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.
18 #include "dpif-netdev.h"
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
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. */
81 /* Protects against changes to 'dp_netdevs'. */
82 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
84 /* Contains all 'struct dp_netdev's. */
85 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
86 = SHASH_INITIALIZER(&dp_netdevs);
88 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
90 /* Datapath based on the network device interface from netdev.h.
96 * Some members, marked 'const', are immutable. Accessing other members
97 * requires synchronization, as noted in more detail below.
99 * Acquisition order is, from outermost to innermost:
101 * dp_netdev_mutex (global)
106 const struct dpif_class *const class;
107 const char *const name;
109 struct ovs_refcount ref_cnt;
110 atomic_flag destroyed;
114 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
115 * changes to 'cls' must be made while still holding the 'flow_mutex'.
117 struct ovs_mutex flow_mutex;
118 struct classifier cls;
119 struct cmap flow_table OVS_GUARDED; /* Flow table. */
123 * ovsthread_stats is internally synchronized. */
124 struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */
128 * Protected by RCU. Take the mutex to add or remove ports. */
129 struct ovs_mutex port_mutex;
131 struct seq *port_seq; /* Incremented whenever a port changes. */
133 /* Protects access to ofproto-dpif-upcall interface during revalidator
134 * thread synchronization. */
135 struct fat_rwlock upcall_rwlock;
136 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
139 /* Forwarding threads. */
140 struct latch exit_latch;
141 struct pmd_thread *pmd_threads;
142 size_t n_pmd_threads;
146 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
150 DP_STAT_HIT, /* Packets that matched in the flow table. */
151 DP_STAT_MISS, /* Packets that did not match. */
152 DP_STAT_LOST, /* Packets not passed up to the client. */
156 /* Contained by struct dp_netdev's 'stats' member. */
157 struct dp_netdev_stats {
158 struct ovs_mutex mutex; /* Protects 'n'. */
160 /* Indexed by DP_STAT_*, protected by 'mutex'. */
161 unsigned long long int n[DP_N_STATS] OVS_GUARDED;
165 /* A port in a netdev-based datapath. */
166 struct dp_netdev_port {
167 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
169 struct netdev *netdev;
170 struct netdev_saved_flags *sf;
171 struct netdev_rxq **rxq;
172 struct ovs_refcount ref_cnt;
173 char *type; /* Port type as requested by user. */
177 /* Stores a miniflow */
179 /* There are fields in the flow structure that we never use. Therefore we can
180 * save a few words of memory */
181 #define NETDEV_KEY_BUF_SIZE_U32 (FLOW_U32S - MINI_N_INLINE \
182 - FLOW_U32_SIZE(regs) \
183 - FLOW_U32_SIZE(metadata) \
185 struct netdev_flow_key {
186 struct miniflow flow;
187 uint32_t buf[NETDEV_KEY_BUF_SIZE_U32];
190 /* A flow in dp_netdev's 'flow_table'.
196 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
197 * its dp_netdev's classifier. The text below calls this classifier 'cls'.
202 * The thread safety rules described here for "struct dp_netdev_flow" are
203 * motivated by two goals:
205 * - Prevent threads that read members of "struct dp_netdev_flow" from
206 * reading bad data due to changes by some thread concurrently modifying
209 * - Prevent two threads making changes to members of a given "struct
210 * dp_netdev_flow" from interfering with each other.
216 * A flow 'flow' may be accessed without a risk of being freed during an RCU
217 * grace period. Code that needs to hold onto a flow for a while
218 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
220 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
221 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
224 * Some members, marked 'const', are immutable. Accessing other members
225 * requires synchronization, as noted in more detail below.
227 struct dp_netdev_flow {
228 /* Packet classification. */
229 const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
231 /* Hash table index by unmasked flow. */
232 const struct cmap_node node; /* In owning dp_netdev's 'flow_table'. */
233 const struct flow flow; /* The flow that created this entry. */
235 /* Number of references.
236 * The classifier owns one reference.
237 * Any thread trying to keep a rule from being freed should hold its own
239 struct ovs_refcount ref_cnt;
243 * Reading or writing these members requires 'mutex'. */
244 struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
247 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
250 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
252 /* Contained by struct dp_netdev_flow's 'stats' member. */
253 struct dp_netdev_flow_stats {
254 struct ovs_mutex mutex; /* Guards all the other members. */
256 long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
257 long long int packet_count OVS_GUARDED; /* Number of packets matched. */
258 long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
259 uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
262 /* A set of datapath actions within a "struct dp_netdev_flow".
268 * A struct dp_netdev_actions 'actions' is protected with RCU. */
269 struct dp_netdev_actions {
270 /* These members are immutable: they do not change during the struct's
272 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
273 unsigned int size; /* Size of 'actions', in bytes. */
276 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
278 struct dp_netdev_actions *dp_netdev_flow_get_actions(
279 const struct dp_netdev_flow *);
280 static void dp_netdev_actions_free(struct dp_netdev_actions *);
282 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
283 * the performance overhead of interrupt processing. Therefore netdev can
284 * not implement rx-wait for these devices. dpif-netdev needs to poll
285 * these device to check for recv buffer. pmd-thread does polling for
286 * devices assigned to itself thread.
288 * DPDK used PMD for accessing NIC.
290 * A thread that receives packets from PMD ports, looks them up in the flow
291 * table, and executes the actions it finds.
294 struct dp_netdev *dp;
297 atomic_uint change_seq;
300 /* Interface to netdev-based datapath. */
303 struct dp_netdev *dp;
304 uint64_t last_port_seq;
307 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
308 struct dp_netdev_port **portp);
309 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
310 struct dp_netdev_port **portp);
311 static void dp_netdev_free(struct dp_netdev *)
312 OVS_REQUIRES(dp_netdev_mutex);
313 static void dp_netdev_flow_flush(struct dp_netdev *);
314 static int do_add_port(struct dp_netdev *dp, const char *devname,
315 const char *type, odp_port_t port_no)
316 OVS_REQUIRES(dp->port_mutex);
317 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
318 OVS_REQUIRES(dp->port_mutex);
319 static int dpif_netdev_open(const struct dpif_class *, const char *name,
320 bool create, struct dpif **);
321 static void dp_netdev_execute_actions(struct dp_netdev *dp,
322 struct dpif_packet **, int c,
323 bool may_steal, struct pkt_metadata *,
324 const struct nlattr *actions,
326 static void dp_netdev_port_input(struct dp_netdev *dp,
327 struct dpif_packet **packets, int cnt,
330 static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
331 static void dp_netdev_disable_upcall(struct dp_netdev *);
333 static struct dpif_netdev *
334 dpif_netdev_cast(const struct dpif *dpif)
336 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
337 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
340 static struct dp_netdev *
341 get_dp_netdev(const struct dpif *dpif)
343 return dpif_netdev_cast(dpif)->dp;
347 dpif_netdev_enumerate(struct sset *all_dps,
348 const struct dpif_class *dpif_class)
350 struct shash_node *node;
352 ovs_mutex_lock(&dp_netdev_mutex);
353 SHASH_FOR_EACH(node, &dp_netdevs) {
354 struct dp_netdev *dp = node->data;
355 if (dpif_class != dp->class) {
356 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
357 * If the class doesn't match, skip this dpif. */
360 sset_add(all_dps, node->name);
362 ovs_mutex_unlock(&dp_netdev_mutex);
368 dpif_netdev_class_is_dummy(const struct dpif_class *class)
370 return class != &dpif_netdev_class;
374 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
376 return strcmp(type, "internal") ? type
377 : dpif_netdev_class_is_dummy(class) ? "dummy"
382 create_dpif_netdev(struct dp_netdev *dp)
384 uint16_t netflow_id = hash_string(dp->name, 0);
385 struct dpif_netdev *dpif;
387 ovs_refcount_ref(&dp->ref_cnt);
389 dpif = xmalloc(sizeof *dpif);
390 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
392 dpif->last_port_seq = seq_read(dp->port_seq);
397 /* Choose an unused, non-zero port number and return it on success.
398 * Return ODPP_NONE on failure. */
400 choose_port(struct dp_netdev *dp, const char *name)
401 OVS_REQUIRES(dp->port_mutex)
405 if (dp->class != &dpif_netdev_class) {
409 /* If the port name begins with "br", start the number search at
410 * 100 to make writing tests easier. */
411 if (!strncmp(name, "br", 2)) {
415 /* If the port name contains a number, try to assign that port number.
416 * This can make writing unit tests easier because port numbers are
418 for (p = name; *p != '\0'; p++) {
419 if (isdigit((unsigned char) *p)) {
420 port_no = start_no + strtol(p, NULL, 10);
421 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
422 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
423 return u32_to_odp(port_no);
430 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
431 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
432 return u32_to_odp(port_no);
440 create_dp_netdev(const char *name, const struct dpif_class *class,
441 struct dp_netdev **dpp)
442 OVS_REQUIRES(dp_netdev_mutex)
444 struct dp_netdev *dp;
447 dp = xzalloc(sizeof *dp);
448 shash_add(&dp_netdevs, name, dp);
450 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
451 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
452 ovs_refcount_init(&dp->ref_cnt);
453 atomic_flag_clear(&dp->destroyed);
455 ovs_mutex_init(&dp->flow_mutex);
456 classifier_init(&dp->cls, NULL);
457 cmap_init(&dp->flow_table);
459 ovsthread_stats_init(&dp->stats);
461 ovs_mutex_init(&dp->port_mutex);
462 cmap_init(&dp->ports);
463 dp->port_seq = seq_create();
464 latch_init(&dp->exit_latch);
465 fat_rwlock_init(&dp->upcall_rwlock);
467 /* Disable upcalls by default. */
468 dp_netdev_disable_upcall(dp);
469 dp->upcall_aux = NULL;
470 dp->upcall_cb = NULL;
472 ovs_mutex_lock(&dp->port_mutex);
473 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
474 ovs_mutex_unlock(&dp->port_mutex);
485 dpif_netdev_open(const struct dpif_class *class, const char *name,
486 bool create, struct dpif **dpifp)
488 struct dp_netdev *dp;
491 ovs_mutex_lock(&dp_netdev_mutex);
492 dp = shash_find_data(&dp_netdevs, name);
494 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
496 error = (dp->class != class ? EINVAL
501 *dpifp = create_dpif_netdev(dp);
504 ovs_mutex_unlock(&dp_netdev_mutex);
509 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
510 * through the 'dp_netdevs' shash while freeing 'dp'. */
512 dp_netdev_free(struct dp_netdev *dp)
513 OVS_REQUIRES(dp_netdev_mutex)
515 struct dp_netdev_port *port;
516 struct dp_netdev_stats *bucket;
519 shash_find_and_delete(&dp_netdevs, dp->name);
521 dp_netdev_set_pmd_threads(dp, 0);
522 free(dp->pmd_threads);
524 dp_netdev_flow_flush(dp);
525 ovs_mutex_lock(&dp->port_mutex);
526 CMAP_FOR_EACH (port, node, &dp->ports) {
527 do_del_port(dp, port);
529 ovs_mutex_unlock(&dp->port_mutex);
531 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
532 ovs_mutex_destroy(&bucket->mutex);
533 free_cacheline(bucket);
535 ovsthread_stats_destroy(&dp->stats);
537 classifier_destroy(&dp->cls);
538 cmap_destroy(&dp->flow_table);
539 ovs_mutex_destroy(&dp->flow_mutex);
540 seq_destroy(dp->port_seq);
541 cmap_destroy(&dp->ports);
542 fat_rwlock_destroy(&dp->upcall_rwlock);
543 latch_destroy(&dp->exit_latch);
544 free(CONST_CAST(char *, dp->name));
549 dp_netdev_unref(struct dp_netdev *dp)
552 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
553 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
554 ovs_mutex_lock(&dp_netdev_mutex);
555 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
558 ovs_mutex_unlock(&dp_netdev_mutex);
563 dpif_netdev_close(struct dpif *dpif)
565 struct dp_netdev *dp = get_dp_netdev(dpif);
572 dpif_netdev_destroy(struct dpif *dpif)
574 struct dp_netdev *dp = get_dp_netdev(dpif);
576 if (!atomic_flag_test_and_set(&dp->destroyed)) {
577 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
578 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
587 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
589 struct dp_netdev *dp = get_dp_netdev(dpif);
590 struct dp_netdev_stats *bucket;
593 stats->n_flows = cmap_count(&dp->flow_table);
595 stats->n_hit = stats->n_missed = stats->n_lost = 0;
596 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
597 ovs_mutex_lock(&bucket->mutex);
598 stats->n_hit += bucket->n[DP_STAT_HIT];
599 stats->n_missed += bucket->n[DP_STAT_MISS];
600 stats->n_lost += bucket->n[DP_STAT_LOST];
601 ovs_mutex_unlock(&bucket->mutex);
603 stats->n_masks = UINT32_MAX;
604 stats->n_mask_hit = UINT64_MAX;
610 dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
614 for (i = 0; i < dp->n_pmd_threads; i++) {
615 struct pmd_thread *f = &dp->pmd_threads[i];
618 atomic_add(&f->change_seq, 1, &id);
623 hash_port_no(odp_port_t port_no)
625 return hash_int(odp_to_u32(port_no), 0);
629 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
631 OVS_REQUIRES(dp->port_mutex)
633 struct netdev_saved_flags *sf;
634 struct dp_netdev_port *port;
635 struct netdev *netdev;
636 enum netdev_flags flags;
637 const char *open_type;
641 /* XXX reject devices already in some dp_netdev. */
643 /* Open and validate network device. */
644 open_type = dpif_netdev_port_open_type(dp->class, type);
645 error = netdev_open(devname, open_type, &netdev);
649 /* XXX reject non-Ethernet devices */
651 netdev_get_flags(netdev, &flags);
652 if (flags & NETDEV_LOOPBACK) {
653 VLOG_ERR("%s: cannot add a loopback device", devname);
654 netdev_close(netdev);
658 port = xzalloc(sizeof *port);
659 port->port_no = port_no;
660 port->netdev = netdev;
661 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
662 port->type = xstrdup(type);
663 for (i = 0; i < netdev_n_rxq(netdev); i++) {
664 error = netdev_rxq_open(netdev, &port->rxq[i], i);
666 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
667 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
668 devname, ovs_strerror(errno));
669 netdev_close(netdev);
674 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
676 for (i = 0; i < netdev_n_rxq(netdev); i++) {
677 netdev_rxq_close(port->rxq[i]);
679 netdev_close(netdev);
686 if (netdev_is_pmd(netdev)) {
688 dp_netdev_set_pmd_threads(dp, NR_PMD_THREADS);
689 dp_netdev_reload_pmd_threads(dp);
691 ovs_refcount_init(&port->ref_cnt);
693 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
694 seq_change(dp->port_seq);
700 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
701 odp_port_t *port_nop)
703 struct dp_netdev *dp = get_dp_netdev(dpif);
704 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
705 const char *dpif_port;
709 ovs_mutex_lock(&dp->port_mutex);
710 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
711 if (*port_nop != ODPP_NONE) {
713 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
715 port_no = choose_port(dp, dpif_port);
716 error = port_no == ODPP_NONE ? EFBIG : 0;
720 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
722 ovs_mutex_unlock(&dp->port_mutex);
728 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
730 struct dp_netdev *dp = get_dp_netdev(dpif);
733 ovs_mutex_lock(&dp->port_mutex);
734 if (port_no == ODPP_LOCAL) {
737 struct dp_netdev_port *port;
739 error = get_port_by_number(dp, port_no, &port);
741 do_del_port(dp, port);
744 ovs_mutex_unlock(&dp->port_mutex);
750 is_valid_port_number(odp_port_t port_no)
752 return port_no != ODPP_NONE;
755 static struct dp_netdev_port *
756 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
758 struct dp_netdev_port *port;
760 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
761 if (port->port_no == port_no) {
769 get_port_by_number(struct dp_netdev *dp,
770 odp_port_t port_no, struct dp_netdev_port **portp)
772 if (!is_valid_port_number(port_no)) {
776 *portp = dp_netdev_lookup_port(dp, port_no);
777 return *portp ? 0 : ENOENT;
782 port_ref(struct dp_netdev_port *port)
785 ovs_refcount_ref(&port->ref_cnt);
790 port_destroy__(struct dp_netdev_port *port)
792 int n_rxq = netdev_n_rxq(port->netdev);
795 netdev_close(port->netdev);
796 netdev_restore_flags(port->sf);
798 for (i = 0; i < n_rxq; i++) {
799 netdev_rxq_close(port->rxq[i]);
807 port_unref(struct dp_netdev_port *port)
809 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
810 ovsrcu_postpone(port_destroy__, port);
815 get_port_by_name(struct dp_netdev *dp,
816 const char *devname, struct dp_netdev_port **portp)
817 OVS_REQUIRES(dp->port_mutex)
819 struct dp_netdev_port *port;
821 CMAP_FOR_EACH (port, node, &dp->ports) {
822 if (!strcmp(netdev_get_name(port->netdev), devname)) {
831 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
832 OVS_REQUIRES(dp->port_mutex)
834 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
835 seq_change(dp->port_seq);
836 if (netdev_is_pmd(port->netdev)) {
837 dp_netdev_reload_pmd_threads(dp);
844 answer_port_query(const struct dp_netdev_port *port,
845 struct dpif_port *dpif_port)
847 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
848 dpif_port->type = xstrdup(port->type);
849 dpif_port->port_no = port->port_no;
853 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
854 struct dpif_port *dpif_port)
856 struct dp_netdev *dp = get_dp_netdev(dpif);
857 struct dp_netdev_port *port;
860 error = get_port_by_number(dp, port_no, &port);
861 if (!error && dpif_port) {
862 answer_port_query(port, dpif_port);
869 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
870 struct dpif_port *dpif_port)
872 struct dp_netdev *dp = get_dp_netdev(dpif);
873 struct dp_netdev_port *port;
876 ovs_mutex_lock(&dp->port_mutex);
877 error = get_port_by_name(dp, devname, &port);
878 if (!error && dpif_port) {
879 answer_port_query(port, dpif_port);
881 ovs_mutex_unlock(&dp->port_mutex);
887 dp_netdev_flow_free(struct dp_netdev_flow *flow)
889 struct dp_netdev_flow_stats *bucket;
892 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
893 ovs_mutex_destroy(&bucket->mutex);
894 free_cacheline(bucket);
896 ovsthread_stats_destroy(&flow->stats);
898 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
899 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
903 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
905 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
906 ovsrcu_postpone(dp_netdev_flow_free, flow);
911 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
912 OVS_REQUIRES(dp->flow_mutex)
914 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
915 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
917 classifier_remove(&dp->cls, cr);
918 cmap_remove(&dp->flow_table, node, flow_hash(&flow->flow, 0));
920 dp_netdev_flow_unref(flow);
924 dp_netdev_flow_flush(struct dp_netdev *dp)
926 struct dp_netdev_flow *netdev_flow;
928 ovs_mutex_lock(&dp->flow_mutex);
929 CMAP_FOR_EACH (netdev_flow, node, &dp->flow_table) {
930 dp_netdev_remove_flow(dp, netdev_flow);
932 ovs_mutex_unlock(&dp->flow_mutex);
936 dpif_netdev_flow_flush(struct dpif *dpif)
938 struct dp_netdev *dp = get_dp_netdev(dpif);
940 dp_netdev_flow_flush(dp);
944 struct dp_netdev_port_state {
945 struct cmap_position position;
950 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
952 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
957 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
958 struct dpif_port *dpif_port)
960 struct dp_netdev_port_state *state = state_;
961 struct dp_netdev *dp = get_dp_netdev(dpif);
962 struct cmap_node *node;
965 node = cmap_next_position(&dp->ports, &state->position);
967 struct dp_netdev_port *port;
969 port = CONTAINER_OF(node, struct dp_netdev_port, node);
972 state->name = xstrdup(netdev_get_name(port->netdev));
973 dpif_port->name = state->name;
974 dpif_port->type = port->type;
975 dpif_port->port_no = port->port_no;
986 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
988 struct dp_netdev_port_state *state = state_;
995 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
997 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
998 uint64_t new_port_seq;
1001 new_port_seq = seq_read(dpif->dp->port_seq);
1002 if (dpif->last_port_seq != new_port_seq) {
1003 dpif->last_port_seq = new_port_seq;
1013 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1015 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1017 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1020 static struct dp_netdev_flow *
1021 dp_netdev_flow_cast(const struct cls_rule *cr)
1023 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1026 static struct dp_netdev_flow *
1027 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
1029 struct dp_netdev_flow *netdev_flow;
1030 struct cls_rule *rule;
1032 classifier_lookup_miniflow_batch(&dp->cls, &key, &rule, 1);
1033 netdev_flow = dp_netdev_flow_cast(rule);
1038 static struct dp_netdev_flow *
1039 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1041 struct dp_netdev_flow *netdev_flow;
1043 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1045 if (flow_equal(&netdev_flow->flow, flow)) {
1054 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow,
1055 struct dpif_flow_stats *stats)
1057 struct dp_netdev_flow_stats *bucket;
1060 memset(stats, 0, sizeof *stats);
1061 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1062 ovs_mutex_lock(&bucket->mutex);
1063 stats->n_packets += bucket->packet_count;
1064 stats->n_bytes += bucket->byte_count;
1065 stats->used = MAX(stats->used, bucket->used);
1066 stats->tcp_flags |= bucket->tcp_flags;
1067 ovs_mutex_unlock(&bucket->mutex);
1072 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1073 struct ofpbuf *buffer, struct dpif_flow *flow)
1075 struct flow_wildcards wc;
1076 struct dp_netdev_actions *actions;
1078 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1079 odp_flow_key_from_mask(buffer, &wc.masks, &netdev_flow->flow,
1080 odp_to_u32(wc.masks.in_port.odp_port),
1082 flow->mask = ofpbuf_data(buffer);
1083 flow->mask_len = ofpbuf_size(buffer);
1085 actions = dp_netdev_flow_get_actions(netdev_flow);
1086 flow->actions = actions->actions;
1087 flow->actions_len = actions->size;
1089 get_dpif_flow_stats(netdev_flow, &flow->stats);
1093 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1094 const struct nlattr *mask_key,
1095 uint32_t mask_key_len, const struct flow *flow,
1099 enum odp_key_fitness fitness;
1101 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1103 /* This should not happen: it indicates that
1104 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1105 * disagree on the acceptable form of a mask. Log the problem
1106 * as an error, with enough details to enable debugging. */
1107 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1109 if (!VLOG_DROP_ERR(&rl)) {
1113 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1115 VLOG_ERR("internal error parsing flow mask %s (%s)",
1116 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1123 enum mf_field_id id;
1124 /* No mask key, unwildcard everything except fields whose
1125 * prerequisities are not met. */
1126 memset(mask, 0x0, sizeof *mask);
1128 for (id = 0; id < MFF_N_IDS; ++id) {
1129 /* Skip registers and metadata. */
1130 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1131 && id != MFF_METADATA) {
1132 const struct mf_field *mf = mf_from_id(id);
1133 if (mf_are_prereqs_ok(mf, flow)) {
1134 mf_mask_field(mf, mask);
1140 /* Force unwildcard the in_port.
1142 * We need to do this even in the case where we unwildcard "everything"
1143 * above because "everything" only includes the 16-bit OpenFlow port number
1144 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1145 * port number mask->in_port.odp_port. */
1146 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1152 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1157 if (odp_flow_key_to_flow(key, key_len, flow)) {
1158 /* This should not happen: it indicates that odp_flow_key_from_flow()
1159 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1160 * flow. Log the problem as an error, with enough details to enable
1162 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1164 if (!VLOG_DROP_ERR(&rl)) {
1168 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1169 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1176 in_port = flow->in_port.odp_port;
1177 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1185 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1187 struct dp_netdev *dp = get_dp_netdev(dpif);
1188 struct dp_netdev_flow *netdev_flow;
1192 error = dpif_netdev_flow_from_nlattrs(get->key, get->key_len, &key);
1197 netdev_flow = dp_netdev_find_flow(dp, &key);
1200 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->flow);
1209 dp_netdev_flow_add(struct dp_netdev *dp, struct match *match,
1210 const struct nlattr *actions, size_t actions_len)
1211 OVS_REQUIRES(dp->flow_mutex)
1213 struct dp_netdev_flow *netdev_flow;
1215 netdev_flow = xzalloc(sizeof *netdev_flow);
1216 *CONST_CAST(struct flow *, &netdev_flow->flow) = match->flow;
1218 ovs_refcount_init(&netdev_flow->ref_cnt);
1220 ovsthread_stats_init(&netdev_flow->stats);
1222 ovsrcu_set(&netdev_flow->actions,
1223 dp_netdev_actions_create(actions, actions_len));
1225 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1226 match, NETDEV_RULE_PRIORITY);
1227 cmap_insert(&dp->flow_table,
1228 CONST_CAST(struct cmap_node *, &netdev_flow->node),
1229 flow_hash(&match->flow, 0));
1230 classifier_insert(&dp->cls,
1231 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1233 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1234 struct ds ds = DS_EMPTY_INITIALIZER;
1236 ds_put_cstr(&ds, "flow_add: ");
1237 match_format(match, &ds, OFP_DEFAULT_PRIORITY);
1238 ds_put_cstr(&ds, ", actions:");
1239 format_odp_actions(&ds, actions, actions_len);
1241 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
1250 clear_stats(struct dp_netdev_flow *netdev_flow)
1252 struct dp_netdev_flow_stats *bucket;
1255 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1256 ovs_mutex_lock(&bucket->mutex);
1258 bucket->packet_count = 0;
1259 bucket->byte_count = 0;
1260 bucket->tcp_flags = 0;
1261 ovs_mutex_unlock(&bucket->mutex);
1266 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1268 struct dp_netdev *dp = get_dp_netdev(dpif);
1269 struct dp_netdev_flow *netdev_flow;
1270 struct miniflow miniflow;
1274 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
1278 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1279 put->mask, put->mask_len,
1280 &match.flow, &match.wc.masks);
1284 miniflow_init(&miniflow, &match.flow);
1286 ovs_mutex_lock(&dp->flow_mutex);
1287 netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
1289 if (put->flags & DPIF_FP_CREATE) {
1290 if (cmap_count(&dp->flow_table) < MAX_FLOWS) {
1292 memset(put->stats, 0, sizeof *put->stats);
1294 error = dp_netdev_flow_add(dp, &match, put->actions,
1303 if (put->flags & DPIF_FP_MODIFY
1304 && flow_equal(&match.flow, &netdev_flow->flow)) {
1305 struct dp_netdev_actions *new_actions;
1306 struct dp_netdev_actions *old_actions;
1308 new_actions = dp_netdev_actions_create(put->actions,
1311 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1312 ovsrcu_set(&netdev_flow->actions, new_actions);
1315 get_dpif_flow_stats(netdev_flow, put->stats);
1317 if (put->flags & DPIF_FP_ZERO_STATS) {
1318 clear_stats(netdev_flow);
1321 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1322 } else if (put->flags & DPIF_FP_CREATE) {
1325 /* Overlapping flow. */
1329 ovs_mutex_unlock(&dp->flow_mutex);
1330 miniflow_destroy(&miniflow);
1336 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1338 struct dp_netdev *dp = get_dp_netdev(dpif);
1339 struct dp_netdev_flow *netdev_flow;
1343 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1348 ovs_mutex_lock(&dp->flow_mutex);
1349 netdev_flow = dp_netdev_find_flow(dp, &key);
1352 get_dpif_flow_stats(netdev_flow, del->stats);
1354 dp_netdev_remove_flow(dp, netdev_flow);
1358 ovs_mutex_unlock(&dp->flow_mutex);
1363 struct dpif_netdev_flow_dump {
1364 struct dpif_flow_dump up;
1365 struct cmap_position pos;
1367 struct ovs_mutex mutex;
1370 static struct dpif_netdev_flow_dump *
1371 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1373 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1376 static struct dpif_flow_dump *
1377 dpif_netdev_flow_dump_create(const struct dpif *dpif_)
1379 struct dpif_netdev_flow_dump *dump;
1381 dump = xmalloc(sizeof *dump);
1382 dpif_flow_dump_init(&dump->up, dpif_);
1383 memset(&dump->pos, 0, sizeof dump->pos);
1385 ovs_mutex_init(&dump->mutex);
1391 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1393 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1395 ovs_mutex_destroy(&dump->mutex);
1400 struct dpif_netdev_flow_dump_thread {
1401 struct dpif_flow_dump_thread up;
1402 struct dpif_netdev_flow_dump *dump;
1403 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1404 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1407 static struct dpif_netdev_flow_dump_thread *
1408 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1410 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1413 static struct dpif_flow_dump_thread *
1414 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1416 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1417 struct dpif_netdev_flow_dump_thread *thread;
1419 thread = xmalloc(sizeof *thread);
1420 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1421 thread->dump = dump;
1426 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1428 struct dpif_netdev_flow_dump_thread *thread
1429 = dpif_netdev_flow_dump_thread_cast(thread_);
1435 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
1436 struct dpif_flow *flows, int max_flows)
1438 struct dpif_netdev_flow_dump_thread *thread
1439 = dpif_netdev_flow_dump_thread_cast(thread_);
1440 struct dpif_netdev_flow_dump *dump = thread->dump;
1441 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
1442 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
1443 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
1447 ovs_mutex_lock(&dump->mutex);
1448 if (!dump->status) {
1449 for (n_flows = 0; n_flows < MIN(max_flows, FLOW_DUMP_MAX_BATCH);
1451 struct cmap_node *node;
1453 node = cmap_next_position(&dp->flow_table, &dump->pos);
1458 netdev_flows[n_flows] = CONTAINER_OF(node, struct dp_netdev_flow,
1462 ovs_mutex_unlock(&dump->mutex);
1464 for (i = 0; i < n_flows; i++) {
1465 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
1466 struct odputil_keybuf *keybuf = &thread->keybuf[i];
1467 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
1468 struct dpif_flow *f = &flows[i];
1469 struct dp_netdev_actions *dp_actions;
1470 struct flow_wildcards wc;
1473 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1476 ofpbuf_use_stack(&buf, keybuf, sizeof *keybuf);
1477 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1478 netdev_flow->flow.in_port.odp_port, true);
1479 f->key = ofpbuf_data(&buf);
1480 f->key_len = ofpbuf_size(&buf);
1483 ofpbuf_use_stack(&buf, maskbuf, sizeof *maskbuf);
1484 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1485 odp_to_u32(wc.masks.in_port.odp_port),
1487 f->mask = ofpbuf_data(&buf);
1488 f->mask_len = ofpbuf_size(&buf);
1491 dp_actions = dp_netdev_flow_get_actions(netdev_flow);
1492 f->actions = dp_actions->actions;
1493 f->actions_len = dp_actions->size;
1496 get_dpif_flow_stats(netdev_flow, &f->stats);
1503 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1505 struct dp_netdev *dp = get_dp_netdev(dpif);
1506 struct dpif_packet packet, *pp;
1507 struct pkt_metadata *md = &execute->md;
1509 if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
1510 ofpbuf_size(execute->packet) > UINT16_MAX) {
1514 packet.ofpbuf = *execute->packet;
1517 dp_netdev_execute_actions(dp, &pp, 1, false, md,
1518 execute->actions, execute->actions_len);
1520 /* Even though may_steal is set to false, some actions could modify or
1521 * reallocate the ofpbuf memory. We need to pass those changes to the
1523 *execute->packet = packet.ofpbuf;
1529 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
1533 for (i = 0; i < n_ops; i++) {
1534 struct dpif_op *op = ops[i];
1537 case DPIF_OP_FLOW_PUT:
1538 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
1541 case DPIF_OP_FLOW_DEL:
1542 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
1545 case DPIF_OP_EXECUTE:
1546 op->error = dpif_netdev_execute(dpif, &op->u.execute);
1549 case DPIF_OP_FLOW_GET:
1550 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
1557 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1558 uint32_t queue_id, uint32_t *priority)
1560 *priority = queue_id;
1565 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1566 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1568 struct dp_netdev_actions *
1569 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1571 struct dp_netdev_actions *netdev_actions;
1573 netdev_actions = xmalloc(sizeof *netdev_actions);
1574 netdev_actions->actions = xmemdup(actions, size);
1575 netdev_actions->size = size;
1577 return netdev_actions;
1580 struct dp_netdev_actions *
1581 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
1583 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
1587 dp_netdev_actions_free(struct dp_netdev_actions *actions)
1589 free(actions->actions);
1595 dp_netdev_process_rxq_port(struct dp_netdev *dp,
1596 struct dp_netdev_port *port,
1597 struct netdev_rxq *rxq)
1599 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1602 error = netdev_rxq_recv(rxq, packets, &cnt);
1604 dp_netdev_port_input(dp, packets, cnt, port->port_no);
1605 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1606 static struct vlog_rate_limit rl
1607 = VLOG_RATE_LIMIT_INIT(1, 5);
1609 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1610 netdev_get_name(port->netdev),
1611 ovs_strerror(error));
1616 dpif_netdev_run(struct dpif *dpif)
1618 struct dp_netdev_port *port;
1619 struct dp_netdev *dp = get_dp_netdev(dpif);
1621 CMAP_FOR_EACH (port, node, &dp->ports) {
1622 if (!netdev_is_pmd(port->netdev)) {
1625 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1626 dp_netdev_process_rxq_port(dp, port, port->rxq[i]);
1633 dpif_netdev_wait(struct dpif *dpif)
1635 struct dp_netdev_port *port;
1636 struct dp_netdev *dp = get_dp_netdev(dpif);
1638 ovs_mutex_lock(&dp_netdev_mutex);
1639 CMAP_FOR_EACH (port, node, &dp->ports) {
1640 if (!netdev_is_pmd(port->netdev)) {
1643 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1644 netdev_rxq_wait(port->rxq[i]);
1648 ovs_mutex_unlock(&dp_netdev_mutex);
1652 struct dp_netdev_port *port;
1653 struct netdev_rxq *rx;
1657 pmd_load_queues(struct pmd_thread *f,
1658 struct rxq_poll **ppoll_list, int poll_cnt)
1660 struct dp_netdev *dp = f->dp;
1661 struct rxq_poll *poll_list = *ppoll_list;
1662 struct dp_netdev_port *port;
1667 /* Simple scheduler for netdev rx polling. */
1668 for (i = 0; i < poll_cnt; i++) {
1669 port_unref(poll_list[i].port);
1675 CMAP_FOR_EACH (port, node, &f->dp->ports) {
1676 if (netdev_is_pmd(port->netdev)) {
1679 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1680 if ((index % dp->n_pmd_threads) == id) {
1681 poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1));
1684 poll_list[poll_cnt].port = port;
1685 poll_list[poll_cnt].rx = port->rxq[i];
1693 *ppoll_list = poll_list;
1698 pmd_thread_main(void *f_)
1700 struct pmd_thread *f = f_;
1701 struct dp_netdev *dp = f->dp;
1702 unsigned int lc = 0;
1703 struct rxq_poll *poll_list;
1704 unsigned int port_seq;
1711 pmd_thread_setaffinity_cpu(f->id);
1713 poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
1714 atomic_read(&f->change_seq, &port_seq);
1717 unsigned int c_port_seq;
1720 for (i = 0; i < poll_cnt; i++) {
1721 dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx);
1727 /* XXX: need completely userspace based signaling method.
1728 * to keep this thread entirely in userspace.
1729 * For now using atomic counter. */
1731 atomic_read_explicit(&f->change_seq, &c_port_seq, memory_order_consume);
1732 if (c_port_seq != port_seq) {
1738 if (!latch_is_set(&f->dp->exit_latch)){
1742 for (i = 0; i < poll_cnt; i++) {
1743 port_unref(poll_list[i].port);
1751 dp_netdev_disable_upcall(struct dp_netdev *dp)
1752 OVS_ACQUIRES(dp->upcall_rwlock)
1754 fat_rwlock_wrlock(&dp->upcall_rwlock);
1758 dpif_netdev_disable_upcall(struct dpif *dpif)
1759 OVS_NO_THREAD_SAFETY_ANALYSIS
1761 struct dp_netdev *dp = get_dp_netdev(dpif);
1762 dp_netdev_disable_upcall(dp);
1766 dp_netdev_enable_upcall(struct dp_netdev *dp)
1767 OVS_RELEASES(dp->upcall_rwlock)
1769 fat_rwlock_unlock(&dp->upcall_rwlock);
1773 dpif_netdev_enable_upcall(struct dpif *dpif)
1774 OVS_NO_THREAD_SAFETY_ANALYSIS
1776 struct dp_netdev *dp = get_dp_netdev(dpif);
1777 dp_netdev_enable_upcall(dp);
1781 dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n)
1785 if (n == dp->n_pmd_threads) {
1789 /* Stop existing threads. */
1790 latch_set(&dp->exit_latch);
1791 dp_netdev_reload_pmd_threads(dp);
1792 for (i = 0; i < dp->n_pmd_threads; i++) {
1793 struct pmd_thread *f = &dp->pmd_threads[i];
1795 xpthread_join(f->thread, NULL);
1797 latch_poll(&dp->exit_latch);
1798 free(dp->pmd_threads);
1800 /* Start new threads. */
1801 dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
1802 dp->n_pmd_threads = n;
1804 for (i = 0; i < n; i++) {
1805 struct pmd_thread *f = &dp->pmd_threads[i];
1809 atomic_store(&f->change_seq, 1);
1811 /* Each thread will distribute all devices rx-queues among
1813 f->thread = ovs_thread_create("pmd", pmd_thread_main, f);
1819 dp_netdev_flow_stats_new_cb(void)
1821 struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1822 ovs_mutex_init(&bucket->mutex);
1827 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
1831 long long int now = time_msec();
1832 struct dp_netdev_flow_stats *bucket;
1834 bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
1835 dp_netdev_flow_stats_new_cb);
1837 ovs_mutex_lock(&bucket->mutex);
1838 bucket->used = MAX(now, bucket->used);
1839 bucket->packet_count += cnt;
1840 bucket->byte_count += size;
1841 bucket->tcp_flags |= tcp_flags;
1842 ovs_mutex_unlock(&bucket->mutex);
1846 dp_netdev_stats_new_cb(void)
1848 struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1849 ovs_mutex_init(&bucket->mutex);
1854 dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type, int cnt)
1856 struct dp_netdev_stats *bucket;
1858 bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
1859 ovs_mutex_lock(&bucket->mutex);
1860 bucket->n[type] += cnt;
1861 ovs_mutex_unlock(&bucket->mutex);
1865 dp_netdev_upcall(struct dp_netdev *dp, struct dpif_packet *packet_,
1866 struct flow *flow, struct flow_wildcards *wc,
1867 enum dpif_upcall_type type, const struct nlattr *userdata,
1868 struct ofpbuf *actions, struct ofpbuf *put_actions)
1870 struct ofpbuf *packet = &packet_->ofpbuf;
1872 if (type == DPIF_UC_MISS) {
1873 dp_netdev_count_packet(dp, DP_STAT_MISS, 1);
1876 if (OVS_UNLIKELY(!dp->upcall_cb)) {
1880 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
1881 struct ds ds = DS_EMPTY_INITIALIZER;
1885 ofpbuf_init(&key, 0);
1886 odp_flow_key_from_flow(&key, flow, &wc->masks, flow->in_port.odp_port,
1889 packet_str = ofp_packet_to_string(ofpbuf_data(packet),
1890 ofpbuf_size(packet));
1892 odp_flow_key_format(ofpbuf_data(&key), ofpbuf_size(&key), &ds);
1894 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
1895 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
1897 ofpbuf_uninit(&key);
1902 return dp->upcall_cb(packet, flow, type, userdata, actions, wc,
1903 put_actions, dp->upcall_aux);
1906 struct packet_batch {
1907 unsigned int packet_count;
1908 unsigned int byte_count;
1911 struct dp_netdev_flow *flow;
1913 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1914 struct pkt_metadata md;
1918 packet_batch_update(struct packet_batch *batch,
1919 struct dpif_packet *packet, const struct miniflow *mf)
1921 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
1922 batch->packets[batch->packet_count++] = packet;
1923 batch->byte_count += ofpbuf_size(&packet->ofpbuf);
1927 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow,
1928 struct pkt_metadata *md)
1933 batch->packet_count = 0;
1934 batch->byte_count = 0;
1935 batch->tcp_flags = 0;
1939 packet_batch_execute(struct packet_batch *batch, struct dp_netdev *dp)
1941 struct dp_netdev_actions *actions;
1942 struct dp_netdev_flow *flow = batch->flow;
1944 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
1947 actions = dp_netdev_flow_get_actions(flow);
1949 dp_netdev_execute_actions(dp, batch->packets,
1950 batch->packet_count, true, &batch->md,
1951 actions->actions, actions->size);
1953 dp_netdev_count_packet(dp, DP_STAT_HIT, batch->packet_count);
1957 dp_netdev_input(struct dp_netdev *dp, struct dpif_packet **packets, int cnt,
1958 struct pkt_metadata *md)
1960 struct packet_batch batches[NETDEV_MAX_RX_BATCH];
1961 struct netdev_flow_key keys[NETDEV_MAX_RX_BATCH];
1962 const struct miniflow *mfs[NETDEV_MAX_RX_BATCH]; /* NULL at bad packets. */
1963 struct cls_rule *rules[NETDEV_MAX_RX_BATCH];
1964 size_t n_batches, i;
1967 for (i = 0; i < cnt; i++) {
1968 if (OVS_UNLIKELY(ofpbuf_size(&packets[i]->ofpbuf) < ETH_HEADER_LEN)) {
1969 dpif_packet_delete(packets[i]);
1974 miniflow_initialize(&keys[i].flow, keys[i].buf);
1975 miniflow_extract(&packets[i]->ofpbuf, md, &keys[i].flow);
1976 mfs[i] = &keys[i].flow;
1979 any_miss = !classifier_lookup_miniflow_batch(&dp->cls, mfs, rules, cnt);
1980 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
1981 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
1982 struct ofpbuf actions, put_actions;
1985 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
1986 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
1988 for (i = 0; i < cnt; i++) {
1989 const struct dp_netdev_flow *netdev_flow;
1990 struct ofpbuf *add_actions;
1993 if (OVS_LIKELY(rules[i] || !mfs[i])) {
1997 /* It's possible that an earlier slow path execution installed
1998 * the rule this flow needs. In this case, it's a lot cheaper
1999 * to catch it here than execute a miss. */
2000 netdev_flow = dp_netdev_lookup_flow(dp, mfs[i]);
2002 rules[i] = CONST_CAST(struct cls_rule *, &netdev_flow->cr);
2006 miniflow_expand(mfs[i], &match.flow);
2008 ofpbuf_clear(&actions);
2009 ofpbuf_clear(&put_actions);
2011 error = dp_netdev_upcall(dp, packets[i], &match.flow, &match.wc,
2012 DPIF_UC_MISS, NULL, &actions,
2014 if (OVS_UNLIKELY(error && error != ENOSPC)) {
2018 /* We can't allow the packet batching in the next loop to execute
2019 * the actions. Otherwise, if there are any slow path actions,
2020 * we'll send the packet up twice. */
2021 dp_netdev_execute_actions(dp, &packets[i], 1, false, md,
2022 ofpbuf_data(&actions),
2023 ofpbuf_size(&actions));
2025 add_actions = ofpbuf_size(&put_actions)
2029 ovs_mutex_lock(&dp->flow_mutex);
2030 /* XXX: There's a brief race where this flow could have already
2031 * been installed since we last did the flow lookup. This could be
2032 * solved by moving the mutex lock outside the loop, but that's an
2033 * awful long time to be locking everyone out of making flow
2034 * installs. If we move to a per-core classifier, it would be
2036 if (OVS_LIKELY(error != ENOSPC)
2037 && !dp_netdev_lookup_flow(dp, mfs[i])) {
2038 dp_netdev_flow_add(dp, &match, ofpbuf_data(add_actions),
2039 ofpbuf_size(add_actions));
2041 ovs_mutex_unlock(&dp->flow_mutex);
2044 ofpbuf_uninit(&actions);
2045 ofpbuf_uninit(&put_actions);
2046 fat_rwlock_unlock(&dp->upcall_rwlock);
2050 for (i = 0; i < cnt; i++) {
2051 struct dp_netdev_flow *flow;
2052 struct packet_batch *batch;
2055 if (OVS_UNLIKELY(!rules[i] || !mfs[i])) {
2059 /* XXX: This O(n^2) algortihm makes sense if we're operating under the
2060 * assumption that the number of distinct flows (and therefore the
2061 * number of distinct batches) is quite small. If this turns out not
2062 * to be the case, it may make sense to pre sort based on the
2063 * netdev_flow pointer. That done we can get the appropriate batching
2064 * in O(n * log(n)) instead. */
2066 flow = dp_netdev_flow_cast(rules[i]);
2067 for (j = 0; j < n_batches; j++) {
2068 if (batches[j].flow == flow) {
2069 batch = &batches[j];
2075 batch = &batches[n_batches++];
2076 packet_batch_init(batch, flow, md);
2078 packet_batch_update(batch, packets[i], mfs[i]);
2081 for (i = 0; i < n_batches; i++) {
2082 packet_batch_execute(&batches[i], dp);
2087 dp_netdev_port_input(struct dp_netdev *dp, struct dpif_packet **packets,
2088 int cnt, odp_port_t port_no)
2090 uint32_t *recirc_depth = recirc_depth_get();
2091 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port_no);
2094 dp_netdev_input(dp, packets, cnt, &md);
2097 struct dp_netdev_execute_aux {
2098 struct dp_netdev *dp;
2102 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
2105 struct dp_netdev *dp = get_dp_netdev(dpif);
2106 dp->upcall_aux = aux;
2111 dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
2112 struct pkt_metadata *md,
2113 const struct nlattr *a, bool may_steal)
2114 OVS_NO_THREAD_SAFETY_ANALYSIS
2116 struct dp_netdev_execute_aux *aux = aux_;
2117 uint32_t *depth = recirc_depth_get();
2118 struct dp_netdev *dp = aux->dp;
2119 int type = nl_attr_type(a);
2120 struct dp_netdev_port *p;
2123 switch ((enum ovs_action_attr)type) {
2124 case OVS_ACTION_ATTR_OUTPUT:
2125 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
2126 if (OVS_LIKELY(p)) {
2127 netdev_send(p->netdev, packets, cnt, may_steal);
2128 } else if (may_steal) {
2129 for (i = 0; i < cnt; i++) {
2130 dpif_packet_delete(packets[i]);
2135 case OVS_ACTION_ATTR_USERSPACE:
2136 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
2137 const struct nlattr *userdata;
2138 struct ofpbuf actions;
2141 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2142 ofpbuf_init(&actions, 0);
2144 for (i = 0; i < cnt; i++) {
2147 ofpbuf_clear(&actions);
2149 flow_extract(&packets[i]->ofpbuf, md, &flow);
2150 error = dp_netdev_upcall(dp, packets[i], &flow, NULL,
2151 DPIF_UC_ACTION, userdata, &actions,
2153 if (!error || error == ENOSPC) {
2154 dp_netdev_execute_actions(dp, &packets[i], 1, false, md,
2155 ofpbuf_data(&actions),
2156 ofpbuf_size(&actions));
2160 dpif_packet_delete(packets[i]);
2163 ofpbuf_uninit(&actions);
2164 fat_rwlock_unlock(&dp->upcall_rwlock);
2169 case OVS_ACTION_ATTR_HASH: {
2170 const struct ovs_action_hash *hash_act;
2171 struct netdev_flow_key key;
2174 hash_act = nl_attr_get(a);
2176 miniflow_initialize(&key.flow, key.buf);
2178 for (i = 0; i < cnt; i++) {
2180 /* XXX: this is slow. Use RSS hash in the future */
2181 miniflow_extract(&packets[i]->ofpbuf, md, &key.flow);
2183 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
2184 /* Hash need not be symmetric, nor does it need to include
2186 hash = miniflow_hash_5tuple(&key.flow, hash_act->hash_basis);
2188 VLOG_WARN("Unknown hash algorithm specified "
2189 "for the hash action.");
2194 hash = 1; /* 0 is not valid */
2200 packets[i]->dp_hash = hash;
2205 case OVS_ACTION_ATTR_RECIRC:
2206 if (*depth < MAX_RECIRC_DEPTH) {
2209 for (i = 0; i < cnt; i++) {
2210 struct dpif_packet *recirc_pkt;
2211 struct pkt_metadata recirc_md = *md;
2213 recirc_pkt = (may_steal) ? packets[i]
2214 : dpif_packet_clone(packets[i]);
2216 recirc_md.recirc_id = nl_attr_get_u32(a);
2218 /* Hash is private to each packet */
2219 recirc_md.dp_hash = packets[i]->dp_hash;
2221 dp_netdev_input(dp, &recirc_pkt, 1, &recirc_md);
2227 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
2229 for (i = 0; i < cnt; i++) {
2230 dpif_packet_delete(packets[i]);
2236 case OVS_ACTION_ATTR_PUSH_VLAN:
2237 case OVS_ACTION_ATTR_POP_VLAN:
2238 case OVS_ACTION_ATTR_PUSH_MPLS:
2239 case OVS_ACTION_ATTR_POP_MPLS:
2240 case OVS_ACTION_ATTR_SET:
2241 case OVS_ACTION_ATTR_SAMPLE:
2242 case OVS_ACTION_ATTR_UNSPEC:
2243 case __OVS_ACTION_ATTR_MAX:
2249 dp_netdev_execute_actions(struct dp_netdev *dp,
2250 struct dpif_packet **packets, int cnt,
2251 bool may_steal, struct pkt_metadata *md,
2252 const struct nlattr *actions, size_t actions_len)
2254 struct dp_netdev_execute_aux aux = {dp};
2256 odp_execute_actions(&aux, packets, cnt, may_steal, md, actions,
2257 actions_len, dp_execute_cb);
2260 const struct dpif_class dpif_netdev_class = {
2262 dpif_netdev_enumerate,
2263 dpif_netdev_port_open_type,
2266 dpif_netdev_destroy,
2269 dpif_netdev_get_stats,
2270 dpif_netdev_port_add,
2271 dpif_netdev_port_del,
2272 dpif_netdev_port_query_by_number,
2273 dpif_netdev_port_query_by_name,
2274 NULL, /* port_get_pid */
2275 dpif_netdev_port_dump_start,
2276 dpif_netdev_port_dump_next,
2277 dpif_netdev_port_dump_done,
2278 dpif_netdev_port_poll,
2279 dpif_netdev_port_poll_wait,
2280 dpif_netdev_flow_flush,
2281 dpif_netdev_flow_dump_create,
2282 dpif_netdev_flow_dump_destroy,
2283 dpif_netdev_flow_dump_thread_create,
2284 dpif_netdev_flow_dump_thread_destroy,
2285 dpif_netdev_flow_dump_next,
2286 dpif_netdev_operate,
2287 NULL, /* recv_set */
2288 NULL, /* handlers_set */
2289 dpif_netdev_queue_to_priority,
2291 NULL, /* recv_wait */
2292 NULL, /* recv_purge */
2293 dpif_netdev_register_upcall_cb,
2294 dpif_netdev_enable_upcall,
2295 dpif_netdev_disable_upcall,
2299 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2300 const char *argv[], void *aux OVS_UNUSED)
2302 struct dp_netdev_port *old_port;
2303 struct dp_netdev_port *new_port;
2304 struct dp_netdev *dp;
2307 ovs_mutex_lock(&dp_netdev_mutex);
2308 dp = shash_find_data(&dp_netdevs, argv[1]);
2309 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2310 ovs_mutex_unlock(&dp_netdev_mutex);
2311 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2314 ovs_refcount_ref(&dp->ref_cnt);
2315 ovs_mutex_unlock(&dp_netdev_mutex);
2317 ovs_mutex_lock(&dp->port_mutex);
2318 if (get_port_by_name(dp, argv[2], &old_port)) {
2319 unixctl_command_reply_error(conn, "unknown port");
2323 port_no = u32_to_odp(atoi(argv[3]));
2324 if (!port_no || port_no == ODPP_NONE) {
2325 unixctl_command_reply_error(conn, "bad port number");
2328 if (dp_netdev_lookup_port(dp, port_no)) {
2329 unixctl_command_reply_error(conn, "port number already in use");
2333 /* Remove old port. */
2334 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
2335 ovsrcu_postpone(free, old_port);
2337 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
2338 new_port = xmemdup(old_port, sizeof *old_port);
2339 new_port->port_no = port_no;
2340 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
2342 seq_change(dp->port_seq);
2343 unixctl_command_reply(conn, NULL);
2346 ovs_mutex_unlock(&dp->port_mutex);
2347 dp_netdev_unref(dp);
2351 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
2352 const char *argv[], void *aux OVS_UNUSED)
2354 struct dp_netdev_port *port;
2355 struct dp_netdev *dp;
2357 ovs_mutex_lock(&dp_netdev_mutex);
2358 dp = shash_find_data(&dp_netdevs, argv[1]);
2359 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2360 ovs_mutex_unlock(&dp_netdev_mutex);
2361 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2364 ovs_refcount_ref(&dp->ref_cnt);
2365 ovs_mutex_unlock(&dp_netdev_mutex);
2367 ovs_mutex_lock(&dp->port_mutex);
2368 if (get_port_by_name(dp, argv[2], &port)) {
2369 unixctl_command_reply_error(conn, "unknown port");
2370 } else if (port->port_no == ODPP_LOCAL) {
2371 unixctl_command_reply_error(conn, "can't delete local port");
2373 do_del_port(dp, port);
2374 unixctl_command_reply(conn, NULL);
2376 ovs_mutex_unlock(&dp->port_mutex);
2378 dp_netdev_unref(dp);
2382 dpif_dummy_register__(const char *type)
2384 struct dpif_class *class;
2386 class = xmalloc(sizeof *class);
2387 *class = dpif_netdev_class;
2388 class->type = xstrdup(type);
2389 dp_register_provider(class);
2393 dpif_dummy_register(bool override)
2400 dp_enumerate_types(&types);
2401 SSET_FOR_EACH (type, &types) {
2402 if (!dp_unregister_provider(type)) {
2403 dpif_dummy_register__(type);
2406 sset_destroy(&types);
2409 dpif_dummy_register__("dummy");
2411 unixctl_command_register("dpif-dummy/change-port-number",
2412 "DP PORT NEW-NUMBER",
2413 3, 3, dpif_dummy_change_port_number, NULL);
2414 unixctl_command_register("dpif-dummy/delete-port", "DP PORT",
2415 2, 2, dpif_dummy_delete_port, NULL);