2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "dpif-netdev.h"
24 #include <netinet/in.h>
25 #include <sys/socket.h>
30 #include <sys/ioctl.h>
37 #include "dp-packet.h"
39 #include "dpif-provider.h"
41 #include "dynamic-string.h"
42 #include "fat-rwlock.h"
50 #include "netdev-dpdk.h"
51 #include "netdev-vport.h"
53 #include "odp-execute.h"
55 #include "ofp-print.h"
60 #include "poll-loop.h"
67 #include "tnl-neigh-cache.h"
68 #include "tnl-ports.h"
71 #include "openvswitch/vlog.h"
73 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
75 #define FLOW_DUMP_MAX_BATCH 50
76 /* Use per thread recirc_depth to prevent recirculation loop. */
77 #define MAX_RECIRC_DEPTH 5
78 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
80 /* Configuration parameters. */
81 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
83 /* Protects against changes to 'dp_netdevs'. */
84 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
86 /* Contains all 'struct dp_netdev's. */
87 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
88 = SHASH_INITIALIZER(&dp_netdevs);
90 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
92 static struct odp_support dp_netdev_support = {
93 .max_mpls_depth = SIZE_MAX,
97 /* Stores a miniflow with inline values */
99 struct netdev_flow_key {
100 uint32_t hash; /* Hash function differs for different users. */
101 uint32_t len; /* Length of the following miniflow (incl. map). */
103 uint64_t buf[FLOW_MAX_PACKET_U64S];
106 /* Exact match cache for frequently used flows
108 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
109 * search its entries for a miniflow that matches exactly the miniflow of the
110 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
112 * A cache entry holds a reference to its 'dp_netdev_flow'.
114 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
115 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
116 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
117 * value is the index of a cache entry where the miniflow could be.
123 * Each pmd_thread has its own private exact match cache.
124 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
127 #define EM_FLOW_HASH_SHIFT 13
128 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
129 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
130 #define EM_FLOW_HASH_SEGS 2
133 struct dp_netdev_flow *flow;
134 struct netdev_flow_key key; /* key.hash used for emc hash value. */
138 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
139 int sweep_idx; /* For emc_cache_slow_sweep(). */
142 /* Iterate in the exact match cache through every entry that might contain a
143 * miniflow with hash 'HASH'. */
144 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
145 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
146 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
147 i__ < EM_FLOW_HASH_SEGS; \
148 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
150 /* Simple non-wildcarding single-priority classifier. */
153 struct cmap subtables_map;
154 struct pvector subtables;
157 /* A rule to be inserted to the classifier. */
159 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
160 struct netdev_flow_key *mask; /* Subtable's mask. */
161 struct netdev_flow_key flow; /* Matching key. */
162 /* 'flow' must be the last field, additional space is allocated here. */
165 static void dpcls_init(struct dpcls *);
166 static void dpcls_destroy(struct dpcls *);
167 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
168 const struct netdev_flow_key *mask);
169 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
170 static bool dpcls_lookup(const struct dpcls *cls,
171 const struct netdev_flow_key keys[],
172 struct dpcls_rule **rules, size_t cnt);
174 /* Datapath based on the network device interface from netdev.h.
180 * Some members, marked 'const', are immutable. Accessing other members
181 * requires synchronization, as noted in more detail below.
183 * Acquisition order is, from outermost to innermost:
185 * dp_netdev_mutex (global)
189 const struct dpif_class *const class;
190 const char *const name;
192 struct ovs_refcount ref_cnt;
193 atomic_flag destroyed;
197 * Protected by RCU. Take the mutex to add or remove ports. */
198 struct ovs_mutex port_mutex;
200 struct seq *port_seq; /* Incremented whenever a port changes. */
202 /* Protects access to ofproto-dpif-upcall interface during revalidator
203 * thread synchronization. */
204 struct fat_rwlock upcall_rwlock;
205 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
208 /* Callback function for notifying the purging of dp flows (during
209 * reseting pmd deletion). */
210 dp_purge_callback *dp_purge_cb;
213 /* Stores all 'struct dp_netdev_pmd_thread's. */
214 struct cmap poll_threads;
216 /* Protects the access of the 'struct dp_netdev_pmd_thread'
217 * instance for non-pmd thread. */
218 struct ovs_mutex non_pmd_mutex;
220 /* Each pmd thread will store its pointer to
221 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
222 ovsthread_key_t per_pmd_key;
224 /* Number of rx queues for each dpdk interface and the cpu mask
225 * for pin of pmd threads. */
228 uint64_t last_tnl_conf_seq;
231 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
235 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
236 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
237 DP_STAT_MISS, /* Packets that did not match. */
238 DP_STAT_LOST, /* Packets not passed up to the client. */
242 enum pmd_cycles_counter_type {
243 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
244 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
248 /* A port in a netdev-based datapath. */
249 struct dp_netdev_port {
251 struct netdev *netdev;
252 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
253 struct netdev_saved_flags *sf;
254 struct netdev_rxq **rxq;
255 struct ovs_refcount ref_cnt;
256 char *type; /* Port type as requested by user. */
259 /* Contained by struct dp_netdev_flow's 'stats' member. */
260 struct dp_netdev_flow_stats {
261 atomic_llong used; /* Last used time, in monotonic msecs. */
262 atomic_ullong packet_count; /* Number of packets matched. */
263 atomic_ullong byte_count; /* Number of bytes matched. */
264 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
267 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
273 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
274 * its pmd thread's classifier. The text below calls this classifier 'cls'.
279 * The thread safety rules described here for "struct dp_netdev_flow" are
280 * motivated by two goals:
282 * - Prevent threads that read members of "struct dp_netdev_flow" from
283 * reading bad data due to changes by some thread concurrently modifying
286 * - Prevent two threads making changes to members of a given "struct
287 * dp_netdev_flow" from interfering with each other.
293 * A flow 'flow' may be accessed without a risk of being freed during an RCU
294 * grace period. Code that needs to hold onto a flow for a while
295 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
297 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
298 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
301 * Some members, marked 'const', are immutable. Accessing other members
302 * requires synchronization, as noted in more detail below.
304 struct dp_netdev_flow {
305 const struct flow flow; /* Unmasked flow that created this entry. */
306 /* Hash table index by unmasked flow. */
307 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
309 const ovs_u128 ufid; /* Unique flow identifier. */
310 const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */
313 /* Number of references.
314 * The classifier owns one reference.
315 * Any thread trying to keep a rule from being freed should hold its own
317 struct ovs_refcount ref_cnt;
322 struct dp_netdev_flow_stats stats;
325 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
327 /* While processing a group of input packets, the datapath uses the next
328 * member to store a pointer to the output batch for the flow. It is
329 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
330 * packet_batch_init() and packet_batch_execute()). */
331 struct packet_batch *batch;
333 /* Packet classification. */
334 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
335 /* 'cr' must be the last member. */
338 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
339 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
340 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
343 /* A set of datapath actions within a "struct dp_netdev_flow".
349 * A struct dp_netdev_actions 'actions' is protected with RCU. */
350 struct dp_netdev_actions {
351 /* These members are immutable: they do not change during the struct's
353 unsigned int size; /* Size of 'actions', in bytes. */
354 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
357 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
359 struct dp_netdev_actions *dp_netdev_flow_get_actions(
360 const struct dp_netdev_flow *);
361 static void dp_netdev_actions_free(struct dp_netdev_actions *);
363 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
364 struct dp_netdev_pmd_stats {
365 /* Indexed by DP_STAT_*. */
366 atomic_ullong n[DP_N_STATS];
369 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
370 struct dp_netdev_pmd_cycles {
371 /* Indexed by PMD_CYCLES_*. */
372 atomic_ullong n[PMD_N_CYCLES];
375 /* 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 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_on_numa(struct dp_netdev *dp, int numa_id)
1290 struct dp_netdev_pmd_thread *pmd;
1293 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1294 if (pmd->numa_id == numa_id) {
1302 /* Returns 'true' if there is a port with pmd netdev and the netdev
1303 * is on numa node 'numa_id'. */
1305 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1307 struct dp_netdev_port *port;
1309 CMAP_FOR_EACH (port, node, &dp->ports) {
1310 if (netdev_is_pmd(port->netdev)
1311 && netdev_get_numa_id(port->netdev) == numa_id) {
1321 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1322 OVS_REQUIRES(dp->port_mutex)
1324 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1325 seq_change(dp->port_seq);
1326 if (netdev_is_pmd(port->netdev)) {
1327 int numa_id = netdev_get_numa_id(port->netdev);
1329 /* PMD threads can not be on invalid numa node. */
1330 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
1331 /* If there is no netdev on the numa node, deletes the pmd threads
1332 * for that numa. Else, deletes the queues from polling lists. */
1333 if (!has_pmd_port_for_numa(dp, numa_id)) {
1334 dp_netdev_del_pmds_on_numa(dp, numa_id);
1336 struct dp_netdev_pmd_thread *pmd;
1337 struct rxq_poll *poll, *next;
1339 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1340 if (pmd->numa_id == numa_id) {
1343 ovs_mutex_lock(&pmd->poll_mutex);
1344 LIST_FOR_EACH_SAFE (poll, next, node, &pmd->poll_list) {
1345 if (poll->port == port) {
1347 port_unref(poll->port);
1348 list_remove(&poll->node);
1353 ovs_mutex_unlock(&pmd->poll_mutex);
1355 dp_netdev_reload_pmd__(pmd);
1366 answer_port_query(const struct dp_netdev_port *port,
1367 struct dpif_port *dpif_port)
1369 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1370 dpif_port->type = xstrdup(port->type);
1371 dpif_port->port_no = port->port_no;
1375 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1376 struct dpif_port *dpif_port)
1378 struct dp_netdev *dp = get_dp_netdev(dpif);
1379 struct dp_netdev_port *port;
1382 error = get_port_by_number(dp, port_no, &port);
1383 if (!error && dpif_port) {
1384 answer_port_query(port, dpif_port);
1391 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1392 struct dpif_port *dpif_port)
1394 struct dp_netdev *dp = get_dp_netdev(dpif);
1395 struct dp_netdev_port *port;
1398 ovs_mutex_lock(&dp->port_mutex);
1399 error = get_port_by_name(dp, devname, &port);
1400 if (!error && dpif_port) {
1401 answer_port_query(port, dpif_port);
1403 ovs_mutex_unlock(&dp->port_mutex);
1409 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1411 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1415 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1417 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1418 ovsrcu_postpone(dp_netdev_flow_free, flow);
1423 dp_netdev_flow_hash(const ovs_u128 *ufid)
1425 return ufid->u32[0];
1429 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1430 struct dp_netdev_flow *flow)
1431 OVS_REQUIRES(pmd->flow_mutex)
1433 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1435 dpcls_remove(&pmd->cls, &flow->cr);
1436 flow->cr.mask = NULL; /* Accessing rule's mask after this is not safe. */
1438 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1441 dp_netdev_flow_unref(flow);
1445 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1447 struct dp_netdev_flow *netdev_flow;
1449 ovs_mutex_lock(&pmd->flow_mutex);
1450 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1451 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1453 ovs_mutex_unlock(&pmd->flow_mutex);
1457 dpif_netdev_flow_flush(struct dpif *dpif)
1459 struct dp_netdev *dp = get_dp_netdev(dpif);
1460 struct dp_netdev_pmd_thread *pmd;
1462 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1463 dp_netdev_pmd_flow_flush(pmd);
1469 struct dp_netdev_port_state {
1470 struct cmap_position position;
1475 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1477 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1482 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1483 struct dpif_port *dpif_port)
1485 struct dp_netdev_port_state *state = state_;
1486 struct dp_netdev *dp = get_dp_netdev(dpif);
1487 struct cmap_node *node;
1490 node = cmap_next_position(&dp->ports, &state->position);
1492 struct dp_netdev_port *port;
1494 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1497 state->name = xstrdup(netdev_get_name(port->netdev));
1498 dpif_port->name = state->name;
1499 dpif_port->type = port->type;
1500 dpif_port->port_no = port->port_no;
1511 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1513 struct dp_netdev_port_state *state = state_;
1520 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1522 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1523 uint64_t new_port_seq;
1526 new_port_seq = seq_read(dpif->dp->port_seq);
1527 if (dpif->last_port_seq != new_port_seq) {
1528 dpif->last_port_seq = new_port_seq;
1538 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1540 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1542 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1545 static struct dp_netdev_flow *
1546 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1548 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1551 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1553 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1556 /* netdev_flow_key utilities.
1558 * netdev_flow_key is basically a miniflow. We use these functions
1559 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1560 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1562 * - Since we are dealing exclusively with miniflows created by
1563 * miniflow_extract(), if the map is different the miniflow is different.
1564 * Therefore we can be faster by comparing the map and the miniflow in a
1566 * - These functions can be inlined by the compiler. */
1568 /* Given the number of bits set in miniflow's maps, returns the size of the
1569 * 'netdev_flow_key.mf' */
1570 static inline size_t
1571 netdev_flow_key_size(size_t flow_u64s)
1573 return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s);
1577 netdev_flow_key_equal(const struct netdev_flow_key *a,
1578 const struct netdev_flow_key *b)
1580 /* 'b->len' may be not set yet. */
1581 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1584 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1585 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1586 * generated by miniflow_extract. */
1588 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1589 const struct miniflow *mf)
1591 return !memcmp(&key->mf, mf, key->len);
1595 netdev_flow_key_clone(struct netdev_flow_key *dst,
1596 const struct netdev_flow_key *src)
1599 offsetof(struct netdev_flow_key, mf) + src->len);
1604 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1605 const struct flow *src)
1607 struct dp_packet packet;
1608 uint64_t buf_stub[512 / 8];
1610 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1611 pkt_metadata_from_flow(&packet.md, src);
1612 flow_compose(&packet, src);
1613 miniflow_extract(&packet, &dst->mf);
1614 dp_packet_uninit(&packet);
1616 dst->len = netdev_flow_key_size(miniflow_n_values(&dst->mf));
1617 dst->hash = 0; /* Not computed yet. */
1620 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1622 netdev_flow_mask_init(struct netdev_flow_key *mask,
1623 const struct match *match)
1625 uint64_t *dst = miniflow_values(&mask->mf);
1626 struct flowmap fmap;
1630 /* Only check masks that make sense for the flow. */
1631 flow_wc_map(&match->flow, &fmap);
1632 flowmap_init(&mask->mf.map);
1634 FLOWMAP_FOR_EACH_INDEX(idx, fmap) {
1635 uint64_t mask_u64 = flow_u64_value(&match->wc.masks, idx);
1638 flowmap_set(&mask->mf.map, idx, 1);
1640 hash = hash_add64(hash, mask_u64);
1646 FLOWMAP_FOR_EACH_MAP (map, mask->mf.map) {
1647 hash = hash_add64(hash, map);
1650 size_t n = dst - miniflow_get_values(&mask->mf);
1652 mask->hash = hash_finish(hash, n * 8);
1653 mask->len = netdev_flow_key_size(n);
1656 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1658 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1659 const struct flow *flow,
1660 const struct netdev_flow_key *mask)
1662 uint64_t *dst_u64 = miniflow_values(&dst->mf);
1663 const uint64_t *mask_u64 = miniflow_get_values(&mask->mf);
1667 dst->len = mask->len;
1668 dst->mf = mask->mf; /* Copy maps. */
1670 FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf.map) {
1671 *dst_u64 = value & *mask_u64++;
1672 hash = hash_add64(hash, *dst_u64++);
1674 dst->hash = hash_finish(hash,
1675 (dst_u64 - miniflow_get_values(&dst->mf)) * 8);
1678 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
1679 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
1680 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
1682 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1684 static inline uint32_t
1685 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1686 const struct netdev_flow_key *mask)
1688 const uint64_t *p = miniflow_get_values(&mask->mf);
1692 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, key, mask->mf.map) {
1693 hash = hash_add64(hash, value & *p++);
1696 return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8);
1700 emc_entry_alive(struct emc_entry *ce)
1702 return ce->flow && !ce->flow->dead;
1706 emc_clear_entry(struct emc_entry *ce)
1709 dp_netdev_flow_unref(ce->flow);
1715 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1716 const struct netdev_flow_key *key)
1718 if (ce->flow != flow) {
1720 dp_netdev_flow_unref(ce->flow);
1723 if (dp_netdev_flow_ref(flow)) {
1730 netdev_flow_key_clone(&ce->key, key);
1735 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1736 struct dp_netdev_flow *flow)
1738 struct emc_entry *to_be_replaced = NULL;
1739 struct emc_entry *current_entry;
1741 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1742 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1743 /* We found the entry with the 'mf' miniflow */
1744 emc_change_entry(current_entry, flow, NULL);
1748 /* Replacement policy: put the flow in an empty (not alive) entry, or
1749 * in the first entry where it can be */
1751 || (emc_entry_alive(to_be_replaced)
1752 && !emc_entry_alive(current_entry))
1753 || current_entry->key.hash < to_be_replaced->key.hash) {
1754 to_be_replaced = current_entry;
1757 /* We didn't find the miniflow in the cache.
1758 * The 'to_be_replaced' entry is where the new flow will be stored */
1760 emc_change_entry(to_be_replaced, flow, key);
1763 static inline struct dp_netdev_flow *
1764 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1766 struct emc_entry *current_entry;
1768 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1769 if (current_entry->key.hash == key->hash
1770 && emc_entry_alive(current_entry)
1771 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1773 /* We found the entry with the 'key->mf' miniflow */
1774 return current_entry->flow;
1781 static struct dp_netdev_flow *
1782 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1783 const struct netdev_flow_key *key)
1785 struct dp_netdev_flow *netdev_flow;
1786 struct dpcls_rule *rule;
1788 dpcls_lookup(&pmd->cls, key, &rule, 1);
1789 netdev_flow = dp_netdev_flow_cast(rule);
1794 static struct dp_netdev_flow *
1795 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1796 const ovs_u128 *ufidp, const struct nlattr *key,
1799 struct dp_netdev_flow *netdev_flow;
1803 /* If a UFID is not provided, determine one based on the key. */
1804 if (!ufidp && key && key_len
1805 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1806 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1811 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1813 if (ovs_u128_equals(&netdev_flow->ufid, ufidp)) {
1823 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1824 struct dpif_flow_stats *stats)
1826 struct dp_netdev_flow *netdev_flow;
1827 unsigned long long n;
1831 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1833 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1834 stats->n_packets = n;
1835 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1837 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1839 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1840 stats->tcp_flags = flags;
1843 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1844 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1845 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1848 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1849 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1850 struct dpif_flow *flow, bool terse)
1853 memset(flow, 0, sizeof *flow);
1855 struct flow_wildcards wc;
1856 struct dp_netdev_actions *actions;
1858 struct odp_flow_key_parms odp_parms = {
1859 .flow = &netdev_flow->flow,
1861 .support = dp_netdev_support,
1864 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1867 offset = key_buf->size;
1868 flow->key = ofpbuf_tail(key_buf);
1869 odp_parms.odp_in_port = netdev_flow->flow.in_port.odp_port;
1870 odp_flow_key_from_flow(&odp_parms, key_buf);
1871 flow->key_len = key_buf->size - offset;
1874 offset = mask_buf->size;
1875 flow->mask = ofpbuf_tail(mask_buf);
1876 odp_parms.odp_in_port = wc.masks.in_port.odp_port;
1877 odp_parms.key_buf = key_buf;
1878 odp_flow_key_from_mask(&odp_parms, mask_buf);
1879 flow->mask_len = mask_buf->size - offset;
1882 actions = dp_netdev_flow_get_actions(netdev_flow);
1883 flow->actions = actions->actions;
1884 flow->actions_len = actions->size;
1887 flow->ufid = netdev_flow->ufid;
1888 flow->ufid_present = true;
1889 flow->pmd_id = netdev_flow->pmd_id;
1890 get_dpif_flow_stats(netdev_flow, &flow->stats);
1894 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1895 const struct nlattr *mask_key,
1896 uint32_t mask_key_len, const struct flow *flow,
1897 struct flow_wildcards *wc)
1899 enum odp_key_fitness fitness;
1901 fitness = odp_flow_key_to_mask_udpif(mask_key, mask_key_len, key,
1904 /* This should not happen: it indicates that
1905 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1906 * disagree on the acceptable form of a mask. Log the problem
1907 * as an error, with enough details to enable debugging. */
1908 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1910 if (!VLOG_DROP_ERR(&rl)) {
1914 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1916 VLOG_ERR("internal error parsing flow mask %s (%s)",
1917 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1928 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1933 if (odp_flow_key_to_flow_udpif(key, key_len, flow)) {
1934 /* This should not happen: it indicates that odp_flow_key_from_flow()
1935 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1936 * flow. Log the problem as an error, with enough details to enable
1938 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1940 if (!VLOG_DROP_ERR(&rl)) {
1944 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1945 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1952 in_port = flow->in_port.odp_port;
1953 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1957 /* Userspace datapath doesn't support conntrack. */
1958 if (flow->ct_state || flow->ct_zone || flow->ct_mark
1959 || !ovs_u128_is_zero(&flow->ct_label)) {
1967 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1969 struct dp_netdev *dp = get_dp_netdev(dpif);
1970 struct dp_netdev_flow *netdev_flow;
1971 struct dp_netdev_pmd_thread *pmd;
1972 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1973 ? NON_PMD_CORE_ID : get->pmd_id;
1976 pmd = dp_netdev_get_pmd(dp, pmd_id);
1981 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1984 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1989 dp_netdev_pmd_unref(pmd);
1995 static struct dp_netdev_flow *
1996 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
1997 struct match *match, const ovs_u128 *ufid,
1998 const struct nlattr *actions, size_t actions_len)
1999 OVS_REQUIRES(pmd->flow_mutex)
2001 struct dp_netdev_flow *flow;
2002 struct netdev_flow_key mask;
2004 netdev_flow_mask_init(&mask, match);
2005 /* Make sure wc does not have metadata. */
2006 ovs_assert(!FLOWMAP_HAS_FIELD(&mask.mf.map, metadata)
2007 && !FLOWMAP_HAS_FIELD(&mask.mf.map, regs));
2009 /* Do not allocate extra space. */
2010 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
2011 memset(&flow->stats, 0, sizeof flow->stats);
2014 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
2015 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
2016 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
2017 ovs_refcount_init(&flow->ref_cnt);
2018 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
2020 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
2021 dpcls_insert(&pmd->cls, &flow->cr, &mask);
2023 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
2024 dp_netdev_flow_hash(&flow->ufid));
2026 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2028 struct ds ds = DS_EMPTY_INITIALIZER;
2030 match.tun_md.valid = false;
2031 match.flow = flow->flow;
2032 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2034 ds_put_cstr(&ds, "flow_add: ");
2035 odp_format_ufid(ufid, &ds);
2036 ds_put_cstr(&ds, " ");
2037 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2038 ds_put_cstr(&ds, ", actions:");
2039 format_odp_actions(&ds, actions, actions_len);
2041 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2050 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2052 struct dp_netdev *dp = get_dp_netdev(dpif);
2053 struct dp_netdev_flow *netdev_flow;
2054 struct netdev_flow_key key;
2055 struct dp_netdev_pmd_thread *pmd;
2058 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2059 ? NON_PMD_CORE_ID : put->pmd_id;
2062 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2066 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2067 put->mask, put->mask_len,
2068 &match.flow, &match.wc);
2073 pmd = dp_netdev_get_pmd(dp, pmd_id);
2078 /* Must produce a netdev_flow_key for lookup.
2079 * This interface is no longer performance critical, since it is not used
2080 * for upcall processing any more. */
2081 netdev_flow_key_from_flow(&key, &match.flow);
2086 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2089 ovs_mutex_lock(&pmd->flow_mutex);
2090 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2092 if (put->flags & DPIF_FP_CREATE) {
2093 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2095 memset(put->stats, 0, sizeof *put->stats);
2097 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2107 if (put->flags & DPIF_FP_MODIFY
2108 && flow_equal(&match.flow, &netdev_flow->flow)) {
2109 struct dp_netdev_actions *new_actions;
2110 struct dp_netdev_actions *old_actions;
2112 new_actions = dp_netdev_actions_create(put->actions,
2115 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2116 ovsrcu_set(&netdev_flow->actions, new_actions);
2119 get_dpif_flow_stats(netdev_flow, put->stats);
2121 if (put->flags & DPIF_FP_ZERO_STATS) {
2122 /* XXX: The userspace datapath uses thread local statistics
2123 * (for flows), which should be updated only by the owning
2124 * thread. Since we cannot write on stats memory here,
2125 * we choose not to support this flag. Please note:
2126 * - This feature is currently used only by dpctl commands with
2128 * - Should the need arise, this operation can be implemented
2129 * by keeping a base value (to be update here) for each
2130 * counter, and subtracting it before outputting the stats */
2134 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2135 } else if (put->flags & DPIF_FP_CREATE) {
2138 /* Overlapping flow. */
2142 ovs_mutex_unlock(&pmd->flow_mutex);
2143 dp_netdev_pmd_unref(pmd);
2149 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2151 struct dp_netdev *dp = get_dp_netdev(dpif);
2152 struct dp_netdev_flow *netdev_flow;
2153 struct dp_netdev_pmd_thread *pmd;
2154 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2155 ? NON_PMD_CORE_ID : del->pmd_id;
2158 pmd = dp_netdev_get_pmd(dp, pmd_id);
2163 ovs_mutex_lock(&pmd->flow_mutex);
2164 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2168 get_dpif_flow_stats(netdev_flow, del->stats);
2170 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2174 ovs_mutex_unlock(&pmd->flow_mutex);
2175 dp_netdev_pmd_unref(pmd);
2180 struct dpif_netdev_flow_dump {
2181 struct dpif_flow_dump up;
2182 struct cmap_position poll_thread_pos;
2183 struct cmap_position flow_pos;
2184 struct dp_netdev_pmd_thread *cur_pmd;
2186 struct ovs_mutex mutex;
2189 static struct dpif_netdev_flow_dump *
2190 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2192 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2195 static struct dpif_flow_dump *
2196 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2198 struct dpif_netdev_flow_dump *dump;
2200 dump = xzalloc(sizeof *dump);
2201 dpif_flow_dump_init(&dump->up, dpif_);
2202 dump->up.terse = terse;
2203 ovs_mutex_init(&dump->mutex);
2209 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2211 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2213 ovs_mutex_destroy(&dump->mutex);
2218 struct dpif_netdev_flow_dump_thread {
2219 struct dpif_flow_dump_thread up;
2220 struct dpif_netdev_flow_dump *dump;
2221 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2222 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2225 static struct dpif_netdev_flow_dump_thread *
2226 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2228 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2231 static struct dpif_flow_dump_thread *
2232 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2234 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2235 struct dpif_netdev_flow_dump_thread *thread;
2237 thread = xmalloc(sizeof *thread);
2238 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2239 thread->dump = dump;
2244 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2246 struct dpif_netdev_flow_dump_thread *thread
2247 = dpif_netdev_flow_dump_thread_cast(thread_);
2253 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2254 struct dpif_flow *flows, int max_flows)
2256 struct dpif_netdev_flow_dump_thread *thread
2257 = dpif_netdev_flow_dump_thread_cast(thread_);
2258 struct dpif_netdev_flow_dump *dump = thread->dump;
2259 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2263 ovs_mutex_lock(&dump->mutex);
2264 if (!dump->status) {
2265 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2266 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2267 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2268 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2270 /* First call to dump_next(), extracts the first pmd thread.
2271 * If there is no pmd thread, returns immediately. */
2273 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2275 ovs_mutex_unlock(&dump->mutex);
2282 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2283 struct cmap_node *node;
2285 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2289 netdev_flows[n_flows] = CONTAINER_OF(node,
2290 struct dp_netdev_flow,
2293 /* When finishing dumping the current pmd thread, moves to
2295 if (n_flows < flow_limit) {
2296 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2297 dp_netdev_pmd_unref(pmd);
2298 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2304 /* Keeps the reference to next caller. */
2305 dump->cur_pmd = pmd;
2307 /* If the current dump is empty, do not exit the loop, since the
2308 * remaining pmds could have flows to be dumped. Just dumps again
2309 * on the new 'pmd'. */
2312 ovs_mutex_unlock(&dump->mutex);
2314 for (i = 0; i < n_flows; i++) {
2315 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2316 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2317 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2318 struct dpif_flow *f = &flows[i];
2319 struct ofpbuf key, mask;
2321 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2322 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2323 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2331 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2332 OVS_NO_THREAD_SAFETY_ANALYSIS
2334 struct dp_netdev *dp = get_dp_netdev(dpif);
2335 struct dp_netdev_pmd_thread *pmd;
2336 struct dp_packet *pp;
2338 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2339 dp_packet_size(execute->packet) > UINT16_MAX) {
2343 /* Tries finding the 'pmd'. If NULL is returned, that means
2344 * the current thread is a non-pmd thread and should use
2345 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2346 pmd = ovsthread_getspecific(dp->per_pmd_key);
2348 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2351 /* If the current thread is non-pmd thread, acquires
2352 * the 'non_pmd_mutex'. */
2353 if (pmd->core_id == NON_PMD_CORE_ID) {
2354 ovs_mutex_lock(&dp->non_pmd_mutex);
2355 ovs_mutex_lock(&dp->port_mutex);
2358 pp = execute->packet;
2359 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2360 execute->actions_len);
2361 if (pmd->core_id == NON_PMD_CORE_ID) {
2362 dp_netdev_pmd_unref(pmd);
2363 ovs_mutex_unlock(&dp->port_mutex);
2364 ovs_mutex_unlock(&dp->non_pmd_mutex);
2371 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2375 for (i = 0; i < n_ops; i++) {
2376 struct dpif_op *op = ops[i];
2379 case DPIF_OP_FLOW_PUT:
2380 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2383 case DPIF_OP_FLOW_DEL:
2384 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2387 case DPIF_OP_EXECUTE:
2388 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2391 case DPIF_OP_FLOW_GET:
2392 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2398 /* Returns true if the configuration for rx queues or cpu mask
2401 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2403 if (dp->n_dpdk_rxqs != rxqs) {
2406 if (dp->pmd_cmask != NULL && cmask != NULL) {
2407 return strcmp(dp->pmd_cmask, cmask);
2409 return (dp->pmd_cmask != NULL || cmask != NULL);
2414 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2416 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2418 struct dp_netdev *dp = get_dp_netdev(dpif);
2420 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2421 struct dp_netdev_port *port;
2423 dp_netdev_destroy_all_pmds(dp);
2425 CMAP_FOR_EACH (port, node, &dp->ports) {
2426 if (netdev_is_pmd(port->netdev)) {
2429 /* Closes the existing 'rxq's. */
2430 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2431 netdev_rxq_close(port->rxq[i]);
2432 port->rxq[i] = NULL;
2435 /* Sets the new rx queue config. */
2436 err = netdev_set_multiq(port->netdev,
2437 ovs_numa_get_n_cores() + 1,
2439 if (err && (err != EOPNOTSUPP)) {
2440 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2441 " %u", netdev_get_name(port->netdev),
2446 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2447 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2448 * netdev_n_rxq(port->netdev));
2449 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2450 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2454 dp->n_dpdk_rxqs = n_rxqs;
2456 /* Reconfigures the cpu mask. */
2457 ovs_numa_set_cpu_mask(cmask);
2458 free(dp->pmd_cmask);
2459 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2461 /* Restores the non-pmd. */
2462 dp_netdev_set_nonpmd(dp);
2463 /* Restores all pmd threads. */
2464 dp_netdev_reset_pmd_threads(dp);
2471 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2472 uint32_t queue_id, uint32_t *priority)
2474 *priority = queue_id;
2479 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2480 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2481 struct dp_netdev_actions *
2482 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2484 struct dp_netdev_actions *netdev_actions;
2486 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2487 memcpy(netdev_actions->actions, actions, size);
2488 netdev_actions->size = size;
2490 return netdev_actions;
2493 struct dp_netdev_actions *
2494 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2496 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2500 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2505 static inline unsigned long long
2506 cycles_counter(void)
2509 return rte_get_tsc_cycles();
2515 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2516 extern struct ovs_mutex cycles_counter_fake_mutex;
2518 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2520 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2521 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2522 OVS_NO_THREAD_SAFETY_ANALYSIS
2524 pmd->last_cycles = cycles_counter();
2527 /* Stop counting cycles and add them to the counter 'type' */
2529 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2530 enum pmd_cycles_counter_type type)
2531 OVS_RELEASES(&cycles_counter_fake_mutex)
2532 OVS_NO_THREAD_SAFETY_ANALYSIS
2534 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2536 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2540 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2541 struct dp_netdev_port *port,
2542 struct netdev_rxq *rxq)
2544 struct dp_packet *packets[NETDEV_MAX_BURST];
2547 cycles_count_start(pmd);
2548 error = netdev_rxq_recv(rxq, packets, &cnt);
2549 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2553 *recirc_depth_get() = 0;
2555 /* XXX: initialize md in netdev implementation. */
2556 for (i = 0; i < cnt; i++) {
2557 pkt_metadata_init(&packets[i]->md, port->port_no);
2559 cycles_count_start(pmd);
2560 dp_netdev_input(pmd, packets, cnt);
2561 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2562 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2563 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2565 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2566 netdev_get_name(port->netdev), ovs_strerror(error));
2570 /* Return true if needs to revalidate datapath flows. */
2572 dpif_netdev_run(struct dpif *dpif)
2574 struct dp_netdev_port *port;
2575 struct dp_netdev *dp = get_dp_netdev(dpif);
2576 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2578 uint64_t new_tnl_seq;
2580 ovs_mutex_lock(&dp->non_pmd_mutex);
2581 CMAP_FOR_EACH (port, node, &dp->ports) {
2582 if (!netdev_is_pmd(port->netdev)) {
2585 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2586 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2590 ovs_mutex_unlock(&dp->non_pmd_mutex);
2591 dp_netdev_pmd_unref(non_pmd);
2593 tnl_neigh_cache_run();
2595 new_tnl_seq = seq_read(tnl_conf_seq);
2597 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2598 dp->last_tnl_conf_seq = new_tnl_seq;
2605 dpif_netdev_wait(struct dpif *dpif)
2607 struct dp_netdev_port *port;
2608 struct dp_netdev *dp = get_dp_netdev(dpif);
2610 ovs_mutex_lock(&dp_netdev_mutex);
2611 CMAP_FOR_EACH (port, node, &dp->ports) {
2612 if (!netdev_is_pmd(port->netdev)) {
2615 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2616 netdev_rxq_wait(port->rxq[i]);
2620 ovs_mutex_unlock(&dp_netdev_mutex);
2621 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2625 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2626 struct rxq_poll **ppoll_list, int poll_cnt)
2627 OVS_REQUIRES(pmd->poll_mutex)
2629 struct rxq_poll *poll_list = *ppoll_list;
2630 struct rxq_poll *poll;
2633 for (i = 0; i < poll_cnt; i++) {
2634 port_unref(poll_list[i].port);
2637 poll_list = xrealloc(poll_list, pmd->poll_cnt * sizeof *poll_list);
2640 LIST_FOR_EACH (poll, node, &pmd->poll_list) {
2641 port_ref(poll->port);
2642 poll_list[i++] = *poll;
2645 *ppoll_list = poll_list;
2646 return pmd->poll_cnt;
2650 pmd_thread_main(void *f_)
2652 struct dp_netdev_pmd_thread *pmd = f_;
2653 unsigned int lc = 0;
2654 struct rxq_poll *poll_list;
2655 unsigned int port_seq = PMD_INITIAL_SEQ;
2662 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2663 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2664 pmd_thread_setaffinity_cpu(pmd->core_id);
2666 emc_cache_init(&pmd->flow_cache);
2668 ovs_mutex_lock(&pmd->poll_mutex);
2669 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2670 ovs_mutex_unlock(&pmd->poll_mutex);
2672 /* List port/core affinity */
2673 for (i = 0; i < poll_cnt; i++) {
2674 VLOG_INFO("Core %d processing port \'%s\'\n", pmd->core_id,
2675 netdev_get_name(poll_list[i].port->netdev));
2678 /* Signal here to make sure the pmd finishes
2679 * reloading the updated configuration. */
2680 dp_netdev_pmd_reload_done(pmd);
2683 for (i = 0; i < poll_cnt; i++) {
2684 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2692 emc_cache_slow_sweep(&pmd->flow_cache);
2693 coverage_try_clear();
2696 atomic_read_relaxed(&pmd->change_seq, &seq);
2697 if (seq != port_seq) {
2704 emc_cache_uninit(&pmd->flow_cache);
2706 if (!latch_is_set(&pmd->exit_latch)){
2710 for (i = 0; i < poll_cnt; i++) {
2711 port_unref(poll_list[i].port);
2714 dp_netdev_pmd_reload_done(pmd);
2721 dp_netdev_disable_upcall(struct dp_netdev *dp)
2722 OVS_ACQUIRES(dp->upcall_rwlock)
2724 fat_rwlock_wrlock(&dp->upcall_rwlock);
2728 dpif_netdev_disable_upcall(struct dpif *dpif)
2729 OVS_NO_THREAD_SAFETY_ANALYSIS
2731 struct dp_netdev *dp = get_dp_netdev(dpif);
2732 dp_netdev_disable_upcall(dp);
2736 dp_netdev_enable_upcall(struct dp_netdev *dp)
2737 OVS_RELEASES(dp->upcall_rwlock)
2739 fat_rwlock_unlock(&dp->upcall_rwlock);
2743 dpif_netdev_enable_upcall(struct dpif *dpif)
2744 OVS_NO_THREAD_SAFETY_ANALYSIS
2746 struct dp_netdev *dp = get_dp_netdev(dpif);
2747 dp_netdev_enable_upcall(dp);
2751 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2753 ovs_mutex_lock(&pmd->cond_mutex);
2754 xpthread_cond_signal(&pmd->cond);
2755 ovs_mutex_unlock(&pmd->cond_mutex);
2758 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2759 * the pointer if succeeds, otherwise, NULL.
2761 * Caller must unrefs the returned reference. */
2762 static struct dp_netdev_pmd_thread *
2763 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2765 struct dp_netdev_pmd_thread *pmd;
2766 const struct cmap_node *pnode;
2768 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2772 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2774 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2777 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2779 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2781 struct dp_netdev_pmd_thread *non_pmd;
2783 non_pmd = xzalloc(sizeof *non_pmd);
2784 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2788 /* Caller must have valid pointer to 'pmd'. */
2790 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2792 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2796 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2798 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2799 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2803 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2804 * fails, keeps checking for next node until reaching the end of cmap.
2806 * Caller must unrefs the returned reference. */
2807 static struct dp_netdev_pmd_thread *
2808 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2810 struct dp_netdev_pmd_thread *next;
2813 struct cmap_node *node;
2815 node = cmap_next_position(&dp->poll_threads, pos);
2816 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2818 } while (next && !dp_netdev_pmd_try_ref(next));
2824 core_id_to_qid(unsigned core_id)
2826 if (core_id != NON_PMD_CORE_ID) {
2829 return ovs_numa_get_n_cores();
2833 /* Configures the 'pmd' based on the input argument. */
2835 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2836 int index, unsigned core_id, int numa_id)
2840 pmd->core_id = core_id;
2841 pmd->tx_qid = core_id_to_qid(core_id);
2842 pmd->numa_id = numa_id;
2845 ovs_refcount_init(&pmd->ref_cnt);
2846 latch_init(&pmd->exit_latch);
2847 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2848 xpthread_cond_init(&pmd->cond, NULL);
2849 ovs_mutex_init(&pmd->cond_mutex);
2850 ovs_mutex_init(&pmd->flow_mutex);
2851 ovs_mutex_init(&pmd->poll_mutex);
2852 dpcls_init(&pmd->cls);
2853 cmap_init(&pmd->flow_table);
2854 list_init(&pmd->poll_list);
2855 /* init the 'flow_cache' since there is no
2856 * actual thread created for NON_PMD_CORE_ID. */
2857 if (core_id == NON_PMD_CORE_ID) {
2858 emc_cache_init(&pmd->flow_cache);
2860 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2861 hash_int(core_id, 0));
2865 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2867 dp_netdev_pmd_flow_flush(pmd);
2868 dpcls_destroy(&pmd->cls);
2869 cmap_destroy(&pmd->flow_table);
2870 ovs_mutex_destroy(&pmd->flow_mutex);
2871 latch_destroy(&pmd->exit_latch);
2872 xpthread_cond_destroy(&pmd->cond);
2873 ovs_mutex_destroy(&pmd->cond_mutex);
2874 ovs_mutex_destroy(&pmd->poll_mutex);
2878 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2879 * and unrefs the struct. */
2881 dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd)
2883 struct rxq_poll *poll;
2885 /* Uninit the 'flow_cache' since there is
2886 * no actual thread uninit it for NON_PMD_CORE_ID. */
2887 if (pmd->core_id == NON_PMD_CORE_ID) {
2888 emc_cache_uninit(&pmd->flow_cache);
2890 latch_set(&pmd->exit_latch);
2891 dp_netdev_reload_pmd__(pmd);
2892 ovs_numa_unpin_core(pmd->core_id);
2893 xpthread_join(pmd->thread, NULL);
2896 /* Unref all ports and free poll_list. */
2897 LIST_FOR_EACH_POP (poll, node, &pmd->poll_list) {
2898 port_unref(poll->port);
2902 /* Purges the 'pmd''s flows after stopping the thread, but before
2903 * destroying the flows, so that the flow stats can be collected. */
2904 if (dp->dp_purge_cb) {
2905 dp->dp_purge_cb(dp->dp_purge_aux, pmd->core_id);
2907 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2908 dp_netdev_pmd_unref(pmd);
2911 /* Destroys all pmd threads. */
2913 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2915 struct dp_netdev_pmd_thread *pmd;
2917 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2918 dp_netdev_del_pmd(dp, pmd);
2922 /* Deletes all pmd threads on numa node 'numa_id'. */
2924 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2926 struct dp_netdev_pmd_thread *pmd;
2928 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2929 if (pmd->numa_id == numa_id) {
2930 dp_netdev_del_pmd(dp, pmd);
2935 /* Returns PMD thread from this numa node with fewer rx queues to poll.
2936 * Returns NULL if there is no PMD threads on this numa node.
2937 * Can be called safely only by main thread. */
2938 static struct dp_netdev_pmd_thread *
2939 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id)
2942 struct dp_netdev_pmd_thread *pmd, *res = NULL;
2944 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2945 if (pmd->numa_id == numa_id
2946 && (min_cnt > pmd->poll_cnt || res == NULL)) {
2947 min_cnt = pmd->poll_cnt;
2955 /* Adds rx queue to poll_list of PMD thread. */
2957 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
2958 struct dp_netdev_port *port, struct netdev_rxq *rx)
2959 OVS_REQUIRES(pmd->poll_mutex)
2961 struct rxq_poll *poll = xmalloc(sizeof *poll);
2967 list_push_back(&pmd->poll_list, &poll->node);
2971 /* Checks the numa node id of 'netdev' and starts pmd threads for
2974 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2978 if (!ovs_numa_numa_id_is_valid(numa_id)) {
2979 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2980 "invalid", numa_id);
2984 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
2986 /* If there are already pmd threads created for the numa node
2987 * in which 'netdev' is on, do nothing. Else, creates the
2988 * pmd threads for the numa node. */
2990 int can_have, n_unpinned, i, index = 0;
2991 struct dp_netdev_pmd_thread **pmds;
2992 struct dp_netdev_port *port;
2994 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
2996 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2997 "cores on numa node");
3001 /* If cpu mask is specified, uses all unpinned cores, otherwise
3002 * tries creating NR_PMD_THREADS pmd threads. */
3003 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
3004 pmds = xzalloc(can_have * sizeof *pmds);
3005 for (i = 0; i < can_have; i++) {
3006 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
3007 pmds[i] = xzalloc(sizeof **pmds);
3008 dp_netdev_configure_pmd(pmds[i], dp, i, core_id, numa_id);
3011 /* Distributes rx queues of this numa node between new pmd threads. */
3012 CMAP_FOR_EACH (port, node, &dp->ports) {
3013 if (netdev_is_pmd(port->netdev)
3014 && netdev_get_numa_id(port->netdev) == numa_id) {
3015 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
3016 /* Make thread-safety analyser happy. */
3017 ovs_mutex_lock(&pmds[index]->poll_mutex);
3018 dp_netdev_add_rxq_to_pmd(pmds[index], port, port->rxq[i]);
3019 ovs_mutex_unlock(&pmds[index]->poll_mutex);
3020 index = (index + 1) % can_have;
3025 /* Actual start of pmd threads. */
3026 for (i = 0; i < can_have; i++) {
3027 pmds[i]->thread = ovs_thread_create("pmd", pmd_thread_main, pmds[i]);
3030 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
3035 /* Called after pmd threads config change. Restarts pmd threads with
3036 * new configuration. */
3038 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
3040 struct dp_netdev_port *port;
3042 CMAP_FOR_EACH (port, node, &dp->ports) {
3043 if (netdev_is_pmd(port->netdev)) {
3044 int numa_id = netdev_get_numa_id(port->netdev);
3046 dp_netdev_set_pmds_on_numa(dp, numa_id);
3052 dpif_netdev_get_datapath_version(void)
3054 return xstrdup("<built-in>");
3058 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
3059 uint16_t tcp_flags, long long now)
3063 atomic_store_relaxed(&netdev_flow->stats.used, now);
3064 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
3065 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
3066 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
3068 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
3072 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
3073 enum dp_stat_type type, int cnt)
3075 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3079 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3080 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3081 enum dpif_upcall_type type, const struct nlattr *userdata,
3082 struct ofpbuf *actions, struct ofpbuf *put_actions)
3084 struct dp_netdev *dp = pmd->dp;
3085 struct flow_tnl orig_tunnel;
3088 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3092 /* Upcall processing expects the Geneve options to be in the translated
3093 * format but we need to retain the raw format for datapath use. */
3094 orig_tunnel.flags = flow->tunnel.flags;
3095 if (flow->tunnel.flags & FLOW_TNL_F_UDPIF) {
3096 orig_tunnel.metadata.present.len = flow->tunnel.metadata.present.len;
3097 memcpy(orig_tunnel.metadata.opts.gnv, flow->tunnel.metadata.opts.gnv,
3098 flow->tunnel.metadata.present.len);
3099 err = tun_metadata_from_geneve_udpif(&orig_tunnel, &orig_tunnel,
3106 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3107 struct ds ds = DS_EMPTY_INITIALIZER;
3110 struct odp_flow_key_parms odp_parms = {
3113 .odp_in_port = flow->in_port.odp_port,
3114 .support = dp_netdev_support,
3117 ofpbuf_init(&key, 0);
3118 odp_flow_key_from_flow(&odp_parms, &key);
3119 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3120 dp_packet_size(packet_));
3122 odp_flow_key_format(key.data, key.size, &ds);
3124 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3125 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3127 ofpbuf_uninit(&key);
3133 err = dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3134 actions, wc, put_actions, dp->upcall_aux);
3135 if (err && err != ENOSPC) {
3139 /* Translate tunnel metadata masks to datapath format. */
3141 if (wc->masks.tunnel.metadata.present.map) {
3142 struct geneve_opt opts[TLV_TOT_OPT_SIZE /
3143 sizeof(struct geneve_opt)];
3145 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3146 tun_metadata_to_geneve_udpif_mask(&flow->tunnel,
3148 orig_tunnel.metadata.opts.gnv,
3149 orig_tunnel.metadata.present.len,
3152 orig_tunnel.metadata.present.len = 0;
3155 memset(&wc->masks.tunnel.metadata, 0,
3156 sizeof wc->masks.tunnel.metadata);
3157 memcpy(&wc->masks.tunnel.metadata.opts.gnv, opts,
3158 orig_tunnel.metadata.present.len);
3160 wc->masks.tunnel.metadata.present.len = 0xff;
3163 /* Restore tunnel metadata. We need to use the saved options to ensure
3164 * that any unknown options are not lost. The generated mask will have
3165 * the same structure, matching on types and lengths but wildcarding
3166 * option data we don't care about. */
3167 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3168 memcpy(&flow->tunnel.metadata.opts.gnv, orig_tunnel.metadata.opts.gnv,
3169 orig_tunnel.metadata.present.len);
3170 flow->tunnel.metadata.present.len = orig_tunnel.metadata.present.len;
3171 flow->tunnel.flags |= FLOW_TNL_F_UDPIF;
3177 static inline uint32_t
3178 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3179 const struct miniflow *mf)
3181 uint32_t hash, recirc_depth;
3183 if (OVS_LIKELY(dp_packet_rss_valid(packet))) {
3184 hash = dp_packet_get_rss_hash(packet);
3186 hash = miniflow_hash_5tuple(mf, 0);
3187 dp_packet_set_rss_hash(packet, hash);
3190 /* The RSS hash must account for the recirculation depth to avoid
3191 * collisions in the exact match cache */
3192 recirc_depth = *recirc_depth_get_unsafe();
3193 if (OVS_UNLIKELY(recirc_depth)) {
3194 hash = hash_finish(hash, recirc_depth);
3195 dp_packet_set_rss_hash(packet, hash);
3200 struct packet_batch {
3201 unsigned int packet_count;
3202 unsigned int byte_count;
3205 struct dp_netdev_flow *flow;
3207 struct dp_packet *packets[NETDEV_MAX_BURST];
3211 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
3212 const struct miniflow *mf)
3214 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3215 batch->packets[batch->packet_count++] = packet;
3216 batch->byte_count += dp_packet_size(packet);
3220 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
3222 flow->batch = batch;
3225 batch->packet_count = 0;
3226 batch->byte_count = 0;
3227 batch->tcp_flags = 0;
3231 packet_batch_execute(struct packet_batch *batch,
3232 struct dp_netdev_pmd_thread *pmd,
3235 struct dp_netdev_actions *actions;
3236 struct dp_netdev_flow *flow = batch->flow;
3238 dp_netdev_flow_used(flow, batch->packet_count, batch->byte_count,
3239 batch->tcp_flags, now);
3241 actions = dp_netdev_flow_get_actions(flow);
3243 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
3244 actions->actions, actions->size);
3248 dp_netdev_queue_batches(struct dp_packet *pkt,
3249 struct dp_netdev_flow *flow, const struct miniflow *mf,
3250 struct packet_batch *batches, size_t *n_batches)
3252 struct packet_batch *batch = flow->batch;
3254 if (OVS_LIKELY(batch)) {
3255 packet_batch_update(batch, pkt, mf);
3259 batch = &batches[(*n_batches)++];
3260 packet_batch_init(batch, flow);
3261 packet_batch_update(batch, pkt, mf);
3265 dp_packet_swap(struct dp_packet **a, struct dp_packet **b)
3267 struct dp_packet *tmp = *a;
3272 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3273 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', the
3274 * miniflow is copied into 'keys' and the packet pointer is moved at the
3275 * beginning of the 'packets' array.
3277 * The function returns the number of packets that needs to be processed in the
3278 * 'packets' array (they have been moved to the beginning of the vector).
3280 static inline size_t
3281 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
3282 size_t cnt, struct netdev_flow_key *keys,
3283 struct packet_batch batches[], size_t *n_batches)
3285 struct emc_cache *flow_cache = &pmd->flow_cache;
3286 struct netdev_flow_key key;
3287 size_t i, notfound_cnt = 0;
3289 for (i = 0; i < cnt; i++) {
3290 struct dp_netdev_flow *flow;
3292 if (OVS_UNLIKELY(dp_packet_size(packets[i]) < ETH_HEADER_LEN)) {
3293 dp_packet_delete(packets[i]);
3298 /* Prefetch next packet data */
3299 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3302 miniflow_extract(packets[i], &key.mf);
3303 key.len = 0; /* Not computed yet. */
3304 key.hash = dpif_netdev_packet_get_rss_hash(packets[i], &key.mf);
3306 flow = emc_lookup(flow_cache, &key);
3307 if (OVS_LIKELY(flow)) {
3308 dp_netdev_queue_batches(packets[i], flow, &key.mf, batches,
3311 if (i != notfound_cnt) {
3312 dp_packet_swap(&packets[i], &packets[notfound_cnt]);
3315 keys[notfound_cnt++] = key;
3319 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - notfound_cnt);
3321 return notfound_cnt;
3325 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3326 struct dp_packet **packets, size_t cnt,
3327 struct netdev_flow_key *keys,
3328 struct packet_batch batches[], size_t *n_batches)
3330 #if !defined(__CHECKER__) && !defined(_WIN32)
3331 const size_t PKT_ARRAY_SIZE = cnt;
3333 /* Sparse or MSVC doesn't like variable length array. */
3334 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3336 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3337 struct dp_netdev *dp = pmd->dp;
3338 struct emc_cache *flow_cache = &pmd->flow_cache;
3339 int miss_cnt = 0, lost_cnt = 0;
3343 for (i = 0; i < cnt; i++) {
3344 /* Key length is needed in all the cases, hash computed on demand. */
3345 keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf));
3347 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3348 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3349 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3350 struct ofpbuf actions, put_actions;
3353 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3354 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3356 for (i = 0; i < cnt; i++) {
3357 struct dp_netdev_flow *netdev_flow;
3358 struct ofpbuf *add_actions;
3362 if (OVS_LIKELY(rules[i])) {
3366 /* It's possible that an earlier slow path execution installed
3367 * a rule covering this flow. In this case, it's a lot cheaper
3368 * to catch it here than execute a miss. */
3369 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3371 rules[i] = &netdev_flow->cr;
3377 match.tun_md.valid = false;
3378 miniflow_expand(&keys[i].mf, &match.flow);
3380 ofpbuf_clear(&actions);
3381 ofpbuf_clear(&put_actions);
3383 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
3384 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
3385 &ufid, DPIF_UC_MISS, NULL, &actions,
3387 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3388 dp_packet_delete(packets[i]);
3393 /* The Netlink encoding of datapath flow keys cannot express
3394 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3395 * tag is interpreted as exact match on the fact that there is no
3396 * VLAN. Unless we refactor a lot of code that translates between
3397 * Netlink and struct flow representations, we have to do the same
3399 if (!match.wc.masks.vlan_tci) {
3400 match.wc.masks.vlan_tci = htons(0xffff);
3403 /* We can't allow the packet batching in the next loop to execute
3404 * the actions. Otherwise, if there are any slow path actions,
3405 * we'll send the packet up twice. */
3406 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
3407 actions.data, actions.size);
3409 add_actions = put_actions.size ? &put_actions : &actions;
3410 if (OVS_LIKELY(error != ENOSPC)) {
3411 /* XXX: There's a race window where a flow covering this packet
3412 * could have already been installed since we last did the flow
3413 * lookup before upcall. This could be solved by moving the
3414 * mutex lock outside the loop, but that's an awful long time
3415 * to be locking everyone out of making flow installs. If we
3416 * move to a per-core classifier, it would be reasonable. */
3417 ovs_mutex_lock(&pmd->flow_mutex);
3418 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3419 if (OVS_LIKELY(!netdev_flow)) {
3420 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3424 ovs_mutex_unlock(&pmd->flow_mutex);
3426 emc_insert(flow_cache, &keys[i], netdev_flow);
3430 ofpbuf_uninit(&actions);
3431 ofpbuf_uninit(&put_actions);
3432 fat_rwlock_unlock(&dp->upcall_rwlock);
3433 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3434 } else if (OVS_UNLIKELY(any_miss)) {
3435 for (i = 0; i < cnt; i++) {
3436 if (OVS_UNLIKELY(!rules[i])) {
3437 dp_packet_delete(packets[i]);
3444 for (i = 0; i < cnt; i++) {
3445 struct dp_packet *packet = packets[i];
3446 struct dp_netdev_flow *flow;
3448 if (OVS_UNLIKELY(!rules[i])) {
3452 flow = dp_netdev_flow_cast(rules[i]);
3454 emc_insert(flow_cache, &keys[i], flow);
3455 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3458 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3459 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3460 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3464 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3465 struct dp_packet **packets, int cnt)
3467 #if !defined(__CHECKER__) && !defined(_WIN32)
3468 const size_t PKT_ARRAY_SIZE = cnt;
3470 /* Sparse or MSVC doesn't like variable length array. */
3471 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3473 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3474 struct packet_batch batches[PKT_ARRAY_SIZE];
3475 long long now = time_msec();
3476 size_t newcnt, n_batches, i;
3479 newcnt = emc_processing(pmd, packets, cnt, keys, batches, &n_batches);
3480 if (OVS_UNLIKELY(newcnt)) {
3481 fast_path_processing(pmd, packets, newcnt, keys, batches, &n_batches);
3484 for (i = 0; i < n_batches; i++) {
3485 batches[i].flow->batch = NULL;
3488 for (i = 0; i < n_batches; i++) {
3489 packet_batch_execute(&batches[i], pmd, now);
3493 struct dp_netdev_execute_aux {
3494 struct dp_netdev_pmd_thread *pmd;
3498 dpif_netdev_register_dp_purge_cb(struct dpif *dpif, dp_purge_callback *cb,
3501 struct dp_netdev *dp = get_dp_netdev(dpif);
3502 dp->dp_purge_aux = aux;
3503 dp->dp_purge_cb = cb;
3507 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3510 struct dp_netdev *dp = get_dp_netdev(dpif);
3511 dp->upcall_aux = aux;
3516 dp_netdev_drop_packets(struct dp_packet **packets, int cnt, bool may_steal)
3521 for (i = 0; i < cnt; i++) {
3522 dp_packet_delete(packets[i]);
3528 push_tnl_action(const struct dp_netdev *dp,
3529 const struct nlattr *attr,
3530 struct dp_packet **packets, int cnt)
3532 struct dp_netdev_port *tun_port;
3533 const struct ovs_action_push_tnl *data;
3535 data = nl_attr_get(attr);
3537 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3541 netdev_push_header(tun_port->netdev, packets, cnt, data);
3547 dp_netdev_clone_pkt_batch(struct dp_packet **dst_pkts,
3548 struct dp_packet **src_pkts, int cnt)
3552 for (i = 0; i < cnt; i++) {
3553 dst_pkts[i] = dp_packet_clone(src_pkts[i]);
3558 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3559 const struct nlattr *a, bool may_steal)
3560 OVS_NO_THREAD_SAFETY_ANALYSIS
3562 struct dp_netdev_execute_aux *aux = aux_;
3563 uint32_t *depth = recirc_depth_get();
3564 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3565 struct dp_netdev *dp = pmd->dp;
3566 int type = nl_attr_type(a);
3567 struct dp_netdev_port *p;
3570 switch ((enum ovs_action_attr)type) {
3571 case OVS_ACTION_ATTR_OUTPUT:
3572 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3573 if (OVS_LIKELY(p)) {
3574 netdev_send(p->netdev, pmd->tx_qid, packets, cnt, may_steal);
3579 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3580 if (*depth < MAX_RECIRC_DEPTH) {
3581 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3585 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3589 err = push_tnl_action(dp, a, packets, cnt);
3592 dp_netdev_input(pmd, packets, cnt);
3595 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3601 case OVS_ACTION_ATTR_TUNNEL_POP:
3602 if (*depth < MAX_RECIRC_DEPTH) {
3603 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3605 p = dp_netdev_lookup_port(dp, portno);
3607 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3611 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3615 err = netdev_pop_header(p->netdev, packets, cnt);
3618 for (i = 0; i < cnt; i++) {
3619 packets[i]->md.in_port.odp_port = portno;
3623 dp_netdev_input(pmd, packets, cnt);
3626 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3633 case OVS_ACTION_ATTR_USERSPACE:
3634 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3635 const struct nlattr *userdata;
3636 struct ofpbuf actions;
3640 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3641 ofpbuf_init(&actions, 0);
3643 for (i = 0; i < cnt; i++) {
3646 ofpbuf_clear(&actions);
3648 flow_extract(packets[i], &flow);
3649 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3650 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3651 DPIF_UC_ACTION, userdata,&actions,
3653 if (!error || error == ENOSPC) {
3654 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3655 actions.data, actions.size);
3656 } else if (may_steal) {
3657 dp_packet_delete(packets[i]);
3660 ofpbuf_uninit(&actions);
3661 fat_rwlock_unlock(&dp->upcall_rwlock);
3667 case OVS_ACTION_ATTR_RECIRC:
3668 if (*depth < MAX_RECIRC_DEPTH) {
3669 struct dp_packet *recirc_pkts[NETDEV_MAX_BURST];
3672 dp_netdev_clone_pkt_batch(recirc_pkts, packets, cnt);
3673 packets = recirc_pkts;
3676 for (i = 0; i < cnt; i++) {
3677 packets[i]->md.recirc_id = nl_attr_get_u32(a);
3681 dp_netdev_input(pmd, packets, cnt);
3687 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3690 case OVS_ACTION_ATTR_CT:
3691 /* If a flow with this action is slow-pathed, datapath assistance is
3692 * required to implement it. However, we don't support this action
3693 * in the userspace datapath. */
3694 VLOG_WARN("Cannot execute conntrack action in userspace.");
3697 case OVS_ACTION_ATTR_PUSH_VLAN:
3698 case OVS_ACTION_ATTR_POP_VLAN:
3699 case OVS_ACTION_ATTR_PUSH_MPLS:
3700 case OVS_ACTION_ATTR_POP_MPLS:
3701 case OVS_ACTION_ATTR_SET:
3702 case OVS_ACTION_ATTR_SET_MASKED:
3703 case OVS_ACTION_ATTR_SAMPLE:
3704 case OVS_ACTION_ATTR_HASH:
3705 case OVS_ACTION_ATTR_UNSPEC:
3706 case __OVS_ACTION_ATTR_MAX:
3710 dp_netdev_drop_packets(packets, cnt, may_steal);
3714 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3715 struct dp_packet **packets, int cnt,
3717 const struct nlattr *actions, size_t actions_len)
3719 struct dp_netdev_execute_aux aux = { pmd };
3721 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3722 actions_len, dp_execute_cb);
3725 const struct dpif_class dpif_netdev_class = {
3728 dpif_netdev_enumerate,
3729 dpif_netdev_port_open_type,
3732 dpif_netdev_destroy,
3735 dpif_netdev_get_stats,
3736 dpif_netdev_port_add,
3737 dpif_netdev_port_del,
3738 dpif_netdev_port_query_by_number,
3739 dpif_netdev_port_query_by_name,
3740 NULL, /* port_get_pid */
3741 dpif_netdev_port_dump_start,
3742 dpif_netdev_port_dump_next,
3743 dpif_netdev_port_dump_done,
3744 dpif_netdev_port_poll,
3745 dpif_netdev_port_poll_wait,
3746 dpif_netdev_flow_flush,
3747 dpif_netdev_flow_dump_create,
3748 dpif_netdev_flow_dump_destroy,
3749 dpif_netdev_flow_dump_thread_create,
3750 dpif_netdev_flow_dump_thread_destroy,
3751 dpif_netdev_flow_dump_next,
3752 dpif_netdev_operate,
3753 NULL, /* recv_set */
3754 NULL, /* handlers_set */
3755 dpif_netdev_pmd_set,
3756 dpif_netdev_queue_to_priority,
3758 NULL, /* recv_wait */
3759 NULL, /* recv_purge */
3760 dpif_netdev_register_dp_purge_cb,
3761 dpif_netdev_register_upcall_cb,
3762 dpif_netdev_enable_upcall,
3763 dpif_netdev_disable_upcall,
3764 dpif_netdev_get_datapath_version,
3765 NULL, /* ct_dump_start */
3766 NULL, /* ct_dump_next */
3767 NULL, /* ct_dump_done */
3768 NULL, /* ct_flush */
3772 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3773 const char *argv[], void *aux OVS_UNUSED)
3775 struct dp_netdev_port *old_port;
3776 struct dp_netdev_port *new_port;
3777 struct dp_netdev *dp;
3780 ovs_mutex_lock(&dp_netdev_mutex);
3781 dp = shash_find_data(&dp_netdevs, argv[1]);
3782 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3783 ovs_mutex_unlock(&dp_netdev_mutex);
3784 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3787 ovs_refcount_ref(&dp->ref_cnt);
3788 ovs_mutex_unlock(&dp_netdev_mutex);
3790 ovs_mutex_lock(&dp->port_mutex);
3791 if (get_port_by_name(dp, argv[2], &old_port)) {
3792 unixctl_command_reply_error(conn, "unknown port");
3796 port_no = u32_to_odp(atoi(argv[3]));
3797 if (!port_no || port_no == ODPP_NONE) {
3798 unixctl_command_reply_error(conn, "bad port number");
3801 if (dp_netdev_lookup_port(dp, port_no)) {
3802 unixctl_command_reply_error(conn, "port number already in use");
3806 /* Remove old port. */
3807 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3808 ovsrcu_postpone(free, old_port);
3810 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3811 new_port = xmemdup(old_port, sizeof *old_port);
3812 new_port->port_no = port_no;
3813 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3815 seq_change(dp->port_seq);
3816 unixctl_command_reply(conn, NULL);
3819 ovs_mutex_unlock(&dp->port_mutex);
3820 dp_netdev_unref(dp);
3824 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3825 const char *argv[], void *aux OVS_UNUSED)
3827 struct dp_netdev_port *port;
3828 struct dp_netdev *dp;
3830 ovs_mutex_lock(&dp_netdev_mutex);
3831 dp = shash_find_data(&dp_netdevs, argv[1]);
3832 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3833 ovs_mutex_unlock(&dp_netdev_mutex);
3834 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3837 ovs_refcount_ref(&dp->ref_cnt);
3838 ovs_mutex_unlock(&dp_netdev_mutex);
3840 ovs_mutex_lock(&dp->port_mutex);
3841 if (get_port_by_name(dp, argv[2], &port)) {
3842 unixctl_command_reply_error(conn, "unknown port");
3843 } else if (port->port_no == ODPP_LOCAL) {
3844 unixctl_command_reply_error(conn, "can't delete local port");
3846 do_del_port(dp, port);
3847 unixctl_command_reply(conn, NULL);
3849 ovs_mutex_unlock(&dp->port_mutex);
3851 dp_netdev_unref(dp);
3855 dpif_dummy_register__(const char *type)
3857 struct dpif_class *class;
3859 class = xmalloc(sizeof *class);
3860 *class = dpif_netdev_class;
3861 class->type = xstrdup(type);
3862 dp_register_provider(class);
3866 dpif_dummy_override(const char *type)
3871 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
3872 * a userland-only build. It's useful for testsuite.
3874 error = dp_unregister_provider(type);
3875 if (error == 0 || error == EAFNOSUPPORT) {
3876 dpif_dummy_register__(type);
3881 dpif_dummy_register(enum dummy_level level)
3883 if (level == DUMMY_OVERRIDE_ALL) {
3888 dp_enumerate_types(&types);
3889 SSET_FOR_EACH (type, &types) {
3890 dpif_dummy_override(type);
3892 sset_destroy(&types);
3893 } else if (level == DUMMY_OVERRIDE_SYSTEM) {
3894 dpif_dummy_override("system");
3897 dpif_dummy_register__("dummy");
3899 unixctl_command_register("dpif-dummy/change-port-number",
3900 "dp port new-number",
3901 3, 3, dpif_dummy_change_port_number, NULL);
3902 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3903 2, 2, dpif_dummy_delete_port, NULL);
3906 /* Datapath Classifier. */
3908 /* A set of rules that all have the same fields wildcarded. */
3909 struct dpcls_subtable {
3910 /* The fields are only used by writers. */
3911 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3913 /* These fields are accessed by readers. */
3914 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3915 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3916 /* 'mask' must be the last field, additional space is allocated here. */
3919 /* Initializes 'cls' as a classifier that initially contains no classification
3922 dpcls_init(struct dpcls *cls)
3924 cmap_init(&cls->subtables_map);
3925 pvector_init(&cls->subtables);
3929 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3931 pvector_remove(&cls->subtables, subtable);
3932 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3933 subtable->mask.hash);
3934 cmap_destroy(&subtable->rules);
3935 ovsrcu_postpone(free, subtable);
3938 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3939 * caller's responsibility.
3940 * May only be called after all the readers have been terminated. */
3942 dpcls_destroy(struct dpcls *cls)
3945 struct dpcls_subtable *subtable;
3947 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3948 ovs_assert(cmap_count(&subtable->rules) == 0);
3949 dpcls_destroy_subtable(cls, subtable);
3951 cmap_destroy(&cls->subtables_map);
3952 pvector_destroy(&cls->subtables);
3956 static struct dpcls_subtable *
3957 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3959 struct dpcls_subtable *subtable;
3961 /* Need to add one. */
3962 subtable = xmalloc(sizeof *subtable
3963 - sizeof subtable->mask.mf + mask->len);
3964 cmap_init(&subtable->rules);
3965 netdev_flow_key_clone(&subtable->mask, mask);
3966 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3967 pvector_insert(&cls->subtables, subtable, 0);
3968 pvector_publish(&cls->subtables);
3973 static inline struct dpcls_subtable *
3974 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3976 struct dpcls_subtable *subtable;
3978 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3979 &cls->subtables_map) {
3980 if (netdev_flow_key_equal(&subtable->mask, mask)) {
3984 return dpcls_create_subtable(cls, mask);
3987 /* Insert 'rule' into 'cls'. */
3989 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
3990 const struct netdev_flow_key *mask)
3992 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
3994 rule->mask = &subtable->mask;
3995 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
3998 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
4000 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
4002 struct dpcls_subtable *subtable;
4004 ovs_assert(rule->mask);
4006 INIT_CONTAINER(subtable, rule->mask, mask);
4008 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
4010 dpcls_destroy_subtable(cls, subtable);
4011 pvector_publish(&cls->subtables);
4015 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
4016 * in 'mask' the values in 'key' and 'target' are the same. */
4018 dpcls_rule_matches_key(const struct dpcls_rule *rule,
4019 const struct netdev_flow_key *target)
4021 const uint64_t *keyp = miniflow_get_values(&rule->flow.mf);
4022 const uint64_t *maskp = miniflow_get_values(&rule->mask->mf);
4025 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, target, rule->flow.mf.map) {
4026 if (OVS_UNLIKELY((value & *maskp++) != *keyp++)) {
4033 /* For each miniflow in 'flows' performs a classifier lookup writing the result
4034 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
4035 * NULL it is skipped.
4037 * This function is optimized for use in the userspace datapath and therefore
4038 * does not implement a lot of features available in the standard
4039 * classifier_lookup() function. Specifically, it does not implement
4040 * priorities, instead returning any rule which matches the flow.
4042 * Returns true if all flows found a corresponding rule. */
4044 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
4045 struct dpcls_rule **rules, const size_t cnt)
4047 /* The batch size 16 was experimentally found faster than 8 or 32. */
4048 typedef uint16_t map_type;
4049 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
4051 #if !defined(__CHECKER__) && !defined(_WIN32)
4052 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
4054 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
4056 map_type maps[N_MAPS];
4057 struct dpcls_subtable *subtable;
4059 memset(maps, 0xff, sizeof maps);
4060 if (cnt % MAP_BITS) {
4061 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
4063 memset(rules, 0, cnt * sizeof *rules);
4065 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
4066 const struct netdev_flow_key *mkeys = keys;
4067 struct dpcls_rule **mrules = rules;
4068 map_type remains = 0;
4071 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
4073 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
4074 uint32_t hashes[MAP_BITS];
4075 const struct cmap_node *nodes[MAP_BITS];
4076 unsigned long map = maps[m];
4080 continue; /* Skip empty maps. */
4083 /* Compute hashes for the remaining keys. */
4084 ULLONG_FOR_EACH_1(i, map) {
4085 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
4089 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
4090 /* Check results. */
4091 ULLONG_FOR_EACH_1(i, map) {
4092 struct dpcls_rule *rule;
4094 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
4095 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
4100 ULLONG_SET0(map, i); /* Did not match. */
4102 ; /* Keep Sparse happy. */
4104 maps[m] &= ~map; /* Clear the found rules. */
4108 return true; /* All found. */
4111 return false; /* Some misses. */