2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "dpif-netdev.h"
24 #include <netinet/in.h>
25 #include <sys/socket.h>
30 #include <sys/ioctl.h>
37 #include "dp-packet.h"
39 #include "dpif-provider.h"
41 #include "dynamic-string.h"
42 #include "fat-rwlock.h"
50 #include "netdev-dpdk.h"
51 #include "netdev-vport.h"
53 #include "odp-execute.h"
55 #include "ofp-print.h"
60 #include "poll-loop.h"
67 #include "tnl-neigh-cache.h"
68 #include "tnl-ports.h"
71 #include "openvswitch/vlog.h"
73 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
75 #define FLOW_DUMP_MAX_BATCH 50
76 /* Use per thread recirc_depth to prevent recirculation loop. */
77 #define MAX_RECIRC_DEPTH 5
78 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
80 /* Configuration parameters. */
81 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
83 /* Protects against changes to 'dp_netdevs'. */
84 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
86 /* Contains all 'struct dp_netdev's. */
87 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
88 = SHASH_INITIALIZER(&dp_netdevs);
90 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
92 static struct odp_support dp_netdev_support = {
93 .max_mpls_depth = SIZE_MAX,
97 /* Stores a miniflow with inline values */
99 struct netdev_flow_key {
100 uint32_t hash; /* Hash function differs for different users. */
101 uint32_t len; /* Length of the following miniflow (incl. map). */
103 uint64_t buf[FLOW_MAX_PACKET_U64S];
106 /* Exact match cache for frequently used flows
108 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
109 * search its entries for a miniflow that matches exactly the miniflow of the
110 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
112 * A cache entry holds a reference to its 'dp_netdev_flow'.
114 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
115 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
116 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
117 * value is the index of a cache entry where the miniflow could be.
123 * Each pmd_thread has its own private exact match cache.
124 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
127 #define EM_FLOW_HASH_SHIFT 13
128 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
129 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
130 #define EM_FLOW_HASH_SEGS 2
133 struct dp_netdev_flow *flow;
134 struct netdev_flow_key key; /* key.hash used for emc hash value. */
138 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
139 int sweep_idx; /* For emc_cache_slow_sweep(). */
142 /* Iterate in the exact match cache through every entry that might contain a
143 * miniflow with hash 'HASH'. */
144 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
145 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
146 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
147 i__ < EM_FLOW_HASH_SEGS; \
148 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
150 /* Simple non-wildcarding single-priority classifier. */
153 struct cmap subtables_map;
154 struct pvector subtables;
157 /* A rule to be inserted to the classifier. */
159 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
160 struct netdev_flow_key *mask; /* Subtable's mask. */
161 struct netdev_flow_key flow; /* Matching key. */
162 /* 'flow' must be the last field, additional space is allocated here. */
165 static void dpcls_init(struct dpcls *);
166 static void dpcls_destroy(struct dpcls *);
167 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
168 const struct netdev_flow_key *mask);
169 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
170 static bool dpcls_lookup(const struct dpcls *cls,
171 const struct netdev_flow_key keys[],
172 struct dpcls_rule **rules, size_t cnt);
174 /* Datapath based on the network device interface from netdev.h.
180 * Some members, marked 'const', are immutable. Accessing other members
181 * requires synchronization, as noted in more detail below.
183 * Acquisition order is, from outermost to innermost:
185 * dp_netdev_mutex (global)
189 const struct dpif_class *const class;
190 const char *const name;
192 struct ovs_refcount ref_cnt;
193 atomic_flag destroyed;
197 * Protected by RCU. Take the mutex to add or remove ports. */
198 struct ovs_mutex port_mutex;
200 struct seq *port_seq; /* Incremented whenever a port changes. */
202 /* Protects access to ofproto-dpif-upcall interface during revalidator
203 * thread synchronization. */
204 struct fat_rwlock upcall_rwlock;
205 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
208 /* Callback function for notifying the purging of dp flows (during
209 * reseting pmd deletion). */
210 dp_purge_callback *dp_purge_cb;
213 /* Stores all 'struct dp_netdev_pmd_thread's. */
214 struct cmap poll_threads;
216 /* Protects the access of the 'struct dp_netdev_pmd_thread'
217 * instance for non-pmd thread. */
218 struct ovs_mutex non_pmd_mutex;
220 /* Each pmd thread will store its pointer to
221 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
222 ovsthread_key_t per_pmd_key;
224 /* Number of rx queues for each dpdk interface and the cpu mask
225 * for pin of pmd threads. */
228 uint64_t last_tnl_conf_seq;
231 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
235 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
236 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
237 DP_STAT_MISS, /* Packets that did not match. */
238 DP_STAT_LOST, /* Packets not passed up to the client. */
242 enum pmd_cycles_counter_type {
243 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
244 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
248 /* A port in a netdev-based datapath. */
249 struct dp_netdev_port {
251 struct netdev *netdev;
252 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
253 struct netdev_saved_flags *sf;
254 struct netdev_rxq **rxq;
255 struct ovs_refcount ref_cnt;
256 char *type; /* Port type as requested by user. */
259 /* Contained by struct dp_netdev_flow's 'stats' member. */
260 struct dp_netdev_flow_stats {
261 atomic_llong used; /* Last used time, in monotonic msecs. */
262 atomic_ullong packet_count; /* Number of packets matched. */
263 atomic_ullong byte_count; /* Number of bytes matched. */
264 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
267 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
273 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
274 * its pmd thread's classifier. The text below calls this classifier 'cls'.
279 * The thread safety rules described here for "struct dp_netdev_flow" are
280 * motivated by two goals:
282 * - Prevent threads that read members of "struct dp_netdev_flow" from
283 * reading bad data due to changes by some thread concurrently modifying
286 * - Prevent two threads making changes to members of a given "struct
287 * dp_netdev_flow" from interfering with each other.
293 * A flow 'flow' may be accessed without a risk of being freed during an RCU
294 * grace period. Code that needs to hold onto a flow for a while
295 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
297 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
298 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
301 * Some members, marked 'const', are immutable. Accessing other members
302 * requires synchronization, as noted in more detail below.
304 struct dp_netdev_flow {
305 const struct flow flow; /* Unmasked flow that created this entry. */
306 /* Hash table index by unmasked flow. */
307 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
309 const ovs_u128 ufid; /* Unique flow identifier. */
310 const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */
313 /* Number of references.
314 * The classifier owns one reference.
315 * Any thread trying to keep a rule from being freed should hold its own
317 struct ovs_refcount ref_cnt;
322 struct dp_netdev_flow_stats stats;
325 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
327 /* While processing a group of input packets, the datapath uses the next
328 * member to store a pointer to the output batch for the flow. It is
329 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
330 * packet_batch_init() and packet_batch_execute()). */
331 struct packet_batch *batch;
333 /* Packet classification. */
334 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
335 /* 'cr' must be the last member. */
338 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
339 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
340 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
343 /* A set of datapath actions within a "struct dp_netdev_flow".
349 * A struct dp_netdev_actions 'actions' is protected with RCU. */
350 struct dp_netdev_actions {
351 /* These members are immutable: they do not change during the struct's
353 unsigned int size; /* Size of 'actions', in bytes. */
354 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
357 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
359 struct dp_netdev_actions *dp_netdev_flow_get_actions(
360 const struct dp_netdev_flow *);
361 static void dp_netdev_actions_free(struct dp_netdev_actions *);
363 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
364 struct dp_netdev_pmd_stats {
365 /* Indexed by DP_STAT_*. */
366 atomic_ullong n[DP_N_STATS];
369 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
370 struct dp_netdev_pmd_cycles {
371 /* Indexed by PMD_CYCLES_*. */
372 atomic_ullong n[PMD_N_CYCLES];
375 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
376 * the performance overhead of interrupt processing. Therefore netdev can
377 * not implement rx-wait for these devices. dpif-netdev needs to poll
378 * these device to check for recv buffer. pmd-thread does polling for
379 * devices assigned to itself.
381 * DPDK used PMD for accessing NIC.
383 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
384 * I/O of all non-pmd threads. There will be no actual thread created
387 * Each struct has its own flow table and classifier. Packets received
388 * from managed ports are looked up in the corresponding pmd thread's
389 * flow table, and are executed with the found actions.
391 struct dp_netdev_pmd_thread {
392 struct dp_netdev *dp;
393 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
394 struct cmap_node node; /* In 'dp->poll_threads'. */
396 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
397 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
399 /* Per thread exact-match cache. Note, the instance for cpu core
400 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
401 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
402 * instances will only be accessed by its own pmd thread. */
403 struct emc_cache flow_cache;
405 /* Classifier and Flow-Table.
407 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
408 * changes to 'cls' must be made while still holding the 'flow_mutex'.
410 struct ovs_mutex flow_mutex;
412 struct cmap flow_table OVS_GUARDED; /* Flow table. */
415 struct dp_netdev_pmd_stats stats;
417 /* Cycles counters */
418 struct dp_netdev_pmd_cycles cycles;
420 /* Used to count cicles. See 'cycles_counter_end()' */
421 unsigned long long last_cycles;
423 struct latch exit_latch; /* For terminating the pmd thread. */
424 atomic_uint change_seq; /* For reloading pmd ports. */
426 int index; /* Idx of this pmd thread among pmd*/
427 /* threads on same numa node. */
428 unsigned core_id; /* CPU core id of this pmd thread. */
429 int numa_id; /* numa node id of this pmd thread. */
430 int tx_qid; /* Queue id used by this pmd thread to
431 * send packets on all netdevs */
433 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
434 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
435 * values and subtracts them from 'stats' and 'cycles' before
436 * reporting to the user */
437 unsigned long long stats_zero[DP_N_STATS];
438 uint64_t cycles_zero[PMD_N_CYCLES];
441 #define PMD_INITIAL_SEQ 1
443 /* Interface to netdev-based datapath. */
446 struct dp_netdev *dp;
447 uint64_t last_port_seq;
450 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
451 struct dp_netdev_port **portp);
452 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
453 struct dp_netdev_port **portp);
454 static void dp_netdev_free(struct dp_netdev *)
455 OVS_REQUIRES(dp_netdev_mutex);
456 static int do_add_port(struct dp_netdev *dp, const char *devname,
457 const char *type, odp_port_t port_no)
458 OVS_REQUIRES(dp->port_mutex);
459 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
460 OVS_REQUIRES(dp->port_mutex);
461 static int dpif_netdev_open(const struct dpif_class *, const char *name,
462 bool create, struct dpif **);
463 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
464 struct dp_packet **, int c,
466 const struct nlattr *actions,
468 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
469 struct dp_packet **, int cnt);
471 static void dp_netdev_disable_upcall(struct dp_netdev *);
472 void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
473 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
474 struct dp_netdev *dp, int index,
475 unsigned core_id, int numa_id);
476 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
477 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
478 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
480 static struct dp_netdev_pmd_thread *
481 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
482 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
483 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
484 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
485 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
486 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
487 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
488 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
490 static inline bool emc_entry_alive(struct emc_entry *ce);
491 static void emc_clear_entry(struct emc_entry *ce);
494 emc_cache_init(struct emc_cache *flow_cache)
498 flow_cache->sweep_idx = 0;
499 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
500 flow_cache->entries[i].flow = NULL;
501 flow_cache->entries[i].key.hash = 0;
502 flow_cache->entries[i].key.len = sizeof(struct miniflow);
503 flowmap_init(&flow_cache->entries[i].key.mf.map);
508 emc_cache_uninit(struct emc_cache *flow_cache)
512 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
513 emc_clear_entry(&flow_cache->entries[i]);
517 /* Check and clear dead flow references slowly (one entry at each
520 emc_cache_slow_sweep(struct emc_cache *flow_cache)
522 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
524 if (!emc_entry_alive(entry)) {
525 emc_clear_entry(entry);
527 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
530 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
532 dpif_is_netdev(const struct dpif *dpif)
534 return dpif->dpif_class->open == dpif_netdev_open;
537 static struct dpif_netdev *
538 dpif_netdev_cast(const struct dpif *dpif)
540 ovs_assert(dpif_is_netdev(dpif));
541 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
544 static struct dp_netdev *
545 get_dp_netdev(const struct dpif *dpif)
547 return dpif_netdev_cast(dpif)->dp;
551 PMD_INFO_SHOW_STATS, /* show how cpu cycles are spent */
552 PMD_INFO_CLEAR_STATS /* set the cycles count to 0 */
556 pmd_info_show_stats(struct ds *reply,
557 struct dp_netdev_pmd_thread *pmd,
558 unsigned long long stats[DP_N_STATS],
559 uint64_t cycles[PMD_N_CYCLES])
561 unsigned long long total_packets = 0;
562 uint64_t total_cycles = 0;
565 /* These loops subtracts reference values ('*_zero') from the counters.
566 * Since loads and stores are relaxed, it might be possible for a '*_zero'
567 * value to be more recent than the current value we're reading from the
568 * counter. This is not a big problem, since these numbers are not
569 * supposed to be too accurate, but we should at least make sure that
570 * the result is not negative. */
571 for (i = 0; i < DP_N_STATS; i++) {
572 if (stats[i] > pmd->stats_zero[i]) {
573 stats[i] -= pmd->stats_zero[i];
578 if (i != DP_STAT_LOST) {
579 /* Lost packets are already included in DP_STAT_MISS */
580 total_packets += stats[i];
584 for (i = 0; i < PMD_N_CYCLES; i++) {
585 if (cycles[i] > pmd->cycles_zero[i]) {
586 cycles[i] -= pmd->cycles_zero[i];
591 total_cycles += cycles[i];
594 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
595 ? "main thread" : "pmd thread");
597 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
598 ds_put_format(reply, " numa_id %d", pmd->numa_id);
600 if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) {
601 ds_put_format(reply, " core_id %u", pmd->core_id);
603 ds_put_cstr(reply, ":\n");
606 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
607 "\tmiss:%llu\n\tlost:%llu\n",
608 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
609 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
611 if (total_cycles == 0) {
616 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
617 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
618 cycles[PMD_CYCLES_POLLING],
619 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
620 cycles[PMD_CYCLES_PROCESSING],
621 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
623 if (total_packets == 0) {
628 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
629 total_cycles / (double)total_packets,
630 total_cycles, total_packets);
633 "\tavg processing cycles per packet: "
634 "%.02f (%"PRIu64"/%llu)\n",
635 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
636 cycles[PMD_CYCLES_PROCESSING], total_packets);
640 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
641 struct dp_netdev_pmd_thread *pmd,
642 unsigned long long stats[DP_N_STATS],
643 uint64_t cycles[PMD_N_CYCLES])
647 /* We cannot write 'stats' and 'cycles' (because they're written by other
648 * threads) and we shouldn't change 'stats' (because they're used to count
649 * datapath stats, which must not be cleared here). Instead, we save the
650 * current values and subtract them from the values to be displayed in the
652 for (i = 0; i < DP_N_STATS; i++) {
653 pmd->stats_zero[i] = stats[i];
655 for (i = 0; i < PMD_N_CYCLES; i++) {
656 pmd->cycles_zero[i] = cycles[i];
661 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
664 struct ds reply = DS_EMPTY_INITIALIZER;
665 struct dp_netdev_pmd_thread *pmd;
666 struct dp_netdev *dp = NULL;
667 enum pmd_info_type type = *(enum pmd_info_type *) aux;
669 ovs_mutex_lock(&dp_netdev_mutex);
672 dp = shash_find_data(&dp_netdevs, argv[1]);
673 } else if (shash_count(&dp_netdevs) == 1) {
674 /* There's only one datapath */
675 dp = shash_first(&dp_netdevs)->data;
679 ovs_mutex_unlock(&dp_netdev_mutex);
680 unixctl_command_reply_error(conn,
681 "please specify an existing datapath");
685 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
686 unsigned long long stats[DP_N_STATS];
687 uint64_t cycles[PMD_N_CYCLES];
690 /* Read current stats and cycle counters */
691 for (i = 0; i < ARRAY_SIZE(stats); i++) {
692 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
694 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
695 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
698 if (type == PMD_INFO_CLEAR_STATS) {
699 pmd_info_clear_stats(&reply, pmd, stats, cycles);
700 } else if (type == PMD_INFO_SHOW_STATS) {
701 pmd_info_show_stats(&reply, pmd, stats, cycles);
705 ovs_mutex_unlock(&dp_netdev_mutex);
707 unixctl_command_reply(conn, ds_cstr(&reply));
712 dpif_netdev_init(void)
714 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
715 clear_aux = PMD_INFO_CLEAR_STATS;
717 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
718 0, 1, dpif_netdev_pmd_info,
720 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
721 0, 1, dpif_netdev_pmd_info,
727 dpif_netdev_enumerate(struct sset *all_dps,
728 const struct dpif_class *dpif_class)
730 struct shash_node *node;
732 ovs_mutex_lock(&dp_netdev_mutex);
733 SHASH_FOR_EACH(node, &dp_netdevs) {
734 struct dp_netdev *dp = node->data;
735 if (dpif_class != dp->class) {
736 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
737 * If the class doesn't match, skip this dpif. */
740 sset_add(all_dps, node->name);
742 ovs_mutex_unlock(&dp_netdev_mutex);
748 dpif_netdev_class_is_dummy(const struct dpif_class *class)
750 return class != &dpif_netdev_class;
754 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
756 return strcmp(type, "internal") ? type
757 : dpif_netdev_class_is_dummy(class) ? "dummy"
762 create_dpif_netdev(struct dp_netdev *dp)
764 uint16_t netflow_id = hash_string(dp->name, 0);
765 struct dpif_netdev *dpif;
767 ovs_refcount_ref(&dp->ref_cnt);
769 dpif = xmalloc(sizeof *dpif);
770 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
772 dpif->last_port_seq = seq_read(dp->port_seq);
777 /* Choose an unused, non-zero port number and return it on success.
778 * Return ODPP_NONE on failure. */
780 choose_port(struct dp_netdev *dp, const char *name)
781 OVS_REQUIRES(dp->port_mutex)
785 if (dp->class != &dpif_netdev_class) {
789 /* If the port name begins with "br", start the number search at
790 * 100 to make writing tests easier. */
791 if (!strncmp(name, "br", 2)) {
795 /* If the port name contains a number, try to assign that port number.
796 * This can make writing unit tests easier because port numbers are
798 for (p = name; *p != '\0'; p++) {
799 if (isdigit((unsigned char) *p)) {
800 port_no = start_no + strtol(p, NULL, 10);
801 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
802 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
803 return u32_to_odp(port_no);
810 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
811 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
812 return u32_to_odp(port_no);
820 create_dp_netdev(const char *name, const struct dpif_class *class,
821 struct dp_netdev **dpp)
822 OVS_REQUIRES(dp_netdev_mutex)
824 struct dp_netdev *dp;
827 dp = xzalloc(sizeof *dp);
828 shash_add(&dp_netdevs, name, dp);
830 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
831 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
832 ovs_refcount_init(&dp->ref_cnt);
833 atomic_flag_clear(&dp->destroyed);
835 ovs_mutex_init(&dp->port_mutex);
836 cmap_init(&dp->ports);
837 dp->port_seq = seq_create();
838 fat_rwlock_init(&dp->upcall_rwlock);
840 /* Disable upcalls by default. */
841 dp_netdev_disable_upcall(dp);
842 dp->upcall_aux = NULL;
843 dp->upcall_cb = NULL;
845 cmap_init(&dp->poll_threads);
846 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
847 ovsthread_key_create(&dp->per_pmd_key, NULL);
849 dp_netdev_set_nonpmd(dp);
850 dp->n_dpdk_rxqs = NR_QUEUE;
852 ovs_mutex_lock(&dp->port_mutex);
853 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
854 ovs_mutex_unlock(&dp->port_mutex);
860 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
866 dpif_netdev_open(const struct dpif_class *class, const char *name,
867 bool create, struct dpif **dpifp)
869 struct dp_netdev *dp;
872 ovs_mutex_lock(&dp_netdev_mutex);
873 dp = shash_find_data(&dp_netdevs, name);
875 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
877 error = (dp->class != class ? EINVAL
882 *dpifp = create_dpif_netdev(dp);
885 ovs_mutex_unlock(&dp_netdev_mutex);
891 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
892 OVS_NO_THREAD_SAFETY_ANALYSIS
894 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
895 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
897 /* Before freeing a lock we should release it */
898 fat_rwlock_unlock(&dp->upcall_rwlock);
899 fat_rwlock_destroy(&dp->upcall_rwlock);
902 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
903 * through the 'dp_netdevs' shash while freeing 'dp'. */
905 dp_netdev_free(struct dp_netdev *dp)
906 OVS_REQUIRES(dp_netdev_mutex)
908 struct dp_netdev_port *port;
910 shash_find_and_delete(&dp_netdevs, dp->name);
912 dp_netdev_destroy_all_pmds(dp);
913 cmap_destroy(&dp->poll_threads);
914 ovs_mutex_destroy(&dp->non_pmd_mutex);
915 ovsthread_key_delete(dp->per_pmd_key);
917 ovs_mutex_lock(&dp->port_mutex);
918 CMAP_FOR_EACH (port, node, &dp->ports) {
919 do_del_port(dp, port);
921 ovs_mutex_unlock(&dp->port_mutex);
923 seq_destroy(dp->port_seq);
924 cmap_destroy(&dp->ports);
926 /* Upcalls must be disabled at this point */
927 dp_netdev_destroy_upcall_lock(dp);
930 free(CONST_CAST(char *, dp->name));
935 dp_netdev_unref(struct dp_netdev *dp)
938 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
939 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
940 ovs_mutex_lock(&dp_netdev_mutex);
941 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
944 ovs_mutex_unlock(&dp_netdev_mutex);
949 dpif_netdev_close(struct dpif *dpif)
951 struct dp_netdev *dp = get_dp_netdev(dpif);
958 dpif_netdev_destroy(struct dpif *dpif)
960 struct dp_netdev *dp = get_dp_netdev(dpif);
962 if (!atomic_flag_test_and_set(&dp->destroyed)) {
963 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
964 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
972 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
973 * load/store semantics. While the increment is not atomic, the load and
974 * store operations are, making it impossible to read inconsistent values.
976 * This is used to update thread local stats counters. */
978 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
980 unsigned long long tmp;
982 atomic_read_relaxed(var, &tmp);
984 atomic_store_relaxed(var, tmp);
988 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
990 struct dp_netdev *dp = get_dp_netdev(dpif);
991 struct dp_netdev_pmd_thread *pmd;
993 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
994 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
995 unsigned long long n;
996 stats->n_flows += cmap_count(&pmd->flow_table);
998 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
1000 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
1002 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
1003 stats->n_missed += n;
1004 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
1007 stats->n_masks = UINT32_MAX;
1008 stats->n_mask_hit = UINT64_MAX;
1014 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1018 if (pmd->core_id == NON_PMD_CORE_ID) {
1022 ovs_mutex_lock(&pmd->cond_mutex);
1023 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1024 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1025 ovs_mutex_unlock(&pmd->cond_mutex);
1028 /* Causes all pmd threads to reload its tx/rx devices.
1029 * Must be called after adding/removing ports. */
1031 dp_netdev_reload_pmds(struct dp_netdev *dp)
1033 struct dp_netdev_pmd_thread *pmd;
1035 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1036 dp_netdev_reload_pmd__(pmd);
1041 hash_port_no(odp_port_t port_no)
1043 return hash_int(odp_to_u32(port_no), 0);
1047 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1049 OVS_REQUIRES(dp->port_mutex)
1051 struct netdev_saved_flags *sf;
1052 struct dp_netdev_port *port;
1053 struct netdev *netdev;
1054 enum netdev_flags flags;
1055 const char *open_type;
1059 /* Reject devices already in 'dp'. */
1060 if (!get_port_by_name(dp, devname, &port)) {
1064 /* Open and validate network device. */
1065 open_type = dpif_netdev_port_open_type(dp->class, type);
1066 error = netdev_open(devname, open_type, &netdev);
1070 /* XXX reject non-Ethernet devices */
1072 netdev_get_flags(netdev, &flags);
1073 if (flags & NETDEV_LOOPBACK) {
1074 VLOG_ERR("%s: cannot add a loopback device", devname);
1075 netdev_close(netdev);
1079 if (netdev_is_pmd(netdev)) {
1080 int n_cores = ovs_numa_get_n_cores();
1082 if (n_cores == OVS_CORE_UNSPEC) {
1083 VLOG_ERR("%s, cannot get cpu core info", devname);
1086 /* There can only be ovs_numa_get_n_cores() pmd threads,
1087 * so creates a txq for each, and one extra for the non
1089 error = netdev_set_multiq(netdev, n_cores + 1, dp->n_dpdk_rxqs);
1090 if (error && (error != EOPNOTSUPP)) {
1091 VLOG_ERR("%s, cannot set multiq", devname);
1095 port = xzalloc(sizeof *port);
1096 port->port_no = port_no;
1097 port->netdev = netdev;
1098 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
1099 port->type = xstrdup(type);
1100 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1101 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1103 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
1104 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1105 devname, ovs_strerror(errno));
1106 netdev_close(netdev);
1114 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1116 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1117 netdev_rxq_close(port->rxq[i]);
1119 netdev_close(netdev);
1127 ovs_refcount_init(&port->ref_cnt);
1128 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1130 if (netdev_is_pmd(netdev)) {
1131 dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
1132 dp_netdev_reload_pmds(dp);
1134 seq_change(dp->port_seq);
1140 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1141 odp_port_t *port_nop)
1143 struct dp_netdev *dp = get_dp_netdev(dpif);
1144 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1145 const char *dpif_port;
1149 ovs_mutex_lock(&dp->port_mutex);
1150 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1151 if (*port_nop != ODPP_NONE) {
1152 port_no = *port_nop;
1153 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1155 port_no = choose_port(dp, dpif_port);
1156 error = port_no == ODPP_NONE ? EFBIG : 0;
1159 *port_nop = port_no;
1160 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1162 ovs_mutex_unlock(&dp->port_mutex);
1168 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1170 struct dp_netdev *dp = get_dp_netdev(dpif);
1173 ovs_mutex_lock(&dp->port_mutex);
1174 if (port_no == ODPP_LOCAL) {
1177 struct dp_netdev_port *port;
1179 error = get_port_by_number(dp, port_no, &port);
1181 do_del_port(dp, port);
1184 ovs_mutex_unlock(&dp->port_mutex);
1190 is_valid_port_number(odp_port_t port_no)
1192 return port_no != ODPP_NONE;
1195 static struct dp_netdev_port *
1196 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1198 struct dp_netdev_port *port;
1200 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1201 if (port->port_no == port_no) {
1209 get_port_by_number(struct dp_netdev *dp,
1210 odp_port_t port_no, struct dp_netdev_port **portp)
1212 if (!is_valid_port_number(port_no)) {
1216 *portp = dp_netdev_lookup_port(dp, port_no);
1217 return *portp ? 0 : ENOENT;
1222 port_ref(struct dp_netdev_port *port)
1225 ovs_refcount_ref(&port->ref_cnt);
1230 port_try_ref(struct dp_netdev_port *port)
1233 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
1240 port_unref(struct dp_netdev_port *port)
1242 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1243 int n_rxq = netdev_n_rxq(port->netdev);
1246 netdev_close(port->netdev);
1247 netdev_restore_flags(port->sf);
1249 for (i = 0; i < n_rxq; i++) {
1250 netdev_rxq_close(port->rxq[i]);
1259 get_port_by_name(struct dp_netdev *dp,
1260 const char *devname, struct dp_netdev_port **portp)
1261 OVS_REQUIRES(dp->port_mutex)
1263 struct dp_netdev_port *port;
1265 CMAP_FOR_EACH (port, node, &dp->ports) {
1266 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1275 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1277 struct dp_netdev_pmd_thread *pmd;
1280 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1281 if (pmd->numa_id == numa_id) {
1289 /* Returns 'true' if there is a port with pmd netdev and the netdev
1290 * is on numa node 'numa_id'. */
1292 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1294 struct dp_netdev_port *port;
1296 CMAP_FOR_EACH (port, node, &dp->ports) {
1297 if (netdev_is_pmd(port->netdev)
1298 && netdev_get_numa_id(port->netdev) == numa_id) {
1308 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1309 OVS_REQUIRES(dp->port_mutex)
1311 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1312 seq_change(dp->port_seq);
1313 if (netdev_is_pmd(port->netdev)) {
1314 int numa_id = netdev_get_numa_id(port->netdev);
1316 /* If there is no netdev on the numa node, deletes the pmd threads
1317 * for that numa. Else, just reloads the queues. */
1318 if (!has_pmd_port_for_numa(dp, numa_id)) {
1319 dp_netdev_del_pmds_on_numa(dp, numa_id);
1321 dp_netdev_reload_pmds(dp);
1328 answer_port_query(const struct dp_netdev_port *port,
1329 struct dpif_port *dpif_port)
1331 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1332 dpif_port->type = xstrdup(port->type);
1333 dpif_port->port_no = port->port_no;
1337 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1338 struct dpif_port *dpif_port)
1340 struct dp_netdev *dp = get_dp_netdev(dpif);
1341 struct dp_netdev_port *port;
1344 error = get_port_by_number(dp, port_no, &port);
1345 if (!error && dpif_port) {
1346 answer_port_query(port, dpif_port);
1353 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1354 struct dpif_port *dpif_port)
1356 struct dp_netdev *dp = get_dp_netdev(dpif);
1357 struct dp_netdev_port *port;
1360 ovs_mutex_lock(&dp->port_mutex);
1361 error = get_port_by_name(dp, devname, &port);
1362 if (!error && dpif_port) {
1363 answer_port_query(port, dpif_port);
1365 ovs_mutex_unlock(&dp->port_mutex);
1371 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1373 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1377 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1379 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1380 ovsrcu_postpone(dp_netdev_flow_free, flow);
1385 dp_netdev_flow_hash(const ovs_u128 *ufid)
1387 return ufid->u32[0];
1391 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1392 struct dp_netdev_flow *flow)
1393 OVS_REQUIRES(pmd->flow_mutex)
1395 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1397 dpcls_remove(&pmd->cls, &flow->cr);
1398 flow->cr.mask = NULL; /* Accessing rule's mask after this is not safe. */
1400 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1403 dp_netdev_flow_unref(flow);
1407 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1409 struct dp_netdev_flow *netdev_flow;
1411 ovs_mutex_lock(&pmd->flow_mutex);
1412 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1413 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1415 ovs_mutex_unlock(&pmd->flow_mutex);
1419 dpif_netdev_flow_flush(struct dpif *dpif)
1421 struct dp_netdev *dp = get_dp_netdev(dpif);
1422 struct dp_netdev_pmd_thread *pmd;
1424 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1425 dp_netdev_pmd_flow_flush(pmd);
1431 struct dp_netdev_port_state {
1432 struct cmap_position position;
1437 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1439 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1444 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1445 struct dpif_port *dpif_port)
1447 struct dp_netdev_port_state *state = state_;
1448 struct dp_netdev *dp = get_dp_netdev(dpif);
1449 struct cmap_node *node;
1452 node = cmap_next_position(&dp->ports, &state->position);
1454 struct dp_netdev_port *port;
1456 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1459 state->name = xstrdup(netdev_get_name(port->netdev));
1460 dpif_port->name = state->name;
1461 dpif_port->type = port->type;
1462 dpif_port->port_no = port->port_no;
1473 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1475 struct dp_netdev_port_state *state = state_;
1482 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1484 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1485 uint64_t new_port_seq;
1488 new_port_seq = seq_read(dpif->dp->port_seq);
1489 if (dpif->last_port_seq != new_port_seq) {
1490 dpif->last_port_seq = new_port_seq;
1500 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1502 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1504 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1507 static struct dp_netdev_flow *
1508 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1510 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1513 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1515 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1518 /* netdev_flow_key utilities.
1520 * netdev_flow_key is basically a miniflow. We use these functions
1521 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1522 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1524 * - Since we are dealing exclusively with miniflows created by
1525 * miniflow_extract(), if the map is different the miniflow is different.
1526 * Therefore we can be faster by comparing the map and the miniflow in a
1528 * - These functions can be inlined by the compiler. */
1530 /* Given the number of bits set in miniflow's maps, returns the size of the
1531 * 'netdev_flow_key.mf' */
1532 static inline size_t
1533 netdev_flow_key_size(size_t flow_u64s)
1535 return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s);
1539 netdev_flow_key_equal(const struct netdev_flow_key *a,
1540 const struct netdev_flow_key *b)
1542 /* 'b->len' may be not set yet. */
1543 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1546 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1547 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1548 * generated by miniflow_extract. */
1550 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1551 const struct miniflow *mf)
1553 return !memcmp(&key->mf, mf, key->len);
1557 netdev_flow_key_clone(struct netdev_flow_key *dst,
1558 const struct netdev_flow_key *src)
1561 offsetof(struct netdev_flow_key, mf) + src->len);
1566 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1567 const struct flow *src)
1569 struct dp_packet packet;
1570 uint64_t buf_stub[512 / 8];
1572 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1573 pkt_metadata_from_flow(&packet.md, src);
1574 flow_compose(&packet, src);
1575 miniflow_extract(&packet, &dst->mf);
1576 dp_packet_uninit(&packet);
1578 dst->len = netdev_flow_key_size(miniflow_n_values(&dst->mf));
1579 dst->hash = 0; /* Not computed yet. */
1582 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1584 netdev_flow_mask_init(struct netdev_flow_key *mask,
1585 const struct match *match)
1587 uint64_t *dst = miniflow_values(&mask->mf);
1588 struct flowmap fmap;
1592 /* Only check masks that make sense for the flow. */
1593 flow_wc_map(&match->flow, &fmap);
1594 flowmap_init(&mask->mf.map);
1596 FLOWMAP_FOR_EACH_INDEX(idx, fmap) {
1597 uint64_t mask_u64 = flow_u64_value(&match->wc.masks, idx);
1600 flowmap_set(&mask->mf.map, idx, 1);
1602 hash = hash_add64(hash, mask_u64);
1608 FLOWMAP_FOR_EACH_MAP (map, mask->mf.map) {
1609 hash = hash_add64(hash, map);
1612 size_t n = dst - miniflow_get_values(&mask->mf);
1614 mask->hash = hash_finish(hash, n * 8);
1615 mask->len = netdev_flow_key_size(n);
1618 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1620 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1621 const struct flow *flow,
1622 const struct netdev_flow_key *mask)
1624 uint64_t *dst_u64 = miniflow_values(&dst->mf);
1625 const uint64_t *mask_u64 = miniflow_get_values(&mask->mf);
1629 dst->len = mask->len;
1630 dst->mf = mask->mf; /* Copy maps. */
1632 FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf.map) {
1633 *dst_u64 = value & *mask_u64++;
1634 hash = hash_add64(hash, *dst_u64++);
1636 dst->hash = hash_finish(hash,
1637 (dst_u64 - miniflow_get_values(&dst->mf)) * 8);
1640 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
1641 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
1642 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
1644 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1646 static inline uint32_t
1647 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1648 const struct netdev_flow_key *mask)
1650 const uint64_t *p = miniflow_get_values(&mask->mf);
1654 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, key, mask->mf.map) {
1655 hash = hash_add64(hash, value & *p++);
1658 return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8);
1662 emc_entry_alive(struct emc_entry *ce)
1664 return ce->flow && !ce->flow->dead;
1668 emc_clear_entry(struct emc_entry *ce)
1671 dp_netdev_flow_unref(ce->flow);
1677 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1678 const struct netdev_flow_key *key)
1680 if (ce->flow != flow) {
1682 dp_netdev_flow_unref(ce->flow);
1685 if (dp_netdev_flow_ref(flow)) {
1692 netdev_flow_key_clone(&ce->key, key);
1697 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1698 struct dp_netdev_flow *flow)
1700 struct emc_entry *to_be_replaced = NULL;
1701 struct emc_entry *current_entry;
1703 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1704 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1705 /* We found the entry with the 'mf' miniflow */
1706 emc_change_entry(current_entry, flow, NULL);
1710 /* Replacement policy: put the flow in an empty (not alive) entry, or
1711 * in the first entry where it can be */
1713 || (emc_entry_alive(to_be_replaced)
1714 && !emc_entry_alive(current_entry))
1715 || current_entry->key.hash < to_be_replaced->key.hash) {
1716 to_be_replaced = current_entry;
1719 /* We didn't find the miniflow in the cache.
1720 * The 'to_be_replaced' entry is where the new flow will be stored */
1722 emc_change_entry(to_be_replaced, flow, key);
1725 static inline struct dp_netdev_flow *
1726 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1728 struct emc_entry *current_entry;
1730 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1731 if (current_entry->key.hash == key->hash
1732 && emc_entry_alive(current_entry)
1733 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1735 /* We found the entry with the 'key->mf' miniflow */
1736 return current_entry->flow;
1743 static struct dp_netdev_flow *
1744 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1745 const struct netdev_flow_key *key)
1747 struct dp_netdev_flow *netdev_flow;
1748 struct dpcls_rule *rule;
1750 dpcls_lookup(&pmd->cls, key, &rule, 1);
1751 netdev_flow = dp_netdev_flow_cast(rule);
1756 static struct dp_netdev_flow *
1757 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1758 const ovs_u128 *ufidp, const struct nlattr *key,
1761 struct dp_netdev_flow *netdev_flow;
1765 /* If a UFID is not provided, determine one based on the key. */
1766 if (!ufidp && key && key_len
1767 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1768 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1773 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1775 if (ovs_u128_equals(&netdev_flow->ufid, ufidp)) {
1785 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1786 struct dpif_flow_stats *stats)
1788 struct dp_netdev_flow *netdev_flow;
1789 unsigned long long n;
1793 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1795 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1796 stats->n_packets = n;
1797 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1799 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1801 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1802 stats->tcp_flags = flags;
1805 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1806 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1807 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1810 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1811 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1812 struct dpif_flow *flow, bool terse)
1815 memset(flow, 0, sizeof *flow);
1817 struct flow_wildcards wc;
1818 struct dp_netdev_actions *actions;
1820 struct odp_flow_key_parms odp_parms = {
1821 .flow = &netdev_flow->flow,
1823 .support = dp_netdev_support,
1826 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1829 offset = key_buf->size;
1830 flow->key = ofpbuf_tail(key_buf);
1831 odp_parms.odp_in_port = netdev_flow->flow.in_port.odp_port;
1832 odp_flow_key_from_flow(&odp_parms, key_buf);
1833 flow->key_len = key_buf->size - offset;
1836 offset = mask_buf->size;
1837 flow->mask = ofpbuf_tail(mask_buf);
1838 odp_parms.odp_in_port = wc.masks.in_port.odp_port;
1839 odp_parms.key_buf = key_buf;
1840 odp_flow_key_from_mask(&odp_parms, mask_buf);
1841 flow->mask_len = mask_buf->size - offset;
1844 actions = dp_netdev_flow_get_actions(netdev_flow);
1845 flow->actions = actions->actions;
1846 flow->actions_len = actions->size;
1849 flow->ufid = netdev_flow->ufid;
1850 flow->ufid_present = true;
1851 flow->pmd_id = netdev_flow->pmd_id;
1852 get_dpif_flow_stats(netdev_flow, &flow->stats);
1856 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1857 const struct nlattr *mask_key,
1858 uint32_t mask_key_len, const struct flow *flow,
1859 struct flow_wildcards *wc)
1861 enum odp_key_fitness fitness;
1863 fitness = odp_flow_key_to_mask_udpif(mask_key, mask_key_len, key,
1866 /* This should not happen: it indicates that
1867 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1868 * disagree on the acceptable form of a mask. Log the problem
1869 * as an error, with enough details to enable debugging. */
1870 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1872 if (!VLOG_DROP_ERR(&rl)) {
1876 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1878 VLOG_ERR("internal error parsing flow mask %s (%s)",
1879 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1890 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1895 if (odp_flow_key_to_flow_udpif(key, key_len, flow)) {
1896 /* This should not happen: it indicates that odp_flow_key_from_flow()
1897 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1898 * flow. Log the problem as an error, with enough details to enable
1900 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1902 if (!VLOG_DROP_ERR(&rl)) {
1906 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1907 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1914 in_port = flow->in_port.odp_port;
1915 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1919 /* Userspace datapath doesn't support conntrack. */
1920 if (flow->ct_state || flow->ct_zone || flow->ct_mark
1921 || !ovs_u128_is_zero(&flow->ct_label)) {
1929 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1931 struct dp_netdev *dp = get_dp_netdev(dpif);
1932 struct dp_netdev_flow *netdev_flow;
1933 struct dp_netdev_pmd_thread *pmd;
1934 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1935 ? NON_PMD_CORE_ID : get->pmd_id;
1938 pmd = dp_netdev_get_pmd(dp, pmd_id);
1943 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1946 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1951 dp_netdev_pmd_unref(pmd);
1957 static struct dp_netdev_flow *
1958 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
1959 struct match *match, const ovs_u128 *ufid,
1960 const struct nlattr *actions, size_t actions_len)
1961 OVS_REQUIRES(pmd->flow_mutex)
1963 struct dp_netdev_flow *flow;
1964 struct netdev_flow_key mask;
1966 netdev_flow_mask_init(&mask, match);
1967 /* Make sure wc does not have metadata. */
1968 ovs_assert(!FLOWMAP_HAS_FIELD(&mask.mf.map, metadata)
1969 && !FLOWMAP_HAS_FIELD(&mask.mf.map, regs));
1971 /* Do not allocate extra space. */
1972 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
1973 memset(&flow->stats, 0, sizeof flow->stats);
1976 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
1977 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
1978 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
1979 ovs_refcount_init(&flow->ref_cnt);
1980 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
1982 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
1983 dpcls_insert(&pmd->cls, &flow->cr, &mask);
1985 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
1986 dp_netdev_flow_hash(&flow->ufid));
1988 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1990 struct ds ds = DS_EMPTY_INITIALIZER;
1992 match.tun_md.valid = false;
1993 match.flow = flow->flow;
1994 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
1996 ds_put_cstr(&ds, "flow_add: ");
1997 odp_format_ufid(ufid, &ds);
1998 ds_put_cstr(&ds, " ");
1999 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2000 ds_put_cstr(&ds, ", actions:");
2001 format_odp_actions(&ds, actions, actions_len);
2003 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2012 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2014 struct dp_netdev *dp = get_dp_netdev(dpif);
2015 struct dp_netdev_flow *netdev_flow;
2016 struct netdev_flow_key key;
2017 struct dp_netdev_pmd_thread *pmd;
2020 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2021 ? NON_PMD_CORE_ID : put->pmd_id;
2024 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2028 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2029 put->mask, put->mask_len,
2030 &match.flow, &match.wc);
2035 pmd = dp_netdev_get_pmd(dp, pmd_id);
2040 /* Must produce a netdev_flow_key for lookup.
2041 * This interface is no longer performance critical, since it is not used
2042 * for upcall processing any more. */
2043 netdev_flow_key_from_flow(&key, &match.flow);
2048 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2051 ovs_mutex_lock(&pmd->flow_mutex);
2052 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2054 if (put->flags & DPIF_FP_CREATE) {
2055 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2057 memset(put->stats, 0, sizeof *put->stats);
2059 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2069 if (put->flags & DPIF_FP_MODIFY
2070 && flow_equal(&match.flow, &netdev_flow->flow)) {
2071 struct dp_netdev_actions *new_actions;
2072 struct dp_netdev_actions *old_actions;
2074 new_actions = dp_netdev_actions_create(put->actions,
2077 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2078 ovsrcu_set(&netdev_flow->actions, new_actions);
2081 get_dpif_flow_stats(netdev_flow, put->stats);
2083 if (put->flags & DPIF_FP_ZERO_STATS) {
2084 /* XXX: The userspace datapath uses thread local statistics
2085 * (for flows), which should be updated only by the owning
2086 * thread. Since we cannot write on stats memory here,
2087 * we choose not to support this flag. Please note:
2088 * - This feature is currently used only by dpctl commands with
2090 * - Should the need arise, this operation can be implemented
2091 * by keeping a base value (to be update here) for each
2092 * counter, and subtracting it before outputting the stats */
2096 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2097 } else if (put->flags & DPIF_FP_CREATE) {
2100 /* Overlapping flow. */
2104 ovs_mutex_unlock(&pmd->flow_mutex);
2105 dp_netdev_pmd_unref(pmd);
2111 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2113 struct dp_netdev *dp = get_dp_netdev(dpif);
2114 struct dp_netdev_flow *netdev_flow;
2115 struct dp_netdev_pmd_thread *pmd;
2116 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2117 ? NON_PMD_CORE_ID : del->pmd_id;
2120 pmd = dp_netdev_get_pmd(dp, pmd_id);
2125 ovs_mutex_lock(&pmd->flow_mutex);
2126 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2130 get_dpif_flow_stats(netdev_flow, del->stats);
2132 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2136 ovs_mutex_unlock(&pmd->flow_mutex);
2137 dp_netdev_pmd_unref(pmd);
2142 struct dpif_netdev_flow_dump {
2143 struct dpif_flow_dump up;
2144 struct cmap_position poll_thread_pos;
2145 struct cmap_position flow_pos;
2146 struct dp_netdev_pmd_thread *cur_pmd;
2148 struct ovs_mutex mutex;
2151 static struct dpif_netdev_flow_dump *
2152 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2154 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2157 static struct dpif_flow_dump *
2158 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2160 struct dpif_netdev_flow_dump *dump;
2162 dump = xzalloc(sizeof *dump);
2163 dpif_flow_dump_init(&dump->up, dpif_);
2164 dump->up.terse = terse;
2165 ovs_mutex_init(&dump->mutex);
2171 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2173 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2175 ovs_mutex_destroy(&dump->mutex);
2180 struct dpif_netdev_flow_dump_thread {
2181 struct dpif_flow_dump_thread up;
2182 struct dpif_netdev_flow_dump *dump;
2183 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2184 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2187 static struct dpif_netdev_flow_dump_thread *
2188 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2190 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2193 static struct dpif_flow_dump_thread *
2194 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2196 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2197 struct dpif_netdev_flow_dump_thread *thread;
2199 thread = xmalloc(sizeof *thread);
2200 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2201 thread->dump = dump;
2206 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2208 struct dpif_netdev_flow_dump_thread *thread
2209 = dpif_netdev_flow_dump_thread_cast(thread_);
2215 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2216 struct dpif_flow *flows, int max_flows)
2218 struct dpif_netdev_flow_dump_thread *thread
2219 = dpif_netdev_flow_dump_thread_cast(thread_);
2220 struct dpif_netdev_flow_dump *dump = thread->dump;
2221 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2225 ovs_mutex_lock(&dump->mutex);
2226 if (!dump->status) {
2227 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2228 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2229 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2230 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2232 /* First call to dump_next(), extracts the first pmd thread.
2233 * If there is no pmd thread, returns immediately. */
2235 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2237 ovs_mutex_unlock(&dump->mutex);
2244 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2245 struct cmap_node *node;
2247 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2251 netdev_flows[n_flows] = CONTAINER_OF(node,
2252 struct dp_netdev_flow,
2255 /* When finishing dumping the current pmd thread, moves to
2257 if (n_flows < flow_limit) {
2258 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2259 dp_netdev_pmd_unref(pmd);
2260 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2266 /* Keeps the reference to next caller. */
2267 dump->cur_pmd = pmd;
2269 /* If the current dump is empty, do not exit the loop, since the
2270 * remaining pmds could have flows to be dumped. Just dumps again
2271 * on the new 'pmd'. */
2274 ovs_mutex_unlock(&dump->mutex);
2276 for (i = 0; i < n_flows; i++) {
2277 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2278 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2279 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2280 struct dpif_flow *f = &flows[i];
2281 struct ofpbuf key, mask;
2283 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2284 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2285 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2293 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2294 OVS_NO_THREAD_SAFETY_ANALYSIS
2296 struct dp_netdev *dp = get_dp_netdev(dpif);
2297 struct dp_netdev_pmd_thread *pmd;
2298 struct dp_packet *pp;
2300 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2301 dp_packet_size(execute->packet) > UINT16_MAX) {
2305 /* Tries finding the 'pmd'. If NULL is returned, that means
2306 * the current thread is a non-pmd thread and should use
2307 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2308 pmd = ovsthread_getspecific(dp->per_pmd_key);
2310 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2313 /* If the current thread is non-pmd thread, acquires
2314 * the 'non_pmd_mutex'. */
2315 if (pmd->core_id == NON_PMD_CORE_ID) {
2316 ovs_mutex_lock(&dp->non_pmd_mutex);
2317 ovs_mutex_lock(&dp->port_mutex);
2320 pp = execute->packet;
2321 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2322 execute->actions_len);
2323 if (pmd->core_id == NON_PMD_CORE_ID) {
2324 dp_netdev_pmd_unref(pmd);
2325 ovs_mutex_unlock(&dp->port_mutex);
2326 ovs_mutex_unlock(&dp->non_pmd_mutex);
2333 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2337 for (i = 0; i < n_ops; i++) {
2338 struct dpif_op *op = ops[i];
2341 case DPIF_OP_FLOW_PUT:
2342 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2345 case DPIF_OP_FLOW_DEL:
2346 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2349 case DPIF_OP_EXECUTE:
2350 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2353 case DPIF_OP_FLOW_GET:
2354 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2360 /* Returns true if the configuration for rx queues or cpu mask
2363 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2365 if (dp->n_dpdk_rxqs != rxqs) {
2368 if (dp->pmd_cmask != NULL && cmask != NULL) {
2369 return strcmp(dp->pmd_cmask, cmask);
2371 return (dp->pmd_cmask != NULL || cmask != NULL);
2376 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2378 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2380 struct dp_netdev *dp = get_dp_netdev(dpif);
2382 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2383 struct dp_netdev_port *port;
2385 dp_netdev_destroy_all_pmds(dp);
2387 CMAP_FOR_EACH (port, node, &dp->ports) {
2388 if (netdev_is_pmd(port->netdev)) {
2391 /* Closes the existing 'rxq's. */
2392 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2393 netdev_rxq_close(port->rxq[i]);
2394 port->rxq[i] = NULL;
2397 /* Sets the new rx queue config. */
2398 err = netdev_set_multiq(port->netdev,
2399 ovs_numa_get_n_cores() + 1,
2401 if (err && (err != EOPNOTSUPP)) {
2402 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2403 " %u", netdev_get_name(port->netdev),
2408 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2409 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2410 * netdev_n_rxq(port->netdev));
2411 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2412 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2416 dp->n_dpdk_rxqs = n_rxqs;
2418 /* Reconfigures the cpu mask. */
2419 ovs_numa_set_cpu_mask(cmask);
2420 free(dp->pmd_cmask);
2421 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2423 /* Restores the non-pmd. */
2424 dp_netdev_set_nonpmd(dp);
2425 /* Restores all pmd threads. */
2426 dp_netdev_reset_pmd_threads(dp);
2433 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2434 uint32_t queue_id, uint32_t *priority)
2436 *priority = queue_id;
2441 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2442 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2443 struct dp_netdev_actions *
2444 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2446 struct dp_netdev_actions *netdev_actions;
2448 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2449 memcpy(netdev_actions->actions, actions, size);
2450 netdev_actions->size = size;
2452 return netdev_actions;
2455 struct dp_netdev_actions *
2456 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2458 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2462 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2467 static inline unsigned long long
2468 cycles_counter(void)
2471 return rte_get_tsc_cycles();
2477 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2478 extern struct ovs_mutex cycles_counter_fake_mutex;
2480 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2482 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2483 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2484 OVS_NO_THREAD_SAFETY_ANALYSIS
2486 pmd->last_cycles = cycles_counter();
2489 /* Stop counting cycles and add them to the counter 'type' */
2491 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2492 enum pmd_cycles_counter_type type)
2493 OVS_RELEASES(&cycles_counter_fake_mutex)
2494 OVS_NO_THREAD_SAFETY_ANALYSIS
2496 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2498 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2502 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2503 struct dp_netdev_port *port,
2504 struct netdev_rxq *rxq)
2506 struct dp_packet *packets[NETDEV_MAX_BURST];
2509 cycles_count_start(pmd);
2510 error = netdev_rxq_recv(rxq, packets, &cnt);
2511 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2515 *recirc_depth_get() = 0;
2517 /* XXX: initialize md in netdev implementation. */
2518 for (i = 0; i < cnt; i++) {
2519 pkt_metadata_init(&packets[i]->md, port->port_no);
2521 cycles_count_start(pmd);
2522 dp_netdev_input(pmd, packets, cnt);
2523 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2524 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2525 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2527 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2528 netdev_get_name(port->netdev), ovs_strerror(error));
2532 /* Return true if needs to revalidate datapath flows. */
2534 dpif_netdev_run(struct dpif *dpif)
2536 struct dp_netdev_port *port;
2537 struct dp_netdev *dp = get_dp_netdev(dpif);
2538 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2540 uint64_t new_tnl_seq;
2542 ovs_mutex_lock(&dp->non_pmd_mutex);
2543 CMAP_FOR_EACH (port, node, &dp->ports) {
2544 if (!netdev_is_pmd(port->netdev)) {
2547 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2548 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2552 ovs_mutex_unlock(&dp->non_pmd_mutex);
2553 dp_netdev_pmd_unref(non_pmd);
2555 tnl_neigh_cache_run();
2557 new_tnl_seq = seq_read(tnl_conf_seq);
2559 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2560 dp->last_tnl_conf_seq = new_tnl_seq;
2567 dpif_netdev_wait(struct dpif *dpif)
2569 struct dp_netdev_port *port;
2570 struct dp_netdev *dp = get_dp_netdev(dpif);
2572 ovs_mutex_lock(&dp_netdev_mutex);
2573 CMAP_FOR_EACH (port, node, &dp->ports) {
2574 if (!netdev_is_pmd(port->netdev)) {
2577 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2578 netdev_rxq_wait(port->rxq[i]);
2582 ovs_mutex_unlock(&dp_netdev_mutex);
2583 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2587 struct dp_netdev_port *port;
2588 struct netdev_rxq *rx;
2592 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2593 struct rxq_poll **ppoll_list, int poll_cnt)
2595 struct rxq_poll *poll_list = *ppoll_list;
2596 struct dp_netdev_port *port;
2597 int n_pmds_on_numa, index, i;
2599 /* Simple scheduler for netdev rx polling. */
2600 for (i = 0; i < poll_cnt; i++) {
2601 port_unref(poll_list[i].port);
2605 n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
2608 CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
2609 /* Calls port_try_ref() to prevent the main thread
2610 * from deleting the port. */
2611 if (port_try_ref(port)) {
2612 if (netdev_is_pmd(port->netdev)
2613 && netdev_get_numa_id(port->netdev) == pmd->numa_id) {
2616 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2617 if ((index % n_pmds_on_numa) == pmd->index) {
2618 poll_list = xrealloc(poll_list,
2619 sizeof *poll_list * (poll_cnt + 1));
2622 poll_list[poll_cnt].port = port;
2623 poll_list[poll_cnt].rx = port->rxq[i];
2629 /* Unrefs the port_try_ref(). */
2634 *ppoll_list = poll_list;
2639 pmd_thread_main(void *f_)
2641 struct dp_netdev_pmd_thread *pmd = f_;
2642 unsigned int lc = 0;
2643 struct rxq_poll *poll_list;
2644 unsigned int port_seq = PMD_INITIAL_SEQ;
2651 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2652 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2653 pmd_thread_setaffinity_cpu(pmd->core_id);
2655 emc_cache_init(&pmd->flow_cache);
2656 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2658 /* List port/core affinity */
2659 for (i = 0; i < poll_cnt; i++) {
2660 VLOG_INFO("Core %d processing port \'%s\'\n", pmd->core_id, netdev_get_name(poll_list[i].port->netdev));
2663 /* Signal here to make sure the pmd finishes
2664 * reloading the updated configuration. */
2665 dp_netdev_pmd_reload_done(pmd);
2670 for (i = 0; i < poll_cnt; i++) {
2671 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2679 emc_cache_slow_sweep(&pmd->flow_cache);
2680 coverage_try_clear();
2683 atomic_read_relaxed(&pmd->change_seq, &seq);
2684 if (seq != port_seq) {
2691 emc_cache_uninit(&pmd->flow_cache);
2693 if (!latch_is_set(&pmd->exit_latch)){
2697 for (i = 0; i < poll_cnt; i++) {
2698 port_unref(poll_list[i].port);
2701 dp_netdev_pmd_reload_done(pmd);
2708 dp_netdev_disable_upcall(struct dp_netdev *dp)
2709 OVS_ACQUIRES(dp->upcall_rwlock)
2711 fat_rwlock_wrlock(&dp->upcall_rwlock);
2715 dpif_netdev_disable_upcall(struct dpif *dpif)
2716 OVS_NO_THREAD_SAFETY_ANALYSIS
2718 struct dp_netdev *dp = get_dp_netdev(dpif);
2719 dp_netdev_disable_upcall(dp);
2723 dp_netdev_enable_upcall(struct dp_netdev *dp)
2724 OVS_RELEASES(dp->upcall_rwlock)
2726 fat_rwlock_unlock(&dp->upcall_rwlock);
2730 dpif_netdev_enable_upcall(struct dpif *dpif)
2731 OVS_NO_THREAD_SAFETY_ANALYSIS
2733 struct dp_netdev *dp = get_dp_netdev(dpif);
2734 dp_netdev_enable_upcall(dp);
2738 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2740 ovs_mutex_lock(&pmd->cond_mutex);
2741 xpthread_cond_signal(&pmd->cond);
2742 ovs_mutex_unlock(&pmd->cond_mutex);
2745 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2746 * the pointer if succeeds, otherwise, NULL.
2748 * Caller must unrefs the returned reference. */
2749 static struct dp_netdev_pmd_thread *
2750 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2752 struct dp_netdev_pmd_thread *pmd;
2753 const struct cmap_node *pnode;
2755 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2759 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2761 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2764 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2766 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2768 struct dp_netdev_pmd_thread *non_pmd;
2770 non_pmd = xzalloc(sizeof *non_pmd);
2771 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2775 /* Caller must have valid pointer to 'pmd'. */
2777 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2779 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2783 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2785 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2786 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2790 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2791 * fails, keeps checking for next node until reaching the end of cmap.
2793 * Caller must unrefs the returned reference. */
2794 static struct dp_netdev_pmd_thread *
2795 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2797 struct dp_netdev_pmd_thread *next;
2800 struct cmap_node *node;
2802 node = cmap_next_position(&dp->poll_threads, pos);
2803 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2805 } while (next && !dp_netdev_pmd_try_ref(next));
2811 core_id_to_qid(unsigned core_id)
2813 if (core_id != NON_PMD_CORE_ID) {
2816 return ovs_numa_get_n_cores();
2820 /* Configures the 'pmd' based on the input argument. */
2822 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2823 int index, unsigned core_id, int numa_id)
2827 pmd->core_id = core_id;
2828 pmd->tx_qid = core_id_to_qid(core_id);
2829 pmd->numa_id = numa_id;
2831 ovs_refcount_init(&pmd->ref_cnt);
2832 latch_init(&pmd->exit_latch);
2833 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2834 xpthread_cond_init(&pmd->cond, NULL);
2835 ovs_mutex_init(&pmd->cond_mutex);
2836 ovs_mutex_init(&pmd->flow_mutex);
2837 dpcls_init(&pmd->cls);
2838 cmap_init(&pmd->flow_table);
2839 /* init the 'flow_cache' since there is no
2840 * actual thread created for NON_PMD_CORE_ID. */
2841 if (core_id == NON_PMD_CORE_ID) {
2842 emc_cache_init(&pmd->flow_cache);
2844 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2845 hash_int(core_id, 0));
2849 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2851 dp_netdev_pmd_flow_flush(pmd);
2852 dpcls_destroy(&pmd->cls);
2853 cmap_destroy(&pmd->flow_table);
2854 ovs_mutex_destroy(&pmd->flow_mutex);
2855 latch_destroy(&pmd->exit_latch);
2856 xpthread_cond_destroy(&pmd->cond);
2857 ovs_mutex_destroy(&pmd->cond_mutex);
2861 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2862 * and unrefs the struct. */
2864 dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd)
2866 /* Uninit the 'flow_cache' since there is
2867 * no actual thread uninit it for NON_PMD_CORE_ID. */
2868 if (pmd->core_id == NON_PMD_CORE_ID) {
2869 emc_cache_uninit(&pmd->flow_cache);
2871 latch_set(&pmd->exit_latch);
2872 dp_netdev_reload_pmd__(pmd);
2873 ovs_numa_unpin_core(pmd->core_id);
2874 xpthread_join(pmd->thread, NULL);
2876 /* Purges the 'pmd''s flows after stopping the thread, but before
2877 * destroying the flows, so that the flow stats can be collected. */
2878 if (dp->dp_purge_cb) {
2879 dp->dp_purge_cb(dp->dp_purge_aux, pmd->core_id);
2881 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2882 dp_netdev_pmd_unref(pmd);
2885 /* Destroys all pmd threads. */
2887 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2889 struct dp_netdev_pmd_thread *pmd;
2891 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2892 dp_netdev_del_pmd(dp, pmd);
2896 /* Deletes all pmd threads on numa node 'numa_id'. */
2898 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2900 struct dp_netdev_pmd_thread *pmd;
2902 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2903 if (pmd->numa_id == numa_id) {
2904 dp_netdev_del_pmd(dp, pmd);
2909 /* Checks the numa node id of 'netdev' and starts pmd threads for
2912 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2916 if (!ovs_numa_numa_id_is_valid(numa_id)) {
2917 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2918 "invalid", numa_id);
2922 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
2924 /* If there are already pmd threads created for the numa node
2925 * in which 'netdev' is on, do nothing. Else, creates the
2926 * pmd threads for the numa node. */
2928 int can_have, n_unpinned, i;
2929 struct dp_netdev_pmd_thread **pmds;
2931 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
2933 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2934 "cores on numa node");
2938 /* If cpu mask is specified, uses all unpinned cores, otherwise
2939 * tries creating NR_PMD_THREADS pmd threads. */
2940 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
2941 pmds = xzalloc(can_have * sizeof *pmds);
2942 for (i = 0; i < can_have; i++) {
2943 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
2944 pmds[i] = xzalloc(sizeof **pmds);
2945 dp_netdev_configure_pmd(pmds[i], dp, i, core_id, numa_id);
2947 /* The pmd thread code needs to see all the others configured pmd
2948 * threads on the same numa node. That's why we call
2949 * 'dp_netdev_configure_pmd()' on all the threads and then we actually
2951 for (i = 0; i < can_have; i++) {
2952 /* Each thread will distribute all devices rx-queues among
2954 pmds[i]->thread = ovs_thread_create("pmd", pmd_thread_main, pmds[i]);
2957 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
2962 /* Called after pmd threads config change. Restarts pmd threads with
2963 * new configuration. */
2965 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
2967 struct dp_netdev_port *port;
2969 CMAP_FOR_EACH (port, node, &dp->ports) {
2970 if (netdev_is_pmd(port->netdev)) {
2971 int numa_id = netdev_get_numa_id(port->netdev);
2973 dp_netdev_set_pmds_on_numa(dp, numa_id);
2979 dpif_netdev_get_datapath_version(void)
2981 return xstrdup("<built-in>");
2985 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
2986 uint16_t tcp_flags, long long now)
2990 atomic_store_relaxed(&netdev_flow->stats.used, now);
2991 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
2992 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
2993 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
2995 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
2999 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
3000 enum dp_stat_type type, int cnt)
3002 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3006 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3007 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3008 enum dpif_upcall_type type, const struct nlattr *userdata,
3009 struct ofpbuf *actions, struct ofpbuf *put_actions)
3011 struct dp_netdev *dp = pmd->dp;
3012 struct flow_tnl orig_tunnel;
3015 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3019 /* Upcall processing expects the Geneve options to be in the translated
3020 * format but we need to retain the raw format for datapath use. */
3021 orig_tunnel.flags = flow->tunnel.flags;
3022 if (flow->tunnel.flags & FLOW_TNL_F_UDPIF) {
3023 orig_tunnel.metadata.present.len = flow->tunnel.metadata.present.len;
3024 memcpy(orig_tunnel.metadata.opts.gnv, flow->tunnel.metadata.opts.gnv,
3025 flow->tunnel.metadata.present.len);
3026 err = tun_metadata_from_geneve_udpif(&orig_tunnel, &orig_tunnel,
3033 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3034 struct ds ds = DS_EMPTY_INITIALIZER;
3037 struct odp_flow_key_parms odp_parms = {
3040 .odp_in_port = flow->in_port.odp_port,
3041 .support = dp_netdev_support,
3044 ofpbuf_init(&key, 0);
3045 odp_flow_key_from_flow(&odp_parms, &key);
3046 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3047 dp_packet_size(packet_));
3049 odp_flow_key_format(key.data, key.size, &ds);
3051 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3052 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3054 ofpbuf_uninit(&key);
3060 err = dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3061 actions, wc, put_actions, dp->upcall_aux);
3062 if (err && err != ENOSPC) {
3066 /* Translate tunnel metadata masks to datapath format. */
3068 if (wc->masks.tunnel.metadata.present.map) {
3069 struct geneve_opt opts[TLV_TOT_OPT_SIZE /
3070 sizeof(struct geneve_opt)];
3072 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3073 tun_metadata_to_geneve_udpif_mask(&flow->tunnel,
3075 orig_tunnel.metadata.opts.gnv,
3076 orig_tunnel.metadata.present.len,
3079 orig_tunnel.metadata.present.len = 0;
3082 memset(&wc->masks.tunnel.metadata, 0,
3083 sizeof wc->masks.tunnel.metadata);
3084 memcpy(&wc->masks.tunnel.metadata.opts.gnv, opts,
3085 orig_tunnel.metadata.present.len);
3087 wc->masks.tunnel.metadata.present.len = 0xff;
3090 /* Restore tunnel metadata. We need to use the saved options to ensure
3091 * that any unknown options are not lost. The generated mask will have
3092 * the same structure, matching on types and lengths but wildcarding
3093 * option data we don't care about. */
3094 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3095 memcpy(&flow->tunnel.metadata.opts.gnv, orig_tunnel.metadata.opts.gnv,
3096 orig_tunnel.metadata.present.len);
3097 flow->tunnel.metadata.present.len = orig_tunnel.metadata.present.len;
3098 flow->tunnel.flags |= FLOW_TNL_F_UDPIF;
3104 static inline uint32_t
3105 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3106 const struct miniflow *mf)
3108 uint32_t hash, recirc_depth;
3110 if (OVS_LIKELY(dp_packet_rss_valid(packet))) {
3111 hash = dp_packet_get_rss_hash(packet);
3113 hash = miniflow_hash_5tuple(mf, 0);
3114 dp_packet_set_rss_hash(packet, hash);
3117 /* The RSS hash must account for the recirculation depth to avoid
3118 * collisions in the exact match cache */
3119 recirc_depth = *recirc_depth_get_unsafe();
3120 if (OVS_UNLIKELY(recirc_depth)) {
3121 hash = hash_finish(hash, recirc_depth);
3122 dp_packet_set_rss_hash(packet, hash);
3127 struct packet_batch {
3128 unsigned int packet_count;
3129 unsigned int byte_count;
3132 struct dp_netdev_flow *flow;
3134 struct dp_packet *packets[NETDEV_MAX_BURST];
3138 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
3139 const struct miniflow *mf)
3141 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3142 batch->packets[batch->packet_count++] = packet;
3143 batch->byte_count += dp_packet_size(packet);
3147 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
3149 flow->batch = batch;
3152 batch->packet_count = 0;
3153 batch->byte_count = 0;
3154 batch->tcp_flags = 0;
3158 packet_batch_execute(struct packet_batch *batch,
3159 struct dp_netdev_pmd_thread *pmd,
3162 struct dp_netdev_actions *actions;
3163 struct dp_netdev_flow *flow = batch->flow;
3165 dp_netdev_flow_used(flow, batch->packet_count, batch->byte_count,
3166 batch->tcp_flags, now);
3168 actions = dp_netdev_flow_get_actions(flow);
3170 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
3171 actions->actions, actions->size);
3175 dp_netdev_queue_batches(struct dp_packet *pkt,
3176 struct dp_netdev_flow *flow, const struct miniflow *mf,
3177 struct packet_batch *batches, size_t *n_batches)
3179 struct packet_batch *batch = flow->batch;
3181 if (OVS_LIKELY(batch)) {
3182 packet_batch_update(batch, pkt, mf);
3186 batch = &batches[(*n_batches)++];
3187 packet_batch_init(batch, flow);
3188 packet_batch_update(batch, pkt, mf);
3192 dp_packet_swap(struct dp_packet **a, struct dp_packet **b)
3194 struct dp_packet *tmp = *a;
3199 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3200 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', the
3201 * miniflow is copied into 'keys' and the packet pointer is moved at the
3202 * beginning of the 'packets' array.
3204 * The function returns the number of packets that needs to be processed in the
3205 * 'packets' array (they have been moved to the beginning of the vector).
3207 static inline size_t
3208 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
3209 size_t cnt, struct netdev_flow_key *keys,
3210 struct packet_batch batches[], size_t *n_batches)
3212 struct emc_cache *flow_cache = &pmd->flow_cache;
3213 struct netdev_flow_key key;
3214 size_t i, notfound_cnt = 0;
3216 for (i = 0; i < cnt; i++) {
3217 struct dp_netdev_flow *flow;
3219 if (OVS_UNLIKELY(dp_packet_size(packets[i]) < ETH_HEADER_LEN)) {
3220 dp_packet_delete(packets[i]);
3225 /* Prefetch next packet data */
3226 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3229 miniflow_extract(packets[i], &key.mf);
3230 key.len = 0; /* Not computed yet. */
3231 key.hash = dpif_netdev_packet_get_rss_hash(packets[i], &key.mf);
3233 flow = emc_lookup(flow_cache, &key);
3234 if (OVS_LIKELY(flow)) {
3235 dp_netdev_queue_batches(packets[i], flow, &key.mf, batches,
3238 if (i != notfound_cnt) {
3239 dp_packet_swap(&packets[i], &packets[notfound_cnt]);
3242 keys[notfound_cnt++] = key;
3246 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - notfound_cnt);
3248 return notfound_cnt;
3252 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3253 struct dp_packet **packets, size_t cnt,
3254 struct netdev_flow_key *keys,
3255 struct packet_batch batches[], size_t *n_batches)
3257 #if !defined(__CHECKER__) && !defined(_WIN32)
3258 const size_t PKT_ARRAY_SIZE = cnt;
3260 /* Sparse or MSVC doesn't like variable length array. */
3261 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3263 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3264 struct dp_netdev *dp = pmd->dp;
3265 struct emc_cache *flow_cache = &pmd->flow_cache;
3266 int miss_cnt = 0, lost_cnt = 0;
3270 for (i = 0; i < cnt; i++) {
3271 /* Key length is needed in all the cases, hash computed on demand. */
3272 keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf));
3274 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3275 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3276 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3277 struct ofpbuf actions, put_actions;
3280 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3281 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3283 for (i = 0; i < cnt; i++) {
3284 struct dp_netdev_flow *netdev_flow;
3285 struct ofpbuf *add_actions;
3289 if (OVS_LIKELY(rules[i])) {
3293 /* It's possible that an earlier slow path execution installed
3294 * a rule covering this flow. In this case, it's a lot cheaper
3295 * to catch it here than execute a miss. */
3296 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3298 rules[i] = &netdev_flow->cr;
3304 match.tun_md.valid = false;
3305 miniflow_expand(&keys[i].mf, &match.flow);
3307 ofpbuf_clear(&actions);
3308 ofpbuf_clear(&put_actions);
3310 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
3311 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
3312 &ufid, DPIF_UC_MISS, NULL, &actions,
3314 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3315 dp_packet_delete(packets[i]);
3320 /* The Netlink encoding of datapath flow keys cannot express
3321 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3322 * tag is interpreted as exact match on the fact that there is no
3323 * VLAN. Unless we refactor a lot of code that translates between
3324 * Netlink and struct flow representations, we have to do the same
3326 if (!match.wc.masks.vlan_tci) {
3327 match.wc.masks.vlan_tci = htons(0xffff);
3330 /* We can't allow the packet batching in the next loop to execute
3331 * the actions. Otherwise, if there are any slow path actions,
3332 * we'll send the packet up twice. */
3333 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
3334 actions.data, actions.size);
3336 add_actions = put_actions.size ? &put_actions : &actions;
3337 if (OVS_LIKELY(error != ENOSPC)) {
3338 /* XXX: There's a race window where a flow covering this packet
3339 * could have already been installed since we last did the flow
3340 * lookup before upcall. This could be solved by moving the
3341 * mutex lock outside the loop, but that's an awful long time
3342 * to be locking everyone out of making flow installs. If we
3343 * move to a per-core classifier, it would be reasonable. */
3344 ovs_mutex_lock(&pmd->flow_mutex);
3345 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3346 if (OVS_LIKELY(!netdev_flow)) {
3347 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3351 ovs_mutex_unlock(&pmd->flow_mutex);
3353 emc_insert(flow_cache, &keys[i], netdev_flow);
3357 ofpbuf_uninit(&actions);
3358 ofpbuf_uninit(&put_actions);
3359 fat_rwlock_unlock(&dp->upcall_rwlock);
3360 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3361 } else if (OVS_UNLIKELY(any_miss)) {
3362 for (i = 0; i < cnt; i++) {
3363 if (OVS_UNLIKELY(!rules[i])) {
3364 dp_packet_delete(packets[i]);
3371 for (i = 0; i < cnt; i++) {
3372 struct dp_packet *packet = packets[i];
3373 struct dp_netdev_flow *flow;
3375 if (OVS_UNLIKELY(!rules[i])) {
3379 flow = dp_netdev_flow_cast(rules[i]);
3381 emc_insert(flow_cache, &keys[i], flow);
3382 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3385 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3386 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3387 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3391 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3392 struct dp_packet **packets, int cnt)
3394 #if !defined(__CHECKER__) && !defined(_WIN32)
3395 const size_t PKT_ARRAY_SIZE = cnt;
3397 /* Sparse or MSVC doesn't like variable length array. */
3398 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3400 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3401 struct packet_batch batches[PKT_ARRAY_SIZE];
3402 long long now = time_msec();
3403 size_t newcnt, n_batches, i;
3406 newcnt = emc_processing(pmd, packets, cnt, keys, batches, &n_batches);
3407 if (OVS_UNLIKELY(newcnt)) {
3408 fast_path_processing(pmd, packets, newcnt, keys, batches, &n_batches);
3411 for (i = 0; i < n_batches; i++) {
3412 batches[i].flow->batch = NULL;
3415 for (i = 0; i < n_batches; i++) {
3416 packet_batch_execute(&batches[i], pmd, now);
3420 struct dp_netdev_execute_aux {
3421 struct dp_netdev_pmd_thread *pmd;
3425 dpif_netdev_register_dp_purge_cb(struct dpif *dpif, dp_purge_callback *cb,
3428 struct dp_netdev *dp = get_dp_netdev(dpif);
3429 dp->dp_purge_aux = aux;
3430 dp->dp_purge_cb = cb;
3434 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3437 struct dp_netdev *dp = get_dp_netdev(dpif);
3438 dp->upcall_aux = aux;
3443 dp_netdev_drop_packets(struct dp_packet **packets, int cnt, bool may_steal)
3448 for (i = 0; i < cnt; i++) {
3449 dp_packet_delete(packets[i]);
3455 push_tnl_action(const struct dp_netdev *dp,
3456 const struct nlattr *attr,
3457 struct dp_packet **packets, int cnt)
3459 struct dp_netdev_port *tun_port;
3460 const struct ovs_action_push_tnl *data;
3462 data = nl_attr_get(attr);
3464 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3468 netdev_push_header(tun_port->netdev, packets, cnt, data);
3474 dp_netdev_clone_pkt_batch(struct dp_packet **dst_pkts,
3475 struct dp_packet **src_pkts, int cnt)
3479 for (i = 0; i < cnt; i++) {
3480 dst_pkts[i] = dp_packet_clone(src_pkts[i]);
3485 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3486 const struct nlattr *a, bool may_steal)
3487 OVS_NO_THREAD_SAFETY_ANALYSIS
3489 struct dp_netdev_execute_aux *aux = aux_;
3490 uint32_t *depth = recirc_depth_get();
3491 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3492 struct dp_netdev *dp = pmd->dp;
3493 int type = nl_attr_type(a);
3494 struct dp_netdev_port *p;
3497 switch ((enum ovs_action_attr)type) {
3498 case OVS_ACTION_ATTR_OUTPUT:
3499 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3500 if (OVS_LIKELY(p)) {
3501 netdev_send(p->netdev, pmd->tx_qid, packets, cnt, may_steal);
3506 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3507 if (*depth < MAX_RECIRC_DEPTH) {
3508 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3512 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3516 err = push_tnl_action(dp, a, packets, cnt);
3519 dp_netdev_input(pmd, packets, cnt);
3522 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3528 case OVS_ACTION_ATTR_TUNNEL_POP:
3529 if (*depth < MAX_RECIRC_DEPTH) {
3530 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3532 p = dp_netdev_lookup_port(dp, portno);
3534 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3538 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3542 err = netdev_pop_header(p->netdev, packets, cnt);
3545 for (i = 0; i < cnt; i++) {
3546 packets[i]->md.in_port.odp_port = portno;
3550 dp_netdev_input(pmd, packets, cnt);
3553 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3560 case OVS_ACTION_ATTR_USERSPACE:
3561 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3562 const struct nlattr *userdata;
3563 struct ofpbuf actions;
3567 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3568 ofpbuf_init(&actions, 0);
3570 for (i = 0; i < cnt; i++) {
3573 ofpbuf_clear(&actions);
3575 flow_extract(packets[i], &flow);
3576 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3577 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3578 DPIF_UC_ACTION, userdata,&actions,
3580 if (!error || error == ENOSPC) {
3581 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3582 actions.data, actions.size);
3583 } else if (may_steal) {
3584 dp_packet_delete(packets[i]);
3587 ofpbuf_uninit(&actions);
3588 fat_rwlock_unlock(&dp->upcall_rwlock);
3594 case OVS_ACTION_ATTR_RECIRC:
3595 if (*depth < MAX_RECIRC_DEPTH) {
3596 struct dp_packet *recirc_pkts[NETDEV_MAX_BURST];
3599 dp_netdev_clone_pkt_batch(recirc_pkts, packets, cnt);
3600 packets = recirc_pkts;
3603 for (i = 0; i < cnt; i++) {
3604 packets[i]->md.recirc_id = nl_attr_get_u32(a);
3608 dp_netdev_input(pmd, packets, cnt);
3614 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3617 case OVS_ACTION_ATTR_CT:
3618 /* If a flow with this action is slow-pathed, datapath assistance is
3619 * required to implement it. However, we don't support this action
3620 * in the userspace datapath. */
3621 VLOG_WARN("Cannot execute conntrack action in userspace.");
3624 case OVS_ACTION_ATTR_PUSH_VLAN:
3625 case OVS_ACTION_ATTR_POP_VLAN:
3626 case OVS_ACTION_ATTR_PUSH_MPLS:
3627 case OVS_ACTION_ATTR_POP_MPLS:
3628 case OVS_ACTION_ATTR_SET:
3629 case OVS_ACTION_ATTR_SET_MASKED:
3630 case OVS_ACTION_ATTR_SAMPLE:
3631 case OVS_ACTION_ATTR_HASH:
3632 case OVS_ACTION_ATTR_UNSPEC:
3633 case __OVS_ACTION_ATTR_MAX:
3637 dp_netdev_drop_packets(packets, cnt, may_steal);
3641 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3642 struct dp_packet **packets, int cnt,
3644 const struct nlattr *actions, size_t actions_len)
3646 struct dp_netdev_execute_aux aux = { pmd };
3648 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3649 actions_len, dp_execute_cb);
3652 const struct dpif_class dpif_netdev_class = {
3655 dpif_netdev_enumerate,
3656 dpif_netdev_port_open_type,
3659 dpif_netdev_destroy,
3662 dpif_netdev_get_stats,
3663 dpif_netdev_port_add,
3664 dpif_netdev_port_del,
3665 dpif_netdev_port_query_by_number,
3666 dpif_netdev_port_query_by_name,
3667 NULL, /* port_get_pid */
3668 dpif_netdev_port_dump_start,
3669 dpif_netdev_port_dump_next,
3670 dpif_netdev_port_dump_done,
3671 dpif_netdev_port_poll,
3672 dpif_netdev_port_poll_wait,
3673 dpif_netdev_flow_flush,
3674 dpif_netdev_flow_dump_create,
3675 dpif_netdev_flow_dump_destroy,
3676 dpif_netdev_flow_dump_thread_create,
3677 dpif_netdev_flow_dump_thread_destroy,
3678 dpif_netdev_flow_dump_next,
3679 dpif_netdev_operate,
3680 NULL, /* recv_set */
3681 NULL, /* handlers_set */
3682 dpif_netdev_pmd_set,
3683 dpif_netdev_queue_to_priority,
3685 NULL, /* recv_wait */
3686 NULL, /* recv_purge */
3687 dpif_netdev_register_dp_purge_cb,
3688 dpif_netdev_register_upcall_cb,
3689 dpif_netdev_enable_upcall,
3690 dpif_netdev_disable_upcall,
3691 dpif_netdev_get_datapath_version,
3692 NULL, /* ct_dump_start */
3693 NULL, /* ct_dump_next */
3694 NULL, /* ct_dump_done */
3695 NULL, /* ct_flush */
3699 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3700 const char *argv[], void *aux OVS_UNUSED)
3702 struct dp_netdev_port *old_port;
3703 struct dp_netdev_port *new_port;
3704 struct dp_netdev *dp;
3707 ovs_mutex_lock(&dp_netdev_mutex);
3708 dp = shash_find_data(&dp_netdevs, argv[1]);
3709 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3710 ovs_mutex_unlock(&dp_netdev_mutex);
3711 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3714 ovs_refcount_ref(&dp->ref_cnt);
3715 ovs_mutex_unlock(&dp_netdev_mutex);
3717 ovs_mutex_lock(&dp->port_mutex);
3718 if (get_port_by_name(dp, argv[2], &old_port)) {
3719 unixctl_command_reply_error(conn, "unknown port");
3723 port_no = u32_to_odp(atoi(argv[3]));
3724 if (!port_no || port_no == ODPP_NONE) {
3725 unixctl_command_reply_error(conn, "bad port number");
3728 if (dp_netdev_lookup_port(dp, port_no)) {
3729 unixctl_command_reply_error(conn, "port number already in use");
3733 /* Remove old port. */
3734 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3735 ovsrcu_postpone(free, old_port);
3737 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3738 new_port = xmemdup(old_port, sizeof *old_port);
3739 new_port->port_no = port_no;
3740 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3742 seq_change(dp->port_seq);
3743 unixctl_command_reply(conn, NULL);
3746 ovs_mutex_unlock(&dp->port_mutex);
3747 dp_netdev_unref(dp);
3751 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3752 const char *argv[], void *aux OVS_UNUSED)
3754 struct dp_netdev_port *port;
3755 struct dp_netdev *dp;
3757 ovs_mutex_lock(&dp_netdev_mutex);
3758 dp = shash_find_data(&dp_netdevs, argv[1]);
3759 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3760 ovs_mutex_unlock(&dp_netdev_mutex);
3761 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3764 ovs_refcount_ref(&dp->ref_cnt);
3765 ovs_mutex_unlock(&dp_netdev_mutex);
3767 ovs_mutex_lock(&dp->port_mutex);
3768 if (get_port_by_name(dp, argv[2], &port)) {
3769 unixctl_command_reply_error(conn, "unknown port");
3770 } else if (port->port_no == ODPP_LOCAL) {
3771 unixctl_command_reply_error(conn, "can't delete local port");
3773 do_del_port(dp, port);
3774 unixctl_command_reply(conn, NULL);
3776 ovs_mutex_unlock(&dp->port_mutex);
3778 dp_netdev_unref(dp);
3782 dpif_dummy_register__(const char *type)
3784 struct dpif_class *class;
3786 class = xmalloc(sizeof *class);
3787 *class = dpif_netdev_class;
3788 class->type = xstrdup(type);
3789 dp_register_provider(class);
3793 dpif_dummy_override(const char *type)
3798 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
3799 * a userland-only build. It's useful for testsuite.
3801 error = dp_unregister_provider(type);
3802 if (error == 0 || error == EAFNOSUPPORT) {
3803 dpif_dummy_register__(type);
3808 dpif_dummy_register(enum dummy_level level)
3810 if (level == DUMMY_OVERRIDE_ALL) {
3815 dp_enumerate_types(&types);
3816 SSET_FOR_EACH (type, &types) {
3817 dpif_dummy_override(type);
3819 sset_destroy(&types);
3820 } else if (level == DUMMY_OVERRIDE_SYSTEM) {
3821 dpif_dummy_override("system");
3824 dpif_dummy_register__("dummy");
3826 unixctl_command_register("dpif-dummy/change-port-number",
3827 "dp port new-number",
3828 3, 3, dpif_dummy_change_port_number, NULL);
3829 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3830 2, 2, dpif_dummy_delete_port, NULL);
3833 /* Datapath Classifier. */
3835 /* A set of rules that all have the same fields wildcarded. */
3836 struct dpcls_subtable {
3837 /* The fields are only used by writers. */
3838 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3840 /* These fields are accessed by readers. */
3841 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3842 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3843 /* 'mask' must be the last field, additional space is allocated here. */
3846 /* Initializes 'cls' as a classifier that initially contains no classification
3849 dpcls_init(struct dpcls *cls)
3851 cmap_init(&cls->subtables_map);
3852 pvector_init(&cls->subtables);
3856 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3858 pvector_remove(&cls->subtables, subtable);
3859 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3860 subtable->mask.hash);
3861 cmap_destroy(&subtable->rules);
3862 ovsrcu_postpone(free, subtable);
3865 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3866 * caller's responsibility.
3867 * May only be called after all the readers have been terminated. */
3869 dpcls_destroy(struct dpcls *cls)
3872 struct dpcls_subtable *subtable;
3874 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3875 ovs_assert(cmap_count(&subtable->rules) == 0);
3876 dpcls_destroy_subtable(cls, subtable);
3878 cmap_destroy(&cls->subtables_map);
3879 pvector_destroy(&cls->subtables);
3883 static struct dpcls_subtable *
3884 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3886 struct dpcls_subtable *subtable;
3888 /* Need to add one. */
3889 subtable = xmalloc(sizeof *subtable
3890 - sizeof subtable->mask.mf + mask->len);
3891 cmap_init(&subtable->rules);
3892 netdev_flow_key_clone(&subtable->mask, mask);
3893 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3894 pvector_insert(&cls->subtables, subtable, 0);
3895 pvector_publish(&cls->subtables);
3900 static inline struct dpcls_subtable *
3901 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3903 struct dpcls_subtable *subtable;
3905 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3906 &cls->subtables_map) {
3907 if (netdev_flow_key_equal(&subtable->mask, mask)) {
3911 return dpcls_create_subtable(cls, mask);
3914 /* Insert 'rule' into 'cls'. */
3916 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
3917 const struct netdev_flow_key *mask)
3919 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
3921 rule->mask = &subtable->mask;
3922 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
3925 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3927 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
3929 struct dpcls_subtable *subtable;
3931 ovs_assert(rule->mask);
3933 INIT_CONTAINER(subtable, rule->mask, mask);
3935 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
3937 dpcls_destroy_subtable(cls, subtable);
3938 pvector_publish(&cls->subtables);
3942 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
3943 * in 'mask' the values in 'key' and 'target' are the same. */
3945 dpcls_rule_matches_key(const struct dpcls_rule *rule,
3946 const struct netdev_flow_key *target)
3948 const uint64_t *keyp = miniflow_get_values(&rule->flow.mf);
3949 const uint64_t *maskp = miniflow_get_values(&rule->mask->mf);
3952 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, target, rule->flow.mf.map) {
3953 if (OVS_UNLIKELY((value & *maskp++) != *keyp++)) {
3960 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3961 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3962 * NULL it is skipped.
3964 * This function is optimized for use in the userspace datapath and therefore
3965 * does not implement a lot of features available in the standard
3966 * classifier_lookup() function. Specifically, it does not implement
3967 * priorities, instead returning any rule which matches the flow.
3969 * Returns true if all flows found a corresponding rule. */
3971 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
3972 struct dpcls_rule **rules, const size_t cnt)
3974 /* The batch size 16 was experimentally found faster than 8 or 32. */
3975 typedef uint16_t map_type;
3976 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3978 #if !defined(__CHECKER__) && !defined(_WIN32)
3979 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
3981 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
3983 map_type maps[N_MAPS];
3984 struct dpcls_subtable *subtable;
3986 memset(maps, 0xff, sizeof maps);
3987 if (cnt % MAP_BITS) {
3988 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
3990 memset(rules, 0, cnt * sizeof *rules);
3992 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
3993 const struct netdev_flow_key *mkeys = keys;
3994 struct dpcls_rule **mrules = rules;
3995 map_type remains = 0;
3998 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
4000 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
4001 uint32_t hashes[MAP_BITS];
4002 const struct cmap_node *nodes[MAP_BITS];
4003 unsigned long map = maps[m];
4007 continue; /* Skip empty maps. */
4010 /* Compute hashes for the remaining keys. */
4011 ULLONG_FOR_EACH_1(i, map) {
4012 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
4016 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
4017 /* Check results. */
4018 ULLONG_FOR_EACH_1(i, map) {
4019 struct dpcls_rule *rule;
4021 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
4022 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
4027 ULLONG_SET0(map, i); /* Did not match. */
4029 ; /* Keep Sparse happy. */
4031 maps[m] &= ~map; /* Clear the found rules. */
4035 return true; /* All found. */
4038 return false; /* Some misses. */