2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2016 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "dpif-netdev.h"
24 #include <netinet/in.h>
25 #include <sys/socket.h>
30 #include <sys/ioctl.h>
37 #include "dp-packet.h"
39 #include "dpif-provider.h"
41 #include "dynamic-string.h"
42 #include "fat-rwlock.h"
50 #include "netdev-dpdk.h"
51 #include "netdev-vport.h"
53 #include "odp-execute.h"
55 #include "ofp-print.h"
60 #include "poll-loop.h"
67 #include "tnl-neigh-cache.h"
68 #include "tnl-ports.h"
71 #include "openvswitch/vlog.h"
73 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
75 #define FLOW_DUMP_MAX_BATCH 50
76 /* Use per thread recirc_depth to prevent recirculation loop. */
77 #define MAX_RECIRC_DEPTH 5
78 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
80 /* Configuration parameters. */
81 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
83 /* Protects against changes to 'dp_netdevs'. */
84 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
86 /* Contains all 'struct dp_netdev's. */
87 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
88 = SHASH_INITIALIZER(&dp_netdevs);
90 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
92 static struct odp_support dp_netdev_support = {
93 .max_mpls_depth = SIZE_MAX,
97 /* Stores a miniflow with inline values */
99 struct netdev_flow_key {
100 uint32_t hash; /* Hash function differs for different users. */
101 uint32_t len; /* Length of the following miniflow (incl. map). */
103 uint64_t buf[FLOW_MAX_PACKET_U64S];
106 /* Exact match cache for frequently used flows
108 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
109 * search its entries for a miniflow that matches exactly the miniflow of the
110 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
112 * A cache entry holds a reference to its 'dp_netdev_flow'.
114 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
115 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
116 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
117 * value is the index of a cache entry where the miniflow could be.
123 * Each pmd_thread has its own private exact match cache.
124 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
127 #define EM_FLOW_HASH_SHIFT 13
128 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
129 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
130 #define EM_FLOW_HASH_SEGS 2
133 struct dp_netdev_flow *flow;
134 struct netdev_flow_key key; /* key.hash used for emc hash value. */
138 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
139 int sweep_idx; /* For emc_cache_slow_sweep(). */
142 /* Iterate in the exact match cache through every entry that might contain a
143 * miniflow with hash 'HASH'. */
144 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
145 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
146 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
147 i__ < EM_FLOW_HASH_SEGS; \
148 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
150 /* Simple non-wildcarding single-priority classifier. */
153 struct cmap subtables_map;
154 struct pvector subtables;
157 /* A rule to be inserted to the classifier. */
159 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
160 struct netdev_flow_key *mask; /* Subtable's mask. */
161 struct netdev_flow_key flow; /* Matching key. */
162 /* 'flow' must be the last field, additional space is allocated here. */
165 static void dpcls_init(struct dpcls *);
166 static void dpcls_destroy(struct dpcls *);
167 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
168 const struct netdev_flow_key *mask);
169 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
170 static bool dpcls_lookup(const struct dpcls *cls,
171 const struct netdev_flow_key keys[],
172 struct dpcls_rule **rules, size_t cnt);
174 /* Datapath based on the network device interface from netdev.h.
180 * Some members, marked 'const', are immutable. Accessing other members
181 * requires synchronization, as noted in more detail below.
183 * Acquisition order is, from outermost to innermost:
185 * dp_netdev_mutex (global)
189 const struct dpif_class *const class;
190 const char *const name;
192 struct ovs_refcount ref_cnt;
193 atomic_flag destroyed;
197 * Protected by RCU. Take the mutex to add or remove ports. */
198 struct ovs_mutex port_mutex;
200 struct seq *port_seq; /* Incremented whenever a port changes. */
202 /* Protects access to ofproto-dpif-upcall interface during revalidator
203 * thread synchronization. */
204 struct fat_rwlock upcall_rwlock;
205 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
208 /* Callback function for notifying the purging of dp flows (during
209 * reseting pmd deletion). */
210 dp_purge_callback *dp_purge_cb;
213 /* Stores all 'struct dp_netdev_pmd_thread's. */
214 struct cmap poll_threads;
216 /* Protects the access of the 'struct dp_netdev_pmd_thread'
217 * instance for non-pmd thread. */
218 struct ovs_mutex non_pmd_mutex;
220 /* Each pmd thread will store its pointer to
221 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
222 ovsthread_key_t per_pmd_key;
224 /* Number of rx queues for each dpdk interface and the cpu mask
225 * for pin of pmd threads. */
228 uint64_t last_tnl_conf_seq;
231 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
235 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
236 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
237 DP_STAT_MISS, /* Packets that did not match. */
238 DP_STAT_LOST, /* Packets not passed up to the client. */
242 enum pmd_cycles_counter_type {
243 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
244 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
248 /* A port in a netdev-based datapath. */
249 struct dp_netdev_port {
251 struct netdev *netdev;
252 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
253 struct netdev_saved_flags *sf;
254 struct netdev_rxq **rxq;
255 struct ovs_refcount ref_cnt;
256 char *type; /* Port type as requested by user. */
259 /* Contained by struct dp_netdev_flow's 'stats' member. */
260 struct dp_netdev_flow_stats {
261 atomic_llong used; /* Last used time, in monotonic msecs. */
262 atomic_ullong packet_count; /* Number of packets matched. */
263 atomic_ullong byte_count; /* Number of bytes matched. */
264 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
267 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
273 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
274 * its pmd thread's classifier. The text below calls this classifier 'cls'.
279 * The thread safety rules described here for "struct dp_netdev_flow" are
280 * motivated by two goals:
282 * - Prevent threads that read members of "struct dp_netdev_flow" from
283 * reading bad data due to changes by some thread concurrently modifying
286 * - Prevent two threads making changes to members of a given "struct
287 * dp_netdev_flow" from interfering with each other.
293 * A flow 'flow' may be accessed without a risk of being freed during an RCU
294 * grace period. Code that needs to hold onto a flow for a while
295 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
297 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
298 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
301 * Some members, marked 'const', are immutable. Accessing other members
302 * requires synchronization, as noted in more detail below.
304 struct dp_netdev_flow {
305 const struct flow flow; /* Unmasked flow that created this entry. */
306 /* Hash table index by unmasked flow. */
307 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
309 const ovs_u128 ufid; /* Unique flow identifier. */
310 const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */
313 /* Number of references.
314 * The classifier owns one reference.
315 * Any thread trying to keep a rule from being freed should hold its own
317 struct ovs_refcount ref_cnt;
322 struct dp_netdev_flow_stats stats;
325 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
327 /* While processing a group of input packets, the datapath uses the next
328 * member to store a pointer to the output batch for the flow. It is
329 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
330 * packet_batch_init() and packet_batch_execute()). */
331 struct packet_batch *batch;
333 /* Packet classification. */
334 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
335 /* 'cr' must be the last member. */
338 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
339 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
340 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
343 /* A set of datapath actions within a "struct dp_netdev_flow".
349 * A struct dp_netdev_actions 'actions' is protected with RCU. */
350 struct dp_netdev_actions {
351 /* These members are immutable: they do not change during the struct's
353 unsigned int size; /* Size of 'actions', in bytes. */
354 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
357 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
359 struct dp_netdev_actions *dp_netdev_flow_get_actions(
360 const struct dp_netdev_flow *);
361 static void dp_netdev_actions_free(struct dp_netdev_actions *);
363 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
364 struct dp_netdev_pmd_stats {
365 /* Indexed by DP_STAT_*. */
366 atomic_ullong n[DP_N_STATS];
369 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
370 struct dp_netdev_pmd_cycles {
371 /* Indexed by PMD_CYCLES_*. */
372 atomic_ullong n[PMD_N_CYCLES];
375 /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */
377 struct dp_netdev_port *port;
378 struct netdev_rxq *rx;
379 struct ovs_list node;
382 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
383 * the performance overhead of interrupt processing. Therefore netdev can
384 * not implement rx-wait for these devices. dpif-netdev needs to poll
385 * these device to check for recv buffer. pmd-thread does polling for
386 * devices assigned to itself.
388 * DPDK used PMD for accessing NIC.
390 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
391 * I/O of all non-pmd threads. There will be no actual thread created
394 * Each struct has its own flow table and classifier. Packets received
395 * from managed ports are looked up in the corresponding pmd thread's
396 * flow table, and are executed with the found actions.
398 struct dp_netdev_pmd_thread {
399 struct dp_netdev *dp;
400 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
401 struct cmap_node node; /* In 'dp->poll_threads'. */
403 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
404 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
406 /* Per thread exact-match cache. Note, the instance for cpu core
407 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
408 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
409 * instances will only be accessed by its own pmd thread. */
410 struct emc_cache flow_cache;
412 /* Classifier and Flow-Table.
414 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
415 * changes to 'cls' must be made while still holding the 'flow_mutex'.
417 struct ovs_mutex flow_mutex;
419 struct cmap flow_table OVS_GUARDED; /* Flow table. */
422 struct dp_netdev_pmd_stats stats;
424 /* Cycles counters */
425 struct dp_netdev_pmd_cycles cycles;
427 /* Used to count cicles. See 'cycles_counter_end()' */
428 unsigned long long last_cycles;
430 struct latch exit_latch; /* For terminating the pmd thread. */
431 atomic_uint change_seq; /* For reloading pmd ports. */
433 int index; /* Idx of this pmd thread among pmd*/
434 /* threads on same numa node. */
435 unsigned core_id; /* CPU core id of this pmd thread. */
436 int numa_id; /* numa node id of this pmd thread. */
437 atomic_int tx_qid; /* Queue id used by this pmd thread to
438 * send packets on all netdevs */
440 struct ovs_mutex poll_mutex; /* Mutex for poll_list. */
441 /* List of rx queues to poll. */
442 struct ovs_list poll_list OVS_GUARDED;
443 int poll_cnt; /* Number of elemints in poll_list. */
445 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
446 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
447 * values and subtracts them from 'stats' and 'cycles' before
448 * reporting to the user */
449 unsigned long long stats_zero[DP_N_STATS];
450 uint64_t cycles_zero[PMD_N_CYCLES];
453 #define PMD_INITIAL_SEQ 1
455 /* Interface to netdev-based datapath. */
458 struct dp_netdev *dp;
459 uint64_t last_port_seq;
462 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
463 struct dp_netdev_port **portp);
464 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
465 struct dp_netdev_port **portp);
466 static void dp_netdev_free(struct dp_netdev *)
467 OVS_REQUIRES(dp_netdev_mutex);
468 static int do_add_port(struct dp_netdev *dp, const char *devname,
469 const char *type, odp_port_t port_no)
470 OVS_REQUIRES(dp->port_mutex);
471 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
472 OVS_REQUIRES(dp->port_mutex);
473 static int dpif_netdev_open(const struct dpif_class *, const char *name,
474 bool create, struct dpif **);
475 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
476 struct dp_packet **, int c,
478 const struct nlattr *actions,
480 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
481 struct dp_packet **, int cnt);
483 static void dp_netdev_disable_upcall(struct dp_netdev *);
484 static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
485 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
486 struct dp_netdev *dp, int index,
487 unsigned core_id, int numa_id);
488 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
489 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
490 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
492 static struct dp_netdev_pmd_thread *
493 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
494 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
495 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
496 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
498 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
499 struct dp_netdev_port *port, struct netdev_rxq *rx);
500 static struct dp_netdev_pmd_thread *
501 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id);
502 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
503 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
504 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
505 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
507 static inline bool emc_entry_alive(struct emc_entry *ce);
508 static void emc_clear_entry(struct emc_entry *ce);
511 emc_cache_init(struct emc_cache *flow_cache)
515 flow_cache->sweep_idx = 0;
516 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
517 flow_cache->entries[i].flow = NULL;
518 flow_cache->entries[i].key.hash = 0;
519 flow_cache->entries[i].key.len = sizeof(struct miniflow);
520 flowmap_init(&flow_cache->entries[i].key.mf.map);
525 emc_cache_uninit(struct emc_cache *flow_cache)
529 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
530 emc_clear_entry(&flow_cache->entries[i]);
534 /* Check and clear dead flow references slowly (one entry at each
537 emc_cache_slow_sweep(struct emc_cache *flow_cache)
539 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
541 if (!emc_entry_alive(entry)) {
542 emc_clear_entry(entry);
544 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
547 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
549 dpif_is_netdev(const struct dpif *dpif)
551 return dpif->dpif_class->open == dpif_netdev_open;
554 static struct dpif_netdev *
555 dpif_netdev_cast(const struct dpif *dpif)
557 ovs_assert(dpif_is_netdev(dpif));
558 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
561 static struct dp_netdev *
562 get_dp_netdev(const struct dpif *dpif)
564 return dpif_netdev_cast(dpif)->dp;
568 PMD_INFO_SHOW_STATS, /* show how cpu cycles are spent */
569 PMD_INFO_CLEAR_STATS /* set the cycles count to 0 */
573 pmd_info_show_stats(struct ds *reply,
574 struct dp_netdev_pmd_thread *pmd,
575 unsigned long long stats[DP_N_STATS],
576 uint64_t cycles[PMD_N_CYCLES])
578 unsigned long long total_packets = 0;
579 uint64_t total_cycles = 0;
582 /* These loops subtracts reference values ('*_zero') from the counters.
583 * Since loads and stores are relaxed, it might be possible for a '*_zero'
584 * value to be more recent than the current value we're reading from the
585 * counter. This is not a big problem, since these numbers are not
586 * supposed to be too accurate, but we should at least make sure that
587 * the result is not negative. */
588 for (i = 0; i < DP_N_STATS; i++) {
589 if (stats[i] > pmd->stats_zero[i]) {
590 stats[i] -= pmd->stats_zero[i];
595 if (i != DP_STAT_LOST) {
596 /* Lost packets are already included in DP_STAT_MISS */
597 total_packets += stats[i];
601 for (i = 0; i < PMD_N_CYCLES; i++) {
602 if (cycles[i] > pmd->cycles_zero[i]) {
603 cycles[i] -= pmd->cycles_zero[i];
608 total_cycles += cycles[i];
611 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
612 ? "main thread" : "pmd thread");
614 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
615 ds_put_format(reply, " numa_id %d", pmd->numa_id);
617 if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) {
618 ds_put_format(reply, " core_id %u", pmd->core_id);
620 ds_put_cstr(reply, ":\n");
623 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
624 "\tmiss:%llu\n\tlost:%llu\n",
625 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
626 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
628 if (total_cycles == 0) {
633 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
634 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
635 cycles[PMD_CYCLES_POLLING],
636 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
637 cycles[PMD_CYCLES_PROCESSING],
638 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
640 if (total_packets == 0) {
645 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
646 total_cycles / (double)total_packets,
647 total_cycles, total_packets);
650 "\tavg processing cycles per packet: "
651 "%.02f (%"PRIu64"/%llu)\n",
652 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
653 cycles[PMD_CYCLES_PROCESSING], total_packets);
657 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
658 struct dp_netdev_pmd_thread *pmd,
659 unsigned long long stats[DP_N_STATS],
660 uint64_t cycles[PMD_N_CYCLES])
664 /* We cannot write 'stats' and 'cycles' (because they're written by other
665 * threads) and we shouldn't change 'stats' (because they're used to count
666 * datapath stats, which must not be cleared here). Instead, we save the
667 * current values and subtract them from the values to be displayed in the
669 for (i = 0; i < DP_N_STATS; i++) {
670 pmd->stats_zero[i] = stats[i];
672 for (i = 0; i < PMD_N_CYCLES; i++) {
673 pmd->cycles_zero[i] = cycles[i];
678 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
681 struct ds reply = DS_EMPTY_INITIALIZER;
682 struct dp_netdev_pmd_thread *pmd;
683 struct dp_netdev *dp = NULL;
684 enum pmd_info_type type = *(enum pmd_info_type *) aux;
686 ovs_mutex_lock(&dp_netdev_mutex);
689 dp = shash_find_data(&dp_netdevs, argv[1]);
690 } else if (shash_count(&dp_netdevs) == 1) {
691 /* There's only one datapath */
692 dp = shash_first(&dp_netdevs)->data;
696 ovs_mutex_unlock(&dp_netdev_mutex);
697 unixctl_command_reply_error(conn,
698 "please specify an existing datapath");
702 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
703 unsigned long long stats[DP_N_STATS];
704 uint64_t cycles[PMD_N_CYCLES];
707 /* Read current stats and cycle counters */
708 for (i = 0; i < ARRAY_SIZE(stats); i++) {
709 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
711 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
712 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
715 if (type == PMD_INFO_CLEAR_STATS) {
716 pmd_info_clear_stats(&reply, pmd, stats, cycles);
717 } else if (type == PMD_INFO_SHOW_STATS) {
718 pmd_info_show_stats(&reply, pmd, stats, cycles);
722 ovs_mutex_unlock(&dp_netdev_mutex);
724 unixctl_command_reply(conn, ds_cstr(&reply));
729 dpif_netdev_init(void)
731 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
732 clear_aux = PMD_INFO_CLEAR_STATS;
734 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
735 0, 1, dpif_netdev_pmd_info,
737 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
738 0, 1, dpif_netdev_pmd_info,
744 dpif_netdev_enumerate(struct sset *all_dps,
745 const struct dpif_class *dpif_class)
747 struct shash_node *node;
749 ovs_mutex_lock(&dp_netdev_mutex);
750 SHASH_FOR_EACH(node, &dp_netdevs) {
751 struct dp_netdev *dp = node->data;
752 if (dpif_class != dp->class) {
753 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
754 * If the class doesn't match, skip this dpif. */
757 sset_add(all_dps, node->name);
759 ovs_mutex_unlock(&dp_netdev_mutex);
765 dpif_netdev_class_is_dummy(const struct dpif_class *class)
767 return class != &dpif_netdev_class;
771 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
773 return strcmp(type, "internal") ? type
774 : dpif_netdev_class_is_dummy(class) ? "dummy"
779 create_dpif_netdev(struct dp_netdev *dp)
781 uint16_t netflow_id = hash_string(dp->name, 0);
782 struct dpif_netdev *dpif;
784 ovs_refcount_ref(&dp->ref_cnt);
786 dpif = xmalloc(sizeof *dpif);
787 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
789 dpif->last_port_seq = seq_read(dp->port_seq);
794 /* Choose an unused, non-zero port number and return it on success.
795 * Return ODPP_NONE on failure. */
797 choose_port(struct dp_netdev *dp, const char *name)
798 OVS_REQUIRES(dp->port_mutex)
802 if (dp->class != &dpif_netdev_class) {
806 /* If the port name begins with "br", start the number search at
807 * 100 to make writing tests easier. */
808 if (!strncmp(name, "br", 2)) {
812 /* If the port name contains a number, try to assign that port number.
813 * This can make writing unit tests easier because port numbers are
815 for (p = name; *p != '\0'; p++) {
816 if (isdigit((unsigned char) *p)) {
817 port_no = start_no + strtol(p, NULL, 10);
818 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
819 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
820 return u32_to_odp(port_no);
827 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
828 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
829 return u32_to_odp(port_no);
837 create_dp_netdev(const char *name, const struct dpif_class *class,
838 struct dp_netdev **dpp)
839 OVS_REQUIRES(dp_netdev_mutex)
841 struct dp_netdev *dp;
844 dp = xzalloc(sizeof *dp);
845 shash_add(&dp_netdevs, name, dp);
847 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
848 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
849 ovs_refcount_init(&dp->ref_cnt);
850 atomic_flag_clear(&dp->destroyed);
852 ovs_mutex_init(&dp->port_mutex);
853 cmap_init(&dp->ports);
854 dp->port_seq = seq_create();
855 fat_rwlock_init(&dp->upcall_rwlock);
857 /* Disable upcalls by default. */
858 dp_netdev_disable_upcall(dp);
859 dp->upcall_aux = NULL;
860 dp->upcall_cb = NULL;
862 cmap_init(&dp->poll_threads);
863 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
864 ovsthread_key_create(&dp->per_pmd_key, NULL);
866 dp_netdev_set_nonpmd(dp);
867 dp->n_dpdk_rxqs = NR_QUEUE;
869 ovs_mutex_lock(&dp->port_mutex);
870 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
871 ovs_mutex_unlock(&dp->port_mutex);
877 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
883 dpif_netdev_open(const struct dpif_class *class, const char *name,
884 bool create, struct dpif **dpifp)
886 struct dp_netdev *dp;
889 ovs_mutex_lock(&dp_netdev_mutex);
890 dp = shash_find_data(&dp_netdevs, name);
892 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
894 error = (dp->class != class ? EINVAL
899 *dpifp = create_dpif_netdev(dp);
902 ovs_mutex_unlock(&dp_netdev_mutex);
908 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
909 OVS_NO_THREAD_SAFETY_ANALYSIS
911 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
912 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
914 /* Before freeing a lock we should release it */
915 fat_rwlock_unlock(&dp->upcall_rwlock);
916 fat_rwlock_destroy(&dp->upcall_rwlock);
919 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
920 * through the 'dp_netdevs' shash while freeing 'dp'. */
922 dp_netdev_free(struct dp_netdev *dp)
923 OVS_REQUIRES(dp_netdev_mutex)
925 struct dp_netdev_port *port;
927 shash_find_and_delete(&dp_netdevs, dp->name);
929 dp_netdev_destroy_all_pmds(dp);
930 cmap_destroy(&dp->poll_threads);
931 ovs_mutex_destroy(&dp->non_pmd_mutex);
932 ovsthread_key_delete(dp->per_pmd_key);
934 ovs_mutex_lock(&dp->port_mutex);
935 CMAP_FOR_EACH (port, node, &dp->ports) {
936 do_del_port(dp, port);
938 ovs_mutex_unlock(&dp->port_mutex);
940 seq_destroy(dp->port_seq);
941 cmap_destroy(&dp->ports);
943 /* Upcalls must be disabled at this point */
944 dp_netdev_destroy_upcall_lock(dp);
947 free(CONST_CAST(char *, dp->name));
952 dp_netdev_unref(struct dp_netdev *dp)
955 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
956 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
957 ovs_mutex_lock(&dp_netdev_mutex);
958 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
961 ovs_mutex_unlock(&dp_netdev_mutex);
966 dpif_netdev_close(struct dpif *dpif)
968 struct dp_netdev *dp = get_dp_netdev(dpif);
975 dpif_netdev_destroy(struct dpif *dpif)
977 struct dp_netdev *dp = get_dp_netdev(dpif);
979 if (!atomic_flag_test_and_set(&dp->destroyed)) {
980 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
981 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
989 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
990 * load/store semantics. While the increment is not atomic, the load and
991 * store operations are, making it impossible to read inconsistent values.
993 * This is used to update thread local stats counters. */
995 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
997 unsigned long long tmp;
999 atomic_read_relaxed(var, &tmp);
1001 atomic_store_relaxed(var, tmp);
1005 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
1007 struct dp_netdev *dp = get_dp_netdev(dpif);
1008 struct dp_netdev_pmd_thread *pmd;
1010 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
1011 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1012 unsigned long long n;
1013 stats->n_flows += cmap_count(&pmd->flow_table);
1015 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
1017 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
1019 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
1020 stats->n_missed += n;
1021 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
1024 stats->n_masks = UINT32_MAX;
1025 stats->n_mask_hit = UINT64_MAX;
1031 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1035 if (pmd->core_id == NON_PMD_CORE_ID) {
1039 ovs_mutex_lock(&pmd->cond_mutex);
1040 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1041 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1042 ovs_mutex_unlock(&pmd->cond_mutex);
1046 hash_port_no(odp_port_t port_no)
1048 return hash_int(odp_to_u32(port_no), 0);
1052 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1054 OVS_REQUIRES(dp->port_mutex)
1056 struct netdev_saved_flags *sf;
1057 struct dp_netdev_port *port;
1058 struct netdev *netdev;
1059 enum netdev_flags flags;
1060 const char *open_type;
1064 /* Reject devices already in 'dp'. */
1065 if (!get_port_by_name(dp, devname, &port)) {
1069 /* Open and validate network device. */
1070 open_type = dpif_netdev_port_open_type(dp->class, type);
1071 error = netdev_open(devname, open_type, &netdev);
1075 /* XXX reject non-Ethernet devices */
1077 netdev_get_flags(netdev, &flags);
1078 if (flags & NETDEV_LOOPBACK) {
1079 VLOG_ERR("%s: cannot add a loopback device", devname);
1080 netdev_close(netdev);
1084 if (netdev_is_pmd(netdev)) {
1085 int n_cores = ovs_numa_get_n_cores();
1087 if (n_cores == OVS_CORE_UNSPEC) {
1088 VLOG_ERR("%s, cannot get cpu core info", devname);
1091 /* There can only be ovs_numa_get_n_cores() pmd threads,
1092 * so creates a txq for each, and one extra for the non
1094 error = netdev_set_multiq(netdev, n_cores + 1, dp->n_dpdk_rxqs);
1095 if (error && (error != EOPNOTSUPP)) {
1096 VLOG_ERR("%s, cannot set multiq", devname);
1100 port = xzalloc(sizeof *port);
1101 port->port_no = port_no;
1102 port->netdev = netdev;
1103 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
1104 port->type = xstrdup(type);
1105 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1106 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1108 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
1109 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1110 devname, ovs_strerror(errno));
1111 netdev_close(netdev);
1119 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1121 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1122 netdev_rxq_close(port->rxq[i]);
1124 netdev_close(netdev);
1132 ovs_refcount_init(&port->ref_cnt);
1133 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1135 if (netdev_is_pmd(netdev)) {
1136 int numa_id = netdev_get_numa_id(netdev);
1137 struct dp_netdev_pmd_thread *pmd;
1139 /* Cannot create pmd threads for invalid numa node. */
1140 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
1142 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1143 pmd = dp_netdev_less_loaded_pmd_on_numa(dp, numa_id);
1145 /* There is no pmd threads on this numa node. */
1146 dp_netdev_set_pmds_on_numa(dp, numa_id);
1147 /* Assigning of rx queues done. */
1151 ovs_mutex_lock(&pmd->poll_mutex);
1152 dp_netdev_add_rxq_to_pmd(pmd, port, port->rxq[i]);
1153 ovs_mutex_unlock(&pmd->poll_mutex);
1154 dp_netdev_reload_pmd__(pmd);
1157 seq_change(dp->port_seq);
1163 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1164 odp_port_t *port_nop)
1166 struct dp_netdev *dp = get_dp_netdev(dpif);
1167 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1168 const char *dpif_port;
1172 ovs_mutex_lock(&dp->port_mutex);
1173 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1174 if (*port_nop != ODPP_NONE) {
1175 port_no = *port_nop;
1176 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1178 port_no = choose_port(dp, dpif_port);
1179 error = port_no == ODPP_NONE ? EFBIG : 0;
1182 *port_nop = port_no;
1183 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1185 ovs_mutex_unlock(&dp->port_mutex);
1191 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1193 struct dp_netdev *dp = get_dp_netdev(dpif);
1196 ovs_mutex_lock(&dp->port_mutex);
1197 if (port_no == ODPP_LOCAL) {
1200 struct dp_netdev_port *port;
1202 error = get_port_by_number(dp, port_no, &port);
1204 do_del_port(dp, port);
1207 ovs_mutex_unlock(&dp->port_mutex);
1213 is_valid_port_number(odp_port_t port_no)
1215 return port_no != ODPP_NONE;
1218 static struct dp_netdev_port *
1219 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1221 struct dp_netdev_port *port;
1223 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1224 if (port->port_no == port_no) {
1232 get_port_by_number(struct dp_netdev *dp,
1233 odp_port_t port_no, struct dp_netdev_port **portp)
1235 if (!is_valid_port_number(port_no)) {
1239 *portp = dp_netdev_lookup_port(dp, port_no);
1240 return *portp ? 0 : ENOENT;
1245 port_ref(struct dp_netdev_port *port)
1248 ovs_refcount_ref(&port->ref_cnt);
1253 port_unref(struct dp_netdev_port *port)
1255 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1256 int n_rxq = netdev_n_rxq(port->netdev);
1259 netdev_close(port->netdev);
1260 netdev_restore_flags(port->sf);
1262 for (i = 0; i < n_rxq; i++) {
1263 netdev_rxq_close(port->rxq[i]);
1272 get_port_by_name(struct dp_netdev *dp,
1273 const char *devname, struct dp_netdev_port **portp)
1274 OVS_REQUIRES(dp->port_mutex)
1276 struct dp_netdev_port *port;
1278 CMAP_FOR_EACH (port, node, &dp->ports) {
1279 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1288 get_n_pmd_threads(struct dp_netdev *dp)
1290 /* There is one non pmd thread in dp->poll_threads */
1291 return cmap_count(&dp->poll_threads) - 1;
1295 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1297 struct dp_netdev_pmd_thread *pmd;
1300 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1301 if (pmd->numa_id == numa_id) {
1309 /* Returns 'true' if there is a port with pmd netdev and the netdev
1310 * is on numa node 'numa_id'. */
1312 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1314 struct dp_netdev_port *port;
1316 CMAP_FOR_EACH (port, node, &dp->ports) {
1317 if (netdev_is_pmd(port->netdev)
1318 && netdev_get_numa_id(port->netdev) == numa_id) {
1328 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1329 OVS_REQUIRES(dp->port_mutex)
1331 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1332 seq_change(dp->port_seq);
1333 if (netdev_is_pmd(port->netdev)) {
1334 int numa_id = netdev_get_numa_id(port->netdev);
1336 /* PMD threads can not be on invalid numa node. */
1337 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
1338 /* If there is no netdev on the numa node, deletes the pmd threads
1339 * for that numa. Else, deletes the queues from polling lists. */
1340 if (!has_pmd_port_for_numa(dp, numa_id)) {
1341 dp_netdev_del_pmds_on_numa(dp, numa_id);
1343 struct dp_netdev_pmd_thread *pmd;
1344 struct rxq_poll *poll, *next;
1346 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1347 if (pmd->numa_id == numa_id) {
1350 ovs_mutex_lock(&pmd->poll_mutex);
1351 LIST_FOR_EACH_SAFE (poll, next, node, &pmd->poll_list) {
1352 if (poll->port == port) {
1354 port_unref(poll->port);
1355 list_remove(&poll->node);
1360 ovs_mutex_unlock(&pmd->poll_mutex);
1362 dp_netdev_reload_pmd__(pmd);
1373 answer_port_query(const struct dp_netdev_port *port,
1374 struct dpif_port *dpif_port)
1376 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1377 dpif_port->type = xstrdup(port->type);
1378 dpif_port->port_no = port->port_no;
1382 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1383 struct dpif_port *dpif_port)
1385 struct dp_netdev *dp = get_dp_netdev(dpif);
1386 struct dp_netdev_port *port;
1389 error = get_port_by_number(dp, port_no, &port);
1390 if (!error && dpif_port) {
1391 answer_port_query(port, dpif_port);
1398 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1399 struct dpif_port *dpif_port)
1401 struct dp_netdev *dp = get_dp_netdev(dpif);
1402 struct dp_netdev_port *port;
1405 ovs_mutex_lock(&dp->port_mutex);
1406 error = get_port_by_name(dp, devname, &port);
1407 if (!error && dpif_port) {
1408 answer_port_query(port, dpif_port);
1410 ovs_mutex_unlock(&dp->port_mutex);
1416 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1418 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1422 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1424 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1425 ovsrcu_postpone(dp_netdev_flow_free, flow);
1430 dp_netdev_flow_hash(const ovs_u128 *ufid)
1432 return ufid->u32[0];
1436 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1437 struct dp_netdev_flow *flow)
1438 OVS_REQUIRES(pmd->flow_mutex)
1440 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1442 dpcls_remove(&pmd->cls, &flow->cr);
1443 flow->cr.mask = NULL; /* Accessing rule's mask after this is not safe. */
1445 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1448 dp_netdev_flow_unref(flow);
1452 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1454 struct dp_netdev_flow *netdev_flow;
1456 ovs_mutex_lock(&pmd->flow_mutex);
1457 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1458 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1460 ovs_mutex_unlock(&pmd->flow_mutex);
1464 dpif_netdev_flow_flush(struct dpif *dpif)
1466 struct dp_netdev *dp = get_dp_netdev(dpif);
1467 struct dp_netdev_pmd_thread *pmd;
1469 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1470 dp_netdev_pmd_flow_flush(pmd);
1476 struct dp_netdev_port_state {
1477 struct cmap_position position;
1482 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1484 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1489 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1490 struct dpif_port *dpif_port)
1492 struct dp_netdev_port_state *state = state_;
1493 struct dp_netdev *dp = get_dp_netdev(dpif);
1494 struct cmap_node *node;
1497 node = cmap_next_position(&dp->ports, &state->position);
1499 struct dp_netdev_port *port;
1501 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1504 state->name = xstrdup(netdev_get_name(port->netdev));
1505 dpif_port->name = state->name;
1506 dpif_port->type = port->type;
1507 dpif_port->port_no = port->port_no;
1518 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1520 struct dp_netdev_port_state *state = state_;
1527 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1529 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1530 uint64_t new_port_seq;
1533 new_port_seq = seq_read(dpif->dp->port_seq);
1534 if (dpif->last_port_seq != new_port_seq) {
1535 dpif->last_port_seq = new_port_seq;
1545 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1547 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1549 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1552 static struct dp_netdev_flow *
1553 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1555 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1558 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1560 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1563 /* netdev_flow_key utilities.
1565 * netdev_flow_key is basically a miniflow. We use these functions
1566 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1567 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1569 * - Since we are dealing exclusively with miniflows created by
1570 * miniflow_extract(), if the map is different the miniflow is different.
1571 * Therefore we can be faster by comparing the map and the miniflow in a
1573 * - These functions can be inlined by the compiler. */
1575 /* Given the number of bits set in miniflow's maps, returns the size of the
1576 * 'netdev_flow_key.mf' */
1577 static inline size_t
1578 netdev_flow_key_size(size_t flow_u64s)
1580 return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s);
1584 netdev_flow_key_equal(const struct netdev_flow_key *a,
1585 const struct netdev_flow_key *b)
1587 /* 'b->len' may be not set yet. */
1588 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1591 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1592 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1593 * generated by miniflow_extract. */
1595 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1596 const struct miniflow *mf)
1598 return !memcmp(&key->mf, mf, key->len);
1602 netdev_flow_key_clone(struct netdev_flow_key *dst,
1603 const struct netdev_flow_key *src)
1606 offsetof(struct netdev_flow_key, mf) + src->len);
1611 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1612 const struct flow *src)
1614 struct dp_packet packet;
1615 uint64_t buf_stub[512 / 8];
1617 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1618 pkt_metadata_from_flow(&packet.md, src);
1619 flow_compose(&packet, src);
1620 miniflow_extract(&packet, &dst->mf);
1621 dp_packet_uninit(&packet);
1623 dst->len = netdev_flow_key_size(miniflow_n_values(&dst->mf));
1624 dst->hash = 0; /* Not computed yet. */
1627 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1629 netdev_flow_mask_init(struct netdev_flow_key *mask,
1630 const struct match *match)
1632 uint64_t *dst = miniflow_values(&mask->mf);
1633 struct flowmap fmap;
1637 /* Only check masks that make sense for the flow. */
1638 flow_wc_map(&match->flow, &fmap);
1639 flowmap_init(&mask->mf.map);
1641 FLOWMAP_FOR_EACH_INDEX(idx, fmap) {
1642 uint64_t mask_u64 = flow_u64_value(&match->wc.masks, idx);
1645 flowmap_set(&mask->mf.map, idx, 1);
1647 hash = hash_add64(hash, mask_u64);
1653 FLOWMAP_FOR_EACH_MAP (map, mask->mf.map) {
1654 hash = hash_add64(hash, map);
1657 size_t n = dst - miniflow_get_values(&mask->mf);
1659 mask->hash = hash_finish(hash, n * 8);
1660 mask->len = netdev_flow_key_size(n);
1663 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1665 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1666 const struct flow *flow,
1667 const struct netdev_flow_key *mask)
1669 uint64_t *dst_u64 = miniflow_values(&dst->mf);
1670 const uint64_t *mask_u64 = miniflow_get_values(&mask->mf);
1674 dst->len = mask->len;
1675 dst->mf = mask->mf; /* Copy maps. */
1677 FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf.map) {
1678 *dst_u64 = value & *mask_u64++;
1679 hash = hash_add64(hash, *dst_u64++);
1681 dst->hash = hash_finish(hash,
1682 (dst_u64 - miniflow_get_values(&dst->mf)) * 8);
1685 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
1686 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
1687 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
1689 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1691 static inline uint32_t
1692 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1693 const struct netdev_flow_key *mask)
1695 const uint64_t *p = miniflow_get_values(&mask->mf);
1699 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, key, mask->mf.map) {
1700 hash = hash_add64(hash, value & *p++);
1703 return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8);
1707 emc_entry_alive(struct emc_entry *ce)
1709 return ce->flow && !ce->flow->dead;
1713 emc_clear_entry(struct emc_entry *ce)
1716 dp_netdev_flow_unref(ce->flow);
1722 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1723 const struct netdev_flow_key *key)
1725 if (ce->flow != flow) {
1727 dp_netdev_flow_unref(ce->flow);
1730 if (dp_netdev_flow_ref(flow)) {
1737 netdev_flow_key_clone(&ce->key, key);
1742 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1743 struct dp_netdev_flow *flow)
1745 struct emc_entry *to_be_replaced = NULL;
1746 struct emc_entry *current_entry;
1748 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1749 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1750 /* We found the entry with the 'mf' miniflow */
1751 emc_change_entry(current_entry, flow, NULL);
1755 /* Replacement policy: put the flow in an empty (not alive) entry, or
1756 * in the first entry where it can be */
1758 || (emc_entry_alive(to_be_replaced)
1759 && !emc_entry_alive(current_entry))
1760 || current_entry->key.hash < to_be_replaced->key.hash) {
1761 to_be_replaced = current_entry;
1764 /* We didn't find the miniflow in the cache.
1765 * The 'to_be_replaced' entry is where the new flow will be stored */
1767 emc_change_entry(to_be_replaced, flow, key);
1770 static inline struct dp_netdev_flow *
1771 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1773 struct emc_entry *current_entry;
1775 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1776 if (current_entry->key.hash == key->hash
1777 && emc_entry_alive(current_entry)
1778 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1780 /* We found the entry with the 'key->mf' miniflow */
1781 return current_entry->flow;
1788 static struct dp_netdev_flow *
1789 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1790 const struct netdev_flow_key *key)
1792 struct dp_netdev_flow *netdev_flow;
1793 struct dpcls_rule *rule;
1795 dpcls_lookup(&pmd->cls, key, &rule, 1);
1796 netdev_flow = dp_netdev_flow_cast(rule);
1801 static struct dp_netdev_flow *
1802 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1803 const ovs_u128 *ufidp, const struct nlattr *key,
1806 struct dp_netdev_flow *netdev_flow;
1810 /* If a UFID is not provided, determine one based on the key. */
1811 if (!ufidp && key && key_len
1812 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1813 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1818 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1820 if (ovs_u128_equals(&netdev_flow->ufid, ufidp)) {
1830 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1831 struct dpif_flow_stats *stats)
1833 struct dp_netdev_flow *netdev_flow;
1834 unsigned long long n;
1838 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1840 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1841 stats->n_packets = n;
1842 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1844 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1846 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1847 stats->tcp_flags = flags;
1850 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1851 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1852 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1855 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1856 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1857 struct dpif_flow *flow, bool terse)
1860 memset(flow, 0, sizeof *flow);
1862 struct flow_wildcards wc;
1863 struct dp_netdev_actions *actions;
1865 struct odp_flow_key_parms odp_parms = {
1866 .flow = &netdev_flow->flow,
1868 .support = dp_netdev_support,
1871 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1874 offset = key_buf->size;
1875 flow->key = ofpbuf_tail(key_buf);
1876 odp_parms.odp_in_port = netdev_flow->flow.in_port.odp_port;
1877 odp_flow_key_from_flow(&odp_parms, key_buf);
1878 flow->key_len = key_buf->size - offset;
1881 offset = mask_buf->size;
1882 flow->mask = ofpbuf_tail(mask_buf);
1883 odp_parms.odp_in_port = wc.masks.in_port.odp_port;
1884 odp_parms.key_buf = key_buf;
1885 odp_flow_key_from_mask(&odp_parms, mask_buf);
1886 flow->mask_len = mask_buf->size - offset;
1889 actions = dp_netdev_flow_get_actions(netdev_flow);
1890 flow->actions = actions->actions;
1891 flow->actions_len = actions->size;
1894 flow->ufid = netdev_flow->ufid;
1895 flow->ufid_present = true;
1896 flow->pmd_id = netdev_flow->pmd_id;
1897 get_dpif_flow_stats(netdev_flow, &flow->stats);
1901 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1902 const struct nlattr *mask_key,
1903 uint32_t mask_key_len, const struct flow *flow,
1904 struct flow_wildcards *wc)
1906 enum odp_key_fitness fitness;
1908 fitness = odp_flow_key_to_mask_udpif(mask_key, mask_key_len, key,
1911 /* This should not happen: it indicates that
1912 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1913 * disagree on the acceptable form of a mask. Log the problem
1914 * as an error, with enough details to enable debugging. */
1915 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1917 if (!VLOG_DROP_ERR(&rl)) {
1921 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1923 VLOG_ERR("internal error parsing flow mask %s (%s)",
1924 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1935 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1940 if (odp_flow_key_to_flow_udpif(key, key_len, flow)) {
1941 /* This should not happen: it indicates that odp_flow_key_from_flow()
1942 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1943 * flow. Log the problem as an error, with enough details to enable
1945 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1947 if (!VLOG_DROP_ERR(&rl)) {
1951 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1952 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1959 in_port = flow->in_port.odp_port;
1960 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1964 /* Userspace datapath doesn't support conntrack. */
1965 if (flow->ct_state || flow->ct_zone || flow->ct_mark
1966 || !ovs_u128_is_zero(&flow->ct_label)) {
1974 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1976 struct dp_netdev *dp = get_dp_netdev(dpif);
1977 struct dp_netdev_flow *netdev_flow;
1978 struct dp_netdev_pmd_thread *pmd;
1979 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1980 ? NON_PMD_CORE_ID : get->pmd_id;
1983 pmd = dp_netdev_get_pmd(dp, pmd_id);
1988 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1991 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1996 dp_netdev_pmd_unref(pmd);
2002 static struct dp_netdev_flow *
2003 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
2004 struct match *match, const ovs_u128 *ufid,
2005 const struct nlattr *actions, size_t actions_len)
2006 OVS_REQUIRES(pmd->flow_mutex)
2008 struct dp_netdev_flow *flow;
2009 struct netdev_flow_key mask;
2011 netdev_flow_mask_init(&mask, match);
2012 /* Make sure wc does not have metadata. */
2013 ovs_assert(!FLOWMAP_HAS_FIELD(&mask.mf.map, metadata)
2014 && !FLOWMAP_HAS_FIELD(&mask.mf.map, regs));
2016 /* Do not allocate extra space. */
2017 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
2018 memset(&flow->stats, 0, sizeof flow->stats);
2021 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
2022 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
2023 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
2024 ovs_refcount_init(&flow->ref_cnt);
2025 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
2027 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
2028 dpcls_insert(&pmd->cls, &flow->cr, &mask);
2030 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
2031 dp_netdev_flow_hash(&flow->ufid));
2033 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2035 struct ds ds = DS_EMPTY_INITIALIZER;
2037 match.tun_md.valid = false;
2038 match.flow = flow->flow;
2039 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2041 ds_put_cstr(&ds, "flow_add: ");
2042 odp_format_ufid(ufid, &ds);
2043 ds_put_cstr(&ds, " ");
2044 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2045 ds_put_cstr(&ds, ", actions:");
2046 format_odp_actions(&ds, actions, actions_len);
2048 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2057 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2059 struct dp_netdev *dp = get_dp_netdev(dpif);
2060 struct dp_netdev_flow *netdev_flow;
2061 struct netdev_flow_key key;
2062 struct dp_netdev_pmd_thread *pmd;
2065 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2066 ? NON_PMD_CORE_ID : put->pmd_id;
2069 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2073 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2074 put->mask, put->mask_len,
2075 &match.flow, &match.wc);
2080 pmd = dp_netdev_get_pmd(dp, pmd_id);
2085 /* Must produce a netdev_flow_key for lookup.
2086 * This interface is no longer performance critical, since it is not used
2087 * for upcall processing any more. */
2088 netdev_flow_key_from_flow(&key, &match.flow);
2093 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2096 ovs_mutex_lock(&pmd->flow_mutex);
2097 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2099 if (put->flags & DPIF_FP_CREATE) {
2100 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2102 memset(put->stats, 0, sizeof *put->stats);
2104 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2114 if (put->flags & DPIF_FP_MODIFY
2115 && flow_equal(&match.flow, &netdev_flow->flow)) {
2116 struct dp_netdev_actions *new_actions;
2117 struct dp_netdev_actions *old_actions;
2119 new_actions = dp_netdev_actions_create(put->actions,
2122 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2123 ovsrcu_set(&netdev_flow->actions, new_actions);
2126 get_dpif_flow_stats(netdev_flow, put->stats);
2128 if (put->flags & DPIF_FP_ZERO_STATS) {
2129 /* XXX: The userspace datapath uses thread local statistics
2130 * (for flows), which should be updated only by the owning
2131 * thread. Since we cannot write on stats memory here,
2132 * we choose not to support this flag. Please note:
2133 * - This feature is currently used only by dpctl commands with
2135 * - Should the need arise, this operation can be implemented
2136 * by keeping a base value (to be update here) for each
2137 * counter, and subtracting it before outputting the stats */
2141 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2142 } else if (put->flags & DPIF_FP_CREATE) {
2145 /* Overlapping flow. */
2149 ovs_mutex_unlock(&pmd->flow_mutex);
2150 dp_netdev_pmd_unref(pmd);
2156 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2158 struct dp_netdev *dp = get_dp_netdev(dpif);
2159 struct dp_netdev_flow *netdev_flow;
2160 struct dp_netdev_pmd_thread *pmd;
2161 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2162 ? NON_PMD_CORE_ID : del->pmd_id;
2165 pmd = dp_netdev_get_pmd(dp, pmd_id);
2170 ovs_mutex_lock(&pmd->flow_mutex);
2171 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2175 get_dpif_flow_stats(netdev_flow, del->stats);
2177 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2181 ovs_mutex_unlock(&pmd->flow_mutex);
2182 dp_netdev_pmd_unref(pmd);
2187 struct dpif_netdev_flow_dump {
2188 struct dpif_flow_dump up;
2189 struct cmap_position poll_thread_pos;
2190 struct cmap_position flow_pos;
2191 struct dp_netdev_pmd_thread *cur_pmd;
2193 struct ovs_mutex mutex;
2196 static struct dpif_netdev_flow_dump *
2197 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2199 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2202 static struct dpif_flow_dump *
2203 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2205 struct dpif_netdev_flow_dump *dump;
2207 dump = xzalloc(sizeof *dump);
2208 dpif_flow_dump_init(&dump->up, dpif_);
2209 dump->up.terse = terse;
2210 ovs_mutex_init(&dump->mutex);
2216 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2218 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2220 ovs_mutex_destroy(&dump->mutex);
2225 struct dpif_netdev_flow_dump_thread {
2226 struct dpif_flow_dump_thread up;
2227 struct dpif_netdev_flow_dump *dump;
2228 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2229 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2232 static struct dpif_netdev_flow_dump_thread *
2233 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2235 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2238 static struct dpif_flow_dump_thread *
2239 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2241 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2242 struct dpif_netdev_flow_dump_thread *thread;
2244 thread = xmalloc(sizeof *thread);
2245 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2246 thread->dump = dump;
2251 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2253 struct dpif_netdev_flow_dump_thread *thread
2254 = dpif_netdev_flow_dump_thread_cast(thread_);
2260 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2261 struct dpif_flow *flows, int max_flows)
2263 struct dpif_netdev_flow_dump_thread *thread
2264 = dpif_netdev_flow_dump_thread_cast(thread_);
2265 struct dpif_netdev_flow_dump *dump = thread->dump;
2266 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2270 ovs_mutex_lock(&dump->mutex);
2271 if (!dump->status) {
2272 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2273 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2274 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2275 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2277 /* First call to dump_next(), extracts the first pmd thread.
2278 * If there is no pmd thread, returns immediately. */
2280 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2282 ovs_mutex_unlock(&dump->mutex);
2289 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2290 struct cmap_node *node;
2292 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2296 netdev_flows[n_flows] = CONTAINER_OF(node,
2297 struct dp_netdev_flow,
2300 /* When finishing dumping the current pmd thread, moves to
2302 if (n_flows < flow_limit) {
2303 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2304 dp_netdev_pmd_unref(pmd);
2305 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2311 /* Keeps the reference to next caller. */
2312 dump->cur_pmd = pmd;
2314 /* If the current dump is empty, do not exit the loop, since the
2315 * remaining pmds could have flows to be dumped. Just dumps again
2316 * on the new 'pmd'. */
2319 ovs_mutex_unlock(&dump->mutex);
2321 for (i = 0; i < n_flows; i++) {
2322 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2323 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2324 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2325 struct dpif_flow *f = &flows[i];
2326 struct ofpbuf key, mask;
2328 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2329 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2330 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2338 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2339 OVS_NO_THREAD_SAFETY_ANALYSIS
2341 struct dp_netdev *dp = get_dp_netdev(dpif);
2342 struct dp_netdev_pmd_thread *pmd;
2343 struct dp_packet *pp;
2345 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2346 dp_packet_size(execute->packet) > UINT16_MAX) {
2350 /* Tries finding the 'pmd'. If NULL is returned, that means
2351 * the current thread is a non-pmd thread and should use
2352 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2353 pmd = ovsthread_getspecific(dp->per_pmd_key);
2355 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2358 /* If the current thread is non-pmd thread, acquires
2359 * the 'non_pmd_mutex'. */
2360 if (pmd->core_id == NON_PMD_CORE_ID) {
2361 ovs_mutex_lock(&dp->non_pmd_mutex);
2362 ovs_mutex_lock(&dp->port_mutex);
2365 pp = execute->packet;
2366 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2367 execute->actions_len);
2368 if (pmd->core_id == NON_PMD_CORE_ID) {
2369 dp_netdev_pmd_unref(pmd);
2370 ovs_mutex_unlock(&dp->port_mutex);
2371 ovs_mutex_unlock(&dp->non_pmd_mutex);
2378 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2382 for (i = 0; i < n_ops; i++) {
2383 struct dpif_op *op = ops[i];
2386 case DPIF_OP_FLOW_PUT:
2387 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2390 case DPIF_OP_FLOW_DEL:
2391 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2394 case DPIF_OP_EXECUTE:
2395 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2398 case DPIF_OP_FLOW_GET:
2399 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2405 /* Returns true if the configuration for rx queues or cpu mask
2408 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2410 if (dp->n_dpdk_rxqs != rxqs) {
2413 if (dp->pmd_cmask != NULL && cmask != NULL) {
2414 return strcmp(dp->pmd_cmask, cmask);
2416 return (dp->pmd_cmask != NULL || cmask != NULL);
2421 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2423 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2425 struct dp_netdev *dp = get_dp_netdev(dpif);
2427 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2428 struct dp_netdev_port *port;
2430 dp_netdev_destroy_all_pmds(dp);
2432 CMAP_FOR_EACH (port, node, &dp->ports) {
2433 if (netdev_is_pmd(port->netdev)) {
2436 /* Closes the existing 'rxq's. */
2437 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2438 netdev_rxq_close(port->rxq[i]);
2439 port->rxq[i] = NULL;
2442 /* Sets the new rx queue config. */
2443 err = netdev_set_multiq(port->netdev,
2444 ovs_numa_get_n_cores() + 1,
2446 if (err && (err != EOPNOTSUPP)) {
2447 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2448 " %u", netdev_get_name(port->netdev),
2453 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2454 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2455 * netdev_n_rxq(port->netdev));
2456 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2457 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2461 dp->n_dpdk_rxqs = n_rxqs;
2463 /* Reconfigures the cpu mask. */
2464 ovs_numa_set_cpu_mask(cmask);
2465 free(dp->pmd_cmask);
2466 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2468 /* Restores the non-pmd. */
2469 dp_netdev_set_nonpmd(dp);
2470 /* Restores all pmd threads. */
2471 dp_netdev_reset_pmd_threads(dp);
2478 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2479 uint32_t queue_id, uint32_t *priority)
2481 *priority = queue_id;
2486 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2487 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2488 struct dp_netdev_actions *
2489 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2491 struct dp_netdev_actions *netdev_actions;
2493 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2494 memcpy(netdev_actions->actions, actions, size);
2495 netdev_actions->size = size;
2497 return netdev_actions;
2500 struct dp_netdev_actions *
2501 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2503 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2507 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2512 static inline unsigned long long
2513 cycles_counter(void)
2516 return rte_get_tsc_cycles();
2522 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2523 extern struct ovs_mutex cycles_counter_fake_mutex;
2525 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2527 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2528 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2529 OVS_NO_THREAD_SAFETY_ANALYSIS
2531 pmd->last_cycles = cycles_counter();
2534 /* Stop counting cycles and add them to the counter 'type' */
2536 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2537 enum pmd_cycles_counter_type type)
2538 OVS_RELEASES(&cycles_counter_fake_mutex)
2539 OVS_NO_THREAD_SAFETY_ANALYSIS
2541 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2543 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2547 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2548 struct dp_netdev_port *port,
2549 struct netdev_rxq *rxq)
2551 struct dp_packet *packets[NETDEV_MAX_BURST];
2554 cycles_count_start(pmd);
2555 error = netdev_rxq_recv(rxq, packets, &cnt);
2556 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2560 *recirc_depth_get() = 0;
2562 /* XXX: initialize md in netdev implementation. */
2563 for (i = 0; i < cnt; i++) {
2564 pkt_metadata_init(&packets[i]->md, port->port_no);
2566 cycles_count_start(pmd);
2567 dp_netdev_input(pmd, packets, cnt);
2568 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2569 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2570 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2572 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2573 netdev_get_name(port->netdev), ovs_strerror(error));
2577 /* Return true if needs to revalidate datapath flows. */
2579 dpif_netdev_run(struct dpif *dpif)
2581 struct dp_netdev_port *port;
2582 struct dp_netdev *dp = get_dp_netdev(dpif);
2583 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2585 uint64_t new_tnl_seq;
2587 ovs_mutex_lock(&dp->non_pmd_mutex);
2588 CMAP_FOR_EACH (port, node, &dp->ports) {
2589 if (!netdev_is_pmd(port->netdev)) {
2592 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2593 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2597 ovs_mutex_unlock(&dp->non_pmd_mutex);
2598 dp_netdev_pmd_unref(non_pmd);
2600 tnl_neigh_cache_run();
2602 new_tnl_seq = seq_read(tnl_conf_seq);
2604 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2605 dp->last_tnl_conf_seq = new_tnl_seq;
2612 dpif_netdev_wait(struct dpif *dpif)
2614 struct dp_netdev_port *port;
2615 struct dp_netdev *dp = get_dp_netdev(dpif);
2617 ovs_mutex_lock(&dp_netdev_mutex);
2618 CMAP_FOR_EACH (port, node, &dp->ports) {
2619 if (!netdev_is_pmd(port->netdev)) {
2622 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2623 netdev_rxq_wait(port->rxq[i]);
2627 ovs_mutex_unlock(&dp_netdev_mutex);
2628 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2632 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2633 struct rxq_poll **ppoll_list, int poll_cnt)
2634 OVS_REQUIRES(pmd->poll_mutex)
2636 struct rxq_poll *poll_list = *ppoll_list;
2637 struct rxq_poll *poll;
2640 for (i = 0; i < poll_cnt; i++) {
2641 port_unref(poll_list[i].port);
2644 poll_list = xrealloc(poll_list, pmd->poll_cnt * sizeof *poll_list);
2647 LIST_FOR_EACH (poll, node, &pmd->poll_list) {
2648 port_ref(poll->port);
2649 poll_list[i++] = *poll;
2652 *ppoll_list = poll_list;
2653 return pmd->poll_cnt;
2657 pmd_thread_main(void *f_)
2659 struct dp_netdev_pmd_thread *pmd = f_;
2660 unsigned int lc = 0;
2661 struct rxq_poll *poll_list;
2662 unsigned int port_seq = PMD_INITIAL_SEQ;
2669 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2670 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2671 pmd_thread_setaffinity_cpu(pmd->core_id);
2673 emc_cache_init(&pmd->flow_cache);
2675 ovs_mutex_lock(&pmd->poll_mutex);
2676 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2677 ovs_mutex_unlock(&pmd->poll_mutex);
2679 /* List port/core affinity */
2680 for (i = 0; i < poll_cnt; i++) {
2681 VLOG_INFO("Core %d processing port \'%s\'\n", pmd->core_id,
2682 netdev_get_name(poll_list[i].port->netdev));
2685 /* Signal here to make sure the pmd finishes
2686 * reloading the updated configuration. */
2687 dp_netdev_pmd_reload_done(pmd);
2690 for (i = 0; i < poll_cnt; i++) {
2691 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2699 emc_cache_slow_sweep(&pmd->flow_cache);
2700 coverage_try_clear();
2703 atomic_read_relaxed(&pmd->change_seq, &seq);
2704 if (seq != port_seq) {
2711 emc_cache_uninit(&pmd->flow_cache);
2713 if (!latch_is_set(&pmd->exit_latch)){
2717 for (i = 0; i < poll_cnt; i++) {
2718 port_unref(poll_list[i].port);
2721 dp_netdev_pmd_reload_done(pmd);
2728 dp_netdev_disable_upcall(struct dp_netdev *dp)
2729 OVS_ACQUIRES(dp->upcall_rwlock)
2731 fat_rwlock_wrlock(&dp->upcall_rwlock);
2735 dpif_netdev_disable_upcall(struct dpif *dpif)
2736 OVS_NO_THREAD_SAFETY_ANALYSIS
2738 struct dp_netdev *dp = get_dp_netdev(dpif);
2739 dp_netdev_disable_upcall(dp);
2743 dp_netdev_enable_upcall(struct dp_netdev *dp)
2744 OVS_RELEASES(dp->upcall_rwlock)
2746 fat_rwlock_unlock(&dp->upcall_rwlock);
2750 dpif_netdev_enable_upcall(struct dpif *dpif)
2751 OVS_NO_THREAD_SAFETY_ANALYSIS
2753 struct dp_netdev *dp = get_dp_netdev(dpif);
2754 dp_netdev_enable_upcall(dp);
2758 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2760 ovs_mutex_lock(&pmd->cond_mutex);
2761 xpthread_cond_signal(&pmd->cond);
2762 ovs_mutex_unlock(&pmd->cond_mutex);
2765 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2766 * the pointer if succeeds, otherwise, NULL.
2768 * Caller must unrefs the returned reference. */
2769 static struct dp_netdev_pmd_thread *
2770 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2772 struct dp_netdev_pmd_thread *pmd;
2773 const struct cmap_node *pnode;
2775 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2779 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2781 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2784 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2786 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2788 struct dp_netdev_pmd_thread *non_pmd;
2790 non_pmd = xzalloc(sizeof *non_pmd);
2791 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2795 /* Caller must have valid pointer to 'pmd'. */
2797 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2799 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2803 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2805 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2806 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2810 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2811 * fails, keeps checking for next node until reaching the end of cmap.
2813 * Caller must unrefs the returned reference. */
2814 static struct dp_netdev_pmd_thread *
2815 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2817 struct dp_netdev_pmd_thread *next;
2820 struct cmap_node *node;
2822 node = cmap_next_position(&dp->poll_threads, pos);
2823 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2825 } while (next && !dp_netdev_pmd_try_ref(next));
2830 /* Configures the 'pmd' based on the input argument. */
2832 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2833 int index, unsigned core_id, int numa_id)
2837 pmd->core_id = core_id;
2838 pmd->numa_id = numa_id;
2841 atomic_init(&pmd->tx_qid,
2842 (core_id == NON_PMD_CORE_ID)
2843 ? ovs_numa_get_n_cores()
2844 : get_n_pmd_threads(dp));
2846 ovs_refcount_init(&pmd->ref_cnt);
2847 latch_init(&pmd->exit_latch);
2848 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2849 xpthread_cond_init(&pmd->cond, NULL);
2850 ovs_mutex_init(&pmd->cond_mutex);
2851 ovs_mutex_init(&pmd->flow_mutex);
2852 ovs_mutex_init(&pmd->poll_mutex);
2853 dpcls_init(&pmd->cls);
2854 cmap_init(&pmd->flow_table);
2855 list_init(&pmd->poll_list);
2856 /* init the 'flow_cache' since there is no
2857 * actual thread created for NON_PMD_CORE_ID. */
2858 if (core_id == NON_PMD_CORE_ID) {
2859 emc_cache_init(&pmd->flow_cache);
2861 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2862 hash_int(core_id, 0));
2866 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2868 dp_netdev_pmd_flow_flush(pmd);
2869 dpcls_destroy(&pmd->cls);
2870 cmap_destroy(&pmd->flow_table);
2871 ovs_mutex_destroy(&pmd->flow_mutex);
2872 latch_destroy(&pmd->exit_latch);
2873 xpthread_cond_destroy(&pmd->cond);
2874 ovs_mutex_destroy(&pmd->cond_mutex);
2875 ovs_mutex_destroy(&pmd->poll_mutex);
2879 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2880 * and unrefs the struct. */
2882 dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd)
2884 struct rxq_poll *poll;
2886 /* Uninit the 'flow_cache' since there is
2887 * no actual thread uninit it for NON_PMD_CORE_ID. */
2888 if (pmd->core_id == NON_PMD_CORE_ID) {
2889 emc_cache_uninit(&pmd->flow_cache);
2891 latch_set(&pmd->exit_latch);
2892 dp_netdev_reload_pmd__(pmd);
2893 ovs_numa_unpin_core(pmd->core_id);
2894 xpthread_join(pmd->thread, NULL);
2897 /* Unref all ports and free poll_list. */
2898 LIST_FOR_EACH_POP (poll, node, &pmd->poll_list) {
2899 port_unref(poll->port);
2903 /* Purges the 'pmd''s flows after stopping the thread, but before
2904 * destroying the flows, so that the flow stats can be collected. */
2905 if (dp->dp_purge_cb) {
2906 dp->dp_purge_cb(dp->dp_purge_aux, pmd->core_id);
2908 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2909 dp_netdev_pmd_unref(pmd);
2912 /* Destroys all pmd threads. */
2914 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2916 struct dp_netdev_pmd_thread *pmd;
2918 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2919 dp_netdev_del_pmd(dp, pmd);
2923 /* Deletes all pmd threads on numa node 'numa_id' and
2924 * fixes tx_qids of other threads to keep them sequential. */
2926 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2928 struct dp_netdev_pmd_thread *pmd;
2929 int n_pmds_on_numa, n_pmds;
2930 int *free_idx, k = 0;
2932 n_pmds_on_numa = get_n_pmd_threads_on_numa(dp, numa_id);
2933 free_idx = xmalloc(n_pmds_on_numa * sizeof *free_idx);
2935 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2936 if (pmd->numa_id == numa_id) {
2937 atomic_read_relaxed(&pmd->tx_qid, &free_idx[k]);
2939 dp_netdev_del_pmd(dp, pmd);
2943 n_pmds = get_n_pmd_threads(dp);
2944 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2947 atomic_read_relaxed(&pmd->tx_qid, &old_tx_qid);
2949 if (old_tx_qid >= n_pmds) {
2950 int new_tx_qid = free_idx[--k];
2952 atomic_store_relaxed(&pmd->tx_qid, new_tx_qid);
2959 /* Returns PMD thread from this numa node with fewer rx queues to poll.
2960 * Returns NULL if there is no PMD threads on this numa node.
2961 * Can be called safely only by main thread. */
2962 static struct dp_netdev_pmd_thread *
2963 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id)
2966 struct dp_netdev_pmd_thread *pmd, *res = NULL;
2968 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2969 if (pmd->numa_id == numa_id
2970 && (min_cnt > pmd->poll_cnt || res == NULL)) {
2971 min_cnt = pmd->poll_cnt;
2979 /* Adds rx queue to poll_list of PMD thread. */
2981 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
2982 struct dp_netdev_port *port, struct netdev_rxq *rx)
2983 OVS_REQUIRES(pmd->poll_mutex)
2985 struct rxq_poll *poll = xmalloc(sizeof *poll);
2991 list_push_back(&pmd->poll_list, &poll->node);
2995 /* Checks the numa node id of 'netdev' and starts pmd threads for
2998 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
3002 if (!ovs_numa_numa_id_is_valid(numa_id)) {
3003 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
3004 "invalid", numa_id);
3008 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
3010 /* If there are already pmd threads created for the numa node
3011 * in which 'netdev' is on, do nothing. Else, creates the
3012 * pmd threads for the numa node. */
3014 int can_have, n_unpinned, i, index = 0;
3015 struct dp_netdev_pmd_thread **pmds;
3016 struct dp_netdev_port *port;
3018 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
3020 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
3021 "cores on numa node");
3025 /* If cpu mask is specified, uses all unpinned cores, otherwise
3026 * tries creating NR_PMD_THREADS pmd threads. */
3027 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
3028 pmds = xzalloc(can_have * sizeof *pmds);
3029 for (i = 0; i < can_have; i++) {
3030 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
3031 pmds[i] = xzalloc(sizeof **pmds);
3032 dp_netdev_configure_pmd(pmds[i], dp, i, core_id, numa_id);
3035 /* Distributes rx queues of this numa node between new pmd threads. */
3036 CMAP_FOR_EACH (port, node, &dp->ports) {
3037 if (netdev_is_pmd(port->netdev)
3038 && netdev_get_numa_id(port->netdev) == numa_id) {
3039 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
3040 /* Make thread-safety analyser happy. */
3041 ovs_mutex_lock(&pmds[index]->poll_mutex);
3042 dp_netdev_add_rxq_to_pmd(pmds[index], port, port->rxq[i]);
3043 ovs_mutex_unlock(&pmds[index]->poll_mutex);
3044 index = (index + 1) % can_have;
3049 /* Actual start of pmd threads. */
3050 for (i = 0; i < can_have; i++) {
3051 pmds[i]->thread = ovs_thread_create("pmd", pmd_thread_main, pmds[i]);
3054 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
3059 /* Called after pmd threads config change. Restarts pmd threads with
3060 * new configuration. */
3062 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
3064 struct dp_netdev_port *port;
3066 CMAP_FOR_EACH (port, node, &dp->ports) {
3067 if (netdev_is_pmd(port->netdev)) {
3068 int numa_id = netdev_get_numa_id(port->netdev);
3070 dp_netdev_set_pmds_on_numa(dp, numa_id);
3076 dpif_netdev_get_datapath_version(void)
3078 return xstrdup("<built-in>");
3082 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
3083 uint16_t tcp_flags, long long now)
3087 atomic_store_relaxed(&netdev_flow->stats.used, now);
3088 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
3089 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
3090 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
3092 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
3096 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
3097 enum dp_stat_type type, int cnt)
3099 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3103 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3104 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3105 enum dpif_upcall_type type, const struct nlattr *userdata,
3106 struct ofpbuf *actions, struct ofpbuf *put_actions)
3108 struct dp_netdev *dp = pmd->dp;
3109 struct flow_tnl orig_tunnel;
3112 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3116 /* Upcall processing expects the Geneve options to be in the translated
3117 * format but we need to retain the raw format for datapath use. */
3118 orig_tunnel.flags = flow->tunnel.flags;
3119 if (flow->tunnel.flags & FLOW_TNL_F_UDPIF) {
3120 orig_tunnel.metadata.present.len = flow->tunnel.metadata.present.len;
3121 memcpy(orig_tunnel.metadata.opts.gnv, flow->tunnel.metadata.opts.gnv,
3122 flow->tunnel.metadata.present.len);
3123 err = tun_metadata_from_geneve_udpif(&orig_tunnel, &orig_tunnel,
3130 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3131 struct ds ds = DS_EMPTY_INITIALIZER;
3134 struct odp_flow_key_parms odp_parms = {
3137 .odp_in_port = flow->in_port.odp_port,
3138 .support = dp_netdev_support,
3141 ofpbuf_init(&key, 0);
3142 odp_flow_key_from_flow(&odp_parms, &key);
3143 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3144 dp_packet_size(packet_));
3146 odp_flow_key_format(key.data, key.size, &ds);
3148 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3149 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3151 ofpbuf_uninit(&key);
3157 err = dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3158 actions, wc, put_actions, dp->upcall_aux);
3159 if (err && err != ENOSPC) {
3163 /* Translate tunnel metadata masks to datapath format. */
3165 if (wc->masks.tunnel.metadata.present.map) {
3166 struct geneve_opt opts[TLV_TOT_OPT_SIZE /
3167 sizeof(struct geneve_opt)];
3169 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3170 tun_metadata_to_geneve_udpif_mask(&flow->tunnel,
3172 orig_tunnel.metadata.opts.gnv,
3173 orig_tunnel.metadata.present.len,
3176 orig_tunnel.metadata.present.len = 0;
3179 memset(&wc->masks.tunnel.metadata, 0,
3180 sizeof wc->masks.tunnel.metadata);
3181 memcpy(&wc->masks.tunnel.metadata.opts.gnv, opts,
3182 orig_tunnel.metadata.present.len);
3184 wc->masks.tunnel.metadata.present.len = 0xff;
3187 /* Restore tunnel metadata. We need to use the saved options to ensure
3188 * that any unknown options are not lost. The generated mask will have
3189 * the same structure, matching on types and lengths but wildcarding
3190 * option data we don't care about. */
3191 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3192 memcpy(&flow->tunnel.metadata.opts.gnv, orig_tunnel.metadata.opts.gnv,
3193 orig_tunnel.metadata.present.len);
3194 flow->tunnel.metadata.present.len = orig_tunnel.metadata.present.len;
3195 flow->tunnel.flags |= FLOW_TNL_F_UDPIF;
3201 static inline uint32_t
3202 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3203 const struct miniflow *mf)
3205 uint32_t hash, recirc_depth;
3207 if (OVS_LIKELY(dp_packet_rss_valid(packet))) {
3208 hash = dp_packet_get_rss_hash(packet);
3210 hash = miniflow_hash_5tuple(mf, 0);
3211 dp_packet_set_rss_hash(packet, hash);
3214 /* The RSS hash must account for the recirculation depth to avoid
3215 * collisions in the exact match cache */
3216 recirc_depth = *recirc_depth_get_unsafe();
3217 if (OVS_UNLIKELY(recirc_depth)) {
3218 hash = hash_finish(hash, recirc_depth);
3219 dp_packet_set_rss_hash(packet, hash);
3224 struct packet_batch {
3225 unsigned int packet_count;
3226 unsigned int byte_count;
3229 struct dp_netdev_flow *flow;
3231 struct dp_packet *packets[NETDEV_MAX_BURST];
3235 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
3236 const struct miniflow *mf)
3238 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3239 batch->packets[batch->packet_count++] = packet;
3240 batch->byte_count += dp_packet_size(packet);
3244 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
3246 flow->batch = batch;
3249 batch->packet_count = 0;
3250 batch->byte_count = 0;
3251 batch->tcp_flags = 0;
3255 packet_batch_execute(struct packet_batch *batch,
3256 struct dp_netdev_pmd_thread *pmd,
3259 struct dp_netdev_actions *actions;
3260 struct dp_netdev_flow *flow = batch->flow;
3262 dp_netdev_flow_used(flow, batch->packet_count, batch->byte_count,
3263 batch->tcp_flags, now);
3265 actions = dp_netdev_flow_get_actions(flow);
3267 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
3268 actions->actions, actions->size);
3272 dp_netdev_queue_batches(struct dp_packet *pkt,
3273 struct dp_netdev_flow *flow, const struct miniflow *mf,
3274 struct packet_batch *batches, size_t *n_batches)
3276 struct packet_batch *batch = flow->batch;
3278 if (OVS_UNLIKELY(!batch)) {
3279 batch = &batches[(*n_batches)++];
3280 packet_batch_init(batch, flow);
3283 packet_batch_update(batch, pkt, mf);
3286 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3287 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', the
3288 * miniflow is copied into 'keys' and the packet pointer is moved at the
3289 * beginning of the 'packets' array.
3291 * The function returns the number of packets that needs to be processed in the
3292 * 'packets' array (they have been moved to the beginning of the vector).
3294 static inline size_t
3295 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
3296 size_t cnt, struct netdev_flow_key *keys,
3297 struct packet_batch batches[], size_t *n_batches)
3299 struct emc_cache *flow_cache = &pmd->flow_cache;
3300 size_t i, n_missed = 0, n_dropped = 0;
3302 for (i = 0; i < cnt; i++) {
3303 struct dp_netdev_flow *flow;
3305 if (OVS_UNLIKELY(dp_packet_size(packets[i]) < ETH_HEADER_LEN)) {
3306 dp_packet_delete(packets[i]);
3312 /* Prefetch next packet data */
3313 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3316 struct netdev_flow_key *key = &keys[n_missed];
3317 miniflow_extract(packets[i], &key->mf);
3318 key->len = 0; /* Not computed yet. */
3319 key->hash = dpif_netdev_packet_get_rss_hash(packets[i], &key->mf);
3321 flow = emc_lookup(flow_cache, key);
3322 if (OVS_LIKELY(flow)) {
3323 dp_netdev_queue_batches(packets[i], flow, &key->mf, batches,
3326 /* Exact match cache missed. Group missed packets together at
3327 * the beginning of the 'packets' array. */
3328 packets[n_missed++] = packets[i];
3332 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - n_dropped - n_missed);
3338 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3339 struct dp_packet **packets, size_t cnt,
3340 struct netdev_flow_key *keys,
3341 struct packet_batch batches[], size_t *n_batches)
3343 #if !defined(__CHECKER__) && !defined(_WIN32)
3344 const size_t PKT_ARRAY_SIZE = cnt;
3346 /* Sparse or MSVC doesn't like variable length array. */
3347 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3349 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3350 struct dp_netdev *dp = pmd->dp;
3351 struct emc_cache *flow_cache = &pmd->flow_cache;
3352 int miss_cnt = 0, lost_cnt = 0;
3356 for (i = 0; i < cnt; i++) {
3357 /* Key length is needed in all the cases, hash computed on demand. */
3358 keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf));
3360 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3361 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3362 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3363 struct ofpbuf actions, put_actions;
3366 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3367 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3369 for (i = 0; i < cnt; i++) {
3370 struct dp_netdev_flow *netdev_flow;
3371 struct ofpbuf *add_actions;
3375 if (OVS_LIKELY(rules[i])) {
3379 /* It's possible that an earlier slow path execution installed
3380 * a rule covering this flow. In this case, it's a lot cheaper
3381 * to catch it here than execute a miss. */
3382 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3384 rules[i] = &netdev_flow->cr;
3390 match.tun_md.valid = false;
3391 miniflow_expand(&keys[i].mf, &match.flow);
3393 ofpbuf_clear(&actions);
3394 ofpbuf_clear(&put_actions);
3396 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
3397 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
3398 &ufid, DPIF_UC_MISS, NULL, &actions,
3400 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3401 dp_packet_delete(packets[i]);
3406 /* The Netlink encoding of datapath flow keys cannot express
3407 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3408 * tag is interpreted as exact match on the fact that there is no
3409 * VLAN. Unless we refactor a lot of code that translates between
3410 * Netlink and struct flow representations, we have to do the same
3412 if (!match.wc.masks.vlan_tci) {
3413 match.wc.masks.vlan_tci = htons(0xffff);
3416 /* We can't allow the packet batching in the next loop to execute
3417 * the actions. Otherwise, if there are any slow path actions,
3418 * we'll send the packet up twice. */
3419 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
3420 actions.data, actions.size);
3422 add_actions = put_actions.size ? &put_actions : &actions;
3423 if (OVS_LIKELY(error != ENOSPC)) {
3424 /* XXX: There's a race window where a flow covering this packet
3425 * could have already been installed since we last did the flow
3426 * lookup before upcall. This could be solved by moving the
3427 * mutex lock outside the loop, but that's an awful long time
3428 * to be locking everyone out of making flow installs. If we
3429 * move to a per-core classifier, it would be reasonable. */
3430 ovs_mutex_lock(&pmd->flow_mutex);
3431 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3432 if (OVS_LIKELY(!netdev_flow)) {
3433 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3437 ovs_mutex_unlock(&pmd->flow_mutex);
3439 emc_insert(flow_cache, &keys[i], netdev_flow);
3443 ofpbuf_uninit(&actions);
3444 ofpbuf_uninit(&put_actions);
3445 fat_rwlock_unlock(&dp->upcall_rwlock);
3446 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3447 } else if (OVS_UNLIKELY(any_miss)) {
3448 for (i = 0; i < cnt; i++) {
3449 if (OVS_UNLIKELY(!rules[i])) {
3450 dp_packet_delete(packets[i]);
3457 for (i = 0; i < cnt; i++) {
3458 struct dp_packet *packet = packets[i];
3459 struct dp_netdev_flow *flow;
3461 if (OVS_UNLIKELY(!rules[i])) {
3465 flow = dp_netdev_flow_cast(rules[i]);
3467 emc_insert(flow_cache, &keys[i], flow);
3468 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3471 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3472 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3473 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3477 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3478 struct dp_packet **packets, int cnt)
3480 #if !defined(__CHECKER__) && !defined(_WIN32)
3481 const size_t PKT_ARRAY_SIZE = cnt;
3483 /* Sparse or MSVC doesn't like variable length array. */
3484 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3486 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3487 struct packet_batch batches[PKT_ARRAY_SIZE];
3488 long long now = time_msec();
3489 size_t newcnt, n_batches, i;
3492 newcnt = emc_processing(pmd, packets, cnt, keys, batches, &n_batches);
3493 if (OVS_UNLIKELY(newcnt)) {
3494 fast_path_processing(pmd, packets, newcnt, keys, batches, &n_batches);
3497 for (i = 0; i < n_batches; i++) {
3498 batches[i].flow->batch = NULL;
3501 for (i = 0; i < n_batches; i++) {
3502 packet_batch_execute(&batches[i], pmd, now);
3506 struct dp_netdev_execute_aux {
3507 struct dp_netdev_pmd_thread *pmd;
3511 dpif_netdev_register_dp_purge_cb(struct dpif *dpif, dp_purge_callback *cb,
3514 struct dp_netdev *dp = get_dp_netdev(dpif);
3515 dp->dp_purge_aux = aux;
3516 dp->dp_purge_cb = cb;
3520 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3523 struct dp_netdev *dp = get_dp_netdev(dpif);
3524 dp->upcall_aux = aux;
3529 dp_netdev_drop_packets(struct dp_packet **packets, int cnt, bool may_steal)
3534 for (i = 0; i < cnt; i++) {
3535 dp_packet_delete(packets[i]);
3541 push_tnl_action(const struct dp_netdev *dp,
3542 const struct nlattr *attr,
3543 struct dp_packet **packets, int cnt)
3545 struct dp_netdev_port *tun_port;
3546 const struct ovs_action_push_tnl *data;
3548 data = nl_attr_get(attr);
3550 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3554 netdev_push_header(tun_port->netdev, packets, cnt, data);
3560 dp_netdev_clone_pkt_batch(struct dp_packet **dst_pkts,
3561 struct dp_packet **src_pkts, int cnt)
3565 for (i = 0; i < cnt; i++) {
3566 dst_pkts[i] = dp_packet_clone(src_pkts[i]);
3571 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3572 const struct nlattr *a, bool may_steal)
3573 OVS_NO_THREAD_SAFETY_ANALYSIS
3575 struct dp_netdev_execute_aux *aux = aux_;
3576 uint32_t *depth = recirc_depth_get();
3577 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3578 struct dp_netdev *dp = pmd->dp;
3579 int type = nl_attr_type(a);
3580 struct dp_netdev_port *p;
3583 switch ((enum ovs_action_attr)type) {
3584 case OVS_ACTION_ATTR_OUTPUT:
3585 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3586 if (OVS_LIKELY(p)) {
3589 atomic_read_relaxed(&pmd->tx_qid, &tx_qid);
3591 netdev_send(p->netdev, tx_qid, packets, cnt, may_steal);
3596 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3597 if (*depth < MAX_RECIRC_DEPTH) {
3598 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3602 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3606 err = push_tnl_action(dp, a, packets, cnt);
3609 dp_netdev_input(pmd, packets, cnt);
3612 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3618 case OVS_ACTION_ATTR_TUNNEL_POP:
3619 if (*depth < MAX_RECIRC_DEPTH) {
3620 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3622 p = dp_netdev_lookup_port(dp, portno);
3624 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3628 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3632 err = netdev_pop_header(p->netdev, packets, cnt);
3635 for (i = 0; i < cnt; i++) {
3636 packets[i]->md.in_port.odp_port = portno;
3640 dp_netdev_input(pmd, packets, cnt);
3643 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3650 case OVS_ACTION_ATTR_USERSPACE:
3651 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3652 const struct nlattr *userdata;
3653 struct ofpbuf actions;
3657 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3658 ofpbuf_init(&actions, 0);
3660 for (i = 0; i < cnt; i++) {
3663 ofpbuf_clear(&actions);
3665 flow_extract(packets[i], &flow);
3666 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3667 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3668 DPIF_UC_ACTION, userdata,&actions,
3670 if (!error || error == ENOSPC) {
3671 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3672 actions.data, actions.size);
3673 } else if (may_steal) {
3674 dp_packet_delete(packets[i]);
3677 ofpbuf_uninit(&actions);
3678 fat_rwlock_unlock(&dp->upcall_rwlock);
3684 case OVS_ACTION_ATTR_RECIRC:
3685 if (*depth < MAX_RECIRC_DEPTH) {
3686 struct dp_packet *recirc_pkts[NETDEV_MAX_BURST];
3689 dp_netdev_clone_pkt_batch(recirc_pkts, packets, cnt);
3690 packets = recirc_pkts;
3693 for (i = 0; i < cnt; i++) {
3694 packets[i]->md.recirc_id = nl_attr_get_u32(a);
3698 dp_netdev_input(pmd, packets, cnt);
3704 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3707 case OVS_ACTION_ATTR_CT:
3708 /* If a flow with this action is slow-pathed, datapath assistance is
3709 * required to implement it. However, we don't support this action
3710 * in the userspace datapath. */
3711 VLOG_WARN("Cannot execute conntrack action in userspace.");
3714 case OVS_ACTION_ATTR_PUSH_VLAN:
3715 case OVS_ACTION_ATTR_POP_VLAN:
3716 case OVS_ACTION_ATTR_PUSH_MPLS:
3717 case OVS_ACTION_ATTR_POP_MPLS:
3718 case OVS_ACTION_ATTR_SET:
3719 case OVS_ACTION_ATTR_SET_MASKED:
3720 case OVS_ACTION_ATTR_SAMPLE:
3721 case OVS_ACTION_ATTR_HASH:
3722 case OVS_ACTION_ATTR_UNSPEC:
3723 case __OVS_ACTION_ATTR_MAX:
3727 dp_netdev_drop_packets(packets, cnt, may_steal);
3731 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3732 struct dp_packet **packets, int cnt,
3734 const struct nlattr *actions, size_t actions_len)
3736 struct dp_netdev_execute_aux aux = { pmd };
3738 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3739 actions_len, dp_execute_cb);
3742 const struct dpif_class dpif_netdev_class = {
3745 dpif_netdev_enumerate,
3746 dpif_netdev_port_open_type,
3749 dpif_netdev_destroy,
3752 dpif_netdev_get_stats,
3753 dpif_netdev_port_add,
3754 dpif_netdev_port_del,
3755 dpif_netdev_port_query_by_number,
3756 dpif_netdev_port_query_by_name,
3757 NULL, /* port_get_pid */
3758 dpif_netdev_port_dump_start,
3759 dpif_netdev_port_dump_next,
3760 dpif_netdev_port_dump_done,
3761 dpif_netdev_port_poll,
3762 dpif_netdev_port_poll_wait,
3763 dpif_netdev_flow_flush,
3764 dpif_netdev_flow_dump_create,
3765 dpif_netdev_flow_dump_destroy,
3766 dpif_netdev_flow_dump_thread_create,
3767 dpif_netdev_flow_dump_thread_destroy,
3768 dpif_netdev_flow_dump_next,
3769 dpif_netdev_operate,
3770 NULL, /* recv_set */
3771 NULL, /* handlers_set */
3772 dpif_netdev_pmd_set,
3773 dpif_netdev_queue_to_priority,
3775 NULL, /* recv_wait */
3776 NULL, /* recv_purge */
3777 dpif_netdev_register_dp_purge_cb,
3778 dpif_netdev_register_upcall_cb,
3779 dpif_netdev_enable_upcall,
3780 dpif_netdev_disable_upcall,
3781 dpif_netdev_get_datapath_version,
3782 NULL, /* ct_dump_start */
3783 NULL, /* ct_dump_next */
3784 NULL, /* ct_dump_done */
3785 NULL, /* ct_flush */
3789 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3790 const char *argv[], void *aux OVS_UNUSED)
3792 struct dp_netdev_port *old_port;
3793 struct dp_netdev_port *new_port;
3794 struct dp_netdev *dp;
3797 ovs_mutex_lock(&dp_netdev_mutex);
3798 dp = shash_find_data(&dp_netdevs, argv[1]);
3799 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3800 ovs_mutex_unlock(&dp_netdev_mutex);
3801 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3804 ovs_refcount_ref(&dp->ref_cnt);
3805 ovs_mutex_unlock(&dp_netdev_mutex);
3807 ovs_mutex_lock(&dp->port_mutex);
3808 if (get_port_by_name(dp, argv[2], &old_port)) {
3809 unixctl_command_reply_error(conn, "unknown port");
3813 port_no = u32_to_odp(atoi(argv[3]));
3814 if (!port_no || port_no == ODPP_NONE) {
3815 unixctl_command_reply_error(conn, "bad port number");
3818 if (dp_netdev_lookup_port(dp, port_no)) {
3819 unixctl_command_reply_error(conn, "port number already in use");
3823 /* Remove old port. */
3824 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3825 ovsrcu_postpone(free, old_port);
3827 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3828 new_port = xmemdup(old_port, sizeof *old_port);
3829 new_port->port_no = port_no;
3830 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3832 seq_change(dp->port_seq);
3833 unixctl_command_reply(conn, NULL);
3836 ovs_mutex_unlock(&dp->port_mutex);
3837 dp_netdev_unref(dp);
3841 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3842 const char *argv[], void *aux OVS_UNUSED)
3844 struct dp_netdev_port *port;
3845 struct dp_netdev *dp;
3847 ovs_mutex_lock(&dp_netdev_mutex);
3848 dp = shash_find_data(&dp_netdevs, argv[1]);
3849 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3850 ovs_mutex_unlock(&dp_netdev_mutex);
3851 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3854 ovs_refcount_ref(&dp->ref_cnt);
3855 ovs_mutex_unlock(&dp_netdev_mutex);
3857 ovs_mutex_lock(&dp->port_mutex);
3858 if (get_port_by_name(dp, argv[2], &port)) {
3859 unixctl_command_reply_error(conn, "unknown port");
3860 } else if (port->port_no == ODPP_LOCAL) {
3861 unixctl_command_reply_error(conn, "can't delete local port");
3863 do_del_port(dp, port);
3864 unixctl_command_reply(conn, NULL);
3866 ovs_mutex_unlock(&dp->port_mutex);
3868 dp_netdev_unref(dp);
3872 dpif_dummy_register__(const char *type)
3874 struct dpif_class *class;
3876 class = xmalloc(sizeof *class);
3877 *class = dpif_netdev_class;
3878 class->type = xstrdup(type);
3879 dp_register_provider(class);
3883 dpif_dummy_override(const char *type)
3888 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
3889 * a userland-only build. It's useful for testsuite.
3891 error = dp_unregister_provider(type);
3892 if (error == 0 || error == EAFNOSUPPORT) {
3893 dpif_dummy_register__(type);
3898 dpif_dummy_register(enum dummy_level level)
3900 if (level == DUMMY_OVERRIDE_ALL) {
3905 dp_enumerate_types(&types);
3906 SSET_FOR_EACH (type, &types) {
3907 dpif_dummy_override(type);
3909 sset_destroy(&types);
3910 } else if (level == DUMMY_OVERRIDE_SYSTEM) {
3911 dpif_dummy_override("system");
3914 dpif_dummy_register__("dummy");
3916 unixctl_command_register("dpif-dummy/change-port-number",
3917 "dp port new-number",
3918 3, 3, dpif_dummy_change_port_number, NULL);
3919 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3920 2, 2, dpif_dummy_delete_port, NULL);
3923 /* Datapath Classifier. */
3925 /* A set of rules that all have the same fields wildcarded. */
3926 struct dpcls_subtable {
3927 /* The fields are only used by writers. */
3928 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3930 /* These fields are accessed by readers. */
3931 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3932 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3933 /* 'mask' must be the last field, additional space is allocated here. */
3936 /* Initializes 'cls' as a classifier that initially contains no classification
3939 dpcls_init(struct dpcls *cls)
3941 cmap_init(&cls->subtables_map);
3942 pvector_init(&cls->subtables);
3946 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3948 pvector_remove(&cls->subtables, subtable);
3949 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3950 subtable->mask.hash);
3951 cmap_destroy(&subtable->rules);
3952 ovsrcu_postpone(free, subtable);
3955 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3956 * caller's responsibility.
3957 * May only be called after all the readers have been terminated. */
3959 dpcls_destroy(struct dpcls *cls)
3962 struct dpcls_subtable *subtable;
3964 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3965 ovs_assert(cmap_count(&subtable->rules) == 0);
3966 dpcls_destroy_subtable(cls, subtable);
3968 cmap_destroy(&cls->subtables_map);
3969 pvector_destroy(&cls->subtables);
3973 static struct dpcls_subtable *
3974 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3976 struct dpcls_subtable *subtable;
3978 /* Need to add one. */
3979 subtable = xmalloc(sizeof *subtable
3980 - sizeof subtable->mask.mf + mask->len);
3981 cmap_init(&subtable->rules);
3982 netdev_flow_key_clone(&subtable->mask, mask);
3983 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3984 pvector_insert(&cls->subtables, subtable, 0);
3985 pvector_publish(&cls->subtables);
3990 static inline struct dpcls_subtable *
3991 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3993 struct dpcls_subtable *subtable;
3995 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3996 &cls->subtables_map) {
3997 if (netdev_flow_key_equal(&subtable->mask, mask)) {
4001 return dpcls_create_subtable(cls, mask);
4004 /* Insert 'rule' into 'cls'. */
4006 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
4007 const struct netdev_flow_key *mask)
4009 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
4011 rule->mask = &subtable->mask;
4012 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
4015 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
4017 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
4019 struct dpcls_subtable *subtable;
4021 ovs_assert(rule->mask);
4023 INIT_CONTAINER(subtable, rule->mask, mask);
4025 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
4027 dpcls_destroy_subtable(cls, subtable);
4028 pvector_publish(&cls->subtables);
4032 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
4033 * in 'mask' the values in 'key' and 'target' are the same. */
4035 dpcls_rule_matches_key(const struct dpcls_rule *rule,
4036 const struct netdev_flow_key *target)
4038 const uint64_t *keyp = miniflow_get_values(&rule->flow.mf);
4039 const uint64_t *maskp = miniflow_get_values(&rule->mask->mf);
4042 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, target, rule->flow.mf.map) {
4043 if (OVS_UNLIKELY((value & *maskp++) != *keyp++)) {
4050 /* For each miniflow in 'flows' performs a classifier lookup writing the result
4051 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
4052 * NULL it is skipped.
4054 * This function is optimized for use in the userspace datapath and therefore
4055 * does not implement a lot of features available in the standard
4056 * classifier_lookup() function. Specifically, it does not implement
4057 * priorities, instead returning any rule which matches the flow.
4059 * Returns true if all flows found a corresponding rule. */
4061 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
4062 struct dpcls_rule **rules, const size_t cnt)
4064 /* The batch size 16 was experimentally found faster than 8 or 32. */
4065 typedef uint16_t map_type;
4066 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
4068 #if !defined(__CHECKER__) && !defined(_WIN32)
4069 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
4071 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
4073 map_type maps[N_MAPS];
4074 struct dpcls_subtable *subtable;
4076 memset(maps, 0xff, sizeof maps);
4077 if (cnt % MAP_BITS) {
4078 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
4080 memset(rules, 0, cnt * sizeof *rules);
4082 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
4083 const struct netdev_flow_key *mkeys = keys;
4084 struct dpcls_rule **mrules = rules;
4085 map_type remains = 0;
4088 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
4090 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
4091 uint32_t hashes[MAP_BITS];
4092 const struct cmap_node *nodes[MAP_BITS];
4093 unsigned long map = maps[m];
4097 continue; /* Skip empty maps. */
4100 /* Compute hashes for the remaining keys. */
4101 ULLONG_FOR_EACH_1(i, map) {
4102 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
4106 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
4107 /* Check results. */
4108 ULLONG_FOR_EACH_1(i, map) {
4109 struct dpcls_rule *rule;
4111 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
4112 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
4117 ULLONG_SET0(map, i); /* Did not match. */
4119 ; /* Keep Sparse happy. */
4121 maps[m] &= ~map; /* Clear the found rules. */
4125 return true; /* All found. */
4128 return false; /* Some misses. */