dpif_netdev_class) {
271 /* If the port name begins with "br", start the number search at
272 * 100 to make writing tests easier. */
273 if (!strncmp(name, "br", 2)) {
277 /* If the port name contains a number, try to assign that port number.
278 * This can make writing unit tests easier because port numbers are
280 for (p = name; *p != '\0'; p++) {
281 if (isdigit((unsigned char) *p)) {
282 port_no = start_no + strtol(p, NULL, 10);
283 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
284 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
285 return u32_to_odp(port_no);
292 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
293 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
294 return u32_to_odp(port_no);
302 create_dp_netdev(const char *name, const struct dpif_class *class,
303 struct dp_netdev **dpp)
305 struct dp_netdev *dp;
309 dp = xzalloc(sizeof *dp);
311 dp->name = xstrdup(name);
312 ovs_refcount_init(&dp->ref_cnt);
313 atomic_flag_init(&dp->destroyed);
314 for (i = 0; i < N_QUEUES; i++) {
315 dp->queues[i].head = dp->queues[i].tail = 0;
317 dp->queue_seq = seq_create();
318 classifier_init(&dp->cls, NULL);
319 hmap_init(&dp->flow_table);
321 dp->n_hit = ovsthread_counter_create();
322 dp->n_missed = ovsthread_counter_create();
323 dp->n_lost = ovsthread_counter_create();
325 hmap_init(&dp->ports);
326 dp->port_seq = seq_create();
328 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
334 shash_add(&dp_netdevs, name, dp);
341 dpif_netdev_open(const struct dpif_class *class, const char *name,
342 bool create, struct dpif **dpifp)
344 struct dp_netdev *dp;
347 ovs_mutex_lock(&dp_netdev_mutex);
348 dp = shash_find_data(&dp_netdevs, name);
350 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
352 error = (dp->class != class ? EINVAL
357 *dpifp = create_dpif_netdev(dp);
359 ovs_mutex_unlock(&dp_netdev_mutex);
365 dp_netdev_purge_queues(struct dp_netdev *dp)
369 for (i = 0; i < N_QUEUES; i++) {
370 struct dp_netdev_queue *q = &dp->queues[i];
372 while (q->tail != q->head) {
373 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
374 ofpbuf_uninit(&u->upcall.packet);
375 ofpbuf_uninit(&u->buf);
381 dp_netdev_free(struct dp_netdev *dp)
383 struct dp_netdev_port *port, *next;
385 dp_netdev_flow_flush(dp);
386 HMAP_FOR_EACH_SAFE (port, next, node, &dp->ports) {
387 do_del_port(dp, port->port_no);
389 ovsthread_counter_destroy(dp->n_hit);
390 ovsthread_counter_destroy(dp->n_missed);
391 ovsthread_counter_destroy(dp->n_lost);
392 dp_netdev_purge_queues(dp);
393 seq_destroy(dp->queue_seq);
394 classifier_destroy(&dp->cls);
395 hmap_destroy(&dp->flow_table);
396 seq_destroy(dp->port_seq);
397 hmap_destroy(&dp->ports);
398 atomic_flag_destroy(&dp->destroyed);
399 ovs_refcount_destroy(&dp->ref_cnt);
405 dpif_netdev_close(struct dpif *dpif)
407 struct dp_netdev *dp = get_dp_netdev(dpif);
409 ovs_mutex_lock(&dp_netdev_mutex);
411 if (ovs_refcount_unref(&dp->ref_cnt) == 1) {
412 shash_find_and_delete(&dp_netdevs, dp->name);
417 ovs_mutex_unlock(&dp_netdev_mutex);
421 dpif_netdev_destroy(struct dpif *dpif)
423 struct dp_netdev *dp = get_dp_netdev(dpif);
425 if (!atomic_flag_test_and_set(&dp->destroyed)) {
426 if (ovs_refcount_unref(&dp->ref_cnt) == 1) {
427 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
436 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
438 struct dp_netdev *dp = get_dp_netdev(dpif);
440 ovs_mutex_lock(&dp_netdev_mutex);
441 stats->n_flows = hmap_count(&dp->flow_table);
442 stats->n_hit = ovsthread_counter_read(dp->n_hit);
443 stats->n_missed = ovsthread_counter_read(dp->n_missed);
444 stats->n_lost = ovsthread_counter_read(dp->n_lost);
445 stats->n_masks = UINT32_MAX;
446 stats->n_mask_hit = UINT64_MAX;
447 ovs_mutex_unlock(&dp_netdev_mutex);
453 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
456 struct netdev_saved_flags *sf;
457 struct dp_netdev_port *port;
458 struct netdev *netdev;
459 struct netdev_rx *rx;
460 enum netdev_flags flags;
461 const char *open_type;
464 /* XXX reject devices already in some dp_netdev. */
466 /* Open and validate network device. */
467 open_type = dpif_netdev_port_open_type(dp->class, type);
468 error = netdev_open(devname, open_type, &netdev);
472 /* XXX reject non-Ethernet devices */
474 netdev_get_flags(netdev, &flags);
475 if (flags & NETDEV_LOOPBACK) {
476 VLOG_ERR("%s: cannot add a loopback device", devname);
477 netdev_close(netdev);
481 error = netdev_rx_open(netdev, &rx);
483 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
484 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
485 devname, ovs_strerror(errno));
486 netdev_close(netdev);
490 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
493 netdev_close(netdev);
497 port = xmalloc(sizeof *port);
498 port->port_no = port_no;
499 port->netdev = netdev;
502 port->type = xstrdup(type);
504 hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0));
505 seq_change(dp->port_seq);
511 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
512 odp_port_t *port_nop)
514 struct dp_netdev *dp = get_dp_netdev(dpif);
515 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
516 const char *dpif_port;
520 ovs_mutex_lock(&dp_netdev_mutex);
521 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
522 if (*port_nop != ODPP_NONE) {
524 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
526 port_no = choose_port(dp, dpif_port);
527 error = port_no == ODPP_NONE ? EFBIG : 0;
531 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
533 ovs_mutex_unlock(&dp_netdev_mutex);
539 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
541 struct dp_netdev *dp = get_dp_netdev(dpif);
544 ovs_mutex_lock(&dp_netdev_mutex);
545 error = port_no == ODPP_LOCAL ? EINVAL : do_del_port(dp, port_no);
546 ovs_mutex_unlock(&dp_netdev_mutex);
552 is_valid_port_number(odp_port_t port_no)
554 return port_no != ODPP_NONE;
557 static struct dp_netdev_port *
558 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
560 struct dp_netdev_port *port;
562 HMAP_FOR_EACH_IN_BUCKET (port, node, hash_int(odp_to_u32(port_no), 0),
564 if (port->port_no == port_no) {
572 get_port_by_number(struct dp_netdev *dp,
573 odp_port_t port_no, struct dp_netdev_port **portp)
575 if (!is_valid_port_number(port_no)) {
579 *portp = dp_netdev_lookup_port(dp, port_no);
580 return *portp ? 0 : ENOENT;
585 get_port_by_name(struct dp_netdev *dp,
586 const char *devname, struct dp_netdev_port **portp)
588 struct dp_netdev_port *port;
590 HMAP_FOR_EACH (port, node, &dp->ports) {
591 if (!strcmp(netdev_get_name(port->netdev), devname)) {
600 do_del_port(struct dp_netdev *dp, odp_port_t port_no)
602 struct dp_netdev_port *port;
605 error = get_port_by_number(dp, port_no, &port);
610 hmap_remove(&dp->ports, &port->node);
611 seq_change(dp->port_seq);
613 netdev_close(port->netdev);
614 netdev_restore_flags(port->sf);
615 netdev_rx_close(port->rx);
623 answer_port_query(const struct dp_netdev_port *port,
624 struct dpif_port *dpif_port)
626 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
627 dpif_port->type = xstrdup(port->type);
628 dpif_port->port_no = port->port_no;
632 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
633 struct dpif_port *dpif_port)
635 struct dp_netdev *dp = get_dp_netdev(dpif);
636 struct dp_netdev_port *port;
639 ovs_mutex_lock(&dp_netdev_mutex);
640 error = get_port_by_number(dp, port_no, &port);
641 if (!error && dpif_port) {
642 answer_port_query(port, dpif_port);
644 ovs_mutex_unlock(&dp_netdev_mutex);
650 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
651 struct dpif_port *dpif_port)
653 struct dp_netdev *dp = get_dp_netdev(dpif);
654 struct dp_netdev_port *port;
657 ovs_mutex_lock(&dp_netdev_mutex);
658 error = get_port_by_name(dp, devname, &port);
659 if (!error && dpif_port) {
660 answer_port_query(port, dpif_port);
662 ovs_mutex_unlock(&dp_netdev_mutex);
668 dp_netdev_free_flow(struct dp_netdev *dp, struct dp_netdev_flow *netdev_flow)
670 ovs_rwlock_wrlock(&dp->cls.rwlock);
671 classifier_remove(&dp->cls, &netdev_flow->cr);
672 ovs_rwlock_unlock(&dp->cls.rwlock);
673 cls_rule_destroy(&netdev_flow->cr);
675 hmap_remove(&dp->flow_table, &netdev_flow->node);
676 free(netdev_flow->actions);
681 dp_netdev_flow_flush(struct dp_netdev *dp)
683 struct dp_netdev_flow *netdev_flow, *next;
685 HMAP_FOR_EACH_SAFE (netdev_flow, next, node, &dp->flow_table) {
686 dp_netdev_free_flow(dp, netdev_flow);
691 dpif_netdev_flow_flush(struct dpif *dpif)
693 struct dp_netdev *dp = get_dp_netdev(dpif);
695 ovs_mutex_lock(&dp_netdev_mutex);
696 dp_netdev_flow_flush(dp);
697 ovs_mutex_unlock(&dp_netdev_mutex);
702 struct dp_netdev_port_state {
709 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
711 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
716 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
717 struct dpif_port *dpif_port)
719 struct dp_netdev_port_state *state = state_;
720 struct dp_netdev *dp = get_dp_netdev(dpif);
721 struct hmap_node *node;
724 ovs_mutex_lock(&dp_netdev_mutex);
725 node = hmap_at_position(&dp->ports, &state->bucket, &state->offset);
727 struct dp_netdev_port *port;
729 port = CONTAINER_OF(node, struct dp_netdev_port, node);
732 state->name = xstrdup(netdev_get_name(port->netdev));
733 dpif_port->name = state->name;
734 dpif_port->type = port->type;
735 dpif_port->port_no = port->port_no;
741 ovs_mutex_unlock(&dp_netdev_mutex);
747 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
749 struct dp_netdev_port_state *state = state_;
756 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
758 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
759 uint64_t new_port_seq;
762 ovs_mutex_lock(&dp_netdev_mutex);
763 new_port_seq = seq_read(dpif->dp->port_seq);
764 if (dpif->last_port_seq != new_port_seq) {
765 dpif->last_port_seq = new_port_seq;
770 ovs_mutex_unlock(&dp_netdev_mutex);
776 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
778 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
780 ovs_mutex_lock(&dp_netdev_mutex);
781 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
782 ovs_mutex_unlock(&dp_netdev_mutex);
785 static struct dp_netdev_flow *
786 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct flow *flow)
790 ovs_rwlock_wrlock(&dp->cls.rwlock);
791 cr = classifier_lookup(&dp->cls, flow, NULL);
792 ovs_rwlock_unlock(&dp->cls.rwlock);
795 ? CONTAINER_OF(cr, struct dp_netdev_flow, cr)
799 static struct dp_netdev_flow *
800 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
802 struct dp_netdev_flow *netdev_flow;
804 HMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
806 if (flow_equal(&netdev_flow->flow, flow)) {
814 get_dpif_flow_stats(struct dp_netdev_flow *netdev_flow,
815 struct dpif_flow_stats *stats)
817 stats->n_packets = netdev_flow->packet_count;
818 stats->n_bytes = netdev_flow->byte_count;
819 stats->used = netdev_flow->used;
820 stats->tcp_flags = netdev_flow->tcp_flags;
824 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
825 const struct nlattr *mask_key,
826 uint32_t mask_key_len, const struct flow *flow,
830 if (odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow)) {
831 /* This should not happen: it indicates that
832 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
833 * disagree on the acceptable form of a mask. Log the problem
834 * as an error, with enough details to enable debugging. */
835 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
837 if (!VLOG_DROP_ERR(&rl)) {
841 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
843 VLOG_ERR("internal error parsing flow mask %s", ds_cstr(&s));
849 /* Force unwildcard the in_port. */
850 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
853 /* No mask key, unwildcard everything except fields whose
854 * prerequisities are not met. */
855 memset(mask, 0x0, sizeof *mask);
857 for (id = 0; id < MFF_N_IDS; ++id) {
858 /* Skip registers and metadata. */
859 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
860 && id != MFF_METADATA) {
861 const struct mf_field *mf = mf_from_id(id);
862 if (mf_are_prereqs_ok(mf, flow)) {
863 mf_mask_field(mf, mask);
873 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
878 if (odp_flow_key_to_flow(key, key_len, flow)) {
879 /* This should not happen: it indicates that odp_flow_key_from_flow()
880 * and odp_flow_key_to_flow() disagree on the acceptable form of a
881 * flow. Log the problem as an error, with enough details to enable
883 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
885 if (!VLOG_DROP_ERR(&rl)) {
889 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
890 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
897 in_port = flow->in_port.odp_port;
898 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
906 dpif_netdev_flow_get(const struct dpif *dpif,
907 const struct nlattr *nl_key, size_t nl_key_len,
908 struct ofpbuf **actionsp, struct dpif_flow_stats *stats)
910 struct dp_netdev *dp = get_dp_netdev(dpif);
911 struct dp_netdev_flow *netdev_flow;
915 error = dpif_netdev_flow_from_nlattrs(nl_key, nl_key_len, &key);
920 ovs_mutex_lock(&dp_netdev_mutex);
921 netdev_flow = dp_netdev_find_flow(dp, &key);
924 get_dpif_flow_stats(netdev_flow, stats);
927 *actionsp = ofpbuf_clone_data(netdev_flow->actions->actions,
928 netdev_flow->actions->size);
933 ovs_mutex_unlock(&dp_netdev_mutex);
939 dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow,
940 const struct flow_wildcards *wc,
941 const struct nlattr *actions,
944 struct dp_netdev_flow *netdev_flow;
947 netdev_flow = xzalloc(sizeof *netdev_flow);
948 netdev_flow->flow = *flow;
949 netdev_flow->actions = dp_netdev_actions_create(actions, actions_len);
951 match_init(&match, flow, wc);
952 cls_rule_init(&netdev_flow->cr, &match, NETDEV_RULE_PRIORITY);
953 ovs_rwlock_wrlock(&dp->cls.rwlock);
954 classifier_insert(&dp->cls, &netdev_flow->cr);
955 ovs_rwlock_unlock(&dp->cls.rwlock);
957 hmap_insert(&dp->flow_table, &netdev_flow->node, flow_hash(flow, 0));
962 clear_stats(struct dp_netdev_flow *netdev_flow)
964 netdev_flow->used = 0;
965 netdev_flow->packet_count = 0;
966 netdev_flow->byte_count = 0;
967 netdev_flow->tcp_flags = 0;
971 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
973 struct dp_netdev *dp = get_dp_netdev(dpif);
974 struct dp_netdev_flow *netdev_flow;
976 struct flow_wildcards wc;
979 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &flow);
983 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
984 put->mask, put->mask_len,
990 ovs_mutex_lock(&dp_netdev_mutex);
991 netdev_flow = dp_netdev_lookup_flow(dp, &flow);
993 if (put->flags & DPIF_FP_CREATE) {
994 if (hmap_count(&dp->flow_table) < MAX_FLOWS) {
996 memset(put->stats, 0, sizeof *put->stats);
998 error = dp_netdev_flow_add(dp, &flow, &wc, put->actions,
1007 if (put->flags & DPIF_FP_MODIFY
1008 && flow_equal(&flow, &netdev_flow->flow)) {
1009 dp_netdev_actions_unref(netdev_flow->actions);
1010 netdev_flow->actions = dp_netdev_actions_create(put->actions,
1013 get_dpif_flow_stats(netdev_flow, put->stats);
1015 if (put->flags & DPIF_FP_ZERO_STATS) {
1016 clear_stats(netdev_flow);
1018 } else if (put->flags & DPIF_FP_CREATE) {
1021 /* Overlapping flow. */
1025 ovs_mutex_unlock(&dp_netdev_mutex);
1031 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1033 struct dp_netdev *dp = get_dp_netdev(dpif);
1034 struct dp_netdev_flow *netdev_flow;
1038 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1043 ovs_mutex_lock(&dp_netdev_mutex);
1044 netdev_flow = dp_netdev_find_flow(dp, &key);
1047 get_dpif_flow_stats(netdev_flow, del->stats);
1049 dp_netdev_free_flow(dp, netdev_flow);
1053 ovs_mutex_unlock(&dp_netdev_mutex);
1058 struct dp_netdev_flow_state {
1061 struct dp_netdev_actions *actions;
1062 struct odputil_keybuf keybuf;
1063 struct odputil_keybuf maskbuf;
1064 struct dpif_flow_stats stats;
1068 dpif_netdev_flow_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1070 struct dp_netdev_flow_state *state;
1072 *statep = state = xmalloc(sizeof *state);
1075 state->actions = NULL;
1080 dpif_netdev_flow_dump_next(const struct dpif *dpif, void *state_,
1081 const struct nlattr **key, size_t *key_len,
1082 const struct nlattr **mask, size_t *mask_len,
1083 const struct nlattr **actions, size_t *actions_len,
1084 const struct dpif_flow_stats **stats)
1086 struct dp_netdev_flow_state *state = state_;
1087 struct dp_netdev *dp = get_dp_netdev(dpif);
1088 struct dp_netdev_flow *netdev_flow;
1089 struct hmap_node *node;
1091 ovs_mutex_lock(&dp_netdev_mutex);
1092 node = hmap_at_position(&dp->flow_table, &state->bucket, &state->offset);
1094 ovs_mutex_unlock(&dp_netdev_mutex);
1098 netdev_flow = CONTAINER_OF(node, struct dp_netdev_flow, node);
1103 ofpbuf_use_stack(&buf, &state->keybuf, sizeof state->keybuf);
1104 odp_flow_key_from_flow(&buf, &netdev_flow->flow,
1105 netdev_flow->flow.in_port.odp_port);
1108 *key_len = buf.size;
1113 struct flow_wildcards wc;
1115 ofpbuf_use_stack(&buf, &state->maskbuf, sizeof state->maskbuf);
1116 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1117 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1118 odp_to_u32(wc.masks.in_port.odp_port));
1121 *mask_len = buf.size;
1125 dp_netdev_actions_unref(state->actions);
1126 state->actions = dp_netdev_actions_ref(netdev_flow->actions);
1127 *actions = state->actions->actions;
1128 *actions_len = state->actions->size;
1132 get_dpif_flow_stats(netdev_flow, &state->stats);
1133 *stats = &state->stats;
1136 ovs_mutex_unlock(&dp_netdev_mutex);
1141 dpif_netdev_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1143 struct dp_netdev_flow_state *state = state_;
1145 dp_netdev_actions_unref(state->actions);
1151 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1153 struct dp_netdev *dp = get_dp_netdev(dpif);
1154 struct pkt_metadata *md = &execute->md;
1157 if (execute->packet->size < ETH_HEADER_LEN ||
1158 execute->packet->size > UINT16_MAX) {
1162 /* Extract flow key. */
1163 flow_extract(execute->packet, md->skb_priority, md->pkt_mark, &md->tunnel,
1164 (union flow_in_port *)&md->in_port, &key);
1165 ovs_mutex_lock(&dp_netdev_mutex);
1166 dp_netdev_execute_actions(dp, &key, execute->packet, md, execute->actions,
1167 execute->actions_len);
1168 ovs_mutex_unlock(&dp_netdev_mutex);
1173 dpif_netdev_recv_set(struct dpif *dpif OVS_UNUSED, bool enable OVS_UNUSED)
1179 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1180 uint32_t queue_id, uint32_t *priority)
1182 *priority = queue_id;
1186 static struct dp_netdev_queue *
1187 find_nonempty_queue(struct dpif *dpif)
1189 struct dp_netdev *dp = get_dp_netdev(dpif);
1192 for (i = 0; i < N_QUEUES; i++) {
1193 struct dp_netdev_queue *q = &dp->queues[i];
1194 if (q->head != q->tail) {
1202 dpif_netdev_recv(struct dpif *dpif, struct dpif_upcall *upcall,
1205 struct dp_netdev_queue *q;
1208 ovs_mutex_lock(&dp_netdev_mutex);
1209 q = find_nonempty_queue(dpif);
1211 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
1213 *upcall = u->upcall;
1222 ovs_mutex_unlock(&dp_netdev_mutex);
1228 dpif_netdev_recv_wait(struct dpif *dpif)
1230 struct dp_netdev *dp = get_dp_netdev(dpif);
1233 ovs_mutex_lock(&dp_netdev_mutex);
1234 seq = seq_read(dp->queue_seq);
1235 if (find_nonempty_queue(dpif)) {
1236 poll_immediate_wake();
1238 seq_wait(dp->queue_seq, seq);
1240 ovs_mutex_unlock(&dp_netdev_mutex);
1244 dpif_netdev_recv_purge(struct dpif *dpif)
1246 struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif);
1247 ovs_mutex_lock(&dp_netdev_mutex);
1248 dp_netdev_purge_queues(dpif_netdev->dp);
1249 ovs_mutex_unlock(&dp_netdev_mutex);
1252 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1253 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1255 struct dp_netdev_actions *
1256 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1258 struct dp_netdev_actions *netdev_actions;
1260 netdev_actions = xmalloc(sizeof *netdev_actions);
1261 ovs_refcount_init(&netdev_actions->ref_cnt);
1262 netdev_actions->actions = xmemdup(actions, size);
1263 netdev_actions->size = size;
1265 return netdev_actions;
1268 /* Increments 'actions''s refcount. */
1269 struct dp_netdev_actions *
1270 dp_netdev_actions_ref(const struct dp_netdev_actions *actions_)
1272 struct dp_netdev_actions *actions;
1274 actions = CONST_CAST(struct dp_netdev_actions *, actions_);
1276 ovs_refcount_ref(&actions->ref_cnt);
1281 /* Decrements 'actions''s refcount and frees 'actions' if the refcount reaches
1284 dp_netdev_actions_unref(struct dp_netdev_actions *actions)
1286 if (actions && ovs_refcount_unref(&actions->ref_cnt) == 1) {
1287 free(actions->actions);
1293 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
1294 const struct ofpbuf *packet)
1296 netdev_flow->used = time_msec();
1297 netdev_flow->packet_count++;
1298 netdev_flow->byte_count += packet->size;
1299 netdev_flow->tcp_flags |= packet_get_tcp_flags(packet, &netdev_flow->flow);
1303 dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet,
1304 struct pkt_metadata *md)
1306 struct dp_netdev_flow *netdev_flow;
1309 if (packet->size < ETH_HEADER_LEN) {
1312 flow_extract(packet, md->skb_priority, md->pkt_mark, &md->tunnel,
1313 (union flow_in_port *)&md->in_port, &key);
1314 netdev_flow = dp_netdev_lookup_flow(dp, &key);
1316 struct dp_netdev_actions *actions;
1318 dp_netdev_flow_used(netdev_flow, packet);
1319 actions = dp_netdev_actions_ref(netdev_flow->actions);
1320 dp_netdev_execute_actions(dp, &key, packet, md,
1321 actions->actions, actions->size);
1322 dp_netdev_actions_unref(actions);
1323 ovsthread_counter_inc(dp->n_hit, 1);
1325 ovsthread_counter_inc(dp->n_missed, 1);
1326 dp_netdev_output_userspace(dp, packet, DPIF_UC_MISS, &key, NULL);
1331 dpif_netdev_run(struct dpif *dpif)
1333 struct dp_netdev_port *port;
1334 struct dp_netdev *dp;
1335 struct ofpbuf packet;
1337 ovs_mutex_lock(&dp_netdev_mutex);
1338 dp = get_dp_netdev(dpif);
1339 ofpbuf_init(&packet, 0);
1341 HMAP_FOR_EACH (port, node, &dp->ports) {
1346 error = netdev_get_mtu(port->netdev, &mtu);
1348 mtu = ETH_PAYLOAD_MAX;
1350 buf_size = DP_NETDEV_HEADROOM + VLAN_ETH_HEADER_LEN + mtu;
1352 ofpbuf_clear(&packet);
1353 ofpbuf_reserve_with_tailroom(&packet, DP_NETDEV_HEADROOM, buf_size);
1355 error = port->rx ? netdev_rx_recv(port->rx, &packet) : EOPNOTSUPP;
1357 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port->port_no);
1358 dp_netdev_port_input(dp, &packet, &md);
1359 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1360 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1362 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1363 netdev_get_name(port->netdev), ovs_strerror(error));
1366 ofpbuf_uninit(&packet);
1367 ovs_mutex_unlock(&dp_netdev_mutex);
1371 dpif_netdev_wait(struct dpif *dpif)
1373 struct dp_netdev_port *port;
1375 /* There is a race here, if thread A calls dpif_netdev_wait(dpif) and
1376 * thread B calls dpif_port_add(dpif) or dpif_port_remove(dpif) before
1377 * A makes it to poll_block().
1379 * But I think it doesn't matter:
1381 * - In the dpif_port_add() case, A will not wake up when a packet
1382 * arrives on the new port, but this would also happen if the
1383 * ordering were reversed.
1385 * - In the dpif_port_remove() case, A might wake up spuriously, but
1386 * that is harmless. */
1388 ovs_mutex_lock(&dp_netdev_mutex);
1389 HMAP_FOR_EACH (port, node, &get_dp_netdev(dpif)->ports) {
1391 netdev_rx_wait(port->rx);
1394 ovs_mutex_unlock(&dp_netdev_mutex);
1398 dp_netdev_output_port(struct dp_netdev *dp, struct ofpbuf *packet,
1399 odp_port_t out_port)
1401 struct dp_netdev_port *p = dp_netdev_lookup_port(dp, out_port);
1403 netdev_send(p->netdev, packet);
1408 dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet,
1409 int queue_no, const struct flow *flow,
1410 const struct nlattr *userdata)
1412 struct dp_netdev_queue *q = &dp->queues[queue_no];
1413 if (q->head - q->tail < MAX_QUEUE_LEN) {
1414 struct dp_netdev_upcall *u = &q->upcalls[q->head++ & QUEUE_MASK];
1415 struct dpif_upcall *upcall = &u->upcall;
1416 struct ofpbuf *buf = &u->buf;
1419 upcall->type = queue_no;
1421 /* Allocate buffer big enough for everything. */
1422 buf_size = ODPUTIL_FLOW_KEY_BYTES;
1424 buf_size += NLA_ALIGN(userdata->nla_len);
1426 ofpbuf_init(buf, buf_size);
1429 odp_flow_key_from_flow(buf, flow, flow->in_port.odp_port);
1430 upcall->key = buf->data;
1431 upcall->key_len = buf->size;
1435 upcall->userdata = ofpbuf_put(buf, userdata,
1436 NLA_ALIGN(userdata->nla_len));
1439 /* Steal packet data. */
1440 ovs_assert(packet->source == OFPBUF_MALLOC);
1441 upcall->packet = *packet;
1442 ofpbuf_use(packet, NULL, 0);
1444 seq_change(dp->queue_seq);
1448 ovsthread_counter_inc(dp->n_lost, 1);
1453 struct dp_netdev_execute_aux {
1454 struct dp_netdev *dp;
1455 const struct flow *key;
1459 dp_execute_cb(void *aux_, struct ofpbuf *packet,
1460 const struct pkt_metadata *md OVS_UNUSED,
1461 const struct nlattr *a, bool may_steal)
1463 struct dp_netdev_execute_aux *aux = aux_;
1464 int type = nl_attr_type(a);
1466 switch ((enum ovs_action_attr)type) {
1467 case OVS_ACTION_ATTR_OUTPUT:
1468 dp_netdev_output_port(aux->dp, packet, u32_to_odp(nl_attr_get_u32(a)));
1471 case OVS_ACTION_ATTR_USERSPACE: {
1472 const struct nlattr *userdata;
1474 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
1476 /* Make a copy if we are not allowed to steal the packet's data. */
1478 packet = ofpbuf_clone_with_headroom(packet, DP_NETDEV_HEADROOM);
1480 dp_netdev_output_userspace(aux->dp, packet, DPIF_UC_ACTION, aux->key,
1483 ofpbuf_uninit(packet);
1487 case OVS_ACTION_ATTR_PUSH_VLAN:
1488 case OVS_ACTION_ATTR_POP_VLAN:
1489 case OVS_ACTION_ATTR_PUSH_MPLS:
1490 case OVS_ACTION_ATTR_POP_MPLS:
1491 case OVS_ACTION_ATTR_SET:
1492 case OVS_ACTION_ATTR_SAMPLE:
1493 case OVS_ACTION_ATTR_UNSPEC:
1494 case __OVS_ACTION_ATTR_MAX:
1500 dp_netdev_execute_actions(struct dp_netdev *dp, const struct flow *key,
1501 struct ofpbuf *packet, struct pkt_metadata *md,
1502 const struct nlattr *actions, size_t actions_len)
1504 struct dp_netdev_execute_aux aux = {dp, key};
1506 odp_execute_actions(&aux, packet, md, actions, actions_len, dp_execute_cb);
1509 const struct dpif_class dpif_netdev_class = {
1511 dpif_netdev_enumerate,
1512 dpif_netdev_port_open_type,
1515 dpif_netdev_destroy,
1518 dpif_netdev_get_stats,
1519 dpif_netdev_port_add,
1520 dpif_netdev_port_del,
1521 dpif_netdev_port_query_by_number,
1522 dpif_netdev_port_query_by_name,
1523 NULL, /* port_get_pid */
1524 dpif_netdev_port_dump_start,
1525 dpif_netdev_port_dump_next,
1526 dpif_netdev_port_dump_done,
1527 dpif_netdev_port_poll,
1528 dpif_netdev_port_poll_wait,
1529 dpif_netdev_flow_get,
1530 dpif_netdev_flow_put,
1531 dpif_netdev_flow_del,
1532 dpif_netdev_flow_flush,
1533 dpif_netdev_flow_dump_start,
1534 dpif_netdev_flow_dump_next,
1535 dpif_netdev_flow_dump_done,
1536 dpif_netdev_execute,
1538 dpif_netdev_recv_set,
1539 dpif_netdev_queue_to_priority,
1541 dpif_netdev_recv_wait,
1542 dpif_netdev_recv_purge,
1546 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
1547 const char *argv[], void *aux OVS_UNUSED)
1549 struct dp_netdev_port *port;
1550 struct dp_netdev *dp;
1553 dp = shash_find_data(&dp_netdevs, argv[1]);
1554 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
1555 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
1559 if (get_port_by_name(dp, argv[2], &port)) {
1560 unixctl_command_reply_error(conn, "unknown port");
1564 port_no = u32_to_odp(atoi(argv[3]));
1565 if (!port_no || port_no == ODPP_NONE) {
1566 unixctl_command_reply_error(conn, "bad port number");
1569 if (dp_netdev_lookup_port(dp, port_no)) {
1570 unixctl_command_reply_error(conn, "port number already in use");
1573 hmap_remove(&dp->ports, &port->node);
1574 port->port_no = port_no;
1575 hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0));
1576 seq_change(dp->port_seq);
1577 unixctl_command_reply(conn, NULL);
1581 dpif_dummy_register__(const char *type)
1583 struct dpif_class *class;
1585 class = xmalloc(sizeof *class);
1586 *class = dpif_netdev_class;
1587 class->type = xstrdup(type);
1588 dp_register_provider(class);
1592 dpif_dummy_register(bool override)
1599 dp_enumerate_types(&types);
1600 SSET_FOR_EACH (type, &types) {
1601 if (!dp_unregister_provider(type)) {
1602 dpif_dummy_register__(type);
1605 sset_destroy(&types);
1608 dpif_dummy_register__("dummy");
1610 unixctl_command_register("dpif-dummy/change-port-number",
1611 "DP PORT NEW-NUMBER",
1612 3, 3, dpif_dummy_change_port_number, NULL);