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 /* Stores a miniflow */
92 /* There are fields in the flow structure that we never use. Therefore we can
93 * save a few words of memory */
94 #define NETDEV_KEY_BUF_SIZE_U32 (FLOW_U32S \
96 - FLOW_U32_SIZE(regs) \
97 - FLOW_U32_SIZE(metadata) \
99 struct netdev_flow_key {
100 struct miniflow flow;
101 uint32_t buf[NETDEV_KEY_BUF_SIZE_U32];
104 /* Exact match cache for frequently used flows
106 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
107 * search its entries for a miniflow that matches exactly the miniflow of the
108 * packet. It stores the 'cls_rule'(rule) that matches the miniflow.
110 * A cache entry holds a reference to its 'dp_netdev_flow'.
112 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
113 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
114 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
115 * value is the index of a cache entry where the miniflow could be.
121 * Each pmd_thread has its own private exact match cache.
122 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
125 #define EM_FLOW_HASH_SHIFT 10
126 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
127 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
128 #define EM_FLOW_HASH_SEGS 2
132 struct netdev_flow_key mf;
133 struct dp_netdev_flow *flow;
137 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
140 /* Iterate in the exact match cache through every entry that might contain a
141 * miniflow with hash 'HASH'. */
142 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
143 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
144 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
145 i__ < EM_FLOW_HASH_SEGS; \
146 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
148 /* Datapath based on the network device interface from netdev.h.
154 * Some members, marked 'const', are immutable. Accessing other members
155 * requires synchronization, as noted in more detail below.
157 * Acquisition order is, from outermost to innermost:
159 * dp_netdev_mutex (global)
165 const struct dpif_class *const class;
166 const char *const name;
168 struct ovs_refcount ref_cnt;
169 atomic_flag destroyed;
173 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
174 * changes to 'cls' must be made while still holding the 'flow_mutex'.
176 struct ovs_mutex flow_mutex;
177 struct classifier cls;
178 struct cmap flow_table OVS_GUARDED; /* Flow table. */
182 * ovsthread_stats is internally synchronized. */
183 struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */
187 * Protected by RCU. Take the mutex to add or remove ports. */
188 struct ovs_mutex port_mutex;
190 struct seq *port_seq; /* Incremented whenever a port changes. */
192 /* Protects access to ofproto-dpif-upcall interface during revalidator
193 * thread synchronization. */
194 struct fat_rwlock upcall_rwlock;
195 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
198 /* Forwarding threads. */
199 struct latch exit_latch;
200 struct pmd_thread *pmd_threads;
201 size_t n_pmd_threads;
204 /* Exact match cache for non-pmd devices.
205 * Pmd devices use instead each thread's flow_cache for this purpose.
206 * Protected by emc_mutex */
207 struct emc_cache flow_cache OVS_GUARDED;
208 struct ovs_mutex emc_mutex;
211 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
215 DP_STAT_HIT, /* Packets that matched in the flow table. */
216 DP_STAT_MISS, /* Packets that did not match. */
217 DP_STAT_LOST, /* Packets not passed up to the client. */
221 /* Contained by struct dp_netdev's 'stats' member. */
222 struct dp_netdev_stats {
223 struct ovs_mutex mutex; /* Protects 'n'. */
225 /* Indexed by DP_STAT_*, protected by 'mutex'. */
226 unsigned long long int n[DP_N_STATS] OVS_GUARDED;
230 /* A port in a netdev-based datapath. */
231 struct dp_netdev_port {
232 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
234 struct netdev *netdev;
235 struct netdev_saved_flags *sf;
236 struct netdev_rxq **rxq;
237 struct ovs_refcount ref_cnt;
238 char *type; /* Port type as requested by user. */
241 /* A flow in dp_netdev's 'flow_table'.
247 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
248 * its dp_netdev's classifier. The text below calls this classifier 'cls'.
253 * The thread safety rules described here for "struct dp_netdev_flow" are
254 * motivated by two goals:
256 * - Prevent threads that read members of "struct dp_netdev_flow" from
257 * reading bad data due to changes by some thread concurrently modifying
260 * - Prevent two threads making changes to members of a given "struct
261 * dp_netdev_flow" from interfering with each other.
267 * A flow 'flow' may be accessed without a risk of being freed during an RCU
268 * grace period. Code that needs to hold onto a flow for a while
269 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
271 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
272 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
275 * Some members, marked 'const', are immutable. Accessing other members
276 * requires synchronization, as noted in more detail below.
278 struct dp_netdev_flow {
280 /* Packet classification. */
281 const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
283 /* Hash table index by unmasked flow. */
284 const struct cmap_node node; /* In owning dp_netdev's 'flow_table'. */
285 const struct flow flow; /* The flow that created this entry. */
287 /* Number of references.
288 * The classifier owns one reference.
289 * Any thread trying to keep a rule from being freed should hold its own
291 struct ovs_refcount ref_cnt;
295 * Reading or writing these members requires 'mutex'. */
296 struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
299 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
302 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
303 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
305 /* Contained by struct dp_netdev_flow's 'stats' member. */
306 struct dp_netdev_flow_stats {
307 struct ovs_mutex mutex; /* Guards all the other members. */
309 long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
310 long long int packet_count OVS_GUARDED; /* Number of packets matched. */
311 long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
312 uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
315 /* A set of datapath actions within a "struct dp_netdev_flow".
321 * A struct dp_netdev_actions 'actions' is protected with RCU. */
322 struct dp_netdev_actions {
323 /* These members are immutable: they do not change during the struct's
325 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
326 unsigned int size; /* Size of 'actions', in bytes. */
329 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
331 struct dp_netdev_actions *dp_netdev_flow_get_actions(
332 const struct dp_netdev_flow *);
333 static void dp_netdev_actions_free(struct dp_netdev_actions *);
335 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
336 * the performance overhead of interrupt processing. Therefore netdev can
337 * not implement rx-wait for these devices. dpif-netdev needs to poll
338 * these device to check for recv buffer. pmd-thread does polling for
339 * devices assigned to itself thread.
341 * DPDK used PMD for accessing NIC.
343 * A thread that receives packets from PMD ports, looks them up in the flow
344 * table, and executes the actions it finds.
347 struct dp_netdev *dp;
348 struct emc_cache flow_cache;
351 atomic_uint change_seq;
354 #define PMD_INITIAL_SEQ 1
356 /* Interface to netdev-based datapath. */
359 struct dp_netdev *dp;
360 uint64_t last_port_seq;
363 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
364 struct dp_netdev_port **portp);
365 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
366 struct dp_netdev_port **portp);
367 static void dp_netdev_free(struct dp_netdev *)
368 OVS_REQUIRES(dp_netdev_mutex);
369 static void dp_netdev_flow_flush(struct dp_netdev *);
370 static int do_add_port(struct dp_netdev *dp, const char *devname,
371 const char *type, odp_port_t port_no)
372 OVS_REQUIRES(dp->port_mutex);
373 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
374 OVS_REQUIRES(dp->port_mutex);
375 static int dpif_netdev_open(const struct dpif_class *, const char *name,
376 bool create, struct dpif **);
377 static void dp_netdev_execute_actions(struct dp_netdev *dp,
378 struct dpif_packet **, int c,
379 bool may_steal, struct pkt_metadata *,
380 struct emc_cache *flow_cache,
381 const struct nlattr *actions,
383 static void dp_netdev_input(struct dp_netdev *, struct emc_cache *,
384 struct dpif_packet **, int cnt,
385 struct pkt_metadata *);
387 static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
388 static void dp_netdev_disable_upcall(struct dp_netdev *);
390 static void emc_clear_entry(struct emc_entry *ce);
393 emc_cache_init(struct emc_cache *flow_cache)
397 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
398 flow_cache->entries[i].flow = NULL;
399 flow_cache->entries[i].hash = 0;
400 miniflow_initialize(&flow_cache->entries[i].mf.flow,
401 flow_cache->entries[i].mf.buf);
406 emc_cache_uninit(struct emc_cache *flow_cache)
410 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
411 emc_clear_entry(&flow_cache->entries[i]);
415 static struct dpif_netdev *
416 dpif_netdev_cast(const struct dpif *dpif)
418 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
419 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
422 static struct dp_netdev *
423 get_dp_netdev(const struct dpif *dpif)
425 return dpif_netdev_cast(dpif)->dp;
429 dpif_netdev_enumerate(struct sset *all_dps,
430 const struct dpif_class *dpif_class)
432 struct shash_node *node;
434 ovs_mutex_lock(&dp_netdev_mutex);
435 SHASH_FOR_EACH(node, &dp_netdevs) {
436 struct dp_netdev *dp = node->data;
437 if (dpif_class != dp->class) {
438 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
439 * If the class doesn't match, skip this dpif. */
442 sset_add(all_dps, node->name);
444 ovs_mutex_unlock(&dp_netdev_mutex);
450 dpif_netdev_class_is_dummy(const struct dpif_class *class)
452 return class != &dpif_netdev_class;
456 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
458 return strcmp(type, "internal") ? type
459 : dpif_netdev_class_is_dummy(class) ? "dummy"
464 create_dpif_netdev(struct dp_netdev *dp)
466 uint16_t netflow_id = hash_string(dp->name, 0);
467 struct dpif_netdev *dpif;
469 ovs_refcount_ref(&dp->ref_cnt);
471 dpif = xmalloc(sizeof *dpif);
472 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
474 dpif->last_port_seq = seq_read(dp->port_seq);
479 /* Choose an unused, non-zero port number and return it on success.
480 * Return ODPP_NONE on failure. */
482 choose_port(struct dp_netdev *dp, const char *name)
483 OVS_REQUIRES(dp->port_mutex)
487 if (dp->class != &dpif_netdev_class) {
491 /* If the port name begins with "br", start the number search at
492 * 100 to make writing tests easier. */
493 if (!strncmp(name, "br", 2)) {
497 /* If the port name contains a number, try to assign that port number.
498 * This can make writing unit tests easier because port numbers are
500 for (p = name; *p != '\0'; p++) {
501 if (isdigit((unsigned char) *p)) {
502 port_no = start_no + strtol(p, NULL, 10);
503 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
504 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
505 return u32_to_odp(port_no);
512 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
513 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
514 return u32_to_odp(port_no);
522 create_dp_netdev(const char *name, const struct dpif_class *class,
523 struct dp_netdev **dpp)
524 OVS_REQUIRES(dp_netdev_mutex)
526 struct dp_netdev *dp;
529 dp = xzalloc(sizeof *dp);
530 shash_add(&dp_netdevs, name, dp);
532 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
533 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
534 ovs_refcount_init(&dp->ref_cnt);
535 atomic_flag_clear(&dp->destroyed);
537 ovs_mutex_init(&dp->flow_mutex);
538 classifier_init(&dp->cls, NULL);
539 cmap_init(&dp->flow_table);
541 ovsthread_stats_init(&dp->stats);
543 ovs_mutex_init(&dp->port_mutex);
544 cmap_init(&dp->ports);
545 dp->port_seq = seq_create();
546 latch_init(&dp->exit_latch);
547 fat_rwlock_init(&dp->upcall_rwlock);
549 /* Disable upcalls by default. */
550 dp_netdev_disable_upcall(dp);
551 dp->upcall_aux = NULL;
552 dp->upcall_cb = NULL;
554 ovs_mutex_lock(&dp->port_mutex);
555 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
556 ovs_mutex_unlock(&dp->port_mutex);
562 ovs_mutex_init_recursive(&dp->emc_mutex);
563 emc_cache_init(&dp->flow_cache);
570 dpif_netdev_open(const struct dpif_class *class, const char *name,
571 bool create, struct dpif **dpifp)
573 struct dp_netdev *dp;
576 ovs_mutex_lock(&dp_netdev_mutex);
577 dp = shash_find_data(&dp_netdevs, name);
579 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
581 error = (dp->class != class ? EINVAL
586 *dpifp = create_dpif_netdev(dp);
589 ovs_mutex_unlock(&dp_netdev_mutex);
594 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
595 * through the 'dp_netdevs' shash while freeing 'dp'. */
597 dp_netdev_free(struct dp_netdev *dp)
598 OVS_REQUIRES(dp_netdev_mutex)
600 struct dp_netdev_port *port;
601 struct dp_netdev_stats *bucket;
604 shash_find_and_delete(&dp_netdevs, dp->name);
606 dp_netdev_set_pmd_threads(dp, 0);
607 free(dp->pmd_threads);
609 dp_netdev_flow_flush(dp);
610 ovs_mutex_lock(&dp->port_mutex);
611 CMAP_FOR_EACH (port, node, &dp->ports) {
612 do_del_port(dp, port);
614 ovs_mutex_unlock(&dp->port_mutex);
616 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
617 ovs_mutex_destroy(&bucket->mutex);
618 free_cacheline(bucket);
620 ovsthread_stats_destroy(&dp->stats);
622 classifier_destroy(&dp->cls);
623 cmap_destroy(&dp->flow_table);
624 ovs_mutex_destroy(&dp->flow_mutex);
625 seq_destroy(dp->port_seq);
626 cmap_destroy(&dp->ports);
627 fat_rwlock_destroy(&dp->upcall_rwlock);
628 latch_destroy(&dp->exit_latch);
630 emc_cache_uninit(&dp->flow_cache);
631 ovs_mutex_destroy(&dp->emc_mutex);
633 free(CONST_CAST(char *, dp->name));
638 dp_netdev_unref(struct dp_netdev *dp)
641 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
642 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
643 ovs_mutex_lock(&dp_netdev_mutex);
644 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
647 ovs_mutex_unlock(&dp_netdev_mutex);
652 dpif_netdev_close(struct dpif *dpif)
654 struct dp_netdev *dp = get_dp_netdev(dpif);
661 dpif_netdev_destroy(struct dpif *dpif)
663 struct dp_netdev *dp = get_dp_netdev(dpif);
665 if (!atomic_flag_test_and_set(&dp->destroyed)) {
666 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
667 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
676 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
678 struct dp_netdev *dp = get_dp_netdev(dpif);
679 struct dp_netdev_stats *bucket;
682 stats->n_flows = cmap_count(&dp->flow_table);
684 stats->n_hit = stats->n_missed = stats->n_lost = 0;
685 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
686 ovs_mutex_lock(&bucket->mutex);
687 stats->n_hit += bucket->n[DP_STAT_HIT];
688 stats->n_missed += bucket->n[DP_STAT_MISS];
689 stats->n_lost += bucket->n[DP_STAT_LOST];
690 ovs_mutex_unlock(&bucket->mutex);
692 stats->n_masks = UINT32_MAX;
693 stats->n_mask_hit = UINT64_MAX;
699 dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
703 for (i = 0; i < dp->n_pmd_threads; i++) {
704 struct pmd_thread *f = &dp->pmd_threads[i];
707 atomic_add_relaxed(&f->change_seq, 1, &old_seq);
712 hash_port_no(odp_port_t port_no)
714 return hash_int(odp_to_u32(port_no), 0);
718 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
720 OVS_REQUIRES(dp->port_mutex)
722 struct netdev_saved_flags *sf;
723 struct dp_netdev_port *port;
724 struct netdev *netdev;
725 enum netdev_flags flags;
726 const char *open_type;
730 /* XXX reject devices already in some dp_netdev. */
732 /* Open and validate network device. */
733 open_type = dpif_netdev_port_open_type(dp->class, type);
734 error = netdev_open(devname, open_type, &netdev);
738 /* XXX reject non-Ethernet devices */
740 netdev_get_flags(netdev, &flags);
741 if (flags & NETDEV_LOOPBACK) {
742 VLOG_ERR("%s: cannot add a loopback device", devname);
743 netdev_close(netdev);
747 port = xzalloc(sizeof *port);
748 port->port_no = port_no;
749 port->netdev = netdev;
750 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
751 port->type = xstrdup(type);
752 for (i = 0; i < netdev_n_rxq(netdev); i++) {
753 error = netdev_rxq_open(netdev, &port->rxq[i], i);
755 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
756 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
757 devname, ovs_strerror(errno));
758 netdev_close(netdev);
766 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
768 for (i = 0; i < netdev_n_rxq(netdev); i++) {
769 netdev_rxq_close(port->rxq[i]);
771 netdev_close(netdev);
779 if (netdev_is_pmd(netdev)) {
781 dp_netdev_set_pmd_threads(dp, NR_PMD_THREADS);
782 dp_netdev_reload_pmd_threads(dp);
784 ovs_refcount_init(&port->ref_cnt);
786 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
787 seq_change(dp->port_seq);
793 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
794 odp_port_t *port_nop)
796 struct dp_netdev *dp = get_dp_netdev(dpif);
797 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
798 const char *dpif_port;
802 ovs_mutex_lock(&dp->port_mutex);
803 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
804 if (*port_nop != ODPP_NONE) {
806 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
808 port_no = choose_port(dp, dpif_port);
809 error = port_no == ODPP_NONE ? EFBIG : 0;
813 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
815 ovs_mutex_unlock(&dp->port_mutex);
821 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
823 struct dp_netdev *dp = get_dp_netdev(dpif);
826 ovs_mutex_lock(&dp->port_mutex);
827 if (port_no == ODPP_LOCAL) {
830 struct dp_netdev_port *port;
832 error = get_port_by_number(dp, port_no, &port);
834 do_del_port(dp, port);
837 ovs_mutex_unlock(&dp->port_mutex);
843 is_valid_port_number(odp_port_t port_no)
845 return port_no != ODPP_NONE;
848 static struct dp_netdev_port *
849 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
851 struct dp_netdev_port *port;
853 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
854 if (port->port_no == port_no) {
862 get_port_by_number(struct dp_netdev *dp,
863 odp_port_t port_no, struct dp_netdev_port **portp)
865 if (!is_valid_port_number(port_no)) {
869 *portp = dp_netdev_lookup_port(dp, port_no);
870 return *portp ? 0 : ENOENT;
875 port_ref(struct dp_netdev_port *port)
878 ovs_refcount_ref(&port->ref_cnt);
883 port_try_ref(struct dp_netdev_port *port)
886 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
893 port_destroy__(struct dp_netdev_port *port)
895 int n_rxq = netdev_n_rxq(port->netdev);
898 netdev_close(port->netdev);
899 netdev_restore_flags(port->sf);
901 for (i = 0; i < n_rxq; i++) {
902 netdev_rxq_close(port->rxq[i]);
910 port_unref(struct dp_netdev_port *port)
912 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
913 ovsrcu_postpone(port_destroy__, port);
918 get_port_by_name(struct dp_netdev *dp,
919 const char *devname, struct dp_netdev_port **portp)
920 OVS_REQUIRES(dp->port_mutex)
922 struct dp_netdev_port *port;
924 CMAP_FOR_EACH (port, node, &dp->ports) {
925 if (!strcmp(netdev_get_name(port->netdev), devname)) {
934 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
935 OVS_REQUIRES(dp->port_mutex)
937 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
938 seq_change(dp->port_seq);
939 if (netdev_is_pmd(port->netdev)) {
940 dp_netdev_reload_pmd_threads(dp);
947 answer_port_query(const struct dp_netdev_port *port,
948 struct dpif_port *dpif_port)
950 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
951 dpif_port->type = xstrdup(port->type);
952 dpif_port->port_no = port->port_no;
956 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
957 struct dpif_port *dpif_port)
959 struct dp_netdev *dp = get_dp_netdev(dpif);
960 struct dp_netdev_port *port;
963 error = get_port_by_number(dp, port_no, &port);
964 if (!error && dpif_port) {
965 answer_port_query(port, dpif_port);
972 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
973 struct dpif_port *dpif_port)
975 struct dp_netdev *dp = get_dp_netdev(dpif);
976 struct dp_netdev_port *port;
979 ovs_mutex_lock(&dp->port_mutex);
980 error = get_port_by_name(dp, devname, &port);
981 if (!error && dpif_port) {
982 answer_port_query(port, dpif_port);
984 ovs_mutex_unlock(&dp->port_mutex);
990 dp_netdev_flow_free(struct dp_netdev_flow *flow)
992 struct dp_netdev_flow_stats *bucket;
995 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
996 ovs_mutex_destroy(&bucket->mutex);
997 free_cacheline(bucket);
999 ovsthread_stats_destroy(&flow->stats);
1001 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
1002 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1006 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1008 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1009 ovsrcu_postpone(dp_netdev_flow_free, flow);
1014 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
1015 OVS_REQUIRES(dp->flow_mutex)
1017 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
1018 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1020 classifier_remove(&dp->cls, cr);
1021 cmap_remove(&dp->flow_table, node, flow_hash(&flow->flow, 0));
1024 dp_netdev_flow_unref(flow);
1028 dp_netdev_flow_flush(struct dp_netdev *dp)
1030 struct dp_netdev_flow *netdev_flow;
1032 ovs_mutex_lock(&dp->flow_mutex);
1033 CMAP_FOR_EACH (netdev_flow, node, &dp->flow_table) {
1034 dp_netdev_remove_flow(dp, netdev_flow);
1036 ovs_mutex_unlock(&dp->flow_mutex);
1040 dpif_netdev_flow_flush(struct dpif *dpif)
1042 struct dp_netdev *dp = get_dp_netdev(dpif);
1044 dp_netdev_flow_flush(dp);
1048 struct dp_netdev_port_state {
1049 struct cmap_position position;
1054 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1056 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1061 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1062 struct dpif_port *dpif_port)
1064 struct dp_netdev_port_state *state = state_;
1065 struct dp_netdev *dp = get_dp_netdev(dpif);
1066 struct cmap_node *node;
1069 node = cmap_next_position(&dp->ports, &state->position);
1071 struct dp_netdev_port *port;
1073 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1076 state->name = xstrdup(netdev_get_name(port->netdev));
1077 dpif_port->name = state->name;
1078 dpif_port->type = port->type;
1079 dpif_port->port_no = port->port_no;
1090 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1092 struct dp_netdev_port_state *state = state_;
1099 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1101 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1102 uint64_t new_port_seq;
1105 new_port_seq = seq_read(dpif->dp->port_seq);
1106 if (dpif->last_port_seq != new_port_seq) {
1107 dpif->last_port_seq = new_port_seq;
1117 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1119 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1121 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1124 static struct dp_netdev_flow *
1125 dp_netdev_flow_cast(const struct cls_rule *cr)
1127 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1130 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1132 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1136 emc_entry_alive(struct emc_entry *ce)
1138 return ce->flow && !ce->flow->dead;
1142 emc_clear_entry(struct emc_entry *ce)
1145 dp_netdev_flow_unref(ce->flow);
1151 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1152 const struct miniflow *mf, uint32_t hash)
1154 if (ce->flow != flow) {
1156 dp_netdev_flow_unref(ce->flow);
1159 if (dp_netdev_flow_ref(flow)) {
1166 miniflow_clone_inline(&ce->mf.flow, mf, count_1bits(mf->map));
1172 emc_insert(struct emc_cache *cache, const struct miniflow *mf, uint32_t hash,
1173 struct dp_netdev_flow *flow)
1175 struct emc_entry *to_be_replaced = NULL;
1176 struct emc_entry *current_entry;
1178 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, hash) {
1179 if (current_entry->hash == hash
1180 && miniflow_equal(¤t_entry->mf.flow, mf)) {
1182 /* We found the entry with the 'mf' miniflow */
1183 emc_change_entry(current_entry, flow, NULL, 0);
1187 /* Replacement policy: put the flow in an empty (not alive) entry, or
1188 * in the first entry where it can be */
1190 || (emc_entry_alive(to_be_replaced)
1191 && !emc_entry_alive(current_entry))
1192 || current_entry->hash < to_be_replaced->hash) {
1193 to_be_replaced = current_entry;
1196 /* We didn't find the miniflow in the cache.
1197 * The 'to_be_replaced' entry is where the new flow will be stored */
1199 emc_change_entry(to_be_replaced, flow, mf, hash);
1202 static inline struct dp_netdev_flow *
1203 emc_lookup(struct emc_cache *cache, const struct miniflow *mf, uint32_t hash)
1205 struct emc_entry *current_entry;
1207 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, hash) {
1208 if (current_entry->hash == hash && emc_entry_alive(current_entry)
1209 && miniflow_equal(¤t_entry->mf.flow, mf)) {
1211 /* We found the entry with the 'mf' miniflow */
1212 return current_entry->flow;
1219 static struct dp_netdev_flow *
1220 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
1222 struct dp_netdev_flow *netdev_flow;
1223 struct cls_rule *rule;
1225 classifier_lookup_miniflow_batch(&dp->cls, &key, &rule, 1);
1226 netdev_flow = dp_netdev_flow_cast(rule);
1231 static struct dp_netdev_flow *
1232 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1234 struct dp_netdev_flow *netdev_flow;
1236 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1238 if (flow_equal(&netdev_flow->flow, flow)) {
1247 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow,
1248 struct dpif_flow_stats *stats)
1250 struct dp_netdev_flow_stats *bucket;
1253 memset(stats, 0, sizeof *stats);
1254 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1255 ovs_mutex_lock(&bucket->mutex);
1256 stats->n_packets += bucket->packet_count;
1257 stats->n_bytes += bucket->byte_count;
1258 stats->used = MAX(stats->used, bucket->used);
1259 stats->tcp_flags |= bucket->tcp_flags;
1260 ovs_mutex_unlock(&bucket->mutex);
1265 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1266 struct ofpbuf *buffer, struct dpif_flow *flow)
1268 struct flow_wildcards wc;
1269 struct dp_netdev_actions *actions;
1271 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1272 odp_flow_key_from_mask(buffer, &wc.masks, &netdev_flow->flow,
1273 odp_to_u32(wc.masks.in_port.odp_port),
1275 flow->mask = ofpbuf_data(buffer);
1276 flow->mask_len = ofpbuf_size(buffer);
1278 actions = dp_netdev_flow_get_actions(netdev_flow);
1279 flow->actions = actions->actions;
1280 flow->actions_len = actions->size;
1282 get_dpif_flow_stats(netdev_flow, &flow->stats);
1286 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1287 const struct nlattr *mask_key,
1288 uint32_t mask_key_len, const struct flow *flow,
1292 enum odp_key_fitness fitness;
1294 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1296 /* This should not happen: it indicates that
1297 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1298 * disagree on the acceptable form of a mask. Log the problem
1299 * as an error, with enough details to enable debugging. */
1300 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1302 if (!VLOG_DROP_ERR(&rl)) {
1306 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1308 VLOG_ERR("internal error parsing flow mask %s (%s)",
1309 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1316 enum mf_field_id id;
1317 /* No mask key, unwildcard everything except fields whose
1318 * prerequisities are not met. */
1319 memset(mask, 0x0, sizeof *mask);
1321 for (id = 0; id < MFF_N_IDS; ++id) {
1322 /* Skip registers and metadata. */
1323 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1324 && id != MFF_METADATA) {
1325 const struct mf_field *mf = mf_from_id(id);
1326 if (mf_are_prereqs_ok(mf, flow)) {
1327 mf_mask_field(mf, mask);
1333 /* Force unwildcard the in_port.
1335 * We need to do this even in the case where we unwildcard "everything"
1336 * above because "everything" only includes the 16-bit OpenFlow port number
1337 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1338 * port number mask->in_port.odp_port. */
1339 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1345 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1350 if (odp_flow_key_to_flow(key, key_len, flow)) {
1351 /* This should not happen: it indicates that odp_flow_key_from_flow()
1352 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1353 * flow. Log the problem as an error, with enough details to enable
1355 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1357 if (!VLOG_DROP_ERR(&rl)) {
1361 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1362 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1369 in_port = flow->in_port.odp_port;
1370 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1378 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1380 struct dp_netdev *dp = get_dp_netdev(dpif);
1381 struct dp_netdev_flow *netdev_flow;
1385 error = dpif_netdev_flow_from_nlattrs(get->key, get->key_len, &key);
1390 netdev_flow = dp_netdev_find_flow(dp, &key);
1393 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->flow);
1402 dp_netdev_flow_add(struct dp_netdev *dp, struct match *match,
1403 const struct nlattr *actions, size_t actions_len)
1404 OVS_REQUIRES(dp->flow_mutex)
1406 struct dp_netdev_flow *netdev_flow;
1408 netdev_flow = xzalloc(sizeof *netdev_flow);
1409 *CONST_CAST(struct flow *, &netdev_flow->flow) = match->flow;
1411 ovs_refcount_init(&netdev_flow->ref_cnt);
1413 ovsthread_stats_init(&netdev_flow->stats);
1415 ovsrcu_set(&netdev_flow->actions,
1416 dp_netdev_actions_create(actions, actions_len));
1418 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1419 match, NETDEV_RULE_PRIORITY);
1420 cmap_insert(&dp->flow_table,
1421 CONST_CAST(struct cmap_node *, &netdev_flow->node),
1422 flow_hash(&match->flow, 0));
1423 classifier_insert(&dp->cls,
1424 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1426 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1427 struct ds ds = DS_EMPTY_INITIALIZER;
1429 ds_put_cstr(&ds, "flow_add: ");
1430 match_format(match, &ds, OFP_DEFAULT_PRIORITY);
1431 ds_put_cstr(&ds, ", actions:");
1432 format_odp_actions(&ds, actions, actions_len);
1434 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
1443 clear_stats(struct dp_netdev_flow *netdev_flow)
1445 struct dp_netdev_flow_stats *bucket;
1448 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1449 ovs_mutex_lock(&bucket->mutex);
1451 bucket->packet_count = 0;
1452 bucket->byte_count = 0;
1453 bucket->tcp_flags = 0;
1454 ovs_mutex_unlock(&bucket->mutex);
1459 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1461 struct dp_netdev *dp = get_dp_netdev(dpif);
1462 struct dp_netdev_flow *netdev_flow;
1463 struct miniflow miniflow;
1467 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
1471 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1472 put->mask, put->mask_len,
1473 &match.flow, &match.wc.masks);
1477 miniflow_init(&miniflow, &match.flow);
1479 ovs_mutex_lock(&dp->flow_mutex);
1480 netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
1482 if (put->flags & DPIF_FP_CREATE) {
1483 if (cmap_count(&dp->flow_table) < MAX_FLOWS) {
1485 memset(put->stats, 0, sizeof *put->stats);
1487 error = dp_netdev_flow_add(dp, &match, put->actions,
1496 if (put->flags & DPIF_FP_MODIFY
1497 && flow_equal(&match.flow, &netdev_flow->flow)) {
1498 struct dp_netdev_actions *new_actions;
1499 struct dp_netdev_actions *old_actions;
1501 new_actions = dp_netdev_actions_create(put->actions,
1504 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1505 ovsrcu_set(&netdev_flow->actions, new_actions);
1508 get_dpif_flow_stats(netdev_flow, put->stats);
1510 if (put->flags & DPIF_FP_ZERO_STATS) {
1511 clear_stats(netdev_flow);
1514 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1515 } else if (put->flags & DPIF_FP_CREATE) {
1518 /* Overlapping flow. */
1522 ovs_mutex_unlock(&dp->flow_mutex);
1523 miniflow_destroy(&miniflow);
1529 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1531 struct dp_netdev *dp = get_dp_netdev(dpif);
1532 struct dp_netdev_flow *netdev_flow;
1536 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1541 ovs_mutex_lock(&dp->flow_mutex);
1542 netdev_flow = dp_netdev_find_flow(dp, &key);
1545 get_dpif_flow_stats(netdev_flow, del->stats);
1547 dp_netdev_remove_flow(dp, netdev_flow);
1551 ovs_mutex_unlock(&dp->flow_mutex);
1556 struct dpif_netdev_flow_dump {
1557 struct dpif_flow_dump up;
1558 struct cmap_position pos;
1560 struct ovs_mutex mutex;
1563 static struct dpif_netdev_flow_dump *
1564 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1566 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1569 static struct dpif_flow_dump *
1570 dpif_netdev_flow_dump_create(const struct dpif *dpif_)
1572 struct dpif_netdev_flow_dump *dump;
1574 dump = xmalloc(sizeof *dump);
1575 dpif_flow_dump_init(&dump->up, dpif_);
1576 memset(&dump->pos, 0, sizeof dump->pos);
1578 ovs_mutex_init(&dump->mutex);
1584 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1586 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1588 ovs_mutex_destroy(&dump->mutex);
1593 struct dpif_netdev_flow_dump_thread {
1594 struct dpif_flow_dump_thread up;
1595 struct dpif_netdev_flow_dump *dump;
1596 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1597 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1600 static struct dpif_netdev_flow_dump_thread *
1601 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1603 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1606 static struct dpif_flow_dump_thread *
1607 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1609 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1610 struct dpif_netdev_flow_dump_thread *thread;
1612 thread = xmalloc(sizeof *thread);
1613 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1614 thread->dump = dump;
1619 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1621 struct dpif_netdev_flow_dump_thread *thread
1622 = dpif_netdev_flow_dump_thread_cast(thread_);
1628 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
1629 struct dpif_flow *flows, int max_flows)
1631 struct dpif_netdev_flow_dump_thread *thread
1632 = dpif_netdev_flow_dump_thread_cast(thread_);
1633 struct dpif_netdev_flow_dump *dump = thread->dump;
1634 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
1635 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
1636 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
1640 ovs_mutex_lock(&dump->mutex);
1641 if (!dump->status) {
1642 for (n_flows = 0; n_flows < MIN(max_flows, FLOW_DUMP_MAX_BATCH);
1644 struct cmap_node *node;
1646 node = cmap_next_position(&dp->flow_table, &dump->pos);
1651 netdev_flows[n_flows] = CONTAINER_OF(node, struct dp_netdev_flow,
1655 ovs_mutex_unlock(&dump->mutex);
1657 for (i = 0; i < n_flows; i++) {
1658 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
1659 struct odputil_keybuf *keybuf = &thread->keybuf[i];
1660 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
1661 struct dpif_flow *f = &flows[i];
1662 struct dp_netdev_actions *dp_actions;
1663 struct flow_wildcards wc;
1666 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1669 ofpbuf_use_stack(&buf, keybuf, sizeof *keybuf);
1670 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1671 netdev_flow->flow.in_port.odp_port, true);
1672 f->key = ofpbuf_data(&buf);
1673 f->key_len = ofpbuf_size(&buf);
1676 ofpbuf_use_stack(&buf, maskbuf, sizeof *maskbuf);
1677 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1678 odp_to_u32(wc.masks.in_port.odp_port),
1680 f->mask = ofpbuf_data(&buf);
1681 f->mask_len = ofpbuf_size(&buf);
1684 dp_actions = dp_netdev_flow_get_actions(netdev_flow);
1685 f->actions = dp_actions->actions;
1686 f->actions_len = dp_actions->size;
1689 get_dpif_flow_stats(netdev_flow, &f->stats);
1696 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1698 struct dp_netdev *dp = get_dp_netdev(dpif);
1699 struct dpif_packet packet, *pp;
1700 struct pkt_metadata *md = &execute->md;
1702 if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
1703 ofpbuf_size(execute->packet) > UINT16_MAX) {
1707 packet.ofpbuf = *execute->packet;
1710 ovs_mutex_lock(&dp->emc_mutex);
1711 dp_netdev_execute_actions(dp, &pp, 1, false, md,
1712 &dp->flow_cache, execute->actions,
1713 execute->actions_len);
1714 ovs_mutex_unlock(&dp->emc_mutex);
1716 /* Even though may_steal is set to false, some actions could modify or
1717 * reallocate the ofpbuf memory. We need to pass those changes to the
1719 *execute->packet = packet.ofpbuf;
1725 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
1729 for (i = 0; i < n_ops; i++) {
1730 struct dpif_op *op = ops[i];
1733 case DPIF_OP_FLOW_PUT:
1734 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
1737 case DPIF_OP_FLOW_DEL:
1738 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
1741 case DPIF_OP_EXECUTE:
1742 op->error = dpif_netdev_execute(dpif, &op->u.execute);
1745 case DPIF_OP_FLOW_GET:
1746 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
1753 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1754 uint32_t queue_id, uint32_t *priority)
1756 *priority = queue_id;
1761 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1762 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1764 struct dp_netdev_actions *
1765 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1767 struct dp_netdev_actions *netdev_actions;
1769 netdev_actions = xmalloc(sizeof *netdev_actions);
1770 netdev_actions->actions = xmemdup(actions, size);
1771 netdev_actions->size = size;
1773 return netdev_actions;
1776 struct dp_netdev_actions *
1777 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
1779 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
1783 dp_netdev_actions_free(struct dp_netdev_actions *actions)
1785 free(actions->actions);
1791 dp_netdev_process_rxq_port(struct dp_netdev *dp,
1792 struct emc_cache *flow_cache,
1793 struct dp_netdev_port *port,
1794 struct netdev_rxq *rxq)
1796 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1799 error = netdev_rxq_recv(rxq, packets, &cnt);
1801 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port->port_no);
1803 *recirc_depth_get() = 0;
1804 dp_netdev_input(dp, flow_cache, packets, cnt, &md);
1805 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1806 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1808 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1809 netdev_get_name(port->netdev), ovs_strerror(error));
1814 dpif_netdev_run(struct dpif *dpif)
1816 struct dp_netdev_port *port;
1817 struct dp_netdev *dp = get_dp_netdev(dpif);
1819 ovs_mutex_lock(&dp->emc_mutex);
1820 CMAP_FOR_EACH (port, node, &dp->ports) {
1821 if (!netdev_is_pmd(port->netdev)) {
1824 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1825 dp_netdev_process_rxq_port(dp, &dp->flow_cache, port,
1830 ovs_mutex_unlock(&dp->emc_mutex);
1834 dpif_netdev_wait(struct dpif *dpif)
1836 struct dp_netdev_port *port;
1837 struct dp_netdev *dp = get_dp_netdev(dpif);
1839 ovs_mutex_lock(&dp_netdev_mutex);
1840 CMAP_FOR_EACH (port, node, &dp->ports) {
1841 if (!netdev_is_pmd(port->netdev)) {
1844 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1845 netdev_rxq_wait(port->rxq[i]);
1849 ovs_mutex_unlock(&dp_netdev_mutex);
1853 struct dp_netdev_port *port;
1854 struct netdev_rxq *rx;
1858 pmd_load_queues(struct pmd_thread *f,
1859 struct rxq_poll **ppoll_list, int poll_cnt)
1861 struct dp_netdev *dp = f->dp;
1862 struct rxq_poll *poll_list = *ppoll_list;
1863 struct dp_netdev_port *port;
1868 /* Simple scheduler for netdev rx polling. */
1869 for (i = 0; i < poll_cnt; i++) {
1870 port_unref(poll_list[i].port);
1876 CMAP_FOR_EACH (port, node, &f->dp->ports) {
1877 /* Calls port_try_ref() to prevent the main thread
1878 * from deleting the port. */
1879 if (port_try_ref(port)) {
1880 if (netdev_is_pmd(port->netdev)) {
1883 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1884 if ((index % dp->n_pmd_threads) == id) {
1885 poll_list = xrealloc(poll_list,
1886 sizeof *poll_list * (poll_cnt + 1));
1889 poll_list[poll_cnt].port = port;
1890 poll_list[poll_cnt].rx = port->rxq[i];
1896 /* Unrefs the port_try_ref(). */
1901 *ppoll_list = poll_list;
1906 pmd_thread_main(void *f_)
1908 struct pmd_thread *f = f_;
1909 struct dp_netdev *dp = f->dp;
1910 unsigned int lc = 0;
1911 struct rxq_poll *poll_list;
1912 unsigned int port_seq = PMD_INITIAL_SEQ;
1919 pmd_thread_setaffinity_cpu(f->id);
1921 emc_cache_init(&f->flow_cache);
1922 poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
1927 for (i = 0; i < poll_cnt; i++) {
1928 dp_netdev_process_rxq_port(dp, &f->flow_cache, poll_list[i].port,
1939 atomic_read_relaxed(&f->change_seq, &seq);
1940 if (seq != port_seq) {
1947 emc_cache_uninit(&f->flow_cache);
1949 if (!latch_is_set(&f->dp->exit_latch)){
1953 for (i = 0; i < poll_cnt; i++) {
1954 port_unref(poll_list[i].port);
1962 dp_netdev_disable_upcall(struct dp_netdev *dp)
1963 OVS_ACQUIRES(dp->upcall_rwlock)
1965 fat_rwlock_wrlock(&dp->upcall_rwlock);
1969 dpif_netdev_disable_upcall(struct dpif *dpif)
1970 OVS_NO_THREAD_SAFETY_ANALYSIS
1972 struct dp_netdev *dp = get_dp_netdev(dpif);
1973 dp_netdev_disable_upcall(dp);
1977 dp_netdev_enable_upcall(struct dp_netdev *dp)
1978 OVS_RELEASES(dp->upcall_rwlock)
1980 fat_rwlock_unlock(&dp->upcall_rwlock);
1984 dpif_netdev_enable_upcall(struct dpif *dpif)
1985 OVS_NO_THREAD_SAFETY_ANALYSIS
1987 struct dp_netdev *dp = get_dp_netdev(dpif);
1988 dp_netdev_enable_upcall(dp);
1992 dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n)
1996 if (n == dp->n_pmd_threads) {
2000 /* Stop existing threads. */
2001 latch_set(&dp->exit_latch);
2002 dp_netdev_reload_pmd_threads(dp);
2003 for (i = 0; i < dp->n_pmd_threads; i++) {
2004 struct pmd_thread *f = &dp->pmd_threads[i];
2006 xpthread_join(f->thread, NULL);
2008 latch_poll(&dp->exit_latch);
2009 free(dp->pmd_threads);
2011 /* Start new threads. */
2012 dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
2013 dp->n_pmd_threads = n;
2015 for (i = 0; i < n; i++) {
2016 struct pmd_thread *f = &dp->pmd_threads[i];
2020 atomic_init(&f->change_seq, PMD_INITIAL_SEQ);
2022 /* Each thread will distribute all devices rx-queues among
2024 f->thread = ovs_thread_create("pmd", pmd_thread_main, f);
2030 dp_netdev_flow_stats_new_cb(void)
2032 struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
2033 ovs_mutex_init(&bucket->mutex);
2038 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
2042 long long int now = time_msec();
2043 struct dp_netdev_flow_stats *bucket;
2045 bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
2046 dp_netdev_flow_stats_new_cb);
2048 ovs_mutex_lock(&bucket->mutex);
2049 bucket->used = MAX(now, bucket->used);
2050 bucket->packet_count += cnt;
2051 bucket->byte_count += size;
2052 bucket->tcp_flags |= tcp_flags;
2053 ovs_mutex_unlock(&bucket->mutex);
2057 dp_netdev_stats_new_cb(void)
2059 struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
2060 ovs_mutex_init(&bucket->mutex);
2065 dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type, int cnt)
2067 struct dp_netdev_stats *bucket;
2069 bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
2070 ovs_mutex_lock(&bucket->mutex);
2071 bucket->n[type] += cnt;
2072 ovs_mutex_unlock(&bucket->mutex);
2076 dp_netdev_upcall(struct dp_netdev *dp, struct dpif_packet *packet_,
2077 struct flow *flow, struct flow_wildcards *wc,
2078 enum dpif_upcall_type type, const struct nlattr *userdata,
2079 struct ofpbuf *actions, struct ofpbuf *put_actions)
2081 struct ofpbuf *packet = &packet_->ofpbuf;
2083 if (type == DPIF_UC_MISS) {
2084 dp_netdev_count_packet(dp, DP_STAT_MISS, 1);
2087 if (OVS_UNLIKELY(!dp->upcall_cb)) {
2091 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
2092 struct ds ds = DS_EMPTY_INITIALIZER;
2096 ofpbuf_init(&key, 0);
2097 odp_flow_key_from_flow(&key, flow, &wc->masks, flow->in_port.odp_port,
2100 packet_str = ofp_packet_to_string(ofpbuf_data(packet),
2101 ofpbuf_size(packet));
2103 odp_flow_key_format(ofpbuf_data(&key), ofpbuf_size(&key), &ds);
2105 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
2106 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
2108 ofpbuf_uninit(&key);
2113 return dp->upcall_cb(packet, flow, type, userdata, actions, wc,
2114 put_actions, dp->upcall_aux);
2117 static inline uint32_t
2118 dpif_netdev_packet_get_dp_hash(struct dpif_packet *packet,
2119 const struct miniflow *mf)
2123 hash = dpif_packet_get_dp_hash(packet);
2124 if (OVS_UNLIKELY(!hash)) {
2125 hash = miniflow_hash_5tuple(mf, 0);
2126 dpif_packet_set_dp_hash(packet, hash);
2131 struct packet_batch {
2132 unsigned int packet_count;
2133 unsigned int byte_count;
2136 struct dp_netdev_flow *flow;
2138 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
2139 struct pkt_metadata md;
2143 packet_batch_update(struct packet_batch *batch, struct dpif_packet *packet,
2144 const struct miniflow *mf)
2146 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
2147 batch->packets[batch->packet_count++] = packet;
2148 batch->byte_count += ofpbuf_size(&packet->ofpbuf);
2152 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow,
2153 struct pkt_metadata *md)
2158 batch->packet_count = 0;
2159 batch->byte_count = 0;
2160 batch->tcp_flags = 0;
2164 packet_batch_execute(struct packet_batch *batch, struct dp_netdev *dp,
2165 struct emc_cache *flow_cache)
2167 struct dp_netdev_actions *actions;
2168 struct dp_netdev_flow *flow = batch->flow;
2170 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
2173 actions = dp_netdev_flow_get_actions(flow);
2175 dp_netdev_execute_actions(dp, batch->packets, batch->packet_count, true,
2176 &batch->md, flow_cache,
2177 actions->actions, actions->size);
2179 dp_netdev_count_packet(dp, DP_STAT_HIT, batch->packet_count);
2183 dp_netdev_queue_batches(struct dpif_packet *pkt, struct pkt_metadata *md,
2184 struct dp_netdev_flow *flow, const struct miniflow *mf,
2185 struct packet_batch *batches, size_t *n_batches,
2188 struct packet_batch *batch = NULL;
2191 if (OVS_UNLIKELY(!flow)) {
2194 /* XXX: This O(n^2) algortihm makes sense if we're operating under the
2195 * assumption that the number of distinct flows (and therefore the
2196 * number of distinct batches) is quite small. If this turns out not
2197 * to be the case, it may make sense to pre sort based on the
2198 * netdev_flow pointer. That done we can get the appropriate batching
2199 * in O(n * log(n)) instead. */
2200 for (j = *n_batches - 1; j >= 0; j--) {
2201 if (batches[j].flow == flow) {
2202 batch = &batches[j];
2203 packet_batch_update(batch, pkt, mf);
2207 if (OVS_UNLIKELY(*n_batches >= max_batches)) {
2211 batch = &batches[(*n_batches)++];
2212 packet_batch_init(batch, flow, md);
2213 packet_batch_update(batch, pkt, mf);
2218 dpif_packet_swap(struct dpif_packet **a, struct dpif_packet **b)
2220 struct dpif_packet *tmp = *a;
2225 /* Try to process all ('cnt') the 'packets' using only the exact match cache
2226 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', or if there
2227 * is no matching batch for a packet's flow, the miniflow is copied into 'keys'
2228 * and the packet pointer is moved at the beginning of the 'packets' array.
2230 * The function returns the number of packets that needs to be processed in the
2231 * 'packets' array (they have been moved to the beginning of the vector).
2233 static inline size_t
2234 emc_processing(struct dp_netdev *dp, struct emc_cache *flow_cache,
2235 struct dpif_packet **packets, size_t cnt,
2236 struct pkt_metadata *md, struct netdev_flow_key *keys)
2238 struct netdev_flow_key key;
2239 struct packet_batch batches[4];
2240 size_t n_batches, i;
2241 size_t notfound_cnt = 0;
2244 miniflow_initialize(&key.flow, key.buf);
2245 for (i = 0; i < cnt; i++) {
2246 struct dp_netdev_flow *flow;
2249 if (OVS_UNLIKELY(ofpbuf_size(&packets[i]->ofpbuf) < ETH_HEADER_LEN)) {
2250 dpif_packet_delete(packets[i]);
2254 miniflow_extract(&packets[i]->ofpbuf, md, &key.flow);
2256 hash = dpif_netdev_packet_get_dp_hash(packets[i], &key.flow);
2258 flow = emc_lookup(flow_cache, &key.flow, hash);
2259 if (OVS_UNLIKELY(!dp_netdev_queue_batches(packets[i], md,
2261 batches, &n_batches,
2262 ARRAY_SIZE(batches)))) {
2263 if (i != notfound_cnt) {
2264 dpif_packet_swap(&packets[i], &packets[notfound_cnt]);
2267 keys[notfound_cnt++] = key;
2271 for (i = 0; i < n_batches; i++) {
2272 packet_batch_execute(&batches[i], dp, flow_cache);
2275 return notfound_cnt;
2279 fast_path_processing(struct dp_netdev *dp, struct emc_cache *flow_cache,
2280 struct dpif_packet **packets, size_t cnt,
2281 struct pkt_metadata *md, struct netdev_flow_key *keys)
2283 #if !defined(__CHECKER__) && !defined(_WIN32)
2284 const size_t PKT_ARRAY_SIZE = cnt;
2286 /* Sparse or MSVC doesn't like variable length array. */
2287 enum { PKT_ARRAY_SIZE = NETDEV_MAX_RX_BATCH };
2289 struct packet_batch batches[PKT_ARRAY_SIZE];
2290 const struct miniflow *mfs[PKT_ARRAY_SIZE]; /* NULL at bad packets. */
2291 struct cls_rule *rules[PKT_ARRAY_SIZE];
2292 size_t n_batches, i;
2295 for (i = 0; i < cnt; i++) {
2296 mfs[i] = &keys[i].flow;
2298 any_miss = !classifier_lookup_miniflow_batch(&dp->cls, mfs, rules, cnt);
2299 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
2300 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
2301 struct ofpbuf actions, put_actions;
2304 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
2305 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
2307 for (i = 0; i < cnt; i++) {
2308 const struct dp_netdev_flow *netdev_flow;
2309 struct ofpbuf *add_actions;
2312 if (OVS_LIKELY(rules[i] || !mfs[i])) {
2316 /* It's possible that an earlier slow path execution installed
2317 * the rule this flow needs. In this case, it's a lot cheaper
2318 * to catch it here than execute a miss. */
2319 netdev_flow = dp_netdev_lookup_flow(dp, mfs[i]);
2321 rules[i] = CONST_CAST(struct cls_rule *, &netdev_flow->cr);
2325 miniflow_expand(mfs[i], &match.flow);
2327 ofpbuf_clear(&actions);
2328 ofpbuf_clear(&put_actions);
2330 error = dp_netdev_upcall(dp, packets[i], &match.flow, &match.wc,
2331 DPIF_UC_MISS, NULL, &actions,
2333 if (OVS_UNLIKELY(error && error != ENOSPC)) {
2337 /* We can't allow the packet batching in the next loop to execute
2338 * the actions. Otherwise, if there are any slow path actions,
2339 * we'll send the packet up twice. */
2340 dp_netdev_execute_actions(dp, &packets[i], 1, false, md,
2341 flow_cache, ofpbuf_data(&actions),
2342 ofpbuf_size(&actions));
2344 add_actions = ofpbuf_size(&put_actions)
2348 ovs_mutex_lock(&dp->flow_mutex);
2349 /* XXX: There's a brief race where this flow could have already
2350 * been installed since we last did the flow lookup. This could be
2351 * solved by moving the mutex lock outside the loop, but that's an
2352 * awful long time to be locking everyone out of making flow
2353 * installs. If we move to a per-core classifier, it would be
2355 if (OVS_LIKELY(error != ENOSPC)
2356 && !dp_netdev_lookup_flow(dp, mfs[i])) {
2357 dp_netdev_flow_add(dp, &match, ofpbuf_data(add_actions),
2358 ofpbuf_size(add_actions));
2360 ovs_mutex_unlock(&dp->flow_mutex);
2363 ofpbuf_uninit(&actions);
2364 ofpbuf_uninit(&put_actions);
2365 fat_rwlock_unlock(&dp->upcall_rwlock);
2369 for (i = 0; i < cnt; i++) {
2370 struct dpif_packet *packet = packets[i];
2371 struct dp_netdev_flow *flow;
2373 if (OVS_UNLIKELY(!rules[i] || !mfs[i])) {
2377 flow = dp_netdev_flow_cast(rules[i]);
2378 emc_insert(flow_cache, mfs[i], dpif_packet_get_dp_hash(packet), flow);
2379 dp_netdev_queue_batches(packet, md, flow, mfs[i], batches, &n_batches,
2380 ARRAY_SIZE(batches));
2383 for (i = 0; i < n_batches; i++) {
2384 packet_batch_execute(&batches[i], dp, flow_cache);
2389 dp_netdev_input(struct dp_netdev *dp, struct emc_cache *flow_cache,
2390 struct dpif_packet **packets, int cnt, struct pkt_metadata *md)
2392 #if !defined(__CHECKER__) && !defined(_WIN32)
2393 const size_t PKT_ARRAY_SIZE = cnt;
2395 /* Sparse or MSVC doesn't like variable length array. */
2396 enum { PKT_ARRAY_SIZE = NETDEV_MAX_RX_BATCH };
2398 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
2401 newcnt = emc_processing(dp, flow_cache, packets, cnt, md, keys);
2402 if (OVS_UNLIKELY(newcnt)) {
2403 fast_path_processing(dp, flow_cache, packets, newcnt, md, keys);
2407 struct dp_netdev_execute_aux {
2408 struct dp_netdev *dp;
2409 struct emc_cache *flow_cache;
2413 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
2416 struct dp_netdev *dp = get_dp_netdev(dpif);
2417 dp->upcall_aux = aux;
2422 dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
2423 struct pkt_metadata *md,
2424 const struct nlattr *a, bool may_steal)
2425 OVS_NO_THREAD_SAFETY_ANALYSIS
2427 struct dp_netdev_execute_aux *aux = aux_;
2428 uint32_t *depth = recirc_depth_get();
2429 struct dp_netdev *dp = aux->dp;
2430 int type = nl_attr_type(a);
2431 struct dp_netdev_port *p;
2434 switch ((enum ovs_action_attr)type) {
2435 case OVS_ACTION_ATTR_OUTPUT:
2436 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
2437 if (OVS_LIKELY(p)) {
2438 netdev_send(p->netdev, packets, cnt, may_steal);
2439 } else if (may_steal) {
2440 for (i = 0; i < cnt; i++) {
2441 dpif_packet_delete(packets[i]);
2446 case OVS_ACTION_ATTR_USERSPACE:
2447 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
2448 const struct nlattr *userdata;
2449 struct ofpbuf actions;
2452 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2453 ofpbuf_init(&actions, 0);
2455 for (i = 0; i < cnt; i++) {
2458 ofpbuf_clear(&actions);
2460 flow_extract(&packets[i]->ofpbuf, md, &flow);
2461 error = dp_netdev_upcall(dp, packets[i], &flow, NULL,
2462 DPIF_UC_ACTION, userdata, &actions,
2464 if (!error || error == ENOSPC) {
2465 dp_netdev_execute_actions(dp, &packets[i], 1, false, md,
2467 ofpbuf_data(&actions),
2468 ofpbuf_size(&actions));
2472 dpif_packet_delete(packets[i]);
2475 ofpbuf_uninit(&actions);
2476 fat_rwlock_unlock(&dp->upcall_rwlock);
2481 case OVS_ACTION_ATTR_HASH: {
2482 const struct ovs_action_hash *hash_act;
2485 hash_act = nl_attr_get(a);
2487 for (i = 0; i < cnt; i++) {
2489 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
2490 /* Hash need not be symmetric, nor does it need to include
2492 hash = hash_2words(dpif_packet_get_dp_hash(packets[i]),
2493 hash_act->hash_basis);
2495 VLOG_WARN("Unknown hash algorithm specified "
2496 "for the hash action.");
2501 hash = 1; /* 0 is not valid */
2507 dpif_packet_set_dp_hash(packets[i], hash);
2512 case OVS_ACTION_ATTR_RECIRC:
2513 if (*depth < MAX_RECIRC_DEPTH) {
2516 for (i = 0; i < cnt; i++) {
2517 struct dpif_packet *recirc_pkt;
2518 struct pkt_metadata recirc_md = *md;
2520 recirc_pkt = (may_steal) ? packets[i]
2521 : dpif_packet_clone(packets[i]);
2523 recirc_md.recirc_id = nl_attr_get_u32(a);
2525 /* Hash is private to each packet */
2526 recirc_md.dp_hash = dpif_packet_get_dp_hash(packets[i]);
2528 dp_netdev_input(dp, aux->flow_cache, &recirc_pkt, 1,
2535 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
2537 for (i = 0; i < cnt; i++) {
2538 dpif_packet_delete(packets[i]);
2544 case OVS_ACTION_ATTR_PUSH_VLAN:
2545 case OVS_ACTION_ATTR_POP_VLAN:
2546 case OVS_ACTION_ATTR_PUSH_MPLS:
2547 case OVS_ACTION_ATTR_POP_MPLS:
2548 case OVS_ACTION_ATTR_SET:
2549 case OVS_ACTION_ATTR_SET_MASKED:
2550 case OVS_ACTION_ATTR_SAMPLE:
2551 case OVS_ACTION_ATTR_UNSPEC:
2552 case __OVS_ACTION_ATTR_MAX:
2558 dp_netdev_execute_actions(struct dp_netdev *dp,
2559 struct dpif_packet **packets, int cnt,
2560 bool may_steal, struct pkt_metadata *md,
2561 struct emc_cache *flow_cache,
2562 const struct nlattr *actions, size_t actions_len)
2564 struct dp_netdev_execute_aux aux = {dp, flow_cache};
2566 odp_execute_actions(&aux, packets, cnt, may_steal, md, actions,
2567 actions_len, dp_execute_cb);
2570 const struct dpif_class dpif_netdev_class = {
2572 dpif_netdev_enumerate,
2573 dpif_netdev_port_open_type,
2576 dpif_netdev_destroy,
2579 dpif_netdev_get_stats,
2580 dpif_netdev_port_add,
2581 dpif_netdev_port_del,
2582 dpif_netdev_port_query_by_number,
2583 dpif_netdev_port_query_by_name,
2584 NULL, /* port_get_pid */
2585 dpif_netdev_port_dump_start,
2586 dpif_netdev_port_dump_next,
2587 dpif_netdev_port_dump_done,
2588 dpif_netdev_port_poll,
2589 dpif_netdev_port_poll_wait,
2590 dpif_netdev_flow_flush,
2591 dpif_netdev_flow_dump_create,
2592 dpif_netdev_flow_dump_destroy,
2593 dpif_netdev_flow_dump_thread_create,
2594 dpif_netdev_flow_dump_thread_destroy,
2595 dpif_netdev_flow_dump_next,
2596 dpif_netdev_operate,
2597 NULL, /* recv_set */
2598 NULL, /* handlers_set */
2599 dpif_netdev_queue_to_priority,
2601 NULL, /* recv_wait */
2602 NULL, /* recv_purge */
2603 dpif_netdev_register_upcall_cb,
2604 dpif_netdev_enable_upcall,
2605 dpif_netdev_disable_upcall,
2609 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2610 const char *argv[], void *aux OVS_UNUSED)
2612 struct dp_netdev_port *old_port;
2613 struct dp_netdev_port *new_port;
2614 struct dp_netdev *dp;
2617 ovs_mutex_lock(&dp_netdev_mutex);
2618 dp = shash_find_data(&dp_netdevs, argv[1]);
2619 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2620 ovs_mutex_unlock(&dp_netdev_mutex);
2621 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2624 ovs_refcount_ref(&dp->ref_cnt);
2625 ovs_mutex_unlock(&dp_netdev_mutex);
2627 ovs_mutex_lock(&dp->port_mutex);
2628 if (get_port_by_name(dp, argv[2], &old_port)) {
2629 unixctl_command_reply_error(conn, "unknown port");
2633 port_no = u32_to_odp(atoi(argv[3]));
2634 if (!port_no || port_no == ODPP_NONE) {
2635 unixctl_command_reply_error(conn, "bad port number");
2638 if (dp_netdev_lookup_port(dp, port_no)) {
2639 unixctl_command_reply_error(conn, "port number already in use");
2643 /* Remove old port. */
2644 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
2645 ovsrcu_postpone(free, old_port);
2647 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
2648 new_port = xmemdup(old_port, sizeof *old_port);
2649 new_port->port_no = port_no;
2650 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
2652 seq_change(dp->port_seq);
2653 unixctl_command_reply(conn, NULL);
2656 ovs_mutex_unlock(&dp->port_mutex);
2657 dp_netdev_unref(dp);
2661 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
2662 const char *argv[], void *aux OVS_UNUSED)
2664 struct dp_netdev_port *port;
2665 struct dp_netdev *dp;
2667 ovs_mutex_lock(&dp_netdev_mutex);
2668 dp = shash_find_data(&dp_netdevs, argv[1]);
2669 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2670 ovs_mutex_unlock(&dp_netdev_mutex);
2671 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2674 ovs_refcount_ref(&dp->ref_cnt);
2675 ovs_mutex_unlock(&dp_netdev_mutex);
2677 ovs_mutex_lock(&dp->port_mutex);
2678 if (get_port_by_name(dp, argv[2], &port)) {
2679 unixctl_command_reply_error(conn, "unknown port");
2680 } else if (port->port_no == ODPP_LOCAL) {
2681 unixctl_command_reply_error(conn, "can't delete local port");
2683 do_del_port(dp, port);
2684 unixctl_command_reply(conn, NULL);
2686 ovs_mutex_unlock(&dp->port_mutex);
2688 dp_netdev_unref(dp);
2692 dpif_dummy_register__(const char *type)
2694 struct dpif_class *class;
2696 class = xmalloc(sizeof *class);
2697 *class = dpif_netdev_class;
2698 class->type = xstrdup(type);
2699 dp_register_provider(class);
2703 dpif_dummy_register(bool override)
2710 dp_enumerate_types(&types);
2711 SSET_FOR_EACH (type, &types) {
2712 if (!dp_unregister_provider(type)) {
2713 dpif_dummy_register__(type);
2716 sset_destroy(&types);
2719 dpif_dummy_register__("dummy");
2721 unixctl_command_register("dpif-dummy/change-port-number",
2722 "DP PORT NEW-NUMBER",
2723 3, 3, dpif_dummy_change_port_number, NULL);
2724 unixctl_command_register("dpif-dummy/delete-port", "DP PORT",
2725 2, 2, dpif_dummy_delete_port, NULL);