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 13
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_EXACT_HIT, /* Packets that had an exact match (emc). */
224 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
225 DP_STAT_MISS, /* Packets that did not match. */
226 DP_STAT_LOST, /* Packets not passed up to the client. */
230 enum pmd_cycles_counter_type {
231 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
232 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
236 /* A port in a netdev-based datapath. */
237 struct dp_netdev_port {
238 struct pkt_metadata md;
239 struct netdev *netdev;
240 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
241 struct netdev_saved_flags *sf;
242 struct netdev_rxq **rxq;
243 struct ovs_refcount ref_cnt;
244 char *type; /* Port type as requested by user. */
247 /* Contained by struct dp_netdev_flow's 'stats' member. */
248 struct dp_netdev_flow_stats {
249 atomic_llong used; /* Last used time, in monotonic msecs. */
250 atomic_ullong packet_count; /* Number of packets matched. */
251 atomic_ullong byte_count; /* Number of bytes matched. */
252 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
255 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
261 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
262 * its pmd thread's classifier. The text below calls this classifier 'cls'.
267 * The thread safety rules described here for "struct dp_netdev_flow" are
268 * motivated by two goals:
270 * - Prevent threads that read members of "struct dp_netdev_flow" from
271 * reading bad data due to changes by some thread concurrently modifying
274 * - Prevent two threads making changes to members of a given "struct
275 * dp_netdev_flow" from interfering with each other.
281 * A flow 'flow' may be accessed without a risk of being freed during an RCU
282 * grace period. Code that needs to hold onto a flow for a while
283 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
285 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
286 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
289 * Some members, marked 'const', are immutable. Accessing other members
290 * requires synchronization, as noted in more detail below.
292 struct dp_netdev_flow {
293 const struct flow flow; /* Unmasked flow that created this entry. */
294 /* Hash table index by unmasked flow. */
295 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
297 const ovs_u128 ufid; /* Unique flow identifier. */
298 const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */
301 /* Number of references.
302 * The classifier owns one reference.
303 * Any thread trying to keep a rule from being freed should hold its own
305 struct ovs_refcount ref_cnt;
310 struct dp_netdev_flow_stats stats;
313 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
315 /* While processing a group of input packets, the datapath uses the next
316 * member to store a pointer to the output batch for the flow. It is
317 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
318 * packet_batch_init() and packet_batch_execute()). */
319 struct packet_batch *batch;
321 /* Packet classification. */
322 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
323 /* 'cr' must be the last member. */
326 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
327 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
328 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
331 /* A set of datapath actions within a "struct dp_netdev_flow".
337 * A struct dp_netdev_actions 'actions' is protected with RCU. */
338 struct dp_netdev_actions {
339 /* These members are immutable: they do not change during the struct's
341 unsigned int size; /* Size of 'actions', in bytes. */
342 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
345 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
347 struct dp_netdev_actions *dp_netdev_flow_get_actions(
348 const struct dp_netdev_flow *);
349 static void dp_netdev_actions_free(struct dp_netdev_actions *);
351 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
352 struct dp_netdev_pmd_stats {
353 /* Indexed by DP_STAT_*. */
354 atomic_ullong n[DP_N_STATS];
357 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
358 struct dp_netdev_pmd_cycles {
359 /* Indexed by PMD_CYCLES_*. */
360 atomic_ullong n[PMD_N_CYCLES];
363 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
364 * the performance overhead of interrupt processing. Therefore netdev can
365 * not implement rx-wait for these devices. dpif-netdev needs to poll
366 * these device to check for recv buffer. pmd-thread does polling for
367 * devices assigned to itself.
369 * DPDK used PMD for accessing NIC.
371 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
372 * I/O of all non-pmd threads. There will be no actual thread created
375 * Each struct has its own flow table and classifier. Packets received
376 * from managed ports are looked up in the corresponding pmd thread's
377 * flow table, and are executed with the found actions.
379 struct dp_netdev_pmd_thread {
380 struct dp_netdev *dp;
381 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
382 struct cmap_node node; /* In 'dp->poll_threads'. */
384 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
385 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
387 /* Per thread exact-match cache. Note, the instance for cpu core
388 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
389 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
390 * instances will only be accessed by its own pmd thread. */
391 struct emc_cache flow_cache;
393 /* Classifier and Flow-Table.
395 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
396 * changes to 'cls' must be made while still holding the 'flow_mutex'.
398 struct ovs_mutex flow_mutex;
400 struct cmap flow_table OVS_GUARDED; /* Flow table. */
403 struct dp_netdev_pmd_stats stats;
405 /* Cycles counters */
406 struct dp_netdev_pmd_cycles cycles;
408 /* Used to count cicles. See 'cycles_counter_end()' */
409 unsigned long long last_cycles;
411 struct latch exit_latch; /* For terminating the pmd thread. */
412 atomic_uint change_seq; /* For reloading pmd ports. */
414 int index; /* Idx of this pmd thread among pmd*/
415 /* threads on same numa node. */
416 unsigned core_id; /* CPU core id of this pmd thread. */
417 int numa_id; /* numa node id of this pmd thread. */
419 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
420 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
421 * values and subtracts them from 'stats' and 'cycles' before
422 * reporting to the user */
423 unsigned long long stats_zero[DP_N_STATS];
424 uint64_t cycles_zero[PMD_N_CYCLES];
427 #define PMD_INITIAL_SEQ 1
429 /* Interface to netdev-based datapath. */
432 struct dp_netdev *dp;
433 uint64_t last_port_seq;
436 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
437 struct dp_netdev_port **portp);
438 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
439 struct dp_netdev_port **portp);
440 static void dp_netdev_free(struct dp_netdev *)
441 OVS_REQUIRES(dp_netdev_mutex);
442 static int do_add_port(struct dp_netdev *dp, const char *devname,
443 const char *type, odp_port_t port_no)
444 OVS_REQUIRES(dp->port_mutex);
445 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
446 OVS_REQUIRES(dp->port_mutex);
447 static int dpif_netdev_open(const struct dpif_class *, const char *name,
448 bool create, struct dpif **);
449 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
450 struct dp_packet **, int c,
452 const struct nlattr *actions,
454 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
455 struct dp_packet **, int cnt);
457 static void dp_netdev_disable_upcall(struct dp_netdev *);
458 void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
459 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
460 struct dp_netdev *dp, int index,
461 unsigned core_id, int numa_id);
462 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
463 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
464 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
466 static struct dp_netdev_pmd_thread *
467 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
468 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
469 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
470 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
471 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
472 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
473 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
474 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
476 static inline bool emc_entry_alive(struct emc_entry *ce);
477 static void emc_clear_entry(struct emc_entry *ce);
480 emc_cache_init(struct emc_cache *flow_cache)
484 BUILD_ASSERT(offsetof(struct miniflow, inline_values) == sizeof(uint64_t));
486 flow_cache->sweep_idx = 0;
487 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
488 flow_cache->entries[i].flow = NULL;
489 flow_cache->entries[i].key.hash = 0;
490 flow_cache->entries[i].key.len
491 = offsetof(struct miniflow, inline_values);
492 miniflow_initialize(&flow_cache->entries[i].key.mf,
493 flow_cache->entries[i].key.buf);
498 emc_cache_uninit(struct emc_cache *flow_cache)
502 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
503 emc_clear_entry(&flow_cache->entries[i]);
507 /* Check and clear dead flow references slowly (one entry at each
510 emc_cache_slow_sweep(struct emc_cache *flow_cache)
512 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
514 if (!emc_entry_alive(entry)) {
515 emc_clear_entry(entry);
517 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
520 static struct dpif_netdev *
521 dpif_netdev_cast(const struct dpif *dpif)
523 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
524 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
527 static struct dp_netdev *
528 get_dp_netdev(const struct dpif *dpif)
530 return dpif_netdev_cast(dpif)->dp;
534 PMD_INFO_SHOW_STATS, /* show how cpu cycles are spent */
535 PMD_INFO_CLEAR_STATS /* set the cycles count to 0 */
539 pmd_info_show_stats(struct ds *reply,
540 struct dp_netdev_pmd_thread *pmd,
541 unsigned long long stats[DP_N_STATS],
542 uint64_t cycles[PMD_N_CYCLES])
544 unsigned long long total_packets = 0;
545 uint64_t total_cycles = 0;
548 /* These loops subtracts reference values ('*_zero') from the counters.
549 * Since loads and stores are relaxed, it might be possible for a '*_zero'
550 * value to be more recent than the current value we're reading from the
551 * counter. This is not a big problem, since these numbers are not
552 * supposed to be too accurate, but we should at least make sure that
553 * the result is not negative. */
554 for (i = 0; i < DP_N_STATS; i++) {
555 if (stats[i] > pmd->stats_zero[i]) {
556 stats[i] -= pmd->stats_zero[i];
561 if (i != DP_STAT_LOST) {
562 /* Lost packets are already included in DP_STAT_MISS */
563 total_packets += stats[i];
567 for (i = 0; i < PMD_N_CYCLES; i++) {
568 if (cycles[i] > pmd->cycles_zero[i]) {
569 cycles[i] -= pmd->cycles_zero[i];
574 total_cycles += cycles[i];
577 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
578 ? "main thread" : "pmd thread");
580 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
581 ds_put_format(reply, " numa_id %d", pmd->numa_id);
583 if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) {
584 ds_put_format(reply, " core_id %u", pmd->core_id);
586 ds_put_cstr(reply, ":\n");
589 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
590 "\tmiss:%llu\n\tlost:%llu\n",
591 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
592 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
594 if (total_cycles == 0) {
599 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
600 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
601 cycles[PMD_CYCLES_POLLING],
602 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
603 cycles[PMD_CYCLES_PROCESSING],
604 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
606 if (total_packets == 0) {
611 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
612 total_cycles / (double)total_packets,
613 total_cycles, total_packets);
616 "\tavg processing cycles per packet: "
617 "%.02f (%"PRIu64"/%llu)\n",
618 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
619 cycles[PMD_CYCLES_PROCESSING], total_packets);
623 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
624 struct dp_netdev_pmd_thread *pmd,
625 unsigned long long stats[DP_N_STATS],
626 uint64_t cycles[PMD_N_CYCLES])
630 /* We cannot write 'stats' and 'cycles' (because they're written by other
631 * threads) and we shouldn't change 'stats' (because they're used to count
632 * datapath stats, which must not be cleared here). Instead, we save the
633 * current values and subtract them from the values to be displayed in the
635 for (i = 0; i < DP_N_STATS; i++) {
636 pmd->stats_zero[i] = stats[i];
638 for (i = 0; i < PMD_N_CYCLES; i++) {
639 pmd->cycles_zero[i] = cycles[i];
644 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
647 struct ds reply = DS_EMPTY_INITIALIZER;
648 struct dp_netdev_pmd_thread *pmd;
649 struct dp_netdev *dp = NULL;
650 enum pmd_info_type type = *(enum pmd_info_type *) aux;
652 ovs_mutex_lock(&dp_netdev_mutex);
655 dp = shash_find_data(&dp_netdevs, argv[1]);
656 } else if (shash_count(&dp_netdevs) == 1) {
657 /* There's only one datapath */
658 dp = shash_first(&dp_netdevs)->data;
662 ovs_mutex_unlock(&dp_netdev_mutex);
663 unixctl_command_reply_error(conn,
664 "please specify an existing datapath");
668 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
669 unsigned long long stats[DP_N_STATS];
670 uint64_t cycles[PMD_N_CYCLES];
673 /* Read current stats and cycle counters */
674 for (i = 0; i < ARRAY_SIZE(stats); i++) {
675 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
677 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
678 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
681 if (type == PMD_INFO_CLEAR_STATS) {
682 pmd_info_clear_stats(&reply, pmd, stats, cycles);
683 } else if (type == PMD_INFO_SHOW_STATS) {
684 pmd_info_show_stats(&reply, pmd, stats, cycles);
688 ovs_mutex_unlock(&dp_netdev_mutex);
690 unixctl_command_reply(conn, ds_cstr(&reply));
695 dpif_netdev_init(void)
697 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
698 clear_aux = PMD_INFO_CLEAR_STATS;
700 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
701 0, 1, dpif_netdev_pmd_info,
703 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
704 0, 1, dpif_netdev_pmd_info,
710 dpif_netdev_enumerate(struct sset *all_dps,
711 const struct dpif_class *dpif_class)
713 struct shash_node *node;
715 ovs_mutex_lock(&dp_netdev_mutex);
716 SHASH_FOR_EACH(node, &dp_netdevs) {
717 struct dp_netdev *dp = node->data;
718 if (dpif_class != dp->class) {
719 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
720 * If the class doesn't match, skip this dpif. */
723 sset_add(all_dps, node->name);
725 ovs_mutex_unlock(&dp_netdev_mutex);
731 dpif_netdev_class_is_dummy(const struct dpif_class *class)
733 return class != &dpif_netdev_class;
737 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
739 return strcmp(type, "internal") ? type
740 : dpif_netdev_class_is_dummy(class) ? "dummy"
745 create_dpif_netdev(struct dp_netdev *dp)
747 uint16_t netflow_id = hash_string(dp->name, 0);
748 struct dpif_netdev *dpif;
750 ovs_refcount_ref(&dp->ref_cnt);
752 dpif = xmalloc(sizeof *dpif);
753 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
755 dpif->last_port_seq = seq_read(dp->port_seq);
760 /* Choose an unused, non-zero port number and return it on success.
761 * Return ODPP_NONE on failure. */
763 choose_port(struct dp_netdev *dp, const char *name)
764 OVS_REQUIRES(dp->port_mutex)
768 if (dp->class != &dpif_netdev_class) {
772 /* If the port name begins with "br", start the number search at
773 * 100 to make writing tests easier. */
774 if (!strncmp(name, "br", 2)) {
778 /* If the port name contains a number, try to assign that port number.
779 * This can make writing unit tests easier because port numbers are
781 for (p = name; *p != '\0'; p++) {
782 if (isdigit((unsigned char) *p)) {
783 port_no = start_no + strtol(p, NULL, 10);
784 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
785 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
786 return u32_to_odp(port_no);
793 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
794 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
795 return u32_to_odp(port_no);
803 create_dp_netdev(const char *name, const struct dpif_class *class,
804 struct dp_netdev **dpp)
805 OVS_REQUIRES(dp_netdev_mutex)
807 struct dp_netdev *dp;
810 dp = xzalloc(sizeof *dp);
811 shash_add(&dp_netdevs, name, dp);
813 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
814 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
815 ovs_refcount_init(&dp->ref_cnt);
816 atomic_flag_clear(&dp->destroyed);
818 ovs_mutex_init(&dp->port_mutex);
819 cmap_init(&dp->ports);
820 dp->port_seq = seq_create();
821 fat_rwlock_init(&dp->upcall_rwlock);
823 /* Disable upcalls by default. */
824 dp_netdev_disable_upcall(dp);
825 dp->upcall_aux = NULL;
826 dp->upcall_cb = NULL;
828 cmap_init(&dp->poll_threads);
829 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
830 ovsthread_key_create(&dp->per_pmd_key, NULL);
832 dp_netdev_set_nonpmd(dp);
833 dp->n_dpdk_rxqs = NR_QUEUE;
835 ovs_mutex_lock(&dp->port_mutex);
836 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
837 ovs_mutex_unlock(&dp->port_mutex);
843 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
849 dpif_netdev_open(const struct dpif_class *class, const char *name,
850 bool create, struct dpif **dpifp)
852 struct dp_netdev *dp;
855 ovs_mutex_lock(&dp_netdev_mutex);
856 dp = shash_find_data(&dp_netdevs, name);
858 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
860 error = (dp->class != class ? EINVAL
865 *dpifp = create_dpif_netdev(dp);
868 ovs_mutex_unlock(&dp_netdev_mutex);
874 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
875 OVS_NO_THREAD_SAFETY_ANALYSIS
877 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
878 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
880 /* Before freeing a lock we should release it */
881 fat_rwlock_unlock(&dp->upcall_rwlock);
882 fat_rwlock_destroy(&dp->upcall_rwlock);
885 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
886 * through the 'dp_netdevs' shash while freeing 'dp'. */
888 dp_netdev_free(struct dp_netdev *dp)
889 OVS_REQUIRES(dp_netdev_mutex)
891 struct dp_netdev_port *port;
893 shash_find_and_delete(&dp_netdevs, dp->name);
895 dp_netdev_destroy_all_pmds(dp);
896 cmap_destroy(&dp->poll_threads);
897 ovs_mutex_destroy(&dp->non_pmd_mutex);
898 ovsthread_key_delete(dp->per_pmd_key);
900 ovs_mutex_lock(&dp->port_mutex);
901 CMAP_FOR_EACH (port, node, &dp->ports) {
902 do_del_port(dp, port);
904 ovs_mutex_unlock(&dp->port_mutex);
906 seq_destroy(dp->port_seq);
907 cmap_destroy(&dp->ports);
909 /* Upcalls must be disabled at this point */
910 dp_netdev_destroy_upcall_lock(dp);
913 free(CONST_CAST(char *, dp->name));
918 dp_netdev_unref(struct dp_netdev *dp)
921 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
922 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
923 ovs_mutex_lock(&dp_netdev_mutex);
924 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
927 ovs_mutex_unlock(&dp_netdev_mutex);
932 dpif_netdev_close(struct dpif *dpif)
934 struct dp_netdev *dp = get_dp_netdev(dpif);
941 dpif_netdev_destroy(struct dpif *dpif)
943 struct dp_netdev *dp = get_dp_netdev(dpif);
945 if (!atomic_flag_test_and_set(&dp->destroyed)) {
946 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
947 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
955 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
956 * load/store semantics. While the increment is not atomic, the load and
957 * store operations are, making it impossible to read inconsistent values.
959 * This is used to update thread local stats counters. */
961 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
963 unsigned long long tmp;
965 atomic_read_relaxed(var, &tmp);
967 atomic_store_relaxed(var, tmp);
971 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
973 struct dp_netdev *dp = get_dp_netdev(dpif);
974 struct dp_netdev_pmd_thread *pmd;
976 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
977 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
978 unsigned long long n;
979 stats->n_flows += cmap_count(&pmd->flow_table);
981 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
983 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
985 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
986 stats->n_missed += n;
987 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
990 stats->n_masks = UINT32_MAX;
991 stats->n_mask_hit = UINT64_MAX;
997 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1001 if (pmd->core_id == NON_PMD_CORE_ID) {
1005 ovs_mutex_lock(&pmd->cond_mutex);
1006 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1007 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1008 ovs_mutex_unlock(&pmd->cond_mutex);
1011 /* Causes all pmd threads to reload its tx/rx devices.
1012 * Must be called after adding/removing ports. */
1014 dp_netdev_reload_pmds(struct dp_netdev *dp)
1016 struct dp_netdev_pmd_thread *pmd;
1018 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1019 dp_netdev_reload_pmd__(pmd);
1024 hash_port_no(odp_port_t port_no)
1026 return hash_int(odp_to_u32(port_no), 0);
1030 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1032 OVS_REQUIRES(dp->port_mutex)
1034 struct netdev_saved_flags *sf;
1035 struct dp_netdev_port *port;
1036 struct netdev *netdev;
1037 enum netdev_flags flags;
1038 const char *open_type;
1042 /* Reject devices already in 'dp'. */
1043 if (!get_port_by_name(dp, devname, &port)) {
1047 /* Open and validate network device. */
1048 open_type = dpif_netdev_port_open_type(dp->class, type);
1049 error = netdev_open(devname, open_type, &netdev);
1053 /* XXX reject non-Ethernet devices */
1055 netdev_get_flags(netdev, &flags);
1056 if (flags & NETDEV_LOOPBACK) {
1057 VLOG_ERR("%s: cannot add a loopback device", devname);
1058 netdev_close(netdev);
1062 if (netdev_is_pmd(netdev)) {
1063 int n_cores = ovs_numa_get_n_cores();
1065 if (n_cores == OVS_CORE_UNSPEC) {
1066 VLOG_ERR("%s, cannot get cpu core info", devname);
1069 /* There can only be ovs_numa_get_n_cores() pmd threads,
1070 * so creates a txq for each. */
1071 error = netdev_set_multiq(netdev, n_cores, dp->n_dpdk_rxqs);
1072 if (error && (error != EOPNOTSUPP)) {
1073 VLOG_ERR("%s, cannot set multiq", devname);
1077 port = xzalloc(sizeof *port);
1078 port->md = PKT_METADATA_INITIALIZER(port_no);
1079 port->netdev = netdev;
1080 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
1081 port->type = xstrdup(type);
1082 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1083 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1085 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
1086 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1087 devname, ovs_strerror(errno));
1088 netdev_close(netdev);
1096 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1098 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1099 netdev_rxq_close(port->rxq[i]);
1101 netdev_close(netdev);
1109 ovs_refcount_init(&port->ref_cnt);
1110 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1112 if (netdev_is_pmd(netdev)) {
1113 dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
1114 dp_netdev_reload_pmds(dp);
1116 seq_change(dp->port_seq);
1122 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1123 odp_port_t *port_nop)
1125 struct dp_netdev *dp = get_dp_netdev(dpif);
1126 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1127 const char *dpif_port;
1131 ovs_mutex_lock(&dp->port_mutex);
1132 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1133 if (*port_nop != ODPP_NONE) {
1134 port_no = *port_nop;
1135 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1137 port_no = choose_port(dp, dpif_port);
1138 error = port_no == ODPP_NONE ? EFBIG : 0;
1141 *port_nop = port_no;
1142 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1144 ovs_mutex_unlock(&dp->port_mutex);
1150 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1152 struct dp_netdev *dp = get_dp_netdev(dpif);
1155 ovs_mutex_lock(&dp->port_mutex);
1156 if (port_no == ODPP_LOCAL) {
1159 struct dp_netdev_port *port;
1161 error = get_port_by_number(dp, port_no, &port);
1163 do_del_port(dp, port);
1166 ovs_mutex_unlock(&dp->port_mutex);
1172 is_valid_port_number(odp_port_t port_no)
1174 return port_no != ODPP_NONE;
1177 static struct dp_netdev_port *
1178 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1180 struct dp_netdev_port *port;
1182 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1183 if (port->md.in_port.odp_port == port_no) {
1191 get_port_by_number(struct dp_netdev *dp,
1192 odp_port_t port_no, struct dp_netdev_port **portp)
1194 if (!is_valid_port_number(port_no)) {
1198 *portp = dp_netdev_lookup_port(dp, port_no);
1199 return *portp ? 0 : ENOENT;
1204 port_ref(struct dp_netdev_port *port)
1207 ovs_refcount_ref(&port->ref_cnt);
1212 port_try_ref(struct dp_netdev_port *port)
1215 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
1222 port_unref(struct dp_netdev_port *port)
1224 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1225 int n_rxq = netdev_n_rxq(port->netdev);
1228 netdev_close(port->netdev);
1229 netdev_restore_flags(port->sf);
1231 for (i = 0; i < n_rxq; i++) {
1232 netdev_rxq_close(port->rxq[i]);
1241 get_port_by_name(struct dp_netdev *dp,
1242 const char *devname, struct dp_netdev_port **portp)
1243 OVS_REQUIRES(dp->port_mutex)
1245 struct dp_netdev_port *port;
1247 CMAP_FOR_EACH (port, node, &dp->ports) {
1248 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1257 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1259 struct dp_netdev_pmd_thread *pmd;
1262 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1263 if (pmd->numa_id == numa_id) {
1271 /* Returns 'true' if there is a port with pmd netdev and the netdev
1272 * is on numa node 'numa_id'. */
1274 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1276 struct dp_netdev_port *port;
1278 CMAP_FOR_EACH (port, node, &dp->ports) {
1279 if (netdev_is_pmd(port->netdev)
1280 && netdev_get_numa_id(port->netdev) == numa_id) {
1290 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1291 OVS_REQUIRES(dp->port_mutex)
1293 cmap_remove(&dp->ports, &port->node,
1294 hash_odp_port(port->md.in_port.odp_port));
1295 seq_change(dp->port_seq);
1296 if (netdev_is_pmd(port->netdev)) {
1297 int numa_id = netdev_get_numa_id(port->netdev);
1299 /* If there is no netdev on the numa node, deletes the pmd threads
1300 * for that numa. Else, just reloads the queues. */
1301 if (!has_pmd_port_for_numa(dp, numa_id)) {
1302 dp_netdev_del_pmds_on_numa(dp, numa_id);
1304 dp_netdev_reload_pmds(dp);
1311 answer_port_query(const struct dp_netdev_port *port,
1312 struct dpif_port *dpif_port)
1314 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1315 dpif_port->type = xstrdup(port->type);
1316 dpif_port->port_no = port->md.in_port.odp_port;
1320 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1321 struct dpif_port *dpif_port)
1323 struct dp_netdev *dp = get_dp_netdev(dpif);
1324 struct dp_netdev_port *port;
1327 error = get_port_by_number(dp, port_no, &port);
1328 if (!error && dpif_port) {
1329 answer_port_query(port, dpif_port);
1336 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1337 struct dpif_port *dpif_port)
1339 struct dp_netdev *dp = get_dp_netdev(dpif);
1340 struct dp_netdev_port *port;
1343 ovs_mutex_lock(&dp->port_mutex);
1344 error = get_port_by_name(dp, devname, &port);
1345 if (!error && dpif_port) {
1346 answer_port_query(port, dpif_port);
1348 ovs_mutex_unlock(&dp->port_mutex);
1354 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1356 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1360 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1362 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1363 ovsrcu_postpone(dp_netdev_flow_free, flow);
1368 dp_netdev_flow_hash(const ovs_u128 *ufid)
1370 return ufid->u32[0];
1374 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1375 struct dp_netdev_flow *flow)
1376 OVS_REQUIRES(pmd->flow_mutex)
1378 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1380 dpcls_remove(&pmd->cls, &flow->cr);
1381 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1384 dp_netdev_flow_unref(flow);
1388 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1390 struct dp_netdev_flow *netdev_flow;
1392 ovs_mutex_lock(&pmd->flow_mutex);
1393 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1394 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1396 ovs_mutex_unlock(&pmd->flow_mutex);
1400 dpif_netdev_flow_flush(struct dpif *dpif)
1402 struct dp_netdev *dp = get_dp_netdev(dpif);
1403 struct dp_netdev_pmd_thread *pmd;
1405 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1406 dp_netdev_pmd_flow_flush(pmd);
1412 struct dp_netdev_port_state {
1413 struct cmap_position position;
1418 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1420 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1425 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1426 struct dpif_port *dpif_port)
1428 struct dp_netdev_port_state *state = state_;
1429 struct dp_netdev *dp = get_dp_netdev(dpif);
1430 struct cmap_node *node;
1433 node = cmap_next_position(&dp->ports, &state->position);
1435 struct dp_netdev_port *port;
1437 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1440 state->name = xstrdup(netdev_get_name(port->netdev));
1441 dpif_port->name = state->name;
1442 dpif_port->type = port->type;
1443 dpif_port->port_no = port->md.in_port.odp_port;
1454 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1456 struct dp_netdev_port_state *state = state_;
1463 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1465 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1466 uint64_t new_port_seq;
1469 new_port_seq = seq_read(dpif->dp->port_seq);
1470 if (dpif->last_port_seq != new_port_seq) {
1471 dpif->last_port_seq = new_port_seq;
1481 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1483 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1485 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1488 static struct dp_netdev_flow *
1489 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1491 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1494 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1496 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1499 /* netdev_flow_key utilities.
1501 * netdev_flow_key is basically a miniflow. We use these functions
1502 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1503 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1505 * - Since we are dealing exclusively with miniflows created by
1506 * miniflow_extract(), if the map is different the miniflow is different.
1507 * Therefore we can be faster by comparing the map and the miniflow in a
1509 * _ netdev_flow_key's miniflow has always inline values.
1510 * - These functions can be inlined by the compiler.
1512 * The following assertions make sure that what we're doing with miniflow is
1515 BUILD_ASSERT_DECL(offsetof(struct miniflow, inline_values)
1516 == sizeof(uint64_t));
1518 /* Given the number of bits set in the miniflow map, returns the size of the
1519 * 'netdev_flow_key.mf' */
1520 static inline uint32_t
1521 netdev_flow_key_size(uint32_t flow_u32s)
1523 return offsetof(struct miniflow, inline_values) +
1524 MINIFLOW_VALUES_SIZE(flow_u32s);
1528 netdev_flow_key_equal(const struct netdev_flow_key *a,
1529 const struct netdev_flow_key *b)
1531 /* 'b->len' may be not set yet. */
1532 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1535 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1536 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1537 * generated by miniflow_extract. */
1539 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1540 const struct miniflow *mf)
1542 return !memcmp(&key->mf, mf, key->len);
1546 netdev_flow_key_clone(struct netdev_flow_key *dst,
1547 const struct netdev_flow_key *src)
1550 offsetof(struct netdev_flow_key, mf) + src->len);
1555 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1556 const struct flow *src)
1558 struct dp_packet packet;
1559 uint64_t buf_stub[512 / 8];
1561 miniflow_initialize(&dst->mf, dst->buf);
1563 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1564 pkt_metadata_from_flow(&packet.md, src);
1565 flow_compose(&packet, src);
1566 miniflow_extract(&packet, &dst->mf);
1567 dp_packet_uninit(&packet);
1569 dst->len = netdev_flow_key_size(count_1bits(dst->mf.map));
1570 dst->hash = 0; /* Not computed yet. */
1573 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1575 netdev_flow_mask_init(struct netdev_flow_key *mask,
1576 const struct match *match)
1578 const uint64_t *mask_u64 = (const uint64_t *) &match->wc.masks;
1579 uint64_t *dst = mask->mf.inline_values;
1580 uint64_t map, mask_map = 0;
1584 /* Only check masks that make sense for the flow. */
1585 map = flow_wc_map(&match->flow);
1588 uint64_t rm1bit = rightmost_1bit(map);
1589 int i = raw_ctz(map);
1593 *dst++ = mask_u64[i];
1594 hash = hash_add64(hash, mask_u64[i]);
1599 mask->mf.values_inline = true;
1600 mask->mf.map = mask_map;
1602 hash = hash_add64(hash, mask_map);
1604 n = dst - mask->mf.inline_values;
1606 mask->hash = hash_finish(hash, n * 8);
1607 mask->len = netdev_flow_key_size(n);
1610 /* Initializes 'dst' as a copy of 'src' masked with 'mask'. */
1612 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1613 const struct flow *flow,
1614 const struct netdev_flow_key *mask)
1616 uint64_t *dst_u64 = dst->mf.inline_values;
1617 const uint64_t *mask_u64 = mask->mf.inline_values;
1621 dst->len = mask->len;
1622 dst->mf.values_inline = true;
1623 dst->mf.map = mask->mf.map;
1625 FLOW_FOR_EACH_IN_MAP(value, flow, mask->mf.map) {
1626 *dst_u64 = value & *mask_u64++;
1627 hash = hash_add64(hash, *dst_u64++);
1629 dst->hash = hash_finish(hash, (dst_u64 - dst->mf.inline_values) * 8);
1632 /* Iterate through all netdev_flow_key u64 values specified by 'MAP' */
1633 #define NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(VALUE, KEY, MAP) \
1634 for (struct mf_for_each_in_map_aux aux__ \
1635 = { (KEY)->mf.inline_values, (KEY)->mf.map, MAP }; \
1636 mf_get_next_in_map(&aux__, &(VALUE)); \
1639 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1641 static inline uint32_t
1642 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1643 const struct netdev_flow_key *mask)
1645 const uint64_t *p = mask->mf.inline_values;
1649 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(key_u64, key, mask->mf.map) {
1650 hash = hash_add64(hash, key_u64 & *p++);
1653 return hash_finish(hash, (p - mask->mf.inline_values) * 8);
1657 emc_entry_alive(struct emc_entry *ce)
1659 return ce->flow && !ce->flow->dead;
1663 emc_clear_entry(struct emc_entry *ce)
1666 dp_netdev_flow_unref(ce->flow);
1672 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1673 const struct netdev_flow_key *key)
1675 if (ce->flow != flow) {
1677 dp_netdev_flow_unref(ce->flow);
1680 if (dp_netdev_flow_ref(flow)) {
1687 netdev_flow_key_clone(&ce->key, key);
1692 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1693 struct dp_netdev_flow *flow)
1695 struct emc_entry *to_be_replaced = NULL;
1696 struct emc_entry *current_entry;
1698 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1699 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1700 /* We found the entry with the 'mf' miniflow */
1701 emc_change_entry(current_entry, flow, NULL);
1705 /* Replacement policy: put the flow in an empty (not alive) entry, or
1706 * in the first entry where it can be */
1708 || (emc_entry_alive(to_be_replaced)
1709 && !emc_entry_alive(current_entry))
1710 || current_entry->key.hash < to_be_replaced->key.hash) {
1711 to_be_replaced = current_entry;
1714 /* We didn't find the miniflow in the cache.
1715 * The 'to_be_replaced' entry is where the new flow will be stored */
1717 emc_change_entry(to_be_replaced, flow, key);
1720 static inline struct dp_netdev_flow *
1721 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1723 struct emc_entry *current_entry;
1725 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1726 if (current_entry->key.hash == key->hash
1727 && emc_entry_alive(current_entry)
1728 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1730 /* We found the entry with the 'key->mf' miniflow */
1731 return current_entry->flow;
1738 static struct dp_netdev_flow *
1739 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1740 const struct netdev_flow_key *key)
1742 struct dp_netdev_flow *netdev_flow;
1743 struct dpcls_rule *rule;
1745 dpcls_lookup(&pmd->cls, key, &rule, 1);
1746 netdev_flow = dp_netdev_flow_cast(rule);
1751 static struct dp_netdev_flow *
1752 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1753 const ovs_u128 *ufidp, const struct nlattr *key,
1756 struct dp_netdev_flow *netdev_flow;
1760 /* If a UFID is not provided, determine one based on the key. */
1761 if (!ufidp && key && key_len
1762 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1763 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1768 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1770 if (ovs_u128_equal(&netdev_flow->ufid, ufidp)) {
1780 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1781 struct dpif_flow_stats *stats)
1783 struct dp_netdev_flow *netdev_flow;
1784 unsigned long long n;
1788 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1790 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1791 stats->n_packets = n;
1792 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1794 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1796 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1797 stats->tcp_flags = flags;
1800 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1801 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1802 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1805 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1806 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1807 struct dpif_flow *flow, bool terse)
1810 memset(flow, 0, sizeof *flow);
1812 struct flow_wildcards wc;
1813 struct dp_netdev_actions *actions;
1816 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1819 offset = key_buf->size;
1820 flow->key = ofpbuf_tail(key_buf);
1821 odp_flow_key_from_flow(key_buf, &netdev_flow->flow, &wc.masks,
1822 netdev_flow->flow.in_port.odp_port, true);
1823 flow->key_len = key_buf->size - offset;
1826 offset = mask_buf->size;
1827 flow->mask = ofpbuf_tail(mask_buf);
1828 odp_flow_key_from_mask(mask_buf, &wc.masks, &netdev_flow->flow,
1829 odp_to_u32(wc.masks.in_port.odp_port),
1831 flow->mask_len = mask_buf->size - offset;
1834 actions = dp_netdev_flow_get_actions(netdev_flow);
1835 flow->actions = actions->actions;
1836 flow->actions_len = actions->size;
1839 flow->ufid = netdev_flow->ufid;
1840 flow->ufid_present = true;
1841 flow->pmd_id = netdev_flow->pmd_id;
1842 get_dpif_flow_stats(netdev_flow, &flow->stats);
1846 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1847 const struct nlattr *mask_key,
1848 uint32_t mask_key_len, const struct flow *flow,
1852 enum odp_key_fitness fitness;
1854 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1856 /* This should not happen: it indicates that
1857 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1858 * disagree on the acceptable form of a mask. Log the problem
1859 * as an error, with enough details to enable debugging. */
1860 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1862 if (!VLOG_DROP_ERR(&rl)) {
1866 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1868 VLOG_ERR("internal error parsing flow mask %s (%s)",
1869 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1876 enum mf_field_id id;
1877 /* No mask key, unwildcard everything except fields whose
1878 * prerequisities are not met. */
1879 memset(mask, 0x0, sizeof *mask);
1881 for (id = 0; id < MFF_N_IDS; ++id) {
1882 /* Skip registers and metadata. */
1883 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1884 && id != MFF_METADATA) {
1885 const struct mf_field *mf = mf_from_id(id);
1886 if (mf_are_prereqs_ok(mf, flow)) {
1887 mf_mask_field(mf, mask);
1893 /* Force unwildcard the in_port.
1895 * We need to do this even in the case where we unwildcard "everything"
1896 * above because "everything" only includes the 16-bit OpenFlow port number
1897 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1898 * port number mask->in_port.odp_port. */
1899 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1905 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1910 if (odp_flow_key_to_flow(key, key_len, flow)) {
1911 /* This should not happen: it indicates that odp_flow_key_from_flow()
1912 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1913 * flow. Log the problem as an error, with enough details to enable
1915 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1917 if (!VLOG_DROP_ERR(&rl)) {
1921 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1922 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1929 in_port = flow->in_port.odp_port;
1930 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1938 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1940 struct dp_netdev *dp = get_dp_netdev(dpif);
1941 struct dp_netdev_flow *netdev_flow;
1942 struct dp_netdev_pmd_thread *pmd;
1943 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1944 ? NON_PMD_CORE_ID : get->pmd_id;
1947 pmd = dp_netdev_get_pmd(dp, pmd_id);
1952 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1955 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1960 dp_netdev_pmd_unref(pmd);
1966 static struct dp_netdev_flow *
1967 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
1968 struct match *match, const ovs_u128 *ufid,
1969 const struct nlattr *actions, size_t actions_len)
1970 OVS_REQUIRES(pmd->flow_mutex)
1972 struct dp_netdev_flow *flow;
1973 struct netdev_flow_key mask;
1975 netdev_flow_mask_init(&mask, match);
1976 /* Make sure wc does not have metadata. */
1977 ovs_assert(!(mask.mf.map & (MINIFLOW_MAP(metadata) | MINIFLOW_MAP(regs))));
1979 /* Do not allocate extra space. */
1980 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
1981 memset(&flow->stats, 0, sizeof flow->stats);
1984 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
1985 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
1986 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
1987 ovs_refcount_init(&flow->ref_cnt);
1988 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
1990 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
1991 dpcls_insert(&pmd->cls, &flow->cr, &mask);
1993 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
1994 dp_netdev_flow_hash(&flow->ufid));
1996 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1998 struct ds ds = DS_EMPTY_INITIALIZER;
2000 match.flow = flow->flow;
2001 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2003 ds_put_cstr(&ds, "flow_add: ");
2004 odp_format_ufid(ufid, &ds);
2005 ds_put_cstr(&ds, " ");
2006 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2007 ds_put_cstr(&ds, ", actions:");
2008 format_odp_actions(&ds, actions, actions_len);
2010 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2019 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2021 struct dp_netdev *dp = get_dp_netdev(dpif);
2022 struct dp_netdev_flow *netdev_flow;
2023 struct netdev_flow_key key;
2024 struct dp_netdev_pmd_thread *pmd;
2027 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2028 ? NON_PMD_CORE_ID : put->pmd_id;
2031 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2035 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2036 put->mask, put->mask_len,
2037 &match.flow, &match.wc.masks);
2042 pmd = dp_netdev_get_pmd(dp, pmd_id);
2047 /* Must produce a netdev_flow_key for lookup.
2048 * This interface is no longer performance critical, since it is not used
2049 * for upcall processing any more. */
2050 netdev_flow_key_from_flow(&key, &match.flow);
2055 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2058 ovs_mutex_lock(&pmd->flow_mutex);
2059 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2061 if (put->flags & DPIF_FP_CREATE) {
2062 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2064 memset(put->stats, 0, sizeof *put->stats);
2066 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2076 if (put->flags & DPIF_FP_MODIFY
2077 && flow_equal(&match.flow, &netdev_flow->flow)) {
2078 struct dp_netdev_actions *new_actions;
2079 struct dp_netdev_actions *old_actions;
2081 new_actions = dp_netdev_actions_create(put->actions,
2084 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2085 ovsrcu_set(&netdev_flow->actions, new_actions);
2088 get_dpif_flow_stats(netdev_flow, put->stats);
2090 if (put->flags & DPIF_FP_ZERO_STATS) {
2091 /* XXX: The userspace datapath uses thread local statistics
2092 * (for flows), which should be updated only by the owning
2093 * thread. Since we cannot write on stats memory here,
2094 * we choose not to support this flag. Please note:
2095 * - This feature is currently used only by dpctl commands with
2097 * - Should the need arise, this operation can be implemented
2098 * by keeping a base value (to be update here) for each
2099 * counter, and subtracting it before outputting the stats */
2103 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2104 } else if (put->flags & DPIF_FP_CREATE) {
2107 /* Overlapping flow. */
2111 ovs_mutex_unlock(&pmd->flow_mutex);
2112 dp_netdev_pmd_unref(pmd);
2118 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2120 struct dp_netdev *dp = get_dp_netdev(dpif);
2121 struct dp_netdev_flow *netdev_flow;
2122 struct dp_netdev_pmd_thread *pmd;
2123 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2124 ? NON_PMD_CORE_ID : del->pmd_id;
2127 pmd = dp_netdev_get_pmd(dp, pmd_id);
2132 ovs_mutex_lock(&pmd->flow_mutex);
2133 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2137 get_dpif_flow_stats(netdev_flow, del->stats);
2139 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2143 ovs_mutex_unlock(&pmd->flow_mutex);
2144 dp_netdev_pmd_unref(pmd);
2149 struct dpif_netdev_flow_dump {
2150 struct dpif_flow_dump up;
2151 struct cmap_position poll_thread_pos;
2152 struct cmap_position flow_pos;
2153 struct dp_netdev_pmd_thread *cur_pmd;
2155 struct ovs_mutex mutex;
2158 static struct dpif_netdev_flow_dump *
2159 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2161 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2164 static struct dpif_flow_dump *
2165 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2167 struct dpif_netdev_flow_dump *dump;
2169 dump = xzalloc(sizeof *dump);
2170 dpif_flow_dump_init(&dump->up, dpif_);
2171 dump->up.terse = terse;
2172 ovs_mutex_init(&dump->mutex);
2178 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2180 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2182 ovs_mutex_destroy(&dump->mutex);
2187 struct dpif_netdev_flow_dump_thread {
2188 struct dpif_flow_dump_thread up;
2189 struct dpif_netdev_flow_dump *dump;
2190 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2191 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2194 static struct dpif_netdev_flow_dump_thread *
2195 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2197 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2200 static struct dpif_flow_dump_thread *
2201 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2203 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2204 struct dpif_netdev_flow_dump_thread *thread;
2206 thread = xmalloc(sizeof *thread);
2207 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2208 thread->dump = dump;
2213 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2215 struct dpif_netdev_flow_dump_thread *thread
2216 = dpif_netdev_flow_dump_thread_cast(thread_);
2222 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2223 struct dpif_flow *flows, int max_flows)
2225 struct dpif_netdev_flow_dump_thread *thread
2226 = dpif_netdev_flow_dump_thread_cast(thread_);
2227 struct dpif_netdev_flow_dump *dump = thread->dump;
2228 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2232 ovs_mutex_lock(&dump->mutex);
2233 if (!dump->status) {
2234 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2235 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2236 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2237 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2239 /* First call to dump_next(), extracts the first pmd thread.
2240 * If there is no pmd thread, returns immediately. */
2242 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2244 ovs_mutex_unlock(&dump->mutex);
2251 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2252 struct cmap_node *node;
2254 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2258 netdev_flows[n_flows] = CONTAINER_OF(node,
2259 struct dp_netdev_flow,
2262 /* When finishing dumping the current pmd thread, moves to
2264 if (n_flows < flow_limit) {
2265 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2266 dp_netdev_pmd_unref(pmd);
2267 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2273 /* Keeps the reference to next caller. */
2274 dump->cur_pmd = pmd;
2276 /* If the current dump is empty, do not exit the loop, since the
2277 * remaining pmds could have flows to be dumped. Just dumps again
2278 * on the new 'pmd'. */
2281 ovs_mutex_unlock(&dump->mutex);
2283 for (i = 0; i < n_flows; i++) {
2284 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2285 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2286 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2287 struct dpif_flow *f = &flows[i];
2288 struct ofpbuf key, mask;
2290 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2291 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2292 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2300 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2301 OVS_NO_THREAD_SAFETY_ANALYSIS
2303 struct dp_netdev *dp = get_dp_netdev(dpif);
2304 struct dp_netdev_pmd_thread *pmd;
2305 struct dp_packet *pp;
2307 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2308 dp_packet_size(execute->packet) > UINT16_MAX) {
2312 /* Tries finding the 'pmd'. If NULL is returned, that means
2313 * the current thread is a non-pmd thread and should use
2314 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2315 pmd = ovsthread_getspecific(dp->per_pmd_key);
2317 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2320 /* If the current thread is non-pmd thread, acquires
2321 * the 'non_pmd_mutex'. */
2322 if (pmd->core_id == NON_PMD_CORE_ID) {
2323 ovs_mutex_lock(&dp->non_pmd_mutex);
2324 ovs_mutex_lock(&dp->port_mutex);
2327 pp = execute->packet;
2328 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2329 execute->actions_len);
2330 if (pmd->core_id == NON_PMD_CORE_ID) {
2331 dp_netdev_pmd_unref(pmd);
2332 ovs_mutex_unlock(&dp->port_mutex);
2333 ovs_mutex_unlock(&dp->non_pmd_mutex);
2340 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2344 for (i = 0; i < n_ops; i++) {
2345 struct dpif_op *op = ops[i];
2348 case DPIF_OP_FLOW_PUT:
2349 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2352 case DPIF_OP_FLOW_DEL:
2353 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2356 case DPIF_OP_EXECUTE:
2357 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2360 case DPIF_OP_FLOW_GET:
2361 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2367 /* Returns true if the configuration for rx queues or cpu mask
2370 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2372 if (dp->n_dpdk_rxqs != rxqs) {
2375 if (dp->pmd_cmask != NULL && cmask != NULL) {
2376 return strcmp(dp->pmd_cmask, cmask);
2378 return (dp->pmd_cmask != NULL || cmask != NULL);
2383 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2385 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2387 struct dp_netdev *dp = get_dp_netdev(dpif);
2389 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2390 struct dp_netdev_port *port;
2392 dp_netdev_destroy_all_pmds(dp);
2394 CMAP_FOR_EACH (port, node, &dp->ports) {
2395 if (netdev_is_pmd(port->netdev)) {
2398 /* Closes the existing 'rxq's. */
2399 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2400 netdev_rxq_close(port->rxq[i]);
2401 port->rxq[i] = NULL;
2404 /* Sets the new rx queue config. */
2405 err = netdev_set_multiq(port->netdev, ovs_numa_get_n_cores(),
2407 if (err && (err != EOPNOTSUPP)) {
2408 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2409 " %u", netdev_get_name(port->netdev),
2414 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2415 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2416 * netdev_n_rxq(port->netdev));
2417 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2418 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2422 dp->n_dpdk_rxqs = n_rxqs;
2424 /* Reconfigures the cpu mask. */
2425 ovs_numa_set_cpu_mask(cmask);
2426 free(dp->pmd_cmask);
2427 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2429 /* Restores the non-pmd. */
2430 dp_netdev_set_nonpmd(dp);
2431 /* Restores all pmd threads. */
2432 dp_netdev_reset_pmd_threads(dp);
2439 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2440 uint32_t queue_id, uint32_t *priority)
2442 *priority = queue_id;
2447 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2448 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2449 struct dp_netdev_actions *
2450 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2452 struct dp_netdev_actions *netdev_actions;
2454 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2455 memcpy(netdev_actions->actions, actions, size);
2456 netdev_actions->size = size;
2458 return netdev_actions;
2461 struct dp_netdev_actions *
2462 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2464 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2468 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2473 static inline unsigned long long
2474 cycles_counter(void)
2477 return rte_get_tsc_cycles();
2483 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2484 extern struct ovs_mutex cycles_counter_fake_mutex;
2486 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2488 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2489 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2490 OVS_NO_THREAD_SAFETY_ANALYSIS
2492 pmd->last_cycles = cycles_counter();
2495 /* Stop counting cycles and add them to the counter 'type' */
2497 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2498 enum pmd_cycles_counter_type type)
2499 OVS_RELEASES(&cycles_counter_fake_mutex)
2500 OVS_NO_THREAD_SAFETY_ANALYSIS
2502 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2504 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2508 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2509 struct dp_netdev_port *port,
2510 struct netdev_rxq *rxq)
2512 struct dp_packet *packets[NETDEV_MAX_BURST];
2515 cycles_count_start(pmd);
2516 error = netdev_rxq_recv(rxq, packets, &cnt);
2517 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2521 *recirc_depth_get() = 0;
2523 /* XXX: initialize md in netdev implementation. */
2524 for (i = 0; i < cnt; i++) {
2525 packets[i]->md = port->md;
2527 cycles_count_start(pmd);
2528 dp_netdev_input(pmd, packets, cnt);
2529 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2530 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2531 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2533 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2534 netdev_get_name(port->netdev), ovs_strerror(error));
2538 /* Return true if needs to revalidate datapath flows. */
2540 dpif_netdev_run(struct dpif *dpif)
2542 struct dp_netdev_port *port;
2543 struct dp_netdev *dp = get_dp_netdev(dpif);
2544 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2546 uint64_t new_tnl_seq;
2548 ovs_mutex_lock(&dp->non_pmd_mutex);
2549 CMAP_FOR_EACH (port, node, &dp->ports) {
2550 if (!netdev_is_pmd(port->netdev)) {
2553 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2554 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2558 ovs_mutex_unlock(&dp->non_pmd_mutex);
2559 dp_netdev_pmd_unref(non_pmd);
2561 tnl_arp_cache_run();
2562 new_tnl_seq = seq_read(tnl_conf_seq);
2564 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2565 dp->last_tnl_conf_seq = new_tnl_seq;
2572 dpif_netdev_wait(struct dpif *dpif)
2574 struct dp_netdev_port *port;
2575 struct dp_netdev *dp = get_dp_netdev(dpif);
2577 ovs_mutex_lock(&dp_netdev_mutex);
2578 CMAP_FOR_EACH (port, node, &dp->ports) {
2579 if (!netdev_is_pmd(port->netdev)) {
2582 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2583 netdev_rxq_wait(port->rxq[i]);
2587 ovs_mutex_unlock(&dp_netdev_mutex);
2588 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2592 struct dp_netdev_port *port;
2593 struct netdev_rxq *rx;
2597 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2598 struct rxq_poll **ppoll_list, int poll_cnt)
2600 struct rxq_poll *poll_list = *ppoll_list;
2601 struct dp_netdev_port *port;
2602 int n_pmds_on_numa, index, i;
2604 /* Simple scheduler for netdev rx polling. */
2605 for (i = 0; i < poll_cnt; i++) {
2606 port_unref(poll_list[i].port);
2610 n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
2613 CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
2614 /* Calls port_try_ref() to prevent the main thread
2615 * from deleting the port. */
2616 if (port_try_ref(port)) {
2617 if (netdev_is_pmd(port->netdev)
2618 && netdev_get_numa_id(port->netdev) == pmd->numa_id) {
2621 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2622 if ((index % n_pmds_on_numa) == pmd->index) {
2623 poll_list = xrealloc(poll_list,
2624 sizeof *poll_list * (poll_cnt + 1));
2627 poll_list[poll_cnt].port = port;
2628 poll_list[poll_cnt].rx = port->rxq[i];
2634 /* Unrefs the port_try_ref(). */
2639 *ppoll_list = poll_list;
2644 pmd_thread_main(void *f_)
2646 struct dp_netdev_pmd_thread *pmd = f_;
2647 unsigned int lc = 0;
2648 struct rxq_poll *poll_list;
2649 unsigned int port_seq = PMD_INITIAL_SEQ;
2656 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2657 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2658 pmd_thread_setaffinity_cpu(pmd->core_id);
2660 emc_cache_init(&pmd->flow_cache);
2661 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2663 /* Signal here to make sure the pmd finishes
2664 * reloading the updated configuration. */
2665 dp_netdev_pmd_reload_done(pmd);
2670 for (i = 0; i < poll_cnt; i++) {
2671 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2679 emc_cache_slow_sweep(&pmd->flow_cache);
2682 atomic_read_relaxed(&pmd->change_seq, &seq);
2683 if (seq != port_seq) {
2690 emc_cache_uninit(&pmd->flow_cache);
2692 if (!latch_is_set(&pmd->exit_latch)){
2696 for (i = 0; i < poll_cnt; i++) {
2697 port_unref(poll_list[i].port);
2700 dp_netdev_pmd_reload_done(pmd);
2707 dp_netdev_disable_upcall(struct dp_netdev *dp)
2708 OVS_ACQUIRES(dp->upcall_rwlock)
2710 fat_rwlock_wrlock(&dp->upcall_rwlock);
2714 dpif_netdev_disable_upcall(struct dpif *dpif)
2715 OVS_NO_THREAD_SAFETY_ANALYSIS
2717 struct dp_netdev *dp = get_dp_netdev(dpif);
2718 dp_netdev_disable_upcall(dp);
2722 dp_netdev_enable_upcall(struct dp_netdev *dp)
2723 OVS_RELEASES(dp->upcall_rwlock)
2725 fat_rwlock_unlock(&dp->upcall_rwlock);
2729 dpif_netdev_enable_upcall(struct dpif *dpif)
2730 OVS_NO_THREAD_SAFETY_ANALYSIS
2732 struct dp_netdev *dp = get_dp_netdev(dpif);
2733 dp_netdev_enable_upcall(dp);
2737 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2739 ovs_mutex_lock(&pmd->cond_mutex);
2740 xpthread_cond_signal(&pmd->cond);
2741 ovs_mutex_unlock(&pmd->cond_mutex);
2744 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2745 * the pointer if succeeds, otherwise, NULL.
2747 * Caller must unrefs the returned reference. */
2748 static struct dp_netdev_pmd_thread *
2749 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2751 struct dp_netdev_pmd_thread *pmd;
2752 const struct cmap_node *pnode;
2754 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2758 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2760 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2763 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2765 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2767 struct dp_netdev_pmd_thread *non_pmd;
2769 non_pmd = xzalloc(sizeof *non_pmd);
2770 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2774 /* Caller must have valid pointer to 'pmd'. */
2776 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2778 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2782 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2784 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2785 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2789 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2790 * fails, keeps checking for next node until reaching the end of cmap.
2792 * Caller must unrefs the returned reference. */
2793 static struct dp_netdev_pmd_thread *
2794 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2796 struct dp_netdev_pmd_thread *next;
2799 struct cmap_node *node;
2801 node = cmap_next_position(&dp->poll_threads, pos);
2802 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2804 } while (next && !dp_netdev_pmd_try_ref(next));
2809 /* Configures the 'pmd' based on the input argument. */
2811 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2812 int index, unsigned core_id, int numa_id)
2816 pmd->core_id = core_id;
2817 pmd->numa_id = numa_id;
2819 ovs_refcount_init(&pmd->ref_cnt);
2820 latch_init(&pmd->exit_latch);
2821 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2822 xpthread_cond_init(&pmd->cond, NULL);
2823 ovs_mutex_init(&pmd->cond_mutex);
2824 ovs_mutex_init(&pmd->flow_mutex);
2825 dpcls_init(&pmd->cls);
2826 cmap_init(&pmd->flow_table);
2827 /* init the 'flow_cache' since there is no
2828 * actual thread created for NON_PMD_CORE_ID. */
2829 if (core_id == NON_PMD_CORE_ID) {
2830 emc_cache_init(&pmd->flow_cache);
2832 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2833 hash_int(core_id, 0));
2837 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2839 dp_netdev_pmd_flow_flush(pmd);
2840 dpcls_destroy(&pmd->cls);
2841 cmap_destroy(&pmd->flow_table);
2842 ovs_mutex_destroy(&pmd->flow_mutex);
2843 latch_destroy(&pmd->exit_latch);
2844 xpthread_cond_destroy(&pmd->cond);
2845 ovs_mutex_destroy(&pmd->cond_mutex);
2849 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2850 * and unrefs the struct. */
2852 dp_netdev_del_pmd(struct dp_netdev_pmd_thread *pmd)
2854 /* Uninit the 'flow_cache' since there is
2855 * no actual thread uninit it for NON_PMD_CORE_ID. */
2856 if (pmd->core_id == NON_PMD_CORE_ID) {
2857 emc_cache_uninit(&pmd->flow_cache);
2859 latch_set(&pmd->exit_latch);
2860 dp_netdev_reload_pmd__(pmd);
2861 ovs_numa_unpin_core(pmd->core_id);
2862 xpthread_join(pmd->thread, NULL);
2864 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2865 dp_netdev_pmd_unref(pmd);
2868 /* Destroys all pmd threads. */
2870 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2872 struct dp_netdev_pmd_thread *pmd;
2874 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2875 dp_netdev_del_pmd(pmd);
2879 /* Deletes all pmd threads on numa node 'numa_id'. */
2881 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2883 struct dp_netdev_pmd_thread *pmd;
2885 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2886 if (pmd->numa_id == numa_id) {
2887 dp_netdev_del_pmd(pmd);
2892 /* Checks the numa node id of 'netdev' and starts pmd threads for
2895 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2899 if (!ovs_numa_numa_id_is_valid(numa_id)) {
2900 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2901 "invalid", numa_id);
2905 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
2907 /* If there are already pmd threads created for the numa node
2908 * in which 'netdev' is on, do nothing. Else, creates the
2909 * pmd threads for the numa node. */
2911 int can_have, n_unpinned, i;
2913 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
2915 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2916 "cores on numa node");
2920 /* If cpu mask is specified, uses all unpinned cores, otherwise
2921 * tries creating NR_PMD_THREADS pmd threads. */
2922 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
2923 for (i = 0; i < can_have; i++) {
2924 struct dp_netdev_pmd_thread *pmd = xzalloc(sizeof *pmd);
2925 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
2927 dp_netdev_configure_pmd(pmd, dp, i, core_id, numa_id);
2928 /* Each thread will distribute all devices rx-queues among
2930 pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd);
2932 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
2937 /* Called after pmd threads config change. Restarts pmd threads with
2938 * new configuration. */
2940 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
2942 struct dp_netdev_port *port;
2944 CMAP_FOR_EACH (port, node, &dp->ports) {
2945 if (netdev_is_pmd(port->netdev)) {
2946 int numa_id = netdev_get_numa_id(port->netdev);
2948 dp_netdev_set_pmds_on_numa(dp, numa_id);
2954 dpif_netdev_get_datapath_version(void)
2956 return xstrdup("<built-in>");
2960 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
2961 uint16_t tcp_flags, long long now)
2965 atomic_store_relaxed(&netdev_flow->stats.used, now);
2966 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
2967 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
2968 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
2970 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
2974 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
2975 enum dp_stat_type type, int cnt)
2977 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
2981 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
2982 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
2983 enum dpif_upcall_type type, const struct nlattr *userdata,
2984 struct ofpbuf *actions, struct ofpbuf *put_actions)
2986 struct dp_netdev *dp = pmd->dp;
2988 if (OVS_UNLIKELY(!dp->upcall_cb)) {
2992 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
2993 struct ds ds = DS_EMPTY_INITIALIZER;
2997 ofpbuf_init(&key, 0);
2998 odp_flow_key_from_flow(&key, flow, &wc->masks, flow->in_port.odp_port,
3000 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3001 dp_packet_size(packet_));
3003 odp_flow_key_format(key.data, key.size, &ds);
3005 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3006 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3008 ofpbuf_uninit(&key);
3014 return dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3015 actions, wc, put_actions, dp->upcall_aux);
3018 static inline uint32_t
3019 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3020 const struct miniflow *mf)
3022 uint32_t hash, recirc_depth;
3024 hash = dp_packet_get_rss_hash(packet);
3025 if (OVS_UNLIKELY(!hash)) {
3026 hash = miniflow_hash_5tuple(mf, 0);
3027 dp_packet_set_rss_hash(packet, hash);
3030 /* The RSS hash must account for the recirculation depth to avoid
3031 * collisions in the exact match cache */
3032 recirc_depth = *recirc_depth_get_unsafe();
3033 if (OVS_UNLIKELY(recirc_depth)) {
3034 hash = hash_finish(hash, recirc_depth);
3035 dp_packet_set_rss_hash(packet, hash);
3040 struct packet_batch {
3041 unsigned int packet_count;
3042 unsigned int byte_count;
3045 struct dp_netdev_flow *flow;
3047 struct dp_packet *packets[NETDEV_MAX_BURST];
3051 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
3052 const struct miniflow *mf)
3054 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3055 batch->packets[batch->packet_count++] = packet;
3056 batch->byte_count += dp_packet_size(packet);
3060 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
3062 flow->batch = batch;
3065 batch->packet_count = 0;
3066 batch->byte_count = 0;
3067 batch->tcp_flags = 0;
3071 packet_batch_execute(struct packet_batch *batch,
3072 struct dp_netdev_pmd_thread *pmd,
3075 struct dp_netdev_actions *actions;
3076 struct dp_netdev_flow *flow = batch->flow;
3078 dp_netdev_flow_used(flow, batch->packet_count, batch->byte_count,
3079 batch->tcp_flags, now);
3081 actions = dp_netdev_flow_get_actions(flow);
3083 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
3084 actions->actions, actions->size);
3088 dp_netdev_queue_batches(struct dp_packet *pkt,
3089 struct dp_netdev_flow *flow, const struct miniflow *mf,
3090 struct packet_batch *batches, size_t *n_batches)
3092 struct packet_batch *batch = flow->batch;
3094 if (OVS_LIKELY(batch)) {
3095 packet_batch_update(batch, pkt, mf);
3099 batch = &batches[(*n_batches)++];
3100 packet_batch_init(batch, flow);
3101 packet_batch_update(batch, pkt, mf);
3105 dp_packet_swap(struct dp_packet **a, struct dp_packet **b)
3107 struct dp_packet *tmp = *a;
3112 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3113 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', the
3114 * miniflow is copied into 'keys' and the packet pointer is moved at the
3115 * beginning of the 'packets' array.
3117 * The function returns the number of packets that needs to be processed in the
3118 * 'packets' array (they have been moved to the beginning of the vector).
3120 static inline size_t
3121 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
3122 size_t cnt, struct netdev_flow_key *keys,
3123 struct packet_batch batches[], size_t *n_batches)
3125 struct emc_cache *flow_cache = &pmd->flow_cache;
3126 struct netdev_flow_key key;
3127 size_t i, notfound_cnt = 0;
3129 miniflow_initialize(&key.mf, key.buf);
3130 for (i = 0; i < cnt; i++) {
3131 struct dp_netdev_flow *flow;
3133 if (OVS_UNLIKELY(dp_packet_size(packets[i]) < ETH_HEADER_LEN)) {
3134 dp_packet_delete(packets[i]);
3138 miniflow_extract(packets[i], &key.mf);
3139 key.len = 0; /* Not computed yet. */
3140 key.hash = dpif_netdev_packet_get_rss_hash(packets[i], &key.mf);
3142 flow = emc_lookup(flow_cache, &key);
3143 if (OVS_LIKELY(flow)) {
3144 dp_netdev_queue_batches(packets[i], flow, &key.mf, batches,
3147 if (i != notfound_cnt) {
3148 dp_packet_swap(&packets[i], &packets[notfound_cnt]);
3151 keys[notfound_cnt++] = key;
3155 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - notfound_cnt);
3157 return notfound_cnt;
3161 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3162 struct dp_packet **packets, size_t cnt,
3163 struct netdev_flow_key *keys,
3164 struct packet_batch batches[], size_t *n_batches)
3166 #if !defined(__CHECKER__) && !defined(_WIN32)
3167 const size_t PKT_ARRAY_SIZE = cnt;
3169 /* Sparse or MSVC doesn't like variable length array. */
3170 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3172 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3173 struct dp_netdev *dp = pmd->dp;
3174 struct emc_cache *flow_cache = &pmd->flow_cache;
3175 int miss_cnt = 0, lost_cnt = 0;
3179 for (i = 0; i < cnt; i++) {
3180 /* Key length is needed in all the cases, hash computed on demand. */
3181 keys[i].len = netdev_flow_key_size(count_1bits(keys[i].mf.map));
3183 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3184 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3185 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3186 struct ofpbuf actions, put_actions;
3189 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3190 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3192 for (i = 0; i < cnt; i++) {
3193 struct dp_netdev_flow *netdev_flow;
3194 struct ofpbuf *add_actions;
3198 if (OVS_LIKELY(rules[i])) {
3202 /* It's possible that an earlier slow path execution installed
3203 * a rule covering this flow. In this case, it's a lot cheaper
3204 * to catch it here than execute a miss. */
3205 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3207 rules[i] = &netdev_flow->cr;
3213 miniflow_expand(&keys[i].mf, &match.flow);
3215 ofpbuf_clear(&actions);
3216 ofpbuf_clear(&put_actions);
3218 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
3219 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
3220 &ufid, DPIF_UC_MISS, NULL, &actions,
3222 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3223 dp_packet_delete(packets[i]);
3228 /* We can't allow the packet batching in the next loop to execute
3229 * the actions. Otherwise, if there are any slow path actions,
3230 * we'll send the packet up twice. */
3231 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
3232 actions.data, actions.size);
3234 add_actions = put_actions.size ? &put_actions : &actions;
3235 if (OVS_LIKELY(error != ENOSPC)) {
3236 /* XXX: There's a race window where a flow covering this packet
3237 * could have already been installed since we last did the flow
3238 * lookup before upcall. This could be solved by moving the
3239 * mutex lock outside the loop, but that's an awful long time
3240 * to be locking everyone out of making flow installs. If we
3241 * move to a per-core classifier, it would be reasonable. */
3242 ovs_mutex_lock(&pmd->flow_mutex);
3243 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3244 if (OVS_LIKELY(!netdev_flow)) {
3245 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3249 ovs_mutex_unlock(&pmd->flow_mutex);
3251 emc_insert(flow_cache, &keys[i], netdev_flow);
3255 ofpbuf_uninit(&actions);
3256 ofpbuf_uninit(&put_actions);
3257 fat_rwlock_unlock(&dp->upcall_rwlock);
3258 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3259 } else if (OVS_UNLIKELY(any_miss)) {
3260 for (i = 0; i < cnt; i++) {
3261 if (OVS_UNLIKELY(!rules[i])) {
3262 dp_packet_delete(packets[i]);
3269 for (i = 0; i < cnt; i++) {
3270 struct dp_packet *packet = packets[i];
3271 struct dp_netdev_flow *flow;
3273 if (OVS_UNLIKELY(!rules[i])) {
3277 flow = dp_netdev_flow_cast(rules[i]);
3279 emc_insert(flow_cache, &keys[i], flow);
3280 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3283 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3284 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3285 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3289 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3290 struct dp_packet **packets, int cnt)
3292 #if !defined(__CHECKER__) && !defined(_WIN32)
3293 const size_t PKT_ARRAY_SIZE = cnt;
3295 /* Sparse or MSVC doesn't like variable length array. */
3296 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3298 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3299 struct packet_batch batches[PKT_ARRAY_SIZE];
3300 long long now = time_msec();
3301 size_t newcnt, n_batches, i;
3304 newcnt = emc_processing(pmd, packets, cnt, keys, batches, &n_batches);
3305 if (OVS_UNLIKELY(newcnt)) {
3306 fast_path_processing(pmd, packets, newcnt, keys, batches, &n_batches);
3309 for (i = 0; i < n_batches; i++) {
3310 batches[i].flow->batch = NULL;
3313 for (i = 0; i < n_batches; i++) {
3314 packet_batch_execute(&batches[i], pmd, now);
3318 struct dp_netdev_execute_aux {
3319 struct dp_netdev_pmd_thread *pmd;
3323 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3326 struct dp_netdev *dp = get_dp_netdev(dpif);
3327 dp->upcall_aux = aux;
3332 dp_netdev_drop_packets(struct dp_packet ** packets, int cnt, bool may_steal)
3337 for (i = 0; i < cnt; i++) {
3338 dp_packet_delete(packets[i]);
3344 push_tnl_action(const struct dp_netdev *dp,
3345 const struct nlattr *attr,
3346 struct dp_packet **packets, int cnt)
3348 struct dp_netdev_port *tun_port;
3349 const struct ovs_action_push_tnl *data;
3351 data = nl_attr_get(attr);
3353 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3357 netdev_push_header(tun_port->netdev, packets, cnt, data);
3363 dp_netdev_clone_pkt_batch(struct dp_packet **dst_pkts,
3364 struct dp_packet **src_pkts, int cnt)
3368 for (i = 0; i < cnt; i++) {
3369 dst_pkts[i] = dp_packet_clone(src_pkts[i]);
3374 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3375 const struct nlattr *a, bool may_steal)
3376 OVS_NO_THREAD_SAFETY_ANALYSIS
3378 struct dp_netdev_execute_aux *aux = aux_;
3379 uint32_t *depth = recirc_depth_get();
3380 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3381 struct dp_netdev *dp = pmd->dp;
3382 int type = nl_attr_type(a);
3383 struct dp_netdev_port *p;
3386 switch ((enum ovs_action_attr)type) {
3387 case OVS_ACTION_ATTR_OUTPUT:
3388 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3389 if (OVS_LIKELY(p)) {
3390 netdev_send(p->netdev, pmd->core_id, packets, cnt, may_steal);
3395 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3396 if (*depth < MAX_RECIRC_DEPTH) {
3397 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3401 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3405 err = push_tnl_action(dp, a, packets, cnt);
3408 dp_netdev_input(pmd, packets, cnt);
3411 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3417 case OVS_ACTION_ATTR_TUNNEL_POP:
3418 if (*depth < MAX_RECIRC_DEPTH) {
3419 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3421 p = dp_netdev_lookup_port(dp, portno);
3423 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3427 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3431 err = netdev_pop_header(p->netdev, packets, cnt);
3434 for (i = 0; i < cnt; i++) {
3435 packets[i]->md.in_port.odp_port = portno;
3439 dp_netdev_input(pmd, packets, cnt);
3442 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3449 case OVS_ACTION_ATTR_USERSPACE:
3450 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3451 const struct nlattr *userdata;
3452 struct ofpbuf actions;
3456 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3457 ofpbuf_init(&actions, 0);
3459 for (i = 0; i < cnt; i++) {
3462 ofpbuf_clear(&actions);
3464 flow_extract(packets[i], &flow);
3465 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3466 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3467 DPIF_UC_ACTION, userdata,&actions,
3469 if (!error || error == ENOSPC) {
3470 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3471 actions.data, actions.size);
3472 } else if (may_steal) {
3473 dp_packet_delete(packets[i]);
3476 ofpbuf_uninit(&actions);
3477 fat_rwlock_unlock(&dp->upcall_rwlock);
3483 case OVS_ACTION_ATTR_RECIRC:
3484 if (*depth < MAX_RECIRC_DEPTH) {
3485 struct dp_packet *recirc_pkts[NETDEV_MAX_BURST];
3488 dp_netdev_clone_pkt_batch(recirc_pkts, packets, cnt);
3489 packets = recirc_pkts;
3492 for (i = 0; i < cnt; i++) {
3493 packets[i]->md.recirc_id = nl_attr_get_u32(a);
3497 dp_netdev_input(pmd, packets, cnt);
3503 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3506 case OVS_ACTION_ATTR_PUSH_VLAN:
3507 case OVS_ACTION_ATTR_POP_VLAN:
3508 case OVS_ACTION_ATTR_PUSH_MPLS:
3509 case OVS_ACTION_ATTR_POP_MPLS:
3510 case OVS_ACTION_ATTR_SET:
3511 case OVS_ACTION_ATTR_SET_MASKED:
3512 case OVS_ACTION_ATTR_SAMPLE:
3513 case OVS_ACTION_ATTR_HASH:
3514 case OVS_ACTION_ATTR_UNSPEC:
3515 case __OVS_ACTION_ATTR_MAX:
3519 dp_netdev_drop_packets(packets, cnt, may_steal);
3523 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3524 struct dp_packet **packets, int cnt,
3526 const struct nlattr *actions, size_t actions_len)
3528 struct dp_netdev_execute_aux aux = { pmd };
3530 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3531 actions_len, dp_execute_cb);
3534 const struct dpif_class dpif_netdev_class = {
3537 dpif_netdev_enumerate,
3538 dpif_netdev_port_open_type,
3541 dpif_netdev_destroy,
3544 dpif_netdev_get_stats,
3545 dpif_netdev_port_add,
3546 dpif_netdev_port_del,
3547 dpif_netdev_port_query_by_number,
3548 dpif_netdev_port_query_by_name,
3549 NULL, /* port_get_pid */
3550 dpif_netdev_port_dump_start,
3551 dpif_netdev_port_dump_next,
3552 dpif_netdev_port_dump_done,
3553 dpif_netdev_port_poll,
3554 dpif_netdev_port_poll_wait,
3555 dpif_netdev_flow_flush,
3556 dpif_netdev_flow_dump_create,
3557 dpif_netdev_flow_dump_destroy,
3558 dpif_netdev_flow_dump_thread_create,
3559 dpif_netdev_flow_dump_thread_destroy,
3560 dpif_netdev_flow_dump_next,
3561 dpif_netdev_operate,
3562 NULL, /* recv_set */
3563 NULL, /* handlers_set */
3564 dpif_netdev_pmd_set,
3565 dpif_netdev_queue_to_priority,
3567 NULL, /* recv_wait */
3568 NULL, /* recv_purge */
3569 dpif_netdev_register_upcall_cb,
3570 dpif_netdev_enable_upcall,
3571 dpif_netdev_disable_upcall,
3572 dpif_netdev_get_datapath_version,
3576 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3577 const char *argv[], void *aux OVS_UNUSED)
3579 struct dp_netdev_port *old_port;
3580 struct dp_netdev_port *new_port;
3581 struct dp_netdev *dp;
3584 ovs_mutex_lock(&dp_netdev_mutex);
3585 dp = shash_find_data(&dp_netdevs, argv[1]);
3586 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3587 ovs_mutex_unlock(&dp_netdev_mutex);
3588 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3591 ovs_refcount_ref(&dp->ref_cnt);
3592 ovs_mutex_unlock(&dp_netdev_mutex);
3594 ovs_mutex_lock(&dp->port_mutex);
3595 if (get_port_by_name(dp, argv[2], &old_port)) {
3596 unixctl_command_reply_error(conn, "unknown port");
3600 port_no = u32_to_odp(atoi(argv[3]));
3601 if (!port_no || port_no == ODPP_NONE) {
3602 unixctl_command_reply_error(conn, "bad port number");
3605 if (dp_netdev_lookup_port(dp, port_no)) {
3606 unixctl_command_reply_error(conn, "port number already in use");
3610 /* Remove old port. */
3611 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->md.in_port.odp_port));
3612 ovsrcu_postpone(free, old_port);
3614 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3615 new_port = xmemdup(old_port, sizeof *old_port);
3616 new_port->md.in_port.odp_port = port_no;
3617 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3619 seq_change(dp->port_seq);
3620 unixctl_command_reply(conn, NULL);
3623 ovs_mutex_unlock(&dp->port_mutex);
3624 dp_netdev_unref(dp);
3628 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3629 const char *argv[], void *aux OVS_UNUSED)
3631 struct dp_netdev_port *port;
3632 struct dp_netdev *dp;
3634 ovs_mutex_lock(&dp_netdev_mutex);
3635 dp = shash_find_data(&dp_netdevs, argv[1]);
3636 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3637 ovs_mutex_unlock(&dp_netdev_mutex);
3638 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3641 ovs_refcount_ref(&dp->ref_cnt);
3642 ovs_mutex_unlock(&dp_netdev_mutex);
3644 ovs_mutex_lock(&dp->port_mutex);
3645 if (get_port_by_name(dp, argv[2], &port)) {
3646 unixctl_command_reply_error(conn, "unknown port");
3647 } else if (port->md.in_port.odp_port == ODPP_LOCAL) {
3648 unixctl_command_reply_error(conn, "can't delete local port");
3650 do_del_port(dp, port);
3651 unixctl_command_reply(conn, NULL);
3653 ovs_mutex_unlock(&dp->port_mutex);
3655 dp_netdev_unref(dp);
3659 dpif_dummy_register__(const char *type)
3661 struct dpif_class *class;
3663 class = xmalloc(sizeof *class);
3664 *class = dpif_netdev_class;
3665 class->type = xstrdup(type);
3666 dp_register_provider(class);
3670 dpif_dummy_register(bool override)
3677 dp_enumerate_types(&types);
3678 SSET_FOR_EACH (type, &types) {
3679 if (!dp_unregister_provider(type)) {
3680 dpif_dummy_register__(type);
3683 sset_destroy(&types);
3686 dpif_dummy_register__("dummy");
3688 unixctl_command_register("dpif-dummy/change-port-number",
3689 "dp port new-number",
3690 3, 3, dpif_dummy_change_port_number, NULL);
3691 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3692 2, 2, dpif_dummy_delete_port, NULL);
3695 /* Datapath Classifier. */
3697 /* A set of rules that all have the same fields wildcarded. */
3698 struct dpcls_subtable {
3699 /* The fields are only used by writers. */
3700 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3702 /* These fields are accessed by readers. */
3703 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3704 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3705 /* 'mask' must be the last field, additional space is allocated here. */
3708 /* Initializes 'cls' as a classifier that initially contains no classification
3711 dpcls_init(struct dpcls *cls)
3713 cmap_init(&cls->subtables_map);
3714 pvector_init(&cls->subtables);
3718 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3720 pvector_remove(&cls->subtables, subtable);
3721 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3722 subtable->mask.hash);
3723 cmap_destroy(&subtable->rules);
3724 ovsrcu_postpone(free, subtable);
3727 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3728 * caller's responsibility.
3729 * May only be called after all the readers have been terminated. */
3731 dpcls_destroy(struct dpcls *cls)
3734 struct dpcls_subtable *subtable;
3736 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3737 dpcls_destroy_subtable(cls, subtable);
3739 cmap_destroy(&cls->subtables_map);
3740 pvector_destroy(&cls->subtables);
3744 static struct dpcls_subtable *
3745 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3747 struct dpcls_subtable *subtable;
3749 /* Need to add one. */
3750 subtable = xmalloc(sizeof *subtable
3751 - sizeof subtable->mask.mf + mask->len);
3752 cmap_init(&subtable->rules);
3753 netdev_flow_key_clone(&subtable->mask, mask);
3754 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3755 pvector_insert(&cls->subtables, subtable, 0);
3756 pvector_publish(&cls->subtables);
3761 static inline struct dpcls_subtable *
3762 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3764 struct dpcls_subtable *subtable;
3766 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3767 &cls->subtables_map) {
3768 if (netdev_flow_key_equal(&subtable->mask, mask)) {
3772 return dpcls_create_subtable(cls, mask);
3775 /* Insert 'rule' into 'cls'. */
3777 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
3778 const struct netdev_flow_key *mask)
3780 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
3782 rule->mask = &subtable->mask;
3783 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
3786 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3788 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
3790 struct dpcls_subtable *subtable;
3792 ovs_assert(rule->mask);
3794 INIT_CONTAINER(subtable, rule->mask, mask);
3796 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
3798 dpcls_destroy_subtable(cls, subtable);
3799 pvector_publish(&cls->subtables);
3803 /* Returns true if 'target' satisifies 'key' in 'mask', that is, if each 1-bit
3804 * in 'mask' the values in 'key' and 'target' are the same.
3806 * Note: 'key' and 'mask' have the same mask, and 'key' is already masked. */
3808 dpcls_rule_matches_key(const struct dpcls_rule *rule,
3809 const struct netdev_flow_key *target)
3811 const uint64_t *keyp = rule->flow.mf.inline_values;
3812 const uint64_t *maskp = rule->mask->mf.inline_values;
3813 uint64_t target_u64;
3815 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(target_u64, target, rule->flow.mf.map) {
3816 if (OVS_UNLIKELY((target_u64 & *maskp++) != *keyp++)) {
3823 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3824 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3825 * NULL it is skipped.
3827 * This function is optimized for use in the userspace datapath and therefore
3828 * does not implement a lot of features available in the standard
3829 * classifier_lookup() function. Specifically, it does not implement
3830 * priorities, instead returning any rule which matches the flow.
3832 * Returns true if all flows found a corresponding rule. */
3834 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
3835 struct dpcls_rule **rules, const size_t cnt)
3837 /* The batch size 16 was experimentally found faster than 8 or 32. */
3838 typedef uint16_t map_type;
3839 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3841 #if !defined(__CHECKER__) && !defined(_WIN32)
3842 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
3844 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
3846 map_type maps[N_MAPS];
3847 struct dpcls_subtable *subtable;
3849 memset(maps, 0xff, sizeof maps);
3850 if (cnt % MAP_BITS) {
3851 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
3853 memset(rules, 0, cnt * sizeof *rules);
3855 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
3856 const struct netdev_flow_key *mkeys = keys;
3857 struct dpcls_rule **mrules = rules;
3858 map_type remains = 0;
3861 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
3863 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
3864 uint32_t hashes[MAP_BITS];
3865 const struct cmap_node *nodes[MAP_BITS];
3866 unsigned long map = maps[m];
3870 continue; /* Skip empty maps. */
3873 /* Compute hashes for the remaining keys. */
3874 ULONG_FOR_EACH_1(i, map) {
3875 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
3879 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
3880 /* Check results. */
3881 ULONG_FOR_EACH_1(i, map) {
3882 struct dpcls_rule *rule;
3884 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
3885 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
3890 ULONG_SET0(map, i); /* Did not match. */
3892 ; /* Keep Sparse happy. */
3894 maps[m] &= ~map; /* Clear the found rules. */
3898 return true; /* All found. */
3901 return false; /* Some misses. */