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 for (int i
= 0; i
< N_UMAPS
; i
++) {
438 cmap_destroy(&udpif
->ukeys
[i
].cmap
);
439 ovs_mutex_destroy(&udpif
->ukeys
[i
].mutex
);
444 ovs_list_remove(&udpif
->list_node
);
445 latch_destroy(&udpif
->exit_latch
);
446 latch_destroy(&udpif
->pause_latch
);
447 seq_destroy(udpif
->reval_seq
);
448 seq_destroy(udpif
->dump_seq
);
449 ovs_mutex_destroy(&udpif
->n_flows_mutex
);
453 /* Stops the handler and revalidator threads, must be enclosed in
454 * ovsrcu quiescent state unless when destroying udpif. */
456 udpif_stop_threads(struct udpif
*udpif
)
458 if (udpif
&& (udpif
->n_handlers
!= 0 || udpif
->n_revalidators
!= 0)) {
461 latch_set(&udpif
->exit_latch
);
463 for (i
= 0; i
< udpif
->n_handlers
; i
++) {
464 struct handler
*handler
= &udpif
->handlers
[i
];
466 xpthread_join(handler
->thread
, NULL
);
469 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
470 xpthread_join(udpif
->revalidators
[i
].thread
, NULL
);
473 dpif_disable_upcall(udpif
->dpif
);
475 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
476 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
478 /* Delete ukeys, and delete all flows from the datapath to prevent
479 * double-counting stats. */
480 revalidator_purge(revalidator
);
483 latch_poll(&udpif
->exit_latch
);
485 ovs_barrier_destroy(&udpif
->reval_barrier
);
486 ovs_barrier_destroy(&udpif
->pause_barrier
);
488 free(udpif
->revalidators
);
489 udpif
->revalidators
= NULL
;
490 udpif
->n_revalidators
= 0;
492 free(udpif
->handlers
);
493 udpif
->handlers
= NULL
;
494 udpif
->n_handlers
= 0;
498 /* Starts the handler and revalidator threads, must be enclosed in
499 * ovsrcu quiescent state. */
501 udpif_start_threads(struct udpif
*udpif
, size_t n_handlers
,
502 size_t n_revalidators
)
504 if (udpif
&& n_handlers
&& n_revalidators
) {
508 udpif
->n_handlers
= n_handlers
;
509 udpif
->n_revalidators
= n_revalidators
;
511 udpif
->handlers
= xzalloc(udpif
->n_handlers
* sizeof *udpif
->handlers
);
512 for (i
= 0; i
< udpif
->n_handlers
; i
++) {
513 struct handler
*handler
= &udpif
->handlers
[i
];
515 handler
->udpif
= udpif
;
516 handler
->handler_id
= i
;
517 handler
->thread
= ovs_thread_create(
518 "handler", udpif_upcall_handler
, handler
);
521 enable_ufid
= ofproto_dpif_get_enable_ufid(udpif
->backer
);
522 atomic_init(&udpif
->enable_ufid
, enable_ufid
);
523 dpif_enable_upcall(udpif
->dpif
);
525 ovs_barrier_init(&udpif
->reval_barrier
, udpif
->n_revalidators
);
526 ovs_barrier_init(&udpif
->pause_barrier
, udpif
->n_revalidators
+ 1);
527 udpif
->reval_exit
= false;
528 udpif
->pause
= false;
529 udpif
->revalidators
= xzalloc(udpif
->n_revalidators
530 * sizeof *udpif
->revalidators
);
531 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
532 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
534 revalidator
->udpif
= udpif
;
535 revalidator
->thread
= ovs_thread_create(
536 "revalidator", udpif_revalidator
, revalidator
);
541 /* Pauses all revalidators. Should only be called by the main thread.
542 * When function returns, all revalidators are paused and will proceed
543 * only after udpif_resume_revalidators() is called. */
545 udpif_pause_revalidators(struct udpif
*udpif
)
547 if (ofproto_dpif_backer_enabled(udpif
->backer
)) {
548 latch_set(&udpif
->pause_latch
);
549 ovs_barrier_block(&udpif
->pause_barrier
);
553 /* Resumes the pausing of revalidators. Should only be called by the
556 udpif_resume_revalidators(struct udpif
*udpif
)
558 if (ofproto_dpif_backer_enabled(udpif
->backer
)) {
559 latch_poll(&udpif
->pause_latch
);
560 ovs_barrier_block(&udpif
->pause_barrier
);
564 /* Tells 'udpif' how many threads it should use to handle upcalls.
565 * 'n_handlers' and 'n_revalidators' can never be zero. 'udpif''s
566 * datapath handle must have packet reception enabled before starting
569 udpif_set_threads(struct udpif
*udpif
, size_t n_handlers
,
570 size_t n_revalidators
)
573 ovs_assert(n_handlers
&& n_revalidators
);
575 ovsrcu_quiesce_start();
576 if (udpif
->n_handlers
!= n_handlers
577 || udpif
->n_revalidators
!= n_revalidators
) {
578 udpif_stop_threads(udpif
);
581 if (!udpif
->handlers
&& !udpif
->revalidators
) {
584 error
= dpif_handlers_set(udpif
->dpif
, n_handlers
);
586 VLOG_ERR("failed to configure handlers in dpif %s: %s",
587 dpif_name(udpif
->dpif
), ovs_strerror(error
));
591 udpif_start_threads(udpif
, n_handlers
, n_revalidators
);
593 ovsrcu_quiesce_end();
596 /* Waits for all ongoing upcall translations to complete. This ensures that
597 * there are no transient references to any removed ofprotos (or other
598 * objects). In particular, this should be called after an ofproto is removed
599 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
601 udpif_synchronize(struct udpif
*udpif
)
603 /* This is stronger than necessary. It would be sufficient to ensure
604 * (somehow) that each handler and revalidator thread had passed through
605 * its main loop once. */
606 size_t n_handlers
= udpif
->n_handlers
;
607 size_t n_revalidators
= udpif
->n_revalidators
;
609 ovsrcu_quiesce_start();
610 udpif_stop_threads(udpif
);
611 udpif_start_threads(udpif
, n_handlers
, n_revalidators
);
612 ovsrcu_quiesce_end();
615 /* Notifies 'udpif' that something changed which may render previous
616 * xlate_actions() results invalid. */
618 udpif_revalidate(struct udpif
*udpif
)
620 seq_change(udpif
->reval_seq
);
623 /* Returns a seq which increments every time 'udpif' pulls stats from the
624 * datapath. Callers can use this to get a sense of when might be a good time
625 * to do periodic work which relies on relatively up to date statistics. */
627 udpif_dump_seq(struct udpif
*udpif
)
629 return udpif
->dump_seq
;
633 udpif_get_memory_usage(struct udpif
*udpif
, struct simap
*usage
)
637 simap_increase(usage
, "handlers", udpif
->n_handlers
);
639 simap_increase(usage
, "revalidators", udpif
->n_revalidators
);
640 for (i
= 0; i
< N_UMAPS
; i
++) {
641 simap_increase(usage
, "udpif keys", cmap_count(&udpif
->ukeys
[i
].cmap
));
645 /* Remove flows from a single datapath. */
647 udpif_flush(struct udpif
*udpif
)
649 size_t n_handlers
, n_revalidators
;
651 n_handlers
= udpif
->n_handlers
;
652 n_revalidators
= udpif
->n_revalidators
;
654 ovsrcu_quiesce_start();
656 udpif_stop_threads(udpif
);
657 dpif_flow_flush(udpif
->dpif
);
658 udpif_start_threads(udpif
, n_handlers
, n_revalidators
);
660 ovsrcu_quiesce_end();
663 /* Removes all flows from all datapaths. */
665 udpif_flush_all_datapaths(void)
669 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
675 udpif_use_ufid(struct udpif
*udpif
)
679 atomic_read_relaxed(&enable_ufid
, &enable
);
680 return enable
&& ofproto_dpif_get_enable_ufid(udpif
->backer
);
685 udpif_get_n_flows(struct udpif
*udpif
)
687 long long int time
, now
;
688 unsigned long flow_count
;
691 atomic_read_relaxed(&udpif
->n_flows_timestamp
, &time
);
692 if (time
< now
- 100 && !ovs_mutex_trylock(&udpif
->n_flows_mutex
)) {
693 struct dpif_dp_stats stats
;
695 atomic_store_relaxed(&udpif
->n_flows_timestamp
, now
);
696 dpif_get_dp_stats(udpif
->dpif
, &stats
);
697 flow_count
= stats
.n_flows
;
698 atomic_store_relaxed(&udpif
->n_flows
, flow_count
);
699 ovs_mutex_unlock(&udpif
->n_flows_mutex
);
701 atomic_read_relaxed(&udpif
->n_flows
, &flow_count
);
706 /* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH
707 * upcalls from dpif, processes the batch and installs corresponding flows
710 udpif_upcall_handler(void *arg
)
712 struct handler
*handler
= arg
;
713 struct udpif
*udpif
= handler
->udpif
;
715 while (!latch_is_set(&handler
->udpif
->exit_latch
)) {
716 if (recv_upcalls(handler
)) {
717 poll_immediate_wake();
719 dpif_recv_wait(udpif
->dpif
, handler
->handler_id
);
720 latch_wait(&udpif
->exit_latch
);
729 recv_upcalls(struct handler
*handler
)
731 struct udpif
*udpif
= handler
->udpif
;
732 uint64_t recv_stubs
[UPCALL_MAX_BATCH
][512 / 8];
733 struct ofpbuf recv_bufs
[UPCALL_MAX_BATCH
];
734 struct dpif_upcall dupcalls
[UPCALL_MAX_BATCH
];
735 struct upcall upcalls
[UPCALL_MAX_BATCH
];
736 struct flow flows
[UPCALL_MAX_BATCH
];
740 while (n_upcalls
< UPCALL_MAX_BATCH
) {
741 struct ofpbuf
*recv_buf
= &recv_bufs
[n_upcalls
];
742 struct dpif_upcall
*dupcall
= &dupcalls
[n_upcalls
];
743 struct upcall
*upcall
= &upcalls
[n_upcalls
];
744 struct flow
*flow
= &flows
[n_upcalls
];
748 ofpbuf_use_stub(recv_buf
, recv_stubs
[n_upcalls
],
749 sizeof recv_stubs
[n_upcalls
]);
750 if (dpif_recv(udpif
->dpif
, handler
->handler_id
, dupcall
, recv_buf
)) {
751 ofpbuf_uninit(recv_buf
);
755 if (odp_flow_key_to_flow(dupcall
->key
, dupcall
->key_len
, flow
)
761 mru
= nl_attr_get_u16(dupcall
->mru
);
766 error
= upcall_receive(upcall
, udpif
->backer
, &dupcall
->packet
,
767 dupcall
->type
, dupcall
->userdata
, flow
, mru
,
768 &dupcall
->ufid
, PMD_ID_NULL
);
770 if (error
== ENODEV
) {
771 /* Received packet on datapath port for which we couldn't
772 * associate an ofproto. This can happen if a port is removed
773 * while traffic is being received. Print a rate-limited
774 * message in case it happens frequently. */
775 dpif_flow_put(udpif
->dpif
, DPIF_FP_CREATE
, dupcall
->key
,
776 dupcall
->key_len
, NULL
, 0, NULL
, 0,
777 &dupcall
->ufid
, PMD_ID_NULL
, NULL
);
778 VLOG_INFO_RL(&rl
, "received packet on unassociated datapath "
779 "port %"PRIu32
, flow
->in_port
.odp_port
);
784 upcall
->key
= dupcall
->key
;
785 upcall
->key_len
= dupcall
->key_len
;
786 upcall
->ufid
= &dupcall
->ufid
;
788 upcall
->out_tun_key
= dupcall
->out_tun_key
;
789 upcall
->actions
= dupcall
->actions
;
791 pkt_metadata_from_flow(&dupcall
->packet
.md
, flow
);
792 flow_extract(&dupcall
->packet
, flow
);
794 error
= process_upcall(udpif
, upcall
,
795 &upcall
->odp_actions
, &upcall
->wc
);
804 upcall_uninit(upcall
);
806 dp_packet_uninit(&dupcall
->packet
);
807 ofpbuf_uninit(recv_buf
);
811 handle_upcalls(handler
->udpif
, upcalls
, n_upcalls
);
812 for (i
= 0; i
< n_upcalls
; i
++) {
813 dp_packet_uninit(&dupcalls
[i
].packet
);
814 ofpbuf_uninit(&recv_bufs
[i
]);
815 upcall_uninit(&upcalls
[i
]);
823 udpif_revalidator(void *arg
)
825 /* Used by all revalidators. */
826 struct revalidator
*revalidator
= arg
;
827 struct udpif
*udpif
= revalidator
->udpif
;
828 bool leader
= revalidator
== &udpif
->revalidators
[0];
830 /* Used only by the leader. */
831 long long int start_time
= 0;
832 uint64_t last_reval_seq
= 0;
835 revalidator
->id
= ovsthread_id_self();
840 recirc_run(); /* Recirculation cleanup. */
842 reval_seq
= seq_read(udpif
->reval_seq
);
843 last_reval_seq
= reval_seq
;
845 n_flows
= udpif_get_n_flows(udpif
);
846 udpif
->max_n_flows
= MAX(n_flows
, udpif
->max_n_flows
);
847 udpif
->avg_n_flows
= (udpif
->avg_n_flows
+ n_flows
) / 2;
849 /* Only the leader checks the pause latch to prevent a race where
850 * some threads think it's false and proceed to block on
851 * reval_barrier and others think it's true and block indefinitely
852 * on the pause_barrier */
853 udpif
->pause
= latch_is_set(&udpif
->pause_latch
);
855 /* Only the leader checks the exit latch to prevent a race where
856 * some threads think it's true and exit and others think it's
857 * false and block indefinitely on the reval_barrier */
858 udpif
->reval_exit
= latch_is_set(&udpif
->exit_latch
);
860 start_time
= time_msec();
861 if (!udpif
->reval_exit
) {
864 terse_dump
= udpif_use_ufid(udpif
);
865 udpif
->dump
= dpif_flow_dump_create(udpif
->dpif
, terse_dump
);
869 /* Wait for the leader to start the flow dump. */
870 ovs_barrier_block(&udpif
->reval_barrier
);
872 revalidator_pause(revalidator
);
875 if (udpif
->reval_exit
) {
878 revalidate(revalidator
);
880 /* Wait for all flows to have been dumped before we garbage collect. */
881 ovs_barrier_block(&udpif
->reval_barrier
);
882 revalidator_sweep(revalidator
);
884 /* Wait for all revalidators to finish garbage collection. */
885 ovs_barrier_block(&udpif
->reval_barrier
);
888 unsigned int flow_limit
;
889 long long int duration
;
891 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
893 dpif_flow_dump_destroy(udpif
->dump
);
894 seq_change(udpif
->dump_seq
);
896 duration
= MAX(time_msec() - start_time
, 1);
897 udpif
->dump_duration
= duration
;
898 if (duration
> 2000) {
899 flow_limit
/= duration
/ 1000;
900 } else if (duration
> 1300) {
901 flow_limit
= flow_limit
* 3 / 4;
902 } else if (duration
< 1000 && n_flows
> 2000
903 && flow_limit
< n_flows
* 1000 / duration
) {
906 flow_limit
= MIN(ofproto_flow_limit
, MAX(flow_limit
, 1000));
907 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
909 if (duration
> 2000) {
910 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
914 poll_timer_wait_until(start_time
+ MIN(ofproto_max_idle
, 500));
915 seq_wait(udpif
->reval_seq
, last_reval_seq
);
916 latch_wait(&udpif
->exit_latch
);
917 latch_wait(&udpif
->pause_latch
);
925 static enum upcall_type
926 classify_upcall(enum dpif_upcall_type type
, const struct nlattr
*userdata
)
928 union user_action_cookie cookie
;
931 /* First look at the upcall type. */
939 case DPIF_N_UC_TYPES
:
941 VLOG_WARN_RL(&rl
, "upcall has unexpected type %"PRIu32
, type
);
945 /* "action" upcalls need a closer look. */
947 VLOG_WARN_RL(&rl
, "action upcall missing cookie");
950 userdata_len
= nl_attr_get_size(userdata
);
951 if (userdata_len
< sizeof cookie
.type
952 || userdata_len
> sizeof cookie
) {
953 VLOG_WARN_RL(&rl
, "action upcall cookie has unexpected size %"PRIuSIZE
,
957 memset(&cookie
, 0, sizeof cookie
);
958 memcpy(&cookie
, nl_attr_get(userdata
), userdata_len
);
959 if (userdata_len
== MAX(8, sizeof cookie
.sflow
)
960 && cookie
.type
== USER_ACTION_COOKIE_SFLOW
) {
962 } else if (userdata_len
== MAX(8, sizeof cookie
.slow_path
)
963 && cookie
.type
== USER_ACTION_COOKIE_SLOW_PATH
) {
965 } else if (userdata_len
== MAX(8, sizeof cookie
.flow_sample
)
966 && cookie
.type
== USER_ACTION_COOKIE_FLOW_SAMPLE
) {
967 return FLOW_SAMPLE_UPCALL
;
968 } else if (userdata_len
== MAX(8, sizeof cookie
.ipfix
)
969 && cookie
.type
== USER_ACTION_COOKIE_IPFIX
) {
972 VLOG_WARN_RL(&rl
, "invalid user cookie of type %"PRIu16
973 " and size %"PRIuSIZE
, cookie
.type
, userdata_len
);
978 /* Calculates slow path actions for 'xout'. 'buf' must statically be
979 * initialized with at least 128 bytes of space. */
981 compose_slow_path(struct udpif
*udpif
, struct xlate_out
*xout
,
982 const struct flow
*flow
, odp_port_t odp_in_port
,
985 union user_action_cookie cookie
;
989 cookie
.type
= USER_ACTION_COOKIE_SLOW_PATH
;
990 cookie
.slow_path
.unused
= 0;
991 cookie
.slow_path
.reason
= xout
->slow
;
993 port
= xout
->slow
& (SLOW_CFM
| SLOW_BFD
| SLOW_LACP
| SLOW_STP
)
996 pid
= dpif_port_get_pid(udpif
->dpif
, port
, flow_hash_5tuple(flow
, 0));
997 odp_put_userspace_action(pid
, &cookie
, sizeof cookie
.slow_path
,
998 ODPP_NONE
, false, buf
);
1001 /* If there is no error, the upcall must be destroyed with upcall_uninit()
1002 * before quiescing, as the referred objects are guaranteed to exist only
1003 * until the calling thread quiesces. Otherwise, do not call upcall_uninit()
1004 * since the 'upcall->put_actions' remains uninitialized. */
1006 upcall_receive(struct upcall
*upcall
, const struct dpif_backer
*backer
,
1007 const struct dp_packet
*packet
, enum dpif_upcall_type type
,
1008 const struct nlattr
*userdata
, const struct flow
*flow
,
1009 const unsigned int mru
,
1010 const ovs_u128
*ufid
, const unsigned pmd_id
)
1014 error
= xlate_lookup(backer
, flow
, &upcall
->ofproto
, &upcall
->ipfix
,
1015 &upcall
->sflow
, NULL
, &upcall
->in_port
);
1020 upcall
->recirc
= NULL
;
1021 upcall
->have_recirc_ref
= false;
1022 upcall
->flow
= flow
;
1023 upcall
->packet
= packet
;
1024 upcall
->ufid
= ufid
;
1025 upcall
->pmd_id
= pmd_id
;
1026 upcall
->type
= type
;
1027 upcall
->userdata
= userdata
;
1028 ofpbuf_use_stub(&upcall
->odp_actions
, upcall
->odp_actions_stub
,
1029 sizeof upcall
->odp_actions_stub
);
1030 ofpbuf_init(&upcall
->put_actions
, 0);
1032 upcall
->xout_initialized
= false;
1033 upcall
->ukey_persists
= false;
1035 upcall
->ukey
= NULL
;
1037 upcall
->key_len
= 0;
1040 upcall
->out_tun_key
= NULL
;
1041 upcall
->actions
= NULL
;
1047 upcall_xlate(struct udpif
*udpif
, struct upcall
*upcall
,
1048 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1050 struct dpif_flow_stats stats
;
1051 struct xlate_in xin
;
1053 stats
.n_packets
= 1;
1054 stats
.n_bytes
= dp_packet_size(upcall
->packet
);
1055 stats
.used
= time_msec();
1056 stats
.tcp_flags
= ntohs(upcall
->flow
->tcp_flags
);
1058 xlate_in_init(&xin
, upcall
->ofproto
, upcall
->flow
, upcall
->in_port
, NULL
,
1059 stats
.tcp_flags
, upcall
->packet
, wc
, odp_actions
);
1061 if (upcall
->type
== DPIF_UC_MISS
) {
1062 xin
.resubmit_stats
= &stats
;
1064 if (xin
.frozen_state
) {
1065 /* We may install a datapath flow only if we get a reference to the
1066 * recirculation context (otherwise we could have recirculation
1067 * upcalls using recirculation ID for which no context can be
1068 * found). We may still execute the flow's actions even if we
1069 * don't install the flow. */
1070 upcall
->recirc
= recirc_id_node_from_state(xin
.frozen_state
);
1071 upcall
->have_recirc_ref
= recirc_id_node_try_ref_rcu(upcall
->recirc
);
1074 /* For non-miss upcalls, we are either executing actions (one of which
1075 * is an userspace action) for an upcall, in which case the stats have
1076 * already been taken care of, or there's a flow in the datapath which
1077 * this packet was accounted to. Presumably the revalidators will deal
1078 * with pushing its stats eventually. */
1081 upcall
->dump_seq
= seq_read(udpif
->dump_seq
);
1082 upcall
->reval_seq
= seq_read(udpif
->reval_seq
);
1083 xlate_actions(&xin
, &upcall
->xout
);
1084 upcall
->xout_initialized
= true;
1086 if (!upcall
->xout
.slow
) {
1087 ofpbuf_use_const(&upcall
->put_actions
,
1088 odp_actions
->data
, odp_actions
->size
);
1090 /* upcall->put_actions already initialized by upcall_receive(). */
1091 compose_slow_path(udpif
, &upcall
->xout
, upcall
->flow
,
1092 upcall
->flow
->in_port
.odp_port
,
1093 &upcall
->put_actions
);
1096 /* This function is also called for slow-pathed flows. As we are only
1097 * going to create new datapath flows for actual datapath misses, there is
1098 * no point in creating a ukey otherwise. */
1099 if (upcall
->type
== DPIF_UC_MISS
) {
1100 upcall
->ukey
= ukey_create_from_upcall(upcall
, wc
);
1105 upcall_uninit(struct upcall
*upcall
)
1108 if (upcall
->xout_initialized
) {
1109 xlate_out_uninit(&upcall
->xout
);
1111 ofpbuf_uninit(&upcall
->odp_actions
);
1112 ofpbuf_uninit(&upcall
->put_actions
);
1114 if (!upcall
->ukey_persists
) {
1115 ukey_delete__(upcall
->ukey
);
1117 } else if (upcall
->have_recirc_ref
) {
1118 /* The reference was transferred to the ukey if one was created. */
1119 recirc_id_node_unref(upcall
->recirc
);
1125 upcall_cb(const struct dp_packet
*packet
, const struct flow
*flow
, ovs_u128
*ufid
,
1126 unsigned pmd_id
, enum dpif_upcall_type type
,
1127 const struct nlattr
*userdata
, struct ofpbuf
*actions
,
1128 struct flow_wildcards
*wc
, struct ofpbuf
*put_actions
, void *aux
)
1130 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 1);
1131 struct udpif
*udpif
= aux
;
1132 unsigned int flow_limit
;
1133 struct upcall upcall
;
1137 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1138 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
1140 error
= upcall_receive(&upcall
, udpif
->backer
, packet
, type
, userdata
,
1141 flow
, 0, ufid
, pmd_id
);
1146 error
= process_upcall(udpif
, &upcall
, actions
, wc
);
1151 if (upcall
.xout
.slow
&& put_actions
) {
1152 ofpbuf_put(put_actions
, upcall
.put_actions
.data
,
1153 upcall
.put_actions
.size
);
1156 if (OVS_UNLIKELY(!megaflow
)) {
1157 flow_wildcards_init_for_packet(wc
, flow
);
1160 if (udpif_get_n_flows(udpif
) >= flow_limit
) {
1161 VLOG_WARN_RL(&rl
, "upcall_cb failure: datapath flow limit reached");
1166 /* Prevent miss flow installation if the key has recirculation ID but we
1167 * were not able to get a reference on it. */
1168 if (type
== DPIF_UC_MISS
&& upcall
.recirc
&& !upcall
.have_recirc_ref
) {
1169 VLOG_WARN_RL(&rl
, "upcall_cb failure: no reference for recirc flow");
1174 if (upcall
.ukey
&& !ukey_install(udpif
, upcall
.ukey
)) {
1175 VLOG_WARN_RL(&rl
, "upcall_cb failure: ukey installation fails");
1180 upcall
.ukey_persists
= true;
1182 upcall_uninit(&upcall
);
1187 process_upcall(struct udpif
*udpif
, struct upcall
*upcall
,
1188 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1190 const struct nlattr
*userdata
= upcall
->userdata
;
1191 const struct dp_packet
*packet
= upcall
->packet
;
1192 const struct flow
*flow
= upcall
->flow
;
1194 switch (classify_upcall(upcall
->type
, userdata
)) {
1196 upcall_xlate(udpif
, upcall
, odp_actions
, wc
);
1200 if (upcall
->sflow
) {
1201 union user_action_cookie cookie
;
1202 const struct nlattr
*actions
;
1203 size_t actions_len
= 0;
1204 struct dpif_sflow_actions sflow_actions
;
1205 memset(&sflow_actions
, 0, sizeof sflow_actions
);
1206 memset(&cookie
, 0, sizeof cookie
);
1207 memcpy(&cookie
, nl_attr_get(userdata
), sizeof cookie
.sflow
);
1208 if (upcall
->actions
) {
1209 /* Actions were passed up from datapath. */
1210 actions
= nl_attr_get(upcall
->actions
);
1211 actions_len
= nl_attr_get_size(upcall
->actions
);
1212 if (actions
&& actions_len
) {
1213 dpif_sflow_read_actions(flow
, actions
, actions_len
,
1217 if (actions_len
== 0) {
1218 /* Lookup actions in userspace cache. */
1219 struct udpif_key
*ukey
= ukey_lookup(udpif
, upcall
->ufid
,
1222 ukey_get_actions(ukey
, &actions
, &actions_len
);
1223 dpif_sflow_read_actions(flow
, actions
, actions_len
,
1227 dpif_sflow_received(upcall
->sflow
, packet
, flow
,
1228 flow
->in_port
.odp_port
, &cookie
,
1229 actions_len
> 0 ? &sflow_actions
: NULL
);
1234 if (upcall
->ipfix
) {
1235 union user_action_cookie cookie
;
1236 struct flow_tnl output_tunnel_key
;
1238 memset(&cookie
, 0, sizeof cookie
);
1239 memcpy(&cookie
, nl_attr_get(userdata
), sizeof cookie
.ipfix
);
1241 if (upcall
->out_tun_key
) {
1242 odp_tun_key_from_attr(upcall
->out_tun_key
, false,
1243 &output_tunnel_key
);
1245 dpif_ipfix_bridge_sample(upcall
->ipfix
, packet
, flow
,
1246 flow
->in_port
.odp_port
,
1247 cookie
.ipfix
.output_odp_port
,
1248 upcall
->out_tun_key
?
1249 &output_tunnel_key
: NULL
);
1253 case FLOW_SAMPLE_UPCALL
:
1254 if (upcall
->ipfix
) {
1255 union user_action_cookie cookie
;
1257 memset(&cookie
, 0, sizeof cookie
);
1258 memcpy(&cookie
, nl_attr_get(userdata
), sizeof cookie
.flow_sample
);
1260 /* The flow reflects exactly the contents of the packet.
1261 * Sample the packet using it. */
1262 dpif_ipfix_flow_sample(upcall
->ipfix
, packet
, flow
,
1263 cookie
.flow_sample
.collector_set_id
,
1264 cookie
.flow_sample
.probability
,
1265 cookie
.flow_sample
.obs_domain_id
,
1266 cookie
.flow_sample
.obs_point_id
);
1278 handle_upcalls(struct udpif
*udpif
, struct upcall
*upcalls
,
1281 struct dpif_op
*opsp
[UPCALL_MAX_BATCH
* 2];
1282 struct ukey_op ops
[UPCALL_MAX_BATCH
* 2];
1283 unsigned int flow_limit
;
1284 size_t n_ops
, n_opsp
, i
;
1287 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
1289 may_put
= udpif_get_n_flows(udpif
) < flow_limit
;
1291 /* Handle the packets individually in order of arrival.
1293 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1294 * processes received packets for these protocols.
1296 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1299 * The loop fills 'ops' with an array of operations to execute in the
1302 for (i
= 0; i
< n_upcalls
; i
++) {
1303 struct upcall
*upcall
= &upcalls
[i
];
1304 const struct dp_packet
*packet
= upcall
->packet
;
1307 /* Do not install a flow into the datapath if:
1309 * - The datapath already has too many flows.
1311 * - We received this packet via some flow installed in the kernel
1314 * - Upcall was a recirculation but we do not have a reference to
1315 * to the recirculation ID. */
1316 if (may_put
&& upcall
->type
== DPIF_UC_MISS
&&
1317 (!upcall
->recirc
|| upcall
->have_recirc_ref
)) {
1318 struct udpif_key
*ukey
= upcall
->ukey
;
1320 upcall
->ukey_persists
= true;
1324 op
->dop
.type
= DPIF_OP_FLOW_PUT
;
1325 op
->dop
.u
.flow_put
.flags
= DPIF_FP_CREATE
;
1326 op
->dop
.u
.flow_put
.key
= ukey
->key
;
1327 op
->dop
.u
.flow_put
.key_len
= ukey
->key_len
;
1328 op
->dop
.u
.flow_put
.mask
= ukey
->mask
;
1329 op
->dop
.u
.flow_put
.mask_len
= ukey
->mask_len
;
1330 op
->dop
.u
.flow_put
.ufid
= upcall
->ufid
;
1331 op
->dop
.u
.flow_put
.stats
= NULL
;
1332 ukey_get_actions(ukey
, &op
->dop
.u
.flow_put
.actions
,
1333 &op
->dop
.u
.flow_put
.actions_len
);
1336 if (upcall
->odp_actions
.size
) {
1339 op
->dop
.type
= DPIF_OP_EXECUTE
;
1340 op
->dop
.u
.execute
.packet
= CONST_CAST(struct dp_packet
*, packet
);
1341 op
->dop
.u
.execute
.flow
= upcall
->flow
;
1342 odp_key_to_pkt_metadata(upcall
->key
, upcall
->key_len
,
1343 &op
->dop
.u
.execute
.packet
->md
);
1344 op
->dop
.u
.execute
.actions
= upcall
->odp_actions
.data
;
1345 op
->dop
.u
.execute
.actions_len
= upcall
->odp_actions
.size
;
1346 op
->dop
.u
.execute
.needs_help
= (upcall
->xout
.slow
& SLOW_ACTION
) != 0;
1347 op
->dop
.u
.execute
.probe
= false;
1348 op
->dop
.u
.execute
.mtu
= upcall
->mru
;
1354 * We install ukeys before installing the flows, locking them for exclusive
1355 * access by this thread for the period of installation. This ensures that
1356 * other threads won't attempt to delete the flows as we are creating them.
1359 for (i
= 0; i
< n_ops
; i
++) {
1360 struct udpif_key
*ukey
= ops
[i
].ukey
;
1363 /* If we can't install the ukey, don't install the flow. */
1364 if (!ukey_install_start(udpif
, ukey
)) {
1365 ukey_delete__(ukey
);
1370 opsp
[n_opsp
++] = &ops
[i
].dop
;
1372 dpif_operate(udpif
->dpif
, opsp
, n_opsp
);
1373 for (i
= 0; i
< n_ops
; i
++) {
1375 ukey_install_finish(ops
[i
].ukey
, ops
[i
].dop
.error
);
1381 get_ukey_hash(const ovs_u128
*ufid
, const unsigned pmd_id
)
1383 return hash_2words(ufid
->u32
[0], pmd_id
);
1386 static struct udpif_key
*
1387 ukey_lookup(struct udpif
*udpif
, const ovs_u128
*ufid
, const unsigned pmd_id
)
1389 struct udpif_key
*ukey
;
1390 int idx
= get_ukey_hash(ufid
, pmd_id
) % N_UMAPS
;
1391 struct cmap
*cmap
= &udpif
->ukeys
[idx
].cmap
;
1393 CMAP_FOR_EACH_WITH_HASH (ukey
, cmap_node
,
1394 get_ukey_hash(ufid
, pmd_id
), cmap
) {
1395 if (ovs_u128_equals(ukey
->ufid
, *ufid
)) {
1402 /* Provides safe lockless access of RCU protected 'ukey->actions'. Callers may
1403 * alternatively access the field directly if they take 'ukey->mutex'. */
1405 ukey_get_actions(struct udpif_key
*ukey
, const struct nlattr
**actions
, size_t *size
)
1407 const struct ofpbuf
*buf
= ovsrcu_get(struct ofpbuf
*, &ukey
->actions
);
1408 *actions
= buf
->data
;
1413 ukey_set_actions(struct udpif_key
*ukey
, const struct ofpbuf
*actions
)
1415 ovsrcu_postpone(ofpbuf_delete
,
1416 ovsrcu_get_protected(struct ofpbuf
*, &ukey
->actions
));
1417 ovsrcu_set(&ukey
->actions
, ofpbuf_clone(actions
));
1420 static struct udpif_key
*
1421 ukey_create__(const struct nlattr
*key
, size_t key_len
,
1422 const struct nlattr
*mask
, size_t mask_len
,
1423 bool ufid_present
, const ovs_u128
*ufid
,
1424 const unsigned pmd_id
, const struct ofpbuf
*actions
,
1425 uint64_t dump_seq
, uint64_t reval_seq
, long long int used
,
1426 uint32_t key_recirc_id
, struct xlate_out
*xout
)
1427 OVS_NO_THREAD_SAFETY_ANALYSIS
1429 struct udpif_key
*ukey
= xmalloc(sizeof *ukey
);
1431 memcpy(&ukey
->keybuf
, key
, key_len
);
1432 ukey
->key
= &ukey
->keybuf
.nla
;
1433 ukey
->key_len
= key_len
;
1434 memcpy(&ukey
->maskbuf
, mask
, mask_len
);
1435 ukey
->mask
= &ukey
->maskbuf
.nla
;
1436 ukey
->mask_len
= mask_len
;
1437 ukey
->ufid_present
= ufid_present
;
1439 ukey
->pmd_id
= pmd_id
;
1440 ukey
->hash
= get_ukey_hash(&ukey
->ufid
, pmd_id
);
1442 ovsrcu_init(&ukey
->actions
, NULL
);
1443 ukey_set_actions(ukey
, actions
);
1445 ovs_mutex_init(&ukey
->mutex
);
1446 ukey
->dump_seq
= dump_seq
;
1447 ukey
->reval_seq
= reval_seq
;
1448 ukey
->flow_exists
= false;
1449 ukey
->created
= time_msec();
1450 memset(&ukey
->stats
, 0, sizeof ukey
->stats
);
1451 ukey
->stats
.used
= used
;
1452 ukey
->xcache
= NULL
;
1454 ukey
->key_recirc_id
= key_recirc_id
;
1455 recirc_refs_init(&ukey
->recircs
);
1457 /* Take ownership of the action recirc id references. */
1458 recirc_refs_swap(&ukey
->recircs
, &xout
->recircs
);
1464 static struct udpif_key
*
1465 ukey_create_from_upcall(struct upcall
*upcall
, struct flow_wildcards
*wc
)
1467 struct odputil_keybuf keystub
, maskstub
;
1468 struct ofpbuf keybuf
, maskbuf
;
1470 struct odp_flow_key_parms odp_parms
= {
1471 .flow
= upcall
->flow
,
1475 odp_parms
.support
= ofproto_dpif_get_support(upcall
->ofproto
)->odp
;
1476 if (upcall
->key_len
) {
1477 ofpbuf_use_const(&keybuf
, upcall
->key
, upcall
->key_len
);
1479 /* dpif-netdev doesn't provide a netlink-formatted flow key in the
1480 * upcall, so convert the upcall's flow here. */
1481 ofpbuf_use_stack(&keybuf
, &keystub
, sizeof keystub
);
1482 odp_parms
.odp_in_port
= upcall
->flow
->in_port
.odp_port
;
1483 odp_flow_key_from_flow(&odp_parms
, &keybuf
);
1486 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1487 ofpbuf_use_stack(&maskbuf
, &maskstub
, sizeof maskstub
);
1489 odp_parms
.odp_in_port
= ODPP_NONE
;
1490 odp_parms
.key_buf
= &keybuf
;
1492 odp_flow_key_from_mask(&odp_parms
, &maskbuf
);
1495 return ukey_create__(keybuf
.data
, keybuf
.size
, maskbuf
.data
, maskbuf
.size
,
1496 true, upcall
->ufid
, upcall
->pmd_id
,
1497 &upcall
->put_actions
, upcall
->dump_seq
,
1498 upcall
->reval_seq
, 0,
1499 upcall
->have_recirc_ref
? upcall
->recirc
->id
: 0,
1504 ukey_create_from_dpif_flow(const struct udpif
*udpif
,
1505 const struct dpif_flow
*flow
,
1506 struct udpif_key
**ukey
)
1508 struct dpif_flow full_flow
;
1509 struct ofpbuf actions
;
1510 uint64_t dump_seq
, reval_seq
;
1511 uint64_t stub
[DPIF_FLOW_BUFSIZE
/ 8];
1512 const struct nlattr
*a
;
1515 if (!flow
->key_len
|| !flow
->actions_len
) {
1519 /* If the key or actions were not provided by the datapath, fetch the
1521 ofpbuf_use_stack(&buf
, &stub
, sizeof stub
);
1522 err
= dpif_flow_get(udpif
->dpif
, flow
->key
, flow
->key_len
,
1523 flow
->ufid_present
? &flow
->ufid
: NULL
,
1524 flow
->pmd_id
, &buf
, &full_flow
);
1531 /* Check the flow actions for recirculation action. As recirculation
1532 * relies on OVS userspace internal state, we need to delete all old
1533 * datapath flows with either a non-zero recirc_id in the key, or any
1534 * recirculation actions upon OVS restart. */
1535 NL_ATTR_FOR_EACH_UNSAFE (a
, left
, flow
->key
, flow
->key_len
) {
1536 if (nl_attr_type(a
) == OVS_KEY_ATTR_RECIRC_ID
1537 && nl_attr_get_u32(a
) != 0) {
1541 NL_ATTR_FOR_EACH_UNSAFE (a
, left
, flow
->actions
, flow
->actions_len
) {
1542 if (nl_attr_type(a
) == OVS_ACTION_ATTR_RECIRC
) {
1547 dump_seq
= seq_read(udpif
->dump_seq
);
1548 reval_seq
= seq_read(udpif
->reval_seq
);
1549 ofpbuf_use_const(&actions
, &flow
->actions
, flow
->actions_len
);
1550 *ukey
= ukey_create__(flow
->key
, flow
->key_len
,
1551 flow
->mask
, flow
->mask_len
, flow
->ufid_present
,
1552 &flow
->ufid
, flow
->pmd_id
, &actions
, dump_seq
,
1553 reval_seq
, flow
->stats
.used
, 0, NULL
);
1558 /* Attempts to insert a ukey into the shared ukey maps.
1560 * On success, returns true, installs the ukey and returns it in a locked
1561 * state. Otherwise, returns false. */
1563 ukey_install_start(struct udpif
*udpif
, struct udpif_key
*new_ukey
)
1564 OVS_TRY_LOCK(true, new_ukey
->mutex
)
1567 struct udpif_key
*old_ukey
;
1569 bool locked
= false;
1571 idx
= new_ukey
->hash
% N_UMAPS
;
1572 umap
= &udpif
->ukeys
[idx
];
1573 ovs_mutex_lock(&umap
->mutex
);
1574 old_ukey
= ukey_lookup(udpif
, &new_ukey
->ufid
, new_ukey
->pmd_id
);
1576 /* Uncommon case: A ukey is already installed with the same UFID. */
1577 if (old_ukey
->key_len
== new_ukey
->key_len
1578 && !memcmp(old_ukey
->key
, new_ukey
->key
, new_ukey
->key_len
)) {
1579 COVERAGE_INC(handler_duplicate_upcall
);
1581 struct ds ds
= DS_EMPTY_INITIALIZER
;
1583 odp_format_ufid(&old_ukey
->ufid
, &ds
);
1584 ds_put_cstr(&ds
, " ");
1585 odp_flow_key_format(old_ukey
->key
, old_ukey
->key_len
, &ds
);
1586 ds_put_cstr(&ds
, "\n");
1587 odp_format_ufid(&new_ukey
->ufid
, &ds
);
1588 ds_put_cstr(&ds
, " ");
1589 odp_flow_key_format(new_ukey
->key
, new_ukey
->key_len
, &ds
);
1591 VLOG_WARN_RL(&rl
, "Conflicting ukey for flows:\n%s", ds_cstr(&ds
));
1595 ovs_mutex_lock(&new_ukey
->mutex
);
1596 cmap_insert(&umap
->cmap
, &new_ukey
->cmap_node
, new_ukey
->hash
);
1599 ovs_mutex_unlock(&umap
->mutex
);
1605 ukey_install_finish__(struct udpif_key
*ukey
) OVS_REQUIRES(ukey
->mutex
)
1607 ukey
->flow_exists
= true;
1611 ukey_install_finish(struct udpif_key
*ukey
, int error
)
1612 OVS_RELEASES(ukey
->mutex
)
1615 ukey_install_finish__(ukey
);
1617 ovs_mutex_unlock(&ukey
->mutex
);
1623 ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
)
1625 /* The usual way to keep 'ukey->flow_exists' in sync with the datapath is
1626 * to call ukey_install_start(), install the corresponding datapath flow,
1627 * then call ukey_install_finish(). The netdev interface using upcall_cb()
1628 * doesn't provide a function to separately finish the flow installation,
1629 * so we perform the operations together here.
1631 * This is fine currently, as revalidator threads will only delete this
1632 * ukey during revalidator_sweep() and only if the dump_seq is mismatched.
1633 * It is unlikely for a revalidator thread to advance dump_seq and reach
1634 * the next GC phase between ukey creation and flow installation. */
1635 return ukey_install_start(udpif
, ukey
) && ukey_install_finish(ukey
, 0);
1638 /* Searches for a ukey in 'udpif->ukeys' that matches 'flow' and attempts to
1639 * lock the ukey. If the ukey does not exist, create it.
1641 * Returns 0 on success, setting *result to the matching ukey and returning it
1642 * in a locked state. Otherwise, returns an errno and clears *result. EBUSY
1643 * indicates that another thread is handling this flow. Other errors indicate
1644 * an unexpected condition creating a new ukey.
1646 * *error is an output parameter provided to appease the threadsafety analyser,
1647 * and its value matches the return value. */
1649 ukey_acquire(struct udpif
*udpif
, const struct dpif_flow
*flow
,
1650 struct udpif_key
**result
, int *error
)
1651 OVS_TRY_LOCK(0, (*result
)->mutex
)
1653 struct udpif_key
*ukey
;
1656 ukey
= ukey_lookup(udpif
, &flow
->ufid
, flow
->pmd_id
);
1658 retval
= ovs_mutex_trylock(&ukey
->mutex
);
1660 /* Usually we try to avoid installing flows from revalidator threads,
1661 * because locking on a umap may cause handler threads to block.
1662 * However there are certain cases, like when ovs-vswitchd is
1663 * restarted, where it is desirable to handle flows that exist in the
1664 * datapath gracefully (ie, don't just clear the datapath). */
1667 retval
= ukey_create_from_dpif_flow(udpif
, flow
, &ukey
);
1671 install
= ukey_install_start(udpif
, ukey
);
1673 ukey_install_finish__(ukey
);
1676 ukey_delete__(ukey
);
1692 ukey_delete__(struct udpif_key
*ukey
)
1693 OVS_NO_THREAD_SAFETY_ANALYSIS
1696 if (ukey
->key_recirc_id
) {
1697 recirc_free_id(ukey
->key_recirc_id
);
1699 recirc_refs_unref(&ukey
->recircs
);
1700 xlate_cache_delete(ukey
->xcache
);
1701 ofpbuf_delete(ovsrcu_get(struct ofpbuf
*, &ukey
->actions
));
1702 ovs_mutex_destroy(&ukey
->mutex
);
1708 ukey_delete(struct umap
*umap
, struct udpif_key
*ukey
)
1709 OVS_REQUIRES(umap
->mutex
)
1711 cmap_remove(&umap
->cmap
, &ukey
->cmap_node
, ukey
->hash
);
1712 ovsrcu_postpone(ukey_delete__
, ukey
);
1716 should_revalidate(const struct udpif
*udpif
, uint64_t packets
,
1719 long long int metric
, now
, duration
;
1721 if (udpif
->dump_duration
< 200) {
1722 /* We are likely to handle full revalidation for the flows. */
1726 /* Calculate the mean time between seeing these packets. If this
1727 * exceeds the threshold, then delete the flow rather than performing
1728 * costly revalidation for flows that aren't being hit frequently.
1730 * This is targeted at situations where the dump_duration is high (~1s),
1731 * and revalidation is triggered by a call to udpif_revalidate(). In
1732 * these situations, revalidation of all flows causes fluctuations in the
1733 * flow_limit due to the interaction with the dump_duration and max_idle.
1734 * This tends to result in deletion of low-throughput flows anyway, so
1735 * skip the revalidation and just delete those flows. */
1736 packets
= MAX(packets
, 1);
1737 now
= MAX(used
, time_msec());
1738 duration
= now
- used
;
1739 metric
= duration
/ packets
;
1742 /* The flow is receiving more than ~5pps, so keep it. */
1748 /* Verifies that the datapath actions of 'ukey' are still correct, and pushes
1751 * Returns a recommended action for 'ukey', options include:
1752 * UKEY_DELETE The ukey should be deleted.
1753 * UKEY_KEEP The ukey is fine as is.
1754 * UKEY_MODIFY The ukey's actions should be changed but is otherwise
1755 * fine. Callers should change the actions to those found
1756 * in the caller supplied 'odp_actions' buffer. The
1757 * recirculation references can be found in 'recircs' and
1758 * must be handled by the caller.
1760 * If the result is UKEY_MODIFY, then references to all recirc_ids used by the
1761 * new flow will be held within 'recircs' (which may be none).
1763 * The caller is responsible for both initializing 'recircs' prior this call,
1764 * and ensuring any references are eventually freed.
1766 static enum reval_result
1767 revalidate_ukey(struct udpif
*udpif
, struct udpif_key
*ukey
,
1768 const struct dpif_flow_stats
*stats
,
1769 struct ofpbuf
*odp_actions
, uint64_t reval_seq
,
1770 struct recirc_refs
*recircs
)
1771 OVS_REQUIRES(ukey
->mutex
)
1773 struct xlate_out xout
, *xoutp
;
1774 struct netflow
*netflow
;
1775 struct ofproto_dpif
*ofproto
;
1776 struct dpif_flow_stats push
;
1778 struct flow_wildcards dp_mask
, wc
;
1779 enum reval_result result
;
1780 ofp_port_t ofp_in_port
;
1781 struct xlate_in xin
;
1782 long long int last_used
;
1784 bool need_revalidate
;
1786 result
= UKEY_DELETE
;
1790 ofpbuf_clear(odp_actions
);
1791 need_revalidate
= (ukey
->reval_seq
!= reval_seq
);
1792 last_used
= ukey
->stats
.used
;
1793 push
.used
= stats
->used
;
1794 push
.tcp_flags
= stats
->tcp_flags
;
1795 push
.n_packets
= (stats
->n_packets
> ukey
->stats
.n_packets
1796 ? stats
->n_packets
- ukey
->stats
.n_packets
1798 push
.n_bytes
= (stats
->n_bytes
> ukey
->stats
.n_bytes
1799 ? stats
->n_bytes
- ukey
->stats
.n_bytes
1802 if (need_revalidate
&& last_used
1803 && !should_revalidate(udpif
, push
.n_packets
, last_used
)) {
1807 /* We will push the stats, so update the ukey stats cache. */
1808 ukey
->stats
= *stats
;
1809 if (!push
.n_packets
&& !need_revalidate
) {
1814 if (ukey
->xcache
&& !need_revalidate
) {
1815 xlate_push_stats(ukey
->xcache
, &push
);
1820 if (odp_flow_key_to_flow(ukey
->key
, ukey
->key_len
, &flow
)
1825 error
= xlate_lookup(udpif
->backer
, &flow
, &ofproto
, NULL
, NULL
, &netflow
,
1831 if (need_revalidate
) {
1832 xlate_cache_clear(ukey
->xcache
);
1834 if (!ukey
->xcache
) {
1835 ukey
->xcache
= xlate_cache_new();
1838 xlate_in_init(&xin
, ofproto
, &flow
, ofp_in_port
, NULL
, push
.tcp_flags
,
1839 NULL
, need_revalidate
? &wc
: NULL
, odp_actions
);
1840 if (push
.n_packets
) {
1841 xin
.resubmit_stats
= &push
;
1842 xin
.may_learn
= true;
1844 xin
.xcache
= ukey
->xcache
;
1845 xlate_actions(&xin
, &xout
);
1848 if (!need_revalidate
) {
1854 ofpbuf_clear(odp_actions
);
1855 compose_slow_path(udpif
, &xout
, &flow
, flow
.in_port
.odp_port
,
1859 if (odp_flow_key_to_mask(ukey
->mask
, ukey
->mask_len
, ukey
->key
,
1860 ukey
->key_len
, &dp_mask
, &flow
)
1865 /* Do not modify if any bit is wildcarded by the installed datapath flow,
1866 * but not the newly revalidated wildcard mask (wc), i.e., if revalidation
1867 * tells that the datapath flow is now too generic and must be narrowed
1868 * down. Note that we do not know if the datapath has ignored any of the
1869 * wildcarded bits, so we may be overtly conservative here. */
1870 if (flow_wildcards_has_extra(&dp_mask
, &wc
)) {
1874 if (!ofpbuf_equal(odp_actions
,
1875 ovsrcu_get(struct ofpbuf
*, &ukey
->actions
))) {
1876 /* The datapath mask was OK, but the actions seem to have changed.
1877 * Let's modify it in place. */
1878 result
= UKEY_MODIFY
;
1879 /* Transfer recirc action ID references to the caller. */
1880 recirc_refs_swap(recircs
, &xoutp
->recircs
);
1887 if (result
!= UKEY_DELETE
) {
1888 ukey
->reval_seq
= reval_seq
;
1890 if (netflow
&& result
== UKEY_DELETE
) {
1891 netflow_flow_clear(netflow
, &flow
);
1893 xlate_out_uninit(xoutp
);
1898 delete_op_init__(struct udpif
*udpif
, struct ukey_op
*op
,
1899 const struct dpif_flow
*flow
)
1902 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
1903 op
->dop
.u
.flow_del
.key
= flow
->key
;
1904 op
->dop
.u
.flow_del
.key_len
= flow
->key_len
;
1905 op
->dop
.u
.flow_del
.ufid
= flow
->ufid_present
? &flow
->ufid
: NULL
;
1906 op
->dop
.u
.flow_del
.pmd_id
= flow
->pmd_id
;
1907 op
->dop
.u
.flow_del
.stats
= &op
->stats
;
1908 op
->dop
.u
.flow_del
.terse
= udpif_use_ufid(udpif
);
1912 delete_op_init(struct udpif
*udpif
, struct ukey_op
*op
, struct udpif_key
*ukey
)
1915 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
1916 op
->dop
.u
.flow_del
.key
= ukey
->key
;
1917 op
->dop
.u
.flow_del
.key_len
= ukey
->key_len
;
1918 op
->dop
.u
.flow_del
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
1919 op
->dop
.u
.flow_del
.pmd_id
= ukey
->pmd_id
;
1920 op
->dop
.u
.flow_del
.stats
= &op
->stats
;
1921 op
->dop
.u
.flow_del
.terse
= udpif_use_ufid(udpif
);
1925 modify_op_init(struct ukey_op
*op
, struct udpif_key
*ukey
)
1928 op
->dop
.type
= DPIF_OP_FLOW_PUT
;
1929 op
->dop
.u
.flow_put
.flags
= DPIF_FP_MODIFY
;
1930 op
->dop
.u
.flow_put
.key
= ukey
->key
;
1931 op
->dop
.u
.flow_put
.key_len
= ukey
->key_len
;
1932 op
->dop
.u
.flow_put
.mask
= ukey
->mask
;
1933 op
->dop
.u
.flow_put
.mask_len
= ukey
->mask_len
;
1934 op
->dop
.u
.flow_put
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
1935 op
->dop
.u
.flow_put
.pmd_id
= ukey
->pmd_id
;
1936 op
->dop
.u
.flow_put
.stats
= NULL
;
1937 ukey_get_actions(ukey
, &op
->dop
.u
.flow_put
.actions
,
1938 &op
->dop
.u
.flow_put
.actions_len
);
1941 /* Executes datapath operations 'ops' and attributes stats retrieved from the
1942 * datapath as part of those operations. */
1944 push_dp_ops(struct udpif
*udpif
, struct ukey_op
*ops
, size_t n_ops
)
1946 struct dpif_op
*opsp
[REVALIDATE_MAX_BATCH
];
1949 ovs_assert(n_ops
<= REVALIDATE_MAX_BATCH
);
1950 for (i
= 0; i
< n_ops
; i
++) {
1951 opsp
[i
] = &ops
[i
].dop
;
1953 dpif_operate(udpif
->dpif
, opsp
, n_ops
);
1955 for (i
= 0; i
< n_ops
; i
++) {
1956 struct ukey_op
*op
= &ops
[i
];
1957 struct dpif_flow_stats
*push
, *stats
, push_buf
;
1959 stats
= op
->dop
.u
.flow_del
.stats
;
1962 if (op
->dop
.type
!= DPIF_OP_FLOW_DEL
) {
1963 /* Only deleted flows need their stats pushed. */
1967 if (op
->dop
.error
) {
1968 /* flow_del error, 'stats' is unusable. */
1973 ovs_mutex_lock(&op
->ukey
->mutex
);
1974 push
->used
= MAX(stats
->used
, op
->ukey
->stats
.used
);
1975 push
->tcp_flags
= stats
->tcp_flags
| op
->ukey
->stats
.tcp_flags
;
1976 push
->n_packets
= stats
->n_packets
- op
->ukey
->stats
.n_packets
;
1977 push
->n_bytes
= stats
->n_bytes
- op
->ukey
->stats
.n_bytes
;
1978 ovs_mutex_unlock(&op
->ukey
->mutex
);
1983 if (push
->n_packets
|| netflow_exists()) {
1984 const struct nlattr
*key
= op
->dop
.u
.flow_del
.key
;
1985 size_t key_len
= op
->dop
.u
.flow_del
.key_len
;
1986 struct ofproto_dpif
*ofproto
;
1987 struct netflow
*netflow
;
1988 ofp_port_t ofp_in_port
;
1993 ovs_mutex_lock(&op
->ukey
->mutex
);
1994 if (op
->ukey
->xcache
) {
1995 xlate_push_stats(op
->ukey
->xcache
, push
);
1996 ovs_mutex_unlock(&op
->ukey
->mutex
);
1999 ovs_mutex_unlock(&op
->ukey
->mutex
);
2000 key
= op
->ukey
->key
;
2001 key_len
= op
->ukey
->key_len
;
2004 if (odp_flow_key_to_flow(key
, key_len
, &flow
)
2009 error
= xlate_lookup(udpif
->backer
, &flow
, &ofproto
, NULL
, NULL
,
2010 &netflow
, &ofp_in_port
);
2012 struct xlate_in xin
;
2014 xlate_in_init(&xin
, ofproto
, &flow
, ofp_in_port
, NULL
,
2015 push
->tcp_flags
, NULL
, NULL
, NULL
);
2016 xin
.resubmit_stats
= push
->n_packets
? push
: NULL
;
2017 xin
.may_learn
= push
->n_packets
> 0;
2018 xlate_actions_for_side_effects(&xin
);
2021 netflow_flow_clear(netflow
, &flow
);
2028 /* Executes datapath operations 'ops', attributes stats retrieved from the
2029 * datapath, and deletes ukeys corresponding to deleted flows. */
2031 push_ukey_ops(struct udpif
*udpif
, struct umap
*umap
,
2032 struct ukey_op
*ops
, size_t n_ops
)
2036 push_dp_ops(udpif
, ops
, n_ops
);
2037 ovs_mutex_lock(&umap
->mutex
);
2038 for (i
= 0; i
< n_ops
; i
++) {
2039 if (ops
[i
].dop
.type
== DPIF_OP_FLOW_DEL
) {
2040 ukey_delete(umap
, ops
[i
].ukey
);
2043 ovs_mutex_unlock(&umap
->mutex
);
2047 log_unexpected_flow(const struct dpif_flow
*flow
, int error
)
2049 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(10, 60);
2050 struct ds ds
= DS_EMPTY_INITIALIZER
;
2052 ds_put_format(&ds
, "Failed to acquire udpif_key corresponding to "
2053 "unexpected flow (%s): ", ovs_strerror(error
));
2054 odp_format_ufid(&flow
->ufid
, &ds
);
2055 VLOG_WARN_RL(&rl
, "%s", ds_cstr(&ds
));
2059 reval_op_init(struct ukey_op
*op
, enum reval_result result
,
2060 struct udpif
*udpif
, struct udpif_key
*ukey
,
2061 struct recirc_refs
*recircs
, struct ofpbuf
*odp_actions
)
2063 if (result
== UKEY_DELETE
) {
2064 delete_op_init(udpif
, op
, ukey
);
2065 } else if (result
== UKEY_MODIFY
) {
2066 /* Store the new recircs. */
2067 recirc_refs_swap(&ukey
->recircs
, recircs
);
2068 /* Release old recircs. */
2069 recirc_refs_unref(recircs
);
2070 /* ukey->key_recirc_id remains, as the key is the same as before. */
2072 ukey_set_actions(ukey
, odp_actions
);
2073 modify_op_init(op
, ukey
);
2078 revalidate(struct revalidator
*revalidator
)
2080 uint64_t odp_actions_stub
[1024 / 8];
2081 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2083 struct udpif
*udpif
= revalidator
->udpif
;
2084 struct dpif_flow_dump_thread
*dump_thread
;
2085 uint64_t dump_seq
, reval_seq
;
2086 unsigned int flow_limit
;
2088 dump_seq
= seq_read(udpif
->dump_seq
);
2089 reval_seq
= seq_read(udpif
->reval_seq
);
2090 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2091 dump_thread
= dpif_flow_dump_thread_create(udpif
->dump
);
2093 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2096 struct dpif_flow flows
[REVALIDATE_MAX_BATCH
];
2097 const struct dpif_flow
*f
;
2100 long long int max_idle
;
2105 n_dumped
= dpif_flow_dump_next(dump_thread
, flows
, ARRAY_SIZE(flows
));
2112 /* In normal operation we want to keep flows around until they have
2113 * been idle for 'ofproto_max_idle' milliseconds. However:
2115 * - If the number of datapath flows climbs above 'flow_limit',
2116 * drop that down to 100 ms to try to bring the flows down to
2119 * - If the number of datapath flows climbs above twice
2120 * 'flow_limit', delete all the datapath flows as an emergency
2121 * measure. (We reassess this condition for the next batch of
2122 * datapath flows, so we will recover before all the flows are
2124 n_dp_flows
= udpif_get_n_flows(udpif
);
2125 kill_them_all
= n_dp_flows
> flow_limit
* 2;
2126 max_idle
= n_dp_flows
> flow_limit
? 100 : ofproto_max_idle
;
2128 for (f
= flows
; f
< &flows
[n_dumped
]; f
++) {
2129 long long int used
= f
->stats
.used
;
2130 struct recirc_refs recircs
= RECIRC_REFS_EMPTY_INITIALIZER
;
2131 enum reval_result result
;
2132 struct udpif_key
*ukey
;
2133 bool already_dumped
;
2136 if (ukey_acquire(udpif
, f
, &ukey
, &error
)) {
2137 if (error
== EBUSY
) {
2138 /* Another thread is processing this flow, so don't bother
2140 COVERAGE_INC(upcall_ukey_contention
);
2142 log_unexpected_flow(f
, error
);
2143 if (error
!= ENOENT
) {
2144 delete_op_init__(udpif
, &ops
[n_ops
++], f
);
2150 already_dumped
= ukey
->dump_seq
== dump_seq
;
2151 if (already_dumped
) {
2152 /* The flow has already been handled during this flow dump
2153 * operation. Skip it. */
2155 COVERAGE_INC(dumped_duplicate_flow
);
2157 COVERAGE_INC(dumped_new_flow
);
2159 ovs_mutex_unlock(&ukey
->mutex
);
2164 used
= ukey
->created
;
2166 if (kill_them_all
|| (used
&& used
< now
- max_idle
)) {
2167 result
= UKEY_DELETE
;
2169 result
= revalidate_ukey(udpif
, ukey
, &f
->stats
, &odp_actions
,
2170 reval_seq
, &recircs
);
2172 ukey
->dump_seq
= dump_seq
;
2173 ukey
->flow_exists
= result
!= UKEY_DELETE
;
2175 if (result
!= UKEY_KEEP
) {
2176 /* Takes ownership of 'recircs'. */
2177 reval_op_init(&ops
[n_ops
++], result
, udpif
, ukey
, &recircs
,
2180 ovs_mutex_unlock(&ukey
->mutex
);
2184 /* Push datapath ops but defer ukey deletion to 'sweep' phase. */
2185 push_dp_ops(udpif
, ops
, n_ops
);
2189 dpif_flow_dump_thread_destroy(dump_thread
);
2190 ofpbuf_uninit(&odp_actions
);
2193 /* Pauses the 'revalidator', can only proceed after main thread
2194 * calls udpif_resume_revalidators(). */
2196 revalidator_pause(struct revalidator
*revalidator
)
2198 /* The first block is for sync'ing the pause with main thread. */
2199 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2200 /* The second block is for pausing until main thread resumes. */
2201 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2205 revalidator_sweep__(struct revalidator
*revalidator
, bool purge
)
2207 struct udpif
*udpif
;
2208 uint64_t dump_seq
, reval_seq
;
2211 udpif
= revalidator
->udpif
;
2212 dump_seq
= seq_read(udpif
->dump_seq
);
2213 reval_seq
= seq_read(udpif
->reval_seq
);
2214 slice
= revalidator
- udpif
->revalidators
;
2215 ovs_assert(slice
< udpif
->n_revalidators
);
2217 for (int i
= slice
; i
< N_UMAPS
; i
+= udpif
->n_revalidators
) {
2218 uint64_t odp_actions_stub
[1024 / 8];
2219 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2221 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2222 struct udpif_key
*ukey
;
2223 struct umap
*umap
= &udpif
->ukeys
[i
];
2226 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2229 /* Handler threads could be holding a ukey lock while it installs a
2230 * new flow, so don't hang around waiting for access to it. */
2231 if (ovs_mutex_trylock(&ukey
->mutex
)) {
2234 flow_exists
= ukey
->flow_exists
;
2236 struct recirc_refs recircs
= RECIRC_REFS_EMPTY_INITIALIZER
;
2237 bool seq_mismatch
= (ukey
->dump_seq
!= dump_seq
2238 && ukey
->reval_seq
!= reval_seq
);
2239 enum reval_result result
;
2242 result
= UKEY_DELETE
;
2243 } else if (!seq_mismatch
) {
2246 struct dpif_flow_stats stats
;
2247 COVERAGE_INC(revalidate_missed_dp_flow
);
2248 memset(&stats
, 0, sizeof stats
);
2249 result
= revalidate_ukey(udpif
, ukey
, &stats
, &odp_actions
,
2250 reval_seq
, &recircs
);
2252 if (result
!= UKEY_KEEP
) {
2253 /* Clears 'recircs' if filled by revalidate_ukey(). */
2254 reval_op_init(&ops
[n_ops
++], result
, udpif
, ukey
, &recircs
,
2258 ovs_mutex_unlock(&ukey
->mutex
);
2261 /* The common flow deletion case involves deletion of the flow
2262 * during the dump phase and ukey deletion here. */
2263 ovs_mutex_lock(&umap
->mutex
);
2264 ukey_delete(umap
, ukey
);
2265 ovs_mutex_unlock(&umap
->mutex
);
2268 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2269 /* Update/delete missed flows and clean up corresponding ukeys
2271 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2277 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2280 ofpbuf_uninit(&odp_actions
);
2286 revalidator_sweep(struct revalidator
*revalidator
)
2288 revalidator_sweep__(revalidator
, false);
2292 revalidator_purge(struct revalidator
*revalidator
)
2294 revalidator_sweep__(revalidator
, true);
2297 /* In reaction to dpif purge, purges all 'ukey's with same 'pmd_id'. */
2299 dp_purge_cb(void *aux
, unsigned pmd_id
)
2301 struct udpif
*udpif
= aux
;
2304 udpif_pause_revalidators(udpif
);
2305 for (i
= 0; i
< N_UMAPS
; i
++) {
2306 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2307 struct udpif_key
*ukey
;
2308 struct umap
*umap
= &udpif
->ukeys
[i
];
2311 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2312 if (ukey
->pmd_id
== pmd_id
) {
2313 delete_op_init(udpif
, &ops
[n_ops
++], ukey
);
2314 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2315 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2322 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2327 udpif_resume_revalidators(udpif
);
2331 upcall_unixctl_show(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2332 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2334 struct ds ds
= DS_EMPTY_INITIALIZER
;
2335 struct udpif
*udpif
;
2337 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2338 unsigned int flow_limit
;
2342 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2343 ufid_enabled
= udpif_use_ufid(udpif
);
2345 ds_put_format(&ds
, "%s:\n", dpif_name(udpif
->dpif
));
2346 ds_put_format(&ds
, "\tflows : (current %lu)"
2347 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif
),
2348 udpif
->avg_n_flows
, udpif
->max_n_flows
, flow_limit
);
2349 ds_put_format(&ds
, "\tdump duration : %lldms\n", udpif
->dump_duration
);
2350 ds_put_format(&ds
, "\tufid enabled : ");
2352 ds_put_format(&ds
, "true\n");
2354 ds_put_format(&ds
, "false\n");
2356 ds_put_char(&ds
, '\n');
2358 for (i
= 0; i
< n_revalidators
; i
++) {
2359 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
2360 int j
, elements
= 0;
2362 for (j
= i
; j
< N_UMAPS
; j
+= n_revalidators
) {
2363 elements
+= cmap_count(&udpif
->ukeys
[j
].cmap
);
2365 ds_put_format(&ds
, "\t%u: (keys %d)\n", revalidator
->id
, elements
);
2369 unixctl_command_reply(conn
, ds_cstr(&ds
));
2373 /* Disable using the megaflows.
2375 * This command is only needed for advanced debugging, so it's not
2376 * documented in the man page. */
2378 upcall_unixctl_disable_megaflows(struct unixctl_conn
*conn
,
2379 int argc OVS_UNUSED
,
2380 const char *argv
[] OVS_UNUSED
,
2381 void *aux OVS_UNUSED
)
2383 atomic_store_relaxed(&enable_megaflows
, false);
2384 udpif_flush_all_datapaths();
2385 unixctl_command_reply(conn
, "megaflows disabled");
2388 /* Re-enable using megaflows.
2390 * This command is only needed for advanced debugging, so it's not
2391 * documented in the man page. */
2393 upcall_unixctl_enable_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
, true);
2399 udpif_flush_all_datapaths();
2400 unixctl_command_reply(conn
, "megaflows enabled");
2403 /* Disable skipping flow attributes during flow dump.
2405 * This command is only needed for advanced debugging, so it's not
2406 * documented in the man page. */
2408 upcall_unixctl_disable_ufid(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2409 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2411 atomic_store_relaxed(&enable_ufid
, false);
2412 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID disabled");
2415 /* Re-enable skipping flow attributes during flow dump.
2417 * This command is only needed for advanced debugging, so it's not documented
2418 * in the man page. */
2420 upcall_unixctl_enable_ufid(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2421 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2423 atomic_store_relaxed(&enable_ufid
, true);
2424 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID enabled "
2425 "for supported datapaths");
2428 /* Set the flow limit.
2430 * This command is only needed for advanced debugging, so it's not
2431 * documented in the man page. */
2433 upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
,
2434 int argc OVS_UNUSED
,
2435 const char *argv
[] OVS_UNUSED
,
2436 void *aux OVS_UNUSED
)
2438 struct ds ds
= DS_EMPTY_INITIALIZER
;
2439 struct udpif
*udpif
;
2440 unsigned int flow_limit
= atoi(argv
[1]);
2442 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2443 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
2445 ds_put_format(&ds
, "set flow_limit to %u\n", flow_limit
);
2446 unixctl_command_reply(conn
, ds_cstr(&ds
));
2451 upcall_unixctl_dump_wait(struct unixctl_conn
*conn
,
2452 int argc OVS_UNUSED
,
2453 const char *argv
[] OVS_UNUSED
,
2454 void *aux OVS_UNUSED
)
2456 if (ovs_list_is_singleton(&all_udpifs
)) {
2457 struct udpif
*udpif
= NULL
;
2460 udpif
= OBJECT_CONTAINING(ovs_list_front(&all_udpifs
), udpif
, list_node
);
2461 len
= (udpif
->n_conns
+ 1) * sizeof *udpif
->conns
;
2462 udpif
->conn_seq
= seq_read(udpif
->dump_seq
);
2463 udpif
->conns
= xrealloc(udpif
->conns
, len
);
2464 udpif
->conns
[udpif
->n_conns
++] = conn
;
2466 unixctl_command_reply_error(conn
, "can't wait on multiple udpifs.");
2471 upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2472 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2474 struct udpif
*udpif
;
2476 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2479 for (n
= 0; n
< udpif
->n_revalidators
; n
++) {
2480 revalidator_purge(&udpif
->revalidators
[n
]);
2483 unixctl_command_reply(conn
, "");