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"
36 #include "ofproto-dpif-xlate-cache.h"
39 #include "poll-loop.h"
42 #include "openvswitch/vlog.h"
44 #define MAX_QUEUE_LENGTH 512
45 #define UPCALL_MAX_BATCH 64
46 #define REVALIDATE_MAX_BATCH 50
48 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall
);
50 COVERAGE_DEFINE(dumped_duplicate_flow
);
51 COVERAGE_DEFINE(dumped_new_flow
);
52 COVERAGE_DEFINE(handler_duplicate_upcall
);
53 COVERAGE_DEFINE(upcall_ukey_contention
);
54 COVERAGE_DEFINE(upcall_ukey_replace
);
55 COVERAGE_DEFINE(revalidate_missed_dp_flow
);
57 /* A thread that reads upcalls from dpif, forwards each upcall's packet,
58 * and possibly sets up a kernel flow as a cache. */
60 struct udpif
*udpif
; /* Parent udpif. */
61 pthread_t thread
; /* Thread ID. */
62 uint32_t handler_id
; /* Handler id. */
65 /* In the absence of a multiple-writer multiple-reader datastructure for
66 * storing udpif_keys ("ukeys"), we use a large number of cmaps, each with its
67 * own lock for writing. */
68 #define N_UMAPS 512 /* per udpif. */
70 struct ovs_mutex mutex
; /* Take for writing to the following. */
71 struct cmap cmap
; /* Datapath flow keys. */
74 /* A thread that processes datapath flows, updates OpenFlow statistics, and
75 * updates or removes them if necessary.
77 * Revalidator threads operate in two phases: "dump" and "sweep". In between
78 * each phase, all revalidators sync up so that all revalidator threads are
79 * either in one phase or the other, but not a combination.
81 * During the dump phase, revalidators fetch flows from the datapath and
82 * attribute the statistics to OpenFlow rules. Each datapath flow has a
83 * corresponding ukey which caches the most recently seen statistics. If
84 * a flow needs to be deleted (for example, because it is unused over a
85 * period of time), revalidator threads may delete the flow during the
86 * dump phase. The datapath is not guaranteed to reliably dump all flows
87 * from the datapath, and there is no mapping between datapath flows to
88 * revalidators, so a particular flow may be handled by zero or more
89 * revalidators during a single dump phase. To avoid duplicate attribution
90 * of statistics, ukeys are never deleted during this phase.
92 * During the sweep phase, each revalidator takes ownership of a different
93 * slice of umaps and sweeps through all ukeys in those umaps to figure out
94 * whether they need to be deleted. During this phase, revalidators may
95 * fetch individual flows which were not dumped during the dump phase to
96 * validate them and attribute statistics.
99 struct udpif
*udpif
; /* Parent udpif. */
100 pthread_t thread
; /* Thread ID. */
101 unsigned int id
; /* ovsthread_id_self(). */
104 /* An upcall handler for ofproto_dpif.
106 * udpif keeps records of two kind of logically separate units:
111 * - An array of 'struct handler's for upcall handling and flow
117 * - Revalidation threads which read the datapath flow table and maintains
121 struct ovs_list list_node
; /* In all_udpifs list. */
123 struct dpif
*dpif
; /* Datapath handle. */
124 struct dpif_backer
*backer
; /* Opaque dpif_backer pointer. */
126 struct handler
*handlers
; /* Upcall handlers. */
129 struct revalidator
*revalidators
; /* Flow revalidators. */
130 size_t n_revalidators
;
132 struct latch exit_latch
; /* Tells child threads to exit. */
135 struct seq
*reval_seq
; /* Incremented to force revalidation. */
136 bool reval_exit
; /* Set by leader on 'exit_latch. */
137 struct ovs_barrier reval_barrier
; /* Barrier used by revalidators. */
138 struct dpif_flow_dump
*dump
; /* DPIF flow dump state. */
139 long long int dump_duration
; /* Duration of the last flow dump. */
140 struct seq
*dump_seq
; /* Increments each dump iteration. */
141 atomic_bool enable_ufid
; /* If true, skip dumping flow attrs. */
143 /* These variables provide a mechanism for the main thread to pause
144 * all revalidation without having to completely shut the threads down.
145 * 'pause_latch' is shared between the main thread and the lead
146 * revalidator thread, so when it is desirable to halt revalidation, the
147 * main thread will set the latch. 'pause' and 'pause_barrier' are shared
148 * by revalidator threads. The lead revalidator will set 'pause' when it
149 * observes the latch has been set, and this will cause all revalidator
150 * threads to wait on 'pause_barrier' at the beginning of the next
151 * revalidation round. */
152 bool pause
; /* Set by leader on 'pause_latch. */
153 struct latch pause_latch
; /* Set to force revalidators pause. */
154 struct ovs_barrier pause_barrier
; /* Barrier used to pause all */
155 /* revalidators by main thread. */
157 /* There are 'N_UMAPS' maps containing 'struct udpif_key' elements.
159 * During the flow dump phase, revalidators insert into these with a random
160 * distribution. During the garbage collection phase, each revalidator
161 * takes care of garbage collecting a slice of these maps. */
164 /* Datapath flow statistics. */
165 unsigned int max_n_flows
;
166 unsigned int avg_n_flows
;
168 /* Following fields are accessed and modified by different threads. */
169 atomic_uint flow_limit
; /* Datapath flow hard limit. */
171 /* n_flows_mutex prevents multiple threads updating these concurrently. */
172 atomic_uint n_flows
; /* Number of flows in the datapath. */
173 atomic_llong n_flows_timestamp
; /* Last time n_flows was updated. */
174 struct ovs_mutex n_flows_mutex
;
176 /* Following fields are accessed and modified only from the main thread. */
177 struct unixctl_conn
**conns
; /* Connections waiting on dump_seq. */
178 uint64_t conn_seq
; /* Corresponds to 'dump_seq' when
179 conns[n_conns-1] was stored. */
180 size_t n_conns
; /* Number of connections waiting. */
184 BAD_UPCALL
, /* Some kind of bug somewhere. */
185 MISS_UPCALL
, /* A flow miss. */
186 SFLOW_UPCALL
, /* sFlow sample. */
187 FLOW_SAMPLE_UPCALL
, /* Per-flow sampling. */
188 IPFIX_UPCALL
/* Per-bridge sampling. */
198 struct ofproto_dpif
*ofproto
; /* Parent ofproto. */
199 const struct recirc_id_node
*recirc
; /* Recirculation context. */
200 bool have_recirc_ref
; /* Reference held on recirc ctx? */
202 /* The flow and packet are only required to be constant when using
203 * dpif-netdev. If a modification is absolutely necessary, a const cast
204 * may be used with other datapaths. */
205 const struct flow
*flow
; /* Parsed representation of the packet. */
206 const ovs_u128
*ufid
; /* Unique identifier for 'flow'. */
207 unsigned pmd_id
; /* Datapath poll mode driver id. */
208 const struct dp_packet
*packet
; /* Packet associated with this upcall. */
209 ofp_port_t in_port
; /* OpenFlow in port, or OFPP_NONE. */
210 uint16_t mru
; /* If !0, Maximum receive unit of
211 fragmented IP packet */
213 enum dpif_upcall_type type
; /* Datapath type of the upcall. */
214 const struct nlattr
*userdata
; /* Userdata for DPIF_UC_ACTION Upcalls. */
215 const struct nlattr
*actions
; /* Flow actions in DPIF_UC_ACTION Upcalls. */
217 bool xout_initialized
; /* True if 'xout' must be uninitialized. */
218 struct xlate_out xout
; /* Result of xlate_actions(). */
219 struct ofpbuf odp_actions
; /* Datapath actions from xlate_actions(). */
220 struct flow_wildcards wc
; /* Dependencies that megaflow must match. */
221 struct ofpbuf put_actions
; /* Actions 'put' in the fastpath. */
223 struct dpif_ipfix
*ipfix
; /* IPFIX pointer or NULL. */
224 struct dpif_sflow
*sflow
; /* SFlow pointer or NULL. */
226 struct udpif_key
*ukey
; /* Revalidator flow cache. */
227 bool ukey_persists
; /* Set true to keep 'ukey' beyond the
228 lifetime of this upcall. */
230 uint64_t dump_seq
; /* udpif->dump_seq at translation time. */
231 uint64_t reval_seq
; /* udpif->reval_seq at translation time. */
233 /* Not used by the upcall callback interface. */
234 const struct nlattr
*key
; /* Datapath flow key. */
235 size_t key_len
; /* Datapath flow key length. */
236 const struct nlattr
*out_tun_key
; /* Datapath output tunnel key. */
238 uint64_t odp_actions_stub
[1024 / 8]; /* Stub for odp_actions. */
241 /* Ukeys must transition through these states using transition_ukey(). */
244 UKEY_VISIBLE
, /* Ukey is in umap, datapath flow install is queued. */
245 UKEY_OPERATIONAL
, /* Ukey is in umap, datapath flow is installed. */
246 UKEY_EVICTING
, /* Ukey is in umap, datapath flow delete is queued. */
247 UKEY_EVICTED
, /* Ukey is in umap, datapath flow is deleted. */
248 UKEY_DELETED
, /* Ukey removed from umap, ukey free is deferred. */
250 #define N_UKEY_STATES (UKEY_DELETED + 1)
252 /* 'udpif_key's are responsible for tracking the little bit of state udpif
253 * needs to do flow expiration which can't be pulled directly from the
254 * datapath. They may be created by any handler or revalidator thread at any
255 * time, and read by any revalidator during the dump phase. They are however
256 * each owned by a single revalidator which takes care of destroying them
257 * during the garbage-collection phase.
259 * The mutex within the ukey protects some members of the ukey. The ukey
260 * itself is protected by RCU and is held within a umap in the parent udpif.
261 * Adding or removing a ukey from a umap is only safe when holding the
262 * corresponding umap lock. */
264 struct cmap_node cmap_node
; /* In parent revalidator 'ukeys' map. */
266 /* These elements are read only once created, and therefore aren't
267 * protected by a mutex. */
268 const struct nlattr
*key
; /* Datapath flow key. */
269 size_t key_len
; /* Length of 'key'. */
270 const struct nlattr
*mask
; /* Datapath flow mask. */
271 size_t mask_len
; /* Length of 'mask'. */
272 ovs_u128 ufid
; /* Unique flow identifier. */
273 bool ufid_present
; /* True if 'ufid' is in datapath. */
274 uint32_t hash
; /* Pre-computed hash for 'key'. */
275 unsigned pmd_id
; /* Datapath poll mode driver id. */
277 struct ovs_mutex mutex
; /* Guards the following. */
278 struct dpif_flow_stats stats OVS_GUARDED
; /* Last known stats.*/
279 long long int created OVS_GUARDED
; /* Estimate of creation time. */
280 uint64_t dump_seq OVS_GUARDED
; /* Tracks udpif->dump_seq. */
281 uint64_t reval_seq OVS_GUARDED
; /* Tracks udpif->reval_seq. */
282 enum ukey_state state OVS_GUARDED
; /* Tracks ukey lifetime. */
284 /* Datapath flow actions as nlattrs. Protected by RCU. Read with
285 * ukey_get_actions(), and write with ukey_set_actions(). */
286 OVSRCU_TYPE(struct ofpbuf
*) actions
;
288 struct xlate_cache
*xcache OVS_GUARDED
; /* Cache for xlate entries that
289 * are affected by this ukey.
290 * Used for stats and learning.*/
292 struct odputil_keybuf buf
;
296 uint32_t key_recirc_id
; /* Non-zero if reference is held by the ukey. */
297 struct recirc_refs recircs
; /* Action recirc IDs with references held. */
300 /* Datapath operation with optional ukey attached. */
302 struct udpif_key
*ukey
;
303 struct dpif_flow_stats stats
; /* Stats for 'op'. */
304 struct dpif_op dop
; /* Flow operation. */
307 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
308 static struct ovs_list all_udpifs
= OVS_LIST_INITIALIZER(&all_udpifs
);
310 static size_t recv_upcalls(struct handler
*);
311 static int process_upcall(struct udpif
*, struct upcall
*,
312 struct ofpbuf
*odp_actions
, struct flow_wildcards
*);
313 static void handle_upcalls(struct udpif
*, struct upcall
*, size_t n_upcalls
);
314 static void udpif_stop_threads(struct udpif
*);
315 static void udpif_start_threads(struct udpif
*, size_t n_handlers
,
316 size_t n_revalidators
);
317 static void udpif_pause_revalidators(struct udpif
*);
318 static void udpif_resume_revalidators(struct udpif
*);
319 static void *udpif_upcall_handler(void *);
320 static void *udpif_revalidator(void *);
321 static unsigned long udpif_get_n_flows(struct udpif
*);
322 static void revalidate(struct revalidator
*);
323 static void revalidator_pause(struct revalidator
*);
324 static void revalidator_sweep(struct revalidator
*);
325 static void revalidator_purge(struct revalidator
*);
326 static void upcall_unixctl_show(struct unixctl_conn
*conn
, int argc
,
327 const char *argv
[], void *aux
);
328 static void upcall_unixctl_disable_megaflows(struct unixctl_conn
*, int argc
,
329 const char *argv
[], void *aux
);
330 static void upcall_unixctl_enable_megaflows(struct unixctl_conn
*, int argc
,
331 const char *argv
[], void *aux
);
332 static void upcall_unixctl_disable_ufid(struct unixctl_conn
*, int argc
,
333 const char *argv
[], void *aux
);
334 static void upcall_unixctl_enable_ufid(struct unixctl_conn
*, int argc
,
335 const char *argv
[], void *aux
);
336 static void upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
, int argc
,
337 const char *argv
[], void *aux
);
338 static void upcall_unixctl_dump_wait(struct unixctl_conn
*conn
, int argc
,
339 const char *argv
[], void *aux
);
340 static void upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc
,
341 const char *argv
[], void *aux
);
343 static struct udpif_key
*ukey_create_from_upcall(struct upcall
*,
344 struct flow_wildcards
*);
345 static int ukey_create_from_dpif_flow(const struct udpif
*,
346 const struct dpif_flow
*,
347 struct udpif_key
**);
348 static void ukey_get_actions(struct udpif_key
*, const struct nlattr
**actions
,
350 static bool ukey_install__(struct udpif
*, struct udpif_key
*ukey
)
351 OVS_TRY_LOCK(true, ukey
->mutex
);
352 static bool ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
);
353 static void transition_ukey(struct udpif_key
*ukey
, enum ukey_state dst
)
354 OVS_REQUIRES(ukey
->mutex
);
355 static struct udpif_key
*ukey_lookup(struct udpif
*udpif
,
356 const ovs_u128
*ufid
,
357 const unsigned pmd_id
);
358 static int ukey_acquire(struct udpif
*, const struct dpif_flow
*,
359 struct udpif_key
**result
, int *error
);
360 static void ukey_delete__(struct udpif_key
*);
361 static void ukey_delete(struct umap
*, struct udpif_key
*);
362 static enum upcall_type
classify_upcall(enum dpif_upcall_type type
,
363 const struct nlattr
*userdata
);
365 static void put_op_init(struct ukey_op
*op
, struct udpif_key
*ukey
,
366 enum dpif_flow_put_flags flags
);
367 static void delete_op_init(struct udpif
*udpif
, struct ukey_op
*op
,
368 struct udpif_key
*ukey
);
370 static int upcall_receive(struct upcall
*, const struct dpif_backer
*,
371 const struct dp_packet
*packet
, enum dpif_upcall_type
,
372 const struct nlattr
*userdata
, const struct flow
*,
373 const unsigned int mru
,
374 const ovs_u128
*ufid
, const unsigned pmd_id
);
375 static void upcall_uninit(struct upcall
*);
377 static upcall_callback upcall_cb
;
378 static dp_purge_callback dp_purge_cb
;
380 static atomic_bool enable_megaflows
= ATOMIC_VAR_INIT(true);
381 static atomic_bool enable_ufid
= ATOMIC_VAR_INIT(true);
386 static struct ovsthread_once once
= OVSTHREAD_ONCE_INITIALIZER
;
387 if (ovsthread_once_start(&once
)) {
388 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show
,
390 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
391 upcall_unixctl_disable_megaflows
, NULL
);
392 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
393 upcall_unixctl_enable_megaflows
, NULL
);
394 unixctl_command_register("upcall/disable-ufid", "", 0, 0,
395 upcall_unixctl_disable_ufid
, NULL
);
396 unixctl_command_register("upcall/enable-ufid", "", 0, 0,
397 upcall_unixctl_enable_ufid
, NULL
);
398 unixctl_command_register("upcall/set-flow-limit", "", 1, 1,
399 upcall_unixctl_set_flow_limit
, NULL
);
400 unixctl_command_register("revalidator/wait", "", 0, 0,
401 upcall_unixctl_dump_wait
, NULL
);
402 unixctl_command_register("revalidator/purge", "", 0, 0,
403 upcall_unixctl_purge
, NULL
);
404 ovsthread_once_done(&once
);
409 udpif_create(struct dpif_backer
*backer
, struct dpif
*dpif
)
411 struct udpif
*udpif
= xzalloc(sizeof *udpif
);
414 udpif
->backer
= backer
;
415 atomic_init(&udpif
->flow_limit
, MIN(ofproto_flow_limit
, 10000));
416 udpif
->reval_seq
= seq_create();
417 udpif
->dump_seq
= seq_create();
418 latch_init(&udpif
->exit_latch
);
419 latch_init(&udpif
->pause_latch
);
420 ovs_list_push_back(&all_udpifs
, &udpif
->list_node
);
421 atomic_init(&udpif
->enable_ufid
, false);
422 atomic_init(&udpif
->n_flows
, 0);
423 atomic_init(&udpif
->n_flows_timestamp
, LLONG_MIN
);
424 ovs_mutex_init(&udpif
->n_flows_mutex
);
425 udpif
->ukeys
= xmalloc(N_UMAPS
* sizeof *udpif
->ukeys
);
426 for (int i
= 0; i
< N_UMAPS
; i
++) {
427 cmap_init(&udpif
->ukeys
[i
].cmap
);
428 ovs_mutex_init(&udpif
->ukeys
[i
].mutex
);
431 dpif_register_upcall_cb(dpif
, upcall_cb
, udpif
);
432 dpif_register_dp_purge_cb(dpif
, dp_purge_cb
, udpif
);
438 udpif_run(struct udpif
*udpif
)
440 if (udpif
->conns
&& udpif
->conn_seq
!= seq_read(udpif
->dump_seq
)) {
443 for (i
= 0; i
< udpif
->n_conns
; i
++) {
444 unixctl_command_reply(udpif
->conns
[i
], NULL
);
453 udpif_destroy(struct udpif
*udpif
)
455 udpif_stop_threads(udpif
);
457 dpif_register_dp_purge_cb(udpif
->dpif
, NULL
, udpif
);
458 dpif_register_upcall_cb(udpif
->dpif
, NULL
, udpif
);
460 for (int i
= 0; i
< N_UMAPS
; i
++) {
461 cmap_destroy(&udpif
->ukeys
[i
].cmap
);
462 ovs_mutex_destroy(&udpif
->ukeys
[i
].mutex
);
467 ovs_list_remove(&udpif
->list_node
);
468 latch_destroy(&udpif
->exit_latch
);
469 latch_destroy(&udpif
->pause_latch
);
470 seq_destroy(udpif
->reval_seq
);
471 seq_destroy(udpif
->dump_seq
);
472 ovs_mutex_destroy(&udpif
->n_flows_mutex
);
476 /* Stops the handler and revalidator threads, must be enclosed in
477 * ovsrcu quiescent state unless when destroying udpif. */
479 udpif_stop_threads(struct udpif
*udpif
)
481 if (udpif
&& (udpif
->n_handlers
!= 0 || udpif
->n_revalidators
!= 0)) {
484 latch_set(&udpif
->exit_latch
);
486 for (i
= 0; i
< udpif
->n_handlers
; i
++) {
487 struct handler
*handler
= &udpif
->handlers
[i
];
489 xpthread_join(handler
->thread
, NULL
);
492 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
493 xpthread_join(udpif
->revalidators
[i
].thread
, NULL
);
496 dpif_disable_upcall(udpif
->dpif
);
498 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
499 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
501 /* Delete ukeys, and delete all flows from the datapath to prevent
502 * double-counting stats. */
503 revalidator_purge(revalidator
);
506 latch_poll(&udpif
->exit_latch
);
508 ovs_barrier_destroy(&udpif
->reval_barrier
);
509 ovs_barrier_destroy(&udpif
->pause_barrier
);
511 free(udpif
->revalidators
);
512 udpif
->revalidators
= NULL
;
513 udpif
->n_revalidators
= 0;
515 free(udpif
->handlers
);
516 udpif
->handlers
= NULL
;
517 udpif
->n_handlers
= 0;
521 /* Starts the handler and revalidator threads, must be enclosed in
522 * ovsrcu quiescent state. */
524 udpif_start_threads(struct udpif
*udpif
, size_t n_handlers
,
525 size_t n_revalidators
)
527 if (udpif
&& n_handlers
&& n_revalidators
) {
531 udpif
->n_handlers
= n_handlers
;
532 udpif
->n_revalidators
= n_revalidators
;
534 udpif
->handlers
= xzalloc(udpif
->n_handlers
* sizeof *udpif
->handlers
);
535 for (i
= 0; i
< udpif
->n_handlers
; i
++) {
536 struct handler
*handler
= &udpif
->handlers
[i
];
538 handler
->udpif
= udpif
;
539 handler
->handler_id
= i
;
540 handler
->thread
= ovs_thread_create(
541 "handler", udpif_upcall_handler
, handler
);
544 enable_ufid
= ofproto_dpif_get_enable_ufid(udpif
->backer
);
545 atomic_init(&udpif
->enable_ufid
, enable_ufid
);
546 dpif_enable_upcall(udpif
->dpif
);
548 ovs_barrier_init(&udpif
->reval_barrier
, udpif
->n_revalidators
);
549 ovs_barrier_init(&udpif
->pause_barrier
, udpif
->n_revalidators
+ 1);
550 udpif
->reval_exit
= false;
551 udpif
->pause
= false;
552 udpif
->revalidators
= xzalloc(udpif
->n_revalidators
553 * sizeof *udpif
->revalidators
);
554 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
555 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
557 revalidator
->udpif
= udpif
;
558 revalidator
->thread
= ovs_thread_create(
559 "revalidator", udpif_revalidator
, revalidator
);
564 /* Pauses all revalidators. Should only be called by the main thread.
565 * When function returns, all revalidators are paused and will proceed
566 * only after udpif_resume_revalidators() is called. */
568 udpif_pause_revalidators(struct udpif
*udpif
)
570 if (ofproto_dpif_backer_enabled(udpif
->backer
)) {
571 latch_set(&udpif
->pause_latch
);
572 ovs_barrier_block(&udpif
->pause_barrier
);
576 /* Resumes the pausing of revalidators. Should only be called by the
579 udpif_resume_revalidators(struct udpif
*udpif
)
581 if (ofproto_dpif_backer_enabled(udpif
->backer
)) {
582 latch_poll(&udpif
->pause_latch
);
583 ovs_barrier_block(&udpif
->pause_barrier
);
587 /* Tells 'udpif' how many threads it should use to handle upcalls.
588 * 'n_handlers' and 'n_revalidators' can never be zero. 'udpif''s
589 * datapath handle must have packet reception enabled before starting
592 udpif_set_threads(struct udpif
*udpif
, size_t n_handlers
,
593 size_t n_revalidators
)
596 ovs_assert(n_handlers
&& n_revalidators
);
598 ovsrcu_quiesce_start();
599 if (udpif
->n_handlers
!= n_handlers
600 || udpif
->n_revalidators
!= n_revalidators
) {
601 udpif_stop_threads(udpif
);
604 if (!udpif
->handlers
&& !udpif
->revalidators
) {
607 error
= dpif_handlers_set(udpif
->dpif
, n_handlers
);
609 VLOG_ERR("failed to configure handlers in dpif %s: %s",
610 dpif_name(udpif
->dpif
), ovs_strerror(error
));
614 udpif_start_threads(udpif
, n_handlers
, n_revalidators
);
616 ovsrcu_quiesce_end();
619 /* Waits for all ongoing upcall translations to complete. This ensures that
620 * there are no transient references to any removed ofprotos (or other
621 * objects). In particular, this should be called after an ofproto is removed
622 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
624 udpif_synchronize(struct udpif
*udpif
)
626 /* This is stronger than necessary. It would be sufficient to ensure
627 * (somehow) that each handler and revalidator thread had passed through
628 * its main loop once. */
629 size_t n_handlers
= udpif
->n_handlers
;
630 size_t n_revalidators
= udpif
->n_revalidators
;
632 ovsrcu_quiesce_start();
633 udpif_stop_threads(udpif
);
634 udpif_start_threads(udpif
, n_handlers
, n_revalidators
);
635 ovsrcu_quiesce_end();
638 /* Notifies 'udpif' that something changed which may render previous
639 * xlate_actions() results invalid. */
641 udpif_revalidate(struct udpif
*udpif
)
643 seq_change(udpif
->reval_seq
);
646 /* Returns a seq which increments every time 'udpif' pulls stats from the
647 * datapath. Callers can use this to get a sense of when might be a good time
648 * to do periodic work which relies on relatively up to date statistics. */
650 udpif_dump_seq(struct udpif
*udpif
)
652 return udpif
->dump_seq
;
656 udpif_get_memory_usage(struct udpif
*udpif
, struct simap
*usage
)
660 simap_increase(usage
, "handlers", udpif
->n_handlers
);
662 simap_increase(usage
, "revalidators", udpif
->n_revalidators
);
663 for (i
= 0; i
< N_UMAPS
; i
++) {
664 simap_increase(usage
, "udpif keys", cmap_count(&udpif
->ukeys
[i
].cmap
));
668 /* Remove flows from a single datapath. */
670 udpif_flush(struct udpif
*udpif
)
672 size_t n_handlers
, n_revalidators
;
674 n_handlers
= udpif
->n_handlers
;
675 n_revalidators
= udpif
->n_revalidators
;
677 ovsrcu_quiesce_start();
679 udpif_stop_threads(udpif
);
680 dpif_flow_flush(udpif
->dpif
);
681 udpif_start_threads(udpif
, n_handlers
, n_revalidators
);
683 ovsrcu_quiesce_end();
686 /* Removes all flows from all datapaths. */
688 udpif_flush_all_datapaths(void)
692 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
698 udpif_use_ufid(struct udpif
*udpif
)
702 atomic_read_relaxed(&enable_ufid
, &enable
);
703 return enable
&& ofproto_dpif_get_enable_ufid(udpif
->backer
);
708 udpif_get_n_flows(struct udpif
*udpif
)
710 long long int time
, now
;
711 unsigned long flow_count
;
714 atomic_read_relaxed(&udpif
->n_flows_timestamp
, &time
);
715 if (time
< now
- 100 && !ovs_mutex_trylock(&udpif
->n_flows_mutex
)) {
716 struct dpif_dp_stats stats
;
718 atomic_store_relaxed(&udpif
->n_flows_timestamp
, now
);
719 dpif_get_dp_stats(udpif
->dpif
, &stats
);
720 flow_count
= stats
.n_flows
;
721 atomic_store_relaxed(&udpif
->n_flows
, flow_count
);
722 ovs_mutex_unlock(&udpif
->n_flows_mutex
);
724 atomic_read_relaxed(&udpif
->n_flows
, &flow_count
);
729 /* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH
730 * upcalls from dpif, processes the batch and installs corresponding flows
733 udpif_upcall_handler(void *arg
)
735 struct handler
*handler
= arg
;
736 struct udpif
*udpif
= handler
->udpif
;
738 while (!latch_is_set(&handler
->udpif
->exit_latch
)) {
739 if (recv_upcalls(handler
)) {
740 poll_immediate_wake();
742 dpif_recv_wait(udpif
->dpif
, handler
->handler_id
);
743 latch_wait(&udpif
->exit_latch
);
752 recv_upcalls(struct handler
*handler
)
754 struct udpif
*udpif
= handler
->udpif
;
755 uint64_t recv_stubs
[UPCALL_MAX_BATCH
][512 / 8];
756 struct ofpbuf recv_bufs
[UPCALL_MAX_BATCH
];
757 struct dpif_upcall dupcalls
[UPCALL_MAX_BATCH
];
758 struct upcall upcalls
[UPCALL_MAX_BATCH
];
759 struct flow flows
[UPCALL_MAX_BATCH
];
763 while (n_upcalls
< UPCALL_MAX_BATCH
) {
764 struct ofpbuf
*recv_buf
= &recv_bufs
[n_upcalls
];
765 struct dpif_upcall
*dupcall
= &dupcalls
[n_upcalls
];
766 struct upcall
*upcall
= &upcalls
[n_upcalls
];
767 struct flow
*flow
= &flows
[n_upcalls
];
771 ofpbuf_use_stub(recv_buf
, recv_stubs
[n_upcalls
],
772 sizeof recv_stubs
[n_upcalls
]);
773 if (dpif_recv(udpif
->dpif
, handler
->handler_id
, dupcall
, recv_buf
)) {
774 ofpbuf_uninit(recv_buf
);
778 if (odp_flow_key_to_flow(dupcall
->key
, dupcall
->key_len
, flow
)
784 mru
= nl_attr_get_u16(dupcall
->mru
);
789 error
= upcall_receive(upcall
, udpif
->backer
, &dupcall
->packet
,
790 dupcall
->type
, dupcall
->userdata
, flow
, mru
,
791 &dupcall
->ufid
, PMD_ID_NULL
);
793 if (error
== ENODEV
) {
794 /* Received packet on datapath port for which we couldn't
795 * associate an ofproto. This can happen if a port is removed
796 * while traffic is being received. Print a rate-limited
797 * message in case it happens frequently. */
798 dpif_flow_put(udpif
->dpif
, DPIF_FP_CREATE
, dupcall
->key
,
799 dupcall
->key_len
, NULL
, 0, NULL
, 0,
800 &dupcall
->ufid
, PMD_ID_NULL
, NULL
);
801 VLOG_INFO_RL(&rl
, "received packet on unassociated datapath "
802 "port %"PRIu32
, flow
->in_port
.odp_port
);
807 upcall
->key
= dupcall
->key
;
808 upcall
->key_len
= dupcall
->key_len
;
809 upcall
->ufid
= &dupcall
->ufid
;
811 upcall
->out_tun_key
= dupcall
->out_tun_key
;
812 upcall
->actions
= dupcall
->actions
;
814 pkt_metadata_from_flow(&dupcall
->packet
.md
, flow
);
815 flow_extract(&dupcall
->packet
, flow
);
817 error
= process_upcall(udpif
, upcall
,
818 &upcall
->odp_actions
, &upcall
->wc
);
827 upcall_uninit(upcall
);
829 dp_packet_uninit(&dupcall
->packet
);
830 ofpbuf_uninit(recv_buf
);
834 handle_upcalls(handler
->udpif
, upcalls
, n_upcalls
);
835 for (i
= 0; i
< n_upcalls
; i
++) {
836 dp_packet_uninit(&dupcalls
[i
].packet
);
837 ofpbuf_uninit(&recv_bufs
[i
]);
838 upcall_uninit(&upcalls
[i
]);
846 udpif_revalidator(void *arg
)
848 /* Used by all revalidators. */
849 struct revalidator
*revalidator
= arg
;
850 struct udpif
*udpif
= revalidator
->udpif
;
851 bool leader
= revalidator
== &udpif
->revalidators
[0];
853 /* Used only by the leader. */
854 long long int start_time
= 0;
855 uint64_t last_reval_seq
= 0;
858 revalidator
->id
= ovsthread_id_self();
863 recirc_run(); /* Recirculation cleanup. */
865 reval_seq
= seq_read(udpif
->reval_seq
);
866 last_reval_seq
= reval_seq
;
868 n_flows
= udpif_get_n_flows(udpif
);
869 udpif
->max_n_flows
= MAX(n_flows
, udpif
->max_n_flows
);
870 udpif
->avg_n_flows
= (udpif
->avg_n_flows
+ n_flows
) / 2;
872 /* Only the leader checks the pause latch to prevent a race where
873 * some threads think it's false and proceed to block on
874 * reval_barrier and others think it's true and block indefinitely
875 * on the pause_barrier */
876 udpif
->pause
= latch_is_set(&udpif
->pause_latch
);
878 /* Only the leader checks the exit latch to prevent a race where
879 * some threads think it's true and exit and others think it's
880 * false and block indefinitely on the reval_barrier */
881 udpif
->reval_exit
= latch_is_set(&udpif
->exit_latch
);
883 start_time
= time_msec();
884 if (!udpif
->reval_exit
) {
887 terse_dump
= udpif_use_ufid(udpif
);
888 udpif
->dump
= dpif_flow_dump_create(udpif
->dpif
, terse_dump
);
892 /* Wait for the leader to start the flow dump. */
893 ovs_barrier_block(&udpif
->reval_barrier
);
895 revalidator_pause(revalidator
);
898 if (udpif
->reval_exit
) {
901 revalidate(revalidator
);
903 /* Wait for all flows to have been dumped before we garbage collect. */
904 ovs_barrier_block(&udpif
->reval_barrier
);
905 revalidator_sweep(revalidator
);
907 /* Wait for all revalidators to finish garbage collection. */
908 ovs_barrier_block(&udpif
->reval_barrier
);
911 unsigned int flow_limit
;
912 long long int duration
;
914 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
916 dpif_flow_dump_destroy(udpif
->dump
);
917 seq_change(udpif
->dump_seq
);
919 duration
= MAX(time_msec() - start_time
, 1);
920 udpif
->dump_duration
= duration
;
921 if (duration
> 2000) {
922 flow_limit
/= duration
/ 1000;
923 } else if (duration
> 1300) {
924 flow_limit
= flow_limit
* 3 / 4;
925 } else if (duration
< 1000 && n_flows
> 2000
926 && flow_limit
< n_flows
* 1000 / duration
) {
929 flow_limit
= MIN(ofproto_flow_limit
, MAX(flow_limit
, 1000));
930 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
932 if (duration
> 2000) {
933 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
937 poll_timer_wait_until(start_time
+ MIN(ofproto_max_idle
, 500));
938 seq_wait(udpif
->reval_seq
, last_reval_seq
);
939 latch_wait(&udpif
->exit_latch
);
940 latch_wait(&udpif
->pause_latch
);
943 if (!latch_is_set(&udpif
->pause_latch
) &&
944 !latch_is_set(&udpif
->exit_latch
)) {
945 long long int now
= time_msec();
946 /* Block again if we are woken up within 5ms of the last start
950 if (now
< start_time
) {
951 poll_timer_wait_until(start_time
);
952 latch_wait(&udpif
->exit_latch
);
953 latch_wait(&udpif
->pause_latch
);
963 static enum upcall_type
964 classify_upcall(enum dpif_upcall_type type
, const struct nlattr
*userdata
)
966 union user_action_cookie cookie
;
969 /* First look at the upcall type. */
977 case DPIF_N_UC_TYPES
:
979 VLOG_WARN_RL(&rl
, "upcall has unexpected type %"PRIu32
, type
);
983 /* "action" upcalls need a closer look. */
985 VLOG_WARN_RL(&rl
, "action upcall missing cookie");
988 userdata_len
= nl_attr_get_size(userdata
);
989 if (userdata_len
< sizeof cookie
.type
990 || userdata_len
> sizeof cookie
) {
991 VLOG_WARN_RL(&rl
, "action upcall cookie has unexpected size %"PRIuSIZE
,
995 memset(&cookie
, 0, sizeof cookie
);
996 memcpy(&cookie
, nl_attr_get(userdata
), userdata_len
);
997 if (userdata_len
== MAX(8, sizeof cookie
.sflow
)
998 && cookie
.type
== USER_ACTION_COOKIE_SFLOW
) {
1000 } else if (userdata_len
== MAX(8, sizeof cookie
.slow_path
)
1001 && cookie
.type
== USER_ACTION_COOKIE_SLOW_PATH
) {
1003 } else if (userdata_len
== MAX(8, sizeof cookie
.flow_sample
)
1004 && cookie
.type
== USER_ACTION_COOKIE_FLOW_SAMPLE
) {
1005 return FLOW_SAMPLE_UPCALL
;
1006 } else if (userdata_len
== MAX(8, sizeof cookie
.ipfix
)
1007 && cookie
.type
== USER_ACTION_COOKIE_IPFIX
) {
1008 return IPFIX_UPCALL
;
1010 VLOG_WARN_RL(&rl
, "invalid user cookie of type %"PRIu16
1011 " and size %"PRIuSIZE
, cookie
.type
, userdata_len
);
1016 /* Calculates slow path actions for 'xout'. 'buf' must statically be
1017 * initialized with at least 128 bytes of space. */
1019 compose_slow_path(struct udpif
*udpif
, struct xlate_out
*xout
,
1020 const struct flow
*flow
, odp_port_t odp_in_port
,
1023 union user_action_cookie cookie
;
1027 cookie
.type
= USER_ACTION_COOKIE_SLOW_PATH
;
1028 cookie
.slow_path
.unused
= 0;
1029 cookie
.slow_path
.reason
= xout
->slow
;
1031 port
= xout
->slow
& (SLOW_CFM
| SLOW_BFD
| SLOW_LACP
| SLOW_STP
)
1034 pid
= dpif_port_get_pid(udpif
->dpif
, port
, flow_hash_5tuple(flow
, 0));
1035 odp_put_userspace_action(pid
, &cookie
, sizeof cookie
.slow_path
,
1036 ODPP_NONE
, false, buf
);
1039 /* If there is no error, the upcall must be destroyed with upcall_uninit()
1040 * before quiescing, as the referred objects are guaranteed to exist only
1041 * until the calling thread quiesces. Otherwise, do not call upcall_uninit()
1042 * since the 'upcall->put_actions' remains uninitialized. */
1044 upcall_receive(struct upcall
*upcall
, const struct dpif_backer
*backer
,
1045 const struct dp_packet
*packet
, enum dpif_upcall_type type
,
1046 const struct nlattr
*userdata
, const struct flow
*flow
,
1047 const unsigned int mru
,
1048 const ovs_u128
*ufid
, const unsigned pmd_id
)
1052 error
= xlate_lookup(backer
, flow
, &upcall
->ofproto
, &upcall
->ipfix
,
1053 &upcall
->sflow
, NULL
, &upcall
->in_port
);
1058 upcall
->recirc
= NULL
;
1059 upcall
->have_recirc_ref
= false;
1060 upcall
->flow
= flow
;
1061 upcall
->packet
= packet
;
1062 upcall
->ufid
= ufid
;
1063 upcall
->pmd_id
= pmd_id
;
1064 upcall
->type
= type
;
1065 upcall
->userdata
= userdata
;
1066 ofpbuf_use_stub(&upcall
->odp_actions
, upcall
->odp_actions_stub
,
1067 sizeof upcall
->odp_actions_stub
);
1068 ofpbuf_init(&upcall
->put_actions
, 0);
1070 upcall
->xout_initialized
= false;
1071 upcall
->ukey_persists
= false;
1073 upcall
->ukey
= NULL
;
1075 upcall
->key_len
= 0;
1078 upcall
->out_tun_key
= NULL
;
1079 upcall
->actions
= NULL
;
1085 upcall_xlate(struct udpif
*udpif
, struct upcall
*upcall
,
1086 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1088 struct dpif_flow_stats stats
;
1089 struct xlate_in xin
;
1091 stats
.n_packets
= 1;
1092 stats
.n_bytes
= dp_packet_size(upcall
->packet
);
1093 stats
.used
= time_msec();
1094 stats
.tcp_flags
= ntohs(upcall
->flow
->tcp_flags
);
1096 xlate_in_init(&xin
, upcall
->ofproto
,
1097 ofproto_dpif_get_tables_version(upcall
->ofproto
),
1098 upcall
->flow
, upcall
->in_port
, NULL
,
1099 stats
.tcp_flags
, upcall
->packet
, wc
, odp_actions
);
1101 if (upcall
->type
== DPIF_UC_MISS
) {
1102 xin
.resubmit_stats
= &stats
;
1104 if (xin
.frozen_state
) {
1105 /* We may install a datapath flow only if we get a reference to the
1106 * recirculation context (otherwise we could have recirculation
1107 * upcalls using recirculation ID for which no context can be
1108 * found). We may still execute the flow's actions even if we
1109 * don't install the flow. */
1110 upcall
->recirc
= recirc_id_node_from_state(xin
.frozen_state
);
1111 upcall
->have_recirc_ref
= recirc_id_node_try_ref_rcu(upcall
->recirc
);
1114 /* For non-miss upcalls, we are either executing actions (one of which
1115 * is an userspace action) for an upcall, in which case the stats have
1116 * already been taken care of, or there's a flow in the datapath which
1117 * this packet was accounted to. Presumably the revalidators will deal
1118 * with pushing its stats eventually. */
1121 upcall
->dump_seq
= seq_read(udpif
->dump_seq
);
1122 upcall
->reval_seq
= seq_read(udpif
->reval_seq
);
1124 xlate_actions(&xin
, &upcall
->xout
);
1126 /* Convert the input port wildcard from OFP to ODP format. There's no
1127 * real way to do this for arbitrary bitmasks since the numbering spaces
1128 * aren't the same. However, flow translation always exact matches the
1129 * whole thing, so we can do the same here. */
1130 WC_MASK_FIELD(wc
, in_port
.odp_port
);
1133 upcall
->xout_initialized
= true;
1135 if (!upcall
->xout
.slow
) {
1136 ofpbuf_use_const(&upcall
->put_actions
,
1137 odp_actions
->data
, odp_actions
->size
);
1139 /* upcall->put_actions already initialized by upcall_receive(). */
1140 compose_slow_path(udpif
, &upcall
->xout
, upcall
->flow
,
1141 upcall
->flow
->in_port
.odp_port
,
1142 &upcall
->put_actions
);
1145 /* This function is also called for slow-pathed flows. As we are only
1146 * going to create new datapath flows for actual datapath misses, there is
1147 * no point in creating a ukey otherwise. */
1148 if (upcall
->type
== DPIF_UC_MISS
) {
1149 upcall
->ukey
= ukey_create_from_upcall(upcall
, wc
);
1154 upcall_uninit(struct upcall
*upcall
)
1157 if (upcall
->xout_initialized
) {
1158 xlate_out_uninit(&upcall
->xout
);
1160 ofpbuf_uninit(&upcall
->odp_actions
);
1161 ofpbuf_uninit(&upcall
->put_actions
);
1163 if (!upcall
->ukey_persists
) {
1164 ukey_delete__(upcall
->ukey
);
1166 } else if (upcall
->have_recirc_ref
) {
1167 /* The reference was transferred to the ukey if one was created. */
1168 recirc_id_node_unref(upcall
->recirc
);
1173 /* If there are less flows than the limit, and this is a miss upcall which
1175 * - Has no recirc_id, OR
1176 * - Has a recirc_id and we can get a reference on the recirc ctx,
1178 * Then we should install the flow (true). Otherwise, return false. */
1180 should_install_flow(struct udpif
*udpif
, struct upcall
*upcall
)
1182 unsigned int flow_limit
;
1184 if (upcall
->type
!= DPIF_UC_MISS
) {
1186 } else if (upcall
->recirc
&& !upcall
->have_recirc_ref
) {
1187 VLOG_DBG_RL(&rl
, "upcall: no reference for recirc flow");
1191 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
1192 if (udpif_get_n_flows(udpif
) >= flow_limit
) {
1193 VLOG_WARN_RL(&rl
, "upcall: datapath flow limit reached");
1201 upcall_cb(const struct dp_packet
*packet
, const struct flow
*flow
, ovs_u128
*ufid
,
1202 unsigned pmd_id
, enum dpif_upcall_type type
,
1203 const struct nlattr
*userdata
, struct ofpbuf
*actions
,
1204 struct flow_wildcards
*wc
, struct ofpbuf
*put_actions
, void *aux
)
1206 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 1);
1207 struct udpif
*udpif
= aux
;
1208 struct upcall upcall
;
1212 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1214 error
= upcall_receive(&upcall
, udpif
->backer
, packet
, type
, userdata
,
1215 flow
, 0, ufid
, pmd_id
);
1220 error
= process_upcall(udpif
, &upcall
, actions
, wc
);
1225 if (upcall
.xout
.slow
&& put_actions
) {
1226 ofpbuf_put(put_actions
, upcall
.put_actions
.data
,
1227 upcall
.put_actions
.size
);
1230 if (OVS_UNLIKELY(!megaflow
)) {
1231 flow_wildcards_init_for_packet(wc
, flow
);
1234 if (!should_install_flow(udpif
, &upcall
)) {
1239 if (upcall
.ukey
&& !ukey_install(udpif
, upcall
.ukey
)) {
1240 VLOG_WARN_RL(&rl
, "upcall_cb failure: ukey installation fails");
1245 upcall
.ukey_persists
= true;
1247 upcall_uninit(&upcall
);
1252 process_upcall(struct udpif
*udpif
, struct upcall
*upcall
,
1253 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1255 const struct nlattr
*userdata
= upcall
->userdata
;
1256 const struct dp_packet
*packet
= upcall
->packet
;
1257 const struct flow
*flow
= upcall
->flow
;
1259 switch (classify_upcall(upcall
->type
, userdata
)) {
1261 upcall_xlate(udpif
, upcall
, odp_actions
, wc
);
1265 if (upcall
->sflow
) {
1266 union user_action_cookie cookie
;
1267 const struct nlattr
*actions
;
1268 size_t actions_len
= 0;
1269 struct dpif_sflow_actions sflow_actions
;
1270 memset(&sflow_actions
, 0, sizeof sflow_actions
);
1271 memset(&cookie
, 0, sizeof cookie
);
1272 memcpy(&cookie
, nl_attr_get(userdata
), sizeof cookie
.sflow
);
1273 if (upcall
->actions
) {
1274 /* Actions were passed up from datapath. */
1275 actions
= nl_attr_get(upcall
->actions
);
1276 actions_len
= nl_attr_get_size(upcall
->actions
);
1277 if (actions
&& actions_len
) {
1278 dpif_sflow_read_actions(flow
, actions
, actions_len
,
1282 if (actions_len
== 0) {
1283 /* Lookup actions in userspace cache. */
1284 struct udpif_key
*ukey
= ukey_lookup(udpif
, upcall
->ufid
,
1287 ukey_get_actions(ukey
, &actions
, &actions_len
);
1288 dpif_sflow_read_actions(flow
, actions
, actions_len
,
1292 dpif_sflow_received(upcall
->sflow
, packet
, flow
,
1293 flow
->in_port
.odp_port
, &cookie
,
1294 actions_len
> 0 ? &sflow_actions
: NULL
);
1299 if (upcall
->ipfix
) {
1300 union user_action_cookie cookie
;
1301 struct flow_tnl output_tunnel_key
;
1303 memset(&cookie
, 0, sizeof cookie
);
1304 memcpy(&cookie
, nl_attr_get(userdata
), sizeof cookie
.ipfix
);
1306 if (upcall
->out_tun_key
) {
1307 odp_tun_key_from_attr(upcall
->out_tun_key
, &output_tunnel_key
);
1309 dpif_ipfix_bridge_sample(upcall
->ipfix
, packet
, flow
,
1310 flow
->in_port
.odp_port
,
1311 cookie
.ipfix
.output_odp_port
,
1312 upcall
->out_tun_key
?
1313 &output_tunnel_key
: NULL
);
1317 case FLOW_SAMPLE_UPCALL
:
1318 if (upcall
->ipfix
) {
1319 union user_action_cookie cookie
;
1320 struct flow_tnl output_tunnel_key
;
1322 memset(&cookie
, 0, sizeof cookie
);
1323 memcpy(&cookie
, nl_attr_get(userdata
), sizeof cookie
.flow_sample
);
1325 if (upcall
->out_tun_key
) {
1326 odp_tun_key_from_attr(upcall
->out_tun_key
, &output_tunnel_key
);
1329 /* The flow reflects exactly the contents of the packet.
1330 * Sample the packet using it. */
1331 dpif_ipfix_flow_sample(upcall
->ipfix
, packet
, flow
,
1332 &cookie
, flow
->in_port
.odp_port
,
1333 upcall
->out_tun_key
?
1334 &output_tunnel_key
: NULL
);
1346 handle_upcalls(struct udpif
*udpif
, struct upcall
*upcalls
,
1349 struct dpif_op
*opsp
[UPCALL_MAX_BATCH
* 2];
1350 struct ukey_op ops
[UPCALL_MAX_BATCH
* 2];
1351 size_t n_ops
, n_opsp
, i
;
1353 /* Handle the packets individually in order of arrival.
1355 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1356 * processes received packets for these protocols.
1358 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1361 * The loop fills 'ops' with an array of operations to execute in the
1364 for (i
= 0; i
< n_upcalls
; i
++) {
1365 struct upcall
*upcall
= &upcalls
[i
];
1366 const struct dp_packet
*packet
= upcall
->packet
;
1369 if (should_install_flow(udpif
, upcall
)) {
1370 struct udpif_key
*ukey
= upcall
->ukey
;
1372 if (ukey_install(udpif
, ukey
)) {
1373 upcall
->ukey_persists
= true;
1374 put_op_init(&ops
[n_ops
++], ukey
, DPIF_FP_CREATE
);
1378 if (upcall
->odp_actions
.size
) {
1381 op
->dop
.type
= DPIF_OP_EXECUTE
;
1382 op
->dop
.u
.execute
.packet
= CONST_CAST(struct dp_packet
*, packet
);
1383 op
->dop
.u
.execute
.flow
= upcall
->flow
;
1384 odp_key_to_pkt_metadata(upcall
->key
, upcall
->key_len
,
1385 &op
->dop
.u
.execute
.packet
->md
);
1386 op
->dop
.u
.execute
.actions
= upcall
->odp_actions
.data
;
1387 op
->dop
.u
.execute
.actions_len
= upcall
->odp_actions
.size
;
1388 op
->dop
.u
.execute
.needs_help
= (upcall
->xout
.slow
& SLOW_ACTION
) != 0;
1389 op
->dop
.u
.execute
.probe
= false;
1390 op
->dop
.u
.execute
.mtu
= upcall
->mru
;
1394 /* Execute batch. */
1396 for (i
= 0; i
< n_ops
; i
++) {
1397 opsp
[n_opsp
++] = &ops
[i
].dop
;
1399 dpif_operate(udpif
->dpif
, opsp
, n_opsp
);
1400 for (i
= 0; i
< n_ops
; i
++) {
1401 struct udpif_key
*ukey
= ops
[i
].ukey
;
1404 ovs_mutex_lock(&ukey
->mutex
);
1405 if (ops
[i
].dop
.error
) {
1406 transition_ukey(ukey
, UKEY_EVICTED
);
1408 transition_ukey(ukey
, UKEY_OPERATIONAL
);
1410 ovs_mutex_unlock(&ukey
->mutex
);
1416 get_ukey_hash(const ovs_u128
*ufid
, const unsigned pmd_id
)
1418 return hash_2words(ufid
->u32
[0], pmd_id
);
1421 static struct udpif_key
*
1422 ukey_lookup(struct udpif
*udpif
, const ovs_u128
*ufid
, const unsigned pmd_id
)
1424 struct udpif_key
*ukey
;
1425 int idx
= get_ukey_hash(ufid
, pmd_id
) % N_UMAPS
;
1426 struct cmap
*cmap
= &udpif
->ukeys
[idx
].cmap
;
1428 CMAP_FOR_EACH_WITH_HASH (ukey
, cmap_node
,
1429 get_ukey_hash(ufid
, pmd_id
), cmap
) {
1430 if (ovs_u128_equals(ukey
->ufid
, *ufid
)) {
1437 /* Provides safe lockless access of RCU protected 'ukey->actions'. Callers may
1438 * alternatively access the field directly if they take 'ukey->mutex'. */
1440 ukey_get_actions(struct udpif_key
*ukey
, const struct nlattr
**actions
, size_t *size
)
1442 const struct ofpbuf
*buf
= ovsrcu_get(struct ofpbuf
*, &ukey
->actions
);
1443 *actions
= buf
->data
;
1448 ukey_set_actions(struct udpif_key
*ukey
, const struct ofpbuf
*actions
)
1450 ovsrcu_postpone(ofpbuf_delete
,
1451 ovsrcu_get_protected(struct ofpbuf
*, &ukey
->actions
));
1452 ovsrcu_set(&ukey
->actions
, ofpbuf_clone(actions
));
1455 static struct udpif_key
*
1456 ukey_create__(const struct nlattr
*key
, size_t key_len
,
1457 const struct nlattr
*mask
, size_t mask_len
,
1458 bool ufid_present
, const ovs_u128
*ufid
,
1459 const unsigned pmd_id
, const struct ofpbuf
*actions
,
1460 uint64_t dump_seq
, uint64_t reval_seq
, long long int used
,
1461 uint32_t key_recirc_id
, struct xlate_out
*xout
)
1462 OVS_NO_THREAD_SAFETY_ANALYSIS
1464 struct udpif_key
*ukey
= xmalloc(sizeof *ukey
);
1466 memcpy(&ukey
->keybuf
, key
, key_len
);
1467 ukey
->key
= &ukey
->keybuf
.nla
;
1468 ukey
->key_len
= key_len
;
1469 memcpy(&ukey
->maskbuf
, mask
, mask_len
);
1470 ukey
->mask
= &ukey
->maskbuf
.nla
;
1471 ukey
->mask_len
= mask_len
;
1472 ukey
->ufid_present
= ufid_present
;
1474 ukey
->pmd_id
= pmd_id
;
1475 ukey
->hash
= get_ukey_hash(&ukey
->ufid
, pmd_id
);
1477 ovsrcu_init(&ukey
->actions
, NULL
);
1478 ukey_set_actions(ukey
, actions
);
1480 ovs_mutex_init(&ukey
->mutex
);
1481 ukey
->dump_seq
= dump_seq
;
1482 ukey
->reval_seq
= reval_seq
;
1483 ukey
->state
= UKEY_CREATED
;
1484 ukey
->created
= time_msec();
1485 memset(&ukey
->stats
, 0, sizeof ukey
->stats
);
1486 ukey
->stats
.used
= used
;
1487 ukey
->xcache
= NULL
;
1489 ukey
->key_recirc_id
= key_recirc_id
;
1490 recirc_refs_init(&ukey
->recircs
);
1492 /* Take ownership of the action recirc id references. */
1493 recirc_refs_swap(&ukey
->recircs
, &xout
->recircs
);
1499 static struct udpif_key
*
1500 ukey_create_from_upcall(struct upcall
*upcall
, struct flow_wildcards
*wc
)
1502 struct odputil_keybuf keystub
, maskstub
;
1503 struct ofpbuf keybuf
, maskbuf
;
1505 struct odp_flow_key_parms odp_parms
= {
1506 .flow
= upcall
->flow
,
1510 odp_parms
.support
= ofproto_dpif_get_support(upcall
->ofproto
)->odp
;
1511 if (upcall
->key_len
) {
1512 ofpbuf_use_const(&keybuf
, upcall
->key
, upcall
->key_len
);
1514 /* dpif-netdev doesn't provide a netlink-formatted flow key in the
1515 * upcall, so convert the upcall's flow here. */
1516 ofpbuf_use_stack(&keybuf
, &keystub
, sizeof keystub
);
1517 odp_flow_key_from_flow(&odp_parms
, &keybuf
);
1520 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1521 ofpbuf_use_stack(&maskbuf
, &maskstub
, sizeof maskstub
);
1523 odp_parms
.key_buf
= &keybuf
;
1524 odp_flow_key_from_mask(&odp_parms
, &maskbuf
);
1527 return ukey_create__(keybuf
.data
, keybuf
.size
, maskbuf
.data
, maskbuf
.size
,
1528 true, upcall
->ufid
, upcall
->pmd_id
,
1529 &upcall
->put_actions
, upcall
->dump_seq
,
1530 upcall
->reval_seq
, 0,
1531 upcall
->have_recirc_ref
? upcall
->recirc
->id
: 0,
1536 ukey_create_from_dpif_flow(const struct udpif
*udpif
,
1537 const struct dpif_flow
*flow
,
1538 struct udpif_key
**ukey
)
1540 struct dpif_flow full_flow
;
1541 struct ofpbuf actions
;
1542 uint64_t dump_seq
, reval_seq
;
1543 uint64_t stub
[DPIF_FLOW_BUFSIZE
/ 8];
1544 const struct nlattr
*a
;
1547 if (!flow
->key_len
|| !flow
->actions_len
) {
1551 /* If the key or actions were not provided by the datapath, fetch the
1553 ofpbuf_use_stack(&buf
, &stub
, sizeof stub
);
1554 err
= dpif_flow_get(udpif
->dpif
, flow
->key
, flow
->key_len
,
1555 flow
->ufid_present
? &flow
->ufid
: NULL
,
1556 flow
->pmd_id
, &buf
, &full_flow
);
1563 /* Check the flow actions for recirculation action. As recirculation
1564 * relies on OVS userspace internal state, we need to delete all old
1565 * datapath flows with either a non-zero recirc_id in the key, or any
1566 * recirculation actions upon OVS restart. */
1567 NL_ATTR_FOR_EACH_UNSAFE (a
, left
, flow
->key
, flow
->key_len
) {
1568 if (nl_attr_type(a
) == OVS_KEY_ATTR_RECIRC_ID
1569 && nl_attr_get_u32(a
) != 0) {
1573 NL_ATTR_FOR_EACH_UNSAFE (a
, left
, flow
->actions
, flow
->actions_len
) {
1574 if (nl_attr_type(a
) == OVS_ACTION_ATTR_RECIRC
) {
1579 dump_seq
= seq_read(udpif
->dump_seq
);
1580 reval_seq
= seq_read(udpif
->reval_seq
);
1581 ofpbuf_use_const(&actions
, &flow
->actions
, flow
->actions_len
);
1582 *ukey
= ukey_create__(flow
->key
, flow
->key_len
,
1583 flow
->mask
, flow
->mask_len
, flow
->ufid_present
,
1584 &flow
->ufid
, flow
->pmd_id
, &actions
, dump_seq
,
1585 reval_seq
, flow
->stats
.used
, 0, NULL
);
1591 try_ukey_replace(struct umap
*umap
, struct udpif_key
*old_ukey
,
1592 struct udpif_key
*new_ukey
)
1593 OVS_REQUIRES(umap
->mutex
)
1594 OVS_TRY_LOCK(true, new_ukey
->mutex
)
1596 bool replaced
= false;
1598 if (!ovs_mutex_trylock(&old_ukey
->mutex
)) {
1599 if (old_ukey
->state
== UKEY_EVICTED
) {
1600 /* The flow was deleted during the current revalidator dump,
1601 * but its ukey won't be fully cleaned up until the sweep phase.
1602 * In the mean time, we are receiving upcalls for this traffic.
1603 * Expedite the (new) flow install by replacing the ukey. */
1604 ovs_mutex_lock(&new_ukey
->mutex
);
1605 cmap_replace(&umap
->cmap
, &old_ukey
->cmap_node
,
1606 &new_ukey
->cmap_node
, new_ukey
->hash
);
1607 ovsrcu_postpone(ukey_delete__
, old_ukey
);
1608 transition_ukey(old_ukey
, UKEY_DELETED
);
1609 transition_ukey(new_ukey
, UKEY_VISIBLE
);
1612 ovs_mutex_unlock(&old_ukey
->mutex
);
1616 COVERAGE_INC(upcall_ukey_replace
);
1618 COVERAGE_INC(handler_duplicate_upcall
);
1623 /* Attempts to insert a ukey into the shared ukey maps.
1625 * On success, returns true, installs the ukey and returns it in a locked
1626 * state. Otherwise, returns false. */
1628 ukey_install__(struct udpif
*udpif
, struct udpif_key
*new_ukey
)
1629 OVS_TRY_LOCK(true, new_ukey
->mutex
)
1632 struct udpif_key
*old_ukey
;
1634 bool locked
= false;
1636 idx
= new_ukey
->hash
% N_UMAPS
;
1637 umap
= &udpif
->ukeys
[idx
];
1638 ovs_mutex_lock(&umap
->mutex
);
1639 old_ukey
= ukey_lookup(udpif
, &new_ukey
->ufid
, new_ukey
->pmd_id
);
1641 /* Uncommon case: A ukey is already installed with the same UFID. */
1642 if (old_ukey
->key_len
== new_ukey
->key_len
1643 && !memcmp(old_ukey
->key
, new_ukey
->key
, new_ukey
->key_len
)) {
1644 locked
= try_ukey_replace(umap
, old_ukey
, new_ukey
);
1646 struct ds ds
= DS_EMPTY_INITIALIZER
;
1648 odp_format_ufid(&old_ukey
->ufid
, &ds
);
1649 ds_put_cstr(&ds
, " ");
1650 odp_flow_key_format(old_ukey
->key
, old_ukey
->key_len
, &ds
);
1651 ds_put_cstr(&ds
, "\n");
1652 odp_format_ufid(&new_ukey
->ufid
, &ds
);
1653 ds_put_cstr(&ds
, " ");
1654 odp_flow_key_format(new_ukey
->key
, new_ukey
->key_len
, &ds
);
1656 VLOG_WARN_RL(&rl
, "Conflicting ukey for flows:\n%s", ds_cstr(&ds
));
1660 ovs_mutex_lock(&new_ukey
->mutex
);
1661 cmap_insert(&umap
->cmap
, &new_ukey
->cmap_node
, new_ukey
->hash
);
1662 transition_ukey(new_ukey
, UKEY_VISIBLE
);
1665 ovs_mutex_unlock(&umap
->mutex
);
1671 transition_ukey(struct udpif_key
*ukey
, enum ukey_state dst
)
1672 OVS_REQUIRES(ukey
->mutex
)
1674 ovs_assert(dst
>= ukey
->state
);
1675 if (ukey
->state
== dst
) {
1679 /* Valid state transitions:
1680 * UKEY_CREATED -> UKEY_VISIBLE
1681 * Ukey is now visible in the umap.
1682 * UKEY_VISIBLE -> UKEY_OPERATIONAL
1683 * A handler has installed the flow, and the flow is in the datapath.
1684 * UKEY_VISIBLE -> UKEY_EVICTING
1685 * A handler installs the flow, then revalidator sweeps the ukey before
1686 * the flow is dumped. Most likely the flow was installed; start trying
1688 * UKEY_VISIBLE -> UKEY_EVICTED
1689 * A handler attempts to install the flow, but the datapath rejects it.
1690 * Consider that the datapath has already destroyed it.
1691 * UKEY_OPERATIONAL -> UKEY_EVICTING
1692 * A revalidator decides to evict the datapath flow.
1693 * UKEY_EVICTING -> UKEY_EVICTED
1694 * A revalidator has evicted the datapath flow.
1695 * UKEY_EVICTED -> UKEY_DELETED
1696 * A revalidator has removed the ukey from the umap and is deleting it.
1698 if (ukey
->state
== dst
- 1 || (ukey
->state
== UKEY_VISIBLE
&&
1699 dst
< UKEY_DELETED
)) {
1702 struct ds ds
= DS_EMPTY_INITIALIZER
;
1704 odp_format_ufid(&ukey
->ufid
, &ds
);
1705 VLOG_WARN_RL(&rl
, "Invalid state transition for ukey %s: %d -> %d",
1706 ds_cstr(&ds
), ukey
->state
, dst
);
1712 ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
)
1716 installed
= ukey_install__(udpif
, ukey
);
1718 ovs_mutex_unlock(&ukey
->mutex
);
1724 /* Searches for a ukey in 'udpif->ukeys' that matches 'flow' and attempts to
1725 * lock the ukey. If the ukey does not exist, create it.
1727 * Returns 0 on success, setting *result to the matching ukey and returning it
1728 * in a locked state. Otherwise, returns an errno and clears *result. EBUSY
1729 * indicates that another thread is handling this flow. Other errors indicate
1730 * an unexpected condition creating a new ukey.
1732 * *error is an output parameter provided to appease the threadsafety analyser,
1733 * and its value matches the return value. */
1735 ukey_acquire(struct udpif
*udpif
, const struct dpif_flow
*flow
,
1736 struct udpif_key
**result
, int *error
)
1737 OVS_TRY_LOCK(0, (*result
)->mutex
)
1739 struct udpif_key
*ukey
;
1742 ukey
= ukey_lookup(udpif
, &flow
->ufid
, flow
->pmd_id
);
1744 retval
= ovs_mutex_trylock(&ukey
->mutex
);
1746 /* Usually we try to avoid installing flows from revalidator threads,
1747 * because locking on a umap may cause handler threads to block.
1748 * However there are certain cases, like when ovs-vswitchd is
1749 * restarted, where it is desirable to handle flows that exist in the
1750 * datapath gracefully (ie, don't just clear the datapath). */
1753 retval
= ukey_create_from_dpif_flow(udpif
, flow
, &ukey
);
1757 install
= ukey_install__(udpif
, ukey
);
1761 ukey_delete__(ukey
);
1777 ukey_delete__(struct udpif_key
*ukey
)
1778 OVS_NO_THREAD_SAFETY_ANALYSIS
1781 if (ukey
->key_recirc_id
) {
1782 recirc_free_id(ukey
->key_recirc_id
);
1784 recirc_refs_unref(&ukey
->recircs
);
1785 xlate_cache_delete(ukey
->xcache
);
1786 ofpbuf_delete(ovsrcu_get(struct ofpbuf
*, &ukey
->actions
));
1787 ovs_mutex_destroy(&ukey
->mutex
);
1793 ukey_delete(struct umap
*umap
, struct udpif_key
*ukey
)
1794 OVS_REQUIRES(umap
->mutex
)
1796 ovs_mutex_lock(&ukey
->mutex
);
1797 cmap_remove(&umap
->cmap
, &ukey
->cmap_node
, ukey
->hash
);
1798 ovsrcu_postpone(ukey_delete__
, ukey
);
1799 transition_ukey(ukey
, UKEY_DELETED
);
1800 ovs_mutex_unlock(&ukey
->mutex
);
1804 should_revalidate(const struct udpif
*udpif
, uint64_t packets
,
1807 long long int metric
, now
, duration
;
1809 if (udpif
->dump_duration
< 200) {
1810 /* We are likely to handle full revalidation for the flows. */
1814 /* Calculate the mean time between seeing these packets. If this
1815 * exceeds the threshold, then delete the flow rather than performing
1816 * costly revalidation for flows that aren't being hit frequently.
1818 * This is targeted at situations where the dump_duration is high (~1s),
1819 * and revalidation is triggered by a call to udpif_revalidate(). In
1820 * these situations, revalidation of all flows causes fluctuations in the
1821 * flow_limit due to the interaction with the dump_duration and max_idle.
1822 * This tends to result in deletion of low-throughput flows anyway, so
1823 * skip the revalidation and just delete those flows. */
1824 packets
= MAX(packets
, 1);
1825 now
= MAX(used
, time_msec());
1826 duration
= now
- used
;
1827 metric
= duration
/ packets
;
1830 /* The flow is receiving more than ~5pps, so keep it. */
1836 /* Verifies that the datapath actions of 'ukey' are still correct, and pushes
1839 * Returns a recommended action for 'ukey', options include:
1840 * UKEY_DELETE The ukey should be deleted.
1841 * UKEY_KEEP The ukey is fine as is.
1842 * UKEY_MODIFY The ukey's actions should be changed but is otherwise
1843 * fine. Callers should change the actions to those found
1844 * in the caller supplied 'odp_actions' buffer. The
1845 * recirculation references can be found in 'recircs' and
1846 * must be handled by the caller.
1848 * If the result is UKEY_MODIFY, then references to all recirc_ids used by the
1849 * new flow will be held within 'recircs' (which may be none).
1851 * The caller is responsible for both initializing 'recircs' prior this call,
1852 * and ensuring any references are eventually freed.
1854 static enum reval_result
1855 revalidate_ukey(struct udpif
*udpif
, struct udpif_key
*ukey
,
1856 const struct dpif_flow_stats
*stats
,
1857 struct ofpbuf
*odp_actions
, uint64_t reval_seq
,
1858 struct recirc_refs
*recircs
)
1859 OVS_REQUIRES(ukey
->mutex
)
1861 struct xlate_out xout
, *xoutp
;
1862 struct netflow
*netflow
;
1863 struct ofproto_dpif
*ofproto
;
1864 struct dpif_flow_stats push
;
1866 struct flow_wildcards dp_mask
, wc
;
1867 enum reval_result result
;
1868 ofp_port_t ofp_in_port
;
1869 struct xlate_in xin
;
1870 long long int last_used
;
1872 bool need_revalidate
;
1874 result
= UKEY_DELETE
;
1878 ofpbuf_clear(odp_actions
);
1879 need_revalidate
= (ukey
->reval_seq
!= reval_seq
);
1880 last_used
= ukey
->stats
.used
;
1881 push
.used
= stats
->used
;
1882 push
.tcp_flags
= stats
->tcp_flags
;
1883 push
.n_packets
= (stats
->n_packets
> ukey
->stats
.n_packets
1884 ? stats
->n_packets
- ukey
->stats
.n_packets
1886 push
.n_bytes
= (stats
->n_bytes
> ukey
->stats
.n_bytes
1887 ? stats
->n_bytes
- ukey
->stats
.n_bytes
1890 if (need_revalidate
&& last_used
1891 && !should_revalidate(udpif
, push
.n_packets
, last_used
)) {
1895 /* We will push the stats, so update the ukey stats cache. */
1896 ukey
->stats
= *stats
;
1897 if (!push
.n_packets
&& !need_revalidate
) {
1902 if (ukey
->xcache
&& !need_revalidate
) {
1903 xlate_push_stats(ukey
->xcache
, &push
);
1908 if (odp_flow_key_to_flow(ukey
->key
, ukey
->key_len
, &flow
)
1913 error
= xlate_lookup(udpif
->backer
, &flow
, &ofproto
, NULL
, NULL
, &netflow
,
1919 if (need_revalidate
) {
1920 xlate_cache_clear(ukey
->xcache
);
1922 if (!ukey
->xcache
) {
1923 ukey
->xcache
= xlate_cache_new();
1926 xlate_in_init(&xin
, ofproto
, ofproto_dpif_get_tables_version(ofproto
),
1927 &flow
, ofp_in_port
, NULL
, push
.tcp_flags
,
1928 NULL
, need_revalidate
? &wc
: NULL
, odp_actions
);
1929 if (push
.n_packets
) {
1930 xin
.resubmit_stats
= &push
;
1931 xin
.allow_side_effects
= true;
1933 xin
.xcache
= ukey
->xcache
;
1934 xlate_actions(&xin
, &xout
);
1937 if (!need_revalidate
) {
1943 ofpbuf_clear(odp_actions
);
1944 compose_slow_path(udpif
, &xout
, &flow
, flow
.in_port
.odp_port
,
1948 if (odp_flow_key_to_mask(ukey
->mask
, ukey
->mask_len
, &dp_mask
, &flow
)
1953 /* Do not modify if any bit is wildcarded by the installed datapath flow,
1954 * but not the newly revalidated wildcard mask (wc), i.e., if revalidation
1955 * tells that the datapath flow is now too generic and must be narrowed
1956 * down. Note that we do not know if the datapath has ignored any of the
1957 * wildcarded bits, so we may be overtly conservative here. */
1958 if (flow_wildcards_has_extra(&dp_mask
, &wc
)) {
1962 if (!ofpbuf_equal(odp_actions
,
1963 ovsrcu_get(struct ofpbuf
*, &ukey
->actions
))) {
1964 /* The datapath mask was OK, but the actions seem to have changed.
1965 * Let's modify it in place. */
1966 result
= UKEY_MODIFY
;
1967 /* Transfer recirc action ID references to the caller. */
1968 recirc_refs_swap(recircs
, &xoutp
->recircs
);
1975 if (result
!= UKEY_DELETE
) {
1976 ukey
->reval_seq
= reval_seq
;
1978 if (netflow
&& result
== UKEY_DELETE
) {
1979 netflow_flow_clear(netflow
, &flow
);
1981 xlate_out_uninit(xoutp
);
1986 delete_op_init__(struct udpif
*udpif
, struct ukey_op
*op
,
1987 const struct dpif_flow
*flow
)
1990 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
1991 op
->dop
.u
.flow_del
.key
= flow
->key
;
1992 op
->dop
.u
.flow_del
.key_len
= flow
->key_len
;
1993 op
->dop
.u
.flow_del
.ufid
= flow
->ufid_present
? &flow
->ufid
: NULL
;
1994 op
->dop
.u
.flow_del
.pmd_id
= flow
->pmd_id
;
1995 op
->dop
.u
.flow_del
.stats
= &op
->stats
;
1996 op
->dop
.u
.flow_del
.terse
= udpif_use_ufid(udpif
);
2000 delete_op_init(struct udpif
*udpif
, struct ukey_op
*op
, struct udpif_key
*ukey
)
2003 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
2004 op
->dop
.u
.flow_del
.key
= ukey
->key
;
2005 op
->dop
.u
.flow_del
.key_len
= ukey
->key_len
;
2006 op
->dop
.u
.flow_del
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
2007 op
->dop
.u
.flow_del
.pmd_id
= ukey
->pmd_id
;
2008 op
->dop
.u
.flow_del
.stats
= &op
->stats
;
2009 op
->dop
.u
.flow_del
.terse
= udpif_use_ufid(udpif
);
2013 put_op_init(struct ukey_op
*op
, struct udpif_key
*ukey
,
2014 enum dpif_flow_put_flags flags
)
2017 op
->dop
.type
= DPIF_OP_FLOW_PUT
;
2018 op
->dop
.u
.flow_put
.flags
= flags
;
2019 op
->dop
.u
.flow_put
.key
= ukey
->key
;
2020 op
->dop
.u
.flow_put
.key_len
= ukey
->key_len
;
2021 op
->dop
.u
.flow_put
.mask
= ukey
->mask
;
2022 op
->dop
.u
.flow_put
.mask_len
= ukey
->mask_len
;
2023 op
->dop
.u
.flow_put
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
2024 op
->dop
.u
.flow_put
.pmd_id
= ukey
->pmd_id
;
2025 op
->dop
.u
.flow_put
.stats
= NULL
;
2026 ukey_get_actions(ukey
, &op
->dop
.u
.flow_put
.actions
,
2027 &op
->dop
.u
.flow_put
.actions_len
);
2030 /* Executes datapath operations 'ops' and attributes stats retrieved from the
2031 * datapath as part of those operations. */
2033 push_dp_ops(struct udpif
*udpif
, struct ukey_op
*ops
, size_t n_ops
)
2035 struct dpif_op
*opsp
[REVALIDATE_MAX_BATCH
];
2038 ovs_assert(n_ops
<= REVALIDATE_MAX_BATCH
);
2039 for (i
= 0; i
< n_ops
; i
++) {
2040 opsp
[i
] = &ops
[i
].dop
;
2042 dpif_operate(udpif
->dpif
, opsp
, n_ops
);
2044 for (i
= 0; i
< n_ops
; i
++) {
2045 struct ukey_op
*op
= &ops
[i
];
2046 struct dpif_flow_stats
*push
, *stats
, push_buf
;
2048 stats
= op
->dop
.u
.flow_del
.stats
;
2051 if (op
->dop
.type
!= DPIF_OP_FLOW_DEL
) {
2052 /* Only deleted flows need their stats pushed. */
2056 if (op
->dop
.error
) {
2057 /* flow_del error, 'stats' is unusable. */
2062 ovs_mutex_lock(&op
->ukey
->mutex
);
2063 transition_ukey(op
->ukey
, UKEY_EVICTED
);
2064 push
->used
= MAX(stats
->used
, op
->ukey
->stats
.used
);
2065 push
->tcp_flags
= stats
->tcp_flags
| op
->ukey
->stats
.tcp_flags
;
2066 push
->n_packets
= stats
->n_packets
- op
->ukey
->stats
.n_packets
;
2067 push
->n_bytes
= stats
->n_bytes
- op
->ukey
->stats
.n_bytes
;
2068 ovs_mutex_unlock(&op
->ukey
->mutex
);
2073 if (push
->n_packets
|| netflow_exists()) {
2074 const struct nlattr
*key
= op
->dop
.u
.flow_del
.key
;
2075 size_t key_len
= op
->dop
.u
.flow_del
.key_len
;
2076 struct ofproto_dpif
*ofproto
;
2077 struct netflow
*netflow
;
2078 ofp_port_t ofp_in_port
;
2083 ovs_mutex_lock(&op
->ukey
->mutex
);
2084 if (op
->ukey
->xcache
) {
2085 xlate_push_stats(op
->ukey
->xcache
, push
);
2086 ovs_mutex_unlock(&op
->ukey
->mutex
);
2089 ovs_mutex_unlock(&op
->ukey
->mutex
);
2090 key
= op
->ukey
->key
;
2091 key_len
= op
->ukey
->key_len
;
2094 if (odp_flow_key_to_flow(key
, key_len
, &flow
)
2099 error
= xlate_lookup(udpif
->backer
, &flow
, &ofproto
, NULL
, NULL
,
2100 &netflow
, &ofp_in_port
);
2102 struct xlate_in xin
;
2104 xlate_in_init(&xin
, ofproto
,
2105 ofproto_dpif_get_tables_version(ofproto
),
2106 &flow
, ofp_in_port
, NULL
,
2107 push
->tcp_flags
, NULL
, NULL
, NULL
);
2108 xin
.resubmit_stats
= push
->n_packets
? push
: NULL
;
2109 xin
.allow_side_effects
= push
->n_packets
> 0;
2110 xlate_actions_for_side_effects(&xin
);
2113 netflow_flow_clear(netflow
, &flow
);
2120 /* Executes datapath operations 'ops', attributes stats retrieved from the
2121 * datapath, and deletes ukeys corresponding to deleted flows. */
2123 push_ukey_ops(struct udpif
*udpif
, struct umap
*umap
,
2124 struct ukey_op
*ops
, size_t n_ops
)
2128 push_dp_ops(udpif
, ops
, n_ops
);
2129 ovs_mutex_lock(&umap
->mutex
);
2130 for (i
= 0; i
< n_ops
; i
++) {
2131 if (ops
[i
].dop
.type
== DPIF_OP_FLOW_DEL
) {
2132 ukey_delete(umap
, ops
[i
].ukey
);
2135 ovs_mutex_unlock(&umap
->mutex
);
2139 log_unexpected_flow(const struct dpif_flow
*flow
, int error
)
2141 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(10, 60);
2142 struct ds ds
= DS_EMPTY_INITIALIZER
;
2144 ds_put_format(&ds
, "Failed to acquire udpif_key corresponding to "
2145 "unexpected flow (%s): ", ovs_strerror(error
));
2146 odp_format_ufid(&flow
->ufid
, &ds
);
2147 VLOG_WARN_RL(&rl
, "%s", ds_cstr(&ds
));
2152 reval_op_init(struct ukey_op
*op
, enum reval_result result
,
2153 struct udpif
*udpif
, struct udpif_key
*ukey
,
2154 struct recirc_refs
*recircs
, struct ofpbuf
*odp_actions
)
2155 OVS_REQUIRES(ukey
->mutex
)
2157 if (result
== UKEY_DELETE
) {
2158 delete_op_init(udpif
, op
, ukey
);
2159 transition_ukey(ukey
, UKEY_EVICTING
);
2160 } else if (result
== UKEY_MODIFY
) {
2161 /* Store the new recircs. */
2162 recirc_refs_swap(&ukey
->recircs
, recircs
);
2163 /* Release old recircs. */
2164 recirc_refs_unref(recircs
);
2165 /* ukey->key_recirc_id remains, as the key is the same as before. */
2167 ukey_set_actions(ukey
, odp_actions
);
2168 put_op_init(op
, ukey
, DPIF_FP_MODIFY
);
2173 revalidate(struct revalidator
*revalidator
)
2175 uint64_t odp_actions_stub
[1024 / 8];
2176 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2178 struct udpif
*udpif
= revalidator
->udpif
;
2179 struct dpif_flow_dump_thread
*dump_thread
;
2180 uint64_t dump_seq
, reval_seq
;
2181 unsigned int flow_limit
;
2183 dump_seq
= seq_read(udpif
->dump_seq
);
2184 reval_seq
= seq_read(udpif
->reval_seq
);
2185 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2186 dump_thread
= dpif_flow_dump_thread_create(udpif
->dump
);
2188 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2191 struct dpif_flow flows
[REVALIDATE_MAX_BATCH
];
2192 const struct dpif_flow
*f
;
2195 long long int max_idle
;
2200 n_dumped
= dpif_flow_dump_next(dump_thread
, flows
, ARRAY_SIZE(flows
));
2207 /* In normal operation we want to keep flows around until they have
2208 * been idle for 'ofproto_max_idle' milliseconds. However:
2210 * - If the number of datapath flows climbs above 'flow_limit',
2211 * drop that down to 100 ms to try to bring the flows down to
2214 * - If the number of datapath flows climbs above twice
2215 * 'flow_limit', delete all the datapath flows as an emergency
2216 * measure. (We reassess this condition for the next batch of
2217 * datapath flows, so we will recover before all the flows are
2219 n_dp_flows
= udpif_get_n_flows(udpif
);
2220 kill_them_all
= n_dp_flows
> flow_limit
* 2;
2221 max_idle
= n_dp_flows
> flow_limit
? 100 : ofproto_max_idle
;
2223 for (f
= flows
; f
< &flows
[n_dumped
]; f
++) {
2224 long long int used
= f
->stats
.used
;
2225 struct recirc_refs recircs
= RECIRC_REFS_EMPTY_INITIALIZER
;
2226 enum reval_result result
;
2227 struct udpif_key
*ukey
;
2228 bool already_dumped
;
2231 if (ukey_acquire(udpif
, f
, &ukey
, &error
)) {
2232 if (error
== EBUSY
) {
2233 /* Another thread is processing this flow, so don't bother
2235 COVERAGE_INC(upcall_ukey_contention
);
2237 log_unexpected_flow(f
, error
);
2238 if (error
!= ENOENT
) {
2239 delete_op_init__(udpif
, &ops
[n_ops
++], f
);
2245 already_dumped
= ukey
->dump_seq
== dump_seq
;
2246 if (already_dumped
) {
2247 /* The flow has already been handled during this flow dump
2248 * operation. Skip it. */
2250 COVERAGE_INC(dumped_duplicate_flow
);
2252 COVERAGE_INC(dumped_new_flow
);
2254 ovs_mutex_unlock(&ukey
->mutex
);
2258 /* The flow is now confirmed to be in the datapath. */
2259 transition_ukey(ukey
, UKEY_OPERATIONAL
);
2262 used
= ukey
->created
;
2264 if (kill_them_all
|| (used
&& used
< now
- max_idle
)) {
2265 result
= UKEY_DELETE
;
2267 result
= revalidate_ukey(udpif
, ukey
, &f
->stats
, &odp_actions
,
2268 reval_seq
, &recircs
);
2270 ukey
->dump_seq
= dump_seq
;
2272 if (result
!= UKEY_KEEP
) {
2273 /* Takes ownership of 'recircs'. */
2274 reval_op_init(&ops
[n_ops
++], result
, udpif
, ukey
, &recircs
,
2277 ovs_mutex_unlock(&ukey
->mutex
);
2281 /* Push datapath ops but defer ukey deletion to 'sweep' phase. */
2282 push_dp_ops(udpif
, ops
, n_ops
);
2286 dpif_flow_dump_thread_destroy(dump_thread
);
2287 ofpbuf_uninit(&odp_actions
);
2290 /* Pauses the 'revalidator', can only proceed after main thread
2291 * calls udpif_resume_revalidators(). */
2293 revalidator_pause(struct revalidator
*revalidator
)
2295 /* The first block is for sync'ing the pause with main thread. */
2296 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2297 /* The second block is for pausing until main thread resumes. */
2298 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2302 revalidator_sweep__(struct revalidator
*revalidator
, bool purge
)
2304 struct udpif
*udpif
;
2305 uint64_t dump_seq
, reval_seq
;
2308 udpif
= revalidator
->udpif
;
2309 dump_seq
= seq_read(udpif
->dump_seq
);
2310 reval_seq
= seq_read(udpif
->reval_seq
);
2311 slice
= revalidator
- udpif
->revalidators
;
2312 ovs_assert(slice
< udpif
->n_revalidators
);
2314 for (int i
= slice
; i
< N_UMAPS
; i
+= udpif
->n_revalidators
) {
2315 uint64_t odp_actions_stub
[1024 / 8];
2316 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2318 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2319 struct udpif_key
*ukey
;
2320 struct umap
*umap
= &udpif
->ukeys
[i
];
2323 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2324 enum ukey_state ukey_state
;
2326 /* Handler threads could be holding a ukey lock while it installs a
2327 * new flow, so don't hang around waiting for access to it. */
2328 if (ovs_mutex_trylock(&ukey
->mutex
)) {
2331 ukey_state
= ukey
->state
;
2332 if (ukey_state
== UKEY_OPERATIONAL
2333 || (ukey_state
== UKEY_VISIBLE
&& purge
)) {
2334 struct recirc_refs recircs
= RECIRC_REFS_EMPTY_INITIALIZER
;
2335 bool seq_mismatch
= (ukey
->dump_seq
!= dump_seq
2336 && ukey
->reval_seq
!= reval_seq
);
2337 enum reval_result result
;
2340 result
= UKEY_DELETE
;
2341 } else if (!seq_mismatch
) {
2344 struct dpif_flow_stats stats
;
2345 COVERAGE_INC(revalidate_missed_dp_flow
);
2346 memset(&stats
, 0, sizeof stats
);
2347 result
= revalidate_ukey(udpif
, ukey
, &stats
, &odp_actions
,
2348 reval_seq
, &recircs
);
2350 if (result
!= UKEY_KEEP
) {
2351 /* Clears 'recircs' if filled by revalidate_ukey(). */
2352 reval_op_init(&ops
[n_ops
++], result
, udpif
, ukey
, &recircs
,
2356 ovs_mutex_unlock(&ukey
->mutex
);
2358 if (ukey_state
== UKEY_EVICTED
) {
2359 /* The common flow deletion case involves deletion of the flow
2360 * during the dump phase and ukey deletion here. */
2361 ovs_mutex_lock(&umap
->mutex
);
2362 ukey_delete(umap
, ukey
);
2363 ovs_mutex_unlock(&umap
->mutex
);
2366 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2367 /* Update/delete missed flows and clean up corresponding ukeys
2369 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2375 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2378 ofpbuf_uninit(&odp_actions
);
2384 revalidator_sweep(struct revalidator
*revalidator
)
2386 revalidator_sweep__(revalidator
, false);
2390 revalidator_purge(struct revalidator
*revalidator
)
2392 revalidator_sweep__(revalidator
, true);
2395 /* In reaction to dpif purge, purges all 'ukey's with same 'pmd_id'. */
2397 dp_purge_cb(void *aux
, unsigned pmd_id
)
2398 OVS_NO_THREAD_SAFETY_ANALYSIS
2400 struct udpif
*udpif
= aux
;
2403 udpif_pause_revalidators(udpif
);
2404 for (i
= 0; i
< N_UMAPS
; i
++) {
2405 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2406 struct udpif_key
*ukey
;
2407 struct umap
*umap
= &udpif
->ukeys
[i
];
2410 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2411 if (ukey
->pmd_id
== pmd_id
) {
2412 delete_op_init(udpif
, &ops
[n_ops
++], ukey
);
2413 transition_ukey(ukey
, UKEY_EVICTING
);
2415 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2416 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2423 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2428 udpif_resume_revalidators(udpif
);
2432 upcall_unixctl_show(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2433 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2435 struct ds ds
= DS_EMPTY_INITIALIZER
;
2436 struct udpif
*udpif
;
2438 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2439 unsigned int flow_limit
;
2443 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2444 ufid_enabled
= udpif_use_ufid(udpif
);
2446 ds_put_format(&ds
, "%s:\n", dpif_name(udpif
->dpif
));
2447 ds_put_format(&ds
, "\tflows : (current %lu)"
2448 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif
),
2449 udpif
->avg_n_flows
, udpif
->max_n_flows
, flow_limit
);
2450 ds_put_format(&ds
, "\tdump duration : %lldms\n", udpif
->dump_duration
);
2451 ds_put_format(&ds
, "\tufid enabled : ");
2453 ds_put_format(&ds
, "true\n");
2455 ds_put_format(&ds
, "false\n");
2457 ds_put_char(&ds
, '\n');
2459 for (i
= 0; i
< n_revalidators
; i
++) {
2460 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
2461 int j
, elements
= 0;
2463 for (j
= i
; j
< N_UMAPS
; j
+= n_revalidators
) {
2464 elements
+= cmap_count(&udpif
->ukeys
[j
].cmap
);
2466 ds_put_format(&ds
, "\t%u: (keys %d)\n", revalidator
->id
, elements
);
2470 unixctl_command_reply(conn
, ds_cstr(&ds
));
2474 /* Disable using the megaflows.
2476 * This command is only needed for advanced debugging, so it's not
2477 * documented in the man page. */
2479 upcall_unixctl_disable_megaflows(struct unixctl_conn
*conn
,
2480 int argc OVS_UNUSED
,
2481 const char *argv
[] OVS_UNUSED
,
2482 void *aux OVS_UNUSED
)
2484 atomic_store_relaxed(&enable_megaflows
, false);
2485 udpif_flush_all_datapaths();
2486 unixctl_command_reply(conn
, "megaflows disabled");
2489 /* Re-enable using megaflows.
2491 * This command is only needed for advanced debugging, so it's not
2492 * documented in the man page. */
2494 upcall_unixctl_enable_megaflows(struct unixctl_conn
*conn
,
2495 int argc OVS_UNUSED
,
2496 const char *argv
[] OVS_UNUSED
,
2497 void *aux OVS_UNUSED
)
2499 atomic_store_relaxed(&enable_megaflows
, true);
2500 udpif_flush_all_datapaths();
2501 unixctl_command_reply(conn
, "megaflows enabled");
2504 /* Disable skipping flow attributes during flow dump.
2506 * This command is only needed for advanced debugging, so it's not
2507 * documented in the man page. */
2509 upcall_unixctl_disable_ufid(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2510 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2512 atomic_store_relaxed(&enable_ufid
, false);
2513 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID disabled");
2516 /* Re-enable skipping flow attributes during flow dump.
2518 * This command is only needed for advanced debugging, so it's not documented
2519 * in the man page. */
2521 upcall_unixctl_enable_ufid(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2522 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2524 atomic_store_relaxed(&enable_ufid
, true);
2525 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID enabled "
2526 "for supported datapaths");
2529 /* Set the flow limit.
2531 * This command is only needed for advanced debugging, so it's not
2532 * documented in the man page. */
2534 upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
,
2535 int argc OVS_UNUSED
,
2536 const char *argv
[] OVS_UNUSED
,
2537 void *aux OVS_UNUSED
)
2539 struct ds ds
= DS_EMPTY_INITIALIZER
;
2540 struct udpif
*udpif
;
2541 unsigned int flow_limit
= atoi(argv
[1]);
2543 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2544 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
2546 ds_put_format(&ds
, "set flow_limit to %u\n", flow_limit
);
2547 unixctl_command_reply(conn
, ds_cstr(&ds
));
2552 upcall_unixctl_dump_wait(struct unixctl_conn
*conn
,
2553 int argc OVS_UNUSED
,
2554 const char *argv
[] OVS_UNUSED
,
2555 void *aux OVS_UNUSED
)
2557 if (ovs_list_is_singleton(&all_udpifs
)) {
2558 struct udpif
*udpif
= NULL
;
2561 udpif
= OBJECT_CONTAINING(ovs_list_front(&all_udpifs
), udpif
, list_node
);
2562 len
= (udpif
->n_conns
+ 1) * sizeof *udpif
->conns
;
2563 udpif
->conn_seq
= seq_read(udpif
->dump_seq
);
2564 udpif
->conns
= xrealloc(udpif
->conns
, len
);
2565 udpif
->conns
[udpif
->n_conns
++] = conn
;
2567 unixctl_command_reply_error(conn
, "can't wait on multiple udpifs.");
2572 upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2573 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2575 struct udpif
*udpif
;
2577 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2580 for (n
= 0; n
< udpif
->n_revalidators
; n
++) {
2581 revalidator_purge(&udpif
->revalidators
[n
]);
2584 unixctl_command_reply(conn
, "");