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"
25 #include <netinet/in.h>
29 #include <sys/ioctl.h>
30 #include <sys/socket.h>
38 #include "dp-packet.h"
40 #include "dpif-provider.h"
42 #include "fat-rwlock.h"
47 #include "netdev-dpdk.h"
48 #include "netdev-vport.h"
50 #include "odp-execute.h"
52 #include "openvswitch/dynamic-string.h"
53 #include "openvswitch/list.h"
54 #include "openvswitch/match.h"
55 #include "openvswitch/ofp-print.h"
56 #include "openvswitch/ofpbuf.h"
57 #include "openvswitch/vlog.h"
61 #include "poll-loop.h"
65 #include "openvswitch/shash.h"
68 #include "tnl-neigh-cache.h"
69 #include "tnl-ports.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)
190 const struct dpif_class *const class;
191 const char *const name;
193 struct ovs_refcount ref_cnt;
194 atomic_flag destroyed;
198 * Any lookup into 'ports' or any access to the dp_netdev_ports found
199 * through 'ports' requires taking 'port_mutex'. */
200 struct ovs_mutex port_mutex;
202 struct seq *port_seq; /* Incremented whenever a port changes. */
204 /* Protects access to ofproto-dpif-upcall interface during revalidator
205 * thread synchronization. */
206 struct fat_rwlock upcall_rwlock;
207 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
210 /* Callback function for notifying the purging of dp flows (during
211 * reseting pmd deletion). */
212 dp_purge_callback *dp_purge_cb;
215 /* Stores all 'struct dp_netdev_pmd_thread's. */
216 struct cmap poll_threads;
218 /* Protects the access of the 'struct dp_netdev_pmd_thread'
219 * instance for non-pmd thread. */
220 struct ovs_mutex non_pmd_mutex;
222 /* Each pmd thread will store its pointer to
223 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
224 ovsthread_key_t per_pmd_key;
226 /* Cpu mask for pin of pmd threads. */
227 char *requested_pmd_cmask;
230 uint64_t last_tnl_conf_seq;
233 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
235 OVS_REQUIRES(dp->port_mutex);
238 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
239 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
240 DP_STAT_MISS, /* Packets that did not match. */
241 DP_STAT_LOST, /* Packets not passed up to the client. */
245 enum pmd_cycles_counter_type {
246 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
247 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
251 /* A port in a netdev-based datapath. */
252 struct dp_netdev_port {
254 struct netdev *netdev;
255 struct hmap_node node; /* Node in dp_netdev's 'ports'. */
256 struct netdev_saved_flags *sf;
257 unsigned n_rxq; /* Number of elements in 'rxq' */
258 struct netdev_rxq **rxq;
259 char *type; /* Port type as requested by user. */
262 /* Contained by struct dp_netdev_flow's 'stats' member. */
263 struct dp_netdev_flow_stats {
264 atomic_llong used; /* Last used time, in monotonic msecs. */
265 atomic_ullong packet_count; /* Number of packets matched. */
266 atomic_ullong byte_count; /* Number of bytes matched. */
267 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
270 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
276 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
277 * its pmd thread's classifier. The text below calls this classifier 'cls'.
282 * The thread safety rules described here for "struct dp_netdev_flow" are
283 * motivated by two goals:
285 * - Prevent threads that read members of "struct dp_netdev_flow" from
286 * reading bad data due to changes by some thread concurrently modifying
289 * - Prevent two threads making changes to members of a given "struct
290 * dp_netdev_flow" from interfering with each other.
296 * A flow 'flow' may be accessed without a risk of being freed during an RCU
297 * grace period. Code that needs to hold onto a flow for a while
298 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
300 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
301 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
304 * Some members, marked 'const', are immutable. Accessing other members
305 * requires synchronization, as noted in more detail below.
307 struct dp_netdev_flow {
308 const struct flow flow; /* Unmasked flow that created this entry. */
309 /* Hash table index by unmasked flow. */
310 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
312 const ovs_u128 ufid; /* Unique flow identifier. */
313 const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */
316 /* Number of references.
317 * The classifier owns one reference.
318 * Any thread trying to keep a rule from being freed should hold its own
320 struct ovs_refcount ref_cnt;
325 struct dp_netdev_flow_stats stats;
328 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
330 /* While processing a group of input packets, the datapath uses the next
331 * member to store a pointer to the output batch for the flow. It is
332 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
333 * packet_batch_per_flow_init() and packet_batch_per_flow_execute()). */
334 struct packet_batch_per_flow *batch;
336 /* Packet classification. */
337 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
338 /* 'cr' must be the last member. */
341 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
342 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
343 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
346 /* A set of datapath actions within a "struct dp_netdev_flow".
352 * A struct dp_netdev_actions 'actions' is protected with RCU. */
353 struct dp_netdev_actions {
354 /* These members are immutable: they do not change during the struct's
356 unsigned int size; /* Size of 'actions', in bytes. */
357 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
360 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
362 struct dp_netdev_actions *dp_netdev_flow_get_actions(
363 const struct dp_netdev_flow *);
364 static void dp_netdev_actions_free(struct dp_netdev_actions *);
366 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
367 struct dp_netdev_pmd_stats {
368 /* Indexed by DP_STAT_*. */
369 atomic_ullong n[DP_N_STATS];
372 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
373 struct dp_netdev_pmd_cycles {
374 /* Indexed by PMD_CYCLES_*. */
375 atomic_ullong n[PMD_N_CYCLES];
378 /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */
380 struct dp_netdev_port *port;
381 struct netdev_rxq *rx;
382 struct ovs_list node;
385 /* Contained by struct dp_netdev_pmd_thread's 'port_cache' or 'tx_ports'. */
388 struct netdev *netdev;
389 struct hmap_node node;
392 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
393 * the performance overhead of interrupt processing. Therefore netdev can
394 * not implement rx-wait for these devices. dpif-netdev needs to poll
395 * these device to check for recv buffer. pmd-thread does polling for
396 * devices assigned to itself.
398 * DPDK used PMD for accessing NIC.
400 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
401 * I/O of all non-pmd threads. There will be no actual thread created
404 * Each struct has its own flow table and classifier. Packets received
405 * from managed ports are looked up in the corresponding pmd thread's
406 * flow table, and are executed with the found actions.
408 struct dp_netdev_pmd_thread {
409 struct dp_netdev *dp;
410 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
411 struct cmap_node node; /* In 'dp->poll_threads'. */
413 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
414 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
416 /* Per thread exact-match cache. Note, the instance for cpu core
417 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
418 * need to be protected by 'non_pmd_mutex'. Every other instance
419 * will only be accessed by its own pmd thread. */
420 struct emc_cache flow_cache;
422 /* Classifier and Flow-Table.
424 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
425 * changes to 'cls' must be made while still holding the 'flow_mutex'.
427 struct ovs_mutex flow_mutex;
429 struct cmap flow_table OVS_GUARDED; /* Flow table. */
432 struct dp_netdev_pmd_stats stats;
434 /* Cycles counters */
435 struct dp_netdev_pmd_cycles cycles;
437 /* Used to count cicles. See 'cycles_counter_end()' */
438 unsigned long long last_cycles;
440 struct latch exit_latch; /* For terminating the pmd thread. */
441 atomic_uint change_seq; /* For reloading pmd ports. */
443 unsigned core_id; /* CPU core id of this pmd thread. */
444 int numa_id; /* numa node id of this pmd thread. */
446 /* Queue id used by this pmd thread to send packets on all netdevs.
447 * All tx_qid's are unique and less than 'ovs_numa_get_n_cores() + 1'. */
450 struct ovs_mutex port_mutex; /* Mutex for 'poll_list' and 'tx_ports'. */
451 /* List of rx queues to poll. */
452 struct ovs_list poll_list OVS_GUARDED;
453 /* Number of elements in 'poll_list' */
455 /* Map of 'tx_port's used for transmission. Written by the main thread,
456 * read by the pmd thread. */
457 struct hmap tx_ports OVS_GUARDED;
459 /* Map of 'tx_port' used in the fast path. This is a thread-local copy of
460 * 'tx_ports'. The instance for cpu core NON_PMD_CORE_ID can be accessed
461 * by multiple threads, and thusly need to be protected by 'non_pmd_mutex'.
462 * Every other instance will only be accessed by its own pmd thread. */
463 struct hmap port_cache;
465 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
466 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
467 * values and subtracts them from 'stats' and 'cycles' before
468 * reporting to the user */
469 unsigned long long stats_zero[DP_N_STATS];
470 uint64_t cycles_zero[PMD_N_CYCLES];
473 #define PMD_INITIAL_SEQ 1
475 /* Interface to netdev-based datapath. */
478 struct dp_netdev *dp;
479 uint64_t last_port_seq;
482 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
483 struct dp_netdev_port **portp)
484 OVS_REQUIRES(dp->port_mutex);
485 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
486 struct dp_netdev_port **portp)
487 OVS_REQUIRES(dp->port_mutex);
488 static void dp_netdev_free(struct dp_netdev *)
489 OVS_REQUIRES(dp_netdev_mutex);
490 static int do_add_port(struct dp_netdev *dp, const char *devname,
491 const char *type, odp_port_t port_no)
492 OVS_REQUIRES(dp->port_mutex);
493 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
494 OVS_REQUIRES(dp->port_mutex);
495 static int dpif_netdev_open(const struct dpif_class *, const char *name,
496 bool create, struct dpif **);
497 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
498 struct dp_packet_batch *,
500 const struct nlattr *actions,
502 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
503 struct dp_packet_batch *, odp_port_t port_no);
504 static void dp_netdev_recirculate(struct dp_netdev_pmd_thread *,
505 struct dp_packet_batch *);
507 static void dp_netdev_disable_upcall(struct dp_netdev *);
508 static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
509 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
510 struct dp_netdev *dp, unsigned core_id,
512 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
513 static void dp_netdev_set_nonpmd(struct dp_netdev *dp)
514 OVS_REQUIRES(dp->port_mutex);
516 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
518 static struct dp_netdev_pmd_thread *
519 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
520 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
521 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
522 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
523 OVS_REQUIRES(dp->port_mutex);
524 static void dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread *pmd);
525 static void dp_netdev_del_port_from_all_pmds(struct dp_netdev *dp,
526 struct dp_netdev_port *port);
527 static void dp_netdev_add_port_to_pmds(struct dp_netdev *dp,
528 struct dp_netdev_port *port);
529 static void dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread *pmd,
530 struct dp_netdev_port *port);
531 static void dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
532 struct dp_netdev_port *port,
533 struct netdev_rxq *rx);
534 static struct dp_netdev_pmd_thread *
535 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id);
536 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
537 OVS_REQUIRES(dp->port_mutex);
538 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
539 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
540 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
541 static void pmd_load_cached_ports(struct dp_netdev_pmd_thread *pmd)
542 OVS_REQUIRES(pmd->port_mutex);
544 static inline bool emc_entry_alive(struct emc_entry *ce);
545 static void emc_clear_entry(struct emc_entry *ce);
548 emc_cache_init(struct emc_cache *flow_cache)
552 flow_cache->sweep_idx = 0;
553 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
554 flow_cache->entries[i].flow = NULL;
555 flow_cache->entries[i].key.hash = 0;
556 flow_cache->entries[i].key.len = sizeof(struct miniflow);
557 flowmap_init(&flow_cache->entries[i].key.mf.map);
562 emc_cache_uninit(struct emc_cache *flow_cache)
566 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
567 emc_clear_entry(&flow_cache->entries[i]);
571 /* Check and clear dead flow references slowly (one entry at each
574 emc_cache_slow_sweep(struct emc_cache *flow_cache)
576 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
578 if (!emc_entry_alive(entry)) {
579 emc_clear_entry(entry);
581 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
584 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
586 dpif_is_netdev(const struct dpif *dpif)
588 return dpif->dpif_class->open == dpif_netdev_open;
591 static struct dpif_netdev *
592 dpif_netdev_cast(const struct dpif *dpif)
594 ovs_assert(dpif_is_netdev(dpif));
595 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
598 static struct dp_netdev *
599 get_dp_netdev(const struct dpif *dpif)
601 return dpif_netdev_cast(dpif)->dp;
605 PMD_INFO_SHOW_STATS, /* Show how cpu cycles are spent. */
606 PMD_INFO_CLEAR_STATS, /* Set the cycles count to 0. */
607 PMD_INFO_SHOW_RXQ /* Show poll-lists of pmd threads. */
611 pmd_info_show_stats(struct ds *reply,
612 struct dp_netdev_pmd_thread *pmd,
613 unsigned long long stats[DP_N_STATS],
614 uint64_t cycles[PMD_N_CYCLES])
616 unsigned long long total_packets = 0;
617 uint64_t total_cycles = 0;
620 /* These loops subtracts reference values ('*_zero') from the counters.
621 * Since loads and stores are relaxed, it might be possible for a '*_zero'
622 * value to be more recent than the current value we're reading from the
623 * counter. This is not a big problem, since these numbers are not
624 * supposed to be too accurate, but we should at least make sure that
625 * the result is not negative. */
626 for (i = 0; i < DP_N_STATS; i++) {
627 if (stats[i] > pmd->stats_zero[i]) {
628 stats[i] -= pmd->stats_zero[i];
633 if (i != DP_STAT_LOST) {
634 /* Lost packets are already included in DP_STAT_MISS */
635 total_packets += stats[i];
639 for (i = 0; i < PMD_N_CYCLES; i++) {
640 if (cycles[i] > pmd->cycles_zero[i]) {
641 cycles[i] -= pmd->cycles_zero[i];
646 total_cycles += cycles[i];
649 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
650 ? "main thread" : "pmd thread");
652 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
653 ds_put_format(reply, " numa_id %d", pmd->numa_id);
655 if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) {
656 ds_put_format(reply, " core_id %u", pmd->core_id);
658 ds_put_cstr(reply, ":\n");
661 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
662 "\tmiss:%llu\n\tlost:%llu\n",
663 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
664 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
666 if (total_cycles == 0) {
671 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
672 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
673 cycles[PMD_CYCLES_POLLING],
674 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
675 cycles[PMD_CYCLES_PROCESSING],
676 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
678 if (total_packets == 0) {
683 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
684 total_cycles / (double)total_packets,
685 total_cycles, total_packets);
688 "\tavg processing cycles per packet: "
689 "%.02f (%"PRIu64"/%llu)\n",
690 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
691 cycles[PMD_CYCLES_PROCESSING], total_packets);
695 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
696 struct dp_netdev_pmd_thread *pmd,
697 unsigned long long stats[DP_N_STATS],
698 uint64_t cycles[PMD_N_CYCLES])
702 /* We cannot write 'stats' and 'cycles' (because they're written by other
703 * threads) and we shouldn't change 'stats' (because they're used to count
704 * datapath stats, which must not be cleared here). Instead, we save the
705 * current values and subtract them from the values to be displayed in the
707 for (i = 0; i < DP_N_STATS; i++) {
708 pmd->stats_zero[i] = stats[i];
710 for (i = 0; i < PMD_N_CYCLES; i++) {
711 pmd->cycles_zero[i] = cycles[i];
716 pmd_info_show_rxq(struct ds *reply, struct dp_netdev_pmd_thread *pmd)
718 if (pmd->core_id != NON_PMD_CORE_ID) {
719 struct rxq_poll *poll;
720 const char *prev_name = NULL;
722 ds_put_format(reply, "pmd thread numa_id %d core_id %u:\n",
723 pmd->numa_id, pmd->core_id);
725 ovs_mutex_lock(&pmd->port_mutex);
726 LIST_FOR_EACH (poll, node, &pmd->poll_list) {
727 const char *name = netdev_get_name(poll->port->netdev);
729 if (!prev_name || strcmp(name, prev_name)) {
731 ds_put_cstr(reply, "\n");
733 ds_put_format(reply, "\tport: %s\tqueue-id:",
734 netdev_get_name(poll->port->netdev));
736 ds_put_format(reply, " %d", netdev_rxq_get_queue_id(poll->rx));
739 ovs_mutex_unlock(&pmd->port_mutex);
740 ds_put_cstr(reply, "\n");
745 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
748 struct ds reply = DS_EMPTY_INITIALIZER;
749 struct dp_netdev_pmd_thread *pmd;
750 struct dp_netdev *dp = NULL;
751 enum pmd_info_type type = *(enum pmd_info_type *) aux;
753 ovs_mutex_lock(&dp_netdev_mutex);
756 dp = shash_find_data(&dp_netdevs, argv[1]);
757 } else if (shash_count(&dp_netdevs) == 1) {
758 /* There's only one datapath */
759 dp = shash_first(&dp_netdevs)->data;
763 ovs_mutex_unlock(&dp_netdev_mutex);
764 unixctl_command_reply_error(conn,
765 "please specify an existing datapath");
769 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
770 if (type == PMD_INFO_SHOW_RXQ) {
771 pmd_info_show_rxq(&reply, pmd);
773 unsigned long long stats[DP_N_STATS];
774 uint64_t cycles[PMD_N_CYCLES];
777 /* Read current stats and cycle counters */
778 for (i = 0; i < ARRAY_SIZE(stats); i++) {
779 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
781 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
782 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
785 if (type == PMD_INFO_CLEAR_STATS) {
786 pmd_info_clear_stats(&reply, pmd, stats, cycles);
787 } else if (type == PMD_INFO_SHOW_STATS) {
788 pmd_info_show_stats(&reply, pmd, stats, cycles);
793 ovs_mutex_unlock(&dp_netdev_mutex);
795 unixctl_command_reply(conn, ds_cstr(&reply));
800 dpif_netdev_init(void)
802 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
803 clear_aux = PMD_INFO_CLEAR_STATS,
804 poll_aux = PMD_INFO_SHOW_RXQ;
806 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
807 0, 1, dpif_netdev_pmd_info,
809 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
810 0, 1, dpif_netdev_pmd_info,
812 unixctl_command_register("dpif-netdev/pmd-rxq-show", "[dp]",
813 0, 1, dpif_netdev_pmd_info,
819 dpif_netdev_enumerate(struct sset *all_dps,
820 const struct dpif_class *dpif_class)
822 struct shash_node *node;
824 ovs_mutex_lock(&dp_netdev_mutex);
825 SHASH_FOR_EACH(node, &dp_netdevs) {
826 struct dp_netdev *dp = node->data;
827 if (dpif_class != dp->class) {
828 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
829 * If the class doesn't match, skip this dpif. */
832 sset_add(all_dps, node->name);
834 ovs_mutex_unlock(&dp_netdev_mutex);
840 dpif_netdev_class_is_dummy(const struct dpif_class *class)
842 return class != &dpif_netdev_class;
846 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
848 return strcmp(type, "internal") ? type
849 : dpif_netdev_class_is_dummy(class) ? "dummy"
854 create_dpif_netdev(struct dp_netdev *dp)
856 uint16_t netflow_id = hash_string(dp->name, 0);
857 struct dpif_netdev *dpif;
859 ovs_refcount_ref(&dp->ref_cnt);
861 dpif = xmalloc(sizeof *dpif);
862 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
864 dpif->last_port_seq = seq_read(dp->port_seq);
869 /* Choose an unused, non-zero port number and return it on success.
870 * Return ODPP_NONE on failure. */
872 choose_port(struct dp_netdev *dp, const char *name)
873 OVS_REQUIRES(dp->port_mutex)
877 if (dp->class != &dpif_netdev_class) {
881 /* If the port name begins with "br", start the number search at
882 * 100 to make writing tests easier. */
883 if (!strncmp(name, "br", 2)) {
887 /* If the port name contains a number, try to assign that port number.
888 * This can make writing unit tests easier because port numbers are
890 for (p = name; *p != '\0'; p++) {
891 if (isdigit((unsigned char) *p)) {
892 port_no = start_no + strtol(p, NULL, 10);
893 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
894 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
895 return u32_to_odp(port_no);
902 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
903 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
904 return u32_to_odp(port_no);
912 create_dp_netdev(const char *name, const struct dpif_class *class,
913 struct dp_netdev **dpp)
914 OVS_REQUIRES(dp_netdev_mutex)
916 struct dp_netdev *dp;
919 dp = xzalloc(sizeof *dp);
920 shash_add(&dp_netdevs, name, dp);
922 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
923 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
924 ovs_refcount_init(&dp->ref_cnt);
925 atomic_flag_clear(&dp->destroyed);
927 ovs_mutex_init(&dp->port_mutex);
928 hmap_init(&dp->ports);
929 dp->port_seq = seq_create();
930 fat_rwlock_init(&dp->upcall_rwlock);
932 /* Disable upcalls by default. */
933 dp_netdev_disable_upcall(dp);
934 dp->upcall_aux = NULL;
935 dp->upcall_cb = NULL;
937 cmap_init(&dp->poll_threads);
938 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
939 ovsthread_key_create(&dp->per_pmd_key, NULL);
941 ovs_mutex_lock(&dp->port_mutex);
942 dp_netdev_set_nonpmd(dp);
944 error = do_add_port(dp, name, dpif_netdev_port_open_type(dp->class,
947 ovs_mutex_unlock(&dp->port_mutex);
953 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
959 dpif_netdev_open(const struct dpif_class *class, const char *name,
960 bool create, struct dpif **dpifp)
962 struct dp_netdev *dp;
965 ovs_mutex_lock(&dp_netdev_mutex);
966 dp = shash_find_data(&dp_netdevs, name);
968 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
970 error = (dp->class != class ? EINVAL
975 *dpifp = create_dpif_netdev(dp);
978 ovs_mutex_unlock(&dp_netdev_mutex);
984 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
985 OVS_NO_THREAD_SAFETY_ANALYSIS
987 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
988 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
990 /* Before freeing a lock we should release it */
991 fat_rwlock_unlock(&dp->upcall_rwlock);
992 fat_rwlock_destroy(&dp->upcall_rwlock);
995 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
996 * through the 'dp_netdevs' shash while freeing 'dp'. */
998 dp_netdev_free(struct dp_netdev *dp)
999 OVS_REQUIRES(dp_netdev_mutex)
1001 struct dp_netdev_port *port, *next;
1003 shash_find_and_delete(&dp_netdevs, dp->name);
1005 dp_netdev_destroy_all_pmds(dp);
1006 ovs_mutex_destroy(&dp->non_pmd_mutex);
1007 ovsthread_key_delete(dp->per_pmd_key);
1009 ovs_mutex_lock(&dp->port_mutex);
1010 HMAP_FOR_EACH_SAFE (port, next, node, &dp->ports) {
1011 do_del_port(dp, port);
1013 ovs_mutex_unlock(&dp->port_mutex);
1014 cmap_destroy(&dp->poll_threads);
1016 seq_destroy(dp->port_seq);
1017 hmap_destroy(&dp->ports);
1018 ovs_mutex_destroy(&dp->port_mutex);
1020 /* Upcalls must be disabled at this point */
1021 dp_netdev_destroy_upcall_lock(dp);
1023 free(dp->pmd_cmask);
1024 free(CONST_CAST(char *, dp->name));
1029 dp_netdev_unref(struct dp_netdev *dp)
1032 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
1033 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
1034 ovs_mutex_lock(&dp_netdev_mutex);
1035 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
1038 ovs_mutex_unlock(&dp_netdev_mutex);
1043 dpif_netdev_close(struct dpif *dpif)
1045 struct dp_netdev *dp = get_dp_netdev(dpif);
1047 dp_netdev_unref(dp);
1052 dpif_netdev_destroy(struct dpif *dpif)
1054 struct dp_netdev *dp = get_dp_netdev(dpif);
1056 if (!atomic_flag_test_and_set(&dp->destroyed)) {
1057 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
1058 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
1066 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
1067 * load/store semantics. While the increment is not atomic, the load and
1068 * store operations are, making it impossible to read inconsistent values.
1070 * This is used to update thread local stats counters. */
1072 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
1074 unsigned long long tmp;
1076 atomic_read_relaxed(var, &tmp);
1078 atomic_store_relaxed(var, tmp);
1082 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
1084 struct dp_netdev *dp = get_dp_netdev(dpif);
1085 struct dp_netdev_pmd_thread *pmd;
1087 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
1088 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1089 unsigned long long n;
1090 stats->n_flows += cmap_count(&pmd->flow_table);
1092 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
1094 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
1096 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
1097 stats->n_missed += n;
1098 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
1101 stats->n_masks = UINT32_MAX;
1102 stats->n_mask_hit = UINT64_MAX;
1108 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1112 if (pmd->core_id == NON_PMD_CORE_ID) {
1113 ovs_mutex_lock(&pmd->dp->non_pmd_mutex);
1114 ovs_mutex_lock(&pmd->port_mutex);
1115 pmd_load_cached_ports(pmd);
1116 ovs_mutex_unlock(&pmd->port_mutex);
1117 ovs_mutex_unlock(&pmd->dp->non_pmd_mutex);
1121 ovs_mutex_lock(&pmd->cond_mutex);
1122 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1123 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1124 ovs_mutex_unlock(&pmd->cond_mutex);
1128 hash_port_no(odp_port_t port_no)
1130 return hash_int(odp_to_u32(port_no), 0);
1134 port_create(const char *devname, const char *type,
1135 odp_port_t port_no, struct dp_netdev_port **portp)
1137 struct netdev_saved_flags *sf;
1138 struct dp_netdev_port *port;
1139 enum netdev_flags flags;
1140 struct netdev *netdev;
1141 int n_open_rxqs = 0;
1146 /* Open and validate network device. */
1147 error = netdev_open(devname, type, &netdev);
1151 /* XXX reject non-Ethernet devices */
1153 netdev_get_flags(netdev, &flags);
1154 if (flags & NETDEV_LOOPBACK) {
1155 VLOG_ERR("%s: cannot add a loopback device", devname);
1160 if (netdev_is_pmd(netdev)) {
1161 int n_cores = ovs_numa_get_n_cores();
1163 if (n_cores == OVS_CORE_UNSPEC) {
1164 VLOG_ERR("%s, cannot get cpu core info", devname);
1168 /* There can only be ovs_numa_get_n_cores() pmd threads,
1169 * so creates a txq for each, and one extra for the non
1171 error = netdev_set_tx_multiq(netdev, n_cores + 1);
1172 if (error && (error != EOPNOTSUPP)) {
1173 VLOG_ERR("%s, cannot set multiq", devname);
1178 if (netdev_is_reconf_required(netdev)) {
1179 error = netdev_reconfigure(netdev);
1185 port = xzalloc(sizeof *port);
1186 port->port_no = port_no;
1187 port->netdev = netdev;
1188 port->n_rxq = netdev_n_rxq(netdev);
1189 port->rxq = xcalloc(port->n_rxq, sizeof *port->rxq);
1190 port->type = xstrdup(type);
1192 for (i = 0; i < port->n_rxq; i++) {
1193 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1195 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1196 devname, ovs_strerror(errno));
1202 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1213 for (i = 0; i < n_open_rxqs; i++) {
1214 netdev_rxq_close(port->rxq[i]);
1221 netdev_close(netdev);
1226 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1228 OVS_REQUIRES(dp->port_mutex)
1230 struct dp_netdev_port *port;
1233 /* Reject devices already in 'dp'. */
1234 if (!get_port_by_name(dp, devname, &port)) {
1238 error = port_create(devname, type, port_no, &port);
1243 if (netdev_is_pmd(port->netdev)) {
1244 int numa_id = netdev_get_numa_id(port->netdev);
1246 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
1247 dp_netdev_set_pmds_on_numa(dp, numa_id);
1250 dp_netdev_add_port_to_pmds(dp, port);
1252 hmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1253 seq_change(dp->port_seq);
1259 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1260 odp_port_t *port_nop)
1262 struct dp_netdev *dp = get_dp_netdev(dpif);
1263 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1264 const char *dpif_port;
1268 ovs_mutex_lock(&dp->port_mutex);
1269 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1270 if (*port_nop != ODPP_NONE) {
1271 port_no = *port_nop;
1272 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1274 port_no = choose_port(dp, dpif_port);
1275 error = port_no == ODPP_NONE ? EFBIG : 0;
1278 *port_nop = port_no;
1279 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1281 ovs_mutex_unlock(&dp->port_mutex);
1287 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1289 struct dp_netdev *dp = get_dp_netdev(dpif);
1292 ovs_mutex_lock(&dp->port_mutex);
1293 if (port_no == ODPP_LOCAL) {
1296 struct dp_netdev_port *port;
1298 error = get_port_by_number(dp, port_no, &port);
1300 do_del_port(dp, port);
1303 ovs_mutex_unlock(&dp->port_mutex);
1309 is_valid_port_number(odp_port_t port_no)
1311 return port_no != ODPP_NONE;
1314 static struct dp_netdev_port *
1315 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1316 OVS_REQUIRES(dp->port_mutex)
1318 struct dp_netdev_port *port;
1320 HMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1321 if (port->port_no == port_no) {
1329 get_port_by_number(struct dp_netdev *dp,
1330 odp_port_t port_no, struct dp_netdev_port **portp)
1331 OVS_REQUIRES(dp->port_mutex)
1333 if (!is_valid_port_number(port_no)) {
1337 *portp = dp_netdev_lookup_port(dp, port_no);
1338 return *portp ? 0 : ENOENT;
1343 port_destroy(struct dp_netdev_port *port)
1349 netdev_close(port->netdev);
1350 netdev_restore_flags(port->sf);
1352 for (unsigned i = 0; i < port->n_rxq; i++) {
1353 netdev_rxq_close(port->rxq[i]);
1362 get_port_by_name(struct dp_netdev *dp,
1363 const char *devname, struct dp_netdev_port **portp)
1364 OVS_REQUIRES(dp->port_mutex)
1366 struct dp_netdev_port *port;
1368 HMAP_FOR_EACH (port, node, &dp->ports) {
1369 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1378 get_n_pmd_threads(struct dp_netdev *dp)
1380 /* There is one non pmd thread in dp->poll_threads */
1381 return cmap_count(&dp->poll_threads) - 1;
1385 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1387 struct dp_netdev_pmd_thread *pmd;
1390 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1391 if (pmd->numa_id == numa_id) {
1399 /* Returns 'true' if there is a port with pmd netdev and the netdev
1400 * is on numa node 'numa_id'. */
1402 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1403 OVS_REQUIRES(dp->port_mutex)
1405 struct dp_netdev_port *port;
1407 HMAP_FOR_EACH (port, node, &dp->ports) {
1408 if (netdev_is_pmd(port->netdev)
1409 && netdev_get_numa_id(port->netdev) == numa_id) {
1419 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1420 OVS_REQUIRES(dp->port_mutex)
1422 hmap_remove(&dp->ports, &port->node);
1423 seq_change(dp->port_seq);
1425 dp_netdev_del_port_from_all_pmds(dp, port);
1427 if (netdev_is_pmd(port->netdev)) {
1428 int numa_id = netdev_get_numa_id(port->netdev);
1430 /* PMD threads can not be on invalid numa node. */
1431 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
1432 /* If there is no netdev on the numa node, deletes the pmd threads
1434 if (!has_pmd_port_for_numa(dp, numa_id)) {
1435 dp_netdev_del_pmds_on_numa(dp, numa_id);
1443 answer_port_query(const struct dp_netdev_port *port,
1444 struct dpif_port *dpif_port)
1446 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1447 dpif_port->type = xstrdup(port->type);
1448 dpif_port->port_no = port->port_no;
1452 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1453 struct dpif_port *dpif_port)
1455 struct dp_netdev *dp = get_dp_netdev(dpif);
1456 struct dp_netdev_port *port;
1459 ovs_mutex_lock(&dp->port_mutex);
1460 error = get_port_by_number(dp, port_no, &port);
1461 if (!error && dpif_port) {
1462 answer_port_query(port, dpif_port);
1464 ovs_mutex_unlock(&dp->port_mutex);
1470 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1471 struct dpif_port *dpif_port)
1473 struct dp_netdev *dp = get_dp_netdev(dpif);
1474 struct dp_netdev_port *port;
1477 ovs_mutex_lock(&dp->port_mutex);
1478 error = get_port_by_name(dp, devname, &port);
1479 if (!error && dpif_port) {
1480 answer_port_query(port, dpif_port);
1482 ovs_mutex_unlock(&dp->port_mutex);
1488 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1490 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1494 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1496 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1497 ovsrcu_postpone(dp_netdev_flow_free, flow);
1502 dp_netdev_flow_hash(const ovs_u128 *ufid)
1504 return ufid->u32[0];
1508 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1509 struct dp_netdev_flow *flow)
1510 OVS_REQUIRES(pmd->flow_mutex)
1512 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1514 dpcls_remove(&pmd->cls, &flow->cr);
1515 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1518 dp_netdev_flow_unref(flow);
1522 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1524 struct dp_netdev_flow *netdev_flow;
1526 ovs_mutex_lock(&pmd->flow_mutex);
1527 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1528 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1530 ovs_mutex_unlock(&pmd->flow_mutex);
1534 dpif_netdev_flow_flush(struct dpif *dpif)
1536 struct dp_netdev *dp = get_dp_netdev(dpif);
1537 struct dp_netdev_pmd_thread *pmd;
1539 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1540 dp_netdev_pmd_flow_flush(pmd);
1546 struct dp_netdev_port_state {
1547 struct hmap_position position;
1552 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1554 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1559 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1560 struct dpif_port *dpif_port)
1562 struct dp_netdev_port_state *state = state_;
1563 struct dp_netdev *dp = get_dp_netdev(dpif);
1564 struct hmap_node *node;
1567 ovs_mutex_lock(&dp->port_mutex);
1568 node = hmap_at_position(&dp->ports, &state->position);
1570 struct dp_netdev_port *port;
1572 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1575 state->name = xstrdup(netdev_get_name(port->netdev));
1576 dpif_port->name = state->name;
1577 dpif_port->type = port->type;
1578 dpif_port->port_no = port->port_no;
1584 ovs_mutex_unlock(&dp->port_mutex);
1590 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1592 struct dp_netdev_port_state *state = state_;
1599 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1601 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1602 uint64_t new_port_seq;
1605 new_port_seq = seq_read(dpif->dp->port_seq);
1606 if (dpif->last_port_seq != new_port_seq) {
1607 dpif->last_port_seq = new_port_seq;
1617 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1619 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1621 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1624 static struct dp_netdev_flow *
1625 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1627 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1630 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1632 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1635 /* netdev_flow_key utilities.
1637 * netdev_flow_key is basically a miniflow. We use these functions
1638 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1639 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1641 * - Since we are dealing exclusively with miniflows created by
1642 * miniflow_extract(), if the map is different the miniflow is different.
1643 * Therefore we can be faster by comparing the map and the miniflow in a
1645 * - These functions can be inlined by the compiler. */
1647 /* Given the number of bits set in miniflow's maps, returns the size of the
1648 * 'netdev_flow_key.mf' */
1649 static inline size_t
1650 netdev_flow_key_size(size_t flow_u64s)
1652 return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s);
1656 netdev_flow_key_equal(const struct netdev_flow_key *a,
1657 const struct netdev_flow_key *b)
1659 /* 'b->len' may be not set yet. */
1660 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1663 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1664 * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been
1665 * generated by miniflow_extract. */
1667 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1668 const struct miniflow *mf)
1670 return !memcmp(&key->mf, mf, key->len);
1674 netdev_flow_key_clone(struct netdev_flow_key *dst,
1675 const struct netdev_flow_key *src)
1678 offsetof(struct netdev_flow_key, mf) + src->len);
1683 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1684 const struct flow *src)
1686 struct dp_packet packet;
1687 uint64_t buf_stub[512 / 8];
1689 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1690 pkt_metadata_from_flow(&packet.md, src);
1691 flow_compose(&packet, src);
1692 miniflow_extract(&packet, &dst->mf);
1693 dp_packet_uninit(&packet);
1695 dst->len = netdev_flow_key_size(miniflow_n_values(&dst->mf));
1696 dst->hash = 0; /* Not computed yet. */
1699 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1701 netdev_flow_mask_init(struct netdev_flow_key *mask,
1702 const struct match *match)
1704 uint64_t *dst = miniflow_values(&mask->mf);
1705 struct flowmap fmap;
1709 /* Only check masks that make sense for the flow. */
1710 flow_wc_map(&match->flow, &fmap);
1711 flowmap_init(&mask->mf.map);
1713 FLOWMAP_FOR_EACH_INDEX(idx, fmap) {
1714 uint64_t mask_u64 = flow_u64_value(&match->wc.masks, idx);
1717 flowmap_set(&mask->mf.map, idx, 1);
1719 hash = hash_add64(hash, mask_u64);
1725 FLOWMAP_FOR_EACH_MAP (map, mask->mf.map) {
1726 hash = hash_add64(hash, map);
1729 size_t n = dst - miniflow_get_values(&mask->mf);
1731 mask->hash = hash_finish(hash, n * 8);
1732 mask->len = netdev_flow_key_size(n);
1735 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1737 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1738 const struct flow *flow,
1739 const struct netdev_flow_key *mask)
1741 uint64_t *dst_u64 = miniflow_values(&dst->mf);
1742 const uint64_t *mask_u64 = miniflow_get_values(&mask->mf);
1746 dst->len = mask->len;
1747 dst->mf = mask->mf; /* Copy maps. */
1749 FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf.map) {
1750 *dst_u64 = value & *mask_u64++;
1751 hash = hash_add64(hash, *dst_u64++);
1753 dst->hash = hash_finish(hash,
1754 (dst_u64 - miniflow_get_values(&dst->mf)) * 8);
1757 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
1758 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
1759 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
1761 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1763 static inline uint32_t
1764 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1765 const struct netdev_flow_key *mask)
1767 const uint64_t *p = miniflow_get_values(&mask->mf);
1771 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, key, mask->mf.map) {
1772 hash = hash_add64(hash, value & *p++);
1775 return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8);
1779 emc_entry_alive(struct emc_entry *ce)
1781 return ce->flow && !ce->flow->dead;
1785 emc_clear_entry(struct emc_entry *ce)
1788 dp_netdev_flow_unref(ce->flow);
1794 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1795 const struct netdev_flow_key *key)
1797 if (ce->flow != flow) {
1799 dp_netdev_flow_unref(ce->flow);
1802 if (dp_netdev_flow_ref(flow)) {
1809 netdev_flow_key_clone(&ce->key, key);
1814 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1815 struct dp_netdev_flow *flow)
1817 struct emc_entry *to_be_replaced = NULL;
1818 struct emc_entry *current_entry;
1820 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1821 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1822 /* We found the entry with the 'mf' miniflow */
1823 emc_change_entry(current_entry, flow, NULL);
1827 /* Replacement policy: put the flow in an empty (not alive) entry, or
1828 * in the first entry where it can be */
1830 || (emc_entry_alive(to_be_replaced)
1831 && !emc_entry_alive(current_entry))
1832 || current_entry->key.hash < to_be_replaced->key.hash) {
1833 to_be_replaced = current_entry;
1836 /* We didn't find the miniflow in the cache.
1837 * The 'to_be_replaced' entry is where the new flow will be stored */
1839 emc_change_entry(to_be_replaced, flow, key);
1842 static inline struct dp_netdev_flow *
1843 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1845 struct emc_entry *current_entry;
1847 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1848 if (current_entry->key.hash == key->hash
1849 && emc_entry_alive(current_entry)
1850 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1852 /* We found the entry with the 'key->mf' miniflow */
1853 return current_entry->flow;
1860 static struct dp_netdev_flow *
1861 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1862 const struct netdev_flow_key *key)
1864 struct dp_netdev_flow *netdev_flow;
1865 struct dpcls_rule *rule;
1867 dpcls_lookup(&pmd->cls, key, &rule, 1);
1868 netdev_flow = dp_netdev_flow_cast(rule);
1873 static struct dp_netdev_flow *
1874 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1875 const ovs_u128 *ufidp, const struct nlattr *key,
1878 struct dp_netdev_flow *netdev_flow;
1882 /* If a UFID is not provided, determine one based on the key. */
1883 if (!ufidp && key && key_len
1884 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1885 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1890 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1892 if (ovs_u128_equals(netdev_flow->ufid, *ufidp)) {
1902 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1903 struct dpif_flow_stats *stats)
1905 struct dp_netdev_flow *netdev_flow;
1906 unsigned long long n;
1910 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1912 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1913 stats->n_packets = n;
1914 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1916 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1918 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1919 stats->tcp_flags = flags;
1922 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1923 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1924 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1927 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1928 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1929 struct dpif_flow *flow, bool terse)
1932 memset(flow, 0, sizeof *flow);
1934 struct flow_wildcards wc;
1935 struct dp_netdev_actions *actions;
1937 struct odp_flow_key_parms odp_parms = {
1938 .flow = &netdev_flow->flow,
1940 .support = dp_netdev_support,
1943 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1946 offset = key_buf->size;
1947 flow->key = ofpbuf_tail(key_buf);
1948 odp_flow_key_from_flow(&odp_parms, key_buf);
1949 flow->key_len = key_buf->size - offset;
1952 offset = mask_buf->size;
1953 flow->mask = ofpbuf_tail(mask_buf);
1954 odp_parms.key_buf = key_buf;
1955 odp_flow_key_from_mask(&odp_parms, mask_buf);
1956 flow->mask_len = mask_buf->size - offset;
1959 actions = dp_netdev_flow_get_actions(netdev_flow);
1960 flow->actions = actions->actions;
1961 flow->actions_len = actions->size;
1964 flow->ufid = netdev_flow->ufid;
1965 flow->ufid_present = true;
1966 flow->pmd_id = netdev_flow->pmd_id;
1967 get_dpif_flow_stats(netdev_flow, &flow->stats);
1971 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1972 const struct nlattr *mask_key,
1973 uint32_t mask_key_len, const struct flow *flow,
1974 struct flow_wildcards *wc)
1976 enum odp_key_fitness fitness;
1978 fitness = odp_flow_key_to_mask_udpif(mask_key, mask_key_len, key,
1981 /* This should not happen: it indicates that
1982 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1983 * disagree on the acceptable form of a mask. Log the problem
1984 * as an error, with enough details to enable debugging. */
1985 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1987 if (!VLOG_DROP_ERR(&rl)) {
1991 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1993 VLOG_ERR("internal error parsing flow mask %s (%s)",
1994 ds_cstr(&s), odp_key_fitness_to_string(fitness));
2005 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
2010 if (odp_flow_key_to_flow_udpif(key, key_len, flow)) {
2011 /* This should not happen: it indicates that odp_flow_key_from_flow()
2012 * and odp_flow_key_to_flow() disagree on the acceptable form of a
2013 * flow. Log the problem as an error, with enough details to enable
2015 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2017 if (!VLOG_DROP_ERR(&rl)) {
2021 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
2022 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
2029 in_port = flow->in_port.odp_port;
2030 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
2034 /* Userspace datapath doesn't support conntrack. */
2035 if (flow->ct_state || flow->ct_zone || flow->ct_mark
2036 || !ovs_u128_is_zero(flow->ct_label)) {
2044 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
2046 struct dp_netdev *dp = get_dp_netdev(dpif);
2047 struct dp_netdev_flow *netdev_flow;
2048 struct dp_netdev_pmd_thread *pmd;
2049 struct hmapx to_find = HMAPX_INITIALIZER(&to_find);
2050 struct hmapx_node *node;
2053 if (get->pmd_id == PMD_ID_NULL) {
2054 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2055 if (dp_netdev_pmd_try_ref(pmd) && !hmapx_add(&to_find, pmd)) {
2056 dp_netdev_pmd_unref(pmd);
2060 pmd = dp_netdev_get_pmd(dp, get->pmd_id);
2064 hmapx_add(&to_find, pmd);
2067 if (!hmapx_count(&to_find)) {
2071 HMAPX_FOR_EACH (node, &to_find) {
2072 pmd = (struct dp_netdev_pmd_thread *) node->data;
2073 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
2076 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
2085 HMAPX_FOR_EACH (node, &to_find) {
2086 pmd = (struct dp_netdev_pmd_thread *) node->data;
2087 dp_netdev_pmd_unref(pmd);
2090 hmapx_destroy(&to_find);
2094 static struct dp_netdev_flow *
2095 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
2096 struct match *match, const ovs_u128 *ufid,
2097 const struct nlattr *actions, size_t actions_len)
2098 OVS_REQUIRES(pmd->flow_mutex)
2100 struct dp_netdev_flow *flow;
2101 struct netdev_flow_key mask;
2103 netdev_flow_mask_init(&mask, match);
2104 /* Make sure wc does not have metadata. */
2105 ovs_assert(!FLOWMAP_HAS_FIELD(&mask.mf.map, metadata)
2106 && !FLOWMAP_HAS_FIELD(&mask.mf.map, regs));
2108 /* Do not allocate extra space. */
2109 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
2110 memset(&flow->stats, 0, sizeof flow->stats);
2113 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
2114 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
2115 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
2116 ovs_refcount_init(&flow->ref_cnt);
2117 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
2119 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
2120 dpcls_insert(&pmd->cls, &flow->cr, &mask);
2122 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
2123 dp_netdev_flow_hash(&flow->ufid));
2125 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2126 struct ds ds = DS_EMPTY_INITIALIZER;
2127 struct ofpbuf key_buf, mask_buf;
2128 struct odp_flow_key_parms odp_parms = {
2129 .flow = &match->flow,
2130 .mask = &match->wc.masks,
2131 .support = dp_netdev_support,
2134 ofpbuf_init(&key_buf, 0);
2135 ofpbuf_init(&mask_buf, 0);
2137 odp_flow_key_from_flow(&odp_parms, &key_buf);
2138 odp_parms.key_buf = &key_buf;
2139 odp_flow_key_from_mask(&odp_parms, &mask_buf);
2141 ds_put_cstr(&ds, "flow_add: ");
2142 odp_format_ufid(ufid, &ds);
2143 ds_put_cstr(&ds, " ");
2144 odp_flow_format(key_buf.data, key_buf.size,
2145 mask_buf.data, mask_buf.size,
2147 ds_put_cstr(&ds, ", actions:");
2148 format_odp_actions(&ds, actions, actions_len);
2150 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2152 ofpbuf_uninit(&key_buf);
2153 ofpbuf_uninit(&mask_buf);
2161 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2163 struct dp_netdev *dp = get_dp_netdev(dpif);
2164 struct dp_netdev_flow *netdev_flow;
2165 struct netdev_flow_key key;
2166 struct dp_netdev_pmd_thread *pmd;
2169 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2170 ? NON_PMD_CORE_ID : put->pmd_id;
2173 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2177 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2178 put->mask, put->mask_len,
2179 &match.flow, &match.wc);
2184 pmd = dp_netdev_get_pmd(dp, pmd_id);
2189 /* Must produce a netdev_flow_key for lookup.
2190 * This interface is no longer performance critical, since it is not used
2191 * for upcall processing any more. */
2192 netdev_flow_key_from_flow(&key, &match.flow);
2197 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2200 ovs_mutex_lock(&pmd->flow_mutex);
2201 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2203 if (put->flags & DPIF_FP_CREATE) {
2204 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2206 memset(put->stats, 0, sizeof *put->stats);
2208 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2218 if (put->flags & DPIF_FP_MODIFY
2219 && flow_equal(&match.flow, &netdev_flow->flow)) {
2220 struct dp_netdev_actions *new_actions;
2221 struct dp_netdev_actions *old_actions;
2223 new_actions = dp_netdev_actions_create(put->actions,
2226 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2227 ovsrcu_set(&netdev_flow->actions, new_actions);
2230 get_dpif_flow_stats(netdev_flow, put->stats);
2232 if (put->flags & DPIF_FP_ZERO_STATS) {
2233 /* XXX: The userspace datapath uses thread local statistics
2234 * (for flows), which should be updated only by the owning
2235 * thread. Since we cannot write on stats memory here,
2236 * we choose not to support this flag. Please note:
2237 * - This feature is currently used only by dpctl commands with
2239 * - Should the need arise, this operation can be implemented
2240 * by keeping a base value (to be update here) for each
2241 * counter, and subtracting it before outputting the stats */
2245 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2246 } else if (put->flags & DPIF_FP_CREATE) {
2249 /* Overlapping flow. */
2253 ovs_mutex_unlock(&pmd->flow_mutex);
2254 dp_netdev_pmd_unref(pmd);
2260 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2262 struct dp_netdev *dp = get_dp_netdev(dpif);
2263 struct dp_netdev_flow *netdev_flow;
2264 struct dp_netdev_pmd_thread *pmd;
2265 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2266 ? NON_PMD_CORE_ID : del->pmd_id;
2269 pmd = dp_netdev_get_pmd(dp, pmd_id);
2274 ovs_mutex_lock(&pmd->flow_mutex);
2275 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2279 get_dpif_flow_stats(netdev_flow, del->stats);
2281 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2285 ovs_mutex_unlock(&pmd->flow_mutex);
2286 dp_netdev_pmd_unref(pmd);
2291 struct dpif_netdev_flow_dump {
2292 struct dpif_flow_dump up;
2293 struct cmap_position poll_thread_pos;
2294 struct cmap_position flow_pos;
2295 struct dp_netdev_pmd_thread *cur_pmd;
2297 struct ovs_mutex mutex;
2300 static struct dpif_netdev_flow_dump *
2301 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2303 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2306 static struct dpif_flow_dump *
2307 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2309 struct dpif_netdev_flow_dump *dump;
2311 dump = xzalloc(sizeof *dump);
2312 dpif_flow_dump_init(&dump->up, dpif_);
2313 dump->up.terse = terse;
2314 ovs_mutex_init(&dump->mutex);
2320 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2322 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2324 ovs_mutex_destroy(&dump->mutex);
2329 struct dpif_netdev_flow_dump_thread {
2330 struct dpif_flow_dump_thread up;
2331 struct dpif_netdev_flow_dump *dump;
2332 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2333 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2336 static struct dpif_netdev_flow_dump_thread *
2337 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2339 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2342 static struct dpif_flow_dump_thread *
2343 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2345 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2346 struct dpif_netdev_flow_dump_thread *thread;
2348 thread = xmalloc(sizeof *thread);
2349 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2350 thread->dump = dump;
2355 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2357 struct dpif_netdev_flow_dump_thread *thread
2358 = dpif_netdev_flow_dump_thread_cast(thread_);
2364 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2365 struct dpif_flow *flows, int max_flows)
2367 struct dpif_netdev_flow_dump_thread *thread
2368 = dpif_netdev_flow_dump_thread_cast(thread_);
2369 struct dpif_netdev_flow_dump *dump = thread->dump;
2370 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2374 ovs_mutex_lock(&dump->mutex);
2375 if (!dump->status) {
2376 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2377 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2378 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2379 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2381 /* First call to dump_next(), extracts the first pmd thread.
2382 * If there is no pmd thread, returns immediately. */
2384 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2386 ovs_mutex_unlock(&dump->mutex);
2393 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2394 struct cmap_node *node;
2396 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2400 netdev_flows[n_flows] = CONTAINER_OF(node,
2401 struct dp_netdev_flow,
2404 /* When finishing dumping the current pmd thread, moves to
2406 if (n_flows < flow_limit) {
2407 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2408 dp_netdev_pmd_unref(pmd);
2409 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2415 /* Keeps the reference to next caller. */
2416 dump->cur_pmd = pmd;
2418 /* If the current dump is empty, do not exit the loop, since the
2419 * remaining pmds could have flows to be dumped. Just dumps again
2420 * on the new 'pmd'. */
2423 ovs_mutex_unlock(&dump->mutex);
2425 for (i = 0; i < n_flows; i++) {
2426 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2427 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2428 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2429 struct dpif_flow *f = &flows[i];
2430 struct ofpbuf key, mask;
2432 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2433 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2434 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2442 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2443 OVS_NO_THREAD_SAFETY_ANALYSIS
2445 struct dp_netdev *dp = get_dp_netdev(dpif);
2446 struct dp_netdev_pmd_thread *pmd;
2447 struct dp_packet_batch pp;
2449 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2450 dp_packet_size(execute->packet) > UINT16_MAX) {
2454 /* Tries finding the 'pmd'. If NULL is returned, that means
2455 * the current thread is a non-pmd thread and should use
2456 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2457 pmd = ovsthread_getspecific(dp->per_pmd_key);
2459 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2462 /* If the current thread is non-pmd thread, acquires
2463 * the 'non_pmd_mutex'. */
2464 if (pmd->core_id == NON_PMD_CORE_ID) {
2465 ovs_mutex_lock(&dp->non_pmd_mutex);
2468 /* The action processing expects the RSS hash to be valid, because
2469 * it's always initialized at the beginning of datapath processing.
2470 * In this case, though, 'execute->packet' may not have gone through
2471 * the datapath at all, it may have been generated by the upper layer
2472 * (OpenFlow packet-out, BFD frame, ...). */
2473 if (!dp_packet_rss_valid(execute->packet)) {
2474 dp_packet_set_rss_hash(execute->packet,
2475 flow_hash_5tuple(execute->flow, 0));
2478 packet_batch_init_packet(&pp, execute->packet);
2479 dp_netdev_execute_actions(pmd, &pp, false, execute->actions,
2480 execute->actions_len);
2482 if (pmd->core_id == NON_PMD_CORE_ID) {
2483 ovs_mutex_unlock(&dp->non_pmd_mutex);
2484 dp_netdev_pmd_unref(pmd);
2491 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2495 for (i = 0; i < n_ops; i++) {
2496 struct dpif_op *op = ops[i];
2499 case DPIF_OP_FLOW_PUT:
2500 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2503 case DPIF_OP_FLOW_DEL:
2504 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2507 case DPIF_OP_EXECUTE:
2508 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2511 case DPIF_OP_FLOW_GET:
2512 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2518 /* Changes the number or the affinity of pmd threads. The changes are actually
2519 * applied in dpif_netdev_run(). */
2521 dpif_netdev_pmd_set(struct dpif *dpif, const char *cmask)
2523 struct dp_netdev *dp = get_dp_netdev(dpif);
2525 if (!nullable_string_is_equal(dp->requested_pmd_cmask, cmask)) {
2526 free(dp->requested_pmd_cmask);
2527 dp->requested_pmd_cmask = nullable_xstrdup(cmask);
2534 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2535 uint32_t queue_id, uint32_t *priority)
2537 *priority = queue_id;
2542 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2543 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2544 struct dp_netdev_actions *
2545 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2547 struct dp_netdev_actions *netdev_actions;
2549 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2550 memcpy(netdev_actions->actions, actions, size);
2551 netdev_actions->size = size;
2553 return netdev_actions;
2556 struct dp_netdev_actions *
2557 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2559 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2563 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2568 static inline unsigned long long
2569 cycles_counter(void)
2572 return rte_get_tsc_cycles();
2578 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2579 extern struct ovs_mutex cycles_counter_fake_mutex;
2581 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2583 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2584 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2585 OVS_NO_THREAD_SAFETY_ANALYSIS
2587 pmd->last_cycles = cycles_counter();
2590 /* Stop counting cycles and add them to the counter 'type' */
2592 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2593 enum pmd_cycles_counter_type type)
2594 OVS_RELEASES(&cycles_counter_fake_mutex)
2595 OVS_NO_THREAD_SAFETY_ANALYSIS
2597 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2599 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2603 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2604 struct dp_netdev_port *port,
2605 struct netdev_rxq *rxq)
2607 struct dp_packet_batch batch;
2610 dp_packet_batch_init(&batch);
2611 cycles_count_start(pmd);
2612 error = netdev_rxq_recv(rxq, &batch);
2613 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2615 *recirc_depth_get() = 0;
2617 cycles_count_start(pmd);
2618 dp_netdev_input(pmd, &batch, port->port_no);
2619 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2620 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2621 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2623 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2624 netdev_get_name(port->netdev), ovs_strerror(error));
2629 port_reconfigure(struct dp_netdev_port *port)
2631 struct netdev *netdev = port->netdev;
2634 if (!netdev_is_reconf_required(netdev)) {
2638 /* Closes the existing 'rxq's. */
2639 for (i = 0; i < port->n_rxq; i++) {
2640 netdev_rxq_close(port->rxq[i]);
2641 port->rxq[i] = NULL;
2645 /* Allows 'netdev' to apply the pending configuration changes. */
2646 err = netdev_reconfigure(netdev);
2647 if (err && (err != EOPNOTSUPP)) {
2648 VLOG_ERR("Failed to set interface %s new configuration",
2649 netdev_get_name(netdev));
2652 /* If the netdev_reconfigure() above succeeds, reopens the 'rxq's. */
2653 port->rxq = xrealloc(port->rxq, sizeof *port->rxq * netdev_n_rxq(netdev));
2654 for (i = 0; i < netdev_n_rxq(netdev); i++) {
2655 err = netdev_rxq_open(netdev, &port->rxq[i], i);
2666 reconfigure_pmd_threads(struct dp_netdev *dp)
2667 OVS_REQUIRES(dp->port_mutex)
2669 struct dp_netdev_port *port, *next;
2671 dp_netdev_destroy_all_pmds(dp);
2673 HMAP_FOR_EACH_SAFE (port, next, node, &dp->ports) {
2676 err = port_reconfigure(port);
2678 hmap_remove(&dp->ports, &port->node);
2679 seq_change(dp->port_seq);
2683 /* Reconfigures the cpu mask. */
2684 ovs_numa_set_cpu_mask(dp->requested_pmd_cmask);
2685 free(dp->pmd_cmask);
2686 dp->pmd_cmask = nullable_xstrdup(dp->requested_pmd_cmask);
2688 /* Restores the non-pmd. */
2689 dp_netdev_set_nonpmd(dp);
2690 /* Restores all pmd threads. */
2691 dp_netdev_reset_pmd_threads(dp);
2694 /* Returns true if one of the netdevs in 'dp' requires a reconfiguration */
2696 ports_require_restart(const struct dp_netdev *dp)
2697 OVS_REQUIRES(dp->port_mutex)
2699 struct dp_netdev_port *port;
2701 HMAP_FOR_EACH (port, node, &dp->ports) {
2702 if (netdev_is_reconf_required(port->netdev)) {
2710 /* Return true if needs to revalidate datapath flows. */
2712 dpif_netdev_run(struct dpif *dpif)
2714 struct dp_netdev_port *port;
2715 struct dp_netdev *dp = get_dp_netdev(dpif);
2716 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2718 uint64_t new_tnl_seq;
2720 ovs_mutex_lock(&dp->port_mutex);
2721 ovs_mutex_lock(&dp->non_pmd_mutex);
2722 HMAP_FOR_EACH (port, node, &dp->ports) {
2723 if (!netdev_is_pmd(port->netdev)) {
2726 for (i = 0; i < port->n_rxq; i++) {
2727 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2731 ovs_mutex_unlock(&dp->non_pmd_mutex);
2733 dp_netdev_pmd_unref(non_pmd);
2735 if (!nullable_string_is_equal(dp->pmd_cmask, dp->requested_pmd_cmask)
2736 || ports_require_restart(dp)) {
2737 reconfigure_pmd_threads(dp);
2739 ovs_mutex_unlock(&dp->port_mutex);
2741 tnl_neigh_cache_run();
2743 new_tnl_seq = seq_read(tnl_conf_seq);
2745 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2746 dp->last_tnl_conf_seq = new_tnl_seq;
2753 dpif_netdev_wait(struct dpif *dpif)
2755 struct dp_netdev_port *port;
2756 struct dp_netdev *dp = get_dp_netdev(dpif);
2758 ovs_mutex_lock(&dp_netdev_mutex);
2759 ovs_mutex_lock(&dp->port_mutex);
2760 HMAP_FOR_EACH (port, node, &dp->ports) {
2761 netdev_wait_reconf_required(port->netdev);
2762 if (!netdev_is_pmd(port->netdev)) {
2765 for (i = 0; i < port->n_rxq; i++) {
2766 netdev_rxq_wait(port->rxq[i]);
2770 ovs_mutex_unlock(&dp->port_mutex);
2771 ovs_mutex_unlock(&dp_netdev_mutex);
2772 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2776 pmd_free_cached_ports(struct dp_netdev_pmd_thread *pmd)
2778 struct tx_port *tx_port_cached;
2780 HMAP_FOR_EACH_POP (tx_port_cached, node, &pmd->port_cache) {
2781 free(tx_port_cached);
2785 /* Copies ports from 'pmd->tx_ports' (shared with the main thread) to
2786 * 'pmd->port_cache' (thread local) */
2788 pmd_load_cached_ports(struct dp_netdev_pmd_thread *pmd)
2789 OVS_REQUIRES(pmd->port_mutex)
2791 struct tx_port *tx_port, *tx_port_cached;
2793 pmd_free_cached_ports(pmd);
2794 hmap_shrink(&pmd->port_cache);
2796 HMAP_FOR_EACH (tx_port, node, &pmd->tx_ports) {
2797 tx_port_cached = xmemdup(tx_port, sizeof *tx_port_cached);
2798 hmap_insert(&pmd->port_cache, &tx_port_cached->node,
2799 hash_port_no(tx_port_cached->port_no));
2804 pmd_load_queues_and_ports(struct dp_netdev_pmd_thread *pmd,
2805 struct rxq_poll **ppoll_list)
2807 struct rxq_poll *poll_list = *ppoll_list;
2808 struct rxq_poll *poll;
2811 ovs_mutex_lock(&pmd->port_mutex);
2812 poll_list = xrealloc(poll_list, pmd->poll_cnt * sizeof *poll_list);
2815 LIST_FOR_EACH (poll, node, &pmd->poll_list) {
2816 poll_list[i++] = *poll;
2819 pmd_load_cached_ports(pmd);
2821 ovs_mutex_unlock(&pmd->port_mutex);
2823 *ppoll_list = poll_list;
2828 pmd_thread_main(void *f_)
2830 struct dp_netdev_pmd_thread *pmd = f_;
2831 unsigned int lc = 0;
2832 struct rxq_poll *poll_list;
2833 unsigned int port_seq = PMD_INITIAL_SEQ;
2840 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2841 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2842 ovs_numa_thread_setaffinity_core(pmd->core_id);
2843 dpdk_set_lcore_id(pmd->core_id);
2844 poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list);
2846 emc_cache_init(&pmd->flow_cache);
2848 /* List port/core affinity */
2849 for (i = 0; i < poll_cnt; i++) {
2850 VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n",
2851 pmd->core_id, netdev_get_name(poll_list[i].port->netdev),
2852 netdev_rxq_get_queue_id(poll_list[i].rx));
2856 for (i = 0; i < poll_cnt; i++) {
2857 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2865 coverage_try_clear();
2866 if (!ovsrcu_try_quiesce()) {
2867 emc_cache_slow_sweep(&pmd->flow_cache);
2870 atomic_read_relaxed(&pmd->change_seq, &seq);
2871 if (seq != port_seq) {
2878 poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list);
2879 exiting = latch_is_set(&pmd->exit_latch);
2880 /* Signal here to make sure the pmd finishes
2881 * reloading the updated configuration. */
2882 dp_netdev_pmd_reload_done(pmd);
2884 emc_cache_uninit(&pmd->flow_cache);
2891 pmd_free_cached_ports(pmd);
2896 dp_netdev_disable_upcall(struct dp_netdev *dp)
2897 OVS_ACQUIRES(dp->upcall_rwlock)
2899 fat_rwlock_wrlock(&dp->upcall_rwlock);
2903 dpif_netdev_disable_upcall(struct dpif *dpif)
2904 OVS_NO_THREAD_SAFETY_ANALYSIS
2906 struct dp_netdev *dp = get_dp_netdev(dpif);
2907 dp_netdev_disable_upcall(dp);
2911 dp_netdev_enable_upcall(struct dp_netdev *dp)
2912 OVS_RELEASES(dp->upcall_rwlock)
2914 fat_rwlock_unlock(&dp->upcall_rwlock);
2918 dpif_netdev_enable_upcall(struct dpif *dpif)
2919 OVS_NO_THREAD_SAFETY_ANALYSIS
2921 struct dp_netdev *dp = get_dp_netdev(dpif);
2922 dp_netdev_enable_upcall(dp);
2926 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2928 ovs_mutex_lock(&pmd->cond_mutex);
2929 xpthread_cond_signal(&pmd->cond);
2930 ovs_mutex_unlock(&pmd->cond_mutex);
2933 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2934 * the pointer if succeeds, otherwise, NULL.
2936 * Caller must unrefs the returned reference. */
2937 static struct dp_netdev_pmd_thread *
2938 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2940 struct dp_netdev_pmd_thread *pmd;
2941 const struct cmap_node *pnode;
2943 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2947 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2949 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2952 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2954 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2955 OVS_REQUIRES(dp->port_mutex)
2957 struct dp_netdev_pmd_thread *non_pmd;
2958 struct dp_netdev_port *port;
2960 non_pmd = xzalloc(sizeof *non_pmd);
2961 dp_netdev_configure_pmd(non_pmd, dp, NON_PMD_CORE_ID, OVS_NUMA_UNSPEC);
2963 HMAP_FOR_EACH (port, node, &dp->ports) {
2964 dp_netdev_add_port_tx_to_pmd(non_pmd, port);
2967 dp_netdev_reload_pmd__(non_pmd);
2970 /* Caller must have valid pointer to 'pmd'. */
2972 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2974 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2978 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2980 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2981 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2985 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2986 * fails, keeps checking for next node until reaching the end of cmap.
2988 * Caller must unrefs the returned reference. */
2989 static struct dp_netdev_pmd_thread *
2990 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2992 struct dp_netdev_pmd_thread *next;
2995 struct cmap_node *node;
2997 node = cmap_next_position(&dp->poll_threads, pos);
2998 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
3000 } while (next && !dp_netdev_pmd_try_ref(next));
3005 /* Configures the 'pmd' based on the input argument. */
3007 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
3008 unsigned core_id, int numa_id)
3011 pmd->core_id = core_id;
3012 pmd->numa_id = numa_id;
3015 atomic_init(&pmd->tx_qid,
3016 (core_id == NON_PMD_CORE_ID)
3017 ? ovs_numa_get_n_cores()
3018 : get_n_pmd_threads(dp));
3020 ovs_refcount_init(&pmd->ref_cnt);
3021 latch_init(&pmd->exit_latch);
3022 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
3023 xpthread_cond_init(&pmd->cond, NULL);
3024 ovs_mutex_init(&pmd->cond_mutex);
3025 ovs_mutex_init(&pmd->flow_mutex);
3026 ovs_mutex_init(&pmd->port_mutex);
3027 dpcls_init(&pmd->cls);
3028 cmap_init(&pmd->flow_table);
3029 ovs_list_init(&pmd->poll_list);
3030 hmap_init(&pmd->tx_ports);
3031 hmap_init(&pmd->port_cache);
3032 /* init the 'flow_cache' since there is no
3033 * actual thread created for NON_PMD_CORE_ID. */
3034 if (core_id == NON_PMD_CORE_ID) {
3035 emc_cache_init(&pmd->flow_cache);
3037 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
3038 hash_int(core_id, 0));
3042 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
3044 dp_netdev_pmd_flow_flush(pmd);
3045 dpcls_destroy(&pmd->cls);
3046 hmap_destroy(&pmd->port_cache);
3047 hmap_destroy(&pmd->tx_ports);
3048 cmap_destroy(&pmd->flow_table);
3049 ovs_mutex_destroy(&pmd->flow_mutex);
3050 latch_destroy(&pmd->exit_latch);
3051 xpthread_cond_destroy(&pmd->cond);
3052 ovs_mutex_destroy(&pmd->cond_mutex);
3053 ovs_mutex_destroy(&pmd->port_mutex);
3057 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
3058 * and unrefs the struct. */
3060 dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd)
3062 /* NON_PMD_CORE_ID doesn't have a thread, so we don't have to synchronize,
3063 * but extra cleanup is necessary */
3064 if (pmd->core_id == NON_PMD_CORE_ID) {
3065 emc_cache_uninit(&pmd->flow_cache);
3066 pmd_free_cached_ports(pmd);
3068 latch_set(&pmd->exit_latch);
3069 dp_netdev_reload_pmd__(pmd);
3070 ovs_numa_unpin_core(pmd->core_id);
3071 xpthread_join(pmd->thread, NULL);
3074 dp_netdev_pmd_clear_ports(pmd);
3076 /* Purges the 'pmd''s flows after stopping the thread, but before
3077 * destroying the flows, so that the flow stats can be collected. */
3078 if (dp->dp_purge_cb) {
3079 dp->dp_purge_cb(dp->dp_purge_aux, pmd->core_id);
3081 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
3082 dp_netdev_pmd_unref(pmd);
3085 /* Destroys all pmd threads. */
3087 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
3089 struct dp_netdev_pmd_thread *pmd;
3090 struct dp_netdev_pmd_thread **pmd_list;
3091 size_t k = 0, n_pmds;
3093 n_pmds = cmap_count(&dp->poll_threads);
3094 pmd_list = xcalloc(n_pmds, sizeof *pmd_list);
3096 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3097 /* We cannot call dp_netdev_del_pmd(), since it alters
3098 * 'dp->poll_threads' (while we're iterating it) and it
3100 ovs_assert(k < n_pmds);
3101 pmd_list[k++] = pmd;
3104 for (size_t i = 0; i < k; i++) {
3105 dp_netdev_del_pmd(dp, pmd_list[i]);
3110 /* Deletes all pmd threads on numa node 'numa_id' and
3111 * fixes tx_qids of other threads to keep them sequential. */
3113 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
3115 struct dp_netdev_pmd_thread *pmd;
3116 int n_pmds_on_numa, n_pmds;
3117 int *free_idx, k = 0;
3118 struct dp_netdev_pmd_thread **pmd_list;
3120 n_pmds_on_numa = get_n_pmd_threads_on_numa(dp, numa_id);
3121 free_idx = xcalloc(n_pmds_on_numa, sizeof *free_idx);
3122 pmd_list = xcalloc(n_pmds_on_numa, sizeof *pmd_list);
3124 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3125 /* We cannot call dp_netdev_del_pmd(), since it alters
3126 * 'dp->poll_threads' (while we're iterating it) and it
3128 if (pmd->numa_id == numa_id) {
3129 atomic_read_relaxed(&pmd->tx_qid, &free_idx[k]);
3131 ovs_assert(k < n_pmds_on_numa);
3136 for (int i = 0; i < k; i++) {
3137 dp_netdev_del_pmd(dp, pmd_list[i]);
3140 n_pmds = get_n_pmd_threads(dp);
3141 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3144 atomic_read_relaxed(&pmd->tx_qid, &old_tx_qid);
3146 if (old_tx_qid >= n_pmds) {
3147 int new_tx_qid = free_idx[--k];
3149 atomic_store_relaxed(&pmd->tx_qid, new_tx_qid);
3157 /* Deletes all rx queues from pmd->poll_list and all the ports from
3160 dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread *pmd)
3162 struct rxq_poll *poll;
3163 struct tx_port *port;
3165 ovs_mutex_lock(&pmd->port_mutex);
3166 LIST_FOR_EACH_POP (poll, node, &pmd->poll_list) {
3170 HMAP_FOR_EACH_POP (port, node, &pmd->tx_ports) {
3173 ovs_mutex_unlock(&pmd->port_mutex);
3176 static struct tx_port *
3177 tx_port_lookup(const struct hmap *hmap, odp_port_t port_no)
3181 HMAP_FOR_EACH_IN_BUCKET (tx, node, hash_port_no(port_no), hmap) {
3182 if (tx->port_no == port_no) {
3190 /* Deletes all rx queues of 'port' from 'poll_list', and the 'port' from
3191 * 'tx_ports' of 'pmd' thread. Returns true if 'port' was found in 'pmd'
3192 * (therefore a restart is required). */
3194 dp_netdev_del_port_from_pmd__(struct dp_netdev_port *port,
3195 struct dp_netdev_pmd_thread *pmd)
3197 struct rxq_poll *poll, *next;
3201 ovs_mutex_lock(&pmd->port_mutex);
3202 LIST_FOR_EACH_SAFE (poll, next, node, &pmd->poll_list) {
3203 if (poll->port == port) {
3205 ovs_list_remove(&poll->node);
3211 tx = tx_port_lookup(&pmd->tx_ports, port->port_no);
3213 hmap_remove(&pmd->tx_ports, &tx->node);
3217 ovs_mutex_unlock(&pmd->port_mutex);
3222 /* Deletes 'port' from the 'poll_list' and from the 'tx_ports' of all the pmd
3223 * threads. The pmd threads that need to be restarted are inserted in
3226 dp_netdev_del_port_from_all_pmds__(struct dp_netdev *dp,
3227 struct dp_netdev_port *port,
3228 struct hmapx *to_reload)
3230 struct dp_netdev_pmd_thread *pmd;
3232 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3235 found = dp_netdev_del_port_from_pmd__(port, pmd);
3238 hmapx_add(to_reload, pmd);
3243 /* Deletes 'port' from the 'poll_list' and from the 'tx_ports' of all the pmd
3244 * threads. Reloads the threads if needed. */
3246 dp_netdev_del_port_from_all_pmds(struct dp_netdev *dp,
3247 struct dp_netdev_port *port)
3249 struct dp_netdev_pmd_thread *pmd;
3250 struct hmapx to_reload = HMAPX_INITIALIZER(&to_reload);
3251 struct hmapx_node *node;
3253 dp_netdev_del_port_from_all_pmds__(dp, port, &to_reload);
3255 HMAPX_FOR_EACH (node, &to_reload) {
3256 pmd = (struct dp_netdev_pmd_thread *) node->data;
3257 dp_netdev_reload_pmd__(pmd);
3260 hmapx_destroy(&to_reload);
3264 /* Returns PMD thread from this numa node with fewer rx queues to poll.
3265 * Returns NULL if there is no PMD threads on this numa node.
3266 * Can be called safely only by main thread. */
3267 static struct dp_netdev_pmd_thread *
3268 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id)
3271 struct dp_netdev_pmd_thread *pmd, *res = NULL;
3273 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3274 if (pmd->numa_id == numa_id
3275 && (min_cnt > pmd->poll_cnt || res == NULL)) {
3276 min_cnt = pmd->poll_cnt;
3284 /* Adds rx queue to poll_list of PMD thread. */
3286 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
3287 struct dp_netdev_port *port, struct netdev_rxq *rx)
3288 OVS_REQUIRES(pmd->port_mutex)
3290 struct rxq_poll *poll = xmalloc(sizeof *poll);
3295 ovs_list_push_back(&pmd->poll_list, &poll->node);
3299 /* Add 'port' to the tx port cache of 'pmd', which must be reloaded for the
3300 * changes to take effect. */
3302 dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread *pmd,
3303 struct dp_netdev_port *port)
3305 struct tx_port *tx = xzalloc(sizeof *tx);
3307 tx->netdev = port->netdev;
3308 tx->port_no = port->port_no;
3310 ovs_mutex_lock(&pmd->port_mutex);
3311 hmap_insert(&pmd->tx_ports, &tx->node, hash_port_no(tx->port_no));
3312 ovs_mutex_unlock(&pmd->port_mutex);
3315 /* Distribute all rx queues of 'port' between PMD threads in 'dp'. The pmd
3316 * threads that need to be restarted are inserted in 'to_reload'. */
3318 dp_netdev_add_port_rx_to_pmds(struct dp_netdev *dp,
3319 struct dp_netdev_port *port,
3320 struct hmapx *to_reload)
3322 int numa_id = netdev_get_numa_id(port->netdev);
3325 if (!netdev_is_pmd(port->netdev)) {
3329 for (i = 0; i < port->n_rxq; i++) {
3330 struct dp_netdev_pmd_thread *pmd;
3332 pmd = dp_netdev_less_loaded_pmd_on_numa(dp, numa_id);
3334 VLOG_WARN("There's no pmd thread on numa node %d", numa_id);
3338 ovs_mutex_lock(&pmd->port_mutex);
3339 dp_netdev_add_rxq_to_pmd(pmd, port, port->rxq[i]);
3340 ovs_mutex_unlock(&pmd->port_mutex);
3342 hmapx_add(to_reload, pmd);
3346 /* Distributes all rx queues of 'port' between all PMD threads in 'dp' and
3347 * inserts 'port' in the PMD threads 'tx_ports'. The pmd threads that need to
3348 * be restarted are inserted in 'to_reload'. */
3350 dp_netdev_add_port_to_pmds__(struct dp_netdev *dp, struct dp_netdev_port *port,
3351 struct hmapx *to_reload)
3353 struct dp_netdev_pmd_thread *pmd;
3355 dp_netdev_add_port_rx_to_pmds(dp, port, to_reload);
3357 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3358 dp_netdev_add_port_tx_to_pmd(pmd, port);
3359 hmapx_add(to_reload, pmd);
3363 /* Distributes all rx queues of 'port' between all PMD threads in 'dp', inserts
3364 * 'port' in the PMD threads 'tx_ports' and reloads them, if needed. */
3366 dp_netdev_add_port_to_pmds(struct dp_netdev *dp, struct dp_netdev_port *port)
3368 struct dp_netdev_pmd_thread *pmd;
3369 struct hmapx to_reload = HMAPX_INITIALIZER(&to_reload);
3370 struct hmapx_node *node;
3372 dp_netdev_add_port_to_pmds__(dp, port, &to_reload);
3374 HMAPX_FOR_EACH (node, &to_reload) {
3375 pmd = (struct dp_netdev_pmd_thread *) node->data;
3376 dp_netdev_reload_pmd__(pmd);
3379 hmapx_destroy(&to_reload);
3382 /* Starts pmd threads for the numa node 'numa_id', if not already started.
3383 * The function takes care of filling the threads tx port cache. */
3385 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
3386 OVS_REQUIRES(dp->port_mutex)
3390 if (!ovs_numa_numa_id_is_valid(numa_id)) {
3391 VLOG_WARN("Cannot create pmd threads due to numa id (%d) invalid",
3396 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
3398 /* If there are already pmd threads created for the numa node
3399 * in which 'netdev' is on, do nothing. Else, creates the
3400 * pmd threads for the numa node. */
3402 int can_have, n_unpinned, i;
3404 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
3406 VLOG_WARN("Cannot create pmd threads due to out of unpinned "
3407 "cores on numa node %d", numa_id);
3411 /* If cpu mask is specified, uses all unpinned cores, otherwise
3412 * tries creating NR_PMD_THREADS pmd threads. */
3413 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
3414 for (i = 0; i < can_have; i++) {
3415 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
3416 struct dp_netdev_pmd_thread *pmd = xzalloc(sizeof *pmd);
3417 struct dp_netdev_port *port;
3419 dp_netdev_configure_pmd(pmd, dp, core_id, numa_id);
3421 HMAP_FOR_EACH (port, node, &dp->ports) {
3422 dp_netdev_add_port_tx_to_pmd(pmd, port);
3425 pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd);
3427 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
3432 /* Called after pmd threads config change. Restarts pmd threads with
3433 * new configuration. */
3435 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
3436 OVS_REQUIRES(dp->port_mutex)
3438 struct hmapx to_reload = HMAPX_INITIALIZER(&to_reload);
3439 struct dp_netdev_pmd_thread *pmd;
3440 struct dp_netdev_port *port;
3441 struct hmapx_node *node;
3443 HMAP_FOR_EACH (port, node, &dp->ports) {
3444 if (netdev_is_pmd(port->netdev)) {
3445 int numa_id = netdev_get_numa_id(port->netdev);
3447 dp_netdev_set_pmds_on_numa(dp, numa_id);
3449 dp_netdev_add_port_rx_to_pmds(dp, port, &to_reload);
3452 HMAPX_FOR_EACH (node, &to_reload) {
3453 pmd = (struct dp_netdev_pmd_thread *) node->data;
3454 dp_netdev_reload_pmd__(pmd);
3457 hmapx_destroy(&to_reload);
3461 dpif_netdev_get_datapath_version(void)
3463 return xstrdup("<built-in>");
3467 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
3468 uint16_t tcp_flags, long long now)
3472 atomic_store_relaxed(&netdev_flow->stats.used, now);
3473 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
3474 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
3475 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
3477 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
3481 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
3482 enum dp_stat_type type, int cnt)
3484 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3488 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3489 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3490 enum dpif_upcall_type type, const struct nlattr *userdata,
3491 struct ofpbuf *actions, struct ofpbuf *put_actions)
3493 struct dp_netdev *dp = pmd->dp;
3494 struct flow_tnl orig_tunnel;
3497 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3501 /* Upcall processing expects the Geneve options to be in the translated
3502 * format but we need to retain the raw format for datapath use. */
3503 orig_tunnel.flags = flow->tunnel.flags;
3504 if (flow->tunnel.flags & FLOW_TNL_F_UDPIF) {
3505 orig_tunnel.metadata.present.len = flow->tunnel.metadata.present.len;
3506 memcpy(orig_tunnel.metadata.opts.gnv, flow->tunnel.metadata.opts.gnv,
3507 flow->tunnel.metadata.present.len);
3508 err = tun_metadata_from_geneve_udpif(&orig_tunnel, &orig_tunnel,
3515 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3516 struct ds ds = DS_EMPTY_INITIALIZER;
3519 struct odp_flow_key_parms odp_parms = {
3522 .support = dp_netdev_support,
3525 ofpbuf_init(&key, 0);
3526 odp_flow_key_from_flow(&odp_parms, &key);
3527 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3528 dp_packet_size(packet_));
3530 odp_flow_key_format(key.data, key.size, &ds);
3532 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3533 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3535 ofpbuf_uninit(&key);
3541 err = dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3542 actions, wc, put_actions, dp->upcall_aux);
3543 if (err && err != ENOSPC) {
3547 /* Translate tunnel metadata masks to datapath format. */
3549 if (wc->masks.tunnel.metadata.present.map) {
3550 struct geneve_opt opts[TLV_TOT_OPT_SIZE /
3551 sizeof(struct geneve_opt)];
3553 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3554 tun_metadata_to_geneve_udpif_mask(&flow->tunnel,
3556 orig_tunnel.metadata.opts.gnv,
3557 orig_tunnel.metadata.present.len,
3560 orig_tunnel.metadata.present.len = 0;
3563 memset(&wc->masks.tunnel.metadata, 0,
3564 sizeof wc->masks.tunnel.metadata);
3565 memcpy(&wc->masks.tunnel.metadata.opts.gnv, opts,
3566 orig_tunnel.metadata.present.len);
3568 wc->masks.tunnel.metadata.present.len = 0xff;
3571 /* Restore tunnel metadata. We need to use the saved options to ensure
3572 * that any unknown options are not lost. The generated mask will have
3573 * the same structure, matching on types and lengths but wildcarding
3574 * option data we don't care about. */
3575 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3576 memcpy(&flow->tunnel.metadata.opts.gnv, orig_tunnel.metadata.opts.gnv,
3577 orig_tunnel.metadata.present.len);
3578 flow->tunnel.metadata.present.len = orig_tunnel.metadata.present.len;
3579 flow->tunnel.flags |= FLOW_TNL_F_UDPIF;
3585 static inline uint32_t
3586 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3587 const struct miniflow *mf)
3589 uint32_t hash, recirc_depth;
3591 if (OVS_LIKELY(dp_packet_rss_valid(packet))) {
3592 hash = dp_packet_get_rss_hash(packet);
3594 hash = miniflow_hash_5tuple(mf, 0);
3595 dp_packet_set_rss_hash(packet, hash);
3598 /* The RSS hash must account for the recirculation depth to avoid
3599 * collisions in the exact match cache */
3600 recirc_depth = *recirc_depth_get_unsafe();
3601 if (OVS_UNLIKELY(recirc_depth)) {
3602 hash = hash_finish(hash, recirc_depth);
3603 dp_packet_set_rss_hash(packet, hash);
3608 struct packet_batch_per_flow {
3609 unsigned int byte_count;
3611 struct dp_netdev_flow *flow;
3613 struct dp_packet_batch array;
3617 packet_batch_per_flow_update(struct packet_batch_per_flow *batch,
3618 struct dp_packet *packet,
3619 const struct miniflow *mf)
3621 batch->byte_count += dp_packet_size(packet);
3622 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3623 batch->array.packets[batch->array.count++] = packet;
3627 packet_batch_per_flow_init(struct packet_batch_per_flow *batch,
3628 struct dp_netdev_flow *flow)
3630 flow->batch = batch;
3633 dp_packet_batch_init(&batch->array);
3634 batch->byte_count = 0;
3635 batch->tcp_flags = 0;
3639 packet_batch_per_flow_execute(struct packet_batch_per_flow *batch,
3640 struct dp_netdev_pmd_thread *pmd,
3643 struct dp_netdev_actions *actions;
3644 struct dp_netdev_flow *flow = batch->flow;
3646 dp_netdev_flow_used(flow, batch->array.count, batch->byte_count,
3647 batch->tcp_flags, now);
3649 actions = dp_netdev_flow_get_actions(flow);
3651 dp_netdev_execute_actions(pmd, &batch->array, true,
3652 actions->actions, actions->size);
3656 dp_netdev_queue_batches(struct dp_packet *pkt,
3657 struct dp_netdev_flow *flow, const struct miniflow *mf,
3658 struct packet_batch_per_flow *batches, size_t *n_batches)
3660 struct packet_batch_per_flow *batch = flow->batch;
3662 if (OVS_UNLIKELY(!batch)) {
3663 batch = &batches[(*n_batches)++];
3664 packet_batch_per_flow_init(batch, flow);
3667 packet_batch_per_flow_update(batch, pkt, mf);
3670 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3671 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
3672 * miniflow is copied into 'keys' and the packet pointer is moved at the
3673 * beginning of the 'packets' array.
3675 * The function returns the number of packets that needs to be processed in the
3676 * 'packets' array (they have been moved to the beginning of the vector).
3678 * If 'md_is_valid' is false, the metadata in 'packets' is not valid and must be
3679 * initialized by this function using 'port_no'.
3681 static inline size_t
3682 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet_batch *packets_,
3683 struct netdev_flow_key *keys,
3684 struct packet_batch_per_flow batches[], size_t *n_batches,
3685 bool md_is_valid, odp_port_t port_no)
3687 struct emc_cache *flow_cache = &pmd->flow_cache;
3688 struct netdev_flow_key *key = &keys[0];
3689 size_t i, n_missed = 0, n_dropped = 0;
3690 struct dp_packet **packets = packets_->packets;
3691 int cnt = packets_->count;
3693 for (i = 0; i < cnt; i++) {
3694 struct dp_netdev_flow *flow;
3695 struct dp_packet *packet = packets[i];
3697 if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) {
3698 dp_packet_delete(packet);
3704 /* Prefetch next packet data and metadata. */
3705 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3706 pkt_metadata_prefetch_init(&packets[i+1]->md);
3710 pkt_metadata_init(&packet->md, port_no);
3712 miniflow_extract(packet, &key->mf);
3713 key->len = 0; /* Not computed yet. */
3714 key->hash = dpif_netdev_packet_get_rss_hash(packet, &key->mf);
3716 flow = emc_lookup(flow_cache, key);
3717 if (OVS_LIKELY(flow)) {
3718 dp_netdev_queue_batches(packet, flow, &key->mf, batches,
3721 /* Exact match cache missed. Group missed packets together at
3722 * the beginning of the 'packets' array. */
3723 packets[n_missed] = packet;
3724 /* 'key[n_missed]' contains the key of the current packet and it
3725 * must be returned to the caller. The next key should be extracted
3726 * to 'keys[n_missed + 1]'. */
3727 key = &keys[++n_missed];
3731 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - n_dropped - n_missed);
3737 handle_packet_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet,
3738 const struct netdev_flow_key *key,
3739 struct ofpbuf *actions, struct ofpbuf *put_actions,
3742 struct ofpbuf *add_actions;
3743 struct dp_packet_batch b;
3748 match.tun_md.valid = false;
3749 miniflow_expand(&key->mf, &match.flow);
3751 ofpbuf_clear(actions);
3752 ofpbuf_clear(put_actions);
3754 dpif_flow_hash(pmd->dp->dpif, &match.flow, sizeof match.flow, &ufid);
3755 error = dp_netdev_upcall(pmd, packet, &match.flow, &match.wc,
3756 &ufid, DPIF_UC_MISS, NULL, actions,
3758 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3759 dp_packet_delete(packet);
3764 /* The Netlink encoding of datapath flow keys cannot express
3765 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3766 * tag is interpreted as exact match on the fact that there is no
3767 * VLAN. Unless we refactor a lot of code that translates between
3768 * Netlink and struct flow representations, we have to do the same
3770 if (!match.wc.masks.vlan_tci) {
3771 match.wc.masks.vlan_tci = htons(0xffff);
3774 /* We can't allow the packet batching in the next loop to execute
3775 * the actions. Otherwise, if there are any slow path actions,
3776 * we'll send the packet up twice. */
3777 packet_batch_init_packet(&b, packet);
3778 dp_netdev_execute_actions(pmd, &b, true,
3779 actions->data, actions->size);
3781 add_actions = put_actions->size ? put_actions : actions;
3782 if (OVS_LIKELY(error != ENOSPC)) {
3783 struct dp_netdev_flow *netdev_flow;
3785 /* XXX: There's a race window where a flow covering this packet
3786 * could have already been installed since we last did the flow
3787 * lookup before upcall. This could be solved by moving the
3788 * mutex lock outside the loop, but that's an awful long time
3789 * to be locking everyone out of making flow installs. If we
3790 * move to a per-core classifier, it would be reasonable. */
3791 ovs_mutex_lock(&pmd->flow_mutex);
3792 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, key);
3793 if (OVS_LIKELY(!netdev_flow)) {
3794 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3798 ovs_mutex_unlock(&pmd->flow_mutex);
3800 emc_insert(&pmd->flow_cache, key, netdev_flow);
3805 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3806 struct dp_packet_batch *packets_,
3807 struct netdev_flow_key *keys,
3808 struct packet_batch_per_flow batches[], size_t *n_batches)
3810 int cnt = packets_->count;
3811 #if !defined(__CHECKER__) && !defined(_WIN32)
3812 const size_t PKT_ARRAY_SIZE = cnt;
3814 /* Sparse or MSVC doesn't like variable length array. */
3815 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3817 struct dp_packet **packets = packets_->packets;
3818 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3819 struct dp_netdev *dp = pmd->dp;
3820 struct emc_cache *flow_cache = &pmd->flow_cache;
3821 int miss_cnt = 0, lost_cnt = 0;
3825 for (i = 0; i < cnt; i++) {
3826 /* Key length is needed in all the cases, hash computed on demand. */
3827 keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf));
3829 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3830 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3831 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3832 struct ofpbuf actions, put_actions;
3834 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3835 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3837 for (i = 0; i < cnt; i++) {
3838 struct dp_netdev_flow *netdev_flow;
3840 if (OVS_LIKELY(rules[i])) {
3844 /* It's possible that an earlier slow path execution installed
3845 * a rule covering this flow. In this case, it's a lot cheaper
3846 * to catch it here than execute a miss. */
3847 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3849 rules[i] = &netdev_flow->cr;
3854 handle_packet_upcall(pmd, packets[i], &keys[i], &actions, &put_actions,
3858 ofpbuf_uninit(&actions);
3859 ofpbuf_uninit(&put_actions);
3860 fat_rwlock_unlock(&dp->upcall_rwlock);
3861 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3862 } else if (OVS_UNLIKELY(any_miss)) {
3863 for (i = 0; i < cnt; i++) {
3864 if (OVS_UNLIKELY(!rules[i])) {
3865 dp_packet_delete(packets[i]);
3872 for (i = 0; i < cnt; i++) {
3873 struct dp_packet *packet = packets[i];
3874 struct dp_netdev_flow *flow;
3876 if (OVS_UNLIKELY(!rules[i])) {
3880 flow = dp_netdev_flow_cast(rules[i]);
3882 emc_insert(flow_cache, &keys[i], flow);
3883 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3886 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3887 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3888 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3891 /* Packets enter the datapath from a port (or from recirculation) here.
3893 * For performance reasons a caller may choose not to initialize the metadata
3894 * in 'packets': in this case 'mdinit' is false and this function needs to
3895 * initialize it using 'port_no'. If the metadata in 'packets' is already
3896 * valid, 'md_is_valid' must be true and 'port_no' will be ignored. */
3898 dp_netdev_input__(struct dp_netdev_pmd_thread *pmd,
3899 struct dp_packet_batch *packets,
3900 bool md_is_valid, odp_port_t port_no)
3902 int cnt = packets->count;
3903 #if !defined(__CHECKER__) && !defined(_WIN32)
3904 const size_t PKT_ARRAY_SIZE = cnt;
3906 /* Sparse or MSVC doesn't like variable length array. */
3907 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3909 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3910 struct packet_batch_per_flow batches[PKT_ARRAY_SIZE];
3911 long long now = time_msec();
3912 size_t newcnt, n_batches, i;
3915 newcnt = emc_processing(pmd, packets, keys, batches, &n_batches,
3916 md_is_valid, port_no);
3917 if (OVS_UNLIKELY(newcnt)) {
3918 packets->count = newcnt;
3919 fast_path_processing(pmd, packets, keys, batches, &n_batches);
3922 for (i = 0; i < n_batches; i++) {
3923 batches[i].flow->batch = NULL;
3926 for (i = 0; i < n_batches; i++) {
3927 packet_batch_per_flow_execute(&batches[i], pmd, now);
3932 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3933 struct dp_packet_batch *packets,
3936 dp_netdev_input__(pmd, packets, false, port_no);
3940 dp_netdev_recirculate(struct dp_netdev_pmd_thread *pmd,
3941 struct dp_packet_batch *packets)
3943 dp_netdev_input__(pmd, packets, true, 0);
3946 struct dp_netdev_execute_aux {
3947 struct dp_netdev_pmd_thread *pmd;
3951 dpif_netdev_register_dp_purge_cb(struct dpif *dpif, dp_purge_callback *cb,
3954 struct dp_netdev *dp = get_dp_netdev(dpif);
3955 dp->dp_purge_aux = aux;
3956 dp->dp_purge_cb = cb;
3960 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3963 struct dp_netdev *dp = get_dp_netdev(dpif);
3964 dp->upcall_aux = aux;
3968 static struct tx_port *
3969 pmd_tx_port_cache_lookup(const struct dp_netdev_pmd_thread *pmd,
3972 return tx_port_lookup(&pmd->port_cache, port_no);
3976 push_tnl_action(const struct dp_netdev_pmd_thread *pmd,
3977 const struct nlattr *attr,
3978 struct dp_packet_batch *batch)
3980 struct tx_port *tun_port;
3981 const struct ovs_action_push_tnl *data;
3984 data = nl_attr_get(attr);
3986 tun_port = pmd_tx_port_cache_lookup(pmd, u32_to_odp(data->tnl_port));
3991 err = netdev_push_header(tun_port->netdev, batch, data);
3996 dp_packet_delete_batch(batch, true);
4001 dp_execute_userspace_action(struct dp_netdev_pmd_thread *pmd,
4002 struct dp_packet *packet, bool may_steal,
4003 struct flow *flow, ovs_u128 *ufid,
4004 struct ofpbuf *actions,
4005 const struct nlattr *userdata)
4007 struct dp_packet_batch b;
4010 ofpbuf_clear(actions);
4012 error = dp_netdev_upcall(pmd, packet, flow, NULL, ufid,
4013 DPIF_UC_ACTION, userdata, actions,
4015 if (!error || error == ENOSPC) {
4016 packet_batch_init_packet(&b, packet);
4017 dp_netdev_execute_actions(pmd, &b, may_steal,
4018 actions->data, actions->size);
4019 } else if (may_steal) {
4020 dp_packet_delete(packet);
4025 dp_execute_cb(void *aux_, struct dp_packet_batch *packets_,
4026 const struct nlattr *a, bool may_steal)
4028 struct dp_netdev_execute_aux *aux = aux_;
4029 uint32_t *depth = recirc_depth_get();
4030 struct dp_netdev_pmd_thread *pmd = aux->pmd;
4031 struct dp_netdev *dp = pmd->dp;
4032 int type = nl_attr_type(a);
4035 switch ((enum ovs_action_attr)type) {
4036 case OVS_ACTION_ATTR_OUTPUT:
4037 p = pmd_tx_port_cache_lookup(pmd, u32_to_odp(nl_attr_get_u32(a)));
4038 if (OVS_LIKELY(p)) {
4041 atomic_read_relaxed(&pmd->tx_qid, &tx_qid);
4043 netdev_send(p->netdev, tx_qid, packets_, may_steal);
4048 case OVS_ACTION_ATTR_TUNNEL_PUSH:
4049 if (*depth < MAX_RECIRC_DEPTH) {
4050 struct dp_packet_batch tnl_pkt;
4051 struct dp_packet_batch *orig_packets_ = packets_;
4055 dp_packet_batch_clone(&tnl_pkt, packets_);
4056 packets_ = &tnl_pkt;
4057 dp_packet_batch_reset_cutlen(orig_packets_);
4060 dp_packet_batch_apply_cutlen(packets_);
4062 err = push_tnl_action(pmd, a, packets_);
4065 dp_netdev_recirculate(pmd, packets_);
4072 case OVS_ACTION_ATTR_TUNNEL_POP:
4073 if (*depth < MAX_RECIRC_DEPTH) {
4074 struct dp_packet_batch *orig_packets_ = packets_;
4075 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
4077 p = pmd_tx_port_cache_lookup(pmd, portno);
4079 struct dp_packet_batch tnl_pkt;
4083 dp_packet_batch_clone(&tnl_pkt, packets_);
4084 packets_ = &tnl_pkt;
4085 dp_packet_batch_reset_cutlen(orig_packets_);
4088 dp_packet_batch_apply_cutlen(packets_);
4090 netdev_pop_header(p->netdev, packets_);
4091 if (!packets_->count) {
4095 for (i = 0; i < packets_->count; i++) {
4096 packets_->packets[i]->md.in_port.odp_port = portno;
4100 dp_netdev_recirculate(pmd, packets_);
4107 case OVS_ACTION_ATTR_USERSPACE:
4108 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
4109 struct dp_packet_batch *orig_packets_ = packets_;
4110 struct dp_packet **packets = packets_->packets;
4111 const struct nlattr *userdata;
4112 struct dp_packet_batch usr_pkt;
4113 struct ofpbuf actions;
4119 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
4120 ofpbuf_init(&actions, 0);
4122 if (packets_->trunc) {
4124 dp_packet_batch_clone(&usr_pkt, packets_);
4125 packets_ = &usr_pkt;
4126 packets = packets_->packets;
4128 dp_packet_batch_reset_cutlen(orig_packets_);
4131 dp_packet_batch_apply_cutlen(packets_);
4134 for (i = 0; i < packets_->count; i++) {
4135 flow_extract(packets[i], &flow);
4136 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
4137 dp_execute_userspace_action(pmd, packets[i], may_steal, &flow,
4138 &ufid, &actions, userdata);
4142 dp_packet_delete_batch(packets_, true);
4145 ofpbuf_uninit(&actions);
4146 fat_rwlock_unlock(&dp->upcall_rwlock);
4152 case OVS_ACTION_ATTR_RECIRC:
4153 if (*depth < MAX_RECIRC_DEPTH) {
4154 struct dp_packet_batch recirc_pkts;
4158 dp_packet_batch_clone(&recirc_pkts, packets_);
4159 packets_ = &recirc_pkts;
4162 for (i = 0; i < packets_->count; i++) {
4163 packets_->packets[i]->md.recirc_id = nl_attr_get_u32(a);
4167 dp_netdev_recirculate(pmd, packets_);
4173 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
4176 case OVS_ACTION_ATTR_CT:
4177 /* If a flow with this action is slow-pathed, datapath assistance is
4178 * required to implement it. However, we don't support this action
4179 * in the userspace datapath. */
4180 VLOG_WARN("Cannot execute conntrack action in userspace.");
4183 case OVS_ACTION_ATTR_PUSH_VLAN:
4184 case OVS_ACTION_ATTR_POP_VLAN:
4185 case OVS_ACTION_ATTR_PUSH_MPLS:
4186 case OVS_ACTION_ATTR_POP_MPLS:
4187 case OVS_ACTION_ATTR_SET:
4188 case OVS_ACTION_ATTR_SET_MASKED:
4189 case OVS_ACTION_ATTR_SAMPLE:
4190 case OVS_ACTION_ATTR_HASH:
4191 case OVS_ACTION_ATTR_UNSPEC:
4192 case OVS_ACTION_ATTR_TRUNC:
4193 case __OVS_ACTION_ATTR_MAX:
4197 dp_packet_delete_batch(packets_, may_steal);
4201 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
4202 struct dp_packet_batch *packets,
4204 const struct nlattr *actions, size_t actions_len)
4206 struct dp_netdev_execute_aux aux = { pmd };
4208 odp_execute_actions(&aux, packets, may_steal, actions,
4209 actions_len, dp_execute_cb);
4212 const struct dpif_class dpif_netdev_class = {
4215 dpif_netdev_enumerate,
4216 dpif_netdev_port_open_type,
4219 dpif_netdev_destroy,
4222 dpif_netdev_get_stats,
4223 dpif_netdev_port_add,
4224 dpif_netdev_port_del,
4225 dpif_netdev_port_query_by_number,
4226 dpif_netdev_port_query_by_name,
4227 NULL, /* port_get_pid */
4228 dpif_netdev_port_dump_start,
4229 dpif_netdev_port_dump_next,
4230 dpif_netdev_port_dump_done,
4231 dpif_netdev_port_poll,
4232 dpif_netdev_port_poll_wait,
4233 dpif_netdev_flow_flush,
4234 dpif_netdev_flow_dump_create,
4235 dpif_netdev_flow_dump_destroy,
4236 dpif_netdev_flow_dump_thread_create,
4237 dpif_netdev_flow_dump_thread_destroy,
4238 dpif_netdev_flow_dump_next,
4239 dpif_netdev_operate,
4240 NULL, /* recv_set */
4241 NULL, /* handlers_set */
4242 dpif_netdev_pmd_set,
4243 dpif_netdev_queue_to_priority,
4245 NULL, /* recv_wait */
4246 NULL, /* recv_purge */
4247 dpif_netdev_register_dp_purge_cb,
4248 dpif_netdev_register_upcall_cb,
4249 dpif_netdev_enable_upcall,
4250 dpif_netdev_disable_upcall,
4251 dpif_netdev_get_datapath_version,
4252 NULL, /* ct_dump_start */
4253 NULL, /* ct_dump_next */
4254 NULL, /* ct_dump_done */
4255 NULL, /* ct_flush */
4259 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
4260 const char *argv[], void *aux OVS_UNUSED)
4262 struct dp_netdev_port *port;
4263 struct dp_netdev *dp;
4266 ovs_mutex_lock(&dp_netdev_mutex);
4267 dp = shash_find_data(&dp_netdevs, argv[1]);
4268 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
4269 ovs_mutex_unlock(&dp_netdev_mutex);
4270 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
4273 ovs_refcount_ref(&dp->ref_cnt);
4274 ovs_mutex_unlock(&dp_netdev_mutex);
4276 ovs_mutex_lock(&dp->port_mutex);
4277 if (get_port_by_name(dp, argv[2], &port)) {
4278 unixctl_command_reply_error(conn, "unknown port");
4282 port_no = u32_to_odp(atoi(argv[3]));
4283 if (!port_no || port_no == ODPP_NONE) {
4284 unixctl_command_reply_error(conn, "bad port number");
4287 if (dp_netdev_lookup_port(dp, port_no)) {
4288 unixctl_command_reply_error(conn, "port number already in use");
4293 hmap_remove(&dp->ports, &port->node);
4294 dp_netdev_del_port_from_all_pmds(dp, port);
4296 /* Reinsert with new port number. */
4297 port->port_no = port_no;
4298 hmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
4299 dp_netdev_add_port_to_pmds(dp, port);
4301 seq_change(dp->port_seq);
4302 unixctl_command_reply(conn, NULL);
4305 ovs_mutex_unlock(&dp->port_mutex);
4306 dp_netdev_unref(dp);
4310 dpif_dummy_register__(const char *type)
4312 struct dpif_class *class;
4314 class = xmalloc(sizeof *class);
4315 *class = dpif_netdev_class;
4316 class->type = xstrdup(type);
4317 dp_register_provider(class);
4321 dpif_dummy_override(const char *type)
4326 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
4327 * a userland-only build. It's useful for testsuite.
4329 error = dp_unregister_provider(type);
4330 if (error == 0 || error == EAFNOSUPPORT) {
4331 dpif_dummy_register__(type);
4336 dpif_dummy_register(enum dummy_level level)
4338 if (level == DUMMY_OVERRIDE_ALL) {
4343 dp_enumerate_types(&types);
4344 SSET_FOR_EACH (type, &types) {
4345 dpif_dummy_override(type);
4347 sset_destroy(&types);
4348 } else if (level == DUMMY_OVERRIDE_SYSTEM) {
4349 dpif_dummy_override("system");
4352 dpif_dummy_register__("dummy");
4354 unixctl_command_register("dpif-dummy/change-port-number",
4355 "dp port new-number",
4356 3, 3, dpif_dummy_change_port_number, NULL);
4359 /* Datapath Classifier. */
4361 /* A set of rules that all have the same fields wildcarded. */
4362 struct dpcls_subtable {
4363 /* The fields are only used by writers. */
4364 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
4366 /* These fields are accessed by readers. */
4367 struct cmap rules; /* Contains "struct dpcls_rule"s. */
4368 struct netdev_flow_key mask; /* Wildcards for fields (const). */
4369 /* 'mask' must be the last field, additional space is allocated here. */
4372 /* Initializes 'cls' as a classifier that initially contains no classification
4375 dpcls_init(struct dpcls *cls)
4377 cmap_init(&cls->subtables_map);
4378 pvector_init(&cls->subtables);
4382 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
4384 pvector_remove(&cls->subtables, subtable);
4385 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
4386 subtable->mask.hash);
4387 cmap_destroy(&subtable->rules);
4388 ovsrcu_postpone(free, subtable);
4391 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
4392 * caller's responsibility.
4393 * May only be called after all the readers have been terminated. */
4395 dpcls_destroy(struct dpcls *cls)
4398 struct dpcls_subtable *subtable;
4400 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
4401 ovs_assert(cmap_count(&subtable->rules) == 0);
4402 dpcls_destroy_subtable(cls, subtable);
4404 cmap_destroy(&cls->subtables_map);
4405 pvector_destroy(&cls->subtables);
4409 static struct dpcls_subtable *
4410 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
4412 struct dpcls_subtable *subtable;
4414 /* Need to add one. */
4415 subtable = xmalloc(sizeof *subtable
4416 - sizeof subtable->mask.mf + mask->len);
4417 cmap_init(&subtable->rules);
4418 netdev_flow_key_clone(&subtable->mask, mask);
4419 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
4420 pvector_insert(&cls->subtables, subtable, 0);
4421 pvector_publish(&cls->subtables);
4426 static inline struct dpcls_subtable *
4427 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
4429 struct dpcls_subtable *subtable;
4431 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
4432 &cls->subtables_map) {
4433 if (netdev_flow_key_equal(&subtable->mask, mask)) {
4437 return dpcls_create_subtable(cls, mask);
4440 /* Insert 'rule' into 'cls'. */
4442 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
4443 const struct netdev_flow_key *mask)
4445 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
4447 rule->mask = &subtable->mask;
4448 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
4451 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
4453 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
4455 struct dpcls_subtable *subtable;
4457 ovs_assert(rule->mask);
4459 INIT_CONTAINER(subtable, rule->mask, mask);
4461 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
4463 dpcls_destroy_subtable(cls, subtable);
4464 pvector_publish(&cls->subtables);
4468 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
4469 * in 'mask' the values in 'key' and 'target' are the same. */
4471 dpcls_rule_matches_key(const struct dpcls_rule *rule,
4472 const struct netdev_flow_key *target)
4474 const uint64_t *keyp = miniflow_get_values(&rule->flow.mf);
4475 const uint64_t *maskp = miniflow_get_values(&rule->mask->mf);
4478 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, target, rule->flow.mf.map) {
4479 if (OVS_UNLIKELY((value & *maskp++) != *keyp++)) {
4486 /* For each miniflow in 'flows' performs a classifier lookup writing the result
4487 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
4488 * NULL it is skipped.
4490 * This function is optimized for use in the userspace datapath and therefore
4491 * does not implement a lot of features available in the standard
4492 * classifier_lookup() function. Specifically, it does not implement
4493 * priorities, instead returning any rule which matches the flow.
4495 * Returns true if all flows found a corresponding rule. */
4497 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
4498 struct dpcls_rule **rules, const size_t cnt)
4500 /* The batch size 16 was experimentally found faster than 8 or 32. */
4501 typedef uint16_t map_type;
4502 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
4504 #if !defined(__CHECKER__) && !defined(_WIN32)
4505 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
4507 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
4509 map_type maps[N_MAPS];
4510 struct dpcls_subtable *subtable;
4512 memset(maps, 0xff, sizeof maps);
4513 if (cnt % MAP_BITS) {
4514 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
4516 memset(rules, 0, cnt * sizeof *rules);
4518 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
4519 const struct netdev_flow_key *mkeys = keys;
4520 struct dpcls_rule **mrules = rules;
4521 map_type remains = 0;
4524 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
4526 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
4527 uint32_t hashes[MAP_BITS];
4528 const struct cmap_node *nodes[MAP_BITS];
4529 unsigned long map = maps[m];
4533 continue; /* Skip empty maps. */
4536 /* Compute hashes for the remaining keys. */
4537 ULLONG_FOR_EACH_1(i, map) {
4538 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
4542 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
4543 /* Check results. */
4544 ULLONG_FOR_EACH_1(i, map) {
4545 struct dpcls_rule *rule;
4547 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
4548 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
4553 ULLONG_SET0(map, i); /* Did not match. */
4555 ; /* Keep Sparse happy. */
4557 maps[m] &= ~map; /* Clear the found rules. */
4561 return true; /* All found. */
4564 return false; /* Some misses. */