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 /* Number of rx queues for each dpdk interface and the cpu mask
225 * for pin of pmd threads. */
228 uint64_t last_tnl_conf_seq;
231 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
235 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
236 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
237 DP_STAT_MISS, /* Packets that did not match. */
238 DP_STAT_LOST, /* Packets not passed up to the client. */
242 enum pmd_cycles_counter_type {
243 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
244 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
248 /* A port in a netdev-based datapath. */
249 struct dp_netdev_port {
251 struct netdev *netdev;
252 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
253 struct netdev_saved_flags *sf;
254 struct netdev_rxq **rxq;
255 struct ovs_refcount ref_cnt;
256 char *type; /* Port type as requested by user. */
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);
869 dp->n_dpdk_rxqs = NR_QUEUE;
871 ovs_mutex_lock(&dp->port_mutex);
872 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
873 ovs_mutex_unlock(&dp->port_mutex);
879 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
885 dpif_netdev_open(const struct dpif_class *class, const char *name,
886 bool create, struct dpif **dpifp)
888 struct dp_netdev *dp;
891 ovs_mutex_lock(&dp_netdev_mutex);
892 dp = shash_find_data(&dp_netdevs, name);
894 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
896 error = (dp->class != class ? EINVAL
901 *dpifp = create_dpif_netdev(dp);
904 ovs_mutex_unlock(&dp_netdev_mutex);
910 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
911 OVS_NO_THREAD_SAFETY_ANALYSIS
913 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
914 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
916 /* Before freeing a lock we should release it */
917 fat_rwlock_unlock(&dp->upcall_rwlock);
918 fat_rwlock_destroy(&dp->upcall_rwlock);
921 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
922 * through the 'dp_netdevs' shash while freeing 'dp'. */
924 dp_netdev_free(struct dp_netdev *dp)
925 OVS_REQUIRES(dp_netdev_mutex)
927 struct dp_netdev_port *port;
929 shash_find_and_delete(&dp_netdevs, dp->name);
931 dp_netdev_destroy_all_pmds(dp);
932 ovs_mutex_destroy(&dp->non_pmd_mutex);
933 ovsthread_key_delete(dp->per_pmd_key);
935 ovs_mutex_lock(&dp->port_mutex);
936 CMAP_FOR_EACH (port, node, &dp->ports) {
937 /* PMD threads are destroyed here. do_del_port() cannot quiesce */
938 do_del_port(dp, port);
940 ovs_mutex_unlock(&dp->port_mutex);
941 cmap_destroy(&dp->poll_threads);
943 seq_destroy(dp->port_seq);
944 cmap_destroy(&dp->ports);
946 /* Upcalls must be disabled at this point */
947 dp_netdev_destroy_upcall_lock(dp);
950 free(CONST_CAST(char *, dp->name));
955 dp_netdev_unref(struct dp_netdev *dp)
958 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
959 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
960 ovs_mutex_lock(&dp_netdev_mutex);
961 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
964 ovs_mutex_unlock(&dp_netdev_mutex);
969 dpif_netdev_close(struct dpif *dpif)
971 struct dp_netdev *dp = get_dp_netdev(dpif);
978 dpif_netdev_destroy(struct dpif *dpif)
980 struct dp_netdev *dp = get_dp_netdev(dpif);
982 if (!atomic_flag_test_and_set(&dp->destroyed)) {
983 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
984 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
992 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
993 * load/store semantics. While the increment is not atomic, the load and
994 * store operations are, making it impossible to read inconsistent values.
996 * This is used to update thread local stats counters. */
998 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
1000 unsigned long long tmp;
1002 atomic_read_relaxed(var, &tmp);
1004 atomic_store_relaxed(var, tmp);
1008 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
1010 struct dp_netdev *dp = get_dp_netdev(dpif);
1011 struct dp_netdev_pmd_thread *pmd;
1013 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
1014 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1015 unsigned long long n;
1016 stats->n_flows += cmap_count(&pmd->flow_table);
1018 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
1020 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
1022 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
1023 stats->n_missed += n;
1024 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
1027 stats->n_masks = UINT32_MAX;
1028 stats->n_mask_hit = UINT64_MAX;
1034 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1038 if (pmd->core_id == NON_PMD_CORE_ID) {
1042 ovs_mutex_lock(&pmd->cond_mutex);
1043 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1044 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1045 ovs_mutex_unlock(&pmd->cond_mutex);
1049 hash_port_no(odp_port_t port_no)
1051 return hash_int(odp_to_u32(port_no), 0);
1055 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1057 OVS_REQUIRES(dp->port_mutex)
1059 struct netdev_saved_flags *sf;
1060 struct dp_netdev_port *port;
1061 struct netdev *netdev;
1062 enum netdev_flags flags;
1063 const char *open_type;
1067 /* Reject devices already in 'dp'. */
1068 if (!get_port_by_name(dp, devname, &port)) {
1072 /* Open and validate network device. */
1073 open_type = dpif_netdev_port_open_type(dp->class, type);
1074 error = netdev_open(devname, open_type, &netdev);
1078 /* XXX reject non-Ethernet devices */
1080 netdev_get_flags(netdev, &flags);
1081 if (flags & NETDEV_LOOPBACK) {
1082 VLOG_ERR("%s: cannot add a loopback device", devname);
1083 netdev_close(netdev);
1087 if (netdev_is_pmd(netdev)) {
1088 int n_cores = ovs_numa_get_n_cores();
1090 if (n_cores == OVS_CORE_UNSPEC) {
1091 VLOG_ERR("%s, cannot get cpu core info", devname);
1094 /* There can only be ovs_numa_get_n_cores() pmd threads,
1095 * so creates a txq for each, and one extra for the non
1097 error = netdev_set_multiq(netdev, n_cores + 1, dp->n_dpdk_rxqs);
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 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1109 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1111 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
1112 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1113 devname, ovs_strerror(errno));
1114 netdev_close(netdev);
1122 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1124 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1125 netdev_rxq_close(port->rxq[i]);
1127 netdev_close(netdev);
1135 ovs_refcount_init(&port->ref_cnt);
1136 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1138 if (netdev_is_pmd(netdev)) {
1139 int numa_id = netdev_get_numa_id(netdev);
1140 struct dp_netdev_pmd_thread *pmd;
1142 /* Cannot create pmd threads for invalid numa node. */
1143 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
1145 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1146 pmd = dp_netdev_less_loaded_pmd_on_numa(dp, numa_id);
1148 /* There is no pmd threads on this numa node. */
1149 dp_netdev_set_pmds_on_numa(dp, numa_id);
1150 /* Assigning of rx queues done. */
1154 ovs_mutex_lock(&pmd->poll_mutex);
1155 dp_netdev_add_rxq_to_pmd(pmd, port, port->rxq[i]);
1156 ovs_mutex_unlock(&pmd->poll_mutex);
1157 dp_netdev_reload_pmd__(pmd);
1160 seq_change(dp->port_seq);
1166 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1167 odp_port_t *port_nop)
1169 struct dp_netdev *dp = get_dp_netdev(dpif);
1170 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1171 const char *dpif_port;
1175 ovs_mutex_lock(&dp->port_mutex);
1176 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1177 if (*port_nop != ODPP_NONE) {
1178 port_no = *port_nop;
1179 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1181 port_no = choose_port(dp, dpif_port);
1182 error = port_no == ODPP_NONE ? EFBIG : 0;
1185 *port_nop = port_no;
1186 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1188 ovs_mutex_unlock(&dp->port_mutex);
1194 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1196 struct dp_netdev *dp = get_dp_netdev(dpif);
1199 ovs_mutex_lock(&dp->port_mutex);
1200 if (port_no == ODPP_LOCAL) {
1203 struct dp_netdev_port *port;
1205 error = get_port_by_number(dp, port_no, &port);
1207 do_del_port(dp, port);
1210 ovs_mutex_unlock(&dp->port_mutex);
1216 is_valid_port_number(odp_port_t port_no)
1218 return port_no != ODPP_NONE;
1221 static struct dp_netdev_port *
1222 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1224 struct dp_netdev_port *port;
1226 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1227 if (port->port_no == port_no) {
1235 get_port_by_number(struct dp_netdev *dp,
1236 odp_port_t port_no, struct dp_netdev_port **portp)
1238 if (!is_valid_port_number(port_no)) {
1242 *portp = dp_netdev_lookup_port(dp, port_no);
1243 return *portp ? 0 : ENOENT;
1248 port_ref(struct dp_netdev_port *port)
1251 ovs_refcount_ref(&port->ref_cnt);
1256 port_unref(struct dp_netdev_port *port)
1258 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1259 int n_rxq = netdev_n_rxq(port->netdev);
1262 netdev_close(port->netdev);
1263 netdev_restore_flags(port->sf);
1265 for (i = 0; i < n_rxq; i++) {
1266 netdev_rxq_close(port->rxq[i]);
1275 get_port_by_name(struct dp_netdev *dp,
1276 const char *devname, struct dp_netdev_port **portp)
1277 OVS_REQUIRES(dp->port_mutex)
1279 struct dp_netdev_port *port;
1281 CMAP_FOR_EACH (port, node, &dp->ports) {
1282 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1291 get_n_pmd_threads(struct dp_netdev *dp)
1293 /* There is one non pmd thread in dp->poll_threads */
1294 return cmap_count(&dp->poll_threads) - 1;
1298 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1300 struct dp_netdev_pmd_thread *pmd;
1303 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1304 if (pmd->numa_id == numa_id) {
1312 /* Returns 'true' if there is a port with pmd netdev and the netdev
1313 * is on numa node 'numa_id'. */
1315 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1317 struct dp_netdev_port *port;
1319 CMAP_FOR_EACH (port, node, &dp->ports) {
1320 if (netdev_is_pmd(port->netdev)
1321 && netdev_get_numa_id(port->netdev) == numa_id) {
1331 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1332 OVS_REQUIRES(dp->port_mutex)
1334 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1335 seq_change(dp->port_seq);
1336 if (netdev_is_pmd(port->netdev)) {
1337 int numa_id = netdev_get_numa_id(port->netdev);
1339 /* PMD threads can not be on invalid numa node. */
1340 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
1341 /* If there is no netdev on the numa node, deletes the pmd threads
1342 * for that numa. Else, deletes the queues from polling lists. */
1343 if (!has_pmd_port_for_numa(dp, numa_id)) {
1344 dp_netdev_del_pmds_on_numa(dp, numa_id);
1346 struct dp_netdev_pmd_thread *pmd;
1347 struct rxq_poll *poll, *next;
1349 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1350 if (pmd->numa_id == numa_id) {
1353 ovs_mutex_lock(&pmd->poll_mutex);
1354 LIST_FOR_EACH_SAFE (poll, next, node, &pmd->poll_list) {
1355 if (poll->port == port) {
1357 port_unref(poll->port);
1358 list_remove(&poll->node);
1363 ovs_mutex_unlock(&pmd->poll_mutex);
1365 dp_netdev_reload_pmd__(pmd);
1376 answer_port_query(const struct dp_netdev_port *port,
1377 struct dpif_port *dpif_port)
1379 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1380 dpif_port->type = xstrdup(port->type);
1381 dpif_port->port_no = port->port_no;
1385 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1386 struct dpif_port *dpif_port)
1388 struct dp_netdev *dp = get_dp_netdev(dpif);
1389 struct dp_netdev_port *port;
1392 error = get_port_by_number(dp, port_no, &port);
1393 if (!error && dpif_port) {
1394 answer_port_query(port, dpif_port);
1401 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1402 struct dpif_port *dpif_port)
1404 struct dp_netdev *dp = get_dp_netdev(dpif);
1405 struct dp_netdev_port *port;
1408 ovs_mutex_lock(&dp->port_mutex);
1409 error = get_port_by_name(dp, devname, &port);
1410 if (!error && dpif_port) {
1411 answer_port_query(port, dpif_port);
1413 ovs_mutex_unlock(&dp->port_mutex);
1419 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1421 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1425 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1427 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1428 ovsrcu_postpone(dp_netdev_flow_free, flow);
1433 dp_netdev_flow_hash(const ovs_u128 *ufid)
1435 return ufid->u32[0];
1439 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1440 struct dp_netdev_flow *flow)
1441 OVS_REQUIRES(pmd->flow_mutex)
1443 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1445 dpcls_remove(&pmd->cls, &flow->cr);
1446 flow->cr.mask = NULL; /* Accessing rule's mask after this is not safe. */
1448 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1451 dp_netdev_flow_unref(flow);
1455 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1457 struct dp_netdev_flow *netdev_flow;
1459 ovs_mutex_lock(&pmd->flow_mutex);
1460 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1461 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1463 ovs_mutex_unlock(&pmd->flow_mutex);
1467 dpif_netdev_flow_flush(struct dpif *dpif)
1469 struct dp_netdev *dp = get_dp_netdev(dpif);
1470 struct dp_netdev_pmd_thread *pmd;
1472 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1473 dp_netdev_pmd_flow_flush(pmd);
1479 struct dp_netdev_port_state {
1480 struct cmap_position position;
1485 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1487 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1492 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1493 struct dpif_port *dpif_port)
1495 struct dp_netdev_port_state *state = state_;
1496 struct dp_netdev *dp = get_dp_netdev(dpif);
1497 struct cmap_node *node;
1500 node = cmap_next_position(&dp->ports, &state->position);
1502 struct dp_netdev_port *port;
1504 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1507 state->name = xstrdup(netdev_get_name(port->netdev));
1508 dpif_port->name = state->name;
1509 dpif_port->type = port->type;
1510 dpif_port->port_no = port->port_no;
1521 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1523 struct dp_netdev_port_state *state = state_;
1530 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1532 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1533 uint64_t new_port_seq;
1536 new_port_seq = seq_read(dpif->dp->port_seq);
1537 if (dpif->last_port_seq != new_port_seq) {
1538 dpif->last_port_seq = new_port_seq;
1548 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1550 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1552 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1555 static struct dp_netdev_flow *
1556 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1558 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1561 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1563 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1566 /* netdev_flow_key utilities.
1568 * netdev_flow_key is basically a miniflow. We use these functions
1569 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1570 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1572 * - Since we are dealing exclusively with miniflows created by
1573 * miniflow_extract(), if the map is different the miniflow is different.
1574 * Therefore we can be faster by comparing the map and the miniflow in a
1576 * - These functions can be inlined by the compiler. */
1578 /* Given the number of bits set in miniflow's maps, returns the size of the
1579 * 'netdev_flow_key.mf' */
1580 static inline size_t
1581 netdev_flow_key_size(size_t flow_u64s)
1583 return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s);
1587 netdev_flow_key_equal(const struct netdev_flow_key *a,
1588 const struct netdev_flow_key *b)
1590 /* 'b->len' may be not set yet. */
1591 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1594 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1595 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1596 * generated by miniflow_extract. */
1598 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1599 const struct miniflow *mf)
1601 return !memcmp(&key->mf, mf, key->len);
1605 netdev_flow_key_clone(struct netdev_flow_key *dst,
1606 const struct netdev_flow_key *src)
1609 offsetof(struct netdev_flow_key, mf) + src->len);
1614 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1615 const struct flow *src)
1617 struct dp_packet packet;
1618 uint64_t buf_stub[512 / 8];
1620 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1621 pkt_metadata_from_flow(&packet.md, src);
1622 flow_compose(&packet, src);
1623 miniflow_extract(&packet, &dst->mf);
1624 dp_packet_uninit(&packet);
1626 dst->len = netdev_flow_key_size(miniflow_n_values(&dst->mf));
1627 dst->hash = 0; /* Not computed yet. */
1630 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1632 netdev_flow_mask_init(struct netdev_flow_key *mask,
1633 const struct match *match)
1635 uint64_t *dst = miniflow_values(&mask->mf);
1636 struct flowmap fmap;
1640 /* Only check masks that make sense for the flow. */
1641 flow_wc_map(&match->flow, &fmap);
1642 flowmap_init(&mask->mf.map);
1644 FLOWMAP_FOR_EACH_INDEX(idx, fmap) {
1645 uint64_t mask_u64 = flow_u64_value(&match->wc.masks, idx);
1648 flowmap_set(&mask->mf.map, idx, 1);
1650 hash = hash_add64(hash, mask_u64);
1656 FLOWMAP_FOR_EACH_MAP (map, mask->mf.map) {
1657 hash = hash_add64(hash, map);
1660 size_t n = dst - miniflow_get_values(&mask->mf);
1662 mask->hash = hash_finish(hash, n * 8);
1663 mask->len = netdev_flow_key_size(n);
1666 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1668 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1669 const struct flow *flow,
1670 const struct netdev_flow_key *mask)
1672 uint64_t *dst_u64 = miniflow_values(&dst->mf);
1673 const uint64_t *mask_u64 = miniflow_get_values(&mask->mf);
1677 dst->len = mask->len;
1678 dst->mf = mask->mf; /* Copy maps. */
1680 FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf.map) {
1681 *dst_u64 = value & *mask_u64++;
1682 hash = hash_add64(hash, *dst_u64++);
1684 dst->hash = hash_finish(hash,
1685 (dst_u64 - miniflow_get_values(&dst->mf)) * 8);
1688 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
1689 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
1690 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
1692 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1694 static inline uint32_t
1695 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1696 const struct netdev_flow_key *mask)
1698 const uint64_t *p = miniflow_get_values(&mask->mf);
1702 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, key, mask->mf.map) {
1703 hash = hash_add64(hash, value & *p++);
1706 return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8);
1710 emc_entry_alive(struct emc_entry *ce)
1712 return ce->flow && !ce->flow->dead;
1716 emc_clear_entry(struct emc_entry *ce)
1719 dp_netdev_flow_unref(ce->flow);
1725 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1726 const struct netdev_flow_key *key)
1728 if (ce->flow != flow) {
1730 dp_netdev_flow_unref(ce->flow);
1733 if (dp_netdev_flow_ref(flow)) {
1740 netdev_flow_key_clone(&ce->key, key);
1745 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1746 struct dp_netdev_flow *flow)
1748 struct emc_entry *to_be_replaced = NULL;
1749 struct emc_entry *current_entry;
1751 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1752 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1753 /* We found the entry with the 'mf' miniflow */
1754 emc_change_entry(current_entry, flow, NULL);
1758 /* Replacement policy: put the flow in an empty (not alive) entry, or
1759 * in the first entry where it can be */
1761 || (emc_entry_alive(to_be_replaced)
1762 && !emc_entry_alive(current_entry))
1763 || current_entry->key.hash < to_be_replaced->key.hash) {
1764 to_be_replaced = current_entry;
1767 /* We didn't find the miniflow in the cache.
1768 * The 'to_be_replaced' entry is where the new flow will be stored */
1770 emc_change_entry(to_be_replaced, flow, key);
1773 static inline struct dp_netdev_flow *
1774 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1776 struct emc_entry *current_entry;
1778 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1779 if (current_entry->key.hash == key->hash
1780 && emc_entry_alive(current_entry)
1781 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1783 /* We found the entry with the 'key->mf' miniflow */
1784 return current_entry->flow;
1791 static struct dp_netdev_flow *
1792 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1793 const struct netdev_flow_key *key)
1795 struct dp_netdev_flow *netdev_flow;
1796 struct dpcls_rule *rule;
1798 dpcls_lookup(&pmd->cls, key, &rule, 1);
1799 netdev_flow = dp_netdev_flow_cast(rule);
1804 static struct dp_netdev_flow *
1805 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1806 const ovs_u128 *ufidp, const struct nlattr *key,
1809 struct dp_netdev_flow *netdev_flow;
1813 /* If a UFID is not provided, determine one based on the key. */
1814 if (!ufidp && key && key_len
1815 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1816 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1821 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1823 if (ovs_u128_equals(&netdev_flow->ufid, ufidp)) {
1833 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1834 struct dpif_flow_stats *stats)
1836 struct dp_netdev_flow *netdev_flow;
1837 unsigned long long n;
1841 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1843 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1844 stats->n_packets = n;
1845 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1847 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1849 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1850 stats->tcp_flags = flags;
1853 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1854 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1855 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1858 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1859 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1860 struct dpif_flow *flow, bool terse)
1863 memset(flow, 0, sizeof *flow);
1865 struct flow_wildcards wc;
1866 struct dp_netdev_actions *actions;
1868 struct odp_flow_key_parms odp_parms = {
1869 .flow = &netdev_flow->flow,
1871 .support = dp_netdev_support,
1874 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1877 offset = key_buf->size;
1878 flow->key = ofpbuf_tail(key_buf);
1879 odp_parms.odp_in_port = netdev_flow->flow.in_port.odp_port;
1880 odp_flow_key_from_flow(&odp_parms, key_buf);
1881 flow->key_len = key_buf->size - offset;
1884 offset = mask_buf->size;
1885 flow->mask = ofpbuf_tail(mask_buf);
1886 odp_parms.odp_in_port = wc.masks.in_port.odp_port;
1887 odp_parms.key_buf = key_buf;
1888 odp_flow_key_from_mask(&odp_parms, mask_buf);
1889 flow->mask_len = mask_buf->size - offset;
1892 actions = dp_netdev_flow_get_actions(netdev_flow);
1893 flow->actions = actions->actions;
1894 flow->actions_len = actions->size;
1897 flow->ufid = netdev_flow->ufid;
1898 flow->ufid_present = true;
1899 flow->pmd_id = netdev_flow->pmd_id;
1900 get_dpif_flow_stats(netdev_flow, &flow->stats);
1904 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1905 const struct nlattr *mask_key,
1906 uint32_t mask_key_len, const struct flow *flow,
1907 struct flow_wildcards *wc)
1909 enum odp_key_fitness fitness;
1911 fitness = odp_flow_key_to_mask_udpif(mask_key, mask_key_len, key,
1914 /* This should not happen: it indicates that
1915 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1916 * disagree on the acceptable form of a mask. Log the problem
1917 * as an error, with enough details to enable debugging. */
1918 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1920 if (!VLOG_DROP_ERR(&rl)) {
1924 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1926 VLOG_ERR("internal error parsing flow mask %s (%s)",
1927 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1938 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1943 if (odp_flow_key_to_flow_udpif(key, key_len, flow)) {
1944 /* This should not happen: it indicates that odp_flow_key_from_flow()
1945 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1946 * flow. Log the problem as an error, with enough details to enable
1948 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1950 if (!VLOG_DROP_ERR(&rl)) {
1954 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1955 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1962 in_port = flow->in_port.odp_port;
1963 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1967 /* Userspace datapath doesn't support conntrack. */
1968 if (flow->ct_state || flow->ct_zone || flow->ct_mark
1969 || !ovs_u128_is_zero(&flow->ct_label)) {
1977 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1979 struct dp_netdev *dp = get_dp_netdev(dpif);
1980 struct dp_netdev_flow *netdev_flow;
1981 struct dp_netdev_pmd_thread *pmd;
1982 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1983 ? NON_PMD_CORE_ID : get->pmd_id;
1986 pmd = dp_netdev_get_pmd(dp, pmd_id);
1991 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1994 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1999 dp_netdev_pmd_unref(pmd);
2005 static struct dp_netdev_flow *
2006 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
2007 struct match *match, const ovs_u128 *ufid,
2008 const struct nlattr *actions, size_t actions_len)
2009 OVS_REQUIRES(pmd->flow_mutex)
2011 struct dp_netdev_flow *flow;
2012 struct netdev_flow_key mask;
2014 netdev_flow_mask_init(&mask, match);
2015 /* Make sure wc does not have metadata. */
2016 ovs_assert(!FLOWMAP_HAS_FIELD(&mask.mf.map, metadata)
2017 && !FLOWMAP_HAS_FIELD(&mask.mf.map, regs));
2019 /* Do not allocate extra space. */
2020 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
2021 memset(&flow->stats, 0, sizeof flow->stats);
2024 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
2025 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
2026 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
2027 ovs_refcount_init(&flow->ref_cnt);
2028 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
2030 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
2031 dpcls_insert(&pmd->cls, &flow->cr, &mask);
2033 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
2034 dp_netdev_flow_hash(&flow->ufid));
2036 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2038 struct ds ds = DS_EMPTY_INITIALIZER;
2040 match.tun_md.valid = false;
2041 match.flow = flow->flow;
2042 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2044 ds_put_cstr(&ds, "flow_add: ");
2045 odp_format_ufid(ufid, &ds);
2046 ds_put_cstr(&ds, " ");
2047 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2048 ds_put_cstr(&ds, ", actions:");
2049 format_odp_actions(&ds, actions, actions_len);
2051 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2060 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2062 struct dp_netdev *dp = get_dp_netdev(dpif);
2063 struct dp_netdev_flow *netdev_flow;
2064 struct netdev_flow_key key;
2065 struct dp_netdev_pmd_thread *pmd;
2068 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2069 ? NON_PMD_CORE_ID : put->pmd_id;
2072 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2076 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2077 put->mask, put->mask_len,
2078 &match.flow, &match.wc);
2083 pmd = dp_netdev_get_pmd(dp, pmd_id);
2088 /* Must produce a netdev_flow_key for lookup.
2089 * This interface is no longer performance critical, since it is not used
2090 * for upcall processing any more. */
2091 netdev_flow_key_from_flow(&key, &match.flow);
2096 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2099 ovs_mutex_lock(&pmd->flow_mutex);
2100 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2102 if (put->flags & DPIF_FP_CREATE) {
2103 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2105 memset(put->stats, 0, sizeof *put->stats);
2107 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2117 if (put->flags & DPIF_FP_MODIFY
2118 && flow_equal(&match.flow, &netdev_flow->flow)) {
2119 struct dp_netdev_actions *new_actions;
2120 struct dp_netdev_actions *old_actions;
2122 new_actions = dp_netdev_actions_create(put->actions,
2125 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2126 ovsrcu_set(&netdev_flow->actions, new_actions);
2129 get_dpif_flow_stats(netdev_flow, put->stats);
2131 if (put->flags & DPIF_FP_ZERO_STATS) {
2132 /* XXX: The userspace datapath uses thread local statistics
2133 * (for flows), which should be updated only by the owning
2134 * thread. Since we cannot write on stats memory here,
2135 * we choose not to support this flag. Please note:
2136 * - This feature is currently used only by dpctl commands with
2138 * - Should the need arise, this operation can be implemented
2139 * by keeping a base value (to be update here) for each
2140 * counter, and subtracting it before outputting the stats */
2144 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2145 } else if (put->flags & DPIF_FP_CREATE) {
2148 /* Overlapping flow. */
2152 ovs_mutex_unlock(&pmd->flow_mutex);
2153 dp_netdev_pmd_unref(pmd);
2159 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2161 struct dp_netdev *dp = get_dp_netdev(dpif);
2162 struct dp_netdev_flow *netdev_flow;
2163 struct dp_netdev_pmd_thread *pmd;
2164 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2165 ? NON_PMD_CORE_ID : del->pmd_id;
2168 pmd = dp_netdev_get_pmd(dp, pmd_id);
2173 ovs_mutex_lock(&pmd->flow_mutex);
2174 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2178 get_dpif_flow_stats(netdev_flow, del->stats);
2180 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2184 ovs_mutex_unlock(&pmd->flow_mutex);
2185 dp_netdev_pmd_unref(pmd);
2190 struct dpif_netdev_flow_dump {
2191 struct dpif_flow_dump up;
2192 struct cmap_position poll_thread_pos;
2193 struct cmap_position flow_pos;
2194 struct dp_netdev_pmd_thread *cur_pmd;
2196 struct ovs_mutex mutex;
2199 static struct dpif_netdev_flow_dump *
2200 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2202 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2205 static struct dpif_flow_dump *
2206 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2208 struct dpif_netdev_flow_dump *dump;
2210 dump = xzalloc(sizeof *dump);
2211 dpif_flow_dump_init(&dump->up, dpif_);
2212 dump->up.terse = terse;
2213 ovs_mutex_init(&dump->mutex);
2219 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2221 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2223 ovs_mutex_destroy(&dump->mutex);
2228 struct dpif_netdev_flow_dump_thread {
2229 struct dpif_flow_dump_thread up;
2230 struct dpif_netdev_flow_dump *dump;
2231 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2232 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2235 static struct dpif_netdev_flow_dump_thread *
2236 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2238 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2241 static struct dpif_flow_dump_thread *
2242 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2244 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2245 struct dpif_netdev_flow_dump_thread *thread;
2247 thread = xmalloc(sizeof *thread);
2248 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2249 thread->dump = dump;
2254 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2256 struct dpif_netdev_flow_dump_thread *thread
2257 = dpif_netdev_flow_dump_thread_cast(thread_);
2263 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2264 struct dpif_flow *flows, int max_flows)
2266 struct dpif_netdev_flow_dump_thread *thread
2267 = dpif_netdev_flow_dump_thread_cast(thread_);
2268 struct dpif_netdev_flow_dump *dump = thread->dump;
2269 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2273 ovs_mutex_lock(&dump->mutex);
2274 if (!dump->status) {
2275 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2276 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2277 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2278 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2280 /* First call to dump_next(), extracts the first pmd thread.
2281 * If there is no pmd thread, returns immediately. */
2283 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2285 ovs_mutex_unlock(&dump->mutex);
2292 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2293 struct cmap_node *node;
2295 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2299 netdev_flows[n_flows] = CONTAINER_OF(node,
2300 struct dp_netdev_flow,
2303 /* When finishing dumping the current pmd thread, moves to
2305 if (n_flows < flow_limit) {
2306 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2307 dp_netdev_pmd_unref(pmd);
2308 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2314 /* Keeps the reference to next caller. */
2315 dump->cur_pmd = pmd;
2317 /* If the current dump is empty, do not exit the loop, since the
2318 * remaining pmds could have flows to be dumped. Just dumps again
2319 * on the new 'pmd'. */
2322 ovs_mutex_unlock(&dump->mutex);
2324 for (i = 0; i < n_flows; i++) {
2325 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2326 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2327 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2328 struct dpif_flow *f = &flows[i];
2329 struct ofpbuf key, mask;
2331 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2332 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2333 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2341 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2342 OVS_NO_THREAD_SAFETY_ANALYSIS
2344 struct dp_netdev *dp = get_dp_netdev(dpif);
2345 struct dp_netdev_pmd_thread *pmd;
2346 struct dp_packet *pp;
2348 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2349 dp_packet_size(execute->packet) > UINT16_MAX) {
2353 /* Tries finding the 'pmd'. If NULL is returned, that means
2354 * the current thread is a non-pmd thread and should use
2355 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2356 pmd = ovsthread_getspecific(dp->per_pmd_key);
2358 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2361 /* If the current thread is non-pmd thread, acquires
2362 * the 'non_pmd_mutex'. */
2363 if (pmd->core_id == NON_PMD_CORE_ID) {
2364 ovs_mutex_lock(&dp->non_pmd_mutex);
2365 ovs_mutex_lock(&dp->port_mutex);
2368 pp = execute->packet;
2369 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2370 execute->actions_len);
2371 if (pmd->core_id == NON_PMD_CORE_ID) {
2372 dp_netdev_pmd_unref(pmd);
2373 ovs_mutex_unlock(&dp->port_mutex);
2374 ovs_mutex_unlock(&dp->non_pmd_mutex);
2381 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2385 for (i = 0; i < n_ops; i++) {
2386 struct dpif_op *op = ops[i];
2389 case DPIF_OP_FLOW_PUT:
2390 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2393 case DPIF_OP_FLOW_DEL:
2394 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2397 case DPIF_OP_EXECUTE:
2398 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2401 case DPIF_OP_FLOW_GET:
2402 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2408 /* Returns true if the configuration for rx queues or cpu mask
2411 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2413 if (dp->n_dpdk_rxqs != rxqs) {
2416 if (dp->pmd_cmask != NULL && cmask != NULL) {
2417 return strcmp(dp->pmd_cmask, cmask);
2419 return (dp->pmd_cmask != NULL || cmask != NULL);
2424 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2426 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2428 struct dp_netdev *dp = get_dp_netdev(dpif);
2430 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2431 struct dp_netdev_port *port;
2433 dp_netdev_destroy_all_pmds(dp);
2435 CMAP_FOR_EACH (port, node, &dp->ports) {
2436 if (netdev_is_pmd(port->netdev)) {
2439 /* Closes the existing 'rxq's. */
2440 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2441 netdev_rxq_close(port->rxq[i]);
2442 port->rxq[i] = NULL;
2445 /* Sets the new rx queue config. */
2446 err = netdev_set_multiq(port->netdev,
2447 ovs_numa_get_n_cores() + 1,
2449 if (err && (err != EOPNOTSUPP)) {
2450 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2451 " %u", netdev_get_name(port->netdev),
2456 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2457 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2458 * netdev_n_rxq(port->netdev));
2459 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2460 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2464 dp->n_dpdk_rxqs = n_rxqs;
2466 /* Reconfigures the cpu mask. */
2467 ovs_numa_set_cpu_mask(cmask);
2468 free(dp->pmd_cmask);
2469 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2471 /* Restores the non-pmd. */
2472 dp_netdev_set_nonpmd(dp);
2473 /* Restores all pmd threads. */
2474 dp_netdev_reset_pmd_threads(dp);
2481 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2482 uint32_t queue_id, uint32_t *priority)
2484 *priority = queue_id;
2489 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2490 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2491 struct dp_netdev_actions *
2492 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2494 struct dp_netdev_actions *netdev_actions;
2496 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2497 memcpy(netdev_actions->actions, actions, size);
2498 netdev_actions->size = size;
2500 return netdev_actions;
2503 struct dp_netdev_actions *
2504 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2506 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2510 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2515 static inline unsigned long long
2516 cycles_counter(void)
2519 return rte_get_tsc_cycles();
2525 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2526 extern struct ovs_mutex cycles_counter_fake_mutex;
2528 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2530 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2531 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2532 OVS_NO_THREAD_SAFETY_ANALYSIS
2534 pmd->last_cycles = cycles_counter();
2537 /* Stop counting cycles and add them to the counter 'type' */
2539 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2540 enum pmd_cycles_counter_type type)
2541 OVS_RELEASES(&cycles_counter_fake_mutex)
2542 OVS_NO_THREAD_SAFETY_ANALYSIS
2544 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2546 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2550 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2551 struct dp_netdev_port *port,
2552 struct netdev_rxq *rxq)
2554 struct dp_packet *packets[NETDEV_MAX_BURST];
2557 cycles_count_start(pmd);
2558 error = netdev_rxq_recv(rxq, packets, &cnt);
2559 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2561 *recirc_depth_get() = 0;
2563 cycles_count_start(pmd);
2564 dp_netdev_input(pmd, packets, cnt, port->port_no);
2565 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2566 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2567 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2569 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2570 netdev_get_name(port->netdev), ovs_strerror(error));
2574 /* Return true if needs to revalidate datapath flows. */
2576 dpif_netdev_run(struct dpif *dpif)
2578 struct dp_netdev_port *port;
2579 struct dp_netdev *dp = get_dp_netdev(dpif);
2580 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2582 uint64_t new_tnl_seq;
2584 ovs_mutex_lock(&dp->non_pmd_mutex);
2585 CMAP_FOR_EACH (port, node, &dp->ports) {
2586 if (!netdev_is_pmd(port->netdev)) {
2589 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2590 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2594 ovs_mutex_unlock(&dp->non_pmd_mutex);
2595 dp_netdev_pmd_unref(non_pmd);
2597 tnl_neigh_cache_run();
2599 new_tnl_seq = seq_read(tnl_conf_seq);
2601 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2602 dp->last_tnl_conf_seq = new_tnl_seq;
2609 dpif_netdev_wait(struct dpif *dpif)
2611 struct dp_netdev_port *port;
2612 struct dp_netdev *dp = get_dp_netdev(dpif);
2614 ovs_mutex_lock(&dp_netdev_mutex);
2615 CMAP_FOR_EACH (port, node, &dp->ports) {
2616 if (!netdev_is_pmd(port->netdev)) {
2619 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2620 netdev_rxq_wait(port->rxq[i]);
2624 ovs_mutex_unlock(&dp_netdev_mutex);
2625 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2629 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2630 struct rxq_poll **ppoll_list, int poll_cnt)
2631 OVS_REQUIRES(pmd->poll_mutex)
2633 struct rxq_poll *poll_list = *ppoll_list;
2634 struct rxq_poll *poll;
2637 for (i = 0; i < poll_cnt; i++) {
2638 port_unref(poll_list[i].port);
2641 poll_list = xrealloc(poll_list, pmd->poll_cnt * sizeof *poll_list);
2644 LIST_FOR_EACH (poll, node, &pmd->poll_list) {
2645 port_ref(poll->port);
2646 poll_list[i++] = *poll;
2649 *ppoll_list = poll_list;
2650 return pmd->poll_cnt;
2654 pmd_thread_main(void *f_)
2656 struct dp_netdev_pmd_thread *pmd = f_;
2657 unsigned int lc = 0;
2658 struct rxq_poll *poll_list;
2659 unsigned int port_seq = PMD_INITIAL_SEQ;
2666 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2667 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2668 pmd_thread_setaffinity_cpu(pmd->core_id);
2670 emc_cache_init(&pmd->flow_cache);
2672 ovs_mutex_lock(&pmd->poll_mutex);
2673 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2674 ovs_mutex_unlock(&pmd->poll_mutex);
2676 /* List port/core affinity */
2677 for (i = 0; i < poll_cnt; i++) {
2678 VLOG_INFO("Core %d processing port \'%s\'\n", pmd->core_id,
2679 netdev_get_name(poll_list[i].port->netdev));
2682 /* Signal here to make sure the pmd finishes
2683 * reloading the updated configuration. */
2684 dp_netdev_pmd_reload_done(pmd);
2687 for (i = 0; i < poll_cnt; i++) {
2688 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2696 emc_cache_slow_sweep(&pmd->flow_cache);
2697 coverage_try_clear();
2700 atomic_read_relaxed(&pmd->change_seq, &seq);
2701 if (seq != port_seq) {
2708 emc_cache_uninit(&pmd->flow_cache);
2710 if (!latch_is_set(&pmd->exit_latch)){
2714 for (i = 0; i < poll_cnt; i++) {
2715 port_unref(poll_list[i].port);
2718 dp_netdev_pmd_reload_done(pmd);
2725 dp_netdev_disable_upcall(struct dp_netdev *dp)
2726 OVS_ACQUIRES(dp->upcall_rwlock)
2728 fat_rwlock_wrlock(&dp->upcall_rwlock);
2732 dpif_netdev_disable_upcall(struct dpif *dpif)
2733 OVS_NO_THREAD_SAFETY_ANALYSIS
2735 struct dp_netdev *dp = get_dp_netdev(dpif);
2736 dp_netdev_disable_upcall(dp);
2740 dp_netdev_enable_upcall(struct dp_netdev *dp)
2741 OVS_RELEASES(dp->upcall_rwlock)
2743 fat_rwlock_unlock(&dp->upcall_rwlock);
2747 dpif_netdev_enable_upcall(struct dpif *dpif)
2748 OVS_NO_THREAD_SAFETY_ANALYSIS
2750 struct dp_netdev *dp = get_dp_netdev(dpif);
2751 dp_netdev_enable_upcall(dp);
2755 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2757 ovs_mutex_lock(&pmd->cond_mutex);
2758 xpthread_cond_signal(&pmd->cond);
2759 ovs_mutex_unlock(&pmd->cond_mutex);
2762 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2763 * the pointer if succeeds, otherwise, NULL.
2765 * Caller must unrefs the returned reference. */
2766 static struct dp_netdev_pmd_thread *
2767 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2769 struct dp_netdev_pmd_thread *pmd;
2770 const struct cmap_node *pnode;
2772 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2776 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2778 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2781 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2783 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2785 struct dp_netdev_pmd_thread *non_pmd;
2787 non_pmd = xzalloc(sizeof *non_pmd);
2788 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2792 /* Caller must have valid pointer to 'pmd'. */
2794 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2796 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2800 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2802 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2803 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2807 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2808 * fails, keeps checking for next node until reaching the end of cmap.
2810 * Caller must unrefs the returned reference. */
2811 static struct dp_netdev_pmd_thread *
2812 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2814 struct dp_netdev_pmd_thread *next;
2817 struct cmap_node *node;
2819 node = cmap_next_position(&dp->poll_threads, pos);
2820 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2822 } while (next && !dp_netdev_pmd_try_ref(next));
2827 /* Configures the 'pmd' based on the input argument. */
2829 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2830 int index, unsigned core_id, int numa_id)
2834 pmd->core_id = core_id;
2835 pmd->numa_id = numa_id;
2838 atomic_init(&pmd->tx_qid,
2839 (core_id == NON_PMD_CORE_ID)
2840 ? ovs_numa_get_n_cores()
2841 : get_n_pmd_threads(dp));
2843 ovs_refcount_init(&pmd->ref_cnt);
2844 latch_init(&pmd->exit_latch);
2845 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2846 xpthread_cond_init(&pmd->cond, NULL);
2847 ovs_mutex_init(&pmd->cond_mutex);
2848 ovs_mutex_init(&pmd->flow_mutex);
2849 ovs_mutex_init(&pmd->poll_mutex);
2850 dpcls_init(&pmd->cls);
2851 cmap_init(&pmd->flow_table);
2852 list_init(&pmd->poll_list);
2853 /* init the 'flow_cache' since there is no
2854 * actual thread created for NON_PMD_CORE_ID. */
2855 if (core_id == NON_PMD_CORE_ID) {
2856 emc_cache_init(&pmd->flow_cache);
2858 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2859 hash_int(core_id, 0));
2863 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2865 dp_netdev_pmd_flow_flush(pmd);
2866 dpcls_destroy(&pmd->cls);
2867 cmap_destroy(&pmd->flow_table);
2868 ovs_mutex_destroy(&pmd->flow_mutex);
2869 latch_destroy(&pmd->exit_latch);
2870 xpthread_cond_destroy(&pmd->cond);
2871 ovs_mutex_destroy(&pmd->cond_mutex);
2872 ovs_mutex_destroy(&pmd->poll_mutex);
2876 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2877 * and unrefs the struct. */
2879 dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd)
2881 struct rxq_poll *poll;
2883 /* Uninit the 'flow_cache' since there is
2884 * no actual thread uninit it for NON_PMD_CORE_ID. */
2885 if (pmd->core_id == NON_PMD_CORE_ID) {
2886 emc_cache_uninit(&pmd->flow_cache);
2888 latch_set(&pmd->exit_latch);
2889 dp_netdev_reload_pmd__(pmd);
2890 ovs_numa_unpin_core(pmd->core_id);
2891 xpthread_join(pmd->thread, NULL);
2894 /* Unref all ports and free poll_list. */
2895 LIST_FOR_EACH_POP (poll, node, &pmd->poll_list) {
2896 port_unref(poll->port);
2900 /* Purges the 'pmd''s flows after stopping the thread, but before
2901 * destroying the flows, so that the flow stats can be collected. */
2902 if (dp->dp_purge_cb) {
2903 dp->dp_purge_cb(dp->dp_purge_aux, pmd->core_id);
2905 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2906 dp_netdev_pmd_unref(pmd);
2909 /* Destroys all pmd threads. */
2911 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2913 struct dp_netdev_pmd_thread *pmd;
2914 struct dp_netdev_pmd_thread **pmd_list;
2915 size_t k = 0, n_pmds;
2917 n_pmds = cmap_count(&dp->poll_threads);
2918 pmd_list = xcalloc(n_pmds, sizeof *pmd_list);
2920 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2921 /* We cannot call dp_netdev_del_pmd(), since it alters
2922 * 'dp->poll_threads' (while we're iterating it) and it
2924 ovs_assert(k < n_pmds);
2925 pmd_list[k++] = pmd;
2928 for (size_t i = 0; i < k; i++) {
2929 dp_netdev_del_pmd(dp, pmd_list[i]);
2934 /* Deletes all pmd threads on numa node 'numa_id' and
2935 * fixes tx_qids of other threads to keep them sequential. */
2937 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2939 struct dp_netdev_pmd_thread *pmd;
2940 int n_pmds_on_numa, n_pmds;
2941 int *free_idx, k = 0;
2942 struct dp_netdev_pmd_thread **pmd_list;
2944 n_pmds_on_numa = get_n_pmd_threads_on_numa(dp, numa_id);
2945 free_idx = xcalloc(n_pmds_on_numa, sizeof *free_idx);
2946 pmd_list = xcalloc(n_pmds_on_numa, sizeof *pmd_list);
2948 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2949 /* We cannot call dp_netdev_del_pmd(), since it alters
2950 * 'dp->poll_threads' (while we're iterating it) and it
2952 if (pmd->numa_id == numa_id) {
2953 atomic_read_relaxed(&pmd->tx_qid, &free_idx[k]);
2955 ovs_assert(k < n_pmds_on_numa);
2960 for (int i = 0; i < k; i++) {
2961 dp_netdev_del_pmd(dp, pmd_list[i]);
2964 n_pmds = get_n_pmd_threads(dp);
2965 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2968 atomic_read_relaxed(&pmd->tx_qid, &old_tx_qid);
2970 if (old_tx_qid >= n_pmds) {
2971 int new_tx_qid = free_idx[--k];
2973 atomic_store_relaxed(&pmd->tx_qid, new_tx_qid);
2981 /* Returns PMD thread from this numa node with fewer rx queues to poll.
2982 * Returns NULL if there is no PMD threads on this numa node.
2983 * Can be called safely only by main thread. */
2984 static struct dp_netdev_pmd_thread *
2985 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id)
2988 struct dp_netdev_pmd_thread *pmd, *res = NULL;
2990 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2991 if (pmd->numa_id == numa_id
2992 && (min_cnt > pmd->poll_cnt || res == NULL)) {
2993 min_cnt = pmd->poll_cnt;
3001 /* Adds rx queue to poll_list of PMD thread. */
3003 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
3004 struct dp_netdev_port *port, struct netdev_rxq *rx)
3005 OVS_REQUIRES(pmd->poll_mutex)
3007 struct rxq_poll *poll = xmalloc(sizeof *poll);
3013 list_push_back(&pmd->poll_list, &poll->node);
3017 /* Checks the numa node id of 'netdev' and starts pmd threads for
3020 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
3024 if (!ovs_numa_numa_id_is_valid(numa_id)) {
3025 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
3026 "invalid", numa_id);
3030 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
3032 /* If there are already pmd threads created for the numa node
3033 * in which 'netdev' is on, do nothing. Else, creates the
3034 * pmd threads for the numa node. */
3036 int can_have, n_unpinned, i, index = 0;
3037 struct dp_netdev_pmd_thread **pmds;
3038 struct dp_netdev_port *port;
3040 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
3042 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
3043 "cores on numa node");
3047 /* If cpu mask is specified, uses all unpinned cores, otherwise
3048 * tries creating NR_PMD_THREADS pmd threads. */
3049 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
3050 pmds = xzalloc(can_have * sizeof *pmds);
3051 for (i = 0; i < can_have; i++) {
3052 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
3053 pmds[i] = xzalloc(sizeof **pmds);
3054 dp_netdev_configure_pmd(pmds[i], dp, i, core_id, numa_id);
3057 /* Distributes rx queues of this numa node between new pmd threads. */
3058 CMAP_FOR_EACH (port, node, &dp->ports) {
3059 if (netdev_is_pmd(port->netdev)
3060 && netdev_get_numa_id(port->netdev) == numa_id) {
3061 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
3062 /* Make thread-safety analyser happy. */
3063 ovs_mutex_lock(&pmds[index]->poll_mutex);
3064 dp_netdev_add_rxq_to_pmd(pmds[index], port, port->rxq[i]);
3065 ovs_mutex_unlock(&pmds[index]->poll_mutex);
3066 index = (index + 1) % can_have;
3071 /* Actual start of pmd threads. */
3072 for (i = 0; i < can_have; i++) {
3073 pmds[i]->thread = ovs_thread_create("pmd", pmd_thread_main, pmds[i]);
3076 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
3081 /* Called after pmd threads config change. Restarts pmd threads with
3082 * new configuration. */
3084 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
3086 struct dp_netdev_port *port;
3088 CMAP_FOR_EACH (port, node, &dp->ports) {
3089 if (netdev_is_pmd(port->netdev)) {
3090 int numa_id = netdev_get_numa_id(port->netdev);
3092 dp_netdev_set_pmds_on_numa(dp, numa_id);
3098 dpif_netdev_get_datapath_version(void)
3100 return xstrdup("<built-in>");
3104 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
3105 uint16_t tcp_flags, long long now)
3109 atomic_store_relaxed(&netdev_flow->stats.used, now);
3110 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
3111 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
3112 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
3114 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
3118 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
3119 enum dp_stat_type type, int cnt)
3121 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3125 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3126 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3127 enum dpif_upcall_type type, const struct nlattr *userdata,
3128 struct ofpbuf *actions, struct ofpbuf *put_actions)
3130 struct dp_netdev *dp = pmd->dp;
3131 struct flow_tnl orig_tunnel;
3134 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3138 /* Upcall processing expects the Geneve options to be in the translated
3139 * format but we need to retain the raw format for datapath use. */
3140 orig_tunnel.flags = flow->tunnel.flags;
3141 if (flow->tunnel.flags & FLOW_TNL_F_UDPIF) {
3142 orig_tunnel.metadata.present.len = flow->tunnel.metadata.present.len;
3143 memcpy(orig_tunnel.metadata.opts.gnv, flow->tunnel.metadata.opts.gnv,
3144 flow->tunnel.metadata.present.len);
3145 err = tun_metadata_from_geneve_udpif(&orig_tunnel, &orig_tunnel,
3152 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3153 struct ds ds = DS_EMPTY_INITIALIZER;
3156 struct odp_flow_key_parms odp_parms = {
3159 .odp_in_port = flow->in_port.odp_port,
3160 .support = dp_netdev_support,
3163 ofpbuf_init(&key, 0);
3164 odp_flow_key_from_flow(&odp_parms, &key);
3165 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3166 dp_packet_size(packet_));
3168 odp_flow_key_format(key.data, key.size, &ds);
3170 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3171 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3173 ofpbuf_uninit(&key);
3179 err = dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3180 actions, wc, put_actions, dp->upcall_aux);
3181 if (err && err != ENOSPC) {
3185 /* Translate tunnel metadata masks to datapath format. */
3187 if (wc->masks.tunnel.metadata.present.map) {
3188 struct geneve_opt opts[TLV_TOT_OPT_SIZE /
3189 sizeof(struct geneve_opt)];
3191 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3192 tun_metadata_to_geneve_udpif_mask(&flow->tunnel,
3194 orig_tunnel.metadata.opts.gnv,
3195 orig_tunnel.metadata.present.len,
3198 orig_tunnel.metadata.present.len = 0;
3201 memset(&wc->masks.tunnel.metadata, 0,
3202 sizeof wc->masks.tunnel.metadata);
3203 memcpy(&wc->masks.tunnel.metadata.opts.gnv, opts,
3204 orig_tunnel.metadata.present.len);
3206 wc->masks.tunnel.metadata.present.len = 0xff;
3209 /* Restore tunnel metadata. We need to use the saved options to ensure
3210 * that any unknown options are not lost. The generated mask will have
3211 * the same structure, matching on types and lengths but wildcarding
3212 * option data we don't care about. */
3213 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3214 memcpy(&flow->tunnel.metadata.opts.gnv, orig_tunnel.metadata.opts.gnv,
3215 orig_tunnel.metadata.present.len);
3216 flow->tunnel.metadata.present.len = orig_tunnel.metadata.present.len;
3217 flow->tunnel.flags |= FLOW_TNL_F_UDPIF;
3223 static inline uint32_t
3224 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3225 const struct miniflow *mf)
3227 uint32_t hash, recirc_depth;
3229 if (OVS_LIKELY(dp_packet_rss_valid(packet))) {
3230 hash = dp_packet_get_rss_hash(packet);
3232 hash = miniflow_hash_5tuple(mf, 0);
3233 dp_packet_set_rss_hash(packet, hash);
3236 /* The RSS hash must account for the recirculation depth to avoid
3237 * collisions in the exact match cache */
3238 recirc_depth = *recirc_depth_get_unsafe();
3239 if (OVS_UNLIKELY(recirc_depth)) {
3240 hash = hash_finish(hash, recirc_depth);
3241 dp_packet_set_rss_hash(packet, hash);
3246 struct packet_batch {
3247 unsigned int packet_count;
3248 unsigned int byte_count;
3251 struct dp_netdev_flow *flow;
3253 struct dp_packet *packets[NETDEV_MAX_BURST];
3257 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
3258 const struct miniflow *mf)
3260 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3261 batch->packets[batch->packet_count++] = packet;
3262 batch->byte_count += dp_packet_size(packet);
3266 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
3268 flow->batch = batch;
3271 batch->packet_count = 0;
3272 batch->byte_count = 0;
3273 batch->tcp_flags = 0;
3277 packet_batch_execute(struct packet_batch *batch,
3278 struct dp_netdev_pmd_thread *pmd,
3281 struct dp_netdev_actions *actions;
3282 struct dp_netdev_flow *flow = batch->flow;
3284 dp_netdev_flow_used(flow, batch->packet_count, batch->byte_count,
3285 batch->tcp_flags, now);
3287 actions = dp_netdev_flow_get_actions(flow);
3289 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
3290 actions->actions, actions->size);
3294 dp_netdev_queue_batches(struct dp_packet *pkt,
3295 struct dp_netdev_flow *flow, const struct miniflow *mf,
3296 struct packet_batch *batches, size_t *n_batches)
3298 struct packet_batch *batch = flow->batch;
3300 if (OVS_UNLIKELY(!batch)) {
3301 batch = &batches[(*n_batches)++];
3302 packet_batch_init(batch, flow);
3305 packet_batch_update(batch, pkt, mf);
3308 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3309 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
3310 * miniflow is copied into 'keys' and the packet pointer is moved at the
3311 * beginning of the 'packets' array.
3313 * The function returns the number of packets that needs to be processed in the
3314 * 'packets' array (they have been moved to the beginning of the vector).
3316 * If 'md_is_valid' is false, the metadata in 'packets' is not valid and must be
3317 * initialized by this function using 'port_no'.
3319 static inline size_t
3320 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
3321 size_t cnt, struct netdev_flow_key *keys,
3322 struct packet_batch batches[], size_t *n_batches,
3323 bool md_is_valid, odp_port_t port_no)
3325 struct emc_cache *flow_cache = &pmd->flow_cache;
3326 struct netdev_flow_key *key = &keys[0];
3327 size_t i, n_missed = 0, n_dropped = 0;
3329 for (i = 0; i < cnt; i++) {
3330 struct dp_netdev_flow *flow;
3331 struct dp_packet *packet = packets[i];
3333 if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) {
3334 dp_packet_delete(packet);
3340 /* Prefetch next packet data and metadata. */
3341 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3342 pkt_metadata_prefetch_init(&packets[i+1]->md);
3346 pkt_metadata_init(&packet->md, port_no);
3348 miniflow_extract(packet, &key->mf);
3349 key->len = 0; /* Not computed yet. */
3350 key->hash = dpif_netdev_packet_get_rss_hash(packet, &key->mf);
3352 flow = emc_lookup(flow_cache, key);
3353 if (OVS_LIKELY(flow)) {
3354 dp_netdev_queue_batches(packet, flow, &key->mf, batches,
3357 /* Exact match cache missed. Group missed packets together at
3358 * the beginning of the 'packets' array. */
3359 packets[n_missed] = packet;
3360 key = &keys[n_missed++];
3364 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - n_dropped - n_missed);
3370 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3371 struct dp_packet **packets, size_t cnt,
3372 struct netdev_flow_key *keys,
3373 struct packet_batch batches[], size_t *n_batches)
3375 #if !defined(__CHECKER__) && !defined(_WIN32)
3376 const size_t PKT_ARRAY_SIZE = cnt;
3378 /* Sparse or MSVC doesn't like variable length array. */
3379 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3381 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3382 struct dp_netdev *dp = pmd->dp;
3383 struct emc_cache *flow_cache = &pmd->flow_cache;
3384 int miss_cnt = 0, lost_cnt = 0;
3388 for (i = 0; i < cnt; i++) {
3389 /* Key length is needed in all the cases, hash computed on demand. */
3390 keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf));
3392 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3393 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3394 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3395 struct ofpbuf actions, put_actions;
3398 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3399 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3401 for (i = 0; i < cnt; i++) {
3402 struct dp_netdev_flow *netdev_flow;
3403 struct ofpbuf *add_actions;
3407 if (OVS_LIKELY(rules[i])) {
3411 /* It's possible that an earlier slow path execution installed
3412 * a rule covering this flow. In this case, it's a lot cheaper
3413 * to catch it here than execute a miss. */
3414 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3416 rules[i] = &netdev_flow->cr;
3422 match.tun_md.valid = false;
3423 miniflow_expand(&keys[i].mf, &match.flow);
3425 ofpbuf_clear(&actions);
3426 ofpbuf_clear(&put_actions);
3428 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
3429 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
3430 &ufid, DPIF_UC_MISS, NULL, &actions,
3432 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3433 dp_packet_delete(packets[i]);
3438 /* The Netlink encoding of datapath flow keys cannot express
3439 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3440 * tag is interpreted as exact match on the fact that there is no
3441 * VLAN. Unless we refactor a lot of code that translates between
3442 * Netlink and struct flow representations, we have to do the same
3444 if (!match.wc.masks.vlan_tci) {
3445 match.wc.masks.vlan_tci = htons(0xffff);
3448 /* We can't allow the packet batching in the next loop to execute
3449 * the actions. Otherwise, if there are any slow path actions,
3450 * we'll send the packet up twice. */
3451 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
3452 actions.data, actions.size);
3454 add_actions = put_actions.size ? &put_actions : &actions;
3455 if (OVS_LIKELY(error != ENOSPC)) {
3456 /* XXX: There's a race window where a flow covering this packet
3457 * could have already been installed since we last did the flow
3458 * lookup before upcall. This could be solved by moving the
3459 * mutex lock outside the loop, but that's an awful long time
3460 * to be locking everyone out of making flow installs. If we
3461 * move to a per-core classifier, it would be reasonable. */
3462 ovs_mutex_lock(&pmd->flow_mutex);
3463 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3464 if (OVS_LIKELY(!netdev_flow)) {
3465 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3469 ovs_mutex_unlock(&pmd->flow_mutex);
3471 emc_insert(flow_cache, &keys[i], netdev_flow);
3475 ofpbuf_uninit(&actions);
3476 ofpbuf_uninit(&put_actions);
3477 fat_rwlock_unlock(&dp->upcall_rwlock);
3478 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3479 } else if (OVS_UNLIKELY(any_miss)) {
3480 for (i = 0; i < cnt; i++) {
3481 if (OVS_UNLIKELY(!rules[i])) {
3482 dp_packet_delete(packets[i]);
3489 for (i = 0; i < cnt; i++) {
3490 struct dp_packet *packet = packets[i];
3491 struct dp_netdev_flow *flow;
3493 if (OVS_UNLIKELY(!rules[i])) {
3497 flow = dp_netdev_flow_cast(rules[i]);
3499 emc_insert(flow_cache, &keys[i], flow);
3500 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3503 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3504 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3505 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3508 /* Packets enter the datapath from a port (or from recirculation) here.
3510 * For performance reasons a caller may choose not to initialize the metadata
3511 * in 'packets': in this case 'mdinit' is false and this function needs to
3512 * initialize it using 'port_no'. If the metadata in 'packets' is already
3513 * valid, 'md_is_valid' must be true and 'port_no' will be ignored. */
3515 dp_netdev_input__(struct dp_netdev_pmd_thread *pmd,
3516 struct dp_packet **packets, int cnt,
3517 bool md_is_valid, odp_port_t port_no)
3519 #if !defined(__CHECKER__) && !defined(_WIN32)
3520 const size_t PKT_ARRAY_SIZE = cnt;
3522 /* Sparse or MSVC doesn't like variable length array. */
3523 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3525 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3526 struct packet_batch batches[PKT_ARRAY_SIZE];
3527 long long now = time_msec();
3528 size_t newcnt, n_batches, i;
3531 newcnt = emc_processing(pmd, packets, cnt, keys, batches, &n_batches,
3532 md_is_valid, port_no);
3533 if (OVS_UNLIKELY(newcnt)) {
3534 fast_path_processing(pmd, packets, newcnt, keys, batches, &n_batches);
3537 for (i = 0; i < n_batches; i++) {
3538 batches[i].flow->batch = NULL;
3541 for (i = 0; i < n_batches; i++) {
3542 packet_batch_execute(&batches[i], pmd, now);
3547 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3548 struct dp_packet **packets, int cnt,
3551 dp_netdev_input__(pmd, packets, cnt, false, port_no);
3555 dp_netdev_recirculate(struct dp_netdev_pmd_thread *pmd,
3556 struct dp_packet **packets, int cnt)
3558 dp_netdev_input__(pmd, packets, cnt, true, 0);
3561 struct dp_netdev_execute_aux {
3562 struct dp_netdev_pmd_thread *pmd;
3566 dpif_netdev_register_dp_purge_cb(struct dpif *dpif, dp_purge_callback *cb,
3569 struct dp_netdev *dp = get_dp_netdev(dpif);
3570 dp->dp_purge_aux = aux;
3571 dp->dp_purge_cb = cb;
3575 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3578 struct dp_netdev *dp = get_dp_netdev(dpif);
3579 dp->upcall_aux = aux;
3584 dp_netdev_drop_packets(struct dp_packet **packets, int cnt, bool may_steal)
3589 for (i = 0; i < cnt; i++) {
3590 dp_packet_delete(packets[i]);
3596 push_tnl_action(const struct dp_netdev *dp,
3597 const struct nlattr *attr,
3598 struct dp_packet **packets, int cnt)
3600 struct dp_netdev_port *tun_port;
3601 const struct ovs_action_push_tnl *data;
3603 data = nl_attr_get(attr);
3605 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3609 netdev_push_header(tun_port->netdev, packets, cnt, data);
3615 dp_netdev_clone_pkt_batch(struct dp_packet **dst_pkts,
3616 struct dp_packet **src_pkts, int cnt)
3620 for (i = 0; i < cnt; i++) {
3621 dst_pkts[i] = dp_packet_clone(src_pkts[i]);
3626 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3627 const struct nlattr *a, bool may_steal)
3628 OVS_NO_THREAD_SAFETY_ANALYSIS
3630 struct dp_netdev_execute_aux *aux = aux_;
3631 uint32_t *depth = recirc_depth_get();
3632 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3633 struct dp_netdev *dp = pmd->dp;
3634 int type = nl_attr_type(a);
3635 struct dp_netdev_port *p;
3638 switch ((enum ovs_action_attr)type) {
3639 case OVS_ACTION_ATTR_OUTPUT:
3640 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3641 if (OVS_LIKELY(p)) {
3644 atomic_read_relaxed(&pmd->tx_qid, &tx_qid);
3646 netdev_send(p->netdev, tx_qid, packets, cnt, may_steal);
3651 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3652 if (*depth < MAX_RECIRC_DEPTH) {
3653 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3657 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3661 err = push_tnl_action(dp, a, packets, cnt);
3664 dp_netdev_recirculate(pmd, packets, cnt);
3667 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3673 case OVS_ACTION_ATTR_TUNNEL_POP:
3674 if (*depth < MAX_RECIRC_DEPTH) {
3675 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3677 p = dp_netdev_lookup_port(dp, portno);
3679 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3683 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3687 err = netdev_pop_header(p->netdev, packets, cnt);
3690 for (i = 0; i < cnt; i++) {
3691 packets[i]->md.in_port.odp_port = portno;
3695 dp_netdev_recirculate(pmd, packets, cnt);
3698 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3705 case OVS_ACTION_ATTR_USERSPACE:
3706 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3707 const struct nlattr *userdata;
3708 struct ofpbuf actions;
3712 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3713 ofpbuf_init(&actions, 0);
3715 for (i = 0; i < cnt; i++) {
3718 ofpbuf_clear(&actions);
3720 flow_extract(packets[i], &flow);
3721 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3722 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3723 DPIF_UC_ACTION, userdata,&actions,
3725 if (!error || error == ENOSPC) {
3726 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3727 actions.data, actions.size);
3728 } else if (may_steal) {
3729 dp_packet_delete(packets[i]);
3732 ofpbuf_uninit(&actions);
3733 fat_rwlock_unlock(&dp->upcall_rwlock);
3739 case OVS_ACTION_ATTR_RECIRC:
3740 if (*depth < MAX_RECIRC_DEPTH) {
3741 struct dp_packet *recirc_pkts[NETDEV_MAX_BURST];
3744 dp_netdev_clone_pkt_batch(recirc_pkts, packets, cnt);
3745 packets = recirc_pkts;
3748 for (i = 0; i < cnt; i++) {
3749 packets[i]->md.recirc_id = nl_attr_get_u32(a);
3753 dp_netdev_recirculate(pmd, packets, cnt);
3759 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3762 case OVS_ACTION_ATTR_CT:
3763 /* If a flow with this action is slow-pathed, datapath assistance is
3764 * required to implement it. However, we don't support this action
3765 * in the userspace datapath. */
3766 VLOG_WARN("Cannot execute conntrack action in userspace.");
3769 case OVS_ACTION_ATTR_PUSH_VLAN:
3770 case OVS_ACTION_ATTR_POP_VLAN:
3771 case OVS_ACTION_ATTR_PUSH_MPLS:
3772 case OVS_ACTION_ATTR_POP_MPLS:
3773 case OVS_ACTION_ATTR_SET:
3774 case OVS_ACTION_ATTR_SET_MASKED:
3775 case OVS_ACTION_ATTR_SAMPLE:
3776 case OVS_ACTION_ATTR_HASH:
3777 case OVS_ACTION_ATTR_UNSPEC:
3778 case __OVS_ACTION_ATTR_MAX:
3782 dp_netdev_drop_packets(packets, cnt, may_steal);
3786 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3787 struct dp_packet **packets, int cnt,
3789 const struct nlattr *actions, size_t actions_len)
3791 struct dp_netdev_execute_aux aux = { pmd };
3793 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3794 actions_len, dp_execute_cb);
3797 const struct dpif_class dpif_netdev_class = {
3800 dpif_netdev_enumerate,
3801 dpif_netdev_port_open_type,
3804 dpif_netdev_destroy,
3807 dpif_netdev_get_stats,
3808 dpif_netdev_port_add,
3809 dpif_netdev_port_del,
3810 dpif_netdev_port_query_by_number,
3811 dpif_netdev_port_query_by_name,
3812 NULL, /* port_get_pid */
3813 dpif_netdev_port_dump_start,
3814 dpif_netdev_port_dump_next,
3815 dpif_netdev_port_dump_done,
3816 dpif_netdev_port_poll,
3817 dpif_netdev_port_poll_wait,
3818 dpif_netdev_flow_flush,
3819 dpif_netdev_flow_dump_create,
3820 dpif_netdev_flow_dump_destroy,
3821 dpif_netdev_flow_dump_thread_create,
3822 dpif_netdev_flow_dump_thread_destroy,
3823 dpif_netdev_flow_dump_next,
3824 dpif_netdev_operate,
3825 NULL, /* recv_set */
3826 NULL, /* handlers_set */
3827 dpif_netdev_pmd_set,
3828 dpif_netdev_queue_to_priority,
3830 NULL, /* recv_wait */
3831 NULL, /* recv_purge */
3832 dpif_netdev_register_dp_purge_cb,
3833 dpif_netdev_register_upcall_cb,
3834 dpif_netdev_enable_upcall,
3835 dpif_netdev_disable_upcall,
3836 dpif_netdev_get_datapath_version,
3837 NULL, /* ct_dump_start */
3838 NULL, /* ct_dump_next */
3839 NULL, /* ct_dump_done */
3840 NULL, /* ct_flush */
3844 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3845 const char *argv[], void *aux OVS_UNUSED)
3847 struct dp_netdev_port *old_port;
3848 struct dp_netdev_port *new_port;
3849 struct dp_netdev *dp;
3852 ovs_mutex_lock(&dp_netdev_mutex);
3853 dp = shash_find_data(&dp_netdevs, argv[1]);
3854 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3855 ovs_mutex_unlock(&dp_netdev_mutex);
3856 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3859 ovs_refcount_ref(&dp->ref_cnt);
3860 ovs_mutex_unlock(&dp_netdev_mutex);
3862 ovs_mutex_lock(&dp->port_mutex);
3863 if (get_port_by_name(dp, argv[2], &old_port)) {
3864 unixctl_command_reply_error(conn, "unknown port");
3868 port_no = u32_to_odp(atoi(argv[3]));
3869 if (!port_no || port_no == ODPP_NONE) {
3870 unixctl_command_reply_error(conn, "bad port number");
3873 if (dp_netdev_lookup_port(dp, port_no)) {
3874 unixctl_command_reply_error(conn, "port number already in use");
3878 /* Remove old port. */
3879 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3880 ovsrcu_postpone(free, old_port);
3882 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3883 new_port = xmemdup(old_port, sizeof *old_port);
3884 new_port->port_no = port_no;
3885 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3887 seq_change(dp->port_seq);
3888 unixctl_command_reply(conn, NULL);
3891 ovs_mutex_unlock(&dp->port_mutex);
3892 dp_netdev_unref(dp);
3896 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3897 const char *argv[], void *aux OVS_UNUSED)
3899 struct dp_netdev_port *port;
3900 struct dp_netdev *dp;
3902 ovs_mutex_lock(&dp_netdev_mutex);
3903 dp = shash_find_data(&dp_netdevs, argv[1]);
3904 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3905 ovs_mutex_unlock(&dp_netdev_mutex);
3906 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3909 ovs_refcount_ref(&dp->ref_cnt);
3910 ovs_mutex_unlock(&dp_netdev_mutex);
3912 ovs_mutex_lock(&dp->port_mutex);
3913 if (get_port_by_name(dp, argv[2], &port)) {
3914 unixctl_command_reply_error(conn, "unknown port");
3915 } else if (port->port_no == ODPP_LOCAL) {
3916 unixctl_command_reply_error(conn, "can't delete local port");
3918 do_del_port(dp, port);
3919 unixctl_command_reply(conn, NULL);
3921 ovs_mutex_unlock(&dp->port_mutex);
3923 dp_netdev_unref(dp);
3927 dpif_dummy_register__(const char *type)
3929 struct dpif_class *class;
3931 class = xmalloc(sizeof *class);
3932 *class = dpif_netdev_class;
3933 class->type = xstrdup(type);
3934 dp_register_provider(class);
3938 dpif_dummy_override(const char *type)
3943 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
3944 * a userland-only build. It's useful for testsuite.
3946 error = dp_unregister_provider(type);
3947 if (error == 0 || error == EAFNOSUPPORT) {
3948 dpif_dummy_register__(type);
3953 dpif_dummy_register(enum dummy_level level)
3955 if (level == DUMMY_OVERRIDE_ALL) {
3960 dp_enumerate_types(&types);
3961 SSET_FOR_EACH (type, &types) {
3962 dpif_dummy_override(type);
3964 sset_destroy(&types);
3965 } else if (level == DUMMY_OVERRIDE_SYSTEM) {
3966 dpif_dummy_override("system");
3969 dpif_dummy_register__("dummy");
3971 unixctl_command_register("dpif-dummy/change-port-number",
3972 "dp port new-number",
3973 3, 3, dpif_dummy_change_port_number, NULL);
3974 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3975 2, 2, dpif_dummy_delete_port, NULL);
3978 /* Datapath Classifier. */
3980 /* A set of rules that all have the same fields wildcarded. */
3981 struct dpcls_subtable {
3982 /* The fields are only used by writers. */
3983 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3985 /* These fields are accessed by readers. */
3986 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3987 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3988 /* 'mask' must be the last field, additional space is allocated here. */
3991 /* Initializes 'cls' as a classifier that initially contains no classification
3994 dpcls_init(struct dpcls *cls)
3996 cmap_init(&cls->subtables_map);
3997 pvector_init(&cls->subtables);
4001 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
4003 pvector_remove(&cls->subtables, subtable);
4004 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
4005 subtable->mask.hash);
4006 cmap_destroy(&subtable->rules);
4007 ovsrcu_postpone(free, subtable);
4010 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
4011 * caller's responsibility.
4012 * May only be called after all the readers have been terminated. */
4014 dpcls_destroy(struct dpcls *cls)
4017 struct dpcls_subtable *subtable;
4019 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
4020 ovs_assert(cmap_count(&subtable->rules) == 0);
4021 dpcls_destroy_subtable(cls, subtable);
4023 cmap_destroy(&cls->subtables_map);
4024 pvector_destroy(&cls->subtables);
4028 static struct dpcls_subtable *
4029 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
4031 struct dpcls_subtable *subtable;
4033 /* Need to add one. */
4034 subtable = xmalloc(sizeof *subtable
4035 - sizeof subtable->mask.mf + mask->len);
4036 cmap_init(&subtable->rules);
4037 netdev_flow_key_clone(&subtable->mask, mask);
4038 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
4039 pvector_insert(&cls->subtables, subtable, 0);
4040 pvector_publish(&cls->subtables);
4045 static inline struct dpcls_subtable *
4046 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
4048 struct dpcls_subtable *subtable;
4050 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
4051 &cls->subtables_map) {
4052 if (netdev_flow_key_equal(&subtable->mask, mask)) {
4056 return dpcls_create_subtable(cls, mask);
4059 /* Insert 'rule' into 'cls'. */
4061 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
4062 const struct netdev_flow_key *mask)
4064 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
4066 rule->mask = &subtable->mask;
4067 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
4070 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
4072 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
4074 struct dpcls_subtable *subtable;
4076 ovs_assert(rule->mask);
4078 INIT_CONTAINER(subtable, rule->mask, mask);
4080 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
4082 dpcls_destroy_subtable(cls, subtable);
4083 pvector_publish(&cls->subtables);
4087 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
4088 * in 'mask' the values in 'key' and 'target' are the same. */
4090 dpcls_rule_matches_key(const struct dpcls_rule *rule,
4091 const struct netdev_flow_key *target)
4093 const uint64_t *keyp = miniflow_get_values(&rule->flow.mf);
4094 const uint64_t *maskp = miniflow_get_values(&rule->mask->mf);
4097 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, target, rule->flow.mf.map) {
4098 if (OVS_UNLIKELY((value & *maskp++) != *keyp++)) {
4105 /* For each miniflow in 'flows' performs a classifier lookup writing the result
4106 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
4107 * NULL it is skipped.
4109 * This function is optimized for use in the userspace datapath and therefore
4110 * does not implement a lot of features available in the standard
4111 * classifier_lookup() function. Specifically, it does not implement
4112 * priorities, instead returning any rule which matches the flow.
4114 * Returns true if all flows found a corresponding rule. */
4116 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
4117 struct dpcls_rule **rules, const size_t cnt)
4119 /* The batch size 16 was experimentally found faster than 8 or 32. */
4120 typedef uint16_t map_type;
4121 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
4123 #if !defined(__CHECKER__) && !defined(_WIN32)
4124 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
4126 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
4128 map_type maps[N_MAPS];
4129 struct dpcls_subtable *subtable;
4131 memset(maps, 0xff, sizeof maps);
4132 if (cnt % MAP_BITS) {
4133 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
4135 memset(rules, 0, cnt * sizeof *rules);
4137 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
4138 const struct netdev_flow_key *mkeys = keys;
4139 struct dpcls_rule **mrules = rules;
4140 map_type remains = 0;
4143 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
4145 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
4146 uint32_t hashes[MAP_BITS];
4147 const struct cmap_node *nodes[MAP_BITS];
4148 unsigned long map = maps[m];
4152 continue; /* Skip empty maps. */
4155 /* Compute hashes for the remaining keys. */
4156 ULLONG_FOR_EACH_1(i, map) {
4157 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
4161 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
4162 /* Check results. */
4163 ULLONG_FOR_EACH_1(i, map) {
4164 struct dpcls_rule *rule;
4166 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
4167 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
4172 ULLONG_SET0(map, i); /* Did not match. */
4174 ; /* Keep Sparse happy. */
4176 maps[m] &= ~map; /* Clear the found rules. */
4180 return true; /* All found. */
4183 return false; /* Some misses. */