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(upcall_duplicate_flow);
49 COVERAGE_DEFINE(revalidate_missed_dp_flow);
51 /* A thread that reads upcalls from dpif, forwards each upcall's packet,
52 * and possibly sets up a kernel flow as a cache. */
54 struct udpif *udpif; /* Parent udpif. */
55 pthread_t thread; /* Thread ID. */
56 uint32_t handler_id; /* Handler id. */
59 /* A thread that processes datapath flows, updates OpenFlow statistics, and
60 * updates or removes them if necessary. */
62 struct udpif *udpif; /* Parent udpif. */
63 pthread_t thread; /* Thread ID. */
64 unsigned int id; /* ovsthread_id_self(). */
65 struct hmap *ukeys; /* Points into udpif->ukeys for this
66 revalidator. Used for GC phase. */
69 /* An upcall handler for ofproto_dpif.
71 * udpif keeps records of two kind of logically separate units:
76 * - An array of 'struct handler's for upcall handling and flow
82 * - Revalidation threads which read the datapath flow table and maintains
86 struct list list_node; /* In all_udpifs list. */
88 struct dpif *dpif; /* Datapath handle. */
89 struct dpif_backer *backer; /* Opaque dpif_backer pointer. */
91 uint32_t secret; /* Random seed for upcall hash. */
93 struct handler *handlers; /* Upcall handlers. */
96 struct revalidator *revalidators; /* Flow revalidators. */
97 size_t n_revalidators;
99 struct latch exit_latch; /* Tells child threads to exit. */
102 struct seq *reval_seq; /* Incremented to force revalidation. */
103 bool need_revalidate; /* As indicated by 'reval_seq'. */
104 bool reval_exit; /* Set by leader on 'exit_latch. */
105 struct ovs_barrier reval_barrier; /* Barrier used by revalidators. */
106 struct dpif_flow_dump *dump; /* DPIF flow dump state. */
107 long long int dump_duration; /* Duration of the last flow dump. */
108 struct seq *dump_seq; /* Increments each dump iteration. */
110 /* There are 'n_revalidators' ukey hmaps. Each revalidator retains a
111 * reference to one of these for garbage collection.
113 * During the flow dump phase, revalidators insert into these with a random
114 * distribution. During the garbage collection phase, each revalidator
115 * takes care of garbage collecting one of these hmaps. */
117 struct ovs_mutex mutex; /* Guards the following. */
118 struct hmap hmap OVS_GUARDED; /* Datapath flow keys. */
121 /* Datapath flow statistics. */
122 unsigned int max_n_flows;
123 unsigned int avg_n_flows;
125 /* Following fields are accessed and modified by different threads. */
126 atomic_uint flow_limit; /* Datapath flow hard limit. */
128 /* n_flows_mutex prevents multiple threads updating these concurrently. */
129 atomic_uint n_flows; /* Number of flows in the datapath. */
130 atomic_llong n_flows_timestamp; /* Last time n_flows was updated. */
131 struct ovs_mutex n_flows_mutex;
133 /* Following fields are accessed and modified only from the main thread. */
134 struct unixctl_conn **conns; /* Connections waiting on dump_seq. */
135 uint64_t conn_seq; /* Corresponds to 'dump_seq' when
136 conns[n_conns-1] was stored. */
137 size_t n_conns; /* Number of connections waiting. */
141 BAD_UPCALL, /* Some kind of bug somewhere. */
142 MISS_UPCALL, /* A flow miss. */
143 SFLOW_UPCALL, /* sFlow sample. */
144 FLOW_SAMPLE_UPCALL, /* Per-flow sampling. */
145 IPFIX_UPCALL /* Per-bridge sampling. */
149 struct ofproto_dpif *ofproto; /* Parent ofproto. */
151 /* The flow and packet are only required to be constant when using
152 * dpif-netdev. If a modification is absolutely necessary, a const cast
153 * may be used with other datapaths. */
154 const struct flow *flow; /* Parsed representation of the packet. */
155 const struct ofpbuf *packet; /* Packet associated with this upcall. */
156 ofp_port_t in_port; /* OpenFlow in port, or OFPP_NONE. */
158 enum dpif_upcall_type type; /* Datapath type of the upcall. */
159 const struct nlattr *userdata; /* Userdata for DPIF_UC_ACTION Upcalls. */
161 bool xout_initialized; /* True if 'xout' must be uninitialized. */
162 struct xlate_out xout; /* Result of xlate_actions(). */
163 struct ofpbuf put_actions; /* Actions 'put' in the fastapath. */
165 struct dpif_ipfix *ipfix; /* IPFIX pointer or NULL. */
166 struct dpif_sflow *sflow; /* SFlow pointer or NULL. */
168 bool vsp_adjusted; /* 'packet' and 'flow' were adjusted for
169 VLAN splinters if true. */
171 /* Not used by the upcall callback interface. */
172 const struct nlattr *key; /* Datapath flow key. */
173 size_t key_len; /* Datapath flow key length. */
174 const struct nlattr *out_tun_key; /* Datapath output tunnel key. */
177 /* 'udpif_key's are responsible for tracking the little bit of state udpif
178 * needs to do flow expiration which can't be pulled directly from the
179 * datapath. They may be created or maintained by any revalidator during
180 * the dump phase, but are owned by a single revalidator, and are destroyed
181 * by that revalidator during the garbage-collection phase.
183 * While some elements of a udpif_key are protected by a mutex, the ukey itself
184 * is not. Therefore it is not safe to destroy a udpif_key except when all
185 * revalidators are in garbage collection phase, or they aren't running. */
187 struct hmap_node hmap_node; /* In parent revalidator 'ukeys' map. */
189 /* These elements are read only once created, and therefore aren't
190 * protected by a mutex. */
191 const struct nlattr *key; /* Datapath flow key. */
192 size_t key_len; /* Length of 'key'. */
194 struct ovs_mutex mutex; /* Guards the following. */
195 struct dpif_flow_stats stats OVS_GUARDED; /* Last known stats.*/
196 long long int created OVS_GUARDED; /* Estimate of creation time. */
197 uint64_t dump_seq OVS_GUARDED; /* Tracks udpif->dump_seq. */
198 bool flow_exists OVS_GUARDED; /* Ensures flows are only deleted
201 struct xlate_cache *xcache OVS_GUARDED; /* Cache for xlate entries that
202 * are affected by this ukey.
203 * Used for stats and learning.*/
205 struct odputil_keybuf key_buf; /* Memory for 'key'. */
206 struct nlattr key_buf_nla;
210 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
211 static struct list all_udpifs = LIST_INITIALIZER(&all_udpifs);
213 static size_t recv_upcalls(struct handler *);
214 static int process_upcall(struct udpif *, struct upcall *,
215 struct ofpbuf *odp_actions);
216 static void handle_upcalls(struct udpif *, struct upcall *, size_t n_upcalls);
217 static void udpif_stop_threads(struct udpif *);
218 static void udpif_start_threads(struct udpif *, size_t n_handlers,
219 size_t n_revalidators);
220 static void *udpif_upcall_handler(void *);
221 static void *udpif_revalidator(void *);
222 static unsigned long udpif_get_n_flows(struct udpif *);
223 static void revalidate(struct revalidator *);
224 static void revalidator_sweep(struct revalidator *);
225 static void revalidator_purge(struct revalidator *);
226 static void upcall_unixctl_show(struct unixctl_conn *conn, int argc,
227 const char *argv[], void *aux);
228 static void upcall_unixctl_disable_megaflows(struct unixctl_conn *, int argc,
229 const char *argv[], void *aux);
230 static void upcall_unixctl_enable_megaflows(struct unixctl_conn *, int argc,
231 const char *argv[], void *aux);
232 static void upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, int argc,
233 const char *argv[], void *aux);
234 static void upcall_unixctl_dump_wait(struct unixctl_conn *conn, int argc,
235 const char *argv[], void *aux);
237 static struct udpif_key *ukey_create(const struct nlattr *key, size_t key_len,
239 static struct udpif_key *ukey_lookup(struct udpif *udpif,
240 const struct nlattr *key, size_t key_len,
242 static bool ukey_acquire(struct udpif *udpif, const struct nlattr *key,
243 size_t key_len, long long int used,
244 struct udpif_key **result);
245 static void ukey_delete(struct revalidator *, struct udpif_key *);
246 static enum upcall_type classify_upcall(enum dpif_upcall_type type,
247 const struct nlattr *userdata);
249 static int upcall_receive(struct upcall *, const struct dpif_backer *,
250 const struct ofpbuf *packet, enum dpif_upcall_type,
251 const struct nlattr *userdata, const struct flow *);
252 static void upcall_uninit(struct upcall *);
254 static upcall_callback upcall_cb;
256 static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true);
259 udpif_create(struct dpif_backer *backer, struct dpif *dpif)
261 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
262 struct udpif *udpif = xzalloc(sizeof *udpif);
264 if (ovsthread_once_start(&once)) {
265 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show,
267 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
268 upcall_unixctl_disable_megaflows, NULL);
269 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
270 upcall_unixctl_enable_megaflows, NULL);
271 unixctl_command_register("upcall/set-flow-limit", "", 1, 1,
272 upcall_unixctl_set_flow_limit, NULL);
273 unixctl_command_register("revalidator/wait", "", 0, 0,
274 upcall_unixctl_dump_wait, NULL);
275 ovsthread_once_done(&once);
279 udpif->backer = backer;
280 atomic_init(&udpif->flow_limit, MIN(ofproto_flow_limit, 10000));
281 udpif->secret = random_uint32();
282 udpif->reval_seq = seq_create();
283 udpif->dump_seq = seq_create();
284 latch_init(&udpif->exit_latch);
285 list_push_back(&all_udpifs, &udpif->list_node);
286 atomic_init(&udpif->n_flows, 0);
287 atomic_init(&udpif->n_flows_timestamp, LLONG_MIN);
288 ovs_mutex_init(&udpif->n_flows_mutex);
290 dpif_register_upcall_cb(dpif, upcall_cb, udpif);
296 udpif_run(struct udpif *udpif)
298 if (udpif->conns && udpif->conn_seq != seq_read(udpif->dump_seq)) {
301 for (i = 0; i < udpif->n_conns; i++) {
302 unixctl_command_reply(udpif->conns[i], NULL);
311 udpif_destroy(struct udpif *udpif)
313 udpif_stop_threads(udpif);
315 list_remove(&udpif->list_node);
316 latch_destroy(&udpif->exit_latch);
317 seq_destroy(udpif->reval_seq);
318 seq_destroy(udpif->dump_seq);
319 ovs_mutex_destroy(&udpif->n_flows_mutex);
323 /* Stops the handler and revalidator threads, must be enclosed in
324 * ovsrcu quiescent state unless when destroying udpif. */
326 udpif_stop_threads(struct udpif *udpif)
328 if (udpif && (udpif->n_handlers != 0 || udpif->n_revalidators != 0)) {
331 latch_set(&udpif->exit_latch);
333 for (i = 0; i < udpif->n_handlers; i++) {
334 struct handler *handler = &udpif->handlers[i];
336 xpthread_join(handler->thread, NULL);
339 for (i = 0; i < udpif->n_revalidators; i++) {
340 xpthread_join(udpif->revalidators[i].thread, NULL);
343 dpif_disable_upcall(udpif->dpif);
345 for (i = 0; i < udpif->n_revalidators; i++) {
346 struct revalidator *revalidator = &udpif->revalidators[i];
348 /* Delete ukeys, and delete all flows from the datapath to prevent
349 * double-counting stats. */
350 revalidator_purge(revalidator);
352 hmap_destroy(&udpif->ukeys[i].hmap);
353 ovs_mutex_destroy(&udpif->ukeys[i].mutex);
356 latch_poll(&udpif->exit_latch);
358 ovs_barrier_destroy(&udpif->reval_barrier);
360 free(udpif->revalidators);
361 udpif->revalidators = NULL;
362 udpif->n_revalidators = 0;
364 free(udpif->handlers);
365 udpif->handlers = NULL;
366 udpif->n_handlers = 0;
373 /* Starts the handler and revalidator threads, must be enclosed in
374 * ovsrcu quiescent state. */
376 udpif_start_threads(struct udpif *udpif, size_t n_handlers,
377 size_t n_revalidators)
379 if (udpif && n_handlers && n_revalidators) {
382 udpif->n_handlers = n_handlers;
383 udpif->n_revalidators = n_revalidators;
385 udpif->handlers = xzalloc(udpif->n_handlers * sizeof *udpif->handlers);
386 for (i = 0; i < udpif->n_handlers; i++) {
387 struct handler *handler = &udpif->handlers[i];
389 handler->udpif = udpif;
390 handler->handler_id = i;
391 handler->thread = ovs_thread_create(
392 "handler", udpif_upcall_handler, handler);
395 dpif_enable_upcall(udpif->dpif);
397 ovs_barrier_init(&udpif->reval_barrier, udpif->n_revalidators);
398 udpif->reval_exit = false;
399 udpif->revalidators = xzalloc(udpif->n_revalidators
400 * sizeof *udpif->revalidators);
401 udpif->ukeys = xmalloc(sizeof *udpif->ukeys * n_revalidators);
402 for (i = 0; i < udpif->n_revalidators; i++) {
403 struct revalidator *revalidator = &udpif->revalidators[i];
405 revalidator->udpif = udpif;
406 hmap_init(&udpif->ukeys[i].hmap);
407 ovs_mutex_init(&udpif->ukeys[i].mutex);
408 revalidator->ukeys = &udpif->ukeys[i].hmap;
409 revalidator->thread = ovs_thread_create(
410 "revalidator", udpif_revalidator, revalidator);
415 /* Tells 'udpif' how many threads it should use to handle upcalls.
416 * 'n_handlers' and 'n_revalidators' can never be zero. 'udpif''s
417 * datapath handle must have packet reception enabled before starting
420 udpif_set_threads(struct udpif *udpif, size_t n_handlers,
421 size_t n_revalidators)
424 ovs_assert(n_handlers && n_revalidators);
426 ovsrcu_quiesce_start();
427 if (udpif->n_handlers != n_handlers
428 || udpif->n_revalidators != n_revalidators) {
429 udpif_stop_threads(udpif);
432 if (!udpif->handlers && !udpif->revalidators) {
435 error = dpif_handlers_set(udpif->dpif, n_handlers);
437 VLOG_ERR("failed to configure handlers in dpif %s: %s",
438 dpif_name(udpif->dpif), ovs_strerror(error));
442 udpif_start_threads(udpif, n_handlers, n_revalidators);
444 ovsrcu_quiesce_end();
447 /* Waits for all ongoing upcall translations to complete. This ensures that
448 * there are no transient references to any removed ofprotos (or other
449 * objects). In particular, this should be called after an ofproto is removed
450 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
452 udpif_synchronize(struct udpif *udpif)
454 /* This is stronger than necessary. It would be sufficient to ensure
455 * (somehow) that each handler and revalidator thread had passed through
456 * its main loop once. */
457 size_t n_handlers = udpif->n_handlers;
458 size_t n_revalidators = udpif->n_revalidators;
460 ovsrcu_quiesce_start();
461 udpif_stop_threads(udpif);
462 udpif_start_threads(udpif, n_handlers, n_revalidators);
463 ovsrcu_quiesce_end();
466 /* Notifies 'udpif' that something changed which may render previous
467 * xlate_actions() results invalid. */
469 udpif_revalidate(struct udpif *udpif)
471 seq_change(udpif->reval_seq);
474 /* Returns a seq which increments every time 'udpif' pulls stats from the
475 * datapath. Callers can use this to get a sense of when might be a good time
476 * to do periodic work which relies on relatively up to date statistics. */
478 udpif_dump_seq(struct udpif *udpif)
480 return udpif->dump_seq;
484 udpif_get_memory_usage(struct udpif *udpif, struct simap *usage)
488 simap_increase(usage, "handlers", udpif->n_handlers);
490 simap_increase(usage, "revalidators", udpif->n_revalidators);
491 for (i = 0; i < udpif->n_revalidators; i++) {
492 ovs_mutex_lock(&udpif->ukeys[i].mutex);
493 simap_increase(usage, "udpif keys", hmap_count(&udpif->ukeys[i].hmap));
494 ovs_mutex_unlock(&udpif->ukeys[i].mutex);
498 /* Remove flows from a single datapath. */
500 udpif_flush(struct udpif *udpif)
502 size_t n_handlers, n_revalidators;
504 n_handlers = udpif->n_handlers;
505 n_revalidators = udpif->n_revalidators;
507 ovsrcu_quiesce_start();
509 udpif_stop_threads(udpif);
510 dpif_flow_flush(udpif->dpif);
511 udpif_start_threads(udpif, n_handlers, n_revalidators);
513 ovsrcu_quiesce_end();
516 /* Removes all flows from all datapaths. */
518 udpif_flush_all_datapaths(void)
522 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
529 udpif_get_n_flows(struct udpif *udpif)
531 long long int time, now;
532 unsigned long flow_count;
535 atomic_read_relaxed(&udpif->n_flows_timestamp, &time);
536 if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) {
537 struct dpif_dp_stats stats;
539 atomic_store_relaxed(&udpif->n_flows_timestamp, now);
540 dpif_get_dp_stats(udpif->dpif, &stats);
541 flow_count = stats.n_flows;
542 atomic_store_relaxed(&udpif->n_flows, flow_count);
543 ovs_mutex_unlock(&udpif->n_flows_mutex);
545 atomic_read_relaxed(&udpif->n_flows, &flow_count);
550 /* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH
551 * upcalls from dpif, processes the batch and installs corresponding flows
554 udpif_upcall_handler(void *arg)
556 struct handler *handler = arg;
557 struct udpif *udpif = handler->udpif;
559 while (!latch_is_set(&handler->udpif->exit_latch)) {
560 if (!recv_upcalls(handler)) {
561 dpif_recv_wait(udpif->dpif, handler->handler_id);
562 latch_wait(&udpif->exit_latch);
572 recv_upcalls(struct handler *handler)
574 struct udpif *udpif = handler->udpif;
575 uint64_t recv_stubs[UPCALL_MAX_BATCH][512 / 8];
576 struct ofpbuf recv_bufs[UPCALL_MAX_BATCH];
577 struct dpif_upcall dupcalls[UPCALL_MAX_BATCH];
578 struct upcall upcalls[UPCALL_MAX_BATCH];
582 while (n_upcalls < UPCALL_MAX_BATCH) {
583 struct ofpbuf *recv_buf = &recv_bufs[n_upcalls];
584 struct dpif_upcall *dupcall = &dupcalls[n_upcalls];
585 struct upcall *upcall = &upcalls[n_upcalls];
586 struct pkt_metadata md;
590 ofpbuf_use_stub(recv_buf, recv_stubs[n_upcalls],
591 sizeof recv_stubs[n_upcalls]);
592 if (dpif_recv(udpif->dpif, handler->handler_id, dupcall, recv_buf)) {
593 ofpbuf_uninit(recv_buf);
597 if (odp_flow_key_to_flow(dupcall->key, dupcall->key_len, &flow)
602 error = upcall_receive(upcall, udpif->backer, &dupcall->packet,
603 dupcall->type, dupcall->userdata, &flow);
605 if (error == ENODEV) {
606 /* Received packet on datapath port for which we couldn't
607 * associate an ofproto. This can happen if a port is removed
608 * while traffic is being received. Print a rate-limited
609 * message in case it happens frequently. */
610 dpif_flow_put(udpif->dpif, DPIF_FP_CREATE, dupcall->key,
611 dupcall->key_len, NULL, 0, NULL, 0, NULL);
612 VLOG_INFO_RL(&rl, "received packet on unassociated datapath "
613 "port %"PRIu32, flow.in_port.odp_port);
618 upcall->key = dupcall->key;
619 upcall->key_len = dupcall->key_len;
621 upcall->out_tun_key = dupcall->out_tun_key;
623 if (vsp_adjust_flow(upcall->ofproto, &flow, &dupcall->packet)) {
624 upcall->vsp_adjusted = true;
627 md = pkt_metadata_from_flow(&flow);
628 flow_extract(&dupcall->packet, &md, &flow);
630 error = process_upcall(udpif, upcall, NULL);
639 upcall_uninit(upcall);
641 ofpbuf_uninit(&dupcall->packet);
642 ofpbuf_uninit(recv_buf);
646 handle_upcalls(handler->udpif, upcalls, n_upcalls);
647 for (i = 0; i < n_upcalls; i++) {
648 ofpbuf_uninit(&dupcalls[i].packet);
649 ofpbuf_uninit(&recv_bufs[i]);
650 upcall_uninit(&upcalls[i]);
658 udpif_revalidator(void *arg)
660 /* Used by all revalidators. */
661 struct revalidator *revalidator = arg;
662 struct udpif *udpif = revalidator->udpif;
663 bool leader = revalidator == &udpif->revalidators[0];
665 /* Used only by the leader. */
666 long long int start_time = 0;
667 uint64_t last_reval_seq = 0;
670 revalidator->id = ovsthread_id_self();
675 reval_seq = seq_read(udpif->reval_seq);
676 udpif->need_revalidate = last_reval_seq != reval_seq;
677 last_reval_seq = reval_seq;
679 n_flows = udpif_get_n_flows(udpif);
680 udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows);
681 udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2;
683 /* Only the leader checks the exit latch to prevent a race where
684 * some threads think it's true and exit and others think it's
685 * false and block indefinitely on the reval_barrier */
686 udpif->reval_exit = latch_is_set(&udpif->exit_latch);
688 start_time = time_msec();
689 if (!udpif->reval_exit) {
690 udpif->dump = dpif_flow_dump_create(udpif->dpif);
694 /* Wait for the leader to start the flow dump. */
695 ovs_barrier_block(&udpif->reval_barrier);
696 if (udpif->reval_exit) {
699 revalidate(revalidator);
701 /* Wait for all flows to have been dumped before we garbage collect. */
702 ovs_barrier_block(&udpif->reval_barrier);
703 revalidator_sweep(revalidator);
705 /* Wait for all revalidators to finish garbage collection. */
706 ovs_barrier_block(&udpif->reval_barrier);
709 unsigned int flow_limit;
710 long long int duration;
712 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
714 dpif_flow_dump_destroy(udpif->dump);
715 seq_change(udpif->dump_seq);
717 duration = MAX(time_msec() - start_time, 1);
718 udpif->dump_duration = duration;
719 if (duration > 2000) {
720 flow_limit /= duration / 1000;
721 } else if (duration > 1300) {
722 flow_limit = flow_limit * 3 / 4;
723 } else if (duration < 1000 && n_flows > 2000
724 && flow_limit < n_flows * 1000 / duration) {
727 flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000));
728 atomic_store_relaxed(&udpif->flow_limit, flow_limit);
730 if (duration > 2000) {
731 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
735 poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500));
736 seq_wait(udpif->reval_seq, last_reval_seq);
737 latch_wait(&udpif->exit_latch);
745 static enum upcall_type
746 classify_upcall(enum dpif_upcall_type type, const struct nlattr *userdata)
748 union user_action_cookie cookie;
751 /* First look at the upcall type. */
759 case DPIF_N_UC_TYPES:
761 VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32, type);
765 /* "action" upcalls need a closer look. */
767 VLOG_WARN_RL(&rl, "action upcall missing cookie");
770 userdata_len = nl_attr_get_size(userdata);
771 if (userdata_len < sizeof cookie.type
772 || userdata_len > sizeof cookie) {
773 VLOG_WARN_RL(&rl, "action upcall cookie has unexpected size %"PRIuSIZE,
777 memset(&cookie, 0, sizeof cookie);
778 memcpy(&cookie, nl_attr_get(userdata), userdata_len);
779 if (userdata_len == MAX(8, sizeof cookie.sflow)
780 && cookie.type == USER_ACTION_COOKIE_SFLOW) {
782 } else if (userdata_len == MAX(8, sizeof cookie.slow_path)
783 && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) {
785 } else if (userdata_len == MAX(8, sizeof cookie.flow_sample)
786 && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) {
787 return FLOW_SAMPLE_UPCALL;
788 } else if (userdata_len == MAX(8, sizeof cookie.ipfix)
789 && cookie.type == USER_ACTION_COOKIE_IPFIX) {
792 VLOG_WARN_RL(&rl, "invalid user cookie of type %"PRIu16
793 " and size %"PRIuSIZE, cookie.type, userdata_len);
798 /* Calculates slow path actions for 'xout'. 'buf' must statically be
799 * initialized with at least 128 bytes of space. */
801 compose_slow_path(struct udpif *udpif, struct xlate_out *xout,
802 const struct flow *flow, odp_port_t odp_in_port,
805 union user_action_cookie cookie;
809 cookie.type = USER_ACTION_COOKIE_SLOW_PATH;
810 cookie.slow_path.unused = 0;
811 cookie.slow_path.reason = xout->slow;
813 port = xout->slow & (SLOW_CFM | SLOW_BFD | SLOW_LACP | SLOW_STP)
816 pid = dpif_port_get_pid(udpif->dpif, port, flow_hash_5tuple(flow, 0));
817 odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, ODPP_NONE,
821 /* If there is no error, the upcall must be destroyed with upcall_uninit()
822 * before quiescing, as the referred objects are guaranteed to exist only
823 * until the calling thread quiesces. Otherwise, do not call upcall_uninit()
824 * since the 'upcall->put_actions' remains uninitialized. */
826 upcall_receive(struct upcall *upcall, const struct dpif_backer *backer,
827 const struct ofpbuf *packet, enum dpif_upcall_type type,
828 const struct nlattr *userdata, const struct flow *flow)
832 error = xlate_lookup(backer, flow, &upcall->ofproto, &upcall->ipfix,
833 &upcall->sflow, NULL, &upcall->in_port);
839 upcall->packet = packet;
841 upcall->userdata = userdata;
842 ofpbuf_init(&upcall->put_actions, 0);
844 upcall->xout_initialized = false;
845 upcall->vsp_adjusted = false;
850 upcall->out_tun_key = NULL;
856 upcall_xlate(struct udpif *udpif, struct upcall *upcall,
857 struct ofpbuf *odp_actions)
859 struct dpif_flow_stats stats;
863 stats.n_bytes = ofpbuf_size(upcall->packet);
864 stats.used = time_msec();
865 stats.tcp_flags = ntohs(upcall->flow->tcp_flags);
867 xlate_in_init(&xin, upcall->ofproto, upcall->flow, upcall->in_port, NULL,
868 stats.tcp_flags, upcall->packet);
869 xin.odp_actions = odp_actions;
871 if (upcall->type == DPIF_UC_MISS) {
872 xin.resubmit_stats = &stats;
874 /* For non-miss upcalls, there's a flow in the datapath which this
875 * packet was accounted to. Presumably the revalidators will deal
876 * with pushing its stats eventually. */
879 xlate_actions(&xin, &upcall->xout);
880 upcall->xout_initialized = true;
882 /* Special case for fail-open mode.
884 * If we are in fail-open mode, but we are connected to a controller too,
885 * then we should send the packet up to the controller in the hope that it
886 * will try to set up a flow and thereby allow us to exit fail-open.
888 * See the top-level comment in fail-open.c for more information.
890 * Copy packets before they are modified by execution. */
891 if (upcall->xout.fail_open) {
892 const struct ofpbuf *packet = upcall->packet;
893 struct ofproto_packet_in *pin;
895 pin = xmalloc(sizeof *pin);
896 pin->up.packet = xmemdup(ofpbuf_data(packet), ofpbuf_size(packet));
897 pin->up.packet_len = ofpbuf_size(packet);
898 pin->up.reason = OFPR_NO_MATCH;
899 pin->up.table_id = 0;
900 pin->up.cookie = OVS_BE64_MAX;
901 flow_get_metadata(upcall->flow, &pin->up.fmd);
902 pin->send_len = 0; /* Not used for flow table misses. */
903 pin->miss_type = OFPROTO_PACKET_IN_NO_MISS;
904 ofproto_dpif_send_packet_in(upcall->ofproto, pin);
907 if (!upcall->xout.slow) {
908 ofpbuf_use_const(&upcall->put_actions,
909 ofpbuf_data(upcall->xout.odp_actions),
910 ofpbuf_size(upcall->xout.odp_actions));
912 ofpbuf_init(&upcall->put_actions, 0);
913 compose_slow_path(udpif, &upcall->xout, upcall->flow,
914 upcall->flow->in_port.odp_port,
915 &upcall->put_actions);
920 upcall_uninit(struct upcall *upcall)
923 if (upcall->xout_initialized) {
924 xlate_out_uninit(&upcall->xout);
926 ofpbuf_uninit(&upcall->put_actions);
931 upcall_cb(const struct ofpbuf *packet, const struct flow *flow,
932 enum dpif_upcall_type type, const struct nlattr *userdata,
933 struct ofpbuf *actions, struct flow_wildcards *wc,
934 struct ofpbuf *put_actions, void *aux)
936 struct udpif *udpif = aux;
937 unsigned int flow_limit;
938 struct upcall upcall;
942 atomic_read_relaxed(&enable_megaflows, &megaflow);
943 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
945 error = upcall_receive(&upcall, udpif->backer, packet, type, userdata,
951 error = process_upcall(udpif, &upcall, actions);
956 if (upcall.xout.slow && put_actions) {
957 ofpbuf_put(put_actions, ofpbuf_data(&upcall.put_actions),
958 ofpbuf_size(&upcall.put_actions));
961 if (OVS_LIKELY(wc)) {
963 /* XXX: This could be avoided with sufficient API changes. */
964 *wc = upcall.xout.wc;
966 memset(wc, 0xff, sizeof *wc);
967 flow_wildcards_clear_non_packet_fields(wc);
971 if (udpif_get_n_flows(udpif) >= flow_limit) {
976 upcall_uninit(&upcall);
981 process_upcall(struct udpif *udpif, struct upcall *upcall,
982 struct ofpbuf *odp_actions)
984 const struct nlattr *userdata = upcall->userdata;
985 const struct ofpbuf *packet = upcall->packet;
986 const struct flow *flow = upcall->flow;
988 switch (classify_upcall(upcall->type, userdata)) {
990 upcall_xlate(udpif, upcall, odp_actions);
995 union user_action_cookie cookie;
997 memset(&cookie, 0, sizeof cookie);
998 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.sflow);
999 dpif_sflow_received(upcall->sflow, packet, flow,
1000 flow->in_port.odp_port, &cookie);
1005 if (upcall->ipfix) {
1006 union user_action_cookie cookie;
1007 struct flow_tnl output_tunnel_key;
1009 memset(&cookie, 0, sizeof cookie);
1010 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.ipfix);
1012 if (upcall->out_tun_key) {
1013 memset(&output_tunnel_key, 0, sizeof output_tunnel_key);
1014 odp_tun_key_from_attr(upcall->out_tun_key,
1015 &output_tunnel_key);
1017 dpif_ipfix_bridge_sample(upcall->ipfix, packet, flow,
1018 flow->in_port.odp_port,
1019 cookie.ipfix.output_odp_port,
1020 upcall->out_tun_key ?
1021 &output_tunnel_key : NULL);
1025 case FLOW_SAMPLE_UPCALL:
1026 if (upcall->ipfix) {
1027 union user_action_cookie cookie;
1029 memset(&cookie, 0, sizeof cookie);
1030 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.flow_sample);
1032 /* The flow reflects exactly the contents of the packet.
1033 * Sample the packet using it. */
1034 dpif_ipfix_flow_sample(upcall->ipfix, packet, flow,
1035 cookie.flow_sample.collector_set_id,
1036 cookie.flow_sample.probability,
1037 cookie.flow_sample.obs_domain_id,
1038 cookie.flow_sample.obs_point_id);
1050 handle_upcalls(struct udpif *udpif, struct upcall *upcalls,
1053 struct odputil_keybuf mask_bufs[UPCALL_MAX_BATCH];
1054 struct dpif_op *opsp[UPCALL_MAX_BATCH * 2];
1055 struct dpif_op ops[UPCALL_MAX_BATCH * 2];
1056 unsigned int flow_limit;
1061 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1062 atomic_read_relaxed(&enable_megaflows, &megaflow);
1064 may_put = udpif_get_n_flows(udpif) < flow_limit;
1066 /* Handle the packets individually in order of arrival.
1068 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1069 * processes received packets for these protocols.
1071 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1074 * The loop fills 'ops' with an array of operations to execute in the
1077 for (i = 0; i < n_upcalls; i++) {
1078 struct upcall *upcall = &upcalls[i];
1079 const struct ofpbuf *packet = upcall->packet;
1082 if (upcall->vsp_adjusted) {
1083 /* This packet was received on a VLAN splinter port. We added a
1084 * VLAN to the packet to make the packet resemble the flow, but the
1085 * actions were composed assuming that the packet contained no
1086 * VLAN. So, we must remove the VLAN header from the packet before
1087 * trying to execute the actions. */
1088 if (ofpbuf_size(upcall->xout.odp_actions)) {
1089 eth_pop_vlan(CONST_CAST(struct ofpbuf *, upcall->packet));
1092 /* Remove the flow vlan tags inserted by vlan splinter logic
1093 * to ensure megaflow masks generated match the data path flow. */
1094 CONST_CAST(struct flow *, upcall->flow)->vlan_tci = 0;
1097 /* Do not install a flow into the datapath if:
1099 * - The datapath already has too many flows.
1101 * - We received this packet via some flow installed in the kernel
1103 if (may_put && upcall->type == DPIF_UC_MISS) {
1106 ofpbuf_use_stack(&mask, &mask_bufs[i], sizeof mask_bufs[i]);
1112 recirc = ofproto_dpif_get_enable_recirc(upcall->ofproto);
1113 max_mpls = ofproto_dpif_get_max_mpls_depth(upcall->ofproto);
1114 odp_flow_key_from_mask(&mask, &upcall->xout.wc.masks,
1115 upcall->flow, UINT32_MAX, max_mpls,
1120 op->type = DPIF_OP_FLOW_PUT;
1121 op->u.flow_put.flags = DPIF_FP_CREATE;
1122 op->u.flow_put.key = upcall->key;
1123 op->u.flow_put.key_len = upcall->key_len;
1124 op->u.flow_put.mask = ofpbuf_data(&mask);
1125 op->u.flow_put.mask_len = ofpbuf_size(&mask);
1126 op->u.flow_put.stats = NULL;
1127 op->u.flow_put.actions = ofpbuf_data(&upcall->put_actions);
1128 op->u.flow_put.actions_len = ofpbuf_size(&upcall->put_actions);
1131 if (ofpbuf_size(upcall->xout.odp_actions)) {
1133 op->type = DPIF_OP_EXECUTE;
1134 op->u.execute.packet = CONST_CAST(struct ofpbuf *, packet);
1135 odp_key_to_pkt_metadata(upcall->key, upcall->key_len,
1137 op->u.execute.actions = ofpbuf_data(upcall->xout.odp_actions);
1138 op->u.execute.actions_len = ofpbuf_size(upcall->xout.odp_actions);
1139 op->u.execute.needs_help = (upcall->xout.slow & SLOW_ACTION) != 0;
1143 /* Execute batch. */
1144 for (i = 0; i < n_ops; i++) {
1147 dpif_operate(udpif->dpif, opsp, n_ops);
1150 /* Must be called with udpif->ukeys[hash % udpif->n_revalidators].mutex. */
1151 static struct udpif_key *
1152 ukey_lookup(struct udpif *udpif, const struct nlattr *key, size_t key_len,
1154 OVS_REQUIRES(udpif->ukeys->mutex)
1156 struct udpif_key *ukey;
1157 struct hmap *hmap = &udpif->ukeys[hash % udpif->n_revalidators].hmap;
1159 HMAP_FOR_EACH_WITH_HASH (ukey, hmap_node, hash, hmap) {
1160 if (ukey->key_len == key_len && !memcmp(ukey->key, key, key_len)) {
1167 /* Creates a ukey for 'key' and 'key_len', returning it with ukey->mutex in
1168 * a locked state. */
1169 static struct udpif_key *
1170 ukey_create(const struct nlattr *key, size_t key_len, long long int used)
1171 OVS_NO_THREAD_SAFETY_ANALYSIS
1173 struct udpif_key *ukey = xmalloc(sizeof *ukey);
1175 ovs_mutex_init(&ukey->mutex);
1176 ukey->key = &ukey->key_buf_nla;
1177 memcpy(&ukey->key_buf, key, key_len);
1178 ukey->key_len = key_len;
1180 ovs_mutex_lock(&ukey->mutex);
1182 ukey->flow_exists = true;
1183 ukey->created = used ? used : time_msec();
1184 memset(&ukey->stats, 0, sizeof ukey->stats);
1185 ukey->xcache = NULL;
1190 /* Searches for a ukey in 'udpif->ukeys' that matches 'key' and 'key_len' and
1191 * attempts to lock the ukey. If the ukey does not exist, create it.
1193 * Returns true on success, setting *result to the matching ukey and returning
1194 * it in a locked state. Otherwise, returns false and clears *result. */
1196 ukey_acquire(struct udpif *udpif, const struct nlattr *key, size_t key_len,
1197 long long int used, struct udpif_key **result)
1198 OVS_TRY_LOCK(true, (*result)->mutex)
1200 struct udpif_key *ukey;
1202 bool locked = false;
1204 hash = hash_bytes(key, key_len, udpif->secret);
1205 idx = hash % udpif->n_revalidators;
1207 ovs_mutex_lock(&udpif->ukeys[idx].mutex);
1208 ukey = ukey_lookup(udpif, key, key_len, hash);
1210 ukey = ukey_create(key, key_len, used);
1211 hmap_insert(&udpif->ukeys[idx].hmap, &ukey->hmap_node, hash);
1213 } else if (!ovs_mutex_trylock(&ukey->mutex)) {
1216 ovs_mutex_unlock(&udpif->ukeys[idx].mutex);
1227 ukey_delete(struct revalidator *revalidator, struct udpif_key *ukey)
1228 OVS_NO_THREAD_SAFETY_ANALYSIS
1231 hmap_remove(revalidator->ukeys, &ukey->hmap_node);
1233 xlate_cache_delete(ukey->xcache);
1234 ovs_mutex_destroy(&ukey->mutex);
1239 should_revalidate(const struct udpif *udpif, uint64_t packets,
1242 long long int metric, now, duration;
1244 if (udpif->dump_duration < 200) {
1245 /* We are likely to handle full revalidation for the flows. */
1249 /* Calculate the mean time between seeing these packets. If this
1250 * exceeds the threshold, then delete the flow rather than performing
1251 * costly revalidation for flows that aren't being hit frequently.
1253 * This is targeted at situations where the dump_duration is high (~1s),
1254 * and revalidation is triggered by a call to udpif_revalidate(). In
1255 * these situations, revalidation of all flows causes fluctuations in the
1256 * flow_limit due to the interaction with the dump_duration and max_idle.
1257 * This tends to result in deletion of low-throughput flows anyway, so
1258 * skip the revalidation and just delete those flows. */
1259 packets = MAX(packets, 1);
1260 now = MAX(used, time_msec());
1261 duration = now - used;
1262 metric = duration / packets;
1265 /* The flow is receiving more than ~5pps, so keep it. */
1272 revalidate_ukey(struct udpif *udpif, struct udpif_key *ukey,
1273 const struct dpif_flow *f)
1274 OVS_REQUIRES(ukey->mutex)
1276 uint64_t slow_path_buf[128 / 8];
1277 struct xlate_out xout, *xoutp;
1278 struct netflow *netflow;
1279 struct ofproto_dpif *ofproto;
1280 struct dpif_flow_stats push;
1281 struct ofpbuf xout_actions;
1282 struct flow flow, dp_mask;
1283 uint32_t *dp32, *xout32;
1284 ofp_port_t ofp_in_port;
1285 struct xlate_in xin;
1286 long long int last_used;
1295 last_used = ukey->stats.used;
1296 push.used = f->stats.used;
1297 push.tcp_flags = f->stats.tcp_flags;
1298 push.n_packets = (f->stats.n_packets > ukey->stats.n_packets
1299 ? f->stats.n_packets - ukey->stats.n_packets
1301 push.n_bytes = (f->stats.n_bytes > ukey->stats.n_bytes
1302 ? f->stats.n_bytes - ukey->stats.n_bytes
1305 if (udpif->need_revalidate && last_used
1306 && !should_revalidate(udpif, push.n_packets, last_used)) {
1311 /* We will push the stats, so update the ukey stats cache. */
1312 ukey->stats = f->stats;
1313 if (!push.n_packets && !udpif->need_revalidate) {
1318 if (ukey->xcache && !udpif->need_revalidate) {
1319 xlate_push_stats(ukey->xcache, &push);
1324 if (odp_flow_key_to_flow(ukey->key, ukey->key_len, &flow)
1329 error = xlate_lookup(udpif->backer, &flow, &ofproto, NULL, NULL, &netflow,
1335 if (udpif->need_revalidate) {
1336 xlate_cache_clear(ukey->xcache);
1338 if (!ukey->xcache) {
1339 ukey->xcache = xlate_cache_new();
1342 xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL, push.tcp_flags,
1344 if (push.n_packets) {
1345 xin.resubmit_stats = &push;
1346 xin.may_learn = true;
1348 xin.xcache = ukey->xcache;
1349 xin.skip_wildcards = !udpif->need_revalidate;
1350 xlate_actions(&xin, &xout);
1353 if (!udpif->need_revalidate) {
1359 ofpbuf_use_const(&xout_actions, ofpbuf_data(xout.odp_actions),
1360 ofpbuf_size(xout.odp_actions));
1362 ofpbuf_use_stack(&xout_actions, slow_path_buf, sizeof slow_path_buf);
1363 compose_slow_path(udpif, &xout, &flow, flow.in_port.odp_port,
1367 if (f->actions_len != ofpbuf_size(&xout_actions)
1368 || memcmp(ofpbuf_data(&xout_actions), f->actions, f->actions_len)) {
1372 if (odp_flow_key_to_mask(f->mask, f->mask_len, &dp_mask, &flow)
1377 /* Since the kernel is free to ignore wildcarded bits in the mask, we can't
1378 * directly check that the masks are the same. Instead we check that the
1379 * mask in the kernel is more specific i.e. less wildcarded, than what
1380 * we've calculated here. This guarantees we don't catch any packets we
1381 * shouldn't with the megaflow. */
1382 dp32 = (uint32_t *) &dp_mask;
1383 xout32 = (uint32_t *) &xout.wc.masks;
1384 for (i = 0; i < FLOW_U32S; i++) {
1385 if ((dp32[i] | xout32[i]) != dp32[i]) {
1392 if (netflow && !ok) {
1393 netflow_flow_clear(netflow, &flow);
1395 xlate_out_uninit(xoutp);
1400 struct udpif_key *ukey;
1401 struct dpif_flow_stats stats; /* Stats for 'op'. */
1402 struct dpif_op op; /* Flow del operation. */
1406 dump_op_init(struct dump_op *op, const struct nlattr *key, size_t key_len,
1407 struct udpif_key *ukey)
1410 op->op.type = DPIF_OP_FLOW_DEL;
1411 op->op.u.flow_del.key = key;
1412 op->op.u.flow_del.key_len = key_len;
1413 op->op.u.flow_del.stats = &op->stats;
1417 push_dump_ops__(struct udpif *udpif, struct dump_op *ops, size_t n_ops)
1419 struct dpif_op *opsp[REVALIDATE_MAX_BATCH];
1422 ovs_assert(n_ops <= REVALIDATE_MAX_BATCH);
1423 for (i = 0; i < n_ops; i++) {
1424 opsp[i] = &ops[i].op;
1426 dpif_operate(udpif->dpif, opsp, n_ops);
1428 for (i = 0; i < n_ops; i++) {
1429 struct dump_op *op = &ops[i];
1430 struct dpif_flow_stats *push, *stats, push_buf;
1432 stats = op->op.u.flow_del.stats;
1435 ovs_mutex_lock(&op->ukey->mutex);
1436 push->used = MAX(stats->used, op->ukey->stats.used);
1437 push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags;
1438 push->n_packets = stats->n_packets - op->ukey->stats.n_packets;
1439 push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes;
1440 ovs_mutex_unlock(&op->ukey->mutex);
1442 if (push->n_packets || netflow_exists()) {
1443 struct ofproto_dpif *ofproto;
1444 struct netflow *netflow;
1445 ofp_port_t ofp_in_port;
1449 ovs_mutex_lock(&op->ukey->mutex);
1450 if (op->ukey->xcache) {
1451 xlate_push_stats(op->ukey->xcache, push);
1452 ovs_mutex_unlock(&op->ukey->mutex);
1455 ovs_mutex_unlock(&op->ukey->mutex);
1457 if (odp_flow_key_to_flow(op->op.u.flow_del.key,
1458 op->op.u.flow_del.key_len, &flow)
1463 error = xlate_lookup(udpif->backer, &flow, &ofproto,
1464 NULL, NULL, &netflow, &ofp_in_port);
1466 struct xlate_in xin;
1468 xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL,
1469 push->tcp_flags, NULL);
1470 xin.resubmit_stats = push->n_packets ? push : NULL;
1471 xin.may_learn = push->n_packets > 0;
1472 xin.skip_wildcards = true;
1473 xlate_actions_for_side_effects(&xin);
1476 netflow_flow_clear(netflow, &flow);
1484 push_dump_ops(struct revalidator *revalidator,
1485 struct dump_op *ops, size_t n_ops)
1489 push_dump_ops__(revalidator->udpif, ops, n_ops);
1490 for (i = 0; i < n_ops; i++) {
1491 ukey_delete(revalidator, ops[i].ukey);
1496 revalidate(struct revalidator *revalidator)
1498 struct udpif *udpif = revalidator->udpif;
1499 struct dpif_flow_dump_thread *dump_thread;
1501 unsigned int flow_limit;
1503 dump_seq = seq_read(udpif->dump_seq);
1504 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1505 dump_thread = dpif_flow_dump_thread_create(udpif->dump);
1507 struct dump_op ops[REVALIDATE_MAX_BATCH];
1510 struct dpif_flow flows[REVALIDATE_MAX_BATCH];
1511 const struct dpif_flow *f;
1514 long long int max_idle;
1519 n_dumped = dpif_flow_dump_next(dump_thread, flows, ARRAY_SIZE(flows));
1526 /* In normal operation we want to keep flows around until they have
1527 * been idle for 'ofproto_max_idle' milliseconds. However:
1529 * - If the number of datapath flows climbs above 'flow_limit',
1530 * drop that down to 100 ms to try to bring the flows down to
1533 * - If the number of datapath flows climbs above twice
1534 * 'flow_limit', delete all the datapath flows as an emergency
1535 * measure. (We reassess this condition for the next batch of
1536 * datapath flows, so we will recover before all the flows are
1538 n_dp_flows = udpif_get_n_flows(udpif);
1539 kill_them_all = n_dp_flows > flow_limit * 2;
1540 max_idle = n_dp_flows > flow_limit ? 100 : ofproto_max_idle;
1542 for (f = flows; f < &flows[n_dumped]; f++) {
1543 long long int used = f->stats.used;
1544 struct udpif_key *ukey;
1545 bool already_dumped, keep;
1547 if (!ukey_acquire(udpif, f->key, f->key_len, used, &ukey)) {
1548 /* We couldn't acquire the ukey. This means that
1549 * another revalidator is processing this flow
1550 * concurrently, so don't bother processing it. */
1551 COVERAGE_INC(upcall_duplicate_flow);
1555 already_dumped = ukey->dump_seq == dump_seq;
1556 if (already_dumped) {
1557 /* The flow has already been dumped and handled by another
1558 * revalidator during this flow dump operation. Skip it. */
1559 COVERAGE_INC(upcall_duplicate_flow);
1560 ovs_mutex_unlock(&ukey->mutex);
1565 used = ukey->created;
1567 if (kill_them_all || (used && used < now - max_idle)) {
1570 keep = revalidate_ukey(udpif, ukey, f);
1572 ukey->dump_seq = dump_seq;
1573 ukey->flow_exists = keep;
1576 dump_op_init(&ops[n_ops++], f->key, f->key_len, ukey);
1578 ovs_mutex_unlock(&ukey->mutex);
1582 push_dump_ops__(udpif, ops, n_ops);
1585 dpif_flow_dump_thread_destroy(dump_thread);
1588 /* Called with exclusive access to 'revalidator' and 'ukey'. */
1590 handle_missed_revalidation(struct revalidator *revalidator,
1591 struct udpif_key *ukey)
1592 OVS_NO_THREAD_SAFETY_ANALYSIS
1594 struct udpif *udpif = revalidator->udpif;
1595 struct dpif_flow flow;
1597 uint64_t stub[DPIF_FLOW_BUFSIZE / 8];
1600 COVERAGE_INC(revalidate_missed_dp_flow);
1602 ofpbuf_use_stub(&buf, &stub, sizeof stub);
1603 if (!dpif_flow_get(udpif->dpif, ukey->key, ukey->key_len, &buf, &flow)) {
1604 keep = revalidate_ukey(udpif, ukey, &flow);
1606 ofpbuf_uninit(&buf);
1612 revalidator_sweep__(struct revalidator *revalidator, bool purge)
1613 OVS_NO_THREAD_SAFETY_ANALYSIS
1615 struct dump_op ops[REVALIDATE_MAX_BATCH];
1616 struct udpif_key *ukey, *next;
1621 dump_seq = seq_read(revalidator->udpif->dump_seq);
1623 /* During garbage collection, this revalidator completely owns its ukeys
1624 * map, and therefore doesn't need to do any locking. */
1625 HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, revalidator->ukeys) {
1626 if (ukey->flow_exists
1628 || (ukey->dump_seq != dump_seq
1629 && revalidator->udpif->need_revalidate
1630 && !handle_missed_revalidation(revalidator, ukey)))) {
1631 struct dump_op *op = &ops[n_ops++];
1633 dump_op_init(op, ukey->key, ukey->key_len, ukey);
1634 if (n_ops == REVALIDATE_MAX_BATCH) {
1635 push_dump_ops(revalidator, ops, n_ops);
1638 } else if (!ukey->flow_exists) {
1639 ukey_delete(revalidator, ukey);
1644 push_dump_ops(revalidator, ops, n_ops);
1649 revalidator_sweep(struct revalidator *revalidator)
1651 revalidator_sweep__(revalidator, false);
1655 revalidator_purge(struct revalidator *revalidator)
1657 revalidator_sweep__(revalidator, true);
1661 upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED,
1662 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
1664 struct ds ds = DS_EMPTY_INITIALIZER;
1665 struct udpif *udpif;
1667 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1668 unsigned int flow_limit;
1671 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1673 ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif));
1674 ds_put_format(&ds, "\tflows : (current %lu)"
1675 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif),
1676 udpif->avg_n_flows, udpif->max_n_flows, flow_limit);
1677 ds_put_format(&ds, "\tdump duration : %lldms\n", udpif->dump_duration);
1679 ds_put_char(&ds, '\n');
1680 for (i = 0; i < n_revalidators; i++) {
1681 struct revalidator *revalidator = &udpif->revalidators[i];
1683 ovs_mutex_lock(&udpif->ukeys[i].mutex);
1684 ds_put_format(&ds, "\t%u: (keys %"PRIuSIZE")\n",
1685 revalidator->id, hmap_count(&udpif->ukeys[i].hmap));
1686 ovs_mutex_unlock(&udpif->ukeys[i].mutex);
1690 unixctl_command_reply(conn, ds_cstr(&ds));
1694 /* Disable using the megaflows.
1696 * This command is only needed for advanced debugging, so it's not
1697 * documented in the man page. */
1699 upcall_unixctl_disable_megaflows(struct unixctl_conn *conn,
1700 int argc OVS_UNUSED,
1701 const char *argv[] OVS_UNUSED,
1702 void *aux OVS_UNUSED)
1704 atomic_store_relaxed(&enable_megaflows, false);
1705 udpif_flush_all_datapaths();
1706 unixctl_command_reply(conn, "megaflows disabled");
1709 /* Re-enable using megaflows.
1711 * This command is only needed for advanced debugging, so it's not
1712 * documented in the man page. */
1714 upcall_unixctl_enable_megaflows(struct unixctl_conn *conn,
1715 int argc OVS_UNUSED,
1716 const char *argv[] OVS_UNUSED,
1717 void *aux OVS_UNUSED)
1719 atomic_store_relaxed(&enable_megaflows, true);
1720 udpif_flush_all_datapaths();
1721 unixctl_command_reply(conn, "megaflows enabled");
1724 /* Set the flow limit.
1726 * This command is only needed for advanced debugging, so it's not
1727 * documented in the man page. */
1729 upcall_unixctl_set_flow_limit(struct unixctl_conn *conn,
1730 int argc OVS_UNUSED,
1731 const char *argv[] OVS_UNUSED,
1732 void *aux OVS_UNUSED)
1734 struct ds ds = DS_EMPTY_INITIALIZER;
1735 struct udpif *udpif;
1736 unsigned int flow_limit = atoi(argv[1]);
1738 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1739 atomic_store_relaxed(&udpif->flow_limit, flow_limit);
1741 ds_put_format(&ds, "set flow_limit to %u\n", flow_limit);
1742 unixctl_command_reply(conn, ds_cstr(&ds));
1747 upcall_unixctl_dump_wait(struct unixctl_conn *conn,
1748 int argc OVS_UNUSED,
1749 const char *argv[] OVS_UNUSED,
1750 void *aux OVS_UNUSED)
1752 if (list_is_singleton(&all_udpifs)) {
1753 struct udpif *udpif = NULL;
1756 udpif = OBJECT_CONTAINING(list_front(&all_udpifs), udpif, list_node);
1757 len = (udpif->n_conns + 1) * sizeof *udpif->conns;
1758 udpif->conn_seq = seq_read(udpif->dump_seq);
1759 udpif->conns = xrealloc(udpif->conns, len);
1760 udpif->conns[udpif->n_conns++] = conn;
1762 unixctl_command_reply_error(conn, "can't wait on multiple udpifs.");