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->handlers &&
300 (udpif->n_handlers != n_handlers
301 || udpif->n_revalidators != n_revalidators)) {
304 latch_set(&udpif->exit_latch);
306 for (i = 0; i < udpif->n_handlers; i++) {
307 struct handler *handler = &udpif->handlers[i];
309 xpthread_join(handler->thread, NULL);
312 for (i = 0; i < udpif->n_revalidators; i++) {
313 struct revalidator *revalidator = &udpif->revalidators[i];
315 ovs_mutex_lock(&revalidator->mutex);
316 xpthread_cond_signal(&revalidator->wake_cond);
317 ovs_mutex_unlock(&revalidator->mutex);
318 xpthread_join(revalidator->thread, NULL);
321 xpthread_join(udpif->flow_dumper, NULL);
323 for (i = 0; i < udpif->n_revalidators; i++) {
324 struct revalidator *revalidator = &udpif->revalidators[i];
325 struct udpif_flow_dump *udump, *next_udump;
327 LIST_FOR_EACH_SAFE (udump, next_udump, list_node,
328 &revalidator->udumps) {
329 list_remove(&udump->list_node);
333 /* Delete ukeys, and delete all flows from the datapath to prevent
334 * double-counting stats. */
335 revalidator_purge(revalidator);
336 hmap_destroy(&revalidator->ukeys);
337 ovs_mutex_destroy(&revalidator->mutex);
339 free(revalidator->name);
342 for (i = 0; i < udpif->n_handlers; i++) {
343 free(udpif->handlers[i].name);
345 latch_poll(&udpif->exit_latch);
347 free(udpif->revalidators);
348 udpif->revalidators = NULL;
349 udpif->n_revalidators = 0;
351 free(udpif->handlers);
352 udpif->handlers = NULL;
353 udpif->n_handlers = 0;
357 /* Starts the handler and revalidator threads, must be enclosed in
358 * ovsrcu quiescent state. */
360 udpif_start_threads(struct udpif *udpif, size_t n_handlers,
361 size_t n_revalidators)
363 if (!udpif->handlers && n_handlers) {
366 udpif->n_handlers = n_handlers;
367 udpif->n_revalidators = n_revalidators;
369 udpif->handlers = xzalloc(udpif->n_handlers * sizeof *udpif->handlers);
370 for (i = 0; i < udpif->n_handlers; i++) {
371 struct handler *handler = &udpif->handlers[i];
373 handler->udpif = udpif;
374 handler->handler_id = i;
375 xpthread_create(&handler->thread, NULL, udpif_upcall_handler,
379 udpif->revalidators = xzalloc(udpif->n_revalidators
380 * sizeof *udpif->revalidators);
381 for (i = 0; i < udpif->n_revalidators; i++) {
382 struct revalidator *revalidator = &udpif->revalidators[i];
384 revalidator->udpif = udpif;
385 list_init(&revalidator->udumps);
386 hmap_init(&revalidator->ukeys);
387 ovs_mutex_init(&revalidator->mutex);
388 xpthread_cond_init(&revalidator->wake_cond, NULL);
389 xpthread_create(&revalidator->thread, NULL, udpif_revalidator,
392 xpthread_create(&udpif->flow_dumper, NULL, udpif_flow_dumper, udpif);
396 /* Tells 'udpif' how many threads it should use to handle upcalls.
397 * 'n_handlers' and 'n_revalidators' can never be zero. 'udpif''s
398 * datapath handle must have packet reception enabled before starting
401 udpif_set_threads(struct udpif *udpif, size_t n_handlers,
402 size_t n_revalidators)
406 ovs_assert(n_handlers && n_revalidators);
408 ovsrcu_quiesce_start();
409 udpif_stop_threads(udpif);
411 error = dpif_handlers_set(udpif->dpif, n_handlers);
413 VLOG_ERR("failed to configure handlers in dpif %s: %s",
414 dpif_name(udpif->dpif), ovs_strerror(error));
418 udpif_start_threads(udpif, n_handlers, n_revalidators);
419 ovsrcu_quiesce_end();
422 /* Waits for all ongoing upcall translations to complete. This ensures that
423 * there are no transient references to any removed ofprotos (or other
424 * objects). In particular, this should be called after an ofproto is removed
425 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
427 udpif_synchronize(struct udpif *udpif)
429 /* This is stronger than necessary. It would be sufficient to ensure
430 * (somehow) that each handler and revalidator thread had passed through
431 * its main loop once. */
432 size_t n_handlers = udpif->n_handlers;
433 size_t n_revalidators = udpif->n_revalidators;
435 ovsrcu_quiesce_start();
436 udpif_stop_threads(udpif);
437 udpif_start_threads(udpif, n_handlers, n_revalidators);
438 ovsrcu_quiesce_end();
441 /* Notifies 'udpif' that something changed which may render previous
442 * xlate_actions() results invalid. */
444 udpif_revalidate(struct udpif *udpif)
446 seq_change(udpif->reval_seq);
449 /* Returns a seq which increments every time 'udpif' pulls stats from the
450 * datapath. Callers can use this to get a sense of when might be a good time
451 * to do periodic work which relies on relatively up to date statistics. */
453 udpif_dump_seq(struct udpif *udpif)
455 return udpif->dump_seq;
459 udpif_get_memory_usage(struct udpif *udpif, struct simap *usage)
463 simap_increase(usage, "flow_dumpers", 1);
465 simap_increase(usage, "handlers", udpif->n_handlers);
467 simap_increase(usage, "revalidators", udpif->n_revalidators);
468 for (i = 0; i < udpif->n_revalidators; i++) {
469 struct revalidator *revalidator = &udpif->revalidators[i];
470 ovs_mutex_lock(&revalidator->mutex);
471 simap_increase(usage, "revalidator dumps", revalidator->n_udumps);
473 /* XXX: This isn't technically thread safe because the revalidator
474 * ukeys maps isn't protected by a mutex since it's per thread. */
475 simap_increase(usage, "revalidator keys",
476 hmap_count(&revalidator->ukeys));
477 ovs_mutex_unlock(&revalidator->mutex);
481 /* Remove flows from a single datapath. */
483 udpif_flush(struct udpif *udpif)
485 size_t n_handlers, n_revalidators;
487 n_handlers = udpif->n_handlers;
488 n_revalidators = udpif->n_revalidators;
490 ovsrcu_quiesce_start();
492 udpif_stop_threads(udpif);
493 dpif_flow_flush(udpif->dpif);
494 udpif_start_threads(udpif, n_handlers, n_revalidators);
496 ovsrcu_quiesce_end();
499 /* Removes all flows from all datapaths. */
501 udpif_flush_all_datapaths(void)
505 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
512 udpif_get_n_flows(struct udpif *udpif)
514 long long int time, now;
518 atomic_read(&udpif->n_flows_timestamp, &time);
519 if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) {
520 struct dpif_dp_stats stats;
522 atomic_store(&udpif->n_flows_timestamp, now);
523 dpif_get_dp_stats(udpif->dpif, &stats);
524 flow_count = stats.n_flows;
525 atomic_store(&udpif->n_flows, flow_count);
526 ovs_mutex_unlock(&udpif->n_flows_mutex);
528 atomic_read(&udpif->n_flows, &flow_count);
534 udpif_flow_dumper(void *arg)
536 struct udpif *udpif = arg;
538 set_subprogram_name("flow_dumper");
539 while (!latch_is_set(&udpif->exit_latch)) {
540 const struct dpif_flow_stats *stats;
541 long long int start_time, duration;
542 const struct nlattr *key, *mask;
543 struct dpif_flow_dump dump;
544 size_t key_len, mask_len;
545 unsigned int flow_limit;
546 bool need_revalidate;
552 reval_seq = seq_read(udpif->reval_seq);
553 need_revalidate = udpif->last_reval_seq != reval_seq;
554 udpif->last_reval_seq = reval_seq;
556 n_flows = udpif_get_n_flows(udpif);
557 udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows);
558 udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2;
560 start_time = time_msec();
561 error = dpif_flow_dump_start(&dump, udpif->dpif);
563 VLOG_INFO("Failed to start flow dump (%s)", ovs_strerror(error));
566 dpif_flow_dump_state_init(udpif->dpif, &state);
567 while (dpif_flow_dump_next(&dump, state, &key, &key_len,
568 &mask, &mask_len, NULL, NULL, &stats)
569 && !latch_is_set(&udpif->exit_latch)) {
570 struct udpif_flow_dump *udump = xmalloc(sizeof *udump);
571 struct revalidator *revalidator;
573 udump->key_hash = hash_bytes(key, key_len, udpif->secret);
574 memcpy(&udump->key_buf, key, key_len);
575 udump->key = (struct nlattr *) &udump->key_buf;
576 udump->key_len = key_len;
578 memcpy(&udump->mask_buf, mask, mask_len);
579 udump->mask = (struct nlattr *) &udump->mask_buf;
580 udump->mask_len = mask_len;
582 udump->stats = *stats;
583 udump->need_revalidate = need_revalidate;
585 revalidator = &udpif->revalidators[udump->key_hash
586 % udpif->n_revalidators];
588 ovs_mutex_lock(&revalidator->mutex);
589 while (revalidator->n_udumps >= REVALIDATE_MAX_BATCH * 3
590 && !latch_is_set(&udpif->exit_latch)) {
591 ovs_mutex_cond_wait(&revalidator->wake_cond,
592 &revalidator->mutex);
594 list_push_back(&revalidator->udumps, &udump->list_node);
595 revalidator->n_udumps++;
596 xpthread_cond_signal(&revalidator->wake_cond);
597 ovs_mutex_unlock(&revalidator->mutex);
599 dpif_flow_dump_state_uninit(udpif->dpif, state);
600 dpif_flow_dump_done(&dump);
602 /* Let all the revalidators finish and garbage collect. */
603 seq_change(udpif->dump_seq);
604 for (i = 0; i < udpif->n_revalidators; i++) {
605 struct revalidator *revalidator = &udpif->revalidators[i];
606 ovs_mutex_lock(&revalidator->mutex);
607 xpthread_cond_signal(&revalidator->wake_cond);
608 ovs_mutex_unlock(&revalidator->mutex);
611 for (i = 0; i < udpif->n_revalidators; i++) {
612 struct revalidator *revalidator = &udpif->revalidators[i];
614 ovs_mutex_lock(&revalidator->mutex);
615 while (revalidator->dump_seq != seq_read(udpif->dump_seq)
616 && !latch_is_set(&udpif->exit_latch)) {
617 ovs_mutex_cond_wait(&revalidator->wake_cond,
618 &revalidator->mutex);
620 ovs_mutex_unlock(&revalidator->mutex);
623 duration = MAX(time_msec() - start_time, 1);
624 udpif->dump_duration = duration;
625 atomic_read(&udpif->flow_limit, &flow_limit);
626 if (duration > 2000) {
627 flow_limit /= duration / 1000;
628 } else if (duration > 1300) {
629 flow_limit = flow_limit * 3 / 4;
630 } else if (duration < 1000 && n_flows > 2000
631 && flow_limit < n_flows * 1000 / duration) {
634 flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000));
635 atomic_store(&udpif->flow_limit, flow_limit);
637 if (duration > 2000) {
638 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
643 poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500));
644 seq_wait(udpif->reval_seq, udpif->last_reval_seq);
645 latch_wait(&udpif->exit_latch);
652 /* The upcall handler thread tries to read a batch of FLOW_MISS_MAX_BATCH
653 * upcalls from dpif, processes the batch and installs corresponding flows
656 udpif_upcall_handler(void *arg)
658 struct handler *handler = arg;
659 struct udpif *udpif = handler->udpif;
660 struct hmap misses = HMAP_INITIALIZER(&misses);
662 handler->name = xasprintf("handler_%u", ovsthread_id_self());
663 set_subprogram_name("%s", handler->name);
665 while (!latch_is_set(&handler->udpif->exit_latch)) {
666 struct upcall upcalls[FLOW_MISS_MAX_BATCH];
667 struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH];
668 struct flow_miss *miss;
671 n_upcalls = read_upcalls(handler, upcalls, miss_buf, &misses);
673 dpif_recv_wait(udpif->dpif, handler->handler_id);
674 latch_wait(&udpif->exit_latch);
677 handle_upcalls(handler, &misses, upcalls, n_upcalls);
679 HMAP_FOR_EACH (miss, hmap_node, &misses) {
680 xlate_out_uninit(&miss->xout);
683 for (i = 0; i < n_upcalls; i++) {
684 ofpbuf_uninit(&upcalls[i].dpif_upcall.packet);
685 ofpbuf_uninit(&upcalls[i].upcall_buf);
690 hmap_destroy(&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);
795 /* Calculates slow path actions for 'xout'. 'buf' must statically be
796 * initialized with at least 128 bytes of space. */
798 compose_slow_path(struct udpif *udpif, struct xlate_out *xout,
799 struct flow *flow, odp_port_t odp_in_port,
802 union user_action_cookie cookie;
806 cookie.type = USER_ACTION_COOKIE_SLOW_PATH;
807 cookie.slow_path.unused = 0;
808 cookie.slow_path.reason = xout->slow;
810 port = xout->slow & (SLOW_CFM | SLOW_BFD | SLOW_LACP | SLOW_STP)
813 pid = dpif_port_get_pid(udpif->dpif, port, flow_hash_5tuple(flow, 0));
814 odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, buf);
817 static struct flow_miss *
818 flow_miss_find(struct hmap *todo, const struct ofproto_dpif *ofproto,
819 const struct flow *flow, uint32_t hash)
821 struct flow_miss *miss;
823 HMAP_FOR_EACH_WITH_HASH (miss, hmap_node, hash, todo) {
824 if (miss->ofproto == ofproto && flow_equal(&miss->flow, flow)) {
832 /* Reads and classifies upcalls. Returns the number of upcalls successfully
835 read_upcalls(struct handler *handler,
836 struct upcall upcalls[FLOW_MISS_MAX_BATCH],
837 struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH],
840 struct udpif *udpif = handler->udpif;
843 size_t n_upcalls = 0;
846 * Try reading FLOW_MISS_MAX_BATCH upcalls from dpif.
848 * Extract the flow from each upcall. Construct in 'misses' a hash table
849 * that maps each unique flow to a 'struct flow_miss'.
851 * Most commonly there is a single packet per flow_miss, but there are
852 * several reasons why there might be more than one, e.g.:
854 * - The dpif packet interface does not support TSO (or UFO, etc.), so a
855 * large packet sent to userspace is split into a sequence of smaller
858 * - A stream of quickly arriving packets in an established "slow-pathed"
861 * - Rarely, a stream of quickly arriving packets in a flow not yet
862 * established. (This is rare because most protocols do not send
863 * multiple back-to-back packets before receiving a reply from the
864 * other end of the connection, which gives OVS a chance to set up a
867 for (i = 0; i < FLOW_MISS_MAX_BATCH; i++) {
868 struct upcall *upcall = &upcalls[n_upcalls];
869 struct flow_miss *miss = &miss_buf[n_misses];
870 struct dpif_upcall *dupcall;
871 struct ofpbuf *packet;
872 struct flow_miss *existing_miss;
873 struct ofproto_dpif *ofproto;
874 struct dpif_sflow *sflow;
875 struct dpif_ipfix *ipfix;
877 enum upcall_type type;
878 odp_port_t odp_in_port;
881 ofpbuf_use_stub(&upcall->upcall_buf, upcall->upcall_stub,
882 sizeof upcall->upcall_stub);
883 error = dpif_recv(udpif->dpif, handler->handler_id,
884 &upcall->dpif_upcall, &upcall->upcall_buf);
886 ofpbuf_uninit(&upcall->upcall_buf);
890 dupcall = &upcall->dpif_upcall;
891 packet = &dupcall->packet;
892 error = xlate_receive(udpif->backer, packet, dupcall->key,
893 dupcall->key_len, &flow,
894 &ofproto, &ipfix, &sflow, NULL, &odp_in_port);
896 if (error == ENODEV) {
897 /* Received packet on datapath port for which we couldn't
898 * associate an ofproto. This can happen if a port is removed
899 * while traffic is being received. Print a rate-limited
900 * message in case it happens frequently. Install a drop flow
901 * so that future packets of the flow are inexpensively dropped
903 VLOG_INFO_RL(&rl, "received packet on unassociated datapath "
904 "port %"PRIu32, odp_in_port);
905 dpif_flow_put(udpif->dpif, DPIF_FP_CREATE | DPIF_FP_MODIFY,
906 dupcall->key, dupcall->key_len, NULL, 0, NULL, 0,
912 type = classify_upcall(upcall);
913 if (type == MISS_UPCALL) {
915 struct pkt_metadata md = pkt_metadata_from_flow(&flow);
917 flow_extract(packet, &md, &miss->flow);
918 hash = flow_hash(&miss->flow, 0);
919 existing_miss = flow_miss_find(misses, ofproto, &miss->flow,
921 if (!existing_miss) {
922 hmap_insert(misses, &miss->hmap_node, hash);
923 miss->ofproto = ofproto;
924 miss->key = dupcall->key;
925 miss->key_len = dupcall->key_len;
926 miss->upcall_type = dupcall->type;
927 miss->stats.n_packets = 0;
928 miss->stats.n_bytes = 0;
929 miss->stats.used = time_msec();
930 miss->stats.tcp_flags = 0;
931 miss->odp_in_port = odp_in_port;
935 miss = existing_miss;
937 miss->stats.tcp_flags |= ntohs(miss->flow.tcp_flags);
938 miss->stats.n_bytes += ofpbuf_size(packet);
939 miss->stats.n_packets++;
941 upcall->flow_miss = miss;
949 union user_action_cookie cookie;
951 memset(&cookie, 0, sizeof cookie);
952 memcpy(&cookie, nl_attr_get(dupcall->userdata),
953 sizeof cookie.sflow);
954 dpif_sflow_received(sflow, packet, &flow, odp_in_port,
960 dpif_ipfix_bridge_sample(ipfix, packet, &flow);
963 case FLOW_SAMPLE_UPCALL:
965 union user_action_cookie cookie;
967 memset(&cookie, 0, sizeof cookie);
968 memcpy(&cookie, nl_attr_get(dupcall->userdata),
969 sizeof cookie.flow_sample);
971 /* The flow reflects exactly the contents of the packet.
972 * Sample the packet using it. */
973 dpif_ipfix_flow_sample(ipfix, packet, &flow,
974 cookie.flow_sample.collector_set_id,
975 cookie.flow_sample.probability,
976 cookie.flow_sample.obs_domain_id,
977 cookie.flow_sample.obs_point_id);
986 dpif_ipfix_unref(ipfix);
987 dpif_sflow_unref(sflow);
990 ofpbuf_uninit(&upcall->dpif_upcall.packet);
991 ofpbuf_uninit(&upcall->upcall_buf);
998 handle_upcalls(struct handler *handler, struct hmap *misses,
999 struct upcall *upcalls, size_t n_upcalls)
1001 struct udpif *udpif = handler->udpif;
1002 struct dpif_op *opsp[FLOW_MISS_MAX_BATCH * 2];
1003 struct dpif_op ops[FLOW_MISS_MAX_BATCH * 2];
1004 struct flow_miss *miss;
1006 unsigned int flow_limit;
1007 bool fail_open, may_put;
1009 atomic_read(&udpif->flow_limit, &flow_limit);
1010 may_put = udpif_get_n_flows(udpif) < flow_limit;
1012 /* Initialize each 'struct flow_miss's ->xout.
1014 * We do this per-flow_miss rather than per-packet because, most commonly,
1015 * all the packets in a flow can use the same translation.
1017 * We can't do this in the previous loop because we need the TCP flags for
1018 * all the packets in each miss. */
1020 HMAP_FOR_EACH (miss, hmap_node, misses) {
1021 struct xlate_in xin;
1023 xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL,
1024 miss->stats.tcp_flags, NULL);
1025 xin.may_learn = true;
1027 if (miss->upcall_type == DPIF_UC_MISS) {
1028 xin.resubmit_stats = &miss->stats;
1030 /* For non-miss upcalls, there's a flow in the datapath which this
1031 * packet was accounted to. Presumably the revalidators will deal
1032 * with pushing its stats eventually. */
1035 xlate_actions(&xin, &miss->xout);
1036 fail_open = fail_open || miss->xout.fail_open;
1039 /* Now handle the packets individually in order of arrival. In the common
1040 * case each packet of a miss can share the same actions, but slow-pathed
1041 * packets need to be translated individually:
1043 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1044 * processes received packets for these protocols.
1046 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1049 * The loop fills 'ops' with an array of operations to execute in the
1052 for (i = 0; i < n_upcalls; i++) {
1053 struct upcall *upcall = &upcalls[i];
1054 struct flow_miss *miss = upcall->flow_miss;
1055 struct ofpbuf *packet = &upcall->dpif_upcall.packet;
1057 ovs_be16 flow_vlan_tci;
1059 /* Save a copy of flow.vlan_tci in case it is changed to
1060 * generate proper mega flow masks for VLAN splinter flows. */
1061 flow_vlan_tci = miss->flow.vlan_tci;
1063 if (miss->xout.slow) {
1064 struct xlate_in xin;
1066 xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL, 0, packet);
1067 xlate_actions_for_side_effects(&xin);
1070 if (miss->flow.in_port.ofp_port
1071 != vsp_realdev_to_vlandev(miss->ofproto,
1072 miss->flow.in_port.ofp_port,
1073 miss->flow.vlan_tci)) {
1074 /* This packet was received on a VLAN splinter port. We
1075 * added a VLAN to the packet to make the packet resemble
1076 * the flow, but the actions were composed assuming that
1077 * the packet contained no VLAN. So, we must remove the
1078 * VLAN header from the packet before trying to execute the
1080 if (ofpbuf_size(&miss->xout.odp_actions)) {
1081 eth_pop_vlan(packet);
1084 /* Remove the flow vlan tags inserted by vlan splinter logic
1085 * to ensure megaflow masks generated match the data path flow. */
1086 miss->flow.vlan_tci = 0;
1089 /* Do not install a flow into the datapath if:
1091 * - The datapath already has too many flows.
1093 * - An earlier iteration of this loop already put the same flow.
1095 * - We received this packet via some flow installed in the kernel
1099 && upcall->dpif_upcall.type == DPIF_UC_MISS) {
1105 atomic_read(&enable_megaflows, &megaflow);
1106 ofpbuf_use_stack(&mask, &miss->mask_buf, sizeof miss->mask_buf);
1110 max_mpls = ofproto_dpif_get_max_mpls_depth(miss->ofproto);
1111 odp_flow_key_from_mask(&mask, &miss->xout.wc.masks,
1112 &miss->flow, UINT32_MAX, max_mpls);
1116 op->type = DPIF_OP_FLOW_PUT;
1117 op->u.flow_put.flags = DPIF_FP_CREATE | DPIF_FP_MODIFY;
1118 op->u.flow_put.key = miss->key;
1119 op->u.flow_put.key_len = miss->key_len;
1120 op->u.flow_put.mask = ofpbuf_data(&mask);
1121 op->u.flow_put.mask_len = ofpbuf_size(&mask);
1122 op->u.flow_put.stats = NULL;
1124 if (!miss->xout.slow) {
1125 op->u.flow_put.actions = ofpbuf_data(&miss->xout.odp_actions);
1126 op->u.flow_put.actions_len = ofpbuf_size(&miss->xout.odp_actions);
1130 ofpbuf_use_stack(&buf, miss->slow_path_buf,
1131 sizeof miss->slow_path_buf);
1132 compose_slow_path(udpif, &miss->xout, &miss->flow,
1133 miss->odp_in_port, &buf);
1134 op->u.flow_put.actions = ofpbuf_data(&buf);
1135 op->u.flow_put.actions_len = ofpbuf_size(&buf);
1140 * The 'miss' may be shared by multiple upcalls. Restore
1141 * the saved flow vlan_tci field before processing the next
1143 miss->flow.vlan_tci = flow_vlan_tci;
1145 if (ofpbuf_size(&miss->xout.odp_actions)) {
1148 op->type = DPIF_OP_EXECUTE;
1149 op->u.execute.packet = packet;
1150 odp_key_to_pkt_metadata(miss->key, miss->key_len,
1152 op->u.execute.actions = ofpbuf_data(&miss->xout.odp_actions);
1153 op->u.execute.actions_len = ofpbuf_size(&miss->xout.odp_actions);
1154 op->u.execute.needs_help = (miss->xout.slow & SLOW_ACTION) != 0;
1158 /* Special case for fail-open mode.
1160 * If we are in fail-open mode, but we are connected to a controller too,
1161 * then we should send the packet up to the controller in the hope that it
1162 * will try to set up a flow and thereby allow us to exit fail-open.
1164 * See the top-level comment in fail-open.c for more information.
1166 * Copy packets before they are modified by execution. */
1168 for (i = 0; i < n_upcalls; i++) {
1169 struct upcall *upcall = &upcalls[i];
1170 struct flow_miss *miss = upcall->flow_miss;
1171 struct ofpbuf *packet = &upcall->dpif_upcall.packet;
1172 struct ofproto_packet_in *pin;
1174 pin = xmalloc(sizeof *pin);
1175 pin->up.packet = xmemdup(ofpbuf_data(packet), ofpbuf_size(packet));
1176 pin->up.packet_len = ofpbuf_size(packet);
1177 pin->up.reason = OFPR_NO_MATCH;
1178 pin->up.table_id = 0;
1179 pin->up.cookie = OVS_BE64_MAX;
1180 flow_get_metadata(&miss->flow, &pin->up.fmd);
1181 pin->send_len = 0; /* Not used for flow table misses. */
1182 pin->miss_type = OFPROTO_PACKET_IN_NO_MISS;
1183 ofproto_dpif_send_packet_in(miss->ofproto, pin);
1187 /* Execute batch. */
1188 for (i = 0; i < n_ops; i++) {
1191 dpif_operate(udpif->dpif, opsp, n_ops);
1194 static struct udpif_key *
1195 ukey_lookup(struct revalidator *revalidator, struct udpif_flow_dump *udump)
1197 struct udpif_key *ukey;
1199 HMAP_FOR_EACH_WITH_HASH (ukey, hmap_node, udump->key_hash,
1200 &revalidator->ukeys) {
1201 if (ukey->key_len == udump->key_len
1202 && !memcmp(ukey->key, udump->key, udump->key_len)) {
1209 static struct udpif_key *
1210 ukey_create(const struct nlattr *key, size_t key_len, long long int used)
1212 struct udpif_key *ukey = xmalloc(sizeof *ukey);
1214 ukey->key = (struct nlattr *) &ukey->key_buf;
1215 memcpy(&ukey->key_buf, key, key_len);
1216 ukey->key_len = key_len;
1219 ukey->created = used ? used : time_msec();
1220 memset(&ukey->stats, 0, sizeof ukey->stats);
1221 ukey->xcache = NULL;
1227 ukey_delete(struct revalidator *revalidator, struct udpif_key *ukey)
1229 hmap_remove(&revalidator->ukeys, &ukey->hmap_node);
1230 xlate_cache_delete(ukey->xcache);
1235 should_revalidate(uint64_t packets, long long int used)
1237 long long int metric, now, duration;
1239 /* Calculate the mean time between seeing these packets. If this
1240 * exceeds the threshold, then delete the flow rather than performing
1241 * costly revalidation for flows that aren't being hit frequently.
1243 * This is targeted at situations where the dump_duration is high (~1s),
1244 * and revalidation is triggered by a call to udpif_revalidate(). In
1245 * these situations, revalidation of all flows causes fluctuations in the
1246 * flow_limit due to the interaction with the dump_duration and max_idle.
1247 * This tends to result in deletion of low-throughput flows anyway, so
1248 * skip the revalidation and just delete those flows. */
1249 packets = MAX(packets, 1);
1250 now = MAX(used, time_msec());
1251 duration = now - used;
1252 metric = duration / packets;
1261 revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump,
1262 struct udpif_key *ukey)
1264 struct ofpbuf xout_actions, *actions;
1265 uint64_t slow_path_buf[128 / 8];
1266 struct xlate_out xout, *xoutp;
1267 struct netflow *netflow;
1268 struct flow flow, udump_mask;
1269 struct ofproto_dpif *ofproto;
1270 struct dpif_flow_stats push;
1271 uint32_t *udump32, *xout32;
1272 odp_port_t odp_in_port;
1273 struct xlate_in xin;
1274 long long int last_used;
1284 /* If we don't need to revalidate, we can simply push the stats contained
1285 * in the udump, otherwise we'll have to get the actions so we can check
1287 if (udump->need_revalidate) {
1288 if (dpif_flow_get(udpif->dpif, ukey->key, ukey->key_len, &actions,
1294 last_used = ukey->stats.used;
1295 push.used = udump->stats.used;
1296 push.tcp_flags = udump->stats.tcp_flags;
1297 push.n_packets = udump->stats.n_packets > ukey->stats.n_packets
1298 ? udump->stats.n_packets - ukey->stats.n_packets
1300 push.n_bytes = udump->stats.n_bytes > ukey->stats.n_bytes
1301 ? udump->stats.n_bytes - ukey->stats.n_bytes
1303 ukey->stats = udump->stats;
1305 if (udump->need_revalidate && last_used
1306 && !should_revalidate(push.n_packets, last_used)) {
1311 if (!push.n_packets && !udump->need_revalidate) {
1316 may_learn = push.n_packets > 0;
1317 if (ukey->xcache && !udump->need_revalidate) {
1318 xlate_push_stats(ukey->xcache, may_learn, &push);
1323 error = xlate_receive(udpif->backer, NULL, ukey->key, ukey->key_len, &flow,
1324 &ofproto, NULL, NULL, &netflow, &odp_in_port);
1329 if (udump->need_revalidate) {
1330 xlate_cache_clear(ukey->xcache);
1332 if (!ukey->xcache) {
1333 ukey->xcache = xlate_cache_new();
1336 xlate_in_init(&xin, ofproto, &flow, NULL, push.tcp_flags, NULL);
1337 xin.resubmit_stats = push.n_packets ? &push : NULL;
1338 xin.xcache = ukey->xcache;
1339 xin.may_learn = may_learn;
1340 xin.skip_wildcards = !udump->need_revalidate;
1341 xlate_actions(&xin, &xout);
1344 if (!udump->need_revalidate) {
1350 ofpbuf_use_const(&xout_actions, ofpbuf_data(&xout.odp_actions),
1351 ofpbuf_size(&xout.odp_actions));
1353 ofpbuf_use_stack(&xout_actions, slow_path_buf, sizeof slow_path_buf);
1354 compose_slow_path(udpif, &xout, &flow, odp_in_port, &xout_actions);
1357 if (!ofpbuf_equal(&xout_actions, actions)) {
1361 if (odp_flow_key_to_mask(udump->mask, udump->mask_len, &udump_mask, &flow)
1366 /* Since the kernel is free to ignore wildcarded bits in the mask, we can't
1367 * directly check that the masks are the same. Instead we check that the
1368 * mask in the kernel is more specific i.e. less wildcarded, than what
1369 * we've calculated here. This guarantees we don't catch any packets we
1370 * shouldn't with the megaflow. */
1371 udump32 = (uint32_t *) &udump_mask;
1372 xout32 = (uint32_t *) &xout.wc.masks;
1373 for (i = 0; i < FLOW_U32S; i++) {
1374 if ((udump32[i] | xout32[i]) != udump32[i]) {
1383 netflow_expire(netflow, &flow);
1384 netflow_flow_clear(netflow, &flow);
1386 netflow_unref(netflow);
1388 ofpbuf_delete(actions);
1389 xlate_out_uninit(xoutp);
1394 struct udpif_key *ukey;
1395 struct udpif_flow_dump *udump;
1396 struct dpif_flow_stats stats; /* Stats for 'op'. */
1397 struct dpif_op op; /* Flow del operation. */
1401 dump_op_init(struct dump_op *op, const struct nlattr *key, size_t key_len,
1402 struct udpif_key *ukey, struct udpif_flow_dump *udump)
1406 op->op.type = DPIF_OP_FLOW_DEL;
1407 op->op.u.flow_del.key = key;
1408 op->op.u.flow_del.key_len = key_len;
1409 op->op.u.flow_del.stats = &op->stats;
1413 push_dump_ops(struct revalidator *revalidator,
1414 struct dump_op *ops, size_t n_ops)
1416 struct udpif *udpif = revalidator->udpif;
1417 struct dpif_op *opsp[REVALIDATE_MAX_BATCH];
1420 ovs_assert(n_ops <= REVALIDATE_MAX_BATCH);
1421 for (i = 0; i < n_ops; i++) {
1422 opsp[i] = &ops[i].op;
1424 dpif_operate(udpif->dpif, opsp, n_ops);
1426 for (i = 0; i < n_ops; i++) {
1427 struct dump_op *op = &ops[i];
1428 struct dpif_flow_stats *push, *stats, push_buf;
1430 stats = op->op.u.flow_del.stats;
1433 push->used = MAX(stats->used, op->ukey->stats.used);
1434 push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags;
1435 push->n_packets = stats->n_packets - op->ukey->stats.n_packets;
1436 push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes;
1441 if (push->n_packets || netflow_exists()) {
1442 struct ofproto_dpif *ofproto;
1443 struct netflow *netflow;
1447 may_learn = push->n_packets > 0;
1448 if (op->ukey && op->ukey->xcache) {
1449 xlate_push_stats(op->ukey->xcache, may_learn, push);
1453 if (!xlate_receive(udpif->backer, NULL, op->op.u.flow_del.key,
1454 op->op.u.flow_del.key_len, &flow, &ofproto,
1455 NULL, NULL, &netflow, NULL)) {
1456 struct xlate_in xin;
1458 xlate_in_init(&xin, ofproto, &flow, NULL, push->tcp_flags,
1460 xin.resubmit_stats = push->n_packets ? push : NULL;
1461 xin.may_learn = may_learn;
1462 xin.skip_wildcards = true;
1463 xlate_actions_for_side_effects(&xin);
1466 netflow_expire(netflow, &flow);
1467 netflow_flow_clear(netflow, &flow);
1468 netflow_unref(netflow);
1474 for (i = 0; i < n_ops; i++) {
1475 struct udpif_key *ukey;
1477 /* If there's a udump, this ukey came directly from a datapath flow
1478 * dump. Sometimes a datapath can send duplicates in flow dumps, in
1479 * which case we wouldn't want to double-free a ukey, so avoid that by
1480 * looking up the ukey again.
1482 * If there's no udump then we know what we're doing. */
1483 ukey = (ops[i].udump
1484 ? ukey_lookup(revalidator, ops[i].udump)
1487 ukey_delete(revalidator, ukey);
1493 revalidate_udumps(struct revalidator *revalidator, struct list *udumps)
1495 struct udpif *udpif = revalidator->udpif;
1497 struct dump_op ops[REVALIDATE_MAX_BATCH];
1498 struct udpif_flow_dump *udump, *next_udump;
1499 size_t n_ops, n_flows;
1500 unsigned int flow_limit;
1501 long long int max_idle;
1504 atomic_read(&udpif->flow_limit, &flow_limit);
1506 n_flows = udpif_get_n_flows(udpif);
1509 max_idle = ofproto_max_idle;
1510 if (n_flows > flow_limit) {
1511 must_del = n_flows > 2 * flow_limit;
1516 LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) {
1517 long long int used, now;
1518 struct udpif_key *ukey;
1521 ukey = ukey_lookup(revalidator, udump);
1523 used = udump->stats.used;
1524 if (!used && ukey) {
1525 used = ukey->created;
1528 if (must_del || (used && used < now - max_idle)) {
1529 struct dump_op *dop = &ops[n_ops++];
1531 dump_op_init(dop, udump->key, udump->key_len, ukey, udump);
1536 ukey = ukey_create(udump->key, udump->key_len, used);
1537 hmap_insert(&revalidator->ukeys, &ukey->hmap_node,
1542 if (!revalidate_ukey(udpif, udump, ukey)) {
1543 dpif_flow_del(udpif->dpif, udump->key, udump->key_len, NULL);
1544 ukey_delete(revalidator, ukey);
1547 list_remove(&udump->list_node);
1551 push_dump_ops(revalidator, ops, n_ops);
1553 LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) {
1554 list_remove(&udump->list_node);
1560 revalidator_sweep__(struct revalidator *revalidator, bool purge)
1562 struct dump_op ops[REVALIDATE_MAX_BATCH];
1563 struct udpif_key *ukey, *next;
1568 HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, &revalidator->ukeys) {
1569 if (!purge && ukey->mark) {
1572 struct dump_op *op = &ops[n_ops++];
1574 /* If we have previously seen a flow in the datapath, but didn't
1575 * see it during the most recent dump, delete it. This allows us
1576 * to clean up the ukey and keep the statistics consistent. */
1577 dump_op_init(op, ukey->key, ukey->key_len, ukey, NULL);
1578 if (n_ops == REVALIDATE_MAX_BATCH) {
1579 push_dump_ops(revalidator, ops, n_ops);
1586 push_dump_ops(revalidator, ops, n_ops);
1591 revalidator_sweep(struct revalidator *revalidator)
1593 revalidator_sweep__(revalidator, false);
1597 revalidator_purge(struct revalidator *revalidator)
1599 revalidator_sweep__(revalidator, true);
1603 upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED,
1604 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
1606 struct ds ds = DS_EMPTY_INITIALIZER;
1607 struct udpif *udpif;
1609 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1610 unsigned int flow_limit;
1613 atomic_read(&udpif->flow_limit, &flow_limit);
1615 ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif));
1616 ds_put_format(&ds, "\tflows : (current %"PRIu64")"
1617 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif),
1618 udpif->avg_n_flows, udpif->max_n_flows, flow_limit);
1619 ds_put_format(&ds, "\tdump duration : %lldms\n", udpif->dump_duration);
1621 ds_put_char(&ds, '\n');
1622 for (i = 0; i < n_revalidators; i++) {
1623 struct revalidator *revalidator = &udpif->revalidators[i];
1625 /* XXX: The result of hmap_count(&revalidator->ukeys) may not be
1626 * accurate because it's not protected by the revalidator mutex. */
1627 ovs_mutex_lock(&revalidator->mutex);
1628 ds_put_format(&ds, "\t%s: (dump queue %"PRIuSIZE") (keys %"PRIuSIZE
1629 ")\n", revalidator->name, revalidator->n_udumps,
1630 hmap_count(&revalidator->ukeys));
1631 ovs_mutex_unlock(&revalidator->mutex);
1635 unixctl_command_reply(conn, ds_cstr(&ds));
1639 /* Disable using the megaflows.
1641 * This command is only needed for advanced debugging, so it's not
1642 * documented in the man page. */
1644 upcall_unixctl_disable_megaflows(struct unixctl_conn *conn,
1645 int argc OVS_UNUSED,
1646 const char *argv[] OVS_UNUSED,
1647 void *aux OVS_UNUSED)
1649 atomic_store(&enable_megaflows, false);
1650 udpif_flush_all_datapaths();
1651 unixctl_command_reply(conn, "megaflows disabled");
1654 /* Re-enable using megaflows.
1656 * This command is only needed for advanced debugging, so it's not
1657 * documented in the man page. */
1659 upcall_unixctl_enable_megaflows(struct unixctl_conn *conn,
1660 int argc OVS_UNUSED,
1661 const char *argv[] OVS_UNUSED,
1662 void *aux OVS_UNUSED)
1664 atomic_store(&enable_megaflows, true);
1665 udpif_flush_all_datapaths();
1666 unixctl_command_reply(conn, "megaflows enabled");
1669 /* Set the flow limit.
1671 * This command is only needed for advanced debugging, so it's not
1672 * documented in the man page. */
1674 upcall_unixctl_set_flow_limit(struct unixctl_conn *conn,
1675 int argc OVS_UNUSED,
1676 const char *argv[] OVS_UNUSED,
1677 void *aux OVS_UNUSED)
1679 struct ds ds = DS_EMPTY_INITIALIZER;
1680 struct udpif *udpif;
1681 unsigned int flow_limit = atoi(argv[1]);
1683 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1684 atomic_store(&udpif->flow_limit, flow_limit);
1686 ds_put_format(&ds, "set flow_limit to %u\n", flow_limit);
1687 unixctl_command_reply(conn, ds_cstr(&ds));