2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2016 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-neigh-cache.h"
68 #include "tnl-ports.h"
71 #include "openvswitch/vlog.h"
73 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
75 #define FLOW_DUMP_MAX_BATCH 50
76 /* Use per thread recirc_depth to prevent recirculation loop. */
77 #define MAX_RECIRC_DEPTH 5
78 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
80 /* Configuration parameters. */
81 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
83 /* Protects against changes to 'dp_netdevs'. */
84 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
86 /* Contains all 'struct dp_netdev's. */
87 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
88 = SHASH_INITIALIZER(&dp_netdevs);
90 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
92 static struct odp_support dp_netdev_support = {
93 .max_mpls_depth = SIZE_MAX,
97 /* Stores a miniflow with inline values */
99 struct netdev_flow_key {
100 uint32_t hash; /* Hash function differs for different users. */
101 uint32_t len; /* Length of the following miniflow (incl. map). */
103 uint64_t buf[FLOW_MAX_PACKET_U64S];
106 /* Exact match cache for frequently used flows
108 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
109 * search its entries for a miniflow that matches exactly the miniflow of the
110 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
112 * A cache entry holds a reference to its 'dp_netdev_flow'.
114 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
115 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
116 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
117 * value is the index of a cache entry where the miniflow could be.
123 * Each pmd_thread has its own private exact match cache.
124 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
127 #define EM_FLOW_HASH_SHIFT 13
128 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
129 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
130 #define EM_FLOW_HASH_SEGS 2
133 struct dp_netdev_flow *flow;
134 struct netdev_flow_key key; /* key.hash used for emc hash value. */
138 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
139 int sweep_idx; /* For emc_cache_slow_sweep(). */
142 /* Iterate in the exact match cache through every entry that might contain a
143 * miniflow with hash 'HASH'. */
144 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
145 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
146 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
147 i__ < EM_FLOW_HASH_SEGS; \
148 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
150 /* Simple non-wildcarding single-priority classifier. */
153 struct cmap subtables_map;
154 struct pvector subtables;
157 /* A rule to be inserted to the classifier. */
159 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
160 struct netdev_flow_key *mask; /* Subtable's mask. */
161 struct netdev_flow_key flow; /* Matching key. */
162 /* 'flow' must be the last field, additional space is allocated here. */
165 static void dpcls_init(struct dpcls *);
166 static void dpcls_destroy(struct dpcls *);
167 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
168 const struct netdev_flow_key *mask);
169 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
170 static bool dpcls_lookup(const struct dpcls *cls,
171 const struct netdev_flow_key keys[],
172 struct dpcls_rule **rules, size_t cnt);
174 /* Datapath based on the network device interface from netdev.h.
180 * Some members, marked 'const', are immutable. Accessing other members
181 * requires synchronization, as noted in more detail below.
183 * Acquisition order is, from outermost to innermost:
185 * dp_netdev_mutex (global)
189 const struct dpif_class *const class;
190 const char *const name;
192 struct ovs_refcount ref_cnt;
193 atomic_flag destroyed;
197 * Protected by RCU. Take the mutex to add or remove ports. */
198 struct ovs_mutex port_mutex;
200 struct seq *port_seq; /* Incremented whenever a port changes. */
202 /* Protects access to ofproto-dpif-upcall interface during revalidator
203 * thread synchronization. */
204 struct fat_rwlock upcall_rwlock;
205 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
208 /* Callback function for notifying the purging of dp flows (during
209 * reseting pmd deletion). */
210 dp_purge_callback *dp_purge_cb;
213 /* Stores all 'struct dp_netdev_pmd_thread's. */
214 struct cmap poll_threads;
216 /* Protects the access of the 'struct dp_netdev_pmd_thread'
217 * instance for non-pmd thread. */
218 struct ovs_mutex non_pmd_mutex;
220 /* Each pmd thread will store its pointer to
221 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
222 ovsthread_key_t per_pmd_key;
224 /* Cpu mask for pin of pmd threads. */
226 uint64_t last_tnl_conf_seq;
229 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
233 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
234 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
235 DP_STAT_MISS, /* Packets that did not match. */
236 DP_STAT_LOST, /* Packets not passed up to the client. */
240 enum pmd_cycles_counter_type {
241 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
242 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
246 /* A port in a netdev-based datapath. */
247 struct dp_netdev_port {
249 struct netdev *netdev;
250 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
251 struct netdev_saved_flags *sf;
252 struct netdev_rxq **rxq;
253 struct ovs_refcount ref_cnt;
254 char *type; /* Port type as requested by user. */
255 int latest_requested_n_rxq; /* Latest requested from netdev number
259 /* Contained by struct dp_netdev_flow's 'stats' member. */
260 struct dp_netdev_flow_stats {
261 atomic_llong used; /* Last used time, in monotonic msecs. */
262 atomic_ullong packet_count; /* Number of packets matched. */
263 atomic_ullong byte_count; /* Number of bytes matched. */
264 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
267 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
273 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
274 * its pmd thread's classifier. The text below calls this classifier 'cls'.
279 * The thread safety rules described here for "struct dp_netdev_flow" are
280 * motivated by two goals:
282 * - Prevent threads that read members of "struct dp_netdev_flow" from
283 * reading bad data due to changes by some thread concurrently modifying
286 * - Prevent two threads making changes to members of a given "struct
287 * dp_netdev_flow" from interfering with each other.
293 * A flow 'flow' may be accessed without a risk of being freed during an RCU
294 * grace period. Code that needs to hold onto a flow for a while
295 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
297 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
298 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
301 * Some members, marked 'const', are immutable. Accessing other members
302 * requires synchronization, as noted in more detail below.
304 struct dp_netdev_flow {
305 const struct flow flow; /* Unmasked flow that created this entry. */
306 /* Hash table index by unmasked flow. */
307 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
309 const ovs_u128 ufid; /* Unique flow identifier. */
310 const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */
313 /* Number of references.
314 * The classifier owns one reference.
315 * Any thread trying to keep a rule from being freed should hold its own
317 struct ovs_refcount ref_cnt;
322 struct dp_netdev_flow_stats stats;
325 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
327 /* While processing a group of input packets, the datapath uses the next
328 * member to store a pointer to the output batch for the flow. It is
329 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
330 * packet_batch_init() and packet_batch_execute()). */
331 struct packet_batch *batch;
333 /* Packet classification. */
334 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
335 /* 'cr' must be the last member. */
338 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
339 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
340 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
343 /* A set of datapath actions within a "struct dp_netdev_flow".
349 * A struct dp_netdev_actions 'actions' is protected with RCU. */
350 struct dp_netdev_actions {
351 /* These members are immutable: they do not change during the struct's
353 unsigned int size; /* Size of 'actions', in bytes. */
354 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
357 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
359 struct dp_netdev_actions *dp_netdev_flow_get_actions(
360 const struct dp_netdev_flow *);
361 static void dp_netdev_actions_free(struct dp_netdev_actions *);
363 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
364 struct dp_netdev_pmd_stats {
365 /* Indexed by DP_STAT_*. */
366 atomic_ullong n[DP_N_STATS];
369 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
370 struct dp_netdev_pmd_cycles {
371 /* Indexed by PMD_CYCLES_*. */
372 atomic_ullong n[PMD_N_CYCLES];
375 /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */
377 struct dp_netdev_port *port;
378 struct netdev_rxq *rx;
379 struct ovs_list node;
382 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
383 * the performance overhead of interrupt processing. Therefore netdev can
384 * not implement rx-wait for these devices. dpif-netdev needs to poll
385 * these device to check for recv buffer. pmd-thread does polling for
386 * devices assigned to itself.
388 * DPDK used PMD for accessing NIC.
390 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
391 * I/O of all non-pmd threads. There will be no actual thread created
394 * Each struct has its own flow table and classifier. Packets received
395 * from managed ports are looked up in the corresponding pmd thread's
396 * flow table, and are executed with the found actions.
398 struct dp_netdev_pmd_thread {
399 struct dp_netdev *dp;
400 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
401 struct cmap_node node; /* In 'dp->poll_threads'. */
403 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
404 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
406 /* Per thread exact-match cache. Note, the instance for cpu core
407 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
408 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
409 * instances will only be accessed by its own pmd thread. */
410 struct emc_cache flow_cache;
412 /* Classifier and Flow-Table.
414 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
415 * changes to 'cls' must be made while still holding the 'flow_mutex'.
417 struct ovs_mutex flow_mutex;
419 struct cmap flow_table OVS_GUARDED; /* Flow table. */
422 struct dp_netdev_pmd_stats stats;
424 /* Cycles counters */
425 struct dp_netdev_pmd_cycles cycles;
427 /* Used to count cicles. See 'cycles_counter_end()' */
428 unsigned long long last_cycles;
430 struct latch exit_latch; /* For terminating the pmd thread. */
431 atomic_uint change_seq; /* For reloading pmd ports. */
433 int index; /* Idx of this pmd thread among pmd*/
434 /* threads on same numa node. */
435 unsigned core_id; /* CPU core id of this pmd thread. */
436 int numa_id; /* numa node id of this pmd thread. */
437 atomic_int tx_qid; /* Queue id used by this pmd thread to
438 * send packets on all netdevs */
440 struct ovs_mutex poll_mutex; /* Mutex for poll_list. */
441 /* List of rx queues to poll. */
442 struct ovs_list poll_list OVS_GUARDED;
443 int poll_cnt; /* Number of elemints in poll_list. */
445 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
446 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
447 * values and subtracts them from 'stats' and 'cycles' before
448 * reporting to the user */
449 unsigned long long stats_zero[DP_N_STATS];
450 uint64_t cycles_zero[PMD_N_CYCLES];
453 #define PMD_INITIAL_SEQ 1
455 /* Interface to netdev-based datapath. */
458 struct dp_netdev *dp;
459 uint64_t last_port_seq;
462 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
463 struct dp_netdev_port **portp);
464 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
465 struct dp_netdev_port **portp);
466 static void dp_netdev_free(struct dp_netdev *)
467 OVS_REQUIRES(dp_netdev_mutex);
468 static int do_add_port(struct dp_netdev *dp, const char *devname,
469 const char *type, odp_port_t port_no)
470 OVS_REQUIRES(dp->port_mutex);
471 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
472 OVS_REQUIRES(dp->port_mutex);
473 static int dpif_netdev_open(const struct dpif_class *, const char *name,
474 bool create, struct dpif **);
475 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
476 struct dp_packet **, int c,
478 const struct nlattr *actions,
480 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
481 struct dp_packet **, int cnt, odp_port_t port_no);
482 static void dp_netdev_recirculate(struct dp_netdev_pmd_thread *,
483 struct dp_packet **, int cnt);
485 static void dp_netdev_disable_upcall(struct dp_netdev *);
486 static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
487 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
488 struct dp_netdev *dp, int index,
489 unsigned core_id, int numa_id);
490 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
491 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
492 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
494 static struct dp_netdev_pmd_thread *
495 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
496 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
497 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
498 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
500 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
501 struct dp_netdev_port *port, struct netdev_rxq *rx);
502 static struct dp_netdev_pmd_thread *
503 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id);
504 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
505 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
506 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
507 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
509 static inline bool emc_entry_alive(struct emc_entry *ce);
510 static void emc_clear_entry(struct emc_entry *ce);
513 emc_cache_init(struct emc_cache *flow_cache)
517 flow_cache->sweep_idx = 0;
518 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
519 flow_cache->entries[i].flow = NULL;
520 flow_cache->entries[i].key.hash = 0;
521 flow_cache->entries[i].key.len = sizeof(struct miniflow);
522 flowmap_init(&flow_cache->entries[i].key.mf.map);
527 emc_cache_uninit(struct emc_cache *flow_cache)
531 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
532 emc_clear_entry(&flow_cache->entries[i]);
536 /* Check and clear dead flow references slowly (one entry at each
539 emc_cache_slow_sweep(struct emc_cache *flow_cache)
541 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
543 if (!emc_entry_alive(entry)) {
544 emc_clear_entry(entry);
546 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
549 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
551 dpif_is_netdev(const struct dpif *dpif)
553 return dpif->dpif_class->open == dpif_netdev_open;
556 static struct dpif_netdev *
557 dpif_netdev_cast(const struct dpif *dpif)
559 ovs_assert(dpif_is_netdev(dpif));
560 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
563 static struct dp_netdev *
564 get_dp_netdev(const struct dpif *dpif)
566 return dpif_netdev_cast(dpif)->dp;
570 PMD_INFO_SHOW_STATS, /* show how cpu cycles are spent */
571 PMD_INFO_CLEAR_STATS /* set the cycles count to 0 */
575 pmd_info_show_stats(struct ds *reply,
576 struct dp_netdev_pmd_thread *pmd,
577 unsigned long long stats[DP_N_STATS],
578 uint64_t cycles[PMD_N_CYCLES])
580 unsigned long long total_packets = 0;
581 uint64_t total_cycles = 0;
584 /* These loops subtracts reference values ('*_zero') from the counters.
585 * Since loads and stores are relaxed, it might be possible for a '*_zero'
586 * value to be more recent than the current value we're reading from the
587 * counter. This is not a big problem, since these numbers are not
588 * supposed to be too accurate, but we should at least make sure that
589 * the result is not negative. */
590 for (i = 0; i < DP_N_STATS; i++) {
591 if (stats[i] > pmd->stats_zero[i]) {
592 stats[i] -= pmd->stats_zero[i];
597 if (i != DP_STAT_LOST) {
598 /* Lost packets are already included in DP_STAT_MISS */
599 total_packets += stats[i];
603 for (i = 0; i < PMD_N_CYCLES; i++) {
604 if (cycles[i] > pmd->cycles_zero[i]) {
605 cycles[i] -= pmd->cycles_zero[i];
610 total_cycles += cycles[i];
613 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
614 ? "main thread" : "pmd thread");
616 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
617 ds_put_format(reply, " numa_id %d", pmd->numa_id);
619 if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) {
620 ds_put_format(reply, " core_id %u", pmd->core_id);
622 ds_put_cstr(reply, ":\n");
625 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
626 "\tmiss:%llu\n\tlost:%llu\n",
627 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
628 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
630 if (total_cycles == 0) {
635 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
636 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
637 cycles[PMD_CYCLES_POLLING],
638 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
639 cycles[PMD_CYCLES_PROCESSING],
640 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
642 if (total_packets == 0) {
647 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
648 total_cycles / (double)total_packets,
649 total_cycles, total_packets);
652 "\tavg processing cycles per packet: "
653 "%.02f (%"PRIu64"/%llu)\n",
654 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
655 cycles[PMD_CYCLES_PROCESSING], total_packets);
659 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
660 struct dp_netdev_pmd_thread *pmd,
661 unsigned long long stats[DP_N_STATS],
662 uint64_t cycles[PMD_N_CYCLES])
666 /* We cannot write 'stats' and 'cycles' (because they're written by other
667 * threads) and we shouldn't change 'stats' (because they're used to count
668 * datapath stats, which must not be cleared here). Instead, we save the
669 * current values and subtract them from the values to be displayed in the
671 for (i = 0; i < DP_N_STATS; i++) {
672 pmd->stats_zero[i] = stats[i];
674 for (i = 0; i < PMD_N_CYCLES; i++) {
675 pmd->cycles_zero[i] = cycles[i];
680 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
683 struct ds reply = DS_EMPTY_INITIALIZER;
684 struct dp_netdev_pmd_thread *pmd;
685 struct dp_netdev *dp = NULL;
686 enum pmd_info_type type = *(enum pmd_info_type *) aux;
688 ovs_mutex_lock(&dp_netdev_mutex);
691 dp = shash_find_data(&dp_netdevs, argv[1]);
692 } else if (shash_count(&dp_netdevs) == 1) {
693 /* There's only one datapath */
694 dp = shash_first(&dp_netdevs)->data;
698 ovs_mutex_unlock(&dp_netdev_mutex);
699 unixctl_command_reply_error(conn,
700 "please specify an existing datapath");
704 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
705 unsigned long long stats[DP_N_STATS];
706 uint64_t cycles[PMD_N_CYCLES];
709 /* Read current stats and cycle counters */
710 for (i = 0; i < ARRAY_SIZE(stats); i++) {
711 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
713 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
714 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
717 if (type == PMD_INFO_CLEAR_STATS) {
718 pmd_info_clear_stats(&reply, pmd, stats, cycles);
719 } else if (type == PMD_INFO_SHOW_STATS) {
720 pmd_info_show_stats(&reply, pmd, stats, cycles);
724 ovs_mutex_unlock(&dp_netdev_mutex);
726 unixctl_command_reply(conn, ds_cstr(&reply));
731 dpif_netdev_init(void)
733 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
734 clear_aux = PMD_INFO_CLEAR_STATS;
736 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
737 0, 1, dpif_netdev_pmd_info,
739 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
740 0, 1, dpif_netdev_pmd_info,
746 dpif_netdev_enumerate(struct sset *all_dps,
747 const struct dpif_class *dpif_class)
749 struct shash_node *node;
751 ovs_mutex_lock(&dp_netdev_mutex);
752 SHASH_FOR_EACH(node, &dp_netdevs) {
753 struct dp_netdev *dp = node->data;
754 if (dpif_class != dp->class) {
755 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
756 * If the class doesn't match, skip this dpif. */
759 sset_add(all_dps, node->name);
761 ovs_mutex_unlock(&dp_netdev_mutex);
767 dpif_netdev_class_is_dummy(const struct dpif_class *class)
769 return class != &dpif_netdev_class;
773 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
775 return strcmp(type, "internal") ? type
776 : dpif_netdev_class_is_dummy(class) ? "dummy"
781 create_dpif_netdev(struct dp_netdev *dp)
783 uint16_t netflow_id = hash_string(dp->name, 0);
784 struct dpif_netdev *dpif;
786 ovs_refcount_ref(&dp->ref_cnt);
788 dpif = xmalloc(sizeof *dpif);
789 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
791 dpif->last_port_seq = seq_read(dp->port_seq);
796 /* Choose an unused, non-zero port number and return it on success.
797 * Return ODPP_NONE on failure. */
799 choose_port(struct dp_netdev *dp, const char *name)
800 OVS_REQUIRES(dp->port_mutex)
804 if (dp->class != &dpif_netdev_class) {
808 /* If the port name begins with "br", start the number search at
809 * 100 to make writing tests easier. */
810 if (!strncmp(name, "br", 2)) {
814 /* If the port name contains a number, try to assign that port number.
815 * This can make writing unit tests easier because port numbers are
817 for (p = name; *p != '\0'; p++) {
818 if (isdigit((unsigned char) *p)) {
819 port_no = start_no + strtol(p, NULL, 10);
820 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
821 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
822 return u32_to_odp(port_no);
829 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
830 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
831 return u32_to_odp(port_no);
839 create_dp_netdev(const char *name, const struct dpif_class *class,
840 struct dp_netdev **dpp)
841 OVS_REQUIRES(dp_netdev_mutex)
843 struct dp_netdev *dp;
846 dp = xzalloc(sizeof *dp);
847 shash_add(&dp_netdevs, name, dp);
849 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
850 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
851 ovs_refcount_init(&dp->ref_cnt);
852 atomic_flag_clear(&dp->destroyed);
854 ovs_mutex_init(&dp->port_mutex);
855 cmap_init(&dp->ports);
856 dp->port_seq = seq_create();
857 fat_rwlock_init(&dp->upcall_rwlock);
859 /* Disable upcalls by default. */
860 dp_netdev_disable_upcall(dp);
861 dp->upcall_aux = NULL;
862 dp->upcall_cb = NULL;
864 cmap_init(&dp->poll_threads);
865 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
866 ovsthread_key_create(&dp->per_pmd_key, NULL);
868 dp_netdev_set_nonpmd(dp);
870 ovs_mutex_lock(&dp->port_mutex);
871 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
872 ovs_mutex_unlock(&dp->port_mutex);
878 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
884 dpif_netdev_open(const struct dpif_class *class, const char *name,
885 bool create, struct dpif **dpifp)
887 struct dp_netdev *dp;
890 ovs_mutex_lock(&dp_netdev_mutex);
891 dp = shash_find_data(&dp_netdevs, name);
893 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
895 error = (dp->class != class ? EINVAL
900 *dpifp = create_dpif_netdev(dp);
903 ovs_mutex_unlock(&dp_netdev_mutex);
909 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
910 OVS_NO_THREAD_SAFETY_ANALYSIS
912 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
913 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
915 /* Before freeing a lock we should release it */
916 fat_rwlock_unlock(&dp->upcall_rwlock);
917 fat_rwlock_destroy(&dp->upcall_rwlock);
920 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
921 * through the 'dp_netdevs' shash while freeing 'dp'. */
923 dp_netdev_free(struct dp_netdev *dp)
924 OVS_REQUIRES(dp_netdev_mutex)
926 struct dp_netdev_port *port;
928 shash_find_and_delete(&dp_netdevs, dp->name);
930 dp_netdev_destroy_all_pmds(dp);
931 ovs_mutex_destroy(&dp->non_pmd_mutex);
932 ovsthread_key_delete(dp->per_pmd_key);
934 ovs_mutex_lock(&dp->port_mutex);
935 CMAP_FOR_EACH (port, node, &dp->ports) {
936 /* PMD threads are destroyed here. do_del_port() cannot quiesce */
937 do_del_port(dp, port);
939 ovs_mutex_unlock(&dp->port_mutex);
940 cmap_destroy(&dp->poll_threads);
942 seq_destroy(dp->port_seq);
943 cmap_destroy(&dp->ports);
945 /* Upcalls must be disabled at this point */
946 dp_netdev_destroy_upcall_lock(dp);
949 free(CONST_CAST(char *, dp->name));
954 dp_netdev_unref(struct dp_netdev *dp)
957 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
958 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
959 ovs_mutex_lock(&dp_netdev_mutex);
960 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
963 ovs_mutex_unlock(&dp_netdev_mutex);
968 dpif_netdev_close(struct dpif *dpif)
970 struct dp_netdev *dp = get_dp_netdev(dpif);
977 dpif_netdev_destroy(struct dpif *dpif)
979 struct dp_netdev *dp = get_dp_netdev(dpif);
981 if (!atomic_flag_test_and_set(&dp->destroyed)) {
982 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
983 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
991 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
992 * load/store semantics. While the increment is not atomic, the load and
993 * store operations are, making it impossible to read inconsistent values.
995 * This is used to update thread local stats counters. */
997 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
999 unsigned long long tmp;
1001 atomic_read_relaxed(var, &tmp);
1003 atomic_store_relaxed(var, tmp);
1007 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
1009 struct dp_netdev *dp = get_dp_netdev(dpif);
1010 struct dp_netdev_pmd_thread *pmd;
1012 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
1013 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1014 unsigned long long n;
1015 stats->n_flows += cmap_count(&pmd->flow_table);
1017 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
1019 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
1021 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
1022 stats->n_missed += n;
1023 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
1026 stats->n_masks = UINT32_MAX;
1027 stats->n_mask_hit = UINT64_MAX;
1033 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1037 if (pmd->core_id == NON_PMD_CORE_ID) {
1041 ovs_mutex_lock(&pmd->cond_mutex);
1042 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1043 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1044 ovs_mutex_unlock(&pmd->cond_mutex);
1048 hash_port_no(odp_port_t port_no)
1050 return hash_int(odp_to_u32(port_no), 0);
1054 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1056 OVS_REQUIRES(dp->port_mutex)
1058 struct netdev_saved_flags *sf;
1059 struct dp_netdev_port *port;
1060 struct netdev *netdev;
1061 enum netdev_flags flags;
1062 const char *open_type;
1066 /* Reject devices already in 'dp'. */
1067 if (!get_port_by_name(dp, devname, &port)) {
1071 /* Open and validate network device. */
1072 open_type = dpif_netdev_port_open_type(dp->class, type);
1073 error = netdev_open(devname, open_type, &netdev);
1077 /* XXX reject non-Ethernet devices */
1079 netdev_get_flags(netdev, &flags);
1080 if (flags & NETDEV_LOOPBACK) {
1081 VLOG_ERR("%s: cannot add a loopback device", devname);
1082 netdev_close(netdev);
1086 if (netdev_is_pmd(netdev)) {
1087 int n_cores = ovs_numa_get_n_cores();
1089 if (n_cores == OVS_CORE_UNSPEC) {
1090 VLOG_ERR("%s, cannot get cpu core info", devname);
1093 /* There can only be ovs_numa_get_n_cores() pmd threads,
1094 * so creates a txq for each, and one extra for the non
1096 error = netdev_set_multiq(netdev, n_cores + 1,
1097 netdev_requested_n_rxq(netdev));
1098 if (error && (error != EOPNOTSUPP)) {
1099 VLOG_ERR("%s, cannot set multiq", devname);
1103 port = xzalloc(sizeof *port);
1104 port->port_no = port_no;
1105 port->netdev = netdev;
1106 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
1107 port->type = xstrdup(type);
1108 port->latest_requested_n_rxq = netdev_requested_n_rxq(netdev);
1109 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1110 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1112 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
1113 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1114 devname, ovs_strerror(errno));
1115 netdev_close(netdev);
1123 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1125 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1126 netdev_rxq_close(port->rxq[i]);
1128 netdev_close(netdev);
1136 ovs_refcount_init(&port->ref_cnt);
1137 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1139 if (netdev_is_pmd(netdev)) {
1140 int numa_id = netdev_get_numa_id(netdev);
1141 struct dp_netdev_pmd_thread *pmd;
1143 /* Cannot create pmd threads for invalid numa node. */
1144 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
1146 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1147 pmd = dp_netdev_less_loaded_pmd_on_numa(dp, numa_id);
1149 /* There is no pmd threads on this numa node. */
1150 dp_netdev_set_pmds_on_numa(dp, numa_id);
1151 /* Assigning of rx queues done. */
1155 ovs_mutex_lock(&pmd->poll_mutex);
1156 dp_netdev_add_rxq_to_pmd(pmd, port, port->rxq[i]);
1157 ovs_mutex_unlock(&pmd->poll_mutex);
1158 dp_netdev_reload_pmd__(pmd);
1161 seq_change(dp->port_seq);
1167 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1168 odp_port_t *port_nop)
1170 struct dp_netdev *dp = get_dp_netdev(dpif);
1171 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1172 const char *dpif_port;
1176 ovs_mutex_lock(&dp->port_mutex);
1177 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1178 if (*port_nop != ODPP_NONE) {
1179 port_no = *port_nop;
1180 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1182 port_no = choose_port(dp, dpif_port);
1183 error = port_no == ODPP_NONE ? EFBIG : 0;
1186 *port_nop = port_no;
1187 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1189 ovs_mutex_unlock(&dp->port_mutex);
1195 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1197 struct dp_netdev *dp = get_dp_netdev(dpif);
1200 ovs_mutex_lock(&dp->port_mutex);
1201 if (port_no == ODPP_LOCAL) {
1204 struct dp_netdev_port *port;
1206 error = get_port_by_number(dp, port_no, &port);
1208 do_del_port(dp, port);
1211 ovs_mutex_unlock(&dp->port_mutex);
1217 is_valid_port_number(odp_port_t port_no)
1219 return port_no != ODPP_NONE;
1222 static struct dp_netdev_port *
1223 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1225 struct dp_netdev_port *port;
1227 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1228 if (port->port_no == port_no) {
1236 get_port_by_number(struct dp_netdev *dp,
1237 odp_port_t port_no, struct dp_netdev_port **portp)
1239 if (!is_valid_port_number(port_no)) {
1243 *portp = dp_netdev_lookup_port(dp, port_no);
1244 return *portp ? 0 : ENOENT;
1249 port_ref(struct dp_netdev_port *port)
1252 ovs_refcount_ref(&port->ref_cnt);
1257 port_unref(struct dp_netdev_port *port)
1259 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1260 int n_rxq = netdev_n_rxq(port->netdev);
1263 netdev_close(port->netdev);
1264 netdev_restore_flags(port->sf);
1266 for (i = 0; i < n_rxq; i++) {
1267 netdev_rxq_close(port->rxq[i]);
1276 get_port_by_name(struct dp_netdev *dp,
1277 const char *devname, struct dp_netdev_port **portp)
1278 OVS_REQUIRES(dp->port_mutex)
1280 struct dp_netdev_port *port;
1282 CMAP_FOR_EACH (port, node, &dp->ports) {
1283 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1292 get_n_pmd_threads(struct dp_netdev *dp)
1294 /* There is one non pmd thread in dp->poll_threads */
1295 return cmap_count(&dp->poll_threads) - 1;
1299 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1301 struct dp_netdev_pmd_thread *pmd;
1304 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1305 if (pmd->numa_id == numa_id) {
1313 /* Returns 'true' if there is a port with pmd netdev and the netdev
1314 * is on numa node 'numa_id'. */
1316 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1318 struct dp_netdev_port *port;
1320 CMAP_FOR_EACH (port, node, &dp->ports) {
1321 if (netdev_is_pmd(port->netdev)
1322 && netdev_get_numa_id(port->netdev) == numa_id) {
1332 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1333 OVS_REQUIRES(dp->port_mutex)
1335 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1336 seq_change(dp->port_seq);
1337 if (netdev_is_pmd(port->netdev)) {
1338 int numa_id = netdev_get_numa_id(port->netdev);
1340 /* PMD threads can not be on invalid numa node. */
1341 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
1342 /* If there is no netdev on the numa node, deletes the pmd threads
1343 * for that numa. Else, deletes the queues from polling lists. */
1344 if (!has_pmd_port_for_numa(dp, numa_id)) {
1345 dp_netdev_del_pmds_on_numa(dp, numa_id);
1347 struct dp_netdev_pmd_thread *pmd;
1348 struct rxq_poll *poll, *next;
1350 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1351 if (pmd->numa_id == numa_id) {
1354 ovs_mutex_lock(&pmd->poll_mutex);
1355 LIST_FOR_EACH_SAFE (poll, next, node, &pmd->poll_list) {
1356 if (poll->port == port) {
1358 port_unref(poll->port);
1359 list_remove(&poll->node);
1364 ovs_mutex_unlock(&pmd->poll_mutex);
1366 dp_netdev_reload_pmd__(pmd);
1377 answer_port_query(const struct dp_netdev_port *port,
1378 struct dpif_port *dpif_port)
1380 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1381 dpif_port->type = xstrdup(port->type);
1382 dpif_port->port_no = port->port_no;
1386 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1387 struct dpif_port *dpif_port)
1389 struct dp_netdev *dp = get_dp_netdev(dpif);
1390 struct dp_netdev_port *port;
1393 error = get_port_by_number(dp, port_no, &port);
1394 if (!error && dpif_port) {
1395 answer_port_query(port, dpif_port);
1402 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1403 struct dpif_port *dpif_port)
1405 struct dp_netdev *dp = get_dp_netdev(dpif);
1406 struct dp_netdev_port *port;
1409 ovs_mutex_lock(&dp->port_mutex);
1410 error = get_port_by_name(dp, devname, &port);
1411 if (!error && dpif_port) {
1412 answer_port_query(port, dpif_port);
1414 ovs_mutex_unlock(&dp->port_mutex);
1420 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1422 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1426 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1428 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1429 ovsrcu_postpone(dp_netdev_flow_free, flow);
1434 dp_netdev_flow_hash(const ovs_u128 *ufid)
1436 return ufid->u32[0];
1440 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1441 struct dp_netdev_flow *flow)
1442 OVS_REQUIRES(pmd->flow_mutex)
1444 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1446 dpcls_remove(&pmd->cls, &flow->cr);
1447 flow->cr.mask = NULL; /* Accessing rule's mask after this is not safe. */
1449 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1452 dp_netdev_flow_unref(flow);
1456 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1458 struct dp_netdev_flow *netdev_flow;
1460 ovs_mutex_lock(&pmd->flow_mutex);
1461 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1462 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1464 ovs_mutex_unlock(&pmd->flow_mutex);
1468 dpif_netdev_flow_flush(struct dpif *dpif)
1470 struct dp_netdev *dp = get_dp_netdev(dpif);
1471 struct dp_netdev_pmd_thread *pmd;
1473 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1474 dp_netdev_pmd_flow_flush(pmd);
1480 struct dp_netdev_port_state {
1481 struct cmap_position position;
1486 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1488 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1493 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1494 struct dpif_port *dpif_port)
1496 struct dp_netdev_port_state *state = state_;
1497 struct dp_netdev *dp = get_dp_netdev(dpif);
1498 struct cmap_node *node;
1501 node = cmap_next_position(&dp->ports, &state->position);
1503 struct dp_netdev_port *port;
1505 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1508 state->name = xstrdup(netdev_get_name(port->netdev));
1509 dpif_port->name = state->name;
1510 dpif_port->type = port->type;
1511 dpif_port->port_no = port->port_no;
1522 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1524 struct dp_netdev_port_state *state = state_;
1531 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1533 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1534 uint64_t new_port_seq;
1537 new_port_seq = seq_read(dpif->dp->port_seq);
1538 if (dpif->last_port_seq != new_port_seq) {
1539 dpif->last_port_seq = new_port_seq;
1549 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1551 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1553 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1556 static struct dp_netdev_flow *
1557 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1559 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1562 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1564 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1567 /* netdev_flow_key utilities.
1569 * netdev_flow_key is basically a miniflow. We use these functions
1570 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1571 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1573 * - Since we are dealing exclusively with miniflows created by
1574 * miniflow_extract(), if the map is different the miniflow is different.
1575 * Therefore we can be faster by comparing the map and the miniflow in a
1577 * - These functions can be inlined by the compiler. */
1579 /* Given the number of bits set in miniflow's maps, returns the size of the
1580 * 'netdev_flow_key.mf' */
1581 static inline size_t
1582 netdev_flow_key_size(size_t flow_u64s)
1584 return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s);
1588 netdev_flow_key_equal(const struct netdev_flow_key *a,
1589 const struct netdev_flow_key *b)
1591 /* 'b->len' may be not set yet. */
1592 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1595 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1596 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1597 * generated by miniflow_extract. */
1599 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1600 const struct miniflow *mf)
1602 return !memcmp(&key->mf, mf, key->len);
1606 netdev_flow_key_clone(struct netdev_flow_key *dst,
1607 const struct netdev_flow_key *src)
1610 offsetof(struct netdev_flow_key, mf) + src->len);
1615 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1616 const struct flow *src)
1618 struct dp_packet packet;
1619 uint64_t buf_stub[512 / 8];
1621 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1622 pkt_metadata_from_flow(&packet.md, src);
1623 flow_compose(&packet, src);
1624 miniflow_extract(&packet, &dst->mf);
1625 dp_packet_uninit(&packet);
1627 dst->len = netdev_flow_key_size(miniflow_n_values(&dst->mf));
1628 dst->hash = 0; /* Not computed yet. */
1631 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1633 netdev_flow_mask_init(struct netdev_flow_key *mask,
1634 const struct match *match)
1636 uint64_t *dst = miniflow_values(&mask->mf);
1637 struct flowmap fmap;
1641 /* Only check masks that make sense for the flow. */
1642 flow_wc_map(&match->flow, &fmap);
1643 flowmap_init(&mask->mf.map);
1645 FLOWMAP_FOR_EACH_INDEX(idx, fmap) {
1646 uint64_t mask_u64 = flow_u64_value(&match->wc.masks, idx);
1649 flowmap_set(&mask->mf.map, idx, 1);
1651 hash = hash_add64(hash, mask_u64);
1657 FLOWMAP_FOR_EACH_MAP (map, mask->mf.map) {
1658 hash = hash_add64(hash, map);
1661 size_t n = dst - miniflow_get_values(&mask->mf);
1663 mask->hash = hash_finish(hash, n * 8);
1664 mask->len = netdev_flow_key_size(n);
1667 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1669 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1670 const struct flow *flow,
1671 const struct netdev_flow_key *mask)
1673 uint64_t *dst_u64 = miniflow_values(&dst->mf);
1674 const uint64_t *mask_u64 = miniflow_get_values(&mask->mf);
1678 dst->len = mask->len;
1679 dst->mf = mask->mf; /* Copy maps. */
1681 FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf.map) {
1682 *dst_u64 = value & *mask_u64++;
1683 hash = hash_add64(hash, *dst_u64++);
1685 dst->hash = hash_finish(hash,
1686 (dst_u64 - miniflow_get_values(&dst->mf)) * 8);
1689 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
1690 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
1691 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
1693 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1695 static inline uint32_t
1696 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1697 const struct netdev_flow_key *mask)
1699 const uint64_t *p = miniflow_get_values(&mask->mf);
1703 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, key, mask->mf.map) {
1704 hash = hash_add64(hash, value & *p++);
1707 return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8);
1711 emc_entry_alive(struct emc_entry *ce)
1713 return ce->flow && !ce->flow->dead;
1717 emc_clear_entry(struct emc_entry *ce)
1720 dp_netdev_flow_unref(ce->flow);
1726 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1727 const struct netdev_flow_key *key)
1729 if (ce->flow != flow) {
1731 dp_netdev_flow_unref(ce->flow);
1734 if (dp_netdev_flow_ref(flow)) {
1741 netdev_flow_key_clone(&ce->key, key);
1746 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1747 struct dp_netdev_flow *flow)
1749 struct emc_entry *to_be_replaced = NULL;
1750 struct emc_entry *current_entry;
1752 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1753 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1754 /* We found the entry with the 'mf' miniflow */
1755 emc_change_entry(current_entry, flow, NULL);
1759 /* Replacement policy: put the flow in an empty (not alive) entry, or
1760 * in the first entry where it can be */
1762 || (emc_entry_alive(to_be_replaced)
1763 && !emc_entry_alive(current_entry))
1764 || current_entry->key.hash < to_be_replaced->key.hash) {
1765 to_be_replaced = current_entry;
1768 /* We didn't find the miniflow in the cache.
1769 * The 'to_be_replaced' entry is where the new flow will be stored */
1771 emc_change_entry(to_be_replaced, flow, key);
1774 static inline struct dp_netdev_flow *
1775 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1777 struct emc_entry *current_entry;
1779 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1780 if (current_entry->key.hash == key->hash
1781 && emc_entry_alive(current_entry)
1782 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1784 /* We found the entry with the 'key->mf' miniflow */
1785 return current_entry->flow;
1792 static struct dp_netdev_flow *
1793 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1794 const struct netdev_flow_key *key)
1796 struct dp_netdev_flow *netdev_flow;
1797 struct dpcls_rule *rule;
1799 dpcls_lookup(&pmd->cls, key, &rule, 1);
1800 netdev_flow = dp_netdev_flow_cast(rule);
1805 static struct dp_netdev_flow *
1806 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1807 const ovs_u128 *ufidp, const struct nlattr *key,
1810 struct dp_netdev_flow *netdev_flow;
1814 /* If a UFID is not provided, determine one based on the key. */
1815 if (!ufidp && key && key_len
1816 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1817 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1822 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1824 if (ovs_u128_equals(&netdev_flow->ufid, ufidp)) {
1834 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1835 struct dpif_flow_stats *stats)
1837 struct dp_netdev_flow *netdev_flow;
1838 unsigned long long n;
1842 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1844 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1845 stats->n_packets = n;
1846 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1848 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1850 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1851 stats->tcp_flags = flags;
1854 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1855 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1856 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1859 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1860 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1861 struct dpif_flow *flow, bool terse)
1864 memset(flow, 0, sizeof *flow);
1866 struct flow_wildcards wc;
1867 struct dp_netdev_actions *actions;
1869 struct odp_flow_key_parms odp_parms = {
1870 .flow = &netdev_flow->flow,
1872 .support = dp_netdev_support,
1875 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1878 offset = key_buf->size;
1879 flow->key = ofpbuf_tail(key_buf);
1880 odp_parms.odp_in_port = netdev_flow->flow.in_port.odp_port;
1881 odp_flow_key_from_flow(&odp_parms, key_buf);
1882 flow->key_len = key_buf->size - offset;
1885 offset = mask_buf->size;
1886 flow->mask = ofpbuf_tail(mask_buf);
1887 odp_parms.odp_in_port = wc.masks.in_port.odp_port;
1888 odp_parms.key_buf = key_buf;
1889 odp_flow_key_from_mask(&odp_parms, mask_buf);
1890 flow->mask_len = mask_buf->size - offset;
1893 actions = dp_netdev_flow_get_actions(netdev_flow);
1894 flow->actions = actions->actions;
1895 flow->actions_len = actions->size;
1898 flow->ufid = netdev_flow->ufid;
1899 flow->ufid_present = true;
1900 flow->pmd_id = netdev_flow->pmd_id;
1901 get_dpif_flow_stats(netdev_flow, &flow->stats);
1905 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1906 const struct nlattr *mask_key,
1907 uint32_t mask_key_len, const struct flow *flow,
1908 struct flow_wildcards *wc)
1910 enum odp_key_fitness fitness;
1912 fitness = odp_flow_key_to_mask_udpif(mask_key, mask_key_len, key,
1915 /* This should not happen: it indicates that
1916 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1917 * disagree on the acceptable form of a mask. Log the problem
1918 * as an error, with enough details to enable debugging. */
1919 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1921 if (!VLOG_DROP_ERR(&rl)) {
1925 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1927 VLOG_ERR("internal error parsing flow mask %s (%s)",
1928 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1939 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1944 if (odp_flow_key_to_flow_udpif(key, key_len, flow)) {
1945 /* This should not happen: it indicates that odp_flow_key_from_flow()
1946 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1947 * flow. Log the problem as an error, with enough details to enable
1949 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1951 if (!VLOG_DROP_ERR(&rl)) {
1955 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1956 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1963 in_port = flow->in_port.odp_port;
1964 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1968 /* Userspace datapath doesn't support conntrack. */
1969 if (flow->ct_state || flow->ct_zone || flow->ct_mark
1970 || !ovs_u128_is_zero(&flow->ct_label)) {
1978 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1980 struct dp_netdev *dp = get_dp_netdev(dpif);
1981 struct dp_netdev_flow *netdev_flow;
1982 struct dp_netdev_pmd_thread *pmd;
1983 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1984 ? NON_PMD_CORE_ID : get->pmd_id;
1987 pmd = dp_netdev_get_pmd(dp, pmd_id);
1992 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1995 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
2000 dp_netdev_pmd_unref(pmd);
2006 static struct dp_netdev_flow *
2007 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
2008 struct match *match, const ovs_u128 *ufid,
2009 const struct nlattr *actions, size_t actions_len)
2010 OVS_REQUIRES(pmd->flow_mutex)
2012 struct dp_netdev_flow *flow;
2013 struct netdev_flow_key mask;
2015 netdev_flow_mask_init(&mask, match);
2016 /* Make sure wc does not have metadata. */
2017 ovs_assert(!FLOWMAP_HAS_FIELD(&mask.mf.map, metadata)
2018 && !FLOWMAP_HAS_FIELD(&mask.mf.map, regs));
2020 /* Do not allocate extra space. */
2021 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
2022 memset(&flow->stats, 0, sizeof flow->stats);
2025 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
2026 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
2027 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
2028 ovs_refcount_init(&flow->ref_cnt);
2029 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
2031 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
2032 dpcls_insert(&pmd->cls, &flow->cr, &mask);
2034 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
2035 dp_netdev_flow_hash(&flow->ufid));
2037 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2039 struct ds ds = DS_EMPTY_INITIALIZER;
2041 match.tun_md.valid = false;
2042 match.flow = flow->flow;
2043 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2045 ds_put_cstr(&ds, "flow_add: ");
2046 odp_format_ufid(ufid, &ds);
2047 ds_put_cstr(&ds, " ");
2048 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2049 ds_put_cstr(&ds, ", actions:");
2050 format_odp_actions(&ds, actions, actions_len);
2052 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2061 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2063 struct dp_netdev *dp = get_dp_netdev(dpif);
2064 struct dp_netdev_flow *netdev_flow;
2065 struct netdev_flow_key key;
2066 struct dp_netdev_pmd_thread *pmd;
2069 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2070 ? NON_PMD_CORE_ID : put->pmd_id;
2073 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2077 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2078 put->mask, put->mask_len,
2079 &match.flow, &match.wc);
2084 pmd = dp_netdev_get_pmd(dp, pmd_id);
2089 /* Must produce a netdev_flow_key for lookup.
2090 * This interface is no longer performance critical, since it is not used
2091 * for upcall processing any more. */
2092 netdev_flow_key_from_flow(&key, &match.flow);
2097 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2100 ovs_mutex_lock(&pmd->flow_mutex);
2101 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2103 if (put->flags & DPIF_FP_CREATE) {
2104 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2106 memset(put->stats, 0, sizeof *put->stats);
2108 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2118 if (put->flags & DPIF_FP_MODIFY
2119 && flow_equal(&match.flow, &netdev_flow->flow)) {
2120 struct dp_netdev_actions *new_actions;
2121 struct dp_netdev_actions *old_actions;
2123 new_actions = dp_netdev_actions_create(put->actions,
2126 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2127 ovsrcu_set(&netdev_flow->actions, new_actions);
2130 get_dpif_flow_stats(netdev_flow, put->stats);
2132 if (put->flags & DPIF_FP_ZERO_STATS) {
2133 /* XXX: The userspace datapath uses thread local statistics
2134 * (for flows), which should be updated only by the owning
2135 * thread. Since we cannot write on stats memory here,
2136 * we choose not to support this flag. Please note:
2137 * - This feature is currently used only by dpctl commands with
2139 * - Should the need arise, this operation can be implemented
2140 * by keeping a base value (to be update here) for each
2141 * counter, and subtracting it before outputting the stats */
2145 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2146 } else if (put->flags & DPIF_FP_CREATE) {
2149 /* Overlapping flow. */
2153 ovs_mutex_unlock(&pmd->flow_mutex);
2154 dp_netdev_pmd_unref(pmd);
2160 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2162 struct dp_netdev *dp = get_dp_netdev(dpif);
2163 struct dp_netdev_flow *netdev_flow;
2164 struct dp_netdev_pmd_thread *pmd;
2165 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2166 ? NON_PMD_CORE_ID : del->pmd_id;
2169 pmd = dp_netdev_get_pmd(dp, pmd_id);
2174 ovs_mutex_lock(&pmd->flow_mutex);
2175 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2179 get_dpif_flow_stats(netdev_flow, del->stats);
2181 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2185 ovs_mutex_unlock(&pmd->flow_mutex);
2186 dp_netdev_pmd_unref(pmd);
2191 struct dpif_netdev_flow_dump {
2192 struct dpif_flow_dump up;
2193 struct cmap_position poll_thread_pos;
2194 struct cmap_position flow_pos;
2195 struct dp_netdev_pmd_thread *cur_pmd;
2197 struct ovs_mutex mutex;
2200 static struct dpif_netdev_flow_dump *
2201 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2203 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2206 static struct dpif_flow_dump *
2207 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2209 struct dpif_netdev_flow_dump *dump;
2211 dump = xzalloc(sizeof *dump);
2212 dpif_flow_dump_init(&dump->up, dpif_);
2213 dump->up.terse = terse;
2214 ovs_mutex_init(&dump->mutex);
2220 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2222 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2224 ovs_mutex_destroy(&dump->mutex);
2229 struct dpif_netdev_flow_dump_thread {
2230 struct dpif_flow_dump_thread up;
2231 struct dpif_netdev_flow_dump *dump;
2232 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2233 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2236 static struct dpif_netdev_flow_dump_thread *
2237 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2239 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2242 static struct dpif_flow_dump_thread *
2243 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2245 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2246 struct dpif_netdev_flow_dump_thread *thread;
2248 thread = xmalloc(sizeof *thread);
2249 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2250 thread->dump = dump;
2255 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2257 struct dpif_netdev_flow_dump_thread *thread
2258 = dpif_netdev_flow_dump_thread_cast(thread_);
2264 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2265 struct dpif_flow *flows, int max_flows)
2267 struct dpif_netdev_flow_dump_thread *thread
2268 = dpif_netdev_flow_dump_thread_cast(thread_);
2269 struct dpif_netdev_flow_dump *dump = thread->dump;
2270 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2274 ovs_mutex_lock(&dump->mutex);
2275 if (!dump->status) {
2276 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2277 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2278 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2279 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2281 /* First call to dump_next(), extracts the first pmd thread.
2282 * If there is no pmd thread, returns immediately. */
2284 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2286 ovs_mutex_unlock(&dump->mutex);
2293 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2294 struct cmap_node *node;
2296 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2300 netdev_flows[n_flows] = CONTAINER_OF(node,
2301 struct dp_netdev_flow,
2304 /* When finishing dumping the current pmd thread, moves to
2306 if (n_flows < flow_limit) {
2307 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2308 dp_netdev_pmd_unref(pmd);
2309 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2315 /* Keeps the reference to next caller. */
2316 dump->cur_pmd = pmd;
2318 /* If the current dump is empty, do not exit the loop, since the
2319 * remaining pmds could have flows to be dumped. Just dumps again
2320 * on the new 'pmd'. */
2323 ovs_mutex_unlock(&dump->mutex);
2325 for (i = 0; i < n_flows; i++) {
2326 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2327 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2328 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2329 struct dpif_flow *f = &flows[i];
2330 struct ofpbuf key, mask;
2332 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2333 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2334 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2342 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2343 OVS_NO_THREAD_SAFETY_ANALYSIS
2345 struct dp_netdev *dp = get_dp_netdev(dpif);
2346 struct dp_netdev_pmd_thread *pmd;
2347 struct dp_packet *pp;
2349 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2350 dp_packet_size(execute->packet) > UINT16_MAX) {
2354 /* Tries finding the 'pmd'. If NULL is returned, that means
2355 * the current thread is a non-pmd thread and should use
2356 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2357 pmd = ovsthread_getspecific(dp->per_pmd_key);
2359 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2362 /* If the current thread is non-pmd thread, acquires
2363 * the 'non_pmd_mutex'. */
2364 if (pmd->core_id == NON_PMD_CORE_ID) {
2365 ovs_mutex_lock(&dp->non_pmd_mutex);
2366 ovs_mutex_lock(&dp->port_mutex);
2369 pp = execute->packet;
2370 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2371 execute->actions_len);
2372 if (pmd->core_id == NON_PMD_CORE_ID) {
2373 dp_netdev_pmd_unref(pmd);
2374 ovs_mutex_unlock(&dp->port_mutex);
2375 ovs_mutex_unlock(&dp->non_pmd_mutex);
2382 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2386 for (i = 0; i < n_ops; i++) {
2387 struct dpif_op *op = ops[i];
2390 case DPIF_OP_FLOW_PUT:
2391 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2394 case DPIF_OP_FLOW_DEL:
2395 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2398 case DPIF_OP_EXECUTE:
2399 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2402 case DPIF_OP_FLOW_GET:
2403 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2409 /* Returns true if the configuration for rx queues or cpu mask
2412 pmd_config_changed(const struct dp_netdev *dp, const char *cmask)
2414 struct dp_netdev_port *port;
2416 CMAP_FOR_EACH (port, node, &dp->ports) {
2417 struct netdev *netdev = port->netdev;
2418 int requested_n_rxq = netdev_requested_n_rxq(netdev);
2419 if (netdev_is_pmd(netdev)
2420 && port->latest_requested_n_rxq != requested_n_rxq) {
2425 if (dp->pmd_cmask != NULL && cmask != NULL) {
2426 return strcmp(dp->pmd_cmask, cmask);
2428 return (dp->pmd_cmask != NULL || cmask != NULL);
2432 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2434 dpif_netdev_pmd_set(struct dpif *dpif, const char *cmask)
2436 struct dp_netdev *dp = get_dp_netdev(dpif);
2438 if (pmd_config_changed(dp, cmask)) {
2439 struct dp_netdev_port *port;
2441 dp_netdev_destroy_all_pmds(dp);
2443 CMAP_FOR_EACH (port, node, &dp->ports) {
2444 struct netdev *netdev = port->netdev;
2445 int requested_n_rxq = netdev_requested_n_rxq(netdev);
2446 if (netdev_is_pmd(port->netdev)
2447 && port->latest_requested_n_rxq != requested_n_rxq) {
2450 /* Closes the existing 'rxq's. */
2451 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2452 netdev_rxq_close(port->rxq[i]);
2453 port->rxq[i] = NULL;
2456 /* Sets the new rx queue config. */
2457 err = netdev_set_multiq(port->netdev,
2458 ovs_numa_get_n_cores() + 1,
2460 if (err && (err != EOPNOTSUPP)) {
2461 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2462 " %u", netdev_get_name(port->netdev),
2466 port->latest_requested_n_rxq = requested_n_rxq;
2467 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2468 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2469 * netdev_n_rxq(port->netdev));
2470 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2471 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2475 /* Reconfigures the cpu mask. */
2476 ovs_numa_set_cpu_mask(cmask);
2477 free(dp->pmd_cmask);
2478 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2480 /* Restores the non-pmd. */
2481 dp_netdev_set_nonpmd(dp);
2482 /* Restores all pmd threads. */
2483 dp_netdev_reset_pmd_threads(dp);
2490 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2491 uint32_t queue_id, uint32_t *priority)
2493 *priority = queue_id;
2498 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2499 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2500 struct dp_netdev_actions *
2501 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2503 struct dp_netdev_actions *netdev_actions;
2505 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2506 memcpy(netdev_actions->actions, actions, size);
2507 netdev_actions->size = size;
2509 return netdev_actions;
2512 struct dp_netdev_actions *
2513 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2515 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2519 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2524 static inline unsigned long long
2525 cycles_counter(void)
2528 return rte_get_tsc_cycles();
2534 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2535 extern struct ovs_mutex cycles_counter_fake_mutex;
2537 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2539 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2540 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2541 OVS_NO_THREAD_SAFETY_ANALYSIS
2543 pmd->last_cycles = cycles_counter();
2546 /* Stop counting cycles and add them to the counter 'type' */
2548 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2549 enum pmd_cycles_counter_type type)
2550 OVS_RELEASES(&cycles_counter_fake_mutex)
2551 OVS_NO_THREAD_SAFETY_ANALYSIS
2553 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2555 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2559 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2560 struct dp_netdev_port *port,
2561 struct netdev_rxq *rxq)
2563 struct dp_packet *packets[NETDEV_MAX_BURST];
2566 cycles_count_start(pmd);
2567 error = netdev_rxq_recv(rxq, packets, &cnt);
2568 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2570 *recirc_depth_get() = 0;
2572 cycles_count_start(pmd);
2573 dp_netdev_input(pmd, packets, cnt, port->port_no);
2574 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2575 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2576 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2578 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2579 netdev_get_name(port->netdev), ovs_strerror(error));
2583 /* Return true if needs to revalidate datapath flows. */
2585 dpif_netdev_run(struct dpif *dpif)
2587 struct dp_netdev_port *port;
2588 struct dp_netdev *dp = get_dp_netdev(dpif);
2589 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2591 uint64_t new_tnl_seq;
2593 ovs_mutex_lock(&dp->non_pmd_mutex);
2594 CMAP_FOR_EACH (port, node, &dp->ports) {
2595 if (!netdev_is_pmd(port->netdev)) {
2598 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2599 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2603 ovs_mutex_unlock(&dp->non_pmd_mutex);
2604 dp_netdev_pmd_unref(non_pmd);
2606 tnl_neigh_cache_run();
2608 new_tnl_seq = seq_read(tnl_conf_seq);
2610 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2611 dp->last_tnl_conf_seq = new_tnl_seq;
2618 dpif_netdev_wait(struct dpif *dpif)
2620 struct dp_netdev_port *port;
2621 struct dp_netdev *dp = get_dp_netdev(dpif);
2623 ovs_mutex_lock(&dp_netdev_mutex);
2624 CMAP_FOR_EACH (port, node, &dp->ports) {
2625 if (!netdev_is_pmd(port->netdev)) {
2628 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2629 netdev_rxq_wait(port->rxq[i]);
2633 ovs_mutex_unlock(&dp_netdev_mutex);
2634 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2638 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2639 struct rxq_poll **ppoll_list, int poll_cnt)
2640 OVS_REQUIRES(pmd->poll_mutex)
2642 struct rxq_poll *poll_list = *ppoll_list;
2643 struct rxq_poll *poll;
2646 for (i = 0; i < poll_cnt; i++) {
2647 port_unref(poll_list[i].port);
2650 poll_list = xrealloc(poll_list, pmd->poll_cnt * sizeof *poll_list);
2653 LIST_FOR_EACH (poll, node, &pmd->poll_list) {
2654 port_ref(poll->port);
2655 poll_list[i++] = *poll;
2658 *ppoll_list = poll_list;
2659 return pmd->poll_cnt;
2663 pmd_thread_main(void *f_)
2665 struct dp_netdev_pmd_thread *pmd = f_;
2666 unsigned int lc = 0;
2667 struct rxq_poll *poll_list;
2668 unsigned int port_seq = PMD_INITIAL_SEQ;
2675 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2676 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2677 pmd_thread_setaffinity_cpu(pmd->core_id);
2679 emc_cache_init(&pmd->flow_cache);
2681 ovs_mutex_lock(&pmd->poll_mutex);
2682 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2683 ovs_mutex_unlock(&pmd->poll_mutex);
2685 /* List port/core affinity */
2686 for (i = 0; i < poll_cnt; i++) {
2687 VLOG_INFO("Core %d processing port \'%s\'\n", pmd->core_id,
2688 netdev_get_name(poll_list[i].port->netdev));
2691 /* Signal here to make sure the pmd finishes
2692 * reloading the updated configuration. */
2693 dp_netdev_pmd_reload_done(pmd);
2696 for (i = 0; i < poll_cnt; i++) {
2697 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2705 emc_cache_slow_sweep(&pmd->flow_cache);
2706 coverage_try_clear();
2709 atomic_read_relaxed(&pmd->change_seq, &seq);
2710 if (seq != port_seq) {
2717 emc_cache_uninit(&pmd->flow_cache);
2719 if (!latch_is_set(&pmd->exit_latch)){
2723 for (i = 0; i < poll_cnt; i++) {
2724 port_unref(poll_list[i].port);
2727 dp_netdev_pmd_reload_done(pmd);
2734 dp_netdev_disable_upcall(struct dp_netdev *dp)
2735 OVS_ACQUIRES(dp->upcall_rwlock)
2737 fat_rwlock_wrlock(&dp->upcall_rwlock);
2741 dpif_netdev_disable_upcall(struct dpif *dpif)
2742 OVS_NO_THREAD_SAFETY_ANALYSIS
2744 struct dp_netdev *dp = get_dp_netdev(dpif);
2745 dp_netdev_disable_upcall(dp);
2749 dp_netdev_enable_upcall(struct dp_netdev *dp)
2750 OVS_RELEASES(dp->upcall_rwlock)
2752 fat_rwlock_unlock(&dp->upcall_rwlock);
2756 dpif_netdev_enable_upcall(struct dpif *dpif)
2757 OVS_NO_THREAD_SAFETY_ANALYSIS
2759 struct dp_netdev *dp = get_dp_netdev(dpif);
2760 dp_netdev_enable_upcall(dp);
2764 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2766 ovs_mutex_lock(&pmd->cond_mutex);
2767 xpthread_cond_signal(&pmd->cond);
2768 ovs_mutex_unlock(&pmd->cond_mutex);
2771 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2772 * the pointer if succeeds, otherwise, NULL.
2774 * Caller must unrefs the returned reference. */
2775 static struct dp_netdev_pmd_thread *
2776 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2778 struct dp_netdev_pmd_thread *pmd;
2779 const struct cmap_node *pnode;
2781 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2785 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2787 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2790 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2792 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2794 struct dp_netdev_pmd_thread *non_pmd;
2796 non_pmd = xzalloc(sizeof *non_pmd);
2797 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2801 /* Caller must have valid pointer to 'pmd'. */
2803 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2805 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2809 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2811 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2812 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2816 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2817 * fails, keeps checking for next node until reaching the end of cmap.
2819 * Caller must unrefs the returned reference. */
2820 static struct dp_netdev_pmd_thread *
2821 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2823 struct dp_netdev_pmd_thread *next;
2826 struct cmap_node *node;
2828 node = cmap_next_position(&dp->poll_threads, pos);
2829 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2831 } while (next && !dp_netdev_pmd_try_ref(next));
2836 /* Configures the 'pmd' based on the input argument. */
2838 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2839 int index, unsigned core_id, int numa_id)
2843 pmd->core_id = core_id;
2844 pmd->numa_id = numa_id;
2847 atomic_init(&pmd->tx_qid,
2848 (core_id == NON_PMD_CORE_ID)
2849 ? ovs_numa_get_n_cores()
2850 : get_n_pmd_threads(dp));
2852 ovs_refcount_init(&pmd->ref_cnt);
2853 latch_init(&pmd->exit_latch);
2854 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2855 xpthread_cond_init(&pmd->cond, NULL);
2856 ovs_mutex_init(&pmd->cond_mutex);
2857 ovs_mutex_init(&pmd->flow_mutex);
2858 ovs_mutex_init(&pmd->poll_mutex);
2859 dpcls_init(&pmd->cls);
2860 cmap_init(&pmd->flow_table);
2861 list_init(&pmd->poll_list);
2862 /* init the 'flow_cache' since there is no
2863 * actual thread created for NON_PMD_CORE_ID. */
2864 if (core_id == NON_PMD_CORE_ID) {
2865 emc_cache_init(&pmd->flow_cache);
2867 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2868 hash_int(core_id, 0));
2872 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2874 dp_netdev_pmd_flow_flush(pmd);
2875 dpcls_destroy(&pmd->cls);
2876 cmap_destroy(&pmd->flow_table);
2877 ovs_mutex_destroy(&pmd->flow_mutex);
2878 latch_destroy(&pmd->exit_latch);
2879 xpthread_cond_destroy(&pmd->cond);
2880 ovs_mutex_destroy(&pmd->cond_mutex);
2881 ovs_mutex_destroy(&pmd->poll_mutex);
2885 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2886 * and unrefs the struct. */
2888 dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd)
2890 struct rxq_poll *poll;
2892 /* Uninit the 'flow_cache' since there is
2893 * no actual thread uninit it for NON_PMD_CORE_ID. */
2894 if (pmd->core_id == NON_PMD_CORE_ID) {
2895 emc_cache_uninit(&pmd->flow_cache);
2897 latch_set(&pmd->exit_latch);
2898 dp_netdev_reload_pmd__(pmd);
2899 ovs_numa_unpin_core(pmd->core_id);
2900 xpthread_join(pmd->thread, NULL);
2903 /* Unref all ports and free poll_list. */
2904 LIST_FOR_EACH_POP (poll, node, &pmd->poll_list) {
2905 port_unref(poll->port);
2909 /* Purges the 'pmd''s flows after stopping the thread, but before
2910 * destroying the flows, so that the flow stats can be collected. */
2911 if (dp->dp_purge_cb) {
2912 dp->dp_purge_cb(dp->dp_purge_aux, pmd->core_id);
2914 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2915 dp_netdev_pmd_unref(pmd);
2918 /* Destroys all pmd threads. */
2920 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2922 struct dp_netdev_pmd_thread *pmd;
2923 struct dp_netdev_pmd_thread **pmd_list;
2924 size_t k = 0, n_pmds;
2926 n_pmds = cmap_count(&dp->poll_threads);
2927 pmd_list = xcalloc(n_pmds, sizeof *pmd_list);
2929 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2930 /* We cannot call dp_netdev_del_pmd(), since it alters
2931 * 'dp->poll_threads' (while we're iterating it) and it
2933 ovs_assert(k < n_pmds);
2934 pmd_list[k++] = pmd;
2937 for (size_t i = 0; i < k; i++) {
2938 dp_netdev_del_pmd(dp, pmd_list[i]);
2943 /* Deletes all pmd threads on numa node 'numa_id' and
2944 * fixes tx_qids of other threads to keep them sequential. */
2946 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2948 struct dp_netdev_pmd_thread *pmd;
2949 int n_pmds_on_numa, n_pmds;
2950 int *free_idx, k = 0;
2951 struct dp_netdev_pmd_thread **pmd_list;
2953 n_pmds_on_numa = get_n_pmd_threads_on_numa(dp, numa_id);
2954 free_idx = xcalloc(n_pmds_on_numa, sizeof *free_idx);
2955 pmd_list = xcalloc(n_pmds_on_numa, sizeof *pmd_list);
2957 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2958 /* We cannot call dp_netdev_del_pmd(), since it alters
2959 * 'dp->poll_threads' (while we're iterating it) and it
2961 if (pmd->numa_id == numa_id) {
2962 atomic_read_relaxed(&pmd->tx_qid, &free_idx[k]);
2964 ovs_assert(k < n_pmds_on_numa);
2969 for (int i = 0; i < k; i++) {
2970 dp_netdev_del_pmd(dp, pmd_list[i]);
2973 n_pmds = get_n_pmd_threads(dp);
2974 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2977 atomic_read_relaxed(&pmd->tx_qid, &old_tx_qid);
2979 if (old_tx_qid >= n_pmds) {
2980 int new_tx_qid = free_idx[--k];
2982 atomic_store_relaxed(&pmd->tx_qid, new_tx_qid);
2990 /* Returns PMD thread from this numa node with fewer rx queues to poll.
2991 * Returns NULL if there is no PMD threads on this numa node.
2992 * Can be called safely only by main thread. */
2993 static struct dp_netdev_pmd_thread *
2994 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id)
2997 struct dp_netdev_pmd_thread *pmd, *res = NULL;
2999 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3000 if (pmd->numa_id == numa_id
3001 && (min_cnt > pmd->poll_cnt || res == NULL)) {
3002 min_cnt = pmd->poll_cnt;
3010 /* Adds rx queue to poll_list of PMD thread. */
3012 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
3013 struct dp_netdev_port *port, struct netdev_rxq *rx)
3014 OVS_REQUIRES(pmd->poll_mutex)
3016 struct rxq_poll *poll = xmalloc(sizeof *poll);
3022 list_push_back(&pmd->poll_list, &poll->node);
3026 /* Checks the numa node id of 'netdev' and starts pmd threads for
3029 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
3033 if (!ovs_numa_numa_id_is_valid(numa_id)) {
3034 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
3035 "invalid", numa_id);
3039 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
3041 /* If there are already pmd threads created for the numa node
3042 * in which 'netdev' is on, do nothing. Else, creates the
3043 * pmd threads for the numa node. */
3045 int can_have, n_unpinned, i, index = 0;
3046 struct dp_netdev_pmd_thread **pmds;
3047 struct dp_netdev_port *port;
3049 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
3051 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
3052 "cores on numa node");
3056 /* If cpu mask is specified, uses all unpinned cores, otherwise
3057 * tries creating NR_PMD_THREADS pmd threads. */
3058 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
3059 pmds = xzalloc(can_have * sizeof *pmds);
3060 for (i = 0; i < can_have; i++) {
3061 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
3062 pmds[i] = xzalloc(sizeof **pmds);
3063 dp_netdev_configure_pmd(pmds[i], dp, i, core_id, numa_id);
3066 /* Distributes rx queues of this numa node between new pmd threads. */
3067 CMAP_FOR_EACH (port, node, &dp->ports) {
3068 if (netdev_is_pmd(port->netdev)
3069 && netdev_get_numa_id(port->netdev) == numa_id) {
3070 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
3071 /* Make thread-safety analyser happy. */
3072 ovs_mutex_lock(&pmds[index]->poll_mutex);
3073 dp_netdev_add_rxq_to_pmd(pmds[index], port, port->rxq[i]);
3074 ovs_mutex_unlock(&pmds[index]->poll_mutex);
3075 index = (index + 1) % can_have;
3080 /* Actual start of pmd threads. */
3081 for (i = 0; i < can_have; i++) {
3082 pmds[i]->thread = ovs_thread_create("pmd", pmd_thread_main, pmds[i]);
3085 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
3090 /* Called after pmd threads config change. Restarts pmd threads with
3091 * new configuration. */
3093 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
3095 struct dp_netdev_port *port;
3097 CMAP_FOR_EACH (port, node, &dp->ports) {
3098 if (netdev_is_pmd(port->netdev)) {
3099 int numa_id = netdev_get_numa_id(port->netdev);
3101 dp_netdev_set_pmds_on_numa(dp, numa_id);
3107 dpif_netdev_get_datapath_version(void)
3109 return xstrdup("<built-in>");
3113 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
3114 uint16_t tcp_flags, long long now)
3118 atomic_store_relaxed(&netdev_flow->stats.used, now);
3119 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
3120 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
3121 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
3123 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
3127 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
3128 enum dp_stat_type type, int cnt)
3130 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3134 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3135 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3136 enum dpif_upcall_type type, const struct nlattr *userdata,
3137 struct ofpbuf *actions, struct ofpbuf *put_actions)
3139 struct dp_netdev *dp = pmd->dp;
3140 struct flow_tnl orig_tunnel;
3143 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3147 /* Upcall processing expects the Geneve options to be in the translated
3148 * format but we need to retain the raw format for datapath use. */
3149 orig_tunnel.flags = flow->tunnel.flags;
3150 if (flow->tunnel.flags & FLOW_TNL_F_UDPIF) {
3151 orig_tunnel.metadata.present.len = flow->tunnel.metadata.present.len;
3152 memcpy(orig_tunnel.metadata.opts.gnv, flow->tunnel.metadata.opts.gnv,
3153 flow->tunnel.metadata.present.len);
3154 err = tun_metadata_from_geneve_udpif(&orig_tunnel, &orig_tunnel,
3161 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3162 struct ds ds = DS_EMPTY_INITIALIZER;
3165 struct odp_flow_key_parms odp_parms = {
3168 .odp_in_port = flow->in_port.odp_port,
3169 .support = dp_netdev_support,
3172 ofpbuf_init(&key, 0);
3173 odp_flow_key_from_flow(&odp_parms, &key);
3174 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3175 dp_packet_size(packet_));
3177 odp_flow_key_format(key.data, key.size, &ds);
3179 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3180 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3182 ofpbuf_uninit(&key);
3188 err = dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3189 actions, wc, put_actions, dp->upcall_aux);
3190 if (err && err != ENOSPC) {
3194 /* Translate tunnel metadata masks to datapath format. */
3196 if (wc->masks.tunnel.metadata.present.map) {
3197 struct geneve_opt opts[TLV_TOT_OPT_SIZE /
3198 sizeof(struct geneve_opt)];
3200 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3201 tun_metadata_to_geneve_udpif_mask(&flow->tunnel,
3203 orig_tunnel.metadata.opts.gnv,
3204 orig_tunnel.metadata.present.len,
3207 orig_tunnel.metadata.present.len = 0;
3210 memset(&wc->masks.tunnel.metadata, 0,
3211 sizeof wc->masks.tunnel.metadata);
3212 memcpy(&wc->masks.tunnel.metadata.opts.gnv, opts,
3213 orig_tunnel.metadata.present.len);
3215 wc->masks.tunnel.metadata.present.len = 0xff;
3218 /* Restore tunnel metadata. We need to use the saved options to ensure
3219 * that any unknown options are not lost. The generated mask will have
3220 * the same structure, matching on types and lengths but wildcarding
3221 * option data we don't care about. */
3222 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3223 memcpy(&flow->tunnel.metadata.opts.gnv, orig_tunnel.metadata.opts.gnv,
3224 orig_tunnel.metadata.present.len);
3225 flow->tunnel.metadata.present.len = orig_tunnel.metadata.present.len;
3226 flow->tunnel.flags |= FLOW_TNL_F_UDPIF;
3232 static inline uint32_t
3233 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3234 const struct miniflow *mf)
3236 uint32_t hash, recirc_depth;
3238 if (OVS_LIKELY(dp_packet_rss_valid(packet))) {
3239 hash = dp_packet_get_rss_hash(packet);
3241 hash = miniflow_hash_5tuple(mf, 0);
3242 dp_packet_set_rss_hash(packet, hash);
3245 /* The RSS hash must account for the recirculation depth to avoid
3246 * collisions in the exact match cache */
3247 recirc_depth = *recirc_depth_get_unsafe();
3248 if (OVS_UNLIKELY(recirc_depth)) {
3249 hash = hash_finish(hash, recirc_depth);
3250 dp_packet_set_rss_hash(packet, hash);
3255 struct packet_batch {
3256 unsigned int packet_count;
3257 unsigned int byte_count;
3260 struct dp_netdev_flow *flow;
3262 struct dp_packet *packets[NETDEV_MAX_BURST];
3266 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
3267 const struct miniflow *mf)
3269 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3270 batch->packets[batch->packet_count++] = packet;
3271 batch->byte_count += dp_packet_size(packet);
3275 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
3277 flow->batch = batch;
3280 batch->packet_count = 0;
3281 batch->byte_count = 0;
3282 batch->tcp_flags = 0;
3286 packet_batch_execute(struct packet_batch *batch,
3287 struct dp_netdev_pmd_thread *pmd,
3290 struct dp_netdev_actions *actions;
3291 struct dp_netdev_flow *flow = batch->flow;
3293 dp_netdev_flow_used(flow, batch->packet_count, batch->byte_count,
3294 batch->tcp_flags, now);
3296 actions = dp_netdev_flow_get_actions(flow);
3298 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
3299 actions->actions, actions->size);
3303 dp_netdev_queue_batches(struct dp_packet *pkt,
3304 struct dp_netdev_flow *flow, const struct miniflow *mf,
3305 struct packet_batch *batches, size_t *n_batches)
3307 struct packet_batch *batch = flow->batch;
3309 if (OVS_UNLIKELY(!batch)) {
3310 batch = &batches[(*n_batches)++];
3311 packet_batch_init(batch, flow);
3314 packet_batch_update(batch, pkt, mf);
3317 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3318 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
3319 * miniflow is copied into 'keys' and the packet pointer is moved at the
3320 * beginning of the 'packets' array.
3322 * The function returns the number of packets that needs to be processed in the
3323 * 'packets' array (they have been moved to the beginning of the vector).
3325 * If 'md_is_valid' is false, the metadata in 'packets' is not valid and must be
3326 * initialized by this function using 'port_no'.
3328 static inline size_t
3329 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
3330 size_t cnt, struct netdev_flow_key *keys,
3331 struct packet_batch batches[], size_t *n_batches,
3332 bool md_is_valid, odp_port_t port_no)
3334 struct emc_cache *flow_cache = &pmd->flow_cache;
3335 struct netdev_flow_key *key = &keys[0];
3336 size_t i, n_missed = 0, n_dropped = 0;
3338 for (i = 0; i < cnt; i++) {
3339 struct dp_netdev_flow *flow;
3340 struct dp_packet *packet = packets[i];
3342 if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) {
3343 dp_packet_delete(packet);
3349 /* Prefetch next packet data and metadata. */
3350 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3351 pkt_metadata_prefetch_init(&packets[i+1]->md);
3355 pkt_metadata_init(&packet->md, port_no);
3357 miniflow_extract(packet, &key->mf);
3358 key->len = 0; /* Not computed yet. */
3359 key->hash = dpif_netdev_packet_get_rss_hash(packet, &key->mf);
3361 flow = emc_lookup(flow_cache, key);
3362 if (OVS_LIKELY(flow)) {
3363 dp_netdev_queue_batches(packet, flow, &key->mf, batches,
3366 /* Exact match cache missed. Group missed packets together at
3367 * the beginning of the 'packets' array. */
3368 packets[n_missed] = packet;
3369 /* 'key[n_missed]' contains the key of the current packet and it
3370 * must be returned to the caller. The next key should be extracted
3371 * to 'keys[n_missed + 1]'. */
3372 key = &keys[++n_missed];
3376 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - n_dropped - n_missed);
3382 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3383 struct dp_packet **packets, size_t cnt,
3384 struct netdev_flow_key *keys,
3385 struct packet_batch batches[], size_t *n_batches)
3387 #if !defined(__CHECKER__) && !defined(_WIN32)
3388 const size_t PKT_ARRAY_SIZE = cnt;
3390 /* Sparse or MSVC doesn't like variable length array. */
3391 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3393 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3394 struct dp_netdev *dp = pmd->dp;
3395 struct emc_cache *flow_cache = &pmd->flow_cache;
3396 int miss_cnt = 0, lost_cnt = 0;
3400 for (i = 0; i < cnt; i++) {
3401 /* Key length is needed in all the cases, hash computed on demand. */
3402 keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf));
3404 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3405 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3406 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3407 struct ofpbuf actions, put_actions;
3410 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3411 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3413 for (i = 0; i < cnt; i++) {
3414 struct dp_netdev_flow *netdev_flow;
3415 struct ofpbuf *add_actions;
3419 if (OVS_LIKELY(rules[i])) {
3423 /* It's possible that an earlier slow path execution installed
3424 * a rule covering this flow. In this case, it's a lot cheaper
3425 * to catch it here than execute a miss. */
3426 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3428 rules[i] = &netdev_flow->cr;
3434 match.tun_md.valid = false;
3435 miniflow_expand(&keys[i].mf, &match.flow);
3437 ofpbuf_clear(&actions);
3438 ofpbuf_clear(&put_actions);
3440 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
3441 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
3442 &ufid, DPIF_UC_MISS, NULL, &actions,
3444 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3445 dp_packet_delete(packets[i]);
3450 /* The Netlink encoding of datapath flow keys cannot express
3451 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3452 * tag is interpreted as exact match on the fact that there is no
3453 * VLAN. Unless we refactor a lot of code that translates between
3454 * Netlink and struct flow representations, we have to do the same
3456 if (!match.wc.masks.vlan_tci) {
3457 match.wc.masks.vlan_tci = htons(0xffff);
3460 /* We can't allow the packet batching in the next loop to execute
3461 * the actions. Otherwise, if there are any slow path actions,
3462 * we'll send the packet up twice. */
3463 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
3464 actions.data, actions.size);
3466 add_actions = put_actions.size ? &put_actions : &actions;
3467 if (OVS_LIKELY(error != ENOSPC)) {
3468 /* XXX: There's a race window where a flow covering this packet
3469 * could have already been installed since we last did the flow
3470 * lookup before upcall. This could be solved by moving the
3471 * mutex lock outside the loop, but that's an awful long time
3472 * to be locking everyone out of making flow installs. If we
3473 * move to a per-core classifier, it would be reasonable. */
3474 ovs_mutex_lock(&pmd->flow_mutex);
3475 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3476 if (OVS_LIKELY(!netdev_flow)) {
3477 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3481 ovs_mutex_unlock(&pmd->flow_mutex);
3483 emc_insert(flow_cache, &keys[i], netdev_flow);
3487 ofpbuf_uninit(&actions);
3488 ofpbuf_uninit(&put_actions);
3489 fat_rwlock_unlock(&dp->upcall_rwlock);
3490 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3491 } else if (OVS_UNLIKELY(any_miss)) {
3492 for (i = 0; i < cnt; i++) {
3493 if (OVS_UNLIKELY(!rules[i])) {
3494 dp_packet_delete(packets[i]);
3501 for (i = 0; i < cnt; i++) {
3502 struct dp_packet *packet = packets[i];
3503 struct dp_netdev_flow *flow;
3505 if (OVS_UNLIKELY(!rules[i])) {
3509 flow = dp_netdev_flow_cast(rules[i]);
3511 emc_insert(flow_cache, &keys[i], flow);
3512 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3515 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3516 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3517 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3520 /* Packets enter the datapath from a port (or from recirculation) here.
3522 * For performance reasons a caller may choose not to initialize the metadata
3523 * in 'packets': in this case 'mdinit' is false and this function needs to
3524 * initialize it using 'port_no'. If the metadata in 'packets' is already
3525 * valid, 'md_is_valid' must be true and 'port_no' will be ignored. */
3527 dp_netdev_input__(struct dp_netdev_pmd_thread *pmd,
3528 struct dp_packet **packets, int cnt,
3529 bool md_is_valid, odp_port_t port_no)
3531 #if !defined(__CHECKER__) && !defined(_WIN32)
3532 const size_t PKT_ARRAY_SIZE = cnt;
3534 /* Sparse or MSVC doesn't like variable length array. */
3535 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3537 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3538 struct packet_batch batches[PKT_ARRAY_SIZE];
3539 long long now = time_msec();
3540 size_t newcnt, n_batches, i;
3543 newcnt = emc_processing(pmd, packets, cnt, keys, batches, &n_batches,
3544 md_is_valid, port_no);
3545 if (OVS_UNLIKELY(newcnt)) {
3546 fast_path_processing(pmd, packets, newcnt, keys, batches, &n_batches);
3549 for (i = 0; i < n_batches; i++) {
3550 batches[i].flow->batch = NULL;
3553 for (i = 0; i < n_batches; i++) {
3554 packet_batch_execute(&batches[i], pmd, now);
3559 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3560 struct dp_packet **packets, int cnt,
3563 dp_netdev_input__(pmd, packets, cnt, false, port_no);
3567 dp_netdev_recirculate(struct dp_netdev_pmd_thread *pmd,
3568 struct dp_packet **packets, int cnt)
3570 dp_netdev_input__(pmd, packets, cnt, true, 0);
3573 struct dp_netdev_execute_aux {
3574 struct dp_netdev_pmd_thread *pmd;
3578 dpif_netdev_register_dp_purge_cb(struct dpif *dpif, dp_purge_callback *cb,
3581 struct dp_netdev *dp = get_dp_netdev(dpif);
3582 dp->dp_purge_aux = aux;
3583 dp->dp_purge_cb = cb;
3587 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3590 struct dp_netdev *dp = get_dp_netdev(dpif);
3591 dp->upcall_aux = aux;
3596 dp_netdev_drop_packets(struct dp_packet **packets, int cnt, bool may_steal)
3601 for (i = 0; i < cnt; i++) {
3602 dp_packet_delete(packets[i]);
3608 push_tnl_action(const struct dp_netdev *dp,
3609 const struct nlattr *attr,
3610 struct dp_packet **packets, int cnt)
3612 struct dp_netdev_port *tun_port;
3613 const struct ovs_action_push_tnl *data;
3615 data = nl_attr_get(attr);
3617 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3621 netdev_push_header(tun_port->netdev, packets, cnt, data);
3627 dp_netdev_clone_pkt_batch(struct dp_packet **dst_pkts,
3628 struct dp_packet **src_pkts, int cnt)
3632 for (i = 0; i < cnt; i++) {
3633 dst_pkts[i] = dp_packet_clone(src_pkts[i]);
3638 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3639 const struct nlattr *a, bool may_steal)
3640 OVS_NO_THREAD_SAFETY_ANALYSIS
3642 struct dp_netdev_execute_aux *aux = aux_;
3643 uint32_t *depth = recirc_depth_get();
3644 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3645 struct dp_netdev *dp = pmd->dp;
3646 int type = nl_attr_type(a);
3647 struct dp_netdev_port *p;
3650 switch ((enum ovs_action_attr)type) {
3651 case OVS_ACTION_ATTR_OUTPUT:
3652 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3653 if (OVS_LIKELY(p)) {
3656 atomic_read_relaxed(&pmd->tx_qid, &tx_qid);
3658 netdev_send(p->netdev, tx_qid, packets, cnt, may_steal);
3663 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3664 if (*depth < MAX_RECIRC_DEPTH) {
3665 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3669 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3673 err = push_tnl_action(dp, a, packets, cnt);
3676 dp_netdev_recirculate(pmd, packets, cnt);
3679 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3685 case OVS_ACTION_ATTR_TUNNEL_POP:
3686 if (*depth < MAX_RECIRC_DEPTH) {
3687 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3689 p = dp_netdev_lookup_port(dp, portno);
3691 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3695 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3699 err = netdev_pop_header(p->netdev, packets, cnt);
3702 for (i = 0; i < cnt; i++) {
3703 packets[i]->md.in_port.odp_port = portno;
3707 dp_netdev_recirculate(pmd, packets, cnt);
3710 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3717 case OVS_ACTION_ATTR_USERSPACE:
3718 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3719 const struct nlattr *userdata;
3720 struct ofpbuf actions;
3724 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3725 ofpbuf_init(&actions, 0);
3727 for (i = 0; i < cnt; i++) {
3730 ofpbuf_clear(&actions);
3732 flow_extract(packets[i], &flow);
3733 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3734 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3735 DPIF_UC_ACTION, userdata,&actions,
3737 if (!error || error == ENOSPC) {
3738 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3739 actions.data, actions.size);
3740 } else if (may_steal) {
3741 dp_packet_delete(packets[i]);
3744 ofpbuf_uninit(&actions);
3745 fat_rwlock_unlock(&dp->upcall_rwlock);
3751 case OVS_ACTION_ATTR_RECIRC:
3752 if (*depth < MAX_RECIRC_DEPTH) {
3753 struct dp_packet *recirc_pkts[NETDEV_MAX_BURST];
3756 dp_netdev_clone_pkt_batch(recirc_pkts, packets, cnt);
3757 packets = recirc_pkts;
3760 for (i = 0; i < cnt; i++) {
3761 packets[i]->md.recirc_id = nl_attr_get_u32(a);
3765 dp_netdev_recirculate(pmd, packets, cnt);
3771 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3774 case OVS_ACTION_ATTR_CT:
3775 /* If a flow with this action is slow-pathed, datapath assistance is
3776 * required to implement it. However, we don't support this action
3777 * in the userspace datapath. */
3778 VLOG_WARN("Cannot execute conntrack action in userspace.");
3781 case OVS_ACTION_ATTR_PUSH_VLAN:
3782 case OVS_ACTION_ATTR_POP_VLAN:
3783 case OVS_ACTION_ATTR_PUSH_MPLS:
3784 case OVS_ACTION_ATTR_POP_MPLS:
3785 case OVS_ACTION_ATTR_SET:
3786 case OVS_ACTION_ATTR_SET_MASKED:
3787 case OVS_ACTION_ATTR_SAMPLE:
3788 case OVS_ACTION_ATTR_HASH:
3789 case OVS_ACTION_ATTR_UNSPEC:
3790 case __OVS_ACTION_ATTR_MAX:
3794 dp_netdev_drop_packets(packets, cnt, may_steal);
3798 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3799 struct dp_packet **packets, int cnt,
3801 const struct nlattr *actions, size_t actions_len)
3803 struct dp_netdev_execute_aux aux = { pmd };
3805 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3806 actions_len, dp_execute_cb);
3809 const struct dpif_class dpif_netdev_class = {
3812 dpif_netdev_enumerate,
3813 dpif_netdev_port_open_type,
3816 dpif_netdev_destroy,
3819 dpif_netdev_get_stats,
3820 dpif_netdev_port_add,
3821 dpif_netdev_port_del,
3822 dpif_netdev_port_query_by_number,
3823 dpif_netdev_port_query_by_name,
3824 NULL, /* port_get_pid */
3825 dpif_netdev_port_dump_start,
3826 dpif_netdev_port_dump_next,
3827 dpif_netdev_port_dump_done,
3828 dpif_netdev_port_poll,
3829 dpif_netdev_port_poll_wait,
3830 dpif_netdev_flow_flush,
3831 dpif_netdev_flow_dump_create,
3832 dpif_netdev_flow_dump_destroy,
3833 dpif_netdev_flow_dump_thread_create,
3834 dpif_netdev_flow_dump_thread_destroy,
3835 dpif_netdev_flow_dump_next,
3836 dpif_netdev_operate,
3837 NULL, /* recv_set */
3838 NULL, /* handlers_set */
3839 dpif_netdev_pmd_set,
3840 dpif_netdev_queue_to_priority,
3842 NULL, /* recv_wait */
3843 NULL, /* recv_purge */
3844 dpif_netdev_register_dp_purge_cb,
3845 dpif_netdev_register_upcall_cb,
3846 dpif_netdev_enable_upcall,
3847 dpif_netdev_disable_upcall,
3848 dpif_netdev_get_datapath_version,
3849 NULL, /* ct_dump_start */
3850 NULL, /* ct_dump_next */
3851 NULL, /* ct_dump_done */
3852 NULL, /* ct_flush */
3856 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3857 const char *argv[], void *aux OVS_UNUSED)
3859 struct dp_netdev_port *old_port;
3860 struct dp_netdev_port *new_port;
3861 struct dp_netdev *dp;
3864 ovs_mutex_lock(&dp_netdev_mutex);
3865 dp = shash_find_data(&dp_netdevs, argv[1]);
3866 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3867 ovs_mutex_unlock(&dp_netdev_mutex);
3868 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3871 ovs_refcount_ref(&dp->ref_cnt);
3872 ovs_mutex_unlock(&dp_netdev_mutex);
3874 ovs_mutex_lock(&dp->port_mutex);
3875 if (get_port_by_name(dp, argv[2], &old_port)) {
3876 unixctl_command_reply_error(conn, "unknown port");
3880 port_no = u32_to_odp(atoi(argv[3]));
3881 if (!port_no || port_no == ODPP_NONE) {
3882 unixctl_command_reply_error(conn, "bad port number");
3885 if (dp_netdev_lookup_port(dp, port_no)) {
3886 unixctl_command_reply_error(conn, "port number already in use");
3890 /* Remove old port. */
3891 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3892 ovsrcu_postpone(free, old_port);
3894 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3895 new_port = xmemdup(old_port, sizeof *old_port);
3896 new_port->port_no = port_no;
3897 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3899 seq_change(dp->port_seq);
3900 unixctl_command_reply(conn, NULL);
3903 ovs_mutex_unlock(&dp->port_mutex);
3904 dp_netdev_unref(dp);
3908 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3909 const char *argv[], void *aux OVS_UNUSED)
3911 struct dp_netdev_port *port;
3912 struct dp_netdev *dp;
3914 ovs_mutex_lock(&dp_netdev_mutex);
3915 dp = shash_find_data(&dp_netdevs, argv[1]);
3916 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3917 ovs_mutex_unlock(&dp_netdev_mutex);
3918 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3921 ovs_refcount_ref(&dp->ref_cnt);
3922 ovs_mutex_unlock(&dp_netdev_mutex);
3924 ovs_mutex_lock(&dp->port_mutex);
3925 if (get_port_by_name(dp, argv[2], &port)) {
3926 unixctl_command_reply_error(conn, "unknown port");
3927 } else if (port->port_no == ODPP_LOCAL) {
3928 unixctl_command_reply_error(conn, "can't delete local port");
3930 do_del_port(dp, port);
3931 unixctl_command_reply(conn, NULL);
3933 ovs_mutex_unlock(&dp->port_mutex);
3935 dp_netdev_unref(dp);
3939 dpif_dummy_register__(const char *type)
3941 struct dpif_class *class;
3943 class = xmalloc(sizeof *class);
3944 *class = dpif_netdev_class;
3945 class->type = xstrdup(type);
3946 dp_register_provider(class);
3950 dpif_dummy_override(const char *type)
3955 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
3956 * a userland-only build. It's useful for testsuite.
3958 error = dp_unregister_provider(type);
3959 if (error == 0 || error == EAFNOSUPPORT) {
3960 dpif_dummy_register__(type);
3965 dpif_dummy_register(enum dummy_level level)
3967 if (level == DUMMY_OVERRIDE_ALL) {
3972 dp_enumerate_types(&types);
3973 SSET_FOR_EACH (type, &types) {
3974 dpif_dummy_override(type);
3976 sset_destroy(&types);
3977 } else if (level == DUMMY_OVERRIDE_SYSTEM) {
3978 dpif_dummy_override("system");
3981 dpif_dummy_register__("dummy");
3983 unixctl_command_register("dpif-dummy/change-port-number",
3984 "dp port new-number",
3985 3, 3, dpif_dummy_change_port_number, NULL);
3986 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3987 2, 2, dpif_dummy_delete_port, NULL);
3990 /* Datapath Classifier. */
3992 /* A set of rules that all have the same fields wildcarded. */
3993 struct dpcls_subtable {
3994 /* The fields are only used by writers. */
3995 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3997 /* These fields are accessed by readers. */
3998 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3999 struct netdev_flow_key mask; /* Wildcards for fields (const). */
4000 /* 'mask' must be the last field, additional space is allocated here. */
4003 /* Initializes 'cls' as a classifier that initially contains no classification
4006 dpcls_init(struct dpcls *cls)
4008 cmap_init(&cls->subtables_map);
4009 pvector_init(&cls->subtables);
4013 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
4015 pvector_remove(&cls->subtables, subtable);
4016 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
4017 subtable->mask.hash);
4018 cmap_destroy(&subtable->rules);
4019 ovsrcu_postpone(free, subtable);
4022 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
4023 * caller's responsibility.
4024 * May only be called after all the readers have been terminated. */
4026 dpcls_destroy(struct dpcls *cls)
4029 struct dpcls_subtable *subtable;
4031 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
4032 ovs_assert(cmap_count(&subtable->rules) == 0);
4033 dpcls_destroy_subtable(cls, subtable);
4035 cmap_destroy(&cls->subtables_map);
4036 pvector_destroy(&cls->subtables);
4040 static struct dpcls_subtable *
4041 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
4043 struct dpcls_subtable *subtable;
4045 /* Need to add one. */
4046 subtable = xmalloc(sizeof *subtable
4047 - sizeof subtable->mask.mf + mask->len);
4048 cmap_init(&subtable->rules);
4049 netdev_flow_key_clone(&subtable->mask, mask);
4050 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
4051 pvector_insert(&cls->subtables, subtable, 0);
4052 pvector_publish(&cls->subtables);
4057 static inline struct dpcls_subtable *
4058 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
4060 struct dpcls_subtable *subtable;
4062 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
4063 &cls->subtables_map) {
4064 if (netdev_flow_key_equal(&subtable->mask, mask)) {
4068 return dpcls_create_subtable(cls, mask);
4071 /* Insert 'rule' into 'cls'. */
4073 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
4074 const struct netdev_flow_key *mask)
4076 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
4078 rule->mask = &subtable->mask;
4079 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
4082 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
4084 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
4086 struct dpcls_subtable *subtable;
4088 ovs_assert(rule->mask);
4090 INIT_CONTAINER(subtable, rule->mask, mask);
4092 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
4094 dpcls_destroy_subtable(cls, subtable);
4095 pvector_publish(&cls->subtables);
4099 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
4100 * in 'mask' the values in 'key' and 'target' are the same. */
4102 dpcls_rule_matches_key(const struct dpcls_rule *rule,
4103 const struct netdev_flow_key *target)
4105 const uint64_t *keyp = miniflow_get_values(&rule->flow.mf);
4106 const uint64_t *maskp = miniflow_get_values(&rule->mask->mf);
4109 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, target, rule->flow.mf.map) {
4110 if (OVS_UNLIKELY((value & *maskp++) != *keyp++)) {
4117 /* For each miniflow in 'flows' performs a classifier lookup writing the result
4118 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
4119 * NULL it is skipped.
4121 * This function is optimized for use in the userspace datapath and therefore
4122 * does not implement a lot of features available in the standard
4123 * classifier_lookup() function. Specifically, it does not implement
4124 * priorities, instead returning any rule which matches the flow.
4126 * Returns true if all flows found a corresponding rule. */
4128 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
4129 struct dpcls_rule **rules, const size_t cnt)
4131 /* The batch size 16 was experimentally found faster than 8 or 32. */
4132 typedef uint16_t map_type;
4133 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
4135 #if !defined(__CHECKER__) && !defined(_WIN32)
4136 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
4138 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
4140 map_type maps[N_MAPS];
4141 struct dpcls_subtable *subtable;
4143 memset(maps, 0xff, sizeof maps);
4144 if (cnt % MAP_BITS) {
4145 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
4147 memset(rules, 0, cnt * sizeof *rules);
4149 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
4150 const struct netdev_flow_key *mkeys = keys;
4151 struct dpcls_rule **mrules = rules;
4152 map_type remains = 0;
4155 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
4157 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
4158 uint32_t hashes[MAP_BITS];
4159 const struct cmap_node *nodes[MAP_BITS];
4160 unsigned long map = maps[m];
4164 continue; /* Skip empty maps. */
4167 /* Compute hashes for the remaining keys. */
4168 ULLONG_FOR_EACH_1(i, map) {
4169 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
4173 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
4174 /* Check results. */
4175 ULLONG_FOR_EACH_1(i, map) {
4176 struct dpcls_rule *rule;
4178 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
4179 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
4184 ULLONG_SET0(map, i); /* Did not match. */
4186 ; /* Keep Sparse happy. */
4188 maps[m] &= ~map; /* Clear the found rules. */
4192 return true; /* All found. */
4195 return false; /* Some misses. */