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_flow_dumper(void *);
227 static void *udpif_upcall_handler(void *);
228 static void *udpif_revalidator(void *);
229 static uint64_t udpif_get_n_flows(struct udpif *);
230 static void revalidate_udumps(struct revalidator *, struct list *udumps);
231 static void revalidator_sweep(struct revalidator *);
232 static void revalidator_purge(struct revalidator *);
233 static void upcall_unixctl_show(struct unixctl_conn *conn, int argc,
234 const char *argv[], void *aux);
235 static void upcall_unixctl_disable_megaflows(struct unixctl_conn *, int argc,
236 const char *argv[], void *aux);
237 static void upcall_unixctl_enable_megaflows(struct unixctl_conn *, int argc,
238 const char *argv[], void *aux);
239 static void upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, int argc,
240 const char *argv[], void *aux);
241 static void ukey_delete(struct revalidator *, struct udpif_key *);
243 static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true);
246 udpif_create(struct dpif_backer *backer, struct dpif *dpif)
248 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
249 struct udpif *udpif = xzalloc(sizeof *udpif);
251 if (ovsthread_once_start(&once)) {
252 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show,
254 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
255 upcall_unixctl_disable_megaflows, NULL);
256 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
257 upcall_unixctl_enable_megaflows, NULL);
258 unixctl_command_register("upcall/set-flow-limit", "", 1, 1,
259 upcall_unixctl_set_flow_limit, NULL);
260 ovsthread_once_done(&once);
264 udpif->backer = backer;
265 atomic_init(&udpif->flow_limit, MIN(ofproto_flow_limit, 10000));
266 udpif->secret = random_uint32();
267 udpif->reval_seq = seq_create();
268 udpif->dump_seq = seq_create();
269 latch_init(&udpif->exit_latch);
270 list_push_back(&all_udpifs, &udpif->list_node);
271 atomic_init(&udpif->n_flows, 0);
272 atomic_init(&udpif->n_flows_timestamp, LLONG_MIN);
273 ovs_mutex_init(&udpif->n_flows_mutex);
279 udpif_destroy(struct udpif *udpif)
281 udpif_set_threads(udpif, 0, 0);
284 list_remove(&udpif->list_node);
285 latch_destroy(&udpif->exit_latch);
286 seq_destroy(udpif->reval_seq);
287 seq_destroy(udpif->dump_seq);
288 ovs_mutex_destroy(&udpif->n_flows_mutex);
292 /* Tells 'udpif' how many threads it should use to handle upcalls. Disables
293 * all threads if 'n_handlers' and 'n_revalidators' is zero. 'udpif''s
294 * datapath handle must have packet reception enabled before starting threads.
297 udpif_set_threads(struct udpif *udpif, size_t n_handlers,
298 size_t n_revalidators)
302 ovsrcu_quiesce_start();
303 /* Stop the old threads (if any). */
304 if (udpif->handlers &&
305 (udpif->n_handlers != n_handlers
306 || udpif->n_revalidators != n_revalidators)) {
309 latch_set(&udpif->exit_latch);
311 for (i = 0; i < udpif->n_handlers; i++) {
312 struct handler *handler = &udpif->handlers[i];
314 xpthread_join(handler->thread, NULL);
317 for (i = 0; i < udpif->n_revalidators; i++) {
318 struct revalidator *revalidator = &udpif->revalidators[i];
320 ovs_mutex_lock(&revalidator->mutex);
321 xpthread_cond_signal(&revalidator->wake_cond);
322 ovs_mutex_unlock(&revalidator->mutex);
323 xpthread_join(revalidator->thread, NULL);
326 xpthread_join(udpif->flow_dumper, NULL);
328 for (i = 0; i < udpif->n_revalidators; i++) {
329 struct revalidator *revalidator = &udpif->revalidators[i];
330 struct udpif_flow_dump *udump, *next_udump;
332 LIST_FOR_EACH_SAFE (udump, next_udump, list_node,
333 &revalidator->udumps) {
334 list_remove(&udump->list_node);
338 /* Delete ukeys, and delete all flows from the datapath to prevent
339 * double-counting stats. */
340 revalidator_purge(revalidator);
341 hmap_destroy(&revalidator->ukeys);
342 ovs_mutex_destroy(&revalidator->mutex);
344 free(revalidator->name);
347 for (i = 0; i < udpif->n_handlers; i++) {
348 free(udpif->handlers[i].name);
350 latch_poll(&udpif->exit_latch);
352 free(udpif->revalidators);
353 udpif->revalidators = NULL;
354 udpif->n_revalidators = 0;
356 free(udpif->handlers);
357 udpif->handlers = NULL;
358 udpif->n_handlers = 0;
361 error = dpif_handlers_set(udpif->dpif, n_handlers);
363 VLOG_ERR("failed to configure handlers in dpif %s: %s",
364 dpif_name(udpif->dpif), ovs_strerror(error));
368 /* Start new threads (if necessary). */
369 if (!udpif->handlers && n_handlers) {
372 udpif->n_handlers = n_handlers;
373 udpif->n_revalidators = n_revalidators;
375 udpif->handlers = xzalloc(udpif->n_handlers * sizeof *udpif->handlers);
376 for (i = 0; i < udpif->n_handlers; i++) {
377 struct handler *handler = &udpif->handlers[i];
379 handler->udpif = udpif;
380 handler->handler_id = i;
381 xpthread_create(&handler->thread, NULL, udpif_upcall_handler,
385 udpif->revalidators = xzalloc(udpif->n_revalidators
386 * sizeof *udpif->revalidators);
387 for (i = 0; i < udpif->n_revalidators; i++) {
388 struct revalidator *revalidator = &udpif->revalidators[i];
390 revalidator->udpif = udpif;
391 list_init(&revalidator->udumps);
392 hmap_init(&revalidator->ukeys);
393 ovs_mutex_init(&revalidator->mutex);
394 xpthread_cond_init(&revalidator->wake_cond, NULL);
395 xpthread_create(&revalidator->thread, NULL, udpif_revalidator,
398 xpthread_create(&udpif->flow_dumper, NULL, udpif_flow_dumper, udpif);
401 ovsrcu_quiesce_end();
404 /* Waits for all ongoing upcall translations to complete. This ensures that
405 * there are no transient references to any removed ofprotos (or other
406 * objects). In particular, this should be called after an ofproto is removed
407 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
409 udpif_synchronize(struct udpif *udpif)
411 /* This is stronger than necessary. It would be sufficient to ensure
412 * (somehow) that each handler and revalidator thread had passed through
413 * its main loop once. */
414 size_t n_handlers = udpif->n_handlers;
415 size_t n_revalidators = udpif->n_revalidators;
416 udpif_set_threads(udpif, 0, 0);
417 udpif_set_threads(udpif, n_handlers, n_revalidators);
420 /* Notifies 'udpif' that something changed which may render previous
421 * xlate_actions() results invalid. */
423 udpif_revalidate(struct udpif *udpif)
425 seq_change(udpif->reval_seq);
428 /* Returns a seq which increments every time 'udpif' pulls stats from the
429 * datapath. Callers can use this to get a sense of when might be a good time
430 * to do periodic work which relies on relatively up to date statistics. */
432 udpif_dump_seq(struct udpif *udpif)
434 return udpif->dump_seq;
438 udpif_get_memory_usage(struct udpif *udpif, struct simap *usage)
442 simap_increase(usage, "flow_dumpers", 1);
444 simap_increase(usage, "handlers", udpif->n_handlers);
446 simap_increase(usage, "revalidators", udpif->n_revalidators);
447 for (i = 0; i < udpif->n_revalidators; i++) {
448 struct revalidator *revalidator = &udpif->revalidators[i];
449 ovs_mutex_lock(&revalidator->mutex);
450 simap_increase(usage, "revalidator dumps", revalidator->n_udumps);
452 /* XXX: This isn't technically thread safe because the revalidator
453 * ukeys maps isn't protected by a mutex since it's per thread. */
454 simap_increase(usage, "revalidator keys",
455 hmap_count(&revalidator->ukeys));
456 ovs_mutex_unlock(&revalidator->mutex);
460 /* Remove flows from a single datapath. */
462 udpif_flush(struct udpif *udpif)
464 size_t n_handlers, n_revalidators;
466 n_handlers = udpif->n_handlers;
467 n_revalidators = udpif->n_revalidators;
469 udpif_set_threads(udpif, 0, 0);
470 dpif_flow_flush(udpif->dpif);
471 udpif_set_threads(udpif, n_handlers, n_revalidators);
474 /* Removes all flows from all datapaths. */
476 udpif_flush_all_datapaths(void)
480 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
487 udpif_get_n_flows(struct udpif *udpif)
489 long long int time, now;
493 atomic_read(&udpif->n_flows_timestamp, &time);
494 if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) {
495 struct dpif_dp_stats stats;
497 atomic_store(&udpif->n_flows_timestamp, now);
498 dpif_get_dp_stats(udpif->dpif, &stats);
499 flow_count = stats.n_flows;
500 atomic_store(&udpif->n_flows, flow_count);
501 ovs_mutex_unlock(&udpif->n_flows_mutex);
503 atomic_read(&udpif->n_flows, &flow_count);
509 udpif_flow_dumper(void *arg)
511 struct udpif *udpif = arg;
513 set_subprogram_name("flow_dumper");
514 while (!latch_is_set(&udpif->exit_latch)) {
515 const struct dpif_flow_stats *stats;
516 long long int start_time, duration;
517 const struct nlattr *key, *mask;
518 struct dpif_flow_dump dump;
519 size_t key_len, mask_len;
520 unsigned int flow_limit;
521 bool need_revalidate;
527 reval_seq = seq_read(udpif->reval_seq);
528 need_revalidate = udpif->last_reval_seq != reval_seq;
529 udpif->last_reval_seq = reval_seq;
531 n_flows = udpif_get_n_flows(udpif);
532 udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows);
533 udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2;
535 start_time = time_msec();
536 error = dpif_flow_dump_start(&dump, udpif->dpif);
538 VLOG_INFO("Failed to start flow dump (%s)", ovs_strerror(error));
541 dpif_flow_dump_state_init(udpif->dpif, &state);
542 while (dpif_flow_dump_next(&dump, state, &key, &key_len,
543 &mask, &mask_len, NULL, NULL, &stats)
544 && !latch_is_set(&udpif->exit_latch)) {
545 struct udpif_flow_dump *udump = xmalloc(sizeof *udump);
546 struct revalidator *revalidator;
548 udump->key_hash = hash_bytes(key, key_len, udpif->secret);
549 memcpy(&udump->key_buf, key, key_len);
550 udump->key = (struct nlattr *) &udump->key_buf;
551 udump->key_len = key_len;
553 memcpy(&udump->mask_buf, mask, mask_len);
554 udump->mask = (struct nlattr *) &udump->mask_buf;
555 udump->mask_len = mask_len;
557 udump->stats = *stats;
558 udump->need_revalidate = need_revalidate;
560 revalidator = &udpif->revalidators[udump->key_hash
561 % udpif->n_revalidators];
563 ovs_mutex_lock(&revalidator->mutex);
564 while (revalidator->n_udumps >= REVALIDATE_MAX_BATCH * 3
565 && !latch_is_set(&udpif->exit_latch)) {
566 ovs_mutex_cond_wait(&revalidator->wake_cond,
567 &revalidator->mutex);
569 list_push_back(&revalidator->udumps, &udump->list_node);
570 revalidator->n_udumps++;
571 xpthread_cond_signal(&revalidator->wake_cond);
572 ovs_mutex_unlock(&revalidator->mutex);
574 dpif_flow_dump_state_uninit(udpif->dpif, state);
575 dpif_flow_dump_done(&dump);
577 /* Let all the revalidators finish and garbage collect. */
578 seq_change(udpif->dump_seq);
579 for (i = 0; i < udpif->n_revalidators; i++) {
580 struct revalidator *revalidator = &udpif->revalidators[i];
581 ovs_mutex_lock(&revalidator->mutex);
582 xpthread_cond_signal(&revalidator->wake_cond);
583 ovs_mutex_unlock(&revalidator->mutex);
586 for (i = 0; i < udpif->n_revalidators; i++) {
587 struct revalidator *revalidator = &udpif->revalidators[i];
589 ovs_mutex_lock(&revalidator->mutex);
590 while (revalidator->dump_seq != seq_read(udpif->dump_seq)
591 && !latch_is_set(&udpif->exit_latch)) {
592 ovs_mutex_cond_wait(&revalidator->wake_cond,
593 &revalidator->mutex);
595 ovs_mutex_unlock(&revalidator->mutex);
598 duration = MAX(time_msec() - start_time, 1);
599 udpif->dump_duration = duration;
600 atomic_read(&udpif->flow_limit, &flow_limit);
601 if (duration > 2000) {
602 flow_limit /= duration / 1000;
603 } else if (duration > 1300) {
604 flow_limit = flow_limit * 3 / 4;
605 } else if (duration < 1000 && n_flows > 2000
606 && flow_limit < n_flows * 1000 / duration) {
609 flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000));
610 atomic_store(&udpif->flow_limit, flow_limit);
612 if (duration > 2000) {
613 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
618 poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500));
619 seq_wait(udpif->reval_seq, udpif->last_reval_seq);
620 latch_wait(&udpif->exit_latch);
627 /* The upcall handler thread tries to read a batch of FLOW_MISS_MAX_BATCH
628 * upcalls from dpif, processes the batch and installs corresponding flows
631 udpif_upcall_handler(void *arg)
633 struct handler *handler = arg;
634 struct udpif *udpif = handler->udpif;
635 struct hmap misses = HMAP_INITIALIZER(&misses);
637 handler->name = xasprintf("handler_%u", ovsthread_id_self());
638 set_subprogram_name("%s", handler->name);
640 while (!latch_is_set(&handler->udpif->exit_latch)) {
641 struct upcall upcalls[FLOW_MISS_MAX_BATCH];
642 struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH];
643 struct flow_miss *miss;
646 n_upcalls = read_upcalls(handler, upcalls, miss_buf, &misses);
648 dpif_recv_wait(udpif->dpif, handler->handler_id);
649 latch_wait(&udpif->exit_latch);
652 handle_upcalls(handler, &misses, upcalls, n_upcalls);
654 HMAP_FOR_EACH (miss, hmap_node, &misses) {
655 xlate_out_uninit(&miss->xout);
658 for (i = 0; i < n_upcalls; i++) {
659 ofpbuf_uninit(&upcalls[i].dpif_upcall.packet);
660 ofpbuf_uninit(&upcalls[i].upcall_buf);
665 hmap_destroy(&misses);
671 udpif_revalidator(void *arg)
673 struct revalidator *revalidator = arg;
675 revalidator->name = xasprintf("revalidator_%u", ovsthread_id_self());
676 set_subprogram_name("%s", revalidator->name);
678 struct list udumps = LIST_INITIALIZER(&udumps);
679 struct udpif *udpif = revalidator->udpif;
682 ovs_mutex_lock(&revalidator->mutex);
683 if (latch_is_set(&udpif->exit_latch)) {
684 ovs_mutex_unlock(&revalidator->mutex);
688 if (!revalidator->n_udumps) {
689 if (revalidator->dump_seq != seq_read(udpif->dump_seq)) {
690 revalidator->dump_seq = seq_read(udpif->dump_seq);
691 revalidator_sweep(revalidator);
693 ovs_mutex_cond_wait(&revalidator->wake_cond,
694 &revalidator->mutex);
698 for (i = 0; i < REVALIDATE_MAX_BATCH && revalidator->n_udumps; i++) {
699 list_push_back(&udumps, list_pop_front(&revalidator->udumps));
700 revalidator->n_udumps--;
703 /* Wake up the flow dumper. */
704 xpthread_cond_signal(&revalidator->wake_cond);
705 ovs_mutex_unlock(&revalidator->mutex);
707 if (!list_is_empty(&udumps)) {
708 revalidate_udumps(revalidator, &udumps);
715 static enum upcall_type
716 classify_upcall(const struct upcall *upcall)
718 const struct dpif_upcall *dpif_upcall = &upcall->dpif_upcall;
719 union user_action_cookie cookie;
722 /* First look at the upcall type. */
723 switch (dpif_upcall->type) {
730 case DPIF_N_UC_TYPES:
732 VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32,
737 /* "action" upcalls need a closer look. */
738 if (!dpif_upcall->userdata) {
739 VLOG_WARN_RL(&rl, "action upcall missing cookie");
742 userdata_len = nl_attr_get_size(dpif_upcall->userdata);
743 if (userdata_len < sizeof cookie.type
744 || userdata_len > sizeof cookie) {
745 VLOG_WARN_RL(&rl, "action upcall cookie has unexpected size %"PRIuSIZE,
749 memset(&cookie, 0, sizeof cookie);
750 memcpy(&cookie, nl_attr_get(dpif_upcall->userdata), userdata_len);
751 if (userdata_len == MAX(8, sizeof cookie.sflow)
752 && cookie.type == USER_ACTION_COOKIE_SFLOW) {
754 } else if (userdata_len == MAX(8, sizeof cookie.slow_path)
755 && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) {
757 } else if (userdata_len == MAX(8, sizeof cookie.flow_sample)
758 && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) {
759 return FLOW_SAMPLE_UPCALL;
760 } else if (userdata_len == MAX(8, sizeof cookie.ipfix)
761 && cookie.type == USER_ACTION_COOKIE_IPFIX) {
764 VLOG_WARN_RL(&rl, "invalid user cookie of type %"PRIu16
765 " and size %"PRIuSIZE, cookie.type, userdata_len);
770 /* Calculates slow path actions for 'xout'. 'buf' must statically be
771 * initialized with at least 128 bytes of space. */
773 compose_slow_path(struct udpif *udpif, struct xlate_out *xout,
774 struct flow *flow, odp_port_t odp_in_port,
777 union user_action_cookie cookie;
781 cookie.type = USER_ACTION_COOKIE_SLOW_PATH;
782 cookie.slow_path.unused = 0;
783 cookie.slow_path.reason = xout->slow;
785 port = xout->slow & (SLOW_CFM | SLOW_BFD | SLOW_LACP | SLOW_STP)
788 pid = dpif_port_get_pid(udpif->dpif, port, flow_hash_5tuple(flow, 0));
789 odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, buf);
792 static struct flow_miss *
793 flow_miss_find(struct hmap *todo, const struct ofproto_dpif *ofproto,
794 const struct flow *flow, uint32_t hash)
796 struct flow_miss *miss;
798 HMAP_FOR_EACH_WITH_HASH (miss, hmap_node, hash, todo) {
799 if (miss->ofproto == ofproto && flow_equal(&miss->flow, flow)) {
807 /* Reads and classifies upcalls. Returns the number of upcalls successfully
810 read_upcalls(struct handler *handler,
811 struct upcall upcalls[FLOW_MISS_MAX_BATCH],
812 struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH],
815 struct udpif *udpif = handler->udpif;
818 size_t n_upcalls = 0;
821 * Try reading FLOW_MISS_MAX_BATCH upcalls from dpif.
823 * Extract the flow from each upcall. Construct in 'misses' a hash table
824 * that maps each unique flow to a 'struct flow_miss'.
826 * Most commonly there is a single packet per flow_miss, but there are
827 * several reasons why there might be more than one, e.g.:
829 * - The dpif packet interface does not support TSO (or UFO, etc.), so a
830 * large packet sent to userspace is split into a sequence of smaller
833 * - A stream of quickly arriving packets in an established "slow-pathed"
836 * - Rarely, a stream of quickly arriving packets in a flow not yet
837 * established. (This is rare because most protocols do not send
838 * multiple back-to-back packets before receiving a reply from the
839 * other end of the connection, which gives OVS a chance to set up a
842 for (i = 0; i < FLOW_MISS_MAX_BATCH; i++) {
843 struct upcall *upcall = &upcalls[n_upcalls];
844 struct flow_miss *miss = &miss_buf[n_misses];
845 struct dpif_upcall *dupcall;
846 struct ofpbuf *packet;
847 struct flow_miss *existing_miss;
848 struct ofproto_dpif *ofproto;
849 struct dpif_sflow *sflow;
850 struct dpif_ipfix *ipfix;
852 enum upcall_type type;
853 odp_port_t odp_in_port;
856 ofpbuf_use_stub(&upcall->upcall_buf, upcall->upcall_stub,
857 sizeof upcall->upcall_stub);
858 error = dpif_recv(udpif->dpif, handler->handler_id,
859 &upcall->dpif_upcall, &upcall->upcall_buf);
861 ofpbuf_uninit(&upcall->upcall_buf);
865 dupcall = &upcall->dpif_upcall;
866 packet = &dupcall->packet;
867 error = xlate_receive(udpif->backer, packet, dupcall->key,
868 dupcall->key_len, &flow,
869 &ofproto, &ipfix, &sflow, NULL, &odp_in_port);
871 if (error == ENODEV) {
872 /* Received packet on datapath port for which we couldn't
873 * associate an ofproto. This can happen if a port is removed
874 * while traffic is being received. Print a rate-limited
875 * message in case it happens frequently. Install a drop flow
876 * so that future packets of the flow are inexpensively dropped
878 VLOG_INFO_RL(&rl, "received packet on unassociated datapath "
879 "port %"PRIu32, odp_in_port);
880 dpif_flow_put(udpif->dpif, DPIF_FP_CREATE | DPIF_FP_MODIFY,
881 dupcall->key, dupcall->key_len, NULL, 0, NULL, 0,
887 type = classify_upcall(upcall);
888 if (type == MISS_UPCALL) {
890 struct pkt_metadata md = pkt_metadata_from_flow(&flow);
892 flow_extract(packet, &md, &miss->flow);
893 hash = flow_hash(&miss->flow, 0);
894 existing_miss = flow_miss_find(misses, ofproto, &miss->flow,
896 if (!existing_miss) {
897 hmap_insert(misses, &miss->hmap_node, hash);
898 miss->ofproto = ofproto;
899 miss->key = dupcall->key;
900 miss->key_len = dupcall->key_len;
901 miss->upcall_type = dupcall->type;
902 miss->stats.n_packets = 0;
903 miss->stats.n_bytes = 0;
904 miss->stats.used = time_msec();
905 miss->stats.tcp_flags = 0;
906 miss->odp_in_port = odp_in_port;
910 miss = existing_miss;
912 miss->stats.tcp_flags |= ntohs(miss->flow.tcp_flags);
913 miss->stats.n_bytes += ofpbuf_size(packet);
914 miss->stats.n_packets++;
916 upcall->flow_miss = miss;
924 union user_action_cookie cookie;
926 memset(&cookie, 0, sizeof cookie);
927 memcpy(&cookie, nl_attr_get(dupcall->userdata),
928 sizeof cookie.sflow);
929 dpif_sflow_received(sflow, packet, &flow, odp_in_port,
935 dpif_ipfix_bridge_sample(ipfix, packet, &flow);
938 case FLOW_SAMPLE_UPCALL:
940 union user_action_cookie cookie;
942 memset(&cookie, 0, sizeof cookie);
943 memcpy(&cookie, nl_attr_get(dupcall->userdata),
944 sizeof cookie.flow_sample);
946 /* The flow reflects exactly the contents of the packet.
947 * Sample the packet using it. */
948 dpif_ipfix_flow_sample(ipfix, packet, &flow,
949 cookie.flow_sample.collector_set_id,
950 cookie.flow_sample.probability,
951 cookie.flow_sample.obs_domain_id,
952 cookie.flow_sample.obs_point_id);
961 dpif_ipfix_unref(ipfix);
962 dpif_sflow_unref(sflow);
965 ofpbuf_uninit(&upcall->dpif_upcall.packet);
966 ofpbuf_uninit(&upcall->upcall_buf);
973 handle_upcalls(struct handler *handler, struct hmap *misses,
974 struct upcall *upcalls, size_t n_upcalls)
976 struct udpif *udpif = handler->udpif;
977 struct dpif_op *opsp[FLOW_MISS_MAX_BATCH * 2];
978 struct dpif_op ops[FLOW_MISS_MAX_BATCH * 2];
979 struct flow_miss *miss;
981 unsigned int flow_limit;
982 bool fail_open, may_put;
984 atomic_read(&udpif->flow_limit, &flow_limit);
985 may_put = udpif_get_n_flows(udpif) < flow_limit;
987 /* Initialize each 'struct flow_miss's ->xout.
989 * We do this per-flow_miss rather than per-packet because, most commonly,
990 * all the packets in a flow can use the same translation.
992 * We can't do this in the previous loop because we need the TCP flags for
993 * all the packets in each miss. */
995 HMAP_FOR_EACH (miss, hmap_node, misses) {
998 xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL,
999 miss->stats.tcp_flags, NULL);
1000 xin.may_learn = true;
1002 if (miss->upcall_type == DPIF_UC_MISS) {
1003 xin.resubmit_stats = &miss->stats;
1005 /* For non-miss upcalls, there's a flow in the datapath which this
1006 * packet was accounted to. Presumably the revalidators will deal
1007 * with pushing its stats eventually. */
1010 xlate_actions(&xin, &miss->xout);
1011 fail_open = fail_open || miss->xout.fail_open;
1014 /* Now handle the packets individually in order of arrival. In the common
1015 * case each packet of a miss can share the same actions, but slow-pathed
1016 * packets need to be translated individually:
1018 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1019 * processes received packets for these protocols.
1021 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1024 * The loop fills 'ops' with an array of operations to execute in the
1027 for (i = 0; i < n_upcalls; i++) {
1028 struct upcall *upcall = &upcalls[i];
1029 struct flow_miss *miss = upcall->flow_miss;
1030 struct ofpbuf *packet = &upcall->dpif_upcall.packet;
1032 ovs_be16 flow_vlan_tci;
1034 /* Save a copy of flow.vlan_tci in case it is changed to
1035 * generate proper mega flow masks for VLAN splinter flows. */
1036 flow_vlan_tci = miss->flow.vlan_tci;
1038 if (miss->xout.slow) {
1039 struct xlate_in xin;
1041 xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL, 0, packet);
1042 xlate_actions_for_side_effects(&xin);
1045 if (miss->flow.in_port.ofp_port
1046 != vsp_realdev_to_vlandev(miss->ofproto,
1047 miss->flow.in_port.ofp_port,
1048 miss->flow.vlan_tci)) {
1049 /* This packet was received on a VLAN splinter port. We
1050 * added a VLAN to the packet to make the packet resemble
1051 * the flow, but the actions were composed assuming that
1052 * the packet contained no VLAN. So, we must remove the
1053 * VLAN header from the packet before trying to execute the
1055 if (ofpbuf_size(&miss->xout.odp_actions)) {
1056 eth_pop_vlan(packet);
1059 /* Remove the flow vlan tags inserted by vlan splinter logic
1060 * to ensure megaflow masks generated match the data path flow. */
1061 miss->flow.vlan_tci = 0;
1064 /* Do not install a flow into the datapath if:
1066 * - The datapath already has too many flows.
1068 * - An earlier iteration of this loop already put the same flow.
1070 * - We received this packet via some flow installed in the kernel
1074 && upcall->dpif_upcall.type == DPIF_UC_MISS) {
1080 atomic_read(&enable_megaflows, &megaflow);
1081 ofpbuf_use_stack(&mask, &miss->mask_buf, sizeof miss->mask_buf);
1085 max_mpls = ofproto_dpif_get_max_mpls_depth(miss->ofproto);
1086 odp_flow_key_from_mask(&mask, &miss->xout.wc.masks,
1087 &miss->flow, UINT32_MAX, max_mpls);
1091 op->type = DPIF_OP_FLOW_PUT;
1092 op->u.flow_put.flags = DPIF_FP_CREATE | DPIF_FP_MODIFY;
1093 op->u.flow_put.key = miss->key;
1094 op->u.flow_put.key_len = miss->key_len;
1095 op->u.flow_put.mask = ofpbuf_data(&mask);
1096 op->u.flow_put.mask_len = ofpbuf_size(&mask);
1097 op->u.flow_put.stats = NULL;
1099 if (!miss->xout.slow) {
1100 op->u.flow_put.actions = ofpbuf_data(&miss->xout.odp_actions);
1101 op->u.flow_put.actions_len = ofpbuf_size(&miss->xout.odp_actions);
1105 ofpbuf_use_stack(&buf, miss->slow_path_buf,
1106 sizeof miss->slow_path_buf);
1107 compose_slow_path(udpif, &miss->xout, &miss->flow,
1108 miss->odp_in_port, &buf);
1109 op->u.flow_put.actions = ofpbuf_data(&buf);
1110 op->u.flow_put.actions_len = ofpbuf_size(&buf);
1115 * The 'miss' may be shared by multiple upcalls. Restore
1116 * the saved flow vlan_tci field before processing the next
1118 miss->flow.vlan_tci = flow_vlan_tci;
1120 if (ofpbuf_size(&miss->xout.odp_actions)) {
1123 op->type = DPIF_OP_EXECUTE;
1124 op->u.execute.packet = packet;
1125 odp_key_to_pkt_metadata(miss->key, miss->key_len,
1127 op->u.execute.actions = ofpbuf_data(&miss->xout.odp_actions);
1128 op->u.execute.actions_len = ofpbuf_size(&miss->xout.odp_actions);
1129 op->u.execute.needs_help = (miss->xout.slow & SLOW_ACTION) != 0;
1133 /* Special case for fail-open mode.
1135 * If we are in fail-open mode, but we are connected to a controller too,
1136 * then we should send the packet up to the controller in the hope that it
1137 * will try to set up a flow and thereby allow us to exit fail-open.
1139 * See the top-level comment in fail-open.c for more information.
1141 * Copy packets before they are modified by execution. */
1143 for (i = 0; i < n_upcalls; i++) {
1144 struct upcall *upcall = &upcalls[i];
1145 struct flow_miss *miss = upcall->flow_miss;
1146 struct ofpbuf *packet = &upcall->dpif_upcall.packet;
1147 struct ofproto_packet_in *pin;
1149 pin = xmalloc(sizeof *pin);
1150 pin->up.packet = xmemdup(ofpbuf_data(packet), ofpbuf_size(packet));
1151 pin->up.packet_len = ofpbuf_size(packet);
1152 pin->up.reason = OFPR_NO_MATCH;
1153 pin->up.table_id = 0;
1154 pin->up.cookie = OVS_BE64_MAX;
1155 flow_get_metadata(&miss->flow, &pin->up.fmd);
1156 pin->send_len = 0; /* Not used for flow table misses. */
1157 pin->miss_type = OFPROTO_PACKET_IN_NO_MISS;
1158 ofproto_dpif_send_packet_in(miss->ofproto, pin);
1162 /* Execute batch. */
1163 for (i = 0; i < n_ops; i++) {
1166 dpif_operate(udpif->dpif, opsp, n_ops);
1169 static struct udpif_key *
1170 ukey_lookup(struct revalidator *revalidator, struct udpif_flow_dump *udump)
1172 struct udpif_key *ukey;
1174 HMAP_FOR_EACH_WITH_HASH (ukey, hmap_node, udump->key_hash,
1175 &revalidator->ukeys) {
1176 if (ukey->key_len == udump->key_len
1177 && !memcmp(ukey->key, udump->key, udump->key_len)) {
1184 static struct udpif_key *
1185 ukey_create(const struct nlattr *key, size_t key_len, long long int used)
1187 struct udpif_key *ukey = xmalloc(sizeof *ukey);
1189 ukey->key = (struct nlattr *) &ukey->key_buf;
1190 memcpy(&ukey->key_buf, key, key_len);
1191 ukey->key_len = key_len;
1194 ukey->created = used ? used : time_msec();
1195 memset(&ukey->stats, 0, sizeof ukey->stats);
1196 ukey->xcache = NULL;
1202 ukey_delete(struct revalidator *revalidator, struct udpif_key *ukey)
1204 hmap_remove(&revalidator->ukeys, &ukey->hmap_node);
1205 xlate_cache_delete(ukey->xcache);
1210 should_revalidate(uint64_t packets, long long int used)
1212 long long int metric, now, duration;
1214 /* Calculate the mean time between seeing these packets. If this
1215 * exceeds the threshold, then delete the flow rather than performing
1216 * costly revalidation for flows that aren't being hit frequently.
1218 * This is targeted at situations where the dump_duration is high (~1s),
1219 * and revalidation is triggered by a call to udpif_revalidate(). In
1220 * these situations, revalidation of all flows causes fluctuations in the
1221 * flow_limit due to the interaction with the dump_duration and max_idle.
1222 * This tends to result in deletion of low-throughput flows anyway, so
1223 * skip the revalidation and just delete those flows. */
1224 packets = MAX(packets, 1);
1225 now = MAX(used, time_msec());
1226 duration = now - used;
1227 metric = duration / packets;
1236 revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump,
1237 struct udpif_key *ukey)
1239 struct ofpbuf xout_actions, *actions;
1240 uint64_t slow_path_buf[128 / 8];
1241 struct xlate_out xout, *xoutp;
1242 struct netflow *netflow;
1243 struct flow flow, udump_mask;
1244 struct ofproto_dpif *ofproto;
1245 struct dpif_flow_stats push;
1246 uint32_t *udump32, *xout32;
1247 odp_port_t odp_in_port;
1248 struct xlate_in xin;
1249 long long int last_used;
1258 may_learn = push.n_packets > 0;
1260 /* If we don't need to revalidate, we can simply push the stats contained
1261 * in the udump, otherwise we'll have to get the actions so we can check
1263 if (udump->need_revalidate) {
1264 if (dpif_flow_get(udpif->dpif, ukey->key, ukey->key_len, &actions,
1270 last_used = ukey->stats.used;
1271 push.used = udump->stats.used;
1272 push.tcp_flags = udump->stats.tcp_flags;
1273 push.n_packets = udump->stats.n_packets > ukey->stats.n_packets
1274 ? udump->stats.n_packets - ukey->stats.n_packets
1276 push.n_bytes = udump->stats.n_bytes > ukey->stats.n_bytes
1277 ? udump->stats.n_bytes - ukey->stats.n_bytes
1279 ukey->stats = udump->stats;
1281 if (udump->need_revalidate && last_used
1282 && !should_revalidate(push.n_packets, last_used)) {
1287 if (!push.n_packets && !udump->need_revalidate) {
1292 if (ukey->xcache && !udump->need_revalidate) {
1293 xlate_push_stats(ukey->xcache, may_learn, &push);
1298 error = xlate_receive(udpif->backer, NULL, ukey->key, ukey->key_len, &flow,
1299 &ofproto, NULL, NULL, &netflow, &odp_in_port);
1304 if (udump->need_revalidate) {
1305 xlate_cache_clear(ukey->xcache);
1307 if (!ukey->xcache) {
1308 ukey->xcache = xlate_cache_new();
1311 xlate_in_init(&xin, ofproto, &flow, NULL, push.tcp_flags, NULL);
1312 xin.resubmit_stats = push.n_packets ? &push : NULL;
1313 xin.xcache = ukey->xcache;
1314 xin.may_learn = may_learn;
1315 xin.skip_wildcards = !udump->need_revalidate;
1316 xlate_actions(&xin, &xout);
1319 if (!udump->need_revalidate) {
1325 ofpbuf_use_const(&xout_actions, ofpbuf_data(&xout.odp_actions),
1326 ofpbuf_size(&xout.odp_actions));
1328 ofpbuf_use_stack(&xout_actions, slow_path_buf, sizeof slow_path_buf);
1329 compose_slow_path(udpif, &xout, &flow, odp_in_port, &xout_actions);
1332 if (!ofpbuf_equal(&xout_actions, actions)) {
1336 if (odp_flow_key_to_mask(udump->mask, udump->mask_len, &udump_mask, &flow)
1341 /* Since the kernel is free to ignore wildcarded bits in the mask, we can't
1342 * directly check that the masks are the same. Instead we check that the
1343 * mask in the kernel is more specific i.e. less wildcarded, than what
1344 * we've calculated here. This guarantees we don't catch any packets we
1345 * shouldn't with the megaflow. */
1346 udump32 = (uint32_t *) &udump_mask;
1347 xout32 = (uint32_t *) &xout.wc.masks;
1348 for (i = 0; i < FLOW_U32S; i++) {
1349 if ((udump32[i] | xout32[i]) != udump32[i]) {
1358 netflow_expire(netflow, &flow);
1359 netflow_flow_clear(netflow, &flow);
1361 netflow_unref(netflow);
1363 ofpbuf_delete(actions);
1364 xlate_out_uninit(xoutp);
1369 struct udpif_key *ukey;
1370 struct udpif_flow_dump *udump;
1371 struct dpif_flow_stats stats; /* Stats for 'op'. */
1372 struct dpif_op op; /* Flow del operation. */
1376 dump_op_init(struct dump_op *op, const struct nlattr *key, size_t key_len,
1377 struct udpif_key *ukey, struct udpif_flow_dump *udump)
1381 op->op.type = DPIF_OP_FLOW_DEL;
1382 op->op.u.flow_del.key = key;
1383 op->op.u.flow_del.key_len = key_len;
1384 op->op.u.flow_del.stats = &op->stats;
1388 push_dump_ops(struct revalidator *revalidator,
1389 struct dump_op *ops, size_t n_ops)
1391 struct udpif *udpif = revalidator->udpif;
1392 struct dpif_op *opsp[REVALIDATE_MAX_BATCH];
1395 ovs_assert(n_ops <= REVALIDATE_MAX_BATCH);
1396 for (i = 0; i < n_ops; i++) {
1397 opsp[i] = &ops[i].op;
1399 dpif_operate(udpif->dpif, opsp, n_ops);
1401 for (i = 0; i < n_ops; i++) {
1402 struct dump_op *op = &ops[i];
1403 struct dpif_flow_stats *push, *stats, push_buf;
1405 stats = op->op.u.flow_del.stats;
1408 push->used = MAX(stats->used, op->ukey->stats.used);
1409 push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags;
1410 push->n_packets = stats->n_packets - op->ukey->stats.n_packets;
1411 push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes;
1416 if (push->n_packets || netflow_exists()) {
1417 struct ofproto_dpif *ofproto;
1418 struct netflow *netflow;
1422 may_learn = push->n_packets > 0;
1423 if (op->ukey && op->ukey->xcache) {
1424 xlate_push_stats(op->ukey->xcache, may_learn, push);
1428 if (!xlate_receive(udpif->backer, NULL, op->op.u.flow_del.key,
1429 op->op.u.flow_del.key_len, &flow, &ofproto,
1430 NULL, NULL, &netflow, NULL)) {
1431 struct xlate_in xin;
1433 xlate_in_init(&xin, ofproto, &flow, NULL, push->tcp_flags,
1435 xin.resubmit_stats = push->n_packets ? push : NULL;
1436 xin.may_learn = may_learn;
1437 xin.skip_wildcards = true;
1438 xlate_actions_for_side_effects(&xin);
1441 netflow_expire(netflow, &flow);
1442 netflow_flow_clear(netflow, &flow);
1443 netflow_unref(netflow);
1449 for (i = 0; i < n_ops; i++) {
1450 struct udpif_key *ukey;
1452 /* If there's a udump, this ukey came directly from a datapath flow
1453 * dump. Sometimes a datapath can send duplicates in flow dumps, in
1454 * which case we wouldn't want to double-free a ukey, so avoid that by
1455 * looking up the ukey again.
1457 * If there's no udump then we know what we're doing. */
1458 ukey = (ops[i].udump
1459 ? ukey_lookup(revalidator, ops[i].udump)
1462 ukey_delete(revalidator, ukey);
1468 revalidate_udumps(struct revalidator *revalidator, struct list *udumps)
1470 struct udpif *udpif = revalidator->udpif;
1472 struct dump_op ops[REVALIDATE_MAX_BATCH];
1473 struct udpif_flow_dump *udump, *next_udump;
1474 size_t n_ops, n_flows;
1475 unsigned int flow_limit;
1476 long long int max_idle;
1479 atomic_read(&udpif->flow_limit, &flow_limit);
1481 n_flows = udpif_get_n_flows(udpif);
1484 max_idle = ofproto_max_idle;
1485 if (n_flows > flow_limit) {
1486 must_del = n_flows > 2 * flow_limit;
1491 LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) {
1492 long long int used, now;
1493 struct udpif_key *ukey;
1496 ukey = ukey_lookup(revalidator, udump);
1498 used = udump->stats.used;
1499 if (!used && ukey) {
1500 used = ukey->created;
1503 if (must_del || (used && used < now - max_idle)) {
1504 struct dump_op *dop = &ops[n_ops++];
1506 dump_op_init(dop, udump->key, udump->key_len, ukey, udump);
1511 ukey = ukey_create(udump->key, udump->key_len, used);
1512 hmap_insert(&revalidator->ukeys, &ukey->hmap_node,
1517 if (!revalidate_ukey(udpif, udump, ukey)) {
1518 dpif_flow_del(udpif->dpif, udump->key, udump->key_len, NULL);
1519 ukey_delete(revalidator, ukey);
1522 list_remove(&udump->list_node);
1526 push_dump_ops(revalidator, ops, n_ops);
1528 LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) {
1529 list_remove(&udump->list_node);
1535 revalidator_sweep__(struct revalidator *revalidator, bool purge)
1537 struct dump_op ops[REVALIDATE_MAX_BATCH];
1538 struct udpif_key *ukey, *next;
1543 HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, &revalidator->ukeys) {
1544 if (!purge && ukey->mark) {
1547 struct dump_op *op = &ops[n_ops++];
1549 /* If we have previously seen a flow in the datapath, but didn't
1550 * see it during the most recent dump, delete it. This allows us
1551 * to clean up the ukey and keep the statistics consistent. */
1552 dump_op_init(op, ukey->key, ukey->key_len, ukey, NULL);
1553 if (n_ops == REVALIDATE_MAX_BATCH) {
1554 push_dump_ops(revalidator, ops, n_ops);
1561 push_dump_ops(revalidator, ops, n_ops);
1566 revalidator_sweep(struct revalidator *revalidator)
1568 revalidator_sweep__(revalidator, false);
1572 revalidator_purge(struct revalidator *revalidator)
1574 revalidator_sweep__(revalidator, true);
1578 upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED,
1579 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
1581 struct ds ds = DS_EMPTY_INITIALIZER;
1582 struct udpif *udpif;
1584 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1585 unsigned int flow_limit;
1588 atomic_read(&udpif->flow_limit, &flow_limit);
1590 ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif));
1591 ds_put_format(&ds, "\tflows : (current %"PRIu64")"
1592 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif),
1593 udpif->avg_n_flows, udpif->max_n_flows, flow_limit);
1594 ds_put_format(&ds, "\tdump duration : %lldms\n", udpif->dump_duration);
1596 ds_put_char(&ds, '\n');
1597 for (i = 0; i < n_revalidators; i++) {
1598 struct revalidator *revalidator = &udpif->revalidators[i];
1600 /* XXX: The result of hmap_count(&revalidator->ukeys) may not be
1601 * accurate because it's not protected by the revalidator mutex. */
1602 ovs_mutex_lock(&revalidator->mutex);
1603 ds_put_format(&ds, "\t%s: (dump queue %"PRIuSIZE") (keys %"PRIuSIZE
1604 ")\n", revalidator->name, revalidator->n_udumps,
1605 hmap_count(&revalidator->ukeys));
1606 ovs_mutex_unlock(&revalidator->mutex);
1610 unixctl_command_reply(conn, ds_cstr(&ds));
1614 /* Disable using the megaflows.
1616 * This command is only needed for advanced debugging, so it's not
1617 * documented in the man page. */
1619 upcall_unixctl_disable_megaflows(struct unixctl_conn *conn,
1620 int argc OVS_UNUSED,
1621 const char *argv[] OVS_UNUSED,
1622 void *aux OVS_UNUSED)
1624 atomic_store(&enable_megaflows, false);
1625 udpif_flush_all_datapaths();
1626 unixctl_command_reply(conn, "megaflows disabled");
1629 /* Re-enable using megaflows.
1631 * This command is only needed for advanced debugging, so it's not
1632 * documented in the man page. */
1634 upcall_unixctl_enable_megaflows(struct unixctl_conn *conn,
1635 int argc OVS_UNUSED,
1636 const char *argv[] OVS_UNUSED,
1637 void *aux OVS_UNUSED)
1639 atomic_store(&enable_megaflows, true);
1640 udpif_flush_all_datapaths();
1641 unixctl_command_reply(conn, "megaflows enabled");
1644 /* Set the flow limit.
1646 * This command is only needed for advanced debugging, so it's not
1647 * documented in the man page. */
1649 upcall_unixctl_set_flow_limit(struct unixctl_conn *conn,
1650 int argc OVS_UNUSED,
1651 const char *argv[] OVS_UNUSED,
1652 void *aux OVS_UNUSED)
1654 struct ds ds = DS_EMPTY_INITIALIZER;
1655 struct udpif *udpif;
1656 unsigned int flow_limit = atoi(argv[1]);
1658 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
1659 atomic_store(&udpif->flow_limit, flow_limit);
1661 ds_put_format(&ds, "set flow_limit to %u\n", flow_limit);
1662 unixctl_command_reply(conn, ds_cstr(&ds));