1 /* Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
3 * Licensed under the Apache License, Version 2.0 (the "License");
4 * you may not use this file except in compliance with the License.
5 * You may obtain a copy of the License at:
7 * http://www.apache.org/licenses/LICENSE-2.0
9 * Unless required by applicable law or agreed to in writing, software
10 * distributed under the License is distributed on an "AS IS" BASIS,
11 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 * See the License for the specific language governing permissions and
13 * limitations under the License. */
16 #include "ofproto-dpif-upcall.h"
25 #include "dynamic-string.h"
26 #include "fail-open.h"
27 #include "guarded-list.h"
32 #include "ofproto-dpif-ipfix.h"
33 #include "ofproto-dpif-sflow.h"
34 #include "ofproto-dpif-xlate.h"
37 #include "poll-loop.h"
42 #define MAX_QUEUE_LENGTH 512
43 #define UPCALL_MAX_BATCH 64
44 #define REVALIDATE_MAX_BATCH 50
46 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall);
48 COVERAGE_DEFINE(dumped_duplicate_flow);
49 COVERAGE_DEFINE(dumped_new_flow);
50 COVERAGE_DEFINE(revalidate_missed_dp_flow);
52 /* A thread that reads upcalls from dpif, forwards each upcall's packet,
53 * and possibly sets up a kernel flow as a cache. */
55 struct udpif *udpif; /* Parent udpif. */
56 pthread_t thread; /* Thread ID. */
57 uint32_t handler_id; /* Handler id. */
60 /* A thread that processes datapath flows, updates OpenFlow statistics, and
61 * updates or removes them if necessary. */
63 struct udpif *udpif; /* Parent udpif. */
64 pthread_t thread; /* Thread ID. */
65 unsigned int id; /* ovsthread_id_self(). */
66 struct hmap *ukeys; /* Points into udpif->ukeys for this
67 revalidator. Used for GC phase. */
70 /* An upcall handler for ofproto_dpif.
72 * udpif keeps records of two kind of logically separate units:
77 * - An array of 'struct handler's for upcall handling and flow
83 * - Revalidation threads which read the datapath flow table and maintains
87 struct list list_node; /* In all_udpifs list. */
89 struct dpif *dpif; /* Datapath handle. */
90 struct dpif_backer *backer; /* Opaque dpif_backer pointer. */
92 uint32_t secret; /* Random seed for upcall hash. */
94 struct handler *handlers; /* Upcall handlers. */
97 struct revalidator *revalidators; /* Flow revalidators. */
98 size_t n_revalidators;
100 struct latch exit_latch; /* Tells child threads to exit. */
103 struct seq *reval_seq; /* Incremented to force revalidation. */
104 bool need_revalidate; /* As indicated by 'reval_seq'. */
105 bool reval_exit; /* Set by leader on 'exit_latch. */
106 struct ovs_barrier reval_barrier; /* Barrier used by revalidators. */
107 struct dpif_flow_dump *dump; /* DPIF flow dump state. */
108 long long int dump_duration; /* Duration of the last flow dump. */
109 struct seq *dump_seq; /* Increments each dump iteration. */
111 /* There are 'n_revalidators' ukey hmaps. Each revalidator retains a
112 * reference to one of these for garbage collection.
114 * During the flow dump phase, revalidators insert into these with a random
115 * distribution. During the garbage collection phase, each revalidator
116 * takes care of garbage collecting one of these hmaps. */
118 struct ovs_mutex mutex; /* Guards the following. */
119 struct hmap hmap OVS_GUARDED; /* Datapath flow keys. */
122 /* Datapath flow statistics. */
123 unsigned int max_n_flows;
124 unsigned int avg_n_flows;
126 /* Following fields are accessed and modified by different threads. */
127 atomic_uint flow_limit; /* Datapath flow hard limit. */
129 /* n_flows_mutex prevents multiple threads updating these concurrently. */
130 atomic_uint n_flows; /* Number of flows in the datapath. */
131 atomic_llong n_flows_timestamp; /* Last time n_flows was updated. */
132 struct ovs_mutex n_flows_mutex;
134 /* Following fields are accessed and modified only from the main thread. */
135 struct unixctl_conn **conns; /* Connections waiting on dump_seq. */
136 uint64_t conn_seq; /* Corresponds to 'dump_seq' when
137 conns[n_conns-1] was stored. */
138 size_t n_conns; /* Number of connections waiting. */
142 BAD_UPCALL, /* Some kind of bug somewhere. */
143 MISS_UPCALL, /* A flow miss. */
144 SFLOW_UPCALL, /* sFlow sample. */
145 FLOW_SAMPLE_UPCALL, /* Per-flow sampling. */
146 IPFIX_UPCALL /* Per-bridge sampling. */
150 struct ofproto_dpif *ofproto; /* Parent ofproto. */
152 /* The flow and packet are only required to be constant when using
153 * dpif-netdev. If a modification is absolutely necessary, a const cast
154 * may be used with other datapaths. */
155 const struct flow *flow; /* Parsed representation of the packet. */
156 const struct ofpbuf *packet; /* Packet associated with this upcall. */
157 ofp_port_t in_port; /* OpenFlow in port, or OFPP_NONE. */
159 enum dpif_upcall_type type; /* Datapath type of the upcall. */
160 const struct nlattr *userdata; /* Userdata for DPIF_UC_ACTION Upcalls. */
162 bool xout_initialized; /* True if 'xout' must be uninitialized. */
163 struct xlate_out xout; /* Result of xlate_actions(). */
164 struct ofpbuf put_actions; /* Actions 'put' in the fastapath. */
166 struct dpif_ipfix *ipfix; /* IPFIX pointer or NULL. */
167 struct dpif_sflow *sflow; /* SFlow pointer or NULL. */
169 bool vsp_adjusted; /* 'packet' and 'flow' were adjusted for
170 VLAN splinters if true. */
172 /* Not used by the upcall callback interface. */
173 const struct nlattr *key; /* Datapath flow key. */
174 size_t key_len; /* Datapath flow key length. */
175 const struct nlattr *out_tun_key; /* Datapath output tunnel key. */
178 /* 'udpif_key's are responsible for tracking the little bit of state udpif
179 * needs to do flow expiration which can't be pulled directly from the
180 * datapath. They may be created or maintained by any revalidator during
181 * the dump phase, but are owned by a single revalidator, and are destroyed
182 * by that revalidator during the garbage-collection phase.
184 * While some elements of a udpif_key are protected by a mutex, the ukey itself
185 * is not. Therefore it is not safe to destroy a udpif_key except when all
186 * revalidators are in garbage collection phase, or they aren't running. */
188 struct hmap_node hmap_node; /* In parent revalidator 'ukeys' map. */
190 /* These elements are read only once created, and therefore aren't
191 * protected by a mutex. */
192 const struct nlattr *key; /* Datapath flow key. */
193 size_t key_len; /* Length of 'key'. */
195 struct ovs_mutex mutex; /* Guards the following. */
196 struct dpif_flow_stats stats OVS_GUARDED; /* Last known stats.*/
197 long long int created OVS_GUARDED; /* Estimate of creation time. */
198 uint64_t dump_seq OVS_GUARDED; /* Tracks udpif->dump_seq. */
199 bool flow_exists OVS_GUARDED; /* Ensures flows are only deleted
202 struct xlate_cache *xcache OVS_GUARDED; /* Cache for xlate entries that
203 * are affected by this ukey.
204 * Used for stats and learning.*/
206 struct odputil_keybuf key_buf; /* Memory for 'key'. */
207 struct nlattr key_buf_nla;
211 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
212 static struct list all_udpifs = LIST_INITIALIZER(&all_udpifs);
214 static size_t recv_upcalls(struct handler *);
215 static int process_upcall(struct udpif *, struct upcall *,
216 struct ofpbuf *odp_actions);
217 static void handle_upcalls(struct udpif *, struct upcall *, size_t n_upcalls);
218 static void udpif_stop_threads(struct udpif *);
219 static void udpif_start_threads(struct udpif *, size_t n_handlers,
220 size_t n_revalidators);
221 static void *udpif_upcall_handler(void *);
222 static void *udpif_revalidator(void *);
223 static unsigned long udpif_get_n_flows(struct udpif *);
224 static void revalidate(struct revalidator *);
225 static void revalidator_sweep(struct revalidator *);
226 static void revalidator_purge(struct revalidator *);
227 static void upcall_unixctl_show(struct unixctl_conn *conn, int argc,
228 const char *argv[], void *aux);
229 static void upcall_unixctl_disable_megaflows(struct unixctl_conn *, int argc,
230 const char *argv[], void *aux);
231 static void upcall_unixctl_enable_megaflows(struct unixctl_conn *, int argc,
232 const char *argv[], void *aux);
233 static void upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, int argc,
234 const char *argv[], void *aux);
235 static void upcall_unixctl_dump_wait(struct unixctl_conn *conn, int argc,
236 const char *argv[], void *aux);
238 static struct udpif_key *ukey_create(const struct nlattr *key, size_t key_len,
240 static struct udpif_key *ukey_lookup(struct udpif *udpif,
241 const struct nlattr *key, size_t key_len,
243 static bool ukey_acquire(struct udpif *udpif, const struct nlattr *key,
244 size_t key_len, long long int used,
245 struct udpif_key **result);
246 static void ukey_delete(struct revalidator *, struct udpif_key *);
247 static enum upcall_type classify_upcall(enum dpif_upcall_type type,
248 const struct nlattr *userdata);
250 static int upcall_receive(struct upcall *, const struct dpif_backer *,
251 const struct ofpbuf *packet, enum dpif_upcall_type,
252 const struct nlattr *userdata, const struct flow *);
253 static void upcall_uninit(struct upcall *);
255 static upcall_callback upcall_cb;
257 static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true);
260 udpif_create(struct dpif_backer *backer, struct dpif *dpif)
262 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
263 struct udpif *udpif = xzalloc(sizeof *udpif);
265 if (ovsthread_once_start(&once)) {
266 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show,
268 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
269 upcall_unixctl_disable_megaflows, NULL);
270 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
271 upcall_unixctl_enable_megaflows, NULL);
272 unixctl_command_register("upcall/set-flow-limit", "", 1, 1,
273 upcall_unixctl_set_flow_limit, NULL);
274 unixctl_command_register("revalidator/wait", "", 0, 0,
275 upcall_unixctl_dump_wait, NULL);
276 ovsthread_once_done(&once);
280 udpif->backer = backer;
281 atomic_init(&udpif->flow_limit, MIN(ofproto_flow_limit, 10000));
282 udpif->secret = random_uint32();
283 udpif->reval_seq = seq_create();
284 udpif->dump_seq = seq_create();
285 latch_init(&udpif->exit_latch);
286 list_push_back(&all_udpifs, &udpif->list_node);
287 atomic_init(&udpif->n_flows, 0);
288 atomic_init(&udpif->n_flows_timestamp, LLONG_MIN);
289 ovs_mutex_init(&udpif->n_flows_mutex);
291 dpif_register_upcall_cb(dpif, upcall_cb, udpif);
297 udpif_run(struct udpif *udpif)
299 if (udpif->conns && udpif->conn_seq != seq_read(udpif->dump_seq)) {
302 for (i = 0; i < udpif->n_conns; i++) {
303 unixctl_command_reply(udpif->conns[i], NULL);
312 udpif_destroy(struct udpif *udpif)
314 udpif_stop_threads(udpif);
316 list_remove(&udpif->list_node);
317 latch_destroy(&udpif->exit_latch);
318 seq_destroy(udpif->reval_seq);
319 seq_destroy(udpif->dump_seq);
320 ovs_mutex_destroy(&udpif->n_flows_mutex);
324 /* Stops the handler and revalidator threads, must be enclosed in
325 * ovsrcu quiescent state unless when destroying udpif. */
327 udpif_stop_threads(struct udpif *udpif)
329 if (udpif && (udpif->n_handlers != 0 || udpif->n_revalidators != 0)) {
332 latch_set(&udpif->exit_latch);
334 for (i = 0; i < udpif->n_handlers; i++) {
335 struct handler *handler = &udpif->handlers[i];
337 xpthread_join(handler->thread, NULL);
340 for (i = 0; i < udpif->n_revalidators; i++) {
341 xpthread_join(udpif->revalidators[i].thread, NULL);
344 dpif_disable_upcall(udpif->dpif);
346 for (i = 0; i < udpif->n_revalidators; i++) {
347 struct revalidator *revalidator = &udpif->revalidators[i];
349 /* Delete ukeys, and delete all flows from the datapath to prevent
350 * double-counting stats. */
351 revalidator_purge(revalidator);
353 hmap_destroy(&udpif->ukeys[i].hmap);
354 ovs_mutex_destroy(&udpif->ukeys[i].mutex);
357 latch_poll(&udpif->exit_latch);
359 ovs_barrier_destroy(&udpif->reval_barrier);
361 free(udpif->revalidators);
362 udpif->revalidators = NULL;
363 udpif->n_revalidators = 0;
365 free(udpif->handlers);
366 udpif->handlers = NULL;
367 udpif->n_handlers = 0;
374 /* Starts the handler and revalidator threads, must be enclosed in
375 * ovsrcu quiescent state. */
377 udpif_start_threads(struct udpif *udpif, size_t n_handlers,
378 size_t n_revalidators)
380 if (udpif && n_handlers && n_revalidators) {
383 udpif->n_handlers = n_handlers;
384 udpif->n_revalidators = n_revalidators;
386 udpif->handlers = xzalloc(udpif->n_handlers * sizeof *udpif->handlers);
387 for (i = 0; i < udpif->n_handlers; i++) {
388 struct handler *handler = &udpif->handlers[i];
390 handler->udpif = udpif;
391 handler->handler_id = i;
392 handler->thread = ovs_thread_create(
393 "handler", udpif_upcall_handler, handler);
396 dpif_enable_upcall(udpif->dpif);
398 ovs_barrier_init(&udpif->reval_barrier, udpif->n_revalidators);
399 udpif->reval_exit = false;
400 udpif->revalidators = xzalloc(udpif->n_revalidators
401 * sizeof *udpif->revalidators);
402 udpif->ukeys = xmalloc(sizeof *udpif->ukeys * n_revalidators);
403 for (i = 0; i < udpif->n_revalidators; i++) {
404 struct revalidator *revalidator = &udpif->revalidators[i];
406 revalidator->udpif = udpif;
407 hmap_init(&udpif->ukeys[i].hmap);
408 ovs_mutex_init(&udpif->ukeys[i].mutex);
409 revalidator->ukeys = &udpif->ukeys[i].hmap;
410 revalidator->thread = ovs_thread_create(
411 "revalidator", udpif_revalidator, revalidator);
416 /* Tells 'udpif' how many threads it should use to handle upcalls.
417 * 'n_handlers' and 'n_revalidators' can never be zero. 'udpif''s
418 * datapath handle must have packet reception enabled before starting
421 udpif_set_threads(struct udpif *udpif, size_t n_handlers,
422 size_t n_revalidators)
425 ovs_assert(n_handlers && n_revalidators);
427 ovsrcu_quiesce_start();
428 if (udpif->n_handlers != n_handlers
429 || udpif->n_revalidators != n_revalidators) {
430 udpif_stop_threads(udpif);
433 if (!udpif->handlers && !udpif->revalidators) {
436 error = dpif_handlers_set(udpif->dpif, n_handlers);
438 VLOG_ERR("failed to configure handlers in dpif %s: %s",
439 dpif_name(udpif->dpif), ovs_strerror(error));
443 udpif_start_threads(udpif, n_handlers, n_revalidators);
445 ovsrcu_quiesce_end();
448 /* Waits for all ongoing upcall translations to complete. This ensures that
449 * there are no transient references to any removed ofprotos (or other
450 * objects). In particular, this should be called after an ofproto is removed
451 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
453 udpif_synchronize(struct udpif *udpif)
455 /* This is stronger than necessary. It would be sufficient to ensure
456 * (somehow) that each handler and revalidator thread had passed through
457 * its main loop once. */
458 size_t n_handlers = udpif->n_handlers;
459 size_t n_revalidators = udpif->n_revalidators;
461 ovsrcu_quiesce_start();
462 udpif_stop_threads(udpif);
463 udpif_start_threads(udpif, n_handlers, n_revalidators);
464 ovsrcu_quiesce_end();
467 /* Notifies 'udpif' that something changed which may render previous
468 * xlate_actions() results invalid. */
470 udpif_revalidate(struct udpif *udpif)
472 seq_change(udpif->reval_seq);
475 /* Returns a seq which increments every time 'udpif' pulls stats from the
476 * datapath. Callers can use this to get a sense of when might be a good time
477 * to do periodic work which relies on relatively up to date statistics. */
479 udpif_dump_seq(struct udpif *udpif)
481 return udpif->dump_seq;
485 udpif_get_memory_usage(struct udpif *udpif, struct simap *usage)
489 simap_increase(usage, "handlers", udpif->n_handlers);
491 simap_increase(usage, "revalidators", udpif->n_revalidators);
492 for (i = 0; i < udpif->n_revalidators; i++) {
493 ovs_mutex_lock(&udpif->ukeys[i].mutex);
494 simap_increase(usage, "udpif keys", hmap_count(&udpif->ukeys[i].hmap));
495 ovs_mutex_unlock(&udpif->ukeys[i].mutex);
499 /* Remove flows from a single datapath. */
501 udpif_flush(struct udpif *udpif)
503 size_t n_handlers, n_revalidators;
505 n_handlers = udpif->n_handlers;
506 n_revalidators = udpif->n_revalidators;
508 ovsrcu_quiesce_start();
510 udpif_stop_threads(udpif);
511 dpif_flow_flush(udpif->dpif);
512 udpif_start_threads(udpif, n_handlers, n_revalidators);
514 ovsrcu_quiesce_end();
517 /* Removes all flows from all datapaths. */
519 udpif_flush_all_datapaths(void)
523 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
530 udpif_get_n_flows(struct udpif *udpif)
532 long long int time, now;
533 unsigned long flow_count;
536 atomic_read_relaxed(&udpif->n_flows_timestamp, &time);
537 if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) {
538 struct dpif_dp_stats stats;
540 atomic_store_relaxed(&udpif->n_flows_timestamp, now);
541 dpif_get_dp_stats(udpif->dpif, &stats);
542 flow_count = stats.n_flows;
543 atomic_store_relaxed(&udpif->n_flows, flow_count);
544 ovs_mutex_unlock(&udpif->n_flows_mutex);
546 atomic_read_relaxed(&udpif->n_flows, &flow_count);
551 /* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH
552 * upcalls from dpif, processes the batch and installs corresponding flows
555 udpif_upcall_handler(void *arg)
557 struct handler *handler = arg;
558 struct udpif *udpif = handler->udpif;
560 while (!latch_is_set(&handler->udpif->exit_latch)) {
561 if (!recv_upcalls(handler)) {
562 dpif_recv_wait(udpif->dpif, handler->handler_id);
563 latch_wait(&udpif->exit_latch);
573 recv_upcalls(struct handler *handler)
575 struct udpif *udpif = handler->udpif;
576 uint64_t recv_stubs[UPCALL_MAX_BATCH][512 / 8];
577 struct ofpbuf recv_bufs[UPCALL_MAX_BATCH];
578 struct dpif_upcall dupcalls[UPCALL_MAX_BATCH];
579 struct upcall upcalls[UPCALL_MAX_BATCH];
580 struct flow flows[UPCALL_MAX_BATCH];
584 while (n_upcalls < UPCALL_MAX_BATCH) {
585 struct ofpbuf *recv_buf = &recv_bufs[n_upcalls];
586 struct dpif_upcall *dupcall = &dupcalls[n_upcalls];
587 struct upcall *upcall = &upcalls[n_upcalls];
588 struct flow *flow = &flows[n_upcalls];
589 struct pkt_metadata md;
592 ofpbuf_use_stub(recv_buf, recv_stubs[n_upcalls],
593 sizeof recv_stubs[n_upcalls]);
594 if (dpif_recv(udpif->dpif, handler->handler_id, dupcall, recv_buf)) {
595 ofpbuf_uninit(recv_buf);
599 if (odp_flow_key_to_flow(dupcall->key, dupcall->key_len, flow)
604 error = upcall_receive(upcall, udpif->backer, &dupcall->packet,
605 dupcall->type, dupcall->userdata, flow);
607 if (error == ENODEV) {
608 /* Received packet on datapath port for which we couldn't
609 * associate an ofproto. This can happen if a port is removed
610 * while traffic is being received. Print a rate-limited
611 * message in case it happens frequently. */
612 dpif_flow_put(udpif->dpif, DPIF_FP_CREATE, dupcall->key,
613 dupcall->key_len, NULL, 0, NULL, 0, NULL);
614 VLOG_INFO_RL(&rl, "received packet on unassociated datapath "
615 "port %"PRIu32, flow->in_port.odp_port);
620 upcall->key = dupcall->key;
621 upcall->key_len = dupcall->key_len;
623 upcall->out_tun_key = dupcall->out_tun_key;
625 if (vsp_adjust_flow(upcall->ofproto, flow, &dupcall->packet)) {
626 upcall->vsp_adjusted = true;
629 md = pkt_metadata_from_flow(flow);
630 flow_extract(&dupcall->packet, &md, flow);
632 error = process_upcall(udpif, upcall, NULL);
641 upcall_uninit(upcall);
643 ofpbuf_uninit(&dupcall->packet);
644 ofpbuf_uninit(recv_buf);
648 handle_upcalls(handler->udpif, upcalls, n_upcalls);
649 for (i = 0; i < n_upcalls; i++) {
650 ofpbuf_uninit(&dupcalls[i].packet);
651 ofpbuf_uninit(&recv_bufs[i]);
652 upcall_uninit(&upcalls[i]);
660 udpif_revalidator(void *arg)
662 /* Used by all revalidators. */
663 struct revalidator *revalidator = arg;
664 struct udpif *udpif = revalidator->udpif;
665 bool leader = revalidator == &udpif->revalidators[0];
667 /* Used only by the leader. */
668 long long int start_time = 0;
669 uint64_t last_reval_seq = 0;
672 revalidator->id = ovsthread_id_self();
677 reval_seq = seq_read(udpif->reval_seq);
678 udpif->need_revalidate = last_reval_seq != reval_seq;
679 last_reval_seq = reval_seq;
681 n_flows = udpif_get_n_flows(udpif);
682 udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows);
683 udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2;
685 /* Only the leader checks the exit latch to prevent a race where
686 * some threads think it's true and exit and others think it's
687 * false and block indefinitely on the reval_barrier */
688 udpif->reval_exit = latch_is_set(&udpif->exit_latch);
690 start_time = time_msec();
691 if (!udpif->reval_exit) {
692 udpif->dump = dpif_flow_dump_create(udpif->dpif);
696 /* Wait for the leader to start the flow dump. */
697 ovs_barrier_block(&udpif->reval_barrier);
698 if (udpif->reval_exit) {
701 revalidate(revalidator);
703 /* Wait for all flows to have been dumped before we garbage collect. */
704 ovs_barrier_block(&udpif->reval_barrier);
705 revalidator_sweep(revalidator);
707 /* Wait for all revalidators to finish garbage collection. */
708 ovs_barrier_block(&udpif->reval_barrier);
711 unsigned int flow_limit;
712 long long int duration;
714 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
716 dpif_flow_dump_destroy(udpif->dump);
717 seq_change(udpif->dump_seq);
719 duration = MAX(time_msec() - start_time, 1);
720 udpif->dump_duration = duration;
721 if (duration > 2000) {
722 flow_limit /= duration / 1000;
723 } else if (duration > 1300) {
724 flow_limit = flow_limit * 3 / 4;
725 } else if (duration < 1000 && n_flows > 2000
726 && flow_limit < n_flows * 1000 / duration) {
729 flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000));
730 atomic_store_relaxed(&udpif->flow_limit, flow_limit);
732 if (duration > 2000) {
733 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
737 poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500));
738 seq_wait(udpif->reval_seq, last_reval_seq);
739 latch_wait(&udpif->exit_latch);
747 static enum upcall_type
748 classify_upcall(enum dpif_upcall_type type, const struct nlattr *userdata)
750 union user_action_cookie cookie;
753 /* First look at the upcall type. */
761 case DPIF_N_UC_TYPES:
763 VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32, type);
767 /* "action" upcalls need a closer look. */
769 VLOG_WARN_RL(&rl, "action upcall missing cookie");
772 userdata_len = nl_attr_get_size(userdata);
773 if (userdata_len < sizeof cookie.type
774 || userdata_len > sizeof cookie) {
775 VLOG_WARN_RL(&rl, "action upcall cookie has unexpected size %"PRIuSIZE,
779 memset(&cookie, 0, sizeof cookie);
780 memcpy(&cookie, nl_attr_get(userdata), userdata_len);
781 if (userdata_len == MAX(8, sizeof cookie.sflow)
782 && cookie.type == USER_ACTION_COOKIE_SFLOW) {
784 } else if (userdata_len == MAX(8, sizeof cookie.slow_path)
785 && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) {
787 } else if (userdata_len == MAX(8, sizeof cookie.flow_sample)
788 && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) {
789 return FLOW_SAMPLE_UPCALL;
790 } else if (userdata_len == MAX(8, sizeof cookie.ipfix)
791 && cookie.type == USER_ACTION_COOKIE_IPFIX) {
794 VLOG_WARN_RL(&rl, "invalid user cookie of type %"PRIu16
795 " and size %"PRIuSIZE, cookie.type, userdata_len);
800 /* Calculates slow path actions for 'xout'. 'buf' must statically be
801 * initialized with at least 128 bytes of space. */
803 compose_slow_path(struct udpif *udpif, struct xlate_out *xout,
804 const struct flow *flow, odp_port_t odp_in_port,
807 union user_action_cookie cookie;
811 cookie.type = USER_ACTION_COOKIE_SLOW_PATH;
812 cookie.slow_path.unused = 0;
813 cookie.slow_path.reason = xout->slow;
815 port = xout->slow & (SLOW_CFM | SLOW_BFD | SLOW_LACP | SLOW_STP)
818 pid = dpif_port_get_pid(udpif->dpif, port, flow_hash_5tuple(flow, 0));
819 odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, ODPP_NONE,
823 /* If there is no error, the upcall must be destroyed with upcall_uninit()
824 * before quiescing, as the referred objects are guaranteed to exist only
825 * until the calling thread quiesces. Otherwise, do not call upcall_uninit()
826 * since the 'upcall->put_actions' remains uninitialized. */
828 upcall_receive(struct upcall *upcall, const struct dpif_backer *backer,
829 const struct ofpbuf *packet, enum dpif_upcall_type type,
830 const struct nlattr *userdata, const struct flow *flow)
834 error = xlate_lookup(backer, flow, &upcall->ofproto, &upcall->ipfix,
835 &upcall->sflow, NULL, &upcall->in_port);
841 upcall->packet = packet;
843 upcall->userdata = userdata;
844 ofpbuf_init(&upcall->put_actions, 0);
846 upcall->xout_initialized = false;
847 upcall->vsp_adjusted = false;
852 upcall->out_tun_key = NULL;
858 upcall_xlate(struct udpif *udpif, struct upcall *upcall,
859 struct ofpbuf *odp_actions)
861 struct dpif_flow_stats stats;
865 stats.n_bytes = ofpbuf_size(upcall->packet);
866 stats.used = time_msec();
867 stats.tcp_flags = ntohs(upcall->flow->tcp_flags);
869 xlate_in_init(&xin, upcall->ofproto, upcall->flow, upcall->in_port, NULL,
870 stats.tcp_flags, upcall->packet);
871 xin.odp_actions = odp_actions;
873 if (upcall->type == DPIF_UC_MISS) {
874 xin.resubmit_stats = &stats;
876 /* For non-miss upcalls, there's a flow in the datapath which this
877 * packet was accounted to. Presumably the revalidators will deal
878 * with pushing its stats eventually. */
881 xlate_actions(&xin, &upcall->xout);
882 upcall->xout_initialized = true;
884 /* Special case for fail-open mode.
886 * If we are in fail-open mode, but we are connected to a controller too,
887 * then we should send the packet up to the controller in the hope that it
888 * will try to set up a flow and thereby allow us to exit fail-open.
890 * See the top-level comment in fail-open.c for more information.
892 * Copy packets before they are modified by execution. */
893 if (upcall->xout.fail_open) {
894 const struct ofpbuf *packet = upcall->packet;
895 struct ofproto_packet_in *pin;
897 pin = xmalloc(sizeof *pin);
898 pin->up.packet = xmemdup(ofpbuf_data(packet), ofpbuf_size(packet));
899 pin->up.packet_len = ofpbuf_size(packet);
900 pin->up.reason = OFPR_NO_MATCH;
901 pin->up.table_id = 0;
902 pin->up.cookie = OVS_BE64_MAX;
903 flow_get_metadata(upcall->flow, &pin->up.fmd);
904 pin->send_len = 0; /* Not used for flow table misses. */
905 pin->miss_type = OFPROTO_PACKET_IN_NO_MISS;
906 ofproto_dpif_send_packet_in(upcall->ofproto, pin);
909 if (!upcall->xout.slow) {
910 ofpbuf_use_const(&upcall->put_actions,
911 ofpbuf_data(upcall->xout.odp_actions),
912 ofpbuf_size(upcall->xout.odp_actions));
914 ofpbuf_init(&upcall->put_actions, 0);
915 compose_slow_path(udpif, &upcall->xout, upcall->flow,
916 upcall->flow->in_port.odp_port,
917 &upcall->put_actions);
922 upcall_uninit(struct upcall *upcall)
925 if (upcall->xout_initialized) {
926 xlate_out_uninit(&upcall->xout);
928 ofpbuf_uninit(&upcall->put_actions);
933 upcall_cb(const struct ofpbuf *packet, const struct flow *flow,
934 enum dpif_upcall_type type, const struct nlattr *userdata,
935 struct ofpbuf *actions, struct flow_wildcards *wc,
936 struct ofpbuf *put_actions, void *aux)
938 struct udpif *udpif = aux;
939 unsigned int flow_limit;
940 struct upcall upcall;
944 atomic_read_relaxed(&enable_megaflows, &megaflow);
945 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
947 error = upcall_receive(&upcall, udpif->backer, packet, type, userdata,
953 error = process_upcall(udpif, &upcall, actions);
958 if (upcall.xout.slow && put_actions) {
959 ofpbuf_put(put_actions, ofpbuf_data(&upcall.put_actions),
960 ofpbuf_size(&upcall.put_actions));
963 if (OVS_LIKELY(wc)) {
965 /* XXX: This could be avoided with sufficient API changes. */
966 *wc = upcall.xout.wc;
968 flow_wildcards_init_for_packet(wc, flow);
972 if (udpif_get_n_flows(udpif) >= flow_limit) {
977 upcall_uninit(&upcall);
982 process_upcall(struct udpif *udpif, struct upcall *upcall,
983 struct ofpbuf *odp_actions)
985 const struct nlattr *userdata = upcall->userdata;
986 const struct ofpbuf *packet = upcall->packet;
987 const struct flow *flow = upcall->flow;
989 switch (classify_upcall(upcall->type, userdata)) {
991 upcall_xlate(udpif, upcall, odp_actions);
996 union user_action_cookie cookie;
998 memset(&cookie, 0, sizeof cookie);
999 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.sflow);
1000 dpif_sflow_received(upcall->sflow, packet, flow,
1001 flow->in_port.odp_port, &cookie);
1006 if (upcall->ipfix) {
1007 union user_action_cookie cookie;
1008 struct flow_tnl output_tunnel_key;
1010 memset(&cookie, 0, sizeof cookie);
1011 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.ipfix);
1013 if (upcall->out_tun_key) {
1014 memset(&output_tunnel_key, 0, sizeof output_tunnel_key);
1015 odp_tun_key_from_attr(upcall->out_tun_key,
1016 &output_tunnel_key);
1018 dpif_ipfix_bridge_sample(upcall->ipfix, packet, flow,
1019 flow->in_port.odp_port,
1020 cookie.ipfix.output_odp_port,
1021 upcall->out_tun_key ?
1022 &output_tunnel_key : NULL);
1026 case FLOW_SAMPLE_UPCALL:
1027 if (upcall->ipfix) {
1028 union user_action_cookie cookie;
1030 memset(&cookie, 0, sizeof cookie);
1031 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.flow_sample);
1033 /* The flow reflects exactly the contents of the packet.
1034 * Sample the packet using it. */
1035 dpif_ipfix_flow_sample(upcall->ipfix, packet, flow,
1036 cookie.flow_sample.collector_set_id,
1037 cookie.flow_sample.probability,
1038 cookie.flow_sample.obs_domain_id,
1039 cookie.flow_sample.obs_point_id);
1051 handle_upcalls(struct udpif *udpif, struct upcall *upcalls,
1054 struct odputil_keybuf mask_bufs[UPCALL_MAX_BATCH];
1055 struct dpif_op *opsp[UPCALL_MAX_BATCH * 2];
1056 struct dpif_op ops[UPCALL_MAX_BATCH * 2];
1057 unsigned int flow_limit;
1062 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1063 atomic_read_relaxed(&enable_megaflows, &megaflow);
1065 may_put = udpif_get_n_flows(udpif) < flow_limit;
1067 /* Handle the packets individually in order of arrival.
1069 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1070 * processes received packets for these protocols.
1072 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1075 * The loop fills 'ops' with an array of operations to execute in the
1078 for (i = 0; i < n_upcalls; i++) {
1079 struct upcall *upcall = &upcalls[i];
1080 const struct ofpbuf *packet = upcall->packet;
1083 if (upcall->vsp_adjusted) {
1084 /* This packet was received on a VLAN splinter port. We added a
1085 * VLAN to the packet to make the packet resemble the flow, but the
1086 * actions were composed assuming that the packet contained no
1087 * VLAN. So, we must remove the VLAN header from the packet before
1088 * trying to execute the actions. */
1089 if (ofpbuf_size(upcall->xout.odp_actions)) {
1090 eth_pop_vlan(CONST_CAST(struct ofpbuf *, upcall->packet));
1093 /* Remove the flow vlan tags inserted by vlan splinter logic
1094 * to ensure megaflow masks generated match the data path flow. */
1095 CONST_CAST(struct flow *, upcall->flow)->vlan_tci = 0;
1098 /* Do not install a flow into the datapath if:
1100 * - The datapath already has too many flows.
1102 * - We received this packet via some flow installed in the kernel
1104 if (may_put && upcall->type == DPIF_UC_MISS) {
1107 ofpbuf_use_stack(&mask, &mask_bufs[i], sizeof mask_bufs[i]);
1113 recirc = ofproto_dpif_get_enable_recirc(upcall->ofproto);
1114 max_mpls = ofproto_dpif_get_max_mpls_depth(upcall->ofproto);
1115 odp_flow_key_from_mask(&mask, &upcall->xout.wc.masks,
1116 upcall->flow, UINT32_MAX, max_mpls,
1121 op->type = DPIF_OP_FLOW_PUT;
1122 op->u.flow_put.flags = DPIF_FP_CREATE;
1123 op->u.flow_put.key = upcall->key;
1124 op->u.flow_put.key_len = upcall->key_len;
1125 op->u.flow_put.mask = ofpbuf_data(&mask);
1126 op->u.flow_put.mask_len = ofpbuf_size(&mask);
1127 op->u.flow_put.stats = NULL;
1128 op->u.flow_put.actions = ofpbuf_data(&upcall->put_actions);
1129 op->u.flow_put.actions_len = ofpbuf_size(&upcall->put_actions);
1132 if (ofpbuf_size(upcall->xout.odp_actions)) {
1134 op->type = DPIF_OP_EXECUTE;
1135 op->u.execute.packet = CONST_CAST(struct ofpbuf *, packet);
1136 odp_key_to_pkt_metadata(upcall->key, upcall->key_len,
1138 op->u.execute.actions = ofpbuf_data(upcall->xout.odp_actions);
1139 op->u.execute.actions_len = ofpbuf_size(upcall->xout.odp_actions);
1140 op->u.execute.needs_help = (upcall->xout.slow & SLOW_ACTION) != 0;
1141 op->u.execute.probe = false;
1145 /* Execute batch. */
1146 for (i = 0; i < n_ops; i++) {
1149 dpif_operate(udpif->dpif, opsp, n_ops);
1152 /* Must be called with udpif->ukeys[hash % udpif->n_revalidators].mutex. */
1153 static struct udpif_key *
1154 ukey_lookup(struct udpif *udpif, const struct nlattr *key, size_t key_len,
1156 OVS_REQUIRES(udpif->ukeys->mutex)
1158 struct udpif_key *ukey;
1159 struct hmap *hmap = &udpif->ukeys[hash % udpif->n_revalidators].hmap;
1161 HMAP_FOR_EACH_WITH_HASH (ukey, hmap_node, hash, hmap) {
1162 if (ukey->key_len == key_len && !memcmp(ukey->key, key, key_len)) {
1169 /* Creates a ukey for 'key' and 'key_len', returning it with ukey->mutex in
1170 * a locked state. */
1171 static struct udpif_key *
1172 ukey_create(const struct nlattr *key, size_t key_len, long long int used)
1173 OVS_NO_THREAD_SAFETY_ANALYSIS
1175 struct udpif_key *ukey = xmalloc(sizeof *ukey);
1177 ovs_mutex_init(&ukey->mutex);
1178 ukey->key = &ukey->key_buf_nla;
1179 memcpy(&ukey->key_buf, key, key_len);
1180 ukey->key_len = key_len;
1182 ovs_mutex_lock(&ukey->mutex);
1184 ukey->flow_exists = true;
1185 ukey->created = used ? used : time_msec();
1186 memset(&ukey->stats, 0, sizeof ukey->stats);
1187 ukey->xcache = NULL;
1192 /* Searches for a ukey in 'udpif->ukeys' that matches 'key' and 'key_len' and
1193 * attempts to lock the ukey. If the ukey does not exist, create it.
1195 * Returns true on success, setting *result to the matching ukey and returning
1196 * it in a locked state. Otherwise, returns false and clears *result. */
1198 ukey_acquire(struct udpif *udpif, const struct nlattr *key, size_t key_len,
1199 long long int used, struct udpif_key **result)
1200 OVS_TRY_LOCK(true, (*result)->mutex)
1202 struct udpif_key *ukey;
1204 bool locked = false;
1206 hash = hash_bytes(key, key_len, udpif->secret);
1207 idx = hash % udpif->n_revalidators;
1209 ovs_mutex_lock(&udpif->ukeys[idx].mutex);
1210 ukey = ukey_lookup(udpif, key, key_len, hash);
1212 ukey = ukey_create(key, key_len, used);
1213 hmap_insert(&udpif->ukeys[idx].hmap, &ukey->hmap_node, hash);
1215 } else if (!ovs_mutex_trylock(&ukey->mutex)) {
1218 ovs_mutex_unlock(&udpif->ukeys[idx].mutex);
1229 ukey_delete(struct revalidator *revalidator, struct udpif_key *ukey)
1230 OVS_NO_THREAD_SAFETY_ANALYSIS
1233 hmap_remove(revalidator->ukeys, &ukey->hmap_node);
1235 xlate_cache_delete(ukey->xcache);
1236 ovs_mutex_destroy(&ukey->mutex);
1241 should_revalidate(const struct udpif *udpif, uint64_t packets,
1244 long long int metric, now, duration;
1246 if (udpif->dump_duration < 200) {
1247 /* We are likely to handle full revalidation for the flows. */
1251 /* Calculate the mean time between seeing these packets. If this
1252 * exceeds the threshold, then delete the flow rather than performing
1253 * costly revalidation for flows that aren't being hit frequently.
1255 * This is targeted at situations where the dump_duration is high (~1s),
1256 * and revalidation is triggered by a call to udpif_revalidate(). In
1257 * these situations, revalidation of all flows causes fluctuations in the
1258 * flow_limit due to the interaction with the dump_duration and max_idle.
1259 * This tends to result in deletion of low-throughput flows anyway, so
1260 * skip the revalidation and just delete those flows. */
1261 packets = MAX(packets, 1);
1262 now = MAX(used, time_msec());
1263 duration = now - used;
1264 metric = duration / packets;
1267 /* The flow is receiving more than ~5pps, so keep it. */
1274 revalidate_ukey(struct udpif *udpif, struct udpif_key *ukey,
1275 const struct dpif_flow *f)
1276 OVS_REQUIRES(ukey->mutex)
1278 uint64_t slow_path_buf[128 / 8];
1279 struct xlate_out xout, *xoutp;
1280 struct netflow *netflow;
1281 struct ofproto_dpif *ofproto;
1282 struct dpif_flow_stats push;
1283 struct ofpbuf xout_actions;
1284 struct flow flow, dp_mask;
1285 uint32_t *dp32, *xout32;
1286 ofp_port_t ofp_in_port;
1287 struct xlate_in xin;
1288 long long int last_used;
1297 last_used = ukey->stats.used;
1298 push.used = f->stats.used;
1299 push.tcp_flags = f->stats.tcp_flags;
1300 push.n_packets = (f->stats.n_packets > ukey->stats.n_packets
1301 ? f->stats.n_packets - ukey->stats.n_packets
1303 push.n_bytes = (f->stats.n_bytes > ukey->stats.n_bytes
1304 ? f->stats.n_bytes - ukey->stats.n_bytes
1307 if (udpif->need_revalidate && last_used
1308 && !should_revalidate(udpif, push.n_packets, last_used)) {
1313 /* We will push the stats, so update the ukey stats cache. */
1314 ukey->stats = f->stats;
1315 if (!push.n_packets && !udpif->need_revalidate) {
1320 if (ukey->xcache && !udpif->need_revalidate) {
1321 xlate_push_stats(ukey->xcache, &push);
1326 if (odp_flow_key_to_flow(ukey->key, ukey->key_len, &flow)
1331 error = xlate_lookup(udpif->backer, &flow, &ofproto, NULL, NULL, &netflow,
1337 if (udpif->need_revalidate) {
1338 xlate_cache_clear(ukey->xcache);
1340 if (!ukey->xcache) {
1341 ukey->xcache = xlate_cache_new();
1344 xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL, push.tcp_flags,
1346 if (push.n_packets) {
1347 xin.resubmit_stats = &push;
1348 xin.may_learn = true;
1350 xin.xcache = ukey->xcache;
1351 xin.skip_wildcards = !udpif->need_revalidate;
1352 xlate_actions(&xin, &xout);
1355 if (!udpif->need_revalidate) {
1361 ofpbuf_use_const(&xout_actions, ofpbuf_data(xout.odp_actions),
1362 ofpbuf_size(xout.odp_actions));
1364 ofpbuf_use_stack(&xout_actions, slow_path_buf, sizeof slow_path_buf);
1365 compose_slow_path(udpif, &xout, &flow, flow.in_port.odp_port,
1369 if (f->actions_len != ofpbuf_size(&xout_actions)
1370 || memcmp(ofpbuf_data(&xout_actions), f->actions, f->actions_len)) {
1374 if (odp_flow_key_to_mask(f->mask, f->mask_len, &dp_mask, &flow)
1379 /* Since the kernel is free to ignore wildcarded bits in the mask, we can't
1380 * directly check that the masks are the same. Instead we check that the
1381 * mask in the kernel is more specific i.e. less wildcarded, than what
1382 * we've calculated here. This guarantees we don't catch any packets we
1383 * shouldn't with the megaflow. */
1384 dp32 = (uint32_t *) &dp_mask;
1385 xout32 = (uint32_t *) &xout.wc.masks;
1386 for (i = 0; i < FLOW_U32S; i++) {
1387 if ((dp32[i] | xout32[i]) != dp32[i]) {
1394 if (netflow && !ok) {
1395 netflow_flow_clear(netflow, &flow);
1397 xlate_out_uninit(xoutp);
1402 struct udpif_key *ukey;
1403 struct dpif_flow_stats stats; /* Stats for 'op'. */
1404 struct dpif_op op; /* Flow del operation. */
1408 dump_op_init(struct dump_op *op, const struct nlattr *key, size_t key_len,
1409 struct udpif_key *ukey)
1412 op->op.type = DPIF_OP_FLOW_DEL;
1413 op->op.u.flow_del.key = key;
1414 op->op.u.flow_del.key_len = key_len;
1415 op->op.u.flow_del.stats = &op->stats;
1419 push_dump_ops__(struct udpif *udpif, struct dump_op *ops, size_t n_ops)
1421 struct dpif_op *opsp[REVALIDATE_MAX_BATCH];
1424 ovs_assert(n_ops <= REVALIDATE_MAX_BATCH);
1425 for (i = 0; i < n_ops; i++) {
1426 opsp[i] = &ops[i].op;
1428 dpif_operate(udpif->dpif, opsp, n_ops);
1430 for (i = 0; i < n_ops; i++) {
1431 struct dump_op *op = &ops[i];
1432 struct dpif_flow_stats *push, *stats, push_buf;
1434 stats = op->op.u.flow_del.stats;
1437 ovs_mutex_lock(&op->ukey->mutex);
1438 push->used = MAX(stats->used, op->ukey->stats.used);
1439 push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags;
1440 push->n_packets = stats->n_packets - op->ukey->stats.n_packets;
1441 push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes;
1442 ovs_mutex_unlock(&op->ukey->mutex);
1444 if (push->n_packets || netflow_exists()) {
1445 struct ofproto_dpif *ofproto;
1446 struct netflow *netflow;
1447 ofp_port_t ofp_in_port;
1451 ovs_mutex_lock(&op->ukey->mutex);
1452 if (op->ukey->xcache) {
1453 xlate_push_stats(op->ukey->xcache, push);
1454 ovs_mutex_unlock(&op->ukey->mutex);
1457 ovs_mutex_unlock(&op->ukey->mutex);
1459 if (odp_flow_key_to_flow(op->op.u.flow_del.key,
1460 op->op.u.flow_del.key_len, &flow)
1465 error = xlate_lookup(udpif->backer, &flow, &ofproto,
1466 NULL, NULL, &netflow, &ofp_in_port);
1468 struct xlate_in xin;
1470 xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL,
1471 push->tcp_flags, NULL);
1472 xin.resubmit_stats = push->n_packets ? push : NULL;
1473 xin.may_learn = push->n_packets > 0;
1474 xin.skip_wildcards = true;
1475 xlate_actions_for_side_effects(&xin);
1478 netflow_flow_clear(netflow, &flow);
1486 push_dump_ops(struct revalidator *revalidator,
1487 struct dump_op *ops, size_t n_ops)
1491 push_dump_ops__(revalidator->udpif, ops, n_ops);
1492 for (i = 0; i < n_ops; i++) {
1493 ukey_delete(revalidator, ops[i].ukey);
1498 revalidate(struct revalidator *revalidator)
1500 struct udpif *udpif = revalidator->udpif;
1501 struct dpif_flow_dump_thread *dump_thread;
1503 unsigned int flow_limit;
1505 dump_seq = seq_read(udpif->dump_seq);
1506 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1507 dump_thread = dpif_flow_dump_thread_create(udpif->dump);
1509 struct dump_op ops[REVALIDATE_MAX_BATCH];
1512 struct dpif_flow flows[REVALIDATE_MAX_BATCH];
1513 const struct dpif_flow *f;
1516 long long int max_idle;
1521 n_dumped = dpif_flow_dump_next(dump_thread, flows, ARRAY_SIZE(flows));
1528 /* In normal operation we want to keep flows around until they have
1529 * been idle for 'ofproto_max_idle' milliseconds. However:
1531 * - If the number of datapath flows climbs above 'flow_limit',
1532 * drop that down to 100 ms to try to bring the flows down to
1535 * - If the number of datapath flows climbs above twice
1536 * 'flow_limit', delete all the datapath flows as an emergency
1537 * measure. (We reassess this condition for the next batch of
1538 * datapath flows, so we will recover before all the flows are
1540 n_dp_flows = udpif_get_n_flows(udpif);
1541 kill_them_all = n_dp_flows > flow_limit * 2;
1542 max_idle = n_dp_flows > flow_limit ? 100 : ofproto_max_idle;
1544 for (f = flows; f < &flows[n_dumped]; f++) {
1545 long long int used = f->stats.used;
1546 struct udpif_key *ukey;
1547 bool already_dumped, keep;
1549 if (!ukey_acquire(udpif, f->key, f->key_len, used, &ukey)) {
1550 /* We couldn't acquire the ukey. This means that
1551 * another revalidator is processing this flow
1552 * concurrently, so don't bother processing it. */
1553 COVERAGE_INC(dumped_duplicate_flow);
1557 already_dumped = ukey->dump_seq == dump_seq;
1558 if (already_dumped) {
1559 /* The flow has already been handled during this flow dump
1560 * operation. Skip it. */
1562 COVERAGE_INC(dumped_duplicate_flow);
1564 COVERAGE_INC(dumped_new_flow);
1566 ovs_mutex_unlock(&ukey->mutex);
1571 used = ukey->created;
1573 if (kill_them_all || (used && used < now - max_idle)) {
1576 keep = revalidate_ukey(udpif, ukey, f);
1578 ukey->dump_seq = dump_seq;
1579 ukey->flow_exists = keep;
1582 dump_op_init(&ops[n_ops++], f->key, f->key_len, ukey);
1584 ovs_mutex_unlock(&ukey->mutex);
1588 push_dump_ops__(udpif, ops, n_ops);
1591 dpif_flow_dump_thread_destroy(dump_thread);
1594 /* Called with exclusive access to 'revalidator' and 'ukey'. */
1596 handle_missed_revalidation(struct revalidator *revalidator,
1597 struct udpif_key *ukey)
1598 OVS_NO_THREAD_SAFETY_ANALYSIS
1600 struct udpif *udpif = revalidator->udpif;
1601 struct dpif_flow flow;
1603 uint64_t stub[DPIF_FLOW_BUFSIZE / 8];
1606 COVERAGE_INC(revalidate_missed_dp_flow);
1608 ofpbuf_use_stub(&buf, &stub, sizeof stub);
1609 if (!dpif_flow_get(udpif->dpif, ukey->key, ukey->key_len, &buf, &flow)) {
1610 keep = revalidate_ukey(udpif, ukey, &flow);
1612 ofpbuf_uninit(&buf);
1618 revalidator_sweep__(struct revalidator *revalidator, bool purge)
1619 OVS_NO_THREAD_SAFETY_ANALYSIS
1621 struct dump_op ops[REVALIDATE_MAX_BATCH];
1622 struct udpif_key *ukey, *next;
1627 dump_seq = seq_read(revalidator->udpif->dump_seq);
1629 /* During garbage collection, this revalidator completely owns its ukeys
1630 * map, and therefore doesn't need to do any locking. */
1631 HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, revalidator->ukeys) {
1632 if (ukey->flow_exists
1634 || (ukey->dump_seq != dump_seq
1635 && revalidator->udpif->need_revalidate
1636 && !handle_missed_revalidation(revalidator, ukey)))) {
1637 struct dump_op *op = &ops[n_ops++];
1639 dump_op_init(op, ukey->key, ukey->key_len, ukey);
1640 if (n_ops == REVALIDATE_MAX_BATCH) {
1641 push_dump_ops(revalidator, ops, n_ops);
1644 } else if (!ukey->flow_exists) {
1645 ukey_delete(revalidator, ukey);
1650 push_dump_ops(revalidator, ops, n_ops);
1655 revalidator_sweep(struct revalidator *revalidator)
1657 revalidator_sweep__(revalidator, false);
1661 revalidator_purge(struct revalidator *revalidator)
1663 revalidator_sweep__(revalidator, true);
1667 upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED,
1668 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
1670 struct ds ds = DS_EMPTY_INITIALIZER;
1671 struct udpif *udpif;
1673 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1674 unsigned int flow_limit;
1677 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1679 ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif));
1680 ds_put_format(&ds, "\tflows : (current %lu)"
1681 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif),
1682 udpif->avg_n_flows, udpif->max_n_flows, flow_limit);
1683 ds_put_format(&ds, "\tdump duration : %lldms\n", udpif->dump_duration);
1685 ds_put_char(&ds, '\n');
1686 for (i = 0; i < n_revalidators; i++) {
1687 struct revalidator *revalidator = &udpif->revalidators[i];
1689 ovs_mutex_lock(&udpif->ukeys[i].mutex);
1690 ds_put_format(&ds, "\t%u: (keys %"PRIuSIZE")\n",
1691 revalidator->id, hmap_count(&udpif->ukeys[i].hmap));
1692 ovs_mutex_unlock(&udpif->ukeys[i].mutex);
1696 unixctl_command_reply(conn, ds_cstr(&ds));
1700 /* Disable using the megaflows.
1702 * This command is only needed for advanced debugging, so it's not
1703 * documented in the man page. */
1705 upcall_unixctl_disable_megaflows(struct unixctl_conn *conn,
1706 int argc OVS_UNUSED,
1707 const char *argv[] OVS_UNUSED,
1708 void *aux OVS_UNUSED)
1710 atomic_store_relaxed(&enable_megaflows, false);
1711 udpif_flush_all_datapaths();
1712 unixctl_command_reply(conn, "megaflows disabled");
1715 /* Re-enable using megaflows.
1717 * This command is only needed for advanced debugging, so it's not
1718 * documented in the man page. */
1720 upcall_unixctl_enable_megaflows(struct unixctl_conn *conn,
1721 int argc OVS_UNUSED,
1722 const char *argv[] OVS_UNUSED,
1723 void *aux OVS_UNUSED)
1725 atomic_store_relaxed(&enable_megaflows, true);
1726 udpif_flush_all_datapaths();
1727 unixctl_command_reply(conn, "megaflows enabled");
1730 /* Set the flow limit.
1732 * This command is only needed for advanced debugging, so it's not
1733 * documented in the man page. */
1735 upcall_unixctl_set_flow_limit(struct unixctl_conn *conn,
1736 int argc OVS_UNUSED,
1737 const char *argv[] OVS_UNUSED,
1738 void *aux OVS_UNUSED)
1740 struct ds ds = DS_EMPTY_INITIALIZER;
1741 struct udpif *udpif;
1742 unsigned int flow_limit = atoi(argv[1]);
1744 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1745 atomic_store_relaxed(&udpif->flow_limit, flow_limit);
1747 ds_put_format(&ds, "set flow_limit to %u\n", flow_limit);
1748 unixctl_command_reply(conn, ds_cstr(&ds));
1753 upcall_unixctl_dump_wait(struct unixctl_conn *conn,
1754 int argc OVS_UNUSED,
1755 const char *argv[] OVS_UNUSED,
1756 void *aux OVS_UNUSED)
1758 if (list_is_singleton(&all_udpifs)) {
1759 struct udpif *udpif = NULL;
1762 udpif = OBJECT_CONTAINING(list_front(&all_udpifs), udpif, list_node);
1763 len = (udpif->n_conns + 1) * sizeof *udpif->conns;
1764 udpif->conn_seq = seq_read(udpif->dump_seq);
1765 udpif->conns = xrealloc(udpif->conns, len);
1766 udpif->conns[udpif->n_conns++] = conn;
1768 unixctl_command_reply_error(conn, "can't wait on multiple udpifs.");