1 /* Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
3 * Licensed under the Apache License, Version 2.0 (the "License");
4 * you may not use this file except in compliance with the License.
5 * You may obtain a copy of the License at:
7 * http://www.apache.org/licenses/LICENSE-2.0
9 * Unless required by applicable law or agreed to in writing, software
10 * distributed under the License is distributed on an "AS IS" BASIS,
11 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 * See the License for the specific language governing permissions and
13 * limitations under the License. */
16 #include "ofproto-dpif-upcall.h"
25 #include "dynamic-string.h"
26 #include "fail-open.h"
27 #include "guarded-list.h"
32 #include "ofproto-dpif-ipfix.h"
33 #include "ofproto-dpif-sflow.h"
34 #include "ofproto-dpif-xlate.h"
37 #include "poll-loop.h"
42 #define MAX_QUEUE_LENGTH 512
43 #define FLOW_MISS_MAX_BATCH 50
44 #define REVALIDATE_MAX_BATCH 50
46 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall);
48 /* A thread that reads upcalls from dpif, forwards each upcall's packet,
49 * and possibly sets up a kernel flow as a cache. */
51 struct udpif *udpif; /* Parent udpif. */
52 pthread_t thread; /* Thread ID. */
53 char *name; /* Thread name. */
54 uint32_t handler_id; /* Handler id. */
57 /* A thread that processes each kernel flow handed to it by the flow_dumper
58 * thread, updates OpenFlow statistics, and updates or removes the kernel flow
61 struct udpif *udpif; /* Parent udpif. */
62 char *name; /* Thread name. */
64 pthread_t thread; /* Thread ID. */
65 struct hmap ukeys; /* Datapath flow keys. */
69 struct ovs_mutex mutex; /* Mutex guarding the following. */
70 pthread_cond_t wake_cond;
71 struct list udumps OVS_GUARDED; /* Unprocessed udumps. */
72 size_t n_udumps OVS_GUARDED; /* Number of unprocessed udumps. */
75 /* An upcall handler for ofproto_dpif.
77 * udpif keeps records of two kind of logically separate units:
82 * - An array of 'struct handler's for upcall handling and flow
88 * - An array of 'struct revalidator's for flow revalidation and
91 * - A "flow_dumper" thread that reads the kernel flow table and dispatches
92 * flows to one of several "revalidator" threads (see struct
96 struct list list_node; /* In all_udpifs list. */
98 struct dpif *dpif; /* Datapath handle. */
99 struct dpif_backer *backer; /* Opaque dpif_backer pointer. */
101 uint32_t secret; /* Random seed for upcall hash. */
103 pthread_t flow_dumper; /* Flow dumper thread ID. */
105 struct handler *handlers; /* Upcall handlers. */
108 struct revalidator *revalidators; /* Flow revalidators. */
109 size_t n_revalidators;
111 uint64_t last_reval_seq; /* 'reval_seq' at last revalidation. */
112 struct seq *reval_seq; /* Incremented to force revalidation. */
114 struct seq *dump_seq; /* Increments each dump iteration. */
116 struct latch exit_latch; /* Tells child threads to exit. */
118 long long int dump_duration; /* Duration of the last flow dump. */
120 /* Datapath flow statistics. */
121 unsigned int max_n_flows;
122 unsigned int avg_n_flows;
124 /* Following fields are accessed and modified by different threads. */
125 atomic_uint flow_limit; /* Datapath flow hard limit. */
127 /* n_flows_mutex prevents multiple threads updating these concurrently. */
128 atomic_uint64_t n_flows; /* Number of flows in the datapath. */
129 atomic_llong n_flows_timestamp; /* Last time n_flows was updated. */
130 struct ovs_mutex n_flows_mutex;
134 BAD_UPCALL, /* Some kind of bug somewhere. */
135 MISS_UPCALL, /* A flow miss. */
136 SFLOW_UPCALL, /* sFlow sample. */
137 FLOW_SAMPLE_UPCALL, /* Per-flow sampling. */
138 IPFIX_UPCALL /* Per-bridge sampling. */
142 struct flow_miss *flow_miss; /* This upcall's flow_miss. */
144 /* Raw upcall plus data for keeping track of the memory backing it. */
145 struct dpif_upcall dpif_upcall; /* As returned by dpif_recv() */
146 struct ofpbuf upcall_buf; /* Owns some data in 'dpif_upcall'. */
147 uint64_t upcall_stub[512 / 8]; /* Buffer to reduce need for malloc(). */
150 /* 'udpif_key's are responsible for tracking the little bit of state udpif
151 * needs to do flow expiration which can't be pulled directly from the
152 * datapath. They are owned, created by, maintained, and destroyed by a single
153 * revalidator making them easy to efficiently handle with multiple threads. */
155 struct hmap_node hmap_node; /* In parent revalidator 'ukeys' map. */
157 struct nlattr *key; /* Datapath flow key. */
158 size_t key_len; /* Length of 'key'. */
160 struct dpif_flow_stats stats; /* Stats at most recent flow dump. */
161 long long int created; /* Estimation of creation time. */
163 bool mark; /* Used by mark and sweep GC algorithm. */
165 struct odputil_keybuf key_buf; /* Memory for 'key'. */
166 struct xlate_cache *xcache; /* Cache for xlate entries that
167 * are affected by this ukey.
168 * Used for stats and learning.*/
171 /* 'udpif_flow_dump's hold the state associated with one iteration in a flow
172 * dump operation. This is created by the flow_dumper thread and handed to the
173 * appropriate revalidator thread to be processed. */
174 struct udpif_flow_dump {
175 struct list list_node;
177 struct nlattr *key; /* Datapath flow key. */
178 size_t key_len; /* Length of 'key'. */
179 uint32_t key_hash; /* Hash of 'key'. */
181 struct odputil_keybuf mask_buf;
182 struct nlattr *mask; /* Datapath mask for 'key'. */
183 size_t mask_len; /* Length of 'mask'. */
185 struct dpif_flow_stats stats; /* Stats pulled from the datapath. */
187 bool need_revalidate; /* Key needs revalidation? */
189 struct odputil_keybuf key_buf;
192 /* Flow miss batching.
194 * Some dpifs implement operations faster when you hand them off in a batch.
195 * To allow batching, "struct flow_miss" queues the dpif-related work needed
196 * for a given flow. Each "struct flow_miss" corresponds to sending one or
197 * more packets, plus possibly installing the flow in the dpif. */
199 struct hmap_node hmap_node;
200 struct ofproto_dpif *ofproto;
203 const struct nlattr *key;
205 enum dpif_upcall_type upcall_type;
206 struct dpif_flow_stats stats;
207 odp_port_t odp_in_port;
209 uint64_t slow_path_buf[128 / 8];
210 struct odputil_keybuf mask_buf;
212 struct xlate_out xout;
217 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
218 static struct list all_udpifs = LIST_INITIALIZER(&all_udpifs);
220 static size_t read_upcalls(struct handler *,
221 struct upcall upcalls[FLOW_MISS_MAX_BATCH],
222 struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH],
224 static void handle_upcalls(struct handler *, struct hmap *, struct upcall *,
226 static void udpif_stop_threads(struct udpif *);
227 static void udpif_start_threads(struct udpif *, size_t n_handlers,
228 size_t n_revalidators);
229 static void *udpif_flow_dumper(void *);
230 static void *udpif_upcall_handler(void *);
231 static void *udpif_revalidator(void *);
232 static uint64_t udpif_get_n_flows(struct udpif *);
233 static void revalidate_udumps(struct revalidator *, struct list *udumps);
234 static void revalidator_sweep(struct revalidator *);
235 static void revalidator_purge(struct revalidator *);
236 static void upcall_unixctl_show(struct unixctl_conn *conn, int argc,
237 const char *argv[], void *aux);
238 static void upcall_unixctl_disable_megaflows(struct unixctl_conn *, int argc,
239 const char *argv[], void *aux);
240 static void upcall_unixctl_enable_megaflows(struct unixctl_conn *, int argc,
241 const char *argv[], void *aux);
242 static void upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, int argc,
243 const char *argv[], void *aux);
244 static void ukey_delete(struct revalidator *, struct udpif_key *);
246 static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true);
249 udpif_create(struct dpif_backer *backer, struct dpif *dpif)
251 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
252 struct udpif *udpif = xzalloc(sizeof *udpif);
254 if (ovsthread_once_start(&once)) {
255 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show,
257 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
258 upcall_unixctl_disable_megaflows, NULL);
259 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
260 upcall_unixctl_enable_megaflows, NULL);
261 unixctl_command_register("upcall/set-flow-limit", "", 1, 1,
262 upcall_unixctl_set_flow_limit, NULL);
263 ovsthread_once_done(&once);
267 udpif->backer = backer;
268 atomic_init(&udpif->flow_limit, MIN(ofproto_flow_limit, 10000));
269 udpif->secret = random_uint32();
270 udpif->reval_seq = seq_create();
271 udpif->dump_seq = seq_create();
272 latch_init(&udpif->exit_latch);
273 list_push_back(&all_udpifs, &udpif->list_node);
274 atomic_init(&udpif->n_flows, 0);
275 atomic_init(&udpif->n_flows_timestamp, LLONG_MIN);
276 ovs_mutex_init(&udpif->n_flows_mutex);
282 udpif_destroy(struct udpif *udpif)
284 udpif_stop_threads(udpif);
286 list_remove(&udpif->list_node);
287 latch_destroy(&udpif->exit_latch);
288 seq_destroy(udpif->reval_seq);
289 seq_destroy(udpif->dump_seq);
290 ovs_mutex_destroy(&udpif->n_flows_mutex);
294 /* Stops the handler and revalidator threads, must be enclosed in
295 * ovsrcu quiescent state unless when destroying udpif. */
297 udpif_stop_threads(struct udpif *udpif)
299 if (udpif && (udpif->n_handlers != 0 || udpif->n_revalidators != 0)) {
302 latch_set(&udpif->exit_latch);
304 for (i = 0; i < udpif->n_handlers; i++) {
305 struct handler *handler = &udpif->handlers[i];
307 xpthread_join(handler->thread, NULL);
310 for (i = 0; i < udpif->n_revalidators; i++) {
311 struct revalidator *revalidator = &udpif->revalidators[i];
313 ovs_mutex_lock(&revalidator->mutex);
314 xpthread_cond_signal(&revalidator->wake_cond);
315 ovs_mutex_unlock(&revalidator->mutex);
316 xpthread_join(revalidator->thread, NULL);
319 xpthread_join(udpif->flow_dumper, NULL);
321 for (i = 0; i < udpif->n_revalidators; i++) {
322 struct revalidator *revalidator = &udpif->revalidators[i];
323 struct udpif_flow_dump *udump, *next_udump;
325 LIST_FOR_EACH_SAFE (udump, next_udump, list_node,
326 &revalidator->udumps) {
327 list_remove(&udump->list_node);
331 /* Delete ukeys, and delete all flows from the datapath to prevent
332 * double-counting stats. */
333 revalidator_purge(revalidator);
334 hmap_destroy(&revalidator->ukeys);
335 ovs_mutex_destroy(&revalidator->mutex);
337 free(revalidator->name);
340 for (i = 0; i < udpif->n_handlers; i++) {
341 free(udpif->handlers[i].name);
343 latch_poll(&udpif->exit_latch);
345 free(udpif->revalidators);
346 udpif->revalidators = NULL;
347 udpif->n_revalidators = 0;
349 free(udpif->handlers);
350 udpif->handlers = NULL;
351 udpif->n_handlers = 0;
355 /* Starts the handler and revalidator threads, must be enclosed in
356 * ovsrcu quiescent state. */
358 udpif_start_threads(struct udpif *udpif, size_t n_handlers,
359 size_t n_revalidators)
361 if (udpif && (!udpif->handlers && !udpif->revalidators)) {
364 udpif->n_handlers = n_handlers;
365 udpif->n_revalidators = n_revalidators;
367 udpif->handlers = xzalloc(udpif->n_handlers * sizeof *udpif->handlers);
368 for (i = 0; i < udpif->n_handlers; i++) {
369 struct handler *handler = &udpif->handlers[i];
371 handler->udpif = udpif;
372 handler->handler_id = i;
373 xpthread_create(&handler->thread, NULL, udpif_upcall_handler,
377 udpif->revalidators = xzalloc(udpif->n_revalidators
378 * sizeof *udpif->revalidators);
379 for (i = 0; i < udpif->n_revalidators; i++) {
380 struct revalidator *revalidator = &udpif->revalidators[i];
382 revalidator->udpif = udpif;
383 list_init(&revalidator->udumps);
384 hmap_init(&revalidator->ukeys);
385 ovs_mutex_init(&revalidator->mutex);
386 xpthread_cond_init(&revalidator->wake_cond, NULL);
387 xpthread_create(&revalidator->thread, NULL, udpif_revalidator,
390 xpthread_create(&udpif->flow_dumper, NULL, udpif_flow_dumper, udpif);
394 /* Tells 'udpif' how many threads it should use to handle upcalls.
395 * 'n_handlers' and 'n_revalidators' can never be zero. 'udpif''s
396 * datapath handle must have packet reception enabled before starting
399 udpif_set_threads(struct udpif *udpif, size_t n_handlers,
400 size_t n_revalidators)
405 ovs_assert(n_handlers && n_revalidators);
407 ovsrcu_quiesce_start();
408 if (udpif->n_handlers != n_handlers
409 || udpif->n_revalidators != n_revalidators) {
410 udpif_stop_threads(udpif);
413 error = dpif_handlers_set(udpif->dpif, n_handlers);
415 VLOG_ERR("failed to configure handlers in dpif %s: %s",
416 dpif_name(udpif->dpif), ovs_strerror(error));
420 if (!udpif->handlers && !udpif->revalidators) {
421 udpif_start_threads(udpif, n_handlers, n_revalidators);
423 ovsrcu_quiesce_end();
426 /* Waits for all ongoing upcall translations to complete. This ensures that
427 * there are no transient references to any removed ofprotos (or other
428 * objects). In particular, this should be called after an ofproto is removed
429 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
431 udpif_synchronize(struct udpif *udpif)
433 /* This is stronger than necessary. It would be sufficient to ensure
434 * (somehow) that each handler and revalidator thread had passed through
435 * its main loop once. */
436 size_t n_handlers = udpif->n_handlers;
437 size_t n_revalidators = udpif->n_revalidators;
439 ovsrcu_quiesce_start();
440 udpif_stop_threads(udpif);
441 udpif_start_threads(udpif, n_handlers, n_revalidators);
442 ovsrcu_quiesce_end();
445 /* Notifies 'udpif' that something changed which may render previous
446 * xlate_actions() results invalid. */
448 udpif_revalidate(struct udpif *udpif)
450 seq_change(udpif->reval_seq);
453 /* Returns a seq which increments every time 'udpif' pulls stats from the
454 * datapath. Callers can use this to get a sense of when might be a good time
455 * to do periodic work which relies on relatively up to date statistics. */
457 udpif_dump_seq(struct udpif *udpif)
459 return udpif->dump_seq;
463 udpif_get_memory_usage(struct udpif *udpif, struct simap *usage)
467 simap_increase(usage, "flow_dumpers", 1);
469 simap_increase(usage, "handlers", udpif->n_handlers);
471 simap_increase(usage, "revalidators", udpif->n_revalidators);
472 for (i = 0; i < udpif->n_revalidators; i++) {
473 struct revalidator *revalidator = &udpif->revalidators[i];
474 ovs_mutex_lock(&revalidator->mutex);
475 simap_increase(usage, "revalidator dumps", revalidator->n_udumps);
477 /* XXX: This isn't technically thread safe because the revalidator
478 * ukeys maps isn't protected by a mutex since it's per thread. */
479 simap_increase(usage, "revalidator keys",
480 hmap_count(&revalidator->ukeys));
481 ovs_mutex_unlock(&revalidator->mutex);
485 /* Remove flows from a single datapath. */
487 udpif_flush(struct udpif *udpif)
489 size_t n_handlers, n_revalidators;
491 n_handlers = udpif->n_handlers;
492 n_revalidators = udpif->n_revalidators;
494 ovsrcu_quiesce_start();
496 udpif_stop_threads(udpif);
497 dpif_flow_flush(udpif->dpif);
498 udpif_start_threads(udpif, n_handlers, n_revalidators);
500 ovsrcu_quiesce_end();
503 /* Removes all flows from all datapaths. */
505 udpif_flush_all_datapaths(void)
509 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
516 udpif_get_n_flows(struct udpif *udpif)
518 long long int time, now;
522 atomic_read(&udpif->n_flows_timestamp, &time);
523 if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) {
524 struct dpif_dp_stats stats;
526 atomic_store(&udpif->n_flows_timestamp, now);
527 dpif_get_dp_stats(udpif->dpif, &stats);
528 flow_count = stats.n_flows;
529 atomic_store(&udpif->n_flows, flow_count);
530 ovs_mutex_unlock(&udpif->n_flows_mutex);
532 atomic_read(&udpif->n_flows, &flow_count);
538 udpif_flow_dumper(void *arg)
540 struct udpif *udpif = arg;
542 set_subprogram_name("flow_dumper");
543 while (!latch_is_set(&udpif->exit_latch)) {
544 const struct dpif_flow_stats *stats;
545 long long int start_time, duration;
546 const struct nlattr *key, *mask;
547 struct dpif_flow_dump dump;
548 size_t key_len, mask_len;
549 unsigned int flow_limit;
550 bool need_revalidate;
556 reval_seq = seq_read(udpif->reval_seq);
557 need_revalidate = udpif->last_reval_seq != reval_seq;
558 udpif->last_reval_seq = reval_seq;
560 n_flows = udpif_get_n_flows(udpif);
561 udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows);
562 udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2;
564 start_time = time_msec();
565 error = dpif_flow_dump_start(&dump, udpif->dpif);
567 VLOG_INFO("Failed to start flow dump (%s)", ovs_strerror(error));
570 dpif_flow_dump_state_init(udpif->dpif, &state);
571 while (dpif_flow_dump_next(&dump, state, &key, &key_len,
572 &mask, &mask_len, NULL, NULL, &stats)
573 && !latch_is_set(&udpif->exit_latch)) {
574 struct udpif_flow_dump *udump = xmalloc(sizeof *udump);
575 struct revalidator *revalidator;
577 udump->key_hash = hash_bytes(key, key_len, udpif->secret);
578 memcpy(&udump->key_buf, key, key_len);
579 udump->key = (struct nlattr *) &udump->key_buf;
580 udump->key_len = key_len;
582 memcpy(&udump->mask_buf, mask, mask_len);
583 udump->mask = (struct nlattr *) &udump->mask_buf;
584 udump->mask_len = mask_len;
586 udump->stats = *stats;
587 udump->need_revalidate = need_revalidate;
589 revalidator = &udpif->revalidators[udump->key_hash
590 % udpif->n_revalidators];
592 ovs_mutex_lock(&revalidator->mutex);
593 while (revalidator->n_udumps >= REVALIDATE_MAX_BATCH * 3
594 && !latch_is_set(&udpif->exit_latch)) {
595 ovs_mutex_cond_wait(&revalidator->wake_cond,
596 &revalidator->mutex);
598 list_push_back(&revalidator->udumps, &udump->list_node);
599 revalidator->n_udumps++;
600 xpthread_cond_signal(&revalidator->wake_cond);
601 ovs_mutex_unlock(&revalidator->mutex);
603 dpif_flow_dump_state_uninit(udpif->dpif, state);
604 dpif_flow_dump_done(&dump);
606 /* Let all the revalidators finish and garbage collect. */
607 seq_change(udpif->dump_seq);
608 for (i = 0; i < udpif->n_revalidators; i++) {
609 struct revalidator *revalidator = &udpif->revalidators[i];
610 ovs_mutex_lock(&revalidator->mutex);
611 xpthread_cond_signal(&revalidator->wake_cond);
612 ovs_mutex_unlock(&revalidator->mutex);
615 for (i = 0; i < udpif->n_revalidators; i++) {
616 struct revalidator *revalidator = &udpif->revalidators[i];
618 ovs_mutex_lock(&revalidator->mutex);
619 while (revalidator->dump_seq != seq_read(udpif->dump_seq)
620 && !latch_is_set(&udpif->exit_latch)) {
621 ovs_mutex_cond_wait(&revalidator->wake_cond,
622 &revalidator->mutex);
624 ovs_mutex_unlock(&revalidator->mutex);
627 duration = MAX(time_msec() - start_time, 1);
628 udpif->dump_duration = duration;
629 atomic_read(&udpif->flow_limit, &flow_limit);
630 if (duration > 2000) {
631 flow_limit /= duration / 1000;
632 } else if (duration > 1300) {
633 flow_limit = flow_limit * 3 / 4;
634 } else if (duration < 1000 && n_flows > 2000
635 && flow_limit < n_flows * 1000 / duration) {
638 flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000));
639 atomic_store(&udpif->flow_limit, flow_limit);
641 if (duration > 2000) {
642 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
647 poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500));
648 seq_wait(udpif->reval_seq, udpif->last_reval_seq);
649 latch_wait(&udpif->exit_latch);
656 /* The upcall handler thread tries to read a batch of FLOW_MISS_MAX_BATCH
657 * upcalls from dpif, processes the batch and installs corresponding flows
660 udpif_upcall_handler(void *arg)
662 struct handler *handler = arg;
663 struct udpif *udpif = handler->udpif;
664 struct hmap misses = HMAP_INITIALIZER(&misses);
666 handler->name = xasprintf("handler_%u", ovsthread_id_self());
667 set_subprogram_name("%s", handler->name);
669 while (!latch_is_set(&handler->udpif->exit_latch)) {
670 struct upcall upcalls[FLOW_MISS_MAX_BATCH];
671 struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH];
672 struct flow_miss *miss;
675 n_upcalls = read_upcalls(handler, upcalls, miss_buf, &misses);
677 dpif_recv_wait(udpif->dpif, handler->handler_id);
678 latch_wait(&udpif->exit_latch);
681 handle_upcalls(handler, &misses, upcalls, n_upcalls);
683 HMAP_FOR_EACH (miss, hmap_node, &misses) {
684 xlate_out_uninit(&miss->xout);
687 for (i = 0; i < n_upcalls; i++) {
688 ofpbuf_uninit(&upcalls[i].dpif_upcall.packet);
689 ofpbuf_uninit(&upcalls[i].upcall_buf);
694 hmap_destroy(&misses);
700 udpif_revalidator(void *arg)
702 struct revalidator *revalidator = arg;
704 revalidator->name = xasprintf("revalidator_%u", ovsthread_id_self());
705 set_subprogram_name("%s", revalidator->name);
707 struct list udumps = LIST_INITIALIZER(&udumps);
708 struct udpif *udpif = revalidator->udpif;
711 ovs_mutex_lock(&revalidator->mutex);
712 if (latch_is_set(&udpif->exit_latch)) {
713 ovs_mutex_unlock(&revalidator->mutex);
717 if (!revalidator->n_udumps) {
718 if (revalidator->dump_seq != seq_read(udpif->dump_seq)) {
719 revalidator->dump_seq = seq_read(udpif->dump_seq);
720 revalidator_sweep(revalidator);
722 ovs_mutex_cond_wait(&revalidator->wake_cond,
723 &revalidator->mutex);
727 for (i = 0; i < REVALIDATE_MAX_BATCH && revalidator->n_udumps; i++) {
728 list_push_back(&udumps, list_pop_front(&revalidator->udumps));
729 revalidator->n_udumps--;
732 /* Wake up the flow dumper. */
733 xpthread_cond_signal(&revalidator->wake_cond);
734 ovs_mutex_unlock(&revalidator->mutex);
736 if (!list_is_empty(&udumps)) {
737 revalidate_udumps(revalidator, &udumps);
744 static enum upcall_type
745 classify_upcall(const struct upcall *upcall)
747 const struct dpif_upcall *dpif_upcall = &upcall->dpif_upcall;
748 union user_action_cookie cookie;
751 /* First look at the upcall type. */
752 switch (dpif_upcall->type) {
759 case DPIF_N_UC_TYPES:
761 VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32,
766 /* "action" upcalls need a closer look. */
767 if (!dpif_upcall->userdata) {
768 VLOG_WARN_RL(&rl, "action upcall missing cookie");
771 userdata_len = nl_attr_get_size(dpif_upcall->userdata);
772 if (userdata_len < sizeof cookie.type
773 || userdata_len > sizeof cookie) {
774 VLOG_WARN_RL(&rl, "action upcall cookie has unexpected size %"PRIuSIZE,
778 memset(&cookie, 0, sizeof cookie);
779 memcpy(&cookie, nl_attr_get(dpif_upcall->userdata), userdata_len);
780 if (userdata_len == MAX(8, sizeof cookie.sflow)
781 && cookie.type == USER_ACTION_COOKIE_SFLOW) {
783 } else if (userdata_len == MAX(8, sizeof cookie.slow_path)
784 && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) {
786 } else if (userdata_len == MAX(8, sizeof cookie.flow_sample)
787 && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) {
788 return FLOW_SAMPLE_UPCALL;
789 } else if (userdata_len == MAX(8, sizeof cookie.ipfix)
790 && cookie.type == USER_ACTION_COOKIE_IPFIX) {
793 VLOG_WARN_RL(&rl, "invalid user cookie of type %"PRIu16
794 " and size %"PRIuSIZE, cookie.type, userdata_len);
799 /* Calculates slow path actions for 'xout'. 'buf' must statically be
800 * initialized with at least 128 bytes of space. */
802 compose_slow_path(struct udpif *udpif, struct xlate_out *xout,
803 struct flow *flow, odp_port_t odp_in_port,
806 union user_action_cookie cookie;
810 cookie.type = USER_ACTION_COOKIE_SLOW_PATH;
811 cookie.slow_path.unused = 0;
812 cookie.slow_path.reason = xout->slow;
814 port = xout->slow & (SLOW_CFM | SLOW_BFD | SLOW_LACP | SLOW_STP)
817 pid = dpif_port_get_pid(udpif->dpif, port, flow_hash_5tuple(flow, 0));
818 odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, buf);
821 static struct flow_miss *
822 flow_miss_find(struct hmap *todo, const struct ofproto_dpif *ofproto,
823 const struct flow *flow, uint32_t hash)
825 struct flow_miss *miss;
827 HMAP_FOR_EACH_WITH_HASH (miss, hmap_node, hash, todo) {
828 if (miss->ofproto == ofproto && flow_equal(&miss->flow, flow)) {
836 /* Reads and classifies upcalls. Returns the number of upcalls successfully
839 read_upcalls(struct handler *handler,
840 struct upcall upcalls[FLOW_MISS_MAX_BATCH],
841 struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH],
844 struct udpif *udpif = handler->udpif;
847 size_t n_upcalls = 0;
850 * Try reading FLOW_MISS_MAX_BATCH upcalls from dpif.
852 * Extract the flow from each upcall. Construct in 'misses' a hash table
853 * that maps each unique flow to a 'struct flow_miss'.
855 * Most commonly there is a single packet per flow_miss, but there are
856 * several reasons why there might be more than one, e.g.:
858 * - The dpif packet interface does not support TSO (or UFO, etc.), so a
859 * large packet sent to userspace is split into a sequence of smaller
862 * - A stream of quickly arriving packets in an established "slow-pathed"
865 * - Rarely, a stream of quickly arriving packets in a flow not yet
866 * established. (This is rare because most protocols do not send
867 * multiple back-to-back packets before receiving a reply from the
868 * other end of the connection, which gives OVS a chance to set up a
871 for (i = 0; i < FLOW_MISS_MAX_BATCH; i++) {
872 struct upcall *upcall = &upcalls[n_upcalls];
873 struct flow_miss *miss = &miss_buf[n_misses];
874 struct dpif_upcall *dupcall;
875 struct ofpbuf *packet;
876 struct flow_miss *existing_miss;
877 struct ofproto_dpif *ofproto;
878 struct dpif_sflow *sflow;
879 struct dpif_ipfix *ipfix;
881 enum upcall_type type;
882 odp_port_t odp_in_port;
885 ofpbuf_use_stub(&upcall->upcall_buf, upcall->upcall_stub,
886 sizeof upcall->upcall_stub);
887 error = dpif_recv(udpif->dpif, handler->handler_id,
888 &upcall->dpif_upcall, &upcall->upcall_buf);
890 ofpbuf_uninit(&upcall->upcall_buf);
894 dupcall = &upcall->dpif_upcall;
895 packet = &dupcall->packet;
896 error = xlate_receive(udpif->backer, packet, dupcall->key,
897 dupcall->key_len, &flow,
898 &ofproto, &ipfix, &sflow, NULL, &odp_in_port);
900 if (error == ENODEV) {
901 /* Received packet on datapath port for which we couldn't
902 * associate an ofproto. This can happen if a port is removed
903 * while traffic is being received. Print a rate-limited
904 * message in case it happens frequently. Install a drop flow
905 * so that future packets of the flow are inexpensively dropped
907 VLOG_INFO_RL(&rl, "received packet on unassociated datapath "
908 "port %"PRIu32, odp_in_port);
909 dpif_flow_put(udpif->dpif, DPIF_FP_CREATE | DPIF_FP_MODIFY,
910 dupcall->key, dupcall->key_len, NULL, 0, NULL, 0,
916 type = classify_upcall(upcall);
917 if (type == MISS_UPCALL) {
919 struct pkt_metadata md = pkt_metadata_from_flow(&flow);
921 flow_extract(packet, &md, &miss->flow);
922 hash = flow_hash(&miss->flow, 0);
923 existing_miss = flow_miss_find(misses, ofproto, &miss->flow,
925 if (!existing_miss) {
926 hmap_insert(misses, &miss->hmap_node, hash);
927 miss->ofproto = ofproto;
928 miss->key = dupcall->key;
929 miss->key_len = dupcall->key_len;
930 miss->upcall_type = dupcall->type;
931 miss->stats.n_packets = 0;
932 miss->stats.n_bytes = 0;
933 miss->stats.used = time_msec();
934 miss->stats.tcp_flags = 0;
935 miss->odp_in_port = odp_in_port;
939 miss = existing_miss;
941 miss->stats.tcp_flags |= ntohs(miss->flow.tcp_flags);
942 miss->stats.n_bytes += ofpbuf_size(packet);
943 miss->stats.n_packets++;
945 upcall->flow_miss = miss;
953 union user_action_cookie cookie;
955 memset(&cookie, 0, sizeof cookie);
956 memcpy(&cookie, nl_attr_get(dupcall->userdata),
957 sizeof cookie.sflow);
958 dpif_sflow_received(sflow, packet, &flow, odp_in_port,
964 dpif_ipfix_bridge_sample(ipfix, packet, &flow);
967 case FLOW_SAMPLE_UPCALL:
969 union user_action_cookie cookie;
971 memset(&cookie, 0, sizeof cookie);
972 memcpy(&cookie, nl_attr_get(dupcall->userdata),
973 sizeof cookie.flow_sample);
975 /* The flow reflects exactly the contents of the packet.
976 * Sample the packet using it. */
977 dpif_ipfix_flow_sample(ipfix, packet, &flow,
978 cookie.flow_sample.collector_set_id,
979 cookie.flow_sample.probability,
980 cookie.flow_sample.obs_domain_id,
981 cookie.flow_sample.obs_point_id);
990 dpif_ipfix_unref(ipfix);
991 dpif_sflow_unref(sflow);
994 ofpbuf_uninit(&upcall->dpif_upcall.packet);
995 ofpbuf_uninit(&upcall->upcall_buf);
1002 handle_upcalls(struct handler *handler, struct hmap *misses,
1003 struct upcall *upcalls, size_t n_upcalls)
1005 struct udpif *udpif = handler->udpif;
1006 struct dpif_op *opsp[FLOW_MISS_MAX_BATCH * 2];
1007 struct dpif_op ops[FLOW_MISS_MAX_BATCH * 2];
1008 struct flow_miss *miss;
1010 unsigned int flow_limit;
1011 bool fail_open, may_put;
1013 atomic_read(&udpif->flow_limit, &flow_limit);
1014 may_put = udpif_get_n_flows(udpif) < flow_limit;
1016 /* Initialize each 'struct flow_miss's ->xout.
1018 * We do this per-flow_miss rather than per-packet because, most commonly,
1019 * all the packets in a flow can use the same translation.
1021 * We can't do this in the previous loop because we need the TCP flags for
1022 * all the packets in each miss. */
1024 HMAP_FOR_EACH (miss, hmap_node, misses) {
1025 struct xlate_in xin;
1027 xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL,
1028 miss->stats.tcp_flags, NULL);
1029 xin.may_learn = true;
1031 if (miss->upcall_type == DPIF_UC_MISS) {
1032 xin.resubmit_stats = &miss->stats;
1034 /* For non-miss upcalls, there's a flow in the datapath which this
1035 * packet was accounted to. Presumably the revalidators will deal
1036 * with pushing its stats eventually. */
1039 xlate_actions(&xin, &miss->xout);
1040 fail_open = fail_open || miss->xout.fail_open;
1043 /* Now handle the packets individually in order of arrival. In the common
1044 * case each packet of a miss can share the same actions, but slow-pathed
1045 * packets need to be translated individually:
1047 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1048 * processes received packets for these protocols.
1050 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1053 * The loop fills 'ops' with an array of operations to execute in the
1056 for (i = 0; i < n_upcalls; i++) {
1057 struct upcall *upcall = &upcalls[i];
1058 struct flow_miss *miss = upcall->flow_miss;
1059 struct ofpbuf *packet = &upcall->dpif_upcall.packet;
1061 ovs_be16 flow_vlan_tci;
1063 /* Save a copy of flow.vlan_tci in case it is changed to
1064 * generate proper mega flow masks for VLAN splinter flows. */
1065 flow_vlan_tci = miss->flow.vlan_tci;
1067 if (miss->xout.slow) {
1068 struct xlate_in xin;
1070 xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL, 0, packet);
1071 xlate_actions_for_side_effects(&xin);
1074 if (miss->flow.in_port.ofp_port
1075 != vsp_realdev_to_vlandev(miss->ofproto,
1076 miss->flow.in_port.ofp_port,
1077 miss->flow.vlan_tci)) {
1078 /* This packet was received on a VLAN splinter port. We
1079 * added a VLAN to the packet to make the packet resemble
1080 * the flow, but the actions were composed assuming that
1081 * the packet contained no VLAN. So, we must remove the
1082 * VLAN header from the packet before trying to execute the
1084 if (ofpbuf_size(&miss->xout.odp_actions)) {
1085 eth_pop_vlan(packet);
1088 /* Remove the flow vlan tags inserted by vlan splinter logic
1089 * to ensure megaflow masks generated match the data path flow. */
1090 miss->flow.vlan_tci = 0;
1093 /* Do not install a flow into the datapath if:
1095 * - The datapath already has too many flows.
1097 * - An earlier iteration of this loop already put the same flow.
1099 * - We received this packet via some flow installed in the kernel
1103 && upcall->dpif_upcall.type == DPIF_UC_MISS) {
1109 atomic_read(&enable_megaflows, &megaflow);
1110 ofpbuf_use_stack(&mask, &miss->mask_buf, sizeof miss->mask_buf);
1114 max_mpls = ofproto_dpif_get_max_mpls_depth(miss->ofproto);
1115 odp_flow_key_from_mask(&mask, &miss->xout.wc.masks,
1116 &miss->flow, UINT32_MAX, max_mpls);
1120 op->type = DPIF_OP_FLOW_PUT;
1121 op->u.flow_put.flags = DPIF_FP_CREATE | DPIF_FP_MODIFY;
1122 op->u.flow_put.key = miss->key;
1123 op->u.flow_put.key_len = miss->key_len;
1124 op->u.flow_put.mask = ofpbuf_data(&mask);
1125 op->u.flow_put.mask_len = ofpbuf_size(&mask);
1126 op->u.flow_put.stats = NULL;
1128 if (!miss->xout.slow) {
1129 op->u.flow_put.actions = ofpbuf_data(&miss->xout.odp_actions);
1130 op->u.flow_put.actions_len = ofpbuf_size(&miss->xout.odp_actions);
1134 ofpbuf_use_stack(&buf, miss->slow_path_buf,
1135 sizeof miss->slow_path_buf);
1136 compose_slow_path(udpif, &miss->xout, &miss->flow,
1137 miss->odp_in_port, &buf);
1138 op->u.flow_put.actions = ofpbuf_data(&buf);
1139 op->u.flow_put.actions_len = ofpbuf_size(&buf);
1144 * The 'miss' may be shared by multiple upcalls. Restore
1145 * the saved flow vlan_tci field before processing the next
1147 miss->flow.vlan_tci = flow_vlan_tci;
1149 if (ofpbuf_size(&miss->xout.odp_actions)) {
1152 op->type = DPIF_OP_EXECUTE;
1153 op->u.execute.packet = packet;
1154 odp_key_to_pkt_metadata(miss->key, miss->key_len,
1156 op->u.execute.actions = ofpbuf_data(&miss->xout.odp_actions);
1157 op->u.execute.actions_len = ofpbuf_size(&miss->xout.odp_actions);
1158 op->u.execute.needs_help = (miss->xout.slow & SLOW_ACTION) != 0;
1162 /* Special case for fail-open mode.
1164 * If we are in fail-open mode, but we are connected to a controller too,
1165 * then we should send the packet up to the controller in the hope that it
1166 * will try to set up a flow and thereby allow us to exit fail-open.
1168 * See the top-level comment in fail-open.c for more information.
1170 * Copy packets before they are modified by execution. */
1172 for (i = 0; i < n_upcalls; i++) {
1173 struct upcall *upcall = &upcalls[i];
1174 struct flow_miss *miss = upcall->flow_miss;
1175 struct ofpbuf *packet = &upcall->dpif_upcall.packet;
1176 struct ofproto_packet_in *pin;
1178 pin = xmalloc(sizeof *pin);
1179 pin->up.packet = xmemdup(ofpbuf_data(packet), ofpbuf_size(packet));
1180 pin->up.packet_len = ofpbuf_size(packet);
1181 pin->up.reason = OFPR_NO_MATCH;
1182 pin->up.table_id = 0;
1183 pin->up.cookie = OVS_BE64_MAX;
1184 flow_get_metadata(&miss->flow, &pin->up.fmd);
1185 pin->send_len = 0; /* Not used for flow table misses. */
1186 pin->miss_type = OFPROTO_PACKET_IN_NO_MISS;
1187 ofproto_dpif_send_packet_in(miss->ofproto, pin);
1191 /* Execute batch. */
1192 for (i = 0; i < n_ops; i++) {
1195 dpif_operate(udpif->dpif, opsp, n_ops);
1198 static struct udpif_key *
1199 ukey_lookup(struct revalidator *revalidator, struct udpif_flow_dump *udump)
1201 struct udpif_key *ukey;
1203 HMAP_FOR_EACH_WITH_HASH (ukey, hmap_node, udump->key_hash,
1204 &revalidator->ukeys) {
1205 if (ukey->key_len == udump->key_len
1206 && !memcmp(ukey->key, udump->key, udump->key_len)) {
1213 static struct udpif_key *
1214 ukey_create(const struct nlattr *key, size_t key_len, long long int used)
1216 struct udpif_key *ukey = xmalloc(sizeof *ukey);
1218 ukey->key = (struct nlattr *) &ukey->key_buf;
1219 memcpy(&ukey->key_buf, key, key_len);
1220 ukey->key_len = key_len;
1223 ukey->created = used ? used : time_msec();
1224 memset(&ukey->stats, 0, sizeof ukey->stats);
1225 ukey->xcache = NULL;
1231 ukey_delete(struct revalidator *revalidator, struct udpif_key *ukey)
1233 hmap_remove(&revalidator->ukeys, &ukey->hmap_node);
1234 xlate_cache_delete(ukey->xcache);
1239 should_revalidate(uint64_t packets, long long int used)
1241 long long int metric, now, duration;
1243 /* Calculate the mean time between seeing these packets. If this
1244 * exceeds the threshold, then delete the flow rather than performing
1245 * costly revalidation for flows that aren't being hit frequently.
1247 * This is targeted at situations where the dump_duration is high (~1s),
1248 * and revalidation is triggered by a call to udpif_revalidate(). In
1249 * these situations, revalidation of all flows causes fluctuations in the
1250 * flow_limit due to the interaction with the dump_duration and max_idle.
1251 * This tends to result in deletion of low-throughput flows anyway, so
1252 * skip the revalidation and just delete those flows. */
1253 packets = MAX(packets, 1);
1254 now = MAX(used, time_msec());
1255 duration = now - used;
1256 metric = duration / packets;
1265 revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump,
1266 struct udpif_key *ukey)
1268 struct ofpbuf xout_actions, *actions;
1269 uint64_t slow_path_buf[128 / 8];
1270 struct xlate_out xout, *xoutp;
1271 struct netflow *netflow;
1272 struct flow flow, udump_mask;
1273 struct ofproto_dpif *ofproto;
1274 struct dpif_flow_stats push;
1275 uint32_t *udump32, *xout32;
1276 odp_port_t odp_in_port;
1277 struct xlate_in xin;
1278 long long int last_used;
1288 /* If we don't need to revalidate, we can simply push the stats contained
1289 * in the udump, otherwise we'll have to get the actions so we can check
1291 if (udump->need_revalidate) {
1292 if (dpif_flow_get(udpif->dpif, ukey->key, ukey->key_len, &actions,
1298 last_used = ukey->stats.used;
1299 push.used = udump->stats.used;
1300 push.tcp_flags = udump->stats.tcp_flags;
1301 push.n_packets = udump->stats.n_packets > ukey->stats.n_packets
1302 ? udump->stats.n_packets - ukey->stats.n_packets
1304 push.n_bytes = udump->stats.n_bytes > ukey->stats.n_bytes
1305 ? udump->stats.n_bytes - ukey->stats.n_bytes
1307 ukey->stats = udump->stats;
1309 if (udump->need_revalidate && last_used
1310 && !should_revalidate(push.n_packets, last_used)) {
1315 if (!push.n_packets && !udump->need_revalidate) {
1320 may_learn = push.n_packets > 0;
1321 if (ukey->xcache && !udump->need_revalidate) {
1322 xlate_push_stats(ukey->xcache, may_learn, &push);
1327 error = xlate_receive(udpif->backer, NULL, ukey->key, ukey->key_len, &flow,
1328 &ofproto, NULL, NULL, &netflow, &odp_in_port);
1333 if (udump->need_revalidate) {
1334 xlate_cache_clear(ukey->xcache);
1336 if (!ukey->xcache) {
1337 ukey->xcache = xlate_cache_new();
1340 xlate_in_init(&xin, ofproto, &flow, NULL, push.tcp_flags, NULL);
1341 xin.resubmit_stats = push.n_packets ? &push : NULL;
1342 xin.xcache = ukey->xcache;
1343 xin.may_learn = may_learn;
1344 xin.skip_wildcards = !udump->need_revalidate;
1345 xlate_actions(&xin, &xout);
1348 if (!udump->need_revalidate) {
1354 ofpbuf_use_const(&xout_actions, ofpbuf_data(&xout.odp_actions),
1355 ofpbuf_size(&xout.odp_actions));
1357 ofpbuf_use_stack(&xout_actions, slow_path_buf, sizeof slow_path_buf);
1358 compose_slow_path(udpif, &xout, &flow, odp_in_port, &xout_actions);
1361 if (!ofpbuf_equal(&xout_actions, actions)) {
1365 if (odp_flow_key_to_mask(udump->mask, udump->mask_len, &udump_mask, &flow)
1370 /* Since the kernel is free to ignore wildcarded bits in the mask, we can't
1371 * directly check that the masks are the same. Instead we check that the
1372 * mask in the kernel is more specific i.e. less wildcarded, than what
1373 * we've calculated here. This guarantees we don't catch any packets we
1374 * shouldn't with the megaflow. */
1375 udump32 = (uint32_t *) &udump_mask;
1376 xout32 = (uint32_t *) &xout.wc.masks;
1377 for (i = 0; i < FLOW_U32S; i++) {
1378 if ((udump32[i] | xout32[i]) != udump32[i]) {
1387 netflow_expire(netflow, &flow);
1388 netflow_flow_clear(netflow, &flow);
1390 netflow_unref(netflow);
1392 ofpbuf_delete(actions);
1393 xlate_out_uninit(xoutp);
1398 struct udpif_key *ukey;
1399 struct udpif_flow_dump *udump;
1400 struct dpif_flow_stats stats; /* Stats for 'op'. */
1401 struct dpif_op op; /* Flow del operation. */
1405 dump_op_init(struct dump_op *op, const struct nlattr *key, size_t key_len,
1406 struct udpif_key *ukey, struct udpif_flow_dump *udump)
1410 op->op.type = DPIF_OP_FLOW_DEL;
1411 op->op.u.flow_del.key = key;
1412 op->op.u.flow_del.key_len = key_len;
1413 op->op.u.flow_del.stats = &op->stats;
1417 push_dump_ops(struct revalidator *revalidator,
1418 struct dump_op *ops, size_t n_ops)
1420 struct udpif *udpif = revalidator->udpif;
1421 struct dpif_op *opsp[REVALIDATE_MAX_BATCH];
1424 ovs_assert(n_ops <= REVALIDATE_MAX_BATCH);
1425 for (i = 0; i < n_ops; i++) {
1426 opsp[i] = &ops[i].op;
1428 dpif_operate(udpif->dpif, opsp, n_ops);
1430 for (i = 0; i < n_ops; i++) {
1431 struct dump_op *op = &ops[i];
1432 struct dpif_flow_stats *push, *stats, push_buf;
1434 stats = op->op.u.flow_del.stats;
1437 push->used = MAX(stats->used, op->ukey->stats.used);
1438 push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags;
1439 push->n_packets = stats->n_packets - op->ukey->stats.n_packets;
1440 push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes;
1445 if (push->n_packets || netflow_exists()) {
1446 struct ofproto_dpif *ofproto;
1447 struct netflow *netflow;
1451 may_learn = push->n_packets > 0;
1452 if (op->ukey && op->ukey->xcache) {
1453 xlate_push_stats(op->ukey->xcache, may_learn, push);
1457 if (!xlate_receive(udpif->backer, NULL, op->op.u.flow_del.key,
1458 op->op.u.flow_del.key_len, &flow, &ofproto,
1459 NULL, NULL, &netflow, NULL)) {
1460 struct xlate_in xin;
1462 xlate_in_init(&xin, ofproto, &flow, NULL, push->tcp_flags,
1464 xin.resubmit_stats = push->n_packets ? push : NULL;
1465 xin.may_learn = may_learn;
1466 xin.skip_wildcards = true;
1467 xlate_actions_for_side_effects(&xin);
1470 netflow_expire(netflow, &flow);
1471 netflow_flow_clear(netflow, &flow);
1472 netflow_unref(netflow);
1478 for (i = 0; i < n_ops; i++) {
1479 struct udpif_key *ukey;
1481 /* If there's a udump, this ukey came directly from a datapath flow
1482 * dump. Sometimes a datapath can send duplicates in flow dumps, in
1483 * which case we wouldn't want to double-free a ukey, so avoid that by
1484 * looking up the ukey again.
1486 * If there's no udump then we know what we're doing. */
1487 ukey = (ops[i].udump
1488 ? ukey_lookup(revalidator, ops[i].udump)
1491 ukey_delete(revalidator, ukey);
1497 revalidate_udumps(struct revalidator *revalidator, struct list *udumps)
1499 struct udpif *udpif = revalidator->udpif;
1501 struct dump_op ops[REVALIDATE_MAX_BATCH];
1502 struct udpif_flow_dump *udump, *next_udump;
1503 size_t n_ops, n_flows;
1504 unsigned int flow_limit;
1505 long long int max_idle;
1508 atomic_read(&udpif->flow_limit, &flow_limit);
1510 n_flows = udpif_get_n_flows(udpif);
1513 max_idle = ofproto_max_idle;
1514 if (n_flows > flow_limit) {
1515 must_del = n_flows > 2 * flow_limit;
1520 LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) {
1521 long long int used, now;
1522 struct udpif_key *ukey;
1525 ukey = ukey_lookup(revalidator, udump);
1527 used = udump->stats.used;
1528 if (!used && ukey) {
1529 used = ukey->created;
1532 if (must_del || (used && used < now - max_idle)) {
1533 struct dump_op *dop = &ops[n_ops++];
1535 dump_op_init(dop, udump->key, udump->key_len, ukey, udump);
1540 ukey = ukey_create(udump->key, udump->key_len, used);
1541 hmap_insert(&revalidator->ukeys, &ukey->hmap_node,
1546 if (!revalidate_ukey(udpif, udump, ukey)) {
1547 dpif_flow_del(udpif->dpif, udump->key, udump->key_len, NULL);
1548 ukey_delete(revalidator, ukey);
1551 list_remove(&udump->list_node);
1555 push_dump_ops(revalidator, ops, n_ops);
1557 LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) {
1558 list_remove(&udump->list_node);
1564 revalidator_sweep__(struct revalidator *revalidator, bool purge)
1566 struct dump_op ops[REVALIDATE_MAX_BATCH];
1567 struct udpif_key *ukey, *next;
1572 HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, &revalidator->ukeys) {
1573 if (!purge && ukey->mark) {
1576 struct dump_op *op = &ops[n_ops++];
1578 /* If we have previously seen a flow in the datapath, but didn't
1579 * see it during the most recent dump, delete it. This allows us
1580 * to clean up the ukey and keep the statistics consistent. */
1581 dump_op_init(op, ukey->key, ukey->key_len, ukey, NULL);
1582 if (n_ops == REVALIDATE_MAX_BATCH) {
1583 push_dump_ops(revalidator, ops, n_ops);
1590 push_dump_ops(revalidator, ops, n_ops);
1595 revalidator_sweep(struct revalidator *revalidator)
1597 revalidator_sweep__(revalidator, false);
1601 revalidator_purge(struct revalidator *revalidator)
1603 revalidator_sweep__(revalidator, true);
1607 upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED,
1608 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
1610 struct ds ds = DS_EMPTY_INITIALIZER;
1611 struct udpif *udpif;
1613 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1614 unsigned int flow_limit;
1617 atomic_read(&udpif->flow_limit, &flow_limit);
1619 ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif));
1620 ds_put_format(&ds, "\tflows : (current %"PRIu64")"
1621 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif),
1622 udpif->avg_n_flows, udpif->max_n_flows, flow_limit);
1623 ds_put_format(&ds, "\tdump duration : %lldms\n", udpif->dump_duration);
1625 ds_put_char(&ds, '\n');
1626 for (i = 0; i < n_revalidators; i++) {
1627 struct revalidator *revalidator = &udpif->revalidators[i];
1629 /* XXX: The result of hmap_count(&revalidator->ukeys) may not be
1630 * accurate because it's not protected by the revalidator mutex. */
1631 ovs_mutex_lock(&revalidator->mutex);
1632 ds_put_format(&ds, "\t%s: (dump queue %"PRIuSIZE") (keys %"PRIuSIZE
1633 ")\n", revalidator->name, revalidator->n_udumps,
1634 hmap_count(&revalidator->ukeys));
1635 ovs_mutex_unlock(&revalidator->mutex);
1639 unixctl_command_reply(conn, ds_cstr(&ds));
1643 /* Disable using the megaflows.
1645 * This command is only needed for advanced debugging, so it's not
1646 * documented in the man page. */
1648 upcall_unixctl_disable_megaflows(struct unixctl_conn *conn,
1649 int argc OVS_UNUSED,
1650 const char *argv[] OVS_UNUSED,
1651 void *aux OVS_UNUSED)
1653 atomic_store(&enable_megaflows, false);
1654 udpif_flush_all_datapaths();
1655 unixctl_command_reply(conn, "megaflows disabled");
1658 /* Re-enable using megaflows.
1660 * This command is only needed for advanced debugging, so it's not
1661 * documented in the man page. */
1663 upcall_unixctl_enable_megaflows(struct unixctl_conn *conn,
1664 int argc OVS_UNUSED,
1665 const char *argv[] OVS_UNUSED,
1666 void *aux OVS_UNUSED)
1668 atomic_store(&enable_megaflows, true);
1669 udpif_flush_all_datapaths();
1670 unixctl_command_reply(conn, "megaflows enabled");
1673 /* Set the flow limit.
1675 * This command is only needed for advanced debugging, so it's not
1676 * documented in the man page. */
1678 upcall_unixctl_set_flow_limit(struct unixctl_conn *conn,
1679 int argc OVS_UNUSED,
1680 const char *argv[] OVS_UNUSED,
1681 void *aux OVS_UNUSED)
1683 struct ds ds = DS_EMPTY_INITIALIZER;
1684 struct udpif *udpif;
1685 unsigned int flow_limit = atoi(argv[1]);
1687 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1688 atomic_store(&udpif->flow_limit, flow_limit);
1690 ds_put_format(&ds, "set flow_limit to %u\n", flow_limit);
1691 unixctl_command_reply(conn, ds_cstr(&ds));