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_port_input(struct dp_netdev *dp,
384 struct emc_cache *flow_cache,
385 struct dpif_packet **packets, int cnt,
388 static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
389 static void dp_netdev_disable_upcall(struct dp_netdev *);
391 static void emc_clear_entry(struct emc_entry *ce);
394 emc_cache_init(struct emc_cache *flow_cache)
398 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
399 flow_cache->entries[i].flow = NULL;
400 flow_cache->entries[i].hash = 0;
401 miniflow_initialize(&flow_cache->entries[i].mf.flow,
402 flow_cache->entries[i].mf.buf);
407 emc_cache_uninit(struct emc_cache *flow_cache)
411 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
412 emc_clear_entry(&flow_cache->entries[i]);
416 static struct dpif_netdev *
417 dpif_netdev_cast(const struct dpif *dpif)
419 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
420 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
423 static struct dp_netdev *
424 get_dp_netdev(const struct dpif *dpif)
426 return dpif_netdev_cast(dpif)->dp;
430 dpif_netdev_enumerate(struct sset *all_dps,
431 const struct dpif_class *dpif_class)
433 struct shash_node *node;
435 ovs_mutex_lock(&dp_netdev_mutex);
436 SHASH_FOR_EACH(node, &dp_netdevs) {
437 struct dp_netdev *dp = node->data;
438 if (dpif_class != dp->class) {
439 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
440 * If the class doesn't match, skip this dpif. */
443 sset_add(all_dps, node->name);
445 ovs_mutex_unlock(&dp_netdev_mutex);
451 dpif_netdev_class_is_dummy(const struct dpif_class *class)
453 return class != &dpif_netdev_class;
457 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
459 return strcmp(type, "internal") ? type
460 : dpif_netdev_class_is_dummy(class) ? "dummy"
465 create_dpif_netdev(struct dp_netdev *dp)
467 uint16_t netflow_id = hash_string(dp->name, 0);
468 struct dpif_netdev *dpif;
470 ovs_refcount_ref(&dp->ref_cnt);
472 dpif = xmalloc(sizeof *dpif);
473 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
475 dpif->last_port_seq = seq_read(dp->port_seq);
480 /* Choose an unused, non-zero port number and return it on success.
481 * Return ODPP_NONE on failure. */
483 choose_port(struct dp_netdev *dp, const char *name)
484 OVS_REQUIRES(dp->port_mutex)
488 if (dp->class != &dpif_netdev_class) {
492 /* If the port name begins with "br", start the number search at
493 * 100 to make writing tests easier. */
494 if (!strncmp(name, "br", 2)) {
498 /* If the port name contains a number, try to assign that port number.
499 * This can make writing unit tests easier because port numbers are
501 for (p = name; *p != '\0'; p++) {
502 if (isdigit((unsigned char) *p)) {
503 port_no = start_no + strtol(p, NULL, 10);
504 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
505 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
506 return u32_to_odp(port_no);
513 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
514 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
515 return u32_to_odp(port_no);
523 create_dp_netdev(const char *name, const struct dpif_class *class,
524 struct dp_netdev **dpp)
525 OVS_REQUIRES(dp_netdev_mutex)
527 struct dp_netdev *dp;
530 dp = xzalloc(sizeof *dp);
531 shash_add(&dp_netdevs, name, dp);
533 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
534 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
535 ovs_refcount_init(&dp->ref_cnt);
536 atomic_flag_clear(&dp->destroyed);
538 ovs_mutex_init(&dp->flow_mutex);
539 classifier_init(&dp->cls, NULL);
540 cmap_init(&dp->flow_table);
542 ovsthread_stats_init(&dp->stats);
544 ovs_mutex_init(&dp->port_mutex);
545 cmap_init(&dp->ports);
546 dp->port_seq = seq_create();
547 latch_init(&dp->exit_latch);
548 fat_rwlock_init(&dp->upcall_rwlock);
550 /* Disable upcalls by default. */
551 dp_netdev_disable_upcall(dp);
552 dp->upcall_aux = NULL;
553 dp->upcall_cb = NULL;
555 ovs_mutex_lock(&dp->port_mutex);
556 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
557 ovs_mutex_unlock(&dp->port_mutex);
563 ovs_mutex_init(&dp->emc_mutex);
564 emc_cache_init(&dp->flow_cache);
571 dpif_netdev_open(const struct dpif_class *class, const char *name,
572 bool create, struct dpif **dpifp)
574 struct dp_netdev *dp;
577 ovs_mutex_lock(&dp_netdev_mutex);
578 dp = shash_find_data(&dp_netdevs, name);
580 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
582 error = (dp->class != class ? EINVAL
587 *dpifp = create_dpif_netdev(dp);
590 ovs_mutex_unlock(&dp_netdev_mutex);
595 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
596 * through the 'dp_netdevs' shash while freeing 'dp'. */
598 dp_netdev_free(struct dp_netdev *dp)
599 OVS_REQUIRES(dp_netdev_mutex)
601 struct dp_netdev_port *port;
602 struct dp_netdev_stats *bucket;
605 shash_find_and_delete(&dp_netdevs, dp->name);
607 dp_netdev_set_pmd_threads(dp, 0);
608 free(dp->pmd_threads);
610 dp_netdev_flow_flush(dp);
611 ovs_mutex_lock(&dp->port_mutex);
612 CMAP_FOR_EACH (port, node, &dp->ports) {
613 do_del_port(dp, port);
615 ovs_mutex_unlock(&dp->port_mutex);
617 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
618 ovs_mutex_destroy(&bucket->mutex);
619 free_cacheline(bucket);
621 ovsthread_stats_destroy(&dp->stats);
623 classifier_destroy(&dp->cls);
624 cmap_destroy(&dp->flow_table);
625 ovs_mutex_destroy(&dp->flow_mutex);
626 seq_destroy(dp->port_seq);
627 cmap_destroy(&dp->ports);
628 fat_rwlock_destroy(&dp->upcall_rwlock);
629 latch_destroy(&dp->exit_latch);
631 emc_cache_uninit(&dp->flow_cache);
632 ovs_mutex_destroy(&dp->emc_mutex);
634 free(CONST_CAST(char *, dp->name));
639 dp_netdev_unref(struct dp_netdev *dp)
642 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
643 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
644 ovs_mutex_lock(&dp_netdev_mutex);
645 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
648 ovs_mutex_unlock(&dp_netdev_mutex);
653 dpif_netdev_close(struct dpif *dpif)
655 struct dp_netdev *dp = get_dp_netdev(dpif);
662 dpif_netdev_destroy(struct dpif *dpif)
664 struct dp_netdev *dp = get_dp_netdev(dpif);
666 if (!atomic_flag_test_and_set(&dp->destroyed)) {
667 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
668 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
677 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
679 struct dp_netdev *dp = get_dp_netdev(dpif);
680 struct dp_netdev_stats *bucket;
683 stats->n_flows = cmap_count(&dp->flow_table);
685 stats->n_hit = stats->n_missed = stats->n_lost = 0;
686 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
687 ovs_mutex_lock(&bucket->mutex);
688 stats->n_hit += bucket->n[DP_STAT_HIT];
689 stats->n_missed += bucket->n[DP_STAT_MISS];
690 stats->n_lost += bucket->n[DP_STAT_LOST];
691 ovs_mutex_unlock(&bucket->mutex);
693 stats->n_masks = UINT32_MAX;
694 stats->n_mask_hit = UINT64_MAX;
700 dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
704 for (i = 0; i < dp->n_pmd_threads; i++) {
705 struct pmd_thread *f = &dp->pmd_threads[i];
708 atomic_add_relaxed(&f->change_seq, 1, &old_seq);
713 hash_port_no(odp_port_t port_no)
715 return hash_int(odp_to_u32(port_no), 0);
719 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
721 OVS_REQUIRES(dp->port_mutex)
723 struct netdev_saved_flags *sf;
724 struct dp_netdev_port *port;
725 struct netdev *netdev;
726 enum netdev_flags flags;
727 const char *open_type;
731 /* XXX reject devices already in some dp_netdev. */
733 /* Open and validate network device. */
734 open_type = dpif_netdev_port_open_type(dp->class, type);
735 error = netdev_open(devname, open_type, &netdev);
739 /* XXX reject non-Ethernet devices */
741 netdev_get_flags(netdev, &flags);
742 if (flags & NETDEV_LOOPBACK) {
743 VLOG_ERR("%s: cannot add a loopback device", devname);
744 netdev_close(netdev);
748 port = xzalloc(sizeof *port);
749 port->port_no = port_no;
750 port->netdev = netdev;
751 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
752 port->type = xstrdup(type);
753 for (i = 0; i < netdev_n_rxq(netdev); i++) {
754 error = netdev_rxq_open(netdev, &port->rxq[i], i);
756 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
757 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
758 devname, ovs_strerror(errno));
759 netdev_close(netdev);
767 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
769 for (i = 0; i < netdev_n_rxq(netdev); i++) {
770 netdev_rxq_close(port->rxq[i]);
772 netdev_close(netdev);
780 if (netdev_is_pmd(netdev)) {
782 dp_netdev_set_pmd_threads(dp, NR_PMD_THREADS);
783 dp_netdev_reload_pmd_threads(dp);
785 ovs_refcount_init(&port->ref_cnt);
787 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
788 seq_change(dp->port_seq);
794 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
795 odp_port_t *port_nop)
797 struct dp_netdev *dp = get_dp_netdev(dpif);
798 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
799 const char *dpif_port;
803 ovs_mutex_lock(&dp->port_mutex);
804 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
805 if (*port_nop != ODPP_NONE) {
807 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
809 port_no = choose_port(dp, dpif_port);
810 error = port_no == ODPP_NONE ? EFBIG : 0;
814 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
816 ovs_mutex_unlock(&dp->port_mutex);
822 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
824 struct dp_netdev *dp = get_dp_netdev(dpif);
827 ovs_mutex_lock(&dp->port_mutex);
828 if (port_no == ODPP_LOCAL) {
831 struct dp_netdev_port *port;
833 error = get_port_by_number(dp, port_no, &port);
835 do_del_port(dp, port);
838 ovs_mutex_unlock(&dp->port_mutex);
844 is_valid_port_number(odp_port_t port_no)
846 return port_no != ODPP_NONE;
849 static struct dp_netdev_port *
850 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
852 struct dp_netdev_port *port;
854 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
855 if (port->port_no == port_no) {
863 get_port_by_number(struct dp_netdev *dp,
864 odp_port_t port_no, struct dp_netdev_port **portp)
866 if (!is_valid_port_number(port_no)) {
870 *portp = dp_netdev_lookup_port(dp, port_no);
871 return *portp ? 0 : ENOENT;
876 port_ref(struct dp_netdev_port *port)
879 ovs_refcount_ref(&port->ref_cnt);
884 port_try_ref(struct dp_netdev_port *port)
887 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
894 port_destroy__(struct dp_netdev_port *port)
896 int n_rxq = netdev_n_rxq(port->netdev);
899 netdev_close(port->netdev);
900 netdev_restore_flags(port->sf);
902 for (i = 0; i < n_rxq; i++) {
903 netdev_rxq_close(port->rxq[i]);
911 port_unref(struct dp_netdev_port *port)
913 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
914 ovsrcu_postpone(port_destroy__, port);
919 get_port_by_name(struct dp_netdev *dp,
920 const char *devname, struct dp_netdev_port **portp)
921 OVS_REQUIRES(dp->port_mutex)
923 struct dp_netdev_port *port;
925 CMAP_FOR_EACH (port, node, &dp->ports) {
926 if (!strcmp(netdev_get_name(port->netdev), devname)) {
935 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
936 OVS_REQUIRES(dp->port_mutex)
938 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
939 seq_change(dp->port_seq);
940 if (netdev_is_pmd(port->netdev)) {
941 dp_netdev_reload_pmd_threads(dp);
948 answer_port_query(const struct dp_netdev_port *port,
949 struct dpif_port *dpif_port)
951 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
952 dpif_port->type = xstrdup(port->type);
953 dpif_port->port_no = port->port_no;
957 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
958 struct dpif_port *dpif_port)
960 struct dp_netdev *dp = get_dp_netdev(dpif);
961 struct dp_netdev_port *port;
964 error = get_port_by_number(dp, port_no, &port);
965 if (!error && dpif_port) {
966 answer_port_query(port, dpif_port);
973 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
974 struct dpif_port *dpif_port)
976 struct dp_netdev *dp = get_dp_netdev(dpif);
977 struct dp_netdev_port *port;
980 ovs_mutex_lock(&dp->port_mutex);
981 error = get_port_by_name(dp, devname, &port);
982 if (!error && dpif_port) {
983 answer_port_query(port, dpif_port);
985 ovs_mutex_unlock(&dp->port_mutex);
991 dp_netdev_flow_free(struct dp_netdev_flow *flow)
993 struct dp_netdev_flow_stats *bucket;
996 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
997 ovs_mutex_destroy(&bucket->mutex);
998 free_cacheline(bucket);
1000 ovsthread_stats_destroy(&flow->stats);
1002 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
1003 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1007 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1009 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1010 ovsrcu_postpone(dp_netdev_flow_free, flow);
1015 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
1016 OVS_REQUIRES(dp->flow_mutex)
1018 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
1019 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1021 classifier_remove(&dp->cls, cr);
1022 cmap_remove(&dp->flow_table, node, flow_hash(&flow->flow, 0));
1025 dp_netdev_flow_unref(flow);
1029 dp_netdev_flow_flush(struct dp_netdev *dp)
1031 struct dp_netdev_flow *netdev_flow;
1033 ovs_mutex_lock(&dp->flow_mutex);
1034 CMAP_FOR_EACH (netdev_flow, node, &dp->flow_table) {
1035 dp_netdev_remove_flow(dp, netdev_flow);
1037 ovs_mutex_unlock(&dp->flow_mutex);
1041 dpif_netdev_flow_flush(struct dpif *dpif)
1043 struct dp_netdev *dp = get_dp_netdev(dpif);
1045 dp_netdev_flow_flush(dp);
1049 struct dp_netdev_port_state {
1050 struct cmap_position position;
1055 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1057 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1062 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1063 struct dpif_port *dpif_port)
1065 struct dp_netdev_port_state *state = state_;
1066 struct dp_netdev *dp = get_dp_netdev(dpif);
1067 struct cmap_node *node;
1070 node = cmap_next_position(&dp->ports, &state->position);
1072 struct dp_netdev_port *port;
1074 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1077 state->name = xstrdup(netdev_get_name(port->netdev));
1078 dpif_port->name = state->name;
1079 dpif_port->type = port->type;
1080 dpif_port->port_no = port->port_no;
1091 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1093 struct dp_netdev_port_state *state = state_;
1100 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1102 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1103 uint64_t new_port_seq;
1106 new_port_seq = seq_read(dpif->dp->port_seq);
1107 if (dpif->last_port_seq != new_port_seq) {
1108 dpif->last_port_seq = new_port_seq;
1118 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1120 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1122 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1125 static struct dp_netdev_flow *
1126 dp_netdev_flow_cast(const struct cls_rule *cr)
1128 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1131 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1133 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1137 emc_entry_alive(struct emc_entry *ce)
1139 return ce->flow && !ce->flow->dead;
1143 emc_clear_entry(struct emc_entry *ce)
1146 dp_netdev_flow_unref(ce->flow);
1152 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1153 const struct miniflow *mf, uint32_t hash)
1155 if (ce->flow != flow) {
1157 dp_netdev_flow_unref(ce->flow);
1160 if (dp_netdev_flow_ref(flow)) {
1167 miniflow_clone_inline(&ce->mf.flow, mf, count_1bits(mf->map));
1173 emc_insert(struct emc_cache *cache, const struct miniflow *mf, uint32_t hash,
1174 struct dp_netdev_flow *flow)
1176 struct emc_entry *to_be_replaced = NULL;
1177 struct emc_entry *current_entry;
1179 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, hash) {
1180 if (current_entry->hash == hash
1181 && miniflow_equal(¤t_entry->mf.flow, mf)) {
1183 /* We found the entry with the 'mf' miniflow */
1184 emc_change_entry(current_entry, flow, NULL, 0);
1188 /* Replacement policy: put the flow in an empty (not alive) entry, or
1189 * in the first entry where it can be */
1191 || (emc_entry_alive(to_be_replaced)
1192 && !emc_entry_alive(current_entry))
1193 || current_entry->hash < to_be_replaced->hash) {
1194 to_be_replaced = current_entry;
1197 /* We didn't find the miniflow in the cache.
1198 * The 'to_be_replaced' entry is where the new flow will be stored */
1200 emc_change_entry(to_be_replaced, flow, mf, hash);
1203 static inline struct dp_netdev_flow *
1204 emc_lookup(struct emc_cache *cache, const struct miniflow *mf, uint32_t hash)
1206 struct emc_entry *current_entry;
1208 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, hash) {
1209 if (current_entry->hash == hash && emc_entry_alive(current_entry)
1210 && miniflow_equal(¤t_entry->mf.flow, mf)) {
1212 /* We found the entry with the 'mf' miniflow */
1213 return current_entry->flow;
1220 static struct dp_netdev_flow *
1221 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
1223 struct dp_netdev_flow *netdev_flow;
1224 struct cls_rule *rule;
1226 classifier_lookup_miniflow_batch(&dp->cls, &key, &rule, 1);
1227 netdev_flow = dp_netdev_flow_cast(rule);
1232 static struct dp_netdev_flow *
1233 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1235 struct dp_netdev_flow *netdev_flow;
1237 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1239 if (flow_equal(&netdev_flow->flow, flow)) {
1248 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow,
1249 struct dpif_flow_stats *stats)
1251 struct dp_netdev_flow_stats *bucket;
1254 memset(stats, 0, sizeof *stats);
1255 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1256 ovs_mutex_lock(&bucket->mutex);
1257 stats->n_packets += bucket->packet_count;
1258 stats->n_bytes += bucket->byte_count;
1259 stats->used = MAX(stats->used, bucket->used);
1260 stats->tcp_flags |= bucket->tcp_flags;
1261 ovs_mutex_unlock(&bucket->mutex);
1266 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1267 struct ofpbuf *buffer, struct dpif_flow *flow)
1269 struct flow_wildcards wc;
1270 struct dp_netdev_actions *actions;
1272 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1273 odp_flow_key_from_mask(buffer, &wc.masks, &netdev_flow->flow,
1274 odp_to_u32(wc.masks.in_port.odp_port),
1276 flow->mask = ofpbuf_data(buffer);
1277 flow->mask_len = ofpbuf_size(buffer);
1279 actions = dp_netdev_flow_get_actions(netdev_flow);
1280 flow->actions = actions->actions;
1281 flow->actions_len = actions->size;
1283 get_dpif_flow_stats(netdev_flow, &flow->stats);
1287 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1288 const struct nlattr *mask_key,
1289 uint32_t mask_key_len, const struct flow *flow,
1293 enum odp_key_fitness fitness;
1295 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1297 /* This should not happen: it indicates that
1298 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1299 * disagree on the acceptable form of a mask. Log the problem
1300 * as an error, with enough details to enable debugging. */
1301 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1303 if (!VLOG_DROP_ERR(&rl)) {
1307 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1309 VLOG_ERR("internal error parsing flow mask %s (%s)",
1310 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1317 enum mf_field_id id;
1318 /* No mask key, unwildcard everything except fields whose
1319 * prerequisities are not met. */
1320 memset(mask, 0x0, sizeof *mask);
1322 for (id = 0; id < MFF_N_IDS; ++id) {
1323 /* Skip registers and metadata. */
1324 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1325 && id != MFF_METADATA) {
1326 const struct mf_field *mf = mf_from_id(id);
1327 if (mf_are_prereqs_ok(mf, flow)) {
1328 mf_mask_field(mf, mask);
1334 /* Force unwildcard the in_port.
1336 * We need to do this even in the case where we unwildcard "everything"
1337 * above because "everything" only includes the 16-bit OpenFlow port number
1338 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1339 * port number mask->in_port.odp_port. */
1340 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1346 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1351 if (odp_flow_key_to_flow(key, key_len, flow)) {
1352 /* This should not happen: it indicates that odp_flow_key_from_flow()
1353 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1354 * flow. Log the problem as an error, with enough details to enable
1356 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1358 if (!VLOG_DROP_ERR(&rl)) {
1362 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1363 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1370 in_port = flow->in_port.odp_port;
1371 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1379 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1381 struct dp_netdev *dp = get_dp_netdev(dpif);
1382 struct dp_netdev_flow *netdev_flow;
1386 error = dpif_netdev_flow_from_nlattrs(get->key, get->key_len, &key);
1391 netdev_flow = dp_netdev_find_flow(dp, &key);
1394 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->flow);
1403 dp_netdev_flow_add(struct dp_netdev *dp, struct match *match,
1404 const struct nlattr *actions, size_t actions_len)
1405 OVS_REQUIRES(dp->flow_mutex)
1407 struct dp_netdev_flow *netdev_flow;
1409 netdev_flow = xzalloc(sizeof *netdev_flow);
1410 *CONST_CAST(struct flow *, &netdev_flow->flow) = match->flow;
1412 ovs_refcount_init(&netdev_flow->ref_cnt);
1414 ovsthread_stats_init(&netdev_flow->stats);
1416 ovsrcu_set(&netdev_flow->actions,
1417 dp_netdev_actions_create(actions, actions_len));
1419 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1420 match, NETDEV_RULE_PRIORITY);
1421 cmap_insert(&dp->flow_table,
1422 CONST_CAST(struct cmap_node *, &netdev_flow->node),
1423 flow_hash(&match->flow, 0));
1424 classifier_insert(&dp->cls,
1425 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1427 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1428 struct ds ds = DS_EMPTY_INITIALIZER;
1430 ds_put_cstr(&ds, "flow_add: ");
1431 match_format(match, &ds, OFP_DEFAULT_PRIORITY);
1432 ds_put_cstr(&ds, ", actions:");
1433 format_odp_actions(&ds, actions, actions_len);
1435 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
1444 clear_stats(struct dp_netdev_flow *netdev_flow)
1446 struct dp_netdev_flow_stats *bucket;
1449 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1450 ovs_mutex_lock(&bucket->mutex);
1452 bucket->packet_count = 0;
1453 bucket->byte_count = 0;
1454 bucket->tcp_flags = 0;
1455 ovs_mutex_unlock(&bucket->mutex);
1460 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1462 struct dp_netdev *dp = get_dp_netdev(dpif);
1463 struct dp_netdev_flow *netdev_flow;
1464 struct miniflow miniflow;
1468 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
1472 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1473 put->mask, put->mask_len,
1474 &match.flow, &match.wc.masks);
1478 miniflow_init(&miniflow, &match.flow);
1480 ovs_mutex_lock(&dp->flow_mutex);
1481 netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
1483 if (put->flags & DPIF_FP_CREATE) {
1484 if (cmap_count(&dp->flow_table) < MAX_FLOWS) {
1486 memset(put->stats, 0, sizeof *put->stats);
1488 error = dp_netdev_flow_add(dp, &match, put->actions,
1497 if (put->flags & DPIF_FP_MODIFY
1498 && flow_equal(&match.flow, &netdev_flow->flow)) {
1499 struct dp_netdev_actions *new_actions;
1500 struct dp_netdev_actions *old_actions;
1502 new_actions = dp_netdev_actions_create(put->actions,
1505 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1506 ovsrcu_set(&netdev_flow->actions, new_actions);
1509 get_dpif_flow_stats(netdev_flow, put->stats);
1511 if (put->flags & DPIF_FP_ZERO_STATS) {
1512 clear_stats(netdev_flow);
1515 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1516 } else if (put->flags & DPIF_FP_CREATE) {
1519 /* Overlapping flow. */
1523 ovs_mutex_unlock(&dp->flow_mutex);
1524 miniflow_destroy(&miniflow);
1530 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1532 struct dp_netdev *dp = get_dp_netdev(dpif);
1533 struct dp_netdev_flow *netdev_flow;
1537 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1542 ovs_mutex_lock(&dp->flow_mutex);
1543 netdev_flow = dp_netdev_find_flow(dp, &key);
1546 get_dpif_flow_stats(netdev_flow, del->stats);
1548 dp_netdev_remove_flow(dp, netdev_flow);
1552 ovs_mutex_unlock(&dp->flow_mutex);
1557 struct dpif_netdev_flow_dump {
1558 struct dpif_flow_dump up;
1559 struct cmap_position pos;
1561 struct ovs_mutex mutex;
1564 static struct dpif_netdev_flow_dump *
1565 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1567 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1570 static struct dpif_flow_dump *
1571 dpif_netdev_flow_dump_create(const struct dpif *dpif_)
1573 struct dpif_netdev_flow_dump *dump;
1575 dump = xmalloc(sizeof *dump);
1576 dpif_flow_dump_init(&dump->up, dpif_);
1577 memset(&dump->pos, 0, sizeof dump->pos);
1579 ovs_mutex_init(&dump->mutex);
1585 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1587 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1589 ovs_mutex_destroy(&dump->mutex);
1594 struct dpif_netdev_flow_dump_thread {
1595 struct dpif_flow_dump_thread up;
1596 struct dpif_netdev_flow_dump *dump;
1597 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1598 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1601 static struct dpif_netdev_flow_dump_thread *
1602 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1604 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1607 static struct dpif_flow_dump_thread *
1608 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1610 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1611 struct dpif_netdev_flow_dump_thread *thread;
1613 thread = xmalloc(sizeof *thread);
1614 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1615 thread->dump = dump;
1620 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1622 struct dpif_netdev_flow_dump_thread *thread
1623 = dpif_netdev_flow_dump_thread_cast(thread_);
1629 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
1630 struct dpif_flow *flows, int max_flows)
1632 struct dpif_netdev_flow_dump_thread *thread
1633 = dpif_netdev_flow_dump_thread_cast(thread_);
1634 struct dpif_netdev_flow_dump *dump = thread->dump;
1635 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
1636 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
1637 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
1641 ovs_mutex_lock(&dump->mutex);
1642 if (!dump->status) {
1643 for (n_flows = 0; n_flows < MIN(max_flows, FLOW_DUMP_MAX_BATCH);
1645 struct cmap_node *node;
1647 node = cmap_next_position(&dp->flow_table, &dump->pos);
1652 netdev_flows[n_flows] = CONTAINER_OF(node, struct dp_netdev_flow,
1656 ovs_mutex_unlock(&dump->mutex);
1658 for (i = 0; i < n_flows; i++) {
1659 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
1660 struct odputil_keybuf *keybuf = &thread->keybuf[i];
1661 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
1662 struct dpif_flow *f = &flows[i];
1663 struct dp_netdev_actions *dp_actions;
1664 struct flow_wildcards wc;
1667 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1670 ofpbuf_use_stack(&buf, keybuf, sizeof *keybuf);
1671 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1672 netdev_flow->flow.in_port.odp_port, true);
1673 f->key = ofpbuf_data(&buf);
1674 f->key_len = ofpbuf_size(&buf);
1677 ofpbuf_use_stack(&buf, maskbuf, sizeof *maskbuf);
1678 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1679 odp_to_u32(wc.masks.in_port.odp_port),
1681 f->mask = ofpbuf_data(&buf);
1682 f->mask_len = ofpbuf_size(&buf);
1685 dp_actions = dp_netdev_flow_get_actions(netdev_flow);
1686 f->actions = dp_actions->actions;
1687 f->actions_len = dp_actions->size;
1690 get_dpif_flow_stats(netdev_flow, &f->stats);
1697 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1699 struct dp_netdev *dp = get_dp_netdev(dpif);
1700 struct dpif_packet packet, *pp;
1701 struct pkt_metadata *md = &execute->md;
1703 if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
1704 ofpbuf_size(execute->packet) > UINT16_MAX) {
1708 packet.ofpbuf = *execute->packet;
1711 ovs_mutex_lock(&dp->emc_mutex);
1712 dp_netdev_execute_actions(dp, &pp, 1, false, md,
1713 &dp->flow_cache, execute->actions,
1714 execute->actions_len);
1715 ovs_mutex_unlock(&dp->emc_mutex);
1717 /* Even though may_steal is set to false, some actions could modify or
1718 * reallocate the ofpbuf memory. We need to pass those changes to the
1720 *execute->packet = packet.ofpbuf;
1726 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
1730 for (i = 0; i < n_ops; i++) {
1731 struct dpif_op *op = ops[i];
1734 case DPIF_OP_FLOW_PUT:
1735 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
1738 case DPIF_OP_FLOW_DEL:
1739 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
1742 case DPIF_OP_EXECUTE:
1743 op->error = dpif_netdev_execute(dpif, &op->u.execute);
1746 case DPIF_OP_FLOW_GET:
1747 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
1754 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1755 uint32_t queue_id, uint32_t *priority)
1757 *priority = queue_id;
1762 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1763 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1765 struct dp_netdev_actions *
1766 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1768 struct dp_netdev_actions *netdev_actions;
1770 netdev_actions = xmalloc(sizeof *netdev_actions);
1771 netdev_actions->actions = xmemdup(actions, size);
1772 netdev_actions->size = size;
1774 return netdev_actions;
1777 struct dp_netdev_actions *
1778 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
1780 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
1784 dp_netdev_actions_free(struct dp_netdev_actions *actions)
1786 free(actions->actions);
1792 dp_netdev_process_rxq_port(struct dp_netdev *dp,
1793 struct emc_cache *flow_cache,
1794 struct dp_netdev_port *port,
1795 struct netdev_rxq *rxq)
1797 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1800 error = netdev_rxq_recv(rxq, packets, &cnt);
1802 dp_netdev_port_input(dp, flow_cache, packets, cnt, port->port_no);
1803 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1804 static struct vlog_rate_limit rl
1805 = VLOG_RATE_LIMIT_INIT(1, 5);
1807 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1808 netdev_get_name(port->netdev),
1809 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);
2409 dp_netdev_port_input(struct dp_netdev *dp, struct emc_cache *flow_cache,
2410 struct dpif_packet **packets, int cnt, odp_port_t port_no)
2412 uint32_t *recirc_depth = recirc_depth_get();
2413 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port_no);
2416 dp_netdev_input(dp, flow_cache, packets, cnt, &md);
2419 struct dp_netdev_execute_aux {
2420 struct dp_netdev *dp;
2421 struct emc_cache *flow_cache;
2425 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
2428 struct dp_netdev *dp = get_dp_netdev(dpif);
2429 dp->upcall_aux = aux;
2434 dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
2435 struct pkt_metadata *md,
2436 const struct nlattr *a, bool may_steal)
2437 OVS_NO_THREAD_SAFETY_ANALYSIS
2439 struct dp_netdev_execute_aux *aux = aux_;
2440 uint32_t *depth = recirc_depth_get();
2441 struct dp_netdev *dp = aux->dp;
2442 int type = nl_attr_type(a);
2443 struct dp_netdev_port *p;
2446 switch ((enum ovs_action_attr)type) {
2447 case OVS_ACTION_ATTR_OUTPUT:
2448 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
2449 if (OVS_LIKELY(p)) {
2450 netdev_send(p->netdev, packets, cnt, may_steal);
2451 } else if (may_steal) {
2452 for (i = 0; i < cnt; i++) {
2453 dpif_packet_delete(packets[i]);
2458 case OVS_ACTION_ATTR_USERSPACE:
2459 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
2460 const struct nlattr *userdata;
2461 struct ofpbuf actions;
2464 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2465 ofpbuf_init(&actions, 0);
2467 for (i = 0; i < cnt; i++) {
2470 ofpbuf_clear(&actions);
2472 flow_extract(&packets[i]->ofpbuf, md, &flow);
2473 error = dp_netdev_upcall(dp, packets[i], &flow, NULL,
2474 DPIF_UC_ACTION, userdata, &actions,
2476 if (!error || error == ENOSPC) {
2477 dp_netdev_execute_actions(dp, &packets[i], 1, false, md,
2479 ofpbuf_data(&actions),
2480 ofpbuf_size(&actions));
2484 dpif_packet_delete(packets[i]);
2487 ofpbuf_uninit(&actions);
2488 fat_rwlock_unlock(&dp->upcall_rwlock);
2493 case OVS_ACTION_ATTR_HASH: {
2494 const struct ovs_action_hash *hash_act;
2497 hash_act = nl_attr_get(a);
2499 for (i = 0; i < cnt; i++) {
2501 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
2502 /* Hash need not be symmetric, nor does it need to include
2504 hash = hash_2words(dpif_packet_get_dp_hash(packets[i]),
2505 hash_act->hash_basis);
2507 VLOG_WARN("Unknown hash algorithm specified "
2508 "for the hash action.");
2513 hash = 1; /* 0 is not valid */
2519 dpif_packet_set_dp_hash(packets[i], hash);
2524 case OVS_ACTION_ATTR_RECIRC:
2525 if (*depth < MAX_RECIRC_DEPTH) {
2528 for (i = 0; i < cnt; i++) {
2529 struct dpif_packet *recirc_pkt;
2530 struct pkt_metadata recirc_md = *md;
2532 recirc_pkt = (may_steal) ? packets[i]
2533 : dpif_packet_clone(packets[i]);
2535 recirc_md.recirc_id = nl_attr_get_u32(a);
2537 /* Hash is private to each packet */
2538 recirc_md.dp_hash = dpif_packet_get_dp_hash(packets[i]);
2540 dp_netdev_input(dp, aux->flow_cache, &recirc_pkt, 1,
2547 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
2549 for (i = 0; i < cnt; i++) {
2550 dpif_packet_delete(packets[i]);
2556 case OVS_ACTION_ATTR_PUSH_VLAN:
2557 case OVS_ACTION_ATTR_POP_VLAN:
2558 case OVS_ACTION_ATTR_PUSH_MPLS:
2559 case OVS_ACTION_ATTR_POP_MPLS:
2560 case OVS_ACTION_ATTR_SET:
2561 case OVS_ACTION_ATTR_SET_MASKED:
2562 case OVS_ACTION_ATTR_SAMPLE:
2563 case OVS_ACTION_ATTR_UNSPEC:
2564 case __OVS_ACTION_ATTR_MAX:
2570 dp_netdev_execute_actions(struct dp_netdev *dp,
2571 struct dpif_packet **packets, int cnt,
2572 bool may_steal, struct pkt_metadata *md,
2573 struct emc_cache *flow_cache,
2574 const struct nlattr *actions, size_t actions_len)
2576 struct dp_netdev_execute_aux aux = {dp, flow_cache};
2578 odp_execute_actions(&aux, packets, cnt, may_steal, md, actions,
2579 actions_len, dp_execute_cb);
2582 const struct dpif_class dpif_netdev_class = {
2584 dpif_netdev_enumerate,
2585 dpif_netdev_port_open_type,
2588 dpif_netdev_destroy,
2591 dpif_netdev_get_stats,
2592 dpif_netdev_port_add,
2593 dpif_netdev_port_del,
2594 dpif_netdev_port_query_by_number,
2595 dpif_netdev_port_query_by_name,
2596 NULL, /* port_get_pid */
2597 dpif_netdev_port_dump_start,
2598 dpif_netdev_port_dump_next,
2599 dpif_netdev_port_dump_done,
2600 dpif_netdev_port_poll,
2601 dpif_netdev_port_poll_wait,
2602 dpif_netdev_flow_flush,
2603 dpif_netdev_flow_dump_create,
2604 dpif_netdev_flow_dump_destroy,
2605 dpif_netdev_flow_dump_thread_create,
2606 dpif_netdev_flow_dump_thread_destroy,
2607 dpif_netdev_flow_dump_next,
2608 dpif_netdev_operate,
2609 NULL, /* recv_set */
2610 NULL, /* handlers_set */
2611 dpif_netdev_queue_to_priority,
2613 NULL, /* recv_wait */
2614 NULL, /* recv_purge */
2615 dpif_netdev_register_upcall_cb,
2616 dpif_netdev_enable_upcall,
2617 dpif_netdev_disable_upcall,
2621 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2622 const char *argv[], void *aux OVS_UNUSED)
2624 struct dp_netdev_port *old_port;
2625 struct dp_netdev_port *new_port;
2626 struct dp_netdev *dp;
2629 ovs_mutex_lock(&dp_netdev_mutex);
2630 dp = shash_find_data(&dp_netdevs, argv[1]);
2631 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2632 ovs_mutex_unlock(&dp_netdev_mutex);
2633 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2636 ovs_refcount_ref(&dp->ref_cnt);
2637 ovs_mutex_unlock(&dp_netdev_mutex);
2639 ovs_mutex_lock(&dp->port_mutex);
2640 if (get_port_by_name(dp, argv[2], &old_port)) {
2641 unixctl_command_reply_error(conn, "unknown port");
2645 port_no = u32_to_odp(atoi(argv[3]));
2646 if (!port_no || port_no == ODPP_NONE) {
2647 unixctl_command_reply_error(conn, "bad port number");
2650 if (dp_netdev_lookup_port(dp, port_no)) {
2651 unixctl_command_reply_error(conn, "port number already in use");
2655 /* Remove old port. */
2656 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
2657 ovsrcu_postpone(free, old_port);
2659 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
2660 new_port = xmemdup(old_port, sizeof *old_port);
2661 new_port->port_no = port_no;
2662 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
2664 seq_change(dp->port_seq);
2665 unixctl_command_reply(conn, NULL);
2668 ovs_mutex_unlock(&dp->port_mutex);
2669 dp_netdev_unref(dp);
2673 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
2674 const char *argv[], void *aux OVS_UNUSED)
2676 struct dp_netdev_port *port;
2677 struct dp_netdev *dp;
2679 ovs_mutex_lock(&dp_netdev_mutex);
2680 dp = shash_find_data(&dp_netdevs, argv[1]);
2681 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2682 ovs_mutex_unlock(&dp_netdev_mutex);
2683 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2686 ovs_refcount_ref(&dp->ref_cnt);
2687 ovs_mutex_unlock(&dp_netdev_mutex);
2689 ovs_mutex_lock(&dp->port_mutex);
2690 if (get_port_by_name(dp, argv[2], &port)) {
2691 unixctl_command_reply_error(conn, "unknown port");
2692 } else if (port->port_no == ODPP_LOCAL) {
2693 unixctl_command_reply_error(conn, "can't delete local port");
2695 do_del_port(dp, port);
2696 unixctl_command_reply(conn, NULL);
2698 ovs_mutex_unlock(&dp->port_mutex);
2700 dp_netdev_unref(dp);
2704 dpif_dummy_register__(const char *type)
2706 struct dpif_class *class;
2708 class = xmalloc(sizeof *class);
2709 *class = dpif_netdev_class;
2710 class->type = xstrdup(type);
2711 dp_register_provider(class);
2715 dpif_dummy_register(bool override)
2722 dp_enumerate_types(&types);
2723 SSET_FOR_EACH (type, &types) {
2724 if (!dp_unregister_provider(type)) {
2725 dpif_dummy_register__(type);
2728 sset_destroy(&types);
2731 dpif_dummy_register__("dummy");
2733 unixctl_command_register("dpif-dummy/change-port-number",
2734 "DP PORT NEW-NUMBER",
2735 3, 3, dpif_dummy_change_port_number, NULL);
2736 unixctl_command_register("dpif-dummy/delete-port", "DP PORT",
2737 2, 2, dpif_dummy_delete_port, NULL);