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"
36 #include "poll-loop.h"
41 #define MAX_QUEUE_LENGTH 512
42 #define FLOW_MISS_MAX_BATCH 50
43 #define REVALIDATE_MAX_BATCH 50
45 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall);
47 COVERAGE_DEFINE(upcall_queue_overflow);
49 /* A thread that processes each upcall handed to it by the dispatcher thread,
50 * forwards the upcall's packet, and possibly sets up a kernel flow as a
53 struct udpif *udpif; /* Parent udpif. */
54 pthread_t thread; /* Thread ID. */
55 char *name; /* Thread name. */
57 struct ovs_mutex mutex; /* Mutex guarding the following. */
59 /* Atomic queue of unprocessed upcalls. */
60 struct list upcalls OVS_GUARDED;
61 size_t n_upcalls OVS_GUARDED;
63 bool need_signal; /* Only changed by the dispatcher. */
65 pthread_cond_t wake_cond; /* Wakes 'thread' while holding
69 /* A thread that processes each kernel flow handed to it by the flow_dumper
70 * thread, updates OpenFlow statistics, and updates or removes the kernel flow
73 struct udpif *udpif; /* Parent udpif. */
74 char *name; /* Thread name. */
76 pthread_t thread; /* Thread ID. */
77 struct hmap ukeys; /* Datapath flow keys. */
81 struct ovs_mutex mutex; /* Mutex guarding the following. */
82 pthread_cond_t wake_cond;
83 struct list udumps OVS_GUARDED; /* Unprocessed udumps. */
84 size_t n_udumps OVS_GUARDED; /* Number of unprocessed udumps. */
87 /* An upcall handler for ofproto_dpif.
89 * udpif has two logically separate pieces:
91 * - A "dispatcher" thread that reads upcalls from the kernel and dispatches
92 * them to one of several "handler" threads (see struct handler).
94 * - A "flow_dumper" thread that reads the kernel flow table and dispatches
95 * flows to one of several "revalidator" threads (see struct
98 struct list list_node; /* In all_udpifs list. */
100 struct dpif *dpif; /* Datapath handle. */
101 struct dpif_backer *backer; /* Opaque dpif_backer pointer. */
103 uint32_t secret; /* Random seed for upcall hash. */
105 pthread_t dispatcher; /* Dispatcher thread ID. */
106 pthread_t flow_dumper; /* Flow dumper thread ID. */
108 struct handler *handlers; /* Upcall handlers. */
111 struct revalidator *revalidators; /* Flow revalidators. */
112 size_t n_revalidators;
114 uint64_t last_reval_seq; /* 'reval_seq' at last revalidation. */
115 struct seq *reval_seq; /* Incremented to force revalidation. */
117 struct seq *dump_seq; /* Increments each dump iteration. */
119 struct latch exit_latch; /* Tells child threads to exit. */
121 long long int dump_duration; /* Duration of the last flow dump. */
123 /* Datapath flow statistics. */
124 unsigned int max_n_flows;
125 unsigned int avg_n_flows;
127 /* Following fields are accessed and modified by different threads. */
128 atomic_uint flow_limit; /* Datapath flow hard limit. */
130 /* n_flows_mutex prevents multiple threads updating these concurrently. */
131 atomic_uint64_t n_flows; /* Number of flows in the datapath. */
132 atomic_llong n_flows_timestamp; /* Last time n_flows was updated. */
133 struct ovs_mutex n_flows_mutex;
137 BAD_UPCALL, /* Some kind of bug somewhere. */
138 MISS_UPCALL, /* A flow miss. */
139 SFLOW_UPCALL, /* sFlow sample. */
140 FLOW_SAMPLE_UPCALL, /* Per-flow sampling. */
141 IPFIX_UPCALL /* Per-bridge sampling. */
145 struct list list_node; /* For queuing upcalls. */
146 struct flow_miss *flow_miss; /* This upcall's flow_miss. */
148 /* Raw upcall plus data for keeping track of the memory backing it. */
149 struct dpif_upcall dpif_upcall; /* As returned by dpif_recv() */
150 struct ofpbuf upcall_buf; /* Owns some data in 'dpif_upcall'. */
151 uint64_t upcall_stub[512 / 8]; /* Buffer to reduce need for malloc(). */
154 /* 'udpif_key's are responsible for tracking the little bit of state udpif
155 * needs to do flow expiration which can't be pulled directly from the
156 * datapath. They are owned, created by, maintained, and destroyed by a single
157 * revalidator making them easy to efficiently handle with multiple threads. */
159 struct hmap_node hmap_node; /* In parent revalidator 'ukeys' map. */
161 struct nlattr *key; /* Datapath flow key. */
162 size_t key_len; /* Length of 'key'. */
164 struct dpif_flow_stats stats; /* Stats at most recent flow dump. */
165 long long int created; /* Estimation of creation time. */
167 bool mark; /* Used by mark and sweep GC algorithm. */
169 struct odputil_keybuf key_buf; /* Memory for 'key'. */
172 /* 'udpif_flow_dump's hold the state associated with one iteration in a flow
173 * dump operation. This is created by the flow_dumper thread and handed to the
174 * appropriate revalidator thread to be processed. */
175 struct udpif_flow_dump {
176 struct list list_node;
178 struct nlattr *key; /* Datapath flow key. */
179 size_t key_len; /* Length of 'key'. */
180 uint32_t key_hash; /* Hash of 'key'. */
182 struct odputil_keybuf mask_buf;
183 struct nlattr *mask; /* Datapath mask for 'key'. */
184 size_t mask_len; /* Length of 'mask'. */
186 struct dpif_flow_stats stats; /* Stats pulled from the datapath. */
188 bool need_revalidate; /* Key needs revalidation? */
190 struct odputil_keybuf key_buf;
193 /* Flow miss batching.
195 * Some dpifs implement operations faster when you hand them off in a batch.
196 * To allow batching, "struct flow_miss" queues the dpif-related work needed
197 * for a given flow. Each "struct flow_miss" corresponds to sending one or
198 * more packets, plus possibly installing the flow in the dpif. */
200 struct hmap_node hmap_node;
201 struct ofproto_dpif *ofproto;
204 const struct nlattr *key;
206 enum dpif_upcall_type upcall_type;
207 struct dpif_flow_stats stats;
208 odp_port_t odp_in_port;
210 uint64_t slow_path_buf[128 / 8];
211 struct odputil_keybuf mask_buf;
213 struct xlate_out xout;
218 static void upcall_destroy(struct upcall *);
220 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
221 static struct list all_udpifs = LIST_INITIALIZER(&all_udpifs);
223 static void recv_upcalls(struct udpif *);
224 static void handle_upcalls(struct handler *handler, struct list *upcalls);
225 static void *udpif_flow_dumper(void *);
226 static void *udpif_dispatcher(void *);
227 static void *udpif_upcall_handler(void *);
228 static void *udpif_revalidator(void *);
229 static uint64_t udpif_get_n_flows(struct udpif *);
230 static void revalidate_udumps(struct revalidator *, struct list *udumps);
231 static void revalidator_sweep(struct revalidator *);
232 static void revalidator_purge(struct revalidator *);
233 static void upcall_unixctl_show(struct unixctl_conn *conn, int argc,
234 const char *argv[], void *aux);
235 static void upcall_unixctl_disable_megaflows(struct unixctl_conn *, int argc,
236 const char *argv[], void *aux);
237 static void upcall_unixctl_enable_megaflows(struct unixctl_conn *, int argc,
238 const char *argv[], void *aux);
239 static void upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, int argc,
240 const char *argv[], void *aux);
241 static void ukey_delete(struct revalidator *, struct udpif_key *);
243 static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true);
246 udpif_create(struct dpif_backer *backer, struct dpif *dpif)
248 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
249 struct udpif *udpif = xzalloc(sizeof *udpif);
251 if (ovsthread_once_start(&once)) {
252 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show,
254 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
255 upcall_unixctl_disable_megaflows, NULL);
256 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
257 upcall_unixctl_enable_megaflows, NULL);
258 unixctl_command_register("upcall/set-flow-limit", "", 1, 1,
259 upcall_unixctl_set_flow_limit, NULL);
260 ovsthread_once_done(&once);
264 udpif->backer = backer;
265 atomic_init(&udpif->flow_limit, MIN(ofproto_flow_limit, 10000));
266 udpif->secret = random_uint32();
267 udpif->reval_seq = seq_create();
268 udpif->dump_seq = seq_create();
269 latch_init(&udpif->exit_latch);
270 list_push_back(&all_udpifs, &udpif->list_node);
271 atomic_init(&udpif->n_flows, 0);
272 atomic_init(&udpif->n_flows_timestamp, LLONG_MIN);
273 ovs_mutex_init(&udpif->n_flows_mutex);
279 udpif_destroy(struct udpif *udpif)
281 udpif_set_threads(udpif, 0, 0);
284 list_remove(&udpif->list_node);
285 latch_destroy(&udpif->exit_latch);
286 seq_destroy(udpif->reval_seq);
287 seq_destroy(udpif->dump_seq);
288 ovs_mutex_destroy(&udpif->n_flows_mutex);
292 /* Tells 'udpif' how many threads it should use to handle upcalls. Disables
293 * all threads if 'n_handlers' and 'n_revalidators' is zero. 'udpif''s
294 * datapath handle must have packet reception enabled before starting threads.
297 udpif_set_threads(struct udpif *udpif, size_t n_handlers,
298 size_t n_revalidators)
300 /* Stop the old threads (if any). */
301 if (udpif->handlers &&
302 (udpif->n_handlers != n_handlers
303 || udpif->n_revalidators != n_revalidators)) {
306 latch_set(&udpif->exit_latch);
308 for (i = 0; i < udpif->n_handlers; i++) {
309 struct handler *handler = &udpif->handlers[i];
311 ovs_mutex_lock(&handler->mutex);
312 xpthread_cond_signal(&handler->wake_cond);
313 ovs_mutex_unlock(&handler->mutex);
314 xpthread_join(handler->thread, NULL);
317 for (i = 0; i < udpif->n_revalidators; i++) {
318 struct revalidator *revalidator = &udpif->revalidators[i];
320 ovs_mutex_lock(&revalidator->mutex);
321 xpthread_cond_signal(&revalidator->wake_cond);
322 ovs_mutex_unlock(&revalidator->mutex);
323 xpthread_join(revalidator->thread, NULL);
326 xpthread_join(udpif->flow_dumper, NULL);
327 xpthread_join(udpif->dispatcher, NULL);
329 for (i = 0; i < udpif->n_revalidators; i++) {
330 struct revalidator *revalidator = &udpif->revalidators[i];
331 struct udpif_flow_dump *udump, *next_udump;
333 LIST_FOR_EACH_SAFE (udump, next_udump, list_node,
334 &revalidator->udumps) {
335 list_remove(&udump->list_node);
339 /* Delete ukeys, and delete all flows from the datapath to prevent
340 * double-counting stats. */
341 revalidator_purge(revalidator);
342 hmap_destroy(&revalidator->ukeys);
343 ovs_mutex_destroy(&revalidator->mutex);
345 free(revalidator->name);
348 for (i = 0; i < udpif->n_handlers; i++) {
349 struct handler *handler = &udpif->handlers[i];
350 struct upcall *miss, *next;
352 LIST_FOR_EACH_SAFE (miss, next, list_node, &handler->upcalls) {
353 list_remove(&miss->list_node);
354 upcall_destroy(miss);
356 ovs_mutex_destroy(&handler->mutex);
358 xpthread_cond_destroy(&handler->wake_cond);
361 latch_poll(&udpif->exit_latch);
363 free(udpif->revalidators);
364 udpif->revalidators = NULL;
365 udpif->n_revalidators = 0;
367 free(udpif->handlers);
368 udpif->handlers = NULL;
369 udpif->n_handlers = 0;
372 /* Start new threads (if necessary). */
373 if (!udpif->handlers && n_handlers) {
376 udpif->n_handlers = n_handlers;
377 udpif->n_revalidators = n_revalidators;
379 udpif->handlers = xzalloc(udpif->n_handlers * sizeof *udpif->handlers);
380 for (i = 0; i < udpif->n_handlers; i++) {
381 struct handler *handler = &udpif->handlers[i];
383 handler->udpif = udpif;
384 list_init(&handler->upcalls);
385 handler->need_signal = false;
386 xpthread_cond_init(&handler->wake_cond, NULL);
387 ovs_mutex_init(&handler->mutex);
388 xpthread_create(&handler->thread, NULL, udpif_upcall_handler,
392 udpif->revalidators = xzalloc(udpif->n_revalidators
393 * sizeof *udpif->revalidators);
394 for (i = 0; i < udpif->n_revalidators; i++) {
395 struct revalidator *revalidator = &udpif->revalidators[i];
397 revalidator->udpif = udpif;
398 list_init(&revalidator->udumps);
399 hmap_init(&revalidator->ukeys);
400 ovs_mutex_init(&revalidator->mutex);
401 xpthread_cond_init(&revalidator->wake_cond, NULL);
402 xpthread_create(&revalidator->thread, NULL, udpif_revalidator,
405 xpthread_create(&udpif->dispatcher, NULL, udpif_dispatcher, udpif);
406 xpthread_create(&udpif->flow_dumper, NULL, udpif_flow_dumper, udpif);
410 /* Waits for all ongoing upcall translations to complete. This ensures that
411 * there are no transient references to any removed ofprotos (or other
412 * objects). In particular, this should be called after an ofproto is removed
413 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
415 udpif_synchronize(struct udpif *udpif)
417 /* This is stronger than necessary. It would be sufficient to ensure
418 * (somehow) that each handler and revalidator thread had passed through
419 * its main loop once. */
420 size_t n_handlers = udpif->n_handlers;
421 size_t n_revalidators = udpif->n_revalidators;
422 udpif_set_threads(udpif, 0, 0);
423 udpif_set_threads(udpif, n_handlers, n_revalidators);
426 /* Notifies 'udpif' that something changed which may render previous
427 * xlate_actions() results invalid. */
429 udpif_revalidate(struct udpif *udpif)
431 seq_change(udpif->reval_seq);
434 /* Returns a seq which increments every time 'udpif' pulls stats from the
435 * datapath. Callers can use this to get a sense of when might be a good time
436 * to do periodic work which relies on relatively up to date statistics. */
438 udpif_dump_seq(struct udpif *udpif)
440 return udpif->dump_seq;
444 udpif_get_memory_usage(struct udpif *udpif, struct simap *usage)
448 simap_increase(usage, "dispatchers", 1);
449 simap_increase(usage, "flow_dumpers", 1);
451 simap_increase(usage, "handlers", udpif->n_handlers);
452 for (i = 0; i < udpif->n_handlers; i++) {
453 struct handler *handler = &udpif->handlers[i];
454 ovs_mutex_lock(&handler->mutex);
455 simap_increase(usage, "handler upcalls", handler->n_upcalls);
456 ovs_mutex_unlock(&handler->mutex);
459 simap_increase(usage, "revalidators", udpif->n_revalidators);
460 for (i = 0; i < udpif->n_revalidators; i++) {
461 struct revalidator *revalidator = &udpif->revalidators[i];
462 ovs_mutex_lock(&revalidator->mutex);
463 simap_increase(usage, "revalidator dumps", revalidator->n_udumps);
465 /* XXX: This isn't technically thread safe because the revalidator
466 * ukeys maps isn't protected by a mutex since it's per thread. */
467 simap_increase(usage, "revalidator keys",
468 hmap_count(&revalidator->ukeys));
469 ovs_mutex_unlock(&revalidator->mutex);
473 /* Removes all flows from all datapaths. */
479 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
480 dpif_flow_flush(udpif->dpif);
484 /* Destroys and deallocates 'upcall'. */
486 upcall_destroy(struct upcall *upcall)
489 ofpbuf_uninit(&upcall->dpif_upcall.packet);
490 ofpbuf_uninit(&upcall->upcall_buf);
496 udpif_get_n_flows(struct udpif *udpif)
498 long long int time, now;
502 atomic_read(&udpif->n_flows_timestamp, &time);
503 if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) {
504 struct dpif_dp_stats stats;
506 atomic_store(&udpif->n_flows_timestamp, now);
507 dpif_get_dp_stats(udpif->dpif, &stats);
508 flow_count = stats.n_flows;
509 atomic_store(&udpif->n_flows, flow_count);
510 ovs_mutex_unlock(&udpif->n_flows_mutex);
512 atomic_read(&udpif->n_flows, &flow_count);
517 /* The dispatcher thread is responsible for receiving upcalls from the kernel,
518 * assigning them to a upcall_handler thread. */
520 udpif_dispatcher(void *arg)
522 struct udpif *udpif = arg;
524 set_subprogram_name("dispatcher");
525 while (!latch_is_set(&udpif->exit_latch)) {
527 dpif_recv_wait(udpif->dpif);
528 latch_wait(&udpif->exit_latch);
536 udpif_flow_dumper(void *arg)
538 struct udpif *udpif = arg;
540 set_subprogram_name("flow_dumper");
541 while (!latch_is_set(&udpif->exit_latch)) {
542 const struct dpif_flow_stats *stats;
543 long long int start_time, duration;
544 const struct nlattr *key, *mask;
545 struct dpif_flow_dump dump;
546 size_t key_len, mask_len;
547 unsigned int flow_limit;
548 bool need_revalidate;
554 reval_seq = seq_read(udpif->reval_seq);
555 need_revalidate = udpif->last_reval_seq != reval_seq;
556 udpif->last_reval_seq = reval_seq;
558 n_flows = udpif_get_n_flows(udpif);
559 udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows);
560 udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2;
562 start_time = time_msec();
563 error = dpif_flow_dump_start(&dump, udpif->dpif);
565 VLOG_INFO("Failed to start flow dump (%s)", ovs_strerror(error));
568 dpif_flow_dump_state_init(udpif->dpif, &state);
569 while (dpif_flow_dump_next(&dump, state, &key, &key_len,
570 &mask, &mask_len, NULL, NULL, &stats)
571 && !latch_is_set(&udpif->exit_latch)) {
572 struct udpif_flow_dump *udump = xmalloc(sizeof *udump);
573 struct revalidator *revalidator;
575 udump->key_hash = hash_bytes(key, key_len, udpif->secret);
576 memcpy(&udump->key_buf, key, key_len);
577 udump->key = (struct nlattr *) &udump->key_buf;
578 udump->key_len = key_len;
580 memcpy(&udump->mask_buf, mask, mask_len);
581 udump->mask = (struct nlattr *) &udump->mask_buf;
582 udump->mask_len = mask_len;
584 udump->stats = *stats;
585 udump->need_revalidate = need_revalidate;
587 revalidator = &udpif->revalidators[udump->key_hash
588 % udpif->n_revalidators];
590 ovs_mutex_lock(&revalidator->mutex);
591 while (revalidator->n_udumps >= REVALIDATE_MAX_BATCH * 3
592 && !latch_is_set(&udpif->exit_latch)) {
593 ovs_mutex_cond_wait(&revalidator->wake_cond,
594 &revalidator->mutex);
596 list_push_back(&revalidator->udumps, &udump->list_node);
597 revalidator->n_udumps++;
598 xpthread_cond_signal(&revalidator->wake_cond);
599 ovs_mutex_unlock(&revalidator->mutex);
601 dpif_flow_dump_state_uninit(udpif->dpif, state);
602 dpif_flow_dump_done(&dump);
604 /* Let all the revalidators finish and garbage collect. */
605 seq_change(udpif->dump_seq);
606 for (i = 0; i < udpif->n_revalidators; i++) {
607 struct revalidator *revalidator = &udpif->revalidators[i];
608 ovs_mutex_lock(&revalidator->mutex);
609 xpthread_cond_signal(&revalidator->wake_cond);
610 ovs_mutex_unlock(&revalidator->mutex);
613 for (i = 0; i < udpif->n_revalidators; i++) {
614 struct revalidator *revalidator = &udpif->revalidators[i];
616 ovs_mutex_lock(&revalidator->mutex);
617 while (revalidator->dump_seq != seq_read(udpif->dump_seq)
618 && !latch_is_set(&udpif->exit_latch)) {
619 ovs_mutex_cond_wait(&revalidator->wake_cond,
620 &revalidator->mutex);
622 ovs_mutex_unlock(&revalidator->mutex);
625 duration = MAX(time_msec() - start_time, 1);
626 udpif->dump_duration = duration;
627 atomic_read(&udpif->flow_limit, &flow_limit);
628 if (duration > 2000) {
629 flow_limit /= duration / 1000;
630 } else if (duration > 1300) {
631 flow_limit = flow_limit * 3 / 4;
632 } else if (duration < 1000 && n_flows > 2000
633 && flow_limit < n_flows * 1000 / duration) {
636 flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000));
637 atomic_store(&udpif->flow_limit, flow_limit);
639 if (duration > 2000) {
640 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
645 poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500));
646 seq_wait(udpif->reval_seq, udpif->last_reval_seq);
647 latch_wait(&udpif->exit_latch);
654 /* The miss handler thread is responsible for processing miss upcalls retrieved
655 * by the dispatcher thread. Once finished it passes the processed miss
656 * upcalls to ofproto-dpif where they're installed in the datapath. */
658 udpif_upcall_handler(void *arg)
660 struct handler *handler = arg;
662 handler->name = xasprintf("handler_%u", ovsthread_id_self());
663 set_subprogram_name("%s", handler->name);
665 while (!latch_is_set(&handler->udpif->exit_latch)) {
666 struct list misses = LIST_INITIALIZER(&misses);
669 ovs_mutex_lock(&handler->mutex);
670 if (!handler->n_upcalls) {
671 ovs_mutex_cond_wait(&handler->wake_cond, &handler->mutex);
674 for (i = 0; i < FLOW_MISS_MAX_BATCH; i++) {
675 if (handler->n_upcalls) {
676 handler->n_upcalls--;
677 list_push_back(&misses, list_pop_front(&handler->upcalls));
682 ovs_mutex_unlock(&handler->mutex);
684 handle_upcalls(handler, &misses);
693 udpif_revalidator(void *arg)
695 struct revalidator *revalidator = arg;
697 revalidator->name = xasprintf("revalidator_%u", ovsthread_id_self());
698 set_subprogram_name("%s", revalidator->name);
700 struct list udumps = LIST_INITIALIZER(&udumps);
701 struct udpif *udpif = revalidator->udpif;
704 ovs_mutex_lock(&revalidator->mutex);
705 if (latch_is_set(&udpif->exit_latch)) {
706 ovs_mutex_unlock(&revalidator->mutex);
710 if (!revalidator->n_udumps) {
711 if (revalidator->dump_seq != seq_read(udpif->dump_seq)) {
712 revalidator->dump_seq = seq_read(udpif->dump_seq);
713 revalidator_sweep(revalidator);
715 ovs_mutex_cond_wait(&revalidator->wake_cond,
716 &revalidator->mutex);
720 for (i = 0; i < REVALIDATE_MAX_BATCH && revalidator->n_udumps; i++) {
721 list_push_back(&udumps, list_pop_front(&revalidator->udumps));
722 revalidator->n_udumps--;
725 /* Wake up the flow dumper. */
726 xpthread_cond_signal(&revalidator->wake_cond);
727 ovs_mutex_unlock(&revalidator->mutex);
729 if (!list_is_empty(&udumps)) {
730 revalidate_udumps(revalidator, &udumps);
737 static enum upcall_type
738 classify_upcall(const struct upcall *upcall)
740 const struct dpif_upcall *dpif_upcall = &upcall->dpif_upcall;
741 union user_action_cookie cookie;
744 /* First look at the upcall type. */
745 switch (dpif_upcall->type) {
752 case DPIF_N_UC_TYPES:
754 VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32,
759 /* "action" upcalls need a closer look. */
760 if (!dpif_upcall->userdata) {
761 VLOG_WARN_RL(&rl, "action upcall missing cookie");
764 userdata_len = nl_attr_get_size(dpif_upcall->userdata);
765 if (userdata_len < sizeof cookie.type
766 || userdata_len > sizeof cookie) {
767 VLOG_WARN_RL(&rl, "action upcall cookie has unexpected size %"PRIuSIZE,
771 memset(&cookie, 0, sizeof cookie);
772 memcpy(&cookie, nl_attr_get(dpif_upcall->userdata), userdata_len);
773 if (userdata_len == MAX(8, sizeof cookie.sflow)
774 && cookie.type == USER_ACTION_COOKIE_SFLOW) {
776 } else if (userdata_len == MAX(8, sizeof cookie.slow_path)
777 && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) {
779 } else if (userdata_len == MAX(8, sizeof cookie.flow_sample)
780 && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) {
781 return FLOW_SAMPLE_UPCALL;
782 } else if (userdata_len == MAX(8, sizeof cookie.ipfix)
783 && cookie.type == USER_ACTION_COOKIE_IPFIX) {
786 VLOG_WARN_RL(&rl, "invalid user cookie of type %"PRIu16
787 " and size %"PRIuSIZE, cookie.type, userdata_len);
793 recv_upcalls(struct udpif *udpif)
798 uint32_t hash = udpif->secret;
799 struct handler *handler;
800 struct upcall *upcall;
801 size_t n_bytes, left;
805 upcall = xmalloc(sizeof *upcall);
806 ofpbuf_use_stub(&upcall->upcall_buf, upcall->upcall_stub,
807 sizeof upcall->upcall_stub);
808 error = dpif_recv(udpif->dpif, &upcall->dpif_upcall,
809 &upcall->upcall_buf);
811 /* upcall_destroy() can only be called on successfully received
813 ofpbuf_uninit(&upcall->upcall_buf);
819 NL_ATTR_FOR_EACH (nla, left, upcall->dpif_upcall.key,
820 upcall->dpif_upcall.key_len) {
821 enum ovs_key_attr type = nl_attr_type(nla);
822 if (type == OVS_KEY_ATTR_IN_PORT
823 || type == OVS_KEY_ATTR_TCP
824 || type == OVS_KEY_ATTR_UDP) {
825 if (nl_attr_get_size(nla) == 4) {
826 hash = mhash_add(hash, nl_attr_get_u32(nla));
830 "Netlink attribute with incorrect size.");
834 hash = mhash_finish(hash, n_bytes);
836 handler = &udpif->handlers[hash % udpif->n_handlers];
838 ovs_mutex_lock(&handler->mutex);
839 if (handler->n_upcalls < MAX_QUEUE_LENGTH) {
840 list_push_back(&handler->upcalls, &upcall->list_node);
841 if (handler->n_upcalls == 0) {
842 handler->need_signal = true;
844 handler->n_upcalls++;
845 if (handler->need_signal &&
846 handler->n_upcalls >= FLOW_MISS_MAX_BATCH) {
847 handler->need_signal = false;
848 xpthread_cond_signal(&handler->wake_cond);
850 ovs_mutex_unlock(&handler->mutex);
851 if (!VLOG_DROP_DBG(&rl)) {
852 struct ds ds = DS_EMPTY_INITIALIZER;
854 odp_flow_key_format(upcall->dpif_upcall.key,
855 upcall->dpif_upcall.key_len,
857 VLOG_DBG("dispatcher: enqueue (%s)", ds_cstr(&ds));
861 ovs_mutex_unlock(&handler->mutex);
862 COVERAGE_INC(upcall_queue_overflow);
863 upcall_destroy(upcall);
867 for (n = 0; n < udpif->n_handlers; ++n) {
868 struct handler *handler = &udpif->handlers[n];
870 if (handler->need_signal) {
871 handler->need_signal = false;
872 ovs_mutex_lock(&handler->mutex);
873 xpthread_cond_signal(&handler->wake_cond);
874 ovs_mutex_unlock(&handler->mutex);
879 /* Calculates slow path actions for 'xout'. 'buf' must statically be
880 * initialized with at least 128 bytes of space. */
882 compose_slow_path(struct udpif *udpif, struct xlate_out *xout,
883 odp_port_t odp_in_port, struct ofpbuf *buf)
885 union user_action_cookie cookie;
889 cookie.type = USER_ACTION_COOKIE_SLOW_PATH;
890 cookie.slow_path.unused = 0;
891 cookie.slow_path.reason = xout->slow;
893 port = xout->slow & (SLOW_CFM | SLOW_BFD | SLOW_LACP | SLOW_STP)
896 pid = dpif_port_get_pid(udpif->dpif, port);
897 odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, buf);
900 static struct flow_miss *
901 flow_miss_find(struct hmap *todo, const struct ofproto_dpif *ofproto,
902 const struct flow *flow, uint32_t hash)
904 struct flow_miss *miss;
906 HMAP_FOR_EACH_WITH_HASH (miss, hmap_node, hash, todo) {
907 if (miss->ofproto == ofproto && flow_equal(&miss->flow, flow)) {
916 handle_upcalls(struct handler *handler, struct list *upcalls)
918 struct hmap misses = HMAP_INITIALIZER(&misses);
919 struct udpif *udpif = handler->udpif;
921 struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH];
922 struct dpif_op *opsp[FLOW_MISS_MAX_BATCH * 2];
923 struct dpif_op ops[FLOW_MISS_MAX_BATCH * 2];
924 struct flow_miss *miss, *next_miss;
925 struct upcall *upcall, *next;
926 size_t n_misses, n_ops, i;
927 unsigned int flow_limit;
928 bool fail_open, may_put;
929 enum upcall_type type;
931 atomic_read(&udpif->flow_limit, &flow_limit);
932 may_put = udpif_get_n_flows(udpif) < flow_limit;
934 /* Extract the flow from each upcall. Construct in 'misses' a hash table
935 * that maps each unique flow to a 'struct flow_miss'.
937 * Most commonly there is a single packet per flow_miss, but there are
938 * several reasons why there might be more than one, e.g.:
940 * - The dpif packet interface does not support TSO (or UFO, etc.), so a
941 * large packet sent to userspace is split into a sequence of smaller
944 * - A stream of quickly arriving packets in an established "slow-pathed"
947 * - Rarely, a stream of quickly arriving packets in a flow not yet
948 * established. (This is rare because most protocols do not send
949 * multiple back-to-back packets before receiving a reply from the
950 * other end of the connection, which gives OVS a chance to set up a
954 LIST_FOR_EACH_SAFE (upcall, next, list_node, upcalls) {
955 struct dpif_upcall *dupcall = &upcall->dpif_upcall;
956 struct flow_miss *miss = &miss_buf[n_misses];
957 struct ofpbuf *packet = &dupcall->packet;
958 struct flow_miss *existing_miss;
959 struct ofproto_dpif *ofproto;
960 struct dpif_sflow *sflow;
961 struct dpif_ipfix *ipfix;
962 odp_port_t odp_in_port;
966 error = xlate_receive(udpif->backer, packet, dupcall->key,
967 dupcall->key_len, &flow,
968 &ofproto, &ipfix, &sflow, NULL, &odp_in_port);
970 if (error == ENODEV) {
971 /* Received packet on datapath port for which we couldn't
972 * associate an ofproto. This can happen if a port is removed
973 * while traffic is being received. Print a rate-limited
974 * message in case it happens frequently. Install a drop flow
975 * so that future packets of the flow are inexpensively dropped
977 VLOG_INFO_RL(&rl, "received packet on unassociated datapath "
978 "port %"PRIu32, odp_in_port);
979 dpif_flow_put(udpif->dpif, DPIF_FP_CREATE | DPIF_FP_MODIFY,
980 dupcall->key, dupcall->key_len, NULL, 0, NULL, 0,
983 list_remove(&upcall->list_node);
984 upcall_destroy(upcall);
988 type = classify_upcall(upcall);
989 if (type == MISS_UPCALL) {
991 struct pkt_metadata md;
993 pkt_metadata_from_flow(&md, &flow);
994 flow_extract(packet, &md, &miss->flow);
996 hash = flow_hash(&miss->flow, 0);
997 existing_miss = flow_miss_find(&misses, ofproto, &miss->flow,
999 if (!existing_miss) {
1000 hmap_insert(&misses, &miss->hmap_node, hash);
1001 miss->ofproto = ofproto;
1002 miss->key = dupcall->key;
1003 miss->key_len = dupcall->key_len;
1004 miss->upcall_type = dupcall->type;
1005 miss->stats.n_packets = 0;
1006 miss->stats.n_bytes = 0;
1007 miss->stats.used = time_msec();
1008 miss->stats.tcp_flags = 0;
1009 miss->odp_in_port = odp_in_port;
1014 miss = existing_miss;
1016 miss->stats.tcp_flags |= packet_get_tcp_flags(packet, &miss->flow);
1017 miss->stats.n_bytes += packet->size;
1018 miss->stats.n_packets++;
1020 upcall->flow_miss = miss;
1027 union user_action_cookie cookie;
1029 memset(&cookie, 0, sizeof cookie);
1030 memcpy(&cookie, nl_attr_get(dupcall->userdata),
1031 sizeof cookie.sflow);
1032 dpif_sflow_received(sflow, packet, &flow, odp_in_port,
1038 dpif_ipfix_bridge_sample(ipfix, packet, &flow);
1041 case FLOW_SAMPLE_UPCALL:
1043 union user_action_cookie cookie;
1045 memset(&cookie, 0, sizeof cookie);
1046 memcpy(&cookie, nl_attr_get(dupcall->userdata),
1047 sizeof cookie.flow_sample);
1049 /* The flow reflects exactly the contents of the packet.
1050 * Sample the packet using it. */
1051 dpif_ipfix_flow_sample(ipfix, packet, &flow,
1052 cookie.flow_sample.collector_set_id,
1053 cookie.flow_sample.probability,
1054 cookie.flow_sample.obs_domain_id,
1055 cookie.flow_sample.obs_point_id);
1064 dpif_ipfix_unref(ipfix);
1065 dpif_sflow_unref(sflow);
1067 list_remove(&upcall->list_node);
1068 upcall_destroy(upcall);
1071 /* Initialize each 'struct flow_miss's ->xout.
1073 * We do this per-flow_miss rather than per-packet because, most commonly,
1074 * all the packets in a flow can use the same translation.
1076 * We can't do this in the previous loop because we need the TCP flags for
1077 * all the packets in each miss. */
1079 HMAP_FOR_EACH (miss, hmap_node, &misses) {
1080 struct xlate_in xin;
1082 xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL,
1083 miss->stats.tcp_flags, NULL);
1084 xin.may_learn = true;
1086 if (miss->upcall_type == DPIF_UC_MISS) {
1087 xin.resubmit_stats = &miss->stats;
1089 /* For non-miss upcalls, there's a flow in the datapath which this
1090 * packet was accounted to. Presumably the revalidators will deal
1091 * with pushing its stats eventually. */
1094 xlate_actions(&xin, &miss->xout);
1095 fail_open = fail_open || miss->xout.fail_open;
1098 /* Now handle the packets individually in order of arrival. In the common
1099 * case each packet of a miss can share the same actions, but slow-pathed
1100 * packets need to be translated individually:
1102 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1103 * processes received packets for these protocols.
1105 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1108 * The loop fills 'ops' with an array of operations to execute in the
1111 LIST_FOR_EACH (upcall, list_node, upcalls) {
1112 struct flow_miss *miss = upcall->flow_miss;
1113 struct ofpbuf *packet = &upcall->dpif_upcall.packet;
1115 ovs_be16 flow_vlan_tci;
1117 /* Save a copy of flow.vlan_tci in case it is changed to
1118 * generate proper mega flow masks for VLAN splinter flows. */
1119 flow_vlan_tci = miss->flow.vlan_tci;
1121 if (miss->xout.slow) {
1122 struct xlate_in xin;
1124 xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL, 0, packet);
1125 xlate_actions_for_side_effects(&xin);
1128 if (miss->flow.in_port.ofp_port
1129 != vsp_realdev_to_vlandev(miss->ofproto,
1130 miss->flow.in_port.ofp_port,
1131 miss->flow.vlan_tci)) {
1132 /* This packet was received on a VLAN splinter port. We
1133 * added a VLAN to the packet to make the packet resemble
1134 * the flow, but the actions were composed assuming that
1135 * the packet contained no VLAN. So, we must remove the
1136 * VLAN header from the packet before trying to execute the
1138 if (miss->xout.odp_actions.size) {
1139 eth_pop_vlan(packet);
1142 /* Remove the flow vlan tags inserted by vlan splinter logic
1143 * to ensure megaflow masks generated match the data path flow. */
1144 miss->flow.vlan_tci = 0;
1147 /* Do not install a flow into the datapath if:
1149 * - The datapath already has too many flows.
1151 * - An earlier iteration of this loop already put the same flow.
1153 * - We received this packet via some flow installed in the kernel
1157 && upcall->dpif_upcall.type == DPIF_UC_MISS) {
1163 atomic_read(&enable_megaflows, &megaflow);
1164 ofpbuf_use_stack(&mask, &miss->mask_buf, sizeof miss->mask_buf);
1168 max_mpls = ofproto_dpif_get_max_mpls_depth(miss->ofproto);
1169 odp_flow_key_from_mask(&mask, &miss->xout.wc.masks,
1170 &miss->flow, UINT32_MAX, max_mpls);
1174 op->type = DPIF_OP_FLOW_PUT;
1175 op->u.flow_put.flags = DPIF_FP_CREATE | DPIF_FP_MODIFY;
1176 op->u.flow_put.key = miss->key;
1177 op->u.flow_put.key_len = miss->key_len;
1178 op->u.flow_put.mask = mask.data;
1179 op->u.flow_put.mask_len = mask.size;
1180 op->u.flow_put.stats = NULL;
1182 if (!miss->xout.slow) {
1183 op->u.flow_put.actions = miss->xout.odp_actions.data;
1184 op->u.flow_put.actions_len = miss->xout.odp_actions.size;
1188 ofpbuf_use_stack(&buf, miss->slow_path_buf,
1189 sizeof miss->slow_path_buf);
1190 compose_slow_path(udpif, &miss->xout, miss->odp_in_port, &buf);
1191 op->u.flow_put.actions = buf.data;
1192 op->u.flow_put.actions_len = buf.size;
1197 * The 'miss' may be shared by multiple upcalls. Restore
1198 * the saved flow vlan_tci field before processing the next
1200 miss->flow.vlan_tci = flow_vlan_tci;
1202 if (miss->xout.odp_actions.size) {
1205 op->type = DPIF_OP_EXECUTE;
1206 op->u.execute.packet = packet;
1207 odp_key_to_pkt_metadata(miss->key, miss->key_len,
1209 op->u.execute.actions = miss->xout.odp_actions.data;
1210 op->u.execute.actions_len = miss->xout.odp_actions.size;
1211 op->u.execute.needs_help = (miss->xout.slow & SLOW_ACTION) != 0;
1215 /* Special case for fail-open mode.
1217 * If we are in fail-open mode, but we are connected to a controller too,
1218 * then we should send the packet up to the controller in the hope that it
1219 * will try to set up a flow and thereby allow us to exit fail-open.
1221 * See the top-level comment in fail-open.c for more information.
1223 * Copy packets before they are modified by execution. */
1225 LIST_FOR_EACH (upcall, list_node, upcalls) {
1226 struct flow_miss *miss = upcall->flow_miss;
1227 struct ofpbuf *packet = &upcall->dpif_upcall.packet;
1228 struct ofproto_packet_in *pin;
1230 pin = xmalloc(sizeof *pin);
1231 pin->up.packet = xmemdup(packet->data, packet->size);
1232 pin->up.packet_len = packet->size;
1233 pin->up.reason = OFPR_NO_MATCH;
1234 pin->up.table_id = 0;
1235 pin->up.cookie = OVS_BE64_MAX;
1236 flow_get_metadata(&miss->flow, &pin->up.fmd);
1237 pin->send_len = 0; /* Not used for flow table misses. */
1238 pin->generated_by_table_miss = false;
1239 ofproto_dpif_send_packet_in(miss->ofproto, pin);
1243 /* Execute batch. */
1244 for (i = 0; i < n_ops; i++) {
1247 dpif_operate(udpif->dpif, opsp, n_ops);
1249 HMAP_FOR_EACH_SAFE (miss, next_miss, hmap_node, &misses) {
1250 hmap_remove(&misses, &miss->hmap_node);
1251 xlate_out_uninit(&miss->xout);
1253 hmap_destroy(&misses);
1255 LIST_FOR_EACH_SAFE (upcall, next, list_node, upcalls) {
1256 list_remove(&upcall->list_node);
1257 upcall_destroy(upcall);
1261 static struct udpif_key *
1262 ukey_lookup(struct revalidator *revalidator, struct udpif_flow_dump *udump)
1264 struct udpif_key *ukey;
1266 HMAP_FOR_EACH_WITH_HASH (ukey, hmap_node, udump->key_hash,
1267 &revalidator->ukeys) {
1268 if (ukey->key_len == udump->key_len
1269 && !memcmp(ukey->key, udump->key, udump->key_len)) {
1276 static struct udpif_key *
1277 ukey_create(const struct nlattr *key, size_t key_len, long long int used)
1279 struct udpif_key *ukey = xmalloc(sizeof *ukey);
1281 ukey->key = (struct nlattr *) &ukey->key_buf;
1282 memcpy(&ukey->key_buf, key, key_len);
1283 ukey->key_len = key_len;
1286 ukey->created = used ? used : time_msec();
1287 memset(&ukey->stats, 0, sizeof ukey->stats);
1293 ukey_delete(struct revalidator *revalidator, struct udpif_key *ukey)
1295 hmap_remove(&revalidator->ukeys, &ukey->hmap_node);
1300 revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump,
1301 struct udpif_key *ukey)
1303 struct ofpbuf xout_actions, *actions;
1304 uint64_t slow_path_buf[128 / 8];
1305 struct xlate_out xout, *xoutp;
1306 struct flow flow, udump_mask;
1307 struct ofproto_dpif *ofproto;
1308 struct dpif_flow_stats push;
1309 uint32_t *udump32, *xout32;
1310 odp_port_t odp_in_port;
1311 struct xlate_in xin;
1320 /* If we don't need to revalidate, we can simply push the stats contained
1321 * in the udump, otherwise we'll have to get the actions so we can check
1323 if (udump->need_revalidate) {
1324 if (dpif_flow_get(udpif->dpif, ukey->key, ukey->key_len, &actions,
1330 push.used = udump->stats.used;
1331 push.tcp_flags = udump->stats.tcp_flags;
1332 push.n_packets = udump->stats.n_packets > ukey->stats.n_packets
1333 ? udump->stats.n_packets - ukey->stats.n_packets
1335 push.n_bytes = udump->stats.n_bytes > ukey->stats.n_bytes
1336 ? udump->stats.n_bytes - ukey->stats.n_bytes
1338 ukey->stats = udump->stats;
1340 if (!push.n_packets && !udump->need_revalidate) {
1345 error = xlate_receive(udpif->backer, NULL, ukey->key, ukey->key_len, &flow,
1346 &ofproto, NULL, NULL, NULL, &odp_in_port);
1351 xlate_in_init(&xin, ofproto, &flow, NULL, push.tcp_flags, NULL);
1352 xin.resubmit_stats = push.n_packets ? &push : NULL;
1353 xin.may_learn = push.n_packets > 0;
1354 xin.skip_wildcards = !udump->need_revalidate;
1355 xlate_actions(&xin, &xout);
1358 if (!udump->need_revalidate) {
1364 ofpbuf_use_const(&xout_actions, xout.odp_actions.data,
1365 xout.odp_actions.size);
1367 ofpbuf_use_stack(&xout_actions, slow_path_buf, sizeof slow_path_buf);
1368 compose_slow_path(udpif, &xout, odp_in_port, &xout_actions);
1371 if (!ofpbuf_equal(&xout_actions, actions)) {
1375 if (odp_flow_key_to_mask(udump->mask, udump->mask_len, &udump_mask, &flow)
1380 /* Since the kernel is free to ignore wildcarded bits in the mask, we can't
1381 * directly check that the masks are the same. Instead we check that the
1382 * mask in the kernel is more specific i.e. less wildcarded, than what
1383 * we've calculated here. This guarantees we don't catch any packets we
1384 * shouldn't with the megaflow. */
1385 udump32 = (uint32_t *) &udump_mask;
1386 xout32 = (uint32_t *) &xout.wc.masks;
1387 for (i = 0; i < FLOW_U32S; i++) {
1388 if ((udump32[i] | xout32[i]) != udump32[i]) {
1395 ofpbuf_delete(actions);
1396 xlate_out_uninit(xoutp);
1401 struct udpif_key *ukey;
1402 struct udpif_flow_dump *udump;
1403 struct dpif_flow_stats stats; /* Stats for 'op'. */
1404 struct dpif_op op; /* Flow del operation. */
1408 dump_op_init(struct dump_op *op, const struct nlattr *key, size_t key_len,
1409 struct udpif_key *ukey, struct udpif_flow_dump *udump)
1413 op->op.type = DPIF_OP_FLOW_DEL;
1414 op->op.u.flow_del.key = key;
1415 op->op.u.flow_del.key_len = key_len;
1416 op->op.u.flow_del.stats = &op->stats;
1420 push_dump_ops(struct revalidator *revalidator,
1421 struct dump_op *ops, size_t n_ops)
1423 struct udpif *udpif = revalidator->udpif;
1424 struct dpif_op *opsp[REVALIDATE_MAX_BATCH];
1427 ovs_assert(n_ops <= REVALIDATE_MAX_BATCH);
1428 for (i = 0; i < n_ops; i++) {
1429 opsp[i] = &ops[i].op;
1431 dpif_operate(udpif->dpif, opsp, n_ops);
1433 for (i = 0; i < n_ops; i++) {
1434 struct dump_op *op = &ops[i];
1435 struct dpif_flow_stats *push, *stats, push_buf;
1437 stats = op->op.u.flow_del.stats;
1440 push->used = MAX(stats->used, op->ukey->stats.used);
1441 push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags;
1442 push->n_packets = stats->n_packets - op->ukey->stats.n_packets;
1443 push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes;
1448 if (push->n_packets || netflow_exists()) {
1449 struct ofproto_dpif *ofproto;
1450 struct netflow *netflow;
1453 if (!xlate_receive(udpif->backer, NULL, op->op.u.flow_del.key,
1454 op->op.u.flow_del.key_len, &flow, &ofproto,
1455 NULL, NULL, &netflow, NULL)) {
1456 struct xlate_in xin;
1458 xlate_in_init(&xin, ofproto, &flow, NULL, push->tcp_flags,
1460 xin.resubmit_stats = push->n_packets ? push : NULL;
1461 xin.may_learn = push->n_packets > 0;
1462 xin.skip_wildcards = true;
1463 xlate_actions_for_side_effects(&xin);
1466 netflow_expire(netflow, &flow);
1467 netflow_flow_clear(netflow, &flow);
1468 netflow_unref(netflow);
1474 for (i = 0; i < n_ops; i++) {
1475 struct udpif_key *ukey;
1477 /* If there's a udump, this ukey came directly from a datapath flow
1478 * dump. Sometimes a datapath can send duplicates in flow dumps, in
1479 * which case we wouldn't want to double-free a ukey, so avoid that by
1480 * looking up the ukey again.
1482 * If there's no udump then we know what we're doing. */
1483 ukey = (ops[i].udump
1484 ? ukey_lookup(revalidator, ops[i].udump)
1487 ukey_delete(revalidator, ukey);
1493 revalidate_udumps(struct revalidator *revalidator, struct list *udumps)
1495 struct udpif *udpif = revalidator->udpif;
1497 struct dump_op ops[REVALIDATE_MAX_BATCH];
1498 struct udpif_flow_dump *udump, *next_udump;
1499 size_t n_ops, n_flows;
1500 unsigned int flow_limit;
1501 long long int max_idle;
1504 atomic_read(&udpif->flow_limit, &flow_limit);
1506 n_flows = udpif_get_n_flows(udpif);
1509 max_idle = ofproto_max_idle;
1510 if (n_flows > flow_limit) {
1511 must_del = n_flows > 2 * flow_limit;
1516 LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) {
1517 long long int used, now;
1518 struct udpif_key *ukey;
1521 ukey = ukey_lookup(revalidator, udump);
1523 used = udump->stats.used;
1524 if (!used && ukey) {
1525 used = ukey->created;
1528 if (must_del || (used && used < now - max_idle)) {
1529 struct dump_op *dop = &ops[n_ops++];
1531 dump_op_init(dop, udump->key, udump->key_len, ukey, udump);
1536 ukey = ukey_create(udump->key, udump->key_len, used);
1537 hmap_insert(&revalidator->ukeys, &ukey->hmap_node,
1542 if (!revalidate_ukey(udpif, udump, ukey)) {
1543 dpif_flow_del(udpif->dpif, udump->key, udump->key_len, NULL);
1544 ukey_delete(revalidator, ukey);
1547 list_remove(&udump->list_node);
1551 push_dump_ops(revalidator, ops, n_ops);
1553 LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) {
1554 list_remove(&udump->list_node);
1560 revalidator_sweep__(struct revalidator *revalidator, bool purge)
1562 struct dump_op ops[REVALIDATE_MAX_BATCH];
1563 struct udpif_key *ukey, *next;
1568 HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, &revalidator->ukeys) {
1569 if (!purge && ukey->mark) {
1572 struct dump_op *op = &ops[n_ops++];
1574 /* If we have previously seen a flow in the datapath, but didn't
1575 * see it during the most recent dump, delete it. This allows us
1576 * to clean up the ukey and keep the statistics consistent. */
1577 dump_op_init(op, ukey->key, ukey->key_len, ukey, NULL);
1578 if (n_ops == REVALIDATE_MAX_BATCH) {
1579 push_dump_ops(revalidator, ops, n_ops);
1586 push_dump_ops(revalidator, ops, n_ops);
1591 revalidator_sweep(struct revalidator *revalidator)
1593 revalidator_sweep__(revalidator, false);
1597 revalidator_purge(struct revalidator *revalidator)
1599 revalidator_sweep__(revalidator, true);
1603 upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED,
1604 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
1606 struct ds ds = DS_EMPTY_INITIALIZER;
1607 struct udpif *udpif;
1609 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1610 unsigned int flow_limit;
1613 atomic_read(&udpif->flow_limit, &flow_limit);
1615 ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif));
1616 ds_put_format(&ds, "\tflows : (current %"PRIu64")"
1617 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif),
1618 udpif->avg_n_flows, udpif->max_n_flows, flow_limit);
1619 ds_put_format(&ds, "\tdump duration : %lldms\n", udpif->dump_duration);
1621 ds_put_char(&ds, '\n');
1622 for (i = 0; i < udpif->n_handlers; i++) {
1623 struct handler *handler = &udpif->handlers[i];
1625 ovs_mutex_lock(&handler->mutex);
1626 ds_put_format(&ds, "\t%s: (upcall queue %"PRIuSIZE")\n",
1627 handler->name, handler->n_upcalls);
1628 ovs_mutex_unlock(&handler->mutex);
1631 ds_put_char(&ds, '\n');
1632 for (i = 0; i < n_revalidators; i++) {
1633 struct revalidator *revalidator = &udpif->revalidators[i];
1635 /* XXX: The result of hmap_count(&revalidator->ukeys) may not be
1636 * accurate because it's not protected by the revalidator mutex. */
1637 ovs_mutex_lock(&revalidator->mutex);
1638 ds_put_format(&ds, "\t%s: (dump queue %"PRIuSIZE") (keys %"PRIuSIZE
1639 ")\n", revalidator->name, revalidator->n_udumps,
1640 hmap_count(&revalidator->ukeys));
1641 ovs_mutex_unlock(&revalidator->mutex);
1645 unixctl_command_reply(conn, ds_cstr(&ds));
1649 /* Disable using the megaflows.
1651 * This command is only needed for advanced debugging, so it's not
1652 * documented in the man page. */
1654 upcall_unixctl_disable_megaflows(struct unixctl_conn *conn,
1655 int argc OVS_UNUSED,
1656 const char *argv[] OVS_UNUSED,
1657 void *aux OVS_UNUSED)
1659 atomic_store(&enable_megaflows, false);
1661 unixctl_command_reply(conn, "megaflows disabled");
1664 /* Re-enable using megaflows.
1666 * This command is only needed for advanced debugging, so it's not
1667 * documented in the man page. */
1669 upcall_unixctl_enable_megaflows(struct unixctl_conn *conn,
1670 int argc OVS_UNUSED,
1671 const char *argv[] OVS_UNUSED,
1672 void *aux OVS_UNUSED)
1674 atomic_store(&enable_megaflows, true);
1676 unixctl_command_reply(conn, "megaflows enabled");
1679 /* Set the flow limit.
1681 * This command is only needed for advanced debugging, so it's not
1682 * documented in the man page. */
1684 upcall_unixctl_set_flow_limit(struct unixctl_conn *conn,
1685 int argc OVS_UNUSED,
1686 const char *argv[] OVS_UNUSED,
1687 void *aux OVS_UNUSED)
1689 struct ds ds = DS_EMPTY_INITIALIZER;
1690 struct udpif *udpif;
1691 unsigned int flow_limit = atoi(argv[1]);
1693 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1694 atomic_store(&udpif->flow_limit, flow_limit);
1696 ds_put_format(&ds, "set flow_limit to %u\n", flow_limit);
1697 unixctl_command_reply(conn, ds_cstr(&ds));