1 /* Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2015, 2016 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 /* These variables provide a mechanism for the main thread to pause
120 * all revalidation without having to completely shut the threads down.
121 * 'pause_latch' is shared between the main thread and the lead
122 * revalidator thread, so when it is desirable to halt revalidation, the
123 * main thread will set the latch. 'pause' and 'pause_barrier' are shared
124 * by revalidator threads. The lead revalidator will set 'pause' when it
125 * observes the latch has been set, and this will cause all revalidator
126 * threads to wait on 'pause_barrier' at the beginning of the next
127 * revalidation round. */
128 bool pause; /* Set by leader on 'pause_latch. */
129 struct latch pause_latch; /* Set to force revalidators pause. */
130 struct ovs_barrier pause_barrier; /* Barrier used to pause all */
131 /* revalidators by main thread. */
133 /* There are 'N_UMAPS' maps containing 'struct udpif_key' elements.
135 * During the flow dump phase, revalidators insert into these with a random
136 * distribution. During the garbage collection phase, each revalidator
137 * takes care of garbage collecting a slice of these maps. */
140 /* Datapath flow statistics. */
141 unsigned int max_n_flows;
142 unsigned int avg_n_flows;
144 /* Following fields are accessed and modified by different threads. */
145 atomic_uint flow_limit; /* Datapath flow hard limit. */
147 /* n_flows_mutex prevents multiple threads updating these concurrently. */
148 atomic_uint n_flows; /* Number of flows in the datapath. */
149 atomic_llong n_flows_timestamp; /* Last time n_flows was updated. */
150 struct ovs_mutex n_flows_mutex;
152 /* Following fields are accessed and modified only from the main thread. */
153 struct unixctl_conn **conns; /* Connections waiting on dump_seq. */
154 uint64_t conn_seq; /* Corresponds to 'dump_seq' when
155 conns[n_conns-1] was stored. */
156 size_t n_conns; /* Number of connections waiting. */
160 BAD_UPCALL, /* Some kind of bug somewhere. */
161 MISS_UPCALL, /* A flow miss. */
162 SFLOW_UPCALL, /* sFlow sample. */
163 FLOW_SAMPLE_UPCALL, /* Per-flow sampling. */
164 IPFIX_UPCALL /* Per-bridge sampling. */
174 struct ofproto_dpif *ofproto; /* Parent ofproto. */
175 const struct recirc_id_node *recirc; /* Recirculation context. */
176 bool have_recirc_ref; /* Reference held on recirc ctx? */
178 /* The flow and packet are only required to be constant when using
179 * dpif-netdev. If a modification is absolutely necessary, a const cast
180 * may be used with other datapaths. */
181 const struct flow *flow; /* Parsed representation of the packet. */
182 const ovs_u128 *ufid; /* Unique identifier for 'flow'. */
183 unsigned pmd_id; /* Datapath poll mode driver id. */
184 const struct dp_packet *packet; /* Packet associated with this upcall. */
185 ofp_port_t in_port; /* OpenFlow in port, or OFPP_NONE. */
186 uint16_t mru; /* If !0, Maximum receive unit of
187 fragmented IP packet */
189 enum dpif_upcall_type type; /* Datapath type of the upcall. */
190 const struct nlattr *userdata; /* Userdata for DPIF_UC_ACTION Upcalls. */
191 const struct nlattr *actions; /* Flow actions in DPIF_UC_ACTION Upcalls. */
193 bool xout_initialized; /* True if 'xout' must be uninitialized. */
194 struct xlate_out xout; /* Result of xlate_actions(). */
195 struct ofpbuf odp_actions; /* Datapath actions from xlate_actions(). */
196 struct flow_wildcards wc; /* Dependencies that megaflow must match. */
197 struct ofpbuf put_actions; /* Actions 'put' in the fastpath. */
199 struct dpif_ipfix *ipfix; /* IPFIX pointer or NULL. */
200 struct dpif_sflow *sflow; /* SFlow pointer or NULL. */
202 bool vsp_adjusted; /* 'packet' and 'flow' were adjusted for
203 VLAN splinters if true. */
205 struct udpif_key *ukey; /* Revalidator flow cache. */
206 bool ukey_persists; /* Set true to keep 'ukey' beyond the
207 lifetime of this upcall. */
209 uint64_t dump_seq; /* udpif->dump_seq at translation time. */
210 uint64_t reval_seq; /* udpif->reval_seq at translation time. */
212 /* Not used by the upcall callback interface. */
213 const struct nlattr *key; /* Datapath flow key. */
214 size_t key_len; /* Datapath flow key length. */
215 const struct nlattr *out_tun_key; /* Datapath output tunnel key. */
217 uint64_t odp_actions_stub[1024 / 8]; /* Stub for odp_actions. */
220 /* 'udpif_key's are responsible for tracking the little bit of state udpif
221 * needs to do flow expiration which can't be pulled directly from the
222 * datapath. They may be created by any handler or revalidator thread at any
223 * time, and read by any revalidator during the dump phase. They are however
224 * each owned by a single revalidator which takes care of destroying them
225 * during the garbage-collection phase.
227 * The mutex within the ukey protects some members of the ukey. The ukey
228 * itself is protected by RCU and is held within a umap in the parent udpif.
229 * Adding or removing a ukey from a umap is only safe when holding the
230 * corresponding umap lock. */
232 struct cmap_node cmap_node; /* In parent revalidator 'ukeys' map. */
234 /* These elements are read only once created, and therefore aren't
235 * protected by a mutex. */
236 const struct nlattr *key; /* Datapath flow key. */
237 size_t key_len; /* Length of 'key'. */
238 const struct nlattr *mask; /* Datapath flow mask. */
239 size_t mask_len; /* Length of 'mask'. */
240 ovs_u128 ufid; /* Unique flow identifier. */
241 bool ufid_present; /* True if 'ufid' is in datapath. */
242 uint32_t hash; /* Pre-computed hash for 'key'. */
243 unsigned pmd_id; /* Datapath poll mode driver id. */
245 struct ovs_mutex mutex; /* Guards the following. */
246 struct dpif_flow_stats stats OVS_GUARDED; /* Last known stats.*/
247 long long int created OVS_GUARDED; /* Estimate of creation time. */
248 uint64_t dump_seq OVS_GUARDED; /* Tracks udpif->dump_seq. */
249 uint64_t reval_seq OVS_GUARDED; /* Tracks udpif->reval_seq. */
250 bool flow_exists OVS_GUARDED; /* Ensures flows are only deleted
252 /* Datapath flow actions as nlattrs. Protected by RCU. Read with
253 * ukey_get_actions(), and write with ukey_set_actions(). */
254 OVSRCU_TYPE(struct ofpbuf *) actions;
256 struct xlate_cache *xcache OVS_GUARDED; /* Cache for xlate entries that
257 * are affected by this ukey.
258 * Used for stats and learning.*/
260 struct odputil_keybuf buf;
264 uint32_t key_recirc_id; /* Non-zero if reference is held by the ukey. */
265 struct recirc_refs recircs; /* Action recirc IDs with references held. */
268 /* Datapath operation with optional ukey attached. */
270 struct udpif_key *ukey;
271 struct dpif_flow_stats stats; /* Stats for 'op'. */
272 struct dpif_op dop; /* Flow operation. */
275 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
276 static struct ovs_list all_udpifs = OVS_LIST_INITIALIZER(&all_udpifs);
278 static size_t recv_upcalls(struct handler *);
279 static int process_upcall(struct udpif *, struct upcall *,
280 struct ofpbuf *odp_actions, struct flow_wildcards *);
281 static void handle_upcalls(struct udpif *, struct upcall *, size_t n_upcalls);
282 static void udpif_stop_threads(struct udpif *);
283 static void udpif_start_threads(struct udpif *, size_t n_handlers,
284 size_t n_revalidators);
285 static void udpif_pause_revalidators(struct udpif *);
286 static void udpif_resume_revalidators(struct udpif *);
287 static void *udpif_upcall_handler(void *);
288 static void *udpif_revalidator(void *);
289 static unsigned long udpif_get_n_flows(struct udpif *);
290 static void revalidate(struct revalidator *);
291 static void revalidator_pause(struct revalidator *);
292 static void revalidator_sweep(struct revalidator *);
293 static void revalidator_purge(struct revalidator *);
294 static void upcall_unixctl_show(struct unixctl_conn *conn, int argc,
295 const char *argv[], void *aux);
296 static void upcall_unixctl_disable_megaflows(struct unixctl_conn *, int argc,
297 const char *argv[], void *aux);
298 static void upcall_unixctl_enable_megaflows(struct unixctl_conn *, int argc,
299 const char *argv[], void *aux);
300 static void upcall_unixctl_disable_ufid(struct unixctl_conn *, int argc,
301 const char *argv[], void *aux);
302 static void upcall_unixctl_enable_ufid(struct unixctl_conn *, int argc,
303 const char *argv[], void *aux);
304 static void upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, int argc,
305 const char *argv[], void *aux);
306 static void upcall_unixctl_dump_wait(struct unixctl_conn *conn, int argc,
307 const char *argv[], void *aux);
308 static void upcall_unixctl_purge(struct unixctl_conn *conn, int argc,
309 const char *argv[], void *aux);
311 static struct udpif_key *ukey_create_from_upcall(struct upcall *,
312 struct flow_wildcards *);
313 static int ukey_create_from_dpif_flow(const struct udpif *,
314 const struct dpif_flow *,
315 struct udpif_key **);
316 static void ukey_get_actions(struct udpif_key *, const struct nlattr **actions,
318 static bool ukey_install_start(struct udpif *, struct udpif_key *ukey);
319 static bool ukey_install_finish(struct udpif_key *ukey, int error);
320 static bool ukey_install(struct udpif *udpif, struct udpif_key *ukey);
321 static struct udpif_key *ukey_lookup(struct udpif *udpif,
322 const ovs_u128 *ufid);
323 static int ukey_acquire(struct udpif *, const struct dpif_flow *,
324 struct udpif_key **result, int *error);
325 static void ukey_delete__(struct udpif_key *);
326 static void ukey_delete(struct umap *, struct udpif_key *);
327 static enum upcall_type classify_upcall(enum dpif_upcall_type type,
328 const struct nlattr *userdata);
330 static int upcall_receive(struct upcall *, const struct dpif_backer *,
331 const struct dp_packet *packet, enum dpif_upcall_type,
332 const struct nlattr *userdata, const struct flow *,
333 const unsigned int mru,
334 const ovs_u128 *ufid, const unsigned pmd_id);
335 static void upcall_uninit(struct upcall *);
337 static upcall_callback upcall_cb;
338 static dp_purge_callback dp_purge_cb;
340 static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true);
341 static atomic_bool enable_ufid = ATOMIC_VAR_INIT(true);
346 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
347 if (ovsthread_once_start(&once)) {
348 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show,
350 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
351 upcall_unixctl_disable_megaflows, NULL);
352 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
353 upcall_unixctl_enable_megaflows, NULL);
354 unixctl_command_register("upcall/disable-ufid", "", 0, 0,
355 upcall_unixctl_disable_ufid, NULL);
356 unixctl_command_register("upcall/enable-ufid", "", 0, 0,
357 upcall_unixctl_enable_ufid, NULL);
358 unixctl_command_register("upcall/set-flow-limit", "", 1, 1,
359 upcall_unixctl_set_flow_limit, NULL);
360 unixctl_command_register("revalidator/wait", "", 0, 0,
361 upcall_unixctl_dump_wait, NULL);
362 unixctl_command_register("revalidator/purge", "", 0, 0,
363 upcall_unixctl_purge, NULL);
364 ovsthread_once_done(&once);
369 udpif_create(struct dpif_backer *backer, struct dpif *dpif)
371 struct udpif *udpif = xzalloc(sizeof *udpif);
374 udpif->backer = backer;
375 atomic_init(&udpif->flow_limit, MIN(ofproto_flow_limit, 10000));
376 udpif->reval_seq = seq_create();
377 udpif->dump_seq = seq_create();
378 latch_init(&udpif->exit_latch);
379 latch_init(&udpif->pause_latch);
380 list_push_back(&all_udpifs, &udpif->list_node);
381 atomic_init(&udpif->enable_ufid, false);
382 atomic_init(&udpif->n_flows, 0);
383 atomic_init(&udpif->n_flows_timestamp, LLONG_MIN);
384 ovs_mutex_init(&udpif->n_flows_mutex);
385 udpif->ukeys = xmalloc(N_UMAPS * sizeof *udpif->ukeys);
386 for (int i = 0; i < N_UMAPS; i++) {
387 cmap_init(&udpif->ukeys[i].cmap);
388 ovs_mutex_init(&udpif->ukeys[i].mutex);
391 dpif_register_upcall_cb(dpif, upcall_cb, udpif);
392 dpif_register_dp_purge_cb(dpif, dp_purge_cb, udpif);
398 udpif_run(struct udpif *udpif)
400 if (udpif->conns && udpif->conn_seq != seq_read(udpif->dump_seq)) {
403 for (i = 0; i < udpif->n_conns; i++) {
404 unixctl_command_reply(udpif->conns[i], NULL);
413 udpif_destroy(struct udpif *udpif)
415 udpif_stop_threads(udpif);
417 for (int i = 0; i < N_UMAPS; i++) {
418 cmap_destroy(&udpif->ukeys[i].cmap);
419 ovs_mutex_destroy(&udpif->ukeys[i].mutex);
424 list_remove(&udpif->list_node);
425 latch_destroy(&udpif->exit_latch);
426 latch_destroy(&udpif->pause_latch);
427 seq_destroy(udpif->reval_seq);
428 seq_destroy(udpif->dump_seq);
429 ovs_mutex_destroy(&udpif->n_flows_mutex);
433 /* Stops the handler and revalidator threads, must be enclosed in
434 * ovsrcu quiescent state unless when destroying udpif. */
436 udpif_stop_threads(struct udpif *udpif)
438 if (udpif && (udpif->n_handlers != 0 || udpif->n_revalidators != 0)) {
441 latch_set(&udpif->exit_latch);
443 for (i = 0; i < udpif->n_handlers; i++) {
444 struct handler *handler = &udpif->handlers[i];
446 xpthread_join(handler->thread, NULL);
449 for (i = 0; i < udpif->n_revalidators; i++) {
450 xpthread_join(udpif->revalidators[i].thread, NULL);
453 dpif_disable_upcall(udpif->dpif);
455 for (i = 0; i < udpif->n_revalidators; i++) {
456 struct revalidator *revalidator = &udpif->revalidators[i];
458 /* Delete ukeys, and delete all flows from the datapath to prevent
459 * double-counting stats. */
460 revalidator_purge(revalidator);
463 latch_poll(&udpif->exit_latch);
465 ovs_barrier_destroy(&udpif->reval_barrier);
466 ovs_barrier_destroy(&udpif->pause_barrier);
468 free(udpif->revalidators);
469 udpif->revalidators = NULL;
470 udpif->n_revalidators = 0;
472 free(udpif->handlers);
473 udpif->handlers = NULL;
474 udpif->n_handlers = 0;
478 /* Starts the handler and revalidator threads, must be enclosed in
479 * ovsrcu quiescent state. */
481 udpif_start_threads(struct udpif *udpif, size_t n_handlers,
482 size_t n_revalidators)
484 if (udpif && n_handlers && n_revalidators) {
488 udpif->n_handlers = n_handlers;
489 udpif->n_revalidators = n_revalidators;
491 udpif->handlers = xzalloc(udpif->n_handlers * sizeof *udpif->handlers);
492 for (i = 0; i < udpif->n_handlers; i++) {
493 struct handler *handler = &udpif->handlers[i];
495 handler->udpif = udpif;
496 handler->handler_id = i;
497 handler->thread = ovs_thread_create(
498 "handler", udpif_upcall_handler, handler);
501 enable_ufid = ofproto_dpif_get_enable_ufid(udpif->backer);
502 atomic_init(&udpif->enable_ufid, enable_ufid);
503 dpif_enable_upcall(udpif->dpif);
505 ovs_barrier_init(&udpif->reval_barrier, udpif->n_revalidators);
506 ovs_barrier_init(&udpif->pause_barrier, udpif->n_revalidators + 1);
507 udpif->reval_exit = false;
508 udpif->pause = false;
509 udpif->revalidators = xzalloc(udpif->n_revalidators
510 * sizeof *udpif->revalidators);
511 for (i = 0; i < udpif->n_revalidators; i++) {
512 struct revalidator *revalidator = &udpif->revalidators[i];
514 revalidator->udpif = udpif;
515 revalidator->thread = ovs_thread_create(
516 "revalidator", udpif_revalidator, revalidator);
521 /* Pauses all revalidators. Should only be called by the main thread.
522 * When function returns, all revalidators are paused and will proceed
523 * only after udpif_resume_revalidators() is called. */
525 udpif_pause_revalidators(struct udpif *udpif)
527 if (ofproto_dpif_backer_enabled(udpif->backer)) {
528 latch_set(&udpif->pause_latch);
529 ovs_barrier_block(&udpif->pause_barrier);
533 /* Resumes the pausing of revalidators. Should only be called by the
536 udpif_resume_revalidators(struct udpif *udpif)
538 if (ofproto_dpif_backer_enabled(udpif->backer)) {
539 latch_poll(&udpif->pause_latch);
540 ovs_barrier_block(&udpif->pause_barrier);
544 /* Tells 'udpif' how many threads it should use to handle upcalls.
545 * 'n_handlers' and 'n_revalidators' can never be zero. 'udpif''s
546 * datapath handle must have packet reception enabled before starting
549 udpif_set_threads(struct udpif *udpif, size_t n_handlers,
550 size_t n_revalidators)
553 ovs_assert(n_handlers && n_revalidators);
555 ovsrcu_quiesce_start();
556 if (udpif->n_handlers != n_handlers
557 || udpif->n_revalidators != n_revalidators) {
558 udpif_stop_threads(udpif);
561 if (!udpif->handlers && !udpif->revalidators) {
564 error = dpif_handlers_set(udpif->dpif, n_handlers);
566 VLOG_ERR("failed to configure handlers in dpif %s: %s",
567 dpif_name(udpif->dpif), ovs_strerror(error));
571 udpif_start_threads(udpif, n_handlers, n_revalidators);
573 ovsrcu_quiesce_end();
576 /* Waits for all ongoing upcall translations to complete. This ensures that
577 * there are no transient references to any removed ofprotos (or other
578 * objects). In particular, this should be called after an ofproto is removed
579 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
581 udpif_synchronize(struct udpif *udpif)
583 /* This is stronger than necessary. It would be sufficient to ensure
584 * (somehow) that each handler and revalidator thread had passed through
585 * its main loop once. */
586 size_t n_handlers = udpif->n_handlers;
587 size_t n_revalidators = udpif->n_revalidators;
589 ovsrcu_quiesce_start();
590 udpif_stop_threads(udpif);
591 udpif_start_threads(udpif, n_handlers, n_revalidators);
592 ovsrcu_quiesce_end();
595 /* Notifies 'udpif' that something changed which may render previous
596 * xlate_actions() results invalid. */
598 udpif_revalidate(struct udpif *udpif)
600 seq_change(udpif->reval_seq);
603 /* Returns a seq which increments every time 'udpif' pulls stats from the
604 * datapath. Callers can use this to get a sense of when might be a good time
605 * to do periodic work which relies on relatively up to date statistics. */
607 udpif_dump_seq(struct udpif *udpif)
609 return udpif->dump_seq;
613 udpif_get_memory_usage(struct udpif *udpif, struct simap *usage)
617 simap_increase(usage, "handlers", udpif->n_handlers);
619 simap_increase(usage, "revalidators", udpif->n_revalidators);
620 for (i = 0; i < N_UMAPS; i++) {
621 simap_increase(usage, "udpif keys", cmap_count(&udpif->ukeys[i].cmap));
625 /* Remove flows from a single datapath. */
627 udpif_flush(struct udpif *udpif)
629 size_t n_handlers, n_revalidators;
631 n_handlers = udpif->n_handlers;
632 n_revalidators = udpif->n_revalidators;
634 ovsrcu_quiesce_start();
636 udpif_stop_threads(udpif);
637 dpif_flow_flush(udpif->dpif);
638 udpif_start_threads(udpif, n_handlers, n_revalidators);
640 ovsrcu_quiesce_end();
643 /* Removes all flows from all datapaths. */
645 udpif_flush_all_datapaths(void)
649 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
655 udpif_use_ufid(struct udpif *udpif)
659 atomic_read_relaxed(&enable_ufid, &enable);
660 return enable && ofproto_dpif_get_enable_ufid(udpif->backer);
665 udpif_get_n_flows(struct udpif *udpif)
667 long long int time, now;
668 unsigned long flow_count;
671 atomic_read_relaxed(&udpif->n_flows_timestamp, &time);
672 if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) {
673 struct dpif_dp_stats stats;
675 atomic_store_relaxed(&udpif->n_flows_timestamp, now);
676 dpif_get_dp_stats(udpif->dpif, &stats);
677 flow_count = stats.n_flows;
678 atomic_store_relaxed(&udpif->n_flows, flow_count);
679 ovs_mutex_unlock(&udpif->n_flows_mutex);
681 atomic_read_relaxed(&udpif->n_flows, &flow_count);
686 /* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH
687 * upcalls from dpif, processes the batch and installs corresponding flows
690 udpif_upcall_handler(void *arg)
692 struct handler *handler = arg;
693 struct udpif *udpif = handler->udpif;
695 while (!latch_is_set(&handler->udpif->exit_latch)) {
696 if (recv_upcalls(handler)) {
697 poll_immediate_wake();
699 dpif_recv_wait(udpif->dpif, handler->handler_id);
700 latch_wait(&udpif->exit_latch);
709 recv_upcalls(struct handler *handler)
711 struct udpif *udpif = handler->udpif;
712 uint64_t recv_stubs[UPCALL_MAX_BATCH][512 / 8];
713 struct ofpbuf recv_bufs[UPCALL_MAX_BATCH];
714 struct dpif_upcall dupcalls[UPCALL_MAX_BATCH];
715 struct upcall upcalls[UPCALL_MAX_BATCH];
716 struct flow flows[UPCALL_MAX_BATCH];
720 while (n_upcalls < UPCALL_MAX_BATCH) {
721 struct ofpbuf *recv_buf = &recv_bufs[n_upcalls];
722 struct dpif_upcall *dupcall = &dupcalls[n_upcalls];
723 struct upcall *upcall = &upcalls[n_upcalls];
724 struct flow *flow = &flows[n_upcalls];
728 ofpbuf_use_stub(recv_buf, recv_stubs[n_upcalls],
729 sizeof recv_stubs[n_upcalls]);
730 if (dpif_recv(udpif->dpif, handler->handler_id, dupcall, recv_buf)) {
731 ofpbuf_uninit(recv_buf);
735 if (odp_flow_key_to_flow(dupcall->key, dupcall->key_len, flow)
741 mru = nl_attr_get_u16(dupcall->mru);
746 error = upcall_receive(upcall, udpif->backer, &dupcall->packet,
747 dupcall->type, dupcall->userdata, flow, mru,
748 &dupcall->ufid, PMD_ID_NULL);
750 if (error == ENODEV) {
751 /* Received packet on datapath port for which we couldn't
752 * associate an ofproto. This can happen if a port is removed
753 * while traffic is being received. Print a rate-limited
754 * message in case it happens frequently. */
755 dpif_flow_put(udpif->dpif, DPIF_FP_CREATE, dupcall->key,
756 dupcall->key_len, NULL, 0, NULL, 0,
757 &dupcall->ufid, PMD_ID_NULL, NULL);
758 VLOG_INFO_RL(&rl, "received packet on unassociated datapath "
759 "port %"PRIu32, flow->in_port.odp_port);
764 upcall->key = dupcall->key;
765 upcall->key_len = dupcall->key_len;
766 upcall->ufid = &dupcall->ufid;
768 upcall->out_tun_key = dupcall->out_tun_key;
769 upcall->actions = dupcall->actions;
771 if (vsp_adjust_flow(upcall->ofproto, flow, &dupcall->packet)) {
772 upcall->vsp_adjusted = true;
775 pkt_metadata_from_flow(&dupcall->packet.md, flow);
776 flow_extract(&dupcall->packet, flow);
778 error = process_upcall(udpif, upcall,
779 &upcall->odp_actions, &upcall->wc);
788 upcall_uninit(upcall);
790 dp_packet_uninit(&dupcall->packet);
791 ofpbuf_uninit(recv_buf);
795 handle_upcalls(handler->udpif, upcalls, n_upcalls);
796 for (i = 0; i < n_upcalls; i++) {
797 dp_packet_uninit(&dupcalls[i].packet);
798 ofpbuf_uninit(&recv_bufs[i]);
799 upcall_uninit(&upcalls[i]);
807 udpif_revalidator(void *arg)
809 /* Used by all revalidators. */
810 struct revalidator *revalidator = arg;
811 struct udpif *udpif = revalidator->udpif;
812 bool leader = revalidator == &udpif->revalidators[0];
814 /* Used only by the leader. */
815 long long int start_time = 0;
816 uint64_t last_reval_seq = 0;
819 revalidator->id = ovsthread_id_self();
824 recirc_run(); /* Recirculation cleanup. */
826 reval_seq = seq_read(udpif->reval_seq);
827 last_reval_seq = reval_seq;
829 n_flows = udpif_get_n_flows(udpif);
830 udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows);
831 udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2;
833 /* Only the leader checks the pause latch to prevent a race where
834 * some threads think it's false and proceed to block on
835 * reval_barrier and others think it's true and block indefinitely
836 * on the pause_barrier */
837 udpif->pause = latch_is_set(&udpif->pause_latch);
839 /* Only the leader checks the exit latch to prevent a race where
840 * some threads think it's true and exit and others think it's
841 * false and block indefinitely on the reval_barrier */
842 udpif->reval_exit = latch_is_set(&udpif->exit_latch);
844 start_time = time_msec();
845 if (!udpif->reval_exit) {
848 terse_dump = udpif_use_ufid(udpif);
849 udpif->dump = dpif_flow_dump_create(udpif->dpif, terse_dump);
853 /* Wait for the leader to start the flow dump. */
854 ovs_barrier_block(&udpif->reval_barrier);
856 revalidator_pause(revalidator);
859 if (udpif->reval_exit) {
862 revalidate(revalidator);
864 /* Wait for all flows to have been dumped before we garbage collect. */
865 ovs_barrier_block(&udpif->reval_barrier);
866 revalidator_sweep(revalidator);
868 /* Wait for all revalidators to finish garbage collection. */
869 ovs_barrier_block(&udpif->reval_barrier);
872 unsigned int flow_limit;
873 long long int duration;
875 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
877 dpif_flow_dump_destroy(udpif->dump);
878 seq_change(udpif->dump_seq);
880 duration = MAX(time_msec() - start_time, 1);
881 udpif->dump_duration = duration;
882 if (duration > 2000) {
883 flow_limit /= duration / 1000;
884 } else if (duration > 1300) {
885 flow_limit = flow_limit * 3 / 4;
886 } else if (duration < 1000 && n_flows > 2000
887 && flow_limit < n_flows * 1000 / duration) {
890 flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000));
891 atomic_store_relaxed(&udpif->flow_limit, flow_limit);
893 if (duration > 2000) {
894 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
898 poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500));
899 seq_wait(udpif->reval_seq, last_reval_seq);
900 latch_wait(&udpif->exit_latch);
901 latch_wait(&udpif->pause_latch);
909 static enum upcall_type
910 classify_upcall(enum dpif_upcall_type type, const struct nlattr *userdata)
912 union user_action_cookie cookie;
915 /* First look at the upcall type. */
923 case DPIF_N_UC_TYPES:
925 VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32, type);
929 /* "action" upcalls need a closer look. */
931 VLOG_WARN_RL(&rl, "action upcall missing cookie");
934 userdata_len = nl_attr_get_size(userdata);
935 if (userdata_len < sizeof cookie.type
936 || userdata_len > sizeof cookie) {
937 VLOG_WARN_RL(&rl, "action upcall cookie has unexpected size %"PRIuSIZE,
941 memset(&cookie, 0, sizeof cookie);
942 memcpy(&cookie, nl_attr_get(userdata), userdata_len);
943 if (userdata_len == MAX(8, sizeof cookie.sflow)
944 && cookie.type == USER_ACTION_COOKIE_SFLOW) {
946 } else if (userdata_len == MAX(8, sizeof cookie.slow_path)
947 && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) {
949 } else if (userdata_len == MAX(8, sizeof cookie.flow_sample)
950 && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) {
951 return FLOW_SAMPLE_UPCALL;
952 } else if (userdata_len == MAX(8, sizeof cookie.ipfix)
953 && cookie.type == USER_ACTION_COOKIE_IPFIX) {
956 VLOG_WARN_RL(&rl, "invalid user cookie of type %"PRIu16
957 " and size %"PRIuSIZE, cookie.type, userdata_len);
962 /* Calculates slow path actions for 'xout'. 'buf' must statically be
963 * initialized with at least 128 bytes of space. */
965 compose_slow_path(struct udpif *udpif, struct xlate_out *xout,
966 const struct flow *flow, odp_port_t odp_in_port,
969 union user_action_cookie cookie;
973 cookie.type = USER_ACTION_COOKIE_SLOW_PATH;
974 cookie.slow_path.unused = 0;
975 cookie.slow_path.reason = xout->slow;
977 port = xout->slow & (SLOW_CFM | SLOW_BFD | SLOW_LACP | SLOW_STP)
980 pid = dpif_port_get_pid(udpif->dpif, port, flow_hash_5tuple(flow, 0));
981 odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path,
982 ODPP_NONE, false, buf);
985 /* If there is no error, the upcall must be destroyed with upcall_uninit()
986 * before quiescing, as the referred objects are guaranteed to exist only
987 * until the calling thread quiesces. Otherwise, do not call upcall_uninit()
988 * since the 'upcall->put_actions' remains uninitialized. */
990 upcall_receive(struct upcall *upcall, const struct dpif_backer *backer,
991 const struct dp_packet *packet, enum dpif_upcall_type type,
992 const struct nlattr *userdata, const struct flow *flow,
993 const unsigned int mru,
994 const ovs_u128 *ufid, const unsigned pmd_id)
998 error = xlate_lookup(backer, flow, &upcall->ofproto, &upcall->ipfix,
999 &upcall->sflow, NULL, &upcall->in_port);
1004 upcall->recirc = NULL;
1005 upcall->have_recirc_ref = false;
1006 upcall->flow = flow;
1007 upcall->packet = packet;
1008 upcall->ufid = ufid;
1009 upcall->pmd_id = pmd_id;
1010 upcall->type = type;
1011 upcall->userdata = userdata;
1012 ofpbuf_use_stub(&upcall->odp_actions, upcall->odp_actions_stub,
1013 sizeof upcall->odp_actions_stub);
1014 ofpbuf_init(&upcall->put_actions, 0);
1016 upcall->xout_initialized = false;
1017 upcall->vsp_adjusted = false;
1018 upcall->ukey_persists = false;
1020 upcall->ukey = NULL;
1022 upcall->key_len = 0;
1025 upcall->out_tun_key = NULL;
1026 upcall->actions = NULL;
1032 upcall_xlate(struct udpif *udpif, struct upcall *upcall,
1033 struct ofpbuf *odp_actions, struct flow_wildcards *wc)
1035 struct dpif_flow_stats stats;
1036 struct xlate_in xin;
1038 stats.n_packets = 1;
1039 stats.n_bytes = dp_packet_size(upcall->packet);
1040 stats.used = time_msec();
1041 stats.tcp_flags = ntohs(upcall->flow->tcp_flags);
1043 xlate_in_init(&xin, upcall->ofproto, upcall->flow, upcall->in_port, NULL,
1044 stats.tcp_flags, upcall->packet, wc, odp_actions);
1046 if (upcall->type == DPIF_UC_MISS) {
1047 xin.resubmit_stats = &stats;
1050 /* We may install a datapath flow only if we get a reference to the
1051 * recirculation context (otherwise we could have recirculation
1052 * upcalls using recirculation ID for which no context can be
1053 * found). We may still execute the flow's actions even if we
1054 * don't install the flow. */
1055 upcall->recirc = xin.recirc;
1056 upcall->have_recirc_ref = recirc_id_node_try_ref_rcu(xin.recirc);
1059 /* For non-miss upcalls, we are either executing actions (one of which
1060 * is an userspace action) for an upcall, in which case the stats have
1061 * already been taken care of, or there's a flow in the datapath which
1062 * this packet was accounted to. Presumably the revalidators will deal
1063 * with pushing its stats eventually. */
1066 upcall->dump_seq = seq_read(udpif->dump_seq);
1067 upcall->reval_seq = seq_read(udpif->reval_seq);
1068 xlate_actions(&xin, &upcall->xout);
1069 upcall->xout_initialized = true;
1071 /* Special case for fail-open mode.
1073 * If we are in fail-open mode, but we are connected to a controller too,
1074 * then we should send the packet up to the controller in the hope that it
1075 * will try to set up a flow and thereby allow us to exit fail-open.
1077 * See the top-level comment in fail-open.c for more information.
1079 * Copy packets before they are modified by execution. */
1080 if (upcall->xout.fail_open) {
1081 const struct dp_packet *packet = upcall->packet;
1082 struct ofproto_packet_in *pin;
1084 pin = xmalloc(sizeof *pin);
1085 pin->up.packet = xmemdup(dp_packet_data(packet), dp_packet_size(packet));
1086 pin->up.packet_len = dp_packet_size(packet);
1087 pin->up.reason = OFPR_NO_MATCH;
1088 pin->up.table_id = 0;
1089 pin->up.cookie = OVS_BE64_MAX;
1090 flow_get_metadata(upcall->flow, &pin->up.flow_metadata);
1091 pin->send_len = 0; /* Not used for flow table misses. */
1092 pin->miss_type = OFPROTO_PACKET_IN_NO_MISS;
1093 ofproto_dpif_send_packet_in(upcall->ofproto, pin);
1096 if (!upcall->xout.slow) {
1097 ofpbuf_use_const(&upcall->put_actions,
1098 odp_actions->data, odp_actions->size);
1100 /* upcall->put_actions already initialized by upcall_receive(). */
1101 compose_slow_path(udpif, &upcall->xout, upcall->flow,
1102 upcall->flow->in_port.odp_port,
1103 &upcall->put_actions);
1106 /* This function is also called for slow-pathed flows. As we are only
1107 * going to create new datapath flows for actual datapath misses, there is
1108 * no point in creating a ukey otherwise. */
1109 if (upcall->type == DPIF_UC_MISS) {
1110 upcall->ukey = ukey_create_from_upcall(upcall, wc);
1115 upcall_uninit(struct upcall *upcall)
1118 if (upcall->xout_initialized) {
1119 xlate_out_uninit(&upcall->xout);
1121 ofpbuf_uninit(&upcall->odp_actions);
1122 ofpbuf_uninit(&upcall->put_actions);
1124 if (!upcall->ukey_persists) {
1125 ukey_delete__(upcall->ukey);
1127 } else if (upcall->have_recirc_ref) {
1128 /* The reference was transferred to the ukey if one was created. */
1129 recirc_id_node_unref(upcall->recirc);
1135 upcall_cb(const struct dp_packet *packet, const struct flow *flow, ovs_u128 *ufid,
1136 unsigned pmd_id, enum dpif_upcall_type type,
1137 const struct nlattr *userdata, struct ofpbuf *actions,
1138 struct flow_wildcards *wc, struct ofpbuf *put_actions, void *aux)
1140 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);
1141 struct udpif *udpif = aux;
1142 unsigned int flow_limit;
1143 struct upcall upcall;
1147 atomic_read_relaxed(&enable_megaflows, &megaflow);
1148 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1150 error = upcall_receive(&upcall, udpif->backer, packet, type, userdata,
1151 flow, 0, ufid, pmd_id);
1156 error = process_upcall(udpif, &upcall, actions, wc);
1161 if (upcall.xout.slow && put_actions) {
1162 ofpbuf_put(put_actions, upcall.put_actions.data,
1163 upcall.put_actions.size);
1166 if (OVS_UNLIKELY(!megaflow)) {
1167 flow_wildcards_init_for_packet(wc, flow);
1170 if (udpif_get_n_flows(udpif) >= flow_limit) {
1171 VLOG_WARN_RL(&rl, "upcall_cb failure: datapath flow limit reached");
1176 /* Prevent miss flow installation if the key has recirculation ID but we
1177 * were not able to get a reference on it. */
1178 if (type == DPIF_UC_MISS && upcall.recirc && !upcall.have_recirc_ref) {
1179 VLOG_WARN_RL(&rl, "upcall_cb failure: no reference for recirc flow");
1184 if (upcall.ukey && !ukey_install(udpif, upcall.ukey)) {
1185 VLOG_WARN_RL(&rl, "upcall_cb failure: ukey installation fails");
1190 upcall.ukey_persists = true;
1192 upcall_uninit(&upcall);
1197 process_upcall(struct udpif *udpif, struct upcall *upcall,
1198 struct ofpbuf *odp_actions, struct flow_wildcards *wc)
1200 const struct nlattr *userdata = upcall->userdata;
1201 const struct dp_packet *packet = upcall->packet;
1202 const struct flow *flow = upcall->flow;
1204 switch (classify_upcall(upcall->type, userdata)) {
1206 upcall_xlate(udpif, upcall, odp_actions, wc);
1210 if (upcall->sflow) {
1211 union user_action_cookie cookie;
1212 const struct nlattr *actions;
1213 size_t actions_len = 0;
1214 struct dpif_sflow_actions sflow_actions;
1215 memset(&sflow_actions, 0, sizeof sflow_actions);
1216 memset(&cookie, 0, sizeof cookie);
1217 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.sflow);
1218 if (upcall->actions) {
1219 /* Actions were passed up from datapath. */
1220 actions = nl_attr_get(upcall->actions);
1221 actions_len = nl_attr_get_size(upcall->actions);
1222 if (actions && actions_len) {
1223 dpif_sflow_read_actions(flow, actions, actions_len,
1227 if (actions_len == 0) {
1228 /* Lookup actions in userspace cache. */
1229 struct udpif_key *ukey = ukey_lookup(udpif, upcall->ufid);
1231 ukey_get_actions(ukey, &actions, &actions_len);
1232 dpif_sflow_read_actions(flow, actions, actions_len,
1236 dpif_sflow_received(upcall->sflow, packet, flow,
1237 flow->in_port.odp_port, &cookie,
1238 actions_len > 0 ? &sflow_actions : NULL);
1243 if (upcall->ipfix) {
1244 union user_action_cookie cookie;
1245 struct flow_tnl output_tunnel_key;
1247 memset(&cookie, 0, sizeof cookie);
1248 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.ipfix);
1250 if (upcall->out_tun_key) {
1251 odp_tun_key_from_attr(upcall->out_tun_key, false,
1252 &output_tunnel_key);
1254 dpif_ipfix_bridge_sample(upcall->ipfix, packet, flow,
1255 flow->in_port.odp_port,
1256 cookie.ipfix.output_odp_port,
1257 upcall->out_tun_key ?
1258 &output_tunnel_key : NULL);
1262 case FLOW_SAMPLE_UPCALL:
1263 if (upcall->ipfix) {
1264 union user_action_cookie cookie;
1266 memset(&cookie, 0, sizeof cookie);
1267 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.flow_sample);
1269 /* The flow reflects exactly the contents of the packet.
1270 * Sample the packet using it. */
1271 dpif_ipfix_flow_sample(upcall->ipfix, packet, flow,
1272 cookie.flow_sample.collector_set_id,
1273 cookie.flow_sample.probability,
1274 cookie.flow_sample.obs_domain_id,
1275 cookie.flow_sample.obs_point_id);
1287 handle_upcalls(struct udpif *udpif, struct upcall *upcalls,
1290 struct dpif_op *opsp[UPCALL_MAX_BATCH * 2];
1291 struct ukey_op ops[UPCALL_MAX_BATCH * 2];
1292 unsigned int flow_limit;
1293 size_t n_ops, n_opsp, i;
1296 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1298 may_put = udpif_get_n_flows(udpif) < flow_limit;
1300 /* Handle the packets individually in order of arrival.
1302 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1303 * processes received packets for these protocols.
1305 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1308 * The loop fills 'ops' with an array of operations to execute in the
1311 for (i = 0; i < n_upcalls; i++) {
1312 struct upcall *upcall = &upcalls[i];
1313 const struct dp_packet *packet = upcall->packet;
1316 if (upcall->vsp_adjusted) {
1317 /* This packet was received on a VLAN splinter port. We added a
1318 * VLAN to the packet to make the packet resemble the flow, but the
1319 * actions were composed assuming that the packet contained no
1320 * VLAN. So, we must remove the VLAN header from the packet before
1321 * trying to execute the actions. */
1322 if (upcall->odp_actions.size) {
1323 eth_pop_vlan(CONST_CAST(struct dp_packet *, upcall->packet));
1326 /* Remove the flow vlan tags inserted by vlan splinter logic
1327 * to ensure megaflow masks generated match the data path flow. */
1328 CONST_CAST(struct flow *, upcall->flow)->vlan_tci = 0;
1331 /* Do not install a flow into the datapath if:
1333 * - The datapath already has too many flows.
1335 * - We received this packet via some flow installed in the kernel
1338 * - Upcall was a recirculation but we do not have a reference to
1339 * to the recirculation ID. */
1340 if (may_put && upcall->type == DPIF_UC_MISS &&
1341 (!upcall->recirc || upcall->have_recirc_ref)) {
1342 struct udpif_key *ukey = upcall->ukey;
1344 upcall->ukey_persists = true;
1348 op->dop.type = DPIF_OP_FLOW_PUT;
1349 op->dop.u.flow_put.flags = DPIF_FP_CREATE;
1350 op->dop.u.flow_put.key = ukey->key;
1351 op->dop.u.flow_put.key_len = ukey->key_len;
1352 op->dop.u.flow_put.mask = ukey->mask;
1353 op->dop.u.flow_put.mask_len = ukey->mask_len;
1354 op->dop.u.flow_put.ufid = upcall->ufid;
1355 op->dop.u.flow_put.stats = NULL;
1356 ukey_get_actions(ukey, &op->dop.u.flow_put.actions,
1357 &op->dop.u.flow_put.actions_len);
1360 if (upcall->odp_actions.size) {
1363 op->dop.type = DPIF_OP_EXECUTE;
1364 op->dop.u.execute.packet = CONST_CAST(struct dp_packet *, packet);
1365 odp_key_to_pkt_metadata(upcall->key, upcall->key_len,
1366 &op->dop.u.execute.packet->md);
1367 op->dop.u.execute.actions = upcall->odp_actions.data;
1368 op->dop.u.execute.actions_len = upcall->odp_actions.size;
1369 op->dop.u.execute.needs_help = (upcall->xout.slow & SLOW_ACTION) != 0;
1370 op->dop.u.execute.probe = false;
1371 op->dop.u.execute.mtu = upcall->mru;
1377 * We install ukeys before installing the flows, locking them for exclusive
1378 * access by this thread for the period of installation. This ensures that
1379 * other threads won't attempt to delete the flows as we are creating them.
1382 for (i = 0; i < n_ops; i++) {
1383 struct udpif_key *ukey = ops[i].ukey;
1386 /* If we can't install the ukey, don't install the flow. */
1387 if (!ukey_install_start(udpif, ukey)) {
1388 ukey_delete__(ukey);
1393 opsp[n_opsp++] = &ops[i].dop;
1395 dpif_operate(udpif->dpif, opsp, n_opsp);
1396 for (i = 0; i < n_ops; i++) {
1398 ukey_install_finish(ops[i].ukey, ops[i].dop.error);
1404 get_ufid_hash(const ovs_u128 *ufid)
1406 return ufid->u32[0];
1409 static struct udpif_key *
1410 ukey_lookup(struct udpif *udpif, const ovs_u128 *ufid)
1412 struct udpif_key *ukey;
1413 int idx = get_ufid_hash(ufid) % N_UMAPS;
1414 struct cmap *cmap = &udpif->ukeys[idx].cmap;
1416 CMAP_FOR_EACH_WITH_HASH (ukey, cmap_node, get_ufid_hash(ufid), cmap) {
1417 if (ovs_u128_equals(&ukey->ufid, ufid)) {
1424 /* Provides safe lockless access of RCU protected 'ukey->actions'. Callers may
1425 * alternatively access the field directly if they take 'ukey->mutex'. */
1427 ukey_get_actions(struct udpif_key *ukey, const struct nlattr **actions, size_t *size)
1429 const struct ofpbuf *buf = ovsrcu_get(struct ofpbuf *, &ukey->actions);
1430 *actions = buf->data;
1435 ukey_set_actions(struct udpif_key *ukey, const struct ofpbuf *actions)
1437 ovsrcu_postpone(ofpbuf_delete,
1438 ovsrcu_get_protected(struct ofpbuf *, &ukey->actions));
1439 ovsrcu_set(&ukey->actions, ofpbuf_clone(actions));
1442 static struct udpif_key *
1443 ukey_create__(const struct nlattr *key, size_t key_len,
1444 const struct nlattr *mask, size_t mask_len,
1445 bool ufid_present, const ovs_u128 *ufid,
1446 const unsigned pmd_id, const struct ofpbuf *actions,
1447 uint64_t dump_seq, uint64_t reval_seq, long long int used,
1448 uint32_t key_recirc_id, struct xlate_out *xout)
1449 OVS_NO_THREAD_SAFETY_ANALYSIS
1451 struct udpif_key *ukey = xmalloc(sizeof *ukey);
1453 memcpy(&ukey->keybuf, key, key_len);
1454 ukey->key = &ukey->keybuf.nla;
1455 ukey->key_len = key_len;
1456 memcpy(&ukey->maskbuf, mask, mask_len);
1457 ukey->mask = &ukey->maskbuf.nla;
1458 ukey->mask_len = mask_len;
1459 ukey->ufid_present = ufid_present;
1461 ukey->pmd_id = pmd_id;
1462 ukey->hash = get_ufid_hash(&ukey->ufid);
1464 ovsrcu_init(&ukey->actions, NULL);
1465 ukey_set_actions(ukey, actions);
1467 ovs_mutex_init(&ukey->mutex);
1468 ukey->dump_seq = dump_seq;
1469 ukey->reval_seq = reval_seq;
1470 ukey->flow_exists = false;
1471 ukey->created = time_msec();
1472 memset(&ukey->stats, 0, sizeof ukey->stats);
1473 ukey->stats.used = used;
1474 ukey->xcache = NULL;
1476 ukey->key_recirc_id = key_recirc_id;
1477 recirc_refs_init(&ukey->recircs);
1479 /* Take ownership of the action recirc id references. */
1480 recirc_refs_swap(&ukey->recircs, &xout->recircs);
1486 static struct udpif_key *
1487 ukey_create_from_upcall(struct upcall *upcall, struct flow_wildcards *wc)
1489 struct odputil_keybuf keystub, maskstub;
1490 struct ofpbuf keybuf, maskbuf;
1492 struct odp_flow_key_parms odp_parms = {
1493 .flow = upcall->flow,
1497 odp_parms.support = ofproto_dpif_get_support(upcall->ofproto)->odp;
1498 if (upcall->key_len) {
1499 ofpbuf_use_const(&keybuf, upcall->key, upcall->key_len);
1501 /* dpif-netdev doesn't provide a netlink-formatted flow key in the
1502 * upcall, so convert the upcall's flow here. */
1503 ofpbuf_use_stack(&keybuf, &keystub, sizeof keystub);
1504 odp_parms.odp_in_port = upcall->flow->in_port.odp_port;
1505 odp_flow_key_from_flow(&odp_parms, &keybuf);
1508 atomic_read_relaxed(&enable_megaflows, &megaflow);
1509 ofpbuf_use_stack(&maskbuf, &maskstub, sizeof maskstub);
1511 odp_parms.odp_in_port = ODPP_NONE;
1512 odp_parms.key_buf = &keybuf;
1514 odp_flow_key_from_mask(&odp_parms, &maskbuf);
1517 return ukey_create__(keybuf.data, keybuf.size, maskbuf.data, maskbuf.size,
1518 true, upcall->ufid, upcall->pmd_id,
1519 &upcall->put_actions, upcall->dump_seq,
1520 upcall->reval_seq, 0,
1521 upcall->have_recirc_ref ? upcall->recirc->id : 0,
1526 ukey_create_from_dpif_flow(const struct udpif *udpif,
1527 const struct dpif_flow *flow,
1528 struct udpif_key **ukey)
1530 struct dpif_flow full_flow;
1531 struct ofpbuf actions;
1532 uint64_t dump_seq, reval_seq;
1533 uint64_t stub[DPIF_FLOW_BUFSIZE / 8];
1534 const struct nlattr *a;
1537 if (!flow->key_len || !flow->actions_len) {
1541 /* If the key or actions were not provided by the datapath, fetch the
1543 ofpbuf_use_stack(&buf, &stub, sizeof stub);
1544 err = dpif_flow_get(udpif->dpif, NULL, 0, &flow->ufid,
1545 flow->pmd_id, &buf, &full_flow);
1552 /* Check the flow actions for recirculation action. As recirculation
1553 * relies on OVS userspace internal state, we need to delete all old
1554 * datapath flows with either a non-zero recirc_id in the key, or any
1555 * recirculation actions upon OVS restart. */
1556 NL_ATTR_FOR_EACH_UNSAFE (a, left, flow->key, flow->key_len) {
1557 if (nl_attr_type(a) == OVS_KEY_ATTR_RECIRC_ID
1558 && nl_attr_get_u32(a) != 0) {
1562 NL_ATTR_FOR_EACH_UNSAFE (a, left, flow->actions, flow->actions_len) {
1563 if (nl_attr_type(a) == OVS_ACTION_ATTR_RECIRC) {
1568 dump_seq = seq_read(udpif->dump_seq);
1569 reval_seq = seq_read(udpif->reval_seq);
1570 ofpbuf_use_const(&actions, &flow->actions, flow->actions_len);
1571 *ukey = ukey_create__(flow->key, flow->key_len,
1572 flow->mask, flow->mask_len, flow->ufid_present,
1573 &flow->ufid, flow->pmd_id, &actions, dump_seq,
1574 reval_seq, flow->stats.used, 0, NULL);
1579 /* Attempts to insert a ukey into the shared ukey maps.
1581 * On success, returns true, installs the ukey and returns it in a locked
1582 * state. Otherwise, returns false. */
1584 ukey_install_start(struct udpif *udpif, struct udpif_key *new_ukey)
1585 OVS_TRY_LOCK(true, new_ukey->mutex)
1588 struct udpif_key *old_ukey;
1590 bool locked = false;
1592 idx = new_ukey->hash % N_UMAPS;
1593 umap = &udpif->ukeys[idx];
1594 ovs_mutex_lock(&umap->mutex);
1595 old_ukey = ukey_lookup(udpif, &new_ukey->ufid);
1597 /* Uncommon case: A ukey is already installed with the same UFID. */
1598 if (old_ukey->key_len == new_ukey->key_len
1599 && !memcmp(old_ukey->key, new_ukey->key, new_ukey->key_len)) {
1600 COVERAGE_INC(handler_duplicate_upcall);
1602 struct ds ds = DS_EMPTY_INITIALIZER;
1604 odp_format_ufid(&old_ukey->ufid, &ds);
1605 ds_put_cstr(&ds, " ");
1606 odp_flow_key_format(old_ukey->key, old_ukey->key_len, &ds);
1607 ds_put_cstr(&ds, "\n");
1608 odp_format_ufid(&new_ukey->ufid, &ds);
1609 ds_put_cstr(&ds, " ");
1610 odp_flow_key_format(new_ukey->key, new_ukey->key_len, &ds);
1612 VLOG_WARN_RL(&rl, "Conflicting ukey for flows:\n%s", ds_cstr(&ds));
1616 ovs_mutex_lock(&new_ukey->mutex);
1617 cmap_insert(&umap->cmap, &new_ukey->cmap_node, new_ukey->hash);
1620 ovs_mutex_unlock(&umap->mutex);
1626 ukey_install_finish__(struct udpif_key *ukey) OVS_REQUIRES(ukey->mutex)
1628 ukey->flow_exists = true;
1632 ukey_install_finish(struct udpif_key *ukey, int error)
1633 OVS_RELEASES(ukey->mutex)
1636 ukey_install_finish__(ukey);
1638 ovs_mutex_unlock(&ukey->mutex);
1644 ukey_install(struct udpif *udpif, struct udpif_key *ukey)
1646 /* The usual way to keep 'ukey->flow_exists' in sync with the datapath is
1647 * to call ukey_install_start(), install the corresponding datapath flow,
1648 * then call ukey_install_finish(). The netdev interface using upcall_cb()
1649 * doesn't provide a function to separately finish the flow installation,
1650 * so we perform the operations together here.
1652 * This is fine currently, as revalidator threads will only delete this
1653 * ukey during revalidator_sweep() and only if the dump_seq is mismatched.
1654 * It is unlikely for a revalidator thread to advance dump_seq and reach
1655 * the next GC phase between ukey creation and flow installation. */
1656 return ukey_install_start(udpif, ukey) && ukey_install_finish(ukey, 0);
1659 /* Searches for a ukey in 'udpif->ukeys' that matches 'flow' and attempts to
1660 * lock the ukey. If the ukey does not exist, create it.
1662 * Returns 0 on success, setting *result to the matching ukey and returning it
1663 * in a locked state. Otherwise, returns an errno and clears *result. EBUSY
1664 * indicates that another thread is handling this flow. Other errors indicate
1665 * an unexpected condition creating a new ukey.
1667 * *error is an output parameter provided to appease the threadsafety analyser,
1668 * and its value matches the return value. */
1670 ukey_acquire(struct udpif *udpif, const struct dpif_flow *flow,
1671 struct udpif_key **result, int *error)
1672 OVS_TRY_LOCK(0, (*result)->mutex)
1674 struct udpif_key *ukey;
1677 ukey = ukey_lookup(udpif, &flow->ufid);
1679 retval = ovs_mutex_trylock(&ukey->mutex);
1681 /* Usually we try to avoid installing flows from revalidator threads,
1682 * because locking on a umap may cause handler threads to block.
1683 * However there are certain cases, like when ovs-vswitchd is
1684 * restarted, where it is desirable to handle flows that exist in the
1685 * datapath gracefully (ie, don't just clear the datapath). */
1688 retval = ukey_create_from_dpif_flow(udpif, flow, &ukey);
1692 install = ukey_install_start(udpif, ukey);
1694 ukey_install_finish__(ukey);
1697 ukey_delete__(ukey);
1713 ukey_delete__(struct udpif_key *ukey)
1714 OVS_NO_THREAD_SAFETY_ANALYSIS
1717 if (ukey->key_recirc_id) {
1718 recirc_free_id(ukey->key_recirc_id);
1720 recirc_refs_unref(&ukey->recircs);
1721 xlate_cache_delete(ukey->xcache);
1722 ofpbuf_delete(ovsrcu_get(struct ofpbuf *, &ukey->actions));
1723 ovs_mutex_destroy(&ukey->mutex);
1729 ukey_delete(struct umap *umap, struct udpif_key *ukey)
1730 OVS_REQUIRES(umap->mutex)
1732 cmap_remove(&umap->cmap, &ukey->cmap_node, ukey->hash);
1733 ovsrcu_postpone(ukey_delete__, ukey);
1737 should_revalidate(const struct udpif *udpif, uint64_t packets,
1740 long long int metric, now, duration;
1742 if (udpif->dump_duration < 200) {
1743 /* We are likely to handle full revalidation for the flows. */
1747 /* Calculate the mean time between seeing these packets. If this
1748 * exceeds the threshold, then delete the flow rather than performing
1749 * costly revalidation for flows that aren't being hit frequently.
1751 * This is targeted at situations where the dump_duration is high (~1s),
1752 * and revalidation is triggered by a call to udpif_revalidate(). In
1753 * these situations, revalidation of all flows causes fluctuations in the
1754 * flow_limit due to the interaction with the dump_duration and max_idle.
1755 * This tends to result in deletion of low-throughput flows anyway, so
1756 * skip the revalidation and just delete those flows. */
1757 packets = MAX(packets, 1);
1758 now = MAX(used, time_msec());
1759 duration = now - used;
1760 metric = duration / packets;
1763 /* The flow is receiving more than ~5pps, so keep it. */
1769 /* Verifies that the datapath actions of 'ukey' are still correct, and pushes
1772 * Returns a recommended action for 'ukey', options include:
1773 * UKEY_DELETE The ukey should be deleted.
1774 * UKEY_KEEP The ukey is fine as is.
1775 * UKEY_MODIFY The ukey's actions should be changed but is otherwise
1776 * fine. Callers should change the actions to those found
1777 * in the caller supplied 'odp_actions' buffer. The
1778 * recirculation references can be found in 'recircs' and
1779 * must be handled by the caller.
1781 * If the result is UKEY_MODIFY, then references to all recirc_ids used by the
1782 * new flow will be held within 'recircs' (which may be none).
1784 * The caller is responsible for both initializing 'recircs' prior this call,
1785 * and ensuring any references are eventually freed.
1787 static enum reval_result
1788 revalidate_ukey(struct udpif *udpif, struct udpif_key *ukey,
1789 const struct dpif_flow_stats *stats,
1790 struct ofpbuf *odp_actions, uint64_t reval_seq,
1791 struct recirc_refs *recircs)
1792 OVS_REQUIRES(ukey->mutex)
1794 struct xlate_out xout, *xoutp;
1795 struct netflow *netflow;
1796 struct ofproto_dpif *ofproto;
1797 struct dpif_flow_stats push;
1799 struct flow_wildcards dp_mask, wc;
1800 enum reval_result result;
1801 ofp_port_t ofp_in_port;
1802 struct xlate_in xin;
1803 long long int last_used;
1805 bool need_revalidate;
1807 result = UKEY_DELETE;
1811 ofpbuf_clear(odp_actions);
1812 need_revalidate = (ukey->reval_seq != reval_seq);
1813 last_used = ukey->stats.used;
1814 push.used = stats->used;
1815 push.tcp_flags = stats->tcp_flags;
1816 push.n_packets = (stats->n_packets > ukey->stats.n_packets
1817 ? stats->n_packets - ukey->stats.n_packets
1819 push.n_bytes = (stats->n_bytes > ukey->stats.n_bytes
1820 ? stats->n_bytes - ukey->stats.n_bytes
1823 if (need_revalidate && last_used
1824 && !should_revalidate(udpif, push.n_packets, last_used)) {
1828 /* We will push the stats, so update the ukey stats cache. */
1829 ukey->stats = *stats;
1830 if (!push.n_packets && !need_revalidate) {
1835 if (ukey->xcache && !need_revalidate) {
1836 xlate_push_stats(ukey->xcache, &push);
1841 if (odp_flow_key_to_flow(ukey->key, ukey->key_len, &flow)
1846 error = xlate_lookup(udpif->backer, &flow, &ofproto, NULL, NULL, &netflow,
1852 if (need_revalidate) {
1853 xlate_cache_clear(ukey->xcache);
1855 if (!ukey->xcache) {
1856 ukey->xcache = xlate_cache_new();
1859 xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL, push.tcp_flags,
1860 NULL, need_revalidate ? &wc : NULL, odp_actions);
1861 if (push.n_packets) {
1862 xin.resubmit_stats = &push;
1863 xin.may_learn = true;
1865 xin.xcache = ukey->xcache;
1866 xlate_actions(&xin, &xout);
1869 if (!need_revalidate) {
1875 ofpbuf_clear(odp_actions);
1876 compose_slow_path(udpif, &xout, &flow, flow.in_port.odp_port,
1880 if (odp_flow_key_to_mask(ukey->mask, ukey->mask_len, ukey->key,
1881 ukey->key_len, &dp_mask, &flow)
1886 /* Do not modify if any bit is wildcarded by the installed datapath flow,
1887 * but not the newly revalidated wildcard mask (wc), i.e., if revalidation
1888 * tells that the datapath flow is now too generic and must be narrowed
1889 * down. Note that we do not know if the datapath has ignored any of the
1890 * wildcarded bits, so we may be overtly conservative here. */
1891 if (flow_wildcards_has_extra(&dp_mask, &wc)) {
1895 if (!ofpbuf_equal(odp_actions,
1896 ovsrcu_get(struct ofpbuf *, &ukey->actions))) {
1897 /* The datapath mask was OK, but the actions seem to have changed.
1898 * Let's modify it in place. */
1899 result = UKEY_MODIFY;
1900 /* Transfer recirc action ID references to the caller. */
1901 recirc_refs_swap(recircs, &xoutp->recircs);
1908 if (result != UKEY_DELETE) {
1909 ukey->reval_seq = reval_seq;
1911 if (netflow && result == UKEY_DELETE) {
1912 netflow_flow_clear(netflow, &flow);
1914 xlate_out_uninit(xoutp);
1919 delete_op_init__(struct udpif *udpif, struct ukey_op *op,
1920 const struct dpif_flow *flow)
1923 op->dop.type = DPIF_OP_FLOW_DEL;
1924 op->dop.u.flow_del.key = flow->key;
1925 op->dop.u.flow_del.key_len = flow->key_len;
1926 op->dop.u.flow_del.ufid = flow->ufid_present ? &flow->ufid : NULL;
1927 op->dop.u.flow_del.pmd_id = flow->pmd_id;
1928 op->dop.u.flow_del.stats = &op->stats;
1929 op->dop.u.flow_del.terse = udpif_use_ufid(udpif);
1933 delete_op_init(struct udpif *udpif, struct ukey_op *op, struct udpif_key *ukey)
1936 op->dop.type = DPIF_OP_FLOW_DEL;
1937 op->dop.u.flow_del.key = ukey->key;
1938 op->dop.u.flow_del.key_len = ukey->key_len;
1939 op->dop.u.flow_del.ufid = ukey->ufid_present ? &ukey->ufid : NULL;
1940 op->dop.u.flow_del.pmd_id = ukey->pmd_id;
1941 op->dop.u.flow_del.stats = &op->stats;
1942 op->dop.u.flow_del.terse = udpif_use_ufid(udpif);
1946 modify_op_init(struct ukey_op *op, struct udpif_key *ukey)
1949 op->dop.type = DPIF_OP_FLOW_PUT;
1950 op->dop.u.flow_put.flags = DPIF_FP_MODIFY;
1951 op->dop.u.flow_put.key = ukey->key;
1952 op->dop.u.flow_put.key_len = ukey->key_len;
1953 op->dop.u.flow_put.mask = ukey->mask;
1954 op->dop.u.flow_put.mask_len = ukey->mask_len;
1955 op->dop.u.flow_put.ufid = &ukey->ufid;
1956 op->dop.u.flow_put.pmd_id = ukey->pmd_id;
1957 op->dop.u.flow_put.stats = NULL;
1958 ukey_get_actions(ukey, &op->dop.u.flow_put.actions,
1959 &op->dop.u.flow_put.actions_len);
1963 push_ukey_ops__(struct udpif *udpif, struct ukey_op *ops, size_t n_ops)
1965 struct dpif_op *opsp[REVALIDATE_MAX_BATCH];
1968 ovs_assert(n_ops <= REVALIDATE_MAX_BATCH);
1969 for (i = 0; i < n_ops; i++) {
1970 opsp[i] = &ops[i].dop;
1972 dpif_operate(udpif->dpif, opsp, n_ops);
1974 for (i = 0; i < n_ops; i++) {
1975 struct ukey_op *op = &ops[i];
1976 struct dpif_flow_stats *push, *stats, push_buf;
1978 stats = op->dop.u.flow_del.stats;
1981 if (op->dop.type != DPIF_OP_FLOW_DEL) {
1982 /* Only deleted flows need their stats pushed. */
1986 if (op->dop.error) {
1987 /* flow_del error, 'stats' is unusable. */
1992 ovs_mutex_lock(&op->ukey->mutex);
1993 push->used = MAX(stats->used, op->ukey->stats.used);
1994 push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags;
1995 push->n_packets = stats->n_packets - op->ukey->stats.n_packets;
1996 push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes;
1997 ovs_mutex_unlock(&op->ukey->mutex);
2002 if (push->n_packets || netflow_exists()) {
2003 const struct nlattr *key = op->dop.u.flow_del.key;
2004 size_t key_len = op->dop.u.flow_del.key_len;
2005 struct ofproto_dpif *ofproto;
2006 struct netflow *netflow;
2007 ofp_port_t ofp_in_port;
2012 ovs_mutex_lock(&op->ukey->mutex);
2013 if (op->ukey->xcache) {
2014 xlate_push_stats(op->ukey->xcache, push);
2015 ovs_mutex_unlock(&op->ukey->mutex);
2018 ovs_mutex_unlock(&op->ukey->mutex);
2019 key = op->ukey->key;
2020 key_len = op->ukey->key_len;
2023 if (odp_flow_key_to_flow(key, key_len, &flow)
2028 error = xlate_lookup(udpif->backer, &flow, &ofproto, NULL, NULL,
2029 &netflow, &ofp_in_port);
2031 struct xlate_in xin;
2033 xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL,
2034 push->tcp_flags, NULL, NULL, NULL);
2035 xin.resubmit_stats = push->n_packets ? push : NULL;
2036 xin.may_learn = push->n_packets > 0;
2037 xlate_actions_for_side_effects(&xin);
2040 netflow_flow_clear(netflow, &flow);
2048 push_ukey_ops(struct udpif *udpif, struct umap *umap,
2049 struct ukey_op *ops, size_t n_ops)
2053 push_ukey_ops__(udpif, ops, n_ops);
2054 ovs_mutex_lock(&umap->mutex);
2055 for (i = 0; i < n_ops; i++) {
2056 if (ops[i].dop.type == DPIF_OP_FLOW_DEL) {
2057 ukey_delete(umap, ops[i].ukey);
2060 ovs_mutex_unlock(&umap->mutex);
2064 log_unexpected_flow(const struct dpif_flow *flow, int error)
2066 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(10, 60);
2067 struct ds ds = DS_EMPTY_INITIALIZER;
2069 ds_put_format(&ds, "Failed to acquire udpif_key corresponding to "
2070 "unexpected flow (%s): ", ovs_strerror(error));
2071 odp_format_ufid(&flow->ufid, &ds);
2072 VLOG_WARN_RL(&rl, "%s", ds_cstr(&ds));
2076 reval_op_init(struct ukey_op *op, enum reval_result result,
2077 struct udpif *udpif, struct udpif_key *ukey,
2078 struct recirc_refs *recircs, struct ofpbuf *odp_actions)
2080 if (result == UKEY_DELETE) {
2081 delete_op_init(udpif, op, ukey);
2082 } else if (result == UKEY_MODIFY) {
2083 /* Store the new recircs. */
2084 recirc_refs_swap(&ukey->recircs, recircs);
2085 /* Release old recircs. */
2086 recirc_refs_unref(recircs);
2087 /* ukey->key_recirc_id remains, as the key is the same as before. */
2089 ukey_set_actions(ukey, odp_actions);
2090 modify_op_init(op, ukey);
2095 revalidate(struct revalidator *revalidator)
2097 uint64_t odp_actions_stub[1024 / 8];
2098 struct ofpbuf odp_actions = OFPBUF_STUB_INITIALIZER(odp_actions_stub);
2100 struct udpif *udpif = revalidator->udpif;
2101 struct dpif_flow_dump_thread *dump_thread;
2102 uint64_t dump_seq, reval_seq;
2103 unsigned int flow_limit;
2105 dump_seq = seq_read(udpif->dump_seq);
2106 reval_seq = seq_read(udpif->reval_seq);
2107 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
2108 dump_thread = dpif_flow_dump_thread_create(udpif->dump);
2110 struct ukey_op ops[REVALIDATE_MAX_BATCH];
2113 struct dpif_flow flows[REVALIDATE_MAX_BATCH];
2114 const struct dpif_flow *f;
2117 long long int max_idle;
2122 n_dumped = dpif_flow_dump_next(dump_thread, flows, ARRAY_SIZE(flows));
2129 /* In normal operation we want to keep flows around until they have
2130 * been idle for 'ofproto_max_idle' milliseconds. However:
2132 * - If the number of datapath flows climbs above 'flow_limit',
2133 * drop that down to 100 ms to try to bring the flows down to
2136 * - If the number of datapath flows climbs above twice
2137 * 'flow_limit', delete all the datapath flows as an emergency
2138 * measure. (We reassess this condition for the next batch of
2139 * datapath flows, so we will recover before all the flows are
2141 n_dp_flows = udpif_get_n_flows(udpif);
2142 kill_them_all = n_dp_flows > flow_limit * 2;
2143 max_idle = n_dp_flows > flow_limit ? 100 : ofproto_max_idle;
2145 for (f = flows; f < &flows[n_dumped]; f++) {
2146 long long int used = f->stats.used;
2147 struct recirc_refs recircs = RECIRC_REFS_EMPTY_INITIALIZER;
2148 enum reval_result result;
2149 struct udpif_key *ukey;
2150 bool already_dumped;
2153 if (ukey_acquire(udpif, f, &ukey, &error)) {
2154 if (error == EBUSY) {
2155 /* Another thread is processing this flow, so don't bother
2157 COVERAGE_INC(upcall_ukey_contention);
2159 log_unexpected_flow(f, error);
2160 if (error != ENOENT) {
2161 delete_op_init__(udpif, &ops[n_ops++], f);
2167 already_dumped = ukey->dump_seq == dump_seq;
2168 if (already_dumped) {
2169 /* The flow has already been handled during this flow dump
2170 * operation. Skip it. */
2172 COVERAGE_INC(dumped_duplicate_flow);
2174 COVERAGE_INC(dumped_new_flow);
2176 ovs_mutex_unlock(&ukey->mutex);
2181 used = ukey->created;
2183 if (kill_them_all || (used && used < now - max_idle)) {
2184 result = UKEY_DELETE;
2186 result = revalidate_ukey(udpif, ukey, &f->stats, &odp_actions,
2187 reval_seq, &recircs);
2189 ukey->dump_seq = dump_seq;
2190 ukey->flow_exists = result != UKEY_DELETE;
2192 if (result != UKEY_KEEP) {
2193 /* Takes ownership of 'recircs'. */
2194 reval_op_init(&ops[n_ops++], result, udpif, ukey, &recircs,
2197 ovs_mutex_unlock(&ukey->mutex);
2201 push_ukey_ops__(udpif, ops, n_ops);
2205 dpif_flow_dump_thread_destroy(dump_thread);
2206 ofpbuf_uninit(&odp_actions);
2209 /* Pauses the 'revalidator', can only proceed after main thread
2210 * calls udpif_resume_revalidators(). */
2212 revalidator_pause(struct revalidator *revalidator)
2214 /* The first block is for sync'ing the pause with main thread. */
2215 ovs_barrier_block(&revalidator->udpif->pause_barrier);
2216 /* The second block is for pausing until main thread resumes. */
2217 ovs_barrier_block(&revalidator->udpif->pause_barrier);
2221 revalidator_sweep__(struct revalidator *revalidator, bool purge)
2223 struct udpif *udpif;
2224 uint64_t dump_seq, reval_seq;
2227 udpif = revalidator->udpif;
2228 dump_seq = seq_read(udpif->dump_seq);
2229 reval_seq = seq_read(udpif->reval_seq);
2230 slice = revalidator - udpif->revalidators;
2231 ovs_assert(slice < udpif->n_revalidators);
2233 for (int i = slice; i < N_UMAPS; i += udpif->n_revalidators) {
2234 uint64_t odp_actions_stub[1024 / 8];
2235 struct ofpbuf odp_actions = OFPBUF_STUB_INITIALIZER(odp_actions_stub);
2237 struct ukey_op ops[REVALIDATE_MAX_BATCH];
2238 struct udpif_key *ukey;
2239 struct umap *umap = &udpif->ukeys[i];
2242 CMAP_FOR_EACH(ukey, cmap_node, &umap->cmap) {
2243 bool flow_exists, seq_mismatch;
2244 struct recirc_refs recircs = RECIRC_REFS_EMPTY_INITIALIZER;
2245 enum reval_result result;
2247 /* Handler threads could be holding a ukey lock while it installs a
2248 * new flow, so don't hang around waiting for access to it. */
2249 if (ovs_mutex_trylock(&ukey->mutex)) {
2252 flow_exists = ukey->flow_exists;
2253 seq_mismatch = (ukey->dump_seq != dump_seq
2254 && ukey->reval_seq != reval_seq);
2257 result = UKEY_DELETE;
2258 } else if (!seq_mismatch) {
2261 struct dpif_flow_stats stats;
2262 COVERAGE_INC(revalidate_missed_dp_flow);
2263 memset(&stats, 0, sizeof stats);
2264 result = revalidate_ukey(udpif, ukey, &stats, &odp_actions,
2265 reval_seq, &recircs);
2267 if (flow_exists && result != UKEY_KEEP) {
2268 /* Takes ownership of 'recircs'. */
2269 reval_op_init(&ops[n_ops++], result, udpif, ukey, &recircs,
2272 ovs_mutex_unlock(&ukey->mutex);
2274 if (n_ops == REVALIDATE_MAX_BATCH) {
2275 push_ukey_ops(udpif, umap, ops, n_ops);
2280 ovs_mutex_lock(&umap->mutex);
2281 ukey_delete(umap, ukey);
2282 ovs_mutex_unlock(&umap->mutex);
2287 push_ukey_ops(udpif, umap, ops, n_ops);
2290 ofpbuf_uninit(&odp_actions);
2296 revalidator_sweep(struct revalidator *revalidator)
2298 revalidator_sweep__(revalidator, false);
2302 revalidator_purge(struct revalidator *revalidator)
2304 revalidator_sweep__(revalidator, true);
2307 /* In reaction to dpif purge, purges all 'ukey's with same 'pmd_id'. */
2309 dp_purge_cb(void *aux, unsigned pmd_id)
2311 struct udpif *udpif = aux;
2314 udpif_pause_revalidators(udpif);
2315 for (i = 0; i < N_UMAPS; i++) {
2316 struct ukey_op ops[REVALIDATE_MAX_BATCH];
2317 struct udpif_key *ukey;
2318 struct umap *umap = &udpif->ukeys[i];
2321 CMAP_FOR_EACH(ukey, cmap_node, &umap->cmap) {
2322 if (ukey->pmd_id == pmd_id) {
2323 delete_op_init(udpif, &ops[n_ops++], ukey);
2324 if (n_ops == REVALIDATE_MAX_BATCH) {
2325 push_ukey_ops(udpif, umap, ops, n_ops);
2332 push_ukey_ops(udpif, umap, ops, n_ops);
2337 udpif_resume_revalidators(udpif);
2341 upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED,
2342 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
2344 struct ds ds = DS_EMPTY_INITIALIZER;
2345 struct udpif *udpif;
2347 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
2348 unsigned int flow_limit;
2352 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
2353 ufid_enabled = udpif_use_ufid(udpif);
2355 ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif));
2356 ds_put_format(&ds, "\tflows : (current %lu)"
2357 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif),
2358 udpif->avg_n_flows, udpif->max_n_flows, flow_limit);
2359 ds_put_format(&ds, "\tdump duration : %lldms\n", udpif->dump_duration);
2360 ds_put_format(&ds, "\tufid enabled : ");
2362 ds_put_format(&ds, "true\n");
2364 ds_put_format(&ds, "false\n");
2366 ds_put_char(&ds, '\n');
2368 for (i = 0; i < n_revalidators; i++) {
2369 struct revalidator *revalidator = &udpif->revalidators[i];
2370 int j, elements = 0;
2372 for (j = i; j < N_UMAPS; j += n_revalidators) {
2373 elements += cmap_count(&udpif->ukeys[j].cmap);
2375 ds_put_format(&ds, "\t%u: (keys %d)\n", revalidator->id, elements);
2379 unixctl_command_reply(conn, ds_cstr(&ds));
2383 /* Disable using the megaflows.
2385 * This command is only needed for advanced debugging, so it's not
2386 * documented in the man page. */
2388 upcall_unixctl_disable_megaflows(struct unixctl_conn *conn,
2389 int argc OVS_UNUSED,
2390 const char *argv[] OVS_UNUSED,
2391 void *aux OVS_UNUSED)
2393 atomic_store_relaxed(&enable_megaflows, false);
2394 udpif_flush_all_datapaths();
2395 unixctl_command_reply(conn, "megaflows disabled");
2398 /* Re-enable using megaflows.
2400 * This command is only needed for advanced debugging, so it's not
2401 * documented in the man page. */
2403 upcall_unixctl_enable_megaflows(struct unixctl_conn *conn,
2404 int argc OVS_UNUSED,
2405 const char *argv[] OVS_UNUSED,
2406 void *aux OVS_UNUSED)
2408 atomic_store_relaxed(&enable_megaflows, true);
2409 udpif_flush_all_datapaths();
2410 unixctl_command_reply(conn, "megaflows enabled");
2413 /* Disable skipping flow attributes during flow dump.
2415 * This command is only needed for advanced debugging, so it's not
2416 * documented in the man page. */
2418 upcall_unixctl_disable_ufid(struct unixctl_conn *conn, int argc OVS_UNUSED,
2419 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
2421 atomic_store_relaxed(&enable_ufid, false);
2422 unixctl_command_reply(conn, "Datapath dumping tersely using UFID disabled");
2425 /* Re-enable skipping flow attributes during flow dump.
2427 * This command is only needed for advanced debugging, so it's not documented
2428 * in the man page. */
2430 upcall_unixctl_enable_ufid(struct unixctl_conn *conn, int argc OVS_UNUSED,
2431 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
2433 atomic_store_relaxed(&enable_ufid, true);
2434 unixctl_command_reply(conn, "Datapath dumping tersely using UFID enabled "
2435 "for supported datapaths");
2438 /* Set the flow limit.
2440 * This command is only needed for advanced debugging, so it's not
2441 * documented in the man page. */
2443 upcall_unixctl_set_flow_limit(struct unixctl_conn *conn,
2444 int argc OVS_UNUSED,
2445 const char *argv[] OVS_UNUSED,
2446 void *aux OVS_UNUSED)
2448 struct ds ds = DS_EMPTY_INITIALIZER;
2449 struct udpif *udpif;
2450 unsigned int flow_limit = atoi(argv[1]);
2452 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
2453 atomic_store_relaxed(&udpif->flow_limit, flow_limit);
2455 ds_put_format(&ds, "set flow_limit to %u\n", flow_limit);
2456 unixctl_command_reply(conn, ds_cstr(&ds));
2461 upcall_unixctl_dump_wait(struct unixctl_conn *conn,
2462 int argc OVS_UNUSED,
2463 const char *argv[] OVS_UNUSED,
2464 void *aux OVS_UNUSED)
2466 if (list_is_singleton(&all_udpifs)) {
2467 struct udpif *udpif = NULL;
2470 udpif = OBJECT_CONTAINING(list_front(&all_udpifs), udpif, list_node);
2471 len = (udpif->n_conns + 1) * sizeof *udpif->conns;
2472 udpif->conn_seq = seq_read(udpif->dump_seq);
2473 udpif->conns = xrealloc(udpif->conns, len);
2474 udpif->conns[udpif->n_conns++] = conn;
2476 unixctl_command_reply_error(conn, "can't wait on multiple udpifs.");
2481 upcall_unixctl_purge(struct unixctl_conn *conn, int argc OVS_UNUSED,
2482 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
2484 struct udpif *udpif;
2486 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
2489 for (n = 0; n < udpif->n_revalidators; n++) {
2490 revalidator_purge(&udpif->revalidators[n]);
2493 unixctl_command_reply(conn, "");