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 unsigned core_id; /* CPU core id of this pmd thread. */
434 int numa_id; /* numa node id of this pmd thread. */
435 atomic_int tx_qid; /* Queue id used by this pmd thread to
436 * send packets on all netdevs */
438 struct ovs_mutex poll_mutex; /* Mutex for poll_list. */
439 /* List of rx queues to poll. */
440 struct ovs_list poll_list OVS_GUARDED;
441 int poll_cnt; /* Number of elemints in poll_list. */
443 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
444 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
445 * values and subtracts them from 'stats' and 'cycles' before
446 * reporting to the user */
447 unsigned long long stats_zero[DP_N_STATS];
448 uint64_t cycles_zero[PMD_N_CYCLES];
451 #define PMD_INITIAL_SEQ 1
453 /* Interface to netdev-based datapath. */
456 struct dp_netdev *dp;
457 uint64_t last_port_seq;
460 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
461 struct dp_netdev_port **portp);
462 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
463 struct dp_netdev_port **portp);
464 static void dp_netdev_free(struct dp_netdev *)
465 OVS_REQUIRES(dp_netdev_mutex);
466 static int do_add_port(struct dp_netdev *dp, const char *devname,
467 const char *type, odp_port_t port_no)
468 OVS_REQUIRES(dp->port_mutex);
469 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
470 OVS_REQUIRES(dp->port_mutex);
471 static int dpif_netdev_open(const struct dpif_class *, const char *name,
472 bool create, struct dpif **);
473 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
474 struct dp_packet_batch *,
476 const struct nlattr *actions,
478 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
479 struct dp_packet_batch *, odp_port_t port_no);
480 static void dp_netdev_recirculate(struct dp_netdev_pmd_thread *,
481 struct dp_packet_batch *);
483 static void dp_netdev_disable_upcall(struct dp_netdev *);
484 static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
485 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
486 struct dp_netdev *dp, unsigned core_id,
488 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
489 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
490 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
492 static struct dp_netdev_pmd_thread *
493 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
494 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
495 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
496 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
497 static void dp_netdev_pmd_clear_poll_list(struct dp_netdev_pmd_thread *pmd);
498 static void dp_netdev_del_port_from_all_pmds(struct dp_netdev *dp,
499 struct dp_netdev_port *port);
501 dp_netdev_add_port_to_pmds(struct dp_netdev *dp, struct dp_netdev_port *port);
503 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
504 struct dp_netdev_port *port, struct netdev_rxq *rx);
505 static struct dp_netdev_pmd_thread *
506 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id);
507 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
508 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
509 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
510 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
512 static inline bool emc_entry_alive(struct emc_entry *ce);
513 static void emc_clear_entry(struct emc_entry *ce);
516 emc_cache_init(struct emc_cache *flow_cache)
520 flow_cache->sweep_idx = 0;
521 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
522 flow_cache->entries[i].flow = NULL;
523 flow_cache->entries[i].key.hash = 0;
524 flow_cache->entries[i].key.len = sizeof(struct miniflow);
525 flowmap_init(&flow_cache->entries[i].key.mf.map);
530 emc_cache_uninit(struct emc_cache *flow_cache)
534 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
535 emc_clear_entry(&flow_cache->entries[i]);
539 /* Check and clear dead flow references slowly (one entry at each
542 emc_cache_slow_sweep(struct emc_cache *flow_cache)
544 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
546 if (!emc_entry_alive(entry)) {
547 emc_clear_entry(entry);
549 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
552 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
554 dpif_is_netdev(const struct dpif *dpif)
556 return dpif->dpif_class->open == dpif_netdev_open;
559 static struct dpif_netdev *
560 dpif_netdev_cast(const struct dpif *dpif)
562 ovs_assert(dpif_is_netdev(dpif));
563 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
566 static struct dp_netdev *
567 get_dp_netdev(const struct dpif *dpif)
569 return dpif_netdev_cast(dpif)->dp;
573 PMD_INFO_SHOW_STATS, /* Show how cpu cycles are spent. */
574 PMD_INFO_CLEAR_STATS, /* Set the cycles count to 0. */
575 PMD_INFO_SHOW_RXQ /* Show poll-lists of pmd threads. */
579 pmd_info_show_stats(struct ds *reply,
580 struct dp_netdev_pmd_thread *pmd,
581 unsigned long long stats[DP_N_STATS],
582 uint64_t cycles[PMD_N_CYCLES])
584 unsigned long long total_packets = 0;
585 uint64_t total_cycles = 0;
588 /* These loops subtracts reference values ('*_zero') from the counters.
589 * Since loads and stores are relaxed, it might be possible for a '*_zero'
590 * value to be more recent than the current value we're reading from the
591 * counter. This is not a big problem, since these numbers are not
592 * supposed to be too accurate, but we should at least make sure that
593 * the result is not negative. */
594 for (i = 0; i < DP_N_STATS; i++) {
595 if (stats[i] > pmd->stats_zero[i]) {
596 stats[i] -= pmd->stats_zero[i];
601 if (i != DP_STAT_LOST) {
602 /* Lost packets are already included in DP_STAT_MISS */
603 total_packets += stats[i];
607 for (i = 0; i < PMD_N_CYCLES; i++) {
608 if (cycles[i] > pmd->cycles_zero[i]) {
609 cycles[i] -= pmd->cycles_zero[i];
614 total_cycles += cycles[i];
617 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
618 ? "main thread" : "pmd thread");
620 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
621 ds_put_format(reply, " numa_id %d", pmd->numa_id);
623 if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) {
624 ds_put_format(reply, " core_id %u", pmd->core_id);
626 ds_put_cstr(reply, ":\n");
629 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
630 "\tmiss:%llu\n\tlost:%llu\n",
631 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
632 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
634 if (total_cycles == 0) {
639 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
640 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
641 cycles[PMD_CYCLES_POLLING],
642 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
643 cycles[PMD_CYCLES_PROCESSING],
644 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
646 if (total_packets == 0) {
651 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
652 total_cycles / (double)total_packets,
653 total_cycles, total_packets);
656 "\tavg processing cycles per packet: "
657 "%.02f (%"PRIu64"/%llu)\n",
658 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
659 cycles[PMD_CYCLES_PROCESSING], total_packets);
663 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
664 struct dp_netdev_pmd_thread *pmd,
665 unsigned long long stats[DP_N_STATS],
666 uint64_t cycles[PMD_N_CYCLES])
670 /* We cannot write 'stats' and 'cycles' (because they're written by other
671 * threads) and we shouldn't change 'stats' (because they're used to count
672 * datapath stats, which must not be cleared here). Instead, we save the
673 * current values and subtract them from the values to be displayed in the
675 for (i = 0; i < DP_N_STATS; i++) {
676 pmd->stats_zero[i] = stats[i];
678 for (i = 0; i < PMD_N_CYCLES; i++) {
679 pmd->cycles_zero[i] = cycles[i];
684 pmd_info_show_rxq(struct ds *reply, struct dp_netdev_pmd_thread *pmd)
686 if (pmd->core_id != NON_PMD_CORE_ID) {
687 struct rxq_poll *poll;
688 const char *prev_name = NULL;
690 ds_put_format(reply, "pmd thread numa_id %d core_id %u:\n",
691 pmd->numa_id, pmd->core_id);
693 ovs_mutex_lock(&pmd->poll_mutex);
694 LIST_FOR_EACH (poll, node, &pmd->poll_list) {
695 const char *name = netdev_get_name(poll->port->netdev);
697 if (!prev_name || strcmp(name, prev_name)) {
699 ds_put_cstr(reply, "\n");
701 ds_put_format(reply, "\tport: %s\tqueue-id:",
702 netdev_get_name(poll->port->netdev));
704 ds_put_format(reply, " %d", netdev_rxq_get_queue_id(poll->rx));
707 ovs_mutex_unlock(&pmd->poll_mutex);
708 ds_put_cstr(reply, "\n");
713 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
716 struct ds reply = DS_EMPTY_INITIALIZER;
717 struct dp_netdev_pmd_thread *pmd;
718 struct dp_netdev *dp = NULL;
719 enum pmd_info_type type = *(enum pmd_info_type *) aux;
721 ovs_mutex_lock(&dp_netdev_mutex);
724 dp = shash_find_data(&dp_netdevs, argv[1]);
725 } else if (shash_count(&dp_netdevs) == 1) {
726 /* There's only one datapath */
727 dp = shash_first(&dp_netdevs)->data;
731 ovs_mutex_unlock(&dp_netdev_mutex);
732 unixctl_command_reply_error(conn,
733 "please specify an existing datapath");
737 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
738 if (type == PMD_INFO_SHOW_RXQ) {
739 pmd_info_show_rxq(&reply, pmd);
741 unsigned long long stats[DP_N_STATS];
742 uint64_t cycles[PMD_N_CYCLES];
745 /* Read current stats and cycle counters */
746 for (i = 0; i < ARRAY_SIZE(stats); i++) {
747 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
749 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
750 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
753 if (type == PMD_INFO_CLEAR_STATS) {
754 pmd_info_clear_stats(&reply, pmd, stats, cycles);
755 } else if (type == PMD_INFO_SHOW_STATS) {
756 pmd_info_show_stats(&reply, pmd, stats, cycles);
761 ovs_mutex_unlock(&dp_netdev_mutex);
763 unixctl_command_reply(conn, ds_cstr(&reply));
768 dpif_netdev_init(void)
770 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
771 clear_aux = PMD_INFO_CLEAR_STATS,
772 poll_aux = PMD_INFO_SHOW_RXQ;
774 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
775 0, 1, dpif_netdev_pmd_info,
777 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
778 0, 1, dpif_netdev_pmd_info,
780 unixctl_command_register("dpif-netdev/pmd-rxq-show", "[dp]",
781 0, 1, dpif_netdev_pmd_info,
787 dpif_netdev_enumerate(struct sset *all_dps,
788 const struct dpif_class *dpif_class)
790 struct shash_node *node;
792 ovs_mutex_lock(&dp_netdev_mutex);
793 SHASH_FOR_EACH(node, &dp_netdevs) {
794 struct dp_netdev *dp = node->data;
795 if (dpif_class != dp->class) {
796 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
797 * If the class doesn't match, skip this dpif. */
800 sset_add(all_dps, node->name);
802 ovs_mutex_unlock(&dp_netdev_mutex);
808 dpif_netdev_class_is_dummy(const struct dpif_class *class)
810 return class != &dpif_netdev_class;
814 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
816 return strcmp(type, "internal") ? type
817 : dpif_netdev_class_is_dummy(class) ? "dummy"
822 create_dpif_netdev(struct dp_netdev *dp)
824 uint16_t netflow_id = hash_string(dp->name, 0);
825 struct dpif_netdev *dpif;
827 ovs_refcount_ref(&dp->ref_cnt);
829 dpif = xmalloc(sizeof *dpif);
830 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
832 dpif->last_port_seq = seq_read(dp->port_seq);
837 /* Choose an unused, non-zero port number and return it on success.
838 * Return ODPP_NONE on failure. */
840 choose_port(struct dp_netdev *dp, const char *name)
841 OVS_REQUIRES(dp->port_mutex)
845 if (dp->class != &dpif_netdev_class) {
849 /* If the port name begins with "br", start the number search at
850 * 100 to make writing tests easier. */
851 if (!strncmp(name, "br", 2)) {
855 /* If the port name contains a number, try to assign that port number.
856 * This can make writing unit tests easier because port numbers are
858 for (p = name; *p != '\0'; p++) {
859 if (isdigit((unsigned char) *p)) {
860 port_no = start_no + strtol(p, NULL, 10);
861 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
862 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
863 return u32_to_odp(port_no);
870 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
871 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
872 return u32_to_odp(port_no);
880 create_dp_netdev(const char *name, const struct dpif_class *class,
881 struct dp_netdev **dpp)
882 OVS_REQUIRES(dp_netdev_mutex)
884 struct dp_netdev *dp;
887 dp = xzalloc(sizeof *dp);
888 shash_add(&dp_netdevs, name, dp);
890 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
891 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
892 ovs_refcount_init(&dp->ref_cnt);
893 atomic_flag_clear(&dp->destroyed);
895 ovs_mutex_init(&dp->port_mutex);
896 cmap_init(&dp->ports);
897 dp->port_seq = seq_create();
898 fat_rwlock_init(&dp->upcall_rwlock);
900 /* Disable upcalls by default. */
901 dp_netdev_disable_upcall(dp);
902 dp->upcall_aux = NULL;
903 dp->upcall_cb = NULL;
905 cmap_init(&dp->poll_threads);
906 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
907 ovsthread_key_create(&dp->per_pmd_key, NULL);
909 dp_netdev_set_nonpmd(dp);
911 ovs_mutex_lock(&dp->port_mutex);
912 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
913 ovs_mutex_unlock(&dp->port_mutex);
919 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
925 dpif_netdev_open(const struct dpif_class *class, const char *name,
926 bool create, struct dpif **dpifp)
928 struct dp_netdev *dp;
931 ovs_mutex_lock(&dp_netdev_mutex);
932 dp = shash_find_data(&dp_netdevs, name);
934 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
936 error = (dp->class != class ? EINVAL
941 *dpifp = create_dpif_netdev(dp);
944 ovs_mutex_unlock(&dp_netdev_mutex);
950 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
951 OVS_NO_THREAD_SAFETY_ANALYSIS
953 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
954 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
956 /* Before freeing a lock we should release it */
957 fat_rwlock_unlock(&dp->upcall_rwlock);
958 fat_rwlock_destroy(&dp->upcall_rwlock);
961 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
962 * through the 'dp_netdevs' shash while freeing 'dp'. */
964 dp_netdev_free(struct dp_netdev *dp)
965 OVS_REQUIRES(dp_netdev_mutex)
967 struct dp_netdev_port *port;
969 shash_find_and_delete(&dp_netdevs, dp->name);
971 dp_netdev_destroy_all_pmds(dp);
972 ovs_mutex_destroy(&dp->non_pmd_mutex);
973 ovsthread_key_delete(dp->per_pmd_key);
975 ovs_mutex_lock(&dp->port_mutex);
976 CMAP_FOR_EACH (port, node, &dp->ports) {
977 /* PMD threads are destroyed here. do_del_port() cannot quiesce */
978 do_del_port(dp, port);
980 ovs_mutex_unlock(&dp->port_mutex);
981 cmap_destroy(&dp->poll_threads);
983 seq_destroy(dp->port_seq);
984 cmap_destroy(&dp->ports);
985 ovs_mutex_destroy(&dp->port_mutex);
987 /* Upcalls must be disabled at this point */
988 dp_netdev_destroy_upcall_lock(dp);
991 free(CONST_CAST(char *, dp->name));
996 dp_netdev_unref(struct dp_netdev *dp)
999 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
1000 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
1001 ovs_mutex_lock(&dp_netdev_mutex);
1002 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
1005 ovs_mutex_unlock(&dp_netdev_mutex);
1010 dpif_netdev_close(struct dpif *dpif)
1012 struct dp_netdev *dp = get_dp_netdev(dpif);
1014 dp_netdev_unref(dp);
1019 dpif_netdev_destroy(struct dpif *dpif)
1021 struct dp_netdev *dp = get_dp_netdev(dpif);
1023 if (!atomic_flag_test_and_set(&dp->destroyed)) {
1024 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
1025 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
1033 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
1034 * load/store semantics. While the increment is not atomic, the load and
1035 * store operations are, making it impossible to read inconsistent values.
1037 * This is used to update thread local stats counters. */
1039 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
1041 unsigned long long tmp;
1043 atomic_read_relaxed(var, &tmp);
1045 atomic_store_relaxed(var, tmp);
1049 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
1051 struct dp_netdev *dp = get_dp_netdev(dpif);
1052 struct dp_netdev_pmd_thread *pmd;
1054 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
1055 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1056 unsigned long long n;
1057 stats->n_flows += cmap_count(&pmd->flow_table);
1059 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
1061 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
1063 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
1064 stats->n_missed += n;
1065 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
1068 stats->n_masks = UINT32_MAX;
1069 stats->n_mask_hit = UINT64_MAX;
1075 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1079 if (pmd->core_id == NON_PMD_CORE_ID) {
1083 ovs_mutex_lock(&pmd->cond_mutex);
1084 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1085 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1086 ovs_mutex_unlock(&pmd->cond_mutex);
1090 hash_port_no(odp_port_t port_no)
1092 return hash_int(odp_to_u32(port_no), 0);
1096 port_create(const char *devname, const char *open_type, const char *type,
1097 odp_port_t port_no, struct dp_netdev_port **portp)
1099 struct netdev_saved_flags *sf;
1100 struct dp_netdev_port *port;
1101 enum netdev_flags flags;
1102 struct netdev *netdev;
1103 int n_open_rxqs = 0;
1108 /* Open and validate network device. */
1109 error = netdev_open(devname, open_type, &netdev);
1113 /* XXX reject non-Ethernet devices */
1115 netdev_get_flags(netdev, &flags);
1116 if (flags & NETDEV_LOOPBACK) {
1117 VLOG_ERR("%s: cannot add a loopback device", devname);
1122 if (netdev_is_pmd(netdev)) {
1123 int n_cores = ovs_numa_get_n_cores();
1125 if (n_cores == OVS_CORE_UNSPEC) {
1126 VLOG_ERR("%s, cannot get cpu core info", devname);
1130 /* There can only be ovs_numa_get_n_cores() pmd threads,
1131 * so creates a txq for each, and one extra for the non
1133 error = netdev_set_multiq(netdev, n_cores + 1,
1134 netdev_requested_n_rxq(netdev));
1135 if (error && (error != EOPNOTSUPP)) {
1136 VLOG_ERR("%s, cannot set multiq", devname);
1140 port = xzalloc(sizeof *port);
1141 port->port_no = port_no;
1142 port->netdev = netdev;
1143 port->n_rxq = netdev_n_rxq(netdev);
1144 port->rxq = xcalloc(port->n_rxq, sizeof *port->rxq);
1145 port->type = xstrdup(type);
1146 port->latest_requested_n_rxq = netdev_requested_n_rxq(netdev);
1148 for (i = 0; i < port->n_rxq; i++) {
1149 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1151 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1152 devname, ovs_strerror(errno));
1158 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1169 for (i = 0; i < n_open_rxqs; i++) {
1170 netdev_rxq_close(port->rxq[i]);
1177 netdev_close(netdev);
1182 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1184 OVS_REQUIRES(dp->port_mutex)
1186 struct dp_netdev_port *port;
1189 /* Reject devices already in 'dp'. */
1190 if (!get_port_by_name(dp, devname, &port)) {
1194 error = port_create(devname, dpif_netdev_port_open_type(dp->class, type),
1195 type, port_no, &port);
1200 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1202 if (netdev_is_pmd(port->netdev)) {
1203 dp_netdev_add_port_to_pmds(dp, port);
1205 seq_change(dp->port_seq);
1211 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1212 odp_port_t *port_nop)
1214 struct dp_netdev *dp = get_dp_netdev(dpif);
1215 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1216 const char *dpif_port;
1220 ovs_mutex_lock(&dp->port_mutex);
1221 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1222 if (*port_nop != ODPP_NONE) {
1223 port_no = *port_nop;
1224 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1226 port_no = choose_port(dp, dpif_port);
1227 error = port_no == ODPP_NONE ? EFBIG : 0;
1230 *port_nop = port_no;
1231 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1233 ovs_mutex_unlock(&dp->port_mutex);
1239 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1241 struct dp_netdev *dp = get_dp_netdev(dpif);
1244 ovs_mutex_lock(&dp->port_mutex);
1245 if (port_no == ODPP_LOCAL) {
1248 struct dp_netdev_port *port;
1250 error = get_port_by_number(dp, port_no, &port);
1252 do_del_port(dp, port);
1255 ovs_mutex_unlock(&dp->port_mutex);
1261 is_valid_port_number(odp_port_t port_no)
1263 return port_no != ODPP_NONE;
1266 static struct dp_netdev_port *
1267 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1269 struct dp_netdev_port *port;
1271 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1272 if (port->port_no == port_no) {
1280 get_port_by_number(struct dp_netdev *dp,
1281 odp_port_t port_no, struct dp_netdev_port **portp)
1283 if (!is_valid_port_number(port_no)) {
1287 *portp = dp_netdev_lookup_port(dp, port_no);
1288 return *portp ? 0 : ENOENT;
1293 port_destroy(struct dp_netdev_port *port)
1299 netdev_close(port->netdev);
1300 netdev_restore_flags(port->sf);
1302 for (unsigned i = 0; i < port->n_rxq; i++) {
1303 netdev_rxq_close(port->rxq[i]);
1312 get_port_by_name(struct dp_netdev *dp,
1313 const char *devname, struct dp_netdev_port **portp)
1314 OVS_REQUIRES(dp->port_mutex)
1316 struct dp_netdev_port *port;
1318 CMAP_FOR_EACH (port, node, &dp->ports) {
1319 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1328 get_n_pmd_threads(struct dp_netdev *dp)
1330 /* There is one non pmd thread in dp->poll_threads */
1331 return cmap_count(&dp->poll_threads) - 1;
1335 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1337 struct dp_netdev_pmd_thread *pmd;
1340 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1341 if (pmd->numa_id == numa_id) {
1349 /* Returns 'true' if there is a port with pmd netdev and the netdev
1350 * is on numa node 'numa_id'. */
1352 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1354 struct dp_netdev_port *port;
1356 CMAP_FOR_EACH (port, node, &dp->ports) {
1357 if (netdev_is_pmd(port->netdev)
1358 && netdev_get_numa_id(port->netdev) == numa_id) {
1368 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1369 OVS_REQUIRES(dp->port_mutex)
1371 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1372 seq_change(dp->port_seq);
1373 if (netdev_is_pmd(port->netdev)) {
1374 int numa_id = netdev_get_numa_id(port->netdev);
1376 /* PMD threads can not be on invalid numa node. */
1377 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
1378 /* If there is no netdev on the numa node, deletes the pmd threads
1379 * for that numa. Else, deletes the queues from polling lists. */
1380 if (!has_pmd_port_for_numa(dp, numa_id)) {
1381 dp_netdev_del_pmds_on_numa(dp, numa_id);
1383 dp_netdev_del_port_from_all_pmds(dp, port);
1391 answer_port_query(const struct dp_netdev_port *port,
1392 struct dpif_port *dpif_port)
1394 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1395 dpif_port->type = xstrdup(port->type);
1396 dpif_port->port_no = port->port_no;
1400 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1401 struct dpif_port *dpif_port)
1403 struct dp_netdev *dp = get_dp_netdev(dpif);
1404 struct dp_netdev_port *port;
1407 error = get_port_by_number(dp, port_no, &port);
1408 if (!error && dpif_port) {
1409 answer_port_query(port, dpif_port);
1416 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1417 struct dpif_port *dpif_port)
1419 struct dp_netdev *dp = get_dp_netdev(dpif);
1420 struct dp_netdev_port *port;
1423 ovs_mutex_lock(&dp->port_mutex);
1424 error = get_port_by_name(dp, devname, &port);
1425 if (!error && dpif_port) {
1426 answer_port_query(port, dpif_port);
1428 ovs_mutex_unlock(&dp->port_mutex);
1434 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1436 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1440 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1442 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1443 ovsrcu_postpone(dp_netdev_flow_free, flow);
1448 dp_netdev_flow_hash(const ovs_u128 *ufid)
1450 return ufid->u32[0];
1454 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1455 struct dp_netdev_flow *flow)
1456 OVS_REQUIRES(pmd->flow_mutex)
1458 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1460 dpcls_remove(&pmd->cls, &flow->cr);
1461 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1464 dp_netdev_flow_unref(flow);
1468 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1470 struct dp_netdev_flow *netdev_flow;
1472 ovs_mutex_lock(&pmd->flow_mutex);
1473 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1474 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1476 ovs_mutex_unlock(&pmd->flow_mutex);
1480 dpif_netdev_flow_flush(struct dpif *dpif)
1482 struct dp_netdev *dp = get_dp_netdev(dpif);
1483 struct dp_netdev_pmd_thread *pmd;
1485 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1486 dp_netdev_pmd_flow_flush(pmd);
1492 struct dp_netdev_port_state {
1493 struct cmap_position position;
1498 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1500 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1505 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1506 struct dpif_port *dpif_port)
1508 struct dp_netdev_port_state *state = state_;
1509 struct dp_netdev *dp = get_dp_netdev(dpif);
1510 struct cmap_node *node;
1513 node = cmap_next_position(&dp->ports, &state->position);
1515 struct dp_netdev_port *port;
1517 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1520 state->name = xstrdup(netdev_get_name(port->netdev));
1521 dpif_port->name = state->name;
1522 dpif_port->type = port->type;
1523 dpif_port->port_no = port->port_no;
1534 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1536 struct dp_netdev_port_state *state = state_;
1543 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1545 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1546 uint64_t new_port_seq;
1549 new_port_seq = seq_read(dpif->dp->port_seq);
1550 if (dpif->last_port_seq != new_port_seq) {
1551 dpif->last_port_seq = new_port_seq;
1561 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1563 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1565 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1568 static struct dp_netdev_flow *
1569 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1571 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1574 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1576 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1579 /* netdev_flow_key utilities.
1581 * netdev_flow_key is basically a miniflow. We use these functions
1582 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1583 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1585 * - Since we are dealing exclusively with miniflows created by
1586 * miniflow_extract(), if the map is different the miniflow is different.
1587 * Therefore we can be faster by comparing the map and the miniflow in a
1589 * - These functions can be inlined by the compiler. */
1591 /* Given the number of bits set in miniflow's maps, returns the size of the
1592 * 'netdev_flow_key.mf' */
1593 static inline size_t
1594 netdev_flow_key_size(size_t flow_u64s)
1596 return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s);
1600 netdev_flow_key_equal(const struct netdev_flow_key *a,
1601 const struct netdev_flow_key *b)
1603 /* 'b->len' may be not set yet. */
1604 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1607 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1608 * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been
1609 * generated by miniflow_extract. */
1611 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1612 const struct miniflow *mf)
1614 return !memcmp(&key->mf, mf, key->len);
1618 netdev_flow_key_clone(struct netdev_flow_key *dst,
1619 const struct netdev_flow_key *src)
1622 offsetof(struct netdev_flow_key, mf) + src->len);
1627 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1628 const struct flow *src)
1630 struct dp_packet packet;
1631 uint64_t buf_stub[512 / 8];
1633 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1634 pkt_metadata_from_flow(&packet.md, src);
1635 flow_compose(&packet, src);
1636 miniflow_extract(&packet, &dst->mf);
1637 dp_packet_uninit(&packet);
1639 dst->len = netdev_flow_key_size(miniflow_n_values(&dst->mf));
1640 dst->hash = 0; /* Not computed yet. */
1643 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1645 netdev_flow_mask_init(struct netdev_flow_key *mask,
1646 const struct match *match)
1648 uint64_t *dst = miniflow_values(&mask->mf);
1649 struct flowmap fmap;
1653 /* Only check masks that make sense for the flow. */
1654 flow_wc_map(&match->flow, &fmap);
1655 flowmap_init(&mask->mf.map);
1657 FLOWMAP_FOR_EACH_INDEX(idx, fmap) {
1658 uint64_t mask_u64 = flow_u64_value(&match->wc.masks, idx);
1661 flowmap_set(&mask->mf.map, idx, 1);
1663 hash = hash_add64(hash, mask_u64);
1669 FLOWMAP_FOR_EACH_MAP (map, mask->mf.map) {
1670 hash = hash_add64(hash, map);
1673 size_t n = dst - miniflow_get_values(&mask->mf);
1675 mask->hash = hash_finish(hash, n * 8);
1676 mask->len = netdev_flow_key_size(n);
1679 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1681 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1682 const struct flow *flow,
1683 const struct netdev_flow_key *mask)
1685 uint64_t *dst_u64 = miniflow_values(&dst->mf);
1686 const uint64_t *mask_u64 = miniflow_get_values(&mask->mf);
1690 dst->len = mask->len;
1691 dst->mf = mask->mf; /* Copy maps. */
1693 FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf.map) {
1694 *dst_u64 = value & *mask_u64++;
1695 hash = hash_add64(hash, *dst_u64++);
1697 dst->hash = hash_finish(hash,
1698 (dst_u64 - miniflow_get_values(&dst->mf)) * 8);
1701 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
1702 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
1703 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
1705 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1707 static inline uint32_t
1708 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1709 const struct netdev_flow_key *mask)
1711 const uint64_t *p = miniflow_get_values(&mask->mf);
1715 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, key, mask->mf.map) {
1716 hash = hash_add64(hash, value & *p++);
1719 return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8);
1723 emc_entry_alive(struct emc_entry *ce)
1725 return ce->flow && !ce->flow->dead;
1729 emc_clear_entry(struct emc_entry *ce)
1732 dp_netdev_flow_unref(ce->flow);
1738 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1739 const struct netdev_flow_key *key)
1741 if (ce->flow != flow) {
1743 dp_netdev_flow_unref(ce->flow);
1746 if (dp_netdev_flow_ref(flow)) {
1753 netdev_flow_key_clone(&ce->key, key);
1758 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1759 struct dp_netdev_flow *flow)
1761 struct emc_entry *to_be_replaced = NULL;
1762 struct emc_entry *current_entry;
1764 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1765 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1766 /* We found the entry with the 'mf' miniflow */
1767 emc_change_entry(current_entry, flow, NULL);
1771 /* Replacement policy: put the flow in an empty (not alive) entry, or
1772 * in the first entry where it can be */
1774 || (emc_entry_alive(to_be_replaced)
1775 && !emc_entry_alive(current_entry))
1776 || current_entry->key.hash < to_be_replaced->key.hash) {
1777 to_be_replaced = current_entry;
1780 /* We didn't find the miniflow in the cache.
1781 * The 'to_be_replaced' entry is where the new flow will be stored */
1783 emc_change_entry(to_be_replaced, flow, key);
1786 static inline struct dp_netdev_flow *
1787 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1789 struct emc_entry *current_entry;
1791 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1792 if (current_entry->key.hash == key->hash
1793 && emc_entry_alive(current_entry)
1794 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1796 /* We found the entry with the 'key->mf' miniflow */
1797 return current_entry->flow;
1804 static struct dp_netdev_flow *
1805 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1806 const struct netdev_flow_key *key)
1808 struct dp_netdev_flow *netdev_flow;
1809 struct dpcls_rule *rule;
1811 dpcls_lookup(&pmd->cls, key, &rule, 1);
1812 netdev_flow = dp_netdev_flow_cast(rule);
1817 static struct dp_netdev_flow *
1818 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1819 const ovs_u128 *ufidp, const struct nlattr *key,
1822 struct dp_netdev_flow *netdev_flow;
1826 /* If a UFID is not provided, determine one based on the key. */
1827 if (!ufidp && key && key_len
1828 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1829 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1834 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1836 if (ovs_u128_equals(netdev_flow->ufid, *ufidp)) {
1846 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1847 struct dpif_flow_stats *stats)
1849 struct dp_netdev_flow *netdev_flow;
1850 unsigned long long n;
1854 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1856 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1857 stats->n_packets = n;
1858 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1860 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1862 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1863 stats->tcp_flags = flags;
1866 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1867 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1868 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1871 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1872 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1873 struct dpif_flow *flow, bool terse)
1876 memset(flow, 0, sizeof *flow);
1878 struct flow_wildcards wc;
1879 struct dp_netdev_actions *actions;
1881 struct odp_flow_key_parms odp_parms = {
1882 .flow = &netdev_flow->flow,
1884 .support = dp_netdev_support,
1887 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1890 offset = key_buf->size;
1891 flow->key = ofpbuf_tail(key_buf);
1892 odp_parms.odp_in_port = netdev_flow->flow.in_port.odp_port;
1893 odp_flow_key_from_flow(&odp_parms, key_buf);
1894 flow->key_len = key_buf->size - offset;
1897 offset = mask_buf->size;
1898 flow->mask = ofpbuf_tail(mask_buf);
1899 odp_parms.odp_in_port = wc.masks.in_port.odp_port;
1900 odp_parms.key_buf = key_buf;
1901 odp_flow_key_from_mask(&odp_parms, mask_buf);
1902 flow->mask_len = mask_buf->size - offset;
1905 actions = dp_netdev_flow_get_actions(netdev_flow);
1906 flow->actions = actions->actions;
1907 flow->actions_len = actions->size;
1910 flow->ufid = netdev_flow->ufid;
1911 flow->ufid_present = true;
1912 flow->pmd_id = netdev_flow->pmd_id;
1913 get_dpif_flow_stats(netdev_flow, &flow->stats);
1917 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1918 const struct nlattr *mask_key,
1919 uint32_t mask_key_len, const struct flow *flow,
1920 struct flow_wildcards *wc)
1922 enum odp_key_fitness fitness;
1924 fitness = odp_flow_key_to_mask_udpif(mask_key, mask_key_len, key,
1927 /* This should not happen: it indicates that
1928 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1929 * disagree on the acceptable form of a mask. Log the problem
1930 * as an error, with enough details to enable debugging. */
1931 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1933 if (!VLOG_DROP_ERR(&rl)) {
1937 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1939 VLOG_ERR("internal error parsing flow mask %s (%s)",
1940 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1951 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1956 if (odp_flow_key_to_flow_udpif(key, key_len, flow)) {
1957 /* This should not happen: it indicates that odp_flow_key_from_flow()
1958 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1959 * flow. Log the problem as an error, with enough details to enable
1961 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1963 if (!VLOG_DROP_ERR(&rl)) {
1967 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1968 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1975 in_port = flow->in_port.odp_port;
1976 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1980 /* Userspace datapath doesn't support conntrack. */
1981 if (flow->ct_state || flow->ct_zone || flow->ct_mark
1982 || !ovs_u128_is_zero(flow->ct_label)) {
1990 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1992 struct dp_netdev *dp = get_dp_netdev(dpif);
1993 struct dp_netdev_flow *netdev_flow;
1994 struct dp_netdev_pmd_thread *pmd;
1995 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1996 ? NON_PMD_CORE_ID : get->pmd_id;
1999 pmd = dp_netdev_get_pmd(dp, pmd_id);
2004 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
2007 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
2012 dp_netdev_pmd_unref(pmd);
2018 static struct dp_netdev_flow *
2019 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
2020 struct match *match, const ovs_u128 *ufid,
2021 const struct nlattr *actions, size_t actions_len)
2022 OVS_REQUIRES(pmd->flow_mutex)
2024 struct dp_netdev_flow *flow;
2025 struct netdev_flow_key mask;
2027 netdev_flow_mask_init(&mask, match);
2028 /* Make sure wc does not have metadata. */
2029 ovs_assert(!FLOWMAP_HAS_FIELD(&mask.mf.map, metadata)
2030 && !FLOWMAP_HAS_FIELD(&mask.mf.map, regs));
2032 /* Do not allocate extra space. */
2033 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
2034 memset(&flow->stats, 0, sizeof flow->stats);
2037 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
2038 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
2039 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
2040 ovs_refcount_init(&flow->ref_cnt);
2041 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
2043 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
2044 dpcls_insert(&pmd->cls, &flow->cr, &mask);
2046 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
2047 dp_netdev_flow_hash(&flow->ufid));
2049 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2051 struct ds ds = DS_EMPTY_INITIALIZER;
2053 match.tun_md.valid = false;
2054 match.flow = flow->flow;
2055 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2057 ds_put_cstr(&ds, "flow_add: ");
2058 odp_format_ufid(ufid, &ds);
2059 ds_put_cstr(&ds, " ");
2060 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2061 ds_put_cstr(&ds, ", actions:");
2062 format_odp_actions(&ds, actions, actions_len);
2064 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2073 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2075 struct dp_netdev *dp = get_dp_netdev(dpif);
2076 struct dp_netdev_flow *netdev_flow;
2077 struct netdev_flow_key key;
2078 struct dp_netdev_pmd_thread *pmd;
2081 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2082 ? NON_PMD_CORE_ID : put->pmd_id;
2085 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2089 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2090 put->mask, put->mask_len,
2091 &match.flow, &match.wc);
2096 pmd = dp_netdev_get_pmd(dp, pmd_id);
2101 /* Must produce a netdev_flow_key for lookup.
2102 * This interface is no longer performance critical, since it is not used
2103 * for upcall processing any more. */
2104 netdev_flow_key_from_flow(&key, &match.flow);
2109 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2112 ovs_mutex_lock(&pmd->flow_mutex);
2113 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2115 if (put->flags & DPIF_FP_CREATE) {
2116 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2118 memset(put->stats, 0, sizeof *put->stats);
2120 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2130 if (put->flags & DPIF_FP_MODIFY
2131 && flow_equal(&match.flow, &netdev_flow->flow)) {
2132 struct dp_netdev_actions *new_actions;
2133 struct dp_netdev_actions *old_actions;
2135 new_actions = dp_netdev_actions_create(put->actions,
2138 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2139 ovsrcu_set(&netdev_flow->actions, new_actions);
2142 get_dpif_flow_stats(netdev_flow, put->stats);
2144 if (put->flags & DPIF_FP_ZERO_STATS) {
2145 /* XXX: The userspace datapath uses thread local statistics
2146 * (for flows), which should be updated only by the owning
2147 * thread. Since we cannot write on stats memory here,
2148 * we choose not to support this flag. Please note:
2149 * - This feature is currently used only by dpctl commands with
2151 * - Should the need arise, this operation can be implemented
2152 * by keeping a base value (to be update here) for each
2153 * counter, and subtracting it before outputting the stats */
2157 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2158 } else if (put->flags & DPIF_FP_CREATE) {
2161 /* Overlapping flow. */
2165 ovs_mutex_unlock(&pmd->flow_mutex);
2166 dp_netdev_pmd_unref(pmd);
2172 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2174 struct dp_netdev *dp = get_dp_netdev(dpif);
2175 struct dp_netdev_flow *netdev_flow;
2176 struct dp_netdev_pmd_thread *pmd;
2177 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2178 ? NON_PMD_CORE_ID : del->pmd_id;
2181 pmd = dp_netdev_get_pmd(dp, pmd_id);
2186 ovs_mutex_lock(&pmd->flow_mutex);
2187 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2191 get_dpif_flow_stats(netdev_flow, del->stats);
2193 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2197 ovs_mutex_unlock(&pmd->flow_mutex);
2198 dp_netdev_pmd_unref(pmd);
2203 struct dpif_netdev_flow_dump {
2204 struct dpif_flow_dump up;
2205 struct cmap_position poll_thread_pos;
2206 struct cmap_position flow_pos;
2207 struct dp_netdev_pmd_thread *cur_pmd;
2209 struct ovs_mutex mutex;
2212 static struct dpif_netdev_flow_dump *
2213 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2215 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2218 static struct dpif_flow_dump *
2219 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2221 struct dpif_netdev_flow_dump *dump;
2223 dump = xzalloc(sizeof *dump);
2224 dpif_flow_dump_init(&dump->up, dpif_);
2225 dump->up.terse = terse;
2226 ovs_mutex_init(&dump->mutex);
2232 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2234 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2236 ovs_mutex_destroy(&dump->mutex);
2241 struct dpif_netdev_flow_dump_thread {
2242 struct dpif_flow_dump_thread up;
2243 struct dpif_netdev_flow_dump *dump;
2244 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2245 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2248 static struct dpif_netdev_flow_dump_thread *
2249 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2251 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2254 static struct dpif_flow_dump_thread *
2255 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2257 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2258 struct dpif_netdev_flow_dump_thread *thread;
2260 thread = xmalloc(sizeof *thread);
2261 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2262 thread->dump = dump;
2267 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2269 struct dpif_netdev_flow_dump_thread *thread
2270 = dpif_netdev_flow_dump_thread_cast(thread_);
2276 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2277 struct dpif_flow *flows, int max_flows)
2279 struct dpif_netdev_flow_dump_thread *thread
2280 = dpif_netdev_flow_dump_thread_cast(thread_);
2281 struct dpif_netdev_flow_dump *dump = thread->dump;
2282 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2286 ovs_mutex_lock(&dump->mutex);
2287 if (!dump->status) {
2288 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2289 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2290 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2291 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2293 /* First call to dump_next(), extracts the first pmd thread.
2294 * If there is no pmd thread, returns immediately. */
2296 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2298 ovs_mutex_unlock(&dump->mutex);
2305 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2306 struct cmap_node *node;
2308 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2312 netdev_flows[n_flows] = CONTAINER_OF(node,
2313 struct dp_netdev_flow,
2316 /* When finishing dumping the current pmd thread, moves to
2318 if (n_flows < flow_limit) {
2319 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2320 dp_netdev_pmd_unref(pmd);
2321 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2327 /* Keeps the reference to next caller. */
2328 dump->cur_pmd = pmd;
2330 /* If the current dump is empty, do not exit the loop, since the
2331 * remaining pmds could have flows to be dumped. Just dumps again
2332 * on the new 'pmd'. */
2335 ovs_mutex_unlock(&dump->mutex);
2337 for (i = 0; i < n_flows; i++) {
2338 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2339 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2340 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2341 struct dpif_flow *f = &flows[i];
2342 struct ofpbuf key, mask;
2344 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2345 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2346 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2354 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2355 OVS_NO_THREAD_SAFETY_ANALYSIS
2357 struct dp_netdev *dp = get_dp_netdev(dpif);
2358 struct dp_netdev_pmd_thread *pmd;
2359 struct dp_packet_batch pp;
2361 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2362 dp_packet_size(execute->packet) > UINT16_MAX) {
2366 /* Tries finding the 'pmd'. If NULL is returned, that means
2367 * the current thread is a non-pmd thread and should use
2368 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2369 pmd = ovsthread_getspecific(dp->per_pmd_key);
2371 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2374 /* If the current thread is non-pmd thread, acquires
2375 * the 'non_pmd_mutex'. */
2376 if (pmd->core_id == NON_PMD_CORE_ID) {
2377 ovs_mutex_lock(&dp->non_pmd_mutex);
2378 ovs_mutex_lock(&dp->port_mutex);
2381 /* The action processing expects the RSS hash to be valid, because
2382 * it's always initialized at the beginning of datapath processing.
2383 * In this case, though, 'execute->packet' may not have gone through
2384 * the datapath at all, it may have been generated by the upper layer
2385 * (OpenFlow packet-out, BFD frame, ...). */
2386 if (!dp_packet_rss_valid(execute->packet)) {
2387 dp_packet_set_rss_hash(execute->packet,
2388 flow_hash_5tuple(execute->flow, 0));
2391 packet_batch_init_packet(&pp, execute->packet);
2392 dp_netdev_execute_actions(pmd, &pp, false, execute->actions,
2393 execute->actions_len);
2395 if (pmd->core_id == NON_PMD_CORE_ID) {
2396 dp_netdev_pmd_unref(pmd);
2397 ovs_mutex_unlock(&dp->port_mutex);
2398 ovs_mutex_unlock(&dp->non_pmd_mutex);
2405 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2409 for (i = 0; i < n_ops; i++) {
2410 struct dpif_op *op = ops[i];
2413 case DPIF_OP_FLOW_PUT:
2414 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2417 case DPIF_OP_FLOW_DEL:
2418 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2421 case DPIF_OP_EXECUTE:
2422 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2425 case DPIF_OP_FLOW_GET:
2426 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2432 /* Returns true if the configuration for rx queues or cpu mask
2435 pmd_config_changed(const struct dp_netdev *dp, const char *cmask)
2437 struct dp_netdev_port *port;
2439 CMAP_FOR_EACH (port, node, &dp->ports) {
2440 struct netdev *netdev = port->netdev;
2441 int requested_n_rxq = netdev_requested_n_rxq(netdev);
2442 if (netdev_is_pmd(netdev)
2443 && port->latest_requested_n_rxq != requested_n_rxq) {
2448 if (dp->pmd_cmask != NULL && cmask != NULL) {
2449 return strcmp(dp->pmd_cmask, cmask);
2451 return (dp->pmd_cmask != NULL || cmask != NULL);
2455 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2457 dpif_netdev_pmd_set(struct dpif *dpif, const char *cmask)
2459 struct dp_netdev *dp = get_dp_netdev(dpif);
2461 if (pmd_config_changed(dp, cmask)) {
2462 struct dp_netdev_port *port;
2464 dp_netdev_destroy_all_pmds(dp);
2466 CMAP_FOR_EACH (port, node, &dp->ports) {
2467 struct netdev *netdev = port->netdev;
2468 int requested_n_rxq = netdev_requested_n_rxq(netdev);
2469 if (netdev_is_pmd(port->netdev)
2470 && port->latest_requested_n_rxq != requested_n_rxq) {
2473 /* Closes the existing 'rxq's. */
2474 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2475 netdev_rxq_close(port->rxq[i]);
2476 port->rxq[i] = NULL;
2480 /* Sets the new rx queue config. */
2481 err = netdev_set_multiq(port->netdev,
2482 ovs_numa_get_n_cores() + 1,
2484 if (err && (err != EOPNOTSUPP)) {
2485 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2486 " %u", netdev_get_name(port->netdev),
2490 port->latest_requested_n_rxq = requested_n_rxq;
2491 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2492 port->n_rxq = netdev_n_rxq(port->netdev);
2493 port->rxq = xrealloc(port->rxq, sizeof *port->rxq * port->n_rxq);
2494 for (i = 0; i < port->n_rxq; i++) {
2495 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2499 /* Reconfigures the cpu mask. */
2500 ovs_numa_set_cpu_mask(cmask);
2501 free(dp->pmd_cmask);
2502 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2504 /* Restores the non-pmd. */
2505 dp_netdev_set_nonpmd(dp);
2506 /* Restores all pmd threads. */
2507 dp_netdev_reset_pmd_threads(dp);
2514 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2515 uint32_t queue_id, uint32_t *priority)
2517 *priority = queue_id;
2522 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2523 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2524 struct dp_netdev_actions *
2525 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2527 struct dp_netdev_actions *netdev_actions;
2529 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2530 memcpy(netdev_actions->actions, actions, size);
2531 netdev_actions->size = size;
2533 return netdev_actions;
2536 struct dp_netdev_actions *
2537 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2539 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2543 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2548 static inline unsigned long long
2549 cycles_counter(void)
2552 return rte_get_tsc_cycles();
2558 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2559 extern struct ovs_mutex cycles_counter_fake_mutex;
2561 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2563 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2564 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2565 OVS_NO_THREAD_SAFETY_ANALYSIS
2567 pmd->last_cycles = cycles_counter();
2570 /* Stop counting cycles and add them to the counter 'type' */
2572 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2573 enum pmd_cycles_counter_type type)
2574 OVS_RELEASES(&cycles_counter_fake_mutex)
2575 OVS_NO_THREAD_SAFETY_ANALYSIS
2577 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2579 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2583 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2584 struct dp_netdev_port *port,
2585 struct netdev_rxq *rxq)
2587 struct dp_packet_batch batch;
2590 dp_packet_batch_init(&batch);
2591 cycles_count_start(pmd);
2592 error = netdev_rxq_recv(rxq, &batch);
2593 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2595 *recirc_depth_get() = 0;
2597 cycles_count_start(pmd);
2598 dp_netdev_input(pmd, &batch, port->port_no);
2599 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2600 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2601 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2603 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2604 netdev_get_name(port->netdev), ovs_strerror(error));
2608 /* Return true if needs to revalidate datapath flows. */
2610 dpif_netdev_run(struct dpif *dpif)
2612 struct dp_netdev_port *port;
2613 struct dp_netdev *dp = get_dp_netdev(dpif);
2614 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2616 uint64_t new_tnl_seq;
2618 ovs_mutex_lock(&dp->non_pmd_mutex);
2619 CMAP_FOR_EACH (port, node, &dp->ports) {
2620 if (!netdev_is_pmd(port->netdev)) {
2623 for (i = 0; i < port->n_rxq; i++) {
2624 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2628 ovs_mutex_unlock(&dp->non_pmd_mutex);
2629 dp_netdev_pmd_unref(non_pmd);
2631 tnl_neigh_cache_run();
2633 new_tnl_seq = seq_read(tnl_conf_seq);
2635 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2636 dp->last_tnl_conf_seq = new_tnl_seq;
2643 dpif_netdev_wait(struct dpif *dpif)
2645 struct dp_netdev_port *port;
2646 struct dp_netdev *dp = get_dp_netdev(dpif);
2648 ovs_mutex_lock(&dp_netdev_mutex);
2649 CMAP_FOR_EACH (port, node, &dp->ports) {
2650 if (!netdev_is_pmd(port->netdev)) {
2653 for (i = 0; i < port->n_rxq; i++) {
2654 netdev_rxq_wait(port->rxq[i]);
2658 ovs_mutex_unlock(&dp_netdev_mutex);
2659 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2663 pmd_load_queues(struct dp_netdev_pmd_thread *pmd, struct rxq_poll **ppoll_list)
2665 struct rxq_poll *poll_list = *ppoll_list;
2666 struct rxq_poll *poll;
2669 ovs_mutex_lock(&pmd->poll_mutex);
2670 poll_list = xrealloc(poll_list, pmd->poll_cnt * sizeof *poll_list);
2673 LIST_FOR_EACH (poll, node, &pmd->poll_list) {
2674 poll_list[i++] = *poll;
2676 ovs_mutex_unlock(&pmd->poll_mutex);
2678 *ppoll_list = poll_list;
2683 pmd_thread_main(void *f_)
2685 struct dp_netdev_pmd_thread *pmd = f_;
2686 unsigned int lc = 0;
2687 struct rxq_poll *poll_list;
2688 unsigned int port_seq = PMD_INITIAL_SEQ;
2696 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2697 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2698 pmd_thread_setaffinity_cpu(pmd->core_id);
2699 poll_cnt = pmd_load_queues(pmd, &poll_list);
2701 emc_cache_init(&pmd->flow_cache);
2703 /* List port/core affinity */
2704 for (i = 0; i < poll_cnt; i++) {
2705 VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n",
2706 pmd->core_id, netdev_get_name(poll_list[i].port->netdev),
2707 netdev_rxq_get_queue_id(poll_list[i].rx));
2711 for (i = 0; i < poll_cnt; i++) {
2712 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2720 emc_cache_slow_sweep(&pmd->flow_cache);
2721 coverage_try_clear();
2724 atomic_read_relaxed(&pmd->change_seq, &seq);
2725 if (seq != port_seq) {
2732 poll_cnt = pmd_load_queues(pmd, &poll_list);
2733 exiting = latch_is_set(&pmd->exit_latch);
2734 /* Signal here to make sure the pmd finishes
2735 * reloading the updated configuration. */
2736 dp_netdev_pmd_reload_done(pmd);
2738 emc_cache_uninit(&pmd->flow_cache);
2749 dp_netdev_disable_upcall(struct dp_netdev *dp)
2750 OVS_ACQUIRES(dp->upcall_rwlock)
2752 fat_rwlock_wrlock(&dp->upcall_rwlock);
2756 dpif_netdev_disable_upcall(struct dpif *dpif)
2757 OVS_NO_THREAD_SAFETY_ANALYSIS
2759 struct dp_netdev *dp = get_dp_netdev(dpif);
2760 dp_netdev_disable_upcall(dp);
2764 dp_netdev_enable_upcall(struct dp_netdev *dp)
2765 OVS_RELEASES(dp->upcall_rwlock)
2767 fat_rwlock_unlock(&dp->upcall_rwlock);
2771 dpif_netdev_enable_upcall(struct dpif *dpif)
2772 OVS_NO_THREAD_SAFETY_ANALYSIS
2774 struct dp_netdev *dp = get_dp_netdev(dpif);
2775 dp_netdev_enable_upcall(dp);
2779 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2781 ovs_mutex_lock(&pmd->cond_mutex);
2782 xpthread_cond_signal(&pmd->cond);
2783 ovs_mutex_unlock(&pmd->cond_mutex);
2786 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2787 * the pointer if succeeds, otherwise, NULL.
2789 * Caller must unrefs the returned reference. */
2790 static struct dp_netdev_pmd_thread *
2791 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2793 struct dp_netdev_pmd_thread *pmd;
2794 const struct cmap_node *pnode;
2796 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2800 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2802 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2805 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2807 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2809 struct dp_netdev_pmd_thread *non_pmd;
2811 non_pmd = xzalloc(sizeof *non_pmd);
2812 dp_netdev_configure_pmd(non_pmd, dp, NON_PMD_CORE_ID, OVS_NUMA_UNSPEC);
2815 /* Caller must have valid pointer to 'pmd'. */
2817 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2819 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2823 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2825 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2826 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2830 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2831 * fails, keeps checking for next node until reaching the end of cmap.
2833 * Caller must unrefs the returned reference. */
2834 static struct dp_netdev_pmd_thread *
2835 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2837 struct dp_netdev_pmd_thread *next;
2840 struct cmap_node *node;
2842 node = cmap_next_position(&dp->poll_threads, pos);
2843 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2845 } while (next && !dp_netdev_pmd_try_ref(next));
2850 /* Configures the 'pmd' based on the input argument. */
2852 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2853 unsigned core_id, int numa_id)
2856 pmd->core_id = core_id;
2857 pmd->numa_id = numa_id;
2860 atomic_init(&pmd->tx_qid,
2861 (core_id == NON_PMD_CORE_ID)
2862 ? ovs_numa_get_n_cores()
2863 : get_n_pmd_threads(dp));
2865 ovs_refcount_init(&pmd->ref_cnt);
2866 latch_init(&pmd->exit_latch);
2867 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2868 xpthread_cond_init(&pmd->cond, NULL);
2869 ovs_mutex_init(&pmd->cond_mutex);
2870 ovs_mutex_init(&pmd->flow_mutex);
2871 ovs_mutex_init(&pmd->poll_mutex);
2872 dpcls_init(&pmd->cls);
2873 cmap_init(&pmd->flow_table);
2874 ovs_list_init(&pmd->poll_list);
2875 /* init the 'flow_cache' since there is no
2876 * actual thread created for NON_PMD_CORE_ID. */
2877 if (core_id == NON_PMD_CORE_ID) {
2878 emc_cache_init(&pmd->flow_cache);
2880 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2881 hash_int(core_id, 0));
2885 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2887 dp_netdev_pmd_flow_flush(pmd);
2888 dpcls_destroy(&pmd->cls);
2889 cmap_destroy(&pmd->flow_table);
2890 ovs_mutex_destroy(&pmd->flow_mutex);
2891 latch_destroy(&pmd->exit_latch);
2892 xpthread_cond_destroy(&pmd->cond);
2893 ovs_mutex_destroy(&pmd->cond_mutex);
2894 ovs_mutex_destroy(&pmd->poll_mutex);
2898 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2899 * and unrefs the struct. */
2901 dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd)
2903 /* Uninit the 'flow_cache' since there is
2904 * no actual thread uninit it for NON_PMD_CORE_ID. */
2905 if (pmd->core_id == NON_PMD_CORE_ID) {
2906 emc_cache_uninit(&pmd->flow_cache);
2908 latch_set(&pmd->exit_latch);
2909 dp_netdev_reload_pmd__(pmd);
2910 ovs_numa_unpin_core(pmd->core_id);
2911 xpthread_join(pmd->thread, NULL);
2914 /* Unref all ports and free poll_list. */
2915 dp_netdev_pmd_clear_poll_list(pmd);
2917 /* Purges the 'pmd''s flows after stopping the thread, but before
2918 * destroying the flows, so that the flow stats can be collected. */
2919 if (dp->dp_purge_cb) {
2920 dp->dp_purge_cb(dp->dp_purge_aux, pmd->core_id);
2922 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2923 dp_netdev_pmd_unref(pmd);
2926 /* Destroys all pmd threads. */
2928 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2930 struct dp_netdev_pmd_thread *pmd;
2931 struct dp_netdev_pmd_thread **pmd_list;
2932 size_t k = 0, n_pmds;
2934 n_pmds = cmap_count(&dp->poll_threads);
2935 pmd_list = xcalloc(n_pmds, sizeof *pmd_list);
2937 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2938 /* We cannot call dp_netdev_del_pmd(), since it alters
2939 * 'dp->poll_threads' (while we're iterating it) and it
2941 ovs_assert(k < n_pmds);
2942 pmd_list[k++] = pmd;
2945 for (size_t i = 0; i < k; i++) {
2946 dp_netdev_del_pmd(dp, pmd_list[i]);
2951 /* Deletes all pmd threads on numa node 'numa_id' and
2952 * fixes tx_qids of other threads to keep them sequential. */
2954 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2956 struct dp_netdev_pmd_thread *pmd;
2957 int n_pmds_on_numa, n_pmds;
2958 int *free_idx, k = 0;
2959 struct dp_netdev_pmd_thread **pmd_list;
2961 n_pmds_on_numa = get_n_pmd_threads_on_numa(dp, numa_id);
2962 free_idx = xcalloc(n_pmds_on_numa, sizeof *free_idx);
2963 pmd_list = xcalloc(n_pmds_on_numa, sizeof *pmd_list);
2965 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2966 /* We cannot call dp_netdev_del_pmd(), since it alters
2967 * 'dp->poll_threads' (while we're iterating it) and it
2969 if (pmd->numa_id == numa_id) {
2970 atomic_read_relaxed(&pmd->tx_qid, &free_idx[k]);
2972 ovs_assert(k < n_pmds_on_numa);
2977 for (int i = 0; i < k; i++) {
2978 dp_netdev_del_pmd(dp, pmd_list[i]);
2981 n_pmds = get_n_pmd_threads(dp);
2982 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2985 atomic_read_relaxed(&pmd->tx_qid, &old_tx_qid);
2987 if (old_tx_qid >= n_pmds) {
2988 int new_tx_qid = free_idx[--k];
2990 atomic_store_relaxed(&pmd->tx_qid, new_tx_qid);
2998 /* Deletes all rx queues from pmd->poll_list. */
3000 dp_netdev_pmd_clear_poll_list(struct dp_netdev_pmd_thread *pmd)
3002 struct rxq_poll *poll;
3004 ovs_mutex_lock(&pmd->poll_mutex);
3005 LIST_FOR_EACH_POP (poll, node, &pmd->poll_list) {
3009 ovs_mutex_unlock(&pmd->poll_mutex);
3012 /* Deletes all rx queues of 'port' from poll_list of pmd thread. Returns true
3013 * if 'port' was found in 'pmd' (therefore a restart is required). */
3015 dp_netdev_del_port_from_pmd__(struct dp_netdev_port *port,
3016 struct dp_netdev_pmd_thread *pmd)
3018 struct rxq_poll *poll, *next;
3021 ovs_mutex_lock(&pmd->poll_mutex);
3022 LIST_FOR_EACH_SAFE (poll, next, node, &pmd->poll_list) {
3023 if (poll->port == port) {
3025 ovs_list_remove(&poll->node);
3030 ovs_mutex_unlock(&pmd->poll_mutex);
3035 /* Deletes all rx queues of 'port' from all pmd threads. The pmd threads that
3036 * need to be restarted are inserted in 'to_reload'. */
3038 dp_netdev_del_port_from_all_pmds__(struct dp_netdev *dp,
3039 struct dp_netdev_port *port,
3040 struct hmapx *to_reload)
3042 int numa_id = netdev_get_numa_id(port->netdev);
3043 struct dp_netdev_pmd_thread *pmd;
3045 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3046 if (pmd->numa_id == numa_id) {
3049 found = dp_netdev_del_port_from_pmd__(port, pmd);
3052 hmapx_add(to_reload, pmd);
3058 /* Deletes all rx queues of 'port' from all pmd threads of dp and
3059 * reloads them if needed. */
3061 dp_netdev_del_port_from_all_pmds(struct dp_netdev *dp,
3062 struct dp_netdev_port *port)
3064 struct dp_netdev_pmd_thread *pmd;
3065 struct hmapx to_reload = HMAPX_INITIALIZER(&to_reload);
3066 struct hmapx_node *node;
3068 dp_netdev_del_port_from_all_pmds__(dp, port, &to_reload);
3070 HMAPX_FOR_EACH (node, &to_reload) {
3071 pmd = (struct dp_netdev_pmd_thread *) node->data;
3072 dp_netdev_reload_pmd__(pmd);
3075 hmapx_destroy(&to_reload);
3079 /* Returns PMD thread from this numa node with fewer rx queues to poll.
3080 * Returns NULL if there is no PMD threads on this numa node.
3081 * Can be called safely only by main thread. */
3082 static struct dp_netdev_pmd_thread *
3083 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id)
3086 struct dp_netdev_pmd_thread *pmd, *res = NULL;
3088 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3089 if (pmd->numa_id == numa_id
3090 && (min_cnt > pmd->poll_cnt || res == NULL)) {
3091 min_cnt = pmd->poll_cnt;
3099 /* Adds rx queue to poll_list of PMD thread. */
3101 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
3102 struct dp_netdev_port *port, struct netdev_rxq *rx)
3103 OVS_REQUIRES(pmd->poll_mutex)
3105 struct rxq_poll *poll = xmalloc(sizeof *poll);
3110 ovs_list_push_back(&pmd->poll_list, &poll->node);
3114 /* Distributes all rx queues of 'port' between all PMD threads in 'dp'. The
3115 * pmd threads that need to be restarted are inserted in 'to_reload'. */
3117 dp_netdev_add_port_to_pmds__(struct dp_netdev *dp, struct dp_netdev_port *port,
3118 struct hmapx *to_reload)
3120 int numa_id = netdev_get_numa_id(port->netdev);
3121 struct dp_netdev_pmd_thread *pmd;
3124 /* Cannot create pmd threads for invalid numa node. */
3125 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
3127 for (i = 0; i < port->n_rxq; i++) {
3128 pmd = dp_netdev_less_loaded_pmd_on_numa(dp, numa_id);
3130 /* There is no pmd threads on this numa node. */
3131 dp_netdev_set_pmds_on_numa(dp, numa_id);
3132 /* Assigning of rx queues done. */
3136 ovs_mutex_lock(&pmd->poll_mutex);
3137 dp_netdev_add_rxq_to_pmd(pmd, port, port->rxq[i]);
3138 ovs_mutex_unlock(&pmd->poll_mutex);
3140 hmapx_add(to_reload, pmd);
3144 /* Distributes all rx queues of 'port' between all PMD threads in 'dp' and
3145 * reloads them, if needed. */
3147 dp_netdev_add_port_to_pmds(struct dp_netdev *dp, struct dp_netdev_port *port)
3149 struct dp_netdev_pmd_thread *pmd;
3150 struct hmapx to_reload = HMAPX_INITIALIZER(&to_reload);
3151 struct hmapx_node *node;
3153 dp_netdev_add_port_to_pmds__(dp, port, &to_reload);
3155 HMAPX_FOR_EACH (node, &to_reload) {
3156 pmd = (struct dp_netdev_pmd_thread *) node->data;
3157 dp_netdev_reload_pmd__(pmd);
3160 hmapx_destroy(&to_reload);
3163 /* Checks the numa node id of 'netdev' and starts pmd threads for
3166 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
3170 if (!ovs_numa_numa_id_is_valid(numa_id)) {
3171 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
3172 "invalid", numa_id);
3176 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
3178 /* If there are already pmd threads created for the numa node
3179 * in which 'netdev' is on, do nothing. Else, creates the
3180 * pmd threads for the numa node. */
3182 int can_have, n_unpinned, i, index = 0;
3183 struct dp_netdev_pmd_thread **pmds;
3184 struct dp_netdev_port *port;
3186 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
3188 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
3189 "cores on numa node %d", numa_id);
3193 /* If cpu mask is specified, uses all unpinned cores, otherwise
3194 * tries creating NR_PMD_THREADS pmd threads. */
3195 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
3196 pmds = xzalloc(can_have * sizeof *pmds);
3197 for (i = 0; i < can_have; i++) {
3198 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
3199 pmds[i] = xzalloc(sizeof **pmds);
3200 dp_netdev_configure_pmd(pmds[i], dp, core_id, numa_id);
3203 /* Distributes rx queues of this numa node between new pmd threads. */
3204 CMAP_FOR_EACH (port, node, &dp->ports) {
3205 if (netdev_is_pmd(port->netdev)
3206 && netdev_get_numa_id(port->netdev) == numa_id) {
3207 for (i = 0; i < port->n_rxq; i++) {
3208 /* Make thread-safety analyser happy. */
3209 ovs_mutex_lock(&pmds[index]->poll_mutex);
3210 dp_netdev_add_rxq_to_pmd(pmds[index], port, port->rxq[i]);
3211 ovs_mutex_unlock(&pmds[index]->poll_mutex);
3212 index = (index + 1) % can_have;
3217 /* Actual start of pmd threads. */
3218 for (i = 0; i < can_have; i++) {
3219 pmds[i]->thread = ovs_thread_create("pmd", pmd_thread_main, pmds[i]);
3222 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
3227 /* Called after pmd threads config change. Restarts pmd threads with
3228 * new configuration. */
3230 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
3232 struct dp_netdev_port *port;
3234 CMAP_FOR_EACH (port, node, &dp->ports) {
3235 if (netdev_is_pmd(port->netdev)) {
3236 int numa_id = netdev_get_numa_id(port->netdev);
3238 dp_netdev_set_pmds_on_numa(dp, numa_id);
3244 dpif_netdev_get_datapath_version(void)
3246 return xstrdup("<built-in>");
3250 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
3251 uint16_t tcp_flags, long long now)
3255 atomic_store_relaxed(&netdev_flow->stats.used, now);
3256 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
3257 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
3258 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
3260 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
3264 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
3265 enum dp_stat_type type, int cnt)
3267 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3271 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3272 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3273 enum dpif_upcall_type type, const struct nlattr *userdata,
3274 struct ofpbuf *actions, struct ofpbuf *put_actions)
3276 struct dp_netdev *dp = pmd->dp;
3277 struct flow_tnl orig_tunnel;
3280 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3284 /* Upcall processing expects the Geneve options to be in the translated
3285 * format but we need to retain the raw format for datapath use. */
3286 orig_tunnel.flags = flow->tunnel.flags;
3287 if (flow->tunnel.flags & FLOW_TNL_F_UDPIF) {
3288 orig_tunnel.metadata.present.len = flow->tunnel.metadata.present.len;
3289 memcpy(orig_tunnel.metadata.opts.gnv, flow->tunnel.metadata.opts.gnv,
3290 flow->tunnel.metadata.present.len);
3291 err = tun_metadata_from_geneve_udpif(&orig_tunnel, &orig_tunnel,
3298 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3299 struct ds ds = DS_EMPTY_INITIALIZER;
3302 struct odp_flow_key_parms odp_parms = {
3305 .odp_in_port = flow->in_port.odp_port,
3306 .support = dp_netdev_support,
3309 ofpbuf_init(&key, 0);
3310 odp_flow_key_from_flow(&odp_parms, &key);
3311 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3312 dp_packet_size(packet_));
3314 odp_flow_key_format(key.data, key.size, &ds);
3316 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3317 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3319 ofpbuf_uninit(&key);
3325 err = dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3326 actions, wc, put_actions, dp->upcall_aux);
3327 if (err && err != ENOSPC) {
3331 /* Translate tunnel metadata masks to datapath format. */
3333 if (wc->masks.tunnel.metadata.present.map) {
3334 struct geneve_opt opts[TLV_TOT_OPT_SIZE /
3335 sizeof(struct geneve_opt)];
3337 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3338 tun_metadata_to_geneve_udpif_mask(&flow->tunnel,
3340 orig_tunnel.metadata.opts.gnv,
3341 orig_tunnel.metadata.present.len,
3344 orig_tunnel.metadata.present.len = 0;
3347 memset(&wc->masks.tunnel.metadata, 0,
3348 sizeof wc->masks.tunnel.metadata);
3349 memcpy(&wc->masks.tunnel.metadata.opts.gnv, opts,
3350 orig_tunnel.metadata.present.len);
3352 wc->masks.tunnel.metadata.present.len = 0xff;
3355 /* Restore tunnel metadata. We need to use the saved options to ensure
3356 * that any unknown options are not lost. The generated mask will have
3357 * the same structure, matching on types and lengths but wildcarding
3358 * option data we don't care about. */
3359 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3360 memcpy(&flow->tunnel.metadata.opts.gnv, orig_tunnel.metadata.opts.gnv,
3361 orig_tunnel.metadata.present.len);
3362 flow->tunnel.metadata.present.len = orig_tunnel.metadata.present.len;
3363 flow->tunnel.flags |= FLOW_TNL_F_UDPIF;
3369 static inline uint32_t
3370 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3371 const struct miniflow *mf)
3373 uint32_t hash, recirc_depth;
3375 if (OVS_LIKELY(dp_packet_rss_valid(packet))) {
3376 hash = dp_packet_get_rss_hash(packet);
3378 hash = miniflow_hash_5tuple(mf, 0);
3379 dp_packet_set_rss_hash(packet, hash);
3382 /* The RSS hash must account for the recirculation depth to avoid
3383 * collisions in the exact match cache */
3384 recirc_depth = *recirc_depth_get_unsafe();
3385 if (OVS_UNLIKELY(recirc_depth)) {
3386 hash = hash_finish(hash, recirc_depth);
3387 dp_packet_set_rss_hash(packet, hash);
3392 struct packet_batch_per_flow {
3393 unsigned int byte_count;
3395 struct dp_netdev_flow *flow;
3397 struct dp_packet_batch array;
3401 packet_batch_per_flow_update(struct packet_batch_per_flow *batch,
3402 struct dp_packet *packet,
3403 const struct miniflow *mf)
3405 batch->byte_count += dp_packet_size(packet);
3406 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3407 batch->array.packets[batch->array.count++] = packet;
3411 packet_batch_per_flow_init(struct packet_batch_per_flow *batch,
3412 struct dp_netdev_flow *flow)
3414 flow->batch = batch;
3417 dp_packet_batch_init(&batch->array);
3418 batch->byte_count = 0;
3419 batch->tcp_flags = 0;
3423 packet_batch_per_flow_execute(struct packet_batch_per_flow *batch,
3424 struct dp_netdev_pmd_thread *pmd,
3427 struct dp_netdev_actions *actions;
3428 struct dp_netdev_flow *flow = batch->flow;
3430 dp_netdev_flow_used(flow, batch->array.count, batch->byte_count,
3431 batch->tcp_flags, now);
3433 actions = dp_netdev_flow_get_actions(flow);
3435 dp_netdev_execute_actions(pmd, &batch->array, true,
3436 actions->actions, actions->size);
3440 dp_netdev_queue_batches(struct dp_packet *pkt,
3441 struct dp_netdev_flow *flow, const struct miniflow *mf,
3442 struct packet_batch_per_flow *batches, size_t *n_batches)
3444 struct packet_batch_per_flow *batch = flow->batch;
3446 if (OVS_UNLIKELY(!batch)) {
3447 batch = &batches[(*n_batches)++];
3448 packet_batch_per_flow_init(batch, flow);
3451 packet_batch_per_flow_update(batch, pkt, mf);
3454 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3455 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
3456 * miniflow is copied into 'keys' and the packet pointer is moved at the
3457 * beginning of the 'packets' array.
3459 * The function returns the number of packets that needs to be processed in the
3460 * 'packets' array (they have been moved to the beginning of the vector).
3462 * If 'md_is_valid' is false, the metadata in 'packets' is not valid and must be
3463 * initialized by this function using 'port_no'.
3465 static inline size_t
3466 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet_batch *packets_,
3467 struct netdev_flow_key *keys,
3468 struct packet_batch_per_flow batches[], size_t *n_batches,
3469 bool md_is_valid, odp_port_t port_no)
3471 struct emc_cache *flow_cache = &pmd->flow_cache;
3472 struct netdev_flow_key *key = &keys[0];
3473 size_t i, n_missed = 0, n_dropped = 0;
3474 struct dp_packet **packets = packets_->packets;
3475 int cnt = packets_->count;
3477 for (i = 0; i < cnt; i++) {
3478 struct dp_netdev_flow *flow;
3479 struct dp_packet *packet = packets[i];
3481 if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) {
3482 dp_packet_delete(packet);
3488 /* Prefetch next packet data and metadata. */
3489 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3490 pkt_metadata_prefetch_init(&packets[i+1]->md);
3494 pkt_metadata_init(&packet->md, port_no);
3496 miniflow_extract(packet, &key->mf);
3497 key->len = 0; /* Not computed yet. */
3498 key->hash = dpif_netdev_packet_get_rss_hash(packet, &key->mf);
3500 flow = emc_lookup(flow_cache, key);
3501 if (OVS_LIKELY(flow)) {
3502 dp_netdev_queue_batches(packet, flow, &key->mf, batches,
3505 /* Exact match cache missed. Group missed packets together at
3506 * the beginning of the 'packets' array. */
3507 packets[n_missed] = packet;
3508 /* 'key[n_missed]' contains the key of the current packet and it
3509 * must be returned to the caller. The next key should be extracted
3510 * to 'keys[n_missed + 1]'. */
3511 key = &keys[++n_missed];
3515 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - n_dropped - n_missed);
3521 handle_packet_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet,
3522 const struct netdev_flow_key *key,
3523 struct ofpbuf *actions, struct ofpbuf *put_actions,
3526 struct ofpbuf *add_actions;
3527 struct dp_packet_batch b;
3532 match.tun_md.valid = false;
3533 miniflow_expand(&key->mf, &match.flow);
3535 ofpbuf_clear(actions);
3536 ofpbuf_clear(put_actions);
3538 dpif_flow_hash(pmd->dp->dpif, &match.flow, sizeof match.flow, &ufid);
3539 error = dp_netdev_upcall(pmd, packet, &match.flow, &match.wc,
3540 &ufid, DPIF_UC_MISS, NULL, actions,
3542 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3543 dp_packet_delete(packet);
3548 /* The Netlink encoding of datapath flow keys cannot express
3549 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3550 * tag is interpreted as exact match on the fact that there is no
3551 * VLAN. Unless we refactor a lot of code that translates between
3552 * Netlink and struct flow representations, we have to do the same
3554 if (!match.wc.masks.vlan_tci) {
3555 match.wc.masks.vlan_tci = htons(0xffff);
3558 /* We can't allow the packet batching in the next loop to execute
3559 * the actions. Otherwise, if there are any slow path actions,
3560 * we'll send the packet up twice. */
3561 packet_batch_init_packet(&b, packet);
3562 dp_netdev_execute_actions(pmd, &b, true,
3563 actions->data, actions->size);
3565 add_actions = put_actions->size ? put_actions : actions;
3566 if (OVS_LIKELY(error != ENOSPC)) {
3567 struct dp_netdev_flow *netdev_flow;
3569 /* XXX: There's a race window where a flow covering this packet
3570 * could have already been installed since we last did the flow
3571 * lookup before upcall. This could be solved by moving the
3572 * mutex lock outside the loop, but that's an awful long time
3573 * to be locking everyone out of making flow installs. If we
3574 * move to a per-core classifier, it would be reasonable. */
3575 ovs_mutex_lock(&pmd->flow_mutex);
3576 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, key);
3577 if (OVS_LIKELY(!netdev_flow)) {
3578 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3582 ovs_mutex_unlock(&pmd->flow_mutex);
3584 emc_insert(&pmd->flow_cache, key, netdev_flow);
3589 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3590 struct dp_packet_batch *packets_,
3591 struct netdev_flow_key *keys,
3592 struct packet_batch_per_flow batches[], size_t *n_batches)
3594 int cnt = packets_->count;
3595 #if !defined(__CHECKER__) && !defined(_WIN32)
3596 const size_t PKT_ARRAY_SIZE = cnt;
3598 /* Sparse or MSVC doesn't like variable length array. */
3599 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3601 struct dp_packet **packets = packets_->packets;
3602 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3603 struct dp_netdev *dp = pmd->dp;
3604 struct emc_cache *flow_cache = &pmd->flow_cache;
3605 int miss_cnt = 0, lost_cnt = 0;
3609 for (i = 0; i < cnt; i++) {
3610 /* Key length is needed in all the cases, hash computed on demand. */
3611 keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf));
3613 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3614 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3615 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3616 struct ofpbuf actions, put_actions;
3618 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3619 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3621 for (i = 0; i < cnt; i++) {
3622 struct dp_netdev_flow *netdev_flow;
3624 if (OVS_LIKELY(rules[i])) {
3628 /* It's possible that an earlier slow path execution installed
3629 * a rule covering this flow. In this case, it's a lot cheaper
3630 * to catch it here than execute a miss. */
3631 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3633 rules[i] = &netdev_flow->cr;
3638 handle_packet_upcall(pmd, packets[i], &keys[i], &actions, &put_actions,
3642 ofpbuf_uninit(&actions);
3643 ofpbuf_uninit(&put_actions);
3644 fat_rwlock_unlock(&dp->upcall_rwlock);
3645 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3646 } else if (OVS_UNLIKELY(any_miss)) {
3647 for (i = 0; i < cnt; i++) {
3648 if (OVS_UNLIKELY(!rules[i])) {
3649 dp_packet_delete(packets[i]);
3656 for (i = 0; i < cnt; i++) {
3657 struct dp_packet *packet = packets[i];
3658 struct dp_netdev_flow *flow;
3660 if (OVS_UNLIKELY(!rules[i])) {
3664 flow = dp_netdev_flow_cast(rules[i]);
3666 emc_insert(flow_cache, &keys[i], flow);
3667 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3670 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3671 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3672 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3675 /* Packets enter the datapath from a port (or from recirculation) here.
3677 * For performance reasons a caller may choose not to initialize the metadata
3678 * in 'packets': in this case 'mdinit' is false and this function needs to
3679 * initialize it using 'port_no'. If the metadata in 'packets' is already
3680 * valid, 'md_is_valid' must be true and 'port_no' will be ignored. */
3682 dp_netdev_input__(struct dp_netdev_pmd_thread *pmd,
3683 struct dp_packet_batch *packets,
3684 bool md_is_valid, odp_port_t port_no)
3686 int cnt = packets->count;
3687 #if !defined(__CHECKER__) && !defined(_WIN32)
3688 const size_t PKT_ARRAY_SIZE = cnt;
3690 /* Sparse or MSVC doesn't like variable length array. */
3691 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3693 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3694 struct packet_batch_per_flow batches[PKT_ARRAY_SIZE];
3695 long long now = time_msec();
3696 size_t newcnt, n_batches, i;
3699 newcnt = emc_processing(pmd, packets, keys, batches, &n_batches,
3700 md_is_valid, port_no);
3701 if (OVS_UNLIKELY(newcnt)) {
3702 packets->count = newcnt;
3703 fast_path_processing(pmd, packets, keys, batches, &n_batches);
3706 for (i = 0; i < n_batches; i++) {
3707 batches[i].flow->batch = NULL;
3710 for (i = 0; i < n_batches; i++) {
3711 packet_batch_per_flow_execute(&batches[i], pmd, now);
3716 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3717 struct dp_packet_batch *packets,
3720 dp_netdev_input__(pmd, packets, false, port_no);
3724 dp_netdev_recirculate(struct dp_netdev_pmd_thread *pmd,
3725 struct dp_packet_batch *packets)
3727 dp_netdev_input__(pmd, packets, true, 0);
3730 struct dp_netdev_execute_aux {
3731 struct dp_netdev_pmd_thread *pmd;
3735 dpif_netdev_register_dp_purge_cb(struct dpif *dpif, dp_purge_callback *cb,
3738 struct dp_netdev *dp = get_dp_netdev(dpif);
3739 dp->dp_purge_aux = aux;
3740 dp->dp_purge_cb = cb;
3744 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3747 struct dp_netdev *dp = get_dp_netdev(dpif);
3748 dp->upcall_aux = aux;
3753 push_tnl_action(const struct dp_netdev *dp,
3754 const struct nlattr *attr,
3755 struct dp_packet_batch *batch)
3757 struct dp_netdev_port *tun_port;
3758 const struct ovs_action_push_tnl *data;
3761 data = nl_attr_get(attr);
3763 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3768 err = netdev_push_header(tun_port->netdev, batch, data);
3773 dp_packet_delete_batch(batch, true);
3778 dp_execute_userspace_action(struct dp_netdev_pmd_thread *pmd,
3779 struct dp_packet *packet, bool may_steal,
3780 struct flow *flow, ovs_u128 *ufid,
3781 struct ofpbuf *actions,
3782 const struct nlattr *userdata)
3784 struct dp_packet_batch b;
3787 ofpbuf_clear(actions);
3789 error = dp_netdev_upcall(pmd, packet, flow, NULL, ufid,
3790 DPIF_UC_ACTION, userdata, actions,
3792 if (!error || error == ENOSPC) {
3793 packet_batch_init_packet(&b, packet);
3794 dp_netdev_execute_actions(pmd, &b, may_steal,
3795 actions->data, actions->size);
3796 } else if (may_steal) {
3797 dp_packet_delete(packet);
3802 dp_execute_cb(void *aux_, struct dp_packet_batch *packets_,
3803 const struct nlattr *a, bool may_steal)
3804 OVS_NO_THREAD_SAFETY_ANALYSIS
3806 struct dp_netdev_execute_aux *aux = aux_;
3807 uint32_t *depth = recirc_depth_get();
3808 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3809 struct dp_netdev *dp = pmd->dp;
3810 int type = nl_attr_type(a);
3811 struct dp_netdev_port *p;
3813 switch ((enum ovs_action_attr)type) {
3814 case OVS_ACTION_ATTR_OUTPUT:
3815 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3816 if (OVS_LIKELY(p)) {
3819 atomic_read_relaxed(&pmd->tx_qid, &tx_qid);
3821 netdev_send(p->netdev, tx_qid, packets_, may_steal);
3826 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3827 if (*depth < MAX_RECIRC_DEPTH) {
3828 struct dp_packet_batch tnl_pkt;
3832 dp_packet_batch_clone(&tnl_pkt, packets_);
3833 packets_ = &tnl_pkt;
3836 err = push_tnl_action(dp, a, packets_);
3839 dp_netdev_recirculate(pmd, packets_);
3846 case OVS_ACTION_ATTR_TUNNEL_POP:
3847 if (*depth < MAX_RECIRC_DEPTH) {
3848 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3850 p = dp_netdev_lookup_port(dp, portno);
3852 struct dp_packet_batch tnl_pkt;
3856 dp_packet_batch_clone(&tnl_pkt, packets_);
3857 packets_ = &tnl_pkt;
3860 netdev_pop_header(p->netdev, packets_);
3861 if (!packets_->count) {
3865 for (i = 0; i < packets_->count; i++) {
3866 packets_->packets[i]->md.in_port.odp_port = portno;
3870 dp_netdev_recirculate(pmd, packets_);
3877 case OVS_ACTION_ATTR_USERSPACE:
3878 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3879 struct dp_packet **packets = packets_->packets;
3880 const struct nlattr *userdata;
3881 struct ofpbuf actions;
3886 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3887 ofpbuf_init(&actions, 0);
3889 for (i = 0; i < packets_->count; i++) {
3890 flow_extract(packets[i], &flow);
3891 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3892 dp_execute_userspace_action(pmd, packets[i], may_steal, &flow,
3893 &ufid, &actions, userdata);
3895 ofpbuf_uninit(&actions);
3896 fat_rwlock_unlock(&dp->upcall_rwlock);
3902 case OVS_ACTION_ATTR_RECIRC:
3903 if (*depth < MAX_RECIRC_DEPTH) {
3904 struct dp_packet_batch recirc_pkts;
3908 dp_packet_batch_clone(&recirc_pkts, packets_);
3909 packets_ = &recirc_pkts;
3912 for (i = 0; i < packets_->count; i++) {
3913 packets_->packets[i]->md.recirc_id = nl_attr_get_u32(a);
3917 dp_netdev_recirculate(pmd, packets_);
3923 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3926 case OVS_ACTION_ATTR_CT:
3927 /* If a flow with this action is slow-pathed, datapath assistance is
3928 * required to implement it. However, we don't support this action
3929 * in the userspace datapath. */
3930 VLOG_WARN("Cannot execute conntrack action in userspace.");
3933 case OVS_ACTION_ATTR_PUSH_VLAN:
3934 case OVS_ACTION_ATTR_POP_VLAN:
3935 case OVS_ACTION_ATTR_PUSH_MPLS:
3936 case OVS_ACTION_ATTR_POP_MPLS:
3937 case OVS_ACTION_ATTR_SET:
3938 case OVS_ACTION_ATTR_SET_MASKED:
3939 case OVS_ACTION_ATTR_SAMPLE:
3940 case OVS_ACTION_ATTR_HASH:
3941 case OVS_ACTION_ATTR_UNSPEC:
3942 case __OVS_ACTION_ATTR_MAX:
3946 dp_packet_delete_batch(packets_, may_steal);
3950 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3951 struct dp_packet_batch *packets,
3953 const struct nlattr *actions, size_t actions_len)
3955 struct dp_netdev_execute_aux aux = { pmd };
3957 odp_execute_actions(&aux, packets, may_steal, actions,
3958 actions_len, dp_execute_cb);
3961 const struct dpif_class dpif_netdev_class = {
3964 dpif_netdev_enumerate,
3965 dpif_netdev_port_open_type,
3968 dpif_netdev_destroy,
3971 dpif_netdev_get_stats,
3972 dpif_netdev_port_add,
3973 dpif_netdev_port_del,
3974 dpif_netdev_port_query_by_number,
3975 dpif_netdev_port_query_by_name,
3976 NULL, /* port_get_pid */
3977 dpif_netdev_port_dump_start,
3978 dpif_netdev_port_dump_next,
3979 dpif_netdev_port_dump_done,
3980 dpif_netdev_port_poll,
3981 dpif_netdev_port_poll_wait,
3982 dpif_netdev_flow_flush,
3983 dpif_netdev_flow_dump_create,
3984 dpif_netdev_flow_dump_destroy,
3985 dpif_netdev_flow_dump_thread_create,
3986 dpif_netdev_flow_dump_thread_destroy,
3987 dpif_netdev_flow_dump_next,
3988 dpif_netdev_operate,
3989 NULL, /* recv_set */
3990 NULL, /* handlers_set */
3991 dpif_netdev_pmd_set,
3992 dpif_netdev_queue_to_priority,
3994 NULL, /* recv_wait */
3995 NULL, /* recv_purge */
3996 dpif_netdev_register_dp_purge_cb,
3997 dpif_netdev_register_upcall_cb,
3998 dpif_netdev_enable_upcall,
3999 dpif_netdev_disable_upcall,
4000 dpif_netdev_get_datapath_version,
4001 NULL, /* ct_dump_start */
4002 NULL, /* ct_dump_next */
4003 NULL, /* ct_dump_done */
4004 NULL, /* ct_flush */
4008 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
4009 const char *argv[], void *aux OVS_UNUSED)
4011 struct dp_netdev_port *old_port;
4012 struct dp_netdev_port *new_port;
4013 struct dp_netdev *dp;
4016 ovs_mutex_lock(&dp_netdev_mutex);
4017 dp = shash_find_data(&dp_netdevs, argv[1]);
4018 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
4019 ovs_mutex_unlock(&dp_netdev_mutex);
4020 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
4023 ovs_refcount_ref(&dp->ref_cnt);
4024 ovs_mutex_unlock(&dp_netdev_mutex);
4026 ovs_mutex_lock(&dp->port_mutex);
4027 if (get_port_by_name(dp, argv[2], &old_port)) {
4028 unixctl_command_reply_error(conn, "unknown port");
4032 port_no = u32_to_odp(atoi(argv[3]));
4033 if (!port_no || port_no == ODPP_NONE) {
4034 unixctl_command_reply_error(conn, "bad port number");
4037 if (dp_netdev_lookup_port(dp, port_no)) {
4038 unixctl_command_reply_error(conn, "port number already in use");
4042 /* Remove old port. */
4043 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
4044 ovsrcu_postpone(free, old_port);
4046 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
4047 new_port = xmemdup(old_port, sizeof *old_port);
4048 new_port->port_no = port_no;
4049 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
4051 seq_change(dp->port_seq);
4052 unixctl_command_reply(conn, NULL);
4055 ovs_mutex_unlock(&dp->port_mutex);
4056 dp_netdev_unref(dp);
4060 dpif_dummy_register__(const char *type)
4062 struct dpif_class *class;
4064 class = xmalloc(sizeof *class);
4065 *class = dpif_netdev_class;
4066 class->type = xstrdup(type);
4067 dp_register_provider(class);
4071 dpif_dummy_override(const char *type)
4076 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
4077 * a userland-only build. It's useful for testsuite.
4079 error = dp_unregister_provider(type);
4080 if (error == 0 || error == EAFNOSUPPORT) {
4081 dpif_dummy_register__(type);
4086 dpif_dummy_register(enum dummy_level level)
4088 if (level == DUMMY_OVERRIDE_ALL) {
4093 dp_enumerate_types(&types);
4094 SSET_FOR_EACH (type, &types) {
4095 dpif_dummy_override(type);
4097 sset_destroy(&types);
4098 } else if (level == DUMMY_OVERRIDE_SYSTEM) {
4099 dpif_dummy_override("system");
4102 dpif_dummy_register__("dummy");
4104 unixctl_command_register("dpif-dummy/change-port-number",
4105 "dp port new-number",
4106 3, 3, dpif_dummy_change_port_number, NULL);
4109 /* Datapath Classifier. */
4111 /* A set of rules that all have the same fields wildcarded. */
4112 struct dpcls_subtable {
4113 /* The fields are only used by writers. */
4114 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
4116 /* These fields are accessed by readers. */
4117 struct cmap rules; /* Contains "struct dpcls_rule"s. */
4118 struct netdev_flow_key mask; /* Wildcards for fields (const). */
4119 /* 'mask' must be the last field, additional space is allocated here. */
4122 /* Initializes 'cls' as a classifier that initially contains no classification
4125 dpcls_init(struct dpcls *cls)
4127 cmap_init(&cls->subtables_map);
4128 pvector_init(&cls->subtables);
4132 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
4134 pvector_remove(&cls->subtables, subtable);
4135 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
4136 subtable->mask.hash);
4137 cmap_destroy(&subtable->rules);
4138 ovsrcu_postpone(free, subtable);
4141 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
4142 * caller's responsibility.
4143 * May only be called after all the readers have been terminated. */
4145 dpcls_destroy(struct dpcls *cls)
4148 struct dpcls_subtable *subtable;
4150 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
4151 ovs_assert(cmap_count(&subtable->rules) == 0);
4152 dpcls_destroy_subtable(cls, subtable);
4154 cmap_destroy(&cls->subtables_map);
4155 pvector_destroy(&cls->subtables);
4159 static struct dpcls_subtable *
4160 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
4162 struct dpcls_subtable *subtable;
4164 /* Need to add one. */
4165 subtable = xmalloc(sizeof *subtable
4166 - sizeof subtable->mask.mf + mask->len);
4167 cmap_init(&subtable->rules);
4168 netdev_flow_key_clone(&subtable->mask, mask);
4169 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
4170 pvector_insert(&cls->subtables, subtable, 0);
4171 pvector_publish(&cls->subtables);
4176 static inline struct dpcls_subtable *
4177 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
4179 struct dpcls_subtable *subtable;
4181 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
4182 &cls->subtables_map) {
4183 if (netdev_flow_key_equal(&subtable->mask, mask)) {
4187 return dpcls_create_subtable(cls, mask);
4190 /* Insert 'rule' into 'cls'. */
4192 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
4193 const struct netdev_flow_key *mask)
4195 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
4197 rule->mask = &subtable->mask;
4198 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
4201 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
4203 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
4205 struct dpcls_subtable *subtable;
4207 ovs_assert(rule->mask);
4209 INIT_CONTAINER(subtable, rule->mask, mask);
4211 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
4213 dpcls_destroy_subtable(cls, subtable);
4214 pvector_publish(&cls->subtables);
4218 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
4219 * in 'mask' the values in 'key' and 'target' are the same. */
4221 dpcls_rule_matches_key(const struct dpcls_rule *rule,
4222 const struct netdev_flow_key *target)
4224 const uint64_t *keyp = miniflow_get_values(&rule->flow.mf);
4225 const uint64_t *maskp = miniflow_get_values(&rule->mask->mf);
4228 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, target, rule->flow.mf.map) {
4229 if (OVS_UNLIKELY((value & *maskp++) != *keyp++)) {
4236 /* For each miniflow in 'flows' performs a classifier lookup writing the result
4237 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
4238 * NULL it is skipped.
4240 * This function is optimized for use in the userspace datapath and therefore
4241 * does not implement a lot of features available in the standard
4242 * classifier_lookup() function. Specifically, it does not implement
4243 * priorities, instead returning any rule which matches the flow.
4245 * Returns true if all flows found a corresponding rule. */
4247 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
4248 struct dpcls_rule **rules, const size_t cnt)
4250 /* The batch size 16 was experimentally found faster than 8 or 32. */
4251 typedef uint16_t map_type;
4252 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
4254 #if !defined(__CHECKER__) && !defined(_WIN32)
4255 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
4257 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
4259 map_type maps[N_MAPS];
4260 struct dpcls_subtable *subtable;
4262 memset(maps, 0xff, sizeof maps);
4263 if (cnt % MAP_BITS) {
4264 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
4266 memset(rules, 0, cnt * sizeof *rules);
4268 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
4269 const struct netdev_flow_key *mkeys = keys;
4270 struct dpcls_rule **mrules = rules;
4271 map_type remains = 0;
4274 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
4276 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
4277 uint32_t hashes[MAP_BITS];
4278 const struct cmap_node *nodes[MAP_BITS];
4279 unsigned long map = maps[m];
4283 continue; /* Skip empty maps. */
4286 /* Compute hashes for the remaining keys. */
4287 ULLONG_FOR_EACH_1(i, map) {
4288 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
4292 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
4293 /* Check results. */
4294 ULLONG_FOR_EACH_1(i, map) {
4295 struct dpcls_rule *rule;
4297 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
4298 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
4303 ULLONG_SET0(map, i); /* Did not match. */
4305 ; /* Keep Sparse happy. */
4307 maps[m] &= ~map; /* Clear the found rules. */
4311 return true; /* All found. */
4314 return false; /* Some misses. */