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 atomic_llong used; /* Last used time, in monotonic msecs. */
243 atomic_ullong packet_count; /* Number of packets matched. */
244 atomic_ullong byte_count; /* Number of bytes matched. */
245 atomic_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 atomic_ullong 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'. */
749 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
750 * load/store semantics. While the increment is not atomic, the load and
751 * store operations are, making it impossible to read inconsistent values.
753 * This is used to update thread local stats counters. */
755 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
757 unsigned long long tmp;
759 atomic_read_relaxed(var, &tmp);
761 atomic_store_relaxed(var, tmp);
765 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
767 struct dp_netdev *dp = get_dp_netdev(dpif);
768 struct dp_netdev_pmd_thread *pmd;
770 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
771 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
772 unsigned long long n;
773 stats->n_flows += cmap_count(&pmd->flow_table);
775 atomic_read_relaxed(&pmd->stats.n[DP_STAT_HIT], &n);
777 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
778 stats->n_missed += n;
779 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
782 stats->n_masks = UINT32_MAX;
783 stats->n_mask_hit = UINT64_MAX;
789 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
793 if (pmd->core_id == NON_PMD_CORE_ID) {
797 ovs_mutex_lock(&pmd->cond_mutex);
798 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
799 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
800 ovs_mutex_unlock(&pmd->cond_mutex);
803 /* Causes all pmd threads to reload its tx/rx devices.
804 * Must be called after adding/removing ports. */
806 dp_netdev_reload_pmds(struct dp_netdev *dp)
808 struct dp_netdev_pmd_thread *pmd;
810 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
811 dp_netdev_reload_pmd__(pmd);
816 hash_port_no(odp_port_t port_no)
818 return hash_int(odp_to_u32(port_no), 0);
822 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
824 OVS_REQUIRES(dp->port_mutex)
826 struct netdev_saved_flags *sf;
827 struct dp_netdev_port *port;
828 struct netdev *netdev;
829 enum netdev_flags flags;
830 const char *open_type;
834 /* XXX reject devices already in some dp_netdev. */
836 /* Open and validate network device. */
837 open_type = dpif_netdev_port_open_type(dp->class, type);
838 error = netdev_open(devname, open_type, &netdev);
842 /* XXX reject non-Ethernet devices */
844 netdev_get_flags(netdev, &flags);
845 if (flags & NETDEV_LOOPBACK) {
846 VLOG_ERR("%s: cannot add a loopback device", devname);
847 netdev_close(netdev);
851 if (netdev_is_pmd(netdev)) {
852 int n_cores = ovs_numa_get_n_cores();
854 if (n_cores == OVS_CORE_UNSPEC) {
855 VLOG_ERR("%s, cannot get cpu core info", devname);
858 /* There can only be ovs_numa_get_n_cores() pmd threads,
859 * so creates a txq for each. */
860 error = netdev_set_multiq(netdev, n_cores, dp->n_dpdk_rxqs);
861 if (error && (error != EOPNOTSUPP)) {
862 VLOG_ERR("%s, cannot set multiq", devname);
866 port = xzalloc(sizeof *port);
867 port->port_no = port_no;
868 port->netdev = netdev;
869 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
870 port->type = xstrdup(type);
871 for (i = 0; i < netdev_n_rxq(netdev); i++) {
872 error = netdev_rxq_open(netdev, &port->rxq[i], i);
874 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
875 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
876 devname, ovs_strerror(errno));
877 netdev_close(netdev);
885 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
887 for (i = 0; i < netdev_n_rxq(netdev); i++) {
888 netdev_rxq_close(port->rxq[i]);
890 netdev_close(netdev);
898 ovs_refcount_init(&port->ref_cnt);
899 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
901 if (netdev_is_pmd(netdev)) {
902 dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
903 dp_netdev_reload_pmds(dp);
905 seq_change(dp->port_seq);
911 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
912 odp_port_t *port_nop)
914 struct dp_netdev *dp = get_dp_netdev(dpif);
915 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
916 const char *dpif_port;
920 ovs_mutex_lock(&dp->port_mutex);
921 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
922 if (*port_nop != ODPP_NONE) {
924 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
926 port_no = choose_port(dp, dpif_port);
927 error = port_no == ODPP_NONE ? EFBIG : 0;
931 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
933 ovs_mutex_unlock(&dp->port_mutex);
939 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
941 struct dp_netdev *dp = get_dp_netdev(dpif);
944 ovs_mutex_lock(&dp->port_mutex);
945 if (port_no == ODPP_LOCAL) {
948 struct dp_netdev_port *port;
950 error = get_port_by_number(dp, port_no, &port);
952 do_del_port(dp, port);
955 ovs_mutex_unlock(&dp->port_mutex);
961 is_valid_port_number(odp_port_t port_no)
963 return port_no != ODPP_NONE;
966 static struct dp_netdev_port *
967 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
969 struct dp_netdev_port *port;
971 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
972 if (port->port_no == port_no) {
980 get_port_by_number(struct dp_netdev *dp,
981 odp_port_t port_no, struct dp_netdev_port **portp)
983 if (!is_valid_port_number(port_no)) {
987 *portp = dp_netdev_lookup_port(dp, port_no);
988 return *portp ? 0 : ENOENT;
993 port_ref(struct dp_netdev_port *port)
996 ovs_refcount_ref(&port->ref_cnt);
1001 port_try_ref(struct dp_netdev_port *port)
1004 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
1011 port_unref(struct dp_netdev_port *port)
1013 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1014 int n_rxq = netdev_n_rxq(port->netdev);
1017 netdev_close(port->netdev);
1018 netdev_restore_flags(port->sf);
1020 for (i = 0; i < n_rxq; i++) {
1021 netdev_rxq_close(port->rxq[i]);
1030 get_port_by_name(struct dp_netdev *dp,
1031 const char *devname, struct dp_netdev_port **portp)
1032 OVS_REQUIRES(dp->port_mutex)
1034 struct dp_netdev_port *port;
1036 CMAP_FOR_EACH (port, node, &dp->ports) {
1037 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1046 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1048 struct dp_netdev_pmd_thread *pmd;
1051 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1052 if (pmd->numa_id == numa_id) {
1060 /* Returns 'true' if there is a port with pmd netdev and the netdev
1061 * is on numa node 'numa_id'. */
1063 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1065 struct dp_netdev_port *port;
1067 CMAP_FOR_EACH (port, node, &dp->ports) {
1068 if (netdev_is_pmd(port->netdev)
1069 && netdev_get_numa_id(port->netdev) == numa_id) {
1079 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1080 OVS_REQUIRES(dp->port_mutex)
1082 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1083 seq_change(dp->port_seq);
1084 if (netdev_is_pmd(port->netdev)) {
1085 int numa_id = netdev_get_numa_id(port->netdev);
1087 /* If there is no netdev on the numa node, deletes the pmd threads
1088 * for that numa. Else, just reloads the queues. */
1089 if (!has_pmd_port_for_numa(dp, numa_id)) {
1090 dp_netdev_del_pmds_on_numa(dp, numa_id);
1092 dp_netdev_reload_pmds(dp);
1099 answer_port_query(const struct dp_netdev_port *port,
1100 struct dpif_port *dpif_port)
1102 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1103 dpif_port->type = xstrdup(port->type);
1104 dpif_port->port_no = port->port_no;
1108 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1109 struct dpif_port *dpif_port)
1111 struct dp_netdev *dp = get_dp_netdev(dpif);
1112 struct dp_netdev_port *port;
1115 error = get_port_by_number(dp, port_no, &port);
1116 if (!error && dpif_port) {
1117 answer_port_query(port, dpif_port);
1124 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1125 struct dpif_port *dpif_port)
1127 struct dp_netdev *dp = get_dp_netdev(dpif);
1128 struct dp_netdev_port *port;
1131 ovs_mutex_lock(&dp->port_mutex);
1132 error = get_port_by_name(dp, devname, &port);
1133 if (!error && dpif_port) {
1134 answer_port_query(port, dpif_port);
1136 ovs_mutex_unlock(&dp->port_mutex);
1142 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1144 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1148 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1150 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1151 ovsrcu_postpone(dp_netdev_flow_free, flow);
1156 dp_netdev_flow_hash(const ovs_u128 *ufid)
1158 return ufid->u32[0];
1162 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1163 struct dp_netdev_flow *flow)
1164 OVS_REQUIRES(pmd->flow_mutex)
1166 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1168 dpcls_remove(&pmd->cls, &flow->cr);
1169 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1172 dp_netdev_flow_unref(flow);
1176 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1178 struct dp_netdev_flow *netdev_flow;
1180 ovs_mutex_lock(&pmd->flow_mutex);
1181 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1182 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1184 ovs_mutex_unlock(&pmd->flow_mutex);
1188 dpif_netdev_flow_flush(struct dpif *dpif)
1190 struct dp_netdev *dp = get_dp_netdev(dpif);
1191 struct dp_netdev_pmd_thread *pmd;
1193 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1194 dp_netdev_pmd_flow_flush(pmd);
1200 struct dp_netdev_port_state {
1201 struct cmap_position position;
1206 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1208 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1213 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1214 struct dpif_port *dpif_port)
1216 struct dp_netdev_port_state *state = state_;
1217 struct dp_netdev *dp = get_dp_netdev(dpif);
1218 struct cmap_node *node;
1221 node = cmap_next_position(&dp->ports, &state->position);
1223 struct dp_netdev_port *port;
1225 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1228 state->name = xstrdup(netdev_get_name(port->netdev));
1229 dpif_port->name = state->name;
1230 dpif_port->type = port->type;
1231 dpif_port->port_no = port->port_no;
1242 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1244 struct dp_netdev_port_state *state = state_;
1251 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1253 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1254 uint64_t new_port_seq;
1257 new_port_seq = seq_read(dpif->dp->port_seq);
1258 if (dpif->last_port_seq != new_port_seq) {
1259 dpif->last_port_seq = new_port_seq;
1269 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1271 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1273 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1276 static struct dp_netdev_flow *
1277 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1279 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1282 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1284 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1287 /* netdev_flow_key utilities.
1289 * netdev_flow_key is basically a miniflow. We use these functions
1290 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1291 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1293 * - Since we are dealing exclusively with miniflows created by
1294 * miniflow_extract(), if the map is different the miniflow is different.
1295 * Therefore we can be faster by comparing the map and the miniflow in a
1297 * _ netdev_flow_key's miniflow has always inline values.
1298 * - These functions can be inlined by the compiler.
1300 * The following assertions make sure that what we're doing with miniflow is
1303 BUILD_ASSERT_DECL(offsetof(struct miniflow, inline_values)
1304 == sizeof(uint64_t));
1306 /* Given the number of bits set in the miniflow map, returns the size of the
1307 * 'netdev_flow_key.mf' */
1308 static inline uint32_t
1309 netdev_flow_key_size(uint32_t flow_u32s)
1311 return offsetof(struct miniflow, inline_values) +
1312 MINIFLOW_VALUES_SIZE(flow_u32s);
1316 netdev_flow_key_equal(const struct netdev_flow_key *a,
1317 const struct netdev_flow_key *b)
1319 /* 'b->len' may be not set yet. */
1320 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1323 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1324 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1325 * generated by miniflow_extract. */
1327 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1328 const struct miniflow *mf)
1330 return !memcmp(&key->mf, mf, key->len);
1334 netdev_flow_key_clone(struct netdev_flow_key *dst,
1335 const struct netdev_flow_key *src)
1338 offsetof(struct netdev_flow_key, mf) + src->len);
1343 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1344 const struct flow *src)
1346 struct dp_packet packet;
1347 uint64_t buf_stub[512 / 8];
1349 miniflow_initialize(&dst->mf, dst->buf);
1351 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1352 pkt_metadata_from_flow(&packet.md, src);
1353 flow_compose(&packet, src);
1354 miniflow_extract(&packet, &dst->mf);
1355 dp_packet_uninit(&packet);
1357 dst->len = netdev_flow_key_size(count_1bits(dst->mf.map));
1358 dst->hash = 0; /* Not computed yet. */
1361 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1363 netdev_flow_mask_init(struct netdev_flow_key *mask,
1364 const struct match *match)
1366 const uint64_t *mask_u64 = (const uint64_t *) &match->wc.masks;
1367 uint64_t *dst = mask->mf.inline_values;
1368 uint64_t map, mask_map = 0;
1372 /* Only check masks that make sense for the flow. */
1373 map = flow_wc_map(&match->flow);
1376 uint64_t rm1bit = rightmost_1bit(map);
1377 int i = raw_ctz(map);
1381 *dst++ = mask_u64[i];
1382 hash = hash_add64(hash, mask_u64[i]);
1387 mask->mf.values_inline = true;
1388 mask->mf.map = mask_map;
1390 hash = hash_add64(hash, mask_map);
1392 n = dst - mask->mf.inline_values;
1394 mask->hash = hash_finish(hash, n * 8);
1395 mask->len = netdev_flow_key_size(n);
1398 /* Initializes 'dst' as a copy of 'src' masked with 'mask'. */
1400 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1401 const struct flow *flow,
1402 const struct netdev_flow_key *mask)
1404 uint64_t *dst_u64 = dst->mf.inline_values;
1405 const uint64_t *mask_u64 = mask->mf.inline_values;
1409 dst->len = mask->len;
1410 dst->mf.values_inline = true;
1411 dst->mf.map = mask->mf.map;
1413 FLOW_FOR_EACH_IN_MAP(value, flow, mask->mf.map) {
1414 *dst_u64 = value & *mask_u64++;
1415 hash = hash_add64(hash, *dst_u64++);
1417 dst->hash = hash_finish(hash, (dst_u64 - dst->mf.inline_values) * 8);
1420 /* Iterate through all netdev_flow_key u64 values specified by 'MAP' */
1421 #define NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(VALUE, KEY, MAP) \
1422 for (struct mf_for_each_in_map_aux aux__ \
1423 = { (KEY)->mf.inline_values, (KEY)->mf.map, MAP }; \
1424 mf_get_next_in_map(&aux__, &(VALUE)); \
1427 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1429 static inline uint32_t
1430 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1431 const struct netdev_flow_key *mask)
1433 const uint64_t *p = mask->mf.inline_values;
1437 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(key_u64, key, mask->mf.map) {
1438 hash = hash_add64(hash, key_u64 & *p++);
1441 return hash_finish(hash, (p - mask->mf.inline_values) * 8);
1445 emc_entry_alive(struct emc_entry *ce)
1447 return ce->flow && !ce->flow->dead;
1451 emc_clear_entry(struct emc_entry *ce)
1454 dp_netdev_flow_unref(ce->flow);
1460 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1461 const struct netdev_flow_key *key)
1463 if (ce->flow != flow) {
1465 dp_netdev_flow_unref(ce->flow);
1468 if (dp_netdev_flow_ref(flow)) {
1475 netdev_flow_key_clone(&ce->key, key);
1480 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1481 struct dp_netdev_flow *flow)
1483 struct emc_entry *to_be_replaced = NULL;
1484 struct emc_entry *current_entry;
1486 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1487 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1488 /* We found the entry with the 'mf' miniflow */
1489 emc_change_entry(current_entry, flow, NULL);
1493 /* Replacement policy: put the flow in an empty (not alive) entry, or
1494 * in the first entry where it can be */
1496 || (emc_entry_alive(to_be_replaced)
1497 && !emc_entry_alive(current_entry))
1498 || current_entry->key.hash < to_be_replaced->key.hash) {
1499 to_be_replaced = current_entry;
1502 /* We didn't find the miniflow in the cache.
1503 * The 'to_be_replaced' entry is where the new flow will be stored */
1505 emc_change_entry(to_be_replaced, flow, key);
1508 static inline struct dp_netdev_flow *
1509 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1511 struct emc_entry *current_entry;
1513 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1514 if (current_entry->key.hash == key->hash
1515 && emc_entry_alive(current_entry)
1516 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1518 /* We found the entry with the 'key->mf' miniflow */
1519 return current_entry->flow;
1526 static struct dp_netdev_flow *
1527 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1528 const struct netdev_flow_key *key)
1530 struct dp_netdev_flow *netdev_flow;
1531 struct dpcls_rule *rule;
1533 dpcls_lookup(&pmd->cls, key, &rule, 1);
1534 netdev_flow = dp_netdev_flow_cast(rule);
1539 static struct dp_netdev_flow *
1540 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1541 const ovs_u128 *ufidp, const struct nlattr *key,
1544 struct dp_netdev_flow *netdev_flow;
1548 /* If a UFID is not provided, determine one based on the key. */
1549 if (!ufidp && key && key_len
1550 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1551 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1556 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1558 if (ovs_u128_equal(&netdev_flow->ufid, ufidp)) {
1568 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1569 struct dpif_flow_stats *stats)
1571 struct dp_netdev_flow *netdev_flow;
1572 unsigned long long n;
1576 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1578 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1579 stats->n_packets = n;
1580 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1582 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1584 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1585 stats->tcp_flags = flags;
1588 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1589 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1590 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1593 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1594 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1595 struct dpif_flow *flow, bool terse)
1598 memset(flow, 0, sizeof *flow);
1600 struct flow_wildcards wc;
1601 struct dp_netdev_actions *actions;
1604 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1607 offset = key_buf->size;
1608 flow->key = ofpbuf_tail(key_buf);
1609 odp_flow_key_from_flow(key_buf, &netdev_flow->flow, &wc.masks,
1610 netdev_flow->flow.in_port.odp_port, true);
1611 flow->key_len = key_buf->size - offset;
1614 offset = mask_buf->size;
1615 flow->mask = ofpbuf_tail(mask_buf);
1616 odp_flow_key_from_mask(mask_buf, &wc.masks, &netdev_flow->flow,
1617 odp_to_u32(wc.masks.in_port.odp_port),
1619 flow->mask_len = mask_buf->size - offset;
1622 actions = dp_netdev_flow_get_actions(netdev_flow);
1623 flow->actions = actions->actions;
1624 flow->actions_len = actions->size;
1627 flow->ufid = netdev_flow->ufid;
1628 flow->ufid_present = true;
1629 flow->pmd_id = netdev_flow->pmd_id;
1630 get_dpif_flow_stats(netdev_flow, &flow->stats);
1634 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1635 const struct nlattr *mask_key,
1636 uint32_t mask_key_len, const struct flow *flow,
1640 enum odp_key_fitness fitness;
1642 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1644 /* This should not happen: it indicates that
1645 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1646 * disagree on the acceptable form of a mask. Log the problem
1647 * as an error, with enough details to enable debugging. */
1648 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1650 if (!VLOG_DROP_ERR(&rl)) {
1654 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1656 VLOG_ERR("internal error parsing flow mask %s (%s)",
1657 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1664 enum mf_field_id id;
1665 /* No mask key, unwildcard everything except fields whose
1666 * prerequisities are not met. */
1667 memset(mask, 0x0, sizeof *mask);
1669 for (id = 0; id < MFF_N_IDS; ++id) {
1670 /* Skip registers and metadata. */
1671 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1672 && id != MFF_METADATA) {
1673 const struct mf_field *mf = mf_from_id(id);
1674 if (mf_are_prereqs_ok(mf, flow)) {
1675 mf_mask_field(mf, mask);
1681 /* Force unwildcard the in_port.
1683 * We need to do this even in the case where we unwildcard "everything"
1684 * above because "everything" only includes the 16-bit OpenFlow port number
1685 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1686 * port number mask->in_port.odp_port. */
1687 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1693 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1698 if (odp_flow_key_to_flow(key, key_len, flow)) {
1699 /* This should not happen: it indicates that odp_flow_key_from_flow()
1700 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1701 * flow. Log the problem as an error, with enough details to enable
1703 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1705 if (!VLOG_DROP_ERR(&rl)) {
1709 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1710 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1717 in_port = flow->in_port.odp_port;
1718 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1726 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1728 struct dp_netdev *dp = get_dp_netdev(dpif);
1729 struct dp_netdev_flow *netdev_flow;
1730 struct dp_netdev_pmd_thread *pmd;
1731 int pmd_id = get->pmd_id == PMD_ID_NULL ? NON_PMD_CORE_ID : get->pmd_id;
1734 pmd = dp_netdev_get_pmd(dp, pmd_id);
1739 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1742 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1747 dp_netdev_pmd_unref(pmd);
1753 static struct dp_netdev_flow *
1754 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
1755 struct match *match, const ovs_u128 *ufid,
1756 const struct nlattr *actions, size_t actions_len)
1757 OVS_REQUIRES(pmd->flow_mutex)
1759 struct dp_netdev_flow *flow;
1760 struct netdev_flow_key mask;
1762 netdev_flow_mask_init(&mask, match);
1763 /* Make sure wc does not have metadata. */
1764 ovs_assert(!(mask.mf.map & (MINIFLOW_MAP(metadata) | MINIFLOW_MAP(regs))));
1766 /* Do not allocate extra space. */
1767 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
1768 memset(&flow->stats, 0, sizeof flow->stats);
1770 *CONST_CAST(int *, &flow->pmd_id) = pmd->core_id;
1771 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
1772 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
1773 ovs_refcount_init(&flow->ref_cnt);
1774 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
1776 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
1777 dpcls_insert(&pmd->cls, &flow->cr, &mask);
1779 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
1780 dp_netdev_flow_hash(&flow->ufid));
1782 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1784 struct ds ds = DS_EMPTY_INITIALIZER;
1786 match.flow = flow->flow;
1787 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
1789 ds_put_cstr(&ds, "flow_add: ");
1790 odp_format_ufid(ufid, &ds);
1791 ds_put_cstr(&ds, " ");
1792 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
1793 ds_put_cstr(&ds, ", actions:");
1794 format_odp_actions(&ds, actions, actions_len);
1796 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
1805 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1807 struct dp_netdev *dp = get_dp_netdev(dpif);
1808 struct dp_netdev_flow *netdev_flow;
1809 struct netdev_flow_key key;
1810 struct dp_netdev_pmd_thread *pmd;
1813 int pmd_id = put->pmd_id == PMD_ID_NULL ? NON_PMD_CORE_ID : put->pmd_id;
1816 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
1820 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1821 put->mask, put->mask_len,
1822 &match.flow, &match.wc.masks);
1827 pmd = dp_netdev_get_pmd(dp, pmd_id);
1832 /* Must produce a netdev_flow_key for lookup.
1833 * This interface is no longer performance critical, since it is not used
1834 * for upcall processing any more. */
1835 netdev_flow_key_from_flow(&key, &match.flow);
1840 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
1843 ovs_mutex_lock(&pmd->flow_mutex);
1844 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
1846 if (put->flags & DPIF_FP_CREATE) {
1847 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
1849 memset(put->stats, 0, sizeof *put->stats);
1851 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
1861 if (put->flags & DPIF_FP_MODIFY
1862 && flow_equal(&match.flow, &netdev_flow->flow)) {
1863 struct dp_netdev_actions *new_actions;
1864 struct dp_netdev_actions *old_actions;
1866 new_actions = dp_netdev_actions_create(put->actions,
1869 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1870 ovsrcu_set(&netdev_flow->actions, new_actions);
1873 get_dpif_flow_stats(netdev_flow, put->stats);
1875 if (put->flags & DPIF_FP_ZERO_STATS) {
1876 /* XXX: The userspace datapath uses thread local statistics
1877 * (for flows), which should be updated only by the owning
1878 * thread. Since we cannot write on stats memory here,
1879 * we choose not to support this flag. Please note:
1880 * - This feature is currently used only by dpctl commands with
1882 * - Should the need arise, this operation can be implemented
1883 * by keeping a base value (to be update here) for each
1884 * counter, and subtracting it before outputting the stats */
1888 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1889 } else if (put->flags & DPIF_FP_CREATE) {
1892 /* Overlapping flow. */
1896 ovs_mutex_unlock(&pmd->flow_mutex);
1897 dp_netdev_pmd_unref(pmd);
1903 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1905 struct dp_netdev *dp = get_dp_netdev(dpif);
1906 struct dp_netdev_flow *netdev_flow;
1907 struct dp_netdev_pmd_thread *pmd;
1908 int pmd_id = del->pmd_id == PMD_ID_NULL ? NON_PMD_CORE_ID : del->pmd_id;
1911 pmd = dp_netdev_get_pmd(dp, pmd_id);
1916 ovs_mutex_lock(&pmd->flow_mutex);
1917 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
1921 get_dpif_flow_stats(netdev_flow, del->stats);
1923 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1927 ovs_mutex_unlock(&pmd->flow_mutex);
1928 dp_netdev_pmd_unref(pmd);
1933 struct dpif_netdev_flow_dump {
1934 struct dpif_flow_dump up;
1935 struct cmap_position poll_thread_pos;
1936 struct cmap_position flow_pos;
1937 struct dp_netdev_pmd_thread *cur_pmd;
1939 struct ovs_mutex mutex;
1942 static struct dpif_netdev_flow_dump *
1943 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1945 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1948 static struct dpif_flow_dump *
1949 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
1951 struct dpif_netdev_flow_dump *dump;
1953 dump = xzalloc(sizeof *dump);
1954 dpif_flow_dump_init(&dump->up, dpif_);
1955 dump->up.terse = terse;
1956 ovs_mutex_init(&dump->mutex);
1962 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1964 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1966 ovs_mutex_destroy(&dump->mutex);
1971 struct dpif_netdev_flow_dump_thread {
1972 struct dpif_flow_dump_thread up;
1973 struct dpif_netdev_flow_dump *dump;
1974 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1975 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1978 static struct dpif_netdev_flow_dump_thread *
1979 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1981 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1984 static struct dpif_flow_dump_thread *
1985 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1987 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1988 struct dpif_netdev_flow_dump_thread *thread;
1990 thread = xmalloc(sizeof *thread);
1991 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1992 thread->dump = dump;
1997 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1999 struct dpif_netdev_flow_dump_thread *thread
2000 = dpif_netdev_flow_dump_thread_cast(thread_);
2006 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2007 struct dpif_flow *flows, int max_flows)
2009 struct dpif_netdev_flow_dump_thread *thread
2010 = dpif_netdev_flow_dump_thread_cast(thread_);
2011 struct dpif_netdev_flow_dump *dump = thread->dump;
2012 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2016 ovs_mutex_lock(&dump->mutex);
2017 if (!dump->status) {
2018 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2019 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2020 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2021 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2023 /* First call to dump_next(), extracts the first pmd thread.
2024 * If there is no pmd thread, returns immediately. */
2026 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2028 ovs_mutex_unlock(&dump->mutex);
2035 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2036 struct cmap_node *node;
2038 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2042 netdev_flows[n_flows] = CONTAINER_OF(node,
2043 struct dp_netdev_flow,
2046 /* When finishing dumping the current pmd thread, moves to
2048 if (n_flows < flow_limit) {
2049 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2050 dp_netdev_pmd_unref(pmd);
2051 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2057 /* Keeps the reference to next caller. */
2058 dump->cur_pmd = pmd;
2060 /* If the current dump is empty, do not exit the loop, since the
2061 * remaining pmds could have flows to be dumped. Just dumps again
2062 * on the new 'pmd'. */
2065 ovs_mutex_unlock(&dump->mutex);
2067 for (i = 0; i < n_flows; i++) {
2068 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2069 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2070 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2071 struct dpif_flow *f = &flows[i];
2072 struct ofpbuf key, mask;
2074 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2075 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2076 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2084 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2085 OVS_NO_THREAD_SAFETY_ANALYSIS
2087 struct dp_netdev *dp = get_dp_netdev(dpif);
2088 struct dp_netdev_pmd_thread *pmd;
2089 struct dp_packet *pp;
2091 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2092 dp_packet_size(execute->packet) > UINT16_MAX) {
2096 /* Tries finding the 'pmd'. If NULL is returned, that means
2097 * the current thread is a non-pmd thread and should use
2098 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2099 pmd = ovsthread_getspecific(dp->per_pmd_key);
2101 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2104 /* If the current thread is non-pmd thread, acquires
2105 * the 'non_pmd_mutex'. */
2106 if (pmd->core_id == NON_PMD_CORE_ID) {
2107 ovs_mutex_lock(&dp->non_pmd_mutex);
2108 ovs_mutex_lock(&dp->port_mutex);
2111 pp = execute->packet;
2112 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2113 execute->actions_len);
2114 if (pmd->core_id == NON_PMD_CORE_ID) {
2115 dp_netdev_pmd_unref(pmd);
2116 ovs_mutex_unlock(&dp->port_mutex);
2117 ovs_mutex_unlock(&dp->non_pmd_mutex);
2124 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2128 for (i = 0; i < n_ops; i++) {
2129 struct dpif_op *op = ops[i];
2132 case DPIF_OP_FLOW_PUT:
2133 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2136 case DPIF_OP_FLOW_DEL:
2137 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2140 case DPIF_OP_EXECUTE:
2141 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2144 case DPIF_OP_FLOW_GET:
2145 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2151 /* Returns true if the configuration for rx queues or cpu mask
2154 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2156 if (dp->n_dpdk_rxqs != rxqs) {
2159 if (dp->pmd_cmask != NULL && cmask != NULL) {
2160 return strcmp(dp->pmd_cmask, cmask);
2162 return (dp->pmd_cmask != NULL || cmask != NULL);
2167 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2169 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2171 struct dp_netdev *dp = get_dp_netdev(dpif);
2173 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2174 struct dp_netdev_port *port;
2176 dp_netdev_destroy_all_pmds(dp);
2178 CMAP_FOR_EACH (port, node, &dp->ports) {
2179 if (netdev_is_pmd(port->netdev)) {
2182 /* Closes the existing 'rxq's. */
2183 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2184 netdev_rxq_close(port->rxq[i]);
2185 port->rxq[i] = NULL;
2188 /* Sets the new rx queue config. */
2189 err = netdev_set_multiq(port->netdev, ovs_numa_get_n_cores(),
2191 if (err && (err != EOPNOTSUPP)) {
2192 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2193 " %u", netdev_get_name(port->netdev),
2198 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2199 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2200 * netdev_n_rxq(port->netdev));
2201 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2202 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2206 dp->n_dpdk_rxqs = n_rxqs;
2208 /* Reconfigures the cpu mask. */
2209 ovs_numa_set_cpu_mask(cmask);
2210 free(dp->pmd_cmask);
2211 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2213 /* Restores the non-pmd. */
2214 dp_netdev_set_nonpmd(dp);
2215 /* Restores all pmd threads. */
2216 dp_netdev_reset_pmd_threads(dp);
2223 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2224 uint32_t queue_id, uint32_t *priority)
2226 *priority = queue_id;
2231 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
2232 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
2234 struct dp_netdev_actions *
2235 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2237 struct dp_netdev_actions *netdev_actions;
2239 netdev_actions = xmalloc(sizeof *netdev_actions);
2240 netdev_actions->actions = xmemdup(actions, size);
2241 netdev_actions->size = size;
2243 return netdev_actions;
2246 struct dp_netdev_actions *
2247 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2249 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2253 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2255 free(actions->actions);
2261 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2262 struct dp_netdev_port *port,
2263 struct netdev_rxq *rxq)
2265 struct dp_packet *packets[NETDEV_MAX_RX_BATCH];
2268 error = netdev_rxq_recv(rxq, packets, &cnt);
2272 *recirc_depth_get() = 0;
2274 /* XXX: initialize md in netdev implementation. */
2275 for (i = 0; i < cnt; i++) {
2276 packets[i]->md = PKT_METADATA_INITIALIZER(port->port_no);
2278 dp_netdev_input(pmd, packets, cnt);
2279 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2280 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2282 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2283 netdev_get_name(port->netdev), ovs_strerror(error));
2287 /* Return true if needs to revalidate datapath flows. */
2289 dpif_netdev_run(struct dpif *dpif)
2291 struct dp_netdev_port *port;
2292 struct dp_netdev *dp = get_dp_netdev(dpif);
2293 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2295 uint64_t new_tnl_seq;
2297 ovs_mutex_lock(&dp->non_pmd_mutex);
2298 CMAP_FOR_EACH (port, node, &dp->ports) {
2299 if (!netdev_is_pmd(port->netdev)) {
2302 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2303 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2307 ovs_mutex_unlock(&dp->non_pmd_mutex);
2308 dp_netdev_pmd_unref(non_pmd);
2310 tnl_arp_cache_run();
2311 new_tnl_seq = seq_read(tnl_conf_seq);
2313 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2314 dp->last_tnl_conf_seq = new_tnl_seq;
2321 dpif_netdev_wait(struct dpif *dpif)
2323 struct dp_netdev_port *port;
2324 struct dp_netdev *dp = get_dp_netdev(dpif);
2326 ovs_mutex_lock(&dp_netdev_mutex);
2327 CMAP_FOR_EACH (port, node, &dp->ports) {
2328 if (!netdev_is_pmd(port->netdev)) {
2331 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2332 netdev_rxq_wait(port->rxq[i]);
2336 ovs_mutex_unlock(&dp_netdev_mutex);
2337 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2341 struct dp_netdev_port *port;
2342 struct netdev_rxq *rx;
2346 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2347 struct rxq_poll **ppoll_list, int poll_cnt)
2349 struct rxq_poll *poll_list = *ppoll_list;
2350 struct dp_netdev_port *port;
2351 int n_pmds_on_numa, index, i;
2353 /* Simple scheduler for netdev rx polling. */
2354 for (i = 0; i < poll_cnt; i++) {
2355 port_unref(poll_list[i].port);
2359 n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
2362 CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
2363 /* Calls port_try_ref() to prevent the main thread
2364 * from deleting the port. */
2365 if (port_try_ref(port)) {
2366 if (netdev_is_pmd(port->netdev)
2367 && netdev_get_numa_id(port->netdev) == pmd->numa_id) {
2370 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2371 if ((index % n_pmds_on_numa) == pmd->index) {
2372 poll_list = xrealloc(poll_list,
2373 sizeof *poll_list * (poll_cnt + 1));
2376 poll_list[poll_cnt].port = port;
2377 poll_list[poll_cnt].rx = port->rxq[i];
2383 /* Unrefs the port_try_ref(). */
2388 *ppoll_list = poll_list;
2393 pmd_thread_main(void *f_)
2395 struct dp_netdev_pmd_thread *pmd = f_;
2396 unsigned int lc = 0;
2397 struct rxq_poll *poll_list;
2398 unsigned int port_seq = PMD_INITIAL_SEQ;
2405 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2406 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2407 pmd_thread_setaffinity_cpu(pmd->core_id);
2409 emc_cache_init(&pmd->flow_cache);
2410 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2412 /* Signal here to make sure the pmd finishes
2413 * reloading the updated configuration. */
2414 dp_netdev_pmd_reload_done(pmd);
2419 for (i = 0; i < poll_cnt; i++) {
2420 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2428 emc_cache_slow_sweep(&pmd->flow_cache);
2431 atomic_read_relaxed(&pmd->change_seq, &seq);
2432 if (seq != port_seq) {
2439 emc_cache_uninit(&pmd->flow_cache);
2441 if (!latch_is_set(&pmd->exit_latch)){
2445 for (i = 0; i < poll_cnt; i++) {
2446 port_unref(poll_list[i].port);
2449 dp_netdev_pmd_reload_done(pmd);
2456 dp_netdev_disable_upcall(struct dp_netdev *dp)
2457 OVS_ACQUIRES(dp->upcall_rwlock)
2459 fat_rwlock_wrlock(&dp->upcall_rwlock);
2463 dpif_netdev_disable_upcall(struct dpif *dpif)
2464 OVS_NO_THREAD_SAFETY_ANALYSIS
2466 struct dp_netdev *dp = get_dp_netdev(dpif);
2467 dp_netdev_disable_upcall(dp);
2471 dp_netdev_enable_upcall(struct dp_netdev *dp)
2472 OVS_RELEASES(dp->upcall_rwlock)
2474 fat_rwlock_unlock(&dp->upcall_rwlock);
2478 dpif_netdev_enable_upcall(struct dpif *dpif)
2479 OVS_NO_THREAD_SAFETY_ANALYSIS
2481 struct dp_netdev *dp = get_dp_netdev(dpif);
2482 dp_netdev_enable_upcall(dp);
2486 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2488 ovs_mutex_lock(&pmd->cond_mutex);
2489 xpthread_cond_signal(&pmd->cond);
2490 ovs_mutex_unlock(&pmd->cond_mutex);
2493 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2494 * the pointer if succeeds, otherwise, NULL.
2496 * Caller must unrefs the returned reference. */
2497 static struct dp_netdev_pmd_thread *
2498 dp_netdev_get_pmd(struct dp_netdev *dp, int core_id)
2500 struct dp_netdev_pmd_thread *pmd;
2501 const struct cmap_node *pnode;
2503 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2507 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2509 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2512 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2514 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2516 struct dp_netdev_pmd_thread *non_pmd;
2518 non_pmd = xzalloc(sizeof *non_pmd);
2519 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2523 /* Caller must have valid pointer to 'pmd'. */
2525 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2527 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2531 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2533 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2534 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2538 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2539 * fails, keeps checking for next node until reaching the end of cmap.
2541 * Caller must unrefs the returned reference. */
2542 static struct dp_netdev_pmd_thread *
2543 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2545 struct dp_netdev_pmd_thread *next;
2548 struct cmap_node *node;
2550 node = cmap_next_position(&dp->poll_threads, pos);
2551 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2553 } while (next && !dp_netdev_pmd_try_ref(next));
2558 /* Configures the 'pmd' based on the input argument. */
2560 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2561 int index, int core_id, int numa_id)
2565 pmd->core_id = core_id;
2566 pmd->numa_id = numa_id;
2568 ovs_refcount_init(&pmd->ref_cnt);
2569 latch_init(&pmd->exit_latch);
2570 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2571 xpthread_cond_init(&pmd->cond, NULL);
2572 ovs_mutex_init(&pmd->cond_mutex);
2573 ovs_mutex_init(&pmd->flow_mutex);
2574 dpcls_init(&pmd->cls);
2575 cmap_init(&pmd->flow_table);
2576 /* init the 'flow_cache' since there is no
2577 * actual thread created for NON_PMD_CORE_ID. */
2578 if (core_id == NON_PMD_CORE_ID) {
2579 emc_cache_init(&pmd->flow_cache);
2581 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2582 hash_int(core_id, 0));
2586 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2588 dp_netdev_pmd_flow_flush(pmd);
2589 dpcls_destroy(&pmd->cls);
2590 cmap_destroy(&pmd->flow_table);
2591 ovs_mutex_destroy(&pmd->flow_mutex);
2592 latch_destroy(&pmd->exit_latch);
2593 xpthread_cond_destroy(&pmd->cond);
2594 ovs_mutex_destroy(&pmd->cond_mutex);
2598 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2599 * and unrefs the struct. */
2601 dp_netdev_del_pmd(struct dp_netdev_pmd_thread *pmd)
2603 /* Uninit the 'flow_cache' since there is
2604 * no actual thread uninit it for NON_PMD_CORE_ID. */
2605 if (pmd->core_id == NON_PMD_CORE_ID) {
2606 emc_cache_uninit(&pmd->flow_cache);
2608 latch_set(&pmd->exit_latch);
2609 dp_netdev_reload_pmd__(pmd);
2610 ovs_numa_unpin_core(pmd->core_id);
2611 xpthread_join(pmd->thread, NULL);
2613 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2614 dp_netdev_pmd_unref(pmd);
2617 /* Destroys all pmd threads. */
2619 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2621 struct dp_netdev_pmd_thread *pmd;
2623 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2624 dp_netdev_del_pmd(pmd);
2628 /* Deletes all pmd threads on numa node 'numa_id'. */
2630 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2632 struct dp_netdev_pmd_thread *pmd;
2634 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2635 if (pmd->numa_id == numa_id) {
2636 dp_netdev_del_pmd(pmd);
2641 /* Checks the numa node id of 'netdev' and starts pmd threads for
2644 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2648 if (!ovs_numa_numa_id_is_valid(numa_id)) {
2649 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2650 "invalid", numa_id);
2654 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
2656 /* If there are already pmd threads created for the numa node
2657 * in which 'netdev' is on, do nothing. Else, creates the
2658 * pmd threads for the numa node. */
2660 int can_have, n_unpinned, i;
2662 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
2664 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2665 "cores on numa node");
2669 /* If cpu mask is specified, uses all unpinned cores, otherwise
2670 * tries creating NR_PMD_THREADS pmd threads. */
2671 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
2672 for (i = 0; i < can_have; i++) {
2673 struct dp_netdev_pmd_thread *pmd = xzalloc(sizeof *pmd);
2674 int core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
2676 dp_netdev_configure_pmd(pmd, dp, i, core_id, numa_id);
2677 /* Each thread will distribute all devices rx-queues among
2679 pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd);
2681 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
2686 /* Called after pmd threads config change. Restarts pmd threads with
2687 * new configuration. */
2689 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
2691 struct dp_netdev_port *port;
2693 CMAP_FOR_EACH (port, node, &dp->ports) {
2694 if (netdev_is_pmd(port->netdev)) {
2695 int numa_id = netdev_get_numa_id(port->netdev);
2697 dp_netdev_set_pmds_on_numa(dp, numa_id);
2703 dpif_netdev_get_datapath_version(void)
2705 return xstrdup("<built-in>");
2709 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
2712 long long now = time_msec();
2715 atomic_store_relaxed(&netdev_flow->stats.used, now);
2716 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
2717 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
2718 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
2720 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
2724 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
2725 enum dp_stat_type type, int cnt)
2727 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
2731 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
2732 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
2733 enum dpif_upcall_type type, const struct nlattr *userdata,
2734 struct ofpbuf *actions, struct ofpbuf *put_actions)
2736 struct dp_netdev *dp = pmd->dp;
2738 if (OVS_UNLIKELY(!dp->upcall_cb)) {
2742 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
2743 struct ds ds = DS_EMPTY_INITIALIZER;
2747 ofpbuf_init(&key, 0);
2748 odp_flow_key_from_flow(&key, flow, &wc->masks, flow->in_port.odp_port,
2750 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
2751 dp_packet_size(packet_));
2753 odp_flow_key_format(key.data, key.size, &ds);
2755 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
2756 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
2758 ofpbuf_uninit(&key);
2764 return dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
2765 actions, wc, put_actions, dp->upcall_aux);
2768 static inline uint32_t
2769 dpif_netdev_packet_get_dp_hash(struct dp_packet *packet,
2770 const struct miniflow *mf)
2774 hash = dp_packet_get_dp_hash(packet);
2775 if (OVS_UNLIKELY(!hash)) {
2776 hash = miniflow_hash_5tuple(mf, 0);
2777 dp_packet_set_dp_hash(packet, hash);
2782 struct packet_batch {
2783 unsigned int packet_count;
2784 unsigned int byte_count;
2787 struct dp_netdev_flow *flow;
2789 struct dp_packet *packets[NETDEV_MAX_RX_BATCH];
2793 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
2794 const struct miniflow *mf)
2796 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
2797 batch->packets[batch->packet_count++] = packet;
2798 batch->byte_count += dp_packet_size(packet);
2802 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
2806 batch->packet_count = 0;
2807 batch->byte_count = 0;
2808 batch->tcp_flags = 0;
2812 packet_batch_execute(struct packet_batch *batch,
2813 struct dp_netdev_pmd_thread *pmd)
2815 struct dp_netdev_actions *actions;
2816 struct dp_netdev_flow *flow = batch->flow;
2818 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
2821 actions = dp_netdev_flow_get_actions(flow);
2823 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
2824 actions->actions, actions->size);
2826 dp_netdev_count_packet(pmd, DP_STAT_HIT, batch->packet_count);
2830 dp_netdev_queue_batches(struct dp_packet *pkt,
2831 struct dp_netdev_flow *flow, const struct miniflow *mf,
2832 struct packet_batch *batches, size_t *n_batches,
2835 struct packet_batch *batch = NULL;
2838 if (OVS_UNLIKELY(!flow)) {
2841 /* XXX: This O(n^2) algortihm makes sense if we're operating under the
2842 * assumption that the number of distinct flows (and therefore the
2843 * number of distinct batches) is quite small. If this turns out not
2844 * to be the case, it may make sense to pre sort based on the
2845 * netdev_flow pointer. That done we can get the appropriate batching
2846 * in O(n * log(n)) instead. */
2847 for (j = *n_batches - 1; j >= 0; j--) {
2848 if (batches[j].flow == flow) {
2849 batch = &batches[j];
2850 packet_batch_update(batch, pkt, mf);
2854 if (OVS_UNLIKELY(*n_batches >= max_batches)) {
2858 batch = &batches[(*n_batches)++];
2859 packet_batch_init(batch, flow);
2860 packet_batch_update(batch, pkt, mf);
2865 dp_packet_swap(struct dp_packet **a, struct dp_packet **b)
2867 struct dp_packet *tmp = *a;
2872 /* Try to process all ('cnt') the 'packets' using only the exact match cache
2873 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', or if there
2874 * is no matching batch for a packet's flow, the miniflow is copied into 'keys'
2875 * and the packet pointer is moved at the beginning of the 'packets' array.
2877 * The function returns the number of packets that needs to be processed in the
2878 * 'packets' array (they have been moved to the beginning of the vector).
2880 static inline size_t
2881 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
2882 size_t cnt, struct netdev_flow_key *keys)
2884 struct netdev_flow_key key;
2885 struct packet_batch batches[4];
2886 struct emc_cache *flow_cache = &pmd->flow_cache;
2887 size_t n_batches, i;
2888 size_t notfound_cnt = 0;
2891 miniflow_initialize(&key.mf, key.buf);
2892 for (i = 0; i < cnt; i++) {
2893 struct dp_netdev_flow *flow;
2895 if (OVS_UNLIKELY(dp_packet_size(packets[i]) < ETH_HEADER_LEN)) {
2896 dp_packet_delete(packets[i]);
2900 miniflow_extract(packets[i], &key.mf);
2901 key.len = 0; /* Not computed yet. */
2902 key.hash = dpif_netdev_packet_get_dp_hash(packets[i], &key.mf);
2904 flow = emc_lookup(flow_cache, &key);
2905 if (OVS_UNLIKELY(!dp_netdev_queue_batches(packets[i], flow, &key.mf,
2906 batches, &n_batches,
2907 ARRAY_SIZE(batches)))) {
2908 if (i != notfound_cnt) {
2909 dp_packet_swap(&packets[i], &packets[notfound_cnt]);
2912 keys[notfound_cnt++] = key;
2916 for (i = 0; i < n_batches; i++) {
2917 packet_batch_execute(&batches[i], pmd);
2920 return notfound_cnt;
2924 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
2925 struct dp_packet **packets, size_t cnt,
2926 struct netdev_flow_key *keys)
2928 #if !defined(__CHECKER__) && !defined(_WIN32)
2929 const size_t PKT_ARRAY_SIZE = cnt;
2931 /* Sparse or MSVC doesn't like variable length array. */
2932 enum { PKT_ARRAY_SIZE = NETDEV_MAX_RX_BATCH };
2934 struct packet_batch batches[PKT_ARRAY_SIZE];
2935 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
2936 struct dp_netdev *dp = pmd->dp;
2937 struct emc_cache *flow_cache = &pmd->flow_cache;
2938 size_t n_batches, i;
2941 for (i = 0; i < cnt; i++) {
2942 /* Key length is needed in all the cases, hash computed on demand. */
2943 keys[i].len = netdev_flow_key_size(count_1bits(keys[i].mf.map));
2945 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
2946 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
2947 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
2948 struct ofpbuf actions, put_actions;
2949 int miss_cnt = 0, lost_cnt = 0;
2952 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
2953 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
2955 for (i = 0; i < cnt; i++) {
2956 struct dp_netdev_flow *netdev_flow;
2957 struct ofpbuf *add_actions;
2961 if (OVS_LIKELY(rules[i])) {
2965 /* It's possible that an earlier slow path execution installed
2966 * a rule covering this flow. In this case, it's a lot cheaper
2967 * to catch it here than execute a miss. */
2968 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
2970 rules[i] = &netdev_flow->cr;
2976 miniflow_expand(&keys[i].mf, &match.flow);
2978 ofpbuf_clear(&actions);
2979 ofpbuf_clear(&put_actions);
2981 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
2982 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
2983 &ufid, DPIF_UC_MISS, NULL, &actions,
2985 if (OVS_UNLIKELY(error && error != ENOSPC)) {
2986 dp_packet_delete(packets[i]);
2991 /* We can't allow the packet batching in the next loop to execute
2992 * the actions. Otherwise, if there are any slow path actions,
2993 * we'll send the packet up twice. */
2994 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
2995 actions.data, actions.size);
2997 add_actions = put_actions.size ? &put_actions : &actions;
2998 if (OVS_LIKELY(error != ENOSPC)) {
2999 /* XXX: There's a race window where a flow covering this packet
3000 * could have already been installed since we last did the flow
3001 * lookup before upcall. This could be solved by moving the
3002 * mutex lock outside the loop, but that's an awful long time
3003 * to be locking everyone out of making flow installs. If we
3004 * move to a per-core classifier, it would be reasonable. */
3005 ovs_mutex_lock(&pmd->flow_mutex);
3006 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3007 if (OVS_LIKELY(!netdev_flow)) {
3008 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3012 ovs_mutex_unlock(&pmd->flow_mutex);
3014 emc_insert(flow_cache, &keys[i], netdev_flow);
3018 ofpbuf_uninit(&actions);
3019 ofpbuf_uninit(&put_actions);
3020 fat_rwlock_unlock(&dp->upcall_rwlock);
3021 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3022 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3023 } else if (OVS_UNLIKELY(any_miss)) {
3024 int dropped_cnt = 0;
3026 for (i = 0; i < cnt; i++) {
3027 if (OVS_UNLIKELY(!rules[i])) {
3028 dp_packet_delete(packets[i]);
3033 dp_netdev_count_packet(pmd, DP_STAT_MISS, dropped_cnt);
3034 dp_netdev_count_packet(pmd, DP_STAT_LOST, dropped_cnt);
3038 for (i = 0; i < cnt; i++) {
3039 struct dp_packet *packet = packets[i];
3040 struct dp_netdev_flow *flow;
3042 if (OVS_UNLIKELY(!rules[i])) {
3046 flow = dp_netdev_flow_cast(rules[i]);
3048 emc_insert(flow_cache, &keys[i], flow);
3049 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches,
3050 &n_batches, ARRAY_SIZE(batches));
3053 for (i = 0; i < n_batches; i++) {
3054 packet_batch_execute(&batches[i], pmd);
3059 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3060 struct dp_packet **packets, int cnt)
3062 #if !defined(__CHECKER__) && !defined(_WIN32)
3063 const size_t PKT_ARRAY_SIZE = cnt;
3065 /* Sparse or MSVC doesn't like variable length array. */
3066 enum { PKT_ARRAY_SIZE = NETDEV_MAX_RX_BATCH };
3068 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3071 newcnt = emc_processing(pmd, packets, cnt, keys);
3072 if (OVS_UNLIKELY(newcnt)) {
3073 fast_path_processing(pmd, packets, newcnt, keys);
3077 struct dp_netdev_execute_aux {
3078 struct dp_netdev_pmd_thread *pmd;
3082 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3085 struct dp_netdev *dp = get_dp_netdev(dpif);
3086 dp->upcall_aux = aux;
3091 dp_netdev_drop_packets(struct dp_packet ** packets, int cnt, bool may_steal)
3096 for (i = 0; i < cnt; i++) {
3097 dp_packet_delete(packets[i]);
3103 push_tnl_action(const struct dp_netdev *dp,
3104 const struct nlattr *attr,
3105 struct dp_packet **packets, int cnt)
3107 struct dp_netdev_port *tun_port;
3108 const struct ovs_action_push_tnl *data;
3110 data = nl_attr_get(attr);
3112 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3116 netdev_push_header(tun_port->netdev, packets, cnt, data);
3122 dp_netdev_clone_pkt_batch(struct dp_packet **tnl_pkt,
3123 struct dp_packet **packets, int cnt)
3127 for (i = 0; i < cnt; i++) {
3128 tnl_pkt[i] = dp_packet_clone(packets[i]);
3133 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3134 const struct nlattr *a, bool may_steal)
3135 OVS_NO_THREAD_SAFETY_ANALYSIS
3137 struct dp_netdev_execute_aux *aux = aux_;
3138 uint32_t *depth = recirc_depth_get();
3139 struct dp_netdev_pmd_thread *pmd= aux->pmd;
3140 struct dp_netdev *dp= pmd->dp;
3141 int type = nl_attr_type(a);
3142 struct dp_netdev_port *p;
3145 switch ((enum ovs_action_attr)type) {
3146 case OVS_ACTION_ATTR_OUTPUT:
3147 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3148 if (OVS_LIKELY(p)) {
3149 netdev_send(p->netdev, pmd->core_id, packets, cnt, may_steal);
3154 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3155 if (*depth < MAX_RECIRC_DEPTH) {
3156 struct dp_packet *tnl_pkt[NETDEV_MAX_RX_BATCH];
3160 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3164 err = push_tnl_action(dp, a, packets, cnt);
3167 dp_netdev_input(pmd, packets, cnt);
3170 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3176 case OVS_ACTION_ATTR_TUNNEL_POP:
3177 if (*depth < MAX_RECIRC_DEPTH) {
3178 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3180 p = dp_netdev_lookup_port(dp, portno);
3182 struct dp_packet *tnl_pkt[NETDEV_MAX_RX_BATCH];
3186 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3190 err = netdev_pop_header(p->netdev, packets, cnt);
3193 for (i = 0; i < cnt; i++) {
3194 packets[i]->md.in_port.odp_port = portno;
3198 dp_netdev_input(pmd, packets, cnt);
3201 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3208 case OVS_ACTION_ATTR_USERSPACE:
3209 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3210 const struct nlattr *userdata;
3211 struct ofpbuf actions;
3215 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3216 ofpbuf_init(&actions, 0);
3218 for (i = 0; i < cnt; i++) {
3221 ofpbuf_clear(&actions);
3223 flow_extract(packets[i], &flow);
3224 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3225 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3226 DPIF_UC_ACTION, userdata,&actions,
3228 if (!error || error == ENOSPC) {
3229 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3230 actions.data, actions.size);
3231 } else if (may_steal) {
3232 dp_packet_delete(packets[i]);
3235 ofpbuf_uninit(&actions);
3236 fat_rwlock_unlock(&dp->upcall_rwlock);
3242 case OVS_ACTION_ATTR_RECIRC:
3243 if (*depth < MAX_RECIRC_DEPTH) {
3246 for (i = 0; i < cnt; i++) {
3247 struct dp_packet *recirc_pkt;
3249 recirc_pkt = (may_steal) ? packets[i]
3250 : dp_packet_clone(packets[i]);
3252 recirc_pkt->md.recirc_id = nl_attr_get_u32(a);
3254 /* Hash is private to each packet */
3255 recirc_pkt->md.dp_hash = dp_packet_get_dp_hash(packets[i]);
3257 dp_netdev_input(pmd, &recirc_pkt, 1);
3264 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3267 case OVS_ACTION_ATTR_PUSH_VLAN:
3268 case OVS_ACTION_ATTR_POP_VLAN:
3269 case OVS_ACTION_ATTR_PUSH_MPLS:
3270 case OVS_ACTION_ATTR_POP_MPLS:
3271 case OVS_ACTION_ATTR_SET:
3272 case OVS_ACTION_ATTR_SET_MASKED:
3273 case OVS_ACTION_ATTR_SAMPLE:
3274 case OVS_ACTION_ATTR_HASH:
3275 case OVS_ACTION_ATTR_UNSPEC:
3276 case __OVS_ACTION_ATTR_MAX:
3280 dp_netdev_drop_packets(packets, cnt, may_steal);
3284 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3285 struct dp_packet **packets, int cnt,
3287 const struct nlattr *actions, size_t actions_len)
3289 struct dp_netdev_execute_aux aux = { pmd };
3291 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3292 actions_len, dp_execute_cb);
3295 const struct dpif_class dpif_netdev_class = {
3297 dpif_netdev_enumerate,
3298 dpif_netdev_port_open_type,
3301 dpif_netdev_destroy,
3304 dpif_netdev_get_stats,
3305 dpif_netdev_port_add,
3306 dpif_netdev_port_del,
3307 dpif_netdev_port_query_by_number,
3308 dpif_netdev_port_query_by_name,
3309 NULL, /* port_get_pid */
3310 dpif_netdev_port_dump_start,
3311 dpif_netdev_port_dump_next,
3312 dpif_netdev_port_dump_done,
3313 dpif_netdev_port_poll,
3314 dpif_netdev_port_poll_wait,
3315 dpif_netdev_flow_flush,
3316 dpif_netdev_flow_dump_create,
3317 dpif_netdev_flow_dump_destroy,
3318 dpif_netdev_flow_dump_thread_create,
3319 dpif_netdev_flow_dump_thread_destroy,
3320 dpif_netdev_flow_dump_next,
3321 dpif_netdev_operate,
3322 NULL, /* recv_set */
3323 NULL, /* handlers_set */
3324 dpif_netdev_pmd_set,
3325 dpif_netdev_queue_to_priority,
3327 NULL, /* recv_wait */
3328 NULL, /* recv_purge */
3329 dpif_netdev_register_upcall_cb,
3330 dpif_netdev_enable_upcall,
3331 dpif_netdev_disable_upcall,
3332 dpif_netdev_get_datapath_version,
3336 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3337 const char *argv[], void *aux OVS_UNUSED)
3339 struct dp_netdev_port *old_port;
3340 struct dp_netdev_port *new_port;
3341 struct dp_netdev *dp;
3344 ovs_mutex_lock(&dp_netdev_mutex);
3345 dp = shash_find_data(&dp_netdevs, argv[1]);
3346 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3347 ovs_mutex_unlock(&dp_netdev_mutex);
3348 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3351 ovs_refcount_ref(&dp->ref_cnt);
3352 ovs_mutex_unlock(&dp_netdev_mutex);
3354 ovs_mutex_lock(&dp->port_mutex);
3355 if (get_port_by_name(dp, argv[2], &old_port)) {
3356 unixctl_command_reply_error(conn, "unknown port");
3360 port_no = u32_to_odp(atoi(argv[3]));
3361 if (!port_no || port_no == ODPP_NONE) {
3362 unixctl_command_reply_error(conn, "bad port number");
3365 if (dp_netdev_lookup_port(dp, port_no)) {
3366 unixctl_command_reply_error(conn, "port number already in use");
3370 /* Remove old port. */
3371 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3372 ovsrcu_postpone(free, old_port);
3374 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3375 new_port = xmemdup(old_port, sizeof *old_port);
3376 new_port->port_no = port_no;
3377 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3379 seq_change(dp->port_seq);
3380 unixctl_command_reply(conn, NULL);
3383 ovs_mutex_unlock(&dp->port_mutex);
3384 dp_netdev_unref(dp);
3388 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3389 const char *argv[], void *aux OVS_UNUSED)
3391 struct dp_netdev_port *port;
3392 struct dp_netdev *dp;
3394 ovs_mutex_lock(&dp_netdev_mutex);
3395 dp = shash_find_data(&dp_netdevs, argv[1]);
3396 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3397 ovs_mutex_unlock(&dp_netdev_mutex);
3398 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3401 ovs_refcount_ref(&dp->ref_cnt);
3402 ovs_mutex_unlock(&dp_netdev_mutex);
3404 ovs_mutex_lock(&dp->port_mutex);
3405 if (get_port_by_name(dp, argv[2], &port)) {
3406 unixctl_command_reply_error(conn, "unknown port");
3407 } else if (port->port_no == ODPP_LOCAL) {
3408 unixctl_command_reply_error(conn, "can't delete local port");
3410 do_del_port(dp, port);
3411 unixctl_command_reply(conn, NULL);
3413 ovs_mutex_unlock(&dp->port_mutex);
3415 dp_netdev_unref(dp);
3419 dpif_dummy_register__(const char *type)
3421 struct dpif_class *class;
3423 class = xmalloc(sizeof *class);
3424 *class = dpif_netdev_class;
3425 class->type = xstrdup(type);
3426 dp_register_provider(class);
3430 dpif_dummy_register(bool override)
3437 dp_enumerate_types(&types);
3438 SSET_FOR_EACH (type, &types) {
3439 if (!dp_unregister_provider(type)) {
3440 dpif_dummy_register__(type);
3443 sset_destroy(&types);
3446 dpif_dummy_register__("dummy");
3448 unixctl_command_register("dpif-dummy/change-port-number",
3449 "dp port new-number",
3450 3, 3, dpif_dummy_change_port_number, NULL);
3451 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3452 2, 2, dpif_dummy_delete_port, NULL);
3455 /* Datapath Classifier. */
3457 /* A set of rules that all have the same fields wildcarded. */
3458 struct dpcls_subtable {
3459 /* The fields are only used by writers. */
3460 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3462 /* These fields are accessed by readers. */
3463 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3464 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3465 /* 'mask' must be the last field, additional space is allocated here. */
3468 /* Initializes 'cls' as a classifier that initially contains no classification
3471 dpcls_init(struct dpcls *cls)
3473 cmap_init(&cls->subtables_map);
3474 pvector_init(&cls->subtables);
3478 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3480 pvector_remove(&cls->subtables, subtable);
3481 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3482 subtable->mask.hash);
3483 cmap_destroy(&subtable->rules);
3484 ovsrcu_postpone(free, subtable);
3487 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3488 * caller's responsibility.
3489 * May only be called after all the readers have been terminated. */
3491 dpcls_destroy(struct dpcls *cls)
3494 struct dpcls_subtable *subtable;
3496 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3497 dpcls_destroy_subtable(cls, subtable);
3499 cmap_destroy(&cls->subtables_map);
3500 pvector_destroy(&cls->subtables);
3504 static struct dpcls_subtable *
3505 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3507 struct dpcls_subtable *subtable;
3509 /* Need to add one. */
3510 subtable = xmalloc(sizeof *subtable
3511 - sizeof subtable->mask.mf + mask->len);
3512 cmap_init(&subtable->rules);
3513 netdev_flow_key_clone(&subtable->mask, mask);
3514 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3515 pvector_insert(&cls->subtables, subtable, 0);
3516 pvector_publish(&cls->subtables);
3521 static inline struct dpcls_subtable *
3522 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3524 struct dpcls_subtable *subtable;
3526 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3527 &cls->subtables_map) {
3528 if (netdev_flow_key_equal(&subtable->mask, mask)) {
3532 return dpcls_create_subtable(cls, mask);
3535 /* Insert 'rule' into 'cls'. */
3537 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
3538 const struct netdev_flow_key *mask)
3540 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
3542 rule->mask = &subtable->mask;
3543 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
3546 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3548 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
3550 struct dpcls_subtable *subtable;
3552 ovs_assert(rule->mask);
3554 INIT_CONTAINER(subtable, rule->mask, mask);
3556 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
3558 dpcls_destroy_subtable(cls, subtable);
3559 pvector_publish(&cls->subtables);
3563 /* Returns true if 'target' satisifies 'key' in 'mask', that is, if each 1-bit
3564 * in 'mask' the values in 'key' and 'target' are the same.
3566 * Note: 'key' and 'mask' have the same mask, and 'key' is already masked. */
3568 dpcls_rule_matches_key(const struct dpcls_rule *rule,
3569 const struct netdev_flow_key *target)
3571 const uint64_t *keyp = rule->flow.mf.inline_values;
3572 const uint64_t *maskp = rule->mask->mf.inline_values;
3573 uint64_t target_u64;
3575 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(target_u64, target, rule->flow.mf.map) {
3576 if (OVS_UNLIKELY((target_u64 & *maskp++) != *keyp++)) {
3583 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3584 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3585 * NULL it is skipped.
3587 * This function is optimized for use in the userspace datapath and therefore
3588 * does not implement a lot of features available in the standard
3589 * classifier_lookup() function. Specifically, it does not implement
3590 * priorities, instead returning any rule which matches the flow.
3592 * Returns true if all flows found a corresponding rule. */
3594 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
3595 struct dpcls_rule **rules, const size_t cnt)
3597 /* The batch size 16 was experimentally found faster than 8 or 32. */
3598 typedef uint16_t map_type;
3599 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3601 #if !defined(__CHECKER__) && !defined(_WIN32)
3602 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
3604 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_RX_BATCH, MAP_BITS) };
3606 map_type maps[N_MAPS];
3607 struct dpcls_subtable *subtable;
3609 memset(maps, 0xff, sizeof maps);
3610 if (cnt % MAP_BITS) {
3611 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
3613 memset(rules, 0, cnt * sizeof *rules);
3615 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
3616 const struct netdev_flow_key *mkeys = keys;
3617 struct dpcls_rule **mrules = rules;
3618 map_type remains = 0;
3621 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
3623 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
3624 uint32_t hashes[MAP_BITS];
3625 const struct cmap_node *nodes[MAP_BITS];
3626 unsigned long map = maps[m];
3630 continue; /* Skip empty maps. */
3633 /* Compute hashes for the remaining keys. */
3634 ULONG_FOR_EACH_1(i, map) {
3635 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
3639 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
3640 /* Check results. */
3641 ULONG_FOR_EACH_1(i, map) {
3642 struct dpcls_rule *rule;
3644 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
3645 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
3650 ULONG_SET0(map, i); /* Did not match. */
3652 ; /* Keep Sparse happy. */
3654 maps[m] &= ~map; /* Clear the found rules. */
3658 return true; /* All found. */
3661 return false; /* Some misses. */