2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
24 #include <netinet/in.h>
25 #include <sys/socket.h>
30 #include <sys/ioctl.h>
34 #include "classifier.h"
38 #include "dpif-provider.h"
40 #include "dynamic-string.h"
45 #include "meta-flow.h"
47 #include "netdev-dpdk.h"
48 #include "netdev-vport.h"
50 #include "odp-execute.h"
52 #include "ofp-print.h"
55 #include "packet-dpif.h"
57 #include "poll-loop.h"
67 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
69 /* By default, choose a priority in the middle. */
70 #define NETDEV_RULE_PRIORITY 0x8000
72 #define FLOW_DUMP_MAX_BATCH 50
74 /* Use per thread recirc_depth to prevent recirculation loop. */
75 #define MAX_RECIRC_DEPTH 5
76 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
78 /* Configuration parameters. */
79 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
82 enum { MAX_QUEUE_LEN = 128 }; /* Maximum number of packets per queue. */
83 enum { QUEUE_MASK = MAX_QUEUE_LEN - 1 };
84 BUILD_ASSERT_DECL(IS_POW2(MAX_QUEUE_LEN));
86 /* Protects against changes to 'dp_netdevs'. */
87 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
89 /* Contains all 'struct dp_netdev's. */
90 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
91 = SHASH_INITIALIZER(&dp_netdevs);
93 struct dp_netdev_upcall {
94 struct dpif_upcall upcall; /* Queued upcall information. */
95 struct ofpbuf buf; /* ofpbuf instance for upcall.packet. */
98 /* A queue passing packets from a struct dp_netdev to its clients (handlers).
104 * Any access at all requires the owning 'dp_netdev''s queue_rwlock and
106 struct dp_netdev_queue {
107 struct ovs_mutex mutex;
108 struct seq *seq; /* Incremented whenever a packet is queued. */
109 struct dp_netdev_upcall upcalls[MAX_QUEUE_LEN] OVS_GUARDED;
110 unsigned int head OVS_GUARDED;
111 unsigned int tail OVS_GUARDED;
114 /* Datapath based on the network device interface from netdev.h.
120 * Some members, marked 'const', are immutable. Accessing other members
121 * requires synchronization, as noted in more detail below.
123 * Acquisition order is, from outermost to innermost:
125 * dp_netdev_mutex (global)
132 const struct dpif_class *const class;
133 const char *const name;
134 struct ovs_refcount ref_cnt;
135 atomic_flag destroyed;
139 * Readers of 'cls' and 'flow_table' must take a 'cls->rwlock' read lock.
141 * Writers of 'cls' and 'flow_table' must take the 'flow_mutex' and then
142 * the 'cls->rwlock' write lock. (The outer 'flow_mutex' allows writers to
143 * atomically perform multiple operations on 'cls' and 'flow_table'.)
145 struct ovs_mutex flow_mutex;
146 struct classifier cls; /* Classifier. Protected by cls.rwlock. */
147 struct hmap flow_table OVS_GUARDED; /* Flow table. */
151 * 'queue_rwlock' protects the modification of 'handler_queues' and
152 * 'n_handlers'. The queue elements are protected by its
153 * 'handler_queues''s mutex. */
154 struct fat_rwlock queue_rwlock;
155 struct dp_netdev_queue *handler_queues;
160 * ovsthread_stats is internally synchronized. */
161 struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */
165 * Protected by RCU. Take the mutex to add or remove ports. */
166 struct ovs_mutex port_mutex;
168 struct seq *port_seq; /* Incremented whenever a port changes. */
170 /* Forwarding threads. */
171 struct latch exit_latch;
172 struct pmd_thread *pmd_threads;
173 size_t n_pmd_threads;
177 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
181 DP_STAT_HIT, /* Packets that matched in the flow table. */
182 DP_STAT_MISS, /* Packets that did not match. */
183 DP_STAT_LOST, /* Packets not passed up to the client. */
187 /* Contained by struct dp_netdev's 'stats' member. */
188 struct dp_netdev_stats {
189 struct ovs_mutex mutex; /* Protects 'n'. */
191 /* Indexed by DP_STAT_*, protected by 'mutex'. */
192 unsigned long long int n[DP_N_STATS] OVS_GUARDED;
196 /* A port in a netdev-based datapath. */
197 struct dp_netdev_port {
198 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
200 struct netdev *netdev;
201 struct netdev_saved_flags *sf;
202 struct netdev_rxq **rxq;
203 struct ovs_refcount ref_cnt;
204 char *type; /* Port type as requested by user. */
208 /* Stores a miniflow */
210 /* There are fields in the flow structure that we never use. Therefore we can
211 * save a few words of memory */
212 #define NETDEV_KEY_BUF_SIZE_U32 (FLOW_U32S \
213 - FLOW_U32_SIZE(regs) \
214 - FLOW_U32_SIZE(metadata) \
216 struct netdev_flow_key {
217 struct miniflow flow;
218 uint32_t buf[NETDEV_KEY_BUF_SIZE_U32];
221 /* A flow in dp_netdev's 'flow_table'.
227 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
228 * its dp_netdev's classifier. The text below calls this classifier 'cls'.
233 * The thread safety rules described here for "struct dp_netdev_flow" are
234 * motivated by two goals:
236 * - Prevent threads that read members of "struct dp_netdev_flow" from
237 * reading bad data due to changes by some thread concurrently modifying
240 * - Prevent two threads making changes to members of a given "struct
241 * dp_netdev_flow" from interfering with each other.
247 * A flow 'flow' may be accessed without a risk of being freed by code that
248 * holds a read-lock or write-lock on 'cls->rwlock' or that owns a reference to
249 * 'flow->ref_cnt' (or both). Code that needs to hold onto a flow for a while
250 * should take 'cls->rwlock', find the flow it needs, increment 'flow->ref_cnt'
251 * with dpif_netdev_flow_ref(), and drop 'cls->rwlock'.
253 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
254 * flow from being deleted from 'cls' (that's 'cls->rwlock') and it doesn't
255 * protect members of 'flow' from modification.
257 * Some members, marked 'const', are immutable. Accessing other members
258 * requires synchronization, as noted in more detail below.
260 struct dp_netdev_flow {
261 /* Packet classification. */
262 const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
264 /* Hash table index by unmasked flow. */
265 const struct hmap_node node; /* In owning dp_netdev's 'flow_table'. */
266 const struct flow flow; /* The flow that created this entry. */
270 * Reading or writing these members requires 'mutex'. */
271 struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
274 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
277 static void dp_netdev_flow_free(struct dp_netdev_flow *);
279 /* Contained by struct dp_netdev_flow's 'stats' member. */
280 struct dp_netdev_flow_stats {
281 struct ovs_mutex mutex; /* Guards all the other members. */
283 long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
284 long long int packet_count OVS_GUARDED; /* Number of packets matched. */
285 long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
286 uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
289 /* A set of datapath actions within a "struct dp_netdev_flow".
295 * A struct dp_netdev_actions 'actions' is protected with RCU. */
296 struct dp_netdev_actions {
297 /* These members are immutable: they do not change during the struct's
299 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
300 unsigned int size; /* Size of 'actions', in bytes. */
303 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
305 struct dp_netdev_actions *dp_netdev_flow_get_actions(
306 const struct dp_netdev_flow *);
307 static void dp_netdev_actions_free(struct dp_netdev_actions *);
309 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
310 * the performance overhead of interrupt processing. Therefore netdev can
311 * not implement rx-wait for these devices. dpif-netdev needs to poll
312 * these device to check for recv buffer. pmd-thread does polling for
313 * devices assigned to itself thread.
315 * DPDK used PMD for accessing NIC.
317 * A thread that receives packets from PMD ports, looks them up in the flow
318 * table, and executes the actions it finds.
321 struct dp_netdev *dp;
324 atomic_uint change_seq;
327 /* Interface to netdev-based datapath. */
330 struct dp_netdev *dp;
331 uint64_t last_port_seq;
334 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
335 struct dp_netdev_port **portp);
336 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
337 struct dp_netdev_port **portp);
338 static void dp_netdev_free(struct dp_netdev *)
339 OVS_REQUIRES(dp_netdev_mutex);
340 static void dp_netdev_flow_flush(struct dp_netdev *);
341 static int do_add_port(struct dp_netdev *dp, const char *devname,
342 const char *type, odp_port_t port_no)
343 OVS_REQUIRES(dp->port_mutex);
344 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
345 OVS_REQUIRES(dp->port_mutex);
346 static void dp_netdev_destroy_all_queues(struct dp_netdev *dp)
347 OVS_REQ_WRLOCK(dp->queue_rwlock);
348 static int dpif_netdev_open(const struct dpif_class *, const char *name,
349 bool create, struct dpif **);
350 static int dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf **,
351 int cnt, int queue_no, int type,
352 const struct miniflow *,
353 const struct nlattr *userdata);
354 static void dp_netdev_execute_actions(struct dp_netdev *dp,
355 struct dpif_packet **, int c,
356 bool may_steal, struct pkt_metadata *,
357 const struct nlattr *actions,
359 static void dp_netdev_port_input(struct dp_netdev *dp,
360 struct dpif_packet **packets, int cnt,
363 static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
365 static struct dpif_netdev *
366 dpif_netdev_cast(const struct dpif *dpif)
368 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
369 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
372 static struct dp_netdev *
373 get_dp_netdev(const struct dpif *dpif)
375 return dpif_netdev_cast(dpif)->dp;
379 dpif_netdev_enumerate(struct sset *all_dps,
380 const struct dpif_class *dpif_class)
382 struct shash_node *node;
384 ovs_mutex_lock(&dp_netdev_mutex);
385 SHASH_FOR_EACH(node, &dp_netdevs) {
386 struct dp_netdev *dp = node->data;
387 if (dpif_class != dp->class) {
388 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
389 * If the class doesn't match, skip this dpif. */
392 sset_add(all_dps, node->name);
394 ovs_mutex_unlock(&dp_netdev_mutex);
400 dpif_netdev_class_is_dummy(const struct dpif_class *class)
402 return class != &dpif_netdev_class;
406 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
408 return strcmp(type, "internal") ? type
409 : dpif_netdev_class_is_dummy(class) ? "dummy"
414 create_dpif_netdev(struct dp_netdev *dp)
416 uint16_t netflow_id = hash_string(dp->name, 0);
417 struct dpif_netdev *dpif;
419 ovs_refcount_ref(&dp->ref_cnt);
421 dpif = xmalloc(sizeof *dpif);
422 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
424 dpif->last_port_seq = seq_read(dp->port_seq);
429 /* Choose an unused, non-zero port number and return it on success.
430 * Return ODPP_NONE on failure. */
432 choose_port(struct dp_netdev *dp, const char *name)
433 OVS_REQUIRES(dp->port_mutex)
437 if (dp->class != &dpif_netdev_class) {
441 /* If the port name begins with "br", start the number search at
442 * 100 to make writing tests easier. */
443 if (!strncmp(name, "br", 2)) {
447 /* If the port name contains a number, try to assign that port number.
448 * This can make writing unit tests easier because port numbers are
450 for (p = name; *p != '\0'; p++) {
451 if (isdigit((unsigned char) *p)) {
452 port_no = start_no + strtol(p, NULL, 10);
453 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
454 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
455 return u32_to_odp(port_no);
462 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
463 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
464 return u32_to_odp(port_no);
472 create_dp_netdev(const char *name, const struct dpif_class *class,
473 struct dp_netdev **dpp)
474 OVS_REQUIRES(dp_netdev_mutex)
476 struct dp_netdev *dp;
479 dp = xzalloc(sizeof *dp);
480 shash_add(&dp_netdevs, name, dp);
482 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
483 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
484 ovs_refcount_init(&dp->ref_cnt);
485 atomic_flag_clear(&dp->destroyed);
487 ovs_mutex_init(&dp->flow_mutex);
488 classifier_init(&dp->cls, NULL);
489 hmap_init(&dp->flow_table);
491 fat_rwlock_init(&dp->queue_rwlock);
493 ovsthread_stats_init(&dp->stats);
495 ovs_mutex_init(&dp->port_mutex);
496 cmap_init(&dp->ports);
497 dp->port_seq = seq_create();
498 latch_init(&dp->exit_latch);
500 ovs_mutex_lock(&dp->port_mutex);
501 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
502 ovs_mutex_unlock(&dp->port_mutex);
513 dpif_netdev_open(const struct dpif_class *class, const char *name,
514 bool create, struct dpif **dpifp)
516 struct dp_netdev *dp;
519 ovs_mutex_lock(&dp_netdev_mutex);
520 dp = shash_find_data(&dp_netdevs, name);
522 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
524 error = (dp->class != class ? EINVAL
529 *dpifp = create_dpif_netdev(dp);
531 ovs_mutex_unlock(&dp_netdev_mutex);
537 dp_netdev_purge_queues(struct dp_netdev *dp)
538 OVS_REQ_WRLOCK(dp->queue_rwlock)
542 for (i = 0; i < dp->n_handlers; i++) {
543 struct dp_netdev_queue *q = &dp->handler_queues[i];
545 ovs_mutex_lock(&q->mutex);
546 while (q->tail != q->head) {
547 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
548 ofpbuf_uninit(&u->upcall.packet);
549 ofpbuf_uninit(&u->buf);
551 ovs_mutex_unlock(&q->mutex);
555 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
556 * through the 'dp_netdevs' shash while freeing 'dp'. */
558 dp_netdev_free(struct dp_netdev *dp)
559 OVS_REQUIRES(dp_netdev_mutex)
561 struct dp_netdev_port *port;
562 struct dp_netdev_stats *bucket;
565 shash_find_and_delete(&dp_netdevs, dp->name);
567 dp_netdev_set_pmd_threads(dp, 0);
568 free(dp->pmd_threads);
570 dp_netdev_flow_flush(dp);
571 ovs_mutex_lock(&dp->port_mutex);
572 CMAP_FOR_EACH (port, node, &dp->ports) {
573 do_del_port(dp, port);
575 ovs_mutex_unlock(&dp->port_mutex);
577 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
578 ovs_mutex_destroy(&bucket->mutex);
579 free_cacheline(bucket);
581 ovsthread_stats_destroy(&dp->stats);
583 fat_rwlock_wrlock(&dp->queue_rwlock);
584 dp_netdev_destroy_all_queues(dp);
585 fat_rwlock_unlock(&dp->queue_rwlock);
587 fat_rwlock_destroy(&dp->queue_rwlock);
589 classifier_destroy(&dp->cls);
590 hmap_destroy(&dp->flow_table);
591 ovs_mutex_destroy(&dp->flow_mutex);
592 seq_destroy(dp->port_seq);
593 cmap_destroy(&dp->ports);
594 latch_destroy(&dp->exit_latch);
595 free(CONST_CAST(char *, dp->name));
600 dp_netdev_unref(struct dp_netdev *dp)
603 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
604 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
605 ovs_mutex_lock(&dp_netdev_mutex);
606 if (ovs_refcount_unref(&dp->ref_cnt) == 1) {
609 ovs_mutex_unlock(&dp_netdev_mutex);
614 dpif_netdev_close(struct dpif *dpif)
616 struct dp_netdev *dp = get_dp_netdev(dpif);
623 dpif_netdev_destroy(struct dpif *dpif)
625 struct dp_netdev *dp = get_dp_netdev(dpif);
627 if (!atomic_flag_test_and_set(&dp->destroyed)) {
628 if (ovs_refcount_unref(&dp->ref_cnt) == 1) {
629 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
638 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
640 struct dp_netdev *dp = get_dp_netdev(dpif);
641 struct dp_netdev_stats *bucket;
644 fat_rwlock_rdlock(&dp->cls.rwlock);
645 stats->n_flows = hmap_count(&dp->flow_table);
646 fat_rwlock_unlock(&dp->cls.rwlock);
648 stats->n_hit = stats->n_missed = stats->n_lost = 0;
649 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
650 ovs_mutex_lock(&bucket->mutex);
651 stats->n_hit += bucket->n[DP_STAT_HIT];
652 stats->n_missed += bucket->n[DP_STAT_MISS];
653 stats->n_lost += bucket->n[DP_STAT_LOST];
654 ovs_mutex_unlock(&bucket->mutex);
656 stats->n_masks = UINT32_MAX;
657 stats->n_mask_hit = UINT64_MAX;
663 dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
667 for (i = 0; i < dp->n_pmd_threads; i++) {
668 struct pmd_thread *f = &dp->pmd_threads[i];
671 atomic_add(&f->change_seq, 1, &id);
676 hash_port_no(odp_port_t port_no)
678 return hash_int(odp_to_u32(port_no), 0);
682 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
684 OVS_REQUIRES(dp->port_mutex)
686 struct netdev_saved_flags *sf;
687 struct dp_netdev_port *port;
688 struct netdev *netdev;
689 enum netdev_flags flags;
690 const char *open_type;
694 /* XXX reject devices already in some dp_netdev. */
696 /* Open and validate network device. */
697 open_type = dpif_netdev_port_open_type(dp->class, type);
698 error = netdev_open(devname, open_type, &netdev);
702 /* XXX reject non-Ethernet devices */
704 netdev_get_flags(netdev, &flags);
705 if (flags & NETDEV_LOOPBACK) {
706 VLOG_ERR("%s: cannot add a loopback device", devname);
707 netdev_close(netdev);
711 port = xzalloc(sizeof *port);
712 port->port_no = port_no;
713 port->netdev = netdev;
714 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
715 port->type = xstrdup(type);
716 for (i = 0; i < netdev_n_rxq(netdev); i++) {
717 error = netdev_rxq_open(netdev, &port->rxq[i], i);
719 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
720 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
721 devname, ovs_strerror(errno));
722 netdev_close(netdev);
727 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
729 for (i = 0; i < netdev_n_rxq(netdev); i++) {
730 netdev_rxq_close(port->rxq[i]);
732 netdev_close(netdev);
739 if (netdev_is_pmd(netdev)) {
741 dp_netdev_set_pmd_threads(dp, NR_THREADS);
742 dp_netdev_reload_pmd_threads(dp);
744 ovs_refcount_init(&port->ref_cnt);
746 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
747 seq_change(dp->port_seq);
753 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
754 odp_port_t *port_nop)
756 struct dp_netdev *dp = get_dp_netdev(dpif);
757 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
758 const char *dpif_port;
762 ovs_mutex_lock(&dp->port_mutex);
763 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
764 if (*port_nop != ODPP_NONE) {
766 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
768 port_no = choose_port(dp, dpif_port);
769 error = port_no == ODPP_NONE ? EFBIG : 0;
773 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
775 ovs_mutex_unlock(&dp->port_mutex);
781 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
783 struct dp_netdev *dp = get_dp_netdev(dpif);
786 ovs_mutex_lock(&dp->port_mutex);
787 if (port_no == ODPP_LOCAL) {
790 struct dp_netdev_port *port;
792 error = get_port_by_number(dp, port_no, &port);
794 do_del_port(dp, port);
797 ovs_mutex_unlock(&dp->port_mutex);
803 is_valid_port_number(odp_port_t port_no)
805 return port_no != ODPP_NONE;
808 static struct dp_netdev_port *
809 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
811 struct dp_netdev_port *port;
813 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
814 if (port->port_no == port_no) {
822 get_port_by_number(struct dp_netdev *dp,
823 odp_port_t port_no, struct dp_netdev_port **portp)
825 if (!is_valid_port_number(port_no)) {
829 *portp = dp_netdev_lookup_port(dp, port_no);
830 return *portp ? 0 : ENOENT;
835 port_ref(struct dp_netdev_port *port)
838 ovs_refcount_ref(&port->ref_cnt);
843 port_destroy__(struct dp_netdev_port *port)
845 int n_rxq = netdev_n_rxq(port->netdev);
848 netdev_close(port->netdev);
849 netdev_restore_flags(port->sf);
851 for (i = 0; i < n_rxq; i++) {
852 netdev_rxq_close(port->rxq[i]);
860 port_unref(struct dp_netdev_port *port)
862 if (port && ovs_refcount_unref(&port->ref_cnt) == 1) {
863 ovsrcu_postpone(port_destroy__, port);
868 get_port_by_name(struct dp_netdev *dp,
869 const char *devname, struct dp_netdev_port **portp)
870 OVS_REQUIRES(dp->port_mutex)
872 struct dp_netdev_port *port;
874 CMAP_FOR_EACH (port, node, &dp->ports) {
875 if (!strcmp(netdev_get_name(port->netdev), devname)) {
884 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
885 OVS_REQUIRES(dp->port_mutex)
887 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
888 seq_change(dp->port_seq);
889 if (netdev_is_pmd(port->netdev)) {
890 dp_netdev_reload_pmd_threads(dp);
897 answer_port_query(const struct dp_netdev_port *port,
898 struct dpif_port *dpif_port)
900 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
901 dpif_port->type = xstrdup(port->type);
902 dpif_port->port_no = port->port_no;
906 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
907 struct dpif_port *dpif_port)
909 struct dp_netdev *dp = get_dp_netdev(dpif);
910 struct dp_netdev_port *port;
913 error = get_port_by_number(dp, port_no, &port);
914 if (!error && dpif_port) {
915 answer_port_query(port, dpif_port);
922 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
923 struct dpif_port *dpif_port)
925 struct dp_netdev *dp = get_dp_netdev(dpif);
926 struct dp_netdev_port *port;
929 ovs_mutex_lock(&dp->port_mutex);
930 error = get_port_by_name(dp, devname, &port);
931 if (!error && dpif_port) {
932 answer_port_query(port, dpif_port);
934 ovs_mutex_unlock(&dp->port_mutex);
940 dp_netdev_flow_free(struct dp_netdev_flow *flow)
942 struct dp_netdev_flow_stats *bucket;
945 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
946 ovs_mutex_destroy(&bucket->mutex);
947 free_cacheline(bucket);
949 ovsthread_stats_destroy(&flow->stats);
951 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
952 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
957 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
958 OVS_REQ_WRLOCK(dp->cls.rwlock)
959 OVS_REQUIRES(dp->flow_mutex)
961 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
962 struct hmap_node *node = CONST_CAST(struct hmap_node *, &flow->node);
964 classifier_remove(&dp->cls, cr);
965 hmap_remove(&dp->flow_table, node);
966 ovsrcu_postpone(dp_netdev_flow_free, flow);
970 dp_netdev_flow_flush(struct dp_netdev *dp)
972 struct dp_netdev_flow *netdev_flow, *next;
974 ovs_mutex_lock(&dp->flow_mutex);
975 fat_rwlock_wrlock(&dp->cls.rwlock);
976 HMAP_FOR_EACH_SAFE (netdev_flow, next, node, &dp->flow_table) {
977 dp_netdev_remove_flow(dp, netdev_flow);
979 fat_rwlock_unlock(&dp->cls.rwlock);
980 ovs_mutex_unlock(&dp->flow_mutex);
984 dpif_netdev_flow_flush(struct dpif *dpif)
986 struct dp_netdev *dp = get_dp_netdev(dpif);
988 dp_netdev_flow_flush(dp);
992 struct dp_netdev_port_state {
993 struct cmap_position position;
998 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1000 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1005 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1006 struct dpif_port *dpif_port)
1008 struct dp_netdev_port_state *state = state_;
1009 struct dp_netdev *dp = get_dp_netdev(dpif);
1010 struct cmap_node *node;
1013 node = cmap_next_position(&dp->ports, &state->position);
1015 struct dp_netdev_port *port;
1017 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1020 state->name = xstrdup(netdev_get_name(port->netdev));
1021 dpif_port->name = state->name;
1022 dpif_port->type = port->type;
1023 dpif_port->port_no = port->port_no;
1034 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1036 struct dp_netdev_port_state *state = state_;
1043 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1045 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1046 uint64_t new_port_seq;
1049 new_port_seq = seq_read(dpif->dp->port_seq);
1050 if (dpif->last_port_seq != new_port_seq) {
1051 dpif->last_port_seq = new_port_seq;
1061 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1063 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1065 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1068 static struct dp_netdev_flow *
1069 dp_netdev_flow_cast(const struct cls_rule *cr)
1071 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1074 static struct dp_netdev_flow *
1075 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
1076 OVS_REQ_RDLOCK(dp->cls.rwlock)
1078 struct dp_netdev_flow *netdev_flow;
1079 struct cls_rule *rule;
1081 rule = classifier_lookup_miniflow_first(&dp->cls, key);
1082 netdev_flow = dp_netdev_flow_cast(rule);
1087 static struct dp_netdev_flow *
1088 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1089 OVS_REQ_RDLOCK(dp->cls.rwlock)
1091 struct dp_netdev_flow *netdev_flow;
1093 HMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1095 if (flow_equal(&netdev_flow->flow, flow)) {
1104 get_dpif_flow_stats(struct dp_netdev_flow *netdev_flow,
1105 struct dpif_flow_stats *stats)
1107 struct dp_netdev_flow_stats *bucket;
1110 memset(stats, 0, sizeof *stats);
1111 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1112 ovs_mutex_lock(&bucket->mutex);
1113 stats->n_packets += bucket->packet_count;
1114 stats->n_bytes += bucket->byte_count;
1115 stats->used = MAX(stats->used, bucket->used);
1116 stats->tcp_flags |= bucket->tcp_flags;
1117 ovs_mutex_unlock(&bucket->mutex);
1122 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1123 const struct nlattr *mask_key,
1124 uint32_t mask_key_len, const struct flow *flow,
1128 enum odp_key_fitness fitness;
1130 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1132 /* This should not happen: it indicates that
1133 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1134 * disagree on the acceptable form of a mask. Log the problem
1135 * as an error, with enough details to enable debugging. */
1136 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1138 if (!VLOG_DROP_ERR(&rl)) {
1142 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1144 VLOG_ERR("internal error parsing flow mask %s (%s)",
1145 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1152 enum mf_field_id id;
1153 /* No mask key, unwildcard everything except fields whose
1154 * prerequisities are not met. */
1155 memset(mask, 0x0, sizeof *mask);
1157 for (id = 0; id < MFF_N_IDS; ++id) {
1158 /* Skip registers and metadata. */
1159 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1160 && id != MFF_METADATA) {
1161 const struct mf_field *mf = mf_from_id(id);
1162 if (mf_are_prereqs_ok(mf, flow)) {
1163 mf_mask_field(mf, mask);
1169 /* Force unwildcard the in_port.
1171 * We need to do this even in the case where we unwildcard "everything"
1172 * above because "everything" only includes the 16-bit OpenFlow port number
1173 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1174 * port number mask->in_port.odp_port. */
1175 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1181 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1186 if (odp_flow_key_to_flow(key, key_len, flow)) {
1187 /* This should not happen: it indicates that odp_flow_key_from_flow()
1188 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1189 * flow. Log the problem as an error, with enough details to enable
1191 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1193 if (!VLOG_DROP_ERR(&rl)) {
1197 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1198 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1205 in_port = flow->in_port.odp_port;
1206 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1214 dpif_netdev_flow_get(const struct dpif *dpif,
1215 const struct nlattr *nl_key, size_t nl_key_len,
1216 struct ofpbuf **actionsp, struct dpif_flow_stats *stats)
1218 struct dp_netdev *dp = get_dp_netdev(dpif);
1219 struct dp_netdev_flow *netdev_flow;
1223 error = dpif_netdev_flow_from_nlattrs(nl_key, nl_key_len, &key);
1228 fat_rwlock_rdlock(&dp->cls.rwlock);
1229 netdev_flow = dp_netdev_find_flow(dp, &key);
1230 fat_rwlock_unlock(&dp->cls.rwlock);
1234 get_dpif_flow_stats(netdev_flow, stats);
1238 struct dp_netdev_actions *actions;
1240 actions = dp_netdev_flow_get_actions(netdev_flow);
1241 *actionsp = ofpbuf_clone_data(actions->actions, actions->size);
1251 dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow,
1252 const struct flow_wildcards *wc,
1253 const struct nlattr *actions,
1255 OVS_REQUIRES(dp->flow_mutex)
1257 struct dp_netdev_flow *netdev_flow;
1260 netdev_flow = xzalloc(sizeof *netdev_flow);
1261 *CONST_CAST(struct flow *, &netdev_flow->flow) = *flow;
1263 ovsthread_stats_init(&netdev_flow->stats);
1265 ovsrcu_set(&netdev_flow->actions,
1266 dp_netdev_actions_create(actions, actions_len));
1268 match_init(&match, flow, wc);
1269 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1270 &match, NETDEV_RULE_PRIORITY);
1271 fat_rwlock_wrlock(&dp->cls.rwlock);
1272 classifier_insert(&dp->cls,
1273 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1274 hmap_insert(&dp->flow_table,
1275 CONST_CAST(struct hmap_node *, &netdev_flow->node),
1276 flow_hash(flow, 0));
1277 fat_rwlock_unlock(&dp->cls.rwlock);
1283 clear_stats(struct dp_netdev_flow *netdev_flow)
1285 struct dp_netdev_flow_stats *bucket;
1288 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1289 ovs_mutex_lock(&bucket->mutex);
1291 bucket->packet_count = 0;
1292 bucket->byte_count = 0;
1293 bucket->tcp_flags = 0;
1294 ovs_mutex_unlock(&bucket->mutex);
1299 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1301 struct dp_netdev *dp = get_dp_netdev(dpif);
1302 struct dp_netdev_flow *netdev_flow;
1304 struct miniflow miniflow;
1305 struct flow_wildcards wc;
1308 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &flow);
1312 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1313 put->mask, put->mask_len,
1318 miniflow_init(&miniflow, &flow);
1320 ovs_mutex_lock(&dp->flow_mutex);
1321 fat_rwlock_rdlock(&dp->cls.rwlock);
1322 netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
1323 fat_rwlock_unlock(&dp->cls.rwlock);
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 dpif_netdev_flow_dump {
1402 struct dpif_flow_dump up;
1406 struct ovs_mutex mutex;
1409 static struct dpif_netdev_flow_dump *
1410 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1412 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1415 static struct dpif_flow_dump *
1416 dpif_netdev_flow_dump_create(const struct dpif *dpif_)
1418 struct dpif_netdev_flow_dump *dump;
1420 dump = xmalloc(sizeof *dump);
1421 dpif_flow_dump_init(&dump->up, dpif_);
1425 ovs_mutex_init(&dump->mutex);
1431 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1433 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1435 ovs_mutex_destroy(&dump->mutex);
1440 struct dpif_netdev_flow_dump_thread {
1441 struct dpif_flow_dump_thread up;
1442 struct dpif_netdev_flow_dump *dump;
1443 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1444 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1447 static struct dpif_netdev_flow_dump_thread *
1448 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1450 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1453 static struct dpif_flow_dump_thread *
1454 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1456 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1457 struct dpif_netdev_flow_dump_thread *thread;
1459 thread = xmalloc(sizeof *thread);
1460 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1461 thread->dump = dump;
1466 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1468 struct dpif_netdev_flow_dump_thread *thread
1469 = dpif_netdev_flow_dump_thread_cast(thread_);
1474 /* XXX the caller must use 'actions' without quiescing */
1476 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
1477 struct dpif_flow *flows, int max_flows)
1479 struct dpif_netdev_flow_dump_thread *thread
1480 = dpif_netdev_flow_dump_thread_cast(thread_);
1481 struct dpif_netdev_flow_dump *dump = thread->dump;
1482 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
1483 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
1484 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
1488 ovs_mutex_lock(&dump->mutex);
1489 if (!dump->status) {
1490 fat_rwlock_rdlock(&dp->cls.rwlock);
1491 for (n_flows = 0; n_flows < MIN(max_flows, FLOW_DUMP_MAX_BATCH);
1493 struct hmap_node *node;
1495 node = hmap_at_position(&dp->flow_table, &dump->bucket,
1501 netdev_flows[n_flows] = CONTAINER_OF(node, struct dp_netdev_flow,
1504 fat_rwlock_unlock(&dp->cls.rwlock);
1506 ovs_mutex_unlock(&dump->mutex);
1508 for (i = 0; i < n_flows; i++) {
1509 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
1510 struct odputil_keybuf *keybuf = &thread->keybuf[i];
1511 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
1512 struct dpif_flow *f = &flows[i];
1513 struct dp_netdev_actions *dp_actions;
1514 struct flow_wildcards wc;
1517 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1520 ofpbuf_use_stack(&buf, keybuf, sizeof *keybuf);
1521 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1522 netdev_flow->flow.in_port.odp_port, true);
1523 f->key = ofpbuf_data(&buf);
1524 f->key_len = ofpbuf_size(&buf);
1527 ofpbuf_use_stack(&buf, maskbuf, sizeof *maskbuf);
1528 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1529 odp_to_u32(wc.masks.in_port.odp_port),
1531 f->mask = ofpbuf_data(&buf);
1532 f->mask_len = ofpbuf_size(&buf);
1535 dp_actions = dp_netdev_flow_get_actions(netdev_flow);
1536 f->actions = dp_actions->actions;
1537 f->actions_len = dp_actions->size;
1540 get_dpif_flow_stats(netdev_flow, &f->stats);
1547 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1549 struct dp_netdev *dp = get_dp_netdev(dpif);
1550 struct dpif_packet packet, *pp;
1551 struct pkt_metadata *md = &execute->md;
1553 if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
1554 ofpbuf_size(execute->packet) > UINT16_MAX) {
1558 packet.ofpbuf = *execute->packet;
1561 dp_netdev_execute_actions(dp, &pp, 1, false, md,
1562 execute->actions, execute->actions_len);
1564 /* Even though may_steal is set to false, some actions could modify or
1565 * reallocate the ofpbuf memory. We need to pass those changes to the
1567 *execute->packet = packet.ofpbuf;
1573 dp_netdev_destroy_all_queues(struct dp_netdev *dp)
1574 OVS_REQ_WRLOCK(dp->queue_rwlock)
1578 dp_netdev_purge_queues(dp);
1580 for (i = 0; i < dp->n_handlers; i++) {
1581 struct dp_netdev_queue *q = &dp->handler_queues[i];
1583 ovs_mutex_destroy(&q->mutex);
1584 seq_destroy(q->seq);
1586 free(dp->handler_queues);
1587 dp->handler_queues = NULL;
1592 dp_netdev_refresh_queues(struct dp_netdev *dp, uint32_t n_handlers)
1593 OVS_REQ_WRLOCK(dp->queue_rwlock)
1595 if (dp->n_handlers != n_handlers) {
1598 dp_netdev_destroy_all_queues(dp);
1600 dp->n_handlers = n_handlers;
1601 dp->handler_queues = xzalloc(n_handlers * sizeof *dp->handler_queues);
1603 for (i = 0; i < n_handlers; i++) {
1604 struct dp_netdev_queue *q = &dp->handler_queues[i];
1606 ovs_mutex_init(&q->mutex);
1607 q->seq = seq_create();
1613 dpif_netdev_recv_set(struct dpif *dpif, bool enable)
1615 struct dp_netdev *dp = get_dp_netdev(dpif);
1617 if ((dp->handler_queues != NULL) == enable) {
1621 fat_rwlock_wrlock(&dp->queue_rwlock);
1623 dp_netdev_destroy_all_queues(dp);
1625 dp_netdev_refresh_queues(dp, 1);
1627 fat_rwlock_unlock(&dp->queue_rwlock);
1633 dpif_netdev_handlers_set(struct dpif *dpif, uint32_t n_handlers)
1635 struct dp_netdev *dp = get_dp_netdev(dpif);
1637 fat_rwlock_wrlock(&dp->queue_rwlock);
1638 if (dp->handler_queues) {
1639 dp_netdev_refresh_queues(dp, n_handlers);
1641 fat_rwlock_unlock(&dp->queue_rwlock);
1647 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1648 uint32_t queue_id, uint32_t *priority)
1650 *priority = queue_id;
1655 dp_netdev_recv_check(const struct dp_netdev *dp, const uint32_t handler_id)
1656 OVS_REQ_RDLOCK(dp->queue_rwlock)
1658 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1660 if (!dp->handler_queues) {
1661 VLOG_WARN_RL(&rl, "receiving upcall disabled");
1665 if (handler_id >= dp->n_handlers) {
1666 VLOG_WARN_RL(&rl, "handler index out of bound");
1674 dpif_netdev_recv(struct dpif *dpif, uint32_t handler_id,
1675 struct dpif_upcall *upcall, struct ofpbuf *buf)
1677 struct dp_netdev *dp = get_dp_netdev(dpif);
1678 struct dp_netdev_queue *q;
1681 fat_rwlock_rdlock(&dp->queue_rwlock);
1683 if (!dp_netdev_recv_check(dp, handler_id)) {
1688 q = &dp->handler_queues[handler_id];
1689 ovs_mutex_lock(&q->mutex);
1690 if (q->head != q->tail) {
1691 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
1693 *upcall = u->upcall;
1700 ovs_mutex_unlock(&q->mutex);
1703 fat_rwlock_unlock(&dp->queue_rwlock);
1709 dpif_netdev_recv_wait(struct dpif *dpif, uint32_t handler_id)
1711 struct dp_netdev *dp = get_dp_netdev(dpif);
1712 struct dp_netdev_queue *q;
1715 fat_rwlock_rdlock(&dp->queue_rwlock);
1717 if (!dp_netdev_recv_check(dp, handler_id)) {
1721 q = &dp->handler_queues[handler_id];
1722 ovs_mutex_lock(&q->mutex);
1723 seq = seq_read(q->seq);
1724 if (q->head != q->tail) {
1725 poll_immediate_wake();
1727 seq_wait(q->seq, seq);
1730 ovs_mutex_unlock(&q->mutex);
1733 fat_rwlock_unlock(&dp->queue_rwlock);
1737 dpif_netdev_recv_purge(struct dpif *dpif)
1739 struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif);
1741 fat_rwlock_wrlock(&dpif_netdev->dp->queue_rwlock);
1742 dp_netdev_purge_queues(dpif_netdev->dp);
1743 fat_rwlock_unlock(&dpif_netdev->dp->queue_rwlock);
1746 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1747 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1749 struct dp_netdev_actions *
1750 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1752 struct dp_netdev_actions *netdev_actions;
1754 netdev_actions = xmalloc(sizeof *netdev_actions);
1755 netdev_actions->actions = xmemdup(actions, size);
1756 netdev_actions->size = size;
1758 return netdev_actions;
1761 struct dp_netdev_actions *
1762 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
1764 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
1768 dp_netdev_actions_free(struct dp_netdev_actions *actions)
1770 free(actions->actions);
1776 dp_netdev_process_rxq_port(struct dp_netdev *dp,
1777 struct dp_netdev_port *port,
1778 struct netdev_rxq *rxq)
1780 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1783 error = netdev_rxq_recv(rxq, packets, &cnt);
1785 dp_netdev_port_input(dp, packets, cnt, port->port_no);
1786 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1787 static struct vlog_rate_limit rl
1788 = VLOG_RATE_LIMIT_INIT(1, 5);
1790 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1791 netdev_get_name(port->netdev),
1792 ovs_strerror(error));
1797 dpif_netdev_run(struct dpif *dpif)
1799 struct dp_netdev_port *port;
1800 struct dp_netdev *dp = get_dp_netdev(dpif);
1802 CMAP_FOR_EACH (port, node, &dp->ports) {
1803 if (!netdev_is_pmd(port->netdev)) {
1806 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1807 dp_netdev_process_rxq_port(dp, port, port->rxq[i]);
1814 dpif_netdev_wait(struct dpif *dpif)
1816 struct dp_netdev_port *port;
1817 struct dp_netdev *dp = get_dp_netdev(dpif);
1819 ovs_mutex_lock(&dp_netdev_mutex);
1820 CMAP_FOR_EACH (port, node, &dp->ports) {
1821 if (!netdev_is_pmd(port->netdev)) {
1824 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1825 netdev_rxq_wait(port->rxq[i]);
1829 ovs_mutex_unlock(&dp_netdev_mutex);
1833 struct dp_netdev_port *port;
1834 struct netdev_rxq *rx;
1838 pmd_load_queues(struct pmd_thread *f,
1839 struct rxq_poll **ppoll_list, int poll_cnt)
1841 struct dp_netdev *dp = f->dp;
1842 struct rxq_poll *poll_list = *ppoll_list;
1843 struct dp_netdev_port *port;
1848 /* Simple scheduler for netdev rx polling. */
1849 for (i = 0; i < poll_cnt; i++) {
1850 port_unref(poll_list[i].port);
1856 CMAP_FOR_EACH (port, node, &f->dp->ports) {
1857 if (netdev_is_pmd(port->netdev)) {
1860 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1861 if ((index % dp->n_pmd_threads) == id) {
1862 poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1));
1865 poll_list[poll_cnt].port = port;
1866 poll_list[poll_cnt].rx = port->rxq[i];
1874 *ppoll_list = poll_list;
1879 pmd_thread_main(void *f_)
1881 struct pmd_thread *f = f_;
1882 struct dp_netdev *dp = f->dp;
1883 unsigned int lc = 0;
1884 struct rxq_poll *poll_list;
1885 unsigned int port_seq;
1892 pmd_thread_setaffinity_cpu(f->id);
1894 poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
1895 atomic_read(&f->change_seq, &port_seq);
1898 unsigned int c_port_seq;
1901 for (i = 0; i < poll_cnt; i++) {
1902 dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx);
1908 /* TODO: need completely userspace based signaling method.
1909 * to keep this thread entirely in userspace.
1910 * For now using atomic counter. */
1912 atomic_read_explicit(&f->change_seq, &c_port_seq, memory_order_consume);
1913 if (c_port_seq != port_seq) {
1919 if (!latch_is_set(&f->dp->exit_latch)){
1923 for (i = 0; i < poll_cnt; i++) {
1924 port_unref(poll_list[i].port);
1932 dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n)
1936 if (n == dp->n_pmd_threads) {
1940 /* Stop existing threads. */
1941 latch_set(&dp->exit_latch);
1942 dp_netdev_reload_pmd_threads(dp);
1943 for (i = 0; i < dp->n_pmd_threads; i++) {
1944 struct pmd_thread *f = &dp->pmd_threads[i];
1946 xpthread_join(f->thread, NULL);
1948 latch_poll(&dp->exit_latch);
1949 free(dp->pmd_threads);
1951 /* Start new threads. */
1952 dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
1953 dp->n_pmd_threads = n;
1955 for (i = 0; i < n; i++) {
1956 struct pmd_thread *f = &dp->pmd_threads[i];
1960 atomic_store(&f->change_seq, 1);
1962 /* Each thread will distribute all devices rx-queues among
1964 f->thread = ovs_thread_create("pmd", pmd_thread_main, f);
1970 dp_netdev_flow_stats_new_cb(void)
1972 struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1973 ovs_mutex_init(&bucket->mutex);
1978 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
1982 long long int now = time_msec();
1983 struct dp_netdev_flow_stats *bucket;
1985 bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
1986 dp_netdev_flow_stats_new_cb);
1988 ovs_mutex_lock(&bucket->mutex);
1989 bucket->used = MAX(now, bucket->used);
1990 bucket->packet_count += cnt;
1991 bucket->byte_count += size;
1992 bucket->tcp_flags |= tcp_flags;
1993 ovs_mutex_unlock(&bucket->mutex);
1997 dp_netdev_stats_new_cb(void)
1999 struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
2000 ovs_mutex_init(&bucket->mutex);
2005 dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type, int cnt)
2007 struct dp_netdev_stats *bucket;
2009 bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
2010 ovs_mutex_lock(&bucket->mutex);
2011 bucket->n[type] += cnt;
2012 ovs_mutex_unlock(&bucket->mutex);
2015 struct packet_batch {
2016 unsigned int packet_count;
2017 unsigned int byte_count;
2020 struct dp_netdev_flow *flow;
2022 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
2023 struct pkt_metadata md;
2027 packet_batch_update(struct packet_batch *batch,
2028 struct dpif_packet *packet, const struct miniflow *mf)
2030 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
2031 batch->packets[batch->packet_count++] = packet;
2032 batch->byte_count += ofpbuf_size(&packet->ofpbuf);
2036 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow,
2037 struct dpif_packet *packet, struct pkt_metadata *md,
2038 const struct miniflow *mf)
2042 batch->packets[0] = packet;
2044 batch->packet_count = 0;
2045 batch->byte_count = 0;
2046 batch->tcp_flags = 0;
2048 packet_batch_update(batch, packet, mf);
2052 packet_batch_execute(struct packet_batch *batch, struct dp_netdev *dp)
2054 struct dp_netdev_actions *actions;
2055 struct dp_netdev_flow *flow = batch->flow;
2057 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
2060 actions = dp_netdev_flow_get_actions(flow);
2062 dp_netdev_execute_actions(dp, batch->packets,
2063 batch->packet_count, true, &batch->md,
2064 actions->actions, actions->size);
2066 dp_netdev_count_packet(dp, DP_STAT_HIT, batch->packet_count);
2070 dp_netdev_input(struct dp_netdev *dp, struct dpif_packet **packets, int cnt,
2071 struct pkt_metadata *md)
2073 struct packet_batch batch;
2075 struct netdev_flow_key key;
2081 miniflow_initialize(&key.flow, key.buf);
2083 fat_rwlock_rdlock(&dp->cls.rwlock);
2084 for (i = 0; i < cnt; i++) {
2085 struct dp_netdev_flow *netdev_flow;
2086 struct ofpbuf *buf = &packets[i]->ofpbuf;
2088 if (ofpbuf_size(buf) < ETH_HEADER_LEN) {
2089 dpif_packet_delete(packets[i]);
2093 miniflow_extract(buf, md, &key.flow);
2095 netdev_flow = dp_netdev_lookup_flow(dp, &key.flow);
2099 packet_batch_init(&batch, netdev_flow, packets[i], md,
2101 } else if (batch.flow == netdev_flow) {
2102 packet_batch_update(&batch, packets[i], &key.flow);
2104 packet_batch_execute(&batch, dp);
2105 packet_batch_init(&batch, netdev_flow, packets[i], md,
2109 /* Packet's flow not in datapath */
2110 dp_netdev_count_packet(dp, DP_STAT_MISS, 1);
2112 if (dp->handler_queues) {
2114 dp_netdev_output_userspace(dp, &buf, 1,
2115 miniflow_hash_5tuple(&key.flow, 0)
2117 DPIF_UC_MISS, &key.flow, NULL);
2119 /* No upcall queue. Freeing the packet */
2120 dpif_packet_delete(packets[i]);
2124 fat_rwlock_unlock(&dp->cls.rwlock);
2127 packet_batch_execute(&batch, dp);
2132 dp_netdev_port_input(struct dp_netdev *dp, struct dpif_packet **packets,
2133 int cnt, odp_port_t port_no)
2135 uint32_t *recirc_depth = recirc_depth_get();
2136 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port_no);
2139 dp_netdev_input(dp, packets, cnt, &md);
2143 dp_netdev_queue_userspace_packet(struct dp_netdev_queue *q,
2144 struct ofpbuf *packet, int type,
2145 const struct miniflow *key,
2146 const struct nlattr *userdata)
2147 OVS_REQUIRES(q->mutex)
2149 if (q->head - q->tail < MAX_QUEUE_LEN) {
2150 struct dp_netdev_upcall *u = &q->upcalls[q->head++ & QUEUE_MASK];
2151 struct dpif_upcall *upcall = &u->upcall;
2152 struct ofpbuf *buf = &u->buf;
2156 upcall->type = type;
2158 /* Allocate buffer big enough for everything. */
2159 buf_size = ODPUTIL_FLOW_KEY_BYTES;
2161 buf_size += NLA_ALIGN(userdata->nla_len);
2163 ofpbuf_init(buf, buf_size);
2166 miniflow_expand(key, &flow);
2167 odp_flow_key_from_flow(buf, &flow, NULL, flow.in_port.odp_port, true);
2168 upcall->key = ofpbuf_data(buf);
2169 upcall->key_len = ofpbuf_size(buf);
2173 upcall->userdata = ofpbuf_put(buf, userdata,
2174 NLA_ALIGN(userdata->nla_len));
2177 upcall->packet = *packet;
2183 ofpbuf_delete(packet);
2190 dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf **packets,
2191 int cnt, int queue_no, int type,
2192 const struct miniflow *key,
2193 const struct nlattr *userdata)
2195 struct dp_netdev_queue *q;
2199 fat_rwlock_rdlock(&dp->queue_rwlock);
2200 q = &dp->handler_queues[queue_no];
2201 ovs_mutex_lock(&q->mutex);
2202 for (i = 0; i < cnt; i++) {
2203 struct ofpbuf *packet = packets[i];
2205 error = dp_netdev_queue_userspace_packet(q, packet, type, key,
2207 if (error == ENOBUFS) {
2208 dp_netdev_count_packet(dp, DP_STAT_LOST, 1);
2211 ovs_mutex_unlock(&q->mutex);
2212 fat_rwlock_unlock(&dp->queue_rwlock);
2217 struct dp_netdev_execute_aux {
2218 struct dp_netdev *dp;
2222 dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
2223 struct pkt_metadata *md,
2224 const struct nlattr *a, bool may_steal)
2225 OVS_NO_THREAD_SAFETY_ANALYSIS
2227 struct dp_netdev_execute_aux *aux = aux_;
2228 int type = nl_attr_type(a);
2229 struct dp_netdev_port *p;
2230 uint32_t *depth = recirc_depth_get();
2233 switch ((enum ovs_action_attr)type) {
2234 case OVS_ACTION_ATTR_OUTPUT:
2235 p = dp_netdev_lookup_port(aux->dp, u32_to_odp(nl_attr_get_u32(a)));
2236 if (OVS_LIKELY(p)) {
2237 netdev_send(p->netdev, packets, cnt, may_steal);
2238 } else if (may_steal) {
2239 for (i = 0; i < cnt; i++) {
2240 dpif_packet_delete(packets[i]);
2245 case OVS_ACTION_ATTR_USERSPACE: {
2246 const struct nlattr *userdata;
2247 struct netdev_flow_key key;
2249 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2251 miniflow_initialize(&key.flow, key.buf);
2253 for (i = 0; i < cnt; i++) {
2254 struct ofpbuf *packet, *userspace_packet;
2256 packet = &packets[i]->ofpbuf;
2258 miniflow_extract(packet, md, &key.flow);
2260 userspace_packet = may_steal ? packet : ofpbuf_clone(packet);
2262 dp_netdev_output_userspace(aux->dp, &userspace_packet, 1,
2263 miniflow_hash_5tuple(&key.flow, 0)
2264 % aux->dp->n_handlers,
2265 DPIF_UC_ACTION, &key.flow,
2271 case OVS_ACTION_ATTR_HASH: {
2272 const struct ovs_action_hash *hash_act;
2273 struct netdev_flow_key key;
2276 hash_act = nl_attr_get(a);
2278 miniflow_initialize(&key.flow, key.buf);
2280 for (i = 0; i < cnt; i++) {
2282 /* TODO: this is slow. Use RSS hash in the future */
2283 miniflow_extract(&packets[i]->ofpbuf, md, &key.flow);
2285 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
2286 /* Hash need not be symmetric, nor does it need to include
2288 hash = miniflow_hash_5tuple(&key.flow, hash_act->hash_basis);
2290 VLOG_WARN("Unknown hash algorithm specified "
2291 "for the hash action.");
2296 hash = 1; /* 0 is not valid */
2302 packets[i]->dp_hash = hash;
2307 case OVS_ACTION_ATTR_RECIRC:
2308 if (*depth < MAX_RECIRC_DEPTH) {
2311 for (i = 0; i < cnt; i++) {
2312 struct dpif_packet *recirc_pkt;
2313 struct pkt_metadata recirc_md = *md;
2315 recirc_pkt = (may_steal) ? packets[i]
2316 : dpif_packet_clone(packets[i]);
2318 recirc_md.recirc_id = nl_attr_get_u32(a);
2320 /* Hash is private to each packet */
2321 recirc_md.dp_hash = packets[i]->dp_hash;
2323 dp_netdev_input(aux->dp, &recirc_pkt, 1, &recirc_md);
2329 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
2331 for (i = 0; i < cnt; i++) {
2332 dpif_packet_delete(packets[i]);
2338 case OVS_ACTION_ATTR_PUSH_VLAN:
2339 case OVS_ACTION_ATTR_POP_VLAN:
2340 case OVS_ACTION_ATTR_PUSH_MPLS:
2341 case OVS_ACTION_ATTR_POP_MPLS:
2342 case OVS_ACTION_ATTR_SET:
2343 case OVS_ACTION_ATTR_SAMPLE:
2344 case OVS_ACTION_ATTR_UNSPEC:
2345 case __OVS_ACTION_ATTR_MAX:
2351 dp_netdev_execute_actions(struct dp_netdev *dp,
2352 struct dpif_packet **packets, int cnt,
2353 bool may_steal, struct pkt_metadata *md,
2354 const struct nlattr *actions, size_t actions_len)
2356 struct dp_netdev_execute_aux aux = {dp};
2358 odp_execute_actions(&aux, packets, cnt, may_steal, md, actions,
2359 actions_len, dp_execute_cb);
2362 const struct dpif_class dpif_netdev_class = {
2364 dpif_netdev_enumerate,
2365 dpif_netdev_port_open_type,
2368 dpif_netdev_destroy,
2371 dpif_netdev_get_stats,
2372 dpif_netdev_port_add,
2373 dpif_netdev_port_del,
2374 dpif_netdev_port_query_by_number,
2375 dpif_netdev_port_query_by_name,
2376 NULL, /* port_get_pid */
2377 dpif_netdev_port_dump_start,
2378 dpif_netdev_port_dump_next,
2379 dpif_netdev_port_dump_done,
2380 dpif_netdev_port_poll,
2381 dpif_netdev_port_poll_wait,
2382 dpif_netdev_flow_get,
2383 dpif_netdev_flow_put,
2384 dpif_netdev_flow_del,
2385 dpif_netdev_flow_flush,
2386 dpif_netdev_flow_dump_create,
2387 dpif_netdev_flow_dump_destroy,
2388 dpif_netdev_flow_dump_thread_create,
2389 dpif_netdev_flow_dump_thread_destroy,
2390 dpif_netdev_flow_dump_next,
2391 dpif_netdev_execute,
2393 dpif_netdev_recv_set,
2394 dpif_netdev_handlers_set,
2395 dpif_netdev_queue_to_priority,
2397 dpif_netdev_recv_wait,
2398 dpif_netdev_recv_purge,
2402 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2403 const char *argv[], void *aux OVS_UNUSED)
2405 struct dp_netdev_port *old_port;
2406 struct dp_netdev_port *new_port;
2407 struct dp_netdev *dp;
2410 ovs_mutex_lock(&dp_netdev_mutex);
2411 dp = shash_find_data(&dp_netdevs, argv[1]);
2412 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2413 ovs_mutex_unlock(&dp_netdev_mutex);
2414 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2417 ovs_refcount_ref(&dp->ref_cnt);
2418 ovs_mutex_unlock(&dp_netdev_mutex);
2420 ovs_mutex_lock(&dp->port_mutex);
2421 if (get_port_by_name(dp, argv[2], &old_port)) {
2422 unixctl_command_reply_error(conn, "unknown port");
2426 port_no = u32_to_odp(atoi(argv[3]));
2427 if (!port_no || port_no == ODPP_NONE) {
2428 unixctl_command_reply_error(conn, "bad port number");
2431 if (dp_netdev_lookup_port(dp, port_no)) {
2432 unixctl_command_reply_error(conn, "port number already in use");
2436 /* Remove old port. */
2437 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
2438 ovsrcu_postpone(free, old_port);
2440 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
2441 new_port = xmemdup(old_port, sizeof *old_port);
2442 new_port->port_no = port_no;
2443 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
2445 seq_change(dp->port_seq);
2446 unixctl_command_reply(conn, NULL);
2449 ovs_mutex_unlock(&dp->port_mutex);
2450 dp_netdev_unref(dp);
2454 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
2455 const char *argv[], void *aux OVS_UNUSED)
2457 struct dp_netdev_port *port;
2458 struct dp_netdev *dp;
2460 ovs_mutex_lock(&dp_netdev_mutex);
2461 dp = shash_find_data(&dp_netdevs, argv[1]);
2462 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2463 ovs_mutex_unlock(&dp_netdev_mutex);
2464 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2467 ovs_refcount_ref(&dp->ref_cnt);
2468 ovs_mutex_unlock(&dp_netdev_mutex);
2470 ovs_mutex_lock(&dp->port_mutex);
2471 if (get_port_by_name(dp, argv[2], &port)) {
2472 unixctl_command_reply_error(conn, "unknown port");
2473 } else if (port->port_no == ODPP_LOCAL) {
2474 unixctl_command_reply_error(conn, "can't delete local port");
2476 do_del_port(dp, port);
2477 unixctl_command_reply(conn, NULL);
2479 ovs_mutex_unlock(&dp->port_mutex);
2481 dp_netdev_unref(dp);
2485 dpif_dummy_register__(const char *type)
2487 struct dpif_class *class;
2489 class = xmalloc(sizeof *class);
2490 *class = dpif_netdev_class;
2491 class->type = xstrdup(type);
2492 dp_register_provider(class);
2496 dpif_dummy_register(bool override)
2503 dp_enumerate_types(&types);
2504 SSET_FOR_EACH (type, &types) {
2505 if (!dp_unregister_provider(type)) {
2506 dpif_dummy_register__(type);
2509 sset_destroy(&types);
2512 dpif_dummy_register__("dummy");
2514 unixctl_command_register("dpif-dummy/change-port-number",
2515 "DP PORT NEW-NUMBER",
2516 3, 3, dpif_dummy_change_port_number, NULL);
2517 unixctl_command_register("dpif-dummy/delete-port", "DP PORT",
2518 2, 2, dpif_dummy_delete_port, NULL);