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>
37 #include "dp-packet.h"
39 #include "dpif-provider.h"
41 #include "dynamic-string.h"
42 #include "fat-rwlock.h"
50 #include "netdev-dpdk.h"
51 #include "netdev-vport.h"
53 #include "odp-execute.h"
55 #include "ofp-print.h"
60 #include "poll-loop.h"
67 #include "tnl-arp-cache.h"
70 #include "openvswitch/vlog.h"
72 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
74 #define FLOW_DUMP_MAX_BATCH 50
75 /* Use per thread recirc_depth to prevent recirculation loop. */
76 #define MAX_RECIRC_DEPTH 5
77 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
79 /* Configuration parameters. */
80 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
82 /* Protects against changes to 'dp_netdevs'. */
83 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
85 /* Contains all 'struct dp_netdev's. */
86 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
87 = SHASH_INITIALIZER(&dp_netdevs);
89 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
91 static struct odp_support dp_netdev_support = {
92 .max_mpls_depth = SIZE_MAX,
96 /* Stores a miniflow with inline values */
98 struct netdev_flow_key {
99 uint32_t hash; /* Hash function differs for different users. */
100 uint32_t len; /* Length of the following miniflow (incl. map). */
102 uint64_t buf[FLOW_MAX_PACKET_U64S];
105 /* Exact match cache for frequently used flows
107 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
108 * search its entries for a miniflow that matches exactly the miniflow of the
109 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
111 * A cache entry holds a reference to its 'dp_netdev_flow'.
113 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
114 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
115 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
116 * value is the index of a cache entry where the miniflow could be.
122 * Each pmd_thread has its own private exact match cache.
123 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
126 #define EM_FLOW_HASH_SHIFT 13
127 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
128 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
129 #define EM_FLOW_HASH_SEGS 2
132 struct dp_netdev_flow *flow;
133 struct netdev_flow_key key; /* key.hash used for emc hash value. */
137 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
138 int sweep_idx; /* For emc_cache_slow_sweep(). */
141 /* Iterate in the exact match cache through every entry that might contain a
142 * miniflow with hash 'HASH'. */
143 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
144 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
145 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
146 i__ < EM_FLOW_HASH_SEGS; \
147 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
149 /* Simple non-wildcarding single-priority classifier. */
152 struct cmap subtables_map;
153 struct pvector subtables;
156 /* A rule to be inserted to the classifier. */
158 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
159 struct netdev_flow_key *mask; /* Subtable's mask. */
160 struct netdev_flow_key flow; /* Matching key. */
161 /* 'flow' must be the last field, additional space is allocated here. */
164 static void dpcls_init(struct dpcls *);
165 static void dpcls_destroy(struct dpcls *);
166 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
167 const struct netdev_flow_key *mask);
168 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
169 static bool dpcls_lookup(const struct dpcls *cls,
170 const struct netdev_flow_key keys[],
171 struct dpcls_rule **rules, size_t cnt);
173 /* Datapath based on the network device interface from netdev.h.
179 * Some members, marked 'const', are immutable. Accessing other members
180 * requires synchronization, as noted in more detail below.
182 * Acquisition order is, from outermost to innermost:
184 * dp_netdev_mutex (global)
188 const struct dpif_class *const class;
189 const char *const name;
191 struct ovs_refcount ref_cnt;
192 atomic_flag destroyed;
196 * Protected by RCU. Take the mutex to add or remove ports. */
197 struct ovs_mutex port_mutex;
199 struct seq *port_seq; /* Incremented whenever a port changes. */
201 /* Protects access to ofproto-dpif-upcall interface during revalidator
202 * thread synchronization. */
203 struct fat_rwlock upcall_rwlock;
204 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
207 /* Stores all 'struct dp_netdev_pmd_thread's. */
208 struct cmap poll_threads;
210 /* Protects the access of the 'struct dp_netdev_pmd_thread'
211 * instance for non-pmd thread. */
212 struct ovs_mutex non_pmd_mutex;
214 /* Each pmd thread will store its pointer to
215 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
216 ovsthread_key_t per_pmd_key;
218 /* Number of rx queues for each dpdk interface and the cpu mask
219 * for pin of pmd threads. */
222 uint64_t last_tnl_conf_seq;
225 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
229 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
230 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
231 DP_STAT_MISS, /* Packets that did not match. */
232 DP_STAT_LOST, /* Packets not passed up to the client. */
236 enum pmd_cycles_counter_type {
237 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
238 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
242 /* A port in a netdev-based datapath. */
243 struct dp_netdev_port {
245 struct netdev *netdev;
246 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
247 struct netdev_saved_flags *sf;
248 struct netdev_rxq **rxq;
249 struct ovs_refcount ref_cnt;
250 char *type; /* Port type as requested by user. */
253 /* Contained by struct dp_netdev_flow's 'stats' member. */
254 struct dp_netdev_flow_stats {
255 atomic_llong used; /* Last used time, in monotonic msecs. */
256 atomic_ullong packet_count; /* Number of packets matched. */
257 atomic_ullong byte_count; /* Number of bytes matched. */
258 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
261 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
267 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
268 * its pmd thread's classifier. The text below calls this classifier 'cls'.
273 * The thread safety rules described here for "struct dp_netdev_flow" are
274 * motivated by two goals:
276 * - Prevent threads that read members of "struct dp_netdev_flow" from
277 * reading bad data due to changes by some thread concurrently modifying
280 * - Prevent two threads making changes to members of a given "struct
281 * dp_netdev_flow" from interfering with each other.
287 * A flow 'flow' may be accessed without a risk of being freed during an RCU
288 * grace period. Code that needs to hold onto a flow for a while
289 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
291 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
292 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
295 * Some members, marked 'const', are immutable. Accessing other members
296 * requires synchronization, as noted in more detail below.
298 struct dp_netdev_flow {
299 const struct flow flow; /* Unmasked flow that created this entry. */
300 /* Hash table index by unmasked flow. */
301 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
303 const ovs_u128 ufid; /* Unique flow identifier. */
304 const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */
307 /* Number of references.
308 * The classifier owns one reference.
309 * Any thread trying to keep a rule from being freed should hold its own
311 struct ovs_refcount ref_cnt;
316 struct dp_netdev_flow_stats stats;
319 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
321 /* While processing a group of input packets, the datapath uses the next
322 * member to store a pointer to the output batch for the flow. It is
323 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
324 * packet_batch_init() and packet_batch_execute()). */
325 struct packet_batch *batch;
327 /* Packet classification. */
328 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
329 /* 'cr' must be the last member. */
332 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
333 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
334 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
337 /* A set of datapath actions within a "struct dp_netdev_flow".
343 * A struct dp_netdev_actions 'actions' is protected with RCU. */
344 struct dp_netdev_actions {
345 /* These members are immutable: they do not change during the struct's
347 unsigned int size; /* Size of 'actions', in bytes. */
348 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
351 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
353 struct dp_netdev_actions *dp_netdev_flow_get_actions(
354 const struct dp_netdev_flow *);
355 static void dp_netdev_actions_free(struct dp_netdev_actions *);
357 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
358 struct dp_netdev_pmd_stats {
359 /* Indexed by DP_STAT_*. */
360 atomic_ullong n[DP_N_STATS];
363 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
364 struct dp_netdev_pmd_cycles {
365 /* Indexed by PMD_CYCLES_*. */
366 atomic_ullong n[PMD_N_CYCLES];
369 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
370 * the performance overhead of interrupt processing. Therefore netdev can
371 * not implement rx-wait for these devices. dpif-netdev needs to poll
372 * these device to check for recv buffer. pmd-thread does polling for
373 * devices assigned to itself.
375 * DPDK used PMD for accessing NIC.
377 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
378 * I/O of all non-pmd threads. There will be no actual thread created
381 * Each struct has its own flow table and classifier. Packets received
382 * from managed ports are looked up in the corresponding pmd thread's
383 * flow table, and are executed with the found actions.
385 struct dp_netdev_pmd_thread {
386 struct dp_netdev *dp;
387 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
388 struct cmap_node node; /* In 'dp->poll_threads'. */
390 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
391 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
393 /* Per thread exact-match cache. Note, the instance for cpu core
394 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
395 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
396 * instances will only be accessed by its own pmd thread. */
397 struct emc_cache flow_cache;
399 /* Classifier and Flow-Table.
401 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
402 * changes to 'cls' must be made while still holding the 'flow_mutex'.
404 struct ovs_mutex flow_mutex;
406 struct cmap flow_table OVS_GUARDED; /* Flow table. */
409 struct dp_netdev_pmd_stats stats;
411 /* Cycles counters */
412 struct dp_netdev_pmd_cycles cycles;
414 /* Used to count cicles. See 'cycles_counter_end()' */
415 unsigned long long last_cycles;
417 struct latch exit_latch; /* For terminating the pmd thread. */
418 atomic_uint change_seq; /* For reloading pmd ports. */
420 int index; /* Idx of this pmd thread among pmd*/
421 /* threads on same numa node. */
422 unsigned core_id; /* CPU core id of this pmd thread. */
423 int numa_id; /* numa node id of this pmd thread. */
424 int tx_qid; /* Queue id used by this pmd thread to
425 * send packets on all netdevs */
427 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
428 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
429 * values and subtracts them from 'stats' and 'cycles' before
430 * reporting to the user */
431 unsigned long long stats_zero[DP_N_STATS];
432 uint64_t cycles_zero[PMD_N_CYCLES];
435 #define PMD_INITIAL_SEQ 1
437 /* Interface to netdev-based datapath. */
440 struct dp_netdev *dp;
441 uint64_t last_port_seq;
444 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
445 struct dp_netdev_port **portp);
446 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
447 struct dp_netdev_port **portp);
448 static void dp_netdev_free(struct dp_netdev *)
449 OVS_REQUIRES(dp_netdev_mutex);
450 static int do_add_port(struct dp_netdev *dp, const char *devname,
451 const char *type, odp_port_t port_no)
452 OVS_REQUIRES(dp->port_mutex);
453 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
454 OVS_REQUIRES(dp->port_mutex);
455 static int dpif_netdev_open(const struct dpif_class *, const char *name,
456 bool create, struct dpif **);
457 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
458 struct dp_packet **, int c,
460 const struct nlattr *actions,
462 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
463 struct dp_packet **, int cnt);
465 static void dp_netdev_disable_upcall(struct dp_netdev *);
466 void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
467 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
468 struct dp_netdev *dp, int index,
469 unsigned core_id, int numa_id);
470 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
471 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
472 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
474 static struct dp_netdev_pmd_thread *
475 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
476 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
477 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
478 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
479 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
480 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
481 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
482 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
484 static inline bool emc_entry_alive(struct emc_entry *ce);
485 static void emc_clear_entry(struct emc_entry *ce);
488 emc_cache_init(struct emc_cache *flow_cache)
492 BUILD_ASSERT(sizeof(struct miniflow) == 2 * sizeof(uint64_t));
494 flow_cache->sweep_idx = 0;
495 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
496 flow_cache->entries[i].flow = NULL;
497 flow_cache->entries[i].key.hash = 0;
498 flow_cache->entries[i].key.len = sizeof(struct miniflow);
499 flow_cache->entries[i].key.mf.tnl_map = 0;
500 flow_cache->entries[i].key.mf.pkt_map = 0;
505 emc_cache_uninit(struct emc_cache *flow_cache)
509 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
510 emc_clear_entry(&flow_cache->entries[i]);
514 /* Check and clear dead flow references slowly (one entry at each
517 emc_cache_slow_sweep(struct emc_cache *flow_cache)
519 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
521 if (!emc_entry_alive(entry)) {
522 emc_clear_entry(entry);
524 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
527 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
529 dpif_is_netdev(const struct dpif *dpif)
531 return dpif->dpif_class->open == dpif_netdev_open;
534 static struct dpif_netdev *
535 dpif_netdev_cast(const struct dpif *dpif)
537 ovs_assert(dpif_is_netdev(dpif));
538 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
541 static struct dp_netdev *
542 get_dp_netdev(const struct dpif *dpif)
544 return dpif_netdev_cast(dpif)->dp;
548 PMD_INFO_SHOW_STATS, /* show how cpu cycles are spent */
549 PMD_INFO_CLEAR_STATS /* set the cycles count to 0 */
553 pmd_info_show_stats(struct ds *reply,
554 struct dp_netdev_pmd_thread *pmd,
555 unsigned long long stats[DP_N_STATS],
556 uint64_t cycles[PMD_N_CYCLES])
558 unsigned long long total_packets = 0;
559 uint64_t total_cycles = 0;
562 /* These loops subtracts reference values ('*_zero') from the counters.
563 * Since loads and stores are relaxed, it might be possible for a '*_zero'
564 * value to be more recent than the current value we're reading from the
565 * counter. This is not a big problem, since these numbers are not
566 * supposed to be too accurate, but we should at least make sure that
567 * the result is not negative. */
568 for (i = 0; i < DP_N_STATS; i++) {
569 if (stats[i] > pmd->stats_zero[i]) {
570 stats[i] -= pmd->stats_zero[i];
575 if (i != DP_STAT_LOST) {
576 /* Lost packets are already included in DP_STAT_MISS */
577 total_packets += stats[i];
581 for (i = 0; i < PMD_N_CYCLES; i++) {
582 if (cycles[i] > pmd->cycles_zero[i]) {
583 cycles[i] -= pmd->cycles_zero[i];
588 total_cycles += cycles[i];
591 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
592 ? "main thread" : "pmd thread");
594 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
595 ds_put_format(reply, " numa_id %d", pmd->numa_id);
597 if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) {
598 ds_put_format(reply, " core_id %u", pmd->core_id);
600 ds_put_cstr(reply, ":\n");
603 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
604 "\tmiss:%llu\n\tlost:%llu\n",
605 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
606 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
608 if (total_cycles == 0) {
613 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
614 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
615 cycles[PMD_CYCLES_POLLING],
616 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
617 cycles[PMD_CYCLES_PROCESSING],
618 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
620 if (total_packets == 0) {
625 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
626 total_cycles / (double)total_packets,
627 total_cycles, total_packets);
630 "\tavg processing cycles per packet: "
631 "%.02f (%"PRIu64"/%llu)\n",
632 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
633 cycles[PMD_CYCLES_PROCESSING], total_packets);
637 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
638 struct dp_netdev_pmd_thread *pmd,
639 unsigned long long stats[DP_N_STATS],
640 uint64_t cycles[PMD_N_CYCLES])
644 /* We cannot write 'stats' and 'cycles' (because they're written by other
645 * threads) and we shouldn't change 'stats' (because they're used to count
646 * datapath stats, which must not be cleared here). Instead, we save the
647 * current values and subtract them from the values to be displayed in the
649 for (i = 0; i < DP_N_STATS; i++) {
650 pmd->stats_zero[i] = stats[i];
652 for (i = 0; i < PMD_N_CYCLES; i++) {
653 pmd->cycles_zero[i] = cycles[i];
658 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
661 struct ds reply = DS_EMPTY_INITIALIZER;
662 struct dp_netdev_pmd_thread *pmd;
663 struct dp_netdev *dp = NULL;
664 enum pmd_info_type type = *(enum pmd_info_type *) aux;
666 ovs_mutex_lock(&dp_netdev_mutex);
669 dp = shash_find_data(&dp_netdevs, argv[1]);
670 } else if (shash_count(&dp_netdevs) == 1) {
671 /* There's only one datapath */
672 dp = shash_first(&dp_netdevs)->data;
676 ovs_mutex_unlock(&dp_netdev_mutex);
677 unixctl_command_reply_error(conn,
678 "please specify an existing datapath");
682 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
683 unsigned long long stats[DP_N_STATS];
684 uint64_t cycles[PMD_N_CYCLES];
687 /* Read current stats and cycle counters */
688 for (i = 0; i < ARRAY_SIZE(stats); i++) {
689 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
691 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
692 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
695 if (type == PMD_INFO_CLEAR_STATS) {
696 pmd_info_clear_stats(&reply, pmd, stats, cycles);
697 } else if (type == PMD_INFO_SHOW_STATS) {
698 pmd_info_show_stats(&reply, pmd, stats, cycles);
702 ovs_mutex_unlock(&dp_netdev_mutex);
704 unixctl_command_reply(conn, ds_cstr(&reply));
709 dpif_netdev_init(void)
711 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
712 clear_aux = PMD_INFO_CLEAR_STATS;
714 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
715 0, 1, dpif_netdev_pmd_info,
717 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
718 0, 1, dpif_netdev_pmd_info,
724 dpif_netdev_enumerate(struct sset *all_dps,
725 const struct dpif_class *dpif_class)
727 struct shash_node *node;
729 ovs_mutex_lock(&dp_netdev_mutex);
730 SHASH_FOR_EACH(node, &dp_netdevs) {
731 struct dp_netdev *dp = node->data;
732 if (dpif_class != dp->class) {
733 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
734 * If the class doesn't match, skip this dpif. */
737 sset_add(all_dps, node->name);
739 ovs_mutex_unlock(&dp_netdev_mutex);
745 dpif_netdev_class_is_dummy(const struct dpif_class *class)
747 return class != &dpif_netdev_class;
751 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
753 return strcmp(type, "internal") ? type
754 : dpif_netdev_class_is_dummy(class) ? "dummy"
759 create_dpif_netdev(struct dp_netdev *dp)
761 uint16_t netflow_id = hash_string(dp->name, 0);
762 struct dpif_netdev *dpif;
764 ovs_refcount_ref(&dp->ref_cnt);
766 dpif = xmalloc(sizeof *dpif);
767 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
769 dpif->last_port_seq = seq_read(dp->port_seq);
774 /* Choose an unused, non-zero port number and return it on success.
775 * Return ODPP_NONE on failure. */
777 choose_port(struct dp_netdev *dp, const char *name)
778 OVS_REQUIRES(dp->port_mutex)
782 if (dp->class != &dpif_netdev_class) {
786 /* If the port name begins with "br", start the number search at
787 * 100 to make writing tests easier. */
788 if (!strncmp(name, "br", 2)) {
792 /* If the port name contains a number, try to assign that port number.
793 * This can make writing unit tests easier because port numbers are
795 for (p = name; *p != '\0'; p++) {
796 if (isdigit((unsigned char) *p)) {
797 port_no = start_no + strtol(p, NULL, 10);
798 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
799 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
800 return u32_to_odp(port_no);
807 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
808 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
809 return u32_to_odp(port_no);
817 create_dp_netdev(const char *name, const struct dpif_class *class,
818 struct dp_netdev **dpp)
819 OVS_REQUIRES(dp_netdev_mutex)
821 struct dp_netdev *dp;
824 dp = xzalloc(sizeof *dp);
825 shash_add(&dp_netdevs, name, dp);
827 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
828 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
829 ovs_refcount_init(&dp->ref_cnt);
830 atomic_flag_clear(&dp->destroyed);
832 ovs_mutex_init(&dp->port_mutex);
833 cmap_init(&dp->ports);
834 dp->port_seq = seq_create();
835 fat_rwlock_init(&dp->upcall_rwlock);
837 /* Disable upcalls by default. */
838 dp_netdev_disable_upcall(dp);
839 dp->upcall_aux = NULL;
840 dp->upcall_cb = NULL;
842 cmap_init(&dp->poll_threads);
843 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
844 ovsthread_key_create(&dp->per_pmd_key, NULL);
846 dp_netdev_set_nonpmd(dp);
847 dp->n_dpdk_rxqs = NR_QUEUE;
849 ovs_mutex_lock(&dp->port_mutex);
850 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
851 ovs_mutex_unlock(&dp->port_mutex);
857 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
863 dpif_netdev_open(const struct dpif_class *class, const char *name,
864 bool create, struct dpif **dpifp)
866 struct dp_netdev *dp;
869 ovs_mutex_lock(&dp_netdev_mutex);
870 dp = shash_find_data(&dp_netdevs, name);
872 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
874 error = (dp->class != class ? EINVAL
879 *dpifp = create_dpif_netdev(dp);
882 ovs_mutex_unlock(&dp_netdev_mutex);
888 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
889 OVS_NO_THREAD_SAFETY_ANALYSIS
891 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
892 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
894 /* Before freeing a lock we should release it */
895 fat_rwlock_unlock(&dp->upcall_rwlock);
896 fat_rwlock_destroy(&dp->upcall_rwlock);
899 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
900 * through the 'dp_netdevs' shash while freeing 'dp'. */
902 dp_netdev_free(struct dp_netdev *dp)
903 OVS_REQUIRES(dp_netdev_mutex)
905 struct dp_netdev_port *port;
907 shash_find_and_delete(&dp_netdevs, dp->name);
909 dp_netdev_destroy_all_pmds(dp);
910 cmap_destroy(&dp->poll_threads);
911 ovs_mutex_destroy(&dp->non_pmd_mutex);
912 ovsthread_key_delete(dp->per_pmd_key);
914 ovs_mutex_lock(&dp->port_mutex);
915 CMAP_FOR_EACH (port, node, &dp->ports) {
916 do_del_port(dp, port);
918 ovs_mutex_unlock(&dp->port_mutex);
920 seq_destroy(dp->port_seq);
921 cmap_destroy(&dp->ports);
923 /* Upcalls must be disabled at this point */
924 dp_netdev_destroy_upcall_lock(dp);
927 free(CONST_CAST(char *, dp->name));
932 dp_netdev_unref(struct dp_netdev *dp)
935 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
936 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
937 ovs_mutex_lock(&dp_netdev_mutex);
938 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
941 ovs_mutex_unlock(&dp_netdev_mutex);
946 dpif_netdev_close(struct dpif *dpif)
948 struct dp_netdev *dp = get_dp_netdev(dpif);
955 dpif_netdev_destroy(struct dpif *dpif)
957 struct dp_netdev *dp = get_dp_netdev(dpif);
959 if (!atomic_flag_test_and_set(&dp->destroyed)) {
960 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
961 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
969 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
970 * load/store semantics. While the increment is not atomic, the load and
971 * store operations are, making it impossible to read inconsistent values.
973 * This is used to update thread local stats counters. */
975 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
977 unsigned long long tmp;
979 atomic_read_relaxed(var, &tmp);
981 atomic_store_relaxed(var, tmp);
985 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
987 struct dp_netdev *dp = get_dp_netdev(dpif);
988 struct dp_netdev_pmd_thread *pmd;
990 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
991 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
992 unsigned long long n;
993 stats->n_flows += cmap_count(&pmd->flow_table);
995 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
997 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
999 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
1000 stats->n_missed += n;
1001 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
1004 stats->n_masks = UINT32_MAX;
1005 stats->n_mask_hit = UINT64_MAX;
1011 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1015 if (pmd->core_id == NON_PMD_CORE_ID) {
1019 ovs_mutex_lock(&pmd->cond_mutex);
1020 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1021 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1022 ovs_mutex_unlock(&pmd->cond_mutex);
1025 /* Causes all pmd threads to reload its tx/rx devices.
1026 * Must be called after adding/removing ports. */
1028 dp_netdev_reload_pmds(struct dp_netdev *dp)
1030 struct dp_netdev_pmd_thread *pmd;
1032 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1033 dp_netdev_reload_pmd__(pmd);
1038 hash_port_no(odp_port_t port_no)
1040 return hash_int(odp_to_u32(port_no), 0);
1044 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1046 OVS_REQUIRES(dp->port_mutex)
1048 struct netdev_saved_flags *sf;
1049 struct dp_netdev_port *port;
1050 struct netdev *netdev;
1051 enum netdev_flags flags;
1052 const char *open_type;
1056 /* Reject devices already in 'dp'. */
1057 if (!get_port_by_name(dp, devname, &port)) {
1061 /* Open and validate network device. */
1062 open_type = dpif_netdev_port_open_type(dp->class, type);
1063 error = netdev_open(devname, open_type, &netdev);
1067 /* XXX reject non-Ethernet devices */
1069 netdev_get_flags(netdev, &flags);
1070 if (flags & NETDEV_LOOPBACK) {
1071 VLOG_ERR("%s: cannot add a loopback device", devname);
1072 netdev_close(netdev);
1076 if (netdev_is_pmd(netdev)) {
1077 int n_cores = ovs_numa_get_n_cores();
1079 if (n_cores == OVS_CORE_UNSPEC) {
1080 VLOG_ERR("%s, cannot get cpu core info", devname);
1083 /* There can only be ovs_numa_get_n_cores() pmd threads,
1084 * so creates a txq for each, and one extra for the non
1086 error = netdev_set_multiq(netdev, n_cores + 1, dp->n_dpdk_rxqs);
1087 if (error && (error != EOPNOTSUPP)) {
1088 VLOG_ERR("%s, cannot set multiq", devname);
1092 port = xzalloc(sizeof *port);
1093 port->port_no = port_no;
1094 port->netdev = netdev;
1095 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
1096 port->type = xstrdup(type);
1097 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1098 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1100 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
1101 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1102 devname, ovs_strerror(errno));
1103 netdev_close(netdev);
1111 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1113 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1114 netdev_rxq_close(port->rxq[i]);
1116 netdev_close(netdev);
1124 ovs_refcount_init(&port->ref_cnt);
1125 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1127 if (netdev_is_pmd(netdev)) {
1128 dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
1129 dp_netdev_reload_pmds(dp);
1131 seq_change(dp->port_seq);
1137 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1138 odp_port_t *port_nop)
1140 struct dp_netdev *dp = get_dp_netdev(dpif);
1141 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1142 const char *dpif_port;
1146 ovs_mutex_lock(&dp->port_mutex);
1147 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1148 if (*port_nop != ODPP_NONE) {
1149 port_no = *port_nop;
1150 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1152 port_no = choose_port(dp, dpif_port);
1153 error = port_no == ODPP_NONE ? EFBIG : 0;
1156 *port_nop = port_no;
1157 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1159 ovs_mutex_unlock(&dp->port_mutex);
1165 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1167 struct dp_netdev *dp = get_dp_netdev(dpif);
1170 ovs_mutex_lock(&dp->port_mutex);
1171 if (port_no == ODPP_LOCAL) {
1174 struct dp_netdev_port *port;
1176 error = get_port_by_number(dp, port_no, &port);
1178 do_del_port(dp, port);
1181 ovs_mutex_unlock(&dp->port_mutex);
1187 is_valid_port_number(odp_port_t port_no)
1189 return port_no != ODPP_NONE;
1192 static struct dp_netdev_port *
1193 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1195 struct dp_netdev_port *port;
1197 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1198 if (port->port_no == port_no) {
1206 get_port_by_number(struct dp_netdev *dp,
1207 odp_port_t port_no, struct dp_netdev_port **portp)
1209 if (!is_valid_port_number(port_no)) {
1213 *portp = dp_netdev_lookup_port(dp, port_no);
1214 return *portp ? 0 : ENOENT;
1219 port_ref(struct dp_netdev_port *port)
1222 ovs_refcount_ref(&port->ref_cnt);
1227 port_try_ref(struct dp_netdev_port *port)
1230 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
1237 port_unref(struct dp_netdev_port *port)
1239 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1240 int n_rxq = netdev_n_rxq(port->netdev);
1243 netdev_close(port->netdev);
1244 netdev_restore_flags(port->sf);
1246 for (i = 0; i < n_rxq; i++) {
1247 netdev_rxq_close(port->rxq[i]);
1256 get_port_by_name(struct dp_netdev *dp,
1257 const char *devname, struct dp_netdev_port **portp)
1258 OVS_REQUIRES(dp->port_mutex)
1260 struct dp_netdev_port *port;
1262 CMAP_FOR_EACH (port, node, &dp->ports) {
1263 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1272 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1274 struct dp_netdev_pmd_thread *pmd;
1277 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1278 if (pmd->numa_id == numa_id) {
1286 /* Returns 'true' if there is a port with pmd netdev and the netdev
1287 * is on numa node 'numa_id'. */
1289 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1291 struct dp_netdev_port *port;
1293 CMAP_FOR_EACH (port, node, &dp->ports) {
1294 if (netdev_is_pmd(port->netdev)
1295 && netdev_get_numa_id(port->netdev) == numa_id) {
1305 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1306 OVS_REQUIRES(dp->port_mutex)
1308 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1309 seq_change(dp->port_seq);
1310 if (netdev_is_pmd(port->netdev)) {
1311 int numa_id = netdev_get_numa_id(port->netdev);
1313 /* If there is no netdev on the numa node, deletes the pmd threads
1314 * for that numa. Else, just reloads the queues. */
1315 if (!has_pmd_port_for_numa(dp, numa_id)) {
1316 dp_netdev_del_pmds_on_numa(dp, numa_id);
1318 dp_netdev_reload_pmds(dp);
1325 answer_port_query(const struct dp_netdev_port *port,
1326 struct dpif_port *dpif_port)
1328 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1329 dpif_port->type = xstrdup(port->type);
1330 dpif_port->port_no = port->port_no;
1334 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1335 struct dpif_port *dpif_port)
1337 struct dp_netdev *dp = get_dp_netdev(dpif);
1338 struct dp_netdev_port *port;
1341 error = get_port_by_number(dp, port_no, &port);
1342 if (!error && dpif_port) {
1343 answer_port_query(port, dpif_port);
1350 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1351 struct dpif_port *dpif_port)
1353 struct dp_netdev *dp = get_dp_netdev(dpif);
1354 struct dp_netdev_port *port;
1357 ovs_mutex_lock(&dp->port_mutex);
1358 error = get_port_by_name(dp, devname, &port);
1359 if (!error && dpif_port) {
1360 answer_port_query(port, dpif_port);
1362 ovs_mutex_unlock(&dp->port_mutex);
1368 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1370 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1374 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1376 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1377 ovsrcu_postpone(dp_netdev_flow_free, flow);
1382 dp_netdev_flow_hash(const ovs_u128 *ufid)
1384 return ufid->u32[0];
1388 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1389 struct dp_netdev_flow *flow)
1390 OVS_REQUIRES(pmd->flow_mutex)
1392 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1394 dpcls_remove(&pmd->cls, &flow->cr);
1395 flow->cr.mask = NULL; /* Accessing rule's mask after this is not safe. */
1397 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1400 dp_netdev_flow_unref(flow);
1404 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1406 struct dp_netdev_flow *netdev_flow;
1408 ovs_mutex_lock(&pmd->flow_mutex);
1409 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1410 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1412 ovs_mutex_unlock(&pmd->flow_mutex);
1416 dpif_netdev_flow_flush(struct dpif *dpif)
1418 struct dp_netdev *dp = get_dp_netdev(dpif);
1419 struct dp_netdev_pmd_thread *pmd;
1421 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1422 dp_netdev_pmd_flow_flush(pmd);
1428 struct dp_netdev_port_state {
1429 struct cmap_position position;
1434 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1436 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1441 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1442 struct dpif_port *dpif_port)
1444 struct dp_netdev_port_state *state = state_;
1445 struct dp_netdev *dp = get_dp_netdev(dpif);
1446 struct cmap_node *node;
1449 node = cmap_next_position(&dp->ports, &state->position);
1451 struct dp_netdev_port *port;
1453 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1456 state->name = xstrdup(netdev_get_name(port->netdev));
1457 dpif_port->name = state->name;
1458 dpif_port->type = port->type;
1459 dpif_port->port_no = port->port_no;
1470 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1472 struct dp_netdev_port_state *state = state_;
1479 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1481 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1482 uint64_t new_port_seq;
1485 new_port_seq = seq_read(dpif->dp->port_seq);
1486 if (dpif->last_port_seq != new_port_seq) {
1487 dpif->last_port_seq = new_port_seq;
1497 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1499 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1501 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1504 static struct dp_netdev_flow *
1505 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1507 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1510 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1512 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1515 /* netdev_flow_key utilities.
1517 * netdev_flow_key is basically a miniflow. We use these functions
1518 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1519 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1521 * - Since we are dealing exclusively with miniflows created by
1522 * miniflow_extract(), if the map is different the miniflow is different.
1523 * Therefore we can be faster by comparing the map and the miniflow in a
1525 * - These functions can be inlined by the compiler.
1527 * The following assertions make sure that what we're doing with miniflow is
1530 BUILD_ASSERT_DECL(sizeof(struct miniflow) == 2 * sizeof(uint64_t));
1532 /* Given the number of bits set in miniflow's maps, returns the size of the
1533 * 'netdev_flow_key.mf' */
1534 static inline size_t
1535 netdev_flow_key_size(size_t flow_u64s)
1537 return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s);
1541 netdev_flow_key_equal(const struct netdev_flow_key *a,
1542 const struct netdev_flow_key *b)
1544 /* 'b->len' may be not set yet. */
1545 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1548 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1549 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1550 * generated by miniflow_extract. */
1552 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1553 const struct miniflow *mf)
1555 return !memcmp(&key->mf, mf, key->len);
1559 netdev_flow_key_clone(struct netdev_flow_key *dst,
1560 const struct netdev_flow_key *src)
1563 offsetof(struct netdev_flow_key, mf) + src->len);
1568 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1569 const struct flow *src)
1571 struct dp_packet packet;
1572 uint64_t buf_stub[512 / 8];
1574 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1575 pkt_metadata_from_flow(&packet.md, src);
1576 flow_compose(&packet, src);
1577 miniflow_extract(&packet, &dst->mf);
1578 dp_packet_uninit(&packet);
1580 dst->len = netdev_flow_key_size(miniflow_n_values(&dst->mf));
1581 dst->hash = 0; /* Not computed yet. */
1584 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1586 netdev_flow_mask_init(struct netdev_flow_key *mask,
1587 const struct match *match)
1589 const uint64_t *mask_u64 = (const uint64_t *) &match->wc.masks;
1590 uint64_t *dst = miniflow_values(&mask->mf);
1591 struct miniflow maps;
1596 /* Only check masks that make sense for the flow. */
1597 flow_wc_map(&match->flow, &maps);
1598 memset(&mask->mf, 0, sizeof mask->mf); /* Clear maps. */
1602 uint64_t rm1bit = rightmost_1bit(map);
1603 int i = raw_ctz(map);
1606 mask->mf.tnl_map |= rm1bit;
1607 *dst++ = mask_u64[i];
1608 hash = hash_add64(hash, mask_u64[i]);
1612 mask_u64 += FLOW_TNL_U64S;
1615 uint64_t rm1bit = rightmost_1bit(map);
1616 int i = raw_ctz(map);
1619 mask->mf.pkt_map |= rm1bit;
1620 *dst++ = mask_u64[i];
1621 hash = hash_add64(hash, mask_u64[i]);
1626 hash = hash_add64(hash, mask->mf.tnl_map);
1627 hash = hash_add64(hash, mask->mf.pkt_map);
1629 n = dst - miniflow_get_values(&mask->mf);
1631 mask->hash = hash_finish(hash, n * 8);
1632 mask->len = netdev_flow_key_size(n);
1635 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1637 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1638 const struct flow *flow,
1639 const struct netdev_flow_key *mask)
1641 uint64_t *dst_u64 = miniflow_values(&dst->mf);
1642 const uint64_t *mask_u64 = miniflow_get_values(&mask->mf);
1646 dst->len = mask->len;
1647 dst->mf = mask->mf; /* Copy maps. */
1649 FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf) {
1650 *dst_u64 = value & *mask_u64++;
1651 hash = hash_add64(hash, *dst_u64++);
1653 dst->hash = hash_finish(hash,
1654 (dst_u64 - miniflow_get_values(&dst->mf)) * 8);
1657 /* Iterate through netdev_flow_key TNL u64 values specified by 'MAPS'. */
1658 #define NETDEV_FLOW_KEY_FOR_EACH_IN_TNL_MAP(VALUE, KEY, MAPS) \
1659 MINIFLOW_FOR_EACH_IN_TNL_MAP(VALUE, &(KEY)->mf, MAPS)
1661 /* Iterate through netdev_flow_key PKT u64 values specified by 'MAPS'. */
1662 #define NETDEV_FLOW_KEY_FOR_EACH_IN_PKT_MAP(VALUE, KEY, MAPS) \
1663 MINIFLOW_FOR_EACH_IN_PKT_MAP(VALUE, &(KEY)->mf, MAPS)
1665 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1667 static inline uint32_t
1668 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1669 const struct netdev_flow_key *mask)
1671 const uint64_t *p = miniflow_get_values(&mask->mf);
1675 NETDEV_FLOW_KEY_FOR_EACH_IN_TNL_MAP(key_u64, key, mask->mf) {
1676 hash = hash_add64(hash, key_u64 & *p++);
1678 NETDEV_FLOW_KEY_FOR_EACH_IN_PKT_MAP(key_u64, key, mask->mf) {
1679 hash = hash_add64(hash, key_u64 & *p++);
1682 return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8);
1686 emc_entry_alive(struct emc_entry *ce)
1688 return ce->flow && !ce->flow->dead;
1692 emc_clear_entry(struct emc_entry *ce)
1695 dp_netdev_flow_unref(ce->flow);
1701 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1702 const struct netdev_flow_key *key)
1704 if (ce->flow != flow) {
1706 dp_netdev_flow_unref(ce->flow);
1709 if (dp_netdev_flow_ref(flow)) {
1716 netdev_flow_key_clone(&ce->key, key);
1721 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1722 struct dp_netdev_flow *flow)
1724 struct emc_entry *to_be_replaced = NULL;
1725 struct emc_entry *current_entry;
1727 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1728 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1729 /* We found the entry with the 'mf' miniflow */
1730 emc_change_entry(current_entry, flow, NULL);
1734 /* Replacement policy: put the flow in an empty (not alive) entry, or
1735 * in the first entry where it can be */
1737 || (emc_entry_alive(to_be_replaced)
1738 && !emc_entry_alive(current_entry))
1739 || current_entry->key.hash < to_be_replaced->key.hash) {
1740 to_be_replaced = current_entry;
1743 /* We didn't find the miniflow in the cache.
1744 * The 'to_be_replaced' entry is where the new flow will be stored */
1746 emc_change_entry(to_be_replaced, flow, key);
1749 static inline struct dp_netdev_flow *
1750 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1752 struct emc_entry *current_entry;
1754 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1755 if (current_entry->key.hash == key->hash
1756 && emc_entry_alive(current_entry)
1757 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1759 /* We found the entry with the 'key->mf' miniflow */
1760 return current_entry->flow;
1767 static struct dp_netdev_flow *
1768 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1769 const struct netdev_flow_key *key)
1771 struct dp_netdev_flow *netdev_flow;
1772 struct dpcls_rule *rule;
1774 dpcls_lookup(&pmd->cls, key, &rule, 1);
1775 netdev_flow = dp_netdev_flow_cast(rule);
1780 static struct dp_netdev_flow *
1781 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1782 const ovs_u128 *ufidp, const struct nlattr *key,
1785 struct dp_netdev_flow *netdev_flow;
1789 /* If a UFID is not provided, determine one based on the key. */
1790 if (!ufidp && key && key_len
1791 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1792 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1797 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1799 if (ovs_u128_equals(&netdev_flow->ufid, ufidp)) {
1809 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1810 struct dpif_flow_stats *stats)
1812 struct dp_netdev_flow *netdev_flow;
1813 unsigned long long n;
1817 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1819 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1820 stats->n_packets = n;
1821 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1823 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1825 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1826 stats->tcp_flags = flags;
1829 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1830 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1831 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1834 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1835 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1836 struct dpif_flow *flow, bool terse)
1839 memset(flow, 0, sizeof *flow);
1841 struct flow_wildcards wc;
1842 struct dp_netdev_actions *actions;
1844 struct odp_flow_key_parms odp_parms = {
1845 .flow = &netdev_flow->flow,
1847 .support = dp_netdev_support,
1850 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1853 offset = key_buf->size;
1854 flow->key = ofpbuf_tail(key_buf);
1855 odp_parms.odp_in_port = netdev_flow->flow.in_port.odp_port;
1856 odp_flow_key_from_flow(&odp_parms, key_buf);
1857 flow->key_len = key_buf->size - offset;
1860 offset = mask_buf->size;
1861 flow->mask = ofpbuf_tail(mask_buf);
1862 odp_parms.odp_in_port = wc.masks.in_port.odp_port;
1863 odp_parms.key_buf = key_buf;
1864 odp_flow_key_from_mask(&odp_parms, mask_buf);
1865 flow->mask_len = mask_buf->size - offset;
1868 actions = dp_netdev_flow_get_actions(netdev_flow);
1869 flow->actions = actions->actions;
1870 flow->actions_len = actions->size;
1873 flow->ufid = netdev_flow->ufid;
1874 flow->ufid_present = true;
1875 flow->pmd_id = netdev_flow->pmd_id;
1876 get_dpif_flow_stats(netdev_flow, &flow->stats);
1880 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1881 const struct nlattr *mask_key,
1882 uint32_t mask_key_len, const struct flow *flow,
1883 struct flow_wildcards *wc)
1886 enum odp_key_fitness fitness;
1888 fitness = odp_flow_key_to_mask_udpif(mask_key, mask_key_len, key,
1889 key_len, &wc->masks, flow);
1891 /* This should not happen: it indicates that
1892 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1893 * disagree on the acceptable form of a mask. Log the problem
1894 * as an error, with enough details to enable debugging. */
1895 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1897 if (!VLOG_DROP_ERR(&rl)) {
1901 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1903 VLOG_ERR("internal error parsing flow mask %s (%s)",
1904 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1911 flow_wildcards_init_for_packet(wc, flow);
1918 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1923 if (odp_flow_key_to_flow_udpif(key, key_len, flow)) {
1924 /* This should not happen: it indicates that odp_flow_key_from_flow()
1925 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1926 * flow. Log the problem as an error, with enough details to enable
1928 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1930 if (!VLOG_DROP_ERR(&rl)) {
1934 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1935 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1942 in_port = flow->in_port.odp_port;
1943 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1951 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1953 struct dp_netdev *dp = get_dp_netdev(dpif);
1954 struct dp_netdev_flow *netdev_flow;
1955 struct dp_netdev_pmd_thread *pmd;
1956 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1957 ? NON_PMD_CORE_ID : get->pmd_id;
1960 pmd = dp_netdev_get_pmd(dp, pmd_id);
1965 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1968 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1973 dp_netdev_pmd_unref(pmd);
1979 static struct dp_netdev_flow *
1980 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
1981 struct match *match, const ovs_u128 *ufid,
1982 const struct nlattr *actions, size_t actions_len)
1983 OVS_REQUIRES(pmd->flow_mutex)
1985 struct dp_netdev_flow *flow;
1986 struct netdev_flow_key mask;
1988 netdev_flow_mask_init(&mask, match);
1989 /* Make sure wc does not have metadata. */
1990 ovs_assert(!(mask.mf.pkt_map
1991 & (MINIFLOW_PKT_MAP(metadata) | MINIFLOW_PKT_MAP(regs))));
1993 /* Do not allocate extra space. */
1994 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
1995 memset(&flow->stats, 0, sizeof flow->stats);
1998 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
1999 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
2000 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
2001 ovs_refcount_init(&flow->ref_cnt);
2002 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
2004 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
2005 dpcls_insert(&pmd->cls, &flow->cr, &mask);
2007 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
2008 dp_netdev_flow_hash(&flow->ufid));
2010 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2012 struct ds ds = DS_EMPTY_INITIALIZER;
2014 match.flow = flow->flow;
2015 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2017 ds_put_cstr(&ds, "flow_add: ");
2018 odp_format_ufid(ufid, &ds);
2019 ds_put_cstr(&ds, " ");
2020 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2021 ds_put_cstr(&ds, ", actions:");
2022 format_odp_actions(&ds, actions, actions_len);
2024 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2033 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2035 struct dp_netdev *dp = get_dp_netdev(dpif);
2036 struct dp_netdev_flow *netdev_flow;
2037 struct netdev_flow_key key;
2038 struct dp_netdev_pmd_thread *pmd;
2041 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2042 ? NON_PMD_CORE_ID : put->pmd_id;
2045 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2049 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2050 put->mask, put->mask_len,
2051 &match.flow, &match.wc);
2056 pmd = dp_netdev_get_pmd(dp, pmd_id);
2061 /* Must produce a netdev_flow_key for lookup.
2062 * This interface is no longer performance critical, since it is not used
2063 * for upcall processing any more. */
2064 netdev_flow_key_from_flow(&key, &match.flow);
2069 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2072 ovs_mutex_lock(&pmd->flow_mutex);
2073 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2075 if (put->flags & DPIF_FP_CREATE) {
2076 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2078 memset(put->stats, 0, sizeof *put->stats);
2080 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2090 if (put->flags & DPIF_FP_MODIFY
2091 && flow_equal(&match.flow, &netdev_flow->flow)) {
2092 struct dp_netdev_actions *new_actions;
2093 struct dp_netdev_actions *old_actions;
2095 new_actions = dp_netdev_actions_create(put->actions,
2098 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2099 ovsrcu_set(&netdev_flow->actions, new_actions);
2102 get_dpif_flow_stats(netdev_flow, put->stats);
2104 if (put->flags & DPIF_FP_ZERO_STATS) {
2105 /* XXX: The userspace datapath uses thread local statistics
2106 * (for flows), which should be updated only by the owning
2107 * thread. Since we cannot write on stats memory here,
2108 * we choose not to support this flag. Please note:
2109 * - This feature is currently used only by dpctl commands with
2111 * - Should the need arise, this operation can be implemented
2112 * by keeping a base value (to be update here) for each
2113 * counter, and subtracting it before outputting the stats */
2117 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2118 } else if (put->flags & DPIF_FP_CREATE) {
2121 /* Overlapping flow. */
2125 ovs_mutex_unlock(&pmd->flow_mutex);
2126 dp_netdev_pmd_unref(pmd);
2132 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2134 struct dp_netdev *dp = get_dp_netdev(dpif);
2135 struct dp_netdev_flow *netdev_flow;
2136 struct dp_netdev_pmd_thread *pmd;
2137 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2138 ? NON_PMD_CORE_ID : del->pmd_id;
2141 pmd = dp_netdev_get_pmd(dp, pmd_id);
2146 ovs_mutex_lock(&pmd->flow_mutex);
2147 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2151 get_dpif_flow_stats(netdev_flow, del->stats);
2153 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2157 ovs_mutex_unlock(&pmd->flow_mutex);
2158 dp_netdev_pmd_unref(pmd);
2163 struct dpif_netdev_flow_dump {
2164 struct dpif_flow_dump up;
2165 struct cmap_position poll_thread_pos;
2166 struct cmap_position flow_pos;
2167 struct dp_netdev_pmd_thread *cur_pmd;
2169 struct ovs_mutex mutex;
2172 static struct dpif_netdev_flow_dump *
2173 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2175 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2178 static struct dpif_flow_dump *
2179 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2181 struct dpif_netdev_flow_dump *dump;
2183 dump = xzalloc(sizeof *dump);
2184 dpif_flow_dump_init(&dump->up, dpif_);
2185 dump->up.terse = terse;
2186 ovs_mutex_init(&dump->mutex);
2192 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2194 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2196 ovs_mutex_destroy(&dump->mutex);
2201 struct dpif_netdev_flow_dump_thread {
2202 struct dpif_flow_dump_thread up;
2203 struct dpif_netdev_flow_dump *dump;
2204 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2205 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2208 static struct dpif_netdev_flow_dump_thread *
2209 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2211 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2214 static struct dpif_flow_dump_thread *
2215 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2217 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2218 struct dpif_netdev_flow_dump_thread *thread;
2220 thread = xmalloc(sizeof *thread);
2221 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2222 thread->dump = dump;
2227 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2229 struct dpif_netdev_flow_dump_thread *thread
2230 = dpif_netdev_flow_dump_thread_cast(thread_);
2236 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2237 struct dpif_flow *flows, int max_flows)
2239 struct dpif_netdev_flow_dump_thread *thread
2240 = dpif_netdev_flow_dump_thread_cast(thread_);
2241 struct dpif_netdev_flow_dump *dump = thread->dump;
2242 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2246 ovs_mutex_lock(&dump->mutex);
2247 if (!dump->status) {
2248 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2249 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2250 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2251 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2253 /* First call to dump_next(), extracts the first pmd thread.
2254 * If there is no pmd thread, returns immediately. */
2256 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2258 ovs_mutex_unlock(&dump->mutex);
2265 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2266 struct cmap_node *node;
2268 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2272 netdev_flows[n_flows] = CONTAINER_OF(node,
2273 struct dp_netdev_flow,
2276 /* When finishing dumping the current pmd thread, moves to
2278 if (n_flows < flow_limit) {
2279 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2280 dp_netdev_pmd_unref(pmd);
2281 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2287 /* Keeps the reference to next caller. */
2288 dump->cur_pmd = pmd;
2290 /* If the current dump is empty, do not exit the loop, since the
2291 * remaining pmds could have flows to be dumped. Just dumps again
2292 * on the new 'pmd'. */
2295 ovs_mutex_unlock(&dump->mutex);
2297 for (i = 0; i < n_flows; i++) {
2298 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2299 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2300 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2301 struct dpif_flow *f = &flows[i];
2302 struct ofpbuf key, mask;
2304 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2305 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2306 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2314 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2315 OVS_NO_THREAD_SAFETY_ANALYSIS
2317 struct dp_netdev *dp = get_dp_netdev(dpif);
2318 struct dp_netdev_pmd_thread *pmd;
2319 struct dp_packet *pp;
2321 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2322 dp_packet_size(execute->packet) > UINT16_MAX) {
2326 /* Tries finding the 'pmd'. If NULL is returned, that means
2327 * the current thread is a non-pmd thread and should use
2328 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2329 pmd = ovsthread_getspecific(dp->per_pmd_key);
2331 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2334 /* If the current thread is non-pmd thread, acquires
2335 * the 'non_pmd_mutex'. */
2336 if (pmd->core_id == NON_PMD_CORE_ID) {
2337 ovs_mutex_lock(&dp->non_pmd_mutex);
2338 ovs_mutex_lock(&dp->port_mutex);
2341 pp = execute->packet;
2342 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2343 execute->actions_len);
2344 if (pmd->core_id == NON_PMD_CORE_ID) {
2345 dp_netdev_pmd_unref(pmd);
2346 ovs_mutex_unlock(&dp->port_mutex);
2347 ovs_mutex_unlock(&dp->non_pmd_mutex);
2354 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2358 for (i = 0; i < n_ops; i++) {
2359 struct dpif_op *op = ops[i];
2362 case DPIF_OP_FLOW_PUT:
2363 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2366 case DPIF_OP_FLOW_DEL:
2367 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2370 case DPIF_OP_EXECUTE:
2371 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2374 case DPIF_OP_FLOW_GET:
2375 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2381 /* Returns true if the configuration for rx queues or cpu mask
2384 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2386 if (dp->n_dpdk_rxqs != rxqs) {
2389 if (dp->pmd_cmask != NULL && cmask != NULL) {
2390 return strcmp(dp->pmd_cmask, cmask);
2392 return (dp->pmd_cmask != NULL || cmask != NULL);
2397 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2399 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2401 struct dp_netdev *dp = get_dp_netdev(dpif);
2403 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2404 struct dp_netdev_port *port;
2406 dp_netdev_destroy_all_pmds(dp);
2408 CMAP_FOR_EACH (port, node, &dp->ports) {
2409 if (netdev_is_pmd(port->netdev)) {
2412 /* Closes the existing 'rxq's. */
2413 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2414 netdev_rxq_close(port->rxq[i]);
2415 port->rxq[i] = NULL;
2418 /* Sets the new rx queue config. */
2419 err = netdev_set_multiq(port->netdev,
2420 ovs_numa_get_n_cores() + 1,
2422 if (err && (err != EOPNOTSUPP)) {
2423 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2424 " %u", netdev_get_name(port->netdev),
2429 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2430 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2431 * netdev_n_rxq(port->netdev));
2432 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2433 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2437 dp->n_dpdk_rxqs = n_rxqs;
2439 /* Reconfigures the cpu mask. */
2440 ovs_numa_set_cpu_mask(cmask);
2441 free(dp->pmd_cmask);
2442 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2444 /* Restores the non-pmd. */
2445 dp_netdev_set_nonpmd(dp);
2446 /* Restores all pmd threads. */
2447 dp_netdev_reset_pmd_threads(dp);
2454 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2455 uint32_t queue_id, uint32_t *priority)
2457 *priority = queue_id;
2462 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2463 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2464 struct dp_netdev_actions *
2465 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2467 struct dp_netdev_actions *netdev_actions;
2469 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2470 memcpy(netdev_actions->actions, actions, size);
2471 netdev_actions->size = size;
2473 return netdev_actions;
2476 struct dp_netdev_actions *
2477 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2479 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2483 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2488 static inline unsigned long long
2489 cycles_counter(void)
2492 return rte_get_tsc_cycles();
2498 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2499 extern struct ovs_mutex cycles_counter_fake_mutex;
2501 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2503 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2504 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2505 OVS_NO_THREAD_SAFETY_ANALYSIS
2507 pmd->last_cycles = cycles_counter();
2510 /* Stop counting cycles and add them to the counter 'type' */
2512 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2513 enum pmd_cycles_counter_type type)
2514 OVS_RELEASES(&cycles_counter_fake_mutex)
2515 OVS_NO_THREAD_SAFETY_ANALYSIS
2517 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2519 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2523 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2524 struct dp_netdev_port *port,
2525 struct netdev_rxq *rxq)
2527 struct dp_packet *packets[NETDEV_MAX_BURST];
2530 cycles_count_start(pmd);
2531 error = netdev_rxq_recv(rxq, packets, &cnt);
2532 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2536 *recirc_depth_get() = 0;
2538 /* XXX: initialize md in netdev implementation. */
2539 for (i = 0; i < cnt; i++) {
2540 pkt_metadata_init(&packets[i]->md, port->port_no);
2542 cycles_count_start(pmd);
2543 dp_netdev_input(pmd, packets, cnt);
2544 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2545 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2546 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2548 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2549 netdev_get_name(port->netdev), ovs_strerror(error));
2553 /* Return true if needs to revalidate datapath flows. */
2555 dpif_netdev_run(struct dpif *dpif)
2557 struct dp_netdev_port *port;
2558 struct dp_netdev *dp = get_dp_netdev(dpif);
2559 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2561 uint64_t new_tnl_seq;
2563 ovs_mutex_lock(&dp->non_pmd_mutex);
2564 CMAP_FOR_EACH (port, node, &dp->ports) {
2565 if (!netdev_is_pmd(port->netdev)) {
2568 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2569 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2573 ovs_mutex_unlock(&dp->non_pmd_mutex);
2574 dp_netdev_pmd_unref(non_pmd);
2576 tnl_arp_cache_run();
2577 new_tnl_seq = seq_read(tnl_conf_seq);
2579 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2580 dp->last_tnl_conf_seq = new_tnl_seq;
2587 dpif_netdev_wait(struct dpif *dpif)
2589 struct dp_netdev_port *port;
2590 struct dp_netdev *dp = get_dp_netdev(dpif);
2592 ovs_mutex_lock(&dp_netdev_mutex);
2593 CMAP_FOR_EACH (port, node, &dp->ports) {
2594 if (!netdev_is_pmd(port->netdev)) {
2597 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2598 netdev_rxq_wait(port->rxq[i]);
2602 ovs_mutex_unlock(&dp_netdev_mutex);
2603 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2607 struct dp_netdev_port *port;
2608 struct netdev_rxq *rx;
2612 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2613 struct rxq_poll **ppoll_list, int poll_cnt)
2615 struct rxq_poll *poll_list = *ppoll_list;
2616 struct dp_netdev_port *port;
2617 int n_pmds_on_numa, index, i;
2619 /* Simple scheduler for netdev rx polling. */
2620 for (i = 0; i < poll_cnt; i++) {
2621 port_unref(poll_list[i].port);
2625 n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
2628 CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
2629 /* Calls port_try_ref() to prevent the main thread
2630 * from deleting the port. */
2631 if (port_try_ref(port)) {
2632 if (netdev_is_pmd(port->netdev)
2633 && netdev_get_numa_id(port->netdev) == pmd->numa_id) {
2636 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2637 if ((index % n_pmds_on_numa) == pmd->index) {
2638 poll_list = xrealloc(poll_list,
2639 sizeof *poll_list * (poll_cnt + 1));
2642 poll_list[poll_cnt].port = port;
2643 poll_list[poll_cnt].rx = port->rxq[i];
2649 /* Unrefs the port_try_ref(). */
2654 *ppoll_list = poll_list;
2659 pmd_thread_main(void *f_)
2661 struct dp_netdev_pmd_thread *pmd = f_;
2662 unsigned int lc = 0;
2663 struct rxq_poll *poll_list;
2664 unsigned int port_seq = PMD_INITIAL_SEQ;
2671 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2672 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2673 pmd_thread_setaffinity_cpu(pmd->core_id);
2675 emc_cache_init(&pmd->flow_cache);
2676 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2678 /* List port/core affinity */
2679 for (i = 0; i < poll_cnt; i++) {
2680 VLOG_INFO("Core %d processing port \'%s\'\n", pmd->core_id, netdev_get_name(poll_list[i].port->netdev));
2683 /* Signal here to make sure the pmd finishes
2684 * reloading the updated configuration. */
2685 dp_netdev_pmd_reload_done(pmd);
2690 for (i = 0; i < poll_cnt; i++) {
2691 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2699 emc_cache_slow_sweep(&pmd->flow_cache);
2700 coverage_try_clear();
2703 atomic_read_relaxed(&pmd->change_seq, &seq);
2704 if (seq != port_seq) {
2711 emc_cache_uninit(&pmd->flow_cache);
2713 if (!latch_is_set(&pmd->exit_latch)){
2717 for (i = 0; i < poll_cnt; i++) {
2718 port_unref(poll_list[i].port);
2721 dp_netdev_pmd_reload_done(pmd);
2728 dp_netdev_disable_upcall(struct dp_netdev *dp)
2729 OVS_ACQUIRES(dp->upcall_rwlock)
2731 fat_rwlock_wrlock(&dp->upcall_rwlock);
2735 dpif_netdev_disable_upcall(struct dpif *dpif)
2736 OVS_NO_THREAD_SAFETY_ANALYSIS
2738 struct dp_netdev *dp = get_dp_netdev(dpif);
2739 dp_netdev_disable_upcall(dp);
2743 dp_netdev_enable_upcall(struct dp_netdev *dp)
2744 OVS_RELEASES(dp->upcall_rwlock)
2746 fat_rwlock_unlock(&dp->upcall_rwlock);
2750 dpif_netdev_enable_upcall(struct dpif *dpif)
2751 OVS_NO_THREAD_SAFETY_ANALYSIS
2753 struct dp_netdev *dp = get_dp_netdev(dpif);
2754 dp_netdev_enable_upcall(dp);
2758 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2760 ovs_mutex_lock(&pmd->cond_mutex);
2761 xpthread_cond_signal(&pmd->cond);
2762 ovs_mutex_unlock(&pmd->cond_mutex);
2765 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2766 * the pointer if succeeds, otherwise, NULL.
2768 * Caller must unrefs the returned reference. */
2769 static struct dp_netdev_pmd_thread *
2770 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2772 struct dp_netdev_pmd_thread *pmd;
2773 const struct cmap_node *pnode;
2775 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2779 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2781 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2784 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2786 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2788 struct dp_netdev_pmd_thread *non_pmd;
2790 non_pmd = xzalloc(sizeof *non_pmd);
2791 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2795 /* Caller must have valid pointer to 'pmd'. */
2797 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2799 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2803 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2805 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2806 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2810 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2811 * fails, keeps checking for next node until reaching the end of cmap.
2813 * Caller must unrefs the returned reference. */
2814 static struct dp_netdev_pmd_thread *
2815 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2817 struct dp_netdev_pmd_thread *next;
2820 struct cmap_node *node;
2822 node = cmap_next_position(&dp->poll_threads, pos);
2823 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2825 } while (next && !dp_netdev_pmd_try_ref(next));
2831 core_id_to_qid(unsigned core_id)
2833 if (core_id != NON_PMD_CORE_ID) {
2836 return ovs_numa_get_n_cores();
2840 /* Configures the 'pmd' based on the input argument. */
2842 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2843 int index, unsigned core_id, int numa_id)
2847 pmd->core_id = core_id;
2848 pmd->tx_qid = core_id_to_qid(core_id);
2849 pmd->numa_id = numa_id;
2851 ovs_refcount_init(&pmd->ref_cnt);
2852 latch_init(&pmd->exit_latch);
2853 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2854 xpthread_cond_init(&pmd->cond, NULL);
2855 ovs_mutex_init(&pmd->cond_mutex);
2856 ovs_mutex_init(&pmd->flow_mutex);
2857 dpcls_init(&pmd->cls);
2858 cmap_init(&pmd->flow_table);
2859 /* init the 'flow_cache' since there is no
2860 * actual thread created for NON_PMD_CORE_ID. */
2861 if (core_id == NON_PMD_CORE_ID) {
2862 emc_cache_init(&pmd->flow_cache);
2864 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2865 hash_int(core_id, 0));
2869 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2871 dp_netdev_pmd_flow_flush(pmd);
2872 dpcls_destroy(&pmd->cls);
2873 cmap_destroy(&pmd->flow_table);
2874 ovs_mutex_destroy(&pmd->flow_mutex);
2875 latch_destroy(&pmd->exit_latch);
2876 xpthread_cond_destroy(&pmd->cond);
2877 ovs_mutex_destroy(&pmd->cond_mutex);
2881 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2882 * and unrefs the struct. */
2884 dp_netdev_del_pmd(struct dp_netdev_pmd_thread *pmd)
2886 /* Uninit the 'flow_cache' since there is
2887 * no actual thread uninit it for NON_PMD_CORE_ID. */
2888 if (pmd->core_id == NON_PMD_CORE_ID) {
2889 emc_cache_uninit(&pmd->flow_cache);
2891 latch_set(&pmd->exit_latch);
2892 dp_netdev_reload_pmd__(pmd);
2893 ovs_numa_unpin_core(pmd->core_id);
2894 xpthread_join(pmd->thread, NULL);
2896 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2897 dp_netdev_pmd_unref(pmd);
2900 /* Destroys all pmd threads. */
2902 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2904 struct dp_netdev_pmd_thread *pmd;
2906 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2907 dp_netdev_del_pmd(pmd);
2911 /* Deletes all pmd threads on numa node 'numa_id'. */
2913 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2915 struct dp_netdev_pmd_thread *pmd;
2917 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2918 if (pmd->numa_id == numa_id) {
2919 dp_netdev_del_pmd(pmd);
2924 /* Checks the numa node id of 'netdev' and starts pmd threads for
2927 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2931 if (!ovs_numa_numa_id_is_valid(numa_id)) {
2932 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2933 "invalid", numa_id);
2937 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
2939 /* If there are already pmd threads created for the numa node
2940 * in which 'netdev' is on, do nothing. Else, creates the
2941 * pmd threads for the numa node. */
2943 int can_have, n_unpinned, i;
2944 struct dp_netdev_pmd_thread **pmds;
2946 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
2948 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2949 "cores on numa node");
2953 /* If cpu mask is specified, uses all unpinned cores, otherwise
2954 * tries creating NR_PMD_THREADS pmd threads. */
2955 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
2956 pmds = xzalloc(can_have * sizeof *pmds);
2957 for (i = 0; i < can_have; i++) {
2958 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
2959 pmds[i] = xzalloc(sizeof **pmds);
2960 dp_netdev_configure_pmd(pmds[i], dp, i, core_id, numa_id);
2962 /* The pmd thread code needs to see all the others configured pmd
2963 * threads on the same numa node. That's why we call
2964 * 'dp_netdev_configure_pmd()' on all the threads and then we actually
2966 for (i = 0; i < can_have; i++) {
2967 /* Each thread will distribute all devices rx-queues among
2969 pmds[i]->thread = ovs_thread_create("pmd", pmd_thread_main, pmds[i]);
2972 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
2977 /* Called after pmd threads config change. Restarts pmd threads with
2978 * new configuration. */
2980 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
2982 struct dp_netdev_port *port;
2984 CMAP_FOR_EACH (port, node, &dp->ports) {
2985 if (netdev_is_pmd(port->netdev)) {
2986 int numa_id = netdev_get_numa_id(port->netdev);
2988 dp_netdev_set_pmds_on_numa(dp, numa_id);
2994 dpif_netdev_get_datapath_version(void)
2996 return xstrdup("<built-in>");
3000 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
3001 uint16_t tcp_flags, long long now)
3005 atomic_store_relaxed(&netdev_flow->stats.used, now);
3006 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
3007 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
3008 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
3010 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
3014 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
3015 enum dp_stat_type type, int cnt)
3017 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3021 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3022 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3023 enum dpif_upcall_type type, const struct nlattr *userdata,
3024 struct ofpbuf *actions, struct ofpbuf *put_actions)
3026 struct dp_netdev *dp = pmd->dp;
3027 struct flow_tnl orig_tunnel;
3030 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3034 /* Upcall processing expects the Geneve options to be in the translated
3035 * format but we need to retain the raw format for datapath use. */
3036 orig_tunnel.flags = flow->tunnel.flags;
3037 if (flow->tunnel.flags & FLOW_TNL_F_UDPIF) {
3038 orig_tunnel.metadata.present.len = flow->tunnel.metadata.present.len;
3039 memcpy(orig_tunnel.metadata.opts.gnv, flow->tunnel.metadata.opts.gnv,
3040 flow->tunnel.metadata.present.len);
3041 err = tun_metadata_from_geneve_udpif(&orig_tunnel, &orig_tunnel,
3048 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3049 struct ds ds = DS_EMPTY_INITIALIZER;
3052 struct odp_flow_key_parms odp_parms = {
3055 .odp_in_port = flow->in_port.odp_port,
3056 .support = dp_netdev_support,
3059 ofpbuf_init(&key, 0);
3060 odp_flow_key_from_flow(&odp_parms, &key);
3061 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3062 dp_packet_size(packet_));
3064 odp_flow_key_format(key.data, key.size, &ds);
3066 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3067 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3069 ofpbuf_uninit(&key);
3075 err = dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3076 actions, wc, put_actions, dp->upcall_aux);
3077 if (err && err != ENOSPC) {
3081 /* Translate tunnel metadata masks to datapath format. */
3083 if (wc->masks.tunnel.metadata.present.map) {
3084 struct geneve_opt opts[GENEVE_TOT_OPT_SIZE /
3085 sizeof(struct geneve_opt)];
3087 tun_metadata_to_geneve_udpif_mask(&flow->tunnel,
3089 orig_tunnel.metadata.opts.gnv,
3090 orig_tunnel.metadata.present.len,
3093 memset(&wc->masks.tunnel.metadata, 0,
3094 sizeof wc->masks.tunnel.metadata);
3095 memcpy(&wc->masks.tunnel.metadata.opts.gnv, opts,
3096 orig_tunnel.metadata.present.len);
3098 wc->masks.tunnel.metadata.present.len = 0xff;
3101 /* Restore tunnel metadata. We need to use the saved options to ensure
3102 * that any unknown options are not lost. The generated mask will have
3103 * the same structure, matching on types and lengths but wildcarding
3104 * option data we don't care about. */
3105 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3106 memcpy(&flow->tunnel.metadata.opts.gnv, orig_tunnel.metadata.opts.gnv,
3107 orig_tunnel.metadata.present.len);
3108 flow->tunnel.metadata.present.len = orig_tunnel.metadata.present.len;
3109 flow->tunnel.flags |= FLOW_TNL_F_UDPIF;
3115 static inline uint32_t
3116 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3117 const struct miniflow *mf)
3119 uint32_t hash, recirc_depth;
3121 hash = dp_packet_get_rss_hash(packet);
3122 if (OVS_UNLIKELY(!hash)) {
3123 hash = miniflow_hash_5tuple(mf, 0);
3124 dp_packet_set_rss_hash(packet, hash);
3127 /* The RSS hash must account for the recirculation depth to avoid
3128 * collisions in the exact match cache */
3129 recirc_depth = *recirc_depth_get_unsafe();
3130 if (OVS_UNLIKELY(recirc_depth)) {
3131 hash = hash_finish(hash, recirc_depth);
3132 dp_packet_set_rss_hash(packet, hash);
3137 struct packet_batch {
3138 unsigned int packet_count;
3139 unsigned int byte_count;
3142 struct dp_netdev_flow *flow;
3144 struct dp_packet *packets[NETDEV_MAX_BURST];
3148 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
3149 const struct miniflow *mf)
3151 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3152 batch->packets[batch->packet_count++] = packet;
3153 batch->byte_count += dp_packet_size(packet);
3157 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
3159 flow->batch = batch;
3162 batch->packet_count = 0;
3163 batch->byte_count = 0;
3164 batch->tcp_flags = 0;
3168 packet_batch_execute(struct packet_batch *batch,
3169 struct dp_netdev_pmd_thread *pmd,
3172 struct dp_netdev_actions *actions;
3173 struct dp_netdev_flow *flow = batch->flow;
3175 dp_netdev_flow_used(flow, batch->packet_count, batch->byte_count,
3176 batch->tcp_flags, now);
3178 actions = dp_netdev_flow_get_actions(flow);
3180 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
3181 actions->actions, actions->size);
3185 dp_netdev_queue_batches(struct dp_packet *pkt,
3186 struct dp_netdev_flow *flow, const struct miniflow *mf,
3187 struct packet_batch *batches, size_t *n_batches)
3189 struct packet_batch *batch = flow->batch;
3191 if (OVS_LIKELY(batch)) {
3192 packet_batch_update(batch, pkt, mf);
3196 batch = &batches[(*n_batches)++];
3197 packet_batch_init(batch, flow);
3198 packet_batch_update(batch, pkt, mf);
3202 dp_packet_swap(struct dp_packet **a, struct dp_packet **b)
3204 struct dp_packet *tmp = *a;
3209 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3210 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', the
3211 * miniflow is copied into 'keys' and the packet pointer is moved at the
3212 * beginning of the 'packets' array.
3214 * The function returns the number of packets that needs to be processed in the
3215 * 'packets' array (they have been moved to the beginning of the vector).
3217 static inline size_t
3218 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
3219 size_t cnt, struct netdev_flow_key *keys,
3220 struct packet_batch batches[], size_t *n_batches)
3222 struct emc_cache *flow_cache = &pmd->flow_cache;
3223 struct netdev_flow_key key;
3224 size_t i, notfound_cnt = 0;
3226 for (i = 0; i < cnt; i++) {
3227 struct dp_netdev_flow *flow;
3229 if (OVS_UNLIKELY(dp_packet_size(packets[i]) < ETH_HEADER_LEN)) {
3230 dp_packet_delete(packets[i]);
3235 /* Prefetch next packet data */
3236 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3239 miniflow_extract(packets[i], &key.mf);
3240 key.len = 0; /* Not computed yet. */
3241 key.hash = dpif_netdev_packet_get_rss_hash(packets[i], &key.mf);
3243 flow = emc_lookup(flow_cache, &key);
3244 if (OVS_LIKELY(flow)) {
3245 dp_netdev_queue_batches(packets[i], flow, &key.mf, batches,
3248 if (i != notfound_cnt) {
3249 dp_packet_swap(&packets[i], &packets[notfound_cnt]);
3252 keys[notfound_cnt++] = key;
3256 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - notfound_cnt);
3258 return notfound_cnt;
3262 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3263 struct dp_packet **packets, size_t cnt,
3264 struct netdev_flow_key *keys,
3265 struct packet_batch batches[], size_t *n_batches)
3267 #if !defined(__CHECKER__) && !defined(_WIN32)
3268 const size_t PKT_ARRAY_SIZE = cnt;
3270 /* Sparse or MSVC doesn't like variable length array. */
3271 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3273 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3274 struct dp_netdev *dp = pmd->dp;
3275 struct emc_cache *flow_cache = &pmd->flow_cache;
3276 int miss_cnt = 0, lost_cnt = 0;
3280 for (i = 0; i < cnt; i++) {
3281 /* Key length is needed in all the cases, hash computed on demand. */
3282 keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf));
3284 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3285 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3286 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3287 struct ofpbuf actions, put_actions;
3290 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3291 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3293 for (i = 0; i < cnt; i++) {
3294 struct dp_netdev_flow *netdev_flow;
3295 struct ofpbuf *add_actions;
3299 if (OVS_LIKELY(rules[i])) {
3303 /* It's possible that an earlier slow path execution installed
3304 * a rule covering this flow. In this case, it's a lot cheaper
3305 * to catch it here than execute a miss. */
3306 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3308 rules[i] = &netdev_flow->cr;
3314 miniflow_expand(&keys[i].mf, &match.flow);
3316 ofpbuf_clear(&actions);
3317 ofpbuf_clear(&put_actions);
3319 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
3320 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
3321 &ufid, DPIF_UC_MISS, NULL, &actions,
3323 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3324 dp_packet_delete(packets[i]);
3329 /* We can't allow the packet batching in the next loop to execute
3330 * the actions. Otherwise, if there are any slow path actions,
3331 * we'll send the packet up twice. */
3332 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
3333 actions.data, actions.size);
3335 add_actions = put_actions.size ? &put_actions : &actions;
3336 if (OVS_LIKELY(error != ENOSPC)) {
3337 /* XXX: There's a race window where a flow covering this packet
3338 * could have already been installed since we last did the flow
3339 * lookup before upcall. This could be solved by moving the
3340 * mutex lock outside the loop, but that's an awful long time
3341 * to be locking everyone out of making flow installs. If we
3342 * move to a per-core classifier, it would be reasonable. */
3343 ovs_mutex_lock(&pmd->flow_mutex);
3344 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3345 if (OVS_LIKELY(!netdev_flow)) {
3346 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3350 ovs_mutex_unlock(&pmd->flow_mutex);
3352 emc_insert(flow_cache, &keys[i], netdev_flow);
3356 ofpbuf_uninit(&actions);
3357 ofpbuf_uninit(&put_actions);
3358 fat_rwlock_unlock(&dp->upcall_rwlock);
3359 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3360 } else if (OVS_UNLIKELY(any_miss)) {
3361 for (i = 0; i < cnt; i++) {
3362 if (OVS_UNLIKELY(!rules[i])) {
3363 dp_packet_delete(packets[i]);
3370 for (i = 0; i < cnt; i++) {
3371 struct dp_packet *packet = packets[i];
3372 struct dp_netdev_flow *flow;
3374 if (OVS_UNLIKELY(!rules[i])) {
3378 flow = dp_netdev_flow_cast(rules[i]);
3380 emc_insert(flow_cache, &keys[i], flow);
3381 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3384 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3385 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3386 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3390 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3391 struct dp_packet **packets, int cnt)
3393 #if !defined(__CHECKER__) && !defined(_WIN32)
3394 const size_t PKT_ARRAY_SIZE = cnt;
3396 /* Sparse or MSVC doesn't like variable length array. */
3397 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3399 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3400 struct packet_batch batches[PKT_ARRAY_SIZE];
3401 long long now = time_msec();
3402 size_t newcnt, n_batches, i;
3405 newcnt = emc_processing(pmd, packets, cnt, keys, batches, &n_batches);
3406 if (OVS_UNLIKELY(newcnt)) {
3407 fast_path_processing(pmd, packets, newcnt, keys, batches, &n_batches);
3410 for (i = 0; i < n_batches; i++) {
3411 batches[i].flow->batch = NULL;
3414 for (i = 0; i < n_batches; i++) {
3415 packet_batch_execute(&batches[i], pmd, now);
3419 struct dp_netdev_execute_aux {
3420 struct dp_netdev_pmd_thread *pmd;
3424 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3427 struct dp_netdev *dp = get_dp_netdev(dpif);
3428 dp->upcall_aux = aux;
3433 dp_netdev_drop_packets(struct dp_packet **packets, int cnt, bool may_steal)
3438 for (i = 0; i < cnt; i++) {
3439 dp_packet_delete(packets[i]);
3445 push_tnl_action(const struct dp_netdev *dp,
3446 const struct nlattr *attr,
3447 struct dp_packet **packets, int cnt)
3449 struct dp_netdev_port *tun_port;
3450 const struct ovs_action_push_tnl *data;
3452 data = nl_attr_get(attr);
3454 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3458 netdev_push_header(tun_port->netdev, packets, cnt, data);
3464 dp_netdev_clone_pkt_batch(struct dp_packet **dst_pkts,
3465 struct dp_packet **src_pkts, int cnt)
3469 for (i = 0; i < cnt; i++) {
3470 dst_pkts[i] = dp_packet_clone(src_pkts[i]);
3475 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3476 const struct nlattr *a, bool may_steal)
3477 OVS_NO_THREAD_SAFETY_ANALYSIS
3479 struct dp_netdev_execute_aux *aux = aux_;
3480 uint32_t *depth = recirc_depth_get();
3481 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3482 struct dp_netdev *dp = pmd->dp;
3483 int type = nl_attr_type(a);
3484 struct dp_netdev_port *p;
3487 switch ((enum ovs_action_attr)type) {
3488 case OVS_ACTION_ATTR_OUTPUT:
3489 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3490 if (OVS_LIKELY(p)) {
3491 netdev_send(p->netdev, pmd->tx_qid, packets, cnt, may_steal);
3496 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3497 if (*depth < MAX_RECIRC_DEPTH) {
3498 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3502 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3506 err = push_tnl_action(dp, a, packets, cnt);
3509 dp_netdev_input(pmd, packets, cnt);
3512 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3518 case OVS_ACTION_ATTR_TUNNEL_POP:
3519 if (*depth < MAX_RECIRC_DEPTH) {
3520 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3522 p = dp_netdev_lookup_port(dp, portno);
3524 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3528 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3532 err = netdev_pop_header(p->netdev, packets, cnt);
3535 for (i = 0; i < cnt; i++) {
3536 packets[i]->md.in_port.odp_port = portno;
3540 dp_netdev_input(pmd, packets, cnt);
3543 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3550 case OVS_ACTION_ATTR_USERSPACE:
3551 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3552 const struct nlattr *userdata;
3553 struct ofpbuf actions;
3557 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3558 ofpbuf_init(&actions, 0);
3560 for (i = 0; i < cnt; i++) {
3563 ofpbuf_clear(&actions);
3565 flow_extract(packets[i], &flow);
3566 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3567 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3568 DPIF_UC_ACTION, userdata,&actions,
3570 if (!error || error == ENOSPC) {
3571 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3572 actions.data, actions.size);
3573 } else if (may_steal) {
3574 dp_packet_delete(packets[i]);
3577 ofpbuf_uninit(&actions);
3578 fat_rwlock_unlock(&dp->upcall_rwlock);
3584 case OVS_ACTION_ATTR_RECIRC:
3585 if (*depth < MAX_RECIRC_DEPTH) {
3586 struct dp_packet *recirc_pkts[NETDEV_MAX_BURST];
3589 dp_netdev_clone_pkt_batch(recirc_pkts, packets, cnt);
3590 packets = recirc_pkts;
3593 for (i = 0; i < cnt; i++) {
3594 packets[i]->md.recirc_id = nl_attr_get_u32(a);
3598 dp_netdev_input(pmd, packets, cnt);
3604 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3607 case OVS_ACTION_ATTR_PUSH_VLAN:
3608 case OVS_ACTION_ATTR_POP_VLAN:
3609 case OVS_ACTION_ATTR_PUSH_MPLS:
3610 case OVS_ACTION_ATTR_POP_MPLS:
3611 case OVS_ACTION_ATTR_SET:
3612 case OVS_ACTION_ATTR_SET_MASKED:
3613 case OVS_ACTION_ATTR_SAMPLE:
3614 case OVS_ACTION_ATTR_HASH:
3615 case OVS_ACTION_ATTR_UNSPEC:
3616 case __OVS_ACTION_ATTR_MAX:
3620 dp_netdev_drop_packets(packets, cnt, may_steal);
3624 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3625 struct dp_packet **packets, int cnt,
3627 const struct nlattr *actions, size_t actions_len)
3629 struct dp_netdev_execute_aux aux = { pmd };
3631 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3632 actions_len, dp_execute_cb);
3635 const struct dpif_class dpif_netdev_class = {
3638 dpif_netdev_enumerate,
3639 dpif_netdev_port_open_type,
3642 dpif_netdev_destroy,
3645 dpif_netdev_get_stats,
3646 dpif_netdev_port_add,
3647 dpif_netdev_port_del,
3648 dpif_netdev_port_query_by_number,
3649 dpif_netdev_port_query_by_name,
3650 NULL, /* port_get_pid */
3651 dpif_netdev_port_dump_start,
3652 dpif_netdev_port_dump_next,
3653 dpif_netdev_port_dump_done,
3654 dpif_netdev_port_poll,
3655 dpif_netdev_port_poll_wait,
3656 dpif_netdev_flow_flush,
3657 dpif_netdev_flow_dump_create,
3658 dpif_netdev_flow_dump_destroy,
3659 dpif_netdev_flow_dump_thread_create,
3660 dpif_netdev_flow_dump_thread_destroy,
3661 dpif_netdev_flow_dump_next,
3662 dpif_netdev_operate,
3663 NULL, /* recv_set */
3664 NULL, /* handlers_set */
3665 dpif_netdev_pmd_set,
3666 dpif_netdev_queue_to_priority,
3668 NULL, /* recv_wait */
3669 NULL, /* recv_purge */
3670 dpif_netdev_register_upcall_cb,
3671 dpif_netdev_enable_upcall,
3672 dpif_netdev_disable_upcall,
3673 dpif_netdev_get_datapath_version,
3677 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3678 const char *argv[], void *aux OVS_UNUSED)
3680 struct dp_netdev_port *old_port;
3681 struct dp_netdev_port *new_port;
3682 struct dp_netdev *dp;
3685 ovs_mutex_lock(&dp_netdev_mutex);
3686 dp = shash_find_data(&dp_netdevs, argv[1]);
3687 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3688 ovs_mutex_unlock(&dp_netdev_mutex);
3689 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3692 ovs_refcount_ref(&dp->ref_cnt);
3693 ovs_mutex_unlock(&dp_netdev_mutex);
3695 ovs_mutex_lock(&dp->port_mutex);
3696 if (get_port_by_name(dp, argv[2], &old_port)) {
3697 unixctl_command_reply_error(conn, "unknown port");
3701 port_no = u32_to_odp(atoi(argv[3]));
3702 if (!port_no || port_no == ODPP_NONE) {
3703 unixctl_command_reply_error(conn, "bad port number");
3706 if (dp_netdev_lookup_port(dp, port_no)) {
3707 unixctl_command_reply_error(conn, "port number already in use");
3711 /* Remove old port. */
3712 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3713 ovsrcu_postpone(free, old_port);
3715 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3716 new_port = xmemdup(old_port, sizeof *old_port);
3717 new_port->port_no = port_no;
3718 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3720 seq_change(dp->port_seq);
3721 unixctl_command_reply(conn, NULL);
3724 ovs_mutex_unlock(&dp->port_mutex);
3725 dp_netdev_unref(dp);
3729 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3730 const char *argv[], void *aux OVS_UNUSED)
3732 struct dp_netdev_port *port;
3733 struct dp_netdev *dp;
3735 ovs_mutex_lock(&dp_netdev_mutex);
3736 dp = shash_find_data(&dp_netdevs, argv[1]);
3737 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3738 ovs_mutex_unlock(&dp_netdev_mutex);
3739 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3742 ovs_refcount_ref(&dp->ref_cnt);
3743 ovs_mutex_unlock(&dp_netdev_mutex);
3745 ovs_mutex_lock(&dp->port_mutex);
3746 if (get_port_by_name(dp, argv[2], &port)) {
3747 unixctl_command_reply_error(conn, "unknown port");
3748 } else if (port->port_no == ODPP_LOCAL) {
3749 unixctl_command_reply_error(conn, "can't delete local port");
3751 do_del_port(dp, port);
3752 unixctl_command_reply(conn, NULL);
3754 ovs_mutex_unlock(&dp->port_mutex);
3756 dp_netdev_unref(dp);
3760 dpif_dummy_register__(const char *type)
3762 struct dpif_class *class;
3764 class = xmalloc(sizeof *class);
3765 *class = dpif_netdev_class;
3766 class->type = xstrdup(type);
3767 dp_register_provider(class);
3771 dpif_dummy_override(const char *type)
3773 if (!dp_unregister_provider(type)) {
3774 dpif_dummy_register__(type);
3779 dpif_dummy_register(enum dummy_level level)
3781 if (level == DUMMY_OVERRIDE_ALL) {
3786 dp_enumerate_types(&types);
3787 SSET_FOR_EACH (type, &types) {
3788 dpif_dummy_override(type);
3790 sset_destroy(&types);
3791 } else if (level == DUMMY_OVERRIDE_SYSTEM) {
3792 dpif_dummy_override("system");
3795 dpif_dummy_register__("dummy");
3797 unixctl_command_register("dpif-dummy/change-port-number",
3798 "dp port new-number",
3799 3, 3, dpif_dummy_change_port_number, NULL);
3800 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3801 2, 2, dpif_dummy_delete_port, NULL);
3804 /* Datapath Classifier. */
3806 /* A set of rules that all have the same fields wildcarded. */
3807 struct dpcls_subtable {
3808 /* The fields are only used by writers. */
3809 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3811 /* These fields are accessed by readers. */
3812 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3813 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3814 /* 'mask' must be the last field, additional space is allocated here. */
3817 /* Initializes 'cls' as a classifier that initially contains no classification
3820 dpcls_init(struct dpcls *cls)
3822 cmap_init(&cls->subtables_map);
3823 pvector_init(&cls->subtables);
3827 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3829 pvector_remove(&cls->subtables, subtable);
3830 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3831 subtable->mask.hash);
3832 cmap_destroy(&subtable->rules);
3833 ovsrcu_postpone(free, subtable);
3836 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3837 * caller's responsibility.
3838 * May only be called after all the readers have been terminated. */
3840 dpcls_destroy(struct dpcls *cls)
3843 struct dpcls_subtable *subtable;
3845 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3846 ovs_assert(cmap_count(&subtable->rules) == 0);
3847 dpcls_destroy_subtable(cls, subtable);
3849 cmap_destroy(&cls->subtables_map);
3850 pvector_destroy(&cls->subtables);
3854 static struct dpcls_subtable *
3855 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3857 struct dpcls_subtable *subtable;
3859 /* Need to add one. */
3860 subtable = xmalloc(sizeof *subtable
3861 - sizeof subtable->mask.mf + mask->len);
3862 cmap_init(&subtable->rules);
3863 netdev_flow_key_clone(&subtable->mask, mask);
3864 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3865 pvector_insert(&cls->subtables, subtable, 0);
3866 pvector_publish(&cls->subtables);
3871 static inline struct dpcls_subtable *
3872 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3874 struct dpcls_subtable *subtable;
3876 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3877 &cls->subtables_map) {
3878 if (netdev_flow_key_equal(&subtable->mask, mask)) {
3882 return dpcls_create_subtable(cls, mask);
3885 /* Insert 'rule' into 'cls'. */
3887 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
3888 const struct netdev_flow_key *mask)
3890 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
3892 rule->mask = &subtable->mask;
3893 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
3896 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3898 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
3900 struct dpcls_subtable *subtable;
3902 ovs_assert(rule->mask);
3904 INIT_CONTAINER(subtable, rule->mask, mask);
3906 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
3908 dpcls_destroy_subtable(cls, subtable);
3909 pvector_publish(&cls->subtables);
3913 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
3914 * in 'mask' the values in 'key' and 'target' are the same. */
3916 dpcls_rule_matches_key(const struct dpcls_rule *rule,
3917 const struct netdev_flow_key *target)
3919 const uint64_t *keyp = miniflow_get_values(&rule->flow.mf);
3920 const uint64_t *maskp = miniflow_get_values(&rule->mask->mf);
3921 uint64_t target_u64;
3923 NETDEV_FLOW_KEY_FOR_EACH_IN_TNL_MAP(target_u64, target, rule->flow.mf) {
3924 if (OVS_UNLIKELY((target_u64 & *maskp++) != *keyp++)) {
3928 NETDEV_FLOW_KEY_FOR_EACH_IN_PKT_MAP(target_u64, target, rule->flow.mf) {
3929 if (OVS_UNLIKELY((target_u64 & *maskp++) != *keyp++)) {
3936 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3937 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3938 * NULL it is skipped.
3940 * This function is optimized for use in the userspace datapath and therefore
3941 * does not implement a lot of features available in the standard
3942 * classifier_lookup() function. Specifically, it does not implement
3943 * priorities, instead returning any rule which matches the flow.
3945 * Returns true if all flows found a corresponding rule. */
3947 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
3948 struct dpcls_rule **rules, const size_t cnt)
3950 /* The batch size 16 was experimentally found faster than 8 or 32. */
3951 typedef uint16_t map_type;
3952 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3954 #if !defined(__CHECKER__) && !defined(_WIN32)
3955 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
3957 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
3959 map_type maps[N_MAPS];
3960 struct dpcls_subtable *subtable;
3962 memset(maps, 0xff, sizeof maps);
3963 if (cnt % MAP_BITS) {
3964 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
3966 memset(rules, 0, cnt * sizeof *rules);
3968 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
3969 const struct netdev_flow_key *mkeys = keys;
3970 struct dpcls_rule **mrules = rules;
3971 map_type remains = 0;
3974 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
3976 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
3977 uint32_t hashes[MAP_BITS];
3978 const struct cmap_node *nodes[MAP_BITS];
3979 unsigned long map = maps[m];
3983 continue; /* Skip empty maps. */
3986 /* Compute hashes for the remaining keys. */
3987 ULLONG_FOR_EACH_1(i, map) {
3988 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
3992 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
3993 /* Check results. */
3994 ULLONG_FOR_EACH_1(i, map) {
3995 struct dpcls_rule *rule;
3997 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
3998 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
4003 ULLONG_SET0(map, i); /* Did not match. */
4005 ; /* Keep Sparse happy. */
4007 maps[m] &= ~map; /* Clear the found rules. */
4011 return true; /* All found. */
4014 return false; /* Some misses. */