2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "dpif-netdev.h"
24 #include <netinet/in.h>
25 #include <sys/socket.h>
30 #include <sys/ioctl.h>
36 #include "dp-packet.h"
38 #include "dpif-provider.h"
40 #include "dynamic-string.h"
41 #include "fat-rwlock.h"
47 #include "meta-flow.h"
49 #include "netdev-dpdk.h"
50 #include "netdev-vport.h"
52 #include "odp-execute.h"
54 #include "ofp-print.h"
59 #include "poll-loop.h"
66 #include "tnl-arp-cache.h"
69 #include "openvswitch/vlog.h"
71 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
73 #define FLOW_DUMP_MAX_BATCH 50
74 /* Use per thread recirc_depth to prevent recirculation loop. */
75 #define MAX_RECIRC_DEPTH 5
76 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
78 /* Configuration parameters. */
79 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
81 /* Protects against changes to 'dp_netdevs'. */
82 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
84 /* Contains all 'struct dp_netdev's. */
85 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
86 = SHASH_INITIALIZER(&dp_netdevs);
88 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
90 static struct odp_support dp_netdev_support = {
91 .max_mpls_depth = SIZE_MAX,
95 /* Stores a miniflow with inline values */
97 struct netdev_flow_key {
98 uint32_t hash; /* Hash function differs for different users. */
99 uint32_t len; /* Length of the following miniflow (incl. map). */
101 uint64_t buf[FLOW_MAX_PACKET_U64S];
104 /* Exact match cache for frequently used flows
106 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
107 * search its entries for a miniflow that matches exactly the miniflow of the
108 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
110 * A cache entry holds a reference to its 'dp_netdev_flow'.
112 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
113 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
114 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
115 * value is the index of a cache entry where the miniflow could be.
121 * Each pmd_thread has its own private exact match cache.
122 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
125 #define EM_FLOW_HASH_SHIFT 13
126 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
127 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
128 #define EM_FLOW_HASH_SEGS 2
131 struct dp_netdev_flow *flow;
132 struct netdev_flow_key key; /* key.hash used for emc hash value. */
136 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
137 int sweep_idx; /* For emc_cache_slow_sweep(). */
140 /* Iterate in the exact match cache through every entry that might contain a
141 * miniflow with hash 'HASH'. */
142 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
143 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
144 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
145 i__ < EM_FLOW_HASH_SEGS; \
146 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
148 /* Simple non-wildcarding single-priority classifier. */
151 struct cmap subtables_map;
152 struct pvector subtables;
155 /* A rule to be inserted to the classifier. */
157 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
158 struct netdev_flow_key *mask; /* Subtable's mask. */
159 struct netdev_flow_key flow; /* Matching key. */
160 /* 'flow' must be the last field, additional space is allocated here. */
163 static void dpcls_init(struct dpcls *);
164 static void dpcls_destroy(struct dpcls *);
165 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
166 const struct netdev_flow_key *mask);
167 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
168 static bool dpcls_lookup(const struct dpcls *cls,
169 const struct netdev_flow_key keys[],
170 struct dpcls_rule **rules, size_t cnt);
172 /* Datapath based on the network device interface from netdev.h.
178 * Some members, marked 'const', are immutable. Accessing other members
179 * requires synchronization, as noted in more detail below.
181 * Acquisition order is, from outermost to innermost:
183 * dp_netdev_mutex (global)
187 const struct dpif_class *const class;
188 const char *const name;
190 struct ovs_refcount ref_cnt;
191 atomic_flag destroyed;
195 * Protected by RCU. Take the mutex to add or remove ports. */
196 struct ovs_mutex port_mutex;
198 struct seq *port_seq; /* Incremented whenever a port changes. */
200 /* Protects access to ofproto-dpif-upcall interface during revalidator
201 * thread synchronization. */
202 struct fat_rwlock upcall_rwlock;
203 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
206 /* Stores all 'struct dp_netdev_pmd_thread's. */
207 struct cmap poll_threads;
209 /* Protects the access of the 'struct dp_netdev_pmd_thread'
210 * instance for non-pmd thread. */
211 struct ovs_mutex non_pmd_mutex;
213 /* Each pmd thread will store its pointer to
214 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
215 ovsthread_key_t per_pmd_key;
217 /* Number of rx queues for each dpdk interface and the cpu mask
218 * for pin of pmd threads. */
221 uint64_t last_tnl_conf_seq;
224 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
228 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
229 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
230 DP_STAT_MISS, /* Packets that did not match. */
231 DP_STAT_LOST, /* Packets not passed up to the client. */
235 enum pmd_cycles_counter_type {
236 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
237 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
241 /* A port in a netdev-based datapath. */
242 struct dp_netdev_port {
244 struct netdev *netdev;
245 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
246 struct netdev_saved_flags *sf;
247 struct netdev_rxq **rxq;
248 struct ovs_refcount ref_cnt;
249 char *type; /* Port type as requested by user. */
252 /* Contained by struct dp_netdev_flow's 'stats' member. */
253 struct dp_netdev_flow_stats {
254 atomic_llong used; /* Last used time, in monotonic msecs. */
255 atomic_ullong packet_count; /* Number of packets matched. */
256 atomic_ullong byte_count; /* Number of bytes matched. */
257 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
260 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
266 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
267 * its pmd thread's classifier. The text below calls this classifier 'cls'.
272 * The thread safety rules described here for "struct dp_netdev_flow" are
273 * motivated by two goals:
275 * - Prevent threads that read members of "struct dp_netdev_flow" from
276 * reading bad data due to changes by some thread concurrently modifying
279 * - Prevent two threads making changes to members of a given "struct
280 * dp_netdev_flow" from interfering with each other.
286 * A flow 'flow' may be accessed without a risk of being freed during an RCU
287 * grace period. Code that needs to hold onto a flow for a while
288 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
290 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
291 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
294 * Some members, marked 'const', are immutable. Accessing other members
295 * requires synchronization, as noted in more detail below.
297 struct dp_netdev_flow {
298 const struct flow flow; /* Unmasked flow that created this entry. */
299 /* Hash table index by unmasked flow. */
300 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
302 const ovs_u128 ufid; /* Unique flow identifier. */
303 const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */
306 /* Number of references.
307 * The classifier owns one reference.
308 * Any thread trying to keep a rule from being freed should hold its own
310 struct ovs_refcount ref_cnt;
315 struct dp_netdev_flow_stats stats;
318 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
320 /* While processing a group of input packets, the datapath uses the next
321 * member to store a pointer to the output batch for the flow. It is
322 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
323 * packet_batch_init() and packet_batch_execute()). */
324 struct packet_batch *batch;
326 /* Packet classification. */
327 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
328 /* 'cr' must be the last member. */
331 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
332 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
333 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
336 /* A set of datapath actions within a "struct dp_netdev_flow".
342 * A struct dp_netdev_actions 'actions' is protected with RCU. */
343 struct dp_netdev_actions {
344 /* These members are immutable: they do not change during the struct's
346 unsigned int size; /* Size of 'actions', in bytes. */
347 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
350 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
352 struct dp_netdev_actions *dp_netdev_flow_get_actions(
353 const struct dp_netdev_flow *);
354 static void dp_netdev_actions_free(struct dp_netdev_actions *);
356 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
357 struct dp_netdev_pmd_stats {
358 /* Indexed by DP_STAT_*. */
359 atomic_ullong n[DP_N_STATS];
362 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
363 struct dp_netdev_pmd_cycles {
364 /* Indexed by PMD_CYCLES_*. */
365 atomic_ullong n[PMD_N_CYCLES];
368 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
369 * the performance overhead of interrupt processing. Therefore netdev can
370 * not implement rx-wait for these devices. dpif-netdev needs to poll
371 * these device to check for recv buffer. pmd-thread does polling for
372 * devices assigned to itself.
374 * DPDK used PMD for accessing NIC.
376 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
377 * I/O of all non-pmd threads. There will be no actual thread created
380 * Each struct has its own flow table and classifier. Packets received
381 * from managed ports are looked up in the corresponding pmd thread's
382 * flow table, and are executed with the found actions.
384 struct dp_netdev_pmd_thread {
385 struct dp_netdev *dp;
386 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
387 struct cmap_node node; /* In 'dp->poll_threads'. */
389 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
390 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
392 /* Per thread exact-match cache. Note, the instance for cpu core
393 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
394 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
395 * instances will only be accessed by its own pmd thread. */
396 struct emc_cache flow_cache;
398 /* Classifier and Flow-Table.
400 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
401 * changes to 'cls' must be made while still holding the 'flow_mutex'.
403 struct ovs_mutex flow_mutex;
405 struct cmap flow_table OVS_GUARDED; /* Flow table. */
408 struct dp_netdev_pmd_stats stats;
410 /* Cycles counters */
411 struct dp_netdev_pmd_cycles cycles;
413 /* Used to count cicles. See 'cycles_counter_end()' */
414 unsigned long long last_cycles;
416 struct latch exit_latch; /* For terminating the pmd thread. */
417 atomic_uint change_seq; /* For reloading pmd ports. */
419 int index; /* Idx of this pmd thread among pmd*/
420 /* threads on same numa node. */
421 unsigned core_id; /* CPU core id of this pmd thread. */
422 int numa_id; /* numa node id of this pmd thread. */
423 int tx_qid; /* Queue id used by this pmd thread to
424 * send packets on all netdevs */
426 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
427 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
428 * values and subtracts them from 'stats' and 'cycles' before
429 * reporting to the user */
430 unsigned long long stats_zero[DP_N_STATS];
431 uint64_t cycles_zero[PMD_N_CYCLES];
434 #define PMD_INITIAL_SEQ 1
436 /* Interface to netdev-based datapath. */
439 struct dp_netdev *dp;
440 uint64_t last_port_seq;
443 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
444 struct dp_netdev_port **portp);
445 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
446 struct dp_netdev_port **portp);
447 static void dp_netdev_free(struct dp_netdev *)
448 OVS_REQUIRES(dp_netdev_mutex);
449 static int do_add_port(struct dp_netdev *dp, const char *devname,
450 const char *type, odp_port_t port_no)
451 OVS_REQUIRES(dp->port_mutex);
452 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
453 OVS_REQUIRES(dp->port_mutex);
454 static int dpif_netdev_open(const struct dpif_class *, const char *name,
455 bool create, struct dpif **);
456 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
457 struct dp_packet **, int c,
459 const struct nlattr *actions,
461 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
462 struct dp_packet **, int cnt);
464 static void dp_netdev_disable_upcall(struct dp_netdev *);
465 void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
466 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
467 struct dp_netdev *dp, int index,
468 unsigned core_id, int numa_id);
469 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
470 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
471 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
473 static struct dp_netdev_pmd_thread *
474 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
475 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
476 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
477 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
478 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
479 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
480 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
481 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
483 static inline bool emc_entry_alive(struct emc_entry *ce);
484 static void emc_clear_entry(struct emc_entry *ce);
487 emc_cache_init(struct emc_cache *flow_cache)
491 BUILD_ASSERT(sizeof(struct miniflow) == 2 * sizeof(uint64_t));
493 flow_cache->sweep_idx = 0;
494 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
495 flow_cache->entries[i].flow = NULL;
496 flow_cache->entries[i].key.hash = 0;
497 flow_cache->entries[i].key.len = sizeof(struct miniflow);
498 flow_cache->entries[i].key.mf.tnl_map = 0;
499 flow_cache->entries[i].key.mf.pkt_map = 0;
504 emc_cache_uninit(struct emc_cache *flow_cache)
508 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
509 emc_clear_entry(&flow_cache->entries[i]);
513 /* Check and clear dead flow references slowly (one entry at each
516 emc_cache_slow_sweep(struct emc_cache *flow_cache)
518 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
520 if (!emc_entry_alive(entry)) {
521 emc_clear_entry(entry);
523 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
526 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
528 dpif_is_netdev(const struct dpif *dpif)
530 return dpif->dpif_class->open == dpif_netdev_open;
533 static struct dpif_netdev *
534 dpif_netdev_cast(const struct dpif *dpif)
536 ovs_assert(dpif_is_netdev(dpif));
537 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
540 static struct dp_netdev *
541 get_dp_netdev(const struct dpif *dpif)
543 return dpif_netdev_cast(dpif)->dp;
547 PMD_INFO_SHOW_STATS, /* show how cpu cycles are spent */
548 PMD_INFO_CLEAR_STATS /* set the cycles count to 0 */
552 pmd_info_show_stats(struct ds *reply,
553 struct dp_netdev_pmd_thread *pmd,
554 unsigned long long stats[DP_N_STATS],
555 uint64_t cycles[PMD_N_CYCLES])
557 unsigned long long total_packets = 0;
558 uint64_t total_cycles = 0;
561 /* These loops subtracts reference values ('*_zero') from the counters.
562 * Since loads and stores are relaxed, it might be possible for a '*_zero'
563 * value to be more recent than the current value we're reading from the
564 * counter. This is not a big problem, since these numbers are not
565 * supposed to be too accurate, but we should at least make sure that
566 * the result is not negative. */
567 for (i = 0; i < DP_N_STATS; i++) {
568 if (stats[i] > pmd->stats_zero[i]) {
569 stats[i] -= pmd->stats_zero[i];
574 if (i != DP_STAT_LOST) {
575 /* Lost packets are already included in DP_STAT_MISS */
576 total_packets += stats[i];
580 for (i = 0; i < PMD_N_CYCLES; i++) {
581 if (cycles[i] > pmd->cycles_zero[i]) {
582 cycles[i] -= pmd->cycles_zero[i];
587 total_cycles += cycles[i];
590 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
591 ? "main thread" : "pmd thread");
593 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
594 ds_put_format(reply, " numa_id %d", pmd->numa_id);
596 if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) {
597 ds_put_format(reply, " core_id %u", pmd->core_id);
599 ds_put_cstr(reply, ":\n");
602 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
603 "\tmiss:%llu\n\tlost:%llu\n",
604 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
605 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
607 if (total_cycles == 0) {
612 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
613 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
614 cycles[PMD_CYCLES_POLLING],
615 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
616 cycles[PMD_CYCLES_PROCESSING],
617 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
619 if (total_packets == 0) {
624 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
625 total_cycles / (double)total_packets,
626 total_cycles, total_packets);
629 "\tavg processing cycles per packet: "
630 "%.02f (%"PRIu64"/%llu)\n",
631 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
632 cycles[PMD_CYCLES_PROCESSING], total_packets);
636 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
637 struct dp_netdev_pmd_thread *pmd,
638 unsigned long long stats[DP_N_STATS],
639 uint64_t cycles[PMD_N_CYCLES])
643 /* We cannot write 'stats' and 'cycles' (because they're written by other
644 * threads) and we shouldn't change 'stats' (because they're used to count
645 * datapath stats, which must not be cleared here). Instead, we save the
646 * current values and subtract them from the values to be displayed in the
648 for (i = 0; i < DP_N_STATS; i++) {
649 pmd->stats_zero[i] = stats[i];
651 for (i = 0; i < PMD_N_CYCLES; i++) {
652 pmd->cycles_zero[i] = cycles[i];
657 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
660 struct ds reply = DS_EMPTY_INITIALIZER;
661 struct dp_netdev_pmd_thread *pmd;
662 struct dp_netdev *dp = NULL;
663 enum pmd_info_type type = *(enum pmd_info_type *) aux;
665 ovs_mutex_lock(&dp_netdev_mutex);
668 dp = shash_find_data(&dp_netdevs, argv[1]);
669 } else if (shash_count(&dp_netdevs) == 1) {
670 /* There's only one datapath */
671 dp = shash_first(&dp_netdevs)->data;
675 ovs_mutex_unlock(&dp_netdev_mutex);
676 unixctl_command_reply_error(conn,
677 "please specify an existing datapath");
681 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
682 unsigned long long stats[DP_N_STATS];
683 uint64_t cycles[PMD_N_CYCLES];
686 /* Read current stats and cycle counters */
687 for (i = 0; i < ARRAY_SIZE(stats); i++) {
688 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
690 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
691 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
694 if (type == PMD_INFO_CLEAR_STATS) {
695 pmd_info_clear_stats(&reply, pmd, stats, cycles);
696 } else if (type == PMD_INFO_SHOW_STATS) {
697 pmd_info_show_stats(&reply, pmd, stats, cycles);
701 ovs_mutex_unlock(&dp_netdev_mutex);
703 unixctl_command_reply(conn, ds_cstr(&reply));
708 dpif_netdev_init(void)
710 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
711 clear_aux = PMD_INFO_CLEAR_STATS;
713 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
714 0, 1, dpif_netdev_pmd_info,
716 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
717 0, 1, dpif_netdev_pmd_info,
723 dpif_netdev_enumerate(struct sset *all_dps,
724 const struct dpif_class *dpif_class)
726 struct shash_node *node;
728 ovs_mutex_lock(&dp_netdev_mutex);
729 SHASH_FOR_EACH(node, &dp_netdevs) {
730 struct dp_netdev *dp = node->data;
731 if (dpif_class != dp->class) {
732 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
733 * If the class doesn't match, skip this dpif. */
736 sset_add(all_dps, node->name);
738 ovs_mutex_unlock(&dp_netdev_mutex);
744 dpif_netdev_class_is_dummy(const struct dpif_class *class)
746 return class != &dpif_netdev_class;
750 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
752 return strcmp(type, "internal") ? type
753 : dpif_netdev_class_is_dummy(class) ? "dummy"
758 create_dpif_netdev(struct dp_netdev *dp)
760 uint16_t netflow_id = hash_string(dp->name, 0);
761 struct dpif_netdev *dpif;
763 ovs_refcount_ref(&dp->ref_cnt);
765 dpif = xmalloc(sizeof *dpif);
766 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
768 dpif->last_port_seq = seq_read(dp->port_seq);
773 /* Choose an unused, non-zero port number and return it on success.
774 * Return ODPP_NONE on failure. */
776 choose_port(struct dp_netdev *dp, const char *name)
777 OVS_REQUIRES(dp->port_mutex)
781 if (dp->class != &dpif_netdev_class) {
785 /* If the port name begins with "br", start the number search at
786 * 100 to make writing tests easier. */
787 if (!strncmp(name, "br", 2)) {
791 /* If the port name contains a number, try to assign that port number.
792 * This can make writing unit tests easier because port numbers are
794 for (p = name; *p != '\0'; p++) {
795 if (isdigit((unsigned char) *p)) {
796 port_no = start_no + strtol(p, NULL, 10);
797 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
798 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
799 return u32_to_odp(port_no);
806 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
807 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
808 return u32_to_odp(port_no);
816 create_dp_netdev(const char *name, const struct dpif_class *class,
817 struct dp_netdev **dpp)
818 OVS_REQUIRES(dp_netdev_mutex)
820 struct dp_netdev *dp;
823 dp = xzalloc(sizeof *dp);
824 shash_add(&dp_netdevs, name, dp);
826 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
827 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
828 ovs_refcount_init(&dp->ref_cnt);
829 atomic_flag_clear(&dp->destroyed);
831 ovs_mutex_init(&dp->port_mutex);
832 cmap_init(&dp->ports);
833 dp->port_seq = seq_create();
834 fat_rwlock_init(&dp->upcall_rwlock);
836 /* Disable upcalls by default. */
837 dp_netdev_disable_upcall(dp);
838 dp->upcall_aux = NULL;
839 dp->upcall_cb = NULL;
841 cmap_init(&dp->poll_threads);
842 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
843 ovsthread_key_create(&dp->per_pmd_key, NULL);
845 dp_netdev_set_nonpmd(dp);
846 dp->n_dpdk_rxqs = NR_QUEUE;
848 ovs_mutex_lock(&dp->port_mutex);
849 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
850 ovs_mutex_unlock(&dp->port_mutex);
856 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
862 dpif_netdev_open(const struct dpif_class *class, const char *name,
863 bool create, struct dpif **dpifp)
865 struct dp_netdev *dp;
868 ovs_mutex_lock(&dp_netdev_mutex);
869 dp = shash_find_data(&dp_netdevs, name);
871 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
873 error = (dp->class != class ? EINVAL
878 *dpifp = create_dpif_netdev(dp);
881 ovs_mutex_unlock(&dp_netdev_mutex);
887 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
888 OVS_NO_THREAD_SAFETY_ANALYSIS
890 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
891 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
893 /* Before freeing a lock we should release it */
894 fat_rwlock_unlock(&dp->upcall_rwlock);
895 fat_rwlock_destroy(&dp->upcall_rwlock);
898 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
899 * through the 'dp_netdevs' shash while freeing 'dp'. */
901 dp_netdev_free(struct dp_netdev *dp)
902 OVS_REQUIRES(dp_netdev_mutex)
904 struct dp_netdev_port *port;
906 shash_find_and_delete(&dp_netdevs, dp->name);
908 dp_netdev_destroy_all_pmds(dp);
909 cmap_destroy(&dp->poll_threads);
910 ovs_mutex_destroy(&dp->non_pmd_mutex);
911 ovsthread_key_delete(dp->per_pmd_key);
913 ovs_mutex_lock(&dp->port_mutex);
914 CMAP_FOR_EACH (port, node, &dp->ports) {
915 do_del_port(dp, port);
917 ovs_mutex_unlock(&dp->port_mutex);
919 seq_destroy(dp->port_seq);
920 cmap_destroy(&dp->ports);
922 /* Upcalls must be disabled at this point */
923 dp_netdev_destroy_upcall_lock(dp);
926 free(CONST_CAST(char *, dp->name));
931 dp_netdev_unref(struct dp_netdev *dp)
934 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
935 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
936 ovs_mutex_lock(&dp_netdev_mutex);
937 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
940 ovs_mutex_unlock(&dp_netdev_mutex);
945 dpif_netdev_close(struct dpif *dpif)
947 struct dp_netdev *dp = get_dp_netdev(dpif);
954 dpif_netdev_destroy(struct dpif *dpif)
956 struct dp_netdev *dp = get_dp_netdev(dpif);
958 if (!atomic_flag_test_and_set(&dp->destroyed)) {
959 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
960 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
968 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
969 * load/store semantics. While the increment is not atomic, the load and
970 * store operations are, making it impossible to read inconsistent values.
972 * This is used to update thread local stats counters. */
974 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
976 unsigned long long tmp;
978 atomic_read_relaxed(var, &tmp);
980 atomic_store_relaxed(var, tmp);
984 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
986 struct dp_netdev *dp = get_dp_netdev(dpif);
987 struct dp_netdev_pmd_thread *pmd;
989 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
990 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
991 unsigned long long n;
992 stats->n_flows += cmap_count(&pmd->flow_table);
994 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
996 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
998 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
999 stats->n_missed += n;
1000 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
1003 stats->n_masks = UINT32_MAX;
1004 stats->n_mask_hit = UINT64_MAX;
1010 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1014 if (pmd->core_id == NON_PMD_CORE_ID) {
1018 ovs_mutex_lock(&pmd->cond_mutex);
1019 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1020 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1021 ovs_mutex_unlock(&pmd->cond_mutex);
1024 /* Causes all pmd threads to reload its tx/rx devices.
1025 * Must be called after adding/removing ports. */
1027 dp_netdev_reload_pmds(struct dp_netdev *dp)
1029 struct dp_netdev_pmd_thread *pmd;
1031 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1032 dp_netdev_reload_pmd__(pmd);
1037 hash_port_no(odp_port_t port_no)
1039 return hash_int(odp_to_u32(port_no), 0);
1043 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1045 OVS_REQUIRES(dp->port_mutex)
1047 struct netdev_saved_flags *sf;
1048 struct dp_netdev_port *port;
1049 struct netdev *netdev;
1050 enum netdev_flags flags;
1051 const char *open_type;
1055 /* Reject devices already in 'dp'. */
1056 if (!get_port_by_name(dp, devname, &port)) {
1060 /* Open and validate network device. */
1061 open_type = dpif_netdev_port_open_type(dp->class, type);
1062 error = netdev_open(devname, open_type, &netdev);
1066 /* XXX reject non-Ethernet devices */
1068 netdev_get_flags(netdev, &flags);
1069 if (flags & NETDEV_LOOPBACK) {
1070 VLOG_ERR("%s: cannot add a loopback device", devname);
1071 netdev_close(netdev);
1075 if (netdev_is_pmd(netdev)) {
1076 int n_cores = ovs_numa_get_n_cores();
1078 if (n_cores == OVS_CORE_UNSPEC) {
1079 VLOG_ERR("%s, cannot get cpu core info", devname);
1082 /* There can only be ovs_numa_get_n_cores() pmd threads,
1083 * so creates a txq for each, and one extra for the non
1085 error = netdev_set_multiq(netdev, n_cores + 1, dp->n_dpdk_rxqs);
1086 if (error && (error != EOPNOTSUPP)) {
1087 VLOG_ERR("%s, cannot set multiq", devname);
1091 port = xzalloc(sizeof *port);
1092 port->port_no = port_no;
1093 port->netdev = netdev;
1094 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
1095 port->type = xstrdup(type);
1096 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1097 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1099 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
1100 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1101 devname, ovs_strerror(errno));
1102 netdev_close(netdev);
1110 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1112 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1113 netdev_rxq_close(port->rxq[i]);
1115 netdev_close(netdev);
1123 ovs_refcount_init(&port->ref_cnt);
1124 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1126 if (netdev_is_pmd(netdev)) {
1127 dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
1128 dp_netdev_reload_pmds(dp);
1130 seq_change(dp->port_seq);
1136 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1137 odp_port_t *port_nop)
1139 struct dp_netdev *dp = get_dp_netdev(dpif);
1140 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1141 const char *dpif_port;
1145 ovs_mutex_lock(&dp->port_mutex);
1146 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1147 if (*port_nop != ODPP_NONE) {
1148 port_no = *port_nop;
1149 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1151 port_no = choose_port(dp, dpif_port);
1152 error = port_no == ODPP_NONE ? EFBIG : 0;
1155 *port_nop = port_no;
1156 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1158 ovs_mutex_unlock(&dp->port_mutex);
1164 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1166 struct dp_netdev *dp = get_dp_netdev(dpif);
1169 ovs_mutex_lock(&dp->port_mutex);
1170 if (port_no == ODPP_LOCAL) {
1173 struct dp_netdev_port *port;
1175 error = get_port_by_number(dp, port_no, &port);
1177 do_del_port(dp, port);
1180 ovs_mutex_unlock(&dp->port_mutex);
1186 is_valid_port_number(odp_port_t port_no)
1188 return port_no != ODPP_NONE;
1191 static struct dp_netdev_port *
1192 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1194 struct dp_netdev_port *port;
1196 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1197 if (port->port_no == port_no) {
1205 get_port_by_number(struct dp_netdev *dp,
1206 odp_port_t port_no, struct dp_netdev_port **portp)
1208 if (!is_valid_port_number(port_no)) {
1212 *portp = dp_netdev_lookup_port(dp, port_no);
1213 return *portp ? 0 : ENOENT;
1218 port_ref(struct dp_netdev_port *port)
1221 ovs_refcount_ref(&port->ref_cnt);
1226 port_try_ref(struct dp_netdev_port *port)
1229 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
1236 port_unref(struct dp_netdev_port *port)
1238 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1239 int n_rxq = netdev_n_rxq(port->netdev);
1242 netdev_close(port->netdev);
1243 netdev_restore_flags(port->sf);
1245 for (i = 0; i < n_rxq; i++) {
1246 netdev_rxq_close(port->rxq[i]);
1255 get_port_by_name(struct dp_netdev *dp,
1256 const char *devname, struct dp_netdev_port **portp)
1257 OVS_REQUIRES(dp->port_mutex)
1259 struct dp_netdev_port *port;
1261 CMAP_FOR_EACH (port, node, &dp->ports) {
1262 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1271 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1273 struct dp_netdev_pmd_thread *pmd;
1276 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1277 if (pmd->numa_id == numa_id) {
1285 /* Returns 'true' if there is a port with pmd netdev and the netdev
1286 * is on numa node 'numa_id'. */
1288 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1290 struct dp_netdev_port *port;
1292 CMAP_FOR_EACH (port, node, &dp->ports) {
1293 if (netdev_is_pmd(port->netdev)
1294 && netdev_get_numa_id(port->netdev) == numa_id) {
1304 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1305 OVS_REQUIRES(dp->port_mutex)
1307 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1308 seq_change(dp->port_seq);
1309 if (netdev_is_pmd(port->netdev)) {
1310 int numa_id = netdev_get_numa_id(port->netdev);
1312 /* If there is no netdev on the numa node, deletes the pmd threads
1313 * for that numa. Else, just reloads the queues. */
1314 if (!has_pmd_port_for_numa(dp, numa_id)) {
1315 dp_netdev_del_pmds_on_numa(dp, numa_id);
1317 dp_netdev_reload_pmds(dp);
1324 answer_port_query(const struct dp_netdev_port *port,
1325 struct dpif_port *dpif_port)
1327 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1328 dpif_port->type = xstrdup(port->type);
1329 dpif_port->port_no = port->port_no;
1333 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1334 struct dpif_port *dpif_port)
1336 struct dp_netdev *dp = get_dp_netdev(dpif);
1337 struct dp_netdev_port *port;
1340 error = get_port_by_number(dp, port_no, &port);
1341 if (!error && dpif_port) {
1342 answer_port_query(port, dpif_port);
1349 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1350 struct dpif_port *dpif_port)
1352 struct dp_netdev *dp = get_dp_netdev(dpif);
1353 struct dp_netdev_port *port;
1356 ovs_mutex_lock(&dp->port_mutex);
1357 error = get_port_by_name(dp, devname, &port);
1358 if (!error && dpif_port) {
1359 answer_port_query(port, dpif_port);
1361 ovs_mutex_unlock(&dp->port_mutex);
1367 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1369 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1373 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1375 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1376 ovsrcu_postpone(dp_netdev_flow_free, flow);
1381 dp_netdev_flow_hash(const ovs_u128 *ufid)
1383 return ufid->u32[0];
1387 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1388 struct dp_netdev_flow *flow)
1389 OVS_REQUIRES(pmd->flow_mutex)
1391 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1393 dpcls_remove(&pmd->cls, &flow->cr);
1394 flow->cr.mask = NULL; /* Accessing rule's mask after this is not safe. */
1396 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1399 dp_netdev_flow_unref(flow);
1403 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1405 struct dp_netdev_flow *netdev_flow;
1407 ovs_mutex_lock(&pmd->flow_mutex);
1408 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1409 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1411 ovs_mutex_unlock(&pmd->flow_mutex);
1415 dpif_netdev_flow_flush(struct dpif *dpif)
1417 struct dp_netdev *dp = get_dp_netdev(dpif);
1418 struct dp_netdev_pmd_thread *pmd;
1420 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1421 dp_netdev_pmd_flow_flush(pmd);
1427 struct dp_netdev_port_state {
1428 struct cmap_position position;
1433 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1435 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1440 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1441 struct dpif_port *dpif_port)
1443 struct dp_netdev_port_state *state = state_;
1444 struct dp_netdev *dp = get_dp_netdev(dpif);
1445 struct cmap_node *node;
1448 node = cmap_next_position(&dp->ports, &state->position);
1450 struct dp_netdev_port *port;
1452 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1455 state->name = xstrdup(netdev_get_name(port->netdev));
1456 dpif_port->name = state->name;
1457 dpif_port->type = port->type;
1458 dpif_port->port_no = port->port_no;
1469 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1471 struct dp_netdev_port_state *state = state_;
1478 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1480 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1481 uint64_t new_port_seq;
1484 new_port_seq = seq_read(dpif->dp->port_seq);
1485 if (dpif->last_port_seq != new_port_seq) {
1486 dpif->last_port_seq = new_port_seq;
1496 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1498 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1500 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1503 static struct dp_netdev_flow *
1504 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1506 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1509 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1511 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1514 /* netdev_flow_key utilities.
1516 * netdev_flow_key is basically a miniflow. We use these functions
1517 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1518 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1520 * - Since we are dealing exclusively with miniflows created by
1521 * miniflow_extract(), if the map is different the miniflow is different.
1522 * Therefore we can be faster by comparing the map and the miniflow in a
1524 * - These functions can be inlined by the compiler.
1526 * The following assertions make sure that what we're doing with miniflow is
1529 BUILD_ASSERT_DECL(sizeof(struct miniflow) == 2 * sizeof(uint64_t));
1531 /* Given the number of bits set in miniflow's maps, returns the size of the
1532 * 'netdev_flow_key.mf' */
1533 static inline size_t
1534 netdev_flow_key_size(size_t flow_u64s)
1536 return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s);
1540 netdev_flow_key_equal(const struct netdev_flow_key *a,
1541 const struct netdev_flow_key *b)
1543 /* 'b->len' may be not set yet. */
1544 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1547 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1548 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1549 * generated by miniflow_extract. */
1551 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1552 const struct miniflow *mf)
1554 return !memcmp(&key->mf, mf, key->len);
1558 netdev_flow_key_clone(struct netdev_flow_key *dst,
1559 const struct netdev_flow_key *src)
1562 offsetof(struct netdev_flow_key, mf) + src->len);
1567 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1568 const struct flow *src)
1570 struct dp_packet packet;
1571 uint64_t buf_stub[512 / 8];
1573 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1574 pkt_metadata_from_flow(&packet.md, src);
1575 flow_compose(&packet, src);
1576 miniflow_extract(&packet, &dst->mf);
1577 dp_packet_uninit(&packet);
1579 dst->len = netdev_flow_key_size(miniflow_n_values(&dst->mf));
1580 dst->hash = 0; /* Not computed yet. */
1583 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1585 netdev_flow_mask_init(struct netdev_flow_key *mask,
1586 const struct match *match)
1588 const uint64_t *mask_u64 = (const uint64_t *) &match->wc.masks;
1589 uint64_t *dst = miniflow_values(&mask->mf);
1590 struct miniflow maps;
1595 /* Only check masks that make sense for the flow. */
1596 flow_wc_map(&match->flow, &maps);
1597 memset(&mask->mf, 0, sizeof mask->mf); /* Clear maps. */
1601 uint64_t rm1bit = rightmost_1bit(map);
1602 int i = raw_ctz(map);
1605 mask->mf.tnl_map |= rm1bit;
1606 *dst++ = mask_u64[i];
1607 hash = hash_add64(hash, mask_u64[i]);
1611 mask_u64 += FLOW_TNL_U64S;
1614 uint64_t rm1bit = rightmost_1bit(map);
1615 int i = raw_ctz(map);
1618 mask->mf.pkt_map |= rm1bit;
1619 *dst++ = mask_u64[i];
1620 hash = hash_add64(hash, mask_u64[i]);
1625 hash = hash_add64(hash, mask->mf.tnl_map);
1626 hash = hash_add64(hash, mask->mf.pkt_map);
1628 n = dst - miniflow_get_values(&mask->mf);
1630 mask->hash = hash_finish(hash, n * 8);
1631 mask->len = netdev_flow_key_size(n);
1634 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1636 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1637 const struct flow *flow,
1638 const struct netdev_flow_key *mask)
1640 uint64_t *dst_u64 = miniflow_values(&dst->mf);
1641 const uint64_t *mask_u64 = miniflow_get_values(&mask->mf);
1645 dst->len = mask->len;
1646 dst->mf = mask->mf; /* Copy maps. */
1648 FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf) {
1649 *dst_u64 = value & *mask_u64++;
1650 hash = hash_add64(hash, *dst_u64++);
1652 dst->hash = hash_finish(hash,
1653 (dst_u64 - miniflow_get_values(&dst->mf)) * 8);
1656 /* Iterate through netdev_flow_key TNL u64 values specified by 'MAPS'. */
1657 #define NETDEV_FLOW_KEY_FOR_EACH_IN_TNL_MAP(VALUE, KEY, MAPS) \
1658 MINIFLOW_FOR_EACH_IN_TNL_MAP(VALUE, &(KEY)->mf, MAPS)
1660 /* Iterate through netdev_flow_key PKT u64 values specified by 'MAPS'. */
1661 #define NETDEV_FLOW_KEY_FOR_EACH_IN_PKT_MAP(VALUE, KEY, MAPS) \
1662 MINIFLOW_FOR_EACH_IN_PKT_MAP(VALUE, &(KEY)->mf, MAPS)
1664 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1666 static inline uint32_t
1667 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1668 const struct netdev_flow_key *mask)
1670 const uint64_t *p = miniflow_get_values(&mask->mf);
1674 NETDEV_FLOW_KEY_FOR_EACH_IN_TNL_MAP(key_u64, key, mask->mf) {
1675 hash = hash_add64(hash, key_u64 & *p++);
1677 NETDEV_FLOW_KEY_FOR_EACH_IN_PKT_MAP(key_u64, key, mask->mf) {
1678 hash = hash_add64(hash, key_u64 & *p++);
1681 return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8);
1685 emc_entry_alive(struct emc_entry *ce)
1687 return ce->flow && !ce->flow->dead;
1691 emc_clear_entry(struct emc_entry *ce)
1694 dp_netdev_flow_unref(ce->flow);
1700 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1701 const struct netdev_flow_key *key)
1703 if (ce->flow != flow) {
1705 dp_netdev_flow_unref(ce->flow);
1708 if (dp_netdev_flow_ref(flow)) {
1715 netdev_flow_key_clone(&ce->key, key);
1720 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1721 struct dp_netdev_flow *flow)
1723 struct emc_entry *to_be_replaced = NULL;
1724 struct emc_entry *current_entry;
1726 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1727 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1728 /* We found the entry with the 'mf' miniflow */
1729 emc_change_entry(current_entry, flow, NULL);
1733 /* Replacement policy: put the flow in an empty (not alive) entry, or
1734 * in the first entry where it can be */
1736 || (emc_entry_alive(to_be_replaced)
1737 && !emc_entry_alive(current_entry))
1738 || current_entry->key.hash < to_be_replaced->key.hash) {
1739 to_be_replaced = current_entry;
1742 /* We didn't find the miniflow in the cache.
1743 * The 'to_be_replaced' entry is where the new flow will be stored */
1745 emc_change_entry(to_be_replaced, flow, key);
1748 static inline struct dp_netdev_flow *
1749 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1751 struct emc_entry *current_entry;
1753 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1754 if (current_entry->key.hash == key->hash
1755 && emc_entry_alive(current_entry)
1756 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1758 /* We found the entry with the 'key->mf' miniflow */
1759 return current_entry->flow;
1766 static struct dp_netdev_flow *
1767 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1768 const struct netdev_flow_key *key)
1770 struct dp_netdev_flow *netdev_flow;
1771 struct dpcls_rule *rule;
1773 dpcls_lookup(&pmd->cls, key, &rule, 1);
1774 netdev_flow = dp_netdev_flow_cast(rule);
1779 static struct dp_netdev_flow *
1780 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1781 const ovs_u128 *ufidp, const struct nlattr *key,
1784 struct dp_netdev_flow *netdev_flow;
1788 /* If a UFID is not provided, determine one based on the key. */
1789 if (!ufidp && key && key_len
1790 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1791 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1796 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1798 if (ovs_u128_equals(&netdev_flow->ufid, ufidp)) {
1808 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1809 struct dpif_flow_stats *stats)
1811 struct dp_netdev_flow *netdev_flow;
1812 unsigned long long n;
1816 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1818 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1819 stats->n_packets = n;
1820 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1822 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1824 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1825 stats->tcp_flags = flags;
1828 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1829 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1830 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1833 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1834 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1835 struct dpif_flow *flow, bool terse)
1838 memset(flow, 0, sizeof *flow);
1840 struct flow_wildcards wc;
1841 struct dp_netdev_actions *actions;
1843 struct odp_flow_key_parms odp_parms = {
1844 .flow = &netdev_flow->flow,
1846 .support = dp_netdev_support,
1849 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1852 offset = key_buf->size;
1853 flow->key = ofpbuf_tail(key_buf);
1854 odp_parms.odp_in_port = netdev_flow->flow.in_port.odp_port;
1855 odp_flow_key_from_flow(&odp_parms, key_buf);
1856 flow->key_len = key_buf->size - offset;
1859 offset = mask_buf->size;
1860 flow->mask = ofpbuf_tail(mask_buf);
1861 odp_parms.odp_in_port = wc.masks.in_port.odp_port;
1862 odp_parms.key_buf = key_buf;
1863 odp_flow_key_from_mask(&odp_parms, mask_buf);
1864 flow->mask_len = mask_buf->size - offset;
1867 actions = dp_netdev_flow_get_actions(netdev_flow);
1868 flow->actions = actions->actions;
1869 flow->actions_len = actions->size;
1872 flow->ufid = netdev_flow->ufid;
1873 flow->ufid_present = true;
1874 flow->pmd_id = netdev_flow->pmd_id;
1875 get_dpif_flow_stats(netdev_flow, &flow->stats);
1879 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1880 const struct nlattr *mask_key,
1881 uint32_t mask_key_len, const struct flow *flow,
1885 enum odp_key_fitness fitness;
1887 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, key, key_len,
1890 /* This should not happen: it indicates that
1891 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1892 * disagree on the acceptable form of a mask. Log the problem
1893 * as an error, with enough details to enable debugging. */
1894 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1896 if (!VLOG_DROP_ERR(&rl)) {
1900 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1902 VLOG_ERR("internal error parsing flow mask %s (%s)",
1903 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1910 enum mf_field_id id;
1911 /* No mask key, unwildcard everything except fields whose
1912 * prerequisities are not met. */
1913 memset(mask, 0x0, sizeof *mask);
1915 for (id = 0; id < MFF_N_IDS; ++id) {
1916 /* Skip registers and metadata. */
1917 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1918 && !(id >= MFF_XREG0 && id < MFF_XREG0 + FLOW_N_XREGS)
1919 && id != MFF_METADATA) {
1920 const struct mf_field *mf = mf_from_id(id);
1921 if (mf_are_prereqs_ok(mf, flow)) {
1922 mf_mask_field(mf, mask);
1928 /* Force unwildcard the in_port.
1930 * We need to do this even in the case where we unwildcard "everything"
1931 * above because "everything" only includes the 16-bit OpenFlow port number
1932 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1933 * port number mask->in_port.odp_port. */
1934 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1939 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1944 if (odp_flow_key_to_flow(key, key_len, flow)) {
1945 /* This should not happen: it indicates that odp_flow_key_from_flow()
1946 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1947 * flow. Log the problem as an error, with enough details to enable
1949 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1951 if (!VLOG_DROP_ERR(&rl)) {
1955 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1956 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1963 in_port = flow->in_port.odp_port;
1964 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1972 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1974 struct dp_netdev *dp = get_dp_netdev(dpif);
1975 struct dp_netdev_flow *netdev_flow;
1976 struct dp_netdev_pmd_thread *pmd;
1977 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1978 ? NON_PMD_CORE_ID : get->pmd_id;
1981 pmd = dp_netdev_get_pmd(dp, pmd_id);
1986 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1989 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1994 dp_netdev_pmd_unref(pmd);
2000 static struct dp_netdev_flow *
2001 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
2002 struct match *match, const ovs_u128 *ufid,
2003 const struct nlattr *actions, size_t actions_len)
2004 OVS_REQUIRES(pmd->flow_mutex)
2006 struct dp_netdev_flow *flow;
2007 struct netdev_flow_key mask;
2009 netdev_flow_mask_init(&mask, match);
2010 /* Make sure wc does not have metadata. */
2011 ovs_assert(!(mask.mf.pkt_map
2012 & (MINIFLOW_PKT_MAP(metadata) | MINIFLOW_PKT_MAP(regs))));
2014 /* Do not allocate extra space. */
2015 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
2016 memset(&flow->stats, 0, sizeof flow->stats);
2019 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
2020 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
2021 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
2022 ovs_refcount_init(&flow->ref_cnt);
2023 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
2025 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
2026 dpcls_insert(&pmd->cls, &flow->cr, &mask);
2028 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
2029 dp_netdev_flow_hash(&flow->ufid));
2031 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2033 struct ds ds = DS_EMPTY_INITIALIZER;
2035 match.flow = flow->flow;
2036 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2038 ds_put_cstr(&ds, "flow_add: ");
2039 odp_format_ufid(ufid, &ds);
2040 ds_put_cstr(&ds, " ");
2041 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2042 ds_put_cstr(&ds, ", actions:");
2043 format_odp_actions(&ds, actions, actions_len);
2045 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2054 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2056 struct dp_netdev *dp = get_dp_netdev(dpif);
2057 struct dp_netdev_flow *netdev_flow;
2058 struct netdev_flow_key key;
2059 struct dp_netdev_pmd_thread *pmd;
2062 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2063 ? NON_PMD_CORE_ID : put->pmd_id;
2066 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2070 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2071 put->mask, put->mask_len,
2072 &match.flow, &match.wc.masks);
2077 pmd = dp_netdev_get_pmd(dp, pmd_id);
2082 /* Must produce a netdev_flow_key for lookup.
2083 * This interface is no longer performance critical, since it is not used
2084 * for upcall processing any more. */
2085 netdev_flow_key_from_flow(&key, &match.flow);
2090 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2093 ovs_mutex_lock(&pmd->flow_mutex);
2094 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2096 if (put->flags & DPIF_FP_CREATE) {
2097 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2099 memset(put->stats, 0, sizeof *put->stats);
2101 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2111 if (put->flags & DPIF_FP_MODIFY
2112 && flow_equal(&match.flow, &netdev_flow->flow)) {
2113 struct dp_netdev_actions *new_actions;
2114 struct dp_netdev_actions *old_actions;
2116 new_actions = dp_netdev_actions_create(put->actions,
2119 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2120 ovsrcu_set(&netdev_flow->actions, new_actions);
2123 get_dpif_flow_stats(netdev_flow, put->stats);
2125 if (put->flags & DPIF_FP_ZERO_STATS) {
2126 /* XXX: The userspace datapath uses thread local statistics
2127 * (for flows), which should be updated only by the owning
2128 * thread. Since we cannot write on stats memory here,
2129 * we choose not to support this flag. Please note:
2130 * - This feature is currently used only by dpctl commands with
2132 * - Should the need arise, this operation can be implemented
2133 * by keeping a base value (to be update here) for each
2134 * counter, and subtracting it before outputting the stats */
2138 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2139 } else if (put->flags & DPIF_FP_CREATE) {
2142 /* Overlapping flow. */
2146 ovs_mutex_unlock(&pmd->flow_mutex);
2147 dp_netdev_pmd_unref(pmd);
2153 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2155 struct dp_netdev *dp = get_dp_netdev(dpif);
2156 struct dp_netdev_flow *netdev_flow;
2157 struct dp_netdev_pmd_thread *pmd;
2158 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2159 ? NON_PMD_CORE_ID : del->pmd_id;
2162 pmd = dp_netdev_get_pmd(dp, pmd_id);
2167 ovs_mutex_lock(&pmd->flow_mutex);
2168 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2172 get_dpif_flow_stats(netdev_flow, del->stats);
2174 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2178 ovs_mutex_unlock(&pmd->flow_mutex);
2179 dp_netdev_pmd_unref(pmd);
2184 struct dpif_netdev_flow_dump {
2185 struct dpif_flow_dump up;
2186 struct cmap_position poll_thread_pos;
2187 struct cmap_position flow_pos;
2188 struct dp_netdev_pmd_thread *cur_pmd;
2190 struct ovs_mutex mutex;
2193 static struct dpif_netdev_flow_dump *
2194 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2196 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2199 static struct dpif_flow_dump *
2200 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2202 struct dpif_netdev_flow_dump *dump;
2204 dump = xzalloc(sizeof *dump);
2205 dpif_flow_dump_init(&dump->up, dpif_);
2206 dump->up.terse = terse;
2207 ovs_mutex_init(&dump->mutex);
2213 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2215 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2217 ovs_mutex_destroy(&dump->mutex);
2222 struct dpif_netdev_flow_dump_thread {
2223 struct dpif_flow_dump_thread up;
2224 struct dpif_netdev_flow_dump *dump;
2225 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2226 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2229 static struct dpif_netdev_flow_dump_thread *
2230 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2232 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2235 static struct dpif_flow_dump_thread *
2236 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2238 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2239 struct dpif_netdev_flow_dump_thread *thread;
2241 thread = xmalloc(sizeof *thread);
2242 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2243 thread->dump = dump;
2248 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2250 struct dpif_netdev_flow_dump_thread *thread
2251 = dpif_netdev_flow_dump_thread_cast(thread_);
2257 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2258 struct dpif_flow *flows, int max_flows)
2260 struct dpif_netdev_flow_dump_thread *thread
2261 = dpif_netdev_flow_dump_thread_cast(thread_);
2262 struct dpif_netdev_flow_dump *dump = thread->dump;
2263 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2267 ovs_mutex_lock(&dump->mutex);
2268 if (!dump->status) {
2269 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2270 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2271 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2272 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2274 /* First call to dump_next(), extracts the first pmd thread.
2275 * If there is no pmd thread, returns immediately. */
2277 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2279 ovs_mutex_unlock(&dump->mutex);
2286 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2287 struct cmap_node *node;
2289 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2293 netdev_flows[n_flows] = CONTAINER_OF(node,
2294 struct dp_netdev_flow,
2297 /* When finishing dumping the current pmd thread, moves to
2299 if (n_flows < flow_limit) {
2300 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2301 dp_netdev_pmd_unref(pmd);
2302 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2308 /* Keeps the reference to next caller. */
2309 dump->cur_pmd = pmd;
2311 /* If the current dump is empty, do not exit the loop, since the
2312 * remaining pmds could have flows to be dumped. Just dumps again
2313 * on the new 'pmd'. */
2316 ovs_mutex_unlock(&dump->mutex);
2318 for (i = 0; i < n_flows; i++) {
2319 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2320 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2321 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2322 struct dpif_flow *f = &flows[i];
2323 struct ofpbuf key, mask;
2325 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2326 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2327 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2335 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2336 OVS_NO_THREAD_SAFETY_ANALYSIS
2338 struct dp_netdev *dp = get_dp_netdev(dpif);
2339 struct dp_netdev_pmd_thread *pmd;
2340 struct dp_packet *pp;
2342 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2343 dp_packet_size(execute->packet) > UINT16_MAX) {
2347 /* Tries finding the 'pmd'. If NULL is returned, that means
2348 * the current thread is a non-pmd thread and should use
2349 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2350 pmd = ovsthread_getspecific(dp->per_pmd_key);
2352 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2355 /* If the current thread is non-pmd thread, acquires
2356 * the 'non_pmd_mutex'. */
2357 if (pmd->core_id == NON_PMD_CORE_ID) {
2358 ovs_mutex_lock(&dp->non_pmd_mutex);
2359 ovs_mutex_lock(&dp->port_mutex);
2362 pp = execute->packet;
2363 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2364 execute->actions_len);
2365 if (pmd->core_id == NON_PMD_CORE_ID) {
2366 dp_netdev_pmd_unref(pmd);
2367 ovs_mutex_unlock(&dp->port_mutex);
2368 ovs_mutex_unlock(&dp->non_pmd_mutex);
2375 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2379 for (i = 0; i < n_ops; i++) {
2380 struct dpif_op *op = ops[i];
2383 case DPIF_OP_FLOW_PUT:
2384 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2387 case DPIF_OP_FLOW_DEL:
2388 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2391 case DPIF_OP_EXECUTE:
2392 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2395 case DPIF_OP_FLOW_GET:
2396 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2402 /* Returns true if the configuration for rx queues or cpu mask
2405 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2407 if (dp->n_dpdk_rxqs != rxqs) {
2410 if (dp->pmd_cmask != NULL && cmask != NULL) {
2411 return strcmp(dp->pmd_cmask, cmask);
2413 return (dp->pmd_cmask != NULL || cmask != NULL);
2418 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2420 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2422 struct dp_netdev *dp = get_dp_netdev(dpif);
2424 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2425 struct dp_netdev_port *port;
2427 dp_netdev_destroy_all_pmds(dp);
2429 CMAP_FOR_EACH (port, node, &dp->ports) {
2430 if (netdev_is_pmd(port->netdev)) {
2433 /* Closes the existing 'rxq's. */
2434 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2435 netdev_rxq_close(port->rxq[i]);
2436 port->rxq[i] = NULL;
2439 /* Sets the new rx queue config. */
2440 err = netdev_set_multiq(port->netdev,
2441 ovs_numa_get_n_cores() + 1,
2443 if (err && (err != EOPNOTSUPP)) {
2444 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2445 " %u", netdev_get_name(port->netdev),
2450 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2451 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2452 * netdev_n_rxq(port->netdev));
2453 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2454 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2458 dp->n_dpdk_rxqs = n_rxqs;
2460 /* Reconfigures the cpu mask. */
2461 ovs_numa_set_cpu_mask(cmask);
2462 free(dp->pmd_cmask);
2463 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2465 /* Restores the non-pmd. */
2466 dp_netdev_set_nonpmd(dp);
2467 /* Restores all pmd threads. */
2468 dp_netdev_reset_pmd_threads(dp);
2475 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2476 uint32_t queue_id, uint32_t *priority)
2478 *priority = queue_id;
2483 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2484 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2485 struct dp_netdev_actions *
2486 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2488 struct dp_netdev_actions *netdev_actions;
2490 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2491 memcpy(netdev_actions->actions, actions, size);
2492 netdev_actions->size = size;
2494 return netdev_actions;
2497 struct dp_netdev_actions *
2498 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2500 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2504 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2509 static inline unsigned long long
2510 cycles_counter(void)
2513 return rte_get_tsc_cycles();
2519 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2520 extern struct ovs_mutex cycles_counter_fake_mutex;
2522 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2524 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2525 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2526 OVS_NO_THREAD_SAFETY_ANALYSIS
2528 pmd->last_cycles = cycles_counter();
2531 /* Stop counting cycles and add them to the counter 'type' */
2533 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2534 enum pmd_cycles_counter_type type)
2535 OVS_RELEASES(&cycles_counter_fake_mutex)
2536 OVS_NO_THREAD_SAFETY_ANALYSIS
2538 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2540 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2544 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2545 struct dp_netdev_port *port,
2546 struct netdev_rxq *rxq)
2548 struct dp_packet *packets[NETDEV_MAX_BURST];
2551 cycles_count_start(pmd);
2552 error = netdev_rxq_recv(rxq, packets, &cnt);
2553 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2557 *recirc_depth_get() = 0;
2559 /* XXX: initialize md in netdev implementation. */
2560 for (i = 0; i < cnt; i++) {
2561 pkt_metadata_init(&packets[i]->md, port->port_no);
2563 cycles_count_start(pmd);
2564 dp_netdev_input(pmd, packets, cnt);
2565 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2566 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2567 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2569 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2570 netdev_get_name(port->netdev), ovs_strerror(error));
2574 /* Return true if needs to revalidate datapath flows. */
2576 dpif_netdev_run(struct dpif *dpif)
2578 struct dp_netdev_port *port;
2579 struct dp_netdev *dp = get_dp_netdev(dpif);
2580 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2582 uint64_t new_tnl_seq;
2584 ovs_mutex_lock(&dp->non_pmd_mutex);
2585 CMAP_FOR_EACH (port, node, &dp->ports) {
2586 if (!netdev_is_pmd(port->netdev)) {
2589 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2590 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2594 ovs_mutex_unlock(&dp->non_pmd_mutex);
2595 dp_netdev_pmd_unref(non_pmd);
2597 tnl_arp_cache_run();
2598 new_tnl_seq = seq_read(tnl_conf_seq);
2600 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2601 dp->last_tnl_conf_seq = new_tnl_seq;
2608 dpif_netdev_wait(struct dpif *dpif)
2610 struct dp_netdev_port *port;
2611 struct dp_netdev *dp = get_dp_netdev(dpif);
2613 ovs_mutex_lock(&dp_netdev_mutex);
2614 CMAP_FOR_EACH (port, node, &dp->ports) {
2615 if (!netdev_is_pmd(port->netdev)) {
2618 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2619 netdev_rxq_wait(port->rxq[i]);
2623 ovs_mutex_unlock(&dp_netdev_mutex);
2624 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2628 struct dp_netdev_port *port;
2629 struct netdev_rxq *rx;
2633 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2634 struct rxq_poll **ppoll_list, int poll_cnt)
2636 struct rxq_poll *poll_list = *ppoll_list;
2637 struct dp_netdev_port *port;
2638 int n_pmds_on_numa, index, i;
2640 /* Simple scheduler for netdev rx polling. */
2641 for (i = 0; i < poll_cnt; i++) {
2642 port_unref(poll_list[i].port);
2646 n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
2649 CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
2650 /* Calls port_try_ref() to prevent the main thread
2651 * from deleting the port. */
2652 if (port_try_ref(port)) {
2653 if (netdev_is_pmd(port->netdev)
2654 && netdev_get_numa_id(port->netdev) == pmd->numa_id) {
2657 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2658 if ((index % n_pmds_on_numa) == pmd->index) {
2659 poll_list = xrealloc(poll_list,
2660 sizeof *poll_list * (poll_cnt + 1));
2663 poll_list[poll_cnt].port = port;
2664 poll_list[poll_cnt].rx = port->rxq[i];
2670 /* Unrefs the port_try_ref(). */
2675 *ppoll_list = poll_list;
2680 pmd_thread_main(void *f_)
2682 struct dp_netdev_pmd_thread *pmd = f_;
2683 unsigned int lc = 0;
2684 struct rxq_poll *poll_list;
2685 unsigned int port_seq = PMD_INITIAL_SEQ;
2692 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2693 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2694 pmd_thread_setaffinity_cpu(pmd->core_id);
2696 emc_cache_init(&pmd->flow_cache);
2697 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2699 /* List port/core affinity */
2700 for (i = 0; i < poll_cnt; i++) {
2701 VLOG_INFO("Core %d processing port \'%s\'\n", pmd->core_id, netdev_get_name(poll_list[i].port->netdev));
2704 /* Signal here to make sure the pmd finishes
2705 * reloading the updated configuration. */
2706 dp_netdev_pmd_reload_done(pmd);
2711 for (i = 0; i < poll_cnt; i++) {
2712 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2720 emc_cache_slow_sweep(&pmd->flow_cache);
2723 atomic_read_relaxed(&pmd->change_seq, &seq);
2724 if (seq != port_seq) {
2731 emc_cache_uninit(&pmd->flow_cache);
2733 if (!latch_is_set(&pmd->exit_latch)){
2737 for (i = 0; i < poll_cnt; i++) {
2738 port_unref(poll_list[i].port);
2741 dp_netdev_pmd_reload_done(pmd);
2748 dp_netdev_disable_upcall(struct dp_netdev *dp)
2749 OVS_ACQUIRES(dp->upcall_rwlock)
2751 fat_rwlock_wrlock(&dp->upcall_rwlock);
2755 dpif_netdev_disable_upcall(struct dpif *dpif)
2756 OVS_NO_THREAD_SAFETY_ANALYSIS
2758 struct dp_netdev *dp = get_dp_netdev(dpif);
2759 dp_netdev_disable_upcall(dp);
2763 dp_netdev_enable_upcall(struct dp_netdev *dp)
2764 OVS_RELEASES(dp->upcall_rwlock)
2766 fat_rwlock_unlock(&dp->upcall_rwlock);
2770 dpif_netdev_enable_upcall(struct dpif *dpif)
2771 OVS_NO_THREAD_SAFETY_ANALYSIS
2773 struct dp_netdev *dp = get_dp_netdev(dpif);
2774 dp_netdev_enable_upcall(dp);
2778 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2780 ovs_mutex_lock(&pmd->cond_mutex);
2781 xpthread_cond_signal(&pmd->cond);
2782 ovs_mutex_unlock(&pmd->cond_mutex);
2785 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2786 * the pointer if succeeds, otherwise, NULL.
2788 * Caller must unrefs the returned reference. */
2789 static struct dp_netdev_pmd_thread *
2790 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2792 struct dp_netdev_pmd_thread *pmd;
2793 const struct cmap_node *pnode;
2795 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2799 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2801 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2804 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2806 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2808 struct dp_netdev_pmd_thread *non_pmd;
2810 non_pmd = xzalloc(sizeof *non_pmd);
2811 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2815 /* Caller must have valid pointer to 'pmd'. */
2817 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2819 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2823 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2825 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2826 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2830 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2831 * fails, keeps checking for next node until reaching the end of cmap.
2833 * Caller must unrefs the returned reference. */
2834 static struct dp_netdev_pmd_thread *
2835 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2837 struct dp_netdev_pmd_thread *next;
2840 struct cmap_node *node;
2842 node = cmap_next_position(&dp->poll_threads, pos);
2843 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2845 } while (next && !dp_netdev_pmd_try_ref(next));
2851 core_id_to_qid(unsigned core_id)
2853 if (core_id != NON_PMD_CORE_ID) {
2856 return ovs_numa_get_n_cores();
2860 /* Configures the 'pmd' based on the input argument. */
2862 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2863 int index, unsigned core_id, int numa_id)
2867 pmd->core_id = core_id;
2868 pmd->tx_qid = core_id_to_qid(core_id);
2869 pmd->numa_id = numa_id;
2871 ovs_refcount_init(&pmd->ref_cnt);
2872 latch_init(&pmd->exit_latch);
2873 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2874 xpthread_cond_init(&pmd->cond, NULL);
2875 ovs_mutex_init(&pmd->cond_mutex);
2876 ovs_mutex_init(&pmd->flow_mutex);
2877 dpcls_init(&pmd->cls);
2878 cmap_init(&pmd->flow_table);
2879 /* init the 'flow_cache' since there is no
2880 * actual thread created for NON_PMD_CORE_ID. */
2881 if (core_id == NON_PMD_CORE_ID) {
2882 emc_cache_init(&pmd->flow_cache);
2884 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2885 hash_int(core_id, 0));
2889 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2891 dp_netdev_pmd_flow_flush(pmd);
2892 dpcls_destroy(&pmd->cls);
2893 cmap_destroy(&pmd->flow_table);
2894 ovs_mutex_destroy(&pmd->flow_mutex);
2895 latch_destroy(&pmd->exit_latch);
2896 xpthread_cond_destroy(&pmd->cond);
2897 ovs_mutex_destroy(&pmd->cond_mutex);
2901 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2902 * and unrefs the struct. */
2904 dp_netdev_del_pmd(struct dp_netdev_pmd_thread *pmd)
2906 /* Uninit the 'flow_cache' since there is
2907 * no actual thread uninit it for NON_PMD_CORE_ID. */
2908 if (pmd->core_id == NON_PMD_CORE_ID) {
2909 emc_cache_uninit(&pmd->flow_cache);
2911 latch_set(&pmd->exit_latch);
2912 dp_netdev_reload_pmd__(pmd);
2913 ovs_numa_unpin_core(pmd->core_id);
2914 xpthread_join(pmd->thread, NULL);
2916 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2917 dp_netdev_pmd_unref(pmd);
2920 /* Destroys all pmd threads. */
2922 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2924 struct dp_netdev_pmd_thread *pmd;
2926 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2927 dp_netdev_del_pmd(pmd);
2931 /* Deletes all pmd threads on numa node 'numa_id'. */
2933 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2935 struct dp_netdev_pmd_thread *pmd;
2937 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2938 if (pmd->numa_id == numa_id) {
2939 dp_netdev_del_pmd(pmd);
2944 /* Checks the numa node id of 'netdev' and starts pmd threads for
2947 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2951 if (!ovs_numa_numa_id_is_valid(numa_id)) {
2952 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2953 "invalid", numa_id);
2957 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
2959 /* If there are already pmd threads created for the numa node
2960 * in which 'netdev' is on, do nothing. Else, creates the
2961 * pmd threads for the numa node. */
2963 int can_have, n_unpinned, i;
2965 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
2967 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2968 "cores on numa node");
2972 /* If cpu mask is specified, uses all unpinned cores, otherwise
2973 * tries creating NR_PMD_THREADS pmd threads. */
2974 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
2975 for (i = 0; i < can_have; i++) {
2976 struct dp_netdev_pmd_thread *pmd = xzalloc(sizeof *pmd);
2977 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
2979 dp_netdev_configure_pmd(pmd, dp, i, core_id, numa_id);
2980 /* Each thread will distribute all devices rx-queues among
2982 pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd);
2984 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
2989 /* Called after pmd threads config change. Restarts pmd threads with
2990 * new configuration. */
2992 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
2994 struct dp_netdev_port *port;
2996 CMAP_FOR_EACH (port, node, &dp->ports) {
2997 if (netdev_is_pmd(port->netdev)) {
2998 int numa_id = netdev_get_numa_id(port->netdev);
3000 dp_netdev_set_pmds_on_numa(dp, numa_id);
3006 dpif_netdev_get_datapath_version(void)
3008 return xstrdup("<built-in>");
3012 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
3013 uint16_t tcp_flags, long long now)
3017 atomic_store_relaxed(&netdev_flow->stats.used, now);
3018 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
3019 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
3020 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
3022 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
3026 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
3027 enum dp_stat_type type, int cnt)
3029 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3033 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3034 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3035 enum dpif_upcall_type type, const struct nlattr *userdata,
3036 struct ofpbuf *actions, struct ofpbuf *put_actions)
3038 struct dp_netdev *dp = pmd->dp;
3040 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3044 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3045 struct ds ds = DS_EMPTY_INITIALIZER;
3048 struct odp_flow_key_parms odp_parms = {
3051 .odp_in_port = flow->in_port.odp_port,
3052 .support = dp_netdev_support,
3055 ofpbuf_init(&key, 0);
3056 odp_flow_key_from_flow(&odp_parms, &key);
3057 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3058 dp_packet_size(packet_));
3060 odp_flow_key_format(key.data, key.size, &ds);
3062 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3063 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3065 ofpbuf_uninit(&key);
3071 return dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3072 actions, wc, put_actions, dp->upcall_aux);
3075 static inline uint32_t
3076 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3077 const struct miniflow *mf)
3079 uint32_t hash, recirc_depth;
3081 hash = dp_packet_get_rss_hash(packet);
3082 if (OVS_UNLIKELY(!hash)) {
3083 hash = miniflow_hash_5tuple(mf, 0);
3084 dp_packet_set_rss_hash(packet, hash);
3087 /* The RSS hash must account for the recirculation depth to avoid
3088 * collisions in the exact match cache */
3089 recirc_depth = *recirc_depth_get_unsafe();
3090 if (OVS_UNLIKELY(recirc_depth)) {
3091 hash = hash_finish(hash, recirc_depth);
3092 dp_packet_set_rss_hash(packet, hash);
3097 struct packet_batch {
3098 unsigned int packet_count;
3099 unsigned int byte_count;
3102 struct dp_netdev_flow *flow;
3104 struct dp_packet *packets[NETDEV_MAX_BURST];
3108 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
3109 const struct miniflow *mf)
3111 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3112 batch->packets[batch->packet_count++] = packet;
3113 batch->byte_count += dp_packet_size(packet);
3117 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
3119 flow->batch = batch;
3122 batch->packet_count = 0;
3123 batch->byte_count = 0;
3124 batch->tcp_flags = 0;
3128 packet_batch_execute(struct packet_batch *batch,
3129 struct dp_netdev_pmd_thread *pmd,
3132 struct dp_netdev_actions *actions;
3133 struct dp_netdev_flow *flow = batch->flow;
3135 dp_netdev_flow_used(flow, batch->packet_count, batch->byte_count,
3136 batch->tcp_flags, now);
3138 actions = dp_netdev_flow_get_actions(flow);
3140 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
3141 actions->actions, actions->size);
3145 dp_netdev_queue_batches(struct dp_packet *pkt,
3146 struct dp_netdev_flow *flow, const struct miniflow *mf,
3147 struct packet_batch *batches, size_t *n_batches)
3149 struct packet_batch *batch = flow->batch;
3151 if (OVS_LIKELY(batch)) {
3152 packet_batch_update(batch, pkt, mf);
3156 batch = &batches[(*n_batches)++];
3157 packet_batch_init(batch, flow);
3158 packet_batch_update(batch, pkt, mf);
3162 dp_packet_swap(struct dp_packet **a, struct dp_packet **b)
3164 struct dp_packet *tmp = *a;
3169 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3170 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', the
3171 * miniflow is copied into 'keys' and the packet pointer is moved at the
3172 * beginning of the 'packets' array.
3174 * The function returns the number of packets that needs to be processed in the
3175 * 'packets' array (they have been moved to the beginning of the vector).
3177 static inline size_t
3178 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
3179 size_t cnt, struct netdev_flow_key *keys,
3180 struct packet_batch batches[], size_t *n_batches)
3182 struct emc_cache *flow_cache = &pmd->flow_cache;
3183 struct netdev_flow_key key;
3184 size_t i, notfound_cnt = 0;
3186 for (i = 0; i < cnt; i++) {
3187 struct dp_netdev_flow *flow;
3189 if (OVS_UNLIKELY(dp_packet_size(packets[i]) < ETH_HEADER_LEN)) {
3190 dp_packet_delete(packets[i]);
3195 /* Prefetch next packet data */
3196 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3199 miniflow_extract(packets[i], &key.mf);
3200 key.len = 0; /* Not computed yet. */
3201 key.hash = dpif_netdev_packet_get_rss_hash(packets[i], &key.mf);
3203 flow = emc_lookup(flow_cache, &key);
3204 if (OVS_LIKELY(flow)) {
3205 dp_netdev_queue_batches(packets[i], flow, &key.mf, batches,
3208 if (i != notfound_cnt) {
3209 dp_packet_swap(&packets[i], &packets[notfound_cnt]);
3212 keys[notfound_cnt++] = key;
3216 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - notfound_cnt);
3218 return notfound_cnt;
3222 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3223 struct dp_packet **packets, size_t cnt,
3224 struct netdev_flow_key *keys,
3225 struct packet_batch batches[], size_t *n_batches)
3227 #if !defined(__CHECKER__) && !defined(_WIN32)
3228 const size_t PKT_ARRAY_SIZE = cnt;
3230 /* Sparse or MSVC doesn't like variable length array. */
3231 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3233 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3234 struct dp_netdev *dp = pmd->dp;
3235 struct emc_cache *flow_cache = &pmd->flow_cache;
3236 int miss_cnt = 0, lost_cnt = 0;
3240 for (i = 0; i < cnt; i++) {
3241 /* Key length is needed in all the cases, hash computed on demand. */
3242 keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf));
3244 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3245 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3246 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3247 struct ofpbuf actions, put_actions;
3250 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3251 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3253 for (i = 0; i < cnt; i++) {
3254 struct dp_netdev_flow *netdev_flow;
3255 struct ofpbuf *add_actions;
3259 if (OVS_LIKELY(rules[i])) {
3263 /* It's possible that an earlier slow path execution installed
3264 * a rule covering this flow. In this case, it's a lot cheaper
3265 * to catch it here than execute a miss. */
3266 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3268 rules[i] = &netdev_flow->cr;
3274 miniflow_expand(&keys[i].mf, &match.flow);
3276 ofpbuf_clear(&actions);
3277 ofpbuf_clear(&put_actions);
3279 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
3280 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
3281 &ufid, DPIF_UC_MISS, NULL, &actions,
3283 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3284 dp_packet_delete(packets[i]);
3289 /* We can't allow the packet batching in the next loop to execute
3290 * the actions. Otherwise, if there are any slow path actions,
3291 * we'll send the packet up twice. */
3292 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
3293 actions.data, actions.size);
3295 add_actions = put_actions.size ? &put_actions : &actions;
3296 if (OVS_LIKELY(error != ENOSPC)) {
3297 /* XXX: There's a race window where a flow covering this packet
3298 * could have already been installed since we last did the flow
3299 * lookup before upcall. This could be solved by moving the
3300 * mutex lock outside the loop, but that's an awful long time
3301 * to be locking everyone out of making flow installs. If we
3302 * move to a per-core classifier, it would be reasonable. */
3303 ovs_mutex_lock(&pmd->flow_mutex);
3304 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3305 if (OVS_LIKELY(!netdev_flow)) {
3306 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3310 ovs_mutex_unlock(&pmd->flow_mutex);
3312 emc_insert(flow_cache, &keys[i], netdev_flow);
3316 ofpbuf_uninit(&actions);
3317 ofpbuf_uninit(&put_actions);
3318 fat_rwlock_unlock(&dp->upcall_rwlock);
3319 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3320 } else if (OVS_UNLIKELY(any_miss)) {
3321 for (i = 0; i < cnt; i++) {
3322 if (OVS_UNLIKELY(!rules[i])) {
3323 dp_packet_delete(packets[i]);
3330 for (i = 0; i < cnt; i++) {
3331 struct dp_packet *packet = packets[i];
3332 struct dp_netdev_flow *flow;
3334 if (OVS_UNLIKELY(!rules[i])) {
3338 flow = dp_netdev_flow_cast(rules[i]);
3340 emc_insert(flow_cache, &keys[i], flow);
3341 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3344 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3345 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3346 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3350 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3351 struct dp_packet **packets, int cnt)
3353 #if !defined(__CHECKER__) && !defined(_WIN32)
3354 const size_t PKT_ARRAY_SIZE = cnt;
3356 /* Sparse or MSVC doesn't like variable length array. */
3357 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3359 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3360 struct packet_batch batches[PKT_ARRAY_SIZE];
3361 long long now = time_msec();
3362 size_t newcnt, n_batches, i;
3365 newcnt = emc_processing(pmd, packets, cnt, keys, batches, &n_batches);
3366 if (OVS_UNLIKELY(newcnt)) {
3367 fast_path_processing(pmd, packets, newcnt, keys, batches, &n_batches);
3370 for (i = 0; i < n_batches; i++) {
3371 batches[i].flow->batch = NULL;
3374 for (i = 0; i < n_batches; i++) {
3375 packet_batch_execute(&batches[i], pmd, now);
3379 struct dp_netdev_execute_aux {
3380 struct dp_netdev_pmd_thread *pmd;
3384 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3387 struct dp_netdev *dp = get_dp_netdev(dpif);
3388 dp->upcall_aux = aux;
3393 dp_netdev_drop_packets(struct dp_packet **packets, int cnt, bool may_steal)
3398 for (i = 0; i < cnt; i++) {
3399 dp_packet_delete(packets[i]);
3405 push_tnl_action(const struct dp_netdev *dp,
3406 const struct nlattr *attr,
3407 struct dp_packet **packets, int cnt)
3409 struct dp_netdev_port *tun_port;
3410 const struct ovs_action_push_tnl *data;
3412 data = nl_attr_get(attr);
3414 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3418 netdev_push_header(tun_port->netdev, packets, cnt, data);
3424 dp_netdev_clone_pkt_batch(struct dp_packet **dst_pkts,
3425 struct dp_packet **src_pkts, int cnt)
3429 for (i = 0; i < cnt; i++) {
3430 dst_pkts[i] = dp_packet_clone(src_pkts[i]);
3435 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3436 const struct nlattr *a, bool may_steal)
3437 OVS_NO_THREAD_SAFETY_ANALYSIS
3439 struct dp_netdev_execute_aux *aux = aux_;
3440 uint32_t *depth = recirc_depth_get();
3441 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3442 struct dp_netdev *dp = pmd->dp;
3443 int type = nl_attr_type(a);
3444 struct dp_netdev_port *p;
3447 switch ((enum ovs_action_attr)type) {
3448 case OVS_ACTION_ATTR_OUTPUT:
3449 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3450 if (OVS_LIKELY(p)) {
3451 netdev_send(p->netdev, pmd->tx_qid, packets, cnt, may_steal);
3456 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3457 if (*depth < MAX_RECIRC_DEPTH) {
3458 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3462 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3466 err = push_tnl_action(dp, a, packets, cnt);
3469 dp_netdev_input(pmd, packets, cnt);
3472 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3478 case OVS_ACTION_ATTR_TUNNEL_POP:
3479 if (*depth < MAX_RECIRC_DEPTH) {
3480 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3482 p = dp_netdev_lookup_port(dp, portno);
3484 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3488 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3492 err = netdev_pop_header(p->netdev, packets, cnt);
3495 for (i = 0; i < cnt; i++) {
3496 packets[i]->md.in_port.odp_port = portno;
3500 dp_netdev_input(pmd, packets, cnt);
3503 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3510 case OVS_ACTION_ATTR_USERSPACE:
3511 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3512 const struct nlattr *userdata;
3513 struct ofpbuf actions;
3517 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3518 ofpbuf_init(&actions, 0);
3520 for (i = 0; i < cnt; i++) {
3523 ofpbuf_clear(&actions);
3525 flow_extract(packets[i], &flow);
3526 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3527 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3528 DPIF_UC_ACTION, userdata,&actions,
3530 if (!error || error == ENOSPC) {
3531 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3532 actions.data, actions.size);
3533 } else if (may_steal) {
3534 dp_packet_delete(packets[i]);
3537 ofpbuf_uninit(&actions);
3538 fat_rwlock_unlock(&dp->upcall_rwlock);
3544 case OVS_ACTION_ATTR_RECIRC:
3545 if (*depth < MAX_RECIRC_DEPTH) {
3546 struct dp_packet *recirc_pkts[NETDEV_MAX_BURST];
3549 dp_netdev_clone_pkt_batch(recirc_pkts, packets, cnt);
3550 packets = recirc_pkts;
3553 for (i = 0; i < cnt; i++) {
3554 packets[i]->md.recirc_id = nl_attr_get_u32(a);
3558 dp_netdev_input(pmd, packets, cnt);
3564 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3567 case OVS_ACTION_ATTR_PUSH_VLAN:
3568 case OVS_ACTION_ATTR_POP_VLAN:
3569 case OVS_ACTION_ATTR_PUSH_MPLS:
3570 case OVS_ACTION_ATTR_POP_MPLS:
3571 case OVS_ACTION_ATTR_SET:
3572 case OVS_ACTION_ATTR_SET_MASKED:
3573 case OVS_ACTION_ATTR_SAMPLE:
3574 case OVS_ACTION_ATTR_HASH:
3575 case OVS_ACTION_ATTR_UNSPEC:
3576 case __OVS_ACTION_ATTR_MAX:
3580 dp_netdev_drop_packets(packets, cnt, may_steal);
3584 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3585 struct dp_packet **packets, int cnt,
3587 const struct nlattr *actions, size_t actions_len)
3589 struct dp_netdev_execute_aux aux = { pmd };
3591 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3592 actions_len, dp_execute_cb);
3595 const struct dpif_class dpif_netdev_class = {
3598 dpif_netdev_enumerate,
3599 dpif_netdev_port_open_type,
3602 dpif_netdev_destroy,
3605 dpif_netdev_get_stats,
3606 dpif_netdev_port_add,
3607 dpif_netdev_port_del,
3608 dpif_netdev_port_query_by_number,
3609 dpif_netdev_port_query_by_name,
3610 NULL, /* port_get_pid */
3611 dpif_netdev_port_dump_start,
3612 dpif_netdev_port_dump_next,
3613 dpif_netdev_port_dump_done,
3614 dpif_netdev_port_poll,
3615 dpif_netdev_port_poll_wait,
3616 dpif_netdev_flow_flush,
3617 dpif_netdev_flow_dump_create,
3618 dpif_netdev_flow_dump_destroy,
3619 dpif_netdev_flow_dump_thread_create,
3620 dpif_netdev_flow_dump_thread_destroy,
3621 dpif_netdev_flow_dump_next,
3622 dpif_netdev_operate,
3623 NULL, /* recv_set */
3624 NULL, /* handlers_set */
3625 dpif_netdev_pmd_set,
3626 dpif_netdev_queue_to_priority,
3628 NULL, /* recv_wait */
3629 NULL, /* recv_purge */
3630 dpif_netdev_register_upcall_cb,
3631 dpif_netdev_enable_upcall,
3632 dpif_netdev_disable_upcall,
3633 dpif_netdev_get_datapath_version,
3637 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3638 const char *argv[], void *aux OVS_UNUSED)
3640 struct dp_netdev_port *old_port;
3641 struct dp_netdev_port *new_port;
3642 struct dp_netdev *dp;
3645 ovs_mutex_lock(&dp_netdev_mutex);
3646 dp = shash_find_data(&dp_netdevs, argv[1]);
3647 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3648 ovs_mutex_unlock(&dp_netdev_mutex);
3649 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3652 ovs_refcount_ref(&dp->ref_cnt);
3653 ovs_mutex_unlock(&dp_netdev_mutex);
3655 ovs_mutex_lock(&dp->port_mutex);
3656 if (get_port_by_name(dp, argv[2], &old_port)) {
3657 unixctl_command_reply_error(conn, "unknown port");
3661 port_no = u32_to_odp(atoi(argv[3]));
3662 if (!port_no || port_no == ODPP_NONE) {
3663 unixctl_command_reply_error(conn, "bad port number");
3666 if (dp_netdev_lookup_port(dp, port_no)) {
3667 unixctl_command_reply_error(conn, "port number already in use");
3671 /* Remove old port. */
3672 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3673 ovsrcu_postpone(free, old_port);
3675 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3676 new_port = xmemdup(old_port, sizeof *old_port);
3677 new_port->port_no = port_no;
3678 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3680 seq_change(dp->port_seq);
3681 unixctl_command_reply(conn, NULL);
3684 ovs_mutex_unlock(&dp->port_mutex);
3685 dp_netdev_unref(dp);
3689 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3690 const char *argv[], void *aux OVS_UNUSED)
3692 struct dp_netdev_port *port;
3693 struct dp_netdev *dp;
3695 ovs_mutex_lock(&dp_netdev_mutex);
3696 dp = shash_find_data(&dp_netdevs, argv[1]);
3697 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3698 ovs_mutex_unlock(&dp_netdev_mutex);
3699 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3702 ovs_refcount_ref(&dp->ref_cnt);
3703 ovs_mutex_unlock(&dp_netdev_mutex);
3705 ovs_mutex_lock(&dp->port_mutex);
3706 if (get_port_by_name(dp, argv[2], &port)) {
3707 unixctl_command_reply_error(conn, "unknown port");
3708 } else if (port->port_no == ODPP_LOCAL) {
3709 unixctl_command_reply_error(conn, "can't delete local port");
3711 do_del_port(dp, port);
3712 unixctl_command_reply(conn, NULL);
3714 ovs_mutex_unlock(&dp->port_mutex);
3716 dp_netdev_unref(dp);
3720 dpif_dummy_register__(const char *type)
3722 struct dpif_class *class;
3724 class = xmalloc(sizeof *class);
3725 *class = dpif_netdev_class;
3726 class->type = xstrdup(type);
3727 dp_register_provider(class);
3731 dpif_dummy_override(const char *type)
3733 if (!dp_unregister_provider(type)) {
3734 dpif_dummy_register__(type);
3739 dpif_dummy_register(enum dummy_level level)
3741 if (level == DUMMY_OVERRIDE_ALL) {
3746 dp_enumerate_types(&types);
3747 SSET_FOR_EACH (type, &types) {
3748 dpif_dummy_override(type);
3750 sset_destroy(&types);
3751 } else if (level == DUMMY_OVERRIDE_SYSTEM) {
3752 dpif_dummy_override("system");
3755 dpif_dummy_register__("dummy");
3757 unixctl_command_register("dpif-dummy/change-port-number",
3758 "dp port new-number",
3759 3, 3, dpif_dummy_change_port_number, NULL);
3760 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3761 2, 2, dpif_dummy_delete_port, NULL);
3764 /* Datapath Classifier. */
3766 /* A set of rules that all have the same fields wildcarded. */
3767 struct dpcls_subtable {
3768 /* The fields are only used by writers. */
3769 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3771 /* These fields are accessed by readers. */
3772 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3773 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3774 /* 'mask' must be the last field, additional space is allocated here. */
3777 /* Initializes 'cls' as a classifier that initially contains no classification
3780 dpcls_init(struct dpcls *cls)
3782 cmap_init(&cls->subtables_map);
3783 pvector_init(&cls->subtables);
3787 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3789 pvector_remove(&cls->subtables, subtable);
3790 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3791 subtable->mask.hash);
3792 cmap_destroy(&subtable->rules);
3793 ovsrcu_postpone(free, subtable);
3796 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3797 * caller's responsibility.
3798 * May only be called after all the readers have been terminated. */
3800 dpcls_destroy(struct dpcls *cls)
3803 struct dpcls_subtable *subtable;
3805 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3806 ovs_assert(cmap_count(&subtable->rules) == 0);
3807 dpcls_destroy_subtable(cls, subtable);
3809 cmap_destroy(&cls->subtables_map);
3810 pvector_destroy(&cls->subtables);
3814 static struct dpcls_subtable *
3815 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3817 struct dpcls_subtable *subtable;
3819 /* Need to add one. */
3820 subtable = xmalloc(sizeof *subtable
3821 - sizeof subtable->mask.mf + mask->len);
3822 cmap_init(&subtable->rules);
3823 netdev_flow_key_clone(&subtable->mask, mask);
3824 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3825 pvector_insert(&cls->subtables, subtable, 0);
3826 pvector_publish(&cls->subtables);
3831 static inline struct dpcls_subtable *
3832 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3834 struct dpcls_subtable *subtable;
3836 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3837 &cls->subtables_map) {
3838 if (netdev_flow_key_equal(&subtable->mask, mask)) {
3842 return dpcls_create_subtable(cls, mask);
3845 /* Insert 'rule' into 'cls'. */
3847 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
3848 const struct netdev_flow_key *mask)
3850 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
3852 rule->mask = &subtable->mask;
3853 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
3856 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3858 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
3860 struct dpcls_subtable *subtable;
3862 ovs_assert(rule->mask);
3864 INIT_CONTAINER(subtable, rule->mask, mask);
3866 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
3868 dpcls_destroy_subtable(cls, subtable);
3869 pvector_publish(&cls->subtables);
3873 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
3874 * in 'mask' the values in 'key' and 'target' are the same. */
3876 dpcls_rule_matches_key(const struct dpcls_rule *rule,
3877 const struct netdev_flow_key *target)
3879 const uint64_t *keyp = miniflow_get_values(&rule->flow.mf);
3880 const uint64_t *maskp = miniflow_get_values(&rule->mask->mf);
3881 uint64_t target_u64;
3883 NETDEV_FLOW_KEY_FOR_EACH_IN_TNL_MAP(target_u64, target, rule->flow.mf) {
3884 if (OVS_UNLIKELY((target_u64 & *maskp++) != *keyp++)) {
3888 NETDEV_FLOW_KEY_FOR_EACH_IN_PKT_MAP(target_u64, target, rule->flow.mf) {
3889 if (OVS_UNLIKELY((target_u64 & *maskp++) != *keyp++)) {
3896 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3897 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3898 * NULL it is skipped.
3900 * This function is optimized for use in the userspace datapath and therefore
3901 * does not implement a lot of features available in the standard
3902 * classifier_lookup() function. Specifically, it does not implement
3903 * priorities, instead returning any rule which matches the flow.
3905 * Returns true if all flows found a corresponding rule. */
3907 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
3908 struct dpcls_rule **rules, const size_t cnt)
3910 /* The batch size 16 was experimentally found faster than 8 or 32. */
3911 typedef uint16_t map_type;
3912 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3914 #if !defined(__CHECKER__) && !defined(_WIN32)
3915 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
3917 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
3919 map_type maps[N_MAPS];
3920 struct dpcls_subtable *subtable;
3922 memset(maps, 0xff, sizeof maps);
3923 if (cnt % MAP_BITS) {
3924 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
3926 memset(rules, 0, cnt * sizeof *rules);
3928 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
3929 const struct netdev_flow_key *mkeys = keys;
3930 struct dpcls_rule **mrules = rules;
3931 map_type remains = 0;
3934 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
3936 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
3937 uint32_t hashes[MAP_BITS];
3938 const struct cmap_node *nodes[MAP_BITS];
3939 unsigned long map = maps[m];
3943 continue; /* Skip empty maps. */
3946 /* Compute hashes for the remaining keys. */
3947 ULLONG_FOR_EACH_1(i, map) {
3948 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
3952 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
3953 /* Check results. */
3954 ULLONG_FOR_EACH_1(i, map) {
3955 struct dpcls_rule *rule;
3957 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
3958 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
3963 ULLONG_SET0(map, i); /* Did not match. */
3965 ; /* Keep Sparse happy. */
3967 maps[m] &= ~map; /* Clear the found rules. */
3971 return true; /* All found. */
3974 return false; /* Some misses. */