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"
26 #include "dynamic-string.h"
27 #include "fail-open.h"
28 #include "guarded-list.h"
33 #include "ofproto-dpif-ipfix.h"
34 #include "ofproto-dpif-sflow.h"
35 #include "ofproto-dpif-xlate.h"
38 #include "poll-loop.h"
41 #include "openvswitch/vlog.h"
43 #define MAX_QUEUE_LENGTH 512
44 #define UPCALL_MAX_BATCH 64
45 #define REVALIDATE_MAX_BATCH 50
47 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall);
49 COVERAGE_DEFINE(dumped_duplicate_flow);
50 COVERAGE_DEFINE(dumped_new_flow);
51 COVERAGE_DEFINE(handler_duplicate_upcall);
52 COVERAGE_DEFINE(upcall_ukey_contention);
53 COVERAGE_DEFINE(revalidate_missed_dp_flow);
55 /* A thread that reads upcalls from dpif, forwards each upcall's packet,
56 * and possibly sets up a kernel flow as a cache. */
58 struct udpif *udpif; /* Parent udpif. */
59 pthread_t thread; /* Thread ID. */
60 uint32_t handler_id; /* Handler id. */
63 /* In the absence of a multiple-writer multiple-reader datastructure for
64 * storing ukeys, we use a large number of cmaps, each with its own lock for
66 #define N_UMAPS 512 /* per udpif. */
68 struct ovs_mutex mutex; /* Take for writing to the following. */
69 struct cmap cmap; /* Datapath flow keys. */
72 /* A thread that processes datapath flows, updates OpenFlow statistics, and
73 * updates or removes them if necessary. */
75 struct udpif *udpif; /* Parent udpif. */
76 pthread_t thread; /* Thread ID. */
77 unsigned int id; /* ovsthread_id_self(). */
80 /* An upcall handler for ofproto_dpif.
82 * udpif keeps records of two kind of logically separate units:
87 * - An array of 'struct handler's for upcall handling and flow
93 * - Revalidation threads which read the datapath flow table and maintains
97 struct ovs_list list_node; /* In all_udpifs list. */
99 struct dpif *dpif; /* Datapath handle. */
100 struct dpif_backer *backer; /* Opaque dpif_backer pointer. */
102 struct handler *handlers; /* Upcall handlers. */
105 struct revalidator *revalidators; /* Flow revalidators. */
106 size_t n_revalidators;
108 struct latch exit_latch; /* Tells child threads to exit. */
111 struct seq *reval_seq; /* Incremented to force revalidation. */
112 bool reval_exit; /* Set by leader on 'exit_latch. */
113 struct ovs_barrier reval_barrier; /* Barrier used by revalidators. */
114 struct dpif_flow_dump *dump; /* DPIF flow dump state. */
115 long long int dump_duration; /* Duration of the last flow dump. */
116 struct seq *dump_seq; /* Increments each dump iteration. */
117 atomic_bool enable_ufid; /* If true, skip dumping flow attrs. */
119 /* There are 'N_UMAPS' maps containing 'struct udpif_key' elements.
121 * During the flow dump phase, revalidators insert into these with a random
122 * distribution. During the garbage collection phase, each revalidator
123 * takes care of garbage collecting a slice of these maps. */
126 /* Datapath flow statistics. */
127 unsigned int max_n_flows;
128 unsigned int avg_n_flows;
130 /* Following fields are accessed and modified by different threads. */
131 atomic_uint flow_limit; /* Datapath flow hard limit. */
133 /* n_flows_mutex prevents multiple threads updating these concurrently. */
134 atomic_uint n_flows; /* Number of flows in the datapath. */
135 atomic_llong n_flows_timestamp; /* Last time n_flows was updated. */
136 struct ovs_mutex n_flows_mutex;
138 /* Following fields are accessed and modified only from the main thread. */
139 struct unixctl_conn **conns; /* Connections waiting on dump_seq. */
140 uint64_t conn_seq; /* Corresponds to 'dump_seq' when
141 conns[n_conns-1] was stored. */
142 size_t n_conns; /* Number of connections waiting. */
146 BAD_UPCALL, /* Some kind of bug somewhere. */
147 MISS_UPCALL, /* A flow miss. */
148 SFLOW_UPCALL, /* sFlow sample. */
149 FLOW_SAMPLE_UPCALL, /* Per-flow sampling. */
150 IPFIX_UPCALL /* Per-bridge sampling. */
154 struct ofproto_dpif *ofproto; /* Parent ofproto. */
156 /* The flow and packet are only required to be constant when using
157 * dpif-netdev. If a modification is absolutely necessary, a const cast
158 * may be used with other datapaths. */
159 const struct flow *flow; /* Parsed representation of the packet. */
160 const ovs_u128 *ufid; /* Unique identifier for 'flow'. */
161 int pmd_id; /* Datapath poll mode driver id. */
162 const struct ofpbuf *packet; /* Packet associated with this upcall. */
163 ofp_port_t in_port; /* OpenFlow in port, or OFPP_NONE. */
165 enum dpif_upcall_type type; /* Datapath type of the upcall. */
166 const struct nlattr *userdata; /* Userdata for DPIF_UC_ACTION Upcalls. */
168 bool xout_initialized; /* True if 'xout' must be uninitialized. */
169 struct xlate_out xout; /* Result of xlate_actions(). */
170 struct ofpbuf put_actions; /* Actions 'put' in the fastapath. */
172 struct dpif_ipfix *ipfix; /* IPFIX pointer or NULL. */
173 struct dpif_sflow *sflow; /* SFlow pointer or NULL. */
175 bool vsp_adjusted; /* 'packet' and 'flow' were adjusted for
176 VLAN splinters if true. */
178 struct udpif_key *ukey; /* Revalidator flow cache. */
179 bool ukey_persists; /* Set true to keep 'ukey' beyond the
180 lifetime of this upcall. */
182 uint64_t dump_seq; /* udpif->dump_seq at translation time. */
183 uint64_t reval_seq; /* udpif->reval_seq at translation time. */
185 /* Not used by the upcall callback interface. */
186 const struct nlattr *key; /* Datapath flow key. */
187 size_t key_len; /* Datapath flow key length. */
188 const struct nlattr *out_tun_key; /* Datapath output tunnel key. */
191 /* 'udpif_key's are responsible for tracking the little bit of state udpif
192 * needs to do flow expiration which can't be pulled directly from the
193 * datapath. They may be created by any handler or revalidator thread at any
194 * time, and read by any revalidator during the dump phase. They are however
195 * each owned by a single revalidator which takes care of destroying them
196 * during the garbage-collection phase.
198 * The mutex within the ukey protects some members of the ukey. The ukey
199 * itself is protected by RCU and is held within a umap in the parent udpif.
200 * Adding or removing a ukey from a umap is only safe when holding the
201 * corresponding umap lock. */
203 struct cmap_node cmap_node; /* In parent revalidator 'ukeys' map. */
205 /* These elements are read only once created, and therefore aren't
206 * protected by a mutex. */
207 const struct nlattr *key; /* Datapath flow key. */
208 size_t key_len; /* Length of 'key'. */
209 const struct nlattr *mask; /* Datapath flow mask. */
210 size_t mask_len; /* Length of 'mask'. */
211 struct ofpbuf *actions; /* Datapath flow actions as nlattrs. */
212 ovs_u128 ufid; /* Unique flow identifier. */
213 bool ufid_present; /* True if 'ufid' is in datapath. */
214 uint32_t hash; /* Pre-computed hash for 'key'. */
215 int pmd_id; /* Datapath poll mode driver id. */
217 struct ovs_mutex mutex; /* Guards the following. */
218 struct dpif_flow_stats stats OVS_GUARDED; /* Last known stats.*/
219 long long int created OVS_GUARDED; /* Estimate of creation time. */
220 uint64_t dump_seq OVS_GUARDED; /* Tracks udpif->dump_seq. */
221 uint64_t reval_seq OVS_GUARDED; /* Tracks udpif->reval_seq. */
222 bool flow_exists OVS_GUARDED; /* Ensures flows are only deleted
225 struct xlate_cache *xcache OVS_GUARDED; /* Cache for xlate entries that
226 * are affected by this ukey.
227 * Used for stats and learning.*/
229 struct odputil_keybuf buf;
234 /* Datapath operation with optional ukey attached. */
236 struct udpif_key *ukey;
237 struct dpif_flow_stats stats; /* Stats for 'op'. */
238 struct dpif_op dop; /* Flow operation. */
241 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
242 static struct ovs_list all_udpifs = OVS_LIST_INITIALIZER(&all_udpifs);
244 static size_t recv_upcalls(struct handler *);
245 static int process_upcall(struct udpif *, struct upcall *,
246 struct ofpbuf *odp_actions);
247 static void handle_upcalls(struct udpif *, struct upcall *, size_t n_upcalls);
248 static void udpif_stop_threads(struct udpif *);
249 static void udpif_start_threads(struct udpif *, size_t n_handlers,
250 size_t n_revalidators);
251 static void *udpif_upcall_handler(void *);
252 static void *udpif_revalidator(void *);
253 static unsigned long udpif_get_n_flows(struct udpif *);
254 static void revalidate(struct revalidator *);
255 static void revalidator_sweep(struct revalidator *);
256 static void revalidator_purge(struct revalidator *);
257 static void upcall_unixctl_show(struct unixctl_conn *conn, int argc,
258 const char *argv[], void *aux);
259 static void upcall_unixctl_disable_megaflows(struct unixctl_conn *, int argc,
260 const char *argv[], void *aux);
261 static void upcall_unixctl_enable_megaflows(struct unixctl_conn *, int argc,
262 const char *argv[], void *aux);
263 static void upcall_unixctl_disable_ufid(struct unixctl_conn *, int argc,
264 const char *argv[], void *aux);
265 static void upcall_unixctl_enable_ufid(struct unixctl_conn *, int argc,
266 const char *argv[], void *aux);
267 static void upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, int argc,
268 const char *argv[], void *aux);
269 static void upcall_unixctl_dump_wait(struct unixctl_conn *conn, int argc,
270 const char *argv[], void *aux);
271 static void upcall_unixctl_purge(struct unixctl_conn *conn, int argc,
272 const char *argv[], void *aux);
274 static struct udpif_key *ukey_create_from_upcall(const struct upcall *);
275 static int ukey_create_from_dpif_flow(const struct udpif *,
276 const struct dpif_flow *,
277 struct udpif_key **);
278 static bool ukey_install_start(struct udpif *, struct udpif_key *ukey);
279 static bool ukey_install_finish(struct udpif_key *ukey, int error);
280 static bool ukey_install(struct udpif *udpif, struct udpif_key *ukey);
281 static struct udpif_key *ukey_lookup(struct udpif *udpif,
282 const ovs_u128 *ufid);
283 static int ukey_acquire(struct udpif *, const struct dpif_flow *,
284 struct udpif_key **result, int *error);
285 static void ukey_delete__(struct udpif_key *);
286 static void ukey_delete(struct umap *, struct udpif_key *);
287 static enum upcall_type classify_upcall(enum dpif_upcall_type type,
288 const struct nlattr *userdata);
290 static int upcall_receive(struct upcall *, const struct dpif_backer *,
291 const struct ofpbuf *packet, enum dpif_upcall_type,
292 const struct nlattr *userdata, const struct flow *,
293 const ovs_u128 *ufid, const int pmd_id);
294 static void upcall_uninit(struct upcall *);
296 static upcall_callback upcall_cb;
298 static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true);
301 udpif_create(struct dpif_backer *backer, struct dpif *dpif)
303 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
304 struct udpif *udpif = xzalloc(sizeof *udpif);
306 if (ovsthread_once_start(&once)) {
307 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show,
309 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
310 upcall_unixctl_disable_megaflows, NULL);
311 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
312 upcall_unixctl_enable_megaflows, NULL);
313 unixctl_command_register("upcall/disable-ufid", "", 0, 0,
314 upcall_unixctl_disable_ufid, NULL);
315 unixctl_command_register("upcall/enable-ufid", "", 0, 0,
316 upcall_unixctl_enable_ufid, NULL);
317 unixctl_command_register("upcall/set-flow-limit", "", 1, 1,
318 upcall_unixctl_set_flow_limit, NULL);
319 unixctl_command_register("revalidator/wait", "", 0, 0,
320 upcall_unixctl_dump_wait, NULL);
321 unixctl_command_register("revalidator/purge", "", 0, 0,
322 upcall_unixctl_purge, NULL);
323 ovsthread_once_done(&once);
327 udpif->backer = backer;
328 atomic_init(&udpif->flow_limit, MIN(ofproto_flow_limit, 10000));
329 udpif->reval_seq = seq_create();
330 udpif->dump_seq = seq_create();
331 latch_init(&udpif->exit_latch);
332 list_push_back(&all_udpifs, &udpif->list_node);
333 atomic_init(&udpif->enable_ufid, false);
334 atomic_init(&udpif->n_flows, 0);
335 atomic_init(&udpif->n_flows_timestamp, LLONG_MIN);
336 ovs_mutex_init(&udpif->n_flows_mutex);
337 udpif->ukeys = xmalloc(N_UMAPS * sizeof *udpif->ukeys);
338 for (int i = 0; i < N_UMAPS; i++) {
339 cmap_init(&udpif->ukeys[i].cmap);
340 ovs_mutex_init(&udpif->ukeys[i].mutex);
343 dpif_register_upcall_cb(dpif, upcall_cb, udpif);
349 udpif_run(struct udpif *udpif)
351 if (udpif->conns && udpif->conn_seq != seq_read(udpif->dump_seq)) {
354 for (i = 0; i < udpif->n_conns; i++) {
355 unixctl_command_reply(udpif->conns[i], NULL);
364 udpif_destroy(struct udpif *udpif)
366 udpif_stop_threads(udpif);
368 for (int i = 0; i < N_UMAPS; i++) {
369 cmap_destroy(&udpif->ukeys[i].cmap);
370 ovs_mutex_destroy(&udpif->ukeys[i].mutex);
375 list_remove(&udpif->list_node);
376 latch_destroy(&udpif->exit_latch);
377 seq_destroy(udpif->reval_seq);
378 seq_destroy(udpif->dump_seq);
379 ovs_mutex_destroy(&udpif->n_flows_mutex);
383 /* Stops the handler and revalidator threads, must be enclosed in
384 * ovsrcu quiescent state unless when destroying udpif. */
386 udpif_stop_threads(struct udpif *udpif)
388 if (udpif && (udpif->n_handlers != 0 || udpif->n_revalidators != 0)) {
391 latch_set(&udpif->exit_latch);
393 for (i = 0; i < udpif->n_handlers; i++) {
394 struct handler *handler = &udpif->handlers[i];
396 xpthread_join(handler->thread, NULL);
399 for (i = 0; i < udpif->n_revalidators; i++) {
400 xpthread_join(udpif->revalidators[i].thread, NULL);
403 dpif_disable_upcall(udpif->dpif);
405 for (i = 0; i < udpif->n_revalidators; i++) {
406 struct revalidator *revalidator = &udpif->revalidators[i];
408 /* Delete ukeys, and delete all flows from the datapath to prevent
409 * double-counting stats. */
410 revalidator_purge(revalidator);
413 latch_poll(&udpif->exit_latch);
415 ovs_barrier_destroy(&udpif->reval_barrier);
417 free(udpif->revalidators);
418 udpif->revalidators = NULL;
419 udpif->n_revalidators = 0;
421 free(udpif->handlers);
422 udpif->handlers = NULL;
423 udpif->n_handlers = 0;
427 /* Starts the handler and revalidator threads, must be enclosed in
428 * ovsrcu quiescent state. */
430 udpif_start_threads(struct udpif *udpif, size_t n_handlers,
431 size_t n_revalidators)
433 if (udpif && n_handlers && n_revalidators) {
437 udpif->n_handlers = n_handlers;
438 udpif->n_revalidators = n_revalidators;
440 udpif->handlers = xzalloc(udpif->n_handlers * sizeof *udpif->handlers);
441 for (i = 0; i < udpif->n_handlers; i++) {
442 struct handler *handler = &udpif->handlers[i];
444 handler->udpif = udpif;
445 handler->handler_id = i;
446 handler->thread = ovs_thread_create(
447 "handler", udpif_upcall_handler, handler);
450 enable_ufid = ofproto_dpif_get_enable_ufid(udpif->backer);
451 atomic_init(&udpif->enable_ufid, enable_ufid);
452 dpif_enable_upcall(udpif->dpif);
454 ovs_barrier_init(&udpif->reval_barrier, udpif->n_revalidators);
455 udpif->reval_exit = false;
456 udpif->revalidators = xzalloc(udpif->n_revalidators
457 * sizeof *udpif->revalidators);
458 for (i = 0; i < udpif->n_revalidators; i++) {
459 struct revalidator *revalidator = &udpif->revalidators[i];
461 revalidator->udpif = udpif;
462 revalidator->thread = ovs_thread_create(
463 "revalidator", udpif_revalidator, revalidator);
468 /* Tells 'udpif' how many threads it should use to handle upcalls.
469 * 'n_handlers' and 'n_revalidators' can never be zero. 'udpif''s
470 * datapath handle must have packet reception enabled before starting
473 udpif_set_threads(struct udpif *udpif, size_t n_handlers,
474 size_t n_revalidators)
477 ovs_assert(n_handlers && n_revalidators);
479 ovsrcu_quiesce_start();
480 if (udpif->n_handlers != n_handlers
481 || udpif->n_revalidators != n_revalidators) {
482 udpif_stop_threads(udpif);
485 if (!udpif->handlers && !udpif->revalidators) {
488 error = dpif_handlers_set(udpif->dpif, n_handlers);
490 VLOG_ERR("failed to configure handlers in dpif %s: %s",
491 dpif_name(udpif->dpif), ovs_strerror(error));
495 udpif_start_threads(udpif, n_handlers, n_revalidators);
497 ovsrcu_quiesce_end();
500 /* Waits for all ongoing upcall translations to complete. This ensures that
501 * there are no transient references to any removed ofprotos (or other
502 * objects). In particular, this should be called after an ofproto is removed
503 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
505 udpif_synchronize(struct udpif *udpif)
507 /* This is stronger than necessary. It would be sufficient to ensure
508 * (somehow) that each handler and revalidator thread had passed through
509 * its main loop once. */
510 size_t n_handlers = udpif->n_handlers;
511 size_t n_revalidators = udpif->n_revalidators;
513 ovsrcu_quiesce_start();
514 udpif_stop_threads(udpif);
515 udpif_start_threads(udpif, n_handlers, n_revalidators);
516 ovsrcu_quiesce_end();
519 /* Notifies 'udpif' that something changed which may render previous
520 * xlate_actions() results invalid. */
522 udpif_revalidate(struct udpif *udpif)
524 seq_change(udpif->reval_seq);
527 /* Returns a seq which increments every time 'udpif' pulls stats from the
528 * datapath. Callers can use this to get a sense of when might be a good time
529 * to do periodic work which relies on relatively up to date statistics. */
531 udpif_dump_seq(struct udpif *udpif)
533 return udpif->dump_seq;
537 udpif_get_memory_usage(struct udpif *udpif, struct simap *usage)
541 simap_increase(usage, "handlers", udpif->n_handlers);
543 simap_increase(usage, "revalidators", udpif->n_revalidators);
544 for (i = 0; i < N_UMAPS; i++) {
545 simap_increase(usage, "udpif keys", cmap_count(&udpif->ukeys[i].cmap));
549 /* Remove flows from a single datapath. */
551 udpif_flush(struct udpif *udpif)
553 size_t n_handlers, n_revalidators;
555 n_handlers = udpif->n_handlers;
556 n_revalidators = udpif->n_revalidators;
558 ovsrcu_quiesce_start();
560 udpif_stop_threads(udpif);
561 dpif_flow_flush(udpif->dpif);
562 udpif_start_threads(udpif, n_handlers, n_revalidators);
564 ovsrcu_quiesce_end();
567 /* Removes all flows from all datapaths. */
569 udpif_flush_all_datapaths(void)
573 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
580 udpif_get_n_flows(struct udpif *udpif)
582 long long int time, now;
583 unsigned long flow_count;
586 atomic_read_relaxed(&udpif->n_flows_timestamp, &time);
587 if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) {
588 struct dpif_dp_stats stats;
590 atomic_store_relaxed(&udpif->n_flows_timestamp, now);
591 dpif_get_dp_stats(udpif->dpif, &stats);
592 flow_count = stats.n_flows;
593 atomic_store_relaxed(&udpif->n_flows, flow_count);
594 ovs_mutex_unlock(&udpif->n_flows_mutex);
596 atomic_read_relaxed(&udpif->n_flows, &flow_count);
601 /* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH
602 * upcalls from dpif, processes the batch and installs corresponding flows
605 udpif_upcall_handler(void *arg)
607 struct handler *handler = arg;
608 struct udpif *udpif = handler->udpif;
610 while (!latch_is_set(&handler->udpif->exit_latch)) {
611 if (recv_upcalls(handler)) {
612 poll_immediate_wake();
614 dpif_recv_wait(udpif->dpif, handler->handler_id);
615 latch_wait(&udpif->exit_latch);
624 recv_upcalls(struct handler *handler)
626 struct udpif *udpif = handler->udpif;
627 uint64_t recv_stubs[UPCALL_MAX_BATCH][512 / 8];
628 struct ofpbuf recv_bufs[UPCALL_MAX_BATCH];
629 struct dpif_upcall dupcalls[UPCALL_MAX_BATCH];
630 struct upcall upcalls[UPCALL_MAX_BATCH];
631 struct flow flows[UPCALL_MAX_BATCH];
635 while (n_upcalls < UPCALL_MAX_BATCH) {
636 struct ofpbuf *recv_buf = &recv_bufs[n_upcalls];
637 struct dpif_upcall *dupcall = &dupcalls[n_upcalls];
638 struct upcall *upcall = &upcalls[n_upcalls];
639 struct flow *flow = &flows[n_upcalls];
640 struct pkt_metadata md;
643 ofpbuf_use_stub(recv_buf, recv_stubs[n_upcalls],
644 sizeof recv_stubs[n_upcalls]);
645 if (dpif_recv(udpif->dpif, handler->handler_id, dupcall, recv_buf)) {
646 ofpbuf_uninit(recv_buf);
650 if (odp_flow_key_to_flow(dupcall->key, dupcall->key_len, flow)
655 error = upcall_receive(upcall, udpif->backer, &dupcall->packet,
656 dupcall->type, dupcall->userdata, flow,
657 &dupcall->ufid, PMD_ID_NULL);
659 if (error == ENODEV) {
660 /* Received packet on datapath port for which we couldn't
661 * associate an ofproto. This can happen if a port is removed
662 * while traffic is being received. Print a rate-limited
663 * message in case it happens frequently. */
664 dpif_flow_put(udpif->dpif, DPIF_FP_CREATE, dupcall->key,
665 dupcall->key_len, NULL, 0, NULL, 0,
666 &dupcall->ufid, PMD_ID_NULL, NULL);
667 VLOG_INFO_RL(&rl, "received packet on unassociated datapath "
668 "port %"PRIu32, flow->in_port.odp_port);
673 upcall->key = dupcall->key;
674 upcall->key_len = dupcall->key_len;
675 upcall->ufid = &dupcall->ufid;
677 upcall->out_tun_key = dupcall->out_tun_key;
679 if (vsp_adjust_flow(upcall->ofproto, flow, &dupcall->packet)) {
680 upcall->vsp_adjusted = true;
683 md = pkt_metadata_from_flow(flow);
684 flow_extract(&dupcall->packet, &md, flow);
686 error = process_upcall(udpif, upcall, NULL);
695 upcall_uninit(upcall);
697 ofpbuf_uninit(&dupcall->packet);
698 ofpbuf_uninit(recv_buf);
702 handle_upcalls(handler->udpif, upcalls, n_upcalls);
703 for (i = 0; i < n_upcalls; i++) {
704 ofpbuf_uninit(&dupcalls[i].packet);
705 ofpbuf_uninit(&recv_bufs[i]);
706 upcall_uninit(&upcalls[i]);
714 udpif_revalidator(void *arg)
716 /* Used by all revalidators. */
717 struct revalidator *revalidator = arg;
718 struct udpif *udpif = revalidator->udpif;
719 bool leader = revalidator == &udpif->revalidators[0];
721 /* Used only by the leader. */
722 long long int start_time = 0;
723 uint64_t last_reval_seq = 0;
726 revalidator->id = ovsthread_id_self();
731 reval_seq = seq_read(udpif->reval_seq);
732 last_reval_seq = reval_seq;
734 n_flows = udpif_get_n_flows(udpif);
735 udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows);
736 udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2;
738 /* Only the leader checks the exit latch to prevent a race where
739 * some threads think it's true and exit and others think it's
740 * false and block indefinitely on the reval_barrier */
741 udpif->reval_exit = latch_is_set(&udpif->exit_latch);
743 start_time = time_msec();
744 if (!udpif->reval_exit) {
747 atomic_read_relaxed(&udpif->enable_ufid, &terse_dump);
748 udpif->dump = dpif_flow_dump_create(udpif->dpif, terse_dump);
752 /* Wait for the leader to start the flow dump. */
753 ovs_barrier_block(&udpif->reval_barrier);
754 if (udpif->reval_exit) {
757 revalidate(revalidator);
759 /* Wait for all flows to have been dumped before we garbage collect. */
760 ovs_barrier_block(&udpif->reval_barrier);
761 revalidator_sweep(revalidator);
763 /* Wait for all revalidators to finish garbage collection. */
764 ovs_barrier_block(&udpif->reval_barrier);
767 unsigned int flow_limit;
768 long long int duration;
770 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
772 dpif_flow_dump_destroy(udpif->dump);
773 seq_change(udpif->dump_seq);
775 duration = MAX(time_msec() - start_time, 1);
776 udpif->dump_duration = duration;
777 if (duration > 2000) {
778 flow_limit /= duration / 1000;
779 } else if (duration > 1300) {
780 flow_limit = flow_limit * 3 / 4;
781 } else if (duration < 1000 && n_flows > 2000
782 && flow_limit < n_flows * 1000 / duration) {
785 flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000));
786 atomic_store_relaxed(&udpif->flow_limit, flow_limit);
788 if (duration > 2000) {
789 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
793 poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500));
794 seq_wait(udpif->reval_seq, last_reval_seq);
795 latch_wait(&udpif->exit_latch);
803 static enum upcall_type
804 classify_upcall(enum dpif_upcall_type type, const struct nlattr *userdata)
806 union user_action_cookie cookie;
809 /* First look at the upcall type. */
817 case DPIF_N_UC_TYPES:
819 VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32, type);
823 /* "action" upcalls need a closer look. */
825 VLOG_WARN_RL(&rl, "action upcall missing cookie");
828 userdata_len = nl_attr_get_size(userdata);
829 if (userdata_len < sizeof cookie.type
830 || userdata_len > sizeof cookie) {
831 VLOG_WARN_RL(&rl, "action upcall cookie has unexpected size %"PRIuSIZE,
835 memset(&cookie, 0, sizeof cookie);
836 memcpy(&cookie, nl_attr_get(userdata), userdata_len);
837 if (userdata_len == MAX(8, sizeof cookie.sflow)
838 && cookie.type == USER_ACTION_COOKIE_SFLOW) {
840 } else if (userdata_len == MAX(8, sizeof cookie.slow_path)
841 && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) {
843 } else if (userdata_len == MAX(8, sizeof cookie.flow_sample)
844 && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) {
845 return FLOW_SAMPLE_UPCALL;
846 } else if (userdata_len == MAX(8, sizeof cookie.ipfix)
847 && cookie.type == USER_ACTION_COOKIE_IPFIX) {
850 VLOG_WARN_RL(&rl, "invalid user cookie of type %"PRIu16
851 " and size %"PRIuSIZE, cookie.type, userdata_len);
856 /* Calculates slow path actions for 'xout'. 'buf' must statically be
857 * initialized with at least 128 bytes of space. */
859 compose_slow_path(struct udpif *udpif, struct xlate_out *xout,
860 const struct flow *flow, odp_port_t odp_in_port,
863 union user_action_cookie cookie;
867 cookie.type = USER_ACTION_COOKIE_SLOW_PATH;
868 cookie.slow_path.unused = 0;
869 cookie.slow_path.reason = xout->slow;
871 port = xout->slow & (SLOW_CFM | SLOW_BFD | SLOW_LACP | SLOW_STP)
874 pid = dpif_port_get_pid(udpif->dpif, port, flow_hash_5tuple(flow, 0));
875 odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, ODPP_NONE,
879 /* If there is no error, the upcall must be destroyed with upcall_uninit()
880 * before quiescing, as the referred objects are guaranteed to exist only
881 * until the calling thread quiesces. Otherwise, do not call upcall_uninit()
882 * since the 'upcall->put_actions' remains uninitialized. */
884 upcall_receive(struct upcall *upcall, const struct dpif_backer *backer,
885 const struct ofpbuf *packet, enum dpif_upcall_type type,
886 const struct nlattr *userdata, const struct flow *flow,
887 const ovs_u128 *ufid, const int pmd_id)
891 error = xlate_lookup(backer, flow, &upcall->ofproto, &upcall->ipfix,
892 &upcall->sflow, NULL, &upcall->in_port);
898 upcall->packet = packet;
900 upcall->pmd_id = pmd_id;
902 upcall->userdata = userdata;
903 ofpbuf_init(&upcall->put_actions, 0);
905 upcall->xout_initialized = false;
906 upcall->vsp_adjusted = false;
907 upcall->ukey_persists = false;
913 upcall->out_tun_key = NULL;
919 upcall_xlate(struct udpif *udpif, struct upcall *upcall,
920 struct ofpbuf *odp_actions)
922 struct dpif_flow_stats stats;
926 stats.n_bytes = ofpbuf_size(upcall->packet);
927 stats.used = time_msec();
928 stats.tcp_flags = ntohs(upcall->flow->tcp_flags);
930 xlate_in_init(&xin, upcall->ofproto, upcall->flow, upcall->in_port, NULL,
931 stats.tcp_flags, upcall->packet);
932 xin.odp_actions = odp_actions;
934 if (upcall->type == DPIF_UC_MISS) {
935 xin.resubmit_stats = &stats;
937 /* For non-miss upcalls, there's a flow in the datapath which this
938 * packet was accounted to. Presumably the revalidators will deal
939 * with pushing its stats eventually. */
942 upcall->dump_seq = seq_read(udpif->dump_seq);
943 upcall->reval_seq = seq_read(udpif->reval_seq);
944 xlate_actions(&xin, &upcall->xout);
945 upcall->xout_initialized = true;
947 /* Special case for fail-open mode.
949 * If we are in fail-open mode, but we are connected to a controller too,
950 * then we should send the packet up to the controller in the hope that it
951 * will try to set up a flow and thereby allow us to exit fail-open.
953 * See the top-level comment in fail-open.c for more information.
955 * Copy packets before they are modified by execution. */
956 if (upcall->xout.fail_open) {
957 const struct ofpbuf *packet = upcall->packet;
958 struct ofproto_packet_in *pin;
960 pin = xmalloc(sizeof *pin);
961 pin->up.packet = xmemdup(ofpbuf_data(packet), ofpbuf_size(packet));
962 pin->up.packet_len = ofpbuf_size(packet);
963 pin->up.reason = OFPR_NO_MATCH;
964 pin->up.table_id = 0;
965 pin->up.cookie = OVS_BE64_MAX;
966 flow_get_metadata(upcall->flow, &pin->up.fmd);
967 pin->send_len = 0; /* Not used for flow table misses. */
968 pin->miss_type = OFPROTO_PACKET_IN_NO_MISS;
969 ofproto_dpif_send_packet_in(upcall->ofproto, pin);
972 if (!upcall->xout.slow) {
973 ofpbuf_use_const(&upcall->put_actions,
974 ofpbuf_data(upcall->xout.odp_actions),
975 ofpbuf_size(upcall->xout.odp_actions));
977 ofpbuf_init(&upcall->put_actions, 0);
978 compose_slow_path(udpif, &upcall->xout, upcall->flow,
979 upcall->flow->in_port.odp_port,
980 &upcall->put_actions);
983 upcall->ukey = ukey_create_from_upcall(upcall);
987 upcall_uninit(struct upcall *upcall)
990 if (upcall->xout_initialized) {
991 xlate_out_uninit(&upcall->xout);
993 ofpbuf_uninit(&upcall->put_actions);
994 if (!upcall->ukey_persists) {
995 ukey_delete__(upcall->ukey);
1001 upcall_cb(const struct ofpbuf *packet, const struct flow *flow, ovs_u128 *ufid,
1002 int pmd_id, enum dpif_upcall_type type,
1003 const struct nlattr *userdata, struct ofpbuf *actions,
1004 struct flow_wildcards *wc, struct ofpbuf *put_actions, void *aux)
1006 struct udpif *udpif = aux;
1007 unsigned int flow_limit;
1008 struct upcall upcall;
1012 atomic_read_relaxed(&enable_megaflows, &megaflow);
1013 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1015 error = upcall_receive(&upcall, udpif->backer, packet, type, userdata,
1016 flow, ufid, pmd_id);
1021 error = process_upcall(udpif, &upcall, actions);
1026 if (upcall.xout.slow && put_actions) {
1027 ofpbuf_put(put_actions, ofpbuf_data(&upcall.put_actions),
1028 ofpbuf_size(&upcall.put_actions));
1031 if (OVS_LIKELY(wc)) {
1033 /* XXX: This could be avoided with sufficient API changes. */
1034 *wc = upcall.xout.wc;
1036 flow_wildcards_init_for_packet(wc, flow);
1040 if (udpif_get_n_flows(udpif) >= flow_limit) {
1045 if (upcall.ukey && !ukey_install(udpif, upcall.ukey)) {
1051 upcall.ukey_persists = true;
1053 upcall_uninit(&upcall);
1058 process_upcall(struct udpif *udpif, struct upcall *upcall,
1059 struct ofpbuf *odp_actions)
1061 const struct nlattr *userdata = upcall->userdata;
1062 const struct ofpbuf *packet = upcall->packet;
1063 const struct flow *flow = upcall->flow;
1065 switch (classify_upcall(upcall->type, userdata)) {
1067 upcall_xlate(udpif, upcall, odp_actions);
1071 if (upcall->sflow) {
1072 union user_action_cookie cookie;
1074 memset(&cookie, 0, sizeof cookie);
1075 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.sflow);
1076 dpif_sflow_received(upcall->sflow, packet, flow,
1077 flow->in_port.odp_port, &cookie);
1082 if (upcall->ipfix) {
1083 union user_action_cookie cookie;
1084 struct flow_tnl output_tunnel_key;
1086 memset(&cookie, 0, sizeof cookie);
1087 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.ipfix);
1089 if (upcall->out_tun_key) {
1090 memset(&output_tunnel_key, 0, sizeof output_tunnel_key);
1091 odp_tun_key_from_attr(upcall->out_tun_key,
1092 &output_tunnel_key);
1094 dpif_ipfix_bridge_sample(upcall->ipfix, packet, flow,
1095 flow->in_port.odp_port,
1096 cookie.ipfix.output_odp_port,
1097 upcall->out_tun_key ?
1098 &output_tunnel_key : NULL);
1102 case FLOW_SAMPLE_UPCALL:
1103 if (upcall->ipfix) {
1104 union user_action_cookie cookie;
1106 memset(&cookie, 0, sizeof cookie);
1107 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.flow_sample);
1109 /* The flow reflects exactly the contents of the packet.
1110 * Sample the packet using it. */
1111 dpif_ipfix_flow_sample(upcall->ipfix, packet, flow,
1112 cookie.flow_sample.collector_set_id,
1113 cookie.flow_sample.probability,
1114 cookie.flow_sample.obs_domain_id,
1115 cookie.flow_sample.obs_point_id);
1127 handle_upcalls(struct udpif *udpif, struct upcall *upcalls,
1130 struct dpif_op *opsp[UPCALL_MAX_BATCH * 2];
1131 struct ukey_op ops[UPCALL_MAX_BATCH * 2];
1132 unsigned int flow_limit;
1133 size_t n_ops, n_opsp, i;
1137 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1138 atomic_read_relaxed(&enable_megaflows, &megaflow);
1140 may_put = udpif_get_n_flows(udpif) < flow_limit;
1142 /* Handle the packets individually in order of arrival.
1144 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1145 * processes received packets for these protocols.
1147 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1150 * The loop fills 'ops' with an array of operations to execute in the
1153 for (i = 0; i < n_upcalls; i++) {
1154 struct upcall *upcall = &upcalls[i];
1155 const struct ofpbuf *packet = upcall->packet;
1158 if (upcall->vsp_adjusted) {
1159 /* This packet was received on a VLAN splinter port. We added a
1160 * VLAN to the packet to make the packet resemble the flow, but the
1161 * actions were composed assuming that the packet contained no
1162 * VLAN. So, we must remove the VLAN header from the packet before
1163 * trying to execute the actions. */
1164 if (ofpbuf_size(upcall->xout.odp_actions)) {
1165 eth_pop_vlan(CONST_CAST(struct ofpbuf *, upcall->packet));
1168 /* Remove the flow vlan tags inserted by vlan splinter logic
1169 * to ensure megaflow masks generated match the data path flow. */
1170 CONST_CAST(struct flow *, upcall->flow)->vlan_tci = 0;
1173 /* Do not install a flow into the datapath if:
1175 * - The datapath already has too many flows.
1177 * - We received this packet via some flow installed in the kernel
1179 if (may_put && upcall->type == DPIF_UC_MISS) {
1180 struct udpif_key *ukey = upcall->ukey;
1182 upcall->ukey_persists = true;
1186 op->dop.type = DPIF_OP_FLOW_PUT;
1187 op->dop.u.flow_put.flags = DPIF_FP_CREATE;
1188 op->dop.u.flow_put.key = ukey->key;
1189 op->dop.u.flow_put.key_len = ukey->key_len;
1190 op->dop.u.flow_put.mask = ukey->mask;
1191 op->dop.u.flow_put.mask_len = ukey->mask_len;
1192 op->dop.u.flow_put.ufid = upcall->ufid;
1193 op->dop.u.flow_put.stats = NULL;
1194 op->dop.u.flow_put.actions = ofpbuf_data(ukey->actions);
1195 op->dop.u.flow_put.actions_len = ofpbuf_size(ukey->actions);
1198 if (ofpbuf_size(upcall->xout.odp_actions)) {
1201 op->dop.type = DPIF_OP_EXECUTE;
1202 op->dop.u.execute.packet = CONST_CAST(struct ofpbuf *, packet);
1203 odp_key_to_pkt_metadata(upcall->key, upcall->key_len,
1204 &op->dop.u.execute.md);
1205 op->dop.u.execute.actions = ofpbuf_data(upcall->xout.odp_actions);
1206 op->dop.u.execute.actions_len = ofpbuf_size(upcall->xout.odp_actions);
1207 op->dop.u.execute.needs_help = (upcall->xout.slow & SLOW_ACTION) != 0;
1208 op->dop.u.execute.probe = false;
1214 * We install ukeys before installing the flows, locking them for exclusive
1215 * access by this thread for the period of installation. This ensures that
1216 * other threads won't attempt to delete the flows as we are creating them.
1219 for (i = 0; i < n_ops; i++) {
1220 struct udpif_key *ukey = ops[i].ukey;
1223 /* If we can't install the ukey, don't install the flow. */
1224 if (!ukey_install_start(udpif, ukey)) {
1225 ukey_delete__(ukey);
1230 opsp[n_opsp++] = &ops[i].dop;
1232 dpif_operate(udpif->dpif, opsp, n_opsp);
1233 for (i = 0; i < n_ops; i++) {
1235 ukey_install_finish(ops[i].ukey, ops[i].dop.error);
1241 get_ufid_hash(const ovs_u128 *ufid)
1243 return ufid->u32[0];
1246 static struct udpif_key *
1247 ukey_lookup(struct udpif *udpif, const ovs_u128 *ufid)
1249 struct udpif_key *ukey;
1250 int idx = get_ufid_hash(ufid) % N_UMAPS;
1251 struct cmap *cmap = &udpif->ukeys[idx].cmap;
1253 CMAP_FOR_EACH_WITH_HASH (ukey, cmap_node, get_ufid_hash(ufid), cmap) {
1254 if (ovs_u128_equal(&ukey->ufid, ufid)) {
1261 static struct udpif_key *
1262 ukey_create__(const struct nlattr *key, size_t key_len,
1263 const struct nlattr *mask, size_t mask_len,
1264 bool ufid_present, const ovs_u128 *ufid,
1265 const int pmd_id, const struct ofpbuf *actions,
1266 uint64_t dump_seq, uint64_t reval_seq, long long int used)
1267 OVS_NO_THREAD_SAFETY_ANALYSIS
1269 struct udpif_key *ukey = xmalloc(sizeof *ukey);
1271 memcpy(&ukey->keybuf, key, key_len);
1272 ukey->key = &ukey->keybuf.nla;
1273 ukey->key_len = key_len;
1274 memcpy(&ukey->maskbuf, mask, mask_len);
1275 ukey->mask = &ukey->maskbuf.nla;
1276 ukey->mask_len = mask_len;
1277 ukey->ufid_present = ufid_present;
1279 ukey->pmd_id = pmd_id;
1280 ukey->hash = get_ufid_hash(&ukey->ufid);
1281 ukey->actions = ofpbuf_clone(actions);
1283 ovs_mutex_init(&ukey->mutex);
1284 ukey->dump_seq = dump_seq;
1285 ukey->reval_seq = reval_seq;
1286 ukey->flow_exists = false;
1287 ukey->created = time_msec();
1288 memset(&ukey->stats, 0, sizeof ukey->stats);
1289 ukey->stats.used = used;
1290 ukey->xcache = NULL;
1295 static struct udpif_key *
1296 ukey_create_from_upcall(const struct upcall *upcall)
1298 struct odputil_keybuf keystub, maskstub;
1299 struct ofpbuf keybuf, maskbuf;
1300 bool recirc, megaflow;
1302 if (upcall->key_len) {
1303 ofpbuf_use_const(&keybuf, upcall->key, upcall->key_len);
1305 /* dpif-netdev doesn't provide a netlink-formatted flow key in the
1306 * upcall, so convert the upcall's flow here. */
1307 ofpbuf_use_stack(&keybuf, &keystub, sizeof keystub);
1308 odp_flow_key_from_flow(&keybuf, upcall->flow, &upcall->xout.wc.masks,
1309 upcall->flow->in_port.odp_port, true);
1312 atomic_read_relaxed(&enable_megaflows, &megaflow);
1313 recirc = ofproto_dpif_get_enable_recirc(upcall->ofproto);
1314 ofpbuf_use_stack(&maskbuf, &maskstub, sizeof maskstub);
1318 max_mpls = ofproto_dpif_get_max_mpls_depth(upcall->ofproto);
1319 odp_flow_key_from_mask(&maskbuf, &upcall->xout.wc.masks, upcall->flow,
1320 UINT32_MAX, max_mpls, recirc);
1323 return ukey_create__(ofpbuf_data(&keybuf), ofpbuf_size(&keybuf),
1324 ofpbuf_data(&maskbuf), ofpbuf_size(&maskbuf),
1325 true, upcall->ufid, upcall->pmd_id,
1326 &upcall->put_actions, upcall->dump_seq,
1327 upcall->reval_seq, 0);
1331 ukey_create_from_dpif_flow(const struct udpif *udpif,
1332 const struct dpif_flow *flow,
1333 struct udpif_key **ukey)
1335 struct dpif_flow full_flow;
1336 struct ofpbuf actions;
1337 uint64_t dump_seq, reval_seq;
1338 uint64_t stub[DPIF_FLOW_BUFSIZE / 8];
1340 if (!flow->key_len) {
1344 /* If the key was not provided by the datapath, fetch the full flow. */
1345 ofpbuf_use_stack(&buf, &stub, sizeof stub);
1346 err = dpif_flow_get(udpif->dpif, NULL, 0, &flow->ufid,
1347 flow->pmd_id, &buf, &full_flow);
1353 dump_seq = seq_read(udpif->dump_seq);
1354 reval_seq = seq_read(udpif->reval_seq);
1355 ofpbuf_use_const(&actions, &flow->actions, flow->actions_len);
1356 *ukey = ukey_create__(flow->key, flow->key_len,
1357 flow->mask, flow->mask_len, flow->ufid_present,
1358 &flow->ufid, flow->pmd_id, &actions, dump_seq,
1359 reval_seq, flow->stats.used);
1364 /* Attempts to insert a ukey into the shared ukey maps.
1366 * On success, returns true, installs the ukey and returns it in a locked
1367 * state. Otherwise, returns false. */
1369 ukey_install_start(struct udpif *udpif, struct udpif_key *new_ukey)
1370 OVS_TRY_LOCK(true, new_ukey->mutex)
1373 struct udpif_key *old_ukey;
1375 bool locked = false;
1377 idx = new_ukey->hash % N_UMAPS;
1378 umap = &udpif->ukeys[idx];
1379 ovs_mutex_lock(&umap->mutex);
1380 old_ukey = ukey_lookup(udpif, &new_ukey->ufid);
1382 /* Uncommon case: A ukey is already installed with the same UFID. */
1383 if (old_ukey->key_len == new_ukey->key_len
1384 && !memcmp(old_ukey->key, new_ukey->key, new_ukey->key_len)) {
1385 COVERAGE_INC(handler_duplicate_upcall);
1387 struct ds ds = DS_EMPTY_INITIALIZER;
1389 odp_format_ufid(&old_ukey->ufid, &ds);
1390 ds_put_cstr(&ds, " ");
1391 odp_flow_key_format(old_ukey->key, old_ukey->key_len, &ds);
1392 ds_put_cstr(&ds, "\n");
1393 odp_format_ufid(&new_ukey->ufid, &ds);
1394 ds_put_cstr(&ds, " ");
1395 odp_flow_key_format(new_ukey->key, new_ukey->key_len, &ds);
1397 VLOG_WARN_RL(&rl, "Conflicting ukey for flows:\n%s", ds_cstr(&ds));
1401 ovs_mutex_lock(&new_ukey->mutex);
1402 cmap_insert(&umap->cmap, &new_ukey->cmap_node, new_ukey->hash);
1405 ovs_mutex_unlock(&umap->mutex);
1411 ukey_install_finish__(struct udpif_key *ukey) OVS_REQUIRES(ukey->mutex)
1413 ukey->flow_exists = true;
1417 ukey_install_finish(struct udpif_key *ukey, int error)
1418 OVS_RELEASES(ukey->mutex)
1421 ukey_install_finish__(ukey);
1423 ovs_mutex_unlock(&ukey->mutex);
1429 ukey_install(struct udpif *udpif, struct udpif_key *ukey)
1431 /* The usual way to keep 'ukey->flow_exists' in sync with the datapath is
1432 * to call ukey_install_start(), install the corresponding datapath flow,
1433 * then call ukey_install_finish(). The netdev interface using upcall_cb()
1434 * doesn't provide a function to separately finish the flow installation,
1435 * so we perform the operations together here.
1437 * This is fine currently, as revalidator threads will only delete this
1438 * ukey during revalidator_sweep() and only if the dump_seq is mismatched.
1439 * It is unlikely for a revalidator thread to advance dump_seq and reach
1440 * the next GC phase between ukey creation and flow installation. */
1441 return ukey_install_start(udpif, ukey) && ukey_install_finish(ukey, 0);
1444 /* Searches for a ukey in 'udpif->ukeys' that matches 'flow' and attempts to
1445 * lock the ukey. If the ukey does not exist, create it.
1447 * Returns 0 on success, setting *result to the matching ukey and returning it
1448 * in a locked state. Otherwise, returns an errno and clears *result. EBUSY
1449 * indicates that another thread is handling this flow. Other errors indicate
1450 * an unexpected condition creating a new ukey.
1452 * *error is an output parameter provided to appease the threadsafety analyser,
1453 * and its value matches the return value. */
1455 ukey_acquire(struct udpif *udpif, const struct dpif_flow *flow,
1456 struct udpif_key **result, int *error)
1457 OVS_TRY_LOCK(0, (*result)->mutex)
1459 struct udpif_key *ukey;
1462 ukey = ukey_lookup(udpif, &flow->ufid);
1464 retval = ovs_mutex_trylock(&ukey->mutex);
1466 /* Usually we try to avoid installing flows from revalidator threads,
1467 * because locking on a umap may cause handler threads to block.
1468 * However there are certain cases, like when ovs-vswitchd is
1469 * restarted, where it is desirable to handle flows that exist in the
1470 * datapath gracefully (ie, don't just clear the datapath). */
1473 retval = ukey_create_from_dpif_flow(udpif, flow, &ukey);
1477 install = ukey_install_start(udpif, ukey);
1479 ukey_install_finish__(ukey);
1482 ukey_delete__(ukey);
1498 ukey_delete__(struct udpif_key *ukey)
1499 OVS_NO_THREAD_SAFETY_ANALYSIS
1502 xlate_cache_delete(ukey->xcache);
1503 ofpbuf_delete(ukey->actions);
1504 ovs_mutex_destroy(&ukey->mutex);
1510 ukey_delete(struct umap *umap, struct udpif_key *ukey)
1511 OVS_REQUIRES(umap->mutex)
1513 cmap_remove(&umap->cmap, &ukey->cmap_node, ukey->hash);
1514 ovsrcu_postpone(ukey_delete__, ukey);
1518 should_revalidate(const struct udpif *udpif, uint64_t packets,
1521 long long int metric, now, duration;
1523 if (udpif->dump_duration < 200) {
1524 /* We are likely to handle full revalidation for the flows. */
1528 /* Calculate the mean time between seeing these packets. If this
1529 * exceeds the threshold, then delete the flow rather than performing
1530 * costly revalidation for flows that aren't being hit frequently.
1532 * This is targeted at situations where the dump_duration is high (~1s),
1533 * and revalidation is triggered by a call to udpif_revalidate(). In
1534 * these situations, revalidation of all flows causes fluctuations in the
1535 * flow_limit due to the interaction with the dump_duration and max_idle.
1536 * This tends to result in deletion of low-throughput flows anyway, so
1537 * skip the revalidation and just delete those flows. */
1538 packets = MAX(packets, 1);
1539 now = MAX(used, time_msec());
1540 duration = now - used;
1541 metric = duration / packets;
1544 /* The flow is receiving more than ~5pps, so keep it. */
1551 revalidate_ukey(struct udpif *udpif, struct udpif_key *ukey,
1552 const struct dpif_flow_stats *stats, uint64_t reval_seq)
1553 OVS_REQUIRES(ukey->mutex)
1555 uint64_t slow_path_buf[128 / 8];
1556 struct xlate_out xout, *xoutp;
1557 struct netflow *netflow;
1558 struct ofproto_dpif *ofproto;
1559 struct dpif_flow_stats push;
1560 struct ofpbuf xout_actions;
1561 struct flow flow, dp_mask;
1562 uint32_t *dp32, *xout32;
1563 ofp_port_t ofp_in_port;
1564 struct xlate_in xin;
1565 long long int last_used;
1569 bool need_revalidate;
1575 need_revalidate = (ukey->reval_seq != reval_seq);
1576 last_used = ukey->stats.used;
1577 push.used = stats->used;
1578 push.tcp_flags = stats->tcp_flags;
1579 push.n_packets = (stats->n_packets > ukey->stats.n_packets
1580 ? stats->n_packets - ukey->stats.n_packets
1582 push.n_bytes = (stats->n_bytes > ukey->stats.n_bytes
1583 ? stats->n_bytes - ukey->stats.n_bytes
1586 if (need_revalidate && last_used
1587 && !should_revalidate(udpif, push.n_packets, last_used)) {
1592 /* We will push the stats, so update the ukey stats cache. */
1593 ukey->stats = *stats;
1594 if (!push.n_packets && !need_revalidate) {
1599 if (ukey->xcache && !need_revalidate) {
1600 xlate_push_stats(ukey->xcache, &push);
1605 if (odp_flow_key_to_flow(ukey->key, ukey->key_len, &flow)
1610 error = xlate_lookup(udpif->backer, &flow, &ofproto, NULL, NULL, &netflow,
1616 if (need_revalidate) {
1617 xlate_cache_clear(ukey->xcache);
1619 if (!ukey->xcache) {
1620 ukey->xcache = xlate_cache_new();
1623 xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL, push.tcp_flags,
1625 if (push.n_packets) {
1626 xin.resubmit_stats = &push;
1627 xin.may_learn = true;
1629 xin.xcache = ukey->xcache;
1630 xin.skip_wildcards = !need_revalidate;
1631 xlate_actions(&xin, &xout);
1634 if (!need_revalidate) {
1640 ofpbuf_use_const(&xout_actions, ofpbuf_data(xout.odp_actions),
1641 ofpbuf_size(xout.odp_actions));
1643 ofpbuf_use_stack(&xout_actions, slow_path_buf, sizeof slow_path_buf);
1644 compose_slow_path(udpif, &xout, &flow, flow.in_port.odp_port,
1648 if (!ofpbuf_equal(&xout_actions, ukey->actions)) {
1652 if (odp_flow_key_to_mask(ukey->mask, ukey->mask_len, &dp_mask, &flow)
1657 /* Since the kernel is free to ignore wildcarded bits in the mask, we can't
1658 * directly check that the masks are the same. Instead we check that the
1659 * mask in the kernel is more specific i.e. less wildcarded, than what
1660 * we've calculated here. This guarantees we don't catch any packets we
1661 * shouldn't with the megaflow. */
1662 dp32 = (uint32_t *) &dp_mask;
1663 xout32 = (uint32_t *) &xout.wc.masks;
1664 for (i = 0; i < FLOW_U32S; i++) {
1665 if ((dp32[i] | xout32[i]) != dp32[i]) {
1674 ukey->reval_seq = reval_seq;
1676 if (netflow && !ok) {
1677 netflow_flow_clear(netflow, &flow);
1679 xlate_out_uninit(xoutp);
1684 delete_op_init__(struct udpif *udpif, struct ukey_op *op,
1685 const struct dpif_flow *flow)
1688 op->dop.type = DPIF_OP_FLOW_DEL;
1689 op->dop.u.flow_del.key = flow->key;
1690 op->dop.u.flow_del.key_len = flow->key_len;
1691 op->dop.u.flow_del.ufid = flow->ufid_present ? &flow->ufid : NULL;
1692 op->dop.u.flow_del.pmd_id = flow->pmd_id;
1693 op->dop.u.flow_del.stats = &op->stats;
1694 atomic_read_relaxed(&udpif->enable_ufid, &op->dop.u.flow_del.terse);
1698 delete_op_init(struct udpif *udpif, struct ukey_op *op, struct udpif_key *ukey)
1701 op->dop.type = DPIF_OP_FLOW_DEL;
1702 op->dop.u.flow_del.key = ukey->key;
1703 op->dop.u.flow_del.key_len = ukey->key_len;
1704 op->dop.u.flow_del.ufid = ukey->ufid_present ? &ukey->ufid : NULL;
1705 op->dop.u.flow_del.pmd_id = ukey->pmd_id;
1706 op->dop.u.flow_del.stats = &op->stats;
1707 atomic_read_relaxed(&udpif->enable_ufid, &op->dop.u.flow_del.terse);
1711 push_ukey_ops__(struct udpif *udpif, struct ukey_op *ops, size_t n_ops)
1713 struct dpif_op *opsp[REVALIDATE_MAX_BATCH];
1716 ovs_assert(n_ops <= REVALIDATE_MAX_BATCH);
1717 for (i = 0; i < n_ops; i++) {
1718 opsp[i] = &ops[i].dop;
1720 dpif_operate(udpif->dpif, opsp, n_ops);
1722 for (i = 0; i < n_ops; i++) {
1723 struct ukey_op *op = &ops[i];
1724 struct dpif_flow_stats *push, *stats, push_buf;
1726 stats = op->dop.u.flow_del.stats;
1730 ovs_mutex_lock(&op->ukey->mutex);
1731 push->used = MAX(stats->used, op->ukey->stats.used);
1732 push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags;
1733 push->n_packets = stats->n_packets - op->ukey->stats.n_packets;
1734 push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes;
1735 ovs_mutex_unlock(&op->ukey->mutex);
1740 if (push->n_packets || netflow_exists()) {
1741 const struct nlattr *key = op->dop.u.flow_del.key;
1742 size_t key_len = op->dop.u.flow_del.key_len;
1743 struct ofproto_dpif *ofproto;
1744 struct netflow *netflow;
1745 ofp_port_t ofp_in_port;
1750 ovs_mutex_lock(&op->ukey->mutex);
1751 if (op->ukey->xcache) {
1752 xlate_push_stats(op->ukey->xcache, push);
1753 ovs_mutex_unlock(&op->ukey->mutex);
1756 ovs_mutex_unlock(&op->ukey->mutex);
1757 key = op->ukey->key;
1758 key_len = op->ukey->key_len;
1761 if (odp_flow_key_to_flow(key, key_len, &flow)
1766 error = xlate_lookup(udpif->backer, &flow, &ofproto,
1767 NULL, NULL, &netflow, &ofp_in_port);
1769 struct xlate_in xin;
1771 xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL,
1772 push->tcp_flags, NULL);
1773 xin.resubmit_stats = push->n_packets ? push : NULL;
1774 xin.may_learn = push->n_packets > 0;
1775 xin.skip_wildcards = true;
1776 xlate_actions_for_side_effects(&xin);
1779 netflow_flow_clear(netflow, &flow);
1787 push_ukey_ops(struct udpif *udpif, struct umap *umap,
1788 struct ukey_op *ops, size_t n_ops)
1792 push_ukey_ops__(udpif, ops, n_ops);
1793 ovs_mutex_lock(&umap->mutex);
1794 for (i = 0; i < n_ops; i++) {
1795 ukey_delete(umap, ops[i].ukey);
1797 ovs_mutex_unlock(&umap->mutex);
1801 log_unexpected_flow(const struct dpif_flow *flow, int error)
1803 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(10, 60);
1804 struct ds ds = DS_EMPTY_INITIALIZER;
1806 ds_put_format(&ds, "Failed to acquire udpif_key corresponding to "
1807 "unexpected flow (%s): ", ovs_strerror(error));
1808 odp_format_ufid(&flow->ufid, &ds);
1809 VLOG_WARN_RL(&rl, "%s", ds_cstr(&ds));
1813 revalidate(struct revalidator *revalidator)
1815 struct udpif *udpif = revalidator->udpif;
1816 struct dpif_flow_dump_thread *dump_thread;
1817 uint64_t dump_seq, reval_seq;
1818 unsigned int flow_limit;
1820 dump_seq = seq_read(udpif->dump_seq);
1821 reval_seq = seq_read(udpif->reval_seq);
1822 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1823 dump_thread = dpif_flow_dump_thread_create(udpif->dump);
1825 struct ukey_op ops[REVALIDATE_MAX_BATCH];
1828 struct dpif_flow flows[REVALIDATE_MAX_BATCH];
1829 const struct dpif_flow *f;
1832 long long int max_idle;
1837 n_dumped = dpif_flow_dump_next(dump_thread, flows, ARRAY_SIZE(flows));
1844 /* In normal operation we want to keep flows around until they have
1845 * been idle for 'ofproto_max_idle' milliseconds. However:
1847 * - If the number of datapath flows climbs above 'flow_limit',
1848 * drop that down to 100 ms to try to bring the flows down to
1851 * - If the number of datapath flows climbs above twice
1852 * 'flow_limit', delete all the datapath flows as an emergency
1853 * measure. (We reassess this condition for the next batch of
1854 * datapath flows, so we will recover before all the flows are
1856 n_dp_flows = udpif_get_n_flows(udpif);
1857 kill_them_all = n_dp_flows > flow_limit * 2;
1858 max_idle = n_dp_flows > flow_limit ? 100 : ofproto_max_idle;
1860 for (f = flows; f < &flows[n_dumped]; f++) {
1861 long long int used = f->stats.used;
1862 struct udpif_key *ukey;
1863 bool already_dumped, keep;
1866 if (ukey_acquire(udpif, f, &ukey, &error)) {
1867 if (error == EBUSY) {
1868 /* Another thread is processing this flow, so don't bother
1870 COVERAGE_INC(upcall_ukey_contention);
1872 log_unexpected_flow(f, error);
1873 if (error != ENOENT) {
1874 delete_op_init__(udpif, &ops[n_ops++], f);
1880 already_dumped = ukey->dump_seq == dump_seq;
1881 if (already_dumped) {
1882 /* The flow has already been handled during this flow dump
1883 * operation. Skip it. */
1885 COVERAGE_INC(dumped_duplicate_flow);
1887 COVERAGE_INC(dumped_new_flow);
1889 ovs_mutex_unlock(&ukey->mutex);
1894 used = ukey->created;
1896 if (kill_them_all || (used && used < now - max_idle)) {
1899 keep = revalidate_ukey(udpif, ukey, &f->stats, reval_seq);
1901 ukey->dump_seq = dump_seq;
1902 ukey->flow_exists = keep;
1905 delete_op_init(udpif, &ops[n_ops++], ukey);
1907 ovs_mutex_unlock(&ukey->mutex);
1911 push_ukey_ops__(udpif, ops, n_ops);
1915 dpif_flow_dump_thread_destroy(dump_thread);
1919 handle_missed_revalidation(struct udpif *udpif, uint64_t reval_seq,
1920 struct udpif_key *ukey)
1922 struct dpif_flow_stats stats;
1925 COVERAGE_INC(revalidate_missed_dp_flow);
1927 memset(&stats, 0, sizeof stats);
1928 ovs_mutex_lock(&ukey->mutex);
1929 keep = revalidate_ukey(udpif, ukey, &stats, reval_seq);
1930 ovs_mutex_unlock(&ukey->mutex);
1936 revalidator_sweep__(struct revalidator *revalidator, bool purge)
1938 struct udpif *udpif;
1939 uint64_t dump_seq, reval_seq;
1942 udpif = revalidator->udpif;
1943 dump_seq = seq_read(udpif->dump_seq);
1944 reval_seq = seq_read(udpif->reval_seq);
1945 slice = revalidator - udpif->revalidators;
1946 ovs_assert(slice < udpif->n_revalidators);
1948 for (int i = slice; i < N_UMAPS; i += udpif->n_revalidators) {
1949 struct ukey_op ops[REVALIDATE_MAX_BATCH];
1950 struct udpif_key *ukey;
1951 struct umap *umap = &udpif->ukeys[i];
1954 CMAP_FOR_EACH(ukey, cmap_node, &umap->cmap) {
1955 bool flow_exists, seq_mismatch;
1957 /* Handler threads could be holding a ukey lock while it installs a
1958 * new flow, so don't hang around waiting for access to it. */
1959 if (ovs_mutex_trylock(&ukey->mutex)) {
1962 flow_exists = ukey->flow_exists;
1963 seq_mismatch = (ukey->dump_seq != dump_seq
1964 && ukey->reval_seq != reval_seq);
1965 ovs_mutex_unlock(&ukey->mutex);
1970 && !handle_missed_revalidation(udpif, reval_seq,
1972 struct ukey_op *op = &ops[n_ops++];
1974 delete_op_init(udpif, op, ukey);
1975 if (n_ops == REVALIDATE_MAX_BATCH) {
1976 push_ukey_ops(udpif, umap, ops, n_ops);
1979 } else if (!flow_exists) {
1980 ovs_mutex_lock(&umap->mutex);
1981 ukey_delete(umap, ukey);
1982 ovs_mutex_unlock(&umap->mutex);
1987 push_ukey_ops(udpif, umap, ops, n_ops);
1994 revalidator_sweep(struct revalidator *revalidator)
1996 revalidator_sweep__(revalidator, false);
2000 revalidator_purge(struct revalidator *revalidator)
2002 revalidator_sweep__(revalidator, true);
2006 upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED,
2007 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
2009 struct ds ds = DS_EMPTY_INITIALIZER;
2010 struct udpif *udpif;
2012 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
2013 unsigned int flow_limit;
2017 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
2018 atomic_read_relaxed(&udpif->enable_ufid, &ufid_enabled);
2020 ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif));
2021 ds_put_format(&ds, "\tflows : (current %lu)"
2022 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif),
2023 udpif->avg_n_flows, udpif->max_n_flows, flow_limit);
2024 ds_put_format(&ds, "\tdump duration : %lldms\n", udpif->dump_duration);
2025 ds_put_format(&ds, "\tufid enabled : ");
2027 ds_put_format(&ds, "true\n");
2029 ds_put_format(&ds, "false\n");
2031 ds_put_char(&ds, '\n');
2033 for (i = 0; i < n_revalidators; i++) {
2034 struct revalidator *revalidator = &udpif->revalidators[i];
2035 int j, elements = 0;
2037 for (j = i; j < N_UMAPS; j += n_revalidators) {
2038 elements += cmap_count(&udpif->ukeys[j].cmap);
2040 ds_put_format(&ds, "\t%u: (keys %d)\n", revalidator->id, elements);
2044 unixctl_command_reply(conn, ds_cstr(&ds));
2048 /* Disable using the megaflows.
2050 * This command is only needed for advanced debugging, so it's not
2051 * documented in the man page. */
2053 upcall_unixctl_disable_megaflows(struct unixctl_conn *conn,
2054 int argc OVS_UNUSED,
2055 const char *argv[] OVS_UNUSED,
2056 void *aux OVS_UNUSED)
2058 atomic_store_relaxed(&enable_megaflows, false);
2059 udpif_flush_all_datapaths();
2060 unixctl_command_reply(conn, "megaflows disabled");
2063 /* Re-enable using megaflows.
2065 * This command is only needed for advanced debugging, so it's not
2066 * documented in the man page. */
2068 upcall_unixctl_enable_megaflows(struct unixctl_conn *conn,
2069 int argc OVS_UNUSED,
2070 const char *argv[] OVS_UNUSED,
2071 void *aux OVS_UNUSED)
2073 atomic_store_relaxed(&enable_megaflows, true);
2074 udpif_flush_all_datapaths();
2075 unixctl_command_reply(conn, "megaflows enabled");
2078 /* Disable skipping flow attributes during flow dump.
2080 * This command is only needed for advanced debugging, so it's not
2081 * documented in the man page. */
2083 upcall_unixctl_disable_ufid(struct unixctl_conn *conn, int argc OVS_UNUSED,
2084 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
2086 struct udpif *udpif;
2088 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
2089 atomic_store(&udpif->enable_ufid, false);
2091 unixctl_command_reply(conn, "Datapath dumping tersely using UFID disabled");
2094 /* Re-enable skipping flow attributes during flow dump.
2096 * This command is only needed for advanced debugging, so it's not documented
2097 * in the man page. */
2099 upcall_unixctl_enable_ufid(struct unixctl_conn *conn, int argc OVS_UNUSED,
2100 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
2102 struct udpif *udpif;
2104 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
2105 atomic_store(&udpif->enable_ufid, true);
2107 unixctl_command_reply(conn, "Datapath dumping tersely using UFID enabled");
2110 /* Set the flow limit.
2112 * This command is only needed for advanced debugging, so it's not
2113 * documented in the man page. */
2115 upcall_unixctl_set_flow_limit(struct unixctl_conn *conn,
2116 int argc OVS_UNUSED,
2117 const char *argv[] OVS_UNUSED,
2118 void *aux OVS_UNUSED)
2120 struct ds ds = DS_EMPTY_INITIALIZER;
2121 struct udpif *udpif;
2122 unsigned int flow_limit = atoi(argv[1]);
2124 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
2125 atomic_store_relaxed(&udpif->flow_limit, flow_limit);
2127 ds_put_format(&ds, "set flow_limit to %u\n", flow_limit);
2128 unixctl_command_reply(conn, ds_cstr(&ds));
2133 upcall_unixctl_dump_wait(struct unixctl_conn *conn,
2134 int argc OVS_UNUSED,
2135 const char *argv[] OVS_UNUSED,
2136 void *aux OVS_UNUSED)
2138 if (list_is_singleton(&all_udpifs)) {
2139 struct udpif *udpif = NULL;
2142 udpif = OBJECT_CONTAINING(list_front(&all_udpifs), udpif, list_node);
2143 len = (udpif->n_conns + 1) * sizeof *udpif->conns;
2144 udpif->conn_seq = seq_read(udpif->dump_seq);
2145 udpif->conns = xrealloc(udpif->conns, len);
2146 udpif->conns[udpif->n_conns++] = conn;
2148 unixctl_command_reply_error(conn, "can't wait on multiple udpifs.");
2153 upcall_unixctl_purge(struct unixctl_conn *conn, int argc OVS_UNUSED,
2154 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
2156 struct udpif *udpif;
2158 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
2161 for (n = 0; n < udpif->n_revalidators; n++) {
2162 revalidator_purge(&udpif->revalidators[n]);
2165 unixctl_command_reply(conn, "");