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 FLOW_MISS_MAX_BATCH 50
44 #define REVALIDATE_MAX_BATCH 50
46 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall);
48 COVERAGE_DEFINE(upcall_queue_overflow);
50 /* A thread that processes each upcall handed to it by the dispatcher thread,
51 * forwards the upcall's packet, and possibly sets up a kernel flow as a
54 struct udpif *udpif; /* Parent udpif. */
55 pthread_t thread; /* Thread ID. */
56 char *name; /* Thread name. */
58 struct ovs_mutex mutex; /* Mutex guarding the following. */
60 /* Atomic queue of unprocessed upcalls. */
61 struct list upcalls OVS_GUARDED;
62 size_t n_upcalls OVS_GUARDED;
64 bool need_signal; /* Only changed by the dispatcher. */
66 pthread_cond_t wake_cond; /* Wakes 'thread' while holding
70 /* A thread that processes each kernel flow handed to it by the flow_dumper
71 * thread, updates OpenFlow statistics, and updates or removes the kernel flow
74 struct udpif *udpif; /* Parent udpif. */
75 char *name; /* Thread name. */
77 pthread_t thread; /* Thread ID. */
78 struct hmap ukeys; /* Datapath flow keys. */
82 struct ovs_mutex mutex; /* Mutex guarding the following. */
83 pthread_cond_t wake_cond;
84 struct list udumps OVS_GUARDED; /* Unprocessed udumps. */
85 size_t n_udumps OVS_GUARDED; /* Number of unprocessed udumps. */
88 /* An upcall handler for ofproto_dpif.
90 * udpif has two logically separate pieces:
92 * - A "dispatcher" thread that reads upcalls from the kernel and dispatches
93 * them to one of several "handler" threads (see struct handler).
95 * - A "flow_dumper" thread that reads the kernel flow table and dispatches
96 * flows to one of several "revalidator" threads (see struct
99 struct list list_node; /* In all_udpifs list. */
101 struct dpif *dpif; /* Datapath handle. */
102 struct dpif_backer *backer; /* Opaque dpif_backer pointer. */
104 uint32_t secret; /* Random seed for upcall hash. */
106 pthread_t dispatcher; /* Dispatcher thread ID. */
107 pthread_t flow_dumper; /* Flow dumper thread ID. */
109 struct handler *handlers; /* Upcall handlers. */
112 struct revalidator *revalidators; /* Flow revalidators. */
113 size_t n_revalidators;
115 uint64_t last_reval_seq; /* 'reval_seq' at last revalidation. */
116 struct seq *reval_seq; /* Incremented to force revalidation. */
118 struct seq *dump_seq; /* Increments each dump iteration. */
120 struct latch exit_latch; /* Tells child threads to exit. */
122 long long int dump_duration; /* Duration of the last flow dump. */
124 /* Datapath flow statistics. */
125 unsigned int max_n_flows;
126 unsigned int avg_n_flows;
128 /* Following fields are accessed and modified by different threads. */
129 atomic_uint flow_limit; /* Datapath flow hard limit. */
131 /* n_flows_mutex prevents multiple threads updating these concurrently. */
132 atomic_uint64_t n_flows; /* Number of flows in the datapath. */
133 atomic_llong n_flows_timestamp; /* Last time n_flows was updated. */
134 struct ovs_mutex n_flows_mutex;
138 BAD_UPCALL, /* Some kind of bug somewhere. */
139 MISS_UPCALL, /* A flow miss. */
140 SFLOW_UPCALL, /* sFlow sample. */
141 FLOW_SAMPLE_UPCALL, /* Per-flow sampling. */
142 IPFIX_UPCALL /* Per-bridge sampling. */
146 struct list list_node; /* For queuing upcalls. */
147 struct flow_miss *flow_miss; /* This upcall's flow_miss. */
149 /* Raw upcall plus data for keeping track of the memory backing it. */
150 struct dpif_upcall dpif_upcall; /* As returned by dpif_recv() */
151 struct ofpbuf upcall_buf; /* Owns some data in 'dpif_upcall'. */
152 uint64_t upcall_stub[512 / 8]; /* Buffer to reduce need for malloc(). */
155 /* 'udpif_key's are responsible for tracking the little bit of state udpif
156 * needs to do flow expiration which can't be pulled directly from the
157 * datapath. They are owned, created by, maintained, and destroyed by a single
158 * revalidator making them easy to efficiently handle with multiple threads. */
160 struct hmap_node hmap_node; /* In parent revalidator 'ukeys' map. */
162 struct nlattr *key; /* Datapath flow key. */
163 size_t key_len; /* Length of 'key'. */
165 struct dpif_flow_stats stats; /* Stats at most recent flow dump. */
166 long long int created; /* Estimation of creation time. */
168 bool mark; /* Used by mark and sweep GC algorithm. */
170 struct odputil_keybuf key_buf; /* Memory for 'key'. */
173 /* 'udpif_flow_dump's hold the state associated with one iteration in a flow
174 * dump operation. This is created by the flow_dumper thread and handed to the
175 * appropriate revalidator thread to be processed. */
176 struct udpif_flow_dump {
177 struct list list_node;
179 struct nlattr *key; /* Datapath flow key. */
180 size_t key_len; /* Length of 'key'. */
181 uint32_t key_hash; /* Hash of 'key'. */
183 struct odputil_keybuf mask_buf;
184 struct nlattr *mask; /* Datapath mask for 'key'. */
185 size_t mask_len; /* Length of 'mask'. */
187 struct dpif_flow_stats stats; /* Stats pulled from the datapath. */
189 bool need_revalidate; /* Key needs revalidation? */
191 struct odputil_keybuf key_buf;
194 /* Flow miss batching.
196 * Some dpifs implement operations faster when you hand them off in a batch.
197 * To allow batching, "struct flow_miss" queues the dpif-related work needed
198 * for a given flow. Each "struct flow_miss" corresponds to sending one or
199 * more packets, plus possibly installing the flow in the dpif. */
201 struct hmap_node hmap_node;
202 struct ofproto_dpif *ofproto;
205 const struct nlattr *key;
207 enum dpif_upcall_type upcall_type;
208 struct dpif_flow_stats stats;
209 odp_port_t odp_in_port;
211 uint64_t slow_path_buf[128 / 8];
212 struct odputil_keybuf mask_buf;
214 struct xlate_out xout;
219 static void upcall_destroy(struct upcall *);
221 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
222 static struct list all_udpifs = LIST_INITIALIZER(&all_udpifs);
224 static void recv_upcalls(struct udpif *);
225 static void handle_upcalls(struct handler *handler, struct list *upcalls);
226 static void *udpif_flow_dumper(void *);
227 static void *udpif_dispatcher(void *);
228 static void *udpif_upcall_handler(void *);
229 static void *udpif_revalidator(void *);
230 static uint64_t udpif_get_n_flows(struct udpif *);
231 static void revalidate_udumps(struct revalidator *, struct list *udumps);
232 static void revalidator_sweep(struct revalidator *);
233 static void revalidator_purge(struct revalidator *);
234 static void upcall_unixctl_show(struct unixctl_conn *conn, int argc,
235 const char *argv[], void *aux);
236 static void upcall_unixctl_disable_megaflows(struct unixctl_conn *, int argc,
237 const char *argv[], void *aux);
238 static void upcall_unixctl_enable_megaflows(struct unixctl_conn *, int argc,
239 const char *argv[], void *aux);
240 static void upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, int argc,
241 const char *argv[], void *aux);
242 static void ukey_delete(struct revalidator *, struct udpif_key *);
244 static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true);
247 udpif_create(struct dpif_backer *backer, struct dpif *dpif)
249 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
250 struct udpif *udpif = xzalloc(sizeof *udpif);
252 if (ovsthread_once_start(&once)) {
253 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show,
255 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
256 upcall_unixctl_disable_megaflows, NULL);
257 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
258 upcall_unixctl_enable_megaflows, NULL);
259 unixctl_command_register("upcall/set-flow-limit", "", 1, 1,
260 upcall_unixctl_set_flow_limit, NULL);
261 ovsthread_once_done(&once);
265 udpif->backer = backer;
266 atomic_init(&udpif->flow_limit, MIN(ofproto_flow_limit, 10000));
267 udpif->secret = random_uint32();
268 udpif->reval_seq = seq_create();
269 udpif->dump_seq = seq_create();
270 latch_init(&udpif->exit_latch);
271 list_push_back(&all_udpifs, &udpif->list_node);
272 atomic_init(&udpif->n_flows, 0);
273 atomic_init(&udpif->n_flows_timestamp, LLONG_MIN);
274 ovs_mutex_init(&udpif->n_flows_mutex);
280 udpif_destroy(struct udpif *udpif)
282 udpif_set_threads(udpif, 0, 0);
285 list_remove(&udpif->list_node);
286 latch_destroy(&udpif->exit_latch);
287 seq_destroy(udpif->reval_seq);
288 seq_destroy(udpif->dump_seq);
289 ovs_mutex_destroy(&udpif->n_flows_mutex);
293 /* Tells 'udpif' how many threads it should use to handle upcalls. Disables
294 * all threads if 'n_handlers' and 'n_revalidators' is zero. 'udpif''s
295 * datapath handle must have packet reception enabled before starting threads.
298 udpif_set_threads(struct udpif *udpif, size_t n_handlers,
299 size_t n_revalidators)
303 ovsrcu_quiesce_start();
304 /* Stop the old threads (if any). */
305 if (udpif->handlers &&
306 (udpif->n_handlers != n_handlers
307 || udpif->n_revalidators != n_revalidators)) {
310 latch_set(&udpif->exit_latch);
312 for (i = 0; i < udpif->n_handlers; i++) {
313 struct handler *handler = &udpif->handlers[i];
315 ovs_mutex_lock(&handler->mutex);
316 xpthread_cond_signal(&handler->wake_cond);
317 ovs_mutex_unlock(&handler->mutex);
318 xpthread_join(handler->thread, NULL);
321 for (i = 0; i < udpif->n_revalidators; i++) {
322 struct revalidator *revalidator = &udpif->revalidators[i];
324 ovs_mutex_lock(&revalidator->mutex);
325 xpthread_cond_signal(&revalidator->wake_cond);
326 ovs_mutex_unlock(&revalidator->mutex);
327 xpthread_join(revalidator->thread, NULL);
330 xpthread_join(udpif->flow_dumper, NULL);
331 xpthread_join(udpif->dispatcher, NULL);
333 for (i = 0; i < udpif->n_revalidators; i++) {
334 struct revalidator *revalidator = &udpif->revalidators[i];
335 struct udpif_flow_dump *udump, *next_udump;
337 LIST_FOR_EACH_SAFE (udump, next_udump, list_node,
338 &revalidator->udumps) {
339 list_remove(&udump->list_node);
343 /* Delete ukeys, and delete all flows from the datapath to prevent
344 * double-counting stats. */
345 revalidator_purge(revalidator);
346 hmap_destroy(&revalidator->ukeys);
347 ovs_mutex_destroy(&revalidator->mutex);
349 free(revalidator->name);
352 for (i = 0; i < udpif->n_handlers; i++) {
353 struct handler *handler = &udpif->handlers[i];
354 struct upcall *miss, *next;
356 LIST_FOR_EACH_SAFE (miss, next, list_node, &handler->upcalls) {
357 list_remove(&miss->list_node);
358 upcall_destroy(miss);
360 ovs_mutex_destroy(&handler->mutex);
362 xpthread_cond_destroy(&handler->wake_cond);
365 latch_poll(&udpif->exit_latch);
367 free(udpif->revalidators);
368 udpif->revalidators = NULL;
369 udpif->n_revalidators = 0;
371 free(udpif->handlers);
372 udpif->handlers = NULL;
373 udpif->n_handlers = 0;
376 error = dpif_handlers_set(udpif->dpif, 1);
378 VLOG_ERR("failed to configure handlers in dpif %s: %s",
379 dpif_name(udpif->dpif), ovs_strerror(error));
383 /* Start new threads (if necessary). */
384 if (!udpif->handlers && n_handlers) {
387 udpif->n_handlers = n_handlers;
388 udpif->n_revalidators = n_revalidators;
390 udpif->handlers = xzalloc(udpif->n_handlers * sizeof *udpif->handlers);
391 for (i = 0; i < udpif->n_handlers; i++) {
392 struct handler *handler = &udpif->handlers[i];
394 handler->udpif = udpif;
395 list_init(&handler->upcalls);
396 handler->need_signal = false;
397 xpthread_cond_init(&handler->wake_cond, NULL);
398 ovs_mutex_init(&handler->mutex);
399 xpthread_create(&handler->thread, NULL, udpif_upcall_handler,
403 udpif->revalidators = xzalloc(udpif->n_revalidators
404 * sizeof *udpif->revalidators);
405 for (i = 0; i < udpif->n_revalidators; i++) {
406 struct revalidator *revalidator = &udpif->revalidators[i];
408 revalidator->udpif = udpif;
409 list_init(&revalidator->udumps);
410 hmap_init(&revalidator->ukeys);
411 ovs_mutex_init(&revalidator->mutex);
412 xpthread_cond_init(&revalidator->wake_cond, NULL);
413 xpthread_create(&revalidator->thread, NULL, udpif_revalidator,
416 xpthread_create(&udpif->dispatcher, NULL, udpif_dispatcher, udpif);
417 xpthread_create(&udpif->flow_dumper, NULL, udpif_flow_dumper, udpif);
420 ovsrcu_quiesce_end();
423 /* Waits for all ongoing upcall translations to complete. This ensures that
424 * there are no transient references to any removed ofprotos (or other
425 * objects). In particular, this should be called after an ofproto is removed
426 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
428 udpif_synchronize(struct udpif *udpif)
430 /* This is stronger than necessary. It would be sufficient to ensure
431 * (somehow) that each handler and revalidator thread had passed through
432 * its main loop once. */
433 size_t n_handlers = udpif->n_handlers;
434 size_t n_revalidators = udpif->n_revalidators;
435 udpif_set_threads(udpif, 0, 0);
436 udpif_set_threads(udpif, n_handlers, n_revalidators);
439 /* Notifies 'udpif' that something changed which may render previous
440 * xlate_actions() results invalid. */
442 udpif_revalidate(struct udpif *udpif)
444 seq_change(udpif->reval_seq);
447 /* Returns a seq which increments every time 'udpif' pulls stats from the
448 * datapath. Callers can use this to get a sense of when might be a good time
449 * to do periodic work which relies on relatively up to date statistics. */
451 udpif_dump_seq(struct udpif *udpif)
453 return udpif->dump_seq;
457 udpif_get_memory_usage(struct udpif *udpif, struct simap *usage)
461 simap_increase(usage, "dispatchers", 1);
462 simap_increase(usage, "flow_dumpers", 1);
464 simap_increase(usage, "handlers", udpif->n_handlers);
465 for (i = 0; i < udpif->n_handlers; i++) {
466 struct handler *handler = &udpif->handlers[i];
467 ovs_mutex_lock(&handler->mutex);
468 simap_increase(usage, "handler upcalls", handler->n_upcalls);
469 ovs_mutex_unlock(&handler->mutex);
472 simap_increase(usage, "revalidators", udpif->n_revalidators);
473 for (i = 0; i < udpif->n_revalidators; i++) {
474 struct revalidator *revalidator = &udpif->revalidators[i];
475 ovs_mutex_lock(&revalidator->mutex);
476 simap_increase(usage, "revalidator dumps", revalidator->n_udumps);
478 /* XXX: This isn't technically thread safe because the revalidator
479 * ukeys maps isn't protected by a mutex since it's per thread. */
480 simap_increase(usage, "revalidator keys",
481 hmap_count(&revalidator->ukeys));
482 ovs_mutex_unlock(&revalidator->mutex);
486 /* Remove flows from a single datapath. */
488 udpif_flush(struct udpif *udpif)
490 size_t n_handlers, n_revalidators;
492 n_handlers = udpif->n_handlers;
493 n_revalidators = udpif->n_revalidators;
495 udpif_set_threads(udpif, 0, 0);
496 dpif_flow_flush(udpif->dpif);
497 udpif_set_threads(udpif, n_handlers, n_revalidators);
500 /* Removes all flows from all datapaths. */
502 udpif_flush_all_datapaths(void)
506 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
512 /* Destroys and deallocates 'upcall'. */
514 upcall_destroy(struct upcall *upcall)
517 ofpbuf_uninit(&upcall->dpif_upcall.packet);
518 ofpbuf_uninit(&upcall->upcall_buf);
524 udpif_get_n_flows(struct udpif *udpif)
526 long long int time, now;
530 atomic_read(&udpif->n_flows_timestamp, &time);
531 if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) {
532 struct dpif_dp_stats stats;
534 atomic_store(&udpif->n_flows_timestamp, now);
535 dpif_get_dp_stats(udpif->dpif, &stats);
536 flow_count = stats.n_flows;
537 atomic_store(&udpif->n_flows, flow_count);
538 ovs_mutex_unlock(&udpif->n_flows_mutex);
540 atomic_read(&udpif->n_flows, &flow_count);
545 /* The dispatcher thread is responsible for receiving upcalls from the kernel,
546 * assigning them to a upcall_handler thread. */
548 udpif_dispatcher(void *arg)
550 struct udpif *udpif = arg;
552 set_subprogram_name("dispatcher");
553 while (!latch_is_set(&udpif->exit_latch)) {
555 dpif_recv_wait(udpif->dpif, 0);
556 latch_wait(&udpif->exit_latch);
564 udpif_flow_dumper(void *arg)
566 struct udpif *udpif = arg;
568 set_subprogram_name("flow_dumper");
569 while (!latch_is_set(&udpif->exit_latch)) {
570 const struct dpif_flow_stats *stats;
571 long long int start_time, duration;
572 const struct nlattr *key, *mask;
573 struct dpif_flow_dump dump;
574 size_t key_len, mask_len;
575 unsigned int flow_limit;
576 bool need_revalidate;
582 reval_seq = seq_read(udpif->reval_seq);
583 need_revalidate = udpif->last_reval_seq != reval_seq;
584 udpif->last_reval_seq = reval_seq;
586 n_flows = udpif_get_n_flows(udpif);
587 udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows);
588 udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2;
590 start_time = time_msec();
591 error = dpif_flow_dump_start(&dump, udpif->dpif);
593 VLOG_INFO("Failed to start flow dump (%s)", ovs_strerror(error));
596 dpif_flow_dump_state_init(udpif->dpif, &state);
597 while (dpif_flow_dump_next(&dump, state, &key, &key_len,
598 &mask, &mask_len, NULL, NULL, &stats)
599 && !latch_is_set(&udpif->exit_latch)) {
600 struct udpif_flow_dump *udump = xmalloc(sizeof *udump);
601 struct revalidator *revalidator;
603 udump->key_hash = hash_bytes(key, key_len, udpif->secret);
604 memcpy(&udump->key_buf, key, key_len);
605 udump->key = (struct nlattr *) &udump->key_buf;
606 udump->key_len = key_len;
608 memcpy(&udump->mask_buf, mask, mask_len);
609 udump->mask = (struct nlattr *) &udump->mask_buf;
610 udump->mask_len = mask_len;
612 udump->stats = *stats;
613 udump->need_revalidate = need_revalidate;
615 revalidator = &udpif->revalidators[udump->key_hash
616 % udpif->n_revalidators];
618 ovs_mutex_lock(&revalidator->mutex);
619 while (revalidator->n_udumps >= REVALIDATE_MAX_BATCH * 3
620 && !latch_is_set(&udpif->exit_latch)) {
621 ovs_mutex_cond_wait(&revalidator->wake_cond,
622 &revalidator->mutex);
624 list_push_back(&revalidator->udumps, &udump->list_node);
625 revalidator->n_udumps++;
626 xpthread_cond_signal(&revalidator->wake_cond);
627 ovs_mutex_unlock(&revalidator->mutex);
629 dpif_flow_dump_state_uninit(udpif->dpif, state);
630 dpif_flow_dump_done(&dump);
632 /* Let all the revalidators finish and garbage collect. */
633 seq_change(udpif->dump_seq);
634 for (i = 0; i < udpif->n_revalidators; i++) {
635 struct revalidator *revalidator = &udpif->revalidators[i];
636 ovs_mutex_lock(&revalidator->mutex);
637 xpthread_cond_signal(&revalidator->wake_cond);
638 ovs_mutex_unlock(&revalidator->mutex);
641 for (i = 0; i < udpif->n_revalidators; i++) {
642 struct revalidator *revalidator = &udpif->revalidators[i];
644 ovs_mutex_lock(&revalidator->mutex);
645 while (revalidator->dump_seq != seq_read(udpif->dump_seq)
646 && !latch_is_set(&udpif->exit_latch)) {
647 ovs_mutex_cond_wait(&revalidator->wake_cond,
648 &revalidator->mutex);
650 ovs_mutex_unlock(&revalidator->mutex);
653 duration = MAX(time_msec() - start_time, 1);
654 udpif->dump_duration = duration;
655 atomic_read(&udpif->flow_limit, &flow_limit);
656 if (duration > 2000) {
657 flow_limit /= duration / 1000;
658 } else if (duration > 1300) {
659 flow_limit = flow_limit * 3 / 4;
660 } else if (duration < 1000 && n_flows > 2000
661 && flow_limit < n_flows * 1000 / duration) {
664 flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000));
665 atomic_store(&udpif->flow_limit, flow_limit);
667 if (duration > 2000) {
668 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
673 poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500));
674 seq_wait(udpif->reval_seq, udpif->last_reval_seq);
675 latch_wait(&udpif->exit_latch);
682 /* The miss handler thread is responsible for processing miss upcalls retrieved
683 * by the dispatcher thread. Once finished it passes the processed miss
684 * upcalls to ofproto-dpif where they're installed in the datapath. */
686 udpif_upcall_handler(void *arg)
688 struct handler *handler = arg;
690 handler->name = xasprintf("handler_%u", ovsthread_id_self());
691 set_subprogram_name("%s", handler->name);
693 while (!latch_is_set(&handler->udpif->exit_latch)) {
694 struct list misses = LIST_INITIALIZER(&misses);
697 ovs_mutex_lock(&handler->mutex);
698 if (!handler->n_upcalls) {
699 ovs_mutex_cond_wait(&handler->wake_cond, &handler->mutex);
702 for (i = 0; i < FLOW_MISS_MAX_BATCH; i++) {
703 if (handler->n_upcalls) {
704 handler->n_upcalls--;
705 list_push_back(&misses, list_pop_front(&handler->upcalls));
710 ovs_mutex_unlock(&handler->mutex);
712 handle_upcalls(handler, &misses);
721 udpif_revalidator(void *arg)
723 struct revalidator *revalidator = arg;
725 revalidator->name = xasprintf("revalidator_%u", ovsthread_id_self());
726 set_subprogram_name("%s", revalidator->name);
728 struct list udumps = LIST_INITIALIZER(&udumps);
729 struct udpif *udpif = revalidator->udpif;
732 ovs_mutex_lock(&revalidator->mutex);
733 if (latch_is_set(&udpif->exit_latch)) {
734 ovs_mutex_unlock(&revalidator->mutex);
738 if (!revalidator->n_udumps) {
739 if (revalidator->dump_seq != seq_read(udpif->dump_seq)) {
740 revalidator->dump_seq = seq_read(udpif->dump_seq);
741 revalidator_sweep(revalidator);
743 ovs_mutex_cond_wait(&revalidator->wake_cond,
744 &revalidator->mutex);
748 for (i = 0; i < REVALIDATE_MAX_BATCH && revalidator->n_udumps; i++) {
749 list_push_back(&udumps, list_pop_front(&revalidator->udumps));
750 revalidator->n_udumps--;
753 /* Wake up the flow dumper. */
754 xpthread_cond_signal(&revalidator->wake_cond);
755 ovs_mutex_unlock(&revalidator->mutex);
757 if (!list_is_empty(&udumps)) {
758 revalidate_udumps(revalidator, &udumps);
765 static enum upcall_type
766 classify_upcall(const struct upcall *upcall)
768 const struct dpif_upcall *dpif_upcall = &upcall->dpif_upcall;
769 union user_action_cookie cookie;
772 /* First look at the upcall type. */
773 switch (dpif_upcall->type) {
780 case DPIF_N_UC_TYPES:
782 VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32,
787 /* "action" upcalls need a closer look. */
788 if (!dpif_upcall->userdata) {
789 VLOG_WARN_RL(&rl, "action upcall missing cookie");
792 userdata_len = nl_attr_get_size(dpif_upcall->userdata);
793 if (userdata_len < sizeof cookie.type
794 || userdata_len > sizeof cookie) {
795 VLOG_WARN_RL(&rl, "action upcall cookie has unexpected size %"PRIuSIZE,
799 memset(&cookie, 0, sizeof cookie);
800 memcpy(&cookie, nl_attr_get(dpif_upcall->userdata), userdata_len);
801 if (userdata_len == MAX(8, sizeof cookie.sflow)
802 && cookie.type == USER_ACTION_COOKIE_SFLOW) {
804 } else if (userdata_len == MAX(8, sizeof cookie.slow_path)
805 && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) {
807 } else if (userdata_len == MAX(8, sizeof cookie.flow_sample)
808 && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) {
809 return FLOW_SAMPLE_UPCALL;
810 } else if (userdata_len == MAX(8, sizeof cookie.ipfix)
811 && cookie.type == USER_ACTION_COOKIE_IPFIX) {
814 VLOG_WARN_RL(&rl, "invalid user cookie of type %"PRIu16
815 " and size %"PRIuSIZE, cookie.type, userdata_len);
821 recv_upcalls(struct udpif *udpif)
826 uint32_t hash = udpif->secret;
827 struct handler *handler;
828 struct upcall *upcall;
829 size_t n_bytes, left;
833 upcall = xmalloc(sizeof *upcall);
834 ofpbuf_use_stub(&upcall->upcall_buf, upcall->upcall_stub,
835 sizeof upcall->upcall_stub);
836 error = dpif_recv(udpif->dpif, 0, &upcall->dpif_upcall,
837 &upcall->upcall_buf);
839 /* upcall_destroy() can only be called on successfully received
841 ofpbuf_uninit(&upcall->upcall_buf);
847 NL_ATTR_FOR_EACH (nla, left, upcall->dpif_upcall.key,
848 upcall->dpif_upcall.key_len) {
849 enum ovs_key_attr type = nl_attr_type(nla);
850 if (type == OVS_KEY_ATTR_IN_PORT
851 || type == OVS_KEY_ATTR_TCP
852 || type == OVS_KEY_ATTR_UDP) {
853 if (nl_attr_get_size(nla) == 4) {
854 hash = mhash_add(hash, nl_attr_get_u32(nla));
858 "Netlink attribute with incorrect size.");
862 hash = mhash_finish(hash, n_bytes);
864 handler = &udpif->handlers[hash % udpif->n_handlers];
866 ovs_mutex_lock(&handler->mutex);
867 if (handler->n_upcalls < MAX_QUEUE_LENGTH) {
868 list_push_back(&handler->upcalls, &upcall->list_node);
869 if (handler->n_upcalls == 0) {
870 handler->need_signal = true;
872 handler->n_upcalls++;
873 if (handler->need_signal &&
874 handler->n_upcalls >= FLOW_MISS_MAX_BATCH) {
875 handler->need_signal = false;
876 xpthread_cond_signal(&handler->wake_cond);
878 ovs_mutex_unlock(&handler->mutex);
879 if (!VLOG_DROP_DBG(&rl)) {
880 struct ds ds = DS_EMPTY_INITIALIZER;
882 odp_flow_key_format(upcall->dpif_upcall.key,
883 upcall->dpif_upcall.key_len,
885 VLOG_DBG("dispatcher: enqueue (%s)", ds_cstr(&ds));
889 ovs_mutex_unlock(&handler->mutex);
890 COVERAGE_INC(upcall_queue_overflow);
891 upcall_destroy(upcall);
895 for (n = 0; n < udpif->n_handlers; ++n) {
896 struct handler *handler = &udpif->handlers[n];
898 if (handler->need_signal) {
899 handler->need_signal = false;
900 ovs_mutex_lock(&handler->mutex);
901 xpthread_cond_signal(&handler->wake_cond);
902 ovs_mutex_unlock(&handler->mutex);
907 /* Calculates slow path actions for 'xout'. 'buf' must statically be
908 * initialized with at least 128 bytes of space. */
910 compose_slow_path(struct udpif *udpif, struct xlate_out *xout,
911 odp_port_t odp_in_port, struct ofpbuf *buf)
913 union user_action_cookie cookie;
917 cookie.type = USER_ACTION_COOKIE_SLOW_PATH;
918 cookie.slow_path.unused = 0;
919 cookie.slow_path.reason = xout->slow;
921 port = xout->slow & (SLOW_CFM | SLOW_BFD | SLOW_LACP | SLOW_STP)
924 pid = dpif_port_get_pid(udpif->dpif, port, 0);
925 odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, buf);
928 static struct flow_miss *
929 flow_miss_find(struct hmap *todo, const struct ofproto_dpif *ofproto,
930 const struct flow *flow, uint32_t hash)
932 struct flow_miss *miss;
934 HMAP_FOR_EACH_WITH_HASH (miss, hmap_node, hash, todo) {
935 if (miss->ofproto == ofproto && flow_equal(&miss->flow, flow)) {
944 handle_upcalls(struct handler *handler, struct list *upcalls)
946 struct hmap misses = HMAP_INITIALIZER(&misses);
947 struct udpif *udpif = handler->udpif;
949 struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH];
950 struct dpif_op *opsp[FLOW_MISS_MAX_BATCH * 2];
951 struct dpif_op ops[FLOW_MISS_MAX_BATCH * 2];
952 struct flow_miss *miss, *next_miss;
953 struct upcall *upcall, *next;
954 size_t n_misses, n_ops, i;
955 unsigned int flow_limit;
956 bool fail_open, may_put;
957 enum upcall_type type;
959 atomic_read(&udpif->flow_limit, &flow_limit);
960 may_put = udpif_get_n_flows(udpif) < flow_limit;
962 /* Extract the flow from each upcall. Construct in 'misses' a hash table
963 * that maps each unique flow to a 'struct flow_miss'.
965 * Most commonly there is a single packet per flow_miss, but there are
966 * several reasons why there might be more than one, e.g.:
968 * - The dpif packet interface does not support TSO (or UFO, etc.), so a
969 * large packet sent to userspace is split into a sequence of smaller
972 * - A stream of quickly arriving packets in an established "slow-pathed"
975 * - Rarely, a stream of quickly arriving packets in a flow not yet
976 * established. (This is rare because most protocols do not send
977 * multiple back-to-back packets before receiving a reply from the
978 * other end of the connection, which gives OVS a chance to set up a
982 LIST_FOR_EACH_SAFE (upcall, next, list_node, upcalls) {
983 struct dpif_upcall *dupcall = &upcall->dpif_upcall;
984 struct flow_miss *miss = &miss_buf[n_misses];
985 struct ofpbuf *packet = &dupcall->packet;
986 struct flow_miss *existing_miss;
987 struct ofproto_dpif *ofproto;
988 struct dpif_sflow *sflow;
989 struct dpif_ipfix *ipfix;
990 odp_port_t odp_in_port;
994 error = xlate_receive(udpif->backer, packet, dupcall->key,
995 dupcall->key_len, &flow,
996 &ofproto, &ipfix, &sflow, NULL, &odp_in_port);
998 if (error == ENODEV) {
999 /* Received packet on datapath port for which we couldn't
1000 * associate an ofproto. This can happen if a port is removed
1001 * while traffic is being received. Print a rate-limited
1002 * message in case it happens frequently. Install a drop flow
1003 * so that future packets of the flow are inexpensively dropped
1005 VLOG_INFO_RL(&rl, "received packet on unassociated datapath "
1006 "port %"PRIu32, odp_in_port);
1007 dpif_flow_put(udpif->dpif, DPIF_FP_CREATE | DPIF_FP_MODIFY,
1008 dupcall->key, dupcall->key_len, NULL, 0, NULL, 0,
1011 list_remove(&upcall->list_node);
1012 upcall_destroy(upcall);
1016 type = classify_upcall(upcall);
1017 if (type == MISS_UPCALL) {
1019 struct pkt_metadata md = pkt_metadata_from_flow(&flow);
1021 flow_extract(packet, &md, &miss->flow);
1022 hash = flow_hash(&miss->flow, 0);
1023 existing_miss = flow_miss_find(&misses, ofproto, &miss->flow,
1025 if (!existing_miss) {
1026 hmap_insert(&misses, &miss->hmap_node, hash);
1027 miss->ofproto = ofproto;
1028 miss->key = dupcall->key;
1029 miss->key_len = dupcall->key_len;
1030 miss->upcall_type = dupcall->type;
1031 miss->stats.n_packets = 0;
1032 miss->stats.n_bytes = 0;
1033 miss->stats.used = time_msec();
1034 miss->stats.tcp_flags = 0;
1035 miss->odp_in_port = odp_in_port;
1040 miss = existing_miss;
1042 miss->stats.tcp_flags |= ntohs(miss->flow.tcp_flags);
1043 miss->stats.n_bytes += packet->size;
1044 miss->stats.n_packets++;
1046 upcall->flow_miss = miss;
1053 union user_action_cookie cookie;
1055 memset(&cookie, 0, sizeof cookie);
1056 memcpy(&cookie, nl_attr_get(dupcall->userdata),
1057 sizeof cookie.sflow);
1058 dpif_sflow_received(sflow, packet, &flow, odp_in_port,
1064 dpif_ipfix_bridge_sample(ipfix, packet, &flow);
1067 case FLOW_SAMPLE_UPCALL:
1069 union user_action_cookie cookie;
1071 memset(&cookie, 0, sizeof cookie);
1072 memcpy(&cookie, nl_attr_get(dupcall->userdata),
1073 sizeof cookie.flow_sample);
1075 /* The flow reflects exactly the contents of the packet.
1076 * Sample the packet using it. */
1077 dpif_ipfix_flow_sample(ipfix, packet, &flow,
1078 cookie.flow_sample.collector_set_id,
1079 cookie.flow_sample.probability,
1080 cookie.flow_sample.obs_domain_id,
1081 cookie.flow_sample.obs_point_id);
1090 dpif_ipfix_unref(ipfix);
1091 dpif_sflow_unref(sflow);
1093 list_remove(&upcall->list_node);
1094 upcall_destroy(upcall);
1097 /* Initialize each 'struct flow_miss's ->xout.
1099 * We do this per-flow_miss rather than per-packet because, most commonly,
1100 * all the packets in a flow can use the same translation.
1102 * We can't do this in the previous loop because we need the TCP flags for
1103 * all the packets in each miss. */
1105 HMAP_FOR_EACH (miss, hmap_node, &misses) {
1106 struct xlate_in xin;
1108 xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL,
1109 miss->stats.tcp_flags, NULL);
1110 xin.may_learn = true;
1112 if (miss->upcall_type == DPIF_UC_MISS) {
1113 xin.resubmit_stats = &miss->stats;
1115 /* For non-miss upcalls, there's a flow in the datapath which this
1116 * packet was accounted to. Presumably the revalidators will deal
1117 * with pushing its stats eventually. */
1120 xlate_actions(&xin, &miss->xout);
1121 fail_open = fail_open || miss->xout.fail_open;
1124 /* Now handle the packets individually in order of arrival. In the common
1125 * case each packet of a miss can share the same actions, but slow-pathed
1126 * packets need to be translated individually:
1128 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1129 * processes received packets for these protocols.
1131 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1134 * The loop fills 'ops' with an array of operations to execute in the
1137 LIST_FOR_EACH (upcall, list_node, upcalls) {
1138 struct flow_miss *miss = upcall->flow_miss;
1139 struct ofpbuf *packet = &upcall->dpif_upcall.packet;
1141 ovs_be16 flow_vlan_tci;
1143 /* Save a copy of flow.vlan_tci in case it is changed to
1144 * generate proper mega flow masks for VLAN splinter flows. */
1145 flow_vlan_tci = miss->flow.vlan_tci;
1147 if (miss->xout.slow) {
1148 struct xlate_in xin;
1150 xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL, 0, packet);
1151 xlate_actions_for_side_effects(&xin);
1154 if (miss->flow.in_port.ofp_port
1155 != vsp_realdev_to_vlandev(miss->ofproto,
1156 miss->flow.in_port.ofp_port,
1157 miss->flow.vlan_tci)) {
1158 /* This packet was received on a VLAN splinter port. We
1159 * added a VLAN to the packet to make the packet resemble
1160 * the flow, but the actions were composed assuming that
1161 * the packet contained no VLAN. So, we must remove the
1162 * VLAN header from the packet before trying to execute the
1164 if (miss->xout.odp_actions.size) {
1165 eth_pop_vlan(packet);
1168 /* Remove the flow vlan tags inserted by vlan splinter logic
1169 * to ensure megaflow masks generated match the data path flow. */
1170 miss->flow.vlan_tci = 0;
1173 /* Do not install a flow into the datapath if:
1175 * - The datapath already has too many flows.
1177 * - An earlier iteration of this loop already put the same flow.
1179 * - We received this packet via some flow installed in the kernel
1183 && upcall->dpif_upcall.type == DPIF_UC_MISS) {
1189 atomic_read(&enable_megaflows, &megaflow);
1190 ofpbuf_use_stack(&mask, &miss->mask_buf, sizeof miss->mask_buf);
1194 max_mpls = ofproto_dpif_get_max_mpls_depth(miss->ofproto);
1195 odp_flow_key_from_mask(&mask, &miss->xout.wc.masks,
1196 &miss->flow, UINT32_MAX, max_mpls);
1200 op->type = DPIF_OP_FLOW_PUT;
1201 op->u.flow_put.flags = DPIF_FP_CREATE | DPIF_FP_MODIFY;
1202 op->u.flow_put.key = miss->key;
1203 op->u.flow_put.key_len = miss->key_len;
1204 op->u.flow_put.mask = mask.data;
1205 op->u.flow_put.mask_len = mask.size;
1206 op->u.flow_put.stats = NULL;
1208 if (!miss->xout.slow) {
1209 op->u.flow_put.actions = miss->xout.odp_actions.data;
1210 op->u.flow_put.actions_len = miss->xout.odp_actions.size;
1214 ofpbuf_use_stack(&buf, miss->slow_path_buf,
1215 sizeof miss->slow_path_buf);
1216 compose_slow_path(udpif, &miss->xout, miss->odp_in_port, &buf);
1217 op->u.flow_put.actions = buf.data;
1218 op->u.flow_put.actions_len = buf.size;
1223 * The 'miss' may be shared by multiple upcalls. Restore
1224 * the saved flow vlan_tci field before processing the next
1226 miss->flow.vlan_tci = flow_vlan_tci;
1228 if (miss->xout.odp_actions.size) {
1231 op->type = DPIF_OP_EXECUTE;
1232 op->u.execute.packet = packet;
1233 odp_key_to_pkt_metadata(miss->key, miss->key_len,
1235 op->u.execute.actions = miss->xout.odp_actions.data;
1236 op->u.execute.actions_len = miss->xout.odp_actions.size;
1237 op->u.execute.needs_help = (miss->xout.slow & SLOW_ACTION) != 0;
1241 /* Special case for fail-open mode.
1243 * If we are in fail-open mode, but we are connected to a controller too,
1244 * then we should send the packet up to the controller in the hope that it
1245 * will try to set up a flow and thereby allow us to exit fail-open.
1247 * See the top-level comment in fail-open.c for more information.
1249 * Copy packets before they are modified by execution. */
1251 LIST_FOR_EACH (upcall, list_node, upcalls) {
1252 struct flow_miss *miss = upcall->flow_miss;
1253 struct ofpbuf *packet = &upcall->dpif_upcall.packet;
1254 struct ofproto_packet_in *pin;
1256 pin = xmalloc(sizeof *pin);
1257 pin->up.packet = xmemdup(packet->data, packet->size);
1258 pin->up.packet_len = packet->size;
1259 pin->up.reason = OFPR_NO_MATCH;
1260 pin->up.table_id = 0;
1261 pin->up.cookie = OVS_BE64_MAX;
1262 flow_get_metadata(&miss->flow, &pin->up.fmd);
1263 pin->send_len = 0; /* Not used for flow table misses. */
1264 pin->miss_type = OFPROTO_PACKET_IN_NO_MISS;
1265 ofproto_dpif_send_packet_in(miss->ofproto, pin);
1269 /* Execute batch. */
1270 for (i = 0; i < n_ops; i++) {
1273 dpif_operate(udpif->dpif, opsp, n_ops);
1275 HMAP_FOR_EACH_SAFE (miss, next_miss, hmap_node, &misses) {
1276 hmap_remove(&misses, &miss->hmap_node);
1277 xlate_out_uninit(&miss->xout);
1279 hmap_destroy(&misses);
1281 LIST_FOR_EACH_SAFE (upcall, next, list_node, upcalls) {
1282 list_remove(&upcall->list_node);
1283 upcall_destroy(upcall);
1287 static struct udpif_key *
1288 ukey_lookup(struct revalidator *revalidator, struct udpif_flow_dump *udump)
1290 struct udpif_key *ukey;
1292 HMAP_FOR_EACH_WITH_HASH (ukey, hmap_node, udump->key_hash,
1293 &revalidator->ukeys) {
1294 if (ukey->key_len == udump->key_len
1295 && !memcmp(ukey->key, udump->key, udump->key_len)) {
1302 static struct udpif_key *
1303 ukey_create(const struct nlattr *key, size_t key_len, long long int used)
1305 struct udpif_key *ukey = xmalloc(sizeof *ukey);
1307 ukey->key = (struct nlattr *) &ukey->key_buf;
1308 memcpy(&ukey->key_buf, key, key_len);
1309 ukey->key_len = key_len;
1312 ukey->created = used ? used : time_msec();
1313 memset(&ukey->stats, 0, sizeof ukey->stats);
1319 ukey_delete(struct revalidator *revalidator, struct udpif_key *ukey)
1321 hmap_remove(&revalidator->ukeys, &ukey->hmap_node);
1326 revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump,
1327 struct udpif_key *ukey)
1329 struct ofpbuf xout_actions, *actions;
1330 uint64_t slow_path_buf[128 / 8];
1331 struct xlate_out xout, *xoutp;
1332 struct flow flow, udump_mask;
1333 struct ofproto_dpif *ofproto;
1334 struct dpif_flow_stats push;
1335 uint32_t *udump32, *xout32;
1336 odp_port_t odp_in_port;
1337 struct xlate_in xin;
1346 /* If we don't need to revalidate, we can simply push the stats contained
1347 * in the udump, otherwise we'll have to get the actions so we can check
1349 if (udump->need_revalidate) {
1350 if (dpif_flow_get(udpif->dpif, ukey->key, ukey->key_len, &actions,
1356 push.used = udump->stats.used;
1357 push.tcp_flags = udump->stats.tcp_flags;
1358 push.n_packets = udump->stats.n_packets > ukey->stats.n_packets
1359 ? udump->stats.n_packets - ukey->stats.n_packets
1361 push.n_bytes = udump->stats.n_bytes > ukey->stats.n_bytes
1362 ? udump->stats.n_bytes - ukey->stats.n_bytes
1364 ukey->stats = udump->stats;
1366 if (!push.n_packets && !udump->need_revalidate) {
1371 error = xlate_receive(udpif->backer, NULL, ukey->key, ukey->key_len, &flow,
1372 &ofproto, NULL, NULL, NULL, &odp_in_port);
1377 xlate_in_init(&xin, ofproto, &flow, NULL, push.tcp_flags, NULL);
1378 xin.resubmit_stats = push.n_packets ? &push : NULL;
1379 xin.may_learn = push.n_packets > 0;
1380 xin.skip_wildcards = !udump->need_revalidate;
1381 xlate_actions(&xin, &xout);
1384 if (!udump->need_revalidate) {
1390 ofpbuf_use_const(&xout_actions, xout.odp_actions.data,
1391 xout.odp_actions.size);
1393 ofpbuf_use_stack(&xout_actions, slow_path_buf, sizeof slow_path_buf);
1394 compose_slow_path(udpif, &xout, odp_in_port, &xout_actions);
1397 if (!ofpbuf_equal(&xout_actions, actions)) {
1401 if (odp_flow_key_to_mask(udump->mask, udump->mask_len, &udump_mask, &flow)
1406 /* Since the kernel is free to ignore wildcarded bits in the mask, we can't
1407 * directly check that the masks are the same. Instead we check that the
1408 * mask in the kernel is more specific i.e. less wildcarded, than what
1409 * we've calculated here. This guarantees we don't catch any packets we
1410 * shouldn't with the megaflow. */
1411 udump32 = (uint32_t *) &udump_mask;
1412 xout32 = (uint32_t *) &xout.wc.masks;
1413 for (i = 0; i < FLOW_U32S; i++) {
1414 if ((udump32[i] | xout32[i]) != udump32[i]) {
1421 ofpbuf_delete(actions);
1422 xlate_out_uninit(xoutp);
1427 struct udpif_key *ukey;
1428 struct udpif_flow_dump *udump;
1429 struct dpif_flow_stats stats; /* Stats for 'op'. */
1430 struct dpif_op op; /* Flow del operation. */
1434 dump_op_init(struct dump_op *op, const struct nlattr *key, size_t key_len,
1435 struct udpif_key *ukey, struct udpif_flow_dump *udump)
1439 op->op.type = DPIF_OP_FLOW_DEL;
1440 op->op.u.flow_del.key = key;
1441 op->op.u.flow_del.key_len = key_len;
1442 op->op.u.flow_del.stats = &op->stats;
1446 push_dump_ops(struct revalidator *revalidator,
1447 struct dump_op *ops, size_t n_ops)
1449 struct udpif *udpif = revalidator->udpif;
1450 struct dpif_op *opsp[REVALIDATE_MAX_BATCH];
1453 ovs_assert(n_ops <= REVALIDATE_MAX_BATCH);
1454 for (i = 0; i < n_ops; i++) {
1455 opsp[i] = &ops[i].op;
1457 dpif_operate(udpif->dpif, opsp, n_ops);
1459 for (i = 0; i < n_ops; i++) {
1460 struct dump_op *op = &ops[i];
1461 struct dpif_flow_stats *push, *stats, push_buf;
1463 stats = op->op.u.flow_del.stats;
1466 push->used = MAX(stats->used, op->ukey->stats.used);
1467 push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags;
1468 push->n_packets = stats->n_packets - op->ukey->stats.n_packets;
1469 push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes;
1474 if (push->n_packets || netflow_exists()) {
1475 struct ofproto_dpif *ofproto;
1476 struct netflow *netflow;
1479 if (!xlate_receive(udpif->backer, NULL, op->op.u.flow_del.key,
1480 op->op.u.flow_del.key_len, &flow, &ofproto,
1481 NULL, NULL, &netflow, NULL)) {
1482 struct xlate_in xin;
1484 xlate_in_init(&xin, ofproto, &flow, NULL, push->tcp_flags,
1486 xin.resubmit_stats = push->n_packets ? push : NULL;
1487 xin.may_learn = push->n_packets > 0;
1488 xin.skip_wildcards = true;
1489 xlate_actions_for_side_effects(&xin);
1492 netflow_expire(netflow, &flow);
1493 netflow_flow_clear(netflow, &flow);
1494 netflow_unref(netflow);
1500 for (i = 0; i < n_ops; i++) {
1501 struct udpif_key *ukey;
1503 /* If there's a udump, this ukey came directly from a datapath flow
1504 * dump. Sometimes a datapath can send duplicates in flow dumps, in
1505 * which case we wouldn't want to double-free a ukey, so avoid that by
1506 * looking up the ukey again.
1508 * If there's no udump then we know what we're doing. */
1509 ukey = (ops[i].udump
1510 ? ukey_lookup(revalidator, ops[i].udump)
1513 ukey_delete(revalidator, ukey);
1519 revalidate_udumps(struct revalidator *revalidator, struct list *udumps)
1521 struct udpif *udpif = revalidator->udpif;
1523 struct dump_op ops[REVALIDATE_MAX_BATCH];
1524 struct udpif_flow_dump *udump, *next_udump;
1525 size_t n_ops, n_flows;
1526 unsigned int flow_limit;
1527 long long int max_idle;
1530 atomic_read(&udpif->flow_limit, &flow_limit);
1532 n_flows = udpif_get_n_flows(udpif);
1535 max_idle = ofproto_max_idle;
1536 if (n_flows > flow_limit) {
1537 must_del = n_flows > 2 * flow_limit;
1542 LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) {
1543 long long int used, now;
1544 struct udpif_key *ukey;
1547 ukey = ukey_lookup(revalidator, udump);
1549 used = udump->stats.used;
1550 if (!used && ukey) {
1551 used = ukey->created;
1554 if (must_del || (used && used < now - max_idle)) {
1555 struct dump_op *dop = &ops[n_ops++];
1557 dump_op_init(dop, udump->key, udump->key_len, ukey, udump);
1562 ukey = ukey_create(udump->key, udump->key_len, used);
1563 hmap_insert(&revalidator->ukeys, &ukey->hmap_node,
1568 if (!revalidate_ukey(udpif, udump, ukey)) {
1569 dpif_flow_del(udpif->dpif, udump->key, udump->key_len, NULL);
1570 ukey_delete(revalidator, ukey);
1573 list_remove(&udump->list_node);
1577 push_dump_ops(revalidator, ops, n_ops);
1579 LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) {
1580 list_remove(&udump->list_node);
1586 revalidator_sweep__(struct revalidator *revalidator, bool purge)
1588 struct dump_op ops[REVALIDATE_MAX_BATCH];
1589 struct udpif_key *ukey, *next;
1594 HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, &revalidator->ukeys) {
1595 if (!purge && ukey->mark) {
1598 struct dump_op *op = &ops[n_ops++];
1600 /* If we have previously seen a flow in the datapath, but didn't
1601 * see it during the most recent dump, delete it. This allows us
1602 * to clean up the ukey and keep the statistics consistent. */
1603 dump_op_init(op, ukey->key, ukey->key_len, ukey, NULL);
1604 if (n_ops == REVALIDATE_MAX_BATCH) {
1605 push_dump_ops(revalidator, ops, n_ops);
1612 push_dump_ops(revalidator, ops, n_ops);
1617 revalidator_sweep(struct revalidator *revalidator)
1619 revalidator_sweep__(revalidator, false);
1623 revalidator_purge(struct revalidator *revalidator)
1625 revalidator_sweep__(revalidator, true);
1629 upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED,
1630 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
1632 struct ds ds = DS_EMPTY_INITIALIZER;
1633 struct udpif *udpif;
1635 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1636 unsigned int flow_limit;
1639 atomic_read(&udpif->flow_limit, &flow_limit);
1641 ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif));
1642 ds_put_format(&ds, "\tflows : (current %"PRIu64")"
1643 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif),
1644 udpif->avg_n_flows, udpif->max_n_flows, flow_limit);
1645 ds_put_format(&ds, "\tdump duration : %lldms\n", udpif->dump_duration);
1647 ds_put_char(&ds, '\n');
1648 for (i = 0; i < udpif->n_handlers; i++) {
1649 struct handler *handler = &udpif->handlers[i];
1651 ovs_mutex_lock(&handler->mutex);
1652 ds_put_format(&ds, "\t%s: (upcall queue %"PRIuSIZE")\n",
1653 handler->name, handler->n_upcalls);
1654 ovs_mutex_unlock(&handler->mutex);
1657 ds_put_char(&ds, '\n');
1658 for (i = 0; i < n_revalidators; i++) {
1659 struct revalidator *revalidator = &udpif->revalidators[i];
1661 /* XXX: The result of hmap_count(&revalidator->ukeys) may not be
1662 * accurate because it's not protected by the revalidator mutex. */
1663 ovs_mutex_lock(&revalidator->mutex);
1664 ds_put_format(&ds, "\t%s: (dump queue %"PRIuSIZE") (keys %"PRIuSIZE
1665 ")\n", revalidator->name, revalidator->n_udumps,
1666 hmap_count(&revalidator->ukeys));
1667 ovs_mutex_unlock(&revalidator->mutex);
1671 unixctl_command_reply(conn, ds_cstr(&ds));
1675 /* Disable using the megaflows.
1677 * This command is only needed for advanced debugging, so it's not
1678 * documented in the man page. */
1680 upcall_unixctl_disable_megaflows(struct unixctl_conn *conn,
1681 int argc OVS_UNUSED,
1682 const char *argv[] OVS_UNUSED,
1683 void *aux OVS_UNUSED)
1685 atomic_store(&enable_megaflows, false);
1686 udpif_flush_all_datapaths();
1687 unixctl_command_reply(conn, "megaflows disabled");
1690 /* Re-enable using megaflows.
1692 * This command is only needed for advanced debugging, so it's not
1693 * documented in the man page. */
1695 upcall_unixctl_enable_megaflows(struct unixctl_conn *conn,
1696 int argc OVS_UNUSED,
1697 const char *argv[] OVS_UNUSED,
1698 void *aux OVS_UNUSED)
1700 atomic_store(&enable_megaflows, true);
1701 udpif_flush_all_datapaths();
1702 unixctl_command_reply(conn, "megaflows enabled");
1705 /* Set the flow limit.
1707 * This command is only needed for advanced debugging, so it's not
1708 * documented in the man page. */
1710 upcall_unixctl_set_flow_limit(struct unixctl_conn *conn,
1711 int argc OVS_UNUSED,
1712 const char *argv[] OVS_UNUSED,
1713 void *aux OVS_UNUSED)
1715 struct ds ds = DS_EMPTY_INITIALIZER;
1716 struct udpif *udpif;
1717 unsigned int flow_limit = atoi(argv[1]);
1719 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1720 atomic_store(&udpif->flow_limit, flow_limit);
1722 ds_put_format(&ds, "set flow_limit to %u\n", flow_limit);
1723 unixctl_command_reply(conn, ds_cstr(&ds));