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"
41 #include "fat-rwlock.h"
46 #include "meta-flow.h"
48 #include "netdev-dpdk.h"
49 #include "netdev-vport.h"
51 #include "odp-execute.h"
53 #include "ofp-print.h"
56 #include "packet-dpif.h"
58 #include "poll-loop.h"
68 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
70 /* By default, choose a priority in the middle. */
71 #define NETDEV_RULE_PRIORITY 0x8000
73 #define FLOW_DUMP_MAX_BATCH 50
75 /* Use per thread recirc_depth to prevent recirculation loop. */
76 #define MAX_RECIRC_DEPTH 5
77 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
79 /* Configuration parameters. */
80 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
83 enum { MAX_QUEUE_LEN = 128 }; /* Maximum number of packets per queue. */
84 enum { QUEUE_MASK = MAX_QUEUE_LEN - 1 };
85 BUILD_ASSERT_DECL(IS_POW2(MAX_QUEUE_LEN));
87 /* Protects against changes to 'dp_netdevs'. */
88 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
90 /* Contains all 'struct dp_netdev's. */
91 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
92 = SHASH_INITIALIZER(&dp_netdevs);
94 struct dp_netdev_upcall {
95 struct dpif_upcall upcall; /* Queued upcall information. */
96 struct ofpbuf buf; /* ofpbuf instance for upcall.packet. */
99 /* A queue passing packets from a struct dp_netdev to its clients (handlers).
105 * Any access at all requires the owning 'dp_netdev''s queue_rwlock and
107 struct dp_netdev_queue {
108 struct ovs_mutex mutex;
109 struct seq *seq; /* Incremented whenever a packet is queued. */
110 struct dp_netdev_upcall upcalls[MAX_QUEUE_LEN] OVS_GUARDED;
111 unsigned int head OVS_GUARDED;
112 unsigned int tail OVS_GUARDED;
115 /* Datapath based on the network device interface from netdev.h.
121 * Some members, marked 'const', are immutable. Accessing other members
122 * requires synchronization, as noted in more detail below.
124 * Acquisition order is, from outermost to innermost:
126 * 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 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
140 * changes to 'cls' must be made while still holding the 'flow_mutex'.
142 struct ovs_mutex flow_mutex;
143 struct classifier cls;
144 struct cmap flow_table OVS_GUARDED; /* Flow table. */
148 * 'queue_rwlock' protects the modification of 'handler_queues' and
149 * 'n_handlers'. The queue elements are protected by its
150 * 'handler_queues''s mutex. */
151 struct fat_rwlock queue_rwlock;
152 struct dp_netdev_queue *handler_queues;
157 * ovsthread_stats is internally synchronized. */
158 struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */
162 * Protected by RCU. Take the mutex to add or remove ports. */
163 struct ovs_mutex port_mutex;
165 struct seq *port_seq; /* Incremented whenever a port changes. */
167 /* Forwarding threads. */
168 struct latch exit_latch;
169 struct pmd_thread *pmd_threads;
170 size_t n_pmd_threads;
174 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
178 DP_STAT_HIT, /* Packets that matched in the flow table. */
179 DP_STAT_MISS, /* Packets that did not match. */
180 DP_STAT_LOST, /* Packets not passed up to the client. */
184 /* Contained by struct dp_netdev's 'stats' member. */
185 struct dp_netdev_stats {
186 struct ovs_mutex mutex; /* Protects 'n'. */
188 /* Indexed by DP_STAT_*, protected by 'mutex'. */
189 unsigned long long int n[DP_N_STATS] OVS_GUARDED;
193 /* A port in a netdev-based datapath. */
194 struct dp_netdev_port {
195 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
197 struct netdev *netdev;
198 struct netdev_saved_flags *sf;
199 struct netdev_rxq **rxq;
200 struct ovs_refcount ref_cnt;
201 char *type; /* Port type as requested by user. */
205 /* Stores a miniflow */
207 /* There are fields in the flow structure that we never use. Therefore we can
208 * save a few words of memory */
209 #define NETDEV_KEY_BUF_SIZE_U32 (FLOW_U32S \
210 - FLOW_U32_SIZE(regs) \
211 - FLOW_U32_SIZE(metadata) \
213 struct netdev_flow_key {
214 struct miniflow flow;
215 uint32_t buf[NETDEV_KEY_BUF_SIZE_U32];
218 /* A flow in dp_netdev's 'flow_table'.
224 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
225 * its dp_netdev's classifier. The text below calls this classifier 'cls'.
230 * The thread safety rules described here for "struct dp_netdev_flow" are
231 * motivated by two goals:
233 * - Prevent threads that read members of "struct dp_netdev_flow" from
234 * reading bad data due to changes by some thread concurrently modifying
237 * - Prevent two threads making changes to members of a given "struct
238 * dp_netdev_flow" from interfering with each other.
244 * A flow 'flow' may be accessed without a risk of being freed by code that
245 * holds a read-lock or write-lock on 'cls->rwlock' or that owns a reference to
246 * 'flow->ref_cnt' (or both). Code that needs to hold onto a flow for a while
247 * should take 'cls->rwlock', find the flow it needs, increment 'flow->ref_cnt'
248 * with dpif_netdev_flow_ref(), and drop 'cls->rwlock'.
250 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
251 * flow from being deleted from 'cls' (that's 'cls->rwlock') and it doesn't
252 * protect members of 'flow' from modification.
254 * Some members, marked 'const', are immutable. Accessing other members
255 * requires synchronization, as noted in more detail below.
257 struct dp_netdev_flow {
258 /* Packet classification. */
259 const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
261 /* Hash table index by unmasked flow. */
262 const struct cmap_node node; /* In owning dp_netdev's 'flow_table'. */
263 const struct flow flow; /* The flow that created this entry. */
267 * Reading or writing these members requires 'mutex'. */
268 struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
271 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
274 static void dp_netdev_flow_free(struct dp_netdev_flow *);
276 /* Contained by struct dp_netdev_flow's 'stats' member. */
277 struct dp_netdev_flow_stats {
278 struct ovs_mutex mutex; /* Guards all the other members. */
280 long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
281 long long int packet_count OVS_GUARDED; /* Number of packets matched. */
282 long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
283 uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
286 /* A set of datapath actions within a "struct dp_netdev_flow".
292 * A struct dp_netdev_actions 'actions' is protected with RCU. */
293 struct dp_netdev_actions {
294 /* These members are immutable: they do not change during the struct's
296 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
297 unsigned int size; /* Size of 'actions', in bytes. */
300 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
302 struct dp_netdev_actions *dp_netdev_flow_get_actions(
303 const struct dp_netdev_flow *);
304 static void dp_netdev_actions_free(struct dp_netdev_actions *);
306 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
307 * the performance overhead of interrupt processing. Therefore netdev can
308 * not implement rx-wait for these devices. dpif-netdev needs to poll
309 * these device to check for recv buffer. pmd-thread does polling for
310 * devices assigned to itself thread.
312 * DPDK used PMD for accessing NIC.
314 * A thread that receives packets from PMD ports, looks them up in the flow
315 * table, and executes the actions it finds.
318 struct dp_netdev *dp;
321 atomic_uint change_seq;
324 /* Interface to netdev-based datapath. */
327 struct dp_netdev *dp;
328 uint64_t last_port_seq;
331 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
332 struct dp_netdev_port **portp);
333 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
334 struct dp_netdev_port **portp);
335 static void dp_netdev_free(struct dp_netdev *)
336 OVS_REQUIRES(dp_netdev_mutex);
337 static void dp_netdev_flow_flush(struct dp_netdev *);
338 static int do_add_port(struct dp_netdev *dp, const char *devname,
339 const char *type, odp_port_t port_no)
340 OVS_REQUIRES(dp->port_mutex);
341 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
342 OVS_REQUIRES(dp->port_mutex);
343 static void dp_netdev_destroy_all_queues(struct dp_netdev *dp)
344 OVS_REQ_WRLOCK(dp->queue_rwlock);
345 static int dpif_netdev_open(const struct dpif_class *, const char *name,
346 bool create, struct dpif **);
347 static int dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf **,
348 int cnt, int queue_no, int type,
349 const struct miniflow *,
350 const struct nlattr *userdata);
351 static void dp_netdev_execute_actions(struct dp_netdev *dp,
352 struct dpif_packet **, int c,
353 bool may_steal, struct pkt_metadata *,
354 const struct nlattr *actions,
356 static void dp_netdev_port_input(struct dp_netdev *dp,
357 struct dpif_packet **packets, int cnt,
360 static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
362 static struct dpif_netdev *
363 dpif_netdev_cast(const struct dpif *dpif)
365 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
366 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
369 static struct dp_netdev *
370 get_dp_netdev(const struct dpif *dpif)
372 return dpif_netdev_cast(dpif)->dp;
376 dpif_netdev_enumerate(struct sset *all_dps,
377 const struct dpif_class *dpif_class)
379 struct shash_node *node;
381 ovs_mutex_lock(&dp_netdev_mutex);
382 SHASH_FOR_EACH(node, &dp_netdevs) {
383 struct dp_netdev *dp = node->data;
384 if (dpif_class != dp->class) {
385 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
386 * If the class doesn't match, skip this dpif. */
389 sset_add(all_dps, node->name);
391 ovs_mutex_unlock(&dp_netdev_mutex);
397 dpif_netdev_class_is_dummy(const struct dpif_class *class)
399 return class != &dpif_netdev_class;
403 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
405 return strcmp(type, "internal") ? type
406 : dpif_netdev_class_is_dummy(class) ? "dummy"
411 create_dpif_netdev(struct dp_netdev *dp)
413 uint16_t netflow_id = hash_string(dp->name, 0);
414 struct dpif_netdev *dpif;
416 ovs_refcount_ref(&dp->ref_cnt);
418 dpif = xmalloc(sizeof *dpif);
419 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
421 dpif->last_port_seq = seq_read(dp->port_seq);
426 /* Choose an unused, non-zero port number and return it on success.
427 * Return ODPP_NONE on failure. */
429 choose_port(struct dp_netdev *dp, const char *name)
430 OVS_REQUIRES(dp->port_mutex)
434 if (dp->class != &dpif_netdev_class) {
438 /* If the port name begins with "br", start the number search at
439 * 100 to make writing tests easier. */
440 if (!strncmp(name, "br", 2)) {
444 /* If the port name contains a number, try to assign that port number.
445 * This can make writing unit tests easier because port numbers are
447 for (p = name; *p != '\0'; p++) {
448 if (isdigit((unsigned char) *p)) {
449 port_no = start_no + strtol(p, NULL, 10);
450 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
451 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
452 return u32_to_odp(port_no);
459 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
460 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
461 return u32_to_odp(port_no);
469 create_dp_netdev(const char *name, const struct dpif_class *class,
470 struct dp_netdev **dpp)
471 OVS_REQUIRES(dp_netdev_mutex)
473 struct dp_netdev *dp;
476 dp = xzalloc(sizeof *dp);
477 shash_add(&dp_netdevs, name, dp);
479 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
480 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
481 ovs_refcount_init(&dp->ref_cnt);
482 atomic_flag_clear(&dp->destroyed);
484 ovs_mutex_init(&dp->flow_mutex);
485 classifier_init(&dp->cls, NULL);
486 cmap_init(&dp->flow_table);
488 fat_rwlock_init(&dp->queue_rwlock);
490 ovsthread_stats_init(&dp->stats);
492 ovs_mutex_init(&dp->port_mutex);
493 cmap_init(&dp->ports);
494 dp->port_seq = seq_create();
495 latch_init(&dp->exit_latch);
497 ovs_mutex_lock(&dp->port_mutex);
498 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
499 ovs_mutex_unlock(&dp->port_mutex);
510 dpif_netdev_open(const struct dpif_class *class, const char *name,
511 bool create, struct dpif **dpifp)
513 struct dp_netdev *dp;
516 ovs_mutex_lock(&dp_netdev_mutex);
517 dp = shash_find_data(&dp_netdevs, name);
519 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
521 error = (dp->class != class ? EINVAL
526 *dpifp = create_dpif_netdev(dp);
528 ovs_mutex_unlock(&dp_netdev_mutex);
534 dp_netdev_purge_queues(struct dp_netdev *dp)
535 OVS_REQ_WRLOCK(dp->queue_rwlock)
539 for (i = 0; i < dp->n_handlers; i++) {
540 struct dp_netdev_queue *q = &dp->handler_queues[i];
542 ovs_mutex_lock(&q->mutex);
543 while (q->tail != q->head) {
544 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
545 ofpbuf_uninit(&u->upcall.packet);
546 ofpbuf_uninit(&u->buf);
548 ovs_mutex_unlock(&q->mutex);
552 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
553 * through the 'dp_netdevs' shash while freeing 'dp'. */
555 dp_netdev_free(struct dp_netdev *dp)
556 OVS_REQUIRES(dp_netdev_mutex)
558 struct dp_netdev_port *port;
559 struct dp_netdev_stats *bucket;
562 shash_find_and_delete(&dp_netdevs, dp->name);
564 dp_netdev_set_pmd_threads(dp, 0);
565 free(dp->pmd_threads);
567 dp_netdev_flow_flush(dp);
568 ovs_mutex_lock(&dp->port_mutex);
569 CMAP_FOR_EACH (port, node, &dp->ports) {
570 do_del_port(dp, port);
572 ovs_mutex_unlock(&dp->port_mutex);
574 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
575 ovs_mutex_destroy(&bucket->mutex);
576 free_cacheline(bucket);
578 ovsthread_stats_destroy(&dp->stats);
580 fat_rwlock_wrlock(&dp->queue_rwlock);
581 dp_netdev_destroy_all_queues(dp);
582 fat_rwlock_unlock(&dp->queue_rwlock);
584 fat_rwlock_destroy(&dp->queue_rwlock);
586 classifier_destroy(&dp->cls);
587 cmap_destroy(&dp->flow_table);
588 ovs_mutex_destroy(&dp->flow_mutex);
589 seq_destroy(dp->port_seq);
590 cmap_destroy(&dp->ports);
591 latch_destroy(&dp->exit_latch);
592 free(CONST_CAST(char *, dp->name));
597 dp_netdev_unref(struct dp_netdev *dp)
600 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
601 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
602 ovs_mutex_lock(&dp_netdev_mutex);
603 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
606 ovs_mutex_unlock(&dp_netdev_mutex);
611 dpif_netdev_close(struct dpif *dpif)
613 struct dp_netdev *dp = get_dp_netdev(dpif);
620 dpif_netdev_destroy(struct dpif *dpif)
622 struct dp_netdev *dp = get_dp_netdev(dpif);
624 if (!atomic_flag_test_and_set(&dp->destroyed)) {
625 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
626 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
635 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
637 struct dp_netdev *dp = get_dp_netdev(dpif);
638 struct dp_netdev_stats *bucket;
641 stats->n_flows = cmap_count(&dp->flow_table);
643 stats->n_hit = stats->n_missed = stats->n_lost = 0;
644 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
645 ovs_mutex_lock(&bucket->mutex);
646 stats->n_hit += bucket->n[DP_STAT_HIT];
647 stats->n_missed += bucket->n[DP_STAT_MISS];
648 stats->n_lost += bucket->n[DP_STAT_LOST];
649 ovs_mutex_unlock(&bucket->mutex);
651 stats->n_masks = UINT32_MAX;
652 stats->n_mask_hit = UINT64_MAX;
658 dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
662 for (i = 0; i < dp->n_pmd_threads; i++) {
663 struct pmd_thread *f = &dp->pmd_threads[i];
666 atomic_add(&f->change_seq, 1, &id);
671 hash_port_no(odp_port_t port_no)
673 return hash_int(odp_to_u32(port_no), 0);
677 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
679 OVS_REQUIRES(dp->port_mutex)
681 struct netdev_saved_flags *sf;
682 struct dp_netdev_port *port;
683 struct netdev *netdev;
684 enum netdev_flags flags;
685 const char *open_type;
689 /* XXX reject devices already in some dp_netdev. */
691 /* Open and validate network device. */
692 open_type = dpif_netdev_port_open_type(dp->class, type);
693 error = netdev_open(devname, open_type, &netdev);
697 /* XXX reject non-Ethernet devices */
699 netdev_get_flags(netdev, &flags);
700 if (flags & NETDEV_LOOPBACK) {
701 VLOG_ERR("%s: cannot add a loopback device", devname);
702 netdev_close(netdev);
706 port = xzalloc(sizeof *port);
707 port->port_no = port_no;
708 port->netdev = netdev;
709 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
710 port->type = xstrdup(type);
711 for (i = 0; i < netdev_n_rxq(netdev); i++) {
712 error = netdev_rxq_open(netdev, &port->rxq[i], i);
714 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
715 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
716 devname, ovs_strerror(errno));
717 netdev_close(netdev);
722 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
724 for (i = 0; i < netdev_n_rxq(netdev); i++) {
725 netdev_rxq_close(port->rxq[i]);
727 netdev_close(netdev);
734 if (netdev_is_pmd(netdev)) {
736 dp_netdev_set_pmd_threads(dp, NR_THREADS);
737 dp_netdev_reload_pmd_threads(dp);
739 ovs_refcount_init(&port->ref_cnt);
741 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
742 seq_change(dp->port_seq);
748 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
749 odp_port_t *port_nop)
751 struct dp_netdev *dp = get_dp_netdev(dpif);
752 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
753 const char *dpif_port;
757 ovs_mutex_lock(&dp->port_mutex);
758 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
759 if (*port_nop != ODPP_NONE) {
761 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
763 port_no = choose_port(dp, dpif_port);
764 error = port_no == ODPP_NONE ? EFBIG : 0;
768 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
770 ovs_mutex_unlock(&dp->port_mutex);
776 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
778 struct dp_netdev *dp = get_dp_netdev(dpif);
781 ovs_mutex_lock(&dp->port_mutex);
782 if (port_no == ODPP_LOCAL) {
785 struct dp_netdev_port *port;
787 error = get_port_by_number(dp, port_no, &port);
789 do_del_port(dp, port);
792 ovs_mutex_unlock(&dp->port_mutex);
798 is_valid_port_number(odp_port_t port_no)
800 return port_no != ODPP_NONE;
803 static struct dp_netdev_port *
804 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
806 struct dp_netdev_port *port;
808 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
809 if (port->port_no == port_no) {
817 get_port_by_number(struct dp_netdev *dp,
818 odp_port_t port_no, struct dp_netdev_port **portp)
820 if (!is_valid_port_number(port_no)) {
824 *portp = dp_netdev_lookup_port(dp, port_no);
825 return *portp ? 0 : ENOENT;
830 port_ref(struct dp_netdev_port *port)
833 ovs_refcount_ref(&port->ref_cnt);
838 port_destroy__(struct dp_netdev_port *port)
840 int n_rxq = netdev_n_rxq(port->netdev);
843 netdev_close(port->netdev);
844 netdev_restore_flags(port->sf);
846 for (i = 0; i < n_rxq; i++) {
847 netdev_rxq_close(port->rxq[i]);
855 port_unref(struct dp_netdev_port *port)
857 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
858 ovsrcu_postpone(port_destroy__, port);
863 get_port_by_name(struct dp_netdev *dp,
864 const char *devname, struct dp_netdev_port **portp)
865 OVS_REQUIRES(dp->port_mutex)
867 struct dp_netdev_port *port;
869 CMAP_FOR_EACH (port, node, &dp->ports) {
870 if (!strcmp(netdev_get_name(port->netdev), devname)) {
879 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
880 OVS_REQUIRES(dp->port_mutex)
882 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
883 seq_change(dp->port_seq);
884 if (netdev_is_pmd(port->netdev)) {
885 dp_netdev_reload_pmd_threads(dp);
892 answer_port_query(const struct dp_netdev_port *port,
893 struct dpif_port *dpif_port)
895 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
896 dpif_port->type = xstrdup(port->type);
897 dpif_port->port_no = port->port_no;
901 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
902 struct dpif_port *dpif_port)
904 struct dp_netdev *dp = get_dp_netdev(dpif);
905 struct dp_netdev_port *port;
908 error = get_port_by_number(dp, port_no, &port);
909 if (!error && dpif_port) {
910 answer_port_query(port, dpif_port);
917 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
918 struct dpif_port *dpif_port)
920 struct dp_netdev *dp = get_dp_netdev(dpif);
921 struct dp_netdev_port *port;
924 ovs_mutex_lock(&dp->port_mutex);
925 error = get_port_by_name(dp, devname, &port);
926 if (!error && dpif_port) {
927 answer_port_query(port, dpif_port);
929 ovs_mutex_unlock(&dp->port_mutex);
935 dp_netdev_flow_free(struct dp_netdev_flow *flow)
937 struct dp_netdev_flow_stats *bucket;
940 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
941 ovs_mutex_destroy(&bucket->mutex);
942 free_cacheline(bucket);
944 ovsthread_stats_destroy(&flow->stats);
946 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
947 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
952 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
953 OVS_REQUIRES(dp->flow_mutex)
955 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
956 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
958 classifier_remove(&dp->cls, cr);
959 cmap_remove(&dp->flow_table, node, flow_hash(&flow->flow, 0));
960 ovsrcu_postpone(dp_netdev_flow_free, flow);
964 dp_netdev_flow_flush(struct dp_netdev *dp)
966 struct dp_netdev_flow *netdev_flow, *next;
968 ovs_mutex_lock(&dp->flow_mutex);
969 CMAP_FOR_EACH_SAFE (netdev_flow, next, node, &dp->flow_table) {
970 dp_netdev_remove_flow(dp, netdev_flow);
972 ovs_mutex_unlock(&dp->flow_mutex);
976 dpif_netdev_flow_flush(struct dpif *dpif)
978 struct dp_netdev *dp = get_dp_netdev(dpif);
980 dp_netdev_flow_flush(dp);
984 struct dp_netdev_port_state {
985 struct cmap_position position;
990 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
992 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
997 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
998 struct dpif_port *dpif_port)
1000 struct dp_netdev_port_state *state = state_;
1001 struct dp_netdev *dp = get_dp_netdev(dpif);
1002 struct cmap_node *node;
1005 node = cmap_next_position(&dp->ports, &state->position);
1007 struct dp_netdev_port *port;
1009 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1012 state->name = xstrdup(netdev_get_name(port->netdev));
1013 dpif_port->name = state->name;
1014 dpif_port->type = port->type;
1015 dpif_port->port_no = port->port_no;
1026 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1028 struct dp_netdev_port_state *state = state_;
1035 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1037 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1038 uint64_t new_port_seq;
1041 new_port_seq = seq_read(dpif->dp->port_seq);
1042 if (dpif->last_port_seq != new_port_seq) {
1043 dpif->last_port_seq = new_port_seq;
1053 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1055 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1057 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1060 static struct dp_netdev_flow *
1061 dp_netdev_flow_cast(const struct cls_rule *cr)
1063 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1066 static struct dp_netdev_flow *
1067 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
1069 struct dp_netdev_flow *netdev_flow;
1070 struct cls_rule *rule;
1072 classifier_lookup_miniflow_batch(&dp->cls, &key, &rule, 1);
1073 netdev_flow = dp_netdev_flow_cast(rule);
1078 static struct dp_netdev_flow *
1079 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1081 struct dp_netdev_flow *netdev_flow;
1083 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1085 if (flow_equal(&netdev_flow->flow, flow)) {
1094 get_dpif_flow_stats(struct dp_netdev_flow *netdev_flow,
1095 struct dpif_flow_stats *stats)
1097 struct dp_netdev_flow_stats *bucket;
1100 memset(stats, 0, sizeof *stats);
1101 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1102 ovs_mutex_lock(&bucket->mutex);
1103 stats->n_packets += bucket->packet_count;
1104 stats->n_bytes += bucket->byte_count;
1105 stats->used = MAX(stats->used, bucket->used);
1106 stats->tcp_flags |= bucket->tcp_flags;
1107 ovs_mutex_unlock(&bucket->mutex);
1112 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1113 const struct nlattr *mask_key,
1114 uint32_t mask_key_len, const struct flow *flow,
1118 enum odp_key_fitness fitness;
1120 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1122 /* This should not happen: it indicates that
1123 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1124 * disagree on the acceptable form of a mask. Log the problem
1125 * as an error, with enough details to enable debugging. */
1126 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1128 if (!VLOG_DROP_ERR(&rl)) {
1132 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1134 VLOG_ERR("internal error parsing flow mask %s (%s)",
1135 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1142 enum mf_field_id id;
1143 /* No mask key, unwildcard everything except fields whose
1144 * prerequisities are not met. */
1145 memset(mask, 0x0, sizeof *mask);
1147 for (id = 0; id < MFF_N_IDS; ++id) {
1148 /* Skip registers and metadata. */
1149 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1150 && id != MFF_METADATA) {
1151 const struct mf_field *mf = mf_from_id(id);
1152 if (mf_are_prereqs_ok(mf, flow)) {
1153 mf_mask_field(mf, mask);
1159 /* Force unwildcard the in_port.
1161 * We need to do this even in the case where we unwildcard "everything"
1162 * above because "everything" only includes the 16-bit OpenFlow port number
1163 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1164 * port number mask->in_port.odp_port. */
1165 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1171 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1176 if (odp_flow_key_to_flow(key, key_len, flow)) {
1177 /* This should not happen: it indicates that odp_flow_key_from_flow()
1178 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1179 * flow. Log the problem as an error, with enough details to enable
1181 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1183 if (!VLOG_DROP_ERR(&rl)) {
1187 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1188 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1195 in_port = flow->in_port.odp_port;
1196 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1204 dpif_netdev_flow_get(const struct dpif *dpif,
1205 const struct nlattr *nl_key, size_t nl_key_len,
1206 struct ofpbuf **actionsp, struct dpif_flow_stats *stats)
1208 struct dp_netdev *dp = get_dp_netdev(dpif);
1209 struct dp_netdev_flow *netdev_flow;
1213 error = dpif_netdev_flow_from_nlattrs(nl_key, nl_key_len, &key);
1218 netdev_flow = dp_netdev_find_flow(dp, &key);
1222 get_dpif_flow_stats(netdev_flow, stats);
1226 struct dp_netdev_actions *actions;
1228 actions = dp_netdev_flow_get_actions(netdev_flow);
1229 *actionsp = ofpbuf_clone_data(actions->actions, actions->size);
1239 dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow,
1240 const struct flow_wildcards *wc,
1241 const struct nlattr *actions,
1243 OVS_REQUIRES(dp->flow_mutex)
1245 struct dp_netdev_flow *netdev_flow;
1248 netdev_flow = xzalloc(sizeof *netdev_flow);
1249 *CONST_CAST(struct flow *, &netdev_flow->flow) = *flow;
1251 ovsthread_stats_init(&netdev_flow->stats);
1253 ovsrcu_set(&netdev_flow->actions,
1254 dp_netdev_actions_create(actions, actions_len));
1256 match_init(&match, flow, wc);
1257 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1258 &match, NETDEV_RULE_PRIORITY);
1259 cmap_insert(&dp->flow_table,
1260 CONST_CAST(struct cmap_node *, &netdev_flow->node),
1261 flow_hash(flow, 0));
1262 classifier_insert(&dp->cls,
1263 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1269 clear_stats(struct dp_netdev_flow *netdev_flow)
1271 struct dp_netdev_flow_stats *bucket;
1274 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1275 ovs_mutex_lock(&bucket->mutex);
1277 bucket->packet_count = 0;
1278 bucket->byte_count = 0;
1279 bucket->tcp_flags = 0;
1280 ovs_mutex_unlock(&bucket->mutex);
1285 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1287 struct dp_netdev *dp = get_dp_netdev(dpif);
1288 struct dp_netdev_flow *netdev_flow;
1290 struct miniflow miniflow;
1291 struct flow_wildcards wc;
1294 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &flow);
1298 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1299 put->mask, put->mask_len,
1304 miniflow_init(&miniflow, &flow);
1306 ovs_mutex_lock(&dp->flow_mutex);
1307 netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
1309 if (put->flags & DPIF_FP_CREATE) {
1310 if (cmap_count(&dp->flow_table) < MAX_FLOWS) {
1312 memset(put->stats, 0, sizeof *put->stats);
1314 error = dp_netdev_flow_add(dp, &flow, &wc, put->actions,
1323 if (put->flags & DPIF_FP_MODIFY
1324 && flow_equal(&flow, &netdev_flow->flow)) {
1325 struct dp_netdev_actions *new_actions;
1326 struct dp_netdev_actions *old_actions;
1328 new_actions = dp_netdev_actions_create(put->actions,
1331 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1332 ovsrcu_set(&netdev_flow->actions, new_actions);
1335 get_dpif_flow_stats(netdev_flow, put->stats);
1337 if (put->flags & DPIF_FP_ZERO_STATS) {
1338 clear_stats(netdev_flow);
1341 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1342 } else if (put->flags & DPIF_FP_CREATE) {
1345 /* Overlapping flow. */
1349 ovs_mutex_unlock(&dp->flow_mutex);
1350 miniflow_destroy(&miniflow);
1356 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1358 struct dp_netdev *dp = get_dp_netdev(dpif);
1359 struct dp_netdev_flow *netdev_flow;
1363 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1368 ovs_mutex_lock(&dp->flow_mutex);
1369 netdev_flow = dp_netdev_find_flow(dp, &key);
1372 get_dpif_flow_stats(netdev_flow, del->stats);
1374 dp_netdev_remove_flow(dp, netdev_flow);
1378 ovs_mutex_unlock(&dp->flow_mutex);
1383 struct dpif_netdev_flow_dump {
1384 struct dpif_flow_dump up;
1385 struct cmap_position pos;
1387 struct ovs_mutex mutex;
1390 static struct dpif_netdev_flow_dump *
1391 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1393 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1396 static struct dpif_flow_dump *
1397 dpif_netdev_flow_dump_create(const struct dpif *dpif_)
1399 struct dpif_netdev_flow_dump *dump;
1401 dump = xmalloc(sizeof *dump);
1402 dpif_flow_dump_init(&dump->up, dpif_);
1403 memset(&dump->pos, 0, sizeof dump->pos);
1405 ovs_mutex_init(&dump->mutex);
1411 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1413 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1415 ovs_mutex_destroy(&dump->mutex);
1420 struct dpif_netdev_flow_dump_thread {
1421 struct dpif_flow_dump_thread up;
1422 struct dpif_netdev_flow_dump *dump;
1423 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1424 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1427 static struct dpif_netdev_flow_dump_thread *
1428 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1430 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1433 static struct dpif_flow_dump_thread *
1434 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1436 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1437 struct dpif_netdev_flow_dump_thread *thread;
1439 thread = xmalloc(sizeof *thread);
1440 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1441 thread->dump = dump;
1446 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1448 struct dpif_netdev_flow_dump_thread *thread
1449 = dpif_netdev_flow_dump_thread_cast(thread_);
1454 /* XXX the caller must use 'actions' without quiescing */
1456 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
1457 struct dpif_flow *flows, int max_flows)
1459 struct dpif_netdev_flow_dump_thread *thread
1460 = dpif_netdev_flow_dump_thread_cast(thread_);
1461 struct dpif_netdev_flow_dump *dump = thread->dump;
1462 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
1463 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
1464 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
1468 ovs_mutex_lock(&dump->mutex);
1469 if (!dump->status) {
1470 for (n_flows = 0; n_flows < MIN(max_flows, FLOW_DUMP_MAX_BATCH);
1472 struct cmap_node *node;
1474 node = cmap_next_position(&dp->flow_table, &dump->pos);
1479 netdev_flows[n_flows] = CONTAINER_OF(node, struct dp_netdev_flow,
1483 ovs_mutex_unlock(&dump->mutex);
1485 for (i = 0; i < n_flows; i++) {
1486 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
1487 struct odputil_keybuf *keybuf = &thread->keybuf[i];
1488 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
1489 struct dpif_flow *f = &flows[i];
1490 struct dp_netdev_actions *dp_actions;
1491 struct flow_wildcards wc;
1494 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1497 ofpbuf_use_stack(&buf, keybuf, sizeof *keybuf);
1498 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1499 netdev_flow->flow.in_port.odp_port, true);
1500 f->key = ofpbuf_data(&buf);
1501 f->key_len = ofpbuf_size(&buf);
1504 ofpbuf_use_stack(&buf, maskbuf, sizeof *maskbuf);
1505 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1506 odp_to_u32(wc.masks.in_port.odp_port),
1508 f->mask = ofpbuf_data(&buf);
1509 f->mask_len = ofpbuf_size(&buf);
1512 dp_actions = dp_netdev_flow_get_actions(netdev_flow);
1513 f->actions = dp_actions->actions;
1514 f->actions_len = dp_actions->size;
1517 get_dpif_flow_stats(netdev_flow, &f->stats);
1524 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1526 struct dp_netdev *dp = get_dp_netdev(dpif);
1527 struct dpif_packet packet, *pp;
1528 struct pkt_metadata *md = &execute->md;
1530 if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
1531 ofpbuf_size(execute->packet) > UINT16_MAX) {
1535 packet.ofpbuf = *execute->packet;
1538 dp_netdev_execute_actions(dp, &pp, 1, false, md,
1539 execute->actions, execute->actions_len);
1541 /* Even though may_steal is set to false, some actions could modify or
1542 * reallocate the ofpbuf memory. We need to pass those changes to the
1544 *execute->packet = packet.ofpbuf;
1550 dp_netdev_destroy_all_queues(struct dp_netdev *dp)
1551 OVS_REQ_WRLOCK(dp->queue_rwlock)
1555 dp_netdev_purge_queues(dp);
1557 for (i = 0; i < dp->n_handlers; i++) {
1558 struct dp_netdev_queue *q = &dp->handler_queues[i];
1560 ovs_mutex_destroy(&q->mutex);
1561 seq_destroy(q->seq);
1563 free(dp->handler_queues);
1564 dp->handler_queues = NULL;
1569 dp_netdev_refresh_queues(struct dp_netdev *dp, uint32_t n_handlers)
1570 OVS_REQ_WRLOCK(dp->queue_rwlock)
1572 if (dp->n_handlers != n_handlers) {
1575 dp_netdev_destroy_all_queues(dp);
1577 dp->n_handlers = n_handlers;
1578 dp->handler_queues = xzalloc(n_handlers * sizeof *dp->handler_queues);
1580 for (i = 0; i < n_handlers; i++) {
1581 struct dp_netdev_queue *q = &dp->handler_queues[i];
1583 ovs_mutex_init(&q->mutex);
1584 q->seq = seq_create();
1590 dpif_netdev_recv_set(struct dpif *dpif, bool enable)
1592 struct dp_netdev *dp = get_dp_netdev(dpif);
1594 if ((dp->handler_queues != NULL) == enable) {
1598 fat_rwlock_wrlock(&dp->queue_rwlock);
1600 dp_netdev_destroy_all_queues(dp);
1602 dp_netdev_refresh_queues(dp, 1);
1604 fat_rwlock_unlock(&dp->queue_rwlock);
1610 dpif_netdev_handlers_set(struct dpif *dpif, uint32_t n_handlers)
1612 struct dp_netdev *dp = get_dp_netdev(dpif);
1614 fat_rwlock_wrlock(&dp->queue_rwlock);
1615 if (dp->handler_queues) {
1616 dp_netdev_refresh_queues(dp, n_handlers);
1618 fat_rwlock_unlock(&dp->queue_rwlock);
1624 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1625 uint32_t queue_id, uint32_t *priority)
1627 *priority = queue_id;
1632 dp_netdev_recv_check(const struct dp_netdev *dp, const uint32_t handler_id)
1633 OVS_REQ_RDLOCK(dp->queue_rwlock)
1635 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1637 if (!dp->handler_queues) {
1638 VLOG_WARN_RL(&rl, "receiving upcall disabled");
1642 if (handler_id >= dp->n_handlers) {
1643 VLOG_WARN_RL(&rl, "handler index out of bound");
1651 dpif_netdev_recv(struct dpif *dpif, uint32_t handler_id,
1652 struct dpif_upcall *upcall, struct ofpbuf *buf)
1654 struct dp_netdev *dp = get_dp_netdev(dpif);
1655 struct dp_netdev_queue *q;
1658 fat_rwlock_rdlock(&dp->queue_rwlock);
1660 if (!dp_netdev_recv_check(dp, handler_id)) {
1665 q = &dp->handler_queues[handler_id];
1666 ovs_mutex_lock(&q->mutex);
1667 if (q->head != q->tail) {
1668 struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
1670 *upcall = u->upcall;
1677 ovs_mutex_unlock(&q->mutex);
1680 fat_rwlock_unlock(&dp->queue_rwlock);
1686 dpif_netdev_recv_wait(struct dpif *dpif, uint32_t handler_id)
1688 struct dp_netdev *dp = get_dp_netdev(dpif);
1689 struct dp_netdev_queue *q;
1692 fat_rwlock_rdlock(&dp->queue_rwlock);
1694 if (!dp_netdev_recv_check(dp, handler_id)) {
1698 q = &dp->handler_queues[handler_id];
1699 ovs_mutex_lock(&q->mutex);
1700 seq = seq_read(q->seq);
1701 if (q->head != q->tail) {
1702 poll_immediate_wake();
1704 seq_wait(q->seq, seq);
1707 ovs_mutex_unlock(&q->mutex);
1710 fat_rwlock_unlock(&dp->queue_rwlock);
1714 dpif_netdev_recv_purge(struct dpif *dpif)
1716 struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif);
1718 fat_rwlock_wrlock(&dpif_netdev->dp->queue_rwlock);
1719 dp_netdev_purge_queues(dpif_netdev->dp);
1720 fat_rwlock_unlock(&dpif_netdev->dp->queue_rwlock);
1723 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1724 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1726 struct dp_netdev_actions *
1727 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1729 struct dp_netdev_actions *netdev_actions;
1731 netdev_actions = xmalloc(sizeof *netdev_actions);
1732 netdev_actions->actions = xmemdup(actions, size);
1733 netdev_actions->size = size;
1735 return netdev_actions;
1738 struct dp_netdev_actions *
1739 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
1741 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
1745 dp_netdev_actions_free(struct dp_netdev_actions *actions)
1747 free(actions->actions);
1753 dp_netdev_process_rxq_port(struct dp_netdev *dp,
1754 struct dp_netdev_port *port,
1755 struct netdev_rxq *rxq)
1757 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1760 error = netdev_rxq_recv(rxq, packets, &cnt);
1762 dp_netdev_port_input(dp, packets, cnt, port->port_no);
1763 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1764 static struct vlog_rate_limit rl
1765 = VLOG_RATE_LIMIT_INIT(1, 5);
1767 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1768 netdev_get_name(port->netdev),
1769 ovs_strerror(error));
1774 dpif_netdev_run(struct dpif *dpif)
1776 struct dp_netdev_port *port;
1777 struct dp_netdev *dp = get_dp_netdev(dpif);
1779 CMAP_FOR_EACH (port, node, &dp->ports) {
1780 if (!netdev_is_pmd(port->netdev)) {
1783 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1784 dp_netdev_process_rxq_port(dp, port, port->rxq[i]);
1791 dpif_netdev_wait(struct dpif *dpif)
1793 struct dp_netdev_port *port;
1794 struct dp_netdev *dp = get_dp_netdev(dpif);
1796 ovs_mutex_lock(&dp_netdev_mutex);
1797 CMAP_FOR_EACH (port, node, &dp->ports) {
1798 if (!netdev_is_pmd(port->netdev)) {
1801 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1802 netdev_rxq_wait(port->rxq[i]);
1806 ovs_mutex_unlock(&dp_netdev_mutex);
1810 struct dp_netdev_port *port;
1811 struct netdev_rxq *rx;
1815 pmd_load_queues(struct pmd_thread *f,
1816 struct rxq_poll **ppoll_list, int poll_cnt)
1818 struct dp_netdev *dp = f->dp;
1819 struct rxq_poll *poll_list = *ppoll_list;
1820 struct dp_netdev_port *port;
1825 /* Simple scheduler for netdev rx polling. */
1826 for (i = 0; i < poll_cnt; i++) {
1827 port_unref(poll_list[i].port);
1833 CMAP_FOR_EACH (port, node, &f->dp->ports) {
1834 if (netdev_is_pmd(port->netdev)) {
1837 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1838 if ((index % dp->n_pmd_threads) == id) {
1839 poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1));
1842 poll_list[poll_cnt].port = port;
1843 poll_list[poll_cnt].rx = port->rxq[i];
1851 *ppoll_list = poll_list;
1856 pmd_thread_main(void *f_)
1858 struct pmd_thread *f = f_;
1859 struct dp_netdev *dp = f->dp;
1860 unsigned int lc = 0;
1861 struct rxq_poll *poll_list;
1862 unsigned int port_seq;
1869 pmd_thread_setaffinity_cpu(f->id);
1871 poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
1872 atomic_read(&f->change_seq, &port_seq);
1875 unsigned int c_port_seq;
1878 for (i = 0; i < poll_cnt; i++) {
1879 dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx);
1885 /* TODO: need completely userspace based signaling method.
1886 * to keep this thread entirely in userspace.
1887 * For now using atomic counter. */
1889 atomic_read_explicit(&f->change_seq, &c_port_seq, memory_order_consume);
1890 if (c_port_seq != port_seq) {
1896 if (!latch_is_set(&f->dp->exit_latch)){
1900 for (i = 0; i < poll_cnt; i++) {
1901 port_unref(poll_list[i].port);
1909 dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n)
1913 if (n == dp->n_pmd_threads) {
1917 /* Stop existing threads. */
1918 latch_set(&dp->exit_latch);
1919 dp_netdev_reload_pmd_threads(dp);
1920 for (i = 0; i < dp->n_pmd_threads; i++) {
1921 struct pmd_thread *f = &dp->pmd_threads[i];
1923 xpthread_join(f->thread, NULL);
1925 latch_poll(&dp->exit_latch);
1926 free(dp->pmd_threads);
1928 /* Start new threads. */
1929 dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
1930 dp->n_pmd_threads = n;
1932 for (i = 0; i < n; i++) {
1933 struct pmd_thread *f = &dp->pmd_threads[i];
1937 atomic_store(&f->change_seq, 1);
1939 /* Each thread will distribute all devices rx-queues among
1941 f->thread = ovs_thread_create("pmd", pmd_thread_main, f);
1947 dp_netdev_flow_stats_new_cb(void)
1949 struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1950 ovs_mutex_init(&bucket->mutex);
1955 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
1959 long long int now = time_msec();
1960 struct dp_netdev_flow_stats *bucket;
1962 bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
1963 dp_netdev_flow_stats_new_cb);
1965 ovs_mutex_lock(&bucket->mutex);
1966 bucket->used = MAX(now, bucket->used);
1967 bucket->packet_count += cnt;
1968 bucket->byte_count += size;
1969 bucket->tcp_flags |= tcp_flags;
1970 ovs_mutex_unlock(&bucket->mutex);
1974 dp_netdev_stats_new_cb(void)
1976 struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1977 ovs_mutex_init(&bucket->mutex);
1982 dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type, int cnt)
1984 struct dp_netdev_stats *bucket;
1986 bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
1987 ovs_mutex_lock(&bucket->mutex);
1988 bucket->n[type] += cnt;
1989 ovs_mutex_unlock(&bucket->mutex);
1992 struct packet_batch {
1993 unsigned int packet_count;
1994 unsigned int byte_count;
1997 struct dp_netdev_flow *flow;
1999 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
2000 struct pkt_metadata md;
2004 packet_batch_update(struct packet_batch *batch,
2005 struct dpif_packet *packet, const struct miniflow *mf)
2007 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
2008 batch->packets[batch->packet_count++] = packet;
2009 batch->byte_count += ofpbuf_size(&packet->ofpbuf);
2013 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow,
2014 struct pkt_metadata *md)
2019 batch->packet_count = 0;
2020 batch->byte_count = 0;
2021 batch->tcp_flags = 0;
2025 packet_batch_execute(struct packet_batch *batch, struct dp_netdev *dp)
2027 struct dp_netdev_actions *actions;
2028 struct dp_netdev_flow *flow = batch->flow;
2030 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
2033 actions = dp_netdev_flow_get_actions(flow);
2035 dp_netdev_execute_actions(dp, batch->packets,
2036 batch->packet_count, true, &batch->md,
2037 actions->actions, actions->size);
2039 dp_netdev_count_packet(dp, DP_STAT_HIT, batch->packet_count);
2043 dp_netdev_input(struct dp_netdev *dp, struct dpif_packet **packets, int cnt,
2044 struct pkt_metadata *md)
2046 struct packet_batch batches[NETDEV_MAX_RX_BATCH];
2047 struct netdev_flow_key keys[NETDEV_MAX_RX_BATCH];
2048 const struct miniflow *mfs[NETDEV_MAX_RX_BATCH]; /* NULL at bad packets. */
2049 struct cls_rule *rules[NETDEV_MAX_RX_BATCH];
2050 size_t n_batches, i;
2052 for (i = 0; i < cnt; i++) {
2053 if (OVS_UNLIKELY(ofpbuf_size(&packets[i]->ofpbuf) < ETH_HEADER_LEN)) {
2054 dpif_packet_delete(packets[i]);
2059 miniflow_initialize(&keys[i].flow, keys[i].buf);
2060 miniflow_extract(&packets[i]->ofpbuf, md, &keys[i].flow);
2061 mfs[i] = &keys[i].flow;
2064 classifier_lookup_miniflow_batch(&dp->cls, mfs, rules, cnt);
2067 for (i = 0; i < cnt; i++) {
2068 struct dp_netdev_flow *flow;
2069 struct packet_batch *batch;
2072 if (OVS_UNLIKELY(!mfs[i])) {
2076 if (OVS_UNLIKELY(!rules[i])) {
2077 dp_netdev_count_packet(dp, DP_STAT_MISS, 1);
2078 if (OVS_LIKELY(dp->handler_queues)) {
2079 uint32_t hash = miniflow_hash_5tuple(mfs[i], 0);
2080 struct ofpbuf *buf = &packets[i]->ofpbuf;
2082 dp_netdev_output_userspace(dp, &buf, 1, hash % dp->n_handlers,
2083 DPIF_UC_MISS, mfs[i], NULL);
2085 /* No upcall queue. Freeing the packet */
2086 dpif_packet_delete(packets[i]);
2091 /* XXX: This O(n^2) algortihm makes sense if we're operating under the
2092 * assumption that the number of distinct flows (and therefore the
2093 * number of distinct batches) is quite small. If this turns out not
2094 * to be the case, it may make sense to pre sort based on the
2095 * netdev_flow pointer. That done we can get the appropriate batching
2096 * in O(n * log(n)) instead. */
2098 flow = dp_netdev_flow_cast(rules[i]);
2099 for (j = 0; j < n_batches; j++) {
2100 if (batches[j].flow == flow) {
2101 batch = &batches[j];
2107 batch = &batches[n_batches++];
2108 packet_batch_init(batch, flow, md);
2110 packet_batch_update(batch, packets[i], mfs[i]);
2113 for (i = 0; i < n_batches; i++) {
2114 packet_batch_execute(&batches[i], dp);
2119 dp_netdev_port_input(struct dp_netdev *dp, struct dpif_packet **packets,
2120 int cnt, odp_port_t port_no)
2122 uint32_t *recirc_depth = recirc_depth_get();
2123 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port_no);
2126 dp_netdev_input(dp, packets, cnt, &md);
2130 dp_netdev_queue_userspace_packet(struct dp_netdev_queue *q,
2131 struct ofpbuf *packet, int type,
2132 const struct miniflow *key,
2133 const struct nlattr *userdata)
2134 OVS_REQUIRES(q->mutex)
2136 if (q->head - q->tail < MAX_QUEUE_LEN) {
2137 struct dp_netdev_upcall *u = &q->upcalls[q->head++ & QUEUE_MASK];
2138 struct dpif_upcall *upcall = &u->upcall;
2139 struct ofpbuf *buf = &u->buf;
2143 upcall->type = type;
2145 /* Allocate buffer big enough for everything. */
2146 buf_size = ODPUTIL_FLOW_KEY_BYTES;
2148 buf_size += NLA_ALIGN(userdata->nla_len);
2150 ofpbuf_init(buf, buf_size);
2153 miniflow_expand(key, &flow);
2154 odp_flow_key_from_flow(buf, &flow, NULL, flow.in_port.odp_port, true);
2155 upcall->key = ofpbuf_data(buf);
2156 upcall->key_len = ofpbuf_size(buf);
2160 upcall->userdata = ofpbuf_put(buf, userdata,
2161 NLA_ALIGN(userdata->nla_len));
2164 upcall->packet = *packet;
2170 ofpbuf_delete(packet);
2177 dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf **packets,
2178 int cnt, int queue_no, int type,
2179 const struct miniflow *key,
2180 const struct nlattr *userdata)
2182 struct dp_netdev_queue *q;
2186 fat_rwlock_rdlock(&dp->queue_rwlock);
2187 q = &dp->handler_queues[queue_no];
2188 ovs_mutex_lock(&q->mutex);
2189 for (i = 0; i < cnt; i++) {
2190 struct ofpbuf *packet = packets[i];
2192 error = dp_netdev_queue_userspace_packet(q, packet, type, key,
2194 if (error == ENOBUFS) {
2195 dp_netdev_count_packet(dp, DP_STAT_LOST, 1);
2198 ovs_mutex_unlock(&q->mutex);
2199 fat_rwlock_unlock(&dp->queue_rwlock);
2204 struct dp_netdev_execute_aux {
2205 struct dp_netdev *dp;
2209 dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
2210 struct pkt_metadata *md,
2211 const struct nlattr *a, bool may_steal)
2212 OVS_NO_THREAD_SAFETY_ANALYSIS
2214 struct dp_netdev_execute_aux *aux = aux_;
2215 int type = nl_attr_type(a);
2216 struct dp_netdev_port *p;
2217 uint32_t *depth = recirc_depth_get();
2220 switch ((enum ovs_action_attr)type) {
2221 case OVS_ACTION_ATTR_OUTPUT:
2222 p = dp_netdev_lookup_port(aux->dp, u32_to_odp(nl_attr_get_u32(a)));
2223 if (OVS_LIKELY(p)) {
2224 netdev_send(p->netdev, packets, cnt, may_steal);
2225 } else if (may_steal) {
2226 for (i = 0; i < cnt; i++) {
2227 dpif_packet_delete(packets[i]);
2232 case OVS_ACTION_ATTR_USERSPACE: {
2233 const struct nlattr *userdata;
2234 struct netdev_flow_key key;
2236 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2238 miniflow_initialize(&key.flow, key.buf);
2240 for (i = 0; i < cnt; i++) {
2241 struct ofpbuf *packet, *userspace_packet;
2243 packet = &packets[i]->ofpbuf;
2245 miniflow_extract(packet, md, &key.flow);
2247 userspace_packet = may_steal ? packet : ofpbuf_clone(packet);
2249 dp_netdev_output_userspace(aux->dp, &userspace_packet, 1,
2250 miniflow_hash_5tuple(&key.flow, 0)
2251 % aux->dp->n_handlers,
2252 DPIF_UC_ACTION, &key.flow,
2258 case OVS_ACTION_ATTR_HASH: {
2259 const struct ovs_action_hash *hash_act;
2260 struct netdev_flow_key key;
2263 hash_act = nl_attr_get(a);
2265 miniflow_initialize(&key.flow, key.buf);
2267 for (i = 0; i < cnt; i++) {
2269 /* TODO: this is slow. Use RSS hash in the future */
2270 miniflow_extract(&packets[i]->ofpbuf, md, &key.flow);
2272 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
2273 /* Hash need not be symmetric, nor does it need to include
2275 hash = miniflow_hash_5tuple(&key.flow, hash_act->hash_basis);
2277 VLOG_WARN("Unknown hash algorithm specified "
2278 "for the hash action.");
2283 hash = 1; /* 0 is not valid */
2289 packets[i]->dp_hash = hash;
2294 case OVS_ACTION_ATTR_RECIRC:
2295 if (*depth < MAX_RECIRC_DEPTH) {
2298 for (i = 0; i < cnt; i++) {
2299 struct dpif_packet *recirc_pkt;
2300 struct pkt_metadata recirc_md = *md;
2302 recirc_pkt = (may_steal) ? packets[i]
2303 : dpif_packet_clone(packets[i]);
2305 recirc_md.recirc_id = nl_attr_get_u32(a);
2307 /* Hash is private to each packet */
2308 recirc_md.dp_hash = packets[i]->dp_hash;
2310 dp_netdev_input(aux->dp, &recirc_pkt, 1, &recirc_md);
2316 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
2318 for (i = 0; i < cnt; i++) {
2319 dpif_packet_delete(packets[i]);
2325 case OVS_ACTION_ATTR_PUSH_VLAN:
2326 case OVS_ACTION_ATTR_POP_VLAN:
2327 case OVS_ACTION_ATTR_PUSH_MPLS:
2328 case OVS_ACTION_ATTR_POP_MPLS:
2329 case OVS_ACTION_ATTR_SET:
2330 case OVS_ACTION_ATTR_SAMPLE:
2331 case OVS_ACTION_ATTR_UNSPEC:
2332 case __OVS_ACTION_ATTR_MAX:
2338 dp_netdev_execute_actions(struct dp_netdev *dp,
2339 struct dpif_packet **packets, int cnt,
2340 bool may_steal, struct pkt_metadata *md,
2341 const struct nlattr *actions, size_t actions_len)
2343 struct dp_netdev_execute_aux aux = {dp};
2345 odp_execute_actions(&aux, packets, cnt, may_steal, md, actions,
2346 actions_len, dp_execute_cb);
2349 const struct dpif_class dpif_netdev_class = {
2351 dpif_netdev_enumerate,
2352 dpif_netdev_port_open_type,
2355 dpif_netdev_destroy,
2358 dpif_netdev_get_stats,
2359 dpif_netdev_port_add,
2360 dpif_netdev_port_del,
2361 dpif_netdev_port_query_by_number,
2362 dpif_netdev_port_query_by_name,
2363 NULL, /* port_get_pid */
2364 dpif_netdev_port_dump_start,
2365 dpif_netdev_port_dump_next,
2366 dpif_netdev_port_dump_done,
2367 dpif_netdev_port_poll,
2368 dpif_netdev_port_poll_wait,
2369 dpif_netdev_flow_get,
2370 dpif_netdev_flow_put,
2371 dpif_netdev_flow_del,
2372 dpif_netdev_flow_flush,
2373 dpif_netdev_flow_dump_create,
2374 dpif_netdev_flow_dump_destroy,
2375 dpif_netdev_flow_dump_thread_create,
2376 dpif_netdev_flow_dump_thread_destroy,
2377 dpif_netdev_flow_dump_next,
2378 dpif_netdev_execute,
2380 dpif_netdev_recv_set,
2381 dpif_netdev_handlers_set,
2382 dpif_netdev_queue_to_priority,
2384 dpif_netdev_recv_wait,
2385 dpif_netdev_recv_purge,
2389 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2390 const char *argv[], void *aux OVS_UNUSED)
2392 struct dp_netdev_port *old_port;
2393 struct dp_netdev_port *new_port;
2394 struct dp_netdev *dp;
2397 ovs_mutex_lock(&dp_netdev_mutex);
2398 dp = shash_find_data(&dp_netdevs, argv[1]);
2399 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2400 ovs_mutex_unlock(&dp_netdev_mutex);
2401 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2404 ovs_refcount_ref(&dp->ref_cnt);
2405 ovs_mutex_unlock(&dp_netdev_mutex);
2407 ovs_mutex_lock(&dp->port_mutex);
2408 if (get_port_by_name(dp, argv[2], &old_port)) {
2409 unixctl_command_reply_error(conn, "unknown port");
2413 port_no = u32_to_odp(atoi(argv[3]));
2414 if (!port_no || port_no == ODPP_NONE) {
2415 unixctl_command_reply_error(conn, "bad port number");
2418 if (dp_netdev_lookup_port(dp, port_no)) {
2419 unixctl_command_reply_error(conn, "port number already in use");
2423 /* Remove old port. */
2424 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
2425 ovsrcu_postpone(free, old_port);
2427 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
2428 new_port = xmemdup(old_port, sizeof *old_port);
2429 new_port->port_no = port_no;
2430 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
2432 seq_change(dp->port_seq);
2433 unixctl_command_reply(conn, NULL);
2436 ovs_mutex_unlock(&dp->port_mutex);
2437 dp_netdev_unref(dp);
2441 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
2442 const char *argv[], void *aux OVS_UNUSED)
2444 struct dp_netdev_port *port;
2445 struct dp_netdev *dp;
2447 ovs_mutex_lock(&dp_netdev_mutex);
2448 dp = shash_find_data(&dp_netdevs, argv[1]);
2449 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2450 ovs_mutex_unlock(&dp_netdev_mutex);
2451 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2454 ovs_refcount_ref(&dp->ref_cnt);
2455 ovs_mutex_unlock(&dp_netdev_mutex);
2457 ovs_mutex_lock(&dp->port_mutex);
2458 if (get_port_by_name(dp, argv[2], &port)) {
2459 unixctl_command_reply_error(conn, "unknown port");
2460 } else if (port->port_no == ODPP_LOCAL) {
2461 unixctl_command_reply_error(conn, "can't delete local port");
2463 do_del_port(dp, port);
2464 unixctl_command_reply(conn, NULL);
2466 ovs_mutex_unlock(&dp->port_mutex);
2468 dp_netdev_unref(dp);
2472 dpif_dummy_register__(const char *type)
2474 struct dpif_class *class;
2476 class = xmalloc(sizeof *class);
2477 *class = dpif_netdev_class;
2478 class->type = xstrdup(type);
2479 dp_register_provider(class);
2483 dpif_dummy_register(bool override)
2490 dp_enumerate_types(&types);
2491 SSET_FOR_EACH (type, &types) {
2492 if (!dp_unregister_provider(type)) {
2493 dpif_dummy_register__(type);
2496 sset_destroy(&types);
2499 dpif_dummy_register__("dummy");
2501 unixctl_command_register("dpif-dummy/change-port-number",
2502 "DP PORT NEW-NUMBER",
2503 3, 3, dpif_dummy_change_port_number, NULL);
2504 unixctl_command_register("dpif-dummy/delete-port", "DP PORT",
2505 2, 2, dpif_dummy_delete_port, NULL);