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 "openvswitch/dynamic-string.h"
27 #include "fail-open.h"
28 #include "guarded-list.h"
30 #include "openvswitch/list.h"
32 #include "openvswitch/ofpbuf.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 udpif_keys ("ukeys"), we use a large number of cmaps, each with its
65 * own lock for writing. */
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 * Revalidator threads operate in two phases: "dump" and "sweep". In between
76 * each phase, all revalidators sync up so that all revalidator threads are
77 * either in one phase or the other, but not a combination.
79 * During the dump phase, revalidators fetch flows from the datapath and
80 * attribute the statistics to OpenFlow rules. Each datapath flow has a
81 * corresponding ukey which caches the most recently seen statistics. If
82 * a flow needs to be deleted (for example, because it is unused over a
83 * period of time), revalidator threads may delete the flow during the
84 * dump phase. The datapath is not guaranteed to reliably dump all flows
85 * from the datapath, and there is no mapping between datapath flows to
86 * revalidators, so a particular flow may be handled by zero or more
87 * revalidators during a single dump phase. To avoid duplicate attribution
88 * of statistics, ukeys are never deleted during this phase.
90 * During the sweep phase, each revalidator takes ownership of a different
91 * slice of umaps and sweeps through all ukeys in those umaps to figure out
92 * whether they need to be deleted. During this phase, revalidators may
93 * fetch individual flows which were not dumped during the dump phase to
94 * validate them and attribute statistics.
97 struct udpif *udpif; /* Parent udpif. */
98 pthread_t thread; /* Thread ID. */
99 unsigned int id; /* ovsthread_id_self(). */
102 /* An upcall handler for ofproto_dpif.
104 * udpif keeps records of two kind of logically separate units:
109 * - An array of 'struct handler's for upcall handling and flow
115 * - Revalidation threads which read the datapath flow table and maintains
119 struct ovs_list list_node; /* In all_udpifs list. */
121 struct dpif *dpif; /* Datapath handle. */
122 struct dpif_backer *backer; /* Opaque dpif_backer pointer. */
124 struct handler *handlers; /* Upcall handlers. */
127 struct revalidator *revalidators; /* Flow revalidators. */
128 size_t n_revalidators;
130 struct latch exit_latch; /* Tells child threads to exit. */
133 struct seq *reval_seq; /* Incremented to force revalidation. */
134 bool reval_exit; /* Set by leader on 'exit_latch. */
135 struct ovs_barrier reval_barrier; /* Barrier used by revalidators. */
136 struct dpif_flow_dump *dump; /* DPIF flow dump state. */
137 long long int dump_duration; /* Duration of the last flow dump. */
138 struct seq *dump_seq; /* Increments each dump iteration. */
139 atomic_bool enable_ufid; /* If true, skip dumping flow attrs. */
141 /* These variables provide a mechanism for the main thread to pause
142 * all revalidation without having to completely shut the threads down.
143 * 'pause_latch' is shared between the main thread and the lead
144 * revalidator thread, so when it is desirable to halt revalidation, the
145 * main thread will set the latch. 'pause' and 'pause_barrier' are shared
146 * by revalidator threads. The lead revalidator will set 'pause' when it
147 * observes the latch has been set, and this will cause all revalidator
148 * threads to wait on 'pause_barrier' at the beginning of the next
149 * revalidation round. */
150 bool pause; /* Set by leader on 'pause_latch. */
151 struct latch pause_latch; /* Set to force revalidators pause. */
152 struct ovs_barrier pause_barrier; /* Barrier used to pause all */
153 /* revalidators by main thread. */
155 /* There are 'N_UMAPS' maps containing 'struct udpif_key' elements.
157 * During the flow dump phase, revalidators insert into these with a random
158 * distribution. During the garbage collection phase, each revalidator
159 * takes care of garbage collecting a slice of these maps. */
162 /* Datapath flow statistics. */
163 unsigned int max_n_flows;
164 unsigned int avg_n_flows;
166 /* Following fields are accessed and modified by different threads. */
167 atomic_uint flow_limit; /* Datapath flow hard limit. */
169 /* n_flows_mutex prevents multiple threads updating these concurrently. */
170 atomic_uint n_flows; /* Number of flows in the datapath. */
171 atomic_llong n_flows_timestamp; /* Last time n_flows was updated. */
172 struct ovs_mutex n_flows_mutex;
174 /* Following fields are accessed and modified only from the main thread. */
175 struct unixctl_conn **conns; /* Connections waiting on dump_seq. */
176 uint64_t conn_seq; /* Corresponds to 'dump_seq' when
177 conns[n_conns-1] was stored. */
178 size_t n_conns; /* Number of connections waiting. */
182 BAD_UPCALL, /* Some kind of bug somewhere. */
183 MISS_UPCALL, /* A flow miss. */
184 SFLOW_UPCALL, /* sFlow sample. */
185 FLOW_SAMPLE_UPCALL, /* Per-flow sampling. */
186 IPFIX_UPCALL /* Per-bridge sampling. */
196 struct ofproto_dpif *ofproto; /* Parent ofproto. */
197 const struct recirc_id_node *recirc; /* Recirculation context. */
198 bool have_recirc_ref; /* Reference held on recirc ctx? */
200 /* The flow and packet are only required to be constant when using
201 * dpif-netdev. If a modification is absolutely necessary, a const cast
202 * may be used with other datapaths. */
203 const struct flow *flow; /* Parsed representation of the packet. */
204 const ovs_u128 *ufid; /* Unique identifier for 'flow'. */
205 unsigned pmd_id; /* Datapath poll mode driver id. */
206 const struct dp_packet *packet; /* Packet associated with this upcall. */
207 ofp_port_t in_port; /* OpenFlow in port, or OFPP_NONE. */
208 uint16_t mru; /* If !0, Maximum receive unit of
209 fragmented IP packet */
211 enum dpif_upcall_type type; /* Datapath type of the upcall. */
212 const struct nlattr *userdata; /* Userdata for DPIF_UC_ACTION Upcalls. */
213 const struct nlattr *actions; /* Flow actions in DPIF_UC_ACTION Upcalls. */
215 bool xout_initialized; /* True if 'xout' must be uninitialized. */
216 struct xlate_out xout; /* Result of xlate_actions(). */
217 struct ofpbuf odp_actions; /* Datapath actions from xlate_actions(). */
218 struct flow_wildcards wc; /* Dependencies that megaflow must match. */
219 struct ofpbuf put_actions; /* Actions 'put' in the fastpath. */
221 struct dpif_ipfix *ipfix; /* IPFIX pointer or NULL. */
222 struct dpif_sflow *sflow; /* SFlow pointer or NULL. */
224 struct udpif_key *ukey; /* Revalidator flow cache. */
225 bool ukey_persists; /* Set true to keep 'ukey' beyond the
226 lifetime of this upcall. */
228 uint64_t dump_seq; /* udpif->dump_seq at translation time. */
229 uint64_t reval_seq; /* udpif->reval_seq at translation time. */
231 /* Not used by the upcall callback interface. */
232 const struct nlattr *key; /* Datapath flow key. */
233 size_t key_len; /* Datapath flow key length. */
234 const struct nlattr *out_tun_key; /* Datapath output tunnel key. */
236 uint64_t odp_actions_stub[1024 / 8]; /* Stub for odp_actions. */
239 /* 'udpif_key's are responsible for tracking the little bit of state udpif
240 * needs to do flow expiration which can't be pulled directly from the
241 * datapath. They may be created by any handler or revalidator thread at any
242 * time, and read by any revalidator during the dump phase. They are however
243 * each owned by a single revalidator which takes care of destroying them
244 * during the garbage-collection phase.
246 * The mutex within the ukey protects some members of the ukey. The ukey
247 * itself is protected by RCU and is held within a umap in the parent udpif.
248 * Adding or removing a ukey from a umap is only safe when holding the
249 * corresponding umap lock. */
251 struct cmap_node cmap_node; /* In parent revalidator 'ukeys' map. */
253 /* These elements are read only once created, and therefore aren't
254 * protected by a mutex. */
255 const struct nlattr *key; /* Datapath flow key. */
256 size_t key_len; /* Length of 'key'. */
257 const struct nlattr *mask; /* Datapath flow mask. */
258 size_t mask_len; /* Length of 'mask'. */
259 ovs_u128 ufid; /* Unique flow identifier. */
260 bool ufid_present; /* True if 'ufid' is in datapath. */
261 uint32_t hash; /* Pre-computed hash for 'key'. */
262 unsigned pmd_id; /* Datapath poll mode driver id. */
264 struct ovs_mutex mutex; /* Guards the following. */
265 struct dpif_flow_stats stats OVS_GUARDED; /* Last known stats.*/
266 long long int created OVS_GUARDED; /* Estimate of creation time. */
267 uint64_t dump_seq OVS_GUARDED; /* Tracks udpif->dump_seq. */
268 uint64_t reval_seq OVS_GUARDED; /* Tracks udpif->reval_seq. */
269 bool flow_exists OVS_GUARDED; /* Ensures flows are only deleted
271 /* Datapath flow actions as nlattrs. Protected by RCU. Read with
272 * ukey_get_actions(), and write with ukey_set_actions(). */
273 OVSRCU_TYPE(struct ofpbuf *) actions;
275 struct xlate_cache *xcache OVS_GUARDED; /* Cache for xlate entries that
276 * are affected by this ukey.
277 * Used for stats and learning.*/
279 struct odputil_keybuf buf;
283 uint32_t key_recirc_id; /* Non-zero if reference is held by the ukey. */
284 struct recirc_refs recircs; /* Action recirc IDs with references held. */
287 /* Datapath operation with optional ukey attached. */
289 struct udpif_key *ukey;
290 struct dpif_flow_stats stats; /* Stats for 'op'. */
291 struct dpif_op dop; /* Flow operation. */
294 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
295 static struct ovs_list all_udpifs = OVS_LIST_INITIALIZER(&all_udpifs);
297 static size_t recv_upcalls(struct handler *);
298 static int process_upcall(struct udpif *, struct upcall *,
299 struct ofpbuf *odp_actions, struct flow_wildcards *);
300 static void handle_upcalls(struct udpif *, struct upcall *, size_t n_upcalls);
301 static void udpif_stop_threads(struct udpif *);
302 static void udpif_start_threads(struct udpif *, size_t n_handlers,
303 size_t n_revalidators);
304 static void udpif_pause_revalidators(struct udpif *);
305 static void udpif_resume_revalidators(struct udpif *);
306 static void *udpif_upcall_handler(void *);
307 static void *udpif_revalidator(void *);
308 static unsigned long udpif_get_n_flows(struct udpif *);
309 static void revalidate(struct revalidator *);
310 static void revalidator_pause(struct revalidator *);
311 static void revalidator_sweep(struct revalidator *);
312 static void revalidator_purge(struct revalidator *);
313 static void upcall_unixctl_show(struct unixctl_conn *conn, int argc,
314 const char *argv[], void *aux);
315 static void upcall_unixctl_disable_megaflows(struct unixctl_conn *, int argc,
316 const char *argv[], void *aux);
317 static void upcall_unixctl_enable_megaflows(struct unixctl_conn *, int argc,
318 const char *argv[], void *aux);
319 static void upcall_unixctl_disable_ufid(struct unixctl_conn *, int argc,
320 const char *argv[], void *aux);
321 static void upcall_unixctl_enable_ufid(struct unixctl_conn *, int argc,
322 const char *argv[], void *aux);
323 static void upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, int argc,
324 const char *argv[], void *aux);
325 static void upcall_unixctl_dump_wait(struct unixctl_conn *conn, int argc,
326 const char *argv[], void *aux);
327 static void upcall_unixctl_purge(struct unixctl_conn *conn, int argc,
328 const char *argv[], void *aux);
330 static struct udpif_key *ukey_create_from_upcall(struct upcall *,
331 struct flow_wildcards *);
332 static int ukey_create_from_dpif_flow(const struct udpif *,
333 const struct dpif_flow *,
334 struct udpif_key **);
335 static void ukey_get_actions(struct udpif_key *, const struct nlattr **actions,
337 static bool ukey_install_start(struct udpif *, struct udpif_key *ukey);
338 static bool ukey_install_finish(struct udpif_key *ukey, int error);
339 static bool ukey_install(struct udpif *udpif, struct udpif_key *ukey);
340 static struct udpif_key *ukey_lookup(struct udpif *udpif,
341 const ovs_u128 *ufid,
342 const unsigned pmd_id);
343 static int ukey_acquire(struct udpif *, const struct dpif_flow *,
344 struct udpif_key **result, int *error);
345 static void ukey_delete__(struct udpif_key *);
346 static void ukey_delete(struct umap *, struct udpif_key *);
347 static enum upcall_type classify_upcall(enum dpif_upcall_type type,
348 const struct nlattr *userdata);
350 static int upcall_receive(struct upcall *, const struct dpif_backer *,
351 const struct dp_packet *packet, enum dpif_upcall_type,
352 const struct nlattr *userdata, const struct flow *,
353 const unsigned int mru,
354 const ovs_u128 *ufid, const unsigned pmd_id);
355 static void upcall_uninit(struct upcall *);
357 static upcall_callback upcall_cb;
358 static dp_purge_callback dp_purge_cb;
360 static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true);
361 static atomic_bool enable_ufid = ATOMIC_VAR_INIT(true);
366 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
367 if (ovsthread_once_start(&once)) {
368 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show,
370 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
371 upcall_unixctl_disable_megaflows, NULL);
372 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
373 upcall_unixctl_enable_megaflows, NULL);
374 unixctl_command_register("upcall/disable-ufid", "", 0, 0,
375 upcall_unixctl_disable_ufid, NULL);
376 unixctl_command_register("upcall/enable-ufid", "", 0, 0,
377 upcall_unixctl_enable_ufid, NULL);
378 unixctl_command_register("upcall/set-flow-limit", "", 1, 1,
379 upcall_unixctl_set_flow_limit, NULL);
380 unixctl_command_register("revalidator/wait", "", 0, 0,
381 upcall_unixctl_dump_wait, NULL);
382 unixctl_command_register("revalidator/purge", "", 0, 0,
383 upcall_unixctl_purge, NULL);
384 ovsthread_once_done(&once);
389 udpif_create(struct dpif_backer *backer, struct dpif *dpif)
391 struct udpif *udpif = xzalloc(sizeof *udpif);
394 udpif->backer = backer;
395 atomic_init(&udpif->flow_limit, MIN(ofproto_flow_limit, 10000));
396 udpif->reval_seq = seq_create();
397 udpif->dump_seq = seq_create();
398 latch_init(&udpif->exit_latch);
399 latch_init(&udpif->pause_latch);
400 ovs_list_push_back(&all_udpifs, &udpif->list_node);
401 atomic_init(&udpif->enable_ufid, false);
402 atomic_init(&udpif->n_flows, 0);
403 atomic_init(&udpif->n_flows_timestamp, LLONG_MIN);
404 ovs_mutex_init(&udpif->n_flows_mutex);
405 udpif->ukeys = xmalloc(N_UMAPS * sizeof *udpif->ukeys);
406 for (int i = 0; i < N_UMAPS; i++) {
407 cmap_init(&udpif->ukeys[i].cmap);
408 ovs_mutex_init(&udpif->ukeys[i].mutex);
411 dpif_register_upcall_cb(dpif, upcall_cb, udpif);
412 dpif_register_dp_purge_cb(dpif, dp_purge_cb, udpif);
418 udpif_run(struct udpif *udpif)
420 if (udpif->conns && udpif->conn_seq != seq_read(udpif->dump_seq)) {
423 for (i = 0; i < udpif->n_conns; i++) {
424 unixctl_command_reply(udpif->conns[i], NULL);
433 udpif_destroy(struct udpif *udpif)
435 udpif_stop_threads(udpif);
437 dpif_register_dp_purge_cb(udpif->dpif, NULL, udpif);
438 dpif_register_upcall_cb(udpif->dpif, NULL, udpif);
440 for (int i = 0; i < N_UMAPS; i++) {
441 cmap_destroy(&udpif->ukeys[i].cmap);
442 ovs_mutex_destroy(&udpif->ukeys[i].mutex);
447 ovs_list_remove(&udpif->list_node);
448 latch_destroy(&udpif->exit_latch);
449 latch_destroy(&udpif->pause_latch);
450 seq_destroy(udpif->reval_seq);
451 seq_destroy(udpif->dump_seq);
452 ovs_mutex_destroy(&udpif->n_flows_mutex);
456 /* Stops the handler and revalidator threads, must be enclosed in
457 * ovsrcu quiescent state unless when destroying udpif. */
459 udpif_stop_threads(struct udpif *udpif)
461 if (udpif && (udpif->n_handlers != 0 || udpif->n_revalidators != 0)) {
464 latch_set(&udpif->exit_latch);
466 for (i = 0; i < udpif->n_handlers; i++) {
467 struct handler *handler = &udpif->handlers[i];
469 xpthread_join(handler->thread, NULL);
472 for (i = 0; i < udpif->n_revalidators; i++) {
473 xpthread_join(udpif->revalidators[i].thread, NULL);
476 dpif_disable_upcall(udpif->dpif);
478 for (i = 0; i < udpif->n_revalidators; i++) {
479 struct revalidator *revalidator = &udpif->revalidators[i];
481 /* Delete ukeys, and delete all flows from the datapath to prevent
482 * double-counting stats. */
483 revalidator_purge(revalidator);
486 latch_poll(&udpif->exit_latch);
488 ovs_barrier_destroy(&udpif->reval_barrier);
489 ovs_barrier_destroy(&udpif->pause_barrier);
491 free(udpif->revalidators);
492 udpif->revalidators = NULL;
493 udpif->n_revalidators = 0;
495 free(udpif->handlers);
496 udpif->handlers = NULL;
497 udpif->n_handlers = 0;
501 /* Starts the handler and revalidator threads, must be enclosed in
502 * ovsrcu quiescent state. */
504 udpif_start_threads(struct udpif *udpif, size_t n_handlers,
505 size_t n_revalidators)
507 if (udpif && n_handlers && n_revalidators) {
511 udpif->n_handlers = n_handlers;
512 udpif->n_revalidators = n_revalidators;
514 udpif->handlers = xzalloc(udpif->n_handlers * sizeof *udpif->handlers);
515 for (i = 0; i < udpif->n_handlers; i++) {
516 struct handler *handler = &udpif->handlers[i];
518 handler->udpif = udpif;
519 handler->handler_id = i;
520 handler->thread = ovs_thread_create(
521 "handler", udpif_upcall_handler, handler);
524 enable_ufid = ofproto_dpif_get_enable_ufid(udpif->backer);
525 atomic_init(&udpif->enable_ufid, enable_ufid);
526 dpif_enable_upcall(udpif->dpif);
528 ovs_barrier_init(&udpif->reval_barrier, udpif->n_revalidators);
529 ovs_barrier_init(&udpif->pause_barrier, udpif->n_revalidators + 1);
530 udpif->reval_exit = false;
531 udpif->pause = false;
532 udpif->revalidators = xzalloc(udpif->n_revalidators
533 * sizeof *udpif->revalidators);
534 for (i = 0; i < udpif->n_revalidators; i++) {
535 struct revalidator *revalidator = &udpif->revalidators[i];
537 revalidator->udpif = udpif;
538 revalidator->thread = ovs_thread_create(
539 "revalidator", udpif_revalidator, revalidator);
544 /* Pauses all revalidators. Should only be called by the main thread.
545 * When function returns, all revalidators are paused and will proceed
546 * only after udpif_resume_revalidators() is called. */
548 udpif_pause_revalidators(struct udpif *udpif)
550 if (ofproto_dpif_backer_enabled(udpif->backer)) {
551 latch_set(&udpif->pause_latch);
552 ovs_barrier_block(&udpif->pause_barrier);
556 /* Resumes the pausing of revalidators. Should only be called by the
559 udpif_resume_revalidators(struct udpif *udpif)
561 if (ofproto_dpif_backer_enabled(udpif->backer)) {
562 latch_poll(&udpif->pause_latch);
563 ovs_barrier_block(&udpif->pause_barrier);
567 /* Tells 'udpif' how many threads it should use to handle upcalls.
568 * 'n_handlers' and 'n_revalidators' can never be zero. 'udpif''s
569 * datapath handle must have packet reception enabled before starting
572 udpif_set_threads(struct udpif *udpif, size_t n_handlers,
573 size_t n_revalidators)
576 ovs_assert(n_handlers && n_revalidators);
578 ovsrcu_quiesce_start();
579 if (udpif->n_handlers != n_handlers
580 || udpif->n_revalidators != n_revalidators) {
581 udpif_stop_threads(udpif);
584 if (!udpif->handlers && !udpif->revalidators) {
587 error = dpif_handlers_set(udpif->dpif, n_handlers);
589 VLOG_ERR("failed to configure handlers in dpif %s: %s",
590 dpif_name(udpif->dpif), ovs_strerror(error));
594 udpif_start_threads(udpif, n_handlers, n_revalidators);
596 ovsrcu_quiesce_end();
599 /* Waits for all ongoing upcall translations to complete. This ensures that
600 * there are no transient references to any removed ofprotos (or other
601 * objects). In particular, this should be called after an ofproto is removed
602 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
604 udpif_synchronize(struct udpif *udpif)
606 /* This is stronger than necessary. It would be sufficient to ensure
607 * (somehow) that each handler and revalidator thread had passed through
608 * its main loop once. */
609 size_t n_handlers = udpif->n_handlers;
610 size_t n_revalidators = udpif->n_revalidators;
612 ovsrcu_quiesce_start();
613 udpif_stop_threads(udpif);
614 udpif_start_threads(udpif, n_handlers, n_revalidators);
615 ovsrcu_quiesce_end();
618 /* Notifies 'udpif' that something changed which may render previous
619 * xlate_actions() results invalid. */
621 udpif_revalidate(struct udpif *udpif)
623 seq_change(udpif->reval_seq);
626 /* Returns a seq which increments every time 'udpif' pulls stats from the
627 * datapath. Callers can use this to get a sense of when might be a good time
628 * to do periodic work which relies on relatively up to date statistics. */
630 udpif_dump_seq(struct udpif *udpif)
632 return udpif->dump_seq;
636 udpif_get_memory_usage(struct udpif *udpif, struct simap *usage)
640 simap_increase(usage, "handlers", udpif->n_handlers);
642 simap_increase(usage, "revalidators", udpif->n_revalidators);
643 for (i = 0; i < N_UMAPS; i++) {
644 simap_increase(usage, "udpif keys", cmap_count(&udpif->ukeys[i].cmap));
648 /* Remove flows from a single datapath. */
650 udpif_flush(struct udpif *udpif)
652 size_t n_handlers, n_revalidators;
654 n_handlers = udpif->n_handlers;
655 n_revalidators = udpif->n_revalidators;
657 ovsrcu_quiesce_start();
659 udpif_stop_threads(udpif);
660 dpif_flow_flush(udpif->dpif);
661 udpif_start_threads(udpif, n_handlers, n_revalidators);
663 ovsrcu_quiesce_end();
666 /* Removes all flows from all datapaths. */
668 udpif_flush_all_datapaths(void)
672 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
678 udpif_use_ufid(struct udpif *udpif)
682 atomic_read_relaxed(&enable_ufid, &enable);
683 return enable && ofproto_dpif_get_enable_ufid(udpif->backer);
688 udpif_get_n_flows(struct udpif *udpif)
690 long long int time, now;
691 unsigned long flow_count;
694 atomic_read_relaxed(&udpif->n_flows_timestamp, &time);
695 if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) {
696 struct dpif_dp_stats stats;
698 atomic_store_relaxed(&udpif->n_flows_timestamp, now);
699 dpif_get_dp_stats(udpif->dpif, &stats);
700 flow_count = stats.n_flows;
701 atomic_store_relaxed(&udpif->n_flows, flow_count);
702 ovs_mutex_unlock(&udpif->n_flows_mutex);
704 atomic_read_relaxed(&udpif->n_flows, &flow_count);
709 /* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH
710 * upcalls from dpif, processes the batch and installs corresponding flows
713 udpif_upcall_handler(void *arg)
715 struct handler *handler = arg;
716 struct udpif *udpif = handler->udpif;
718 while (!latch_is_set(&handler->udpif->exit_latch)) {
719 if (recv_upcalls(handler)) {
720 poll_immediate_wake();
722 dpif_recv_wait(udpif->dpif, handler->handler_id);
723 latch_wait(&udpif->exit_latch);
732 recv_upcalls(struct handler *handler)
734 struct udpif *udpif = handler->udpif;
735 uint64_t recv_stubs[UPCALL_MAX_BATCH][512 / 8];
736 struct ofpbuf recv_bufs[UPCALL_MAX_BATCH];
737 struct dpif_upcall dupcalls[UPCALL_MAX_BATCH];
738 struct upcall upcalls[UPCALL_MAX_BATCH];
739 struct flow flows[UPCALL_MAX_BATCH];
743 while (n_upcalls < UPCALL_MAX_BATCH) {
744 struct ofpbuf *recv_buf = &recv_bufs[n_upcalls];
745 struct dpif_upcall *dupcall = &dupcalls[n_upcalls];
746 struct upcall *upcall = &upcalls[n_upcalls];
747 struct flow *flow = &flows[n_upcalls];
751 ofpbuf_use_stub(recv_buf, recv_stubs[n_upcalls],
752 sizeof recv_stubs[n_upcalls]);
753 if (dpif_recv(udpif->dpif, handler->handler_id, dupcall, recv_buf)) {
754 ofpbuf_uninit(recv_buf);
758 if (odp_flow_key_to_flow(dupcall->key, dupcall->key_len, flow)
764 mru = nl_attr_get_u16(dupcall->mru);
769 error = upcall_receive(upcall, udpif->backer, &dupcall->packet,
770 dupcall->type, dupcall->userdata, flow, mru,
771 &dupcall->ufid, PMD_ID_NULL);
773 if (error == ENODEV) {
774 /* Received packet on datapath port for which we couldn't
775 * associate an ofproto. This can happen if a port is removed
776 * while traffic is being received. Print a rate-limited
777 * message in case it happens frequently. */
778 dpif_flow_put(udpif->dpif, DPIF_FP_CREATE, dupcall->key,
779 dupcall->key_len, NULL, 0, NULL, 0,
780 &dupcall->ufid, PMD_ID_NULL, NULL);
781 VLOG_INFO_RL(&rl, "received packet on unassociated datapath "
782 "port %"PRIu32, flow->in_port.odp_port);
787 upcall->key = dupcall->key;
788 upcall->key_len = dupcall->key_len;
789 upcall->ufid = &dupcall->ufid;
791 upcall->out_tun_key = dupcall->out_tun_key;
792 upcall->actions = dupcall->actions;
794 pkt_metadata_from_flow(&dupcall->packet.md, flow);
795 flow_extract(&dupcall->packet, flow);
797 error = process_upcall(udpif, upcall,
798 &upcall->odp_actions, &upcall->wc);
807 upcall_uninit(upcall);
809 dp_packet_uninit(&dupcall->packet);
810 ofpbuf_uninit(recv_buf);
814 handle_upcalls(handler->udpif, upcalls, n_upcalls);
815 for (i = 0; i < n_upcalls; i++) {
816 dp_packet_uninit(&dupcalls[i].packet);
817 ofpbuf_uninit(&recv_bufs[i]);
818 upcall_uninit(&upcalls[i]);
826 udpif_revalidator(void *arg)
828 /* Used by all revalidators. */
829 struct revalidator *revalidator = arg;
830 struct udpif *udpif = revalidator->udpif;
831 bool leader = revalidator == &udpif->revalidators[0];
833 /* Used only by the leader. */
834 long long int start_time = 0;
835 uint64_t last_reval_seq = 0;
838 revalidator->id = ovsthread_id_self();
843 recirc_run(); /* Recirculation cleanup. */
845 reval_seq = seq_read(udpif->reval_seq);
846 last_reval_seq = reval_seq;
848 n_flows = udpif_get_n_flows(udpif);
849 udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows);
850 udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2;
852 /* Only the leader checks the pause latch to prevent a race where
853 * some threads think it's false and proceed to block on
854 * reval_barrier and others think it's true and block indefinitely
855 * on the pause_barrier */
856 udpif->pause = latch_is_set(&udpif->pause_latch);
858 /* Only the leader checks the exit latch to prevent a race where
859 * some threads think it's true and exit and others think it's
860 * false and block indefinitely on the reval_barrier */
861 udpif->reval_exit = latch_is_set(&udpif->exit_latch);
863 start_time = time_msec();
864 if (!udpif->reval_exit) {
867 terse_dump = udpif_use_ufid(udpif);
868 udpif->dump = dpif_flow_dump_create(udpif->dpif, terse_dump);
872 /* Wait for the leader to start the flow dump. */
873 ovs_barrier_block(&udpif->reval_barrier);
875 revalidator_pause(revalidator);
878 if (udpif->reval_exit) {
881 revalidate(revalidator);
883 /* Wait for all flows to have been dumped before we garbage collect. */
884 ovs_barrier_block(&udpif->reval_barrier);
885 revalidator_sweep(revalidator);
887 /* Wait for all revalidators to finish garbage collection. */
888 ovs_barrier_block(&udpif->reval_barrier);
891 unsigned int flow_limit;
892 long long int duration;
894 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
896 dpif_flow_dump_destroy(udpif->dump);
897 seq_change(udpif->dump_seq);
899 duration = MAX(time_msec() - start_time, 1);
900 udpif->dump_duration = duration;
901 if (duration > 2000) {
902 flow_limit /= duration / 1000;
903 } else if (duration > 1300) {
904 flow_limit = flow_limit * 3 / 4;
905 } else if (duration < 1000 && n_flows > 2000
906 && flow_limit < n_flows * 1000 / duration) {
909 flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000));
910 atomic_store_relaxed(&udpif->flow_limit, flow_limit);
912 if (duration > 2000) {
913 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
917 poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500));
918 seq_wait(udpif->reval_seq, last_reval_seq);
919 latch_wait(&udpif->exit_latch);
920 latch_wait(&udpif->pause_latch);
928 static enum upcall_type
929 classify_upcall(enum dpif_upcall_type type, const struct nlattr *userdata)
931 union user_action_cookie cookie;
934 /* First look at the upcall type. */
942 case DPIF_N_UC_TYPES:
944 VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32, type);
948 /* "action" upcalls need a closer look. */
950 VLOG_WARN_RL(&rl, "action upcall missing cookie");
953 userdata_len = nl_attr_get_size(userdata);
954 if (userdata_len < sizeof cookie.type
955 || userdata_len > sizeof cookie) {
956 VLOG_WARN_RL(&rl, "action upcall cookie has unexpected size %"PRIuSIZE,
960 memset(&cookie, 0, sizeof cookie);
961 memcpy(&cookie, nl_attr_get(userdata), userdata_len);
962 if (userdata_len == MAX(8, sizeof cookie.sflow)
963 && cookie.type == USER_ACTION_COOKIE_SFLOW) {
965 } else if (userdata_len == MAX(8, sizeof cookie.slow_path)
966 && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) {
968 } else if (userdata_len == MAX(8, sizeof cookie.flow_sample)
969 && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) {
970 return FLOW_SAMPLE_UPCALL;
971 } else if (userdata_len == MAX(8, sizeof cookie.ipfix)
972 && cookie.type == USER_ACTION_COOKIE_IPFIX) {
975 VLOG_WARN_RL(&rl, "invalid user cookie of type %"PRIu16
976 " and size %"PRIuSIZE, cookie.type, userdata_len);
981 /* Calculates slow path actions for 'xout'. 'buf' must statically be
982 * initialized with at least 128 bytes of space. */
984 compose_slow_path(struct udpif *udpif, struct xlate_out *xout,
985 const struct flow *flow, odp_port_t odp_in_port,
988 union user_action_cookie cookie;
992 cookie.type = USER_ACTION_COOKIE_SLOW_PATH;
993 cookie.slow_path.unused = 0;
994 cookie.slow_path.reason = xout->slow;
996 port = xout->slow & (SLOW_CFM | SLOW_BFD | SLOW_LACP | SLOW_STP)
999 pid = dpif_port_get_pid(udpif->dpif, port, flow_hash_5tuple(flow, 0));
1000 odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path,
1001 ODPP_NONE, false, buf);
1004 /* If there is no error, the upcall must be destroyed with upcall_uninit()
1005 * before quiescing, as the referred objects are guaranteed to exist only
1006 * until the calling thread quiesces. Otherwise, do not call upcall_uninit()
1007 * since the 'upcall->put_actions' remains uninitialized. */
1009 upcall_receive(struct upcall *upcall, const struct dpif_backer *backer,
1010 const struct dp_packet *packet, enum dpif_upcall_type type,
1011 const struct nlattr *userdata, const struct flow *flow,
1012 const unsigned int mru,
1013 const ovs_u128 *ufid, const unsigned pmd_id)
1017 error = xlate_lookup(backer, flow, &upcall->ofproto, &upcall->ipfix,
1018 &upcall->sflow, NULL, &upcall->in_port);
1023 upcall->recirc = NULL;
1024 upcall->have_recirc_ref = false;
1025 upcall->flow = flow;
1026 upcall->packet = packet;
1027 upcall->ufid = ufid;
1028 upcall->pmd_id = pmd_id;
1029 upcall->type = type;
1030 upcall->userdata = userdata;
1031 ofpbuf_use_stub(&upcall->odp_actions, upcall->odp_actions_stub,
1032 sizeof upcall->odp_actions_stub);
1033 ofpbuf_init(&upcall->put_actions, 0);
1035 upcall->xout_initialized = false;
1036 upcall->ukey_persists = false;
1038 upcall->ukey = NULL;
1040 upcall->key_len = 0;
1043 upcall->out_tun_key = NULL;
1044 upcall->actions = NULL;
1050 upcall_xlate(struct udpif *udpif, struct upcall *upcall,
1051 struct ofpbuf *odp_actions, struct flow_wildcards *wc)
1053 struct dpif_flow_stats stats;
1054 struct xlate_in xin;
1056 stats.n_packets = 1;
1057 stats.n_bytes = dp_packet_size(upcall->packet);
1058 stats.used = time_msec();
1059 stats.tcp_flags = ntohs(upcall->flow->tcp_flags);
1061 xlate_in_init(&xin, upcall->ofproto, upcall->flow, upcall->in_port, NULL,
1062 stats.tcp_flags, upcall->packet, wc, odp_actions);
1064 if (upcall->type == DPIF_UC_MISS) {
1065 xin.resubmit_stats = &stats;
1067 if (xin.frozen_state) {
1068 /* We may install a datapath flow only if we get a reference to the
1069 * recirculation context (otherwise we could have recirculation
1070 * upcalls using recirculation ID for which no context can be
1071 * found). We may still execute the flow's actions even if we
1072 * don't install the flow. */
1073 upcall->recirc = recirc_id_node_from_state(xin.frozen_state);
1074 upcall->have_recirc_ref = recirc_id_node_try_ref_rcu(upcall->recirc);
1077 /* For non-miss upcalls, we are either executing actions (one of which
1078 * is an userspace action) for an upcall, in which case the stats have
1079 * already been taken care of, or there's a flow in the datapath which
1080 * this packet was accounted to. Presumably the revalidators will deal
1081 * with pushing its stats eventually. */
1084 upcall->dump_seq = seq_read(udpif->dump_seq);
1085 upcall->reval_seq = seq_read(udpif->reval_seq);
1087 xlate_actions(&xin, &upcall->xout);
1089 /* Convert the input port wildcard from OFP to ODP format. There's no
1090 * real way to do this for arbitrary bitmasks since the numbering spaces
1091 * aren't the same. However, flow translation always exact matches the
1092 * whole thing, so we can do the same here. */
1093 WC_MASK_FIELD(wc, in_port.odp_port);
1096 upcall->xout_initialized = true;
1098 if (!upcall->xout.slow) {
1099 ofpbuf_use_const(&upcall->put_actions,
1100 odp_actions->data, odp_actions->size);
1102 /* upcall->put_actions already initialized by upcall_receive(). */
1103 compose_slow_path(udpif, &upcall->xout, upcall->flow,
1104 upcall->flow->in_port.odp_port,
1105 &upcall->put_actions);
1108 /* This function is also called for slow-pathed flows. As we are only
1109 * going to create new datapath flows for actual datapath misses, there is
1110 * no point in creating a ukey otherwise. */
1111 if (upcall->type == DPIF_UC_MISS) {
1112 upcall->ukey = ukey_create_from_upcall(upcall, wc);
1117 upcall_uninit(struct upcall *upcall)
1120 if (upcall->xout_initialized) {
1121 xlate_out_uninit(&upcall->xout);
1123 ofpbuf_uninit(&upcall->odp_actions);
1124 ofpbuf_uninit(&upcall->put_actions);
1126 if (!upcall->ukey_persists) {
1127 ukey_delete__(upcall->ukey);
1129 } else if (upcall->have_recirc_ref) {
1130 /* The reference was transferred to the ukey if one was created. */
1131 recirc_id_node_unref(upcall->recirc);
1137 upcall_cb(const struct dp_packet *packet, const struct flow *flow, ovs_u128 *ufid,
1138 unsigned pmd_id, enum dpif_upcall_type type,
1139 const struct nlattr *userdata, struct ofpbuf *actions,
1140 struct flow_wildcards *wc, struct ofpbuf *put_actions, void *aux)
1142 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);
1143 struct udpif *udpif = aux;
1144 unsigned int flow_limit;
1145 struct upcall upcall;
1149 atomic_read_relaxed(&enable_megaflows, &megaflow);
1150 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1152 error = upcall_receive(&upcall, udpif->backer, packet, type, userdata,
1153 flow, 0, ufid, pmd_id);
1158 error = process_upcall(udpif, &upcall, actions, wc);
1163 if (upcall.xout.slow && put_actions) {
1164 ofpbuf_put(put_actions, upcall.put_actions.data,
1165 upcall.put_actions.size);
1168 if (OVS_UNLIKELY(!megaflow)) {
1169 flow_wildcards_init_for_packet(wc, flow);
1172 if (udpif_get_n_flows(udpif) >= flow_limit) {
1173 VLOG_WARN_RL(&rl, "upcall_cb failure: datapath flow limit reached");
1178 /* Prevent miss flow installation if the key has recirculation ID but we
1179 * were not able to get a reference on it. */
1180 if (type == DPIF_UC_MISS && upcall.recirc && !upcall.have_recirc_ref) {
1181 VLOG_WARN_RL(&rl, "upcall_cb failure: no reference for recirc flow");
1186 if (upcall.ukey && !ukey_install(udpif, upcall.ukey)) {
1187 VLOG_WARN_RL(&rl, "upcall_cb failure: ukey installation fails");
1192 upcall.ukey_persists = true;
1194 upcall_uninit(&upcall);
1199 process_upcall(struct udpif *udpif, struct upcall *upcall,
1200 struct ofpbuf *odp_actions, struct flow_wildcards *wc)
1202 const struct nlattr *userdata = upcall->userdata;
1203 const struct dp_packet *packet = upcall->packet;
1204 const struct flow *flow = upcall->flow;
1206 switch (classify_upcall(upcall->type, userdata)) {
1208 upcall_xlate(udpif, upcall, odp_actions, wc);
1212 if (upcall->sflow) {
1213 union user_action_cookie cookie;
1214 const struct nlattr *actions;
1215 size_t actions_len = 0;
1216 struct dpif_sflow_actions sflow_actions;
1217 memset(&sflow_actions, 0, sizeof sflow_actions);
1218 memset(&cookie, 0, sizeof cookie);
1219 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.sflow);
1220 if (upcall->actions) {
1221 /* Actions were passed up from datapath. */
1222 actions = nl_attr_get(upcall->actions);
1223 actions_len = nl_attr_get_size(upcall->actions);
1224 if (actions && actions_len) {
1225 dpif_sflow_read_actions(flow, actions, actions_len,
1229 if (actions_len == 0) {
1230 /* Lookup actions in userspace cache. */
1231 struct udpif_key *ukey = ukey_lookup(udpif, upcall->ufid,
1234 ukey_get_actions(ukey, &actions, &actions_len);
1235 dpif_sflow_read_actions(flow, actions, actions_len,
1239 dpif_sflow_received(upcall->sflow, packet, flow,
1240 flow->in_port.odp_port, &cookie,
1241 actions_len > 0 ? &sflow_actions : NULL);
1246 if (upcall->ipfix) {
1247 union user_action_cookie cookie;
1248 struct flow_tnl output_tunnel_key;
1250 memset(&cookie, 0, sizeof cookie);
1251 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.ipfix);
1253 if (upcall->out_tun_key) {
1254 odp_tun_key_from_attr(upcall->out_tun_key, false,
1255 &output_tunnel_key);
1257 dpif_ipfix_bridge_sample(upcall->ipfix, packet, flow,
1258 flow->in_port.odp_port,
1259 cookie.ipfix.output_odp_port,
1260 upcall->out_tun_key ?
1261 &output_tunnel_key : NULL);
1265 case FLOW_SAMPLE_UPCALL:
1266 if (upcall->ipfix) {
1267 union user_action_cookie cookie;
1268 struct flow_tnl output_tunnel_key;
1270 memset(&cookie, 0, sizeof cookie);
1271 memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.flow_sample);
1273 if (upcall->out_tun_key) {
1274 odp_tun_key_from_attr(upcall->out_tun_key, false,
1275 &output_tunnel_key);
1278 /* The flow reflects exactly the contents of the packet.
1279 * Sample the packet using it. */
1280 dpif_ipfix_flow_sample(upcall->ipfix, packet, flow,
1281 &cookie, flow->in_port.odp_port,
1282 upcall->out_tun_key ?
1283 &output_tunnel_key : NULL);
1295 handle_upcalls(struct udpif *udpif, struct upcall *upcalls,
1298 struct dpif_op *opsp[UPCALL_MAX_BATCH * 2];
1299 struct ukey_op ops[UPCALL_MAX_BATCH * 2];
1300 unsigned int flow_limit;
1301 size_t n_ops, n_opsp, i;
1304 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
1306 may_put = udpif_get_n_flows(udpif) < flow_limit;
1308 /* Handle the packets individually in order of arrival.
1310 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1311 * processes received packets for these protocols.
1313 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1316 * The loop fills 'ops' with an array of operations to execute in the
1319 for (i = 0; i < n_upcalls; i++) {
1320 struct upcall *upcall = &upcalls[i];
1321 const struct dp_packet *packet = upcall->packet;
1324 /* Do not install a flow into the datapath if:
1326 * - The datapath already has too many flows.
1328 * - We received this packet via some flow installed in the kernel
1331 * - Upcall was a recirculation but we do not have a reference to
1332 * to the recirculation ID. */
1333 if (may_put && upcall->type == DPIF_UC_MISS &&
1334 (!upcall->recirc || upcall->have_recirc_ref)) {
1335 struct udpif_key *ukey = upcall->ukey;
1337 upcall->ukey_persists = true;
1341 op->dop.type = DPIF_OP_FLOW_PUT;
1342 op->dop.u.flow_put.flags = DPIF_FP_CREATE;
1343 op->dop.u.flow_put.key = ukey->key;
1344 op->dop.u.flow_put.key_len = ukey->key_len;
1345 op->dop.u.flow_put.mask = ukey->mask;
1346 op->dop.u.flow_put.mask_len = ukey->mask_len;
1347 op->dop.u.flow_put.ufid = upcall->ufid;
1348 op->dop.u.flow_put.stats = NULL;
1349 ukey_get_actions(ukey, &op->dop.u.flow_put.actions,
1350 &op->dop.u.flow_put.actions_len);
1353 if (upcall->odp_actions.size) {
1356 op->dop.type = DPIF_OP_EXECUTE;
1357 op->dop.u.execute.packet = CONST_CAST(struct dp_packet *, packet);
1358 op->dop.u.execute.flow = upcall->flow;
1359 odp_key_to_pkt_metadata(upcall->key, upcall->key_len,
1360 &op->dop.u.execute.packet->md);
1361 op->dop.u.execute.actions = upcall->odp_actions.data;
1362 op->dop.u.execute.actions_len = upcall->odp_actions.size;
1363 op->dop.u.execute.needs_help = (upcall->xout.slow & SLOW_ACTION) != 0;
1364 op->dop.u.execute.probe = false;
1365 op->dop.u.execute.mtu = upcall->mru;
1371 * We install ukeys before installing the flows, locking them for exclusive
1372 * access by this thread for the period of installation. This ensures that
1373 * other threads won't attempt to delete the flows as we are creating them.
1376 for (i = 0; i < n_ops; i++) {
1377 struct udpif_key *ukey = ops[i].ukey;
1380 /* If we can't install the ukey, don't install the flow. */
1381 if (!ukey_install_start(udpif, ukey)) {
1382 ukey_delete__(ukey);
1387 opsp[n_opsp++] = &ops[i].dop;
1389 dpif_operate(udpif->dpif, opsp, n_opsp);
1390 for (i = 0; i < n_ops; i++) {
1392 ukey_install_finish(ops[i].ukey, ops[i].dop.error);
1398 get_ukey_hash(const ovs_u128 *ufid, const unsigned pmd_id)
1400 return hash_2words(ufid->u32[0], pmd_id);
1403 static struct udpif_key *
1404 ukey_lookup(struct udpif *udpif, const ovs_u128 *ufid, const unsigned pmd_id)
1406 struct udpif_key *ukey;
1407 int idx = get_ukey_hash(ufid, pmd_id) % N_UMAPS;
1408 struct cmap *cmap = &udpif->ukeys[idx].cmap;
1410 CMAP_FOR_EACH_WITH_HASH (ukey, cmap_node,
1411 get_ukey_hash(ufid, pmd_id), cmap) {
1412 if (ovs_u128_equals(ukey->ufid, *ufid)) {
1419 /* Provides safe lockless access of RCU protected 'ukey->actions'. Callers may
1420 * alternatively access the field directly if they take 'ukey->mutex'. */
1422 ukey_get_actions(struct udpif_key *ukey, const struct nlattr **actions, size_t *size)
1424 const struct ofpbuf *buf = ovsrcu_get(struct ofpbuf *, &ukey->actions);
1425 *actions = buf->data;
1430 ukey_set_actions(struct udpif_key *ukey, const struct ofpbuf *actions)
1432 ovsrcu_postpone(ofpbuf_delete,
1433 ovsrcu_get_protected(struct ofpbuf *, &ukey->actions));
1434 ovsrcu_set(&ukey->actions, ofpbuf_clone(actions));
1437 static struct udpif_key *
1438 ukey_create__(const struct nlattr *key, size_t key_len,
1439 const struct nlattr *mask, size_t mask_len,
1440 bool ufid_present, const ovs_u128 *ufid,
1441 const unsigned pmd_id, const struct ofpbuf *actions,
1442 uint64_t dump_seq, uint64_t reval_seq, long long int used,
1443 uint32_t key_recirc_id, struct xlate_out *xout)
1444 OVS_NO_THREAD_SAFETY_ANALYSIS
1446 struct udpif_key *ukey = xmalloc(sizeof *ukey);
1448 memcpy(&ukey->keybuf, key, key_len);
1449 ukey->key = &ukey->keybuf.nla;
1450 ukey->key_len = key_len;
1451 memcpy(&ukey->maskbuf, mask, mask_len);
1452 ukey->mask = &ukey->maskbuf.nla;
1453 ukey->mask_len = mask_len;
1454 ukey->ufid_present = ufid_present;
1456 ukey->pmd_id = pmd_id;
1457 ukey->hash = get_ukey_hash(&ukey->ufid, pmd_id);
1459 ovsrcu_init(&ukey->actions, NULL);
1460 ukey_set_actions(ukey, actions);
1462 ovs_mutex_init(&ukey->mutex);
1463 ukey->dump_seq = dump_seq;
1464 ukey->reval_seq = reval_seq;
1465 ukey->flow_exists = false;
1466 ukey->created = time_msec();
1467 memset(&ukey->stats, 0, sizeof ukey->stats);
1468 ukey->stats.used = used;
1469 ukey->xcache = NULL;
1471 ukey->key_recirc_id = key_recirc_id;
1472 recirc_refs_init(&ukey->recircs);
1474 /* Take ownership of the action recirc id references. */
1475 recirc_refs_swap(&ukey->recircs, &xout->recircs);
1481 static struct udpif_key *
1482 ukey_create_from_upcall(struct upcall *upcall, struct flow_wildcards *wc)
1484 struct odputil_keybuf keystub, maskstub;
1485 struct ofpbuf keybuf, maskbuf;
1487 struct odp_flow_key_parms odp_parms = {
1488 .flow = upcall->flow,
1492 odp_parms.support = ofproto_dpif_get_support(upcall->ofproto)->odp;
1493 if (upcall->key_len) {
1494 ofpbuf_use_const(&keybuf, upcall->key, upcall->key_len);
1496 /* dpif-netdev doesn't provide a netlink-formatted flow key in the
1497 * upcall, so convert the upcall's flow here. */
1498 ofpbuf_use_stack(&keybuf, &keystub, sizeof keystub);
1499 odp_flow_key_from_flow(&odp_parms, &keybuf);
1502 atomic_read_relaxed(&enable_megaflows, &megaflow);
1503 ofpbuf_use_stack(&maskbuf, &maskstub, sizeof maskstub);
1505 odp_parms.key_buf = &keybuf;
1506 odp_flow_key_from_mask(&odp_parms, &maskbuf);
1509 return ukey_create__(keybuf.data, keybuf.size, maskbuf.data, maskbuf.size,
1510 true, upcall->ufid, upcall->pmd_id,
1511 &upcall->put_actions, upcall->dump_seq,
1512 upcall->reval_seq, 0,
1513 upcall->have_recirc_ref ? upcall->recirc->id : 0,
1518 ukey_create_from_dpif_flow(const struct udpif *udpif,
1519 const struct dpif_flow *flow,
1520 struct udpif_key **ukey)
1522 struct dpif_flow full_flow;
1523 struct ofpbuf actions;
1524 uint64_t dump_seq, reval_seq;
1525 uint64_t stub[DPIF_FLOW_BUFSIZE / 8];
1526 const struct nlattr *a;
1529 if (!flow->key_len || !flow->actions_len) {
1533 /* If the key or actions were not provided by the datapath, fetch the
1535 ofpbuf_use_stack(&buf, &stub, sizeof stub);
1536 err = dpif_flow_get(udpif->dpif, flow->key, flow->key_len,
1537 flow->ufid_present ? &flow->ufid : NULL,
1538 flow->pmd_id, &buf, &full_flow);
1545 /* Check the flow actions for recirculation action. As recirculation
1546 * relies on OVS userspace internal state, we need to delete all old
1547 * datapath flows with either a non-zero recirc_id in the key, or any
1548 * recirculation actions upon OVS restart. */
1549 NL_ATTR_FOR_EACH_UNSAFE (a, left, flow->key, flow->key_len) {
1550 if (nl_attr_type(a) == OVS_KEY_ATTR_RECIRC_ID
1551 && nl_attr_get_u32(a) != 0) {
1555 NL_ATTR_FOR_EACH_UNSAFE (a, left, flow->actions, flow->actions_len) {
1556 if (nl_attr_type(a) == OVS_ACTION_ATTR_RECIRC) {
1561 dump_seq = seq_read(udpif->dump_seq);
1562 reval_seq = seq_read(udpif->reval_seq);
1563 ofpbuf_use_const(&actions, &flow->actions, flow->actions_len);
1564 *ukey = ukey_create__(flow->key, flow->key_len,
1565 flow->mask, flow->mask_len, flow->ufid_present,
1566 &flow->ufid, flow->pmd_id, &actions, dump_seq,
1567 reval_seq, flow->stats.used, 0, NULL);
1572 /* Attempts to insert a ukey into the shared ukey maps.
1574 * On success, returns true, installs the ukey and returns it in a locked
1575 * state. Otherwise, returns false. */
1577 ukey_install_start(struct udpif *udpif, struct udpif_key *new_ukey)
1578 OVS_TRY_LOCK(true, new_ukey->mutex)
1581 struct udpif_key *old_ukey;
1583 bool locked = false;
1585 idx = new_ukey->hash % N_UMAPS;
1586 umap = &udpif->ukeys[idx];
1587 ovs_mutex_lock(&umap->mutex);
1588 old_ukey = ukey_lookup(udpif, &new_ukey->ufid, new_ukey->pmd_id);
1590 /* Uncommon case: A ukey is already installed with the same UFID. */
1591 if (old_ukey->key_len == new_ukey->key_len
1592 && !memcmp(old_ukey->key, new_ukey->key, new_ukey->key_len)) {
1593 COVERAGE_INC(handler_duplicate_upcall);
1595 struct ds ds = DS_EMPTY_INITIALIZER;
1597 odp_format_ufid(&old_ukey->ufid, &ds);
1598 ds_put_cstr(&ds, " ");
1599 odp_flow_key_format(old_ukey->key, old_ukey->key_len, &ds);
1600 ds_put_cstr(&ds, "\n");
1601 odp_format_ufid(&new_ukey->ufid, &ds);
1602 ds_put_cstr(&ds, " ");
1603 odp_flow_key_format(new_ukey->key, new_ukey->key_len, &ds);
1605 VLOG_WARN_RL(&rl, "Conflicting ukey for flows:\n%s", ds_cstr(&ds));
1609 ovs_mutex_lock(&new_ukey->mutex);
1610 cmap_insert(&umap->cmap, &new_ukey->cmap_node, new_ukey->hash);
1613 ovs_mutex_unlock(&umap->mutex);
1619 ukey_install_finish__(struct udpif_key *ukey) OVS_REQUIRES(ukey->mutex)
1621 ukey->flow_exists = true;
1625 ukey_install_finish(struct udpif_key *ukey, int error)
1626 OVS_RELEASES(ukey->mutex)
1629 ukey_install_finish__(ukey);
1631 ovs_mutex_unlock(&ukey->mutex);
1637 ukey_install(struct udpif *udpif, struct udpif_key *ukey)
1639 /* The usual way to keep 'ukey->flow_exists' in sync with the datapath is
1640 * to call ukey_install_start(), install the corresponding datapath flow,
1641 * then call ukey_install_finish(). The netdev interface using upcall_cb()
1642 * doesn't provide a function to separately finish the flow installation,
1643 * so we perform the operations together here.
1645 * This is fine currently, as revalidator threads will only delete this
1646 * ukey during revalidator_sweep() and only if the dump_seq is mismatched.
1647 * It is unlikely for a revalidator thread to advance dump_seq and reach
1648 * the next GC phase between ukey creation and flow installation. */
1649 return ukey_install_start(udpif, ukey) && ukey_install_finish(ukey, 0);
1652 /* Searches for a ukey in 'udpif->ukeys' that matches 'flow' and attempts to
1653 * lock the ukey. If the ukey does not exist, create it.
1655 * Returns 0 on success, setting *result to the matching ukey and returning it
1656 * in a locked state. Otherwise, returns an errno and clears *result. EBUSY
1657 * indicates that another thread is handling this flow. Other errors indicate
1658 * an unexpected condition creating a new ukey.
1660 * *error is an output parameter provided to appease the threadsafety analyser,
1661 * and its value matches the return value. */
1663 ukey_acquire(struct udpif *udpif, const struct dpif_flow *flow,
1664 struct udpif_key **result, int *error)
1665 OVS_TRY_LOCK(0, (*result)->mutex)
1667 struct udpif_key *ukey;
1670 ukey = ukey_lookup(udpif, &flow->ufid, flow->pmd_id);
1672 retval = ovs_mutex_trylock(&ukey->mutex);
1674 /* Usually we try to avoid installing flows from revalidator threads,
1675 * because locking on a umap may cause handler threads to block.
1676 * However there are certain cases, like when ovs-vswitchd is
1677 * restarted, where it is desirable to handle flows that exist in the
1678 * datapath gracefully (ie, don't just clear the datapath). */
1681 retval = ukey_create_from_dpif_flow(udpif, flow, &ukey);
1685 install = ukey_install_start(udpif, ukey);
1687 ukey_install_finish__(ukey);
1690 ukey_delete__(ukey);
1706 ukey_delete__(struct udpif_key *ukey)
1707 OVS_NO_THREAD_SAFETY_ANALYSIS
1710 if (ukey->key_recirc_id) {
1711 recirc_free_id(ukey->key_recirc_id);
1713 recirc_refs_unref(&ukey->recircs);
1714 xlate_cache_delete(ukey->xcache);
1715 ofpbuf_delete(ovsrcu_get(struct ofpbuf *, &ukey->actions));
1716 ovs_mutex_destroy(&ukey->mutex);
1722 ukey_delete(struct umap *umap, struct udpif_key *ukey)
1723 OVS_REQUIRES(umap->mutex)
1725 cmap_remove(&umap->cmap, &ukey->cmap_node, ukey->hash);
1726 ovsrcu_postpone(ukey_delete__, ukey);
1730 should_revalidate(const struct udpif *udpif, uint64_t packets,
1733 long long int metric, now, duration;
1735 if (udpif->dump_duration < 200) {
1736 /* We are likely to handle full revalidation for the flows. */
1740 /* Calculate the mean time between seeing these packets. If this
1741 * exceeds the threshold, then delete the flow rather than performing
1742 * costly revalidation for flows that aren't being hit frequently.
1744 * This is targeted at situations where the dump_duration is high (~1s),
1745 * and revalidation is triggered by a call to udpif_revalidate(). In
1746 * these situations, revalidation of all flows causes fluctuations in the
1747 * flow_limit due to the interaction with the dump_duration and max_idle.
1748 * This tends to result in deletion of low-throughput flows anyway, so
1749 * skip the revalidation and just delete those flows. */
1750 packets = MAX(packets, 1);
1751 now = MAX(used, time_msec());
1752 duration = now - used;
1753 metric = duration / packets;
1756 /* The flow is receiving more than ~5pps, so keep it. */
1762 /* Verifies that the datapath actions of 'ukey' are still correct, and pushes
1765 * Returns a recommended action for 'ukey', options include:
1766 * UKEY_DELETE The ukey should be deleted.
1767 * UKEY_KEEP The ukey is fine as is.
1768 * UKEY_MODIFY The ukey's actions should be changed but is otherwise
1769 * fine. Callers should change the actions to those found
1770 * in the caller supplied 'odp_actions' buffer. The
1771 * recirculation references can be found in 'recircs' and
1772 * must be handled by the caller.
1774 * If the result is UKEY_MODIFY, then references to all recirc_ids used by the
1775 * new flow will be held within 'recircs' (which may be none).
1777 * The caller is responsible for both initializing 'recircs' prior this call,
1778 * and ensuring any references are eventually freed.
1780 static enum reval_result
1781 revalidate_ukey(struct udpif *udpif, struct udpif_key *ukey,
1782 const struct dpif_flow_stats *stats,
1783 struct ofpbuf *odp_actions, uint64_t reval_seq,
1784 struct recirc_refs *recircs)
1785 OVS_REQUIRES(ukey->mutex)
1787 struct xlate_out xout, *xoutp;
1788 struct netflow *netflow;
1789 struct ofproto_dpif *ofproto;
1790 struct dpif_flow_stats push;
1792 struct flow_wildcards dp_mask, wc;
1793 enum reval_result result;
1794 ofp_port_t ofp_in_port;
1795 struct xlate_in xin;
1796 long long int last_used;
1798 bool need_revalidate;
1800 result = UKEY_DELETE;
1804 ofpbuf_clear(odp_actions);
1805 need_revalidate = (ukey->reval_seq != reval_seq);
1806 last_used = ukey->stats.used;
1807 push.used = stats->used;
1808 push.tcp_flags = stats->tcp_flags;
1809 push.n_packets = (stats->n_packets > ukey->stats.n_packets
1810 ? stats->n_packets - ukey->stats.n_packets
1812 push.n_bytes = (stats->n_bytes > ukey->stats.n_bytes
1813 ? stats->n_bytes - ukey->stats.n_bytes
1816 if (need_revalidate && last_used
1817 && !should_revalidate(udpif, push.n_packets, last_used)) {
1821 /* We will push the stats, so update the ukey stats cache. */
1822 ukey->stats = *stats;
1823 if (!push.n_packets && !need_revalidate) {
1828 if (ukey->xcache && !need_revalidate) {
1829 xlate_push_stats(ukey->xcache, &push);
1834 if (odp_flow_key_to_flow(ukey->key, ukey->key_len, &flow)
1839 error = xlate_lookup(udpif->backer, &flow, &ofproto, NULL, NULL, &netflow,
1845 if (need_revalidate) {
1846 xlate_cache_clear(ukey->xcache);
1848 if (!ukey->xcache) {
1849 ukey->xcache = xlate_cache_new();
1852 xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL, push.tcp_flags,
1853 NULL, need_revalidate ? &wc : NULL, odp_actions);
1854 if (push.n_packets) {
1855 xin.resubmit_stats = &push;
1856 xin.may_learn = true;
1858 xin.xcache = ukey->xcache;
1859 xlate_actions(&xin, &xout);
1862 if (!need_revalidate) {
1868 ofpbuf_clear(odp_actions);
1869 compose_slow_path(udpif, &xout, &flow, flow.in_port.odp_port,
1873 if (odp_flow_key_to_mask(ukey->mask, ukey->mask_len, ukey->key,
1874 ukey->key_len, &dp_mask, &flow)
1879 /* Do not modify if any bit is wildcarded by the installed datapath flow,
1880 * but not the newly revalidated wildcard mask (wc), i.e., if revalidation
1881 * tells that the datapath flow is now too generic and must be narrowed
1882 * down. Note that we do not know if the datapath has ignored any of the
1883 * wildcarded bits, so we may be overtly conservative here. */
1884 if (flow_wildcards_has_extra(&dp_mask, &wc)) {
1888 if (!ofpbuf_equal(odp_actions,
1889 ovsrcu_get(struct ofpbuf *, &ukey->actions))) {
1890 /* The datapath mask was OK, but the actions seem to have changed.
1891 * Let's modify it in place. */
1892 result = UKEY_MODIFY;
1893 /* Transfer recirc action ID references to the caller. */
1894 recirc_refs_swap(recircs, &xoutp->recircs);
1901 if (result != UKEY_DELETE) {
1902 ukey->reval_seq = reval_seq;
1904 if (netflow && result == UKEY_DELETE) {
1905 netflow_flow_clear(netflow, &flow);
1907 xlate_out_uninit(xoutp);
1912 delete_op_init__(struct udpif *udpif, struct ukey_op *op,
1913 const struct dpif_flow *flow)
1916 op->dop.type = DPIF_OP_FLOW_DEL;
1917 op->dop.u.flow_del.key = flow->key;
1918 op->dop.u.flow_del.key_len = flow->key_len;
1919 op->dop.u.flow_del.ufid = flow->ufid_present ? &flow->ufid : NULL;
1920 op->dop.u.flow_del.pmd_id = flow->pmd_id;
1921 op->dop.u.flow_del.stats = &op->stats;
1922 op->dop.u.flow_del.terse = udpif_use_ufid(udpif);
1926 delete_op_init(struct udpif *udpif, struct ukey_op *op, struct udpif_key *ukey)
1929 op->dop.type = DPIF_OP_FLOW_DEL;
1930 op->dop.u.flow_del.key = ukey->key;
1931 op->dop.u.flow_del.key_len = ukey->key_len;
1932 op->dop.u.flow_del.ufid = ukey->ufid_present ? &ukey->ufid : NULL;
1933 op->dop.u.flow_del.pmd_id = ukey->pmd_id;
1934 op->dop.u.flow_del.stats = &op->stats;
1935 op->dop.u.flow_del.terse = udpif_use_ufid(udpif);
1939 modify_op_init(struct ukey_op *op, struct udpif_key *ukey)
1942 op->dop.type = DPIF_OP_FLOW_PUT;
1943 op->dop.u.flow_put.flags = DPIF_FP_MODIFY;
1944 op->dop.u.flow_put.key = ukey->key;
1945 op->dop.u.flow_put.key_len = ukey->key_len;
1946 op->dop.u.flow_put.mask = ukey->mask;
1947 op->dop.u.flow_put.mask_len = ukey->mask_len;
1948 op->dop.u.flow_put.ufid = ukey->ufid_present ? &ukey->ufid : NULL;
1949 op->dop.u.flow_put.pmd_id = ukey->pmd_id;
1950 op->dop.u.flow_put.stats = NULL;
1951 ukey_get_actions(ukey, &op->dop.u.flow_put.actions,
1952 &op->dop.u.flow_put.actions_len);
1955 /* Executes datapath operations 'ops' and attributes stats retrieved from the
1956 * datapath as part of those operations. */
1958 push_dp_ops(struct udpif *udpif, struct ukey_op *ops, size_t n_ops)
1960 struct dpif_op *opsp[REVALIDATE_MAX_BATCH];
1963 ovs_assert(n_ops <= REVALIDATE_MAX_BATCH);
1964 for (i = 0; i < n_ops; i++) {
1965 opsp[i] = &ops[i].dop;
1967 dpif_operate(udpif->dpif, opsp, n_ops);
1969 for (i = 0; i < n_ops; i++) {
1970 struct ukey_op *op = &ops[i];
1971 struct dpif_flow_stats *push, *stats, push_buf;
1973 stats = op->dop.u.flow_del.stats;
1976 if (op->dop.type != DPIF_OP_FLOW_DEL) {
1977 /* Only deleted flows need their stats pushed. */
1981 if (op->dop.error) {
1982 /* flow_del error, 'stats' is unusable. */
1987 ovs_mutex_lock(&op->ukey->mutex);
1988 push->used = MAX(stats->used, op->ukey->stats.used);
1989 push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags;
1990 push->n_packets = stats->n_packets - op->ukey->stats.n_packets;
1991 push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes;
1992 ovs_mutex_unlock(&op->ukey->mutex);
1997 if (push->n_packets || netflow_exists()) {
1998 const struct nlattr *key = op->dop.u.flow_del.key;
1999 size_t key_len = op->dop.u.flow_del.key_len;
2000 struct ofproto_dpif *ofproto;
2001 struct netflow *netflow;
2002 ofp_port_t ofp_in_port;
2007 ovs_mutex_lock(&op->ukey->mutex);
2008 if (op->ukey->xcache) {
2009 xlate_push_stats(op->ukey->xcache, push);
2010 ovs_mutex_unlock(&op->ukey->mutex);
2013 ovs_mutex_unlock(&op->ukey->mutex);
2014 key = op->ukey->key;
2015 key_len = op->ukey->key_len;
2018 if (odp_flow_key_to_flow(key, key_len, &flow)
2023 error = xlate_lookup(udpif->backer, &flow, &ofproto, NULL, NULL,
2024 &netflow, &ofp_in_port);
2026 struct xlate_in xin;
2028 xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL,
2029 push->tcp_flags, NULL, NULL, NULL);
2030 xin.resubmit_stats = push->n_packets ? push : NULL;
2031 xin.may_learn = push->n_packets > 0;
2032 xlate_actions_for_side_effects(&xin);
2035 netflow_flow_clear(netflow, &flow);
2042 /* Executes datapath operations 'ops', attributes stats retrieved from the
2043 * datapath, and deletes ukeys corresponding to deleted flows. */
2045 push_ukey_ops(struct udpif *udpif, struct umap *umap,
2046 struct ukey_op *ops, size_t n_ops)
2050 push_dp_ops(udpif, ops, n_ops);
2051 ovs_mutex_lock(&umap->mutex);
2052 for (i = 0; i < n_ops; i++) {
2053 if (ops[i].dop.type == DPIF_OP_FLOW_DEL) {
2054 ukey_delete(umap, ops[i].ukey);
2057 ovs_mutex_unlock(&umap->mutex);
2061 log_unexpected_flow(const struct dpif_flow *flow, int error)
2063 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(10, 60);
2064 struct ds ds = DS_EMPTY_INITIALIZER;
2066 ds_put_format(&ds, "Failed to acquire udpif_key corresponding to "
2067 "unexpected flow (%s): ", ovs_strerror(error));
2068 odp_format_ufid(&flow->ufid, &ds);
2069 VLOG_WARN_RL(&rl, "%s", ds_cstr(&ds));
2074 reval_op_init(struct ukey_op *op, enum reval_result result,
2075 struct udpif *udpif, struct udpif_key *ukey,
2076 struct recirc_refs *recircs, struct ofpbuf *odp_actions)
2078 if (result == UKEY_DELETE) {
2079 delete_op_init(udpif, op, ukey);
2080 } else if (result == UKEY_MODIFY) {
2081 /* Store the new recircs. */
2082 recirc_refs_swap(&ukey->recircs, recircs);
2083 /* Release old recircs. */
2084 recirc_refs_unref(recircs);
2085 /* ukey->key_recirc_id remains, as the key is the same as before. */
2087 ukey_set_actions(ukey, odp_actions);
2088 modify_op_init(op, ukey);
2093 revalidate(struct revalidator *revalidator)
2095 uint64_t odp_actions_stub[1024 / 8];
2096 struct ofpbuf odp_actions = OFPBUF_STUB_INITIALIZER(odp_actions_stub);
2098 struct udpif *udpif = revalidator->udpif;
2099 struct dpif_flow_dump_thread *dump_thread;
2100 uint64_t dump_seq, reval_seq;
2101 unsigned int flow_limit;
2103 dump_seq = seq_read(udpif->dump_seq);
2104 reval_seq = seq_read(udpif->reval_seq);
2105 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
2106 dump_thread = dpif_flow_dump_thread_create(udpif->dump);
2108 struct ukey_op ops[REVALIDATE_MAX_BATCH];
2111 struct dpif_flow flows[REVALIDATE_MAX_BATCH];
2112 const struct dpif_flow *f;
2115 long long int max_idle;
2120 n_dumped = dpif_flow_dump_next(dump_thread, flows, ARRAY_SIZE(flows));
2127 /* In normal operation we want to keep flows around until they have
2128 * been idle for 'ofproto_max_idle' milliseconds. However:
2130 * - If the number of datapath flows climbs above 'flow_limit',
2131 * drop that down to 100 ms to try to bring the flows down to
2134 * - If the number of datapath flows climbs above twice
2135 * 'flow_limit', delete all the datapath flows as an emergency
2136 * measure. (We reassess this condition for the next batch of
2137 * datapath flows, so we will recover before all the flows are
2139 n_dp_flows = udpif_get_n_flows(udpif);
2140 kill_them_all = n_dp_flows > flow_limit * 2;
2141 max_idle = n_dp_flows > flow_limit ? 100 : ofproto_max_idle;
2143 for (f = flows; f < &flows[n_dumped]; f++) {
2144 long long int used = f->stats.used;
2145 struct recirc_refs recircs = RECIRC_REFS_EMPTY_INITIALIZER;
2146 enum reval_result result;
2147 struct udpif_key *ukey;
2148 bool already_dumped;
2151 if (ukey_acquire(udpif, f, &ukey, &error)) {
2152 if (error == EBUSY) {
2153 /* Another thread is processing this flow, so don't bother
2155 COVERAGE_INC(upcall_ukey_contention);
2157 log_unexpected_flow(f, error);
2158 if (error != ENOENT) {
2159 delete_op_init__(udpif, &ops[n_ops++], f);
2165 already_dumped = ukey->dump_seq == dump_seq;
2166 if (already_dumped) {
2167 /* The flow has already been handled during this flow dump
2168 * operation. Skip it. */
2170 COVERAGE_INC(dumped_duplicate_flow);
2172 COVERAGE_INC(dumped_new_flow);
2174 ovs_mutex_unlock(&ukey->mutex);
2179 used = ukey->created;
2181 if (kill_them_all || (used && used < now - max_idle)) {
2182 result = UKEY_DELETE;
2184 result = revalidate_ukey(udpif, ukey, &f->stats, &odp_actions,
2185 reval_seq, &recircs);
2187 ukey->dump_seq = dump_seq;
2188 ukey->flow_exists = result != UKEY_DELETE;
2190 if (result != UKEY_KEEP) {
2191 /* Takes ownership of 'recircs'. */
2192 reval_op_init(&ops[n_ops++], result, udpif, ukey, &recircs,
2195 ovs_mutex_unlock(&ukey->mutex);
2199 /* Push datapath ops but defer ukey deletion to 'sweep' phase. */
2200 push_dp_ops(udpif, ops, n_ops);
2204 dpif_flow_dump_thread_destroy(dump_thread);
2205 ofpbuf_uninit(&odp_actions);
2208 /* Pauses the 'revalidator', can only proceed after main thread
2209 * calls udpif_resume_revalidators(). */
2211 revalidator_pause(struct revalidator *revalidator)
2213 /* The first block is for sync'ing the pause with main thread. */
2214 ovs_barrier_block(&revalidator->udpif->pause_barrier);
2215 /* The second block is for pausing until main thread resumes. */
2216 ovs_barrier_block(&revalidator->udpif->pause_barrier);
2220 revalidator_sweep__(struct revalidator *revalidator, bool purge)
2222 struct udpif *udpif;
2223 uint64_t dump_seq, reval_seq;
2226 udpif = revalidator->udpif;
2227 dump_seq = seq_read(udpif->dump_seq);
2228 reval_seq = seq_read(udpif->reval_seq);
2229 slice = revalidator - udpif->revalidators;
2230 ovs_assert(slice < udpif->n_revalidators);
2232 for (int i = slice; i < N_UMAPS; i += udpif->n_revalidators) {
2233 uint64_t odp_actions_stub[1024 / 8];
2234 struct ofpbuf odp_actions = OFPBUF_STUB_INITIALIZER(odp_actions_stub);
2236 struct ukey_op ops[REVALIDATE_MAX_BATCH];
2237 struct udpif_key *ukey;
2238 struct umap *umap = &udpif->ukeys[i];
2241 CMAP_FOR_EACH(ukey, cmap_node, &umap->cmap) {
2244 /* Handler threads could be holding a ukey lock while it installs a
2245 * new flow, so don't hang around waiting for access to it. */
2246 if (ovs_mutex_trylock(&ukey->mutex)) {
2249 flow_exists = ukey->flow_exists;
2251 struct recirc_refs recircs = RECIRC_REFS_EMPTY_INITIALIZER;
2252 bool seq_mismatch = (ukey->dump_seq != dump_seq
2253 && ukey->reval_seq != reval_seq);
2254 enum reval_result result;
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 (result != UKEY_KEEP) {
2268 /* Clears 'recircs' if filled by revalidate_ukey(). */
2269 reval_op_init(&ops[n_ops++], result, udpif, ukey, &recircs,
2273 ovs_mutex_unlock(&ukey->mutex);
2276 /* The common flow deletion case involves deletion of the flow
2277 * during the dump phase and ukey deletion here. */
2278 ovs_mutex_lock(&umap->mutex);
2279 ukey_delete(umap, ukey);
2280 ovs_mutex_unlock(&umap->mutex);
2283 if (n_ops == REVALIDATE_MAX_BATCH) {
2284 /* Update/delete missed flows and clean up corresponding ukeys
2286 push_ukey_ops(udpif, umap, ops, n_ops);
2292 push_ukey_ops(udpif, umap, ops, n_ops);
2295 ofpbuf_uninit(&odp_actions);
2301 revalidator_sweep(struct revalidator *revalidator)
2303 revalidator_sweep__(revalidator, false);
2307 revalidator_purge(struct revalidator *revalidator)
2309 revalidator_sweep__(revalidator, true);
2312 /* In reaction to dpif purge, purges all 'ukey's with same 'pmd_id'. */
2314 dp_purge_cb(void *aux, unsigned pmd_id)
2316 struct udpif *udpif = aux;
2319 udpif_pause_revalidators(udpif);
2320 for (i = 0; i < N_UMAPS; i++) {
2321 struct ukey_op ops[REVALIDATE_MAX_BATCH];
2322 struct udpif_key *ukey;
2323 struct umap *umap = &udpif->ukeys[i];
2326 CMAP_FOR_EACH(ukey, cmap_node, &umap->cmap) {
2327 if (ukey->pmd_id == pmd_id) {
2328 delete_op_init(udpif, &ops[n_ops++], ukey);
2329 if (n_ops == REVALIDATE_MAX_BATCH) {
2330 push_ukey_ops(udpif, umap, ops, n_ops);
2337 push_ukey_ops(udpif, umap, ops, n_ops);
2342 udpif_resume_revalidators(udpif);
2346 upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED,
2347 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
2349 struct ds ds = DS_EMPTY_INITIALIZER;
2350 struct udpif *udpif;
2352 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
2353 unsigned int flow_limit;
2357 atomic_read_relaxed(&udpif->flow_limit, &flow_limit);
2358 ufid_enabled = udpif_use_ufid(udpif);
2360 ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif));
2361 ds_put_format(&ds, "\tflows : (current %lu)"
2362 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif),
2363 udpif->avg_n_flows, udpif->max_n_flows, flow_limit);
2364 ds_put_format(&ds, "\tdump duration : %lldms\n", udpif->dump_duration);
2365 ds_put_format(&ds, "\tufid enabled : ");
2367 ds_put_format(&ds, "true\n");
2369 ds_put_format(&ds, "false\n");
2371 ds_put_char(&ds, '\n');
2373 for (i = 0; i < n_revalidators; i++) {
2374 struct revalidator *revalidator = &udpif->revalidators[i];
2375 int j, elements = 0;
2377 for (j = i; j < N_UMAPS; j += n_revalidators) {
2378 elements += cmap_count(&udpif->ukeys[j].cmap);
2380 ds_put_format(&ds, "\t%u: (keys %d)\n", revalidator->id, elements);
2384 unixctl_command_reply(conn, ds_cstr(&ds));
2388 /* Disable using the megaflows.
2390 * This command is only needed for advanced debugging, so it's not
2391 * documented in the man page. */
2393 upcall_unixctl_disable_megaflows(struct unixctl_conn *conn,
2394 int argc OVS_UNUSED,
2395 const char *argv[] OVS_UNUSED,
2396 void *aux OVS_UNUSED)
2398 atomic_store_relaxed(&enable_megaflows, false);
2399 udpif_flush_all_datapaths();
2400 unixctl_command_reply(conn, "megaflows disabled");
2403 /* Re-enable using megaflows.
2405 * This command is only needed for advanced debugging, so it's not
2406 * documented in the man page. */
2408 upcall_unixctl_enable_megaflows(struct unixctl_conn *conn,
2409 int argc OVS_UNUSED,
2410 const char *argv[] OVS_UNUSED,
2411 void *aux OVS_UNUSED)
2413 atomic_store_relaxed(&enable_megaflows, true);
2414 udpif_flush_all_datapaths();
2415 unixctl_command_reply(conn, "megaflows enabled");
2418 /* Disable skipping flow attributes during flow dump.
2420 * This command is only needed for advanced debugging, so it's not
2421 * documented in the man page. */
2423 upcall_unixctl_disable_ufid(struct unixctl_conn *conn, int argc OVS_UNUSED,
2424 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
2426 atomic_store_relaxed(&enable_ufid, false);
2427 unixctl_command_reply(conn, "Datapath dumping tersely using UFID disabled");
2430 /* Re-enable skipping flow attributes during flow dump.
2432 * This command is only needed for advanced debugging, so it's not documented
2433 * in the man page. */
2435 upcall_unixctl_enable_ufid(struct unixctl_conn *conn, int argc OVS_UNUSED,
2436 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
2438 atomic_store_relaxed(&enable_ufid, true);
2439 unixctl_command_reply(conn, "Datapath dumping tersely using UFID enabled "
2440 "for supported datapaths");
2443 /* Set the flow limit.
2445 * This command is only needed for advanced debugging, so it's not
2446 * documented in the man page. */
2448 upcall_unixctl_set_flow_limit(struct unixctl_conn *conn,
2449 int argc OVS_UNUSED,
2450 const char *argv[] OVS_UNUSED,
2451 void *aux OVS_UNUSED)
2453 struct ds ds = DS_EMPTY_INITIALIZER;
2454 struct udpif *udpif;
2455 unsigned int flow_limit = atoi(argv[1]);
2457 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
2458 atomic_store_relaxed(&udpif->flow_limit, flow_limit);
2460 ds_put_format(&ds, "set flow_limit to %u\n", flow_limit);
2461 unixctl_command_reply(conn, ds_cstr(&ds));
2466 upcall_unixctl_dump_wait(struct unixctl_conn *conn,
2467 int argc OVS_UNUSED,
2468 const char *argv[] OVS_UNUSED,
2469 void *aux OVS_UNUSED)
2471 if (ovs_list_is_singleton(&all_udpifs)) {
2472 struct udpif *udpif = NULL;
2475 udpif = OBJECT_CONTAINING(ovs_list_front(&all_udpifs), udpif, list_node);
2476 len = (udpif->n_conns + 1) * sizeof *udpif->conns;
2477 udpif->conn_seq = seq_read(udpif->dump_seq);
2478 udpif->conns = xrealloc(udpif->conns, len);
2479 udpif->conns[udpif->n_conns++] = conn;
2481 unixctl_command_reply_error(conn, "can't wait on multiple udpifs.");
2486 upcall_unixctl_purge(struct unixctl_conn *conn, int argc OVS_UNUSED,
2487 const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED)
2489 struct udpif *udpif;
2491 LIST_FOR_EACH (udpif, list_node, &all_udpifs) {
2494 for (n = 0; n < udpif->n_revalidators; n++) {
2495 revalidator_purge(&udpif->revalidators[n]);
2498 unixctl_command_reply(conn, "");