2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "dpif-netdev.h"
24 #include <netinet/in.h>
25 #include <sys/socket.h>
30 #include <sys/ioctl.h>
36 #include "dp-packet.h"
38 #include "dpif-provider.h"
40 #include "dynamic-string.h"
41 #include "fat-rwlock.h"
47 #include "meta-flow.h"
49 #include "netdev-dpdk.h"
50 #include "netdev-vport.h"
52 #include "odp-execute.h"
54 #include "ofp-print.h"
59 #include "poll-loop.h"
66 #include "tnl-arp-cache.h"
69 #include "openvswitch/vlog.h"
71 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
73 #define FLOW_DUMP_MAX_BATCH 50
74 /* Use per thread recirc_depth to prevent recirculation loop. */
75 #define MAX_RECIRC_DEPTH 5
76 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
78 /* Configuration parameters. */
79 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
81 /* Protects against changes to 'dp_netdevs'. */
82 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
84 /* Contains all 'struct dp_netdev's. */
85 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
86 = SHASH_INITIALIZER(&dp_netdevs);
88 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
90 /* Stores a miniflow with inline values */
92 struct netdev_flow_key {
93 uint32_t hash; /* Hash function differs for different users. */
94 uint32_t len; /* Length of the following miniflow (incl. map). */
96 uint64_t buf[FLOW_MAX_PACKET_U64S - MINI_N_INLINE];
99 /* Exact match cache for frequently used flows
101 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
102 * search its entries for a miniflow that matches exactly the miniflow of the
103 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
105 * A cache entry holds a reference to its 'dp_netdev_flow'.
107 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
108 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
109 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
110 * value is the index of a cache entry where the miniflow could be.
116 * Each pmd_thread has its own private exact match cache.
117 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
120 #define EM_FLOW_HASH_SHIFT 13
121 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
122 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
123 #define EM_FLOW_HASH_SEGS 2
126 struct dp_netdev_flow *flow;
127 struct netdev_flow_key key; /* key.hash used for emc hash value. */
131 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
132 int sweep_idx; /* For emc_cache_slow_sweep(). */
135 /* Iterate in the exact match cache through every entry that might contain a
136 * miniflow with hash 'HASH'. */
137 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
138 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
139 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
140 i__ < EM_FLOW_HASH_SEGS; \
141 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
143 /* Simple non-wildcarding single-priority classifier. */
146 struct cmap subtables_map;
147 struct pvector subtables;
150 /* A rule to be inserted to the classifier. */
152 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
153 struct netdev_flow_key *mask; /* Subtable's mask. */
154 struct netdev_flow_key flow; /* Matching key. */
155 /* 'flow' must be the last field, additional space is allocated here. */
158 static void dpcls_init(struct dpcls *);
159 static void dpcls_destroy(struct dpcls *);
160 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
161 const struct netdev_flow_key *mask);
162 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
163 static bool dpcls_lookup(const struct dpcls *cls,
164 const struct netdev_flow_key keys[],
165 struct dpcls_rule **rules, size_t cnt);
167 /* Datapath based on the network device interface from netdev.h.
173 * Some members, marked 'const', are immutable. Accessing other members
174 * requires synchronization, as noted in more detail below.
176 * Acquisition order is, from outermost to innermost:
178 * dp_netdev_mutex (global)
182 const struct dpif_class *const class;
183 const char *const name;
185 struct ovs_refcount ref_cnt;
186 atomic_flag destroyed;
190 * Protected by RCU. Take the mutex to add or remove ports. */
191 struct ovs_mutex port_mutex;
193 struct seq *port_seq; /* Incremented whenever a port changes. */
195 /* Protects access to ofproto-dpif-upcall interface during revalidator
196 * thread synchronization. */
197 struct fat_rwlock upcall_rwlock;
198 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
201 /* Stores all 'struct dp_netdev_pmd_thread's. */
202 struct cmap poll_threads;
204 /* Protects the access of the 'struct dp_netdev_pmd_thread'
205 * instance for non-pmd thread. */
206 struct ovs_mutex non_pmd_mutex;
208 /* Each pmd thread will store its pointer to
209 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
210 ovsthread_key_t per_pmd_key;
212 /* Number of rx queues for each dpdk interface and the cpu mask
213 * for pin of pmd threads. */
216 uint64_t last_tnl_conf_seq;
219 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
223 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
224 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
225 DP_STAT_MISS, /* Packets that did not match. */
226 DP_STAT_LOST, /* Packets not passed up to the client. */
230 enum pmd_cycles_counter_type {
231 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
232 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
236 /* A port in a netdev-based datapath. */
237 struct dp_netdev_port {
238 struct pkt_metadata md;
239 struct netdev *netdev;
240 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
241 struct netdev_saved_flags *sf;
242 struct netdev_rxq **rxq;
243 struct ovs_refcount ref_cnt;
244 char *type; /* Port type as requested by user. */
247 /* Contained by struct dp_netdev_flow's 'stats' member. */
248 struct dp_netdev_flow_stats {
249 atomic_llong used; /* Last used time, in monotonic msecs. */
250 atomic_ullong packet_count; /* Number of packets matched. */
251 atomic_ullong byte_count; /* Number of bytes matched. */
252 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
255 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
261 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
262 * its pmd thread's classifier. The text below calls this classifier 'cls'.
267 * The thread safety rules described here for "struct dp_netdev_flow" are
268 * motivated by two goals:
270 * - Prevent threads that read members of "struct dp_netdev_flow" from
271 * reading bad data due to changes by some thread concurrently modifying
274 * - Prevent two threads making changes to members of a given "struct
275 * dp_netdev_flow" from interfering with each other.
281 * A flow 'flow' may be accessed without a risk of being freed during an RCU
282 * grace period. Code that needs to hold onto a flow for a while
283 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
285 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
286 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
289 * Some members, marked 'const', are immutable. Accessing other members
290 * requires synchronization, as noted in more detail below.
292 struct dp_netdev_flow {
293 const struct flow flow; /* Unmasked flow that created this entry. */
294 /* Hash table index by unmasked flow. */
295 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
297 const ovs_u128 ufid; /* Unique flow identifier. */
298 const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */
301 /* Number of references.
302 * The classifier owns one reference.
303 * Any thread trying to keep a rule from being freed should hold its own
305 struct ovs_refcount ref_cnt;
310 struct dp_netdev_flow_stats stats;
313 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
315 /* While processing a group of input packets, the datapath uses the next
316 * member to store a pointer to the output batch for the flow. It is
317 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
318 * packet_batch_init() and packet_batch_execute()). */
319 struct packet_batch *batch;
321 /* Packet classification. */
322 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
323 /* 'cr' must be the last member. */
326 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
327 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
328 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
331 /* A set of datapath actions within a "struct dp_netdev_flow".
337 * A struct dp_netdev_actions 'actions' is protected with RCU. */
338 struct dp_netdev_actions {
339 /* These members are immutable: they do not change during the struct's
341 unsigned int size; /* Size of 'actions', in bytes. */
342 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
345 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
347 struct dp_netdev_actions *dp_netdev_flow_get_actions(
348 const struct dp_netdev_flow *);
349 static void dp_netdev_actions_free(struct dp_netdev_actions *);
351 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
352 struct dp_netdev_pmd_stats {
353 /* Indexed by DP_STAT_*. */
354 atomic_ullong n[DP_N_STATS];
357 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
358 struct dp_netdev_pmd_cycles {
359 /* Indexed by PMD_CYCLES_*. */
360 atomic_ullong n[PMD_N_CYCLES];
363 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
364 * the performance overhead of interrupt processing. Therefore netdev can
365 * not implement rx-wait for these devices. dpif-netdev needs to poll
366 * these device to check for recv buffer. pmd-thread does polling for
367 * devices assigned to itself.
369 * DPDK used PMD for accessing NIC.
371 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
372 * I/O of all non-pmd threads. There will be no actual thread created
375 * Each struct has its own flow table and classifier. Packets received
376 * from managed ports are looked up in the corresponding pmd thread's
377 * flow table, and are executed with the found actions.
379 struct dp_netdev_pmd_thread {
380 struct dp_netdev *dp;
381 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
382 struct cmap_node node; /* In 'dp->poll_threads'. */
384 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
385 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
387 /* Per thread exact-match cache. Note, the instance for cpu core
388 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
389 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
390 * instances will only be accessed by its own pmd thread. */
391 struct emc_cache flow_cache;
393 /* Classifier and Flow-Table.
395 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
396 * changes to 'cls' must be made while still holding the 'flow_mutex'.
398 struct ovs_mutex flow_mutex;
400 struct cmap flow_table OVS_GUARDED; /* Flow table. */
403 struct dp_netdev_pmd_stats stats;
405 /* Cycles counters */
406 struct dp_netdev_pmd_cycles cycles;
408 /* Used to count cicles. See 'cycles_counter_end()' */
409 unsigned long long last_cycles;
411 struct latch exit_latch; /* For terminating the pmd thread. */
412 atomic_uint change_seq; /* For reloading pmd ports. */
414 int index; /* Idx of this pmd thread among pmd*/
415 /* threads on same numa node. */
416 unsigned core_id; /* CPU core id of this pmd thread. */
417 int numa_id; /* numa node id of this pmd thread. */
418 int tx_qid; /* Queue id used by this pmd thread to
419 * send packets on all netdevs */
421 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
422 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
423 * values and subtracts them from 'stats' and 'cycles' before
424 * reporting to the user */
425 unsigned long long stats_zero[DP_N_STATS];
426 uint64_t cycles_zero[PMD_N_CYCLES];
429 #define PMD_INITIAL_SEQ 1
431 /* Interface to netdev-based datapath. */
434 struct dp_netdev *dp;
435 uint64_t last_port_seq;
438 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
439 struct dp_netdev_port **portp);
440 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
441 struct dp_netdev_port **portp);
442 static void dp_netdev_free(struct dp_netdev *)
443 OVS_REQUIRES(dp_netdev_mutex);
444 static int do_add_port(struct dp_netdev *dp, const char *devname,
445 const char *type, odp_port_t port_no)
446 OVS_REQUIRES(dp->port_mutex);
447 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
448 OVS_REQUIRES(dp->port_mutex);
449 static int dpif_netdev_open(const struct dpif_class *, const char *name,
450 bool create, struct dpif **);
451 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
452 struct dp_packet **, int c,
454 const struct nlattr *actions,
456 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
457 struct dp_packet **, int cnt);
459 static void dp_netdev_disable_upcall(struct dp_netdev *);
460 void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
461 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
462 struct dp_netdev *dp, int index,
463 unsigned core_id, int numa_id);
464 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
465 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
466 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
468 static struct dp_netdev_pmd_thread *
469 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
470 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
471 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
472 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
473 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
474 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
475 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
476 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
478 static inline bool emc_entry_alive(struct emc_entry *ce);
479 static void emc_clear_entry(struct emc_entry *ce);
482 emc_cache_init(struct emc_cache *flow_cache)
486 BUILD_ASSERT(offsetof(struct miniflow, inline_values) == sizeof(uint64_t));
488 flow_cache->sweep_idx = 0;
489 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
490 flow_cache->entries[i].flow = NULL;
491 flow_cache->entries[i].key.hash = 0;
492 flow_cache->entries[i].key.len
493 = offsetof(struct miniflow, inline_values);
494 miniflow_initialize(&flow_cache->entries[i].key.mf,
495 flow_cache->entries[i].key.buf);
500 emc_cache_uninit(struct emc_cache *flow_cache)
504 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
505 emc_clear_entry(&flow_cache->entries[i]);
509 /* Check and clear dead flow references slowly (one entry at each
512 emc_cache_slow_sweep(struct emc_cache *flow_cache)
514 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
516 if (!emc_entry_alive(entry)) {
517 emc_clear_entry(entry);
519 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
522 static struct dpif_netdev *
523 dpif_netdev_cast(const struct dpif *dpif)
525 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
526 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
529 static struct dp_netdev *
530 get_dp_netdev(const struct dpif *dpif)
532 return dpif_netdev_cast(dpif)->dp;
536 PMD_INFO_SHOW_STATS, /* show how cpu cycles are spent */
537 PMD_INFO_CLEAR_STATS /* set the cycles count to 0 */
541 pmd_info_show_stats(struct ds *reply,
542 struct dp_netdev_pmd_thread *pmd,
543 unsigned long long stats[DP_N_STATS],
544 uint64_t cycles[PMD_N_CYCLES])
546 unsigned long long total_packets = 0;
547 uint64_t total_cycles = 0;
550 /* These loops subtracts reference values ('*_zero') from the counters.
551 * Since loads and stores are relaxed, it might be possible for a '*_zero'
552 * value to be more recent than the current value we're reading from the
553 * counter. This is not a big problem, since these numbers are not
554 * supposed to be too accurate, but we should at least make sure that
555 * the result is not negative. */
556 for (i = 0; i < DP_N_STATS; i++) {
557 if (stats[i] > pmd->stats_zero[i]) {
558 stats[i] -= pmd->stats_zero[i];
563 if (i != DP_STAT_LOST) {
564 /* Lost packets are already included in DP_STAT_MISS */
565 total_packets += stats[i];
569 for (i = 0; i < PMD_N_CYCLES; i++) {
570 if (cycles[i] > pmd->cycles_zero[i]) {
571 cycles[i] -= pmd->cycles_zero[i];
576 total_cycles += cycles[i];
579 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
580 ? "main thread" : "pmd thread");
582 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
583 ds_put_format(reply, " numa_id %d", pmd->numa_id);
585 if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) {
586 ds_put_format(reply, " core_id %u", pmd->core_id);
588 ds_put_cstr(reply, ":\n");
591 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
592 "\tmiss:%llu\n\tlost:%llu\n",
593 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
594 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
596 if (total_cycles == 0) {
601 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
602 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
603 cycles[PMD_CYCLES_POLLING],
604 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
605 cycles[PMD_CYCLES_PROCESSING],
606 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
608 if (total_packets == 0) {
613 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
614 total_cycles / (double)total_packets,
615 total_cycles, total_packets);
618 "\tavg processing cycles per packet: "
619 "%.02f (%"PRIu64"/%llu)\n",
620 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
621 cycles[PMD_CYCLES_PROCESSING], total_packets);
625 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
626 struct dp_netdev_pmd_thread *pmd,
627 unsigned long long stats[DP_N_STATS],
628 uint64_t cycles[PMD_N_CYCLES])
632 /* We cannot write 'stats' and 'cycles' (because they're written by other
633 * threads) and we shouldn't change 'stats' (because they're used to count
634 * datapath stats, which must not be cleared here). Instead, we save the
635 * current values and subtract them from the values to be displayed in the
637 for (i = 0; i < DP_N_STATS; i++) {
638 pmd->stats_zero[i] = stats[i];
640 for (i = 0; i < PMD_N_CYCLES; i++) {
641 pmd->cycles_zero[i] = cycles[i];
646 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
649 struct ds reply = DS_EMPTY_INITIALIZER;
650 struct dp_netdev_pmd_thread *pmd;
651 struct dp_netdev *dp = NULL;
652 enum pmd_info_type type = *(enum pmd_info_type *) aux;
654 ovs_mutex_lock(&dp_netdev_mutex);
657 dp = shash_find_data(&dp_netdevs, argv[1]);
658 } else if (shash_count(&dp_netdevs) == 1) {
659 /* There's only one datapath */
660 dp = shash_first(&dp_netdevs)->data;
664 ovs_mutex_unlock(&dp_netdev_mutex);
665 unixctl_command_reply_error(conn,
666 "please specify an existing datapath");
670 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
671 unsigned long long stats[DP_N_STATS];
672 uint64_t cycles[PMD_N_CYCLES];
675 /* Read current stats and cycle counters */
676 for (i = 0; i < ARRAY_SIZE(stats); i++) {
677 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
679 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
680 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
683 if (type == PMD_INFO_CLEAR_STATS) {
684 pmd_info_clear_stats(&reply, pmd, stats, cycles);
685 } else if (type == PMD_INFO_SHOW_STATS) {
686 pmd_info_show_stats(&reply, pmd, stats, cycles);
690 ovs_mutex_unlock(&dp_netdev_mutex);
692 unixctl_command_reply(conn, ds_cstr(&reply));
697 dpif_netdev_init(void)
699 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
700 clear_aux = PMD_INFO_CLEAR_STATS;
702 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
703 0, 1, dpif_netdev_pmd_info,
705 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
706 0, 1, dpif_netdev_pmd_info,
712 dpif_netdev_enumerate(struct sset *all_dps,
713 const struct dpif_class *dpif_class)
715 struct shash_node *node;
717 ovs_mutex_lock(&dp_netdev_mutex);
718 SHASH_FOR_EACH(node, &dp_netdevs) {
719 struct dp_netdev *dp = node->data;
720 if (dpif_class != dp->class) {
721 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
722 * If the class doesn't match, skip this dpif. */
725 sset_add(all_dps, node->name);
727 ovs_mutex_unlock(&dp_netdev_mutex);
733 dpif_netdev_class_is_dummy(const struct dpif_class *class)
735 return class != &dpif_netdev_class;
739 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
741 return strcmp(type, "internal") ? type
742 : dpif_netdev_class_is_dummy(class) ? "dummy"
747 create_dpif_netdev(struct dp_netdev *dp)
749 uint16_t netflow_id = hash_string(dp->name, 0);
750 struct dpif_netdev *dpif;
752 ovs_refcount_ref(&dp->ref_cnt);
754 dpif = xmalloc(sizeof *dpif);
755 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
757 dpif->last_port_seq = seq_read(dp->port_seq);
762 /* Choose an unused, non-zero port number and return it on success.
763 * Return ODPP_NONE on failure. */
765 choose_port(struct dp_netdev *dp, const char *name)
766 OVS_REQUIRES(dp->port_mutex)
770 if (dp->class != &dpif_netdev_class) {
774 /* If the port name begins with "br", start the number search at
775 * 100 to make writing tests easier. */
776 if (!strncmp(name, "br", 2)) {
780 /* If the port name contains a number, try to assign that port number.
781 * This can make writing unit tests easier because port numbers are
783 for (p = name; *p != '\0'; p++) {
784 if (isdigit((unsigned char) *p)) {
785 port_no = start_no + strtol(p, NULL, 10);
786 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
787 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
788 return u32_to_odp(port_no);
795 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
796 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
797 return u32_to_odp(port_no);
805 create_dp_netdev(const char *name, const struct dpif_class *class,
806 struct dp_netdev **dpp)
807 OVS_REQUIRES(dp_netdev_mutex)
809 struct dp_netdev *dp;
812 dp = xzalloc(sizeof *dp);
813 shash_add(&dp_netdevs, name, dp);
815 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
816 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
817 ovs_refcount_init(&dp->ref_cnt);
818 atomic_flag_clear(&dp->destroyed);
820 ovs_mutex_init(&dp->port_mutex);
821 cmap_init(&dp->ports);
822 dp->port_seq = seq_create();
823 fat_rwlock_init(&dp->upcall_rwlock);
825 /* Disable upcalls by default. */
826 dp_netdev_disable_upcall(dp);
827 dp->upcall_aux = NULL;
828 dp->upcall_cb = NULL;
830 cmap_init(&dp->poll_threads);
831 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
832 ovsthread_key_create(&dp->per_pmd_key, NULL);
834 dp_netdev_set_nonpmd(dp);
835 dp->n_dpdk_rxqs = NR_QUEUE;
837 ovs_mutex_lock(&dp->port_mutex);
838 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
839 ovs_mutex_unlock(&dp->port_mutex);
845 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
851 dpif_netdev_open(const struct dpif_class *class, const char *name,
852 bool create, struct dpif **dpifp)
854 struct dp_netdev *dp;
857 ovs_mutex_lock(&dp_netdev_mutex);
858 dp = shash_find_data(&dp_netdevs, name);
860 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
862 error = (dp->class != class ? EINVAL
867 *dpifp = create_dpif_netdev(dp);
870 ovs_mutex_unlock(&dp_netdev_mutex);
876 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
877 OVS_NO_THREAD_SAFETY_ANALYSIS
879 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
880 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
882 /* Before freeing a lock we should release it */
883 fat_rwlock_unlock(&dp->upcall_rwlock);
884 fat_rwlock_destroy(&dp->upcall_rwlock);
887 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
888 * through the 'dp_netdevs' shash while freeing 'dp'. */
890 dp_netdev_free(struct dp_netdev *dp)
891 OVS_REQUIRES(dp_netdev_mutex)
893 struct dp_netdev_port *port;
895 shash_find_and_delete(&dp_netdevs, dp->name);
897 dp_netdev_destroy_all_pmds(dp);
898 cmap_destroy(&dp->poll_threads);
899 ovs_mutex_destroy(&dp->non_pmd_mutex);
900 ovsthread_key_delete(dp->per_pmd_key);
902 ovs_mutex_lock(&dp->port_mutex);
903 CMAP_FOR_EACH (port, node, &dp->ports) {
904 do_del_port(dp, port);
906 ovs_mutex_unlock(&dp->port_mutex);
908 seq_destroy(dp->port_seq);
909 cmap_destroy(&dp->ports);
911 /* Upcalls must be disabled at this point */
912 dp_netdev_destroy_upcall_lock(dp);
915 free(CONST_CAST(char *, dp->name));
920 dp_netdev_unref(struct dp_netdev *dp)
923 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
924 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
925 ovs_mutex_lock(&dp_netdev_mutex);
926 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
929 ovs_mutex_unlock(&dp_netdev_mutex);
934 dpif_netdev_close(struct dpif *dpif)
936 struct dp_netdev *dp = get_dp_netdev(dpif);
943 dpif_netdev_destroy(struct dpif *dpif)
945 struct dp_netdev *dp = get_dp_netdev(dpif);
947 if (!atomic_flag_test_and_set(&dp->destroyed)) {
948 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
949 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
957 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
958 * load/store semantics. While the increment is not atomic, the load and
959 * store operations are, making it impossible to read inconsistent values.
961 * This is used to update thread local stats counters. */
963 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
965 unsigned long long tmp;
967 atomic_read_relaxed(var, &tmp);
969 atomic_store_relaxed(var, tmp);
973 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
975 struct dp_netdev *dp = get_dp_netdev(dpif);
976 struct dp_netdev_pmd_thread *pmd;
978 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
979 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
980 unsigned long long n;
981 stats->n_flows += cmap_count(&pmd->flow_table);
983 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
985 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
987 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
988 stats->n_missed += n;
989 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
992 stats->n_masks = UINT32_MAX;
993 stats->n_mask_hit = UINT64_MAX;
999 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1003 if (pmd->core_id == NON_PMD_CORE_ID) {
1007 ovs_mutex_lock(&pmd->cond_mutex);
1008 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1009 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1010 ovs_mutex_unlock(&pmd->cond_mutex);
1013 /* Causes all pmd threads to reload its tx/rx devices.
1014 * Must be called after adding/removing ports. */
1016 dp_netdev_reload_pmds(struct dp_netdev *dp)
1018 struct dp_netdev_pmd_thread *pmd;
1020 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1021 dp_netdev_reload_pmd__(pmd);
1026 hash_port_no(odp_port_t port_no)
1028 return hash_int(odp_to_u32(port_no), 0);
1032 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1034 OVS_REQUIRES(dp->port_mutex)
1036 struct netdev_saved_flags *sf;
1037 struct dp_netdev_port *port;
1038 struct netdev *netdev;
1039 enum netdev_flags flags;
1040 const char *open_type;
1044 /* Reject devices already in 'dp'. */
1045 if (!get_port_by_name(dp, devname, &port)) {
1049 /* Open and validate network device. */
1050 open_type = dpif_netdev_port_open_type(dp->class, type);
1051 error = netdev_open(devname, open_type, &netdev);
1055 /* XXX reject non-Ethernet devices */
1057 netdev_get_flags(netdev, &flags);
1058 if (flags & NETDEV_LOOPBACK) {
1059 VLOG_ERR("%s: cannot add a loopback device", devname);
1060 netdev_close(netdev);
1064 if (netdev_is_pmd(netdev)) {
1065 int n_cores = ovs_numa_get_n_cores();
1067 if (n_cores == OVS_CORE_UNSPEC) {
1068 VLOG_ERR("%s, cannot get cpu core info", devname);
1071 /* There can only be ovs_numa_get_n_cores() pmd threads,
1072 * so creates a txq for each, and one extra for the non
1074 error = netdev_set_multiq(netdev, n_cores + 1, dp->n_dpdk_rxqs);
1075 if (error && (error != EOPNOTSUPP)) {
1076 VLOG_ERR("%s, cannot set multiq", devname);
1080 port = xzalloc(sizeof *port);
1081 port->md = PKT_METADATA_INITIALIZER(port_no);
1082 port->netdev = netdev;
1083 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
1084 port->type = xstrdup(type);
1085 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1086 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1088 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
1089 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1090 devname, ovs_strerror(errno));
1091 netdev_close(netdev);
1099 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1101 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1102 netdev_rxq_close(port->rxq[i]);
1104 netdev_close(netdev);
1112 ovs_refcount_init(&port->ref_cnt);
1113 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1115 if (netdev_is_pmd(netdev)) {
1116 dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
1117 dp_netdev_reload_pmds(dp);
1119 seq_change(dp->port_seq);
1125 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1126 odp_port_t *port_nop)
1128 struct dp_netdev *dp = get_dp_netdev(dpif);
1129 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1130 const char *dpif_port;
1134 ovs_mutex_lock(&dp->port_mutex);
1135 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1136 if (*port_nop != ODPP_NONE) {
1137 port_no = *port_nop;
1138 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1140 port_no = choose_port(dp, dpif_port);
1141 error = port_no == ODPP_NONE ? EFBIG : 0;
1144 *port_nop = port_no;
1145 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1147 ovs_mutex_unlock(&dp->port_mutex);
1153 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1155 struct dp_netdev *dp = get_dp_netdev(dpif);
1158 ovs_mutex_lock(&dp->port_mutex);
1159 if (port_no == ODPP_LOCAL) {
1162 struct dp_netdev_port *port;
1164 error = get_port_by_number(dp, port_no, &port);
1166 do_del_port(dp, port);
1169 ovs_mutex_unlock(&dp->port_mutex);
1175 is_valid_port_number(odp_port_t port_no)
1177 return port_no != ODPP_NONE;
1180 static struct dp_netdev_port *
1181 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1183 struct dp_netdev_port *port;
1185 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1186 if (port->md.in_port.odp_port == port_no) {
1194 get_port_by_number(struct dp_netdev *dp,
1195 odp_port_t port_no, struct dp_netdev_port **portp)
1197 if (!is_valid_port_number(port_no)) {
1201 *portp = dp_netdev_lookup_port(dp, port_no);
1202 return *portp ? 0 : ENOENT;
1207 port_ref(struct dp_netdev_port *port)
1210 ovs_refcount_ref(&port->ref_cnt);
1215 port_try_ref(struct dp_netdev_port *port)
1218 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
1225 port_unref(struct dp_netdev_port *port)
1227 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1228 int n_rxq = netdev_n_rxq(port->netdev);
1231 netdev_close(port->netdev);
1232 netdev_restore_flags(port->sf);
1234 for (i = 0; i < n_rxq; i++) {
1235 netdev_rxq_close(port->rxq[i]);
1244 get_port_by_name(struct dp_netdev *dp,
1245 const char *devname, struct dp_netdev_port **portp)
1246 OVS_REQUIRES(dp->port_mutex)
1248 struct dp_netdev_port *port;
1250 CMAP_FOR_EACH (port, node, &dp->ports) {
1251 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1260 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1262 struct dp_netdev_pmd_thread *pmd;
1265 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1266 if (pmd->numa_id == numa_id) {
1274 /* Returns 'true' if there is a port with pmd netdev and the netdev
1275 * is on numa node 'numa_id'. */
1277 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1279 struct dp_netdev_port *port;
1281 CMAP_FOR_EACH (port, node, &dp->ports) {
1282 if (netdev_is_pmd(port->netdev)
1283 && netdev_get_numa_id(port->netdev) == numa_id) {
1293 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1294 OVS_REQUIRES(dp->port_mutex)
1296 cmap_remove(&dp->ports, &port->node,
1297 hash_odp_port(port->md.in_port.odp_port));
1298 seq_change(dp->port_seq);
1299 if (netdev_is_pmd(port->netdev)) {
1300 int numa_id = netdev_get_numa_id(port->netdev);
1302 /* If there is no netdev on the numa node, deletes the pmd threads
1303 * for that numa. Else, just reloads the queues. */
1304 if (!has_pmd_port_for_numa(dp, numa_id)) {
1305 dp_netdev_del_pmds_on_numa(dp, numa_id);
1307 dp_netdev_reload_pmds(dp);
1314 answer_port_query(const struct dp_netdev_port *port,
1315 struct dpif_port *dpif_port)
1317 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1318 dpif_port->type = xstrdup(port->type);
1319 dpif_port->port_no = port->md.in_port.odp_port;
1323 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1324 struct dpif_port *dpif_port)
1326 struct dp_netdev *dp = get_dp_netdev(dpif);
1327 struct dp_netdev_port *port;
1330 error = get_port_by_number(dp, port_no, &port);
1331 if (!error && dpif_port) {
1332 answer_port_query(port, dpif_port);
1339 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1340 struct dpif_port *dpif_port)
1342 struct dp_netdev *dp = get_dp_netdev(dpif);
1343 struct dp_netdev_port *port;
1346 ovs_mutex_lock(&dp->port_mutex);
1347 error = get_port_by_name(dp, devname, &port);
1348 if (!error && dpif_port) {
1349 answer_port_query(port, dpif_port);
1351 ovs_mutex_unlock(&dp->port_mutex);
1357 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1359 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1363 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1365 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1366 ovsrcu_postpone(dp_netdev_flow_free, flow);
1371 dp_netdev_flow_hash(const ovs_u128 *ufid)
1373 return ufid->u32[0];
1377 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1378 struct dp_netdev_flow *flow)
1379 OVS_REQUIRES(pmd->flow_mutex)
1381 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1383 dpcls_remove(&pmd->cls, &flow->cr);
1384 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1387 dp_netdev_flow_unref(flow);
1391 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1393 struct dp_netdev_flow *netdev_flow;
1395 ovs_mutex_lock(&pmd->flow_mutex);
1396 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1397 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1399 ovs_mutex_unlock(&pmd->flow_mutex);
1403 dpif_netdev_flow_flush(struct dpif *dpif)
1405 struct dp_netdev *dp = get_dp_netdev(dpif);
1406 struct dp_netdev_pmd_thread *pmd;
1408 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1409 dp_netdev_pmd_flow_flush(pmd);
1415 struct dp_netdev_port_state {
1416 struct cmap_position position;
1421 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1423 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1428 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1429 struct dpif_port *dpif_port)
1431 struct dp_netdev_port_state *state = state_;
1432 struct dp_netdev *dp = get_dp_netdev(dpif);
1433 struct cmap_node *node;
1436 node = cmap_next_position(&dp->ports, &state->position);
1438 struct dp_netdev_port *port;
1440 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1443 state->name = xstrdup(netdev_get_name(port->netdev));
1444 dpif_port->name = state->name;
1445 dpif_port->type = port->type;
1446 dpif_port->port_no = port->md.in_port.odp_port;
1457 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1459 struct dp_netdev_port_state *state = state_;
1466 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1468 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1469 uint64_t new_port_seq;
1472 new_port_seq = seq_read(dpif->dp->port_seq);
1473 if (dpif->last_port_seq != new_port_seq) {
1474 dpif->last_port_seq = new_port_seq;
1484 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1486 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1488 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1491 static struct dp_netdev_flow *
1492 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1494 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1497 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1499 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1502 /* netdev_flow_key utilities.
1504 * netdev_flow_key is basically a miniflow. We use these functions
1505 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1506 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1508 * - Since we are dealing exclusively with miniflows created by
1509 * miniflow_extract(), if the map is different the miniflow is different.
1510 * Therefore we can be faster by comparing the map and the miniflow in a
1512 * _ netdev_flow_key's miniflow has always inline values.
1513 * - These functions can be inlined by the compiler.
1515 * The following assertions make sure that what we're doing with miniflow is
1518 BUILD_ASSERT_DECL(offsetof(struct miniflow, inline_values)
1519 == sizeof(uint64_t));
1521 /* Given the number of bits set in the miniflow map, returns the size of the
1522 * 'netdev_flow_key.mf' */
1523 static inline uint32_t
1524 netdev_flow_key_size(uint32_t flow_u32s)
1526 return offsetof(struct miniflow, inline_values) +
1527 MINIFLOW_VALUES_SIZE(flow_u32s);
1531 netdev_flow_key_equal(const struct netdev_flow_key *a,
1532 const struct netdev_flow_key *b)
1534 /* 'b->len' may be not set yet. */
1535 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1538 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1539 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1540 * generated by miniflow_extract. */
1542 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1543 const struct miniflow *mf)
1545 return !memcmp(&key->mf, mf, key->len);
1549 netdev_flow_key_clone(struct netdev_flow_key *dst,
1550 const struct netdev_flow_key *src)
1553 offsetof(struct netdev_flow_key, mf) + src->len);
1558 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1559 const struct flow *src)
1561 struct dp_packet packet;
1562 uint64_t buf_stub[512 / 8];
1564 miniflow_initialize(&dst->mf, dst->buf);
1566 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1567 pkt_metadata_from_flow(&packet.md, src);
1568 flow_compose(&packet, src);
1569 miniflow_extract(&packet, &dst->mf);
1570 dp_packet_uninit(&packet);
1572 dst->len = netdev_flow_key_size(count_1bits(dst->mf.map));
1573 dst->hash = 0; /* Not computed yet. */
1576 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1578 netdev_flow_mask_init(struct netdev_flow_key *mask,
1579 const struct match *match)
1581 const uint64_t *mask_u64 = (const uint64_t *) &match->wc.masks;
1582 uint64_t *dst = mask->mf.inline_values;
1583 uint64_t map, mask_map = 0;
1587 /* Only check masks that make sense for the flow. */
1588 map = flow_wc_map(&match->flow);
1591 uint64_t rm1bit = rightmost_1bit(map);
1592 int i = raw_ctz(map);
1596 *dst++ = mask_u64[i];
1597 hash = hash_add64(hash, mask_u64[i]);
1602 mask->mf.values_inline = true;
1603 mask->mf.map = mask_map;
1605 hash = hash_add64(hash, mask_map);
1607 n = dst - mask->mf.inline_values;
1609 mask->hash = hash_finish(hash, n * 8);
1610 mask->len = netdev_flow_key_size(n);
1613 /* Initializes 'dst' as a copy of 'src' masked with 'mask'. */
1615 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1616 const struct flow *flow,
1617 const struct netdev_flow_key *mask)
1619 uint64_t *dst_u64 = dst->mf.inline_values;
1620 const uint64_t *mask_u64 = mask->mf.inline_values;
1624 dst->len = mask->len;
1625 dst->mf.values_inline = true;
1626 dst->mf.map = mask->mf.map;
1628 FLOW_FOR_EACH_IN_MAP(value, flow, mask->mf.map) {
1629 *dst_u64 = value & *mask_u64++;
1630 hash = hash_add64(hash, *dst_u64++);
1632 dst->hash = hash_finish(hash, (dst_u64 - dst->mf.inline_values) * 8);
1635 /* Iterate through all netdev_flow_key u64 values specified by 'MAP' */
1636 #define NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(VALUE, KEY, MAP) \
1637 for (struct mf_for_each_in_map_aux aux__ \
1638 = { (KEY)->mf.inline_values, (KEY)->mf.map, MAP }; \
1639 mf_get_next_in_map(&aux__, &(VALUE)); \
1642 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1644 static inline uint32_t
1645 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1646 const struct netdev_flow_key *mask)
1648 const uint64_t *p = mask->mf.inline_values;
1652 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(key_u64, key, mask->mf.map) {
1653 hash = hash_add64(hash, key_u64 & *p++);
1656 return hash_finish(hash, (p - mask->mf.inline_values) * 8);
1660 emc_entry_alive(struct emc_entry *ce)
1662 return ce->flow && !ce->flow->dead;
1666 emc_clear_entry(struct emc_entry *ce)
1669 dp_netdev_flow_unref(ce->flow);
1675 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1676 const struct netdev_flow_key *key)
1678 if (ce->flow != flow) {
1680 dp_netdev_flow_unref(ce->flow);
1683 if (dp_netdev_flow_ref(flow)) {
1690 netdev_flow_key_clone(&ce->key, key);
1695 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1696 struct dp_netdev_flow *flow)
1698 struct emc_entry *to_be_replaced = NULL;
1699 struct emc_entry *current_entry;
1701 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1702 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1703 /* We found the entry with the 'mf' miniflow */
1704 emc_change_entry(current_entry, flow, NULL);
1708 /* Replacement policy: put the flow in an empty (not alive) entry, or
1709 * in the first entry where it can be */
1711 || (emc_entry_alive(to_be_replaced)
1712 && !emc_entry_alive(current_entry))
1713 || current_entry->key.hash < to_be_replaced->key.hash) {
1714 to_be_replaced = current_entry;
1717 /* We didn't find the miniflow in the cache.
1718 * The 'to_be_replaced' entry is where the new flow will be stored */
1720 emc_change_entry(to_be_replaced, flow, key);
1723 static inline struct dp_netdev_flow *
1724 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1726 struct emc_entry *current_entry;
1728 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1729 if (current_entry->key.hash == key->hash
1730 && emc_entry_alive(current_entry)
1731 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1733 /* We found the entry with the 'key->mf' miniflow */
1734 return current_entry->flow;
1741 static struct dp_netdev_flow *
1742 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1743 const struct netdev_flow_key *key)
1745 struct dp_netdev_flow *netdev_flow;
1746 struct dpcls_rule *rule;
1748 dpcls_lookup(&pmd->cls, key, &rule, 1);
1749 netdev_flow = dp_netdev_flow_cast(rule);
1754 static struct dp_netdev_flow *
1755 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1756 const ovs_u128 *ufidp, const struct nlattr *key,
1759 struct dp_netdev_flow *netdev_flow;
1763 /* If a UFID is not provided, determine one based on the key. */
1764 if (!ufidp && key && key_len
1765 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1766 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1771 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1773 if (ovs_u128_equals(&netdev_flow->ufid, ufidp)) {
1783 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1784 struct dpif_flow_stats *stats)
1786 struct dp_netdev_flow *netdev_flow;
1787 unsigned long long n;
1791 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1793 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1794 stats->n_packets = n;
1795 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1797 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1799 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1800 stats->tcp_flags = flags;
1803 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1804 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1805 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1808 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1809 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1810 struct dpif_flow *flow, bool terse)
1813 memset(flow, 0, sizeof *flow);
1815 struct flow_wildcards wc;
1816 struct dp_netdev_actions *actions;
1818 struct odp_flow_key_parms odp_parms = {
1819 .flow = &netdev_flow->flow,
1822 .max_mpls_depth = SIZE_MAX,
1825 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1828 offset = key_buf->size;
1829 flow->key = ofpbuf_tail(key_buf);
1830 odp_parms.odp_in_port = netdev_flow->flow.in_port.odp_port;
1831 odp_flow_key_from_flow(&odp_parms, key_buf);
1832 flow->key_len = key_buf->size - offset;
1835 offset = mask_buf->size;
1836 flow->mask = ofpbuf_tail(mask_buf);
1837 odp_parms.odp_in_port = wc.masks.in_port.odp_port;
1838 odp_flow_key_from_mask(&odp_parms, mask_buf);
1839 flow->mask_len = mask_buf->size - offset;
1842 actions = dp_netdev_flow_get_actions(netdev_flow);
1843 flow->actions = actions->actions;
1844 flow->actions_len = actions->size;
1847 flow->ufid = netdev_flow->ufid;
1848 flow->ufid_present = true;
1849 flow->pmd_id = netdev_flow->pmd_id;
1850 get_dpif_flow_stats(netdev_flow, &flow->stats);
1854 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1855 const struct nlattr *mask_key,
1856 uint32_t mask_key_len, const struct flow *flow,
1860 enum odp_key_fitness fitness;
1862 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1864 /* This should not happen: it indicates that
1865 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1866 * disagree on the acceptable form of a mask. Log the problem
1867 * as an error, with enough details to enable debugging. */
1868 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1870 if (!VLOG_DROP_ERR(&rl)) {
1874 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1876 VLOG_ERR("internal error parsing flow mask %s (%s)",
1877 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1884 enum mf_field_id id;
1885 /* No mask key, unwildcard everything except fields whose
1886 * prerequisities are not met. */
1887 memset(mask, 0x0, sizeof *mask);
1889 for (id = 0; id < MFF_N_IDS; ++id) {
1890 /* Skip registers and metadata. */
1891 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1892 && id != MFF_METADATA) {
1893 const struct mf_field *mf = mf_from_id(id);
1894 if (mf_are_prereqs_ok(mf, flow)) {
1895 mf_mask_field(mf, mask);
1901 /* Force unwildcard the in_port.
1903 * We need to do this even in the case where we unwildcard "everything"
1904 * above because "everything" only includes the 16-bit OpenFlow port number
1905 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1906 * port number mask->in_port.odp_port. */
1907 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1913 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1918 if (odp_flow_key_to_flow(key, key_len, flow)) {
1919 /* This should not happen: it indicates that odp_flow_key_from_flow()
1920 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1921 * flow. Log the problem as an error, with enough details to enable
1923 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1925 if (!VLOG_DROP_ERR(&rl)) {
1929 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1930 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1937 in_port = flow->in_port.odp_port;
1938 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1946 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1948 struct dp_netdev *dp = get_dp_netdev(dpif);
1949 struct dp_netdev_flow *netdev_flow;
1950 struct dp_netdev_pmd_thread *pmd;
1951 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1952 ? NON_PMD_CORE_ID : get->pmd_id;
1955 pmd = dp_netdev_get_pmd(dp, pmd_id);
1960 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1963 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1968 dp_netdev_pmd_unref(pmd);
1974 static struct dp_netdev_flow *
1975 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
1976 struct match *match, const ovs_u128 *ufid,
1977 const struct nlattr *actions, size_t actions_len)
1978 OVS_REQUIRES(pmd->flow_mutex)
1980 struct dp_netdev_flow *flow;
1981 struct netdev_flow_key mask;
1983 netdev_flow_mask_init(&mask, match);
1984 /* Make sure wc does not have metadata. */
1985 ovs_assert(!(mask.mf.map & (MINIFLOW_MAP(metadata) | MINIFLOW_MAP(regs))));
1987 /* Do not allocate extra space. */
1988 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
1989 memset(&flow->stats, 0, sizeof flow->stats);
1992 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
1993 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
1994 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
1995 ovs_refcount_init(&flow->ref_cnt);
1996 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
1998 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
1999 dpcls_insert(&pmd->cls, &flow->cr, &mask);
2001 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
2002 dp_netdev_flow_hash(&flow->ufid));
2004 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2006 struct ds ds = DS_EMPTY_INITIALIZER;
2008 match.flow = flow->flow;
2009 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2011 ds_put_cstr(&ds, "flow_add: ");
2012 odp_format_ufid(ufid, &ds);
2013 ds_put_cstr(&ds, " ");
2014 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2015 ds_put_cstr(&ds, ", actions:");
2016 format_odp_actions(&ds, actions, actions_len);
2018 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2027 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2029 struct dp_netdev *dp = get_dp_netdev(dpif);
2030 struct dp_netdev_flow *netdev_flow;
2031 struct netdev_flow_key key;
2032 struct dp_netdev_pmd_thread *pmd;
2035 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2036 ? NON_PMD_CORE_ID : put->pmd_id;
2039 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2043 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2044 put->mask, put->mask_len,
2045 &match.flow, &match.wc.masks);
2050 pmd = dp_netdev_get_pmd(dp, pmd_id);
2055 /* Must produce a netdev_flow_key for lookup.
2056 * This interface is no longer performance critical, since it is not used
2057 * for upcall processing any more. */
2058 netdev_flow_key_from_flow(&key, &match.flow);
2063 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2066 ovs_mutex_lock(&pmd->flow_mutex);
2067 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2069 if (put->flags & DPIF_FP_CREATE) {
2070 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2072 memset(put->stats, 0, sizeof *put->stats);
2074 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2084 if (put->flags & DPIF_FP_MODIFY
2085 && flow_equal(&match.flow, &netdev_flow->flow)) {
2086 struct dp_netdev_actions *new_actions;
2087 struct dp_netdev_actions *old_actions;
2089 new_actions = dp_netdev_actions_create(put->actions,
2092 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2093 ovsrcu_set(&netdev_flow->actions, new_actions);
2096 get_dpif_flow_stats(netdev_flow, put->stats);
2098 if (put->flags & DPIF_FP_ZERO_STATS) {
2099 /* XXX: The userspace datapath uses thread local statistics
2100 * (for flows), which should be updated only by the owning
2101 * thread. Since we cannot write on stats memory here,
2102 * we choose not to support this flag. Please note:
2103 * - This feature is currently used only by dpctl commands with
2105 * - Should the need arise, this operation can be implemented
2106 * by keeping a base value (to be update here) for each
2107 * counter, and subtracting it before outputting the stats */
2111 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2112 } else if (put->flags & DPIF_FP_CREATE) {
2115 /* Overlapping flow. */
2119 ovs_mutex_unlock(&pmd->flow_mutex);
2120 dp_netdev_pmd_unref(pmd);
2126 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2128 struct dp_netdev *dp = get_dp_netdev(dpif);
2129 struct dp_netdev_flow *netdev_flow;
2130 struct dp_netdev_pmd_thread *pmd;
2131 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2132 ? NON_PMD_CORE_ID : del->pmd_id;
2135 pmd = dp_netdev_get_pmd(dp, pmd_id);
2140 ovs_mutex_lock(&pmd->flow_mutex);
2141 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2145 get_dpif_flow_stats(netdev_flow, del->stats);
2147 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2151 ovs_mutex_unlock(&pmd->flow_mutex);
2152 dp_netdev_pmd_unref(pmd);
2157 struct dpif_netdev_flow_dump {
2158 struct dpif_flow_dump up;
2159 struct cmap_position poll_thread_pos;
2160 struct cmap_position flow_pos;
2161 struct dp_netdev_pmd_thread *cur_pmd;
2163 struct ovs_mutex mutex;
2166 static struct dpif_netdev_flow_dump *
2167 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2169 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2172 static struct dpif_flow_dump *
2173 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2175 struct dpif_netdev_flow_dump *dump;
2177 dump = xzalloc(sizeof *dump);
2178 dpif_flow_dump_init(&dump->up, dpif_);
2179 dump->up.terse = terse;
2180 ovs_mutex_init(&dump->mutex);
2186 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2188 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2190 ovs_mutex_destroy(&dump->mutex);
2195 struct dpif_netdev_flow_dump_thread {
2196 struct dpif_flow_dump_thread up;
2197 struct dpif_netdev_flow_dump *dump;
2198 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2199 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2202 static struct dpif_netdev_flow_dump_thread *
2203 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2205 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2208 static struct dpif_flow_dump_thread *
2209 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2211 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2212 struct dpif_netdev_flow_dump_thread *thread;
2214 thread = xmalloc(sizeof *thread);
2215 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2216 thread->dump = dump;
2221 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2223 struct dpif_netdev_flow_dump_thread *thread
2224 = dpif_netdev_flow_dump_thread_cast(thread_);
2230 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2231 struct dpif_flow *flows, int max_flows)
2233 struct dpif_netdev_flow_dump_thread *thread
2234 = dpif_netdev_flow_dump_thread_cast(thread_);
2235 struct dpif_netdev_flow_dump *dump = thread->dump;
2236 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2240 ovs_mutex_lock(&dump->mutex);
2241 if (!dump->status) {
2242 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2243 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2244 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2245 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2247 /* First call to dump_next(), extracts the first pmd thread.
2248 * If there is no pmd thread, returns immediately. */
2250 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2252 ovs_mutex_unlock(&dump->mutex);
2259 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2260 struct cmap_node *node;
2262 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2266 netdev_flows[n_flows] = CONTAINER_OF(node,
2267 struct dp_netdev_flow,
2270 /* When finishing dumping the current pmd thread, moves to
2272 if (n_flows < flow_limit) {
2273 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2274 dp_netdev_pmd_unref(pmd);
2275 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2281 /* Keeps the reference to next caller. */
2282 dump->cur_pmd = pmd;
2284 /* If the current dump is empty, do not exit the loop, since the
2285 * remaining pmds could have flows to be dumped. Just dumps again
2286 * on the new 'pmd'. */
2289 ovs_mutex_unlock(&dump->mutex);
2291 for (i = 0; i < n_flows; i++) {
2292 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2293 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2294 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2295 struct dpif_flow *f = &flows[i];
2296 struct ofpbuf key, mask;
2298 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2299 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2300 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2308 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2309 OVS_NO_THREAD_SAFETY_ANALYSIS
2311 struct dp_netdev *dp = get_dp_netdev(dpif);
2312 struct dp_netdev_pmd_thread *pmd;
2313 struct dp_packet *pp;
2315 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2316 dp_packet_size(execute->packet) > UINT16_MAX) {
2320 /* Tries finding the 'pmd'. If NULL is returned, that means
2321 * the current thread is a non-pmd thread and should use
2322 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2323 pmd = ovsthread_getspecific(dp->per_pmd_key);
2325 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2328 /* If the current thread is non-pmd thread, acquires
2329 * the 'non_pmd_mutex'. */
2330 if (pmd->core_id == NON_PMD_CORE_ID) {
2331 ovs_mutex_lock(&dp->non_pmd_mutex);
2332 ovs_mutex_lock(&dp->port_mutex);
2335 pp = execute->packet;
2336 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2337 execute->actions_len);
2338 if (pmd->core_id == NON_PMD_CORE_ID) {
2339 dp_netdev_pmd_unref(pmd);
2340 ovs_mutex_unlock(&dp->port_mutex);
2341 ovs_mutex_unlock(&dp->non_pmd_mutex);
2348 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2352 for (i = 0; i < n_ops; i++) {
2353 struct dpif_op *op = ops[i];
2356 case DPIF_OP_FLOW_PUT:
2357 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2360 case DPIF_OP_FLOW_DEL:
2361 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2364 case DPIF_OP_EXECUTE:
2365 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2368 case DPIF_OP_FLOW_GET:
2369 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2375 /* Returns true if the configuration for rx queues or cpu mask
2378 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2380 if (dp->n_dpdk_rxqs != rxqs) {
2383 if (dp->pmd_cmask != NULL && cmask != NULL) {
2384 return strcmp(dp->pmd_cmask, cmask);
2386 return (dp->pmd_cmask != NULL || cmask != NULL);
2391 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2393 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2395 struct dp_netdev *dp = get_dp_netdev(dpif);
2397 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2398 struct dp_netdev_port *port;
2400 dp_netdev_destroy_all_pmds(dp);
2402 CMAP_FOR_EACH (port, node, &dp->ports) {
2403 if (netdev_is_pmd(port->netdev)) {
2406 /* Closes the existing 'rxq's. */
2407 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2408 netdev_rxq_close(port->rxq[i]);
2409 port->rxq[i] = NULL;
2412 /* Sets the new rx queue config. */
2413 err = netdev_set_multiq(port->netdev,
2414 ovs_numa_get_n_cores() + 1,
2416 if (err && (err != EOPNOTSUPP)) {
2417 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2418 " %u", netdev_get_name(port->netdev),
2423 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2424 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2425 * netdev_n_rxq(port->netdev));
2426 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2427 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2431 dp->n_dpdk_rxqs = n_rxqs;
2433 /* Reconfigures the cpu mask. */
2434 ovs_numa_set_cpu_mask(cmask);
2435 free(dp->pmd_cmask);
2436 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2438 /* Restores the non-pmd. */
2439 dp_netdev_set_nonpmd(dp);
2440 /* Restores all pmd threads. */
2441 dp_netdev_reset_pmd_threads(dp);
2448 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2449 uint32_t queue_id, uint32_t *priority)
2451 *priority = queue_id;
2456 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2457 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2458 struct dp_netdev_actions *
2459 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2461 struct dp_netdev_actions *netdev_actions;
2463 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2464 memcpy(netdev_actions->actions, actions, size);
2465 netdev_actions->size = size;
2467 return netdev_actions;
2470 struct dp_netdev_actions *
2471 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2473 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2477 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2482 static inline unsigned long long
2483 cycles_counter(void)
2486 return rte_get_tsc_cycles();
2492 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2493 extern struct ovs_mutex cycles_counter_fake_mutex;
2495 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2497 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2498 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2499 OVS_NO_THREAD_SAFETY_ANALYSIS
2501 pmd->last_cycles = cycles_counter();
2504 /* Stop counting cycles and add them to the counter 'type' */
2506 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2507 enum pmd_cycles_counter_type type)
2508 OVS_RELEASES(&cycles_counter_fake_mutex)
2509 OVS_NO_THREAD_SAFETY_ANALYSIS
2511 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2513 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2517 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2518 struct dp_netdev_port *port,
2519 struct netdev_rxq *rxq)
2521 struct dp_packet *packets[NETDEV_MAX_BURST];
2524 cycles_count_start(pmd);
2525 error = netdev_rxq_recv(rxq, packets, &cnt);
2526 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2530 *recirc_depth_get() = 0;
2532 /* XXX: initialize md in netdev implementation. */
2533 for (i = 0; i < cnt; i++) {
2534 packets[i]->md = port->md;
2536 cycles_count_start(pmd);
2537 dp_netdev_input(pmd, packets, cnt);
2538 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2539 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2540 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2542 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2543 netdev_get_name(port->netdev), ovs_strerror(error));
2547 /* Return true if needs to revalidate datapath flows. */
2549 dpif_netdev_run(struct dpif *dpif)
2551 struct dp_netdev_port *port;
2552 struct dp_netdev *dp = get_dp_netdev(dpif);
2553 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2555 uint64_t new_tnl_seq;
2557 ovs_mutex_lock(&dp->non_pmd_mutex);
2558 CMAP_FOR_EACH (port, node, &dp->ports) {
2559 if (!netdev_is_pmd(port->netdev)) {
2562 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2563 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2567 ovs_mutex_unlock(&dp->non_pmd_mutex);
2568 dp_netdev_pmd_unref(non_pmd);
2570 tnl_arp_cache_run();
2571 new_tnl_seq = seq_read(tnl_conf_seq);
2573 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2574 dp->last_tnl_conf_seq = new_tnl_seq;
2581 dpif_netdev_wait(struct dpif *dpif)
2583 struct dp_netdev_port *port;
2584 struct dp_netdev *dp = get_dp_netdev(dpif);
2586 ovs_mutex_lock(&dp_netdev_mutex);
2587 CMAP_FOR_EACH (port, node, &dp->ports) {
2588 if (!netdev_is_pmd(port->netdev)) {
2591 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2592 netdev_rxq_wait(port->rxq[i]);
2596 ovs_mutex_unlock(&dp_netdev_mutex);
2597 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2601 struct dp_netdev_port *port;
2602 struct netdev_rxq *rx;
2606 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2607 struct rxq_poll **ppoll_list, int poll_cnt)
2609 struct rxq_poll *poll_list = *ppoll_list;
2610 struct dp_netdev_port *port;
2611 int n_pmds_on_numa, index, i;
2613 /* Simple scheduler for netdev rx polling. */
2614 for (i = 0; i < poll_cnt; i++) {
2615 port_unref(poll_list[i].port);
2619 n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
2622 CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
2623 /* Calls port_try_ref() to prevent the main thread
2624 * from deleting the port. */
2625 if (port_try_ref(port)) {
2626 if (netdev_is_pmd(port->netdev)
2627 && netdev_get_numa_id(port->netdev) == pmd->numa_id) {
2630 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2631 if ((index % n_pmds_on_numa) == pmd->index) {
2632 poll_list = xrealloc(poll_list,
2633 sizeof *poll_list * (poll_cnt + 1));
2636 poll_list[poll_cnt].port = port;
2637 poll_list[poll_cnt].rx = port->rxq[i];
2643 /* Unrefs the port_try_ref(). */
2648 *ppoll_list = poll_list;
2653 pmd_thread_main(void *f_)
2655 struct dp_netdev_pmd_thread *pmd = f_;
2656 unsigned int lc = 0;
2657 struct rxq_poll *poll_list;
2658 unsigned int port_seq = PMD_INITIAL_SEQ;
2665 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2666 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2667 pmd_thread_setaffinity_cpu(pmd->core_id);
2669 emc_cache_init(&pmd->flow_cache);
2670 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2672 /* Signal here to make sure the pmd finishes
2673 * reloading the updated configuration. */
2674 dp_netdev_pmd_reload_done(pmd);
2679 for (i = 0; i < poll_cnt; i++) {
2680 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2688 emc_cache_slow_sweep(&pmd->flow_cache);
2691 atomic_read_relaxed(&pmd->change_seq, &seq);
2692 if (seq != port_seq) {
2699 emc_cache_uninit(&pmd->flow_cache);
2701 if (!latch_is_set(&pmd->exit_latch)){
2705 for (i = 0; i < poll_cnt; i++) {
2706 port_unref(poll_list[i].port);
2709 dp_netdev_pmd_reload_done(pmd);
2716 dp_netdev_disable_upcall(struct dp_netdev *dp)
2717 OVS_ACQUIRES(dp->upcall_rwlock)
2719 fat_rwlock_wrlock(&dp->upcall_rwlock);
2723 dpif_netdev_disable_upcall(struct dpif *dpif)
2724 OVS_NO_THREAD_SAFETY_ANALYSIS
2726 struct dp_netdev *dp = get_dp_netdev(dpif);
2727 dp_netdev_disable_upcall(dp);
2731 dp_netdev_enable_upcall(struct dp_netdev *dp)
2732 OVS_RELEASES(dp->upcall_rwlock)
2734 fat_rwlock_unlock(&dp->upcall_rwlock);
2738 dpif_netdev_enable_upcall(struct dpif *dpif)
2739 OVS_NO_THREAD_SAFETY_ANALYSIS
2741 struct dp_netdev *dp = get_dp_netdev(dpif);
2742 dp_netdev_enable_upcall(dp);
2746 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2748 ovs_mutex_lock(&pmd->cond_mutex);
2749 xpthread_cond_signal(&pmd->cond);
2750 ovs_mutex_unlock(&pmd->cond_mutex);
2753 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2754 * the pointer if succeeds, otherwise, NULL.
2756 * Caller must unrefs the returned reference. */
2757 static struct dp_netdev_pmd_thread *
2758 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2760 struct dp_netdev_pmd_thread *pmd;
2761 const struct cmap_node *pnode;
2763 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2767 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2769 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2772 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2774 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2776 struct dp_netdev_pmd_thread *non_pmd;
2778 non_pmd = xzalloc(sizeof *non_pmd);
2779 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2783 /* Caller must have valid pointer to 'pmd'. */
2785 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2787 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2791 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2793 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2794 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2798 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2799 * fails, keeps checking for next node until reaching the end of cmap.
2801 * Caller must unrefs the returned reference. */
2802 static struct dp_netdev_pmd_thread *
2803 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2805 struct dp_netdev_pmd_thread *next;
2808 struct cmap_node *node;
2810 node = cmap_next_position(&dp->poll_threads, pos);
2811 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2813 } while (next && !dp_netdev_pmd_try_ref(next));
2819 core_id_to_qid(unsigned core_id)
2821 if (core_id != NON_PMD_CORE_ID) {
2824 return ovs_numa_get_n_cores();
2828 /* Configures the 'pmd' based on the input argument. */
2830 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2831 int index, unsigned core_id, int numa_id)
2835 pmd->core_id = core_id;
2836 pmd->tx_qid = core_id_to_qid(core_id);
2837 pmd->numa_id = numa_id;
2839 ovs_refcount_init(&pmd->ref_cnt);
2840 latch_init(&pmd->exit_latch);
2841 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2842 xpthread_cond_init(&pmd->cond, NULL);
2843 ovs_mutex_init(&pmd->cond_mutex);
2844 ovs_mutex_init(&pmd->flow_mutex);
2845 dpcls_init(&pmd->cls);
2846 cmap_init(&pmd->flow_table);
2847 /* init the 'flow_cache' since there is no
2848 * actual thread created for NON_PMD_CORE_ID. */
2849 if (core_id == NON_PMD_CORE_ID) {
2850 emc_cache_init(&pmd->flow_cache);
2852 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2853 hash_int(core_id, 0));
2857 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2859 dp_netdev_pmd_flow_flush(pmd);
2860 dpcls_destroy(&pmd->cls);
2861 cmap_destroy(&pmd->flow_table);
2862 ovs_mutex_destroy(&pmd->flow_mutex);
2863 latch_destroy(&pmd->exit_latch);
2864 xpthread_cond_destroy(&pmd->cond);
2865 ovs_mutex_destroy(&pmd->cond_mutex);
2869 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2870 * and unrefs the struct. */
2872 dp_netdev_del_pmd(struct dp_netdev_pmd_thread *pmd)
2874 /* Uninit the 'flow_cache' since there is
2875 * no actual thread uninit it for NON_PMD_CORE_ID. */
2876 if (pmd->core_id == NON_PMD_CORE_ID) {
2877 emc_cache_uninit(&pmd->flow_cache);
2879 latch_set(&pmd->exit_latch);
2880 dp_netdev_reload_pmd__(pmd);
2881 ovs_numa_unpin_core(pmd->core_id);
2882 xpthread_join(pmd->thread, NULL);
2884 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2885 dp_netdev_pmd_unref(pmd);
2888 /* Destroys all pmd threads. */
2890 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2892 struct dp_netdev_pmd_thread *pmd;
2894 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2895 dp_netdev_del_pmd(pmd);
2899 /* Deletes all pmd threads on numa node 'numa_id'. */
2901 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2903 struct dp_netdev_pmd_thread *pmd;
2905 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2906 if (pmd->numa_id == numa_id) {
2907 dp_netdev_del_pmd(pmd);
2912 /* Checks the numa node id of 'netdev' and starts pmd threads for
2915 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2919 if (!ovs_numa_numa_id_is_valid(numa_id)) {
2920 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2921 "invalid", numa_id);
2925 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
2927 /* If there are already pmd threads created for the numa node
2928 * in which 'netdev' is on, do nothing. Else, creates the
2929 * pmd threads for the numa node. */
2931 int can_have, n_unpinned, i;
2933 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
2935 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2936 "cores on numa node");
2940 /* If cpu mask is specified, uses all unpinned cores, otherwise
2941 * tries creating NR_PMD_THREADS pmd threads. */
2942 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
2943 for (i = 0; i < can_have; i++) {
2944 struct dp_netdev_pmd_thread *pmd = xzalloc(sizeof *pmd);
2945 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
2947 dp_netdev_configure_pmd(pmd, dp, i, core_id, numa_id);
2948 /* Each thread will distribute all devices rx-queues among
2950 pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd);
2952 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
2957 /* Called after pmd threads config change. Restarts pmd threads with
2958 * new configuration. */
2960 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
2962 struct dp_netdev_port *port;
2964 CMAP_FOR_EACH (port, node, &dp->ports) {
2965 if (netdev_is_pmd(port->netdev)) {
2966 int numa_id = netdev_get_numa_id(port->netdev);
2968 dp_netdev_set_pmds_on_numa(dp, numa_id);
2974 dpif_netdev_get_datapath_version(void)
2976 return xstrdup("<built-in>");
2980 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
2981 uint16_t tcp_flags, long long now)
2985 atomic_store_relaxed(&netdev_flow->stats.used, now);
2986 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
2987 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
2988 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
2990 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
2994 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
2995 enum dp_stat_type type, int cnt)
2997 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3001 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3002 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3003 enum dpif_upcall_type type, const struct nlattr *userdata,
3004 struct ofpbuf *actions, struct ofpbuf *put_actions)
3006 struct dp_netdev *dp = pmd->dp;
3008 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3012 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3013 struct ds ds = DS_EMPTY_INITIALIZER;
3016 struct odp_flow_key_parms odp_parms = {
3019 .odp_in_port = flow->in_port.odp_port,
3023 ofpbuf_init(&key, 0);
3024 odp_flow_key_from_flow(&odp_parms, &key);
3025 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3026 dp_packet_size(packet_));
3028 odp_flow_key_format(key.data, key.size, &ds);
3030 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3031 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3033 ofpbuf_uninit(&key);
3039 return dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3040 actions, wc, put_actions, dp->upcall_aux);
3043 static inline uint32_t
3044 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3045 const struct miniflow *mf)
3047 uint32_t hash, recirc_depth;
3049 hash = dp_packet_get_rss_hash(packet);
3050 if (OVS_UNLIKELY(!hash)) {
3051 hash = miniflow_hash_5tuple(mf, 0);
3052 dp_packet_set_rss_hash(packet, hash);
3055 /* The RSS hash must account for the recirculation depth to avoid
3056 * collisions in the exact match cache */
3057 recirc_depth = *recirc_depth_get_unsafe();
3058 if (OVS_UNLIKELY(recirc_depth)) {
3059 hash = hash_finish(hash, recirc_depth);
3060 dp_packet_set_rss_hash(packet, hash);
3065 struct packet_batch {
3066 unsigned int packet_count;
3067 unsigned int byte_count;
3070 struct dp_netdev_flow *flow;
3072 struct dp_packet *packets[NETDEV_MAX_BURST];
3076 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
3077 const struct miniflow *mf)
3079 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3080 batch->packets[batch->packet_count++] = packet;
3081 batch->byte_count += dp_packet_size(packet);
3085 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
3087 flow->batch = batch;
3090 batch->packet_count = 0;
3091 batch->byte_count = 0;
3092 batch->tcp_flags = 0;
3096 packet_batch_execute(struct packet_batch *batch,
3097 struct dp_netdev_pmd_thread *pmd,
3100 struct dp_netdev_actions *actions;
3101 struct dp_netdev_flow *flow = batch->flow;
3103 dp_netdev_flow_used(flow, batch->packet_count, batch->byte_count,
3104 batch->tcp_flags, now);
3106 actions = dp_netdev_flow_get_actions(flow);
3108 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
3109 actions->actions, actions->size);
3113 dp_netdev_queue_batches(struct dp_packet *pkt,
3114 struct dp_netdev_flow *flow, const struct miniflow *mf,
3115 struct packet_batch *batches, size_t *n_batches)
3117 struct packet_batch *batch = flow->batch;
3119 if (OVS_LIKELY(batch)) {
3120 packet_batch_update(batch, pkt, mf);
3124 batch = &batches[(*n_batches)++];
3125 packet_batch_init(batch, flow);
3126 packet_batch_update(batch, pkt, mf);
3130 dp_packet_swap(struct dp_packet **a, struct dp_packet **b)
3132 struct dp_packet *tmp = *a;
3137 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3138 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', the
3139 * miniflow is copied into 'keys' and the packet pointer is moved at the
3140 * beginning of the 'packets' array.
3142 * The function returns the number of packets that needs to be processed in the
3143 * 'packets' array (they have been moved to the beginning of the vector).
3145 static inline size_t
3146 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
3147 size_t cnt, struct netdev_flow_key *keys,
3148 struct packet_batch batches[], size_t *n_batches)
3150 struct emc_cache *flow_cache = &pmd->flow_cache;
3151 struct netdev_flow_key key;
3152 size_t i, notfound_cnt = 0;
3154 miniflow_initialize(&key.mf, key.buf);
3155 for (i = 0; i < cnt; i++) {
3156 struct dp_netdev_flow *flow;
3158 if (OVS_UNLIKELY(dp_packet_size(packets[i]) < ETH_HEADER_LEN)) {
3159 dp_packet_delete(packets[i]);
3164 /* Prefetch next packet data */
3165 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3168 miniflow_extract(packets[i], &key.mf);
3169 key.len = 0; /* Not computed yet. */
3170 key.hash = dpif_netdev_packet_get_rss_hash(packets[i], &key.mf);
3172 flow = emc_lookup(flow_cache, &key);
3173 if (OVS_LIKELY(flow)) {
3174 dp_netdev_queue_batches(packets[i], flow, &key.mf, batches,
3177 if (i != notfound_cnt) {
3178 dp_packet_swap(&packets[i], &packets[notfound_cnt]);
3181 keys[notfound_cnt++] = key;
3185 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - notfound_cnt);
3187 return notfound_cnt;
3191 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3192 struct dp_packet **packets, size_t cnt,
3193 struct netdev_flow_key *keys,
3194 struct packet_batch batches[], size_t *n_batches)
3196 #if !defined(__CHECKER__) && !defined(_WIN32)
3197 const size_t PKT_ARRAY_SIZE = cnt;
3199 /* Sparse or MSVC doesn't like variable length array. */
3200 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3202 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3203 struct dp_netdev *dp = pmd->dp;
3204 struct emc_cache *flow_cache = &pmd->flow_cache;
3205 int miss_cnt = 0, lost_cnt = 0;
3209 for (i = 0; i < cnt; i++) {
3210 /* Key length is needed in all the cases, hash computed on demand. */
3211 keys[i].len = netdev_flow_key_size(count_1bits(keys[i].mf.map));
3213 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3214 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3215 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3216 struct ofpbuf actions, put_actions;
3219 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3220 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3222 for (i = 0; i < cnt; i++) {
3223 struct dp_netdev_flow *netdev_flow;
3224 struct ofpbuf *add_actions;
3228 if (OVS_LIKELY(rules[i])) {
3232 /* It's possible that an earlier slow path execution installed
3233 * a rule covering this flow. In this case, it's a lot cheaper
3234 * to catch it here than execute a miss. */
3235 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3237 rules[i] = &netdev_flow->cr;
3243 miniflow_expand(&keys[i].mf, &match.flow);
3245 ofpbuf_clear(&actions);
3246 ofpbuf_clear(&put_actions);
3248 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
3249 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
3250 &ufid, DPIF_UC_MISS, NULL, &actions,
3252 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3253 dp_packet_delete(packets[i]);
3258 /* We can't allow the packet batching in the next loop to execute
3259 * the actions. Otherwise, if there are any slow path actions,
3260 * we'll send the packet up twice. */
3261 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
3262 actions.data, actions.size);
3264 add_actions = put_actions.size ? &put_actions : &actions;
3265 if (OVS_LIKELY(error != ENOSPC)) {
3266 /* XXX: There's a race window where a flow covering this packet
3267 * could have already been installed since we last did the flow
3268 * lookup before upcall. This could be solved by moving the
3269 * mutex lock outside the loop, but that's an awful long time
3270 * to be locking everyone out of making flow installs. If we
3271 * move to a per-core classifier, it would be reasonable. */
3272 ovs_mutex_lock(&pmd->flow_mutex);
3273 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3274 if (OVS_LIKELY(!netdev_flow)) {
3275 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3279 ovs_mutex_unlock(&pmd->flow_mutex);
3281 emc_insert(flow_cache, &keys[i], netdev_flow);
3285 ofpbuf_uninit(&actions);
3286 ofpbuf_uninit(&put_actions);
3287 fat_rwlock_unlock(&dp->upcall_rwlock);
3288 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3289 } else if (OVS_UNLIKELY(any_miss)) {
3290 for (i = 0; i < cnt; i++) {
3291 if (OVS_UNLIKELY(!rules[i])) {
3292 dp_packet_delete(packets[i]);
3299 for (i = 0; i < cnt; i++) {
3300 struct dp_packet *packet = packets[i];
3301 struct dp_netdev_flow *flow;
3303 if (OVS_UNLIKELY(!rules[i])) {
3307 flow = dp_netdev_flow_cast(rules[i]);
3309 emc_insert(flow_cache, &keys[i], flow);
3310 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3313 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3314 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3315 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3319 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3320 struct dp_packet **packets, int cnt)
3322 #if !defined(__CHECKER__) && !defined(_WIN32)
3323 const size_t PKT_ARRAY_SIZE = cnt;
3325 /* Sparse or MSVC doesn't like variable length array. */
3326 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3328 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3329 struct packet_batch batches[PKT_ARRAY_SIZE];
3330 long long now = time_msec();
3331 size_t newcnt, n_batches, i;
3334 newcnt = emc_processing(pmd, packets, cnt, keys, batches, &n_batches);
3335 if (OVS_UNLIKELY(newcnt)) {
3336 fast_path_processing(pmd, packets, newcnt, keys, batches, &n_batches);
3339 for (i = 0; i < n_batches; i++) {
3340 batches[i].flow->batch = NULL;
3343 for (i = 0; i < n_batches; i++) {
3344 packet_batch_execute(&batches[i], pmd, now);
3348 struct dp_netdev_execute_aux {
3349 struct dp_netdev_pmd_thread *pmd;
3353 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3356 struct dp_netdev *dp = get_dp_netdev(dpif);
3357 dp->upcall_aux = aux;
3362 dp_netdev_drop_packets(struct dp_packet **packets, int cnt, bool may_steal)
3367 for (i = 0; i < cnt; i++) {
3368 dp_packet_delete(packets[i]);
3374 push_tnl_action(const struct dp_netdev *dp,
3375 const struct nlattr *attr,
3376 struct dp_packet **packets, int cnt)
3378 struct dp_netdev_port *tun_port;
3379 const struct ovs_action_push_tnl *data;
3381 data = nl_attr_get(attr);
3383 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3387 netdev_push_header(tun_port->netdev, packets, cnt, data);
3393 dp_netdev_clone_pkt_batch(struct dp_packet **dst_pkts,
3394 struct dp_packet **src_pkts, int cnt)
3398 for (i = 0; i < cnt; i++) {
3399 dst_pkts[i] = dp_packet_clone(src_pkts[i]);
3404 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3405 const struct nlattr *a, bool may_steal)
3406 OVS_NO_THREAD_SAFETY_ANALYSIS
3408 struct dp_netdev_execute_aux *aux = aux_;
3409 uint32_t *depth = recirc_depth_get();
3410 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3411 struct dp_netdev *dp = pmd->dp;
3412 int type = nl_attr_type(a);
3413 struct dp_netdev_port *p;
3416 switch ((enum ovs_action_attr)type) {
3417 case OVS_ACTION_ATTR_OUTPUT:
3418 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3419 if (OVS_LIKELY(p)) {
3420 netdev_send(p->netdev, pmd->tx_qid, packets, cnt, may_steal);
3425 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3426 if (*depth < MAX_RECIRC_DEPTH) {
3427 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3431 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3435 err = push_tnl_action(dp, a, packets, cnt);
3438 dp_netdev_input(pmd, packets, cnt);
3441 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3447 case OVS_ACTION_ATTR_TUNNEL_POP:
3448 if (*depth < MAX_RECIRC_DEPTH) {
3449 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3451 p = dp_netdev_lookup_port(dp, portno);
3453 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3457 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3461 err = netdev_pop_header(p->netdev, packets, cnt);
3464 for (i = 0; i < cnt; i++) {
3465 packets[i]->md.in_port.odp_port = portno;
3469 dp_netdev_input(pmd, packets, cnt);
3472 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3479 case OVS_ACTION_ATTR_USERSPACE:
3480 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3481 const struct nlattr *userdata;
3482 struct ofpbuf actions;
3486 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3487 ofpbuf_init(&actions, 0);
3489 for (i = 0; i < cnt; i++) {
3492 ofpbuf_clear(&actions);
3494 flow_extract(packets[i], &flow);
3495 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3496 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3497 DPIF_UC_ACTION, userdata,&actions,
3499 if (!error || error == ENOSPC) {
3500 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3501 actions.data, actions.size);
3502 } else if (may_steal) {
3503 dp_packet_delete(packets[i]);
3506 ofpbuf_uninit(&actions);
3507 fat_rwlock_unlock(&dp->upcall_rwlock);
3513 case OVS_ACTION_ATTR_RECIRC:
3514 if (*depth < MAX_RECIRC_DEPTH) {
3515 struct dp_packet *recirc_pkts[NETDEV_MAX_BURST];
3518 dp_netdev_clone_pkt_batch(recirc_pkts, packets, cnt);
3519 packets = recirc_pkts;
3522 for (i = 0; i < cnt; i++) {
3523 packets[i]->md.recirc_id = nl_attr_get_u32(a);
3527 dp_netdev_input(pmd, packets, cnt);
3533 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3536 case OVS_ACTION_ATTR_PUSH_VLAN:
3537 case OVS_ACTION_ATTR_POP_VLAN:
3538 case OVS_ACTION_ATTR_PUSH_MPLS:
3539 case OVS_ACTION_ATTR_POP_MPLS:
3540 case OVS_ACTION_ATTR_SET:
3541 case OVS_ACTION_ATTR_SET_MASKED:
3542 case OVS_ACTION_ATTR_SAMPLE:
3543 case OVS_ACTION_ATTR_HASH:
3544 case OVS_ACTION_ATTR_UNSPEC:
3545 case __OVS_ACTION_ATTR_MAX:
3549 dp_netdev_drop_packets(packets, cnt, may_steal);
3553 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3554 struct dp_packet **packets, int cnt,
3556 const struct nlattr *actions, size_t actions_len)
3558 struct dp_netdev_execute_aux aux = { pmd };
3560 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3561 actions_len, dp_execute_cb);
3564 const struct dpif_class dpif_netdev_class = {
3567 dpif_netdev_enumerate,
3568 dpif_netdev_port_open_type,
3571 dpif_netdev_destroy,
3574 dpif_netdev_get_stats,
3575 dpif_netdev_port_add,
3576 dpif_netdev_port_del,
3577 dpif_netdev_port_query_by_number,
3578 dpif_netdev_port_query_by_name,
3579 NULL, /* port_get_pid */
3580 dpif_netdev_port_dump_start,
3581 dpif_netdev_port_dump_next,
3582 dpif_netdev_port_dump_done,
3583 dpif_netdev_port_poll,
3584 dpif_netdev_port_poll_wait,
3585 dpif_netdev_flow_flush,
3586 dpif_netdev_flow_dump_create,
3587 dpif_netdev_flow_dump_destroy,
3588 dpif_netdev_flow_dump_thread_create,
3589 dpif_netdev_flow_dump_thread_destroy,
3590 dpif_netdev_flow_dump_next,
3591 dpif_netdev_operate,
3592 NULL, /* recv_set */
3593 NULL, /* handlers_set */
3594 dpif_netdev_pmd_set,
3595 dpif_netdev_queue_to_priority,
3597 NULL, /* recv_wait */
3598 NULL, /* recv_purge */
3599 dpif_netdev_register_upcall_cb,
3600 dpif_netdev_enable_upcall,
3601 dpif_netdev_disable_upcall,
3602 dpif_netdev_get_datapath_version,
3606 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3607 const char *argv[], void *aux OVS_UNUSED)
3609 struct dp_netdev_port *old_port;
3610 struct dp_netdev_port *new_port;
3611 struct dp_netdev *dp;
3614 ovs_mutex_lock(&dp_netdev_mutex);
3615 dp = shash_find_data(&dp_netdevs, argv[1]);
3616 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3617 ovs_mutex_unlock(&dp_netdev_mutex);
3618 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3621 ovs_refcount_ref(&dp->ref_cnt);
3622 ovs_mutex_unlock(&dp_netdev_mutex);
3624 ovs_mutex_lock(&dp->port_mutex);
3625 if (get_port_by_name(dp, argv[2], &old_port)) {
3626 unixctl_command_reply_error(conn, "unknown port");
3630 port_no = u32_to_odp(atoi(argv[3]));
3631 if (!port_no || port_no == ODPP_NONE) {
3632 unixctl_command_reply_error(conn, "bad port number");
3635 if (dp_netdev_lookup_port(dp, port_no)) {
3636 unixctl_command_reply_error(conn, "port number already in use");
3640 /* Remove old port. */
3641 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->md.in_port.odp_port));
3642 ovsrcu_postpone(free, old_port);
3644 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3645 new_port = xmemdup(old_port, sizeof *old_port);
3646 new_port->md.in_port.odp_port = port_no;
3647 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3649 seq_change(dp->port_seq);
3650 unixctl_command_reply(conn, NULL);
3653 ovs_mutex_unlock(&dp->port_mutex);
3654 dp_netdev_unref(dp);
3658 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3659 const char *argv[], void *aux OVS_UNUSED)
3661 struct dp_netdev_port *port;
3662 struct dp_netdev *dp;
3664 ovs_mutex_lock(&dp_netdev_mutex);
3665 dp = shash_find_data(&dp_netdevs, argv[1]);
3666 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3667 ovs_mutex_unlock(&dp_netdev_mutex);
3668 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3671 ovs_refcount_ref(&dp->ref_cnt);
3672 ovs_mutex_unlock(&dp_netdev_mutex);
3674 ovs_mutex_lock(&dp->port_mutex);
3675 if (get_port_by_name(dp, argv[2], &port)) {
3676 unixctl_command_reply_error(conn, "unknown port");
3677 } else if (port->md.in_port.odp_port == ODPP_LOCAL) {
3678 unixctl_command_reply_error(conn, "can't delete local port");
3680 do_del_port(dp, port);
3681 unixctl_command_reply(conn, NULL);
3683 ovs_mutex_unlock(&dp->port_mutex);
3685 dp_netdev_unref(dp);
3689 dpif_dummy_register__(const char *type)
3691 struct dpif_class *class;
3693 class = xmalloc(sizeof *class);
3694 *class = dpif_netdev_class;
3695 class->type = xstrdup(type);
3696 dp_register_provider(class);
3700 dpif_dummy_register(bool override)
3707 dp_enumerate_types(&types);
3708 SSET_FOR_EACH (type, &types) {
3709 if (!dp_unregister_provider(type)) {
3710 dpif_dummy_register__(type);
3713 sset_destroy(&types);
3716 dpif_dummy_register__("dummy");
3718 unixctl_command_register("dpif-dummy/change-port-number",
3719 "dp port new-number",
3720 3, 3, dpif_dummy_change_port_number, NULL);
3721 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3722 2, 2, dpif_dummy_delete_port, NULL);
3725 /* Datapath Classifier. */
3727 /* A set of rules that all have the same fields wildcarded. */
3728 struct dpcls_subtable {
3729 /* The fields are only used by writers. */
3730 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3732 /* These fields are accessed by readers. */
3733 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3734 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3735 /* 'mask' must be the last field, additional space is allocated here. */
3738 /* Initializes 'cls' as a classifier that initially contains no classification
3741 dpcls_init(struct dpcls *cls)
3743 cmap_init(&cls->subtables_map);
3744 pvector_init(&cls->subtables);
3748 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3750 pvector_remove(&cls->subtables, subtable);
3751 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3752 subtable->mask.hash);
3753 cmap_destroy(&subtable->rules);
3754 ovsrcu_postpone(free, subtable);
3757 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3758 * caller's responsibility.
3759 * May only be called after all the readers have been terminated. */
3761 dpcls_destroy(struct dpcls *cls)
3764 struct dpcls_subtable *subtable;
3766 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3767 dpcls_destroy_subtable(cls, subtable);
3769 cmap_destroy(&cls->subtables_map);
3770 pvector_destroy(&cls->subtables);
3774 static struct dpcls_subtable *
3775 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3777 struct dpcls_subtable *subtable;
3779 /* Need to add one. */
3780 subtable = xmalloc(sizeof *subtable
3781 - sizeof subtable->mask.mf + mask->len);
3782 cmap_init(&subtable->rules);
3783 netdev_flow_key_clone(&subtable->mask, mask);
3784 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3785 pvector_insert(&cls->subtables, subtable, 0);
3786 pvector_publish(&cls->subtables);
3791 static inline struct dpcls_subtable *
3792 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3794 struct dpcls_subtable *subtable;
3796 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3797 &cls->subtables_map) {
3798 if (netdev_flow_key_equal(&subtable->mask, mask)) {
3802 return dpcls_create_subtable(cls, mask);
3805 /* Insert 'rule' into 'cls'. */
3807 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
3808 const struct netdev_flow_key *mask)
3810 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
3812 rule->mask = &subtable->mask;
3813 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
3816 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3818 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
3820 struct dpcls_subtable *subtable;
3822 ovs_assert(rule->mask);
3824 INIT_CONTAINER(subtable, rule->mask, mask);
3826 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
3828 dpcls_destroy_subtable(cls, subtable);
3829 pvector_publish(&cls->subtables);
3833 /* Returns true if 'target' satisifies 'key' in 'mask', that is, if each 1-bit
3834 * in 'mask' the values in 'key' and 'target' are the same.
3836 * Note: 'key' and 'mask' have the same mask, and 'key' is already masked. */
3838 dpcls_rule_matches_key(const struct dpcls_rule *rule,
3839 const struct netdev_flow_key *target)
3841 const uint64_t *keyp = rule->flow.mf.inline_values;
3842 const uint64_t *maskp = rule->mask->mf.inline_values;
3843 uint64_t target_u64;
3845 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(target_u64, target, rule->flow.mf.map) {
3846 if (OVS_UNLIKELY((target_u64 & *maskp++) != *keyp++)) {
3853 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3854 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3855 * NULL it is skipped.
3857 * This function is optimized for use in the userspace datapath and therefore
3858 * does not implement a lot of features available in the standard
3859 * classifier_lookup() function. Specifically, it does not implement
3860 * priorities, instead returning any rule which matches the flow.
3862 * Returns true if all flows found a corresponding rule. */
3864 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
3865 struct dpcls_rule **rules, const size_t cnt)
3867 /* The batch size 16 was experimentally found faster than 8 or 32. */
3868 typedef uint16_t map_type;
3869 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3871 #if !defined(__CHECKER__) && !defined(_WIN32)
3872 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
3874 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
3876 map_type maps[N_MAPS];
3877 struct dpcls_subtable *subtable;
3879 memset(maps, 0xff, sizeof maps);
3880 if (cnt % MAP_BITS) {
3881 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
3883 memset(rules, 0, cnt * sizeof *rules);
3885 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
3886 const struct netdev_flow_key *mkeys = keys;
3887 struct dpcls_rule **mrules = rules;
3888 map_type remains = 0;
3891 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
3893 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
3894 uint32_t hashes[MAP_BITS];
3895 const struct cmap_node *nodes[MAP_BITS];
3896 unsigned long map = maps[m];
3900 continue; /* Skip empty maps. */
3903 /* Compute hashes for the remaining keys. */
3904 ULONG_FOR_EACH_1(i, map) {
3905 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
3909 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
3910 /* Check results. */
3911 ULONG_FOR_EACH_1(i, map) {
3912 struct dpcls_rule *rule;
3914 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
3915 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
3920 ULONG_SET0(map, i); /* Did not match. */
3922 ; /* Keep Sparse happy. */
3924 maps[m] &= ~map; /* Clear the found rules. */
3928 return true; /* All found. */
3931 return false; /* Some misses. */