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 atomic_destroy(&udpif->flow_limit);
289 atomic_destroy(&udpif->n_flows);
290 atomic_destroy(&udpif->n_flows_timestamp);
291 ovs_mutex_destroy(&udpif->n_flows_mutex);
295 /* Tells 'udpif' how many threads it should use to handle upcalls. Disables
296 * all threads if 'n_handlers' and 'n_revalidators' is zero. 'udpif''s
297 * datapath handle must have packet reception enabled before starting threads.
300 udpif_set_threads(struct udpif *udpif, size_t n_handlers,
301 size_t n_revalidators)
303 /* Stop the old threads (if any). */
304 if (udpif->handlers &&
305 (udpif->n_handlers != n_handlers
306 || udpif->n_revalidators != n_revalidators)) {
309 latch_set(&udpif->exit_latch);
311 for (i = 0; i < udpif->n_handlers; i++) {
312 struct handler *handler = &udpif->handlers[i];
314 ovs_mutex_lock(&handler->mutex);
315 xpthread_cond_signal(&handler->wake_cond);
316 ovs_mutex_unlock(&handler->mutex);
317 xpthread_join(handler->thread, NULL);
320 for (i = 0; i < udpif->n_revalidators; i++) {
321 struct revalidator *revalidator = &udpif->revalidators[i];
323 ovs_mutex_lock(&revalidator->mutex);
324 xpthread_cond_signal(&revalidator->wake_cond);
325 ovs_mutex_unlock(&revalidator->mutex);
326 xpthread_join(revalidator->thread, NULL);
329 xpthread_join(udpif->flow_dumper, NULL);
330 xpthread_join(udpif->dispatcher, NULL);
332 for (i = 0; i < udpif->n_revalidators; i++) {
333 struct revalidator *revalidator = &udpif->revalidators[i];
334 struct udpif_flow_dump *udump, *next_udump;
336 LIST_FOR_EACH_SAFE (udump, next_udump, list_node,
337 &revalidator->udumps) {
338 list_remove(&udump->list_node);
342 /* Delete ukeys, and delete all flows from the datapath to prevent
343 * double-counting stats. */
344 revalidator_purge(revalidator);
345 hmap_destroy(&revalidator->ukeys);
346 ovs_mutex_destroy(&revalidator->mutex);
348 free(revalidator->name);
351 for (i = 0; i < udpif->n_handlers; i++) {
352 struct handler *handler = &udpif->handlers[i];
353 struct upcall *miss, *next;
355 LIST_FOR_EACH_SAFE (miss, next, list_node, &handler->upcalls) {
356 list_remove(&miss->list_node);
357 upcall_destroy(miss);
359 ovs_mutex_destroy(&handler->mutex);
361 xpthread_cond_destroy(&handler->wake_cond);
364 latch_poll(&udpif->exit_latch);
366 free(udpif->revalidators);
367 udpif->revalidators = NULL;
368 udpif->n_revalidators = 0;
370 free(udpif->handlers);
371 udpif->handlers = NULL;
372 udpif->n_handlers = 0;
375 /* Start new threads (if necessary). */
376 if (!udpif->handlers && n_handlers) {
379 udpif->n_handlers = n_handlers;
380 udpif->n_revalidators = n_revalidators;
382 udpif->handlers = xzalloc(udpif->n_handlers * sizeof *udpif->handlers);
383 for (i = 0; i < udpif->n_handlers; i++) {
384 struct handler *handler = &udpif->handlers[i];
386 handler->udpif = udpif;
387 list_init(&handler->upcalls);
388 handler->need_signal = false;
389 xpthread_cond_init(&handler->wake_cond, NULL);
390 ovs_mutex_init(&handler->mutex);
391 xpthread_create(&handler->thread, NULL, udpif_upcall_handler,
395 udpif->revalidators = xzalloc(udpif->n_revalidators
396 * sizeof *udpif->revalidators);
397 for (i = 0; i < udpif->n_revalidators; i++) {
398 struct revalidator *revalidator = &udpif->revalidators[i];
400 revalidator->udpif = udpif;
401 list_init(&revalidator->udumps);
402 hmap_init(&revalidator->ukeys);
403 ovs_mutex_init(&revalidator->mutex);
404 xpthread_cond_init(&revalidator->wake_cond, NULL);
405 xpthread_create(&revalidator->thread, NULL, udpif_revalidator,
408 xpthread_create(&udpif->dispatcher, NULL, udpif_dispatcher, udpif);
409 xpthread_create(&udpif->flow_dumper, NULL, udpif_flow_dumper, udpif);
413 /* Waits for all ongoing upcall translations to complete. This ensures that
414 * there are no transient references to any removed ofprotos (or other
415 * objects). In particular, this should be called after an ofproto is removed
416 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
418 udpif_synchronize(struct udpif *udpif)
420 /* This is stronger than necessary. It would be sufficient to ensure
421 * (somehow) that each handler and revalidator thread had passed through
422 * its main loop once. */
423 size_t n_handlers = udpif->n_handlers;
424 size_t n_revalidators = udpif->n_revalidators;
425 udpif_set_threads(udpif, 0, 0);
426 udpif_set_threads(udpif, n_handlers, n_revalidators);
429 /* Notifies 'udpif' that something changed which may render previous
430 * xlate_actions() results invalid. */
432 udpif_revalidate(struct udpif *udpif)
434 seq_change(udpif->reval_seq);
437 /* Returns a seq which increments every time 'udpif' pulls stats from the
438 * datapath. Callers can use this to get a sense of when might be a good time
439 * to do periodic work which relies on relatively up to date statistics. */
441 udpif_dump_seq(struct udpif *udpif)
443 return udpif->dump_seq;
447 udpif_get_memory_usage(struct udpif *udpif, struct simap *usage)
451 simap_increase(usage, "dispatchers", 1);
452 simap_increase(usage, "flow_dumpers", 1);
454 simap_increase(usage, "handlers", udpif->n_handlers);
455 for (i = 0; i < udpif->n_handlers; i++) {
456 struct handler *handler = &udpif->handlers[i];
457 ovs_mutex_lock(&handler->mutex);
458 simap_increase(usage, "handler upcalls", handler->n_upcalls);
459 ovs_mutex_unlock(&handler->mutex);
462 simap_increase(usage, "revalidators", udpif->n_revalidators);
463 for (i = 0; i < udpif->n_revalidators; i++) {
464 struct revalidator *revalidator = &udpif->revalidators[i];
465 ovs_mutex_lock(&revalidator->mutex);
466 simap_increase(usage, "revalidator dumps", revalidator->n_udumps);
468 /* XXX: This isn't technically thread safe because the revalidator
469 * ukeys maps isn't protected by a mutex since it's per thread. */
470 simap_increase(usage, "revalidator keys",
471 hmap_count(&revalidator->ukeys));
472 ovs_mutex_unlock(&revalidator->mutex);
476 /* Removes all flows from all datapaths. */
482 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
483 dpif_flow_flush(udpif->dpif);
487 /* Destroys and deallocates 'upcall'. */
489 upcall_destroy(struct upcall *upcall)
492 ofpbuf_uninit(&upcall->dpif_upcall.packet);
493 ofpbuf_uninit(&upcall->upcall_buf);
499 udpif_get_n_flows(struct udpif *udpif)
501 long long int time, now;
505 atomic_read(&udpif->n_flows_timestamp, &time);
506 if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) {
507 struct dpif_dp_stats stats;
509 atomic_store(&udpif->n_flows_timestamp, now);
510 dpif_get_dp_stats(udpif->dpif, &stats);
511 flow_count = stats.n_flows;
512 atomic_store(&udpif->n_flows, flow_count);
513 ovs_mutex_unlock(&udpif->n_flows_mutex);
515 atomic_read(&udpif->n_flows, &flow_count);
520 /* The dispatcher thread is responsible for receiving upcalls from the kernel,
521 * assigning them to a upcall_handler thread. */
523 udpif_dispatcher(void *arg)
525 struct udpif *udpif = arg;
527 set_subprogram_name("dispatcher");
528 while (!latch_is_set(&udpif->exit_latch)) {
530 dpif_recv_wait(udpif->dpif);
531 latch_wait(&udpif->exit_latch);
539 udpif_flow_dumper(void *arg)
541 struct udpif *udpif = arg;
543 set_subprogram_name("flow_dumper");
544 while (!latch_is_set(&udpif->exit_latch)) {
545 const struct dpif_flow_stats *stats;
546 long long int start_time, duration;
547 const struct nlattr *key, *mask;
548 struct dpif_flow_dump dump;
549 size_t key_len, mask_len;
550 unsigned int flow_limit;
551 bool need_revalidate;
557 reval_seq = seq_read(udpif->reval_seq);
558 need_revalidate = udpif->last_reval_seq != reval_seq;
559 udpif->last_reval_seq = reval_seq;
561 n_flows = udpif_get_n_flows(udpif);
562 udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows);
563 udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2;
565 start_time = time_msec();
566 error = dpif_flow_dump_start(&dump, udpif->dpif);
568 VLOG_INFO("Failed to start flow dump (%s)", ovs_strerror(error));
571 dpif_flow_dump_state_init(udpif->dpif, &state);
572 while (dpif_flow_dump_next(&dump, state, &key, &key_len,
573 &mask, &mask_len, NULL, NULL, &stats)
574 && !latch_is_set(&udpif->exit_latch)) {
575 struct udpif_flow_dump *udump = xmalloc(sizeof *udump);
576 struct revalidator *revalidator;
578 udump->key_hash = hash_bytes(key, key_len, udpif->secret);
579 memcpy(&udump->key_buf, key, key_len);
580 udump->key = (struct nlattr *) &udump->key_buf;
581 udump->key_len = key_len;
583 memcpy(&udump->mask_buf, mask, mask_len);
584 udump->mask = (struct nlattr *) &udump->mask_buf;
585 udump->mask_len = mask_len;
587 udump->stats = *stats;
588 udump->need_revalidate = need_revalidate;
590 revalidator = &udpif->revalidators[udump->key_hash
591 % udpif->n_revalidators];
593 ovs_mutex_lock(&revalidator->mutex);
594 while (revalidator->n_udumps >= REVALIDATE_MAX_BATCH * 3
595 && !latch_is_set(&udpif->exit_latch)) {
596 ovs_mutex_cond_wait(&revalidator->wake_cond,
597 &revalidator->mutex);
599 list_push_back(&revalidator->udumps, &udump->list_node);
600 revalidator->n_udumps++;
601 xpthread_cond_signal(&revalidator->wake_cond);
602 ovs_mutex_unlock(&revalidator->mutex);
604 dpif_flow_dump_state_uninit(udpif->dpif, state);
605 dpif_flow_dump_done(&dump);
607 /* Let all the revalidators finish and garbage collect. */
608 seq_change(udpif->dump_seq);
609 for (i = 0; i < udpif->n_revalidators; i++) {
610 struct revalidator *revalidator = &udpif->revalidators[i];
611 ovs_mutex_lock(&revalidator->mutex);
612 xpthread_cond_signal(&revalidator->wake_cond);
613 ovs_mutex_unlock(&revalidator->mutex);
616 for (i = 0; i < udpif->n_revalidators; i++) {
617 struct revalidator *revalidator = &udpif->revalidators[i];
619 ovs_mutex_lock(&revalidator->mutex);
620 while (revalidator->dump_seq != seq_read(udpif->dump_seq)
621 && !latch_is_set(&udpif->exit_latch)) {
622 ovs_mutex_cond_wait(&revalidator->wake_cond,
623 &revalidator->mutex);
625 ovs_mutex_unlock(&revalidator->mutex);
628 duration = MAX(time_msec() - start_time, 1);
629 udpif->dump_duration = duration;
630 atomic_read(&udpif->flow_limit, &flow_limit);
631 if (duration > 2000) {
632 flow_limit /= duration / 1000;
633 } else if (duration > 1300) {
634 flow_limit = flow_limit * 3 / 4;
635 } else if (duration < 1000 && n_flows > 2000
636 && flow_limit < n_flows * 1000 / duration) {
639 flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000));
640 atomic_store(&udpif->flow_limit, flow_limit);
642 if (duration > 2000) {
643 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
648 poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500));
649 seq_wait(udpif->reval_seq, udpif->last_reval_seq);
650 latch_wait(&udpif->exit_latch);
657 /* The miss handler thread is responsible for processing miss upcalls retrieved
658 * by the dispatcher thread. Once finished it passes the processed miss
659 * upcalls to ofproto-dpif where they're installed in the datapath. */
661 udpif_upcall_handler(void *arg)
663 struct handler *handler = arg;
665 handler->name = xasprintf("handler_%u", ovsthread_id_self());
666 set_subprogram_name("%s", handler->name);
668 while (!latch_is_set(&handler->udpif->exit_latch)) {
669 struct list misses = LIST_INITIALIZER(&misses);
672 ovs_mutex_lock(&handler->mutex);
673 if (!handler->n_upcalls) {
674 ovs_mutex_cond_wait(&handler->wake_cond, &handler->mutex);
677 for (i = 0; i < FLOW_MISS_MAX_BATCH; i++) {
678 if (handler->n_upcalls) {
679 handler->n_upcalls--;
680 list_push_back(&misses, list_pop_front(&handler->upcalls));
685 ovs_mutex_unlock(&handler->mutex);
687 handle_upcalls(handler, &misses);
696 udpif_revalidator(void *arg)
698 struct revalidator *revalidator = arg;
700 revalidator->name = xasprintf("revalidator_%u", ovsthread_id_self());
701 set_subprogram_name("%s", revalidator->name);
703 struct list udumps = LIST_INITIALIZER(&udumps);
704 struct udpif *udpif = revalidator->udpif;
707 ovs_mutex_lock(&revalidator->mutex);
708 if (latch_is_set(&udpif->exit_latch)) {
709 ovs_mutex_unlock(&revalidator->mutex);
713 if (!revalidator->n_udumps) {
714 if (revalidator->dump_seq != seq_read(udpif->dump_seq)) {
715 revalidator->dump_seq = seq_read(udpif->dump_seq);
716 revalidator_sweep(revalidator);
718 ovs_mutex_cond_wait(&revalidator->wake_cond,
719 &revalidator->mutex);
723 for (i = 0; i < REVALIDATE_MAX_BATCH && revalidator->n_udumps; i++) {
724 list_push_back(&udumps, list_pop_front(&revalidator->udumps));
725 revalidator->n_udumps--;
728 /* Wake up the flow dumper. */
729 xpthread_cond_signal(&revalidator->wake_cond);
730 ovs_mutex_unlock(&revalidator->mutex);
732 if (!list_is_empty(&udumps)) {
733 revalidate_udumps(revalidator, &udumps);
740 static enum upcall_type
741 classify_upcall(const struct upcall *upcall)
743 const struct dpif_upcall *dpif_upcall = &upcall->dpif_upcall;
744 union user_action_cookie cookie;
747 /* First look at the upcall type. */
748 switch (dpif_upcall->type) {
755 case DPIF_N_UC_TYPES:
757 VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32,
762 /* "action" upcalls need a closer look. */
763 if (!dpif_upcall->userdata) {
764 VLOG_WARN_RL(&rl, "action upcall missing cookie");
767 userdata_len = nl_attr_get_size(dpif_upcall->userdata);
768 if (userdata_len < sizeof cookie.type
769 || userdata_len > sizeof cookie) {
770 VLOG_WARN_RL(&rl, "action upcall cookie has unexpected size %"PRIuSIZE,
774 memset(&cookie, 0, sizeof cookie);
775 memcpy(&cookie, nl_attr_get(dpif_upcall->userdata), userdata_len);
776 if (userdata_len == MAX(8, sizeof cookie.sflow)
777 && cookie.type == USER_ACTION_COOKIE_SFLOW) {
779 } else if (userdata_len == MAX(8, sizeof cookie.slow_path)
780 && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) {
782 } else if (userdata_len == MAX(8, sizeof cookie.flow_sample)
783 && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) {
784 return FLOW_SAMPLE_UPCALL;
785 } else if (userdata_len == MAX(8, sizeof cookie.ipfix)
786 && cookie.type == USER_ACTION_COOKIE_IPFIX) {
789 VLOG_WARN_RL(&rl, "invalid user cookie of type %"PRIu16
790 " and size %"PRIuSIZE, cookie.type, userdata_len);
796 recv_upcalls(struct udpif *udpif)
801 uint32_t hash = udpif->secret;
802 struct handler *handler;
803 struct upcall *upcall;
804 size_t n_bytes, left;
808 upcall = xmalloc(sizeof *upcall);
809 ofpbuf_use_stub(&upcall->upcall_buf, upcall->upcall_stub,
810 sizeof upcall->upcall_stub);
811 error = dpif_recv(udpif->dpif, &upcall->dpif_upcall,
812 &upcall->upcall_buf);
814 /* upcall_destroy() can only be called on successfully received
816 ofpbuf_uninit(&upcall->upcall_buf);
822 NL_ATTR_FOR_EACH (nla, left, upcall->dpif_upcall.key,
823 upcall->dpif_upcall.key_len) {
824 enum ovs_key_attr type = nl_attr_type(nla);
825 if (type == OVS_KEY_ATTR_IN_PORT
826 || type == OVS_KEY_ATTR_TCP
827 || type == OVS_KEY_ATTR_UDP) {
828 if (nl_attr_get_size(nla) == 4) {
829 hash = mhash_add(hash, nl_attr_get_u32(nla));
833 "Netlink attribute with incorrect size.");
837 hash = mhash_finish(hash, n_bytes);
839 handler = &udpif->handlers[hash % udpif->n_handlers];
841 ovs_mutex_lock(&handler->mutex);
842 if (handler->n_upcalls < MAX_QUEUE_LENGTH) {
843 list_push_back(&handler->upcalls, &upcall->list_node);
844 if (handler->n_upcalls == 0) {
845 handler->need_signal = true;
847 handler->n_upcalls++;
848 if (handler->need_signal &&
849 handler->n_upcalls >= FLOW_MISS_MAX_BATCH) {
850 handler->need_signal = false;
851 xpthread_cond_signal(&handler->wake_cond);
853 ovs_mutex_unlock(&handler->mutex);
854 if (!VLOG_DROP_DBG(&rl)) {
855 struct ds ds = DS_EMPTY_INITIALIZER;
857 odp_flow_key_format(upcall->dpif_upcall.key,
858 upcall->dpif_upcall.key_len,
860 VLOG_DBG("dispatcher: enqueue (%s)", ds_cstr(&ds));
864 ovs_mutex_unlock(&handler->mutex);
865 COVERAGE_INC(upcall_queue_overflow);
866 upcall_destroy(upcall);
870 for (n = 0; n < udpif->n_handlers; ++n) {
871 struct handler *handler = &udpif->handlers[n];
873 if (handler->need_signal) {
874 handler->need_signal = false;
875 ovs_mutex_lock(&handler->mutex);
876 xpthread_cond_signal(&handler->wake_cond);
877 ovs_mutex_unlock(&handler->mutex);
882 /* Calculates slow path actions for 'xout'. 'buf' must statically be
883 * initialized with at least 128 bytes of space. */
885 compose_slow_path(struct udpif *udpif, struct xlate_out *xout,
886 odp_port_t odp_in_port, struct ofpbuf *buf)
888 union user_action_cookie cookie;
892 cookie.type = USER_ACTION_COOKIE_SLOW_PATH;
893 cookie.slow_path.unused = 0;
894 cookie.slow_path.reason = xout->slow;
896 port = xout->slow & (SLOW_CFM | SLOW_BFD | SLOW_LACP | SLOW_STP)
899 pid = dpif_port_get_pid(udpif->dpif, port);
900 odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, buf);
903 static struct flow_miss *
904 flow_miss_find(struct hmap *todo, const struct ofproto_dpif *ofproto,
905 const struct flow *flow, uint32_t hash)
907 struct flow_miss *miss;
909 HMAP_FOR_EACH_WITH_HASH (miss, hmap_node, hash, todo) {
910 if (miss->ofproto == ofproto && flow_equal(&miss->flow, flow)) {
919 handle_upcalls(struct handler *handler, struct list *upcalls)
921 struct hmap misses = HMAP_INITIALIZER(&misses);
922 struct udpif *udpif = handler->udpif;
924 struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH];
925 struct dpif_op *opsp[FLOW_MISS_MAX_BATCH * 2];
926 struct dpif_op ops[FLOW_MISS_MAX_BATCH * 2];
927 struct flow_miss *miss, *next_miss;
928 struct upcall *upcall, *next;
929 size_t n_misses, n_ops, i;
930 unsigned int flow_limit;
931 bool fail_open, may_put;
932 enum upcall_type type;
934 atomic_read(&udpif->flow_limit, &flow_limit);
935 may_put = udpif_get_n_flows(udpif) < flow_limit;
937 /* Extract the flow from each upcall. Construct in 'misses' a hash table
938 * that maps each unique flow to a 'struct flow_miss'.
940 * Most commonly there is a single packet per flow_miss, but there are
941 * several reasons why there might be more than one, e.g.:
943 * - The dpif packet interface does not support TSO (or UFO, etc.), so a
944 * large packet sent to userspace is split into a sequence of smaller
947 * - A stream of quickly arriving packets in an established "slow-pathed"
950 * - Rarely, a stream of quickly arriving packets in a flow not yet
951 * established. (This is rare because most protocols do not send
952 * multiple back-to-back packets before receiving a reply from the
953 * other end of the connection, which gives OVS a chance to set up a
957 LIST_FOR_EACH_SAFE (upcall, next, list_node, upcalls) {
958 struct dpif_upcall *dupcall = &upcall->dpif_upcall;
959 struct flow_miss *miss = &miss_buf[n_misses];
960 struct ofpbuf *packet = &dupcall->packet;
961 struct flow_miss *existing_miss;
962 struct ofproto_dpif *ofproto;
963 struct dpif_sflow *sflow;
964 struct dpif_ipfix *ipfix;
965 odp_port_t odp_in_port;
969 error = xlate_receive(udpif->backer, packet, dupcall->key,
970 dupcall->key_len, &flow,
971 &ofproto, &ipfix, &sflow, NULL, &odp_in_port);
973 if (error == ENODEV) {
974 /* Received packet on datapath port for which we couldn't
975 * associate an ofproto. This can happen if a port is removed
976 * while traffic is being received. Print a rate-limited
977 * message in case it happens frequently. Install a drop flow
978 * so that future packets of the flow are inexpensively dropped
980 VLOG_INFO_RL(&rl, "received packet on unassociated datapath "
981 "port %"PRIu32, odp_in_port);
982 dpif_flow_put(udpif->dpif, DPIF_FP_CREATE | DPIF_FP_MODIFY,
983 dupcall->key, dupcall->key_len, NULL, 0, NULL, 0,
986 list_remove(&upcall->list_node);
987 upcall_destroy(upcall);
991 type = classify_upcall(upcall);
992 if (type == MISS_UPCALL) {
994 struct pkt_metadata md;
996 pkt_metadata_from_flow(&md, &flow);
997 flow_extract(packet, &md, &miss->flow);
999 hash = flow_hash(&miss->flow, 0);
1000 existing_miss = flow_miss_find(&misses, ofproto, &miss->flow,
1002 if (!existing_miss) {
1003 hmap_insert(&misses, &miss->hmap_node, hash);
1004 miss->ofproto = ofproto;
1005 miss->key = dupcall->key;
1006 miss->key_len = dupcall->key_len;
1007 miss->upcall_type = dupcall->type;
1008 miss->stats.n_packets = 0;
1009 miss->stats.n_bytes = 0;
1010 miss->stats.used = time_msec();
1011 miss->stats.tcp_flags = 0;
1012 miss->odp_in_port = odp_in_port;
1017 miss = existing_miss;
1019 miss->stats.tcp_flags |= packet_get_tcp_flags(packet, &miss->flow);
1020 miss->stats.n_bytes += packet->size;
1021 miss->stats.n_packets++;
1023 upcall->flow_miss = miss;
1030 union user_action_cookie cookie;
1032 memset(&cookie, 0, sizeof cookie);
1033 memcpy(&cookie, nl_attr_get(dupcall->userdata),
1034 sizeof cookie.sflow);
1035 dpif_sflow_received(sflow, packet, &flow, odp_in_port,
1041 dpif_ipfix_bridge_sample(ipfix, packet, &flow);
1044 case FLOW_SAMPLE_UPCALL:
1046 union user_action_cookie cookie;
1048 memset(&cookie, 0, sizeof cookie);
1049 memcpy(&cookie, nl_attr_get(dupcall->userdata),
1050 sizeof cookie.flow_sample);
1052 /* The flow reflects exactly the contents of the packet.
1053 * Sample the packet using it. */
1054 dpif_ipfix_flow_sample(ipfix, packet, &flow,
1055 cookie.flow_sample.collector_set_id,
1056 cookie.flow_sample.probability,
1057 cookie.flow_sample.obs_domain_id,
1058 cookie.flow_sample.obs_point_id);
1067 dpif_ipfix_unref(ipfix);
1068 dpif_sflow_unref(sflow);
1070 list_remove(&upcall->list_node);
1071 upcall_destroy(upcall);
1074 /* Initialize each 'struct flow_miss's ->xout.
1076 * We do this per-flow_miss rather than per-packet because, most commonly,
1077 * all the packets in a flow can use the same translation.
1079 * We can't do this in the previous loop because we need the TCP flags for
1080 * all the packets in each miss. */
1082 HMAP_FOR_EACH (miss, hmap_node, &misses) {
1083 struct xlate_in xin;
1085 xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL,
1086 miss->stats.tcp_flags, NULL);
1087 xin.may_learn = true;
1089 if (miss->upcall_type == DPIF_UC_MISS) {
1090 xin.resubmit_stats = &miss->stats;
1092 /* For non-miss upcalls, there's a flow in the datapath which this
1093 * packet was accounted to. Presumably the revalidators will deal
1094 * with pushing its stats eventually. */
1097 xlate_actions(&xin, &miss->xout);
1098 fail_open = fail_open || miss->xout.fail_open;
1101 /* Now handle the packets individually in order of arrival. In the common
1102 * case each packet of a miss can share the same actions, but slow-pathed
1103 * packets need to be translated individually:
1105 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1106 * processes received packets for these protocols.
1108 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1111 * The loop fills 'ops' with an array of operations to execute in the
1114 LIST_FOR_EACH (upcall, list_node, upcalls) {
1115 struct flow_miss *miss = upcall->flow_miss;
1116 struct ofpbuf *packet = &upcall->dpif_upcall.packet;
1118 ovs_be16 flow_vlan_tci;
1120 /* Save a copy of flow.vlan_tci in case it is changed to
1121 * generate proper mega flow masks for VLAN splinter flows. */
1122 flow_vlan_tci = miss->flow.vlan_tci;
1124 if (miss->xout.slow) {
1125 struct xlate_in xin;
1127 xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL, 0, packet);
1128 xlate_actions_for_side_effects(&xin);
1131 if (miss->flow.in_port.ofp_port
1132 != vsp_realdev_to_vlandev(miss->ofproto,
1133 miss->flow.in_port.ofp_port,
1134 miss->flow.vlan_tci)) {
1135 /* This packet was received on a VLAN splinter port. We
1136 * added a VLAN to the packet to make the packet resemble
1137 * the flow, but the actions were composed assuming that
1138 * the packet contained no VLAN. So, we must remove the
1139 * VLAN header from the packet before trying to execute the
1141 if (miss->xout.odp_actions.size) {
1142 eth_pop_vlan(packet);
1145 /* Remove the flow vlan tags inserted by vlan splinter logic
1146 * to ensure megaflow masks generated match the data path flow. */
1147 miss->flow.vlan_tci = 0;
1150 /* Do not install a flow into the datapath if:
1152 * - The datapath already has too many flows.
1154 * - An earlier iteration of this loop already put the same flow.
1156 * - We received this packet via some flow installed in the kernel
1160 && upcall->dpif_upcall.type == DPIF_UC_MISS) {
1166 atomic_read(&enable_megaflows, &megaflow);
1167 ofpbuf_use_stack(&mask, &miss->mask_buf, sizeof miss->mask_buf);
1171 max_mpls = ofproto_dpif_get_max_mpls_depth(miss->ofproto);
1172 odp_flow_key_from_mask(&mask, &miss->xout.wc.masks,
1173 &miss->flow, UINT32_MAX, max_mpls);
1177 op->type = DPIF_OP_FLOW_PUT;
1178 op->u.flow_put.flags = DPIF_FP_CREATE | DPIF_FP_MODIFY;
1179 op->u.flow_put.key = miss->key;
1180 op->u.flow_put.key_len = miss->key_len;
1181 op->u.flow_put.mask = mask.data;
1182 op->u.flow_put.mask_len = mask.size;
1183 op->u.flow_put.stats = NULL;
1185 if (!miss->xout.slow) {
1186 op->u.flow_put.actions = miss->xout.odp_actions.data;
1187 op->u.flow_put.actions_len = miss->xout.odp_actions.size;
1191 ofpbuf_use_stack(&buf, miss->slow_path_buf,
1192 sizeof miss->slow_path_buf);
1193 compose_slow_path(udpif, &miss->xout, miss->odp_in_port, &buf);
1194 op->u.flow_put.actions = buf.data;
1195 op->u.flow_put.actions_len = buf.size;
1200 * The 'miss' may be shared by multiple upcalls. Restore
1201 * the saved flow vlan_tci field before processing the next
1203 miss->flow.vlan_tci = flow_vlan_tci;
1205 if (miss->xout.odp_actions.size) {
1208 op->type = DPIF_OP_EXECUTE;
1209 op->u.execute.packet = packet;
1210 odp_key_to_pkt_metadata(miss->key, miss->key_len,
1212 op->u.execute.actions = miss->xout.odp_actions.data;
1213 op->u.execute.actions_len = miss->xout.odp_actions.size;
1214 op->u.execute.needs_help = (miss->xout.slow & SLOW_ACTION) != 0;
1218 /* Special case for fail-open mode.
1220 * If we are in fail-open mode, but we are connected to a controller too,
1221 * then we should send the packet up to the controller in the hope that it
1222 * will try to set up a flow and thereby allow us to exit fail-open.
1224 * See the top-level comment in fail-open.c for more information.
1226 * Copy packets before they are modified by execution. */
1228 LIST_FOR_EACH (upcall, list_node, upcalls) {
1229 struct flow_miss *miss = upcall->flow_miss;
1230 struct ofpbuf *packet = &upcall->dpif_upcall.packet;
1231 struct ofproto_packet_in *pin;
1233 pin = xmalloc(sizeof *pin);
1234 pin->up.packet = xmemdup(packet->data, packet->size);
1235 pin->up.packet_len = packet->size;
1236 pin->up.reason = OFPR_NO_MATCH;
1237 pin->up.table_id = 0;
1238 pin->up.cookie = OVS_BE64_MAX;
1239 flow_get_metadata(&miss->flow, &pin->up.fmd);
1240 pin->send_len = 0; /* Not used for flow table misses. */
1241 pin->generated_by_table_miss = false;
1242 ofproto_dpif_send_packet_in(miss->ofproto, pin);
1246 /* Execute batch. */
1247 for (i = 0; i < n_ops; i++) {
1250 dpif_operate(udpif->dpif, opsp, n_ops);
1252 HMAP_FOR_EACH_SAFE (miss, next_miss, hmap_node, &misses) {
1253 hmap_remove(&misses, &miss->hmap_node);
1254 xlate_out_uninit(&miss->xout);
1256 hmap_destroy(&misses);
1258 LIST_FOR_EACH_SAFE (upcall, next, list_node, upcalls) {
1259 list_remove(&upcall->list_node);
1260 upcall_destroy(upcall);
1264 static struct udpif_key *
1265 ukey_lookup(struct revalidator *revalidator, struct udpif_flow_dump *udump)
1267 struct udpif_key *ukey;
1269 HMAP_FOR_EACH_WITH_HASH (ukey, hmap_node, udump->key_hash,
1270 &revalidator->ukeys) {
1271 if (ukey->key_len == udump->key_len
1272 && !memcmp(ukey->key, udump->key, udump->key_len)) {
1279 static struct udpif_key *
1280 ukey_create(const struct nlattr *key, size_t key_len, long long int used)
1282 struct udpif_key *ukey = xmalloc(sizeof *ukey);
1284 ukey->key = (struct nlattr *) &ukey->key_buf;
1285 memcpy(&ukey->key_buf, key, key_len);
1286 ukey->key_len = key_len;
1289 ukey->created = used ? used : time_msec();
1290 memset(&ukey->stats, 0, sizeof ukey->stats);
1296 ukey_delete(struct revalidator *revalidator, struct udpif_key *ukey)
1298 hmap_remove(&revalidator->ukeys, &ukey->hmap_node);
1303 revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump,
1304 struct udpif_key *ukey)
1306 struct ofpbuf xout_actions, *actions;
1307 uint64_t slow_path_buf[128 / 8];
1308 struct xlate_out xout, *xoutp;
1309 struct flow flow, udump_mask;
1310 struct ofproto_dpif *ofproto;
1311 struct dpif_flow_stats push;
1312 uint32_t *udump32, *xout32;
1313 odp_port_t odp_in_port;
1314 struct xlate_in xin;
1323 /* If we don't need to revalidate, we can simply push the stats contained
1324 * in the udump, otherwise we'll have to get the actions so we can check
1326 if (udump->need_revalidate) {
1327 if (dpif_flow_get(udpif->dpif, ukey->key, ukey->key_len, &actions,
1333 push.used = udump->stats.used;
1334 push.tcp_flags = udump->stats.tcp_flags;
1335 push.n_packets = udump->stats.n_packets > ukey->stats.n_packets
1336 ? udump->stats.n_packets - ukey->stats.n_packets
1338 push.n_bytes = udump->stats.n_bytes > ukey->stats.n_bytes
1339 ? udump->stats.n_bytes - ukey->stats.n_bytes
1341 ukey->stats = udump->stats;
1343 if (!push.n_packets && !udump->need_revalidate) {
1348 error = xlate_receive(udpif->backer, NULL, ukey->key, ukey->key_len, &flow,
1349 &ofproto, NULL, NULL, NULL, &odp_in_port);
1354 xlate_in_init(&xin, ofproto, &flow, NULL, push.tcp_flags, NULL);
1355 xin.resubmit_stats = push.n_packets ? &push : NULL;
1356 xin.may_learn = push.n_packets > 0;
1357 xin.skip_wildcards = !udump->need_revalidate;
1358 xlate_actions(&xin, &xout);
1361 if (!udump->need_revalidate) {
1367 ofpbuf_use_const(&xout_actions, xout.odp_actions.data,
1368 xout.odp_actions.size);
1370 ofpbuf_use_stack(&xout_actions, slow_path_buf, sizeof slow_path_buf);
1371 compose_slow_path(udpif, &xout, odp_in_port, &xout_actions);
1374 if (!ofpbuf_equal(&xout_actions, actions)) {
1378 if (odp_flow_key_to_mask(udump->mask, udump->mask_len, &udump_mask, &flow)
1383 /* Since the kernel is free to ignore wildcarded bits in the mask, we can't
1384 * directly check that the masks are the same. Instead we check that the
1385 * mask in the kernel is more specific i.e. less wildcarded, than what
1386 * we've calculated here. This guarantees we don't catch any packets we
1387 * shouldn't with the megaflow. */
1388 udump32 = (uint32_t *) &udump_mask;
1389 xout32 = (uint32_t *) &xout.wc.masks;
1390 for (i = 0; i < FLOW_U32S; i++) {
1391 if ((udump32[i] | xout32[i]) != udump32[i]) {
1398 ofpbuf_delete(actions);
1399 xlate_out_uninit(xoutp);
1404 struct udpif_key *ukey;
1405 struct udpif_flow_dump *udump;
1406 struct dpif_flow_stats stats; /* Stats for 'op'. */
1407 struct dpif_op op; /* Flow del operation. */
1411 dump_op_init(struct dump_op *op, const struct nlattr *key, size_t key_len,
1412 struct udpif_key *ukey, struct udpif_flow_dump *udump)
1416 op->op.type = DPIF_OP_FLOW_DEL;
1417 op->op.u.flow_del.key = key;
1418 op->op.u.flow_del.key_len = key_len;
1419 op->op.u.flow_del.stats = &op->stats;
1423 push_dump_ops(struct revalidator *revalidator,
1424 struct dump_op *ops, size_t n_ops)
1426 struct udpif *udpif = revalidator->udpif;
1427 struct dpif_op *opsp[REVALIDATE_MAX_BATCH];
1430 ovs_assert(n_ops <= REVALIDATE_MAX_BATCH);
1431 for (i = 0; i < n_ops; i++) {
1432 opsp[i] = &ops[i].op;
1434 dpif_operate(udpif->dpif, opsp, n_ops);
1436 for (i = 0; i < n_ops; i++) {
1437 struct dump_op *op = &ops[i];
1438 struct dpif_flow_stats *push, *stats, push_buf;
1440 stats = op->op.u.flow_del.stats;
1443 push->used = MAX(stats->used, op->ukey->stats.used);
1444 push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags;
1445 push->n_packets = stats->n_packets - op->ukey->stats.n_packets;
1446 push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes;
1451 if (push->n_packets || netflow_exists()) {
1452 struct ofproto_dpif *ofproto;
1453 struct netflow *netflow;
1456 if (!xlate_receive(udpif->backer, NULL, op->op.u.flow_del.key,
1457 op->op.u.flow_del.key_len, &flow, &ofproto,
1458 NULL, NULL, &netflow, NULL)) {
1459 struct xlate_in xin;
1461 xlate_in_init(&xin, ofproto, &flow, NULL, push->tcp_flags,
1463 xin.resubmit_stats = push->n_packets ? push : NULL;
1464 xin.may_learn = push->n_packets > 0;
1465 xin.skip_wildcards = true;
1466 xlate_actions_for_side_effects(&xin);
1469 netflow_expire(netflow, &flow);
1470 netflow_flow_clear(netflow, &flow);
1471 netflow_unref(netflow);
1477 for (i = 0; i < n_ops; i++) {
1478 struct udpif_key *ukey;
1480 /* If there's a udump, this ukey came directly from a datapath flow
1481 * dump. Sometimes a datapath can send duplicates in flow dumps, in
1482 * which case we wouldn't want to double-free a ukey, so avoid that by
1483 * looking up the ukey again.
1485 * If there's no udump then we know what we're doing. */
1486 ukey = (ops[i].udump
1487 ? ukey_lookup(revalidator, ops[i].udump)
1490 ukey_delete(revalidator, ukey);
1496 revalidate_udumps(struct revalidator *revalidator, struct list *udumps)
1498 struct udpif *udpif = revalidator->udpif;
1500 struct dump_op ops[REVALIDATE_MAX_BATCH];
1501 struct udpif_flow_dump *udump, *next_udump;
1502 size_t n_ops, n_flows;
1503 unsigned int flow_limit;
1504 long long int max_idle;
1507 atomic_read(&udpif->flow_limit, &flow_limit);
1509 n_flows = udpif_get_n_flows(udpif);
1512 max_idle = ofproto_max_idle;
1513 if (n_flows > flow_limit) {
1514 must_del = n_flows > 2 * flow_limit;
1519 LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) {
1520 long long int used, now;
1521 struct udpif_key *ukey;
1524 ukey = ukey_lookup(revalidator, udump);
1526 used = udump->stats.used;
1527 if (!used && ukey) {
1528 used = ukey->created;
1531 if (must_del || (used && used < now - max_idle)) {
1532 struct dump_op *dop = &ops[n_ops++];
1534 dump_op_init(dop, udump->key, udump->key_len, ukey, udump);
1539 ukey = ukey_create(udump->key, udump->key_len, used);
1540 hmap_insert(&revalidator->ukeys, &ukey->hmap_node,
1545 if (!revalidate_ukey(udpif, udump, ukey)) {
1546 dpif_flow_del(udpif->dpif, udump->key, udump->key_len, NULL);
1547 ukey_delete(revalidator, ukey);
1550 list_remove(&udump->list_node);
1554 push_dump_ops(revalidator, ops, n_ops);
1556 LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) {
1557 list_remove(&udump->list_node);
1563 revalidator_sweep__(struct revalidator *revalidator, bool purge)
1565 struct dump_op ops[REVALIDATE_MAX_BATCH];
1566 struct udpif_key *ukey, *next;
1571 HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, &revalidator->ukeys) {
1572 if (!purge && ukey->mark) {
1575 struct dump_op *op = &ops[n_ops++];
1577 /* If we have previously seen a flow in the datapath, but didn't
1578 * see it during the most recent dump, delete it. This allows us
1579 * to clean up the ukey and keep the statistics consistent. */
1580 dump_op_init(op, ukey->key, ukey->key_len, ukey, NULL);
1581 if (n_ops == REVALIDATE_MAX_BATCH) {
1582 push_dump_ops(revalidator, ops, n_ops);
1589 push_dump_ops(revalidator, ops, n_ops);
1594 revalidator_sweep(struct revalidator *revalidator)
1596 revalidator_sweep__(revalidator, false);
1600 revalidator_purge(struct revalidator *revalidator)
1602 revalidator_sweep__(revalidator, true);
1606 upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED,
1607 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
1609 struct ds ds = DS_EMPTY_INITIALIZER;
1610 struct udpif *udpif;
1612 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1613 unsigned int flow_limit;
1616 atomic_read(&udpif->flow_limit, &flow_limit);
1618 ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif));
1619 ds_put_format(&ds, "\tflows : (current %"PRIu64")"
1620 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif),
1621 udpif->avg_n_flows, udpif->max_n_flows, flow_limit);
1622 ds_put_format(&ds, "\tdump duration : %lldms\n", udpif->dump_duration);
1624 ds_put_char(&ds, '\n');
1625 for (i = 0; i < udpif->n_handlers; i++) {
1626 struct handler *handler = &udpif->handlers[i];
1628 ovs_mutex_lock(&handler->mutex);
1629 ds_put_format(&ds, "\t%s: (upcall queue %"PRIuSIZE")\n",
1630 handler->name, handler->n_upcalls);
1631 ovs_mutex_unlock(&handler->mutex);
1634 ds_put_char(&ds, '\n');
1635 for (i = 0; i < n_revalidators; i++) {
1636 struct revalidator *revalidator = &udpif->revalidators[i];
1638 /* XXX: The result of hmap_count(&revalidator->ukeys) may not be
1639 * accurate because it's not protected by the revalidator mutex. */
1640 ovs_mutex_lock(&revalidator->mutex);
1641 ds_put_format(&ds, "\t%s: (dump queue %"PRIuSIZE") (keys %"PRIuSIZE
1642 ")\n", revalidator->name, revalidator->n_udumps,
1643 hmap_count(&revalidator->ukeys));
1644 ovs_mutex_unlock(&revalidator->mutex);
1648 unixctl_command_reply(conn, ds_cstr(&ds));
1652 /* Disable using the megaflows.
1654 * This command is only needed for advanced debugging, so it's not
1655 * documented in the man page. */
1657 upcall_unixctl_disable_megaflows(struct unixctl_conn *conn,
1658 int argc OVS_UNUSED,
1659 const char *argv[] OVS_UNUSED,
1660 void *aux OVS_UNUSED)
1662 atomic_store(&enable_megaflows, false);
1664 unixctl_command_reply(conn, "megaflows disabled");
1667 /* Re-enable using megaflows.
1669 * This command is only needed for advanced debugging, so it's not
1670 * documented in the man page. */
1672 upcall_unixctl_enable_megaflows(struct unixctl_conn *conn,
1673 int argc OVS_UNUSED,
1674 const char *argv[] OVS_UNUSED,
1675 void *aux OVS_UNUSED)
1677 atomic_store(&enable_megaflows, true);
1679 unixctl_command_reply(conn, "megaflows enabled");
1682 /* Set the flow limit.
1684 * This command is only needed for advanced debugging, so it's not
1685 * documented in the man page. */
1687 upcall_unixctl_set_flow_limit(struct unixctl_conn *conn,
1688 int argc OVS_UNUSED,
1689 const char *argv[] OVS_UNUSED,
1690 void *aux OVS_UNUSED)
1692 struct ds ds = DS_EMPTY_INITIALIZER;
1693 struct udpif *udpif;
1694 unsigned int flow_limit = atoi(argv[1]);
1696 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1697 atomic_store(&udpif->flow_limit, flow_limit);
1699 ds_put_format(&ds, "set flow_limit to %u\n", flow_limit);
1700 unixctl_command_reply(conn, ds_cstr(&ds));