1 /* Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2015, 2016 Nicira, Inc.
3 * Licensed under the Apache License, Version 2.0 (the "License");
4 * you may not use this file except in compliance with the License.
5 * You may obtain a copy of the License at:
7 * http://www.apache.org/licenses/LICENSE-2.0
9 * Unless required by applicable law or agreed to in writing, software
10 * distributed under the License is distributed on an "AS IS" BASIS,
11 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 * See the License for the specific language governing permissions and
13 * limitations under the License. */
16 #include "ofproto-dpif-upcall.h"
26 #include "openvswitch/dynamic-string.h"
27 #include "fail-open.h"
28 #include "guarded-list.h"
30 #include "openvswitch/list.h"
32 #include "openvswitch/ofpbuf.h"
33 #include "ofproto-dpif-ipfix.h"
34 #include "ofproto-dpif-sflow.h"
35 #include "ofproto-dpif-xlate.h"
38 #include "poll-loop.h"
41 #include "openvswitch/vlog.h"
43 #define MAX_QUEUE_LENGTH 512
44 #define UPCALL_MAX_BATCH 64
45 #define REVALIDATE_MAX_BATCH 50
47 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall
);
49 COVERAGE_DEFINE(dumped_duplicate_flow
);
50 COVERAGE_DEFINE(dumped_new_flow
);
51 COVERAGE_DEFINE(handler_duplicate_upcall
);
52 COVERAGE_DEFINE(upcall_ukey_contention
);
53 COVERAGE_DEFINE(revalidate_missed_dp_flow
);
55 /* A thread that reads upcalls from dpif, forwards each upcall's packet,
56 * and possibly sets up a kernel flow as a cache. */
58 struct udpif
*udpif
; /* Parent udpif. */
59 pthread_t thread
; /* Thread ID. */
60 uint32_t handler_id
; /* Handler id. */
63 /* In the absence of a multiple-writer multiple-reader datastructure for
64 * storing udpif_keys ("ukeys"), we use a large number of cmaps, each with its
65 * own lock for writing. */
66 #define N_UMAPS 512 /* per udpif. */
68 struct ovs_mutex mutex
; /* Take for writing to the following. */
69 struct cmap cmap
; /* Datapath flow keys. */
72 /* A thread that processes datapath flows, updates OpenFlow statistics, and
73 * updates or removes them if necessary.
75 * Revalidator threads operate in two phases: "dump" and "sweep". In between
76 * each phase, all revalidators sync up so that all revalidator threads are
77 * either in one phase or the other, but not a combination.
79 * During the dump phase, revalidators fetch flows from the datapath and
80 * attribute the statistics to OpenFlow rules. Each datapath flow has a
81 * corresponding ukey which caches the most recently seen statistics. If
82 * a flow needs to be deleted (for example, because it is unused over a
83 * period of time), revalidator threads may delete the flow during the
84 * dump phase. The datapath is not guaranteed to reliably dump all flows
85 * from the datapath, and there is no mapping between datapath flows to
86 * revalidators, so a particular flow may be handled by zero or more
87 * revalidators during a single dump phase. To avoid duplicate attribution
88 * of statistics, ukeys are never deleted during this phase.
90 * During the sweep phase, each revalidator takes ownership of a different
91 * slice of umaps and sweeps through all ukeys in those umaps to figure out
92 * whether they need to be deleted. During this phase, revalidators may
93 * fetch individual flows which were not dumped during the dump phase to
94 * validate them and attribute statistics.
97 struct udpif
*udpif
; /* Parent udpif. */
98 pthread_t thread
; /* Thread ID. */
99 unsigned int id
; /* ovsthread_id_self(). */
102 /* An upcall handler for ofproto_dpif.
104 * udpif keeps records of two kind of logically separate units:
109 * - An array of 'struct handler's for upcall handling and flow
115 * - Revalidation threads which read the datapath flow table and maintains
119 struct ovs_list list_node
; /* In all_udpifs list. */
121 struct dpif
*dpif
; /* Datapath handle. */
122 struct dpif_backer
*backer
; /* Opaque dpif_backer pointer. */
124 struct handler
*handlers
; /* Upcall handlers. */
127 struct revalidator
*revalidators
; /* Flow revalidators. */
128 size_t n_revalidators
;
130 struct latch exit_latch
; /* Tells child threads to exit. */
133 struct seq
*reval_seq
; /* Incremented to force revalidation. */
134 bool reval_exit
; /* Set by leader on 'exit_latch. */
135 struct ovs_barrier reval_barrier
; /* Barrier used by revalidators. */
136 struct dpif_flow_dump
*dump
; /* DPIF flow dump state. */
137 long long int dump_duration
; /* Duration of the last flow dump. */
138 struct seq
*dump_seq
; /* Increments each dump iteration. */
139 atomic_bool enable_ufid
; /* If true, skip dumping flow attrs. */
141 /* These variables provide a mechanism for the main thread to pause
142 * all revalidation without having to completely shut the threads down.
143 * 'pause_latch' is shared between the main thread and the lead
144 * revalidator thread, so when it is desirable to halt revalidation, the
145 * main thread will set the latch. 'pause' and 'pause_barrier' are shared
146 * by revalidator threads. The lead revalidator will set 'pause' when it
147 * observes the latch has been set, and this will cause all revalidator
148 * threads to wait on 'pause_barrier' at the beginning of the next
149 * revalidation round. */
150 bool pause
; /* Set by leader on 'pause_latch. */
151 struct latch pause_latch
; /* Set to force revalidators pause. */
152 struct ovs_barrier pause_barrier
; /* Barrier used to pause all */
153 /* revalidators by main thread. */
155 /* There are 'N_UMAPS' maps containing 'struct udpif_key' elements.
157 * During the flow dump phase, revalidators insert into these with a random
158 * distribution. During the garbage collection phase, each revalidator
159 * takes care of garbage collecting a slice of these maps. */
162 /* Datapath flow statistics. */
163 unsigned int max_n_flows
;
164 unsigned int avg_n_flows
;
166 /* Following fields are accessed and modified by different threads. */
167 atomic_uint flow_limit
; /* Datapath flow hard limit. */
169 /* n_flows_mutex prevents multiple threads updating these concurrently. */
170 atomic_uint n_flows
; /* Number of flows in the datapath. */
171 atomic_llong n_flows_timestamp
; /* Last time n_flows was updated. */
172 struct ovs_mutex n_flows_mutex
;
174 /* Following fields are accessed and modified only from the main thread. */
175 struct unixctl_conn
**conns
; /* Connections waiting on dump_seq. */
176 uint64_t conn_seq
; /* Corresponds to 'dump_seq' when
177 conns[n_conns-1] was stored. */
178 size_t n_conns
; /* Number of connections waiting. */
182 BAD_UPCALL
, /* Some kind of bug somewhere. */
183 MISS_UPCALL
, /* A flow miss. */
184 SFLOW_UPCALL
, /* sFlow sample. */
185 FLOW_SAMPLE_UPCALL
, /* Per-flow sampling. */
186 IPFIX_UPCALL
/* Per-bridge sampling. */
196 struct ofproto_dpif
*ofproto
; /* Parent ofproto. */
197 const struct recirc_id_node
*recirc
; /* Recirculation context. */
198 bool have_recirc_ref
; /* Reference held on recirc ctx? */
200 /* The flow and packet are only required to be constant when using
201 * dpif-netdev. If a modification is absolutely necessary, a const cast
202 * may be used with other datapaths. */
203 const struct flow
*flow
; /* Parsed representation of the packet. */
204 const ovs_u128
*ufid
; /* Unique identifier for 'flow'. */
205 unsigned pmd_id
; /* Datapath poll mode driver id. */
206 const struct dp_packet
*packet
; /* Packet associated with this upcall. */
207 ofp_port_t in_port
; /* OpenFlow in port, or OFPP_NONE. */
208 uint16_t mru
; /* If !0, Maximum receive unit of
209 fragmented IP packet */
211 enum dpif_upcall_type type
; /* Datapath type of the upcall. */
212 const struct nlattr
*userdata
; /* Userdata for DPIF_UC_ACTION Upcalls. */
213 const struct nlattr
*actions
; /* Flow actions in DPIF_UC_ACTION Upcalls. */
215 bool xout_initialized
; /* True if 'xout' must be uninitialized. */
216 struct xlate_out xout
; /* Result of xlate_actions(). */
217 struct ofpbuf odp_actions
; /* Datapath actions from xlate_actions(). */
218 struct flow_wildcards wc
; /* Dependencies that megaflow must match. */
219 struct ofpbuf put_actions
; /* Actions 'put' in the fastpath. */
221 struct dpif_ipfix
*ipfix
; /* IPFIX pointer or NULL. */
222 struct dpif_sflow
*sflow
; /* SFlow pointer or NULL. */
224 struct udpif_key
*ukey
; /* Revalidator flow cache. */
225 bool ukey_persists
; /* Set true to keep 'ukey' beyond the
226 lifetime of this upcall. */
228 uint64_t dump_seq
; /* udpif->dump_seq at translation time. */
229 uint64_t reval_seq
; /* udpif->reval_seq at translation time. */
231 /* Not used by the upcall callback interface. */
232 const struct nlattr
*key
; /* Datapath flow key. */
233 size_t key_len
; /* Datapath flow key length. */
234 const struct nlattr
*out_tun_key
; /* Datapath output tunnel key. */
236 uint64_t odp_actions_stub
[1024 / 8]; /* Stub for odp_actions. */
239 /* 'udpif_key's are responsible for tracking the little bit of state udpif
240 * needs to do flow expiration which can't be pulled directly from the
241 * datapath. They may be created by any handler or revalidator thread at any
242 * time, and read by any revalidator during the dump phase. They are however
243 * each owned by a single revalidator which takes care of destroying them
244 * during the garbage-collection phase.
246 * The mutex within the ukey protects some members of the ukey. The ukey
247 * itself is protected by RCU and is held within a umap in the parent udpif.
248 * Adding or removing a ukey from a umap is only safe when holding the
249 * corresponding umap lock. */
251 struct cmap_node cmap_node
; /* In parent revalidator 'ukeys' map. */
253 /* These elements are read only once created, and therefore aren't
254 * protected by a mutex. */
255 const struct nlattr
*key
; /* Datapath flow key. */
256 size_t key_len
; /* Length of 'key'. */
257 const struct nlattr
*mask
; /* Datapath flow mask. */
258 size_t mask_len
; /* Length of 'mask'. */
259 ovs_u128 ufid
; /* Unique flow identifier. */
260 bool ufid_present
; /* True if 'ufid' is in datapath. */
261 uint32_t hash
; /* Pre-computed hash for 'key'. */
262 unsigned pmd_id
; /* Datapath poll mode driver id. */
264 struct ovs_mutex mutex
; /* Guards the following. */
265 struct dpif_flow_stats stats OVS_GUARDED
; /* Last known stats.*/
266 long long int created OVS_GUARDED
; /* Estimate of creation time. */
267 uint64_t dump_seq OVS_GUARDED
; /* Tracks udpif->dump_seq. */
268 uint64_t reval_seq OVS_GUARDED
; /* Tracks udpif->reval_seq. */
269 bool flow_exists OVS_GUARDED
; /* Ensures flows are only deleted
271 /* Datapath flow actions as nlattrs. Protected by RCU. Read with
272 * ukey_get_actions(), and write with ukey_set_actions(). */
273 OVSRCU_TYPE(struct ofpbuf
*) actions
;
275 struct xlate_cache
*xcache OVS_GUARDED
; /* Cache for xlate entries that
276 * are affected by this ukey.
277 * Used for stats and learning.*/
279 struct odputil_keybuf buf
;
283 uint32_t key_recirc_id
; /* Non-zero if reference is held by the ukey. */
284 struct recirc_refs recircs
; /* Action recirc IDs with references held. */
287 /* Datapath operation with optional ukey attached. */
289 struct udpif_key
*ukey
;
290 struct dpif_flow_stats stats
; /* Stats for 'op'. */
291 struct dpif_op dop
; /* Flow operation. */
294 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
295 static struct ovs_list all_udpifs
= OVS_LIST_INITIALIZER(&all_udpifs
);
297 static size_t recv_upcalls(struct handler
*);
298 static int process_upcall(struct udpif
*, struct upcall
*,
299 struct ofpbuf
*odp_actions
, struct flow_wildcards
*);
300 static void handle_upcalls(struct udpif
*, struct upcall
*, size_t n_upcalls
);
301 static void udpif_stop_threads(struct udpif
*);
302 static void udpif_start_threads(struct udpif
*, size_t n_handlers
,
303 size_t n_revalidators
);
304 static void udpif_pause_revalidators(struct udpif
*);
305 static void udpif_resume_revalidators(struct udpif
*);
306 static void *udpif_upcall_handler(void *);
307 static void *udpif_revalidator(void *);
308 static unsigned long udpif_get_n_flows(struct udpif
*);
309 static void revalidate(struct revalidator
*);
310 static void revalidator_pause(struct revalidator
*);
311 static void revalidator_sweep(struct revalidator
*);
312 static void revalidator_purge(struct revalidator
*);
313 static void upcall_unixctl_show(struct unixctl_conn
*conn
, int argc
,
314 const char *argv
[], void *aux
);
315 static void upcall_unixctl_disable_megaflows(struct unixctl_conn
*, int argc
,
316 const char *argv
[], void *aux
);
317 static void upcall_unixctl_enable_megaflows(struct unixctl_conn
*, int argc
,
318 const char *argv
[], void *aux
);
319 static void upcall_unixctl_disable_ufid(struct unixctl_conn
*, int argc
,
320 const char *argv
[], void *aux
);
321 static void upcall_unixctl_enable_ufid(struct unixctl_conn
*, int argc
,
322 const char *argv
[], void *aux
);
323 static void upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
, int argc
,
324 const char *argv
[], void *aux
);
325 static void upcall_unixctl_dump_wait(struct unixctl_conn
*conn
, int argc
,
326 const char *argv
[], void *aux
);
327 static void upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc
,
328 const char *argv
[], void *aux
);
330 static struct udpif_key
*ukey_create_from_upcall(struct upcall
*,
331 struct flow_wildcards
*);
332 static int ukey_create_from_dpif_flow(const struct udpif
*,
333 const struct dpif_flow
*,
334 struct udpif_key
**);
335 static void ukey_get_actions(struct udpif_key
*, const struct nlattr
**actions
,
337 static bool ukey_install_start(struct udpif
*, struct udpif_key
*ukey
);
338 static bool ukey_install_finish(struct udpif_key
*ukey
, int error
);
339 static bool ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
);
340 static struct udpif_key
*ukey_lookup(struct udpif
*udpif
,
341 const ovs_u128
*ufid
,
342 const unsigned pmd_id
);
343 static int ukey_acquire(struct udpif
*, const struct dpif_flow
*,
344 struct udpif_key
**result
, int *error
);
345 static void ukey_delete__(struct udpif_key
*);
346 static void ukey_delete(struct umap
*, struct udpif_key
*);
347 static enum upcall_type
classify_upcall(enum dpif_upcall_type type
,
348 const struct nlattr
*userdata
);
350 static int upcall_receive(struct upcall
*, const struct dpif_backer
*,
351 const struct dp_packet
*packet
, enum dpif_upcall_type
,
352 const struct nlattr
*userdata
, const struct flow
*,
353 const unsigned int mru
,
354 const ovs_u128
*ufid
, const unsigned pmd_id
);
355 static void upcall_uninit(struct upcall
*);
357 static upcall_callback upcall_cb
;
358 static dp_purge_callback dp_purge_cb
;
360 static atomic_bool enable_megaflows
= ATOMIC_VAR_INIT(true);
361 static atomic_bool enable_ufid
= ATOMIC_VAR_INIT(true);
366 static struct ovsthread_once once
= OVSTHREAD_ONCE_INITIALIZER
;
367 if (ovsthread_once_start(&once
)) {
368 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show
,
370 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
371 upcall_unixctl_disable_megaflows
, NULL
);
372 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
373 upcall_unixctl_enable_megaflows
, NULL
);
374 unixctl_command_register("upcall/disable-ufid", "", 0, 0,
375 upcall_unixctl_disable_ufid
, NULL
);
376 unixctl_command_register("upcall/enable-ufid", "", 0, 0,
377 upcall_unixctl_enable_ufid
, NULL
);
378 unixctl_command_register("upcall/set-flow-limit", "", 1, 1,
379 upcall_unixctl_set_flow_limit
, NULL
);
380 unixctl_command_register("revalidator/wait", "", 0, 0,
381 upcall_unixctl_dump_wait
, NULL
);
382 unixctl_command_register("revalidator/purge", "", 0, 0,
383 upcall_unixctl_purge
, NULL
);
384 ovsthread_once_done(&once
);
389 udpif_create(struct dpif_backer
*backer
, struct dpif
*dpif
)
391 struct udpif
*udpif
= xzalloc(sizeof *udpif
);
394 udpif
->backer
= backer
;
395 atomic_init(&udpif
->flow_limit
, MIN(ofproto_flow_limit
, 10000));
396 udpif
->reval_seq
= seq_create();
397 udpif
->dump_seq
= seq_create();
398 latch_init(&udpif
->exit_latch
);
399 latch_init(&udpif
->pause_latch
);
400 ovs_list_push_back(&all_udpifs
, &udpif
->list_node
);
401 atomic_init(&udpif
->enable_ufid
, false);
402 atomic_init(&udpif
->n_flows
, 0);
403 atomic_init(&udpif
->n_flows_timestamp
, LLONG_MIN
);
404 ovs_mutex_init(&udpif
->n_flows_mutex
);
405 udpif
->ukeys
= xmalloc(N_UMAPS
* sizeof *udpif
->ukeys
);
406 for (int i
= 0; i
< N_UMAPS
; i
++) {
407 cmap_init(&udpif
->ukeys
[i
].cmap
);
408 ovs_mutex_init(&udpif
->ukeys
[i
].mutex
);
411 dpif_register_upcall_cb(dpif
, upcall_cb
, udpif
);
412 dpif_register_dp_purge_cb(dpif
, dp_purge_cb
, udpif
);
418 udpif_run(struct udpif
*udpif
)
420 if (udpif
->conns
&& udpif
->conn_seq
!= seq_read(udpif
->dump_seq
)) {
423 for (i
= 0; i
< udpif
->n_conns
; i
++) {
424 unixctl_command_reply(udpif
->conns
[i
], NULL
);
433 udpif_destroy(struct udpif
*udpif
)
435 udpif_stop_threads(udpif
);
437 dpif_register_dp_purge_cb(udpif
->dpif
, NULL
, udpif
);
438 dpif_register_upcall_cb(udpif
->dpif
, NULL
, udpif
);
440 for (int i
= 0; i
< N_UMAPS
; i
++) {
441 cmap_destroy(&udpif
->ukeys
[i
].cmap
);
442 ovs_mutex_destroy(&udpif
->ukeys
[i
].mutex
);
447 ovs_list_remove(&udpif
->list_node
);
448 latch_destroy(&udpif
->exit_latch
);
449 latch_destroy(&udpif
->pause_latch
);
450 seq_destroy(udpif
->reval_seq
);
451 seq_destroy(udpif
->dump_seq
);
452 ovs_mutex_destroy(&udpif
->n_flows_mutex
);
456 /* Stops the handler and revalidator threads, must be enclosed in
457 * ovsrcu quiescent state unless when destroying udpif. */
459 udpif_stop_threads(struct udpif
*udpif
)
461 if (udpif
&& (udpif
->n_handlers
!= 0 || udpif
->n_revalidators
!= 0)) {
464 latch_set(&udpif
->exit_latch
);
466 for (i
= 0; i
< udpif
->n_handlers
; i
++) {
467 struct handler
*handler
= &udpif
->handlers
[i
];
469 xpthread_join(handler
->thread
, NULL
);
472 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
473 xpthread_join(udpif
->revalidators
[i
].thread
, NULL
);
476 dpif_disable_upcall(udpif
->dpif
);
478 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
479 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
481 /* Delete ukeys, and delete all flows from the datapath to prevent
482 * double-counting stats. */
483 revalidator_purge(revalidator
);
486 latch_poll(&udpif
->exit_latch
);
488 ovs_barrier_destroy(&udpif
->reval_barrier
);
489 ovs_barrier_destroy(&udpif
->pause_barrier
);
491 free(udpif
->revalidators
);
492 udpif
->revalidators
= NULL
;
493 udpif
->n_revalidators
= 0;
495 free(udpif
->handlers
);
496 udpif
->handlers
= NULL
;
497 udpif
->n_handlers
= 0;
501 /* Starts the handler and revalidator threads, must be enclosed in
502 * ovsrcu quiescent state. */
504 udpif_start_threads(struct udpif
*udpif
, size_t n_handlers
,
505 size_t n_revalidators
)
507 if (udpif
&& n_handlers
&& n_revalidators
) {
511 udpif
->n_handlers
= n_handlers
;
512 udpif
->n_revalidators
= n_revalidators
;
514 udpif
->handlers
= xzalloc(udpif
->n_handlers
* sizeof *udpif
->handlers
);
515 for (i
= 0; i
< udpif
->n_handlers
; i
++) {
516 struct handler
*handler
= &udpif
->handlers
[i
];
518 handler
->udpif
= udpif
;
519 handler
->handler_id
= i
;
520 handler
->thread
= ovs_thread_create(
521 "handler", udpif_upcall_handler
, handler
);
524 enable_ufid
= ofproto_dpif_get_enable_ufid(udpif
->backer
);
525 atomic_init(&udpif
->enable_ufid
, enable_ufid
);
526 dpif_enable_upcall(udpif
->dpif
);
528 ovs_barrier_init(&udpif
->reval_barrier
, udpif
->n_revalidators
);
529 ovs_barrier_init(&udpif
->pause_barrier
, udpif
->n_revalidators
+ 1);
530 udpif
->reval_exit
= false;
531 udpif
->pause
= false;
532 udpif
->revalidators
= xzalloc(udpif
->n_revalidators
533 * sizeof *udpif
->revalidators
);
534 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
535 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
537 revalidator
->udpif
= udpif
;
538 revalidator
->thread
= ovs_thread_create(
539 "revalidator", udpif_revalidator
, revalidator
);
544 /* Pauses all revalidators. Should only be called by the main thread.
545 * When function returns, all revalidators are paused and will proceed
546 * only after udpif_resume_revalidators() is called. */
548 udpif_pause_revalidators(struct udpif
*udpif
)
550 if (ofproto_dpif_backer_enabled(udpif
->backer
)) {
551 latch_set(&udpif
->pause_latch
);
552 ovs_barrier_block(&udpif
->pause_barrier
);
556 /* Resumes the pausing of revalidators. Should only be called by the
559 udpif_resume_revalidators(struct udpif
*udpif
)
561 if (ofproto_dpif_backer_enabled(udpif
->backer
)) {
562 latch_poll(&udpif
->pause_latch
);
563 ovs_barrier_block(&udpif
->pause_barrier
);
567 /* Tells 'udpif' how many threads it should use to handle upcalls.
568 * 'n_handlers' and 'n_revalidators' can never be zero. 'udpif''s
569 * datapath handle must have packet reception enabled before starting
572 udpif_set_threads(struct udpif
*udpif
, size_t n_handlers
,
573 size_t n_revalidators
)
576 ovs_assert(n_handlers
&& n_revalidators
);
578 ovsrcu_quiesce_start();
579 if (udpif
->n_handlers
!= n_handlers
580 || udpif
->n_revalidators
!= n_revalidators
) {
581 udpif_stop_threads(udpif
);
584 if (!udpif
->handlers
&& !udpif
->revalidators
) {
587 error
= dpif_handlers_set(udpif
->dpif
, n_handlers
);
589 VLOG_ERR("failed to configure handlers in dpif %s: %s",
590 dpif_name(udpif
->dpif
), ovs_strerror(error
));
594 udpif_start_threads(udpif
, n_handlers
, n_revalidators
);
596 ovsrcu_quiesce_end();
599 /* Waits for all ongoing upcall translations to complete. This ensures that
600 * there are no transient references to any removed ofprotos (or other
601 * objects). In particular, this should be called after an ofproto is removed
602 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
604 udpif_synchronize(struct udpif
*udpif
)
606 /* This is stronger than necessary. It would be sufficient to ensure
607 * (somehow) that each handler and revalidator thread had passed through
608 * its main loop once. */
609 size_t n_handlers
= udpif
->n_handlers
;
610 size_t n_revalidators
= udpif
->n_revalidators
;
612 ovsrcu_quiesce_start();
613 udpif_stop_threads(udpif
);
614 udpif_start_threads(udpif
, n_handlers
, n_revalidators
);
615 ovsrcu_quiesce_end();
618 /* Notifies 'udpif' that something changed which may render previous
619 * xlate_actions() results invalid. */
621 udpif_revalidate(struct udpif
*udpif
)
623 seq_change(udpif
->reval_seq
);
626 /* Returns a seq which increments every time 'udpif' pulls stats from the
627 * datapath. Callers can use this to get a sense of when might be a good time
628 * to do periodic work which relies on relatively up to date statistics. */
630 udpif_dump_seq(struct udpif
*udpif
)
632 return udpif
->dump_seq
;
636 udpif_get_memory_usage(struct udpif
*udpif
, struct simap
*usage
)
640 simap_increase(usage
, "handlers", udpif
->n_handlers
);
642 simap_increase(usage
, "revalidators", udpif
->n_revalidators
);
643 for (i
= 0; i
< N_UMAPS
; i
++) {
644 simap_increase(usage
, "udpif keys", cmap_count(&udpif
->ukeys
[i
].cmap
));
648 /* Remove flows from a single datapath. */
650 udpif_flush(struct udpif
*udpif
)
652 size_t n_handlers
, n_revalidators
;
654 n_handlers
= udpif
->n_handlers
;
655 n_revalidators
= udpif
->n_revalidators
;
657 ovsrcu_quiesce_start();
659 udpif_stop_threads(udpif
);
660 dpif_flow_flush(udpif
->dpif
);
661 udpif_start_threads(udpif
, n_handlers
, n_revalidators
);
663 ovsrcu_quiesce_end();
666 /* Removes all flows from all datapaths. */
668 udpif_flush_all_datapaths(void)
672 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
678 udpif_use_ufid(struct udpif
*udpif
)
682 atomic_read_relaxed(&enable_ufid
, &enable
);
683 return enable
&& ofproto_dpif_get_enable_ufid(udpif
->backer
);
688 udpif_get_n_flows(struct udpif
*udpif
)
690 long long int time
, now
;
691 unsigned long flow_count
;
694 atomic_read_relaxed(&udpif
->n_flows_timestamp
, &time
);
695 if (time
< now
- 100 && !ovs_mutex_trylock(&udpif
->n_flows_mutex
)) {
696 struct dpif_dp_stats stats
;
698 atomic_store_relaxed(&udpif
->n_flows_timestamp
, now
);
699 dpif_get_dp_stats(udpif
->dpif
, &stats
);
700 flow_count
= stats
.n_flows
;
701 atomic_store_relaxed(&udpif
->n_flows
, flow_count
);
702 ovs_mutex_unlock(&udpif
->n_flows_mutex
);
704 atomic_read_relaxed(&udpif
->n_flows
, &flow_count
);
709 /* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH
710 * upcalls from dpif, processes the batch and installs corresponding flows
713 udpif_upcall_handler(void *arg
)
715 struct handler
*handler
= arg
;
716 struct udpif
*udpif
= handler
->udpif
;
718 while (!latch_is_set(&handler
->udpif
->exit_latch
)) {
719 if (recv_upcalls(handler
)) {
720 poll_immediate_wake();
722 dpif_recv_wait(udpif
->dpif
, handler
->handler_id
);
723 latch_wait(&udpif
->exit_latch
);
732 recv_upcalls(struct handler
*handler
)
734 struct udpif
*udpif
= handler
->udpif
;
735 uint64_t recv_stubs
[UPCALL_MAX_BATCH
][512 / 8];
736 struct ofpbuf recv_bufs
[UPCALL_MAX_BATCH
];
737 struct dpif_upcall dupcalls
[UPCALL_MAX_BATCH
];
738 struct upcall upcalls
[UPCALL_MAX_BATCH
];
739 struct flow flows
[UPCALL_MAX_BATCH
];
743 while (n_upcalls
< UPCALL_MAX_BATCH
) {
744 struct ofpbuf
*recv_buf
= &recv_bufs
[n_upcalls
];
745 struct dpif_upcall
*dupcall
= &dupcalls
[n_upcalls
];
746 struct upcall
*upcall
= &upcalls
[n_upcalls
];
747 struct flow
*flow
= &flows
[n_upcalls
];
751 ofpbuf_use_stub(recv_buf
, recv_stubs
[n_upcalls
],
752 sizeof recv_stubs
[n_upcalls
]);
753 if (dpif_recv(udpif
->dpif
, handler
->handler_id
, dupcall
, recv_buf
)) {
754 ofpbuf_uninit(recv_buf
);
758 if (odp_flow_key_to_flow(dupcall
->key
, dupcall
->key_len
, flow
)
764 mru
= nl_attr_get_u16(dupcall
->mru
);
769 error
= upcall_receive(upcall
, udpif
->backer
, &dupcall
->packet
,
770 dupcall
->type
, dupcall
->userdata
, flow
, mru
,
771 &dupcall
->ufid
, PMD_ID_NULL
);
773 if (error
== ENODEV
) {
774 /* Received packet on datapath port for which we couldn't
775 * associate an ofproto. This can happen if a port is removed
776 * while traffic is being received. Print a rate-limited
777 * message in case it happens frequently. */
778 dpif_flow_put(udpif
->dpif
, DPIF_FP_CREATE
, dupcall
->key
,
779 dupcall
->key_len
, NULL
, 0, NULL
, 0,
780 &dupcall
->ufid
, PMD_ID_NULL
, NULL
);
781 VLOG_INFO_RL(&rl
, "received packet on unassociated datapath "
782 "port %"PRIu32
, flow
->in_port
.odp_port
);
787 upcall
->key
= dupcall
->key
;
788 upcall
->key_len
= dupcall
->key_len
;
789 upcall
->ufid
= &dupcall
->ufid
;
791 upcall
->out_tun_key
= dupcall
->out_tun_key
;
792 upcall
->actions
= dupcall
->actions
;
794 pkt_metadata_from_flow(&dupcall
->packet
.md
, flow
);
795 flow_extract(&dupcall
->packet
, flow
);
797 error
= process_upcall(udpif
, upcall
,
798 &upcall
->odp_actions
, &upcall
->wc
);
807 upcall_uninit(upcall
);
809 dp_packet_uninit(&dupcall
->packet
);
810 ofpbuf_uninit(recv_buf
);
814 handle_upcalls(handler
->udpif
, upcalls
, n_upcalls
);
815 for (i
= 0; i
< n_upcalls
; i
++) {
816 dp_packet_uninit(&dupcalls
[i
].packet
);
817 ofpbuf_uninit(&recv_bufs
[i
]);
818 upcall_uninit(&upcalls
[i
]);
826 udpif_revalidator(void *arg
)
828 /* Used by all revalidators. */
829 struct revalidator
*revalidator
= arg
;
830 struct udpif
*udpif
= revalidator
->udpif
;
831 bool leader
= revalidator
== &udpif
->revalidators
[0];
833 /* Used only by the leader. */
834 long long int start_time
= 0;
835 uint64_t last_reval_seq
= 0;
838 revalidator
->id
= ovsthread_id_self();
843 recirc_run(); /* Recirculation cleanup. */
845 reval_seq
= seq_read(udpif
->reval_seq
);
846 last_reval_seq
= reval_seq
;
848 n_flows
= udpif_get_n_flows(udpif
);
849 udpif
->max_n_flows
= MAX(n_flows
, udpif
->max_n_flows
);
850 udpif
->avg_n_flows
= (udpif
->avg_n_flows
+ n_flows
) / 2;
852 /* Only the leader checks the pause latch to prevent a race where
853 * some threads think it's false and proceed to block on
854 * reval_barrier and others think it's true and block indefinitely
855 * on the pause_barrier */
856 udpif
->pause
= latch_is_set(&udpif
->pause_latch
);
858 /* Only the leader checks the exit latch to prevent a race where
859 * some threads think it's true and exit and others think it's
860 * false and block indefinitely on the reval_barrier */
861 udpif
->reval_exit
= latch_is_set(&udpif
->exit_latch
);
863 start_time
= time_msec();
864 if (!udpif
->reval_exit
) {
867 terse_dump
= udpif_use_ufid(udpif
);
868 udpif
->dump
= dpif_flow_dump_create(udpif
->dpif
, terse_dump
);
872 /* Wait for the leader to start the flow dump. */
873 ovs_barrier_block(&udpif
->reval_barrier
);
875 revalidator_pause(revalidator
);
878 if (udpif
->reval_exit
) {
881 revalidate(revalidator
);
883 /* Wait for all flows to have been dumped before we garbage collect. */
884 ovs_barrier_block(&udpif
->reval_barrier
);
885 revalidator_sweep(revalidator
);
887 /* Wait for all revalidators to finish garbage collection. */
888 ovs_barrier_block(&udpif
->reval_barrier
);
891 unsigned int flow_limit
;
892 long long int duration
;
894 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
896 dpif_flow_dump_destroy(udpif
->dump
);
897 seq_change(udpif
->dump_seq
);
899 duration
= MAX(time_msec() - start_time
, 1);
900 udpif
->dump_duration
= duration
;
901 if (duration
> 2000) {
902 flow_limit
/= duration
/ 1000;
903 } else if (duration
> 1300) {
904 flow_limit
= flow_limit
* 3 / 4;
905 } else if (duration
< 1000 && n_flows
> 2000
906 && flow_limit
< n_flows
* 1000 / duration
) {
909 flow_limit
= MIN(ofproto_flow_limit
, MAX(flow_limit
, 1000));
910 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
912 if (duration
> 2000) {
913 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
917 poll_timer_wait_until(start_time
+ MIN(ofproto_max_idle
, 500));
918 seq_wait(udpif
->reval_seq
, last_reval_seq
);
919 latch_wait(&udpif
->exit_latch
);
920 latch_wait(&udpif
->pause_latch
);
928 static enum upcall_type
929 classify_upcall(enum dpif_upcall_type type
, const struct nlattr
*userdata
)
931 union user_action_cookie cookie
;
934 /* First look at the upcall type. */
942 case DPIF_N_UC_TYPES
:
944 VLOG_WARN_RL(&rl
, "upcall has unexpected type %"PRIu32
, type
);
948 /* "action" upcalls need a closer look. */
950 VLOG_WARN_RL(&rl
, "action upcall missing cookie");
953 userdata_len
= nl_attr_get_size(userdata
);
954 if (userdata_len
< sizeof cookie
.type
955 || userdata_len
> sizeof cookie
) {
956 VLOG_WARN_RL(&rl
, "action upcall cookie has unexpected size %"PRIuSIZE
,
960 memset(&cookie
, 0, sizeof cookie
);
961 memcpy(&cookie
, nl_attr_get(userdata
), userdata_len
);
962 if (userdata_len
== MAX(8, sizeof cookie
.sflow
)
963 && cookie
.type
== USER_ACTION_COOKIE_SFLOW
) {
965 } else if (userdata_len
== MAX(8, sizeof cookie
.slow_path
)
966 && cookie
.type
== USER_ACTION_COOKIE_SLOW_PATH
) {
968 } else if (userdata_len
== MAX(8, sizeof cookie
.flow_sample
)
969 && cookie
.type
== USER_ACTION_COOKIE_FLOW_SAMPLE
) {
970 return FLOW_SAMPLE_UPCALL
;
971 } else if (userdata_len
== MAX(8, sizeof cookie
.ipfix
)
972 && cookie
.type
== USER_ACTION_COOKIE_IPFIX
) {
975 VLOG_WARN_RL(&rl
, "invalid user cookie of type %"PRIu16
976 " and size %"PRIuSIZE
, cookie
.type
, userdata_len
);
981 /* Calculates slow path actions for 'xout'. 'buf' must statically be
982 * initialized with at least 128 bytes of space. */
984 compose_slow_path(struct udpif
*udpif
, struct xlate_out
*xout
,
985 const struct flow
*flow
, odp_port_t odp_in_port
,
988 union user_action_cookie cookie
;
992 cookie
.type
= USER_ACTION_COOKIE_SLOW_PATH
;
993 cookie
.slow_path
.unused
= 0;
994 cookie
.slow_path
.reason
= xout
->slow
;
996 port
= xout
->slow
& (SLOW_CFM
| SLOW_BFD
| SLOW_LACP
| SLOW_STP
)
999 pid
= dpif_port_get_pid(udpif
->dpif
, port
, flow_hash_5tuple(flow
, 0));
1000 odp_put_userspace_action(pid
, &cookie
, sizeof cookie
.slow_path
,
1001 ODPP_NONE
, false, buf
);
1004 /* If there is no error, the upcall must be destroyed with upcall_uninit()
1005 * before quiescing, as the referred objects are guaranteed to exist only
1006 * until the calling thread quiesces. Otherwise, do not call upcall_uninit()
1007 * since the 'upcall->put_actions' remains uninitialized. */
1009 upcall_receive(struct upcall
*upcall
, const struct dpif_backer
*backer
,
1010 const struct dp_packet
*packet
, enum dpif_upcall_type type
,
1011 const struct nlattr
*userdata
, const struct flow
*flow
,
1012 const unsigned int mru
,
1013 const ovs_u128
*ufid
, const unsigned pmd_id
)
1017 error
= xlate_lookup(backer
, flow
, &upcall
->ofproto
, &upcall
->ipfix
,
1018 &upcall
->sflow
, NULL
, &upcall
->in_port
);
1023 upcall
->recirc
= NULL
;
1024 upcall
->have_recirc_ref
= false;
1025 upcall
->flow
= flow
;
1026 upcall
->packet
= packet
;
1027 upcall
->ufid
= ufid
;
1028 upcall
->pmd_id
= pmd_id
;
1029 upcall
->type
= type
;
1030 upcall
->userdata
= userdata
;
1031 ofpbuf_use_stub(&upcall
->odp_actions
, upcall
->odp_actions_stub
,
1032 sizeof upcall
->odp_actions_stub
);
1033 ofpbuf_init(&upcall
->put_actions
, 0);
1035 upcall
->xout_initialized
= false;
1036 upcall
->ukey_persists
= false;
1038 upcall
->ukey
= NULL
;
1040 upcall
->key_len
= 0;
1043 upcall
->out_tun_key
= NULL
;
1044 upcall
->actions
= NULL
;
1050 upcall_xlate(struct udpif
*udpif
, struct upcall
*upcall
,
1051 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1053 struct dpif_flow_stats stats
;
1054 struct xlate_in xin
;
1056 stats
.n_packets
= 1;
1057 stats
.n_bytes
= dp_packet_size(upcall
->packet
);
1058 stats
.used
= time_msec();
1059 stats
.tcp_flags
= ntohs(upcall
->flow
->tcp_flags
);
1061 xlate_in_init(&xin
, upcall
->ofproto
, upcall
->flow
, upcall
->in_port
, NULL
,
1062 stats
.tcp_flags
, upcall
->packet
, wc
, odp_actions
);
1064 if (upcall
->type
== DPIF_UC_MISS
) {
1065 xin
.resubmit_stats
= &stats
;
1067 if (xin
.frozen_state
) {
1068 /* We may install a datapath flow only if we get a reference to the
1069 * recirculation context (otherwise we could have recirculation
1070 * upcalls using recirculation ID for which no context can be
1071 * found). We may still execute the flow's actions even if we
1072 * don't install the flow. */
1073 upcall
->recirc
= recirc_id_node_from_state(xin
.frozen_state
);
1074 upcall
->have_recirc_ref
= recirc_id_node_try_ref_rcu(upcall
->recirc
);
1077 /* For non-miss upcalls, we are either executing actions (one of which
1078 * is an userspace action) for an upcall, in which case the stats have
1079 * already been taken care of, or there's a flow in the datapath which
1080 * this packet was accounted to. Presumably the revalidators will deal
1081 * with pushing its stats eventually. */
1084 upcall
->dump_seq
= seq_read(udpif
->dump_seq
);
1085 upcall
->reval_seq
= seq_read(udpif
->reval_seq
);
1086 xlate_actions(&xin
, &upcall
->xout
);
1087 upcall
->xout_initialized
= true;
1089 if (!upcall
->xout
.slow
) {
1090 ofpbuf_use_const(&upcall
->put_actions
,
1091 odp_actions
->data
, odp_actions
->size
);
1093 /* upcall->put_actions already initialized by upcall_receive(). */
1094 compose_slow_path(udpif
, &upcall
->xout
, upcall
->flow
,
1095 upcall
->flow
->in_port
.odp_port
,
1096 &upcall
->put_actions
);
1099 /* This function is also called for slow-pathed flows. As we are only
1100 * going to create new datapath flows for actual datapath misses, there is
1101 * no point in creating a ukey otherwise. */
1102 if (upcall
->type
== DPIF_UC_MISS
) {
1103 upcall
->ukey
= ukey_create_from_upcall(upcall
, wc
);
1108 upcall_uninit(struct upcall
*upcall
)
1111 if (upcall
->xout_initialized
) {
1112 xlate_out_uninit(&upcall
->xout
);
1114 ofpbuf_uninit(&upcall
->odp_actions
);
1115 ofpbuf_uninit(&upcall
->put_actions
);
1117 if (!upcall
->ukey_persists
) {
1118 ukey_delete__(upcall
->ukey
);
1120 } else if (upcall
->have_recirc_ref
) {
1121 /* The reference was transferred to the ukey if one was created. */
1122 recirc_id_node_unref(upcall
->recirc
);
1128 upcall_cb(const struct dp_packet
*packet
, const struct flow
*flow
, ovs_u128
*ufid
,
1129 unsigned pmd_id
, enum dpif_upcall_type type
,
1130 const struct nlattr
*userdata
, struct ofpbuf
*actions
,
1131 struct flow_wildcards
*wc
, struct ofpbuf
*put_actions
, void *aux
)
1133 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 1);
1134 struct udpif
*udpif
= aux
;
1135 unsigned int flow_limit
;
1136 struct upcall upcall
;
1140 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1141 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
1143 error
= upcall_receive(&upcall
, udpif
->backer
, packet
, type
, userdata
,
1144 flow
, 0, ufid
, pmd_id
);
1149 error
= process_upcall(udpif
, &upcall
, actions
, wc
);
1154 if (upcall
.xout
.slow
&& put_actions
) {
1155 ofpbuf_put(put_actions
, upcall
.put_actions
.data
,
1156 upcall
.put_actions
.size
);
1159 if (OVS_UNLIKELY(!megaflow
)) {
1160 flow_wildcards_init_for_packet(wc
, flow
);
1163 if (udpif_get_n_flows(udpif
) >= flow_limit
) {
1164 VLOG_WARN_RL(&rl
, "upcall_cb failure: datapath flow limit reached");
1169 /* Prevent miss flow installation if the key has recirculation ID but we
1170 * were not able to get a reference on it. */
1171 if (type
== DPIF_UC_MISS
&& upcall
.recirc
&& !upcall
.have_recirc_ref
) {
1172 VLOG_WARN_RL(&rl
, "upcall_cb failure: no reference for recirc flow");
1177 if (upcall
.ukey
&& !ukey_install(udpif
, upcall
.ukey
)) {
1178 VLOG_WARN_RL(&rl
, "upcall_cb failure: ukey installation fails");
1183 upcall
.ukey_persists
= true;
1185 upcall_uninit(&upcall
);
1190 process_upcall(struct udpif
*udpif
, struct upcall
*upcall
,
1191 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1193 const struct nlattr
*userdata
= upcall
->userdata
;
1194 const struct dp_packet
*packet
= upcall
->packet
;
1195 const struct flow
*flow
= upcall
->flow
;
1197 switch (classify_upcall(upcall
->type
, userdata
)) {
1199 upcall_xlate(udpif
, upcall
, odp_actions
, wc
);
1203 if (upcall
->sflow
) {
1204 union user_action_cookie cookie
;
1205 const struct nlattr
*actions
;
1206 size_t actions_len
= 0;
1207 struct dpif_sflow_actions sflow_actions
;
1208 memset(&sflow_actions
, 0, sizeof sflow_actions
);
1209 memset(&cookie
, 0, sizeof cookie
);
1210 memcpy(&cookie
, nl_attr_get(userdata
), sizeof cookie
.sflow
);
1211 if (upcall
->actions
) {
1212 /* Actions were passed up from datapath. */
1213 actions
= nl_attr_get(upcall
->actions
);
1214 actions_len
= nl_attr_get_size(upcall
->actions
);
1215 if (actions
&& actions_len
) {
1216 dpif_sflow_read_actions(flow
, actions
, actions_len
,
1220 if (actions_len
== 0) {
1221 /* Lookup actions in userspace cache. */
1222 struct udpif_key
*ukey
= ukey_lookup(udpif
, upcall
->ufid
,
1225 ukey_get_actions(ukey
, &actions
, &actions_len
);
1226 dpif_sflow_read_actions(flow
, actions
, actions_len
,
1230 dpif_sflow_received(upcall
->sflow
, packet
, flow
,
1231 flow
->in_port
.odp_port
, &cookie
,
1232 actions_len
> 0 ? &sflow_actions
: NULL
);
1237 if (upcall
->ipfix
) {
1238 union user_action_cookie cookie
;
1239 struct flow_tnl output_tunnel_key
;
1241 memset(&cookie
, 0, sizeof cookie
);
1242 memcpy(&cookie
, nl_attr_get(userdata
), sizeof cookie
.ipfix
);
1244 if (upcall
->out_tun_key
) {
1245 odp_tun_key_from_attr(upcall
->out_tun_key
, false,
1246 &output_tunnel_key
);
1248 dpif_ipfix_bridge_sample(upcall
->ipfix
, packet
, flow
,
1249 flow
->in_port
.odp_port
,
1250 cookie
.ipfix
.output_odp_port
,
1251 upcall
->out_tun_key
?
1252 &output_tunnel_key
: NULL
);
1256 case FLOW_SAMPLE_UPCALL
:
1257 if (upcall
->ipfix
) {
1258 union user_action_cookie cookie
;
1260 memset(&cookie
, 0, sizeof cookie
);
1261 memcpy(&cookie
, nl_attr_get(userdata
), sizeof cookie
.flow_sample
);
1263 /* The flow reflects exactly the contents of the packet.
1264 * Sample the packet using it. */
1265 dpif_ipfix_flow_sample(upcall
->ipfix
, packet
, flow
,
1266 cookie
.flow_sample
.collector_set_id
,
1267 cookie
.flow_sample
.probability
,
1268 cookie
.flow_sample
.obs_domain_id
,
1269 cookie
.flow_sample
.obs_point_id
);
1281 handle_upcalls(struct udpif
*udpif
, struct upcall
*upcalls
,
1284 struct dpif_op
*opsp
[UPCALL_MAX_BATCH
* 2];
1285 struct ukey_op ops
[UPCALL_MAX_BATCH
* 2];
1286 unsigned int flow_limit
;
1287 size_t n_ops
, n_opsp
, i
;
1290 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
1292 may_put
= udpif_get_n_flows(udpif
) < flow_limit
;
1294 /* Handle the packets individually in order of arrival.
1296 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1297 * processes received packets for these protocols.
1299 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1302 * The loop fills 'ops' with an array of operations to execute in the
1305 for (i
= 0; i
< n_upcalls
; i
++) {
1306 struct upcall
*upcall
= &upcalls
[i
];
1307 const struct dp_packet
*packet
= upcall
->packet
;
1310 /* Do not install a flow into the datapath if:
1312 * - The datapath already has too many flows.
1314 * - We received this packet via some flow installed in the kernel
1317 * - Upcall was a recirculation but we do not have a reference to
1318 * to the recirculation ID. */
1319 if (may_put
&& upcall
->type
== DPIF_UC_MISS
&&
1320 (!upcall
->recirc
|| upcall
->have_recirc_ref
)) {
1321 struct udpif_key
*ukey
= upcall
->ukey
;
1323 upcall
->ukey_persists
= true;
1327 op
->dop
.type
= DPIF_OP_FLOW_PUT
;
1328 op
->dop
.u
.flow_put
.flags
= DPIF_FP_CREATE
;
1329 op
->dop
.u
.flow_put
.key
= ukey
->key
;
1330 op
->dop
.u
.flow_put
.key_len
= ukey
->key_len
;
1331 op
->dop
.u
.flow_put
.mask
= ukey
->mask
;
1332 op
->dop
.u
.flow_put
.mask_len
= ukey
->mask_len
;
1333 op
->dop
.u
.flow_put
.ufid
= upcall
->ufid
;
1334 op
->dop
.u
.flow_put
.stats
= NULL
;
1335 ukey_get_actions(ukey
, &op
->dop
.u
.flow_put
.actions
,
1336 &op
->dop
.u
.flow_put
.actions_len
);
1339 if (upcall
->odp_actions
.size
) {
1342 op
->dop
.type
= DPIF_OP_EXECUTE
;
1343 op
->dop
.u
.execute
.packet
= CONST_CAST(struct dp_packet
*, packet
);
1344 op
->dop
.u
.execute
.flow
= upcall
->flow
;
1345 odp_key_to_pkt_metadata(upcall
->key
, upcall
->key_len
,
1346 &op
->dop
.u
.execute
.packet
->md
);
1347 op
->dop
.u
.execute
.actions
= upcall
->odp_actions
.data
;
1348 op
->dop
.u
.execute
.actions_len
= upcall
->odp_actions
.size
;
1349 op
->dop
.u
.execute
.needs_help
= (upcall
->xout
.slow
& SLOW_ACTION
) != 0;
1350 op
->dop
.u
.execute
.probe
= false;
1351 op
->dop
.u
.execute
.mtu
= upcall
->mru
;
1357 * We install ukeys before installing the flows, locking them for exclusive
1358 * access by this thread for the period of installation. This ensures that
1359 * other threads won't attempt to delete the flows as we are creating them.
1362 for (i
= 0; i
< n_ops
; i
++) {
1363 struct udpif_key
*ukey
= ops
[i
].ukey
;
1366 /* If we can't install the ukey, don't install the flow. */
1367 if (!ukey_install_start(udpif
, ukey
)) {
1368 ukey_delete__(ukey
);
1373 opsp
[n_opsp
++] = &ops
[i
].dop
;
1375 dpif_operate(udpif
->dpif
, opsp
, n_opsp
);
1376 for (i
= 0; i
< n_ops
; i
++) {
1378 ukey_install_finish(ops
[i
].ukey
, ops
[i
].dop
.error
);
1384 get_ukey_hash(const ovs_u128
*ufid
, const unsigned pmd_id
)
1386 return hash_2words(ufid
->u32
[0], pmd_id
);
1389 static struct udpif_key
*
1390 ukey_lookup(struct udpif
*udpif
, const ovs_u128
*ufid
, const unsigned pmd_id
)
1392 struct udpif_key
*ukey
;
1393 int idx
= get_ukey_hash(ufid
, pmd_id
) % N_UMAPS
;
1394 struct cmap
*cmap
= &udpif
->ukeys
[idx
].cmap
;
1396 CMAP_FOR_EACH_WITH_HASH (ukey
, cmap_node
,
1397 get_ukey_hash(ufid
, pmd_id
), cmap
) {
1398 if (ovs_u128_equals(ukey
->ufid
, *ufid
)) {
1405 /* Provides safe lockless access of RCU protected 'ukey->actions'. Callers may
1406 * alternatively access the field directly if they take 'ukey->mutex'. */
1408 ukey_get_actions(struct udpif_key
*ukey
, const struct nlattr
**actions
, size_t *size
)
1410 const struct ofpbuf
*buf
= ovsrcu_get(struct ofpbuf
*, &ukey
->actions
);
1411 *actions
= buf
->data
;
1416 ukey_set_actions(struct udpif_key
*ukey
, const struct ofpbuf
*actions
)
1418 ovsrcu_postpone(ofpbuf_delete
,
1419 ovsrcu_get_protected(struct ofpbuf
*, &ukey
->actions
));
1420 ovsrcu_set(&ukey
->actions
, ofpbuf_clone(actions
));
1423 static struct udpif_key
*
1424 ukey_create__(const struct nlattr
*key
, size_t key_len
,
1425 const struct nlattr
*mask
, size_t mask_len
,
1426 bool ufid_present
, const ovs_u128
*ufid
,
1427 const unsigned pmd_id
, const struct ofpbuf
*actions
,
1428 uint64_t dump_seq
, uint64_t reval_seq
, long long int used
,
1429 uint32_t key_recirc_id
, struct xlate_out
*xout
)
1430 OVS_NO_THREAD_SAFETY_ANALYSIS
1432 struct udpif_key
*ukey
= xmalloc(sizeof *ukey
);
1434 memcpy(&ukey
->keybuf
, key
, key_len
);
1435 ukey
->key
= &ukey
->keybuf
.nla
;
1436 ukey
->key_len
= key_len
;
1437 memcpy(&ukey
->maskbuf
, mask
, mask_len
);
1438 ukey
->mask
= &ukey
->maskbuf
.nla
;
1439 ukey
->mask_len
= mask_len
;
1440 ukey
->ufid_present
= ufid_present
;
1442 ukey
->pmd_id
= pmd_id
;
1443 ukey
->hash
= get_ukey_hash(&ukey
->ufid
, pmd_id
);
1445 ovsrcu_init(&ukey
->actions
, NULL
);
1446 ukey_set_actions(ukey
, actions
);
1448 ovs_mutex_init(&ukey
->mutex
);
1449 ukey
->dump_seq
= dump_seq
;
1450 ukey
->reval_seq
= reval_seq
;
1451 ukey
->flow_exists
= false;
1452 ukey
->created
= time_msec();
1453 memset(&ukey
->stats
, 0, sizeof ukey
->stats
);
1454 ukey
->stats
.used
= used
;
1455 ukey
->xcache
= NULL
;
1457 ukey
->key_recirc_id
= key_recirc_id
;
1458 recirc_refs_init(&ukey
->recircs
);
1460 /* Take ownership of the action recirc id references. */
1461 recirc_refs_swap(&ukey
->recircs
, &xout
->recircs
);
1467 static struct udpif_key
*
1468 ukey_create_from_upcall(struct upcall
*upcall
, struct flow_wildcards
*wc
)
1470 struct odputil_keybuf keystub
, maskstub
;
1471 struct ofpbuf keybuf
, maskbuf
;
1473 struct odp_flow_key_parms odp_parms
= {
1474 .flow
= upcall
->flow
,
1478 odp_parms
.support
= ofproto_dpif_get_support(upcall
->ofproto
)->odp
;
1479 if (upcall
->key_len
) {
1480 ofpbuf_use_const(&keybuf
, upcall
->key
, upcall
->key_len
);
1482 /* dpif-netdev doesn't provide a netlink-formatted flow key in the
1483 * upcall, so convert the upcall's flow here. */
1484 ofpbuf_use_stack(&keybuf
, &keystub
, sizeof keystub
);
1485 odp_parms
.odp_in_port
= upcall
->flow
->in_port
.odp_port
;
1486 odp_flow_key_from_flow(&odp_parms
, &keybuf
);
1489 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1490 ofpbuf_use_stack(&maskbuf
, &maskstub
, sizeof maskstub
);
1492 odp_parms
.odp_in_port
= ODPP_NONE
;
1493 odp_parms
.key_buf
= &keybuf
;
1495 odp_flow_key_from_mask(&odp_parms
, &maskbuf
);
1498 return ukey_create__(keybuf
.data
, keybuf
.size
, maskbuf
.data
, maskbuf
.size
,
1499 true, upcall
->ufid
, upcall
->pmd_id
,
1500 &upcall
->put_actions
, upcall
->dump_seq
,
1501 upcall
->reval_seq
, 0,
1502 upcall
->have_recirc_ref
? upcall
->recirc
->id
: 0,
1507 ukey_create_from_dpif_flow(const struct udpif
*udpif
,
1508 const struct dpif_flow
*flow
,
1509 struct udpif_key
**ukey
)
1511 struct dpif_flow full_flow
;
1512 struct ofpbuf actions
;
1513 uint64_t dump_seq
, reval_seq
;
1514 uint64_t stub
[DPIF_FLOW_BUFSIZE
/ 8];
1515 const struct nlattr
*a
;
1518 if (!flow
->key_len
|| !flow
->actions_len
) {
1522 /* If the key or actions were not provided by the datapath, fetch the
1524 ofpbuf_use_stack(&buf
, &stub
, sizeof stub
);
1525 err
= dpif_flow_get(udpif
->dpif
, flow
->key
, flow
->key_len
,
1526 flow
->ufid_present
? &flow
->ufid
: NULL
,
1527 flow
->pmd_id
, &buf
, &full_flow
);
1534 /* Check the flow actions for recirculation action. As recirculation
1535 * relies on OVS userspace internal state, we need to delete all old
1536 * datapath flows with either a non-zero recirc_id in the key, or any
1537 * recirculation actions upon OVS restart. */
1538 NL_ATTR_FOR_EACH_UNSAFE (a
, left
, flow
->key
, flow
->key_len
) {
1539 if (nl_attr_type(a
) == OVS_KEY_ATTR_RECIRC_ID
1540 && nl_attr_get_u32(a
) != 0) {
1544 NL_ATTR_FOR_EACH_UNSAFE (a
, left
, flow
->actions
, flow
->actions_len
) {
1545 if (nl_attr_type(a
) == OVS_ACTION_ATTR_RECIRC
) {
1550 dump_seq
= seq_read(udpif
->dump_seq
);
1551 reval_seq
= seq_read(udpif
->reval_seq
);
1552 ofpbuf_use_const(&actions
, &flow
->actions
, flow
->actions_len
);
1553 *ukey
= ukey_create__(flow
->key
, flow
->key_len
,
1554 flow
->mask
, flow
->mask_len
, flow
->ufid_present
,
1555 &flow
->ufid
, flow
->pmd_id
, &actions
, dump_seq
,
1556 reval_seq
, flow
->stats
.used
, 0, NULL
);
1561 /* Attempts to insert a ukey into the shared ukey maps.
1563 * On success, returns true, installs the ukey and returns it in a locked
1564 * state. Otherwise, returns false. */
1566 ukey_install_start(struct udpif
*udpif
, struct udpif_key
*new_ukey
)
1567 OVS_TRY_LOCK(true, new_ukey
->mutex
)
1570 struct udpif_key
*old_ukey
;
1572 bool locked
= false;
1574 idx
= new_ukey
->hash
% N_UMAPS
;
1575 umap
= &udpif
->ukeys
[idx
];
1576 ovs_mutex_lock(&umap
->mutex
);
1577 old_ukey
= ukey_lookup(udpif
, &new_ukey
->ufid
, new_ukey
->pmd_id
);
1579 /* Uncommon case: A ukey is already installed with the same UFID. */
1580 if (old_ukey
->key_len
== new_ukey
->key_len
1581 && !memcmp(old_ukey
->key
, new_ukey
->key
, new_ukey
->key_len
)) {
1582 COVERAGE_INC(handler_duplicate_upcall
);
1584 struct ds ds
= DS_EMPTY_INITIALIZER
;
1586 odp_format_ufid(&old_ukey
->ufid
, &ds
);
1587 ds_put_cstr(&ds
, " ");
1588 odp_flow_key_format(old_ukey
->key
, old_ukey
->key_len
, &ds
);
1589 ds_put_cstr(&ds
, "\n");
1590 odp_format_ufid(&new_ukey
->ufid
, &ds
);
1591 ds_put_cstr(&ds
, " ");
1592 odp_flow_key_format(new_ukey
->key
, new_ukey
->key_len
, &ds
);
1594 VLOG_WARN_RL(&rl
, "Conflicting ukey for flows:\n%s", ds_cstr(&ds
));
1598 ovs_mutex_lock(&new_ukey
->mutex
);
1599 cmap_insert(&umap
->cmap
, &new_ukey
->cmap_node
, new_ukey
->hash
);
1602 ovs_mutex_unlock(&umap
->mutex
);
1608 ukey_install_finish__(struct udpif_key
*ukey
) OVS_REQUIRES(ukey
->mutex
)
1610 ukey
->flow_exists
= true;
1614 ukey_install_finish(struct udpif_key
*ukey
, int error
)
1615 OVS_RELEASES(ukey
->mutex
)
1618 ukey_install_finish__(ukey
);
1620 ovs_mutex_unlock(&ukey
->mutex
);
1626 ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
)
1628 /* The usual way to keep 'ukey->flow_exists' in sync with the datapath is
1629 * to call ukey_install_start(), install the corresponding datapath flow,
1630 * then call ukey_install_finish(). The netdev interface using upcall_cb()
1631 * doesn't provide a function to separately finish the flow installation,
1632 * so we perform the operations together here.
1634 * This is fine currently, as revalidator threads will only delete this
1635 * ukey during revalidator_sweep() and only if the dump_seq is mismatched.
1636 * It is unlikely for a revalidator thread to advance dump_seq and reach
1637 * the next GC phase between ukey creation and flow installation. */
1638 return ukey_install_start(udpif
, ukey
) && ukey_install_finish(ukey
, 0);
1641 /* Searches for a ukey in 'udpif->ukeys' that matches 'flow' and attempts to
1642 * lock the ukey. If the ukey does not exist, create it.
1644 * Returns 0 on success, setting *result to the matching ukey and returning it
1645 * in a locked state. Otherwise, returns an errno and clears *result. EBUSY
1646 * indicates that another thread is handling this flow. Other errors indicate
1647 * an unexpected condition creating a new ukey.
1649 * *error is an output parameter provided to appease the threadsafety analyser,
1650 * and its value matches the return value. */
1652 ukey_acquire(struct udpif
*udpif
, const struct dpif_flow
*flow
,
1653 struct udpif_key
**result
, int *error
)
1654 OVS_TRY_LOCK(0, (*result
)->mutex
)
1656 struct udpif_key
*ukey
;
1659 ukey
= ukey_lookup(udpif
, &flow
->ufid
, flow
->pmd_id
);
1661 retval
= ovs_mutex_trylock(&ukey
->mutex
);
1663 /* Usually we try to avoid installing flows from revalidator threads,
1664 * because locking on a umap may cause handler threads to block.
1665 * However there are certain cases, like when ovs-vswitchd is
1666 * restarted, where it is desirable to handle flows that exist in the
1667 * datapath gracefully (ie, don't just clear the datapath). */
1670 retval
= ukey_create_from_dpif_flow(udpif
, flow
, &ukey
);
1674 install
= ukey_install_start(udpif
, ukey
);
1676 ukey_install_finish__(ukey
);
1679 ukey_delete__(ukey
);
1695 ukey_delete__(struct udpif_key
*ukey
)
1696 OVS_NO_THREAD_SAFETY_ANALYSIS
1699 if (ukey
->key_recirc_id
) {
1700 recirc_free_id(ukey
->key_recirc_id
);
1702 recirc_refs_unref(&ukey
->recircs
);
1703 xlate_cache_delete(ukey
->xcache
);
1704 ofpbuf_delete(ovsrcu_get(struct ofpbuf
*, &ukey
->actions
));
1705 ovs_mutex_destroy(&ukey
->mutex
);
1711 ukey_delete(struct umap
*umap
, struct udpif_key
*ukey
)
1712 OVS_REQUIRES(umap
->mutex
)
1714 cmap_remove(&umap
->cmap
, &ukey
->cmap_node
, ukey
->hash
);
1715 ovsrcu_postpone(ukey_delete__
, ukey
);
1719 should_revalidate(const struct udpif
*udpif
, uint64_t packets
,
1722 long long int metric
, now
, duration
;
1724 if (udpif
->dump_duration
< 200) {
1725 /* We are likely to handle full revalidation for the flows. */
1729 /* Calculate the mean time between seeing these packets. If this
1730 * exceeds the threshold, then delete the flow rather than performing
1731 * costly revalidation for flows that aren't being hit frequently.
1733 * This is targeted at situations where the dump_duration is high (~1s),
1734 * and revalidation is triggered by a call to udpif_revalidate(). In
1735 * these situations, revalidation of all flows causes fluctuations in the
1736 * flow_limit due to the interaction with the dump_duration and max_idle.
1737 * This tends to result in deletion of low-throughput flows anyway, so
1738 * skip the revalidation and just delete those flows. */
1739 packets
= MAX(packets
, 1);
1740 now
= MAX(used
, time_msec());
1741 duration
= now
- used
;
1742 metric
= duration
/ packets
;
1745 /* The flow is receiving more than ~5pps, so keep it. */
1751 /* Verifies that the datapath actions of 'ukey' are still correct, and pushes
1754 * Returns a recommended action for 'ukey', options include:
1755 * UKEY_DELETE The ukey should be deleted.
1756 * UKEY_KEEP The ukey is fine as is.
1757 * UKEY_MODIFY The ukey's actions should be changed but is otherwise
1758 * fine. Callers should change the actions to those found
1759 * in the caller supplied 'odp_actions' buffer. The
1760 * recirculation references can be found in 'recircs' and
1761 * must be handled by the caller.
1763 * If the result is UKEY_MODIFY, then references to all recirc_ids used by the
1764 * new flow will be held within 'recircs' (which may be none).
1766 * The caller is responsible for both initializing 'recircs' prior this call,
1767 * and ensuring any references are eventually freed.
1769 static enum reval_result
1770 revalidate_ukey(struct udpif
*udpif
, struct udpif_key
*ukey
,
1771 const struct dpif_flow_stats
*stats
,
1772 struct ofpbuf
*odp_actions
, uint64_t reval_seq
,
1773 struct recirc_refs
*recircs
)
1774 OVS_REQUIRES(ukey
->mutex
)
1776 struct xlate_out xout
, *xoutp
;
1777 struct netflow
*netflow
;
1778 struct ofproto_dpif
*ofproto
;
1779 struct dpif_flow_stats push
;
1781 struct flow_wildcards dp_mask
, wc
;
1782 enum reval_result result
;
1783 ofp_port_t ofp_in_port
;
1784 struct xlate_in xin
;
1785 long long int last_used
;
1787 bool need_revalidate
;
1789 result
= UKEY_DELETE
;
1793 ofpbuf_clear(odp_actions
);
1794 need_revalidate
= (ukey
->reval_seq
!= reval_seq
);
1795 last_used
= ukey
->stats
.used
;
1796 push
.used
= stats
->used
;
1797 push
.tcp_flags
= stats
->tcp_flags
;
1798 push
.n_packets
= (stats
->n_packets
> ukey
->stats
.n_packets
1799 ? stats
->n_packets
- ukey
->stats
.n_packets
1801 push
.n_bytes
= (stats
->n_bytes
> ukey
->stats
.n_bytes
1802 ? stats
->n_bytes
- ukey
->stats
.n_bytes
1805 if (need_revalidate
&& last_used
1806 && !should_revalidate(udpif
, push
.n_packets
, last_used
)) {
1810 /* We will push the stats, so update the ukey stats cache. */
1811 ukey
->stats
= *stats
;
1812 if (!push
.n_packets
&& !need_revalidate
) {
1817 if (ukey
->xcache
&& !need_revalidate
) {
1818 xlate_push_stats(ukey
->xcache
, &push
);
1823 if (odp_flow_key_to_flow(ukey
->key
, ukey
->key_len
, &flow
)
1828 error
= xlate_lookup(udpif
->backer
, &flow
, &ofproto
, NULL
, NULL
, &netflow
,
1834 if (need_revalidate
) {
1835 xlate_cache_clear(ukey
->xcache
);
1837 if (!ukey
->xcache
) {
1838 ukey
->xcache
= xlate_cache_new();
1841 xlate_in_init(&xin
, ofproto
, &flow
, ofp_in_port
, NULL
, push
.tcp_flags
,
1842 NULL
, need_revalidate
? &wc
: NULL
, odp_actions
);
1843 if (push
.n_packets
) {
1844 xin
.resubmit_stats
= &push
;
1845 xin
.may_learn
= true;
1847 xin
.xcache
= ukey
->xcache
;
1848 xlate_actions(&xin
, &xout
);
1851 if (!need_revalidate
) {
1857 ofpbuf_clear(odp_actions
);
1858 compose_slow_path(udpif
, &xout
, &flow
, flow
.in_port
.odp_port
,
1862 if (odp_flow_key_to_mask(ukey
->mask
, ukey
->mask_len
, ukey
->key
,
1863 ukey
->key_len
, &dp_mask
, &flow
)
1868 /* Do not modify if any bit is wildcarded by the installed datapath flow,
1869 * but not the newly revalidated wildcard mask (wc), i.e., if revalidation
1870 * tells that the datapath flow is now too generic and must be narrowed
1871 * down. Note that we do not know if the datapath has ignored any of the
1872 * wildcarded bits, so we may be overtly conservative here. */
1873 if (flow_wildcards_has_extra(&dp_mask
, &wc
)) {
1877 if (!ofpbuf_equal(odp_actions
,
1878 ovsrcu_get(struct ofpbuf
*, &ukey
->actions
))) {
1879 /* The datapath mask was OK, but the actions seem to have changed.
1880 * Let's modify it in place. */
1881 result
= UKEY_MODIFY
;
1882 /* Transfer recirc action ID references to the caller. */
1883 recirc_refs_swap(recircs
, &xoutp
->recircs
);
1890 if (result
!= UKEY_DELETE
) {
1891 ukey
->reval_seq
= reval_seq
;
1893 if (netflow
&& result
== UKEY_DELETE
) {
1894 netflow_flow_clear(netflow
, &flow
);
1896 xlate_out_uninit(xoutp
);
1901 delete_op_init__(struct udpif
*udpif
, struct ukey_op
*op
,
1902 const struct dpif_flow
*flow
)
1905 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
1906 op
->dop
.u
.flow_del
.key
= flow
->key
;
1907 op
->dop
.u
.flow_del
.key_len
= flow
->key_len
;
1908 op
->dop
.u
.flow_del
.ufid
= flow
->ufid_present
? &flow
->ufid
: NULL
;
1909 op
->dop
.u
.flow_del
.pmd_id
= flow
->pmd_id
;
1910 op
->dop
.u
.flow_del
.stats
= &op
->stats
;
1911 op
->dop
.u
.flow_del
.terse
= udpif_use_ufid(udpif
);
1915 delete_op_init(struct udpif
*udpif
, struct ukey_op
*op
, struct udpif_key
*ukey
)
1918 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
1919 op
->dop
.u
.flow_del
.key
= ukey
->key
;
1920 op
->dop
.u
.flow_del
.key_len
= ukey
->key_len
;
1921 op
->dop
.u
.flow_del
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
1922 op
->dop
.u
.flow_del
.pmd_id
= ukey
->pmd_id
;
1923 op
->dop
.u
.flow_del
.stats
= &op
->stats
;
1924 op
->dop
.u
.flow_del
.terse
= udpif_use_ufid(udpif
);
1928 modify_op_init(struct ukey_op
*op
, struct udpif_key
*ukey
)
1931 op
->dop
.type
= DPIF_OP_FLOW_PUT
;
1932 op
->dop
.u
.flow_put
.flags
= DPIF_FP_MODIFY
;
1933 op
->dop
.u
.flow_put
.key
= ukey
->key
;
1934 op
->dop
.u
.flow_put
.key_len
= ukey
->key_len
;
1935 op
->dop
.u
.flow_put
.mask
= ukey
->mask
;
1936 op
->dop
.u
.flow_put
.mask_len
= ukey
->mask_len
;
1937 op
->dop
.u
.flow_put
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
1938 op
->dop
.u
.flow_put
.pmd_id
= ukey
->pmd_id
;
1939 op
->dop
.u
.flow_put
.stats
= NULL
;
1940 ukey_get_actions(ukey
, &op
->dop
.u
.flow_put
.actions
,
1941 &op
->dop
.u
.flow_put
.actions_len
);
1944 /* Executes datapath operations 'ops' and attributes stats retrieved from the
1945 * datapath as part of those operations. */
1947 push_dp_ops(struct udpif
*udpif
, struct ukey_op
*ops
, size_t n_ops
)
1949 struct dpif_op
*opsp
[REVALIDATE_MAX_BATCH
];
1952 ovs_assert(n_ops
<= REVALIDATE_MAX_BATCH
);
1953 for (i
= 0; i
< n_ops
; i
++) {
1954 opsp
[i
] = &ops
[i
].dop
;
1956 dpif_operate(udpif
->dpif
, opsp
, n_ops
);
1958 for (i
= 0; i
< n_ops
; i
++) {
1959 struct ukey_op
*op
= &ops
[i
];
1960 struct dpif_flow_stats
*push
, *stats
, push_buf
;
1962 stats
= op
->dop
.u
.flow_del
.stats
;
1965 if (op
->dop
.type
!= DPIF_OP_FLOW_DEL
) {
1966 /* Only deleted flows need their stats pushed. */
1970 if (op
->dop
.error
) {
1971 /* flow_del error, 'stats' is unusable. */
1976 ovs_mutex_lock(&op
->ukey
->mutex
);
1977 push
->used
= MAX(stats
->used
, op
->ukey
->stats
.used
);
1978 push
->tcp_flags
= stats
->tcp_flags
| op
->ukey
->stats
.tcp_flags
;
1979 push
->n_packets
= stats
->n_packets
- op
->ukey
->stats
.n_packets
;
1980 push
->n_bytes
= stats
->n_bytes
- op
->ukey
->stats
.n_bytes
;
1981 ovs_mutex_unlock(&op
->ukey
->mutex
);
1986 if (push
->n_packets
|| netflow_exists()) {
1987 const struct nlattr
*key
= op
->dop
.u
.flow_del
.key
;
1988 size_t key_len
= op
->dop
.u
.flow_del
.key_len
;
1989 struct ofproto_dpif
*ofproto
;
1990 struct netflow
*netflow
;
1991 ofp_port_t ofp_in_port
;
1996 ovs_mutex_lock(&op
->ukey
->mutex
);
1997 if (op
->ukey
->xcache
) {
1998 xlate_push_stats(op
->ukey
->xcache
, push
);
1999 ovs_mutex_unlock(&op
->ukey
->mutex
);
2002 ovs_mutex_unlock(&op
->ukey
->mutex
);
2003 key
= op
->ukey
->key
;
2004 key_len
= op
->ukey
->key_len
;
2007 if (odp_flow_key_to_flow(key
, key_len
, &flow
)
2012 error
= xlate_lookup(udpif
->backer
, &flow
, &ofproto
, NULL
, NULL
,
2013 &netflow
, &ofp_in_port
);
2015 struct xlate_in xin
;
2017 xlate_in_init(&xin
, ofproto
, &flow
, ofp_in_port
, NULL
,
2018 push
->tcp_flags
, NULL
, NULL
, NULL
);
2019 xin
.resubmit_stats
= push
->n_packets
? push
: NULL
;
2020 xin
.may_learn
= push
->n_packets
> 0;
2021 xlate_actions_for_side_effects(&xin
);
2024 netflow_flow_clear(netflow
, &flow
);
2031 /* Executes datapath operations 'ops', attributes stats retrieved from the
2032 * datapath, and deletes ukeys corresponding to deleted flows. */
2034 push_ukey_ops(struct udpif
*udpif
, struct umap
*umap
,
2035 struct ukey_op
*ops
, size_t n_ops
)
2039 push_dp_ops(udpif
, ops
, n_ops
);
2040 ovs_mutex_lock(&umap
->mutex
);
2041 for (i
= 0; i
< n_ops
; i
++) {
2042 if (ops
[i
].dop
.type
== DPIF_OP_FLOW_DEL
) {
2043 ukey_delete(umap
, ops
[i
].ukey
);
2046 ovs_mutex_unlock(&umap
->mutex
);
2050 log_unexpected_flow(const struct dpif_flow
*flow
, int error
)
2052 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(10, 60);
2053 struct ds ds
= DS_EMPTY_INITIALIZER
;
2055 ds_put_format(&ds
, "Failed to acquire udpif_key corresponding to "
2056 "unexpected flow (%s): ", ovs_strerror(error
));
2057 odp_format_ufid(&flow
->ufid
, &ds
);
2058 VLOG_WARN_RL(&rl
, "%s", ds_cstr(&ds
));
2062 reval_op_init(struct ukey_op
*op
, enum reval_result result
,
2063 struct udpif
*udpif
, struct udpif_key
*ukey
,
2064 struct recirc_refs
*recircs
, struct ofpbuf
*odp_actions
)
2066 if (result
== UKEY_DELETE
) {
2067 delete_op_init(udpif
, op
, ukey
);
2068 } else if (result
== UKEY_MODIFY
) {
2069 /* Store the new recircs. */
2070 recirc_refs_swap(&ukey
->recircs
, recircs
);
2071 /* Release old recircs. */
2072 recirc_refs_unref(recircs
);
2073 /* ukey->key_recirc_id remains, as the key is the same as before. */
2075 ukey_set_actions(ukey
, odp_actions
);
2076 modify_op_init(op
, ukey
);
2081 revalidate(struct revalidator
*revalidator
)
2083 uint64_t odp_actions_stub
[1024 / 8];
2084 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2086 struct udpif
*udpif
= revalidator
->udpif
;
2087 struct dpif_flow_dump_thread
*dump_thread
;
2088 uint64_t dump_seq
, reval_seq
;
2089 unsigned int flow_limit
;
2091 dump_seq
= seq_read(udpif
->dump_seq
);
2092 reval_seq
= seq_read(udpif
->reval_seq
);
2093 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2094 dump_thread
= dpif_flow_dump_thread_create(udpif
->dump
);
2096 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2099 struct dpif_flow flows
[REVALIDATE_MAX_BATCH
];
2100 const struct dpif_flow
*f
;
2103 long long int max_idle
;
2108 n_dumped
= dpif_flow_dump_next(dump_thread
, flows
, ARRAY_SIZE(flows
));
2115 /* In normal operation we want to keep flows around until they have
2116 * been idle for 'ofproto_max_idle' milliseconds. However:
2118 * - If the number of datapath flows climbs above 'flow_limit',
2119 * drop that down to 100 ms to try to bring the flows down to
2122 * - If the number of datapath flows climbs above twice
2123 * 'flow_limit', delete all the datapath flows as an emergency
2124 * measure. (We reassess this condition for the next batch of
2125 * datapath flows, so we will recover before all the flows are
2127 n_dp_flows
= udpif_get_n_flows(udpif
);
2128 kill_them_all
= n_dp_flows
> flow_limit
* 2;
2129 max_idle
= n_dp_flows
> flow_limit
? 100 : ofproto_max_idle
;
2131 for (f
= flows
; f
< &flows
[n_dumped
]; f
++) {
2132 long long int used
= f
->stats
.used
;
2133 struct recirc_refs recircs
= RECIRC_REFS_EMPTY_INITIALIZER
;
2134 enum reval_result result
;
2135 struct udpif_key
*ukey
;
2136 bool already_dumped
;
2139 if (ukey_acquire(udpif
, f
, &ukey
, &error
)) {
2140 if (error
== EBUSY
) {
2141 /* Another thread is processing this flow, so don't bother
2143 COVERAGE_INC(upcall_ukey_contention
);
2145 log_unexpected_flow(f
, error
);
2146 if (error
!= ENOENT
) {
2147 delete_op_init__(udpif
, &ops
[n_ops
++], f
);
2153 already_dumped
= ukey
->dump_seq
== dump_seq
;
2154 if (already_dumped
) {
2155 /* The flow has already been handled during this flow dump
2156 * operation. Skip it. */
2158 COVERAGE_INC(dumped_duplicate_flow
);
2160 COVERAGE_INC(dumped_new_flow
);
2162 ovs_mutex_unlock(&ukey
->mutex
);
2167 used
= ukey
->created
;
2169 if (kill_them_all
|| (used
&& used
< now
- max_idle
)) {
2170 result
= UKEY_DELETE
;
2172 result
= revalidate_ukey(udpif
, ukey
, &f
->stats
, &odp_actions
,
2173 reval_seq
, &recircs
);
2175 ukey
->dump_seq
= dump_seq
;
2176 ukey
->flow_exists
= result
!= UKEY_DELETE
;
2178 if (result
!= UKEY_KEEP
) {
2179 /* Takes ownership of 'recircs'. */
2180 reval_op_init(&ops
[n_ops
++], result
, udpif
, ukey
, &recircs
,
2183 ovs_mutex_unlock(&ukey
->mutex
);
2187 /* Push datapath ops but defer ukey deletion to 'sweep' phase. */
2188 push_dp_ops(udpif
, ops
, n_ops
);
2192 dpif_flow_dump_thread_destroy(dump_thread
);
2193 ofpbuf_uninit(&odp_actions
);
2196 /* Pauses the 'revalidator', can only proceed after main thread
2197 * calls udpif_resume_revalidators(). */
2199 revalidator_pause(struct revalidator
*revalidator
)
2201 /* The first block is for sync'ing the pause with main thread. */
2202 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2203 /* The second block is for pausing until main thread resumes. */
2204 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2208 revalidator_sweep__(struct revalidator
*revalidator
, bool purge
)
2210 struct udpif
*udpif
;
2211 uint64_t dump_seq
, reval_seq
;
2214 udpif
= revalidator
->udpif
;
2215 dump_seq
= seq_read(udpif
->dump_seq
);
2216 reval_seq
= seq_read(udpif
->reval_seq
);
2217 slice
= revalidator
- udpif
->revalidators
;
2218 ovs_assert(slice
< udpif
->n_revalidators
);
2220 for (int i
= slice
; i
< N_UMAPS
; i
+= udpif
->n_revalidators
) {
2221 uint64_t odp_actions_stub
[1024 / 8];
2222 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2224 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2225 struct udpif_key
*ukey
;
2226 struct umap
*umap
= &udpif
->ukeys
[i
];
2229 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2232 /* Handler threads could be holding a ukey lock while it installs a
2233 * new flow, so don't hang around waiting for access to it. */
2234 if (ovs_mutex_trylock(&ukey
->mutex
)) {
2237 flow_exists
= ukey
->flow_exists
;
2239 struct recirc_refs recircs
= RECIRC_REFS_EMPTY_INITIALIZER
;
2240 bool seq_mismatch
= (ukey
->dump_seq
!= dump_seq
2241 && ukey
->reval_seq
!= reval_seq
);
2242 enum reval_result result
;
2245 result
= UKEY_DELETE
;
2246 } else if (!seq_mismatch
) {
2249 struct dpif_flow_stats stats
;
2250 COVERAGE_INC(revalidate_missed_dp_flow
);
2251 memset(&stats
, 0, sizeof stats
);
2252 result
= revalidate_ukey(udpif
, ukey
, &stats
, &odp_actions
,
2253 reval_seq
, &recircs
);
2255 if (result
!= UKEY_KEEP
) {
2256 /* Clears 'recircs' if filled by revalidate_ukey(). */
2257 reval_op_init(&ops
[n_ops
++], result
, udpif
, ukey
, &recircs
,
2261 ovs_mutex_unlock(&ukey
->mutex
);
2264 /* The common flow deletion case involves deletion of the flow
2265 * during the dump phase and ukey deletion here. */
2266 ovs_mutex_lock(&umap
->mutex
);
2267 ukey_delete(umap
, ukey
);
2268 ovs_mutex_unlock(&umap
->mutex
);
2271 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2272 /* Update/delete missed flows and clean up corresponding ukeys
2274 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2280 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2283 ofpbuf_uninit(&odp_actions
);
2289 revalidator_sweep(struct revalidator
*revalidator
)
2291 revalidator_sweep__(revalidator
, false);
2295 revalidator_purge(struct revalidator
*revalidator
)
2297 revalidator_sweep__(revalidator
, true);
2300 /* In reaction to dpif purge, purges all 'ukey's with same 'pmd_id'. */
2302 dp_purge_cb(void *aux
, unsigned pmd_id
)
2304 struct udpif
*udpif
= aux
;
2307 udpif_pause_revalidators(udpif
);
2308 for (i
= 0; i
< N_UMAPS
; i
++) {
2309 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2310 struct udpif_key
*ukey
;
2311 struct umap
*umap
= &udpif
->ukeys
[i
];
2314 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2315 if (ukey
->pmd_id
== pmd_id
) {
2316 delete_op_init(udpif
, &ops
[n_ops
++], ukey
);
2317 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2318 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2325 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2330 udpif_resume_revalidators(udpif
);
2334 upcall_unixctl_show(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2335 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2337 struct ds ds
= DS_EMPTY_INITIALIZER
;
2338 struct udpif
*udpif
;
2340 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2341 unsigned int flow_limit
;
2345 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2346 ufid_enabled
= udpif_use_ufid(udpif
);
2348 ds_put_format(&ds
, "%s:\n", dpif_name(udpif
->dpif
));
2349 ds_put_format(&ds
, "\tflows : (current %lu)"
2350 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif
),
2351 udpif
->avg_n_flows
, udpif
->max_n_flows
, flow_limit
);
2352 ds_put_format(&ds
, "\tdump duration : %lldms\n", udpif
->dump_duration
);
2353 ds_put_format(&ds
, "\tufid enabled : ");
2355 ds_put_format(&ds
, "true\n");
2357 ds_put_format(&ds
, "false\n");
2359 ds_put_char(&ds
, '\n');
2361 for (i
= 0; i
< n_revalidators
; i
++) {
2362 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
2363 int j
, elements
= 0;
2365 for (j
= i
; j
< N_UMAPS
; j
+= n_revalidators
) {
2366 elements
+= cmap_count(&udpif
->ukeys
[j
].cmap
);
2368 ds_put_format(&ds
, "\t%u: (keys %d)\n", revalidator
->id
, elements
);
2372 unixctl_command_reply(conn
, ds_cstr(&ds
));
2376 /* Disable using the megaflows.
2378 * This command is only needed for advanced debugging, so it's not
2379 * documented in the man page. */
2381 upcall_unixctl_disable_megaflows(struct unixctl_conn
*conn
,
2382 int argc OVS_UNUSED
,
2383 const char *argv
[] OVS_UNUSED
,
2384 void *aux OVS_UNUSED
)
2386 atomic_store_relaxed(&enable_megaflows
, false);
2387 udpif_flush_all_datapaths();
2388 unixctl_command_reply(conn
, "megaflows disabled");
2391 /* Re-enable using megaflows.
2393 * This command is only needed for advanced debugging, so it's not
2394 * documented in the man page. */
2396 upcall_unixctl_enable_megaflows(struct unixctl_conn
*conn
,
2397 int argc OVS_UNUSED
,
2398 const char *argv
[] OVS_UNUSED
,
2399 void *aux OVS_UNUSED
)
2401 atomic_store_relaxed(&enable_megaflows
, true);
2402 udpif_flush_all_datapaths();
2403 unixctl_command_reply(conn
, "megaflows enabled");
2406 /* Disable skipping flow attributes during flow dump.
2408 * This command is only needed for advanced debugging, so it's not
2409 * documented in the man page. */
2411 upcall_unixctl_disable_ufid(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2412 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2414 atomic_store_relaxed(&enable_ufid
, false);
2415 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID disabled");
2418 /* Re-enable skipping flow attributes during flow dump.
2420 * This command is only needed for advanced debugging, so it's not documented
2421 * in the man page. */
2423 upcall_unixctl_enable_ufid(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2424 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2426 atomic_store_relaxed(&enable_ufid
, true);
2427 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID enabled "
2428 "for supported datapaths");
2431 /* Set the flow limit.
2433 * This command is only needed for advanced debugging, so it's not
2434 * documented in the man page. */
2436 upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
,
2437 int argc OVS_UNUSED
,
2438 const char *argv
[] OVS_UNUSED
,
2439 void *aux OVS_UNUSED
)
2441 struct ds ds
= DS_EMPTY_INITIALIZER
;
2442 struct udpif
*udpif
;
2443 unsigned int flow_limit
= atoi(argv
[1]);
2445 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2446 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
2448 ds_put_format(&ds
, "set flow_limit to %u\n", flow_limit
);
2449 unixctl_command_reply(conn
, ds_cstr(&ds
));
2454 upcall_unixctl_dump_wait(struct unixctl_conn
*conn
,
2455 int argc OVS_UNUSED
,
2456 const char *argv
[] OVS_UNUSED
,
2457 void *aux OVS_UNUSED
)
2459 if (ovs_list_is_singleton(&all_udpifs
)) {
2460 struct udpif
*udpif
= NULL
;
2463 udpif
= OBJECT_CONTAINING(ovs_list_front(&all_udpifs
), udpif
, list_node
);
2464 len
= (udpif
->n_conns
+ 1) * sizeof *udpif
->conns
;
2465 udpif
->conn_seq
= seq_read(udpif
->dump_seq
);
2466 udpif
->conns
= xrealloc(udpif
->conns
, len
);
2467 udpif
->conns
[udpif
->n_conns
++] = conn
;
2469 unixctl_command_reply_error(conn
, "can't wait on multiple udpifs.");
2474 upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2475 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2477 struct udpif
*udpif
;
2479 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2482 for (n
= 0; n
< udpif
->n_revalidators
; n
++) {
2483 revalidator_purge(&udpif
->revalidators
[n
]);
2486 unixctl_command_reply(conn
, "");