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 "dynamic-string.h"
43 #include "fat-rwlock.h"
50 #include "netdev-dpdk.h"
51 #include "netdev-vport.h"
53 #include "odp-execute.h"
55 #include "ofp-print.h"
60 #include "poll-loop.h"
67 #include "tnl-neigh-cache.h"
68 #include "tnl-ports.h"
72 #include "openvswitch/vlog.h"
74 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
76 #define FLOW_DUMP_MAX_BATCH 50
77 /* Use per thread recirc_depth to prevent recirculation loop. */
78 #define MAX_RECIRC_DEPTH 5
79 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
81 /* Configuration parameters. */
82 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
84 /* Protects against changes to 'dp_netdevs'. */
85 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
87 /* Contains all 'struct dp_netdev's. */
88 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
89 = SHASH_INITIALIZER(&dp_netdevs);
91 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
93 static struct odp_support dp_netdev_support = {
94 .max_mpls_depth = SIZE_MAX,
98 /* Stores a miniflow with inline values */
100 struct netdev_flow_key {
101 uint32_t hash; /* Hash function differs for different users. */
102 uint32_t len; /* Length of the following miniflow (incl. map). */
104 uint64_t buf[FLOW_MAX_PACKET_U64S];
107 /* Exact match cache for frequently used flows
109 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
110 * search its entries for a miniflow that matches exactly the miniflow of the
111 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
113 * A cache entry holds a reference to its 'dp_netdev_flow'.
115 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
116 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
117 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
118 * value is the index of a cache entry where the miniflow could be.
124 * Each pmd_thread has its own private exact match cache.
125 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
128 #define EM_FLOW_HASH_SHIFT 13
129 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
130 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
131 #define EM_FLOW_HASH_SEGS 2
134 struct dp_netdev_flow *flow;
135 struct netdev_flow_key key; /* key.hash used for emc hash value. */
139 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
140 int sweep_idx; /* For emc_cache_slow_sweep(). */
143 /* Iterate in the exact match cache through every entry that might contain a
144 * miniflow with hash 'HASH'. */
145 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
146 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
147 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
148 i__ < EM_FLOW_HASH_SEGS; \
149 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
151 /* Simple non-wildcarding single-priority classifier. */
154 struct cmap subtables_map;
155 struct pvector subtables;
158 /* A rule to be inserted to the classifier. */
160 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
161 struct netdev_flow_key *mask; /* Subtable's mask. */
162 struct netdev_flow_key flow; /* Matching key. */
163 /* 'flow' must be the last field, additional space is allocated here. */
166 static void dpcls_init(struct dpcls *);
167 static void dpcls_destroy(struct dpcls *);
168 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
169 const struct netdev_flow_key *mask);
170 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
171 static bool dpcls_lookup(const struct dpcls *cls,
172 const struct netdev_flow_key keys[],
173 struct dpcls_rule **rules, size_t cnt);
175 /* Datapath based on the network device interface from netdev.h.
181 * Some members, marked 'const', are immutable. Accessing other members
182 * requires synchronization, as noted in more detail below.
184 * Acquisition order is, from outermost to innermost:
186 * dp_netdev_mutex (global)
190 const struct dpif_class *const class;
191 const char *const name;
193 struct ovs_refcount ref_cnt;
194 atomic_flag destroyed;
198 * Protected by RCU. Take the mutex to add or remove ports. */
199 struct ovs_mutex port_mutex;
201 struct seq *port_seq; /* Incremented whenever a port changes. */
203 /* Protects access to ofproto-dpif-upcall interface during revalidator
204 * thread synchronization. */
205 struct fat_rwlock upcall_rwlock;
206 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
209 /* Callback function for notifying the purging of dp flows (during
210 * reseting pmd deletion). */
211 dp_purge_callback *dp_purge_cb;
214 /* Stores all 'struct dp_netdev_pmd_thread's. */
215 struct cmap poll_threads;
217 /* Protects the access of the 'struct dp_netdev_pmd_thread'
218 * instance for non-pmd thread. */
219 struct ovs_mutex non_pmd_mutex;
221 /* Each pmd thread will store its pointer to
222 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
223 ovsthread_key_t per_pmd_key;
225 /* Cpu mask for pin of pmd threads. */
227 uint64_t last_tnl_conf_seq;
230 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
234 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
235 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
236 DP_STAT_MISS, /* Packets that did not match. */
237 DP_STAT_LOST, /* Packets not passed up to the client. */
241 enum pmd_cycles_counter_type {
242 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
243 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
247 /* A port in a netdev-based datapath. */
248 struct dp_netdev_port {
250 struct netdev *netdev;
251 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
252 struct netdev_saved_flags *sf;
253 struct netdev_rxq **rxq;
254 struct ovs_refcount ref_cnt;
255 char *type; /* Port type as requested by user. */
256 int latest_requested_n_rxq; /* Latest requested from netdev number
260 /* Contained by struct dp_netdev_flow's 'stats' member. */
261 struct dp_netdev_flow_stats {
262 atomic_llong used; /* Last used time, in monotonic msecs. */
263 atomic_ullong packet_count; /* Number of packets matched. */
264 atomic_ullong byte_count; /* Number of bytes matched. */
265 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
268 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
274 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
275 * its pmd thread's classifier. The text below calls this classifier 'cls'.
280 * The thread safety rules described here for "struct dp_netdev_flow" are
281 * motivated by two goals:
283 * - Prevent threads that read members of "struct dp_netdev_flow" from
284 * reading bad data due to changes by some thread concurrently modifying
287 * - Prevent two threads making changes to members of a given "struct
288 * dp_netdev_flow" from interfering with each other.
294 * A flow 'flow' may be accessed without a risk of being freed during an RCU
295 * grace period. Code that needs to hold onto a flow for a while
296 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
298 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
299 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
302 * Some members, marked 'const', are immutable. Accessing other members
303 * requires synchronization, as noted in more detail below.
305 struct dp_netdev_flow {
306 const struct flow flow; /* Unmasked flow that created this entry. */
307 /* Hash table index by unmasked flow. */
308 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
310 const ovs_u128 ufid; /* Unique flow identifier. */
311 const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */
314 /* Number of references.
315 * The classifier owns one reference.
316 * Any thread trying to keep a rule from being freed should hold its own
318 struct ovs_refcount ref_cnt;
323 struct dp_netdev_flow_stats stats;
326 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
328 /* While processing a group of input packets, the datapath uses the next
329 * member to store a pointer to the output batch for the flow. It is
330 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
331 * packet_batch_init() and packet_batch_execute()). */
332 struct packet_batch *batch;
334 /* Packet classification. */
335 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
336 /* 'cr' must be the last member. */
339 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
340 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
341 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
344 /* A set of datapath actions within a "struct dp_netdev_flow".
350 * A struct dp_netdev_actions 'actions' is protected with RCU. */
351 struct dp_netdev_actions {
352 /* These members are immutable: they do not change during the struct's
354 unsigned int size; /* Size of 'actions', in bytes. */
355 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
358 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
360 struct dp_netdev_actions *dp_netdev_flow_get_actions(
361 const struct dp_netdev_flow *);
362 static void dp_netdev_actions_free(struct dp_netdev_actions *);
364 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
365 struct dp_netdev_pmd_stats {
366 /* Indexed by DP_STAT_*. */
367 atomic_ullong n[DP_N_STATS];
370 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
371 struct dp_netdev_pmd_cycles {
372 /* Indexed by PMD_CYCLES_*. */
373 atomic_ullong n[PMD_N_CYCLES];
376 /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */
378 struct dp_netdev_port *port;
379 struct netdev_rxq *rx;
380 struct ovs_list node;
383 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
384 * the performance overhead of interrupt processing. Therefore netdev can
385 * not implement rx-wait for these devices. dpif-netdev needs to poll
386 * these device to check for recv buffer. pmd-thread does polling for
387 * devices assigned to itself.
389 * DPDK used PMD for accessing NIC.
391 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
392 * I/O of all non-pmd threads. There will be no actual thread created
395 * Each struct has its own flow table and classifier. Packets received
396 * from managed ports are looked up in the corresponding pmd thread's
397 * flow table, and are executed with the found actions.
399 struct dp_netdev_pmd_thread {
400 struct dp_netdev *dp;
401 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
402 struct cmap_node node; /* In 'dp->poll_threads'. */
404 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
405 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
407 /* Per thread exact-match cache. Note, the instance for cpu core
408 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
409 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
410 * instances will only be accessed by its own pmd thread. */
411 struct emc_cache flow_cache;
413 /* Classifier and Flow-Table.
415 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
416 * changes to 'cls' must be made while still holding the 'flow_mutex'.
418 struct ovs_mutex flow_mutex;
420 struct cmap flow_table OVS_GUARDED; /* Flow table. */
423 struct dp_netdev_pmd_stats stats;
425 /* Cycles counters */
426 struct dp_netdev_pmd_cycles cycles;
428 /* Used to count cicles. See 'cycles_counter_end()' */
429 unsigned long long last_cycles;
431 struct latch exit_latch; /* For terminating the pmd thread. */
432 atomic_uint change_seq; /* For reloading pmd ports. */
434 int index; /* Idx of this pmd thread among pmd*/
435 /* threads on same numa node. */
436 unsigned core_id; /* CPU core id of this pmd thread. */
437 int numa_id; /* numa node id of this pmd thread. */
438 atomic_int tx_qid; /* Queue id used by this pmd thread to
439 * send packets on all netdevs */
441 struct ovs_mutex poll_mutex; /* Mutex for poll_list. */
442 /* List of rx queues to poll. */
443 struct ovs_list poll_list OVS_GUARDED;
444 int poll_cnt; /* Number of elemints in poll_list. */
446 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
447 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
448 * values and subtracts them from 'stats' and 'cycles' before
449 * reporting to the user */
450 unsigned long long stats_zero[DP_N_STATS];
451 uint64_t cycles_zero[PMD_N_CYCLES];
454 #define PMD_INITIAL_SEQ 1
456 /* Interface to netdev-based datapath. */
459 struct dp_netdev *dp;
460 uint64_t last_port_seq;
463 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
464 struct dp_netdev_port **portp);
465 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
466 struct dp_netdev_port **portp);
467 static void dp_netdev_free(struct dp_netdev *)
468 OVS_REQUIRES(dp_netdev_mutex);
469 static int do_add_port(struct dp_netdev *dp, const char *devname,
470 const char *type, odp_port_t port_no)
471 OVS_REQUIRES(dp->port_mutex);
472 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
473 OVS_REQUIRES(dp->port_mutex);
474 static int dpif_netdev_open(const struct dpif_class *, const char *name,
475 bool create, struct dpif **);
476 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
477 struct dp_packet **, int c,
479 const struct nlattr *actions,
481 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
482 struct dp_packet **, int cnt, odp_port_t port_no);
483 static void dp_netdev_recirculate(struct dp_netdev_pmd_thread *,
484 struct dp_packet **, int cnt);
486 static void dp_netdev_disable_upcall(struct dp_netdev *);
487 static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
488 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
489 struct dp_netdev *dp, int index,
490 unsigned core_id, int numa_id);
491 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
492 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
493 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
495 static struct dp_netdev_pmd_thread *
496 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
497 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
498 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
499 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
500 static void dp_netdev_pmd_clear_poll_list(struct dp_netdev_pmd_thread *pmd);
501 static void dp_netdev_del_port_from_pmd(struct dp_netdev_port *port,
502 struct dp_netdev_pmd_thread *pmd);
503 static void dp_netdev_del_port_from_all_pmds(struct dp_netdev *dp,
504 struct dp_netdev_port *port);
506 dp_netdev_add_port_to_pmds(struct dp_netdev *dp, struct dp_netdev_port *port);
508 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
509 struct dp_netdev_port *port, struct netdev_rxq *rx);
510 static struct dp_netdev_pmd_thread *
511 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id);
512 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
513 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
514 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
515 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
517 static inline bool emc_entry_alive(struct emc_entry *ce);
518 static void emc_clear_entry(struct emc_entry *ce);
521 emc_cache_init(struct emc_cache *flow_cache)
525 flow_cache->sweep_idx = 0;
526 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
527 flow_cache->entries[i].flow = NULL;
528 flow_cache->entries[i].key.hash = 0;
529 flow_cache->entries[i].key.len = sizeof(struct miniflow);
530 flowmap_init(&flow_cache->entries[i].key.mf.map);
535 emc_cache_uninit(struct emc_cache *flow_cache)
539 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
540 emc_clear_entry(&flow_cache->entries[i]);
544 /* Check and clear dead flow references slowly (one entry at each
547 emc_cache_slow_sweep(struct emc_cache *flow_cache)
549 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
551 if (!emc_entry_alive(entry)) {
552 emc_clear_entry(entry);
554 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
557 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
559 dpif_is_netdev(const struct dpif *dpif)
561 return dpif->dpif_class->open == dpif_netdev_open;
564 static struct dpif_netdev *
565 dpif_netdev_cast(const struct dpif *dpif)
567 ovs_assert(dpif_is_netdev(dpif));
568 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
571 static struct dp_netdev *
572 get_dp_netdev(const struct dpif *dpif)
574 return dpif_netdev_cast(dpif)->dp;
578 PMD_INFO_SHOW_STATS, /* Show how cpu cycles are spent. */
579 PMD_INFO_CLEAR_STATS, /* Set the cycles count to 0. */
580 PMD_INFO_SHOW_RXQ /* Show poll-lists of pmd threads. */
584 pmd_info_show_stats(struct ds *reply,
585 struct dp_netdev_pmd_thread *pmd,
586 unsigned long long stats[DP_N_STATS],
587 uint64_t cycles[PMD_N_CYCLES])
589 unsigned long long total_packets = 0;
590 uint64_t total_cycles = 0;
593 /* These loops subtracts reference values ('*_zero') from the counters.
594 * Since loads and stores are relaxed, it might be possible for a '*_zero'
595 * value to be more recent than the current value we're reading from the
596 * counter. This is not a big problem, since these numbers are not
597 * supposed to be too accurate, but we should at least make sure that
598 * the result is not negative. */
599 for (i = 0; i < DP_N_STATS; i++) {
600 if (stats[i] > pmd->stats_zero[i]) {
601 stats[i] -= pmd->stats_zero[i];
606 if (i != DP_STAT_LOST) {
607 /* Lost packets are already included in DP_STAT_MISS */
608 total_packets += stats[i];
612 for (i = 0; i < PMD_N_CYCLES; i++) {
613 if (cycles[i] > pmd->cycles_zero[i]) {
614 cycles[i] -= pmd->cycles_zero[i];
619 total_cycles += cycles[i];
622 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
623 ? "main thread" : "pmd thread");
625 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
626 ds_put_format(reply, " numa_id %d", pmd->numa_id);
628 if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) {
629 ds_put_format(reply, " core_id %u", pmd->core_id);
631 ds_put_cstr(reply, ":\n");
634 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
635 "\tmiss:%llu\n\tlost:%llu\n",
636 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
637 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
639 if (total_cycles == 0) {
644 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
645 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
646 cycles[PMD_CYCLES_POLLING],
647 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
648 cycles[PMD_CYCLES_PROCESSING],
649 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
651 if (total_packets == 0) {
656 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
657 total_cycles / (double)total_packets,
658 total_cycles, total_packets);
661 "\tavg processing cycles per packet: "
662 "%.02f (%"PRIu64"/%llu)\n",
663 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
664 cycles[PMD_CYCLES_PROCESSING], total_packets);
668 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
669 struct dp_netdev_pmd_thread *pmd,
670 unsigned long long stats[DP_N_STATS],
671 uint64_t cycles[PMD_N_CYCLES])
675 /* We cannot write 'stats' and 'cycles' (because they're written by other
676 * threads) and we shouldn't change 'stats' (because they're used to count
677 * datapath stats, which must not be cleared here). Instead, we save the
678 * current values and subtract them from the values to be displayed in the
680 for (i = 0; i < DP_N_STATS; i++) {
681 pmd->stats_zero[i] = stats[i];
683 for (i = 0; i < PMD_N_CYCLES; i++) {
684 pmd->cycles_zero[i] = cycles[i];
689 pmd_info_show_rxq(struct ds *reply, struct dp_netdev_pmd_thread *pmd)
691 if (pmd->core_id != NON_PMD_CORE_ID) {
692 struct rxq_poll *poll;
693 const char *prev_name = NULL;
695 ds_put_format(reply, "pmd thread numa_id %d core_id %u:\n",
696 pmd->numa_id, pmd->core_id);
698 ovs_mutex_lock(&pmd->poll_mutex);
699 LIST_FOR_EACH (poll, node, &pmd->poll_list) {
700 const char *name = netdev_get_name(poll->port->netdev);
702 if (!prev_name || strcmp(name, prev_name)) {
704 ds_put_cstr(reply, "\n");
706 ds_put_format(reply, "\tport: %s\tqueue-id:",
707 netdev_get_name(poll->port->netdev));
709 ds_put_format(reply, " %d", netdev_rxq_get_queue_id(poll->rx));
712 ovs_mutex_unlock(&pmd->poll_mutex);
713 ds_put_cstr(reply, "\n");
718 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
721 struct ds reply = DS_EMPTY_INITIALIZER;
722 struct dp_netdev_pmd_thread *pmd;
723 struct dp_netdev *dp = NULL;
724 enum pmd_info_type type = *(enum pmd_info_type *) aux;
726 ovs_mutex_lock(&dp_netdev_mutex);
729 dp = shash_find_data(&dp_netdevs, argv[1]);
730 } else if (shash_count(&dp_netdevs) == 1) {
731 /* There's only one datapath */
732 dp = shash_first(&dp_netdevs)->data;
736 ovs_mutex_unlock(&dp_netdev_mutex);
737 unixctl_command_reply_error(conn,
738 "please specify an existing datapath");
742 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
743 if (type == PMD_INFO_SHOW_RXQ) {
744 pmd_info_show_rxq(&reply, pmd);
746 unsigned long long stats[DP_N_STATS];
747 uint64_t cycles[PMD_N_CYCLES];
750 /* Read current stats and cycle counters */
751 for (i = 0; i < ARRAY_SIZE(stats); i++) {
752 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
754 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
755 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
758 if (type == PMD_INFO_CLEAR_STATS) {
759 pmd_info_clear_stats(&reply, pmd, stats, cycles);
760 } else if (type == PMD_INFO_SHOW_STATS) {
761 pmd_info_show_stats(&reply, pmd, stats, cycles);
766 ovs_mutex_unlock(&dp_netdev_mutex);
768 unixctl_command_reply(conn, ds_cstr(&reply));
773 dpif_netdev_init(void)
775 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
776 clear_aux = PMD_INFO_CLEAR_STATS,
777 poll_aux = PMD_INFO_SHOW_RXQ;
779 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
780 0, 1, dpif_netdev_pmd_info,
782 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
783 0, 1, dpif_netdev_pmd_info,
785 unixctl_command_register("dpif-netdev/pmd-rxq-show", "[dp]",
786 0, 1, dpif_netdev_pmd_info,
792 dpif_netdev_enumerate(struct sset *all_dps,
793 const struct dpif_class *dpif_class)
795 struct shash_node *node;
797 ovs_mutex_lock(&dp_netdev_mutex);
798 SHASH_FOR_EACH(node, &dp_netdevs) {
799 struct dp_netdev *dp = node->data;
800 if (dpif_class != dp->class) {
801 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
802 * If the class doesn't match, skip this dpif. */
805 sset_add(all_dps, node->name);
807 ovs_mutex_unlock(&dp_netdev_mutex);
813 dpif_netdev_class_is_dummy(const struct dpif_class *class)
815 return class != &dpif_netdev_class;
819 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
821 return strcmp(type, "internal") ? type
822 : dpif_netdev_class_is_dummy(class) ? "dummy"
827 create_dpif_netdev(struct dp_netdev *dp)
829 uint16_t netflow_id = hash_string(dp->name, 0);
830 struct dpif_netdev *dpif;
832 ovs_refcount_ref(&dp->ref_cnt);
834 dpif = xmalloc(sizeof *dpif);
835 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
837 dpif->last_port_seq = seq_read(dp->port_seq);
842 /* Choose an unused, non-zero port number and return it on success.
843 * Return ODPP_NONE on failure. */
845 choose_port(struct dp_netdev *dp, const char *name)
846 OVS_REQUIRES(dp->port_mutex)
850 if (dp->class != &dpif_netdev_class) {
854 /* If the port name begins with "br", start the number search at
855 * 100 to make writing tests easier. */
856 if (!strncmp(name, "br", 2)) {
860 /* If the port name contains a number, try to assign that port number.
861 * This can make writing unit tests easier because port numbers are
863 for (p = name; *p != '\0'; p++) {
864 if (isdigit((unsigned char) *p)) {
865 port_no = start_no + strtol(p, NULL, 10);
866 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
867 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
868 return u32_to_odp(port_no);
875 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
876 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
877 return u32_to_odp(port_no);
885 create_dp_netdev(const char *name, const struct dpif_class *class,
886 struct dp_netdev **dpp)
887 OVS_REQUIRES(dp_netdev_mutex)
889 struct dp_netdev *dp;
892 dp = xzalloc(sizeof *dp);
893 shash_add(&dp_netdevs, name, dp);
895 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
896 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
897 ovs_refcount_init(&dp->ref_cnt);
898 atomic_flag_clear(&dp->destroyed);
900 ovs_mutex_init(&dp->port_mutex);
901 cmap_init(&dp->ports);
902 dp->port_seq = seq_create();
903 fat_rwlock_init(&dp->upcall_rwlock);
905 /* Disable upcalls by default. */
906 dp_netdev_disable_upcall(dp);
907 dp->upcall_aux = NULL;
908 dp->upcall_cb = NULL;
910 cmap_init(&dp->poll_threads);
911 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
912 ovsthread_key_create(&dp->per_pmd_key, NULL);
914 dp_netdev_set_nonpmd(dp);
916 ovs_mutex_lock(&dp->port_mutex);
917 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
918 ovs_mutex_unlock(&dp->port_mutex);
924 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
930 dpif_netdev_open(const struct dpif_class *class, const char *name,
931 bool create, struct dpif **dpifp)
933 struct dp_netdev *dp;
936 ovs_mutex_lock(&dp_netdev_mutex);
937 dp = shash_find_data(&dp_netdevs, name);
939 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
941 error = (dp->class != class ? EINVAL
946 *dpifp = create_dpif_netdev(dp);
949 ovs_mutex_unlock(&dp_netdev_mutex);
955 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
956 OVS_NO_THREAD_SAFETY_ANALYSIS
958 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
959 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
961 /* Before freeing a lock we should release it */
962 fat_rwlock_unlock(&dp->upcall_rwlock);
963 fat_rwlock_destroy(&dp->upcall_rwlock);
966 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
967 * through the 'dp_netdevs' shash while freeing 'dp'. */
969 dp_netdev_free(struct dp_netdev *dp)
970 OVS_REQUIRES(dp_netdev_mutex)
972 struct dp_netdev_port *port;
974 shash_find_and_delete(&dp_netdevs, dp->name);
976 dp_netdev_destroy_all_pmds(dp);
977 ovs_mutex_destroy(&dp->non_pmd_mutex);
978 ovsthread_key_delete(dp->per_pmd_key);
980 ovs_mutex_lock(&dp->port_mutex);
981 CMAP_FOR_EACH (port, node, &dp->ports) {
982 /* PMD threads are destroyed here. do_del_port() cannot quiesce */
983 do_del_port(dp, port);
985 ovs_mutex_unlock(&dp->port_mutex);
986 cmap_destroy(&dp->poll_threads);
988 seq_destroy(dp->port_seq);
989 cmap_destroy(&dp->ports);
991 /* Upcalls must be disabled at this point */
992 dp_netdev_destroy_upcall_lock(dp);
995 free(CONST_CAST(char *, dp->name));
1000 dp_netdev_unref(struct dp_netdev *dp)
1003 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
1004 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
1005 ovs_mutex_lock(&dp_netdev_mutex);
1006 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
1009 ovs_mutex_unlock(&dp_netdev_mutex);
1014 dpif_netdev_close(struct dpif *dpif)
1016 struct dp_netdev *dp = get_dp_netdev(dpif);
1018 dp_netdev_unref(dp);
1023 dpif_netdev_destroy(struct dpif *dpif)
1025 struct dp_netdev *dp = get_dp_netdev(dpif);
1027 if (!atomic_flag_test_and_set(&dp->destroyed)) {
1028 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
1029 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
1037 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
1038 * load/store semantics. While the increment is not atomic, the load and
1039 * store operations are, making it impossible to read inconsistent values.
1041 * This is used to update thread local stats counters. */
1043 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
1045 unsigned long long tmp;
1047 atomic_read_relaxed(var, &tmp);
1049 atomic_store_relaxed(var, tmp);
1053 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
1055 struct dp_netdev *dp = get_dp_netdev(dpif);
1056 struct dp_netdev_pmd_thread *pmd;
1058 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
1059 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1060 unsigned long long n;
1061 stats->n_flows += cmap_count(&pmd->flow_table);
1063 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
1065 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
1067 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
1068 stats->n_missed += n;
1069 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
1072 stats->n_masks = UINT32_MAX;
1073 stats->n_mask_hit = UINT64_MAX;
1079 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1083 if (pmd->core_id == NON_PMD_CORE_ID) {
1087 ovs_mutex_lock(&pmd->cond_mutex);
1088 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1089 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1090 ovs_mutex_unlock(&pmd->cond_mutex);
1094 hash_port_no(odp_port_t port_no)
1096 return hash_int(odp_to_u32(port_no), 0);
1100 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1102 OVS_REQUIRES(dp->port_mutex)
1104 struct netdev_saved_flags *sf;
1105 struct dp_netdev_port *port;
1106 struct netdev *netdev;
1107 enum netdev_flags flags;
1108 const char *open_type;
1112 /* Reject devices already in 'dp'. */
1113 if (!get_port_by_name(dp, devname, &port)) {
1117 /* Open and validate network device. */
1118 open_type = dpif_netdev_port_open_type(dp->class, type);
1119 error = netdev_open(devname, open_type, &netdev);
1123 /* XXX reject non-Ethernet devices */
1125 netdev_get_flags(netdev, &flags);
1126 if (flags & NETDEV_LOOPBACK) {
1127 VLOG_ERR("%s: cannot add a loopback device", devname);
1128 netdev_close(netdev);
1132 if (netdev_is_pmd(netdev)) {
1133 int n_cores = ovs_numa_get_n_cores();
1135 if (n_cores == OVS_CORE_UNSPEC) {
1136 VLOG_ERR("%s, cannot get cpu core info", devname);
1139 /* There can only be ovs_numa_get_n_cores() pmd threads,
1140 * so creates a txq for each, and one extra for the non
1142 error = netdev_set_multiq(netdev, n_cores + 1,
1143 netdev_requested_n_rxq(netdev));
1144 if (error && (error != EOPNOTSUPP)) {
1145 VLOG_ERR("%s, cannot set multiq", devname);
1149 port = xzalloc(sizeof *port);
1150 port->port_no = port_no;
1151 port->netdev = netdev;
1152 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
1153 port->type = xstrdup(type);
1154 port->latest_requested_n_rxq = netdev_requested_n_rxq(netdev);
1155 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1156 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1158 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
1159 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1160 devname, ovs_strerror(errno));
1161 netdev_close(netdev);
1169 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1171 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1172 netdev_rxq_close(port->rxq[i]);
1174 netdev_close(netdev);
1182 ovs_refcount_init(&port->ref_cnt);
1183 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1185 if (netdev_is_pmd(netdev)) {
1186 dp_netdev_add_port_to_pmds(dp, port);
1188 seq_change(dp->port_seq);
1194 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1195 odp_port_t *port_nop)
1197 struct dp_netdev *dp = get_dp_netdev(dpif);
1198 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1199 const char *dpif_port;
1203 ovs_mutex_lock(&dp->port_mutex);
1204 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1205 if (*port_nop != ODPP_NONE) {
1206 port_no = *port_nop;
1207 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1209 port_no = choose_port(dp, dpif_port);
1210 error = port_no == ODPP_NONE ? EFBIG : 0;
1213 *port_nop = port_no;
1214 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1216 ovs_mutex_unlock(&dp->port_mutex);
1222 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1224 struct dp_netdev *dp = get_dp_netdev(dpif);
1227 ovs_mutex_lock(&dp->port_mutex);
1228 if (port_no == ODPP_LOCAL) {
1231 struct dp_netdev_port *port;
1233 error = get_port_by_number(dp, port_no, &port);
1235 do_del_port(dp, port);
1238 ovs_mutex_unlock(&dp->port_mutex);
1244 is_valid_port_number(odp_port_t port_no)
1246 return port_no != ODPP_NONE;
1249 static struct dp_netdev_port *
1250 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1252 struct dp_netdev_port *port;
1254 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1255 if (port->port_no == port_no) {
1263 get_port_by_number(struct dp_netdev *dp,
1264 odp_port_t port_no, struct dp_netdev_port **portp)
1266 if (!is_valid_port_number(port_no)) {
1270 *portp = dp_netdev_lookup_port(dp, port_no);
1271 return *portp ? 0 : ENOENT;
1276 port_ref(struct dp_netdev_port *port)
1279 ovs_refcount_ref(&port->ref_cnt);
1284 port_unref(struct dp_netdev_port *port)
1286 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1287 int n_rxq = netdev_n_rxq(port->netdev);
1290 netdev_close(port->netdev);
1291 netdev_restore_flags(port->sf);
1293 for (i = 0; i < n_rxq; i++) {
1294 netdev_rxq_close(port->rxq[i]);
1303 get_port_by_name(struct dp_netdev *dp,
1304 const char *devname, struct dp_netdev_port **portp)
1305 OVS_REQUIRES(dp->port_mutex)
1307 struct dp_netdev_port *port;
1309 CMAP_FOR_EACH (port, node, &dp->ports) {
1310 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1319 get_n_pmd_threads(struct dp_netdev *dp)
1321 /* There is one non pmd thread in dp->poll_threads */
1322 return cmap_count(&dp->poll_threads) - 1;
1326 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1328 struct dp_netdev_pmd_thread *pmd;
1331 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1332 if (pmd->numa_id == numa_id) {
1340 /* Returns 'true' if there is a port with pmd netdev and the netdev
1341 * is on numa node 'numa_id'. */
1343 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1345 struct dp_netdev_port *port;
1347 CMAP_FOR_EACH (port, node, &dp->ports) {
1348 if (netdev_is_pmd(port->netdev)
1349 && netdev_get_numa_id(port->netdev) == numa_id) {
1359 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1360 OVS_REQUIRES(dp->port_mutex)
1362 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1363 seq_change(dp->port_seq);
1364 if (netdev_is_pmd(port->netdev)) {
1365 int numa_id = netdev_get_numa_id(port->netdev);
1367 /* PMD threads can not be on invalid numa node. */
1368 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
1369 /* If there is no netdev on the numa node, deletes the pmd threads
1370 * for that numa. Else, deletes the queues from polling lists. */
1371 if (!has_pmd_port_for_numa(dp, numa_id)) {
1372 dp_netdev_del_pmds_on_numa(dp, numa_id);
1374 dp_netdev_del_port_from_all_pmds(dp, port);
1382 answer_port_query(const struct dp_netdev_port *port,
1383 struct dpif_port *dpif_port)
1385 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1386 dpif_port->type = xstrdup(port->type);
1387 dpif_port->port_no = port->port_no;
1391 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1392 struct dpif_port *dpif_port)
1394 struct dp_netdev *dp = get_dp_netdev(dpif);
1395 struct dp_netdev_port *port;
1398 error = get_port_by_number(dp, port_no, &port);
1399 if (!error && dpif_port) {
1400 answer_port_query(port, dpif_port);
1407 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1408 struct dpif_port *dpif_port)
1410 struct dp_netdev *dp = get_dp_netdev(dpif);
1411 struct dp_netdev_port *port;
1414 ovs_mutex_lock(&dp->port_mutex);
1415 error = get_port_by_name(dp, devname, &port);
1416 if (!error && dpif_port) {
1417 answer_port_query(port, dpif_port);
1419 ovs_mutex_unlock(&dp->port_mutex);
1425 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1427 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1431 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1433 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1434 ovsrcu_postpone(dp_netdev_flow_free, flow);
1439 dp_netdev_flow_hash(const ovs_u128 *ufid)
1441 return ufid->u32[0];
1445 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1446 struct dp_netdev_flow *flow)
1447 OVS_REQUIRES(pmd->flow_mutex)
1449 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1451 dpcls_remove(&pmd->cls, &flow->cr);
1452 flow->cr.mask = NULL; /* Accessing rule's mask after this is not safe. */
1454 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1457 dp_netdev_flow_unref(flow);
1461 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1463 struct dp_netdev_flow *netdev_flow;
1465 ovs_mutex_lock(&pmd->flow_mutex);
1466 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1467 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1469 ovs_mutex_unlock(&pmd->flow_mutex);
1473 dpif_netdev_flow_flush(struct dpif *dpif)
1475 struct dp_netdev *dp = get_dp_netdev(dpif);
1476 struct dp_netdev_pmd_thread *pmd;
1478 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1479 dp_netdev_pmd_flow_flush(pmd);
1485 struct dp_netdev_port_state {
1486 struct cmap_position position;
1491 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1493 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1498 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1499 struct dpif_port *dpif_port)
1501 struct dp_netdev_port_state *state = state_;
1502 struct dp_netdev *dp = get_dp_netdev(dpif);
1503 struct cmap_node *node;
1506 node = cmap_next_position(&dp->ports, &state->position);
1508 struct dp_netdev_port *port;
1510 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1513 state->name = xstrdup(netdev_get_name(port->netdev));
1514 dpif_port->name = state->name;
1515 dpif_port->type = port->type;
1516 dpif_port->port_no = port->port_no;
1527 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1529 struct dp_netdev_port_state *state = state_;
1536 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1538 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1539 uint64_t new_port_seq;
1542 new_port_seq = seq_read(dpif->dp->port_seq);
1543 if (dpif->last_port_seq != new_port_seq) {
1544 dpif->last_port_seq = new_port_seq;
1554 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1556 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1558 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1561 static struct dp_netdev_flow *
1562 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1564 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1567 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1569 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1572 /* netdev_flow_key utilities.
1574 * netdev_flow_key is basically a miniflow. We use these functions
1575 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1576 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1578 * - Since we are dealing exclusively with miniflows created by
1579 * miniflow_extract(), if the map is different the miniflow is different.
1580 * Therefore we can be faster by comparing the map and the miniflow in a
1582 * - These functions can be inlined by the compiler. */
1584 /* Given the number of bits set in miniflow's maps, returns the size of the
1585 * 'netdev_flow_key.mf' */
1586 static inline size_t
1587 netdev_flow_key_size(size_t flow_u64s)
1589 return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s);
1593 netdev_flow_key_equal(const struct netdev_flow_key *a,
1594 const struct netdev_flow_key *b)
1596 /* 'b->len' may be not set yet. */
1597 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1600 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1601 * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been
1602 * generated by miniflow_extract. */
1604 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1605 const struct miniflow *mf)
1607 return !memcmp(&key->mf, mf, key->len);
1611 netdev_flow_key_clone(struct netdev_flow_key *dst,
1612 const struct netdev_flow_key *src)
1615 offsetof(struct netdev_flow_key, mf) + src->len);
1620 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1621 const struct flow *src)
1623 struct dp_packet packet;
1624 uint64_t buf_stub[512 / 8];
1626 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1627 pkt_metadata_from_flow(&packet.md, src);
1628 flow_compose(&packet, src);
1629 miniflow_extract(&packet, &dst->mf);
1630 dp_packet_uninit(&packet);
1632 dst->len = netdev_flow_key_size(miniflow_n_values(&dst->mf));
1633 dst->hash = 0; /* Not computed yet. */
1636 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1638 netdev_flow_mask_init(struct netdev_flow_key *mask,
1639 const struct match *match)
1641 uint64_t *dst = miniflow_values(&mask->mf);
1642 struct flowmap fmap;
1646 /* Only check masks that make sense for the flow. */
1647 flow_wc_map(&match->flow, &fmap);
1648 flowmap_init(&mask->mf.map);
1650 FLOWMAP_FOR_EACH_INDEX(idx, fmap) {
1651 uint64_t mask_u64 = flow_u64_value(&match->wc.masks, idx);
1654 flowmap_set(&mask->mf.map, idx, 1);
1656 hash = hash_add64(hash, mask_u64);
1662 FLOWMAP_FOR_EACH_MAP (map, mask->mf.map) {
1663 hash = hash_add64(hash, map);
1666 size_t n = dst - miniflow_get_values(&mask->mf);
1668 mask->hash = hash_finish(hash, n * 8);
1669 mask->len = netdev_flow_key_size(n);
1672 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1674 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1675 const struct flow *flow,
1676 const struct netdev_flow_key *mask)
1678 uint64_t *dst_u64 = miniflow_values(&dst->mf);
1679 const uint64_t *mask_u64 = miniflow_get_values(&mask->mf);
1683 dst->len = mask->len;
1684 dst->mf = mask->mf; /* Copy maps. */
1686 FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf.map) {
1687 *dst_u64 = value & *mask_u64++;
1688 hash = hash_add64(hash, *dst_u64++);
1690 dst->hash = hash_finish(hash,
1691 (dst_u64 - miniflow_get_values(&dst->mf)) * 8);
1694 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
1695 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
1696 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
1698 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1700 static inline uint32_t
1701 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1702 const struct netdev_flow_key *mask)
1704 const uint64_t *p = miniflow_get_values(&mask->mf);
1708 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, key, mask->mf.map) {
1709 hash = hash_add64(hash, value & *p++);
1712 return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8);
1716 emc_entry_alive(struct emc_entry *ce)
1718 return ce->flow && !ce->flow->dead;
1722 emc_clear_entry(struct emc_entry *ce)
1725 dp_netdev_flow_unref(ce->flow);
1731 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1732 const struct netdev_flow_key *key)
1734 if (ce->flow != flow) {
1736 dp_netdev_flow_unref(ce->flow);
1739 if (dp_netdev_flow_ref(flow)) {
1746 netdev_flow_key_clone(&ce->key, key);
1751 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1752 struct dp_netdev_flow *flow)
1754 struct emc_entry *to_be_replaced = NULL;
1755 struct emc_entry *current_entry;
1757 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1758 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1759 /* We found the entry with the 'mf' miniflow */
1760 emc_change_entry(current_entry, flow, NULL);
1764 /* Replacement policy: put the flow in an empty (not alive) entry, or
1765 * in the first entry where it can be */
1767 || (emc_entry_alive(to_be_replaced)
1768 && !emc_entry_alive(current_entry))
1769 || current_entry->key.hash < to_be_replaced->key.hash) {
1770 to_be_replaced = current_entry;
1773 /* We didn't find the miniflow in the cache.
1774 * The 'to_be_replaced' entry is where the new flow will be stored */
1776 emc_change_entry(to_be_replaced, flow, key);
1779 static inline struct dp_netdev_flow *
1780 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1782 struct emc_entry *current_entry;
1784 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1785 if (current_entry->key.hash == key->hash
1786 && emc_entry_alive(current_entry)
1787 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1789 /* We found the entry with the 'key->mf' miniflow */
1790 return current_entry->flow;
1797 static struct dp_netdev_flow *
1798 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1799 const struct netdev_flow_key *key)
1801 struct dp_netdev_flow *netdev_flow;
1802 struct dpcls_rule *rule;
1804 dpcls_lookup(&pmd->cls, key, &rule, 1);
1805 netdev_flow = dp_netdev_flow_cast(rule);
1810 static struct dp_netdev_flow *
1811 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1812 const ovs_u128 *ufidp, const struct nlattr *key,
1815 struct dp_netdev_flow *netdev_flow;
1819 /* If a UFID is not provided, determine one based on the key. */
1820 if (!ufidp && key && key_len
1821 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1822 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1827 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1829 if (ovs_u128_equals(&netdev_flow->ufid, ufidp)) {
1839 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1840 struct dpif_flow_stats *stats)
1842 struct dp_netdev_flow *netdev_flow;
1843 unsigned long long n;
1847 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1849 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1850 stats->n_packets = n;
1851 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1853 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1855 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1856 stats->tcp_flags = flags;
1859 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1860 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1861 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1864 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1865 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1866 struct dpif_flow *flow, bool terse)
1869 memset(flow, 0, sizeof *flow);
1871 struct flow_wildcards wc;
1872 struct dp_netdev_actions *actions;
1874 struct odp_flow_key_parms odp_parms = {
1875 .flow = &netdev_flow->flow,
1877 .support = dp_netdev_support,
1880 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1883 offset = key_buf->size;
1884 flow->key = ofpbuf_tail(key_buf);
1885 odp_parms.odp_in_port = netdev_flow->flow.in_port.odp_port;
1886 odp_flow_key_from_flow(&odp_parms, key_buf);
1887 flow->key_len = key_buf->size - offset;
1890 offset = mask_buf->size;
1891 flow->mask = ofpbuf_tail(mask_buf);
1892 odp_parms.odp_in_port = wc.masks.in_port.odp_port;
1893 odp_parms.key_buf = key_buf;
1894 odp_flow_key_from_mask(&odp_parms, mask_buf);
1895 flow->mask_len = mask_buf->size - offset;
1898 actions = dp_netdev_flow_get_actions(netdev_flow);
1899 flow->actions = actions->actions;
1900 flow->actions_len = actions->size;
1903 flow->ufid = netdev_flow->ufid;
1904 flow->ufid_present = true;
1905 flow->pmd_id = netdev_flow->pmd_id;
1906 get_dpif_flow_stats(netdev_flow, &flow->stats);
1910 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1911 const struct nlattr *mask_key,
1912 uint32_t mask_key_len, const struct flow *flow,
1913 struct flow_wildcards *wc)
1915 enum odp_key_fitness fitness;
1917 fitness = odp_flow_key_to_mask_udpif(mask_key, mask_key_len, key,
1920 /* This should not happen: it indicates that
1921 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1922 * disagree on the acceptable form of a mask. Log the problem
1923 * as an error, with enough details to enable debugging. */
1924 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1926 if (!VLOG_DROP_ERR(&rl)) {
1930 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1932 VLOG_ERR("internal error parsing flow mask %s (%s)",
1933 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1944 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1949 if (odp_flow_key_to_flow_udpif(key, key_len, flow)) {
1950 /* This should not happen: it indicates that odp_flow_key_from_flow()
1951 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1952 * flow. Log the problem as an error, with enough details to enable
1954 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1956 if (!VLOG_DROP_ERR(&rl)) {
1960 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1961 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1968 in_port = flow->in_port.odp_port;
1969 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1973 /* Userspace datapath doesn't support conntrack. */
1974 if (flow->ct_state || flow->ct_zone || flow->ct_mark
1975 || !ovs_u128_is_zero(&flow->ct_label)) {
1983 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1985 struct dp_netdev *dp = get_dp_netdev(dpif);
1986 struct dp_netdev_flow *netdev_flow;
1987 struct dp_netdev_pmd_thread *pmd;
1988 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1989 ? NON_PMD_CORE_ID : get->pmd_id;
1992 pmd = dp_netdev_get_pmd(dp, pmd_id);
1997 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
2000 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
2005 dp_netdev_pmd_unref(pmd);
2011 static struct dp_netdev_flow *
2012 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
2013 struct match *match, const ovs_u128 *ufid,
2014 const struct nlattr *actions, size_t actions_len)
2015 OVS_REQUIRES(pmd->flow_mutex)
2017 struct dp_netdev_flow *flow;
2018 struct netdev_flow_key mask;
2020 netdev_flow_mask_init(&mask, match);
2021 /* Make sure wc does not have metadata. */
2022 ovs_assert(!FLOWMAP_HAS_FIELD(&mask.mf.map, metadata)
2023 && !FLOWMAP_HAS_FIELD(&mask.mf.map, regs));
2025 /* Do not allocate extra space. */
2026 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
2027 memset(&flow->stats, 0, sizeof flow->stats);
2030 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
2031 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
2032 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
2033 ovs_refcount_init(&flow->ref_cnt);
2034 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
2036 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
2037 dpcls_insert(&pmd->cls, &flow->cr, &mask);
2039 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
2040 dp_netdev_flow_hash(&flow->ufid));
2042 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2044 struct ds ds = DS_EMPTY_INITIALIZER;
2046 match.tun_md.valid = false;
2047 match.flow = flow->flow;
2048 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2050 ds_put_cstr(&ds, "flow_add: ");
2051 odp_format_ufid(ufid, &ds);
2052 ds_put_cstr(&ds, " ");
2053 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2054 ds_put_cstr(&ds, ", actions:");
2055 format_odp_actions(&ds, actions, actions_len);
2057 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2066 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2068 struct dp_netdev *dp = get_dp_netdev(dpif);
2069 struct dp_netdev_flow *netdev_flow;
2070 struct netdev_flow_key key;
2071 struct dp_netdev_pmd_thread *pmd;
2074 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2075 ? NON_PMD_CORE_ID : put->pmd_id;
2078 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2082 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2083 put->mask, put->mask_len,
2084 &match.flow, &match.wc);
2089 pmd = dp_netdev_get_pmd(dp, pmd_id);
2094 /* Must produce a netdev_flow_key for lookup.
2095 * This interface is no longer performance critical, since it is not used
2096 * for upcall processing any more. */
2097 netdev_flow_key_from_flow(&key, &match.flow);
2102 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2105 ovs_mutex_lock(&pmd->flow_mutex);
2106 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2108 if (put->flags & DPIF_FP_CREATE) {
2109 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2111 memset(put->stats, 0, sizeof *put->stats);
2113 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2123 if (put->flags & DPIF_FP_MODIFY
2124 && flow_equal(&match.flow, &netdev_flow->flow)) {
2125 struct dp_netdev_actions *new_actions;
2126 struct dp_netdev_actions *old_actions;
2128 new_actions = dp_netdev_actions_create(put->actions,
2131 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2132 ovsrcu_set(&netdev_flow->actions, new_actions);
2135 get_dpif_flow_stats(netdev_flow, put->stats);
2137 if (put->flags & DPIF_FP_ZERO_STATS) {
2138 /* XXX: The userspace datapath uses thread local statistics
2139 * (for flows), which should be updated only by the owning
2140 * thread. Since we cannot write on stats memory here,
2141 * we choose not to support this flag. Please note:
2142 * - This feature is currently used only by dpctl commands with
2144 * - Should the need arise, this operation can be implemented
2145 * by keeping a base value (to be update here) for each
2146 * counter, and subtracting it before outputting the stats */
2150 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2151 } else if (put->flags & DPIF_FP_CREATE) {
2154 /* Overlapping flow. */
2158 ovs_mutex_unlock(&pmd->flow_mutex);
2159 dp_netdev_pmd_unref(pmd);
2165 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2167 struct dp_netdev *dp = get_dp_netdev(dpif);
2168 struct dp_netdev_flow *netdev_flow;
2169 struct dp_netdev_pmd_thread *pmd;
2170 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2171 ? NON_PMD_CORE_ID : del->pmd_id;
2174 pmd = dp_netdev_get_pmd(dp, pmd_id);
2179 ovs_mutex_lock(&pmd->flow_mutex);
2180 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2184 get_dpif_flow_stats(netdev_flow, del->stats);
2186 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2190 ovs_mutex_unlock(&pmd->flow_mutex);
2191 dp_netdev_pmd_unref(pmd);
2196 struct dpif_netdev_flow_dump {
2197 struct dpif_flow_dump up;
2198 struct cmap_position poll_thread_pos;
2199 struct cmap_position flow_pos;
2200 struct dp_netdev_pmd_thread *cur_pmd;
2202 struct ovs_mutex mutex;
2205 static struct dpif_netdev_flow_dump *
2206 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2208 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2211 static struct dpif_flow_dump *
2212 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2214 struct dpif_netdev_flow_dump *dump;
2216 dump = xzalloc(sizeof *dump);
2217 dpif_flow_dump_init(&dump->up, dpif_);
2218 dump->up.terse = terse;
2219 ovs_mutex_init(&dump->mutex);
2225 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2227 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2229 ovs_mutex_destroy(&dump->mutex);
2234 struct dpif_netdev_flow_dump_thread {
2235 struct dpif_flow_dump_thread up;
2236 struct dpif_netdev_flow_dump *dump;
2237 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2238 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2241 static struct dpif_netdev_flow_dump_thread *
2242 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2244 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2247 static struct dpif_flow_dump_thread *
2248 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2250 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2251 struct dpif_netdev_flow_dump_thread *thread;
2253 thread = xmalloc(sizeof *thread);
2254 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2255 thread->dump = dump;
2260 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2262 struct dpif_netdev_flow_dump_thread *thread
2263 = dpif_netdev_flow_dump_thread_cast(thread_);
2269 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2270 struct dpif_flow *flows, int max_flows)
2272 struct dpif_netdev_flow_dump_thread *thread
2273 = dpif_netdev_flow_dump_thread_cast(thread_);
2274 struct dpif_netdev_flow_dump *dump = thread->dump;
2275 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2279 ovs_mutex_lock(&dump->mutex);
2280 if (!dump->status) {
2281 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2282 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2283 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2284 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2286 /* First call to dump_next(), extracts the first pmd thread.
2287 * If there is no pmd thread, returns immediately. */
2289 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2291 ovs_mutex_unlock(&dump->mutex);
2298 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2299 struct cmap_node *node;
2301 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2305 netdev_flows[n_flows] = CONTAINER_OF(node,
2306 struct dp_netdev_flow,
2309 /* When finishing dumping the current pmd thread, moves to
2311 if (n_flows < flow_limit) {
2312 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2313 dp_netdev_pmd_unref(pmd);
2314 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2320 /* Keeps the reference to next caller. */
2321 dump->cur_pmd = pmd;
2323 /* If the current dump is empty, do not exit the loop, since the
2324 * remaining pmds could have flows to be dumped. Just dumps again
2325 * on the new 'pmd'. */
2328 ovs_mutex_unlock(&dump->mutex);
2330 for (i = 0; i < n_flows; i++) {
2331 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2332 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2333 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2334 struct dpif_flow *f = &flows[i];
2335 struct ofpbuf key, mask;
2337 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2338 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2339 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2347 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2348 OVS_NO_THREAD_SAFETY_ANALYSIS
2350 struct dp_netdev *dp = get_dp_netdev(dpif);
2351 struct dp_netdev_pmd_thread *pmd;
2352 struct dp_packet *pp;
2354 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2355 dp_packet_size(execute->packet) > UINT16_MAX) {
2359 /* Tries finding the 'pmd'. If NULL is returned, that means
2360 * the current thread is a non-pmd thread and should use
2361 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2362 pmd = ovsthread_getspecific(dp->per_pmd_key);
2364 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2367 /* If the current thread is non-pmd thread, acquires
2368 * the 'non_pmd_mutex'. */
2369 if (pmd->core_id == NON_PMD_CORE_ID) {
2370 ovs_mutex_lock(&dp->non_pmd_mutex);
2371 ovs_mutex_lock(&dp->port_mutex);
2374 pp = execute->packet;
2375 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2376 execute->actions_len);
2377 if (pmd->core_id == NON_PMD_CORE_ID) {
2378 dp_netdev_pmd_unref(pmd);
2379 ovs_mutex_unlock(&dp->port_mutex);
2380 ovs_mutex_unlock(&dp->non_pmd_mutex);
2387 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2391 for (i = 0; i < n_ops; i++) {
2392 struct dpif_op *op = ops[i];
2395 case DPIF_OP_FLOW_PUT:
2396 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2399 case DPIF_OP_FLOW_DEL:
2400 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2403 case DPIF_OP_EXECUTE:
2404 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2407 case DPIF_OP_FLOW_GET:
2408 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2414 /* Returns true if the configuration for rx queues or cpu mask
2417 pmd_config_changed(const struct dp_netdev *dp, const char *cmask)
2419 struct dp_netdev_port *port;
2421 CMAP_FOR_EACH (port, node, &dp->ports) {
2422 struct netdev *netdev = port->netdev;
2423 int requested_n_rxq = netdev_requested_n_rxq(netdev);
2424 if (netdev_is_pmd(netdev)
2425 && port->latest_requested_n_rxq != requested_n_rxq) {
2430 if (dp->pmd_cmask != NULL && cmask != NULL) {
2431 return strcmp(dp->pmd_cmask, cmask);
2433 return (dp->pmd_cmask != NULL || cmask != NULL);
2437 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2439 dpif_netdev_pmd_set(struct dpif *dpif, const char *cmask)
2441 struct dp_netdev *dp = get_dp_netdev(dpif);
2443 if (pmd_config_changed(dp, cmask)) {
2444 struct dp_netdev_port *port;
2446 dp_netdev_destroy_all_pmds(dp);
2448 CMAP_FOR_EACH (port, node, &dp->ports) {
2449 struct netdev *netdev = port->netdev;
2450 int requested_n_rxq = netdev_requested_n_rxq(netdev);
2451 if (netdev_is_pmd(port->netdev)
2452 && port->latest_requested_n_rxq != requested_n_rxq) {
2455 /* Closes the existing 'rxq's. */
2456 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2457 netdev_rxq_close(port->rxq[i]);
2458 port->rxq[i] = NULL;
2461 /* Sets the new rx queue config. */
2462 err = netdev_set_multiq(port->netdev,
2463 ovs_numa_get_n_cores() + 1,
2465 if (err && (err != EOPNOTSUPP)) {
2466 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2467 " %u", netdev_get_name(port->netdev),
2471 port->latest_requested_n_rxq = requested_n_rxq;
2472 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2473 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2474 * netdev_n_rxq(port->netdev));
2475 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2476 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2480 /* Reconfigures the cpu mask. */
2481 ovs_numa_set_cpu_mask(cmask);
2482 free(dp->pmd_cmask);
2483 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2485 /* Restores the non-pmd. */
2486 dp_netdev_set_nonpmd(dp);
2487 /* Restores all pmd threads. */
2488 dp_netdev_reset_pmd_threads(dp);
2495 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2496 uint32_t queue_id, uint32_t *priority)
2498 *priority = queue_id;
2503 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2504 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2505 struct dp_netdev_actions *
2506 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2508 struct dp_netdev_actions *netdev_actions;
2510 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2511 memcpy(netdev_actions->actions, actions, size);
2512 netdev_actions->size = size;
2514 return netdev_actions;
2517 struct dp_netdev_actions *
2518 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2520 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2524 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2529 static inline unsigned long long
2530 cycles_counter(void)
2533 return rte_get_tsc_cycles();
2539 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2540 extern struct ovs_mutex cycles_counter_fake_mutex;
2542 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2544 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2545 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2546 OVS_NO_THREAD_SAFETY_ANALYSIS
2548 pmd->last_cycles = cycles_counter();
2551 /* Stop counting cycles and add them to the counter 'type' */
2553 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2554 enum pmd_cycles_counter_type type)
2555 OVS_RELEASES(&cycles_counter_fake_mutex)
2556 OVS_NO_THREAD_SAFETY_ANALYSIS
2558 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2560 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2564 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2565 struct dp_netdev_port *port,
2566 struct netdev_rxq *rxq)
2568 struct dp_packet *packets[NETDEV_MAX_BURST];
2571 cycles_count_start(pmd);
2572 error = netdev_rxq_recv(rxq, packets, &cnt);
2573 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2575 *recirc_depth_get() = 0;
2577 cycles_count_start(pmd);
2578 dp_netdev_input(pmd, packets, cnt, port->port_no);
2579 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2580 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2581 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2583 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2584 netdev_get_name(port->netdev), ovs_strerror(error));
2588 /* Return true if needs to revalidate datapath flows. */
2590 dpif_netdev_run(struct dpif *dpif)
2592 struct dp_netdev_port *port;
2593 struct dp_netdev *dp = get_dp_netdev(dpif);
2594 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2596 uint64_t new_tnl_seq;
2598 ovs_mutex_lock(&dp->non_pmd_mutex);
2599 CMAP_FOR_EACH (port, node, &dp->ports) {
2600 if (!netdev_is_pmd(port->netdev)) {
2603 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2604 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2608 ovs_mutex_unlock(&dp->non_pmd_mutex);
2609 dp_netdev_pmd_unref(non_pmd);
2611 tnl_neigh_cache_run();
2613 new_tnl_seq = seq_read(tnl_conf_seq);
2615 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2616 dp->last_tnl_conf_seq = new_tnl_seq;
2623 dpif_netdev_wait(struct dpif *dpif)
2625 struct dp_netdev_port *port;
2626 struct dp_netdev *dp = get_dp_netdev(dpif);
2628 ovs_mutex_lock(&dp_netdev_mutex);
2629 CMAP_FOR_EACH (port, node, &dp->ports) {
2630 if (!netdev_is_pmd(port->netdev)) {
2633 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2634 netdev_rxq_wait(port->rxq[i]);
2638 ovs_mutex_unlock(&dp_netdev_mutex);
2639 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2643 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2644 struct rxq_poll **ppoll_list, int poll_cnt)
2645 OVS_REQUIRES(pmd->poll_mutex)
2647 struct rxq_poll *poll_list = *ppoll_list;
2648 struct rxq_poll *poll;
2651 for (i = 0; i < poll_cnt; i++) {
2652 port_unref(poll_list[i].port);
2655 poll_list = xrealloc(poll_list, pmd->poll_cnt * sizeof *poll_list);
2658 LIST_FOR_EACH (poll, node, &pmd->poll_list) {
2659 port_ref(poll->port);
2660 poll_list[i++] = *poll;
2663 *ppoll_list = poll_list;
2664 return pmd->poll_cnt;
2668 pmd_thread_main(void *f_)
2670 struct dp_netdev_pmd_thread *pmd = f_;
2671 unsigned int lc = 0;
2672 struct rxq_poll *poll_list;
2673 unsigned int port_seq = PMD_INITIAL_SEQ;
2680 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2681 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2682 pmd_thread_setaffinity_cpu(pmd->core_id);
2684 emc_cache_init(&pmd->flow_cache);
2686 ovs_mutex_lock(&pmd->poll_mutex);
2687 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2688 ovs_mutex_unlock(&pmd->poll_mutex);
2690 /* List port/core affinity */
2691 for (i = 0; i < poll_cnt; i++) {
2692 VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n",
2693 pmd->core_id, netdev_get_name(poll_list[i].port->netdev),
2694 netdev_rxq_get_queue_id(poll_list[i].rx));
2697 /* Signal here to make sure the pmd finishes
2698 * reloading the updated configuration. */
2699 dp_netdev_pmd_reload_done(pmd);
2702 for (i = 0; i < poll_cnt; i++) {
2703 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2711 emc_cache_slow_sweep(&pmd->flow_cache);
2712 coverage_try_clear();
2715 atomic_read_relaxed(&pmd->change_seq, &seq);
2716 if (seq != port_seq) {
2723 emc_cache_uninit(&pmd->flow_cache);
2725 if (!latch_is_set(&pmd->exit_latch)){
2729 for (i = 0; i < poll_cnt; i++) {
2730 port_unref(poll_list[i].port);
2733 dp_netdev_pmd_reload_done(pmd);
2740 dp_netdev_disable_upcall(struct dp_netdev *dp)
2741 OVS_ACQUIRES(dp->upcall_rwlock)
2743 fat_rwlock_wrlock(&dp->upcall_rwlock);
2747 dpif_netdev_disable_upcall(struct dpif *dpif)
2748 OVS_NO_THREAD_SAFETY_ANALYSIS
2750 struct dp_netdev *dp = get_dp_netdev(dpif);
2751 dp_netdev_disable_upcall(dp);
2755 dp_netdev_enable_upcall(struct dp_netdev *dp)
2756 OVS_RELEASES(dp->upcall_rwlock)
2758 fat_rwlock_unlock(&dp->upcall_rwlock);
2762 dpif_netdev_enable_upcall(struct dpif *dpif)
2763 OVS_NO_THREAD_SAFETY_ANALYSIS
2765 struct dp_netdev *dp = get_dp_netdev(dpif);
2766 dp_netdev_enable_upcall(dp);
2770 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2772 ovs_mutex_lock(&pmd->cond_mutex);
2773 xpthread_cond_signal(&pmd->cond);
2774 ovs_mutex_unlock(&pmd->cond_mutex);
2777 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2778 * the pointer if succeeds, otherwise, NULL.
2780 * Caller must unrefs the returned reference. */
2781 static struct dp_netdev_pmd_thread *
2782 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2784 struct dp_netdev_pmd_thread *pmd;
2785 const struct cmap_node *pnode;
2787 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2791 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2793 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2796 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2798 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2800 struct dp_netdev_pmd_thread *non_pmd;
2802 non_pmd = xzalloc(sizeof *non_pmd);
2803 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2807 /* Caller must have valid pointer to 'pmd'. */
2809 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2811 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2815 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2817 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2818 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2822 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2823 * fails, keeps checking for next node until reaching the end of cmap.
2825 * Caller must unrefs the returned reference. */
2826 static struct dp_netdev_pmd_thread *
2827 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2829 struct dp_netdev_pmd_thread *next;
2832 struct cmap_node *node;
2834 node = cmap_next_position(&dp->poll_threads, pos);
2835 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2837 } while (next && !dp_netdev_pmd_try_ref(next));
2842 /* Configures the 'pmd' based on the input argument. */
2844 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2845 int index, unsigned core_id, int numa_id)
2849 pmd->core_id = core_id;
2850 pmd->numa_id = numa_id;
2853 atomic_init(&pmd->tx_qid,
2854 (core_id == NON_PMD_CORE_ID)
2855 ? ovs_numa_get_n_cores()
2856 : get_n_pmd_threads(dp));
2858 ovs_refcount_init(&pmd->ref_cnt);
2859 latch_init(&pmd->exit_latch);
2860 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2861 xpthread_cond_init(&pmd->cond, NULL);
2862 ovs_mutex_init(&pmd->cond_mutex);
2863 ovs_mutex_init(&pmd->flow_mutex);
2864 ovs_mutex_init(&pmd->poll_mutex);
2865 dpcls_init(&pmd->cls);
2866 cmap_init(&pmd->flow_table);
2867 list_init(&pmd->poll_list);
2868 /* init the 'flow_cache' since there is no
2869 * actual thread created for NON_PMD_CORE_ID. */
2870 if (core_id == NON_PMD_CORE_ID) {
2871 emc_cache_init(&pmd->flow_cache);
2873 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2874 hash_int(core_id, 0));
2878 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2880 dp_netdev_pmd_flow_flush(pmd);
2881 dpcls_destroy(&pmd->cls);
2882 cmap_destroy(&pmd->flow_table);
2883 ovs_mutex_destroy(&pmd->flow_mutex);
2884 latch_destroy(&pmd->exit_latch);
2885 xpthread_cond_destroy(&pmd->cond);
2886 ovs_mutex_destroy(&pmd->cond_mutex);
2887 ovs_mutex_destroy(&pmd->poll_mutex);
2891 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2892 * and unrefs the struct. */
2894 dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd)
2896 /* Uninit the 'flow_cache' since there is
2897 * no actual thread uninit it for NON_PMD_CORE_ID. */
2898 if (pmd->core_id == NON_PMD_CORE_ID) {
2899 emc_cache_uninit(&pmd->flow_cache);
2901 latch_set(&pmd->exit_latch);
2902 dp_netdev_reload_pmd__(pmd);
2903 ovs_numa_unpin_core(pmd->core_id);
2904 xpthread_join(pmd->thread, NULL);
2907 /* Unref all ports and free poll_list. */
2908 dp_netdev_pmd_clear_poll_list(pmd);
2910 /* Purges the 'pmd''s flows after stopping the thread, but before
2911 * destroying the flows, so that the flow stats can be collected. */
2912 if (dp->dp_purge_cb) {
2913 dp->dp_purge_cb(dp->dp_purge_aux, pmd->core_id);
2915 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2916 dp_netdev_pmd_unref(pmd);
2919 /* Destroys all pmd threads. */
2921 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2923 struct dp_netdev_pmd_thread *pmd;
2924 struct dp_netdev_pmd_thread **pmd_list;
2925 size_t k = 0, n_pmds;
2927 n_pmds = cmap_count(&dp->poll_threads);
2928 pmd_list = xcalloc(n_pmds, sizeof *pmd_list);
2930 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2931 /* We cannot call dp_netdev_del_pmd(), since it alters
2932 * 'dp->poll_threads' (while we're iterating it) and it
2934 ovs_assert(k < n_pmds);
2935 pmd_list[k++] = pmd;
2938 for (size_t i = 0; i < k; i++) {
2939 dp_netdev_del_pmd(dp, pmd_list[i]);
2944 /* Deletes all pmd threads on numa node 'numa_id' and
2945 * fixes tx_qids of other threads to keep them sequential. */
2947 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2949 struct dp_netdev_pmd_thread *pmd;
2950 int n_pmds_on_numa, n_pmds;
2951 int *free_idx, k = 0;
2952 struct dp_netdev_pmd_thread **pmd_list;
2954 n_pmds_on_numa = get_n_pmd_threads_on_numa(dp, numa_id);
2955 free_idx = xcalloc(n_pmds_on_numa, sizeof *free_idx);
2956 pmd_list = xcalloc(n_pmds_on_numa, sizeof *pmd_list);
2958 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2959 /* We cannot call dp_netdev_del_pmd(), since it alters
2960 * 'dp->poll_threads' (while we're iterating it) and it
2962 if (pmd->numa_id == numa_id) {
2963 atomic_read_relaxed(&pmd->tx_qid, &free_idx[k]);
2965 ovs_assert(k < n_pmds_on_numa);
2970 for (int i = 0; i < k; i++) {
2971 dp_netdev_del_pmd(dp, pmd_list[i]);
2974 n_pmds = get_n_pmd_threads(dp);
2975 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2978 atomic_read_relaxed(&pmd->tx_qid, &old_tx_qid);
2980 if (old_tx_qid >= n_pmds) {
2981 int new_tx_qid = free_idx[--k];
2983 atomic_store_relaxed(&pmd->tx_qid, new_tx_qid);
2991 /* Deletes all rx queues from pmd->poll_list. */
2993 dp_netdev_pmd_clear_poll_list(struct dp_netdev_pmd_thread *pmd)
2995 struct rxq_poll *poll;
2997 ovs_mutex_lock(&pmd->poll_mutex);
2998 LIST_FOR_EACH_POP (poll, node, &pmd->poll_list) {
2999 port_unref(poll->port);
3003 ovs_mutex_unlock(&pmd->poll_mutex);
3006 /* Deletes all rx queues of 'port' from poll_list of pmd thread and
3007 * reloads it if poll_list was changed. */
3009 dp_netdev_del_port_from_pmd(struct dp_netdev_port *port,
3010 struct dp_netdev_pmd_thread *pmd)
3012 struct rxq_poll *poll, *next;
3015 ovs_mutex_lock(&pmd->poll_mutex);
3016 LIST_FOR_EACH_SAFE (poll, next, node, &pmd->poll_list) {
3017 if (poll->port == port) {
3019 port_unref(poll->port);
3020 list_remove(&poll->node);
3025 ovs_mutex_unlock(&pmd->poll_mutex);
3027 dp_netdev_reload_pmd__(pmd);
3031 /* Deletes all rx queues of 'port' from all pmd threads of dp and
3032 * reloads them if needed. */
3034 dp_netdev_del_port_from_all_pmds(struct dp_netdev *dp,
3035 struct dp_netdev_port *port)
3037 int numa_id = netdev_get_numa_id(port->netdev);
3038 struct dp_netdev_pmd_thread *pmd;
3040 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3041 if (pmd->numa_id == numa_id) {
3042 dp_netdev_del_port_from_pmd(port, pmd);
3047 /* Returns PMD thread from this numa node with fewer rx queues to poll.
3048 * Returns NULL if there is no PMD threads on this numa node.
3049 * Can be called safely only by main thread. */
3050 static struct dp_netdev_pmd_thread *
3051 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev *dp, int numa_id)
3054 struct dp_netdev_pmd_thread *pmd, *res = NULL;
3056 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
3057 if (pmd->numa_id == numa_id
3058 && (min_cnt > pmd->poll_cnt || res == NULL)) {
3059 min_cnt = pmd->poll_cnt;
3067 /* Adds rx queue to poll_list of PMD thread. */
3069 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd,
3070 struct dp_netdev_port *port, struct netdev_rxq *rx)
3071 OVS_REQUIRES(pmd->poll_mutex)
3073 struct rxq_poll *poll = xmalloc(sizeof *poll);
3079 list_push_back(&pmd->poll_list, &poll->node);
3083 /* Distributes all rx queues of 'port' between all PMD threads and reloads
3084 * them if needed. */
3086 dp_netdev_add_port_to_pmds(struct dp_netdev *dp, struct dp_netdev_port *port)
3088 int numa_id = netdev_get_numa_id(port->netdev);
3089 struct dp_netdev_pmd_thread *pmd;
3090 struct hmapx to_reload;
3091 struct hmapx_node *node;
3094 hmapx_init(&to_reload);
3095 /* Cannot create pmd threads for invalid numa node. */
3096 ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
3098 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
3099 pmd = dp_netdev_less_loaded_pmd_on_numa(dp, numa_id);
3101 /* There is no pmd threads on this numa node. */
3102 dp_netdev_set_pmds_on_numa(dp, numa_id);
3103 /* Assigning of rx queues done. */
3107 ovs_mutex_lock(&pmd->poll_mutex);
3108 dp_netdev_add_rxq_to_pmd(pmd, port, port->rxq[i]);
3109 ovs_mutex_unlock(&pmd->poll_mutex);
3111 hmapx_add(&to_reload, pmd);
3114 HMAPX_FOR_EACH (node, &to_reload) {
3115 pmd = (struct dp_netdev_pmd_thread *) node->data;
3116 dp_netdev_reload_pmd__(pmd);
3119 hmapx_destroy(&to_reload);
3122 /* Checks the numa node id of 'netdev' and starts pmd threads for
3125 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
3129 if (!ovs_numa_numa_id_is_valid(numa_id)) {
3130 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
3131 "invalid", numa_id);
3135 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
3137 /* If there are already pmd threads created for the numa node
3138 * in which 'netdev' is on, do nothing. Else, creates the
3139 * pmd threads for the numa node. */
3141 int can_have, n_unpinned, i, index = 0;
3142 struct dp_netdev_pmd_thread **pmds;
3143 struct dp_netdev_port *port;
3145 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
3147 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
3148 "cores on numa node");
3152 /* If cpu mask is specified, uses all unpinned cores, otherwise
3153 * tries creating NR_PMD_THREADS pmd threads. */
3154 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
3155 pmds = xzalloc(can_have * sizeof *pmds);
3156 for (i = 0; i < can_have; i++) {
3157 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
3158 pmds[i] = xzalloc(sizeof **pmds);
3159 dp_netdev_configure_pmd(pmds[i], dp, i, core_id, numa_id);
3162 /* Distributes rx queues of this numa node between new pmd threads. */
3163 CMAP_FOR_EACH (port, node, &dp->ports) {
3164 if (netdev_is_pmd(port->netdev)
3165 && netdev_get_numa_id(port->netdev) == numa_id) {
3166 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
3167 /* Make thread-safety analyser happy. */
3168 ovs_mutex_lock(&pmds[index]->poll_mutex);
3169 dp_netdev_add_rxq_to_pmd(pmds[index], port, port->rxq[i]);
3170 ovs_mutex_unlock(&pmds[index]->poll_mutex);
3171 index = (index + 1) % can_have;
3176 /* Actual start of pmd threads. */
3177 for (i = 0; i < can_have; i++) {
3178 pmds[i]->thread = ovs_thread_create("pmd", pmd_thread_main, pmds[i]);
3181 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
3186 /* Called after pmd threads config change. Restarts pmd threads with
3187 * new configuration. */
3189 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
3191 struct dp_netdev_port *port;
3193 CMAP_FOR_EACH (port, node, &dp->ports) {
3194 if (netdev_is_pmd(port->netdev)) {
3195 int numa_id = netdev_get_numa_id(port->netdev);
3197 dp_netdev_set_pmds_on_numa(dp, numa_id);
3203 dpif_netdev_get_datapath_version(void)
3205 return xstrdup("<built-in>");
3209 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
3210 uint16_t tcp_flags, long long now)
3214 atomic_store_relaxed(&netdev_flow->stats.used, now);
3215 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
3216 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
3217 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
3219 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
3223 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
3224 enum dp_stat_type type, int cnt)
3226 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3230 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3231 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3232 enum dpif_upcall_type type, const struct nlattr *userdata,
3233 struct ofpbuf *actions, struct ofpbuf *put_actions)
3235 struct dp_netdev *dp = pmd->dp;
3236 struct flow_tnl orig_tunnel;
3239 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3243 /* Upcall processing expects the Geneve options to be in the translated
3244 * format but we need to retain the raw format for datapath use. */
3245 orig_tunnel.flags = flow->tunnel.flags;
3246 if (flow->tunnel.flags & FLOW_TNL_F_UDPIF) {
3247 orig_tunnel.metadata.present.len = flow->tunnel.metadata.present.len;
3248 memcpy(orig_tunnel.metadata.opts.gnv, flow->tunnel.metadata.opts.gnv,
3249 flow->tunnel.metadata.present.len);
3250 err = tun_metadata_from_geneve_udpif(&orig_tunnel, &orig_tunnel,
3257 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3258 struct ds ds = DS_EMPTY_INITIALIZER;
3261 struct odp_flow_key_parms odp_parms = {
3264 .odp_in_port = flow->in_port.odp_port,
3265 .support = dp_netdev_support,
3268 ofpbuf_init(&key, 0);
3269 odp_flow_key_from_flow(&odp_parms, &key);
3270 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3271 dp_packet_size(packet_));
3273 odp_flow_key_format(key.data, key.size, &ds);
3275 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3276 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3278 ofpbuf_uninit(&key);
3284 err = dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3285 actions, wc, put_actions, dp->upcall_aux);
3286 if (err && err != ENOSPC) {
3290 /* Translate tunnel metadata masks to datapath format. */
3292 if (wc->masks.tunnel.metadata.present.map) {
3293 struct geneve_opt opts[TLV_TOT_OPT_SIZE /
3294 sizeof(struct geneve_opt)];
3296 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3297 tun_metadata_to_geneve_udpif_mask(&flow->tunnel,
3299 orig_tunnel.metadata.opts.gnv,
3300 orig_tunnel.metadata.present.len,
3303 orig_tunnel.metadata.present.len = 0;
3306 memset(&wc->masks.tunnel.metadata, 0,
3307 sizeof wc->masks.tunnel.metadata);
3308 memcpy(&wc->masks.tunnel.metadata.opts.gnv, opts,
3309 orig_tunnel.metadata.present.len);
3311 wc->masks.tunnel.metadata.present.len = 0xff;
3314 /* Restore tunnel metadata. We need to use the saved options to ensure
3315 * that any unknown options are not lost. The generated mask will have
3316 * the same structure, matching on types and lengths but wildcarding
3317 * option data we don't care about. */
3318 if (orig_tunnel.flags & FLOW_TNL_F_UDPIF) {
3319 memcpy(&flow->tunnel.metadata.opts.gnv, orig_tunnel.metadata.opts.gnv,
3320 orig_tunnel.metadata.present.len);
3321 flow->tunnel.metadata.present.len = orig_tunnel.metadata.present.len;
3322 flow->tunnel.flags |= FLOW_TNL_F_UDPIF;
3328 static inline uint32_t
3329 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3330 const struct miniflow *mf)
3332 uint32_t hash, recirc_depth;
3334 if (OVS_LIKELY(dp_packet_rss_valid(packet))) {
3335 hash = dp_packet_get_rss_hash(packet);
3337 hash = miniflow_hash_5tuple(mf, 0);
3338 dp_packet_set_rss_hash(packet, hash);
3341 /* The RSS hash must account for the recirculation depth to avoid
3342 * collisions in the exact match cache */
3343 recirc_depth = *recirc_depth_get_unsafe();
3344 if (OVS_UNLIKELY(recirc_depth)) {
3345 hash = hash_finish(hash, recirc_depth);
3346 dp_packet_set_rss_hash(packet, hash);
3351 struct packet_batch {
3352 unsigned int packet_count;
3353 unsigned int byte_count;
3356 struct dp_netdev_flow *flow;
3358 struct dp_packet *packets[NETDEV_MAX_BURST];
3362 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
3363 const struct miniflow *mf)
3365 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3366 batch->packets[batch->packet_count++] = packet;
3367 batch->byte_count += dp_packet_size(packet);
3371 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
3373 flow->batch = batch;
3376 batch->packet_count = 0;
3377 batch->byte_count = 0;
3378 batch->tcp_flags = 0;
3382 packet_batch_execute(struct packet_batch *batch,
3383 struct dp_netdev_pmd_thread *pmd,
3386 struct dp_netdev_actions *actions;
3387 struct dp_netdev_flow *flow = batch->flow;
3389 dp_netdev_flow_used(flow, batch->packet_count, batch->byte_count,
3390 batch->tcp_flags, now);
3392 actions = dp_netdev_flow_get_actions(flow);
3394 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
3395 actions->actions, actions->size);
3399 dp_netdev_queue_batches(struct dp_packet *pkt,
3400 struct dp_netdev_flow *flow, const struct miniflow *mf,
3401 struct packet_batch *batches, size_t *n_batches)
3403 struct packet_batch *batch = flow->batch;
3405 if (OVS_UNLIKELY(!batch)) {
3406 batch = &batches[(*n_batches)++];
3407 packet_batch_init(batch, flow);
3410 packet_batch_update(batch, pkt, mf);
3413 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3414 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
3415 * miniflow is copied into 'keys' and the packet pointer is moved at the
3416 * beginning of the 'packets' array.
3418 * The function returns the number of packets that needs to be processed in the
3419 * 'packets' array (they have been moved to the beginning of the vector).
3421 * If 'md_is_valid' is false, the metadata in 'packets' is not valid and must be
3422 * initialized by this function using 'port_no'.
3424 static inline size_t
3425 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
3426 size_t cnt, struct netdev_flow_key *keys,
3427 struct packet_batch batches[], size_t *n_batches,
3428 bool md_is_valid, odp_port_t port_no)
3430 struct emc_cache *flow_cache = &pmd->flow_cache;
3431 struct netdev_flow_key *key = &keys[0];
3432 size_t i, n_missed = 0, n_dropped = 0;
3434 for (i = 0; i < cnt; i++) {
3435 struct dp_netdev_flow *flow;
3436 struct dp_packet *packet = packets[i];
3438 if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) {
3439 dp_packet_delete(packet);
3445 /* Prefetch next packet data and metadata. */
3446 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3447 pkt_metadata_prefetch_init(&packets[i+1]->md);
3451 pkt_metadata_init(&packet->md, port_no);
3453 miniflow_extract(packet, &key->mf);
3454 key->len = 0; /* Not computed yet. */
3455 key->hash = dpif_netdev_packet_get_rss_hash(packet, &key->mf);
3457 flow = emc_lookup(flow_cache, key);
3458 if (OVS_LIKELY(flow)) {
3459 dp_netdev_queue_batches(packet, flow, &key->mf, batches,
3462 /* Exact match cache missed. Group missed packets together at
3463 * the beginning of the 'packets' array. */
3464 packets[n_missed] = packet;
3465 /* 'key[n_missed]' contains the key of the current packet and it
3466 * must be returned to the caller. The next key should be extracted
3467 * to 'keys[n_missed + 1]'. */
3468 key = &keys[++n_missed];
3472 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - n_dropped - n_missed);
3478 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3479 struct dp_packet **packets, size_t cnt,
3480 struct netdev_flow_key *keys,
3481 struct packet_batch batches[], size_t *n_batches)
3483 #if !defined(__CHECKER__) && !defined(_WIN32)
3484 const size_t PKT_ARRAY_SIZE = cnt;
3486 /* Sparse or MSVC doesn't like variable length array. */
3487 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3489 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3490 struct dp_netdev *dp = pmd->dp;
3491 struct emc_cache *flow_cache = &pmd->flow_cache;
3492 int miss_cnt = 0, lost_cnt = 0;
3496 for (i = 0; i < cnt; i++) {
3497 /* Key length is needed in all the cases, hash computed on demand. */
3498 keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf));
3500 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3501 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3502 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3503 struct ofpbuf actions, put_actions;
3506 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3507 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3509 for (i = 0; i < cnt; i++) {
3510 struct dp_netdev_flow *netdev_flow;
3511 struct ofpbuf *add_actions;
3515 if (OVS_LIKELY(rules[i])) {
3519 /* It's possible that an earlier slow path execution installed
3520 * a rule covering this flow. In this case, it's a lot cheaper
3521 * to catch it here than execute a miss. */
3522 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3524 rules[i] = &netdev_flow->cr;
3530 match.tun_md.valid = false;
3531 miniflow_expand(&keys[i].mf, &match.flow);
3533 ofpbuf_clear(&actions);
3534 ofpbuf_clear(&put_actions);
3536 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
3537 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
3538 &ufid, DPIF_UC_MISS, NULL, &actions,
3540 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3541 dp_packet_delete(packets[i]);
3546 /* The Netlink encoding of datapath flow keys cannot express
3547 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3548 * tag is interpreted as exact match on the fact that there is no
3549 * VLAN. Unless we refactor a lot of code that translates between
3550 * Netlink and struct flow representations, we have to do the same
3552 if (!match.wc.masks.vlan_tci) {
3553 match.wc.masks.vlan_tci = htons(0xffff);
3556 /* We can't allow the packet batching in the next loop to execute
3557 * the actions. Otherwise, if there are any slow path actions,
3558 * we'll send the packet up twice. */
3559 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
3560 actions.data, actions.size);
3562 add_actions = put_actions.size ? &put_actions : &actions;
3563 if (OVS_LIKELY(error != ENOSPC)) {
3564 /* XXX: There's a race window where a flow covering this packet
3565 * could have already been installed since we last did the flow
3566 * lookup before upcall. This could be solved by moving the
3567 * mutex lock outside the loop, but that's an awful long time
3568 * to be locking everyone out of making flow installs. If we
3569 * move to a per-core classifier, it would be reasonable. */
3570 ovs_mutex_lock(&pmd->flow_mutex);
3571 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3572 if (OVS_LIKELY(!netdev_flow)) {
3573 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3577 ovs_mutex_unlock(&pmd->flow_mutex);
3579 emc_insert(flow_cache, &keys[i], netdev_flow);
3583 ofpbuf_uninit(&actions);
3584 ofpbuf_uninit(&put_actions);
3585 fat_rwlock_unlock(&dp->upcall_rwlock);
3586 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3587 } else if (OVS_UNLIKELY(any_miss)) {
3588 for (i = 0; i < cnt; i++) {
3589 if (OVS_UNLIKELY(!rules[i])) {
3590 dp_packet_delete(packets[i]);
3597 for (i = 0; i < cnt; i++) {
3598 struct dp_packet *packet = packets[i];
3599 struct dp_netdev_flow *flow;
3601 if (OVS_UNLIKELY(!rules[i])) {
3605 flow = dp_netdev_flow_cast(rules[i]);
3607 emc_insert(flow_cache, &keys[i], flow);
3608 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3611 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3612 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3613 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3616 /* Packets enter the datapath from a port (or from recirculation) here.
3618 * For performance reasons a caller may choose not to initialize the metadata
3619 * in 'packets': in this case 'mdinit' is false and this function needs to
3620 * initialize it using 'port_no'. If the metadata in 'packets' is already
3621 * valid, 'md_is_valid' must be true and 'port_no' will be ignored. */
3623 dp_netdev_input__(struct dp_netdev_pmd_thread *pmd,
3624 struct dp_packet **packets, int cnt,
3625 bool md_is_valid, odp_port_t port_no)
3627 #if !defined(__CHECKER__) && !defined(_WIN32)
3628 const size_t PKT_ARRAY_SIZE = cnt;
3630 /* Sparse or MSVC doesn't like variable length array. */
3631 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3633 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3634 struct packet_batch batches[PKT_ARRAY_SIZE];
3635 long long now = time_msec();
3636 size_t newcnt, n_batches, i;
3639 newcnt = emc_processing(pmd, packets, cnt, keys, batches, &n_batches,
3640 md_is_valid, port_no);
3641 if (OVS_UNLIKELY(newcnt)) {
3642 fast_path_processing(pmd, packets, newcnt, keys, batches, &n_batches);
3645 for (i = 0; i < n_batches; i++) {
3646 batches[i].flow->batch = NULL;
3649 for (i = 0; i < n_batches; i++) {
3650 packet_batch_execute(&batches[i], pmd, now);
3655 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3656 struct dp_packet **packets, int cnt,
3659 dp_netdev_input__(pmd, packets, cnt, false, port_no);
3663 dp_netdev_recirculate(struct dp_netdev_pmd_thread *pmd,
3664 struct dp_packet **packets, int cnt)
3666 dp_netdev_input__(pmd, packets, cnt, true, 0);
3669 struct dp_netdev_execute_aux {
3670 struct dp_netdev_pmd_thread *pmd;
3674 dpif_netdev_register_dp_purge_cb(struct dpif *dpif, dp_purge_callback *cb,
3677 struct dp_netdev *dp = get_dp_netdev(dpif);
3678 dp->dp_purge_aux = aux;
3679 dp->dp_purge_cb = cb;
3683 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3686 struct dp_netdev *dp = get_dp_netdev(dpif);
3687 dp->upcall_aux = aux;
3692 dp_netdev_drop_packets(struct dp_packet **packets, int cnt, bool may_steal)
3697 for (i = 0; i < cnt; i++) {
3698 dp_packet_delete(packets[i]);
3704 push_tnl_action(const struct dp_netdev *dp,
3705 const struct nlattr *attr,
3706 struct dp_packet **packets, int cnt)
3708 struct dp_netdev_port *tun_port;
3709 const struct ovs_action_push_tnl *data;
3711 data = nl_attr_get(attr);
3713 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3717 netdev_push_header(tun_port->netdev, packets, cnt, data);
3723 dp_netdev_clone_pkt_batch(struct dp_packet **dst_pkts,
3724 struct dp_packet **src_pkts, int cnt)
3728 for (i = 0; i < cnt; i++) {
3729 dst_pkts[i] = dp_packet_clone(src_pkts[i]);
3734 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3735 const struct nlattr *a, bool may_steal)
3736 OVS_NO_THREAD_SAFETY_ANALYSIS
3738 struct dp_netdev_execute_aux *aux = aux_;
3739 uint32_t *depth = recirc_depth_get();
3740 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3741 struct dp_netdev *dp = pmd->dp;
3742 int type = nl_attr_type(a);
3743 struct dp_netdev_port *p;
3746 switch ((enum ovs_action_attr)type) {
3747 case OVS_ACTION_ATTR_OUTPUT:
3748 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3749 if (OVS_LIKELY(p)) {
3752 atomic_read_relaxed(&pmd->tx_qid, &tx_qid);
3754 netdev_send(p->netdev, tx_qid, packets, cnt, may_steal);
3759 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3760 if (*depth < MAX_RECIRC_DEPTH) {
3761 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3765 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3769 err = push_tnl_action(dp, a, packets, cnt);
3772 dp_netdev_recirculate(pmd, packets, cnt);
3775 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3781 case OVS_ACTION_ATTR_TUNNEL_POP:
3782 if (*depth < MAX_RECIRC_DEPTH) {
3783 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3785 p = dp_netdev_lookup_port(dp, portno);
3787 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3791 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3795 err = netdev_pop_header(p->netdev, packets, cnt);
3798 for (i = 0; i < cnt; i++) {
3799 packets[i]->md.in_port.odp_port = portno;
3803 dp_netdev_recirculate(pmd, packets, cnt);
3806 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3813 case OVS_ACTION_ATTR_USERSPACE:
3814 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3815 const struct nlattr *userdata;
3816 struct ofpbuf actions;
3820 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3821 ofpbuf_init(&actions, 0);
3823 for (i = 0; i < cnt; i++) {
3826 ofpbuf_clear(&actions);
3828 flow_extract(packets[i], &flow);
3829 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3830 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3831 DPIF_UC_ACTION, userdata,&actions,
3833 if (!error || error == ENOSPC) {
3834 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3835 actions.data, actions.size);
3836 } else if (may_steal) {
3837 dp_packet_delete(packets[i]);
3840 ofpbuf_uninit(&actions);
3841 fat_rwlock_unlock(&dp->upcall_rwlock);
3847 case OVS_ACTION_ATTR_RECIRC:
3848 if (*depth < MAX_RECIRC_DEPTH) {
3849 struct dp_packet *recirc_pkts[NETDEV_MAX_BURST];
3852 dp_netdev_clone_pkt_batch(recirc_pkts, packets, cnt);
3853 packets = recirc_pkts;
3856 for (i = 0; i < cnt; i++) {
3857 packets[i]->md.recirc_id = nl_attr_get_u32(a);
3861 dp_netdev_recirculate(pmd, packets, cnt);
3867 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3870 case OVS_ACTION_ATTR_CT:
3871 /* If a flow with this action is slow-pathed, datapath assistance is
3872 * required to implement it. However, we don't support this action
3873 * in the userspace datapath. */
3874 VLOG_WARN("Cannot execute conntrack action in userspace.");
3877 case OVS_ACTION_ATTR_PUSH_VLAN:
3878 case OVS_ACTION_ATTR_POP_VLAN:
3879 case OVS_ACTION_ATTR_PUSH_MPLS:
3880 case OVS_ACTION_ATTR_POP_MPLS:
3881 case OVS_ACTION_ATTR_SET:
3882 case OVS_ACTION_ATTR_SET_MASKED:
3883 case OVS_ACTION_ATTR_SAMPLE:
3884 case OVS_ACTION_ATTR_HASH:
3885 case OVS_ACTION_ATTR_UNSPEC:
3886 case __OVS_ACTION_ATTR_MAX:
3890 dp_netdev_drop_packets(packets, cnt, may_steal);
3894 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3895 struct dp_packet **packets, int cnt,
3897 const struct nlattr *actions, size_t actions_len)
3899 struct dp_netdev_execute_aux aux = { pmd };
3901 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3902 actions_len, dp_execute_cb);
3905 const struct dpif_class dpif_netdev_class = {
3908 dpif_netdev_enumerate,
3909 dpif_netdev_port_open_type,
3912 dpif_netdev_destroy,
3915 dpif_netdev_get_stats,
3916 dpif_netdev_port_add,
3917 dpif_netdev_port_del,
3918 dpif_netdev_port_query_by_number,
3919 dpif_netdev_port_query_by_name,
3920 NULL, /* port_get_pid */
3921 dpif_netdev_port_dump_start,
3922 dpif_netdev_port_dump_next,
3923 dpif_netdev_port_dump_done,
3924 dpif_netdev_port_poll,
3925 dpif_netdev_port_poll_wait,
3926 dpif_netdev_flow_flush,
3927 dpif_netdev_flow_dump_create,
3928 dpif_netdev_flow_dump_destroy,
3929 dpif_netdev_flow_dump_thread_create,
3930 dpif_netdev_flow_dump_thread_destroy,
3931 dpif_netdev_flow_dump_next,
3932 dpif_netdev_operate,
3933 NULL, /* recv_set */
3934 NULL, /* handlers_set */
3935 dpif_netdev_pmd_set,
3936 dpif_netdev_queue_to_priority,
3938 NULL, /* recv_wait */
3939 NULL, /* recv_purge */
3940 dpif_netdev_register_dp_purge_cb,
3941 dpif_netdev_register_upcall_cb,
3942 dpif_netdev_enable_upcall,
3943 dpif_netdev_disable_upcall,
3944 dpif_netdev_get_datapath_version,
3945 NULL, /* ct_dump_start */
3946 NULL, /* ct_dump_next */
3947 NULL, /* ct_dump_done */
3948 NULL, /* ct_flush */
3952 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3953 const char *argv[], void *aux OVS_UNUSED)
3955 struct dp_netdev_port *old_port;
3956 struct dp_netdev_port *new_port;
3957 struct dp_netdev *dp;
3960 ovs_mutex_lock(&dp_netdev_mutex);
3961 dp = shash_find_data(&dp_netdevs, argv[1]);
3962 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3963 ovs_mutex_unlock(&dp_netdev_mutex);
3964 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3967 ovs_refcount_ref(&dp->ref_cnt);
3968 ovs_mutex_unlock(&dp_netdev_mutex);
3970 ovs_mutex_lock(&dp->port_mutex);
3971 if (get_port_by_name(dp, argv[2], &old_port)) {
3972 unixctl_command_reply_error(conn, "unknown port");
3976 port_no = u32_to_odp(atoi(argv[3]));
3977 if (!port_no || port_no == ODPP_NONE) {
3978 unixctl_command_reply_error(conn, "bad port number");
3981 if (dp_netdev_lookup_port(dp, port_no)) {
3982 unixctl_command_reply_error(conn, "port number already in use");
3986 /* Remove old port. */
3987 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3988 ovsrcu_postpone(free, old_port);
3990 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3991 new_port = xmemdup(old_port, sizeof *old_port);
3992 new_port->port_no = port_no;
3993 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3995 seq_change(dp->port_seq);
3996 unixctl_command_reply(conn, NULL);
3999 ovs_mutex_unlock(&dp->port_mutex);
4000 dp_netdev_unref(dp);
4004 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
4005 const char *argv[], void *aux OVS_UNUSED)
4007 struct dp_netdev_port *port;
4008 struct dp_netdev *dp;
4010 ovs_mutex_lock(&dp_netdev_mutex);
4011 dp = shash_find_data(&dp_netdevs, argv[1]);
4012 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
4013 ovs_mutex_unlock(&dp_netdev_mutex);
4014 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
4017 ovs_refcount_ref(&dp->ref_cnt);
4018 ovs_mutex_unlock(&dp_netdev_mutex);
4020 ovs_mutex_lock(&dp->port_mutex);
4021 if (get_port_by_name(dp, argv[2], &port)) {
4022 unixctl_command_reply_error(conn, "unknown port");
4023 } else if (port->port_no == ODPP_LOCAL) {
4024 unixctl_command_reply_error(conn, "can't delete local port");
4026 do_del_port(dp, port);
4027 unixctl_command_reply(conn, NULL);
4029 ovs_mutex_unlock(&dp->port_mutex);
4031 dp_netdev_unref(dp);
4035 dpif_dummy_register__(const char *type)
4037 struct dpif_class *class;
4039 class = xmalloc(sizeof *class);
4040 *class = dpif_netdev_class;
4041 class->type = xstrdup(type);
4042 dp_register_provider(class);
4046 dpif_dummy_override(const char *type)
4051 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
4052 * a userland-only build. It's useful for testsuite.
4054 error = dp_unregister_provider(type);
4055 if (error == 0 || error == EAFNOSUPPORT) {
4056 dpif_dummy_register__(type);
4061 dpif_dummy_register(enum dummy_level level)
4063 if (level == DUMMY_OVERRIDE_ALL) {
4068 dp_enumerate_types(&types);
4069 SSET_FOR_EACH (type, &types) {
4070 dpif_dummy_override(type);
4072 sset_destroy(&types);
4073 } else if (level == DUMMY_OVERRIDE_SYSTEM) {
4074 dpif_dummy_override("system");
4077 dpif_dummy_register__("dummy");
4079 unixctl_command_register("dpif-dummy/change-port-number",
4080 "dp port new-number",
4081 3, 3, dpif_dummy_change_port_number, NULL);
4082 unixctl_command_register("dpif-dummy/delete-port", "dp port",
4083 2, 2, dpif_dummy_delete_port, NULL);
4086 /* Datapath Classifier. */
4088 /* A set of rules that all have the same fields wildcarded. */
4089 struct dpcls_subtable {
4090 /* The fields are only used by writers. */
4091 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
4093 /* These fields are accessed by readers. */
4094 struct cmap rules; /* Contains "struct dpcls_rule"s. */
4095 struct netdev_flow_key mask; /* Wildcards for fields (const). */
4096 /* 'mask' must be the last field, additional space is allocated here. */
4099 /* Initializes 'cls' as a classifier that initially contains no classification
4102 dpcls_init(struct dpcls *cls)
4104 cmap_init(&cls->subtables_map);
4105 pvector_init(&cls->subtables);
4109 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
4111 pvector_remove(&cls->subtables, subtable);
4112 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
4113 subtable->mask.hash);
4114 cmap_destroy(&subtable->rules);
4115 ovsrcu_postpone(free, subtable);
4118 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
4119 * caller's responsibility.
4120 * May only be called after all the readers have been terminated. */
4122 dpcls_destroy(struct dpcls *cls)
4125 struct dpcls_subtable *subtable;
4127 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
4128 ovs_assert(cmap_count(&subtable->rules) == 0);
4129 dpcls_destroy_subtable(cls, subtable);
4131 cmap_destroy(&cls->subtables_map);
4132 pvector_destroy(&cls->subtables);
4136 static struct dpcls_subtable *
4137 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
4139 struct dpcls_subtable *subtable;
4141 /* Need to add one. */
4142 subtable = xmalloc(sizeof *subtable
4143 - sizeof subtable->mask.mf + mask->len);
4144 cmap_init(&subtable->rules);
4145 netdev_flow_key_clone(&subtable->mask, mask);
4146 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
4147 pvector_insert(&cls->subtables, subtable, 0);
4148 pvector_publish(&cls->subtables);
4153 static inline struct dpcls_subtable *
4154 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
4156 struct dpcls_subtable *subtable;
4158 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
4159 &cls->subtables_map) {
4160 if (netdev_flow_key_equal(&subtable->mask, mask)) {
4164 return dpcls_create_subtable(cls, mask);
4167 /* Insert 'rule' into 'cls'. */
4169 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
4170 const struct netdev_flow_key *mask)
4172 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
4174 rule->mask = &subtable->mask;
4175 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
4178 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
4180 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
4182 struct dpcls_subtable *subtable;
4184 ovs_assert(rule->mask);
4186 INIT_CONTAINER(subtable, rule->mask, mask);
4188 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
4190 dpcls_destroy_subtable(cls, subtable);
4191 pvector_publish(&cls->subtables);
4195 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
4196 * in 'mask' the values in 'key' and 'target' are the same. */
4198 dpcls_rule_matches_key(const struct dpcls_rule *rule,
4199 const struct netdev_flow_key *target)
4201 const uint64_t *keyp = miniflow_get_values(&rule->flow.mf);
4202 const uint64_t *maskp = miniflow_get_values(&rule->mask->mf);
4205 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, target, rule->flow.mf.map) {
4206 if (OVS_UNLIKELY((value & *maskp++) != *keyp++)) {
4213 /* For each miniflow in 'flows' performs a classifier lookup writing the result
4214 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
4215 * NULL it is skipped.
4217 * This function is optimized for use in the userspace datapath and therefore
4218 * does not implement a lot of features available in the standard
4219 * classifier_lookup() function. Specifically, it does not implement
4220 * priorities, instead returning any rule which matches the flow.
4222 * Returns true if all flows found a corresponding rule. */
4224 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
4225 struct dpcls_rule **rules, const size_t cnt)
4227 /* The batch size 16 was experimentally found faster than 8 or 32. */
4228 typedef uint16_t map_type;
4229 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
4231 #if !defined(__CHECKER__) && !defined(_WIN32)
4232 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
4234 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
4236 map_type maps[N_MAPS];
4237 struct dpcls_subtable *subtable;
4239 memset(maps, 0xff, sizeof maps);
4240 if (cnt % MAP_BITS) {
4241 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
4243 memset(rules, 0, cnt * sizeof *rules);
4245 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
4246 const struct netdev_flow_key *mkeys = keys;
4247 struct dpcls_rule **mrules = rules;
4248 map_type remains = 0;
4251 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
4253 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
4254 uint32_t hashes[MAP_BITS];
4255 const struct cmap_node *nodes[MAP_BITS];
4256 unsigned long map = maps[m];
4260 continue; /* Skip empty maps. */
4263 /* Compute hashes for the remaining keys. */
4264 ULLONG_FOR_EACH_1(i, map) {
4265 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
4269 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
4270 /* Check results. */
4271 ULLONG_FOR_EACH_1(i, map) {
4272 struct dpcls_rule *rule;
4274 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
4275 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
4280 ULLONG_SET0(map, i); /* Did not match. */
4282 ; /* Keep Sparse happy. */
4284 maps[m] &= ~map; /* Clear the found rules. */
4288 return true; /* All found. */
4291 return false; /* Some misses. */