2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2016 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "dpif-netdev.h"
25 #include <netinet/in.h>
29 #include <sys/ioctl.h>
30 #include <sys/socket.h>
38 #include "dp-packet.h"
40 #include "dpif-provider.h"
42 #include "fat-rwlock.h"
47 #include "netdev-dpdk.h"
48 #include "netdev-vport.h"
50 #include "odp-execute.h"
52 #include "openvswitch/dynamic-string.h"
53 #include "openvswitch/list.h"
54 #include "openvswitch/match.h"
55 #include "openvswitch/ofp-print.h"
56 #include "openvswitch/ofpbuf.h"
57 #include "openvswitch/vlog.h"
61 #include "poll-loop.h"
68 #include "tnl-neigh-cache.h"
69 #include "tnl-ports.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 /* Cpu mask for pin of pmd threads. */
226 uint64_t last_tnl_conf_seq;
229 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
233 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
234 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
235 DP_STAT_MISS, /* Packets that did not match. */
236 DP_STAT_LOST, /* Packets not passed up to the client. */
240 enum pmd_cycles_counter_type {
241 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
242 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
246 /* A port in a netdev-based datapath. */
247 struct dp_netdev_port {
249 struct netdev *netdev;
250 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
251 struct netdev_saved_flags *sf;
252 unsigned n_rxq; /* Number of elements in 'rxq' */
253 struct netdev_rxq **rxq;
254 char *type; /* Port type as requested by user. */
255 int latest_requested_n_rxq; /* Latest requested from netdev number
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_per_flow_init() and packet_batch_per_flow_execute()). */
331 struct packet_batch_per_flow *batch;
333 /* Packet classification. */
334 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
335 /* 'cr' must be the last member. */
338 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
339 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
340 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
343 /* A set of datapath actions within a "struct dp_netdev_flow".
349 * A struct dp_netdev_actions 'actions' is protected with RCU. */
350 struct dp_netdev_actions {
351 /* These members are immutable: they do not change during the struct's
353 unsigned int size; /* Size of 'actions', in bytes. */
354 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
357 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
359 struct dp_netdev_actions *dp_netdev_flow_get_actions(
360 const struct dp_netdev_flow *);
361 static void dp_netdev_actions_free(struct dp_netdev_actions *);
363 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
364 struct dp_netdev_pmd_stats {
365 /* Indexed by DP_STAT_*. */
366 atomic_ullong n[DP_N_STATS];
369 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
370 struct dp_netdev_pmd_cycles {
371 /* Indexed by PMD_CYCLES_*. */
372 atomic_ullong n[PMD_N_CYCLES];
375 /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */
377 struct dp_netdev_port *port;
378 struct netdev_rxq *rx;
379 struct ovs_list node;
382 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
383 * the performance overhead of interrupt processing. Therefore netdev can
384 * not implement rx-wait for these devices. dpif-netdev needs to poll
385 * these device to check for recv buffer. pmd-thread does polling for
386 * devices assigned to itself.
388 * DPDK used PMD for accessing NIC.
390 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
391 * I/O of all non-pmd threads. There will be no actual thread created
394 * Each struct has its own flow table and classifier. Packets received
395 * from managed ports are looked up in the corresponding pmd thread's
396 * flow table, and are executed with the found actions.
398 struct dp_netdev_pmd_thread {
399 struct dp_netdev *dp;
400 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
401 struct cmap_node node; /* In 'dp->poll_threads'. */
403 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
404 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
406 /* Per thread exact-match cache. Note, the instance for cpu core
407 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
408 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
409 * instances will only be accessed by its own pmd thread. */
410 struct emc_cache flow_cache;
412 /* Classifier and Flow-Table.
414 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
415 * changes to 'cls' must be made while still holding the 'flow_mutex'.
417 struct ovs_mutex flow_mutex;
419 struct cmap flow_table OVS_GUARDED; /* Flow table. */
422 struct dp_netdev_pmd_stats stats;
424 /* Cycles counters */
425 struct dp_netdev_pmd_cycles cycles;
427 /* Used to count cicles. See 'cycles_counter_end()' */
428 unsigned long long last_cycles;
430 struct latch exit_latch; /* For terminating the pmd thread. */
431 atomic_uint change_seq; /* For reloading pmd ports. */
433 int index; /* Idx of this pmd thread among pmd*/
434 /* threads on same numa node. */
435 unsigned core_id; /* CPU core id of this pmd thread. */
436 int numa_id; /* numa node id of this pmd thread. */
437 atomic_int tx_qid; /* Queue id used by this pmd thread to
438 * send packets on all netdevs */
440 struct ovs_mutex poll_mutex; /* Mutex for poll_list. */
441 /* List of rx queues to poll. */
442 struct ovs_list poll_list OVS_GUARDED;
443 int poll_cnt; /* Number of elemints in poll_list. */
445 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
446 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
447 * values and subtracts them from 'stats' and 'cycles' before
448 * reporting to the user */
449 unsigned long long stats_zero[DP_N_STATS];
450 uint64_t cycles_zero[PMD_N_CYCLES];
453 #define PMD_INITIAL_SEQ 1
455 /* Interface to netdev-based datapath. */
458 struct dp_netdev *dp;
459 uint64_t last_port_seq;
462 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
463 struct dp_netdev_port **portp);
464 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
465 struct dp_netdev_port **portp);
466 static void dp_netdev_free(struct dp_netdev *)
467 OVS_REQUIRES(dp_netdev_mutex);
468 static int do_add_port(struct dp_netdev *dp, const char *devname,
469 const char *type, odp_port_t port_no)
470 OVS_REQUIRES(dp->port_mutex);
471 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
472 OVS_REQUIRES(dp->port_mutex);
473 static int dpif_netdev_open(const struct dpif_class *, const char *name,
474 bool create, struct dpif **);
475 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
476 struct dp_packet_batch *,
478 const struct nlattr *actions,
480 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
481 struct dp_packet_batch *, odp_port_t port_no);
482 static void dp_netdev_recirculate(struct dp_netdev_pmd_thread *,
483 struct dp_packet_batch *);
485 static void dp_netdev_disable_upcall(struct dp_netdev *);
486 static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
487 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
488 struct dp_netdev *dp, int index,
489 unsigned core_id, int numa_id);
490 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
491 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
492 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
494 static struct dp_netdev_pmd_thread *
495 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
496 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
497 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
498 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
499 static void dp_netdev_pmd_clear_poll_list(struct dp_netdev_pmd_thread *pmd);
500 static void dp_netdev_del_port_from_pmd(struct dp_netdev_port *port,
501 struct dp_netdev_pmd_thread *pmd);
502 static void dp_netdev_del_port_from_all_pmds(struct dp_netdev *dp,
503 struct dp_netdev_port *port);
505 dp_netdev_add_port_to_pmds(struct dp_netdev *dp, struct dp_netdev_port *port);
507 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
508 struct dp_netdev_port *port, struct netdev_rxq *rx);
509 static struct dp_netdev_pmd_thread *
510 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id);
511 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
512 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
513 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
514 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
516 static inline bool emc_entry_alive(struct emc_entry *ce);
517 static void emc_clear_entry(struct emc_entry *ce);
520 emc_cache_init(struct emc_cache *flow_cache)
524 flow_cache->sweep_idx = 0;
525 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
526 flow_cache->entries[i].flow = NULL;
527 flow_cache->entries[i].key.hash = 0;
528 flow_cache->entries[i].key.len = sizeof(struct miniflow);
529 flowmap_init(&flow_cache->entries[i].key.mf.map);
534 emc_cache_uninit(struct emc_cache *flow_cache)
538 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
539 emc_clear_entry(&flow_cache->entries[i]);
543 /* Check and clear dead flow references slowly (one entry at each
546 emc_cache_slow_sweep(struct emc_cache *flow_cache)
548 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
550 if (!emc_entry_alive(entry)) {
551 emc_clear_entry(entry);
553 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
556 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
558 dpif_is_netdev(const struct dpif *dpif)
560 return dpif->dpif_class->open == dpif_netdev_open;
563 static struct dpif_netdev *
564 dpif_netdev_cast(const struct dpif *dpif)
566 ovs_assert(dpif_is_netdev(dpif));
567 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
570 static struct dp_netdev *
571 get_dp_netdev(const struct dpif *dpif)
573 return dpif_netdev_cast(dpif)->dp;
577 PMD_INFO_SHOW_STATS, /* Show how cpu cycles are spent. */
578 PMD_INFO_CLEAR_STATS, /* Set the cycles count to 0. */
579 PMD_INFO_SHOW_RXQ /* Show poll-lists of pmd threads. */
583 pmd_info_show_stats(struct ds *reply,
584 struct dp_netdev_pmd_thread *pmd,
585 unsigned long long stats[DP_N_STATS],
586 uint64_t cycles[PMD_N_CYCLES])
588 unsigned long long total_packets = 0;
589 uint64_t total_cycles = 0;
592 /* These loops subtracts reference values ('*_zero') from the counters.
593 * Since loads and stores are relaxed, it might be possible for a '*_zero'
594 * value to be more recent than the current value we're reading from the
595 * counter. This is not a big problem, since these numbers are not
596 * supposed to be too accurate, but we should at least make sure that
597 * the result is not negative. */
598 for (i = 0; i < DP_N_STATS; i++) {
599 if (stats[i] > pmd->stats_zero[i]) {
600 stats[i] -= pmd->stats_zero[i];
605 if (i != DP_STAT_LOST) {
606 /* Lost packets are already included in DP_STAT_MISS */
607 total_packets += stats[i];
611 for (i = 0; i < PMD_N_CYCLES; i++) {
612 if (cycles[i] > pmd->cycles_zero[i]) {
613 cycles[i] -= pmd->cycles_zero[i];
618 total_cycles += cycles[i];
621 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
622 ? "main thread" : "pmd thread");
624 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
625 ds_put_format(reply, " numa_id %d", pmd->numa_id);
627 if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) {
628 ds_put_format(reply, " core_id %u", pmd->core_id);
630 ds_put_cstr(reply, ":\n");
633 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
634 "\tmiss:%llu\n\tlost:%llu\n",
635 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
636 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
638 if (total_cycles == 0) {
643 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
644 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
645 cycles[PMD_CYCLES_POLLING],
646 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
647 cycles[PMD_CYCLES_PROCESSING],
648 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
650 if (total_packets == 0) {
655 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
656 total_cycles / (double)total_packets,
657 total_cycles, total_packets);
660 "\tavg processing cycles per packet: "
661 "%.02f (%"PRIu64"/%llu)\n",
662 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
663 cycles[PMD_CYCLES_PROCESSING], total_packets);
667 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
668 struct dp_netdev_pmd_thread *pmd,
669 unsigned long long stats[DP_N_STATS],
670 uint64_t cycles[PMD_N_CYCLES])
674 /* We cannot write 'stats' and 'cycles' (because they're written by other
675 * threads) and we shouldn't change 'stats' (because they're used to count
676 * datapath stats, which must not be cleared here). Instead, we save the
677 * current values and subtract them from the values to be displayed in the
679 for (i = 0; i < DP_N_STATS; i++) {
680 pmd->stats_zero[i] = stats[i];
682 for (i = 0; i < PMD_N_CYCLES; i++) {
683 pmd->cycles_zero[i] = cycles[i];
688 pmd_info_show_rxq(struct ds *reply, struct dp_netdev_pmd_thread *pmd)
690 if (pmd->core_id != NON_PMD_CORE_ID) {
691 struct rxq_poll *poll;
692 const char *prev_name = NULL;
694 ds_put_format(reply, "pmd thread numa_id %d core_id %u:\n",
695 pmd->numa_id, pmd->core_id);
697 ovs_mutex_lock(&pmd->poll_mutex);
698 LIST_FOR_EACH (poll, node, &pmd->poll_list) {
699 const char *name = netdev_get_name(poll->port->netdev);
701 if (!prev_name || strcmp(name, prev_name)) {
703 ds_put_cstr(reply, "\n");
705 ds_put_format(reply, "\tport: %s\tqueue-id:",
706 netdev_get_name(poll->port->netdev));
708 ds_put_format(reply, " %d", netdev_rxq_get_queue_id(poll->rx));
711 ovs_mutex_unlock(&pmd->poll_mutex);
712 ds_put_cstr(reply, "\n");
717 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
720 struct ds reply = DS_EMPTY_INITIALIZER;
721 struct dp_netdev_pmd_thread *pmd;
722 struct dp_netdev *dp = NULL;
723 enum pmd_info_type type = *(enum pmd_info_type *) aux;
725 ovs_mutex_lock(&dp_netdev_mutex);
728 dp = shash_find_data(&dp_netdevs, argv[1]);
729 } else if (shash_count(&dp_netdevs) == 1) {
730 /* There's only one datapath */
731 dp = shash_first(&dp_netdevs)->data;
735 ovs_mutex_unlock(&dp_netdev_mutex);
736 unixctl_command_reply_error(conn,
737 "please specify an existing datapath");
741 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
742 if (type == PMD_INFO_SHOW_RXQ) {
743 pmd_info_show_rxq(&reply, pmd);
745 unsigned long long stats[DP_N_STATS];
746 uint64_t cycles[PMD_N_CYCLES];
749 /* Read current stats and cycle counters */
750 for (i = 0; i < ARRAY_SIZE(stats); i++) {
751 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
753 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
754 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
757 if (type == PMD_INFO_CLEAR_STATS) {
758 pmd_info_clear_stats(&reply, pmd, stats, cycles);
759 } else if (type == PMD_INFO_SHOW_STATS) {
760 pmd_info_show_stats(&reply, pmd, stats, cycles);
765 ovs_mutex_unlock(&dp_netdev_mutex);
767 unixctl_command_reply(conn, ds_cstr(&reply));
772 dpif_netdev_init(void)
774 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
775 clear_aux = PMD_INFO_CLEAR_STATS,
776 poll_aux = PMD_INFO_SHOW_RXQ;
778 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
779 0, 1, dpif_netdev_pmd_info,
781 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
782 0, 1, dpif_netdev_pmd_info,
784 unixctl_command_register("dpif-netdev/pmd-rxq-show", "[dp]",
785 0, 1, dpif_netdev_pmd_info,
791 dpif_netdev_enumerate(struct sset *all_dps,
792 const struct dpif_class *dpif_class)
794 struct shash_node *node;
796 ovs_mutex_lock(&dp_netdev_mutex);
797 SHASH_FOR_EACH(node, &dp_netdevs) {
798 struct dp_netdev *dp = node->data;
799 if (dpif_class != dp->class) {
800 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
801 * If the class doesn't match, skip this dpif. */
804 sset_add(all_dps, node->name);
806 ovs_mutex_unlock(&dp_netdev_mutex);
812 dpif_netdev_class_is_dummy(const struct dpif_class *class)
814 return class != &dpif_netdev_class;
818 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
820 return strcmp(type, "internal") ? type
821 : dpif_netdev_class_is_dummy(class) ? "dummy"
826 create_dpif_netdev(struct dp_netdev *dp)
828 uint16_t netflow_id = hash_string(dp->name, 0);
829 struct dpif_netdev *dpif;
831 ovs_refcount_ref(&dp->ref_cnt);
833 dpif = xmalloc(sizeof *dpif);
834 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
836 dpif->last_port_seq = seq_read(dp->port_seq);
841 /* Choose an unused, non-zero port number and return it on success.
842 * Return ODPP_NONE on failure. */
844 choose_port(struct dp_netdev *dp, const char *name)
845 OVS_REQUIRES(dp->port_mutex)
849 if (dp->class != &dpif_netdev_class) {
853 /* If the port name begins with "br", start the number search at
854 * 100 to make writing tests easier. */
855 if (!strncmp(name, "br", 2)) {
859 /* If the port name contains a number, try to assign that port number.
860 * This can make writing unit tests easier because port numbers are
862 for (p = name; *p != '\0'; p++) {
863 if (isdigit((unsigned char) *p)) {
864 port_no = start_no + strtol(p, NULL, 10);
865 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
866 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
867 return u32_to_odp(port_no);
874 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
875 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
876 return u32_to_odp(port_no);
884 create_dp_netdev(const char *name, const struct dpif_class *class,
885 struct dp_netdev **dpp)
886 OVS_REQUIRES(dp_netdev_mutex)
888 struct dp_netdev *dp;
891 dp = xzalloc(sizeof *dp);
892 shash_add(&dp_netdevs, name, dp);
894 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
895 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
896 ovs_refcount_init(&dp->ref_cnt);
897 atomic_flag_clear(&dp->destroyed);
899 ovs_mutex_init(&dp->port_mutex);
900 cmap_init(&dp->ports);
901 dp->port_seq = seq_create();
902 fat_rwlock_init(&dp->upcall_rwlock);
904 /* Disable upcalls by default. */
905 dp_netdev_disable_upcall(dp);
906 dp->upcall_aux = NULL;
907 dp->upcall_cb = NULL;
909 cmap_init(&dp->poll_threads);
910 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
911 ovsthread_key_create(&dp->per_pmd_key, NULL);
913 dp_netdev_set_nonpmd(dp);
915 ovs_mutex_lock(&dp->port_mutex);
916 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
917 ovs_mutex_unlock(&dp->port_mutex);
923 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
929 dpif_netdev_open(const struct dpif_class *class, const char *name,
930 bool create, struct dpif **dpifp)
932 struct dp_netdev *dp;
935 ovs_mutex_lock(&dp_netdev_mutex);
936 dp = shash_find_data(&dp_netdevs, name);
938 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
940 error = (dp->class != class ? EINVAL
945 *dpifp = create_dpif_netdev(dp);
948 ovs_mutex_unlock(&dp_netdev_mutex);
954 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
955 OVS_NO_THREAD_SAFETY_ANALYSIS
957 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
958 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
960 /* Before freeing a lock we should release it */
961 fat_rwlock_unlock(&dp->upcall_rwlock);
962 fat_rwlock_destroy(&dp->upcall_rwlock);
965 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
966 * through the 'dp_netdevs' shash while freeing 'dp'. */
968 dp_netdev_free(struct dp_netdev *dp)
969 OVS_REQUIRES(dp_netdev_mutex)
971 struct dp_netdev_port *port;
973 shash_find_and_delete(&dp_netdevs, dp->name);
975 dp_netdev_destroy_all_pmds(dp);
976 ovs_mutex_destroy(&dp->non_pmd_mutex);
977 ovsthread_key_delete(dp->per_pmd_key);
979 ovs_mutex_lock(&dp->port_mutex);
980 CMAP_FOR_EACH (port, node, &dp->ports) {
981 /* PMD threads are destroyed here. do_del_port() cannot quiesce */
982 do_del_port(dp, port);
984 ovs_mutex_unlock(&dp->port_mutex);
985 cmap_destroy(&dp->poll_threads);
987 seq_destroy(dp->port_seq);
988 cmap_destroy(&dp->ports);
990 /* Upcalls must be disabled at this point */
991 dp_netdev_destroy_upcall_lock(dp);
994 free(CONST_CAST(char *, dp->name));
999 dp_netdev_unref(struct dp_netdev *dp)
1002 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
1003 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
1004 ovs_mutex_lock(&dp_netdev_mutex);
1005 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
1008 ovs_mutex_unlock(&dp_netdev_mutex);
1013 dpif_netdev_close(struct dpif *dpif)
1015 struct dp_netdev *dp = get_dp_netdev(dpif);
1017 dp_netdev_unref(dp);
1022 dpif_netdev_destroy(struct dpif *dpif)
1024 struct dp_netdev *dp = get_dp_netdev(dpif);
1026 if (!atomic_flag_test_and_set(&dp->destroyed)) {
1027 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
1028 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
1036 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
1037 * load/store semantics. While the increment is not atomic, the load and
1038 * store operations are, making it impossible to read inconsistent values.
1040 * This is used to update thread local stats counters. */
1042 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
1044 unsigned long long tmp;
1046 atomic_read_relaxed(var, &tmp);
1048 atomic_store_relaxed(var, tmp);
1052 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
1054 struct dp_netdev *dp = get_dp_netdev(dpif);
1055 struct dp_netdev_pmd_thread *pmd;
1057 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
1058 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1059 unsigned long long n;
1060 stats->n_flows += cmap_count(&pmd->flow_table);
1062 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
1064 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
1066 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
1067 stats->n_missed += n;
1068 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
1071 stats->n_masks = UINT32_MAX;
1072 stats->n_mask_hit = UINT64_MAX;
1078 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1082 if (pmd->core_id == NON_PMD_CORE_ID) {
1086 ovs_mutex_lock(&pmd->cond_mutex);
1087 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1088 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1089 ovs_mutex_unlock(&pmd->cond_mutex);
1093 hash_port_no(odp_port_t port_no)
1095 return hash_int(odp_to_u32(port_no), 0);
1099 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1101 OVS_REQUIRES(dp->port_mutex)
1103 struct netdev_saved_flags *sf;
1104 struct dp_netdev_port *port;
1105 struct netdev *netdev;
1106 enum netdev_flags flags;
1107 const char *open_type;
1109 int i, n_open_rxqs = 0;
1111 /* Reject devices already in 'dp'. */
1112 if (!get_port_by_name(dp, devname, &port)) {
1117 /* Open and validate network device. */
1118 open_type = dpif_netdev_port_open_type(dp->class, type);
1119 error = netdev_open(devname, open_type, &netdev);
1123 /* XXX reject non-Ethernet devices */
1125 netdev_get_flags(netdev, &flags);
1126 if (flags & NETDEV_LOOPBACK) {
1127 VLOG_ERR("%s: cannot add a loopback device", devname);
1132 if (netdev_is_pmd(netdev)) {
1133 int n_cores = ovs_numa_get_n_cores();
1135 if (n_cores == OVS_CORE_UNSPEC) {
1136 VLOG_ERR("%s, cannot get cpu core info", devname);
1140 /* There can only be ovs_numa_get_n_cores() pmd threads,
1141 * so creates a txq for each, and one extra for the non
1143 error = netdev_set_multiq(netdev, n_cores + 1,
1144 netdev_requested_n_rxq(netdev));
1145 if (error && (error != EOPNOTSUPP)) {
1146 VLOG_ERR("%s, cannot set multiq", devname);
1150 port = xzalloc(sizeof *port);
1151 port->port_no = port_no;
1152 port->netdev = netdev;
1153 port->n_rxq = netdev_n_rxq(netdev);
1154 port->rxq = xmalloc(sizeof *port->rxq * port->n_rxq);
1155 port->type = xstrdup(type);
1156 port->latest_requested_n_rxq = netdev_requested_n_rxq(netdev);
1158 for (i = 0; i < port->n_rxq; i++) {
1159 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1161 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1162 devname, ovs_strerror(errno));
1168 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1174 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1176 if (netdev_is_pmd(netdev)) {
1177 dp_netdev_add_port_to_pmds(dp, port);
1179 seq_change(dp->port_seq);
1184 for (i = 0; i < n_open_rxqs; i++) {
1185 netdev_rxq_close(port->rxq[i]);
1191 netdev_close(netdev);
1197 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1198 odp_port_t *port_nop)
1200 struct dp_netdev *dp = get_dp_netdev(dpif);
1201 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1202 const char *dpif_port;
1206 ovs_mutex_lock(&dp->port_mutex);
1207 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1208 if (*port_nop != ODPP_NONE) {
1209 port_no = *port_nop;
1210 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1212 port_no = choose_port(dp, dpif_port);
1213 error = port_no == ODPP_NONE ? EFBIG : 0;
1216 *port_nop = port_no;
1217 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1219 ovs_mutex_unlock(&dp->port_mutex);
1225 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1227 struct dp_netdev *dp = get_dp_netdev(dpif);
1230 ovs_mutex_lock(&dp->port_mutex);
1231 if (port_no == ODPP_LOCAL) {
1234 struct dp_netdev_port *port;
1236 error = get_port_by_number(dp, port_no, &port);
1238 do_del_port(dp, port);
1241 ovs_mutex_unlock(&dp->port_mutex);
1247 is_valid_port_number(odp_port_t port_no)
1249 return port_no != ODPP_NONE;
1252 static struct dp_netdev_port *
1253 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1255 struct dp_netdev_port *port;
1257 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1258 if (port->port_no == port_no) {
1266 get_port_by_number(struct dp_netdev *dp,
1267 odp_port_t port_no, struct dp_netdev_port **portp)
1269 if (!is_valid_port_number(port_no)) {
1273 *portp = dp_netdev_lookup_port(dp, port_no);
1274 return *portp ? 0 : ENOENT;
1279 port_destroy(struct dp_netdev_port *port)
1285 netdev_close(port->netdev);
1286 netdev_restore_flags(port->sf);
1288 for (unsigned i = 0; i < port->n_rxq; i++) {
1289 netdev_rxq_close(port->rxq[i]);
1298 get_port_by_name(struct dp_netdev *dp,
1299 const char *devname, struct dp_netdev_port **portp)
1300 OVS_REQUIRES(dp->port_mutex)
1302 struct dp_netdev_port *port;
1304 CMAP_FOR_EACH (port, node, &dp->ports) {
1305 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1314 get_n_pmd_threads(struct dp_netdev *dp)
1316 /* There is one non pmd thread in dp->poll_threads */
1317 return cmap_count(&dp->poll_threads) - 1;
1321 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1323 struct dp_netdev_pmd_thread *pmd;
1326 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1327 if (pmd->numa_id == numa_id) {
1335 /* Returns 'true' if there is a port with pmd netdev and the netdev
1336 * is on numa node 'numa_id'. */
1338 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1340 struct dp_netdev_port *port;
1342 CMAP_FOR_EACH (port, node, &dp->ports) {
1343 if (netdev_is_pmd(port->netdev)
1344 && netdev_get_numa_id(port->netdev) == numa_id) {
1354 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1355 OVS_REQUIRES(dp->port_mutex)
1357 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1358 seq_change(dp->port_seq);
1359 if (netdev_is_pmd(port->netdev)) {
1360 int numa_id = netdev_get_numa_id(port->netdev);
1362 /* PMD threads can not be on invalid numa node. */
1363 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
1364 /* If there is no netdev on the numa node, deletes the pmd threads
1365 * for that numa. Else, deletes the queues from polling lists. */
1366 if (!has_pmd_port_for_numa(dp, numa_id)) {
1367 dp_netdev_del_pmds_on_numa(dp, numa_id);
1369 dp_netdev_del_port_from_all_pmds(dp, port);
1377 answer_port_query(const struct dp_netdev_port *port,
1378 struct dpif_port *dpif_port)
1380 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1381 dpif_port->type = xstrdup(port->type);
1382 dpif_port->port_no = port->port_no;
1386 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1387 struct dpif_port *dpif_port)
1389 struct dp_netdev *dp = get_dp_netdev(dpif);
1390 struct dp_netdev_port *port;
1393 error = get_port_by_number(dp, port_no, &port);
1394 if (!error && dpif_port) {
1395 answer_port_query(port, dpif_port);
1402 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1403 struct dpif_port *dpif_port)
1405 struct dp_netdev *dp = get_dp_netdev(dpif);
1406 struct dp_netdev_port *port;
1409 ovs_mutex_lock(&dp->port_mutex);
1410 error = get_port_by_name(dp, devname, &port);
1411 if (!error && dpif_port) {
1412 answer_port_query(port, dpif_port);
1414 ovs_mutex_unlock(&dp->port_mutex);
1420 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1422 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1426 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1428 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1429 ovsrcu_postpone(dp_netdev_flow_free, flow);
1434 dp_netdev_flow_hash(const ovs_u128 *ufid)
1436 return ufid->u32[0];
1440 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1441 struct dp_netdev_flow *flow)
1442 OVS_REQUIRES(pmd->flow_mutex)
1444 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1446 dpcls_remove(&pmd->cls, &flow->cr);
1447 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1450 dp_netdev_flow_unref(flow);
1454 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1456 struct dp_netdev_flow *netdev_flow;
1458 ovs_mutex_lock(&pmd->flow_mutex);
1459 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1460 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1462 ovs_mutex_unlock(&pmd->flow_mutex);
1466 dpif_netdev_flow_flush(struct dpif *dpif)
1468 struct dp_netdev *dp = get_dp_netdev(dpif);
1469 struct dp_netdev_pmd_thread *pmd;
1471 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1472 dp_netdev_pmd_flow_flush(pmd);
1478 struct dp_netdev_port_state {
1479 struct cmap_position position;
1484 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1486 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1491 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1492 struct dpif_port *dpif_port)
1494 struct dp_netdev_port_state *state = state_;
1495 struct dp_netdev *dp = get_dp_netdev(dpif);
1496 struct cmap_node *node;
1499 node = cmap_next_position(&dp->ports, &state->position);
1501 struct dp_netdev_port *port;
1503 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1506 state->name = xstrdup(netdev_get_name(port->netdev));
1507 dpif_port->name = state->name;
1508 dpif_port->type = port->type;
1509 dpif_port->port_no = port->port_no;
1520 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1522 struct dp_netdev_port_state *state = state_;
1529 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1531 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1532 uint64_t new_port_seq;
1535 new_port_seq = seq_read(dpif->dp->port_seq);
1536 if (dpif->last_port_seq != new_port_seq) {
1537 dpif->last_port_seq = new_port_seq;
1547 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1549 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1551 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1554 static struct dp_netdev_flow *
1555 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1557 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1560 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1562 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1565 /* netdev_flow_key utilities.
1567 * netdev_flow_key is basically a miniflow. We use these functions
1568 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1569 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1571 * - Since we are dealing exclusively with miniflows created by
1572 * miniflow_extract(), if the map is different the miniflow is different.
1573 * Therefore we can be faster by comparing the map and the miniflow in a
1575 * - These functions can be inlined by the compiler. */
1577 /* Given the number of bits set in miniflow's maps, returns the size of the
1578 * 'netdev_flow_key.mf' */
1579 static inline size_t
1580 netdev_flow_key_size(size_t flow_u64s)
1582 return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s);
1586 netdev_flow_key_equal(const struct netdev_flow_key *a,
1587 const struct netdev_flow_key *b)
1589 /* 'b->len' may be not set yet. */
1590 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1593 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1594 * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been
1595 * generated by miniflow_extract. */
1597 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1598 const struct miniflow *mf)
1600 return !memcmp(&key->mf, mf, key->len);
1604 netdev_flow_key_clone(struct netdev_flow_key *dst,
1605 const struct netdev_flow_key *src)
1608 offsetof(struct netdev_flow_key, mf) + src->len);
1613 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1614 const struct flow *src)
1616 struct dp_packet packet;
1617 uint64_t buf_stub[512 / 8];
1619 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1620 pkt_metadata_from_flow(&packet.md, src);
1621 flow_compose(&packet, src);
1622 miniflow_extract(&packet, &dst->mf);
1623 dp_packet_uninit(&packet);
1625 dst->len = netdev_flow_key_size(miniflow_n_values(&dst->mf));
1626 dst->hash = 0; /* Not computed yet. */
1629 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1631 netdev_flow_mask_init(struct netdev_flow_key *mask,
1632 const struct match *match)
1634 uint64_t *dst = miniflow_values(&mask->mf);
1635 struct flowmap fmap;
1639 /* Only check masks that make sense for the flow. */
1640 flow_wc_map(&match->flow, &fmap);
1641 flowmap_init(&mask->mf.map);
1643 FLOWMAP_FOR_EACH_INDEX(idx, fmap) {
1644 uint64_t mask_u64 = flow_u64_value(&match->wc.masks, idx);
1647 flowmap_set(&mask->mf.map, idx, 1);
1649 hash = hash_add64(hash, mask_u64);
1655 FLOWMAP_FOR_EACH_MAP (map, mask->mf.map) {
1656 hash = hash_add64(hash, map);
1659 size_t n = dst - miniflow_get_values(&mask->mf);
1661 mask->hash = hash_finish(hash, n * 8);
1662 mask->len = netdev_flow_key_size(n);
1665 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1667 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1668 const struct flow *flow,
1669 const struct netdev_flow_key *mask)
1671 uint64_t *dst_u64 = miniflow_values(&dst->mf);
1672 const uint64_t *mask_u64 = miniflow_get_values(&mask->mf);
1676 dst->len = mask->len;
1677 dst->mf = mask->mf; /* Copy maps. */
1679 FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf.map) {
1680 *dst_u64 = value & *mask_u64++;
1681 hash = hash_add64(hash, *dst_u64++);
1683 dst->hash = hash_finish(hash,
1684 (dst_u64 - miniflow_get_values(&dst->mf)) * 8);
1687 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
1688 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
1689 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
1691 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1693 static inline uint32_t
1694 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1695 const struct netdev_flow_key *mask)
1697 const uint64_t *p = miniflow_get_values(&mask->mf);
1701 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, key, mask->mf.map) {
1702 hash = hash_add64(hash, value & *p++);
1705 return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8);
1709 emc_entry_alive(struct emc_entry *ce)
1711 return ce->flow && !ce->flow->dead;
1715 emc_clear_entry(struct emc_entry *ce)
1718 dp_netdev_flow_unref(ce->flow);
1724 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1725 const struct netdev_flow_key *key)
1727 if (ce->flow != flow) {
1729 dp_netdev_flow_unref(ce->flow);
1732 if (dp_netdev_flow_ref(flow)) {
1739 netdev_flow_key_clone(&ce->key, key);
1744 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1745 struct dp_netdev_flow *flow)
1747 struct emc_entry *to_be_replaced = NULL;
1748 struct emc_entry *current_entry;
1750 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1751 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1752 /* We found the entry with the 'mf' miniflow */
1753 emc_change_entry(current_entry, flow, NULL);
1757 /* Replacement policy: put the flow in an empty (not alive) entry, or
1758 * in the first entry where it can be */
1760 || (emc_entry_alive(to_be_replaced)
1761 && !emc_entry_alive(current_entry))
1762 || current_entry->key.hash < to_be_replaced->key.hash) {
1763 to_be_replaced = current_entry;
1766 /* We didn't find the miniflow in the cache.
1767 * The 'to_be_replaced' entry is where the new flow will be stored */
1769 emc_change_entry(to_be_replaced, flow, key);
1772 static inline struct dp_netdev_flow *
1773 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1775 struct emc_entry *current_entry;
1777 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1778 if (current_entry->key.hash == key->hash
1779 && emc_entry_alive(current_entry)
1780 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1782 /* We found the entry with the 'key->mf' miniflow */
1783 return current_entry->flow;
1790 static struct dp_netdev_flow *
1791 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1792 const struct netdev_flow_key *key)
1794 struct dp_netdev_flow *netdev_flow;
1795 struct dpcls_rule *rule;
1797 dpcls_lookup(&pmd->cls, key, &rule, 1);
1798 netdev_flow = dp_netdev_flow_cast(rule);
1803 static struct dp_netdev_flow *
1804 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1805 const ovs_u128 *ufidp, const struct nlattr *key,
1808 struct dp_netdev_flow *netdev_flow;
1812 /* If a UFID is not provided, determine one based on the key. */
1813 if (!ufidp && key && key_len
1814 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1815 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1820 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1822 if (ovs_u128_equals(netdev_flow->ufid, *ufidp)) {
1832 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1833 struct dpif_flow_stats *stats)
1835 struct dp_netdev_flow *netdev_flow;
1836 unsigned long long n;
1840 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1842 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1843 stats->n_packets = n;
1844 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1846 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1848 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1849 stats->tcp_flags = flags;
1852 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1853 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1854 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1857 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1858 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1859 struct dpif_flow *flow, bool terse)
1862 memset(flow, 0, sizeof *flow);
1864 struct flow_wildcards wc;
1865 struct dp_netdev_actions *actions;
1867 struct odp_flow_key_parms odp_parms = {
1868 .flow = &netdev_flow->flow,
1870 .support = dp_netdev_support,
1873 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1876 offset = key_buf->size;
1877 flow->key = ofpbuf_tail(key_buf);
1878 odp_parms.odp_in_port = netdev_flow->flow.in_port.odp_port;
1879 odp_flow_key_from_flow(&odp_parms, key_buf);
1880 flow->key_len = key_buf->size - offset;
1883 offset = mask_buf->size;
1884 flow->mask = ofpbuf_tail(mask_buf);
1885 odp_parms.odp_in_port = wc.masks.in_port.odp_port;
1886 odp_parms.key_buf = key_buf;
1887 odp_flow_key_from_mask(&odp_parms, mask_buf);
1888 flow->mask_len = mask_buf->size - offset;
1891 actions = dp_netdev_flow_get_actions(netdev_flow);
1892 flow->actions = actions->actions;
1893 flow->actions_len = actions->size;
1896 flow->ufid = netdev_flow->ufid;
1897 flow->ufid_present = true;
1898 flow->pmd_id = netdev_flow->pmd_id;
1899 get_dpif_flow_stats(netdev_flow, &flow->stats);
1903 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1904 const struct nlattr *mask_key,
1905 uint32_t mask_key_len, const struct flow *flow,
1906 struct flow_wildcards *wc)
1908 enum odp_key_fitness fitness;
1910 fitness = odp_flow_key_to_mask_udpif(mask_key, mask_key_len, key,
1913 /* This should not happen: it indicates that
1914 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1915 * disagree on the acceptable form of a mask. Log the problem
1916 * as an error, with enough details to enable debugging. */
1917 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1919 if (!VLOG_DROP_ERR(&rl)) {
1923 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1925 VLOG_ERR("internal error parsing flow mask %s (%s)",
1926 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1937 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1942 if (odp_flow_key_to_flow_udpif(key, key_len, flow)) {
1943 /* This should not happen: it indicates that odp_flow_key_from_flow()
1944 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1945 * flow. Log the problem as an error, with enough details to enable
1947 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1949 if (!VLOG_DROP_ERR(&rl)) {
1953 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1954 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1961 in_port = flow->in_port.odp_port;
1962 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1966 /* Userspace datapath doesn't support conntrack. */
1967 if (flow->ct_state || flow->ct_zone || flow->ct_mark
1968 || !ovs_u128_is_zero(flow->ct_label)) {
1976 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1978 struct dp_netdev *dp = get_dp_netdev(dpif);
1979 struct dp_netdev_flow *netdev_flow;
1980 struct dp_netdev_pmd_thread *pmd;
1981 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1982 ? NON_PMD_CORE_ID : get->pmd_id;
1985 pmd = dp_netdev_get_pmd(dp, pmd_id);
1990 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1993 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1998 dp_netdev_pmd_unref(pmd);
2004 static struct dp_netdev_flow *
2005 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
2006 struct match *match, const ovs_u128 *ufid,
2007 const struct nlattr *actions, size_t actions_len)
2008 OVS_REQUIRES(pmd->flow_mutex)
2010 struct dp_netdev_flow *flow;
2011 struct netdev_flow_key mask;
2013 netdev_flow_mask_init(&mask, match);
2014 /* Make sure wc does not have metadata. */
2015 ovs_assert(!FLOWMAP_HAS_FIELD(&mask.mf.map, metadata)
2016 && !FLOWMAP_HAS_FIELD(&mask.mf.map, regs));
2018 /* Do not allocate extra space. */
2019 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
2020 memset(&flow->stats, 0, sizeof flow->stats);
2023 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
2024 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
2025 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
2026 ovs_refcount_init(&flow->ref_cnt);
2027 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
2029 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
2030 dpcls_insert(&pmd->cls, &flow->cr, &mask);
2032 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
2033 dp_netdev_flow_hash(&flow->ufid));
2035 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2037 struct ds ds = DS_EMPTY_INITIALIZER;
2039 match.tun_md.valid = false;
2040 match.flow = flow->flow;
2041 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2043 ds_put_cstr(&ds, "flow_add: ");
2044 odp_format_ufid(ufid, &ds);
2045 ds_put_cstr(&ds, " ");
2046 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2047 ds_put_cstr(&ds, ", actions:");
2048 format_odp_actions(&ds, actions, actions_len);
2050 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2059 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2061 struct dp_netdev *dp = get_dp_netdev(dpif);
2062 struct dp_netdev_flow *netdev_flow;
2063 struct netdev_flow_key key;
2064 struct dp_netdev_pmd_thread *pmd;
2067 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2068 ? NON_PMD_CORE_ID : put->pmd_id;
2071 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2075 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2076 put->mask, put->mask_len,
2077 &match.flow, &match.wc);
2082 pmd = dp_netdev_get_pmd(dp, pmd_id);
2087 /* Must produce a netdev_flow_key for lookup.
2088 * This interface is no longer performance critical, since it is not used
2089 * for upcall processing any more. */
2090 netdev_flow_key_from_flow(&key, &match.flow);
2095 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2098 ovs_mutex_lock(&pmd->flow_mutex);
2099 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2101 if (put->flags & DPIF_FP_CREATE) {
2102 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2104 memset(put->stats, 0, sizeof *put->stats);
2106 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2116 if (put->flags & DPIF_FP_MODIFY
2117 && flow_equal(&match.flow, &netdev_flow->flow)) {
2118 struct dp_netdev_actions *new_actions;
2119 struct dp_netdev_actions *old_actions;
2121 new_actions = dp_netdev_actions_create(put->actions,
2124 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2125 ovsrcu_set(&netdev_flow->actions, new_actions);
2128 get_dpif_flow_stats(netdev_flow, put->stats);
2130 if (put->flags & DPIF_FP_ZERO_STATS) {
2131 /* XXX: The userspace datapath uses thread local statistics
2132 * (for flows), which should be updated only by the owning
2133 * thread. Since we cannot write on stats memory here,
2134 * we choose not to support this flag. Please note:
2135 * - This feature is currently used only by dpctl commands with
2137 * - Should the need arise, this operation can be implemented
2138 * by keeping a base value (to be update here) for each
2139 * counter, and subtracting it before outputting the stats */
2143 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2144 } else if (put->flags & DPIF_FP_CREATE) {
2147 /* Overlapping flow. */
2151 ovs_mutex_unlock(&pmd->flow_mutex);
2152 dp_netdev_pmd_unref(pmd);
2158 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2160 struct dp_netdev *dp = get_dp_netdev(dpif);
2161 struct dp_netdev_flow *netdev_flow;
2162 struct dp_netdev_pmd_thread *pmd;
2163 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2164 ? NON_PMD_CORE_ID : del->pmd_id;
2167 pmd = dp_netdev_get_pmd(dp, pmd_id);
2172 ovs_mutex_lock(&pmd->flow_mutex);
2173 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2177 get_dpif_flow_stats(netdev_flow, del->stats);
2179 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2183 ovs_mutex_unlock(&pmd->flow_mutex);
2184 dp_netdev_pmd_unref(pmd);
2189 struct dpif_netdev_flow_dump {
2190 struct dpif_flow_dump up;
2191 struct cmap_position poll_thread_pos;
2192 struct cmap_position flow_pos;
2193 struct dp_netdev_pmd_thread *cur_pmd;
2195 struct ovs_mutex mutex;
2198 static struct dpif_netdev_flow_dump *
2199 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2201 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2204 static struct dpif_flow_dump *
2205 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2207 struct dpif_netdev_flow_dump *dump;
2209 dump = xzalloc(sizeof *dump);
2210 dpif_flow_dump_init(&dump->up, dpif_);
2211 dump->up.terse = terse;
2212 ovs_mutex_init(&dump->mutex);
2218 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2220 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2222 ovs_mutex_destroy(&dump->mutex);
2227 struct dpif_netdev_flow_dump_thread {
2228 struct dpif_flow_dump_thread up;
2229 struct dpif_netdev_flow_dump *dump;
2230 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2231 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2234 static struct dpif_netdev_flow_dump_thread *
2235 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2237 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2240 static struct dpif_flow_dump_thread *
2241 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2243 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2244 struct dpif_netdev_flow_dump_thread *thread;
2246 thread = xmalloc(sizeof *thread);
2247 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2248 thread->dump = dump;
2253 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2255 struct dpif_netdev_flow_dump_thread *thread
2256 = dpif_netdev_flow_dump_thread_cast(thread_);
2262 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2263 struct dpif_flow *flows, int max_flows)
2265 struct dpif_netdev_flow_dump_thread *thread
2266 = dpif_netdev_flow_dump_thread_cast(thread_);
2267 struct dpif_netdev_flow_dump *dump = thread->dump;
2268 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2272 ovs_mutex_lock(&dump->mutex);
2273 if (!dump->status) {
2274 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2275 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2276 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2277 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2279 /* First call to dump_next(), extracts the first pmd thread.
2280 * If there is no pmd thread, returns immediately. */
2282 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2284 ovs_mutex_unlock(&dump->mutex);
2291 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2292 struct cmap_node *node;
2294 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2298 netdev_flows[n_flows] = CONTAINER_OF(node,
2299 struct dp_netdev_flow,
2302 /* When finishing dumping the current pmd thread, moves to
2304 if (n_flows < flow_limit) {
2305 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2306 dp_netdev_pmd_unref(pmd);
2307 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2313 /* Keeps the reference to next caller. */
2314 dump->cur_pmd = pmd;
2316 /* If the current dump is empty, do not exit the loop, since the
2317 * remaining pmds could have flows to be dumped. Just dumps again
2318 * on the new 'pmd'. */
2321 ovs_mutex_unlock(&dump->mutex);
2323 for (i = 0; i < n_flows; i++) {
2324 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2325 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2326 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2327 struct dpif_flow *f = &flows[i];
2328 struct ofpbuf key, mask;
2330 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2331 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2332 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2340 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2341 OVS_NO_THREAD_SAFETY_ANALYSIS
2343 struct dp_netdev *dp = get_dp_netdev(dpif);
2344 struct dp_netdev_pmd_thread *pmd;
2345 struct dp_packet_batch pp;
2347 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2348 dp_packet_size(execute->packet) > UINT16_MAX) {
2352 /* Tries finding the 'pmd'. If NULL is returned, that means
2353 * the current thread is a non-pmd thread and should use
2354 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2355 pmd = ovsthread_getspecific(dp->per_pmd_key);
2357 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2360 /* If the current thread is non-pmd thread, acquires
2361 * the 'non_pmd_mutex'. */
2362 if (pmd->core_id == NON_PMD_CORE_ID) {
2363 ovs_mutex_lock(&dp->non_pmd_mutex);
2364 ovs_mutex_lock(&dp->port_mutex);
2367 packet_batch_init_packet(&pp, execute->packet);
2368 dp_netdev_execute_actions(pmd, &pp, false, execute->actions,
2369 execute->actions_len);
2370 if (pmd->core_id == NON_PMD_CORE_ID) {
2371 dp_netdev_pmd_unref(pmd);
2372 ovs_mutex_unlock(&dp->port_mutex);
2373 ovs_mutex_unlock(&dp->non_pmd_mutex);
2380 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2384 for (i = 0; i < n_ops; i++) {
2385 struct dpif_op *op = ops[i];
2388 case DPIF_OP_FLOW_PUT:
2389 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2392 case DPIF_OP_FLOW_DEL:
2393 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2396 case DPIF_OP_EXECUTE:
2397 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2400 case DPIF_OP_FLOW_GET:
2401 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2407 /* Returns true if the configuration for rx queues or cpu mask
2410 pmd_config_changed(const struct dp_netdev *dp, const char *cmask)
2412 struct dp_netdev_port *port;
2414 CMAP_FOR_EACH (port, node, &dp->ports) {
2415 struct netdev *netdev = port->netdev;
2416 int requested_n_rxq = netdev_requested_n_rxq(netdev);
2417 if (netdev_is_pmd(netdev)
2418 && port->latest_requested_n_rxq != requested_n_rxq) {
2423 if (dp->pmd_cmask != NULL && cmask != NULL) {
2424 return strcmp(dp->pmd_cmask, cmask);
2426 return (dp->pmd_cmask != NULL || cmask != NULL);
2430 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2432 dpif_netdev_pmd_set(struct dpif *dpif, const char *cmask)
2434 struct dp_netdev *dp = get_dp_netdev(dpif);
2436 if (pmd_config_changed(dp, cmask)) {
2437 struct dp_netdev_port *port;
2439 dp_netdev_destroy_all_pmds(dp);
2441 CMAP_FOR_EACH (port, node, &dp->ports) {
2442 struct netdev *netdev = port->netdev;
2443 int requested_n_rxq = netdev_requested_n_rxq(netdev);
2444 if (netdev_is_pmd(port->netdev)
2445 && port->latest_requested_n_rxq != requested_n_rxq) {
2448 /* Closes the existing 'rxq's. */
2449 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2450 netdev_rxq_close(port->rxq[i]);
2451 port->rxq[i] = NULL;
2455 /* Sets the new rx queue config. */
2456 err = netdev_set_multiq(port->netdev,
2457 ovs_numa_get_n_cores() + 1,
2459 if (err && (err != EOPNOTSUPP)) {
2460 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2461 " %u", netdev_get_name(port->netdev),
2465 port->latest_requested_n_rxq = requested_n_rxq;
2466 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2467 port->n_rxq = netdev_n_rxq(port->netdev);
2468 port->rxq = xrealloc(port->rxq, sizeof *port->rxq * port->n_rxq);
2469 for (i = 0; i < port->n_rxq; i++) {
2470 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2474 /* Reconfigures the cpu mask. */
2475 ovs_numa_set_cpu_mask(cmask);
2476 free(dp->pmd_cmask);
2477 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2479 /* Restores the non-pmd. */
2480 dp_netdev_set_nonpmd(dp);
2481 /* Restores all pmd threads. */
2482 dp_netdev_reset_pmd_threads(dp);
2489 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2490 uint32_t queue_id, uint32_t *priority)
2492 *priority = queue_id;
2497 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2498 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2499 struct dp_netdev_actions *
2500 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2502 struct dp_netdev_actions *netdev_actions;
2504 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2505 memcpy(netdev_actions->actions, actions, size);
2506 netdev_actions->size = size;
2508 return netdev_actions;
2511 struct dp_netdev_actions *
2512 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2514 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2518 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2523 static inline unsigned long long
2524 cycles_counter(void)
2527 return rte_get_tsc_cycles();
2533 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2534 extern struct ovs_mutex cycles_counter_fake_mutex;
2536 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2538 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2539 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2540 OVS_NO_THREAD_SAFETY_ANALYSIS
2542 pmd->last_cycles = cycles_counter();
2545 /* Stop counting cycles and add them to the counter 'type' */
2547 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2548 enum pmd_cycles_counter_type type)
2549 OVS_RELEASES(&cycles_counter_fake_mutex)
2550 OVS_NO_THREAD_SAFETY_ANALYSIS
2552 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2554 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2558 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2559 struct dp_netdev_port *port,
2560 struct netdev_rxq *rxq)
2562 struct dp_packet_batch batch;
2565 dp_packet_batch_init(&batch);
2566 cycles_count_start(pmd);
2567 error = netdev_rxq_recv(rxq, &batch);
2568 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2570 *recirc_depth_get() = 0;
2572 cycles_count_start(pmd);
2573 dp_netdev_input(pmd, &batch, port->port_no);
2574 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2575 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2576 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2578 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2579 netdev_get_name(port->netdev), ovs_strerror(error));
2583 /* Return true if needs to revalidate datapath flows. */
2585 dpif_netdev_run(struct dpif *dpif)
2587 struct dp_netdev_port *port;
2588 struct dp_netdev *dp = get_dp_netdev(dpif);
2589 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2591 uint64_t new_tnl_seq;
2593 ovs_mutex_lock(&dp->non_pmd_mutex);
2594 CMAP_FOR_EACH (port, node, &dp->ports) {
2595 if (!netdev_is_pmd(port->netdev)) {
2598 for (i = 0; i < port->n_rxq; i++) {
2599 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2603 ovs_mutex_unlock(&dp->non_pmd_mutex);
2604 dp_netdev_pmd_unref(non_pmd);
2606 tnl_neigh_cache_run();
2608 new_tnl_seq = seq_read(tnl_conf_seq);
2610 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2611 dp->last_tnl_conf_seq = new_tnl_seq;
2618 dpif_netdev_wait(struct dpif *dpif)
2620 struct dp_netdev_port *port;
2621 struct dp_netdev *dp = get_dp_netdev(dpif);
2623 ovs_mutex_lock(&dp_netdev_mutex);
2624 CMAP_FOR_EACH (port, node, &dp->ports) {
2625 if (!netdev_is_pmd(port->netdev)) {
2628 for (i = 0; i < port->n_rxq; i++) {
2629 netdev_rxq_wait(port->rxq[i]);
2633 ovs_mutex_unlock(&dp_netdev_mutex);
2634 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2638 pmd_load_queues(struct dp_netdev_pmd_thread *pmd, struct rxq_poll **ppoll_list)
2639 OVS_REQUIRES(pmd->poll_mutex)
2641 struct rxq_poll *poll_list = *ppoll_list;
2642 struct rxq_poll *poll;
2645 poll_list = xrealloc(poll_list, pmd->poll_cnt * sizeof *poll_list);
2648 LIST_FOR_EACH (poll, node, &pmd->poll_list) {
2649 poll_list[i++] = *poll;
2652 *ppoll_list = poll_list;
2653 return pmd->poll_cnt;
2657 pmd_thread_main(void *f_)
2659 struct dp_netdev_pmd_thread *pmd = f_;
2660 unsigned int lc = 0;
2661 struct rxq_poll *poll_list;
2662 unsigned int port_seq = PMD_INITIAL_SEQ;
2669 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2670 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2671 pmd_thread_setaffinity_cpu(pmd->core_id);
2673 emc_cache_init(&pmd->flow_cache);
2675 ovs_mutex_lock(&pmd->poll_mutex);
2676 poll_cnt = pmd_load_queues(pmd, &poll_list);
2677 ovs_mutex_unlock(&pmd->poll_mutex);
2679 /* List port/core affinity */
2680 for (i = 0; i < poll_cnt; i++) {
2681 VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n",
2682 pmd->core_id, netdev_get_name(poll_list[i].port->netdev),
2683 netdev_rxq_get_queue_id(poll_list[i].rx));
2686 /* Signal here to make sure the pmd finishes
2687 * reloading the updated configuration. */
2688 dp_netdev_pmd_reload_done(pmd);
2691 for (i = 0; i < poll_cnt; i++) {
2692 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2700 emc_cache_slow_sweep(&pmd->flow_cache);
2701 coverage_try_clear();
2704 atomic_read_relaxed(&pmd->change_seq, &seq);
2705 if (seq != port_seq) {
2712 emc_cache_uninit(&pmd->flow_cache);
2714 if (!latch_is_set(&pmd->exit_latch)){
2718 dp_netdev_pmd_reload_done(pmd);
2725 dp_netdev_disable_upcall(struct dp_netdev *dp)
2726 OVS_ACQUIRES(dp->upcall_rwlock)
2728 fat_rwlock_wrlock(&dp->upcall_rwlock);
2732 dpif_netdev_disable_upcall(struct dpif *dpif)
2733 OVS_NO_THREAD_SAFETY_ANALYSIS
2735 struct dp_netdev *dp = get_dp_netdev(dpif);
2736 dp_netdev_disable_upcall(dp);
2740 dp_netdev_enable_upcall(struct dp_netdev *dp)
2741 OVS_RELEASES(dp->upcall_rwlock)
2743 fat_rwlock_unlock(&dp->upcall_rwlock);
2747 dpif_netdev_enable_upcall(struct dpif *dpif)
2748 OVS_NO_THREAD_SAFETY_ANALYSIS
2750 struct dp_netdev *dp = get_dp_netdev(dpif);
2751 dp_netdev_enable_upcall(dp);
2755 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2757 ovs_mutex_lock(&pmd->cond_mutex);
2758 xpthread_cond_signal(&pmd->cond);
2759 ovs_mutex_unlock(&pmd->cond_mutex);
2762 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2763 * the pointer if succeeds, otherwise, NULL.
2765 * Caller must unrefs the returned reference. */
2766 static struct dp_netdev_pmd_thread *
2767 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2769 struct dp_netdev_pmd_thread *pmd;
2770 const struct cmap_node *pnode;
2772 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2776 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2778 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2781 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2783 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2785 struct dp_netdev_pmd_thread *non_pmd;
2787 non_pmd = xzalloc(sizeof *non_pmd);
2788 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2792 /* Caller must have valid pointer to 'pmd'. */
2794 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2796 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2800 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2802 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2803 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2807 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2808 * fails, keeps checking for next node until reaching the end of cmap.
2810 * Caller must unrefs the returned reference. */
2811 static struct dp_netdev_pmd_thread *
2812 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2814 struct dp_netdev_pmd_thread *next;
2817 struct cmap_node *node;
2819 node = cmap_next_position(&dp->poll_threads, pos);
2820 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2822 } while (next && !dp_netdev_pmd_try_ref(next));
2827 /* Configures the 'pmd' based on the input argument. */
2829 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2830 int index, unsigned core_id, int numa_id)
2834 pmd->core_id = core_id;
2835 pmd->numa_id = numa_id;
2838 atomic_init(&pmd->tx_qid,
2839 (core_id == NON_PMD_CORE_ID)
2840 ? ovs_numa_get_n_cores()
2841 : get_n_pmd_threads(dp));
2843 ovs_refcount_init(&pmd->ref_cnt);
2844 latch_init(&pmd->exit_latch);
2845 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2846 xpthread_cond_init(&pmd->cond, NULL);
2847 ovs_mutex_init(&pmd->cond_mutex);
2848 ovs_mutex_init(&pmd->flow_mutex);
2849 ovs_mutex_init(&pmd->poll_mutex);
2850 dpcls_init(&pmd->cls);
2851 cmap_init(&pmd->flow_table);
2852 ovs_list_init(&pmd->poll_list);
2853 /* init the 'flow_cache' since there is no
2854 * actual thread created for NON_PMD_CORE_ID. */
2855 if (core_id == NON_PMD_CORE_ID) {
2856 emc_cache_init(&pmd->flow_cache);
2858 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2859 hash_int(core_id, 0));
2863 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2865 dp_netdev_pmd_flow_flush(pmd);
2866 dpcls_destroy(&pmd->cls);
2867 cmap_destroy(&pmd->flow_table);
2868 ovs_mutex_destroy(&pmd->flow_mutex);
2869 latch_destroy(&pmd->exit_latch);
2870 xpthread_cond_destroy(&pmd->cond);
2871 ovs_mutex_destroy(&pmd->cond_mutex);
2872 ovs_mutex_destroy(&pmd->poll_mutex);
2876 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2877 * and unrefs the struct. */
2879 dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd)
2881 /* Uninit the 'flow_cache' since there is
2882 * no actual thread uninit it for NON_PMD_CORE_ID. */
2883 if (pmd->core_id == NON_PMD_CORE_ID) {
2884 emc_cache_uninit(&pmd->flow_cache);
2886 latch_set(&pmd->exit_latch);
2887 dp_netdev_reload_pmd__(pmd);
2888 ovs_numa_unpin_core(pmd->core_id);
2889 xpthread_join(pmd->thread, NULL);
2892 /* Unref all ports and free poll_list. */
2893 dp_netdev_pmd_clear_poll_list(pmd);
2895 /* Purges the 'pmd''s flows after stopping the thread, but before
2896 * destroying the flows, so that the flow stats can be collected. */
2897 if (dp->dp_purge_cb) {
2898 dp->dp_purge_cb(dp->dp_purge_aux, pmd->core_id);
2900 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2901 dp_netdev_pmd_unref(pmd);
2904 /* Destroys all pmd threads. */
2906 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2908 struct dp_netdev_pmd_thread *pmd;
2909 struct dp_netdev_pmd_thread **pmd_list;
2910 size_t k = 0, n_pmds;
2912 n_pmds = cmap_count(&dp->poll_threads);
2913 pmd_list = xcalloc(n_pmds, sizeof *pmd_list);
2915 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2916 /* We cannot call dp_netdev_del_pmd(), since it alters
2917 * 'dp->poll_threads' (while we're iterating it) and it
2919 ovs_assert(k < n_pmds);
2920 pmd_list[k++] = pmd;
2923 for (size_t i = 0; i < k; i++) {
2924 dp_netdev_del_pmd(dp, pmd_list[i]);
2929 /* Deletes all pmd threads on numa node 'numa_id' and
2930 * fixes tx_qids of other threads to keep them sequential. */
2932 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2934 struct dp_netdev_pmd_thread *pmd;
2935 int n_pmds_on_numa, n_pmds;
2936 int *free_idx, k = 0;
2937 struct dp_netdev_pmd_thread **pmd_list;
2939 n_pmds_on_numa = get_n_pmd_threads_on_numa(dp, numa_id);
2940 free_idx = xcalloc(n_pmds_on_numa, sizeof *free_idx);
2941 pmd_list = xcalloc(n_pmds_on_numa, sizeof *pmd_list);
2943 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2944 /* We cannot call dp_netdev_del_pmd(), since it alters
2945 * 'dp->poll_threads' (while we're iterating it) and it
2947 if (pmd->numa_id == numa_id) {
2948 atomic_read_relaxed(&pmd->tx_qid, &free_idx[k]);
2950 ovs_assert(k < n_pmds_on_numa);
2955 for (int i = 0; i < k; i++) {
2956 dp_netdev_del_pmd(dp, pmd_list[i]);
2959 n_pmds = get_n_pmd_threads(dp);
2960 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2963 atomic_read_relaxed(&pmd->tx_qid, &old_tx_qid);
2965 if (old_tx_qid >= n_pmds) {
2966 int new_tx_qid = free_idx[--k];
2968 atomic_store_relaxed(&pmd->tx_qid, new_tx_qid);
2976 /* Deletes all rx queues from pmd->poll_list. */
2978 dp_netdev_pmd_clear_poll_list(struct dp_netdev_pmd_thread *pmd)
2980 struct rxq_poll *poll;
2982 ovs_mutex_lock(&pmd->poll_mutex);
2983 LIST_FOR_EACH_POP (poll, node, &pmd->poll_list) {
2987 ovs_mutex_unlock(&pmd->poll_mutex);
2990 /* Deletes all rx queues of 'port' from poll_list of pmd thread and
2991 * reloads it if poll_list was changed. */
2993 dp_netdev_del_port_from_pmd(struct dp_netdev_port *port,
2994 struct dp_netdev_pmd_thread *pmd)
2996 struct rxq_poll *poll, *next;
2999 ovs_mutex_lock(&pmd->poll_mutex);
3000 LIST_FOR_EACH_SAFE (poll, next, node, &pmd->poll_list) {
3001 if (poll->port == port) {
3003 ovs_list_remove(&poll->node);
3008 ovs_mutex_unlock(&pmd->poll_mutex);
3010 dp_netdev_reload_pmd__(pmd);
3014 /* Deletes all rx queues of 'port' from all pmd threads of dp and
3015 * reloads them if needed. */
3017 dp_netdev_del_port_from_all_pmds(struct dp_netdev *dp,
3018 struct dp_netdev_port *port)
3020 int numa_id = netdev_get_numa_id(port->netdev);
3021 struct dp_netdev_pmd_thread *pmd;
3023 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3024 if (pmd->numa_id == numa_id) {
3025 dp_netdev_del_port_from_pmd(port, pmd);
3030 /* Returns PMD thread from this numa node with fewer rx queues to poll.
3031 * Returns NULL if there is no PMD threads on this numa node.
3032 * Can be called safely only by main thread. */
3033 static struct dp_netdev_pmd_thread *
3034 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id)
3037 struct dp_netdev_pmd_thread *pmd, *res = NULL;
3039 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3040 if (pmd->numa_id == numa_id
3041 && (min_cnt > pmd->poll_cnt || res == NULL)) {
3042 min_cnt = pmd->poll_cnt;
3050 /* Adds rx queue to poll_list of PMD thread. */
3052 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
3053 struct dp_netdev_port *port, struct netdev_rxq *rx)
3054 OVS_REQUIRES(pmd->poll_mutex)
3056 struct rxq_poll *poll = xmalloc(sizeof *poll);
3061 ovs_list_push_back(&pmd->poll_list, &poll->node);
3065 /* Distributes all rx queues of 'port' between all PMD threads and reloads
3066 * them if needed. */
3068 dp_netdev_add_port_to_pmds(struct dp_netdev *dp, struct dp_netdev_port *port)
3070 int numa_id = netdev_get_numa_id(port->netdev);
3071 struct dp_netdev_pmd_thread *pmd;
3072 struct hmapx to_reload;
3073 struct hmapx_node *node;
3076 hmapx_init(&to_reload);
3077 /* Cannot create pmd threads for invalid numa node. */
3078 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
3080 for (i = 0; i < port->n_rxq; i++) {
3081 pmd = dp_netdev_less_loaded_pmd_on_numa(dp, numa_id);
3083 /* There is no pmd threads on this numa node. */
3084 dp_netdev_set_pmds_on_numa(dp, numa_id);
3085 /* Assigning of rx queues done. */
3089 ovs_mutex_lock(&pmd->poll_mutex);
3090 dp_netdev_add_rxq_to_pmd(pmd, port, port->rxq[i]);
3091 ovs_mutex_unlock(&pmd->poll_mutex);
3093 hmapx_add(&to_reload, pmd);
3096 HMAPX_FOR_EACH (node, &to_reload) {
3097 pmd = (struct dp_netdev_pmd_thread *) node->data;
3098 dp_netdev_reload_pmd__(pmd);
3101 hmapx_destroy(&to_reload);
3104 /* Checks the numa node id of 'netdev' and starts pmd threads for
3107 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
3111 if (!ovs_numa_numa_id_is_valid(numa_id)) {
3112 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
3113 "invalid", numa_id);
3117 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
3119 /* If there are already pmd threads created for the numa node
3120 * in which 'netdev' is on, do nothing. Else, creates the
3121 * pmd threads for the numa node. */
3123 int can_have, n_unpinned, i, index = 0;
3124 struct dp_netdev_pmd_thread **pmds;
3125 struct dp_netdev_port *port;
3127 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
3129 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
3130 "cores on numa node %d", numa_id);
3134 /* If cpu mask is specified, uses all unpinned cores, otherwise
3135 * tries creating NR_PMD_THREADS pmd threads. */
3136 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
3137 pmds = xzalloc(can_have * sizeof *pmds);
3138 for (i = 0; i < can_have; i++) {
3139 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
3140 pmds[i] = xzalloc(sizeof **pmds);
3141 dp_netdev_configure_pmd(pmds[i], dp, i, core_id, numa_id);
3144 /* Distributes rx queues of this numa node between new pmd threads. */
3145 CMAP_FOR_EACH (port, node, &dp->ports) {
3146 if (netdev_is_pmd(port->netdev)
3147 && netdev_get_numa_id(port->netdev) == numa_id) {
3148 for (i = 0; i < port->n_rxq; i++) {
3149 /* Make thread-safety analyser happy. */
3150 ovs_mutex_lock(&pmds[index]->poll_mutex);
3151 dp_netdev_add_rxq_to_pmd(pmds[index], port, port->rxq[i]);
3152 ovs_mutex_unlock(&pmds[index]->poll_mutex);
3153 index = (index + 1) % can_have;
3158 /* Actual start of pmd threads. */
3159 for (i = 0; i < can_have; i++) {
3160 pmds[i]->thread = ovs_thread_create("pmd", pmd_thread_main, pmds[i]);
3163 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
3168 /* Called after pmd threads config change. Restarts pmd threads with
3169 * new configuration. */
3171 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
3173 struct dp_netdev_port *port;
3175 CMAP_FOR_EACH (port, node, &dp->ports) {
3176 if (netdev_is_pmd(port->netdev)) {
3177 int numa_id = netdev_get_numa_id(port->netdev);
3179 dp_netdev_set_pmds_on_numa(dp, numa_id);
3185 dpif_netdev_get_datapath_version(void)
3187 return xstrdup("<built-in>");
3191 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
3192 uint16_t tcp_flags, long long now)
3196 atomic_store_relaxed(&netdev_flow->stats.used, now);
3197 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
3198 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
3199 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
3201 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
3205 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
3206 enum dp_stat_type type, int cnt)
3208 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3212 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3213 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3214 enum dpif_upcall_type type, const struct nlattr *userdata,
3215 struct ofpbuf *actions, struct ofpbuf *put_actions)
3217 struct dp_netdev *dp = pmd->dp;
3218 struct flow_tnl orig_tunnel;
3221 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3225 /* Upcall processing expects the Geneve options to be in the translated
3226 * format but we need to retain the raw format for datapath use. */
3227 orig_tunnel.flags = flow->tunnel.flags;
3228 if (flow->tunnel.flags & FLOW_TNL_F_UDPIF) {
3229 orig_tunnel.metadata.present.len = flow->tunnel.metadata.present.len;
3230 memcpy(orig_tunnel.metadata.opts.gnv, flow->tunnel.metadata.opts.gnv,
3231 flow->tunnel.metadata.present.len);
3232 err = tun_metadata_from_geneve_udpif(&orig_tunnel, &orig_tunnel,
3239 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3240 struct ds ds = DS_EMPTY_INITIALIZER;
3243 struct odp_flow_key_parms odp_parms = {
3246 .odp_in_port = flow->in_port.odp_port,
3247 .support = dp_netdev_support,
3250 ofpbuf_init(&key, 0);
3251 odp_flow_key_from_flow(&odp_parms, &key);
3252 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3253 dp_packet_size(packet_));
3255 odp_flow_key_format(key.data, key.size, &ds);
3257 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3258 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3260 ofpbuf_uninit(&key);
3266 err = dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3267 actions, wc, put_actions, dp->upcall_aux);
3268 if (err && err != ENOSPC) {
3272 /* Translate tunnel metadata masks to datapath format. */
3274 if (wc->masks.tunnel.metadata.present.map) {
3275 struct geneve_opt opts[TLV_TOT_OPT_SIZE /
3276 sizeof(struct geneve_opt)];
3278 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3279 tun_metadata_to_geneve_udpif_mask(&flow->tunnel,
3281 orig_tunnel.metadata.opts.gnv,
3282 orig_tunnel.metadata.present.len,
3285 orig_tunnel.metadata.present.len = 0;
3288 memset(&wc->masks.tunnel.metadata, 0,
3289 sizeof wc->masks.tunnel.metadata);
3290 memcpy(&wc->masks.tunnel.metadata.opts.gnv, opts,
3291 orig_tunnel.metadata.present.len);
3293 wc->masks.tunnel.metadata.present.len = 0xff;
3296 /* Restore tunnel metadata. We need to use the saved options to ensure
3297 * that any unknown options are not lost. The generated mask will have
3298 * the same structure, matching on types and lengths but wildcarding
3299 * option data we don't care about. */
3300 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3301 memcpy(&flow->tunnel.metadata.opts.gnv, orig_tunnel.metadata.opts.gnv,
3302 orig_tunnel.metadata.present.len);
3303 flow->tunnel.metadata.present.len = orig_tunnel.metadata.present.len;
3304 flow->tunnel.flags |= FLOW_TNL_F_UDPIF;
3310 static inline uint32_t
3311 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3312 const struct miniflow *mf)
3314 uint32_t hash, recirc_depth;
3316 if (OVS_LIKELY(dp_packet_rss_valid(packet))) {
3317 hash = dp_packet_get_rss_hash(packet);
3319 hash = miniflow_hash_5tuple(mf, 0);
3320 dp_packet_set_rss_hash(packet, hash);
3323 /* The RSS hash must account for the recirculation depth to avoid
3324 * collisions in the exact match cache */
3325 recirc_depth = *recirc_depth_get_unsafe();
3326 if (OVS_UNLIKELY(recirc_depth)) {
3327 hash = hash_finish(hash, recirc_depth);
3328 dp_packet_set_rss_hash(packet, hash);
3333 struct packet_batch_per_flow {
3334 unsigned int byte_count;
3336 struct dp_netdev_flow *flow;
3338 struct dp_packet_batch array;
3342 packet_batch_per_flow_update(struct packet_batch_per_flow *batch,
3343 struct dp_packet *packet,
3344 const struct miniflow *mf)
3346 batch->byte_count += dp_packet_size(packet);
3347 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3348 batch->array.packets[batch->array.count++] = packet;
3352 packet_batch_per_flow_init(struct packet_batch_per_flow *batch,
3353 struct dp_netdev_flow *flow)
3355 flow->batch = batch;
3358 dp_packet_batch_init(&batch->array);
3359 batch->byte_count = 0;
3360 batch->tcp_flags = 0;
3364 packet_batch_per_flow_execute(struct packet_batch_per_flow *batch,
3365 struct dp_netdev_pmd_thread *pmd,
3368 struct dp_netdev_actions *actions;
3369 struct dp_netdev_flow *flow = batch->flow;
3371 dp_netdev_flow_used(flow, batch->array.count, batch->byte_count,
3372 batch->tcp_flags, now);
3374 actions = dp_netdev_flow_get_actions(flow);
3376 dp_netdev_execute_actions(pmd, &batch->array, true,
3377 actions->actions, actions->size);
3381 dp_netdev_queue_batches(struct dp_packet *pkt,
3382 struct dp_netdev_flow *flow, const struct miniflow *mf,
3383 struct packet_batch_per_flow *batches, size_t *n_batches)
3385 struct packet_batch_per_flow *batch = flow->batch;
3387 if (OVS_UNLIKELY(!batch)) {
3388 batch = &batches[(*n_batches)++];
3389 packet_batch_per_flow_init(batch, flow);
3392 packet_batch_per_flow_update(batch, pkt, mf);
3395 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3396 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
3397 * miniflow is copied into 'keys' and the packet pointer is moved at the
3398 * beginning of the 'packets' array.
3400 * The function returns the number of packets that needs to be processed in the
3401 * 'packets' array (they have been moved to the beginning of the vector).
3403 * If 'md_is_valid' is false, the metadata in 'packets' is not valid and must be
3404 * initialized by this function using 'port_no'.
3406 static inline size_t
3407 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet_batch *packets_,
3408 struct netdev_flow_key *keys,
3409 struct packet_batch_per_flow batches[], size_t *n_batches,
3410 bool md_is_valid, odp_port_t port_no)
3412 struct emc_cache *flow_cache = &pmd->flow_cache;
3413 struct netdev_flow_key *key = &keys[0];
3414 size_t i, n_missed = 0, n_dropped = 0;
3415 struct dp_packet **packets = packets_->packets;
3416 int cnt = packets_->count;
3418 for (i = 0; i < cnt; i++) {
3419 struct dp_netdev_flow *flow;
3420 struct dp_packet *packet = packets[i];
3422 if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) {
3423 dp_packet_delete(packet);
3429 /* Prefetch next packet data and metadata. */
3430 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3431 pkt_metadata_prefetch_init(&packets[i+1]->md);
3435 pkt_metadata_init(&packet->md, port_no);
3437 miniflow_extract(packet, &key->mf);
3438 key->len = 0; /* Not computed yet. */
3439 key->hash = dpif_netdev_packet_get_rss_hash(packet, &key->mf);
3441 flow = emc_lookup(flow_cache, key);
3442 if (OVS_LIKELY(flow)) {
3443 dp_netdev_queue_batches(packet, flow, &key->mf, batches,
3446 /* Exact match cache missed. Group missed packets together at
3447 * the beginning of the 'packets' array. */
3448 packets[n_missed] = packet;
3449 /* 'key[n_missed]' contains the key of the current packet and it
3450 * must be returned to the caller. The next key should be extracted
3451 * to 'keys[n_missed + 1]'. */
3452 key = &keys[++n_missed];
3456 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - n_dropped - n_missed);
3462 handle_packet_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet,
3463 const struct netdev_flow_key *key,
3464 struct ofpbuf *actions, struct ofpbuf *put_actions,
3467 struct ofpbuf *add_actions;
3468 struct dp_packet_batch b;
3473 match.tun_md.valid = false;
3474 miniflow_expand(&key->mf, &match.flow);
3476 ofpbuf_clear(actions);
3477 ofpbuf_clear(put_actions);
3479 dpif_flow_hash(pmd->dp->dpif, &match.flow, sizeof match.flow, &ufid);
3480 error = dp_netdev_upcall(pmd, packet, &match.flow, &match.wc,
3481 &ufid, DPIF_UC_MISS, NULL, actions,
3483 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3484 dp_packet_delete(packet);
3489 /* The Netlink encoding of datapath flow keys cannot express
3490 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3491 * tag is interpreted as exact match on the fact that there is no
3492 * VLAN. Unless we refactor a lot of code that translates between
3493 * Netlink and struct flow representations, we have to do the same
3495 if (!match.wc.masks.vlan_tci) {
3496 match.wc.masks.vlan_tci = htons(0xffff);
3499 /* We can't allow the packet batching in the next loop to execute
3500 * the actions. Otherwise, if there are any slow path actions,
3501 * we'll send the packet up twice. */
3502 packet_batch_init_packet(&b, packet);
3503 dp_netdev_execute_actions(pmd, &b, true,
3504 actions->data, actions->size);
3506 add_actions = put_actions->size ? put_actions : actions;
3507 if (OVS_LIKELY(error != ENOSPC)) {
3508 struct dp_netdev_flow *netdev_flow;
3510 /* XXX: There's a race window where a flow covering this packet
3511 * could have already been installed since we last did the flow
3512 * lookup before upcall. This could be solved by moving the
3513 * mutex lock outside the loop, but that's an awful long time
3514 * to be locking everyone out of making flow installs. If we
3515 * move to a per-core classifier, it would be reasonable. */
3516 ovs_mutex_lock(&pmd->flow_mutex);
3517 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, key);
3518 if (OVS_LIKELY(!netdev_flow)) {
3519 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3523 ovs_mutex_unlock(&pmd->flow_mutex);
3525 emc_insert(&pmd->flow_cache, key, netdev_flow);
3530 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3531 struct dp_packet_batch *packets_,
3532 struct netdev_flow_key *keys,
3533 struct packet_batch_per_flow batches[], size_t *n_batches)
3535 int cnt = packets_->count;
3536 #if !defined(__CHECKER__) && !defined(_WIN32)
3537 const size_t PKT_ARRAY_SIZE = cnt;
3539 /* Sparse or MSVC doesn't like variable length array. */
3540 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3542 struct dp_packet **packets = packets_->packets;
3543 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3544 struct dp_netdev *dp = pmd->dp;
3545 struct emc_cache *flow_cache = &pmd->flow_cache;
3546 int miss_cnt = 0, lost_cnt = 0;
3550 for (i = 0; i < cnt; i++) {
3551 /* Key length is needed in all the cases, hash computed on demand. */
3552 keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf));
3554 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3555 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3556 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3557 struct ofpbuf actions, put_actions;
3559 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3560 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3562 for (i = 0; i < cnt; i++) {
3563 struct dp_netdev_flow *netdev_flow;
3565 if (OVS_LIKELY(rules[i])) {
3569 /* It's possible that an earlier slow path execution installed
3570 * a rule covering this flow. In this case, it's a lot cheaper
3571 * to catch it here than execute a miss. */
3572 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3574 rules[i] = &netdev_flow->cr;
3579 handle_packet_upcall(pmd, packets[i], &keys[i], &actions, &put_actions,
3583 ofpbuf_uninit(&actions);
3584 ofpbuf_uninit(&put_actions);
3585 fat_rwlock_unlock(&dp->upcall_rwlock);
3586 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3587 } else if (OVS_UNLIKELY(any_miss)) {
3588 for (i = 0; i < cnt; i++) {
3589 if (OVS_UNLIKELY(!rules[i])) {
3590 dp_packet_delete(packets[i]);
3597 for (i = 0; i < cnt; i++) {
3598 struct dp_packet *packet = packets[i];
3599 struct dp_netdev_flow *flow;
3601 if (OVS_UNLIKELY(!rules[i])) {
3605 flow = dp_netdev_flow_cast(rules[i]);
3607 emc_insert(flow_cache, &keys[i], flow);
3608 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3611 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3612 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3613 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3616 /* Packets enter the datapath from a port (or from recirculation) here.
3618 * For performance reasons a caller may choose not to initialize the metadata
3619 * in 'packets': in this case 'mdinit' is false and this function needs to
3620 * initialize it using 'port_no'. If the metadata in 'packets' is already
3621 * valid, 'md_is_valid' must be true and 'port_no' will be ignored. */
3623 dp_netdev_input__(struct dp_netdev_pmd_thread *pmd,
3624 struct dp_packet_batch *packets,
3625 bool md_is_valid, odp_port_t port_no)
3627 int cnt = packets->count;
3628 #if !defined(__CHECKER__) && !defined(_WIN32)
3629 const size_t PKT_ARRAY_SIZE = cnt;
3631 /* Sparse or MSVC doesn't like variable length array. */
3632 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3634 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3635 struct packet_batch_per_flow batches[PKT_ARRAY_SIZE];
3636 long long now = time_msec();
3637 size_t newcnt, n_batches, i;
3640 newcnt = emc_processing(pmd, packets, keys, batches, &n_batches,
3641 md_is_valid, port_no);
3642 if (OVS_UNLIKELY(newcnt)) {
3643 packets->count = newcnt;
3644 fast_path_processing(pmd, packets, keys, batches, &n_batches);
3647 for (i = 0; i < n_batches; i++) {
3648 batches[i].flow->batch = NULL;
3651 for (i = 0; i < n_batches; i++) {
3652 packet_batch_per_flow_execute(&batches[i], pmd, now);
3657 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3658 struct dp_packet_batch *packets,
3661 dp_netdev_input__(pmd, packets, false, port_no);
3665 dp_netdev_recirculate(struct dp_netdev_pmd_thread *pmd,
3666 struct dp_packet_batch *packets)
3668 dp_netdev_input__(pmd, packets, true, 0);
3671 struct dp_netdev_execute_aux {
3672 struct dp_netdev_pmd_thread *pmd;
3676 dpif_netdev_register_dp_purge_cb(struct dpif *dpif, dp_purge_callback *cb,
3679 struct dp_netdev *dp = get_dp_netdev(dpif);
3680 dp->dp_purge_aux = aux;
3681 dp->dp_purge_cb = cb;
3685 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3688 struct dp_netdev *dp = get_dp_netdev(dpif);
3689 dp->upcall_aux = aux;
3694 push_tnl_action(const struct dp_netdev *dp,
3695 const struct nlattr *attr,
3696 struct dp_packet_batch *batch)
3698 struct dp_netdev_port *tun_port;
3699 const struct ovs_action_push_tnl *data;
3702 data = nl_attr_get(attr);
3704 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3709 err = netdev_push_header(tun_port->netdev, batch, data);
3714 dp_packet_delete_batch(batch, true);
3719 dp_execute_cb(void *aux_, struct dp_packet_batch *packets_,
3720 const struct nlattr *a, bool may_steal)
3721 OVS_NO_THREAD_SAFETY_ANALYSIS
3723 struct dp_netdev_execute_aux *aux = aux_;
3724 uint32_t *depth = recirc_depth_get();
3725 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3726 struct dp_netdev *dp = pmd->dp;
3727 int type = nl_attr_type(a);
3728 struct dp_netdev_port *p;
3730 switch ((enum ovs_action_attr)type) {
3731 case OVS_ACTION_ATTR_OUTPUT:
3732 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3733 if (OVS_LIKELY(p)) {
3736 atomic_read_relaxed(&pmd->tx_qid, &tx_qid);
3738 netdev_send(p->netdev, tx_qid, packets_, may_steal);
3743 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3744 if (*depth < MAX_RECIRC_DEPTH) {
3745 struct dp_packet_batch tnl_pkt;
3749 dp_packet_batch_clone(&tnl_pkt, packets_);
3750 packets_ = &tnl_pkt;
3753 err = push_tnl_action(dp, a, packets_);
3756 dp_netdev_recirculate(pmd, packets_);
3763 case OVS_ACTION_ATTR_TUNNEL_POP:
3764 if (*depth < MAX_RECIRC_DEPTH) {
3765 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3767 p = dp_netdev_lookup_port(dp, portno);
3769 struct dp_packet_batch tnl_pkt;
3773 dp_packet_batch_clone(&tnl_pkt, packets_);
3774 packets_ = &tnl_pkt;
3777 netdev_pop_header(p->netdev, packets_);
3778 if (!packets_->count) {
3782 for (i = 0; i < packets_->count; i++) {
3783 packets_->packets[i]->md.in_port.odp_port = portno;
3787 dp_netdev_recirculate(pmd, packets_);
3794 case OVS_ACTION_ATTR_USERSPACE:
3795 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3796 struct dp_packet **packets = packets_->packets;
3797 const struct nlattr *userdata;
3798 struct ofpbuf actions;
3803 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3804 ofpbuf_init(&actions, 0);
3806 for (i = 0; i < packets_->count; i++) {
3808 struct dp_packet_batch b;
3810 ofpbuf_clear(&actions);
3812 flow_extract(packets[i], &flow);
3813 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3814 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3815 DPIF_UC_ACTION, userdata,&actions,
3817 if (!error || error == ENOSPC) {
3818 packet_batch_init_packet(&b, packets[i]);
3819 dp_netdev_execute_actions(pmd, &b, may_steal,
3820 actions.data, actions.size);
3821 } else if (may_steal) {
3822 dp_packet_delete(packets[i]);
3825 ofpbuf_uninit(&actions);
3826 fat_rwlock_unlock(&dp->upcall_rwlock);
3832 case OVS_ACTION_ATTR_RECIRC:
3833 if (*depth < MAX_RECIRC_DEPTH) {
3834 struct dp_packet_batch recirc_pkts;
3838 dp_packet_batch_clone(&recirc_pkts, packets_);
3839 packets_ = &recirc_pkts;
3842 for (i = 0; i < packets_->count; i++) {
3843 packets_->packets[i]->md.recirc_id = nl_attr_get_u32(a);
3847 dp_netdev_recirculate(pmd, packets_);
3853 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3856 case OVS_ACTION_ATTR_CT:
3857 /* If a flow with this action is slow-pathed, datapath assistance is
3858 * required to implement it. However, we don't support this action
3859 * in the userspace datapath. */
3860 VLOG_WARN("Cannot execute conntrack action in userspace.");
3863 case OVS_ACTION_ATTR_PUSH_VLAN:
3864 case OVS_ACTION_ATTR_POP_VLAN:
3865 case OVS_ACTION_ATTR_PUSH_MPLS:
3866 case OVS_ACTION_ATTR_POP_MPLS:
3867 case OVS_ACTION_ATTR_SET:
3868 case OVS_ACTION_ATTR_SET_MASKED:
3869 case OVS_ACTION_ATTR_SAMPLE:
3870 case OVS_ACTION_ATTR_HASH:
3871 case OVS_ACTION_ATTR_UNSPEC:
3872 case __OVS_ACTION_ATTR_MAX:
3876 dp_packet_delete_batch(packets_, may_steal);
3880 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3881 struct dp_packet_batch *packets,
3883 const struct nlattr *actions, size_t actions_len)
3885 struct dp_netdev_execute_aux aux = { pmd };
3887 odp_execute_actions(&aux, packets, may_steal, actions,
3888 actions_len, dp_execute_cb);
3891 const struct dpif_class dpif_netdev_class = {
3894 dpif_netdev_enumerate,
3895 dpif_netdev_port_open_type,
3898 dpif_netdev_destroy,
3901 dpif_netdev_get_stats,
3902 dpif_netdev_port_add,
3903 dpif_netdev_port_del,
3904 dpif_netdev_port_query_by_number,
3905 dpif_netdev_port_query_by_name,
3906 NULL, /* port_get_pid */
3907 dpif_netdev_port_dump_start,
3908 dpif_netdev_port_dump_next,
3909 dpif_netdev_port_dump_done,
3910 dpif_netdev_port_poll,
3911 dpif_netdev_port_poll_wait,
3912 dpif_netdev_flow_flush,
3913 dpif_netdev_flow_dump_create,
3914 dpif_netdev_flow_dump_destroy,
3915 dpif_netdev_flow_dump_thread_create,
3916 dpif_netdev_flow_dump_thread_destroy,
3917 dpif_netdev_flow_dump_next,
3918 dpif_netdev_operate,
3919 NULL, /* recv_set */
3920 NULL, /* handlers_set */
3921 dpif_netdev_pmd_set,
3922 dpif_netdev_queue_to_priority,
3924 NULL, /* recv_wait */
3925 NULL, /* recv_purge */
3926 dpif_netdev_register_dp_purge_cb,
3927 dpif_netdev_register_upcall_cb,
3928 dpif_netdev_enable_upcall,
3929 dpif_netdev_disable_upcall,
3930 dpif_netdev_get_datapath_version,
3931 NULL, /* ct_dump_start */
3932 NULL, /* ct_dump_next */
3933 NULL, /* ct_dump_done */
3934 NULL, /* ct_flush */
3938 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3939 const char *argv[], void *aux OVS_UNUSED)
3941 struct dp_netdev_port *old_port;
3942 struct dp_netdev_port *new_port;
3943 struct dp_netdev *dp;
3946 ovs_mutex_lock(&dp_netdev_mutex);
3947 dp = shash_find_data(&dp_netdevs, argv[1]);
3948 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3949 ovs_mutex_unlock(&dp_netdev_mutex);
3950 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3953 ovs_refcount_ref(&dp->ref_cnt);
3954 ovs_mutex_unlock(&dp_netdev_mutex);
3956 ovs_mutex_lock(&dp->port_mutex);
3957 if (get_port_by_name(dp, argv[2], &old_port)) {
3958 unixctl_command_reply_error(conn, "unknown port");
3962 port_no = u32_to_odp(atoi(argv[3]));
3963 if (!port_no || port_no == ODPP_NONE) {
3964 unixctl_command_reply_error(conn, "bad port number");
3967 if (dp_netdev_lookup_port(dp, port_no)) {
3968 unixctl_command_reply_error(conn, "port number already in use");
3972 /* Remove old port. */
3973 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3974 ovsrcu_postpone(free, old_port);
3976 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3977 new_port = xmemdup(old_port, sizeof *old_port);
3978 new_port->port_no = port_no;
3979 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3981 seq_change(dp->port_seq);
3982 unixctl_command_reply(conn, NULL);
3985 ovs_mutex_unlock(&dp->port_mutex);
3986 dp_netdev_unref(dp);
3990 dpif_dummy_register__(const char *type)
3992 struct dpif_class *class;
3994 class = xmalloc(sizeof *class);
3995 *class = dpif_netdev_class;
3996 class->type = xstrdup(type);
3997 dp_register_provider(class);
4001 dpif_dummy_override(const char *type)
4006 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
4007 * a userland-only build. It's useful for testsuite.
4009 error = dp_unregister_provider(type);
4010 if (error == 0 || error == EAFNOSUPPORT) {
4011 dpif_dummy_register__(type);
4016 dpif_dummy_register(enum dummy_level level)
4018 if (level == DUMMY_OVERRIDE_ALL) {
4023 dp_enumerate_types(&types);
4024 SSET_FOR_EACH (type, &types) {
4025 dpif_dummy_override(type);
4027 sset_destroy(&types);
4028 } else if (level == DUMMY_OVERRIDE_SYSTEM) {
4029 dpif_dummy_override("system");
4032 dpif_dummy_register__("dummy");
4034 unixctl_command_register("dpif-dummy/change-port-number",
4035 "dp port new-number",
4036 3, 3, dpif_dummy_change_port_number, NULL);
4039 /* Datapath Classifier. */
4041 /* A set of rules that all have the same fields wildcarded. */
4042 struct dpcls_subtable {
4043 /* The fields are only used by writers. */
4044 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
4046 /* These fields are accessed by readers. */
4047 struct cmap rules; /* Contains "struct dpcls_rule"s. */
4048 struct netdev_flow_key mask; /* Wildcards for fields (const). */
4049 /* 'mask' must be the last field, additional space is allocated here. */
4052 /* Initializes 'cls' as a classifier that initially contains no classification
4055 dpcls_init(struct dpcls *cls)
4057 cmap_init(&cls->subtables_map);
4058 pvector_init(&cls->subtables);
4062 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
4064 pvector_remove(&cls->subtables, subtable);
4065 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
4066 subtable->mask.hash);
4067 cmap_destroy(&subtable->rules);
4068 ovsrcu_postpone(free, subtable);
4071 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
4072 * caller's responsibility.
4073 * May only be called after all the readers have been terminated. */
4075 dpcls_destroy(struct dpcls *cls)
4078 struct dpcls_subtable *subtable;
4080 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
4081 ovs_assert(cmap_count(&subtable->rules) == 0);
4082 dpcls_destroy_subtable(cls, subtable);
4084 cmap_destroy(&cls->subtables_map);
4085 pvector_destroy(&cls->subtables);
4089 static struct dpcls_subtable *
4090 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
4092 struct dpcls_subtable *subtable;
4094 /* Need to add one. */
4095 subtable = xmalloc(sizeof *subtable
4096 - sizeof subtable->mask.mf + mask->len);
4097 cmap_init(&subtable->rules);
4098 netdev_flow_key_clone(&subtable->mask, mask);
4099 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
4100 pvector_insert(&cls->subtables, subtable, 0);
4101 pvector_publish(&cls->subtables);
4106 static inline struct dpcls_subtable *
4107 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
4109 struct dpcls_subtable *subtable;
4111 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
4112 &cls->subtables_map) {
4113 if (netdev_flow_key_equal(&subtable->mask, mask)) {
4117 return dpcls_create_subtable(cls, mask);
4120 /* Insert 'rule' into 'cls'. */
4122 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
4123 const struct netdev_flow_key *mask)
4125 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
4127 rule->mask = &subtable->mask;
4128 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
4131 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
4133 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
4135 struct dpcls_subtable *subtable;
4137 ovs_assert(rule->mask);
4139 INIT_CONTAINER(subtable, rule->mask, mask);
4141 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
4143 dpcls_destroy_subtable(cls, subtable);
4144 pvector_publish(&cls->subtables);
4148 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
4149 * in 'mask' the values in 'key' and 'target' are the same. */
4151 dpcls_rule_matches_key(const struct dpcls_rule *rule,
4152 const struct netdev_flow_key *target)
4154 const uint64_t *keyp = miniflow_get_values(&rule->flow.mf);
4155 const uint64_t *maskp = miniflow_get_values(&rule->mask->mf);
4158 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, target, rule->flow.mf.map) {
4159 if (OVS_UNLIKELY((value & *maskp++) != *keyp++)) {
4166 /* For each miniflow in 'flows' performs a classifier lookup writing the result
4167 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
4168 * NULL it is skipped.
4170 * This function is optimized for use in the userspace datapath and therefore
4171 * does not implement a lot of features available in the standard
4172 * classifier_lookup() function. Specifically, it does not implement
4173 * priorities, instead returning any rule which matches the flow.
4175 * Returns true if all flows found a corresponding rule. */
4177 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
4178 struct dpcls_rule **rules, const size_t cnt)
4180 /* The batch size 16 was experimentally found faster than 8 or 32. */
4181 typedef uint16_t map_type;
4182 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
4184 #if !defined(__CHECKER__) && !defined(_WIN32)
4185 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
4187 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
4189 map_type maps[N_MAPS];
4190 struct dpcls_subtable *subtable;
4192 memset(maps, 0xff, sizeof maps);
4193 if (cnt % MAP_BITS) {
4194 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
4196 memset(rules, 0, cnt * sizeof *rules);
4198 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
4199 const struct netdev_flow_key *mkeys = keys;
4200 struct dpcls_rule **mrules = rules;
4201 map_type remains = 0;
4204 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
4206 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
4207 uint32_t hashes[MAP_BITS];
4208 const struct cmap_node *nodes[MAP_BITS];
4209 unsigned long map = maps[m];
4213 continue; /* Skip empty maps. */
4216 /* Compute hashes for the remaining keys. */
4217 ULLONG_FOR_EACH_1(i, map) {
4218 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
4222 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
4223 /* Check results. */
4224 ULLONG_FOR_EACH_1(i, map) {
4225 struct dpcls_rule *rule;
4227 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
4228 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
4233 ULLONG_SET0(map, i); /* Did not match. */
4235 ; /* Keep Sparse happy. */
4237 maps[m] &= ~map; /* Clear the found rules. */
4241 return true; /* All found. */
4244 return false; /* Some misses. */