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>
36 #include "dp-packet.h"
38 #include "dpif-provider.h"
40 #include "dynamic-string.h"
41 #include "fat-rwlock.h"
47 #include "meta-flow.h"
49 #include "netdev-dpdk.h"
50 #include "netdev-vport.h"
52 #include "odp-execute.h"
54 #include "ofp-print.h"
59 #include "poll-loop.h"
66 #include "tnl-arp-cache.h"
69 #include "openvswitch/vlog.h"
71 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
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 with inline values */
92 struct netdev_flow_key {
93 uint32_t hash; /* Hash function differs for different users. */
94 uint32_t len; /* Length of the following miniflow (incl. map). */
96 uint64_t buf[FLOW_MAX_PACKET_U64S - MINI_N_INLINE];
99 /* Exact match cache for frequently used flows
101 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
102 * search its entries for a miniflow that matches exactly the miniflow of the
103 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
105 * A cache entry holds a reference to its 'dp_netdev_flow'.
107 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
108 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
109 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
110 * value is the index of a cache entry where the miniflow could be.
116 * Each pmd_thread has its own private exact match cache.
117 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
120 #define EM_FLOW_HASH_SHIFT 10
121 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
122 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
123 #define EM_FLOW_HASH_SEGS 2
126 struct dp_netdev_flow *flow;
127 struct netdev_flow_key key; /* key.hash used for emc hash value. */
131 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
132 int sweep_idx; /* For emc_cache_slow_sweep(). */
135 /* Iterate in the exact match cache through every entry that might contain a
136 * miniflow with hash 'HASH'. */
137 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
138 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
139 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
140 i__ < EM_FLOW_HASH_SEGS; \
141 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
143 /* Simple non-wildcarding single-priority classifier. */
146 struct cmap subtables_map;
147 struct pvector subtables;
150 /* A rule to be inserted to the classifier. */
152 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
153 struct netdev_flow_key *mask; /* Subtable's mask. */
154 struct netdev_flow_key flow; /* Matching key. */
155 /* 'flow' must be the last field, additional space is allocated here. */
158 static void dpcls_init(struct dpcls *);
159 static void dpcls_destroy(struct dpcls *);
160 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
161 const struct netdev_flow_key *mask);
162 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
163 static bool dpcls_lookup(const struct dpcls *cls,
164 const struct netdev_flow_key keys[],
165 struct dpcls_rule **rules, size_t cnt);
167 /* Datapath based on the network device interface from netdev.h.
173 * Some members, marked 'const', are immutable. Accessing other members
174 * requires synchronization, as noted in more detail below.
176 * Acquisition order is, from outermost to innermost:
178 * dp_netdev_mutex (global)
182 const struct dpif_class *const class;
183 const char *const name;
185 struct ovs_refcount ref_cnt;
186 atomic_flag destroyed;
190 * Protected by RCU. Take the mutex to add or remove ports. */
191 struct ovs_mutex port_mutex;
193 struct seq *port_seq; /* Incremented whenever a port changes. */
195 /* Protects access to ofproto-dpif-upcall interface during revalidator
196 * thread synchronization. */
197 struct fat_rwlock upcall_rwlock;
198 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
201 /* Stores all 'struct dp_netdev_pmd_thread's. */
202 struct cmap poll_threads;
204 /* Protects the access of the 'struct dp_netdev_pmd_thread'
205 * instance for non-pmd thread. */
206 struct ovs_mutex non_pmd_mutex;
208 /* Each pmd thread will store its pointer to
209 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
210 ovsthread_key_t per_pmd_key;
212 /* Number of rx queues for each dpdk interface and the cpu mask
213 * for pin of pmd threads. */
216 uint64_t last_tnl_conf_seq;
219 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
223 DP_STAT_HIT, /* Packets that matched in the flow table. */
224 DP_STAT_MISS, /* Packets that did not match. */
225 DP_STAT_LOST, /* Packets not passed up to the client. */
229 /* A port in a netdev-based datapath. */
230 struct dp_netdev_port {
231 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
233 struct netdev *netdev;
234 struct netdev_saved_flags *sf;
235 struct netdev_rxq **rxq;
236 struct ovs_refcount ref_cnt;
237 char *type; /* Port type as requested by user. */
240 /* Contained by struct dp_netdev_flow's 'stats' member. */
241 struct dp_netdev_flow_stats {
242 long long int used; /* Last used time, in monotonic msecs. */
243 long long int packet_count; /* Number of packets matched. */
244 long long int byte_count; /* Number of bytes matched. */
245 uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
248 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
254 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
255 * its pmd thread's classifier. The text below calls this classifier 'cls'.
260 * The thread safety rules described here for "struct dp_netdev_flow" are
261 * motivated by two goals:
263 * - Prevent threads that read members of "struct dp_netdev_flow" from
264 * reading bad data due to changes by some thread concurrently modifying
267 * - Prevent two threads making changes to members of a given "struct
268 * dp_netdev_flow" from interfering with each other.
274 * A flow 'flow' may be accessed without a risk of being freed during an RCU
275 * grace period. Code that needs to hold onto a flow for a while
276 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
278 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
279 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
282 * Some members, marked 'const', are immutable. Accessing other members
283 * requires synchronization, as noted in more detail below.
285 struct dp_netdev_flow {
288 /* Hash table index by unmasked flow. */
289 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
291 const ovs_u128 ufid; /* Unique flow identifier. */
292 const struct flow flow; /* Unmasked flow that created this entry. */
293 const int pmd_id; /* The 'core_id' of pmd thread owning this */
296 /* Number of references.
297 * The classifier owns one reference.
298 * Any thread trying to keep a rule from being freed should hold its own
300 struct ovs_refcount ref_cnt;
303 struct dp_netdev_flow_stats stats;
306 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
308 /* Packet classification. */
309 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
310 /* 'cr' must be the last member. */
313 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
314 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
315 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
318 /* A set of datapath actions within a "struct dp_netdev_flow".
324 * A struct dp_netdev_actions 'actions' is protected with RCU. */
325 struct dp_netdev_actions {
326 /* These members are immutable: they do not change during the struct's
328 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
329 unsigned int size; /* Size of 'actions', in bytes. */
332 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
334 struct dp_netdev_actions *dp_netdev_flow_get_actions(
335 const struct dp_netdev_flow *);
336 static void dp_netdev_actions_free(struct dp_netdev_actions *);
338 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
339 struct dp_netdev_pmd_stats {
340 /* Indexed by DP_STAT_*. */
341 unsigned long long int n[DP_N_STATS];
344 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
345 * the performance overhead of interrupt processing. Therefore netdev can
346 * not implement rx-wait for these devices. dpif-netdev needs to poll
347 * these device to check for recv buffer. pmd-thread does polling for
348 * devices assigned to itself.
350 * DPDK used PMD for accessing NIC.
352 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
353 * I/O of all non-pmd threads. There will be no actual thread created
356 * Each struct has its own flow table and classifier. Packets received
357 * from managed ports are looked up in the corresponding pmd thread's
358 * flow table, and are executed with the found actions.
360 struct dp_netdev_pmd_thread {
361 struct dp_netdev *dp;
362 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
363 struct cmap_node node; /* In 'dp->poll_threads'. */
365 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
366 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
368 /* Per thread exact-match cache. Note, the instance for cpu core
369 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
370 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
371 * instances will only be accessed by its own pmd thread. */
372 struct emc_cache flow_cache;
374 /* Classifier and Flow-Table.
376 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
377 * changes to 'cls' must be made while still holding the 'flow_mutex'.
379 struct ovs_mutex flow_mutex;
381 struct cmap flow_table OVS_GUARDED; /* Flow table. */
384 struct dp_netdev_pmd_stats stats;
386 struct latch exit_latch; /* For terminating the pmd thread. */
387 atomic_uint change_seq; /* For reloading pmd ports. */
389 int index; /* Idx of this pmd thread among pmd*/
390 /* threads on same numa node. */
391 int core_id; /* CPU core id of this pmd thread. */
392 int numa_id; /* numa node id of this pmd thread. */
395 #define PMD_INITIAL_SEQ 1
397 /* Interface to netdev-based datapath. */
400 struct dp_netdev *dp;
401 uint64_t last_port_seq;
404 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
405 struct dp_netdev_port **portp);
406 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
407 struct dp_netdev_port **portp);
408 static void dp_netdev_free(struct dp_netdev *)
409 OVS_REQUIRES(dp_netdev_mutex);
410 static int do_add_port(struct dp_netdev *dp, const char *devname,
411 const char *type, odp_port_t port_no)
412 OVS_REQUIRES(dp->port_mutex);
413 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
414 OVS_REQUIRES(dp->port_mutex);
415 static int dpif_netdev_open(const struct dpif_class *, const char *name,
416 bool create, struct dpif **);
417 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
418 struct dp_packet **, int c,
420 const struct nlattr *actions,
422 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
423 struct dp_packet **, int cnt);
425 static void dp_netdev_disable_upcall(struct dp_netdev *);
426 void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
427 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
428 struct dp_netdev *dp, int index,
429 int core_id, int numa_id);
430 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
431 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
432 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
434 static struct dp_netdev_pmd_thread *
435 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
436 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
437 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
438 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
439 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
440 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
441 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
442 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
444 static inline bool emc_entry_alive(struct emc_entry *ce);
445 static void emc_clear_entry(struct emc_entry *ce);
448 emc_cache_init(struct emc_cache *flow_cache)
452 BUILD_ASSERT(offsetof(struct miniflow, inline_values) == sizeof(uint64_t));
454 flow_cache->sweep_idx = 0;
455 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
456 flow_cache->entries[i].flow = NULL;
457 flow_cache->entries[i].key.hash = 0;
458 flow_cache->entries[i].key.len
459 = offsetof(struct miniflow, inline_values);
460 miniflow_initialize(&flow_cache->entries[i].key.mf,
461 flow_cache->entries[i].key.buf);
466 emc_cache_uninit(struct emc_cache *flow_cache)
470 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
471 emc_clear_entry(&flow_cache->entries[i]);
475 /* Check and clear dead flow references slowly (one entry at each
478 emc_cache_slow_sweep(struct emc_cache *flow_cache)
480 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
482 if (!emc_entry_alive(entry)) {
483 emc_clear_entry(entry);
485 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
488 static struct dpif_netdev *
489 dpif_netdev_cast(const struct dpif *dpif)
491 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
492 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
495 static struct dp_netdev *
496 get_dp_netdev(const struct dpif *dpif)
498 return dpif_netdev_cast(dpif)->dp;
502 dpif_netdev_enumerate(struct sset *all_dps,
503 const struct dpif_class *dpif_class)
505 struct shash_node *node;
507 ovs_mutex_lock(&dp_netdev_mutex);
508 SHASH_FOR_EACH(node, &dp_netdevs) {
509 struct dp_netdev *dp = node->data;
510 if (dpif_class != dp->class) {
511 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
512 * If the class doesn't match, skip this dpif. */
515 sset_add(all_dps, node->name);
517 ovs_mutex_unlock(&dp_netdev_mutex);
523 dpif_netdev_class_is_dummy(const struct dpif_class *class)
525 return class != &dpif_netdev_class;
529 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
531 return strcmp(type, "internal") ? type
532 : dpif_netdev_class_is_dummy(class) ? "dummy"
537 create_dpif_netdev(struct dp_netdev *dp)
539 uint16_t netflow_id = hash_string(dp->name, 0);
540 struct dpif_netdev *dpif;
542 ovs_refcount_ref(&dp->ref_cnt);
544 dpif = xmalloc(sizeof *dpif);
545 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
547 dpif->last_port_seq = seq_read(dp->port_seq);
552 /* Choose an unused, non-zero port number and return it on success.
553 * Return ODPP_NONE on failure. */
555 choose_port(struct dp_netdev *dp, const char *name)
556 OVS_REQUIRES(dp->port_mutex)
560 if (dp->class != &dpif_netdev_class) {
564 /* If the port name begins with "br", start the number search at
565 * 100 to make writing tests easier. */
566 if (!strncmp(name, "br", 2)) {
570 /* If the port name contains a number, try to assign that port number.
571 * This can make writing unit tests easier because port numbers are
573 for (p = name; *p != '\0'; p++) {
574 if (isdigit((unsigned char) *p)) {
575 port_no = start_no + strtol(p, NULL, 10);
576 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
577 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
578 return u32_to_odp(port_no);
585 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
586 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
587 return u32_to_odp(port_no);
595 create_dp_netdev(const char *name, const struct dpif_class *class,
596 struct dp_netdev **dpp)
597 OVS_REQUIRES(dp_netdev_mutex)
599 struct dp_netdev *dp;
602 dp = xzalloc(sizeof *dp);
603 shash_add(&dp_netdevs, name, dp);
605 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
606 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
607 ovs_refcount_init(&dp->ref_cnt);
608 atomic_flag_clear(&dp->destroyed);
610 ovs_mutex_init(&dp->port_mutex);
611 cmap_init(&dp->ports);
612 dp->port_seq = seq_create();
613 fat_rwlock_init(&dp->upcall_rwlock);
615 /* Disable upcalls by default. */
616 dp_netdev_disable_upcall(dp);
617 dp->upcall_aux = NULL;
618 dp->upcall_cb = NULL;
620 cmap_init(&dp->poll_threads);
621 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
622 ovsthread_key_create(&dp->per_pmd_key, NULL);
624 /* Reserves the core NON_PMD_CORE_ID for all non-pmd threads. */
625 ovs_numa_try_pin_core_specific(NON_PMD_CORE_ID);
626 dp_netdev_set_nonpmd(dp);
627 dp->n_dpdk_rxqs = NR_QUEUE;
629 ovs_mutex_lock(&dp->port_mutex);
630 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
631 ovs_mutex_unlock(&dp->port_mutex);
637 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
643 dpif_netdev_open(const struct dpif_class *class, const char *name,
644 bool create, struct dpif **dpifp)
646 struct dp_netdev *dp;
649 ovs_mutex_lock(&dp_netdev_mutex);
650 dp = shash_find_data(&dp_netdevs, name);
652 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
654 error = (dp->class != class ? EINVAL
659 *dpifp = create_dpif_netdev(dp);
662 ovs_mutex_unlock(&dp_netdev_mutex);
668 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
669 OVS_NO_THREAD_SAFETY_ANALYSIS
671 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
672 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
674 /* Before freeing a lock we should release it */
675 fat_rwlock_unlock(&dp->upcall_rwlock);
676 fat_rwlock_destroy(&dp->upcall_rwlock);
679 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
680 * through the 'dp_netdevs' shash while freeing 'dp'. */
682 dp_netdev_free(struct dp_netdev *dp)
683 OVS_REQUIRES(dp_netdev_mutex)
685 struct dp_netdev_port *port;
687 shash_find_and_delete(&dp_netdevs, dp->name);
689 dp_netdev_destroy_all_pmds(dp);
690 cmap_destroy(&dp->poll_threads);
691 ovs_mutex_destroy(&dp->non_pmd_mutex);
692 ovsthread_key_delete(dp->per_pmd_key);
694 ovs_mutex_lock(&dp->port_mutex);
695 CMAP_FOR_EACH (port, node, &dp->ports) {
696 do_del_port(dp, port);
698 ovs_mutex_unlock(&dp->port_mutex);
700 seq_destroy(dp->port_seq);
701 cmap_destroy(&dp->ports);
703 /* Upcalls must be disabled at this point */
704 dp_netdev_destroy_upcall_lock(dp);
707 free(CONST_CAST(char *, dp->name));
712 dp_netdev_unref(struct dp_netdev *dp)
715 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
716 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
717 ovs_mutex_lock(&dp_netdev_mutex);
718 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
721 ovs_mutex_unlock(&dp_netdev_mutex);
726 dpif_netdev_close(struct dpif *dpif)
728 struct dp_netdev *dp = get_dp_netdev(dpif);
735 dpif_netdev_destroy(struct dpif *dpif)
737 struct dp_netdev *dp = get_dp_netdev(dpif);
739 if (!atomic_flag_test_and_set(&dp->destroyed)) {
740 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
741 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
750 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
752 struct dp_netdev *dp = get_dp_netdev(dpif);
753 struct dp_netdev_pmd_thread *pmd;
755 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
756 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
757 stats->n_flows += cmap_count(&pmd->flow_table);
758 stats->n_hit += pmd->stats.n[DP_STAT_HIT];
759 stats->n_missed += pmd->stats.n[DP_STAT_MISS];
760 stats->n_lost += pmd->stats.n[DP_STAT_LOST];
762 stats->n_masks = UINT32_MAX;
763 stats->n_mask_hit = UINT64_MAX;
769 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
773 if (pmd->core_id == NON_PMD_CORE_ID) {
777 ovs_mutex_lock(&pmd->cond_mutex);
778 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
779 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
780 ovs_mutex_unlock(&pmd->cond_mutex);
783 /* Causes all pmd threads to reload its tx/rx devices.
784 * Must be called after adding/removing ports. */
786 dp_netdev_reload_pmds(struct dp_netdev *dp)
788 struct dp_netdev_pmd_thread *pmd;
790 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
791 dp_netdev_reload_pmd__(pmd);
796 hash_port_no(odp_port_t port_no)
798 return hash_int(odp_to_u32(port_no), 0);
802 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
804 OVS_REQUIRES(dp->port_mutex)
806 struct netdev_saved_flags *sf;
807 struct dp_netdev_port *port;
808 struct netdev *netdev;
809 enum netdev_flags flags;
810 const char *open_type;
814 /* XXX reject devices already in some dp_netdev. */
816 /* Open and validate network device. */
817 open_type = dpif_netdev_port_open_type(dp->class, type);
818 error = netdev_open(devname, open_type, &netdev);
822 /* XXX reject non-Ethernet devices */
824 netdev_get_flags(netdev, &flags);
825 if (flags & NETDEV_LOOPBACK) {
826 VLOG_ERR("%s: cannot add a loopback device", devname);
827 netdev_close(netdev);
831 if (netdev_is_pmd(netdev)) {
832 int n_cores = ovs_numa_get_n_cores();
834 if (n_cores == OVS_CORE_UNSPEC) {
835 VLOG_ERR("%s, cannot get cpu core info", devname);
838 /* There can only be ovs_numa_get_n_cores() pmd threads,
839 * so creates a txq for each. */
840 error = netdev_set_multiq(netdev, n_cores, dp->n_dpdk_rxqs);
841 if (error && (error != EOPNOTSUPP)) {
842 VLOG_ERR("%s, cannot set multiq", devname);
846 port = xzalloc(sizeof *port);
847 port->port_no = port_no;
848 port->netdev = netdev;
849 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
850 port->type = xstrdup(type);
851 for (i = 0; i < netdev_n_rxq(netdev); i++) {
852 error = netdev_rxq_open(netdev, &port->rxq[i], i);
854 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
855 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
856 devname, ovs_strerror(errno));
857 netdev_close(netdev);
865 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
867 for (i = 0; i < netdev_n_rxq(netdev); i++) {
868 netdev_rxq_close(port->rxq[i]);
870 netdev_close(netdev);
878 ovs_refcount_init(&port->ref_cnt);
879 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
881 if (netdev_is_pmd(netdev)) {
882 dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
883 dp_netdev_reload_pmds(dp);
885 seq_change(dp->port_seq);
891 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
892 odp_port_t *port_nop)
894 struct dp_netdev *dp = get_dp_netdev(dpif);
895 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
896 const char *dpif_port;
900 ovs_mutex_lock(&dp->port_mutex);
901 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
902 if (*port_nop != ODPP_NONE) {
904 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
906 port_no = choose_port(dp, dpif_port);
907 error = port_no == ODPP_NONE ? EFBIG : 0;
911 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
913 ovs_mutex_unlock(&dp->port_mutex);
919 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
921 struct dp_netdev *dp = get_dp_netdev(dpif);
924 ovs_mutex_lock(&dp->port_mutex);
925 if (port_no == ODPP_LOCAL) {
928 struct dp_netdev_port *port;
930 error = get_port_by_number(dp, port_no, &port);
932 do_del_port(dp, port);
935 ovs_mutex_unlock(&dp->port_mutex);
941 is_valid_port_number(odp_port_t port_no)
943 return port_no != ODPP_NONE;
946 static struct dp_netdev_port *
947 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
949 struct dp_netdev_port *port;
951 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
952 if (port->port_no == port_no) {
960 get_port_by_number(struct dp_netdev *dp,
961 odp_port_t port_no, struct dp_netdev_port **portp)
963 if (!is_valid_port_number(port_no)) {
967 *portp = dp_netdev_lookup_port(dp, port_no);
968 return *portp ? 0 : ENOENT;
973 port_ref(struct dp_netdev_port *port)
976 ovs_refcount_ref(&port->ref_cnt);
981 port_try_ref(struct dp_netdev_port *port)
984 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
991 port_unref(struct dp_netdev_port *port)
993 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
994 int n_rxq = netdev_n_rxq(port->netdev);
997 netdev_close(port->netdev);
998 netdev_restore_flags(port->sf);
1000 for (i = 0; i < n_rxq; i++) {
1001 netdev_rxq_close(port->rxq[i]);
1010 get_port_by_name(struct dp_netdev *dp,
1011 const char *devname, struct dp_netdev_port **portp)
1012 OVS_REQUIRES(dp->port_mutex)
1014 struct dp_netdev_port *port;
1016 CMAP_FOR_EACH (port, node, &dp->ports) {
1017 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1026 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1028 struct dp_netdev_pmd_thread *pmd;
1031 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1032 if (pmd->numa_id == numa_id) {
1040 /* Returns 'true' if there is a port with pmd netdev and the netdev
1041 * is on numa node 'numa_id'. */
1043 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1045 struct dp_netdev_port *port;
1047 CMAP_FOR_EACH (port, node, &dp->ports) {
1048 if (netdev_is_pmd(port->netdev)
1049 && netdev_get_numa_id(port->netdev) == numa_id) {
1059 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1060 OVS_REQUIRES(dp->port_mutex)
1062 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1063 seq_change(dp->port_seq);
1064 if (netdev_is_pmd(port->netdev)) {
1065 int numa_id = netdev_get_numa_id(port->netdev);
1067 /* If there is no netdev on the numa node, deletes the pmd threads
1068 * for that numa. Else, just reloads the queues. */
1069 if (!has_pmd_port_for_numa(dp, numa_id)) {
1070 dp_netdev_del_pmds_on_numa(dp, numa_id);
1072 dp_netdev_reload_pmds(dp);
1079 answer_port_query(const struct dp_netdev_port *port,
1080 struct dpif_port *dpif_port)
1082 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1083 dpif_port->type = xstrdup(port->type);
1084 dpif_port->port_no = port->port_no;
1088 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1089 struct dpif_port *dpif_port)
1091 struct dp_netdev *dp = get_dp_netdev(dpif);
1092 struct dp_netdev_port *port;
1095 error = get_port_by_number(dp, port_no, &port);
1096 if (!error && dpif_port) {
1097 answer_port_query(port, dpif_port);
1104 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1105 struct dpif_port *dpif_port)
1107 struct dp_netdev *dp = get_dp_netdev(dpif);
1108 struct dp_netdev_port *port;
1111 ovs_mutex_lock(&dp->port_mutex);
1112 error = get_port_by_name(dp, devname, &port);
1113 if (!error && dpif_port) {
1114 answer_port_query(port, dpif_port);
1116 ovs_mutex_unlock(&dp->port_mutex);
1122 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1124 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1128 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1130 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1131 ovsrcu_postpone(dp_netdev_flow_free, flow);
1136 dp_netdev_flow_hash(const ovs_u128 *ufid)
1138 return ufid->u32[0];
1142 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1143 struct dp_netdev_flow *flow)
1144 OVS_REQUIRES(pmd->flow_mutex)
1146 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1148 dpcls_remove(&pmd->cls, &flow->cr);
1149 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1152 dp_netdev_flow_unref(flow);
1156 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1158 struct dp_netdev_flow *netdev_flow;
1160 ovs_mutex_lock(&pmd->flow_mutex);
1161 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1162 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1164 ovs_mutex_unlock(&pmd->flow_mutex);
1168 dpif_netdev_flow_flush(struct dpif *dpif)
1170 struct dp_netdev *dp = get_dp_netdev(dpif);
1171 struct dp_netdev_pmd_thread *pmd;
1173 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1174 dp_netdev_pmd_flow_flush(pmd);
1180 struct dp_netdev_port_state {
1181 struct cmap_position position;
1186 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1188 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1193 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1194 struct dpif_port *dpif_port)
1196 struct dp_netdev_port_state *state = state_;
1197 struct dp_netdev *dp = get_dp_netdev(dpif);
1198 struct cmap_node *node;
1201 node = cmap_next_position(&dp->ports, &state->position);
1203 struct dp_netdev_port *port;
1205 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1208 state->name = xstrdup(netdev_get_name(port->netdev));
1209 dpif_port->name = state->name;
1210 dpif_port->type = port->type;
1211 dpif_port->port_no = port->port_no;
1222 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1224 struct dp_netdev_port_state *state = state_;
1231 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1233 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1234 uint64_t new_port_seq;
1237 new_port_seq = seq_read(dpif->dp->port_seq);
1238 if (dpif->last_port_seq != new_port_seq) {
1239 dpif->last_port_seq = new_port_seq;
1249 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1251 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1253 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1256 static struct dp_netdev_flow *
1257 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1259 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1262 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1264 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1267 /* netdev_flow_key utilities.
1269 * netdev_flow_key is basically a miniflow. We use these functions
1270 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1271 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1273 * - Since we are dealing exclusively with miniflows created by
1274 * miniflow_extract(), if the map is different the miniflow is different.
1275 * Therefore we can be faster by comparing the map and the miniflow in a
1277 * _ netdev_flow_key's miniflow has always inline values.
1278 * - These functions can be inlined by the compiler.
1280 * The following assertions make sure that what we're doing with miniflow is
1283 BUILD_ASSERT_DECL(offsetof(struct miniflow, inline_values)
1284 == sizeof(uint64_t));
1286 /* Given the number of bits set in the miniflow map, returns the size of the
1287 * 'netdev_flow_key.mf' */
1288 static inline uint32_t
1289 netdev_flow_key_size(uint32_t flow_u32s)
1291 return offsetof(struct miniflow, inline_values) +
1292 MINIFLOW_VALUES_SIZE(flow_u32s);
1296 netdev_flow_key_equal(const struct netdev_flow_key *a,
1297 const struct netdev_flow_key *b)
1299 /* 'b->len' may be not set yet. */
1300 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1303 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1304 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1305 * generated by miniflow_extract. */
1307 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1308 const struct miniflow *mf)
1310 return !memcmp(&key->mf, mf, key->len);
1314 netdev_flow_key_clone(struct netdev_flow_key *dst,
1315 const struct netdev_flow_key *src)
1318 offsetof(struct netdev_flow_key, mf) + src->len);
1323 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1324 const struct flow *src)
1326 struct dp_packet packet;
1327 uint64_t buf_stub[512 / 8];
1329 miniflow_initialize(&dst->mf, dst->buf);
1331 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1332 pkt_metadata_from_flow(&packet.md, src);
1333 flow_compose(&packet, src);
1334 miniflow_extract(&packet, &dst->mf);
1335 dp_packet_uninit(&packet);
1337 dst->len = netdev_flow_key_size(count_1bits(dst->mf.map));
1338 dst->hash = 0; /* Not computed yet. */
1341 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1343 netdev_flow_mask_init(struct netdev_flow_key *mask,
1344 const struct match *match)
1346 const uint64_t *mask_u64 = (const uint64_t *) &match->wc.masks;
1347 uint64_t *dst = mask->mf.inline_values;
1348 uint64_t map, mask_map = 0;
1352 /* Only check masks that make sense for the flow. */
1353 map = flow_wc_map(&match->flow);
1356 uint64_t rm1bit = rightmost_1bit(map);
1357 int i = raw_ctz(map);
1361 *dst++ = mask_u64[i];
1362 hash = hash_add64(hash, mask_u64[i]);
1367 mask->mf.values_inline = true;
1368 mask->mf.map = mask_map;
1370 hash = hash_add64(hash, mask_map);
1372 n = dst - mask->mf.inline_values;
1374 mask->hash = hash_finish(hash, n * 8);
1375 mask->len = netdev_flow_key_size(n);
1378 /* Initializes 'dst' as a copy of 'src' masked with 'mask'. */
1380 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1381 const struct flow *flow,
1382 const struct netdev_flow_key *mask)
1384 uint64_t *dst_u64 = dst->mf.inline_values;
1385 const uint64_t *mask_u64 = mask->mf.inline_values;
1389 dst->len = mask->len;
1390 dst->mf.values_inline = true;
1391 dst->mf.map = mask->mf.map;
1393 FLOW_FOR_EACH_IN_MAP(value, flow, mask->mf.map) {
1394 *dst_u64 = value & *mask_u64++;
1395 hash = hash_add64(hash, *dst_u64++);
1397 dst->hash = hash_finish(hash, (dst_u64 - dst->mf.inline_values) * 8);
1400 /* Iterate through all netdev_flow_key u64 values specified by 'MAP' */
1401 #define NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(VALUE, KEY, MAP) \
1402 for (struct mf_for_each_in_map_aux aux__ \
1403 = { (KEY)->mf.inline_values, (KEY)->mf.map, MAP }; \
1404 mf_get_next_in_map(&aux__, &(VALUE)); \
1407 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1409 static inline uint32_t
1410 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1411 const struct netdev_flow_key *mask)
1413 const uint64_t *p = mask->mf.inline_values;
1417 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(key_u64, key, mask->mf.map) {
1418 hash = hash_add64(hash, key_u64 & *p++);
1421 return hash_finish(hash, (p - mask->mf.inline_values) * 8);
1425 emc_entry_alive(struct emc_entry *ce)
1427 return ce->flow && !ce->flow->dead;
1431 emc_clear_entry(struct emc_entry *ce)
1434 dp_netdev_flow_unref(ce->flow);
1440 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1441 const struct netdev_flow_key *key)
1443 if (ce->flow != flow) {
1445 dp_netdev_flow_unref(ce->flow);
1448 if (dp_netdev_flow_ref(flow)) {
1455 netdev_flow_key_clone(&ce->key, key);
1460 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1461 struct dp_netdev_flow *flow)
1463 struct emc_entry *to_be_replaced = NULL;
1464 struct emc_entry *current_entry;
1466 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1467 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1468 /* We found the entry with the 'mf' miniflow */
1469 emc_change_entry(current_entry, flow, NULL);
1473 /* Replacement policy: put the flow in an empty (not alive) entry, or
1474 * in the first entry where it can be */
1476 || (emc_entry_alive(to_be_replaced)
1477 && !emc_entry_alive(current_entry))
1478 || current_entry->key.hash < to_be_replaced->key.hash) {
1479 to_be_replaced = current_entry;
1482 /* We didn't find the miniflow in the cache.
1483 * The 'to_be_replaced' entry is where the new flow will be stored */
1485 emc_change_entry(to_be_replaced, flow, key);
1488 static inline struct dp_netdev_flow *
1489 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1491 struct emc_entry *current_entry;
1493 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1494 if (current_entry->key.hash == key->hash
1495 && emc_entry_alive(current_entry)
1496 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1498 /* We found the entry with the 'key->mf' miniflow */
1499 return current_entry->flow;
1506 static struct dp_netdev_flow *
1507 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1508 const struct netdev_flow_key *key)
1510 struct dp_netdev_flow *netdev_flow;
1511 struct dpcls_rule *rule;
1513 dpcls_lookup(&pmd->cls, key, &rule, 1);
1514 netdev_flow = dp_netdev_flow_cast(rule);
1519 static struct dp_netdev_flow *
1520 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1521 const ovs_u128 *ufidp, const struct nlattr *key,
1524 struct dp_netdev_flow *netdev_flow;
1528 /* If a UFID is not provided, determine one based on the key. */
1529 if (!ufidp && key && key_len
1530 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1531 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1536 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1538 if (ovs_u128_equal(&netdev_flow->ufid, ufidp)) {
1548 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow,
1549 struct dpif_flow_stats *stats)
1551 stats->n_packets = netdev_flow->stats.packet_count;
1552 stats->n_bytes = netdev_flow->stats.byte_count;
1553 stats->used = netdev_flow->stats.used;
1554 stats->tcp_flags = netdev_flow->stats.tcp_flags;
1557 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1558 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1559 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1562 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1563 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1564 struct dpif_flow *flow, bool terse)
1567 memset(flow, 0, sizeof *flow);
1569 struct flow_wildcards wc;
1570 struct dp_netdev_actions *actions;
1573 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1576 offset = key_buf->size;
1577 flow->key = ofpbuf_tail(key_buf);
1578 odp_flow_key_from_flow(key_buf, &netdev_flow->flow, &wc.masks,
1579 netdev_flow->flow.in_port.odp_port, true);
1580 flow->key_len = key_buf->size - offset;
1583 offset = mask_buf->size;
1584 flow->mask = ofpbuf_tail(mask_buf);
1585 odp_flow_key_from_mask(mask_buf, &wc.masks, &netdev_flow->flow,
1586 odp_to_u32(wc.masks.in_port.odp_port),
1588 flow->mask_len = mask_buf->size - offset;
1591 actions = dp_netdev_flow_get_actions(netdev_flow);
1592 flow->actions = actions->actions;
1593 flow->actions_len = actions->size;
1596 flow->ufid = netdev_flow->ufid;
1597 flow->ufid_present = true;
1598 flow->pmd_id = netdev_flow->pmd_id;
1599 get_dpif_flow_stats(netdev_flow, &flow->stats);
1603 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1604 const struct nlattr *mask_key,
1605 uint32_t mask_key_len, const struct flow *flow,
1609 enum odp_key_fitness fitness;
1611 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1613 /* This should not happen: it indicates that
1614 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1615 * disagree on the acceptable form of a mask. Log the problem
1616 * as an error, with enough details to enable debugging. */
1617 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1619 if (!VLOG_DROP_ERR(&rl)) {
1623 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1625 VLOG_ERR("internal error parsing flow mask %s (%s)",
1626 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1633 enum mf_field_id id;
1634 /* No mask key, unwildcard everything except fields whose
1635 * prerequisities are not met. */
1636 memset(mask, 0x0, sizeof *mask);
1638 for (id = 0; id < MFF_N_IDS; ++id) {
1639 /* Skip registers and metadata. */
1640 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1641 && id != MFF_METADATA) {
1642 const struct mf_field *mf = mf_from_id(id);
1643 if (mf_are_prereqs_ok(mf, flow)) {
1644 mf_mask_field(mf, mask);
1650 /* Force unwildcard the in_port.
1652 * We need to do this even in the case where we unwildcard "everything"
1653 * above because "everything" only includes the 16-bit OpenFlow port number
1654 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1655 * port number mask->in_port.odp_port. */
1656 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1662 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1667 if (odp_flow_key_to_flow(key, key_len, flow)) {
1668 /* This should not happen: it indicates that odp_flow_key_from_flow()
1669 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1670 * flow. Log the problem as an error, with enough details to enable
1672 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1674 if (!VLOG_DROP_ERR(&rl)) {
1678 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1679 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1686 in_port = flow->in_port.odp_port;
1687 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1695 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1697 struct dp_netdev *dp = get_dp_netdev(dpif);
1698 struct dp_netdev_flow *netdev_flow;
1699 struct dp_netdev_pmd_thread *pmd;
1700 int pmd_id = get->pmd_id == PMD_ID_NULL ? NON_PMD_CORE_ID : get->pmd_id;
1703 pmd = dp_netdev_get_pmd(dp, pmd_id);
1708 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1711 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1716 dp_netdev_pmd_unref(pmd);
1722 static struct dp_netdev_flow *
1723 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
1724 struct match *match, const ovs_u128 *ufid,
1725 const struct nlattr *actions, size_t actions_len)
1726 OVS_REQUIRES(pmd->flow_mutex)
1728 struct dp_netdev_flow *flow;
1729 struct netdev_flow_key mask;
1731 netdev_flow_mask_init(&mask, match);
1732 /* Make sure wc does not have metadata. */
1733 ovs_assert(!(mask.mf.map & (MINIFLOW_MAP(metadata) | MINIFLOW_MAP(regs))));
1735 /* Do not allocate extra space. */
1736 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
1737 memset(&flow->stats, 0, sizeof flow->stats);
1739 *CONST_CAST(int *, &flow->pmd_id) = pmd->core_id;
1740 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
1741 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
1742 ovs_refcount_init(&flow->ref_cnt);
1743 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
1745 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
1746 dpcls_insert(&pmd->cls, &flow->cr, &mask);
1748 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
1749 dp_netdev_flow_hash(&flow->ufid));
1751 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1753 struct ds ds = DS_EMPTY_INITIALIZER;
1755 match.flow = flow->flow;
1756 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
1758 ds_put_cstr(&ds, "flow_add: ");
1759 odp_format_ufid(ufid, &ds);
1760 ds_put_cstr(&ds, " ");
1761 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
1762 ds_put_cstr(&ds, ", actions:");
1763 format_odp_actions(&ds, actions, actions_len);
1765 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
1774 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1776 struct dp_netdev *dp = get_dp_netdev(dpif);
1777 struct dp_netdev_flow *netdev_flow;
1778 struct netdev_flow_key key;
1779 struct dp_netdev_pmd_thread *pmd;
1782 int pmd_id = put->pmd_id == PMD_ID_NULL ? NON_PMD_CORE_ID : put->pmd_id;
1785 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
1789 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1790 put->mask, put->mask_len,
1791 &match.flow, &match.wc.masks);
1796 pmd = dp_netdev_get_pmd(dp, pmd_id);
1801 /* Must produce a netdev_flow_key for lookup.
1802 * This interface is no longer performance critical, since it is not used
1803 * for upcall processing any more. */
1804 netdev_flow_key_from_flow(&key, &match.flow);
1809 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
1812 ovs_mutex_lock(&pmd->flow_mutex);
1813 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
1815 if (put->flags & DPIF_FP_CREATE) {
1816 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
1818 memset(put->stats, 0, sizeof *put->stats);
1820 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
1830 if (put->flags & DPIF_FP_MODIFY
1831 && flow_equal(&match.flow, &netdev_flow->flow)) {
1832 struct dp_netdev_actions *new_actions;
1833 struct dp_netdev_actions *old_actions;
1835 new_actions = dp_netdev_actions_create(put->actions,
1838 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1839 ovsrcu_set(&netdev_flow->actions, new_actions);
1842 get_dpif_flow_stats(netdev_flow, put->stats);
1844 if (put->flags & DPIF_FP_ZERO_STATS) {
1845 memset(&netdev_flow->stats, 0, sizeof netdev_flow->stats);
1848 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1849 } else if (put->flags & DPIF_FP_CREATE) {
1852 /* Overlapping flow. */
1856 ovs_mutex_unlock(&pmd->flow_mutex);
1857 dp_netdev_pmd_unref(pmd);
1863 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1865 struct dp_netdev *dp = get_dp_netdev(dpif);
1866 struct dp_netdev_flow *netdev_flow;
1867 struct dp_netdev_pmd_thread *pmd;
1868 int pmd_id = del->pmd_id == PMD_ID_NULL ? NON_PMD_CORE_ID : del->pmd_id;
1871 pmd = dp_netdev_get_pmd(dp, pmd_id);
1876 ovs_mutex_lock(&pmd->flow_mutex);
1877 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
1881 get_dpif_flow_stats(netdev_flow, del->stats);
1883 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1887 ovs_mutex_unlock(&pmd->flow_mutex);
1888 dp_netdev_pmd_unref(pmd);
1893 struct dpif_netdev_flow_dump {
1894 struct dpif_flow_dump up;
1895 struct cmap_position poll_thread_pos;
1896 struct cmap_position flow_pos;
1897 struct dp_netdev_pmd_thread *cur_pmd;
1899 struct ovs_mutex mutex;
1902 static struct dpif_netdev_flow_dump *
1903 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1905 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1908 static struct dpif_flow_dump *
1909 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
1911 struct dpif_netdev_flow_dump *dump;
1913 dump = xzalloc(sizeof *dump);
1914 dpif_flow_dump_init(&dump->up, dpif_);
1915 dump->up.terse = terse;
1916 ovs_mutex_init(&dump->mutex);
1922 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1924 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1926 ovs_mutex_destroy(&dump->mutex);
1931 struct dpif_netdev_flow_dump_thread {
1932 struct dpif_flow_dump_thread up;
1933 struct dpif_netdev_flow_dump *dump;
1934 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1935 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1938 static struct dpif_netdev_flow_dump_thread *
1939 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1941 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1944 static struct dpif_flow_dump_thread *
1945 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1947 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1948 struct dpif_netdev_flow_dump_thread *thread;
1950 thread = xmalloc(sizeof *thread);
1951 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1952 thread->dump = dump;
1957 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1959 struct dpif_netdev_flow_dump_thread *thread
1960 = dpif_netdev_flow_dump_thread_cast(thread_);
1966 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
1967 struct dpif_flow *flows, int max_flows)
1969 struct dpif_netdev_flow_dump_thread *thread
1970 = dpif_netdev_flow_dump_thread_cast(thread_);
1971 struct dpif_netdev_flow_dump *dump = thread->dump;
1972 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
1976 ovs_mutex_lock(&dump->mutex);
1977 if (!dump->status) {
1978 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
1979 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
1980 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
1981 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
1983 /* First call to dump_next(), extracts the first pmd thread.
1984 * If there is no pmd thread, returns immediately. */
1986 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
1988 ovs_mutex_unlock(&dump->mutex);
1995 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
1996 struct cmap_node *node;
1998 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2002 netdev_flows[n_flows] = CONTAINER_OF(node,
2003 struct dp_netdev_flow,
2006 /* When finishing dumping the current pmd thread, moves to
2008 if (n_flows < flow_limit) {
2009 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2010 dp_netdev_pmd_unref(pmd);
2011 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2017 /* Keeps the reference to next caller. */
2018 dump->cur_pmd = pmd;
2020 /* If the current dump is empty, do not exit the loop, since the
2021 * remaining pmds could have flows to be dumped. Just dumps again
2022 * on the new 'pmd'. */
2025 ovs_mutex_unlock(&dump->mutex);
2027 for (i = 0; i < n_flows; i++) {
2028 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2029 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2030 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2031 struct dpif_flow *f = &flows[i];
2032 struct ofpbuf key, mask;
2034 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2035 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2036 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2044 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2045 OVS_NO_THREAD_SAFETY_ANALYSIS
2047 struct dp_netdev *dp = get_dp_netdev(dpif);
2048 struct dp_netdev_pmd_thread *pmd;
2049 struct dp_packet *pp;
2051 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2052 dp_packet_size(execute->packet) > UINT16_MAX) {
2056 /* Tries finding the 'pmd'. If NULL is returned, that means
2057 * the current thread is a non-pmd thread and should use
2058 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2059 pmd = ovsthread_getspecific(dp->per_pmd_key);
2061 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2064 /* If the current thread is non-pmd thread, acquires
2065 * the 'non_pmd_mutex'. */
2066 if (pmd->core_id == NON_PMD_CORE_ID) {
2067 ovs_mutex_lock(&dp->non_pmd_mutex);
2068 ovs_mutex_lock(&dp->port_mutex);
2071 pp = execute->packet;
2072 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2073 execute->actions_len);
2074 if (pmd->core_id == NON_PMD_CORE_ID) {
2075 dp_netdev_pmd_unref(pmd);
2076 ovs_mutex_unlock(&dp->port_mutex);
2077 ovs_mutex_unlock(&dp->non_pmd_mutex);
2084 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2088 for (i = 0; i < n_ops; i++) {
2089 struct dpif_op *op = ops[i];
2092 case DPIF_OP_FLOW_PUT:
2093 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2096 case DPIF_OP_FLOW_DEL:
2097 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2100 case DPIF_OP_EXECUTE:
2101 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2104 case DPIF_OP_FLOW_GET:
2105 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2111 /* Returns true if the configuration for rx queues or cpu mask
2114 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2116 if (dp->n_dpdk_rxqs != rxqs) {
2119 if (dp->pmd_cmask != NULL && cmask != NULL) {
2120 return strcmp(dp->pmd_cmask, cmask);
2122 return (dp->pmd_cmask != NULL || cmask != NULL);
2127 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2129 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2131 struct dp_netdev *dp = get_dp_netdev(dpif);
2133 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2134 struct dp_netdev_port *port;
2136 dp_netdev_destroy_all_pmds(dp);
2138 CMAP_FOR_EACH (port, node, &dp->ports) {
2139 if (netdev_is_pmd(port->netdev)) {
2142 /* Closes the existing 'rxq's. */
2143 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2144 netdev_rxq_close(port->rxq[i]);
2145 port->rxq[i] = NULL;
2148 /* Sets the new rx queue config. */
2149 err = netdev_set_multiq(port->netdev, ovs_numa_get_n_cores(),
2151 if (err && (err != EOPNOTSUPP)) {
2152 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2153 " %u", netdev_get_name(port->netdev),
2158 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2159 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2160 * netdev_n_rxq(port->netdev));
2161 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2162 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2166 dp->n_dpdk_rxqs = n_rxqs;
2168 /* Reconfigures the cpu mask. */
2169 ovs_numa_set_cpu_mask(cmask);
2170 free(dp->pmd_cmask);
2171 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2173 /* Restores the non-pmd. */
2174 dp_netdev_set_nonpmd(dp);
2175 /* Restores all pmd threads. */
2176 dp_netdev_reset_pmd_threads(dp);
2183 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2184 uint32_t queue_id, uint32_t *priority)
2186 *priority = queue_id;
2191 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
2192 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
2194 struct dp_netdev_actions *
2195 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2197 struct dp_netdev_actions *netdev_actions;
2199 netdev_actions = xmalloc(sizeof *netdev_actions);
2200 netdev_actions->actions = xmemdup(actions, size);
2201 netdev_actions->size = size;
2203 return netdev_actions;
2206 struct dp_netdev_actions *
2207 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2209 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2213 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2215 free(actions->actions);
2221 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2222 struct dp_netdev_port *port,
2223 struct netdev_rxq *rxq)
2225 struct dp_packet *packets[NETDEV_MAX_RX_BATCH];
2228 error = netdev_rxq_recv(rxq, packets, &cnt);
2232 *recirc_depth_get() = 0;
2234 /* XXX: initialize md in netdev implementation. */
2235 for (i = 0; i < cnt; i++) {
2236 packets[i]->md = PKT_METADATA_INITIALIZER(port->port_no);
2238 dp_netdev_input(pmd, packets, cnt);
2239 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2240 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2242 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2243 netdev_get_name(port->netdev), ovs_strerror(error));
2247 /* Return true if needs to revalidate datapath flows. */
2249 dpif_netdev_run(struct dpif *dpif)
2251 struct dp_netdev_port *port;
2252 struct dp_netdev *dp = get_dp_netdev(dpif);
2253 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2255 uint64_t new_tnl_seq;
2257 ovs_mutex_lock(&dp->non_pmd_mutex);
2258 CMAP_FOR_EACH (port, node, &dp->ports) {
2259 if (!netdev_is_pmd(port->netdev)) {
2262 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2263 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2267 ovs_mutex_unlock(&dp->non_pmd_mutex);
2268 dp_netdev_pmd_unref(non_pmd);
2270 tnl_arp_cache_run();
2271 new_tnl_seq = seq_read(tnl_conf_seq);
2273 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2274 dp->last_tnl_conf_seq = new_tnl_seq;
2281 dpif_netdev_wait(struct dpif *dpif)
2283 struct dp_netdev_port *port;
2284 struct dp_netdev *dp = get_dp_netdev(dpif);
2286 ovs_mutex_lock(&dp_netdev_mutex);
2287 CMAP_FOR_EACH (port, node, &dp->ports) {
2288 if (!netdev_is_pmd(port->netdev)) {
2291 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2292 netdev_rxq_wait(port->rxq[i]);
2296 ovs_mutex_unlock(&dp_netdev_mutex);
2297 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2301 struct dp_netdev_port *port;
2302 struct netdev_rxq *rx;
2306 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2307 struct rxq_poll **ppoll_list, int poll_cnt)
2309 struct rxq_poll *poll_list = *ppoll_list;
2310 struct dp_netdev_port *port;
2311 int n_pmds_on_numa, index, i;
2313 /* Simple scheduler for netdev rx polling. */
2314 for (i = 0; i < poll_cnt; i++) {
2315 port_unref(poll_list[i].port);
2319 n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
2322 CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
2323 /* Calls port_try_ref() to prevent the main thread
2324 * from deleting the port. */
2325 if (port_try_ref(port)) {
2326 if (netdev_is_pmd(port->netdev)
2327 && netdev_get_numa_id(port->netdev) == pmd->numa_id) {
2330 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2331 if ((index % n_pmds_on_numa) == pmd->index) {
2332 poll_list = xrealloc(poll_list,
2333 sizeof *poll_list * (poll_cnt + 1));
2336 poll_list[poll_cnt].port = port;
2337 poll_list[poll_cnt].rx = port->rxq[i];
2343 /* Unrefs the port_try_ref(). */
2348 *ppoll_list = poll_list;
2353 pmd_thread_main(void *f_)
2355 struct dp_netdev_pmd_thread *pmd = f_;
2356 unsigned int lc = 0;
2357 struct rxq_poll *poll_list;
2358 unsigned int port_seq = PMD_INITIAL_SEQ;
2365 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2366 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2367 pmd_thread_setaffinity_cpu(pmd->core_id);
2369 emc_cache_init(&pmd->flow_cache);
2370 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2372 /* Signal here to make sure the pmd finishes
2373 * reloading the updated configuration. */
2374 dp_netdev_pmd_reload_done(pmd);
2379 for (i = 0; i < poll_cnt; i++) {
2380 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2388 emc_cache_slow_sweep(&pmd->flow_cache);
2391 atomic_read_relaxed(&pmd->change_seq, &seq);
2392 if (seq != port_seq) {
2399 emc_cache_uninit(&pmd->flow_cache);
2401 if (!latch_is_set(&pmd->exit_latch)){
2405 for (i = 0; i < poll_cnt; i++) {
2406 port_unref(poll_list[i].port);
2409 dp_netdev_pmd_reload_done(pmd);
2416 dp_netdev_disable_upcall(struct dp_netdev *dp)
2417 OVS_ACQUIRES(dp->upcall_rwlock)
2419 fat_rwlock_wrlock(&dp->upcall_rwlock);
2423 dpif_netdev_disable_upcall(struct dpif *dpif)
2424 OVS_NO_THREAD_SAFETY_ANALYSIS
2426 struct dp_netdev *dp = get_dp_netdev(dpif);
2427 dp_netdev_disable_upcall(dp);
2431 dp_netdev_enable_upcall(struct dp_netdev *dp)
2432 OVS_RELEASES(dp->upcall_rwlock)
2434 fat_rwlock_unlock(&dp->upcall_rwlock);
2438 dpif_netdev_enable_upcall(struct dpif *dpif)
2439 OVS_NO_THREAD_SAFETY_ANALYSIS
2441 struct dp_netdev *dp = get_dp_netdev(dpif);
2442 dp_netdev_enable_upcall(dp);
2446 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2448 ovs_mutex_lock(&pmd->cond_mutex);
2449 xpthread_cond_signal(&pmd->cond);
2450 ovs_mutex_unlock(&pmd->cond_mutex);
2453 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2454 * the pointer if succeeds, otherwise, NULL.
2456 * Caller must unrefs the returned reference. */
2457 static struct dp_netdev_pmd_thread *
2458 dp_netdev_get_pmd(struct dp_netdev *dp, int core_id)
2460 struct dp_netdev_pmd_thread *pmd;
2461 const struct cmap_node *pnode;
2463 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2467 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2469 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2472 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2474 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2476 struct dp_netdev_pmd_thread *non_pmd;
2478 non_pmd = xzalloc(sizeof *non_pmd);
2479 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2483 /* Caller must have valid pointer to 'pmd'. */
2485 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2487 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2491 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2493 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2494 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2498 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2499 * fails, keeps checking for next node until reaching the end of cmap.
2501 * Caller must unrefs the returned reference. */
2502 static struct dp_netdev_pmd_thread *
2503 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2505 struct dp_netdev_pmd_thread *next;
2508 struct cmap_node *node;
2510 node = cmap_next_position(&dp->poll_threads, pos);
2511 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2513 } while (next && !dp_netdev_pmd_try_ref(next));
2518 /* Configures the 'pmd' based on the input argument. */
2520 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2521 int index, int core_id, int numa_id)
2525 pmd->core_id = core_id;
2526 pmd->numa_id = numa_id;
2528 ovs_refcount_init(&pmd->ref_cnt);
2529 latch_init(&pmd->exit_latch);
2530 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2531 xpthread_cond_init(&pmd->cond, NULL);
2532 ovs_mutex_init(&pmd->cond_mutex);
2533 ovs_mutex_init(&pmd->flow_mutex);
2534 dpcls_init(&pmd->cls);
2535 cmap_init(&pmd->flow_table);
2536 /* init the 'flow_cache' since there is no
2537 * actual thread created for NON_PMD_CORE_ID. */
2538 if (core_id == NON_PMD_CORE_ID) {
2539 emc_cache_init(&pmd->flow_cache);
2541 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2542 hash_int(core_id, 0));
2546 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2548 dp_netdev_pmd_flow_flush(pmd);
2549 dpcls_destroy(&pmd->cls);
2550 cmap_destroy(&pmd->flow_table);
2551 ovs_mutex_destroy(&pmd->flow_mutex);
2552 latch_destroy(&pmd->exit_latch);
2553 xpthread_cond_destroy(&pmd->cond);
2554 ovs_mutex_destroy(&pmd->cond_mutex);
2558 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2559 * and unrefs the struct. */
2561 dp_netdev_del_pmd(struct dp_netdev_pmd_thread *pmd)
2563 /* Uninit the 'flow_cache' since there is
2564 * no actual thread uninit it for NON_PMD_CORE_ID. */
2565 if (pmd->core_id == NON_PMD_CORE_ID) {
2566 emc_cache_uninit(&pmd->flow_cache);
2568 latch_set(&pmd->exit_latch);
2569 dp_netdev_reload_pmd__(pmd);
2570 ovs_numa_unpin_core(pmd->core_id);
2571 xpthread_join(pmd->thread, NULL);
2573 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2574 dp_netdev_pmd_unref(pmd);
2577 /* Destroys all pmd threads. */
2579 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2581 struct dp_netdev_pmd_thread *pmd;
2583 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2584 dp_netdev_del_pmd(pmd);
2588 /* Deletes all pmd threads on numa node 'numa_id'. */
2590 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2592 struct dp_netdev_pmd_thread *pmd;
2594 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2595 if (pmd->numa_id == numa_id) {
2596 dp_netdev_del_pmd(pmd);
2601 /* Checks the numa node id of 'netdev' and starts pmd threads for
2604 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2608 if (!ovs_numa_numa_id_is_valid(numa_id)) {
2609 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2610 "invalid", numa_id);
2614 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
2616 /* If there are already pmd threads created for the numa node
2617 * in which 'netdev' is on, do nothing. Else, creates the
2618 * pmd threads for the numa node. */
2620 int can_have, n_unpinned, i;
2622 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
2624 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2625 "cores on numa node");
2629 /* If cpu mask is specified, uses all unpinned cores, otherwise
2630 * tries creating NR_PMD_THREADS pmd threads. */
2631 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
2632 for (i = 0; i < can_have; i++) {
2633 struct dp_netdev_pmd_thread *pmd = xzalloc(sizeof *pmd);
2634 int core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
2636 dp_netdev_configure_pmd(pmd, dp, i, core_id, numa_id);
2637 /* Each thread will distribute all devices rx-queues among
2639 pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd);
2641 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
2646 /* Called after pmd threads config change. Restarts pmd threads with
2647 * new configuration. */
2649 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
2651 struct dp_netdev_port *port;
2653 CMAP_FOR_EACH (port, node, &dp->ports) {
2654 if (netdev_is_pmd(port->netdev)) {
2655 int numa_id = netdev_get_numa_id(port->netdev);
2657 dp_netdev_set_pmds_on_numa(dp, numa_id);
2663 dpif_netdev_get_datapath_version(void)
2665 return xstrdup("<built-in>");
2669 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
2672 long long int now = time_msec();
2674 netdev_flow->stats.used = MAX(now, netdev_flow->stats.used);
2675 netdev_flow->stats.packet_count += cnt;
2676 netdev_flow->stats.byte_count += size;
2677 netdev_flow->stats.tcp_flags |= tcp_flags;
2681 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
2682 enum dp_stat_type type, int cnt)
2684 pmd->stats.n[type] += cnt;
2688 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
2689 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
2690 enum dpif_upcall_type type, const struct nlattr *userdata,
2691 struct ofpbuf *actions, struct ofpbuf *put_actions)
2693 struct dp_netdev *dp = pmd->dp;
2695 if (type == DPIF_UC_MISS) {
2696 dp_netdev_count_packet(pmd, DP_STAT_MISS, 1);
2699 if (OVS_UNLIKELY(!dp->upcall_cb)) {
2703 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
2704 struct ds ds = DS_EMPTY_INITIALIZER;
2708 ofpbuf_init(&key, 0);
2709 odp_flow_key_from_flow(&key, flow, &wc->masks, flow->in_port.odp_port,
2711 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
2712 dp_packet_size(packet_));
2714 odp_flow_key_format(key.data, key.size, &ds);
2716 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
2717 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
2719 ofpbuf_uninit(&key);
2725 return dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
2726 actions, wc, put_actions, dp->upcall_aux);
2729 static inline uint32_t
2730 dpif_netdev_packet_get_dp_hash(struct dp_packet *packet,
2731 const struct miniflow *mf)
2735 hash = dp_packet_get_dp_hash(packet);
2736 if (OVS_UNLIKELY(!hash)) {
2737 hash = miniflow_hash_5tuple(mf, 0);
2738 dp_packet_set_dp_hash(packet, hash);
2743 struct packet_batch {
2744 unsigned int packet_count;
2745 unsigned int byte_count;
2748 struct dp_netdev_flow *flow;
2750 struct dp_packet *packets[NETDEV_MAX_RX_BATCH];
2754 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
2755 const struct miniflow *mf)
2757 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
2758 batch->packets[batch->packet_count++] = packet;
2759 batch->byte_count += dp_packet_size(packet);
2763 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
2767 batch->packet_count = 0;
2768 batch->byte_count = 0;
2769 batch->tcp_flags = 0;
2773 packet_batch_execute(struct packet_batch *batch,
2774 struct dp_netdev_pmd_thread *pmd)
2776 struct dp_netdev_actions *actions;
2777 struct dp_netdev_flow *flow = batch->flow;
2779 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
2782 actions = dp_netdev_flow_get_actions(flow);
2784 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
2785 actions->actions, actions->size);
2787 dp_netdev_count_packet(pmd, DP_STAT_HIT, batch->packet_count);
2791 dp_netdev_queue_batches(struct dp_packet *pkt,
2792 struct dp_netdev_flow *flow, const struct miniflow *mf,
2793 struct packet_batch *batches, size_t *n_batches,
2796 struct packet_batch *batch = NULL;
2799 if (OVS_UNLIKELY(!flow)) {
2802 /* XXX: This O(n^2) algortihm makes sense if we're operating under the
2803 * assumption that the number of distinct flows (and therefore the
2804 * number of distinct batches) is quite small. If this turns out not
2805 * to be the case, it may make sense to pre sort based on the
2806 * netdev_flow pointer. That done we can get the appropriate batching
2807 * in O(n * log(n)) instead. */
2808 for (j = *n_batches - 1; j >= 0; j--) {
2809 if (batches[j].flow == flow) {
2810 batch = &batches[j];
2811 packet_batch_update(batch, pkt, mf);
2815 if (OVS_UNLIKELY(*n_batches >= max_batches)) {
2819 batch = &batches[(*n_batches)++];
2820 packet_batch_init(batch, flow);
2821 packet_batch_update(batch, pkt, mf);
2826 dp_packet_swap(struct dp_packet **a, struct dp_packet **b)
2828 struct dp_packet *tmp = *a;
2833 /* Try to process all ('cnt') the 'packets' using only the exact match cache
2834 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', or if there
2835 * is no matching batch for a packet's flow, the miniflow is copied into 'keys'
2836 * and the packet pointer is moved at the beginning of the 'packets' array.
2838 * The function returns the number of packets that needs to be processed in the
2839 * 'packets' array (they have been moved to the beginning of the vector).
2841 static inline size_t
2842 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
2843 size_t cnt, struct netdev_flow_key *keys)
2845 struct netdev_flow_key key;
2846 struct packet_batch batches[4];
2847 struct emc_cache *flow_cache = &pmd->flow_cache;
2848 size_t n_batches, i;
2849 size_t notfound_cnt = 0;
2852 miniflow_initialize(&key.mf, key.buf);
2853 for (i = 0; i < cnt; i++) {
2854 struct dp_netdev_flow *flow;
2856 if (OVS_UNLIKELY(dp_packet_size(packets[i]) < ETH_HEADER_LEN)) {
2857 dp_packet_delete(packets[i]);
2861 miniflow_extract(packets[i], &key.mf);
2862 key.len = 0; /* Not computed yet. */
2863 key.hash = dpif_netdev_packet_get_dp_hash(packets[i], &key.mf);
2865 flow = emc_lookup(flow_cache, &key);
2866 if (OVS_UNLIKELY(!dp_netdev_queue_batches(packets[i], flow, &key.mf,
2867 batches, &n_batches,
2868 ARRAY_SIZE(batches)))) {
2869 if (i != notfound_cnt) {
2870 dp_packet_swap(&packets[i], &packets[notfound_cnt]);
2873 keys[notfound_cnt++] = key;
2877 for (i = 0; i < n_batches; i++) {
2878 packet_batch_execute(&batches[i], pmd);
2881 return notfound_cnt;
2885 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
2886 struct dp_packet **packets, size_t cnt,
2887 struct netdev_flow_key *keys)
2889 #if !defined(__CHECKER__) && !defined(_WIN32)
2890 const size_t PKT_ARRAY_SIZE = cnt;
2892 /* Sparse or MSVC doesn't like variable length array. */
2893 enum { PKT_ARRAY_SIZE = NETDEV_MAX_RX_BATCH };
2895 struct packet_batch batches[PKT_ARRAY_SIZE];
2896 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
2897 struct dp_netdev *dp = pmd->dp;
2898 struct emc_cache *flow_cache = &pmd->flow_cache;
2899 size_t n_batches, i;
2902 for (i = 0; i < cnt; i++) {
2903 /* Key length is needed in all the cases, hash computed on demand. */
2904 keys[i].len = netdev_flow_key_size(count_1bits(keys[i].mf.map));
2906 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
2907 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
2908 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
2909 struct ofpbuf actions, put_actions;
2912 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
2913 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
2915 for (i = 0; i < cnt; i++) {
2916 struct dp_netdev_flow *netdev_flow;
2917 struct ofpbuf *add_actions;
2921 if (OVS_LIKELY(rules[i])) {
2925 /* It's possible that an earlier slow path execution installed
2926 * a rule covering this flow. In this case, it's a lot cheaper
2927 * to catch it here than execute a miss. */
2928 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
2930 rules[i] = &netdev_flow->cr;
2934 miniflow_expand(&keys[i].mf, &match.flow);
2936 ofpbuf_clear(&actions);
2937 ofpbuf_clear(&put_actions);
2939 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
2940 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
2941 &ufid, DPIF_UC_MISS, NULL, &actions,
2943 if (OVS_UNLIKELY(error && error != ENOSPC)) {
2947 /* We can't allow the packet batching in the next loop to execute
2948 * the actions. Otherwise, if there are any slow path actions,
2949 * we'll send the packet up twice. */
2950 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
2951 actions.data, actions.size);
2953 add_actions = put_actions.size ? &put_actions : &actions;
2954 if (OVS_LIKELY(error != ENOSPC)) {
2955 /* XXX: There's a race window where a flow covering this packet
2956 * could have already been installed since we last did the flow
2957 * lookup before upcall. This could be solved by moving the
2958 * mutex lock outside the loop, but that's an awful long time
2959 * to be locking everyone out of making flow installs. If we
2960 * move to a per-core classifier, it would be reasonable. */
2961 ovs_mutex_lock(&pmd->flow_mutex);
2962 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
2963 if (OVS_LIKELY(!netdev_flow)) {
2964 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
2968 ovs_mutex_unlock(&pmd->flow_mutex);
2970 emc_insert(flow_cache, &keys[i], netdev_flow);
2974 ofpbuf_uninit(&actions);
2975 ofpbuf_uninit(&put_actions);
2976 fat_rwlock_unlock(&dp->upcall_rwlock);
2977 } else if (OVS_UNLIKELY(any_miss)) {
2978 int dropped_cnt = 0;
2980 for (i = 0; i < cnt; i++) {
2981 if (OVS_UNLIKELY(!rules[i])) {
2982 dp_packet_delete(packets[i]);
2987 dp_netdev_count_packet(pmd, DP_STAT_LOST, dropped_cnt);
2991 for (i = 0; i < cnt; i++) {
2992 struct dp_packet *packet = packets[i];
2993 struct dp_netdev_flow *flow;
2995 if (OVS_UNLIKELY(!rules[i])) {
2999 flow = dp_netdev_flow_cast(rules[i]);
3001 emc_insert(flow_cache, &keys[i], flow);
3002 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches,
3003 &n_batches, ARRAY_SIZE(batches));
3006 for (i = 0; i < n_batches; i++) {
3007 packet_batch_execute(&batches[i], pmd);
3012 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3013 struct dp_packet **packets, int cnt)
3015 #if !defined(__CHECKER__) && !defined(_WIN32)
3016 const size_t PKT_ARRAY_SIZE = cnt;
3018 /* Sparse or MSVC doesn't like variable length array. */
3019 enum { PKT_ARRAY_SIZE = NETDEV_MAX_RX_BATCH };
3021 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3024 newcnt = emc_processing(pmd, packets, cnt, keys);
3025 if (OVS_UNLIKELY(newcnt)) {
3026 fast_path_processing(pmd, packets, newcnt, keys);
3030 struct dp_netdev_execute_aux {
3031 struct dp_netdev_pmd_thread *pmd;
3035 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3038 struct dp_netdev *dp = get_dp_netdev(dpif);
3039 dp->upcall_aux = aux;
3044 dp_netdev_drop_packets(struct dp_packet ** packets, int cnt, bool may_steal)
3049 for (i = 0; i < cnt; i++) {
3050 dp_packet_delete(packets[i]);
3056 push_tnl_action(const struct dp_netdev *dp,
3057 const struct nlattr *attr,
3058 struct dp_packet **packets, int cnt)
3060 struct dp_netdev_port *tun_port;
3061 const struct ovs_action_push_tnl *data;
3063 data = nl_attr_get(attr);
3065 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3069 netdev_push_header(tun_port->netdev, packets, cnt, data);
3075 dp_netdev_clone_pkt_batch(struct dp_packet **tnl_pkt,
3076 struct dp_packet **packets, int cnt)
3080 for (i = 0; i < cnt; i++) {
3081 tnl_pkt[i] = dp_packet_clone(packets[i]);
3086 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3087 const struct nlattr *a, bool may_steal)
3088 OVS_NO_THREAD_SAFETY_ANALYSIS
3090 struct dp_netdev_execute_aux *aux = aux_;
3091 uint32_t *depth = recirc_depth_get();
3092 struct dp_netdev_pmd_thread *pmd= aux->pmd;
3093 struct dp_netdev *dp= pmd->dp;
3094 int type = nl_attr_type(a);
3095 struct dp_netdev_port *p;
3098 switch ((enum ovs_action_attr)type) {
3099 case OVS_ACTION_ATTR_OUTPUT:
3100 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3101 if (OVS_LIKELY(p)) {
3102 netdev_send(p->netdev, pmd->core_id, packets, cnt, may_steal);
3107 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3108 if (*depth < MAX_RECIRC_DEPTH) {
3109 struct dp_packet *tnl_pkt[NETDEV_MAX_RX_BATCH];
3113 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3117 err = push_tnl_action(dp, a, packets, cnt);
3120 dp_netdev_input(pmd, packets, cnt);
3123 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3129 case OVS_ACTION_ATTR_TUNNEL_POP:
3130 if (*depth < MAX_RECIRC_DEPTH) {
3131 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3133 p = dp_netdev_lookup_port(dp, portno);
3135 struct dp_packet *tnl_pkt[NETDEV_MAX_RX_BATCH];
3139 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3143 err = netdev_pop_header(p->netdev, packets, cnt);
3146 for (i = 0; i < cnt; i++) {
3147 packets[i]->md.in_port.odp_port = portno;
3151 dp_netdev_input(pmd, packets, cnt);
3154 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3161 case OVS_ACTION_ATTR_USERSPACE:
3162 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3163 const struct nlattr *userdata;
3164 struct ofpbuf actions;
3168 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3169 ofpbuf_init(&actions, 0);
3171 for (i = 0; i < cnt; i++) {
3174 ofpbuf_clear(&actions);
3176 flow_extract(packets[i], &flow);
3177 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3178 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3179 DPIF_UC_ACTION, userdata,&actions,
3181 if (!error || error == ENOSPC) {
3182 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3183 actions.data, actions.size);
3184 } else if (may_steal) {
3185 dp_packet_delete(packets[i]);
3188 ofpbuf_uninit(&actions);
3189 fat_rwlock_unlock(&dp->upcall_rwlock);
3195 case OVS_ACTION_ATTR_RECIRC:
3196 if (*depth < MAX_RECIRC_DEPTH) {
3199 for (i = 0; i < cnt; i++) {
3200 struct dp_packet *recirc_pkt;
3202 recirc_pkt = (may_steal) ? packets[i]
3203 : dp_packet_clone(packets[i]);
3205 recirc_pkt->md.recirc_id = nl_attr_get_u32(a);
3207 /* Hash is private to each packet */
3208 recirc_pkt->md.dp_hash = dp_packet_get_dp_hash(packets[i]);
3210 dp_netdev_input(pmd, &recirc_pkt, 1);
3217 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3220 case OVS_ACTION_ATTR_PUSH_VLAN:
3221 case OVS_ACTION_ATTR_POP_VLAN:
3222 case OVS_ACTION_ATTR_PUSH_MPLS:
3223 case OVS_ACTION_ATTR_POP_MPLS:
3224 case OVS_ACTION_ATTR_SET:
3225 case OVS_ACTION_ATTR_SET_MASKED:
3226 case OVS_ACTION_ATTR_SAMPLE:
3227 case OVS_ACTION_ATTR_HASH:
3228 case OVS_ACTION_ATTR_UNSPEC:
3229 case __OVS_ACTION_ATTR_MAX:
3233 dp_netdev_drop_packets(packets, cnt, may_steal);
3237 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3238 struct dp_packet **packets, int cnt,
3240 const struct nlattr *actions, size_t actions_len)
3242 struct dp_netdev_execute_aux aux = { pmd };
3244 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3245 actions_len, dp_execute_cb);
3248 const struct dpif_class dpif_netdev_class = {
3250 dpif_netdev_enumerate,
3251 dpif_netdev_port_open_type,
3254 dpif_netdev_destroy,
3257 dpif_netdev_get_stats,
3258 dpif_netdev_port_add,
3259 dpif_netdev_port_del,
3260 dpif_netdev_port_query_by_number,
3261 dpif_netdev_port_query_by_name,
3262 NULL, /* port_get_pid */
3263 dpif_netdev_port_dump_start,
3264 dpif_netdev_port_dump_next,
3265 dpif_netdev_port_dump_done,
3266 dpif_netdev_port_poll,
3267 dpif_netdev_port_poll_wait,
3268 dpif_netdev_flow_flush,
3269 dpif_netdev_flow_dump_create,
3270 dpif_netdev_flow_dump_destroy,
3271 dpif_netdev_flow_dump_thread_create,
3272 dpif_netdev_flow_dump_thread_destroy,
3273 dpif_netdev_flow_dump_next,
3274 dpif_netdev_operate,
3275 NULL, /* recv_set */
3276 NULL, /* handlers_set */
3277 dpif_netdev_pmd_set,
3278 dpif_netdev_queue_to_priority,
3280 NULL, /* recv_wait */
3281 NULL, /* recv_purge */
3282 dpif_netdev_register_upcall_cb,
3283 dpif_netdev_enable_upcall,
3284 dpif_netdev_disable_upcall,
3285 dpif_netdev_get_datapath_version,
3289 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3290 const char *argv[], void *aux OVS_UNUSED)
3292 struct dp_netdev_port *old_port;
3293 struct dp_netdev_port *new_port;
3294 struct dp_netdev *dp;
3297 ovs_mutex_lock(&dp_netdev_mutex);
3298 dp = shash_find_data(&dp_netdevs, argv[1]);
3299 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3300 ovs_mutex_unlock(&dp_netdev_mutex);
3301 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3304 ovs_refcount_ref(&dp->ref_cnt);
3305 ovs_mutex_unlock(&dp_netdev_mutex);
3307 ovs_mutex_lock(&dp->port_mutex);
3308 if (get_port_by_name(dp, argv[2], &old_port)) {
3309 unixctl_command_reply_error(conn, "unknown port");
3313 port_no = u32_to_odp(atoi(argv[3]));
3314 if (!port_no || port_no == ODPP_NONE) {
3315 unixctl_command_reply_error(conn, "bad port number");
3318 if (dp_netdev_lookup_port(dp, port_no)) {
3319 unixctl_command_reply_error(conn, "port number already in use");
3323 /* Remove old port. */
3324 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3325 ovsrcu_postpone(free, old_port);
3327 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3328 new_port = xmemdup(old_port, sizeof *old_port);
3329 new_port->port_no = port_no;
3330 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3332 seq_change(dp->port_seq);
3333 unixctl_command_reply(conn, NULL);
3336 ovs_mutex_unlock(&dp->port_mutex);
3337 dp_netdev_unref(dp);
3341 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3342 const char *argv[], void *aux OVS_UNUSED)
3344 struct dp_netdev_port *port;
3345 struct dp_netdev *dp;
3347 ovs_mutex_lock(&dp_netdev_mutex);
3348 dp = shash_find_data(&dp_netdevs, argv[1]);
3349 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3350 ovs_mutex_unlock(&dp_netdev_mutex);
3351 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3354 ovs_refcount_ref(&dp->ref_cnt);
3355 ovs_mutex_unlock(&dp_netdev_mutex);
3357 ovs_mutex_lock(&dp->port_mutex);
3358 if (get_port_by_name(dp, argv[2], &port)) {
3359 unixctl_command_reply_error(conn, "unknown port");
3360 } else if (port->port_no == ODPP_LOCAL) {
3361 unixctl_command_reply_error(conn, "can't delete local port");
3363 do_del_port(dp, port);
3364 unixctl_command_reply(conn, NULL);
3366 ovs_mutex_unlock(&dp->port_mutex);
3368 dp_netdev_unref(dp);
3372 dpif_dummy_register__(const char *type)
3374 struct dpif_class *class;
3376 class = xmalloc(sizeof *class);
3377 *class = dpif_netdev_class;
3378 class->type = xstrdup(type);
3379 dp_register_provider(class);
3383 dpif_dummy_register(bool override)
3390 dp_enumerate_types(&types);
3391 SSET_FOR_EACH (type, &types) {
3392 if (!dp_unregister_provider(type)) {
3393 dpif_dummy_register__(type);
3396 sset_destroy(&types);
3399 dpif_dummy_register__("dummy");
3401 unixctl_command_register("dpif-dummy/change-port-number",
3402 "dp port new-number",
3403 3, 3, dpif_dummy_change_port_number, NULL);
3404 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3405 2, 2, dpif_dummy_delete_port, NULL);
3408 /* Datapath Classifier. */
3410 /* A set of rules that all have the same fields wildcarded. */
3411 struct dpcls_subtable {
3412 /* The fields are only used by writers. */
3413 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3415 /* These fields are accessed by readers. */
3416 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3417 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3418 /* 'mask' must be the last field, additional space is allocated here. */
3421 /* Initializes 'cls' as a classifier that initially contains no classification
3424 dpcls_init(struct dpcls *cls)
3426 cmap_init(&cls->subtables_map);
3427 pvector_init(&cls->subtables);
3431 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3433 pvector_remove(&cls->subtables, subtable);
3434 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3435 subtable->mask.hash);
3436 cmap_destroy(&subtable->rules);
3437 ovsrcu_postpone(free, subtable);
3440 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3441 * caller's responsibility.
3442 * May only be called after all the readers have been terminated. */
3444 dpcls_destroy(struct dpcls *cls)
3447 struct dpcls_subtable *subtable;
3449 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3450 dpcls_destroy_subtable(cls, subtable);
3452 cmap_destroy(&cls->subtables_map);
3453 pvector_destroy(&cls->subtables);
3457 static struct dpcls_subtable *
3458 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3460 struct dpcls_subtable *subtable;
3462 /* Need to add one. */
3463 subtable = xmalloc(sizeof *subtable
3464 - sizeof subtable->mask.mf + mask->len);
3465 cmap_init(&subtable->rules);
3466 netdev_flow_key_clone(&subtable->mask, mask);
3467 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3468 pvector_insert(&cls->subtables, subtable, 0);
3469 pvector_publish(&cls->subtables);
3474 static inline struct dpcls_subtable *
3475 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3477 struct dpcls_subtable *subtable;
3479 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3480 &cls->subtables_map) {
3481 if (netdev_flow_key_equal(&subtable->mask, mask)) {
3485 return dpcls_create_subtable(cls, mask);
3488 /* Insert 'rule' into 'cls'. */
3490 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
3491 const struct netdev_flow_key *mask)
3493 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
3495 rule->mask = &subtable->mask;
3496 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
3499 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3501 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
3503 struct dpcls_subtable *subtable;
3505 ovs_assert(rule->mask);
3507 INIT_CONTAINER(subtable, rule->mask, mask);
3509 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
3511 dpcls_destroy_subtable(cls, subtable);
3512 pvector_publish(&cls->subtables);
3516 /* Returns true if 'target' satisifies 'key' in 'mask', that is, if each 1-bit
3517 * in 'mask' the values in 'key' and 'target' are the same.
3519 * Note: 'key' and 'mask' have the same mask, and 'key' is already masked. */
3521 dpcls_rule_matches_key(const struct dpcls_rule *rule,
3522 const struct netdev_flow_key *target)
3524 const uint64_t *keyp = rule->flow.mf.inline_values;
3525 const uint64_t *maskp = rule->mask->mf.inline_values;
3526 uint64_t target_u64;
3528 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(target_u64, target, rule->flow.mf.map) {
3529 if (OVS_UNLIKELY((target_u64 & *maskp++) != *keyp++)) {
3536 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3537 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3538 * NULL it is skipped.
3540 * This function is optimized for use in the userspace datapath and therefore
3541 * does not implement a lot of features available in the standard
3542 * classifier_lookup() function. Specifically, it does not implement
3543 * priorities, instead returning any rule which matches the flow.
3545 * Returns true if all flows found a corresponding rule. */
3547 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
3548 struct dpcls_rule **rules, const size_t cnt)
3550 /* The batch size 16 was experimentally found faster than 8 or 32. */
3551 typedef uint16_t map_type;
3552 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3554 #if !defined(__CHECKER__) && !defined(_WIN32)
3555 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
3557 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_RX_BATCH, MAP_BITS) };
3559 map_type maps[N_MAPS];
3560 struct dpcls_subtable *subtable;
3562 memset(maps, 0xff, sizeof maps);
3563 if (cnt % MAP_BITS) {
3564 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
3566 memset(rules, 0, cnt * sizeof *rules);
3568 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
3569 const struct netdev_flow_key *mkeys = keys;
3570 struct dpcls_rule **mrules = rules;
3571 map_type remains = 0;
3574 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
3576 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
3577 uint32_t hashes[MAP_BITS];
3578 const struct cmap_node *nodes[MAP_BITS];
3579 unsigned long map = maps[m];
3583 continue; /* Skip empty maps. */
3586 /* Compute hashes for the remaining keys. */
3587 ULONG_FOR_EACH_1(i, map) {
3588 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
3592 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
3593 /* Check results. */
3594 ULONG_FOR_EACH_1(i, map) {
3595 struct dpcls_rule *rule;
3597 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
3598 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
3603 ULONG_SET0(map, i); /* Did not match. */
3605 ; /* Keep Sparse happy. */
3607 maps[m] &= ~map; /* Clear the found rules. */
3611 return true; /* All found. */
3614 return false; /* Some misses. */