1 /* Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2015 Nicira, Inc.
3 * Licensed under the Apache License, Version 2.0 (the "License");
4 * you may not use this file except in compliance with the License.
5 * You may obtain a copy of the License at:
7 * http://www.apache.org/licenses/LICENSE-2.0
9 * Unless required by applicable law or agreed to in writing, software
10 * distributed under the License is distributed on an "AS IS" BASIS,
11 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 * See the License for the specific language governing permissions and
13 * limitations under the License. */
16 #include "ofproto-dpif-upcall.h"
26 #include "dynamic-string.h"
27 #include "fail-open.h"
28 #include "guarded-list.h"
33 #include "ofproto-dpif-ipfix.h"
34 #include "ofproto-dpif-sflow.h"
35 #include "ofproto-dpif-xlate.h"
38 #include "poll-loop.h"
41 #include "openvswitch/vlog.h"
43 #define MAX_QUEUE_LENGTH 512
44 #define UPCALL_MAX_BATCH 64
45 #define REVALIDATE_MAX_BATCH 50
47 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall
);
49 COVERAGE_DEFINE(dumped_duplicate_flow
);
50 COVERAGE_DEFINE(dumped_new_flow
);
51 COVERAGE_DEFINE(handler_duplicate_upcall
);
52 COVERAGE_DEFINE(upcall_ukey_contention
);
53 COVERAGE_DEFINE(revalidate_missed_dp_flow
);
55 /* A thread that reads upcalls from dpif, forwards each upcall's packet,
56 * and possibly sets up a kernel flow as a cache. */
58 struct udpif
*udpif
; /* Parent udpif. */
59 pthread_t thread
; /* Thread ID. */
60 uint32_t handler_id
; /* Handler id. */
63 /* In the absence of a multiple-writer multiple-reader datastructure for
64 * storing ukeys, we use a large number of cmaps, each with its own lock for
66 #define N_UMAPS 512 /* per udpif. */
68 struct ovs_mutex mutex
; /* Take for writing to the following. */
69 struct cmap cmap
; /* Datapath flow keys. */
72 /* A thread that processes datapath flows, updates OpenFlow statistics, and
73 * updates or removes them if necessary. */
75 struct udpif
*udpif
; /* Parent udpif. */
76 pthread_t thread
; /* Thread ID. */
77 unsigned int id
; /* ovsthread_id_self(). */
80 /* An upcall handler for ofproto_dpif.
82 * udpif keeps records of two kind of logically separate units:
87 * - An array of 'struct handler's for upcall handling and flow
93 * - Revalidation threads which read the datapath flow table and maintains
97 struct ovs_list list_node
; /* In all_udpifs list. */
99 struct dpif
*dpif
; /* Datapath handle. */
100 struct dpif_backer
*backer
; /* Opaque dpif_backer pointer. */
102 struct handler
*handlers
; /* Upcall handlers. */
105 struct revalidator
*revalidators
; /* Flow revalidators. */
106 size_t n_revalidators
;
108 struct latch exit_latch
; /* Tells child threads to exit. */
111 struct seq
*reval_seq
; /* Incremented to force revalidation. */
112 bool reval_exit
; /* Set by leader on 'exit_latch. */
113 struct ovs_barrier reval_barrier
; /* Barrier used by revalidators. */
114 struct dpif_flow_dump
*dump
; /* DPIF flow dump state. */
115 long long int dump_duration
; /* Duration of the last flow dump. */
116 struct seq
*dump_seq
; /* Increments each dump iteration. */
117 atomic_bool enable_ufid
; /* If true, skip dumping flow attrs. */
119 /* These variables provide a mechanism for the main thread to pause
120 * all revalidation without having to completely shut the threads down.
121 * 'pause_latch' is shared between the main thread and the lead
122 * revalidator thread, so when it is desirable to halt revalidation, the
123 * main thread will set the latch. 'pause' and 'pause_barrier' are shared
124 * by revalidator threads. The lead revalidator will set 'pause' when it
125 * observes the latch has been set, and this will cause all revalidator
126 * threads to wait on 'pause_barrier' at the beginning of the next
127 * revalidation round. */
128 bool pause
; /* Set by leader on 'pause_latch. */
129 struct latch pause_latch
; /* Set to force revalidators pause. */
130 struct ovs_barrier pause_barrier
; /* Barrier used to pause all */
131 /* revalidators by main thread. */
133 /* There are 'N_UMAPS' maps containing 'struct udpif_key' elements.
135 * During the flow dump phase, revalidators insert into these with a random
136 * distribution. During the garbage collection phase, each revalidator
137 * takes care of garbage collecting a slice of these maps. */
140 /* Datapath flow statistics. */
141 unsigned int max_n_flows
;
142 unsigned int avg_n_flows
;
144 /* Following fields are accessed and modified by different threads. */
145 atomic_uint flow_limit
; /* Datapath flow hard limit. */
147 /* n_flows_mutex prevents multiple threads updating these concurrently. */
148 atomic_uint n_flows
; /* Number of flows in the datapath. */
149 atomic_llong n_flows_timestamp
; /* Last time n_flows was updated. */
150 struct ovs_mutex n_flows_mutex
;
152 /* Following fields are accessed and modified only from the main thread. */
153 struct unixctl_conn
**conns
; /* Connections waiting on dump_seq. */
154 uint64_t conn_seq
; /* Corresponds to 'dump_seq' when
155 conns[n_conns-1] was stored. */
156 size_t n_conns
; /* Number of connections waiting. */
160 BAD_UPCALL
, /* Some kind of bug somewhere. */
161 MISS_UPCALL
, /* A flow miss. */
162 SFLOW_UPCALL
, /* sFlow sample. */
163 FLOW_SAMPLE_UPCALL
, /* Per-flow sampling. */
164 IPFIX_UPCALL
/* Per-bridge sampling. */
174 struct ofproto_dpif
*ofproto
; /* Parent ofproto. */
175 const struct recirc_id_node
*recirc
; /* Recirculation context. */
176 bool have_recirc_ref
; /* Reference held on recirc ctx? */
178 /* The flow and packet are only required to be constant when using
179 * dpif-netdev. If a modification is absolutely necessary, a const cast
180 * may be used with other datapaths. */
181 const struct flow
*flow
; /* Parsed representation of the packet. */
182 const ovs_u128
*ufid
; /* Unique identifier for 'flow'. */
183 unsigned pmd_id
; /* Datapath poll mode driver id. */
184 const struct dp_packet
*packet
; /* Packet associated with this upcall. */
185 ofp_port_t in_port
; /* OpenFlow in port, or OFPP_NONE. */
186 uint16_t mru
; /* If !0, Maximum receive unit of
187 fragmented IP packet */
189 enum dpif_upcall_type type
; /* Datapath type of the upcall. */
190 const struct nlattr
*userdata
; /* Userdata for DPIF_UC_ACTION Upcalls. */
191 const struct nlattr
*actions
; /* Flow actions in DPIF_UC_ACTION Upcalls. */
193 bool xout_initialized
; /* True if 'xout' must be uninitialized. */
194 struct xlate_out xout
; /* Result of xlate_actions(). */
195 struct ofpbuf odp_actions
; /* Datapath actions from xlate_actions(). */
196 struct flow_wildcards wc
; /* Dependencies that megaflow must match. */
197 struct ofpbuf put_actions
; /* Actions 'put' in the fastpath. */
199 struct dpif_ipfix
*ipfix
; /* IPFIX pointer or NULL. */
200 struct dpif_sflow
*sflow
; /* SFlow pointer or NULL. */
202 bool vsp_adjusted
; /* 'packet' and 'flow' were adjusted for
203 VLAN splinters if true. */
205 struct udpif_key
*ukey
; /* Revalidator flow cache. */
206 bool ukey_persists
; /* Set true to keep 'ukey' beyond the
207 lifetime of this upcall. */
209 uint64_t dump_seq
; /* udpif->dump_seq at translation time. */
210 uint64_t reval_seq
; /* udpif->reval_seq at translation time. */
212 /* Not used by the upcall callback interface. */
213 const struct nlattr
*key
; /* Datapath flow key. */
214 size_t key_len
; /* Datapath flow key length. */
215 const struct nlattr
*out_tun_key
; /* Datapath output tunnel key. */
217 uint64_t odp_actions_stub
[1024 / 8]; /* Stub for odp_actions. */
220 /* 'udpif_key's are responsible for tracking the little bit of state udpif
221 * needs to do flow expiration which can't be pulled directly from the
222 * datapath. They may be created by any handler or revalidator thread at any
223 * time, and read by any revalidator during the dump phase. They are however
224 * each owned by a single revalidator which takes care of destroying them
225 * during the garbage-collection phase.
227 * The mutex within the ukey protects some members of the ukey. The ukey
228 * itself is protected by RCU and is held within a umap in the parent udpif.
229 * Adding or removing a ukey from a umap is only safe when holding the
230 * corresponding umap lock. */
232 struct cmap_node cmap_node
; /* In parent revalidator 'ukeys' map. */
234 /* These elements are read only once created, and therefore aren't
235 * protected by a mutex. */
236 const struct nlattr
*key
; /* Datapath flow key. */
237 size_t key_len
; /* Length of 'key'. */
238 const struct nlattr
*mask
; /* Datapath flow mask. */
239 size_t mask_len
; /* Length of 'mask'. */
240 ovs_u128 ufid
; /* Unique flow identifier. */
241 bool ufid_present
; /* True if 'ufid' is in datapath. */
242 uint32_t hash
; /* Pre-computed hash for 'key'. */
243 unsigned pmd_id
; /* Datapath poll mode driver id. */
245 struct ovs_mutex mutex
; /* Guards the following. */
246 struct dpif_flow_stats stats OVS_GUARDED
; /* Last known stats.*/
247 long long int created OVS_GUARDED
; /* Estimate of creation time. */
248 uint64_t dump_seq OVS_GUARDED
; /* Tracks udpif->dump_seq. */
249 uint64_t reval_seq OVS_GUARDED
; /* Tracks udpif->reval_seq. */
250 bool flow_exists OVS_GUARDED
; /* Ensures flows are only deleted
252 /* Datapath flow actions as nlattrs. Protected by RCU. Read with
253 * ukey_get_actions(), and write with ukey_set_actions(). */
254 OVSRCU_TYPE(struct ofpbuf
*) actions
;
256 struct xlate_cache
*xcache OVS_GUARDED
; /* Cache for xlate entries that
257 * are affected by this ukey.
258 * Used for stats and learning.*/
260 struct odputil_keybuf buf
;
264 /* Recirculation IDs with references held by the ukey. */
266 uint32_t recircs
[]; /* 'n_recircs' id's for which references are held. */
269 /* Datapath operation with optional ukey attached. */
271 struct udpif_key
*ukey
;
272 struct dpif_flow_stats stats
; /* Stats for 'op'. */
273 struct dpif_op dop
; /* Flow operation. */
276 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
277 static struct ovs_list all_udpifs
= OVS_LIST_INITIALIZER(&all_udpifs
);
279 static size_t recv_upcalls(struct handler
*);
280 static int process_upcall(struct udpif
*, struct upcall
*,
281 struct ofpbuf
*odp_actions
, struct flow_wildcards
*);
282 static void handle_upcalls(struct udpif
*, struct upcall
*, size_t n_upcalls
);
283 static void udpif_stop_threads(struct udpif
*);
284 static void udpif_start_threads(struct udpif
*, size_t n_handlers
,
285 size_t n_revalidators
);
286 static void udpif_pause_revalidators(struct udpif
*);
287 static void udpif_resume_revalidators(struct udpif
*);
288 static void *udpif_upcall_handler(void *);
289 static void *udpif_revalidator(void *);
290 static unsigned long udpif_get_n_flows(struct udpif
*);
291 static void revalidate(struct revalidator
*);
292 static void revalidator_pause(struct revalidator
*);
293 static void revalidator_sweep(struct revalidator
*);
294 static void revalidator_purge(struct revalidator
*);
295 static void upcall_unixctl_show(struct unixctl_conn
*conn
, int argc
,
296 const char *argv
[], void *aux
);
297 static void upcall_unixctl_disable_megaflows(struct unixctl_conn
*, int argc
,
298 const char *argv
[], void *aux
);
299 static void upcall_unixctl_enable_megaflows(struct unixctl_conn
*, int argc
,
300 const char *argv
[], void *aux
);
301 static void upcall_unixctl_disable_ufid(struct unixctl_conn
*, int argc
,
302 const char *argv
[], void *aux
);
303 static void upcall_unixctl_enable_ufid(struct unixctl_conn
*, int argc
,
304 const char *argv
[], void *aux
);
305 static void upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
, int argc
,
306 const char *argv
[], void *aux
);
307 static void upcall_unixctl_dump_wait(struct unixctl_conn
*conn
, int argc
,
308 const char *argv
[], void *aux
);
309 static void upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc
,
310 const char *argv
[], void *aux
);
312 static struct udpif_key
*ukey_create_from_upcall(struct upcall
*,
313 struct flow_wildcards
*);
314 static int ukey_create_from_dpif_flow(const struct udpif
*,
315 const struct dpif_flow
*,
316 struct udpif_key
**);
317 static void ukey_get_actions(struct udpif_key
*, const struct nlattr
**actions
,
319 static bool ukey_install_start(struct udpif
*, struct udpif_key
*ukey
);
320 static bool ukey_install_finish(struct udpif_key
*ukey
, int error
);
321 static bool ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
);
322 static struct udpif_key
*ukey_lookup(struct udpif
*udpif
,
323 const ovs_u128
*ufid
);
324 static int ukey_acquire(struct udpif
*, const struct dpif_flow
*,
325 struct udpif_key
**result
, int *error
);
326 static void ukey_delete__(struct udpif_key
*);
327 static void ukey_delete(struct umap
*, struct udpif_key
*);
328 static enum upcall_type
classify_upcall(enum dpif_upcall_type type
,
329 const struct nlattr
*userdata
);
331 static int upcall_receive(struct upcall
*, const struct dpif_backer
*,
332 const struct dp_packet
*packet
, enum dpif_upcall_type
,
333 const struct nlattr
*userdata
, const struct flow
*,
334 const unsigned int mru
,
335 const ovs_u128
*ufid
, const unsigned pmd_id
);
336 static void upcall_uninit(struct upcall
*);
338 static upcall_callback upcall_cb
;
339 static dp_purge_callback dp_purge_cb
;
341 static atomic_bool enable_megaflows
= ATOMIC_VAR_INIT(true);
342 static atomic_bool enable_ufid
= ATOMIC_VAR_INIT(true);
347 static struct ovsthread_once once
= OVSTHREAD_ONCE_INITIALIZER
;
348 if (ovsthread_once_start(&once
)) {
349 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show
,
351 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
352 upcall_unixctl_disable_megaflows
, NULL
);
353 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
354 upcall_unixctl_enable_megaflows
, NULL
);
355 unixctl_command_register("upcall/disable-ufid", "", 0, 0,
356 upcall_unixctl_disable_ufid
, NULL
);
357 unixctl_command_register("upcall/enable-ufid", "", 0, 0,
358 upcall_unixctl_enable_ufid
, NULL
);
359 unixctl_command_register("upcall/set-flow-limit", "", 1, 1,
360 upcall_unixctl_set_flow_limit
, NULL
);
361 unixctl_command_register("revalidator/wait", "", 0, 0,
362 upcall_unixctl_dump_wait
, NULL
);
363 unixctl_command_register("revalidator/purge", "", 0, 0,
364 upcall_unixctl_purge
, NULL
);
365 ovsthread_once_done(&once
);
370 udpif_create(struct dpif_backer
*backer
, struct dpif
*dpif
)
372 struct udpif
*udpif
= xzalloc(sizeof *udpif
);
375 udpif
->backer
= backer
;
376 atomic_init(&udpif
->flow_limit
, MIN(ofproto_flow_limit
, 10000));
377 udpif
->reval_seq
= seq_create();
378 udpif
->dump_seq
= seq_create();
379 latch_init(&udpif
->exit_latch
);
380 latch_init(&udpif
->pause_latch
);
381 list_push_back(&all_udpifs
, &udpif
->list_node
);
382 atomic_init(&udpif
->enable_ufid
, false);
383 atomic_init(&udpif
->n_flows
, 0);
384 atomic_init(&udpif
->n_flows_timestamp
, LLONG_MIN
);
385 ovs_mutex_init(&udpif
->n_flows_mutex
);
386 udpif
->ukeys
= xmalloc(N_UMAPS
* sizeof *udpif
->ukeys
);
387 for (int i
= 0; i
< N_UMAPS
; i
++) {
388 cmap_init(&udpif
->ukeys
[i
].cmap
);
389 ovs_mutex_init(&udpif
->ukeys
[i
].mutex
);
392 dpif_register_upcall_cb(dpif
, upcall_cb
, udpif
);
393 dpif_register_dp_purge_cb(dpif
, dp_purge_cb
, udpif
);
399 udpif_run(struct udpif
*udpif
)
401 if (udpif
->conns
&& udpif
->conn_seq
!= seq_read(udpif
->dump_seq
)) {
404 for (i
= 0; i
< udpif
->n_conns
; i
++) {
405 unixctl_command_reply(udpif
->conns
[i
], NULL
);
414 udpif_destroy(struct udpif
*udpif
)
416 udpif_stop_threads(udpif
);
418 for (int i
= 0; i
< N_UMAPS
; i
++) {
419 cmap_destroy(&udpif
->ukeys
[i
].cmap
);
420 ovs_mutex_destroy(&udpif
->ukeys
[i
].mutex
);
425 list_remove(&udpif
->list_node
);
426 latch_destroy(&udpif
->exit_latch
);
427 latch_destroy(&udpif
->pause_latch
);
428 seq_destroy(udpif
->reval_seq
);
429 seq_destroy(udpif
->dump_seq
);
430 ovs_mutex_destroy(&udpif
->n_flows_mutex
);
434 /* Stops the handler and revalidator threads, must be enclosed in
435 * ovsrcu quiescent state unless when destroying udpif. */
437 udpif_stop_threads(struct udpif
*udpif
)
439 if (udpif
&& (udpif
->n_handlers
!= 0 || udpif
->n_revalidators
!= 0)) {
442 latch_set(&udpif
->exit_latch
);
444 for (i
= 0; i
< udpif
->n_handlers
; i
++) {
445 struct handler
*handler
= &udpif
->handlers
[i
];
447 xpthread_join(handler
->thread
, NULL
);
450 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
451 xpthread_join(udpif
->revalidators
[i
].thread
, NULL
);
454 dpif_disable_upcall(udpif
->dpif
);
456 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
457 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
459 /* Delete ukeys, and delete all flows from the datapath to prevent
460 * double-counting stats. */
461 revalidator_purge(revalidator
);
464 latch_poll(&udpif
->exit_latch
);
466 ovs_barrier_destroy(&udpif
->reval_barrier
);
467 ovs_barrier_destroy(&udpif
->pause_barrier
);
469 free(udpif
->revalidators
);
470 udpif
->revalidators
= NULL
;
471 udpif
->n_revalidators
= 0;
473 free(udpif
->handlers
);
474 udpif
->handlers
= NULL
;
475 udpif
->n_handlers
= 0;
479 /* Starts the handler and revalidator threads, must be enclosed in
480 * ovsrcu quiescent state. */
482 udpif_start_threads(struct udpif
*udpif
, size_t n_handlers
,
483 size_t n_revalidators
)
485 if (udpif
&& n_handlers
&& n_revalidators
) {
489 udpif
->n_handlers
= n_handlers
;
490 udpif
->n_revalidators
= n_revalidators
;
492 udpif
->handlers
= xzalloc(udpif
->n_handlers
* sizeof *udpif
->handlers
);
493 for (i
= 0; i
< udpif
->n_handlers
; i
++) {
494 struct handler
*handler
= &udpif
->handlers
[i
];
496 handler
->udpif
= udpif
;
497 handler
->handler_id
= i
;
498 handler
->thread
= ovs_thread_create(
499 "handler", udpif_upcall_handler
, handler
);
502 enable_ufid
= ofproto_dpif_get_enable_ufid(udpif
->backer
);
503 atomic_init(&udpif
->enable_ufid
, enable_ufid
);
504 dpif_enable_upcall(udpif
->dpif
);
506 ovs_barrier_init(&udpif
->reval_barrier
, udpif
->n_revalidators
);
507 ovs_barrier_init(&udpif
->pause_barrier
, udpif
->n_revalidators
+ 1);
508 udpif
->reval_exit
= false;
509 udpif
->pause
= false;
510 udpif
->revalidators
= xzalloc(udpif
->n_revalidators
511 * sizeof *udpif
->revalidators
);
512 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
513 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
515 revalidator
->udpif
= udpif
;
516 revalidator
->thread
= ovs_thread_create(
517 "revalidator", udpif_revalidator
, revalidator
);
522 /* Pauses all revalidators. Should only be called by the main thread.
523 * When function returns, all revalidators are paused and will proceed
524 * only after udpif_resume_revalidators() is called. */
526 udpif_pause_revalidators(struct udpif
*udpif
)
528 if (ofproto_dpif_backer_enabled(udpif
->backer
)) {
529 latch_set(&udpif
->pause_latch
);
530 ovs_barrier_block(&udpif
->pause_barrier
);
534 /* Resumes the pausing of revalidators. Should only be called by the
537 udpif_resume_revalidators(struct udpif
*udpif
)
539 if (ofproto_dpif_backer_enabled(udpif
->backer
)) {
540 latch_poll(&udpif
->pause_latch
);
541 ovs_barrier_block(&udpif
->pause_barrier
);
545 /* Tells 'udpif' how many threads it should use to handle upcalls.
546 * 'n_handlers' and 'n_revalidators' can never be zero. 'udpif''s
547 * datapath handle must have packet reception enabled before starting
550 udpif_set_threads(struct udpif
*udpif
, size_t n_handlers
,
551 size_t n_revalidators
)
554 ovs_assert(n_handlers
&& n_revalidators
);
556 ovsrcu_quiesce_start();
557 if (udpif
->n_handlers
!= n_handlers
558 || udpif
->n_revalidators
!= n_revalidators
) {
559 udpif_stop_threads(udpif
);
562 if (!udpif
->handlers
&& !udpif
->revalidators
) {
565 error
= dpif_handlers_set(udpif
->dpif
, n_handlers
);
567 VLOG_ERR("failed to configure handlers in dpif %s: %s",
568 dpif_name(udpif
->dpif
), ovs_strerror(error
));
572 udpif_start_threads(udpif
, n_handlers
, n_revalidators
);
574 ovsrcu_quiesce_end();
577 /* Waits for all ongoing upcall translations to complete. This ensures that
578 * there are no transient references to any removed ofprotos (or other
579 * objects). In particular, this should be called after an ofproto is removed
580 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
582 udpif_synchronize(struct udpif
*udpif
)
584 /* This is stronger than necessary. It would be sufficient to ensure
585 * (somehow) that each handler and revalidator thread had passed through
586 * its main loop once. */
587 size_t n_handlers
= udpif
->n_handlers
;
588 size_t n_revalidators
= udpif
->n_revalidators
;
590 ovsrcu_quiesce_start();
591 udpif_stop_threads(udpif
);
592 udpif_start_threads(udpif
, n_handlers
, n_revalidators
);
593 ovsrcu_quiesce_end();
596 /* Notifies 'udpif' that something changed which may render previous
597 * xlate_actions() results invalid. */
599 udpif_revalidate(struct udpif
*udpif
)
601 seq_change(udpif
->reval_seq
);
604 /* Returns a seq which increments every time 'udpif' pulls stats from the
605 * datapath. Callers can use this to get a sense of when might be a good time
606 * to do periodic work which relies on relatively up to date statistics. */
608 udpif_dump_seq(struct udpif
*udpif
)
610 return udpif
->dump_seq
;
614 udpif_get_memory_usage(struct udpif
*udpif
, struct simap
*usage
)
618 simap_increase(usage
, "handlers", udpif
->n_handlers
);
620 simap_increase(usage
, "revalidators", udpif
->n_revalidators
);
621 for (i
= 0; i
< N_UMAPS
; i
++) {
622 simap_increase(usage
, "udpif keys", cmap_count(&udpif
->ukeys
[i
].cmap
));
626 /* Remove flows from a single datapath. */
628 udpif_flush(struct udpif
*udpif
)
630 size_t n_handlers
, n_revalidators
;
632 n_handlers
= udpif
->n_handlers
;
633 n_revalidators
= udpif
->n_revalidators
;
635 ovsrcu_quiesce_start();
637 udpif_stop_threads(udpif
);
638 dpif_flow_flush(udpif
->dpif
);
639 udpif_start_threads(udpif
, n_handlers
, n_revalidators
);
641 ovsrcu_quiesce_end();
644 /* Removes all flows from all datapaths. */
646 udpif_flush_all_datapaths(void)
650 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
656 udpif_use_ufid(struct udpif
*udpif
)
660 atomic_read_relaxed(&enable_ufid
, &enable
);
661 return enable
&& ofproto_dpif_get_enable_ufid(udpif
->backer
);
666 udpif_get_n_flows(struct udpif
*udpif
)
668 long long int time
, now
;
669 unsigned long flow_count
;
672 atomic_read_relaxed(&udpif
->n_flows_timestamp
, &time
);
673 if (time
< now
- 100 && !ovs_mutex_trylock(&udpif
->n_flows_mutex
)) {
674 struct dpif_dp_stats stats
;
676 atomic_store_relaxed(&udpif
->n_flows_timestamp
, now
);
677 dpif_get_dp_stats(udpif
->dpif
, &stats
);
678 flow_count
= stats
.n_flows
;
679 atomic_store_relaxed(&udpif
->n_flows
, flow_count
);
680 ovs_mutex_unlock(&udpif
->n_flows_mutex
);
682 atomic_read_relaxed(&udpif
->n_flows
, &flow_count
);
687 /* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH
688 * upcalls from dpif, processes the batch and installs corresponding flows
691 udpif_upcall_handler(void *arg
)
693 struct handler
*handler
= arg
;
694 struct udpif
*udpif
= handler
->udpif
;
696 while (!latch_is_set(&handler
->udpif
->exit_latch
)) {
697 if (recv_upcalls(handler
)) {
698 poll_immediate_wake();
700 dpif_recv_wait(udpif
->dpif
, handler
->handler_id
);
701 latch_wait(&udpif
->exit_latch
);
710 recv_upcalls(struct handler
*handler
)
712 struct udpif
*udpif
= handler
->udpif
;
713 uint64_t recv_stubs
[UPCALL_MAX_BATCH
][512 / 8];
714 struct ofpbuf recv_bufs
[UPCALL_MAX_BATCH
];
715 struct dpif_upcall dupcalls
[UPCALL_MAX_BATCH
];
716 struct upcall upcalls
[UPCALL_MAX_BATCH
];
717 struct flow flows
[UPCALL_MAX_BATCH
];
721 while (n_upcalls
< UPCALL_MAX_BATCH
) {
722 struct ofpbuf
*recv_buf
= &recv_bufs
[n_upcalls
];
723 struct dpif_upcall
*dupcall
= &dupcalls
[n_upcalls
];
724 struct upcall
*upcall
= &upcalls
[n_upcalls
];
725 struct flow
*flow
= &flows
[n_upcalls
];
729 ofpbuf_use_stub(recv_buf
, recv_stubs
[n_upcalls
],
730 sizeof recv_stubs
[n_upcalls
]);
731 if (dpif_recv(udpif
->dpif
, handler
->handler_id
, dupcall
, recv_buf
)) {
732 ofpbuf_uninit(recv_buf
);
736 if (odp_flow_key_to_flow(dupcall
->key
, dupcall
->key_len
, flow
)
742 mru
= nl_attr_get_u16(dupcall
->mru
);
747 error
= upcall_receive(upcall
, udpif
->backer
, &dupcall
->packet
,
748 dupcall
->type
, dupcall
->userdata
, flow
, mru
,
749 &dupcall
->ufid
, PMD_ID_NULL
);
751 if (error
== ENODEV
) {
752 /* Received packet on datapath port for which we couldn't
753 * associate an ofproto. This can happen if a port is removed
754 * while traffic is being received. Print a rate-limited
755 * message in case it happens frequently. */
756 dpif_flow_put(udpif
->dpif
, DPIF_FP_CREATE
, dupcall
->key
,
757 dupcall
->key_len
, NULL
, 0, NULL
, 0,
758 &dupcall
->ufid
, PMD_ID_NULL
, NULL
);
759 VLOG_INFO_RL(&rl
, "received packet on unassociated datapath "
760 "port %"PRIu32
, flow
->in_port
.odp_port
);
765 upcall
->key
= dupcall
->key
;
766 upcall
->key_len
= dupcall
->key_len
;
767 upcall
->ufid
= &dupcall
->ufid
;
769 upcall
->out_tun_key
= dupcall
->out_tun_key
;
770 upcall
->actions
= dupcall
->actions
;
772 if (vsp_adjust_flow(upcall
->ofproto
, flow
, &dupcall
->packet
)) {
773 upcall
->vsp_adjusted
= true;
776 pkt_metadata_from_flow(&dupcall
->packet
.md
, flow
);
777 flow_extract(&dupcall
->packet
, flow
);
779 error
= process_upcall(udpif
, upcall
,
780 &upcall
->odp_actions
, &upcall
->wc
);
789 upcall_uninit(upcall
);
791 dp_packet_uninit(&dupcall
->packet
);
792 ofpbuf_uninit(recv_buf
);
796 handle_upcalls(handler
->udpif
, upcalls
, n_upcalls
);
797 for (i
= 0; i
< n_upcalls
; i
++) {
798 dp_packet_uninit(&dupcalls
[i
].packet
);
799 ofpbuf_uninit(&recv_bufs
[i
]);
800 upcall_uninit(&upcalls
[i
]);
808 udpif_revalidator(void *arg
)
810 /* Used by all revalidators. */
811 struct revalidator
*revalidator
= arg
;
812 struct udpif
*udpif
= revalidator
->udpif
;
813 bool leader
= revalidator
== &udpif
->revalidators
[0];
815 /* Used only by the leader. */
816 long long int start_time
= 0;
817 uint64_t last_reval_seq
= 0;
820 revalidator
->id
= ovsthread_id_self();
825 recirc_run(); /* Recirculation cleanup. */
827 reval_seq
= seq_read(udpif
->reval_seq
);
828 last_reval_seq
= reval_seq
;
830 n_flows
= udpif_get_n_flows(udpif
);
831 udpif
->max_n_flows
= MAX(n_flows
, udpif
->max_n_flows
);
832 udpif
->avg_n_flows
= (udpif
->avg_n_flows
+ n_flows
) / 2;
834 /* Only the leader checks the pause latch to prevent a race where
835 * some threads think it's false and proceed to block on
836 * reval_barrier and others think it's true and block indefinitely
837 * on the pause_barrier */
838 udpif
->pause
= latch_is_set(&udpif
->pause_latch
);
840 /* Only the leader checks the exit latch to prevent a race where
841 * some threads think it's true and exit and others think it's
842 * false and block indefinitely on the reval_barrier */
843 udpif
->reval_exit
= latch_is_set(&udpif
->exit_latch
);
845 start_time
= time_msec();
846 if (!udpif
->reval_exit
) {
849 terse_dump
= udpif_use_ufid(udpif
);
850 udpif
->dump
= dpif_flow_dump_create(udpif
->dpif
, terse_dump
);
854 /* Wait for the leader to start the flow dump. */
855 ovs_barrier_block(&udpif
->reval_barrier
);
857 revalidator_pause(revalidator
);
860 if (udpif
->reval_exit
) {
863 revalidate(revalidator
);
865 /* Wait for all flows to have been dumped before we garbage collect. */
866 ovs_barrier_block(&udpif
->reval_barrier
);
867 revalidator_sweep(revalidator
);
869 /* Wait for all revalidators to finish garbage collection. */
870 ovs_barrier_block(&udpif
->reval_barrier
);
873 unsigned int flow_limit
;
874 long long int duration
;
876 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
878 dpif_flow_dump_destroy(udpif
->dump
);
879 seq_change(udpif
->dump_seq
);
881 duration
= MAX(time_msec() - start_time
, 1);
882 udpif
->dump_duration
= duration
;
883 if (duration
> 2000) {
884 flow_limit
/= duration
/ 1000;
885 } else if (duration
> 1300) {
886 flow_limit
= flow_limit
* 3 / 4;
887 } else if (duration
< 1000 && n_flows
> 2000
888 && flow_limit
< n_flows
* 1000 / duration
) {
891 flow_limit
= MIN(ofproto_flow_limit
, MAX(flow_limit
, 1000));
892 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
894 if (duration
> 2000) {
895 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
899 poll_timer_wait_until(start_time
+ MIN(ofproto_max_idle
, 500));
900 seq_wait(udpif
->reval_seq
, last_reval_seq
);
901 latch_wait(&udpif
->exit_latch
);
902 latch_wait(&udpif
->pause_latch
);
910 static enum upcall_type
911 classify_upcall(enum dpif_upcall_type type
, const struct nlattr
*userdata
)
913 union user_action_cookie cookie
;
916 /* First look at the upcall type. */
924 case DPIF_N_UC_TYPES
:
926 VLOG_WARN_RL(&rl
, "upcall has unexpected type %"PRIu32
, type
);
930 /* "action" upcalls need a closer look. */
932 VLOG_WARN_RL(&rl
, "action upcall missing cookie");
935 userdata_len
= nl_attr_get_size(userdata
);
936 if (userdata_len
< sizeof cookie
.type
937 || userdata_len
> sizeof cookie
) {
938 VLOG_WARN_RL(&rl
, "action upcall cookie has unexpected size %"PRIuSIZE
,
942 memset(&cookie
, 0, sizeof cookie
);
943 memcpy(&cookie
, nl_attr_get(userdata
), userdata_len
);
944 if (userdata_len
== MAX(8, sizeof cookie
.sflow
)
945 && cookie
.type
== USER_ACTION_COOKIE_SFLOW
) {
947 } else if (userdata_len
== MAX(8, sizeof cookie
.slow_path
)
948 && cookie
.type
== USER_ACTION_COOKIE_SLOW_PATH
) {
950 } else if (userdata_len
== MAX(8, sizeof cookie
.flow_sample
)
951 && cookie
.type
== USER_ACTION_COOKIE_FLOW_SAMPLE
) {
952 return FLOW_SAMPLE_UPCALL
;
953 } else if (userdata_len
== MAX(8, sizeof cookie
.ipfix
)
954 && cookie
.type
== USER_ACTION_COOKIE_IPFIX
) {
957 VLOG_WARN_RL(&rl
, "invalid user cookie of type %"PRIu16
958 " and size %"PRIuSIZE
, cookie
.type
, userdata_len
);
963 /* Calculates slow path actions for 'xout'. 'buf' must statically be
964 * initialized with at least 128 bytes of space. */
966 compose_slow_path(struct udpif
*udpif
, struct xlate_out
*xout
,
967 const struct flow
*flow
, odp_port_t odp_in_port
,
970 union user_action_cookie cookie
;
974 cookie
.type
= USER_ACTION_COOKIE_SLOW_PATH
;
975 cookie
.slow_path
.unused
= 0;
976 cookie
.slow_path
.reason
= xout
->slow
;
978 port
= xout
->slow
& (SLOW_CFM
| SLOW_BFD
| SLOW_LACP
| SLOW_STP
)
981 pid
= dpif_port_get_pid(udpif
->dpif
, port
, flow_hash_5tuple(flow
, 0));
982 odp_put_userspace_action(pid
, &cookie
, sizeof cookie
.slow_path
,
983 ODPP_NONE
, false, buf
);
986 /* If there is no error, the upcall must be destroyed with upcall_uninit()
987 * before quiescing, as the referred objects are guaranteed to exist only
988 * until the calling thread quiesces. Otherwise, do not call upcall_uninit()
989 * since the 'upcall->put_actions' remains uninitialized. */
991 upcall_receive(struct upcall
*upcall
, const struct dpif_backer
*backer
,
992 const struct dp_packet
*packet
, enum dpif_upcall_type type
,
993 const struct nlattr
*userdata
, const struct flow
*flow
,
994 const unsigned int mru
,
995 const ovs_u128
*ufid
, const unsigned pmd_id
)
999 error
= xlate_lookup(backer
, flow
, &upcall
->ofproto
, &upcall
->ipfix
,
1000 &upcall
->sflow
, NULL
, &upcall
->in_port
);
1005 upcall
->recirc
= NULL
;
1006 upcall
->have_recirc_ref
= false;
1007 upcall
->flow
= flow
;
1008 upcall
->packet
= packet
;
1009 upcall
->ufid
= ufid
;
1010 upcall
->pmd_id
= pmd_id
;
1011 upcall
->type
= type
;
1012 upcall
->userdata
= userdata
;
1013 ofpbuf_use_stub(&upcall
->odp_actions
, upcall
->odp_actions_stub
,
1014 sizeof upcall
->odp_actions_stub
);
1015 ofpbuf_init(&upcall
->put_actions
, 0);
1017 upcall
->xout_initialized
= false;
1018 upcall
->vsp_adjusted
= false;
1019 upcall
->ukey_persists
= false;
1021 upcall
->ukey
= NULL
;
1023 upcall
->key_len
= 0;
1026 upcall
->out_tun_key
= NULL
;
1027 upcall
->actions
= NULL
;
1033 upcall_xlate(struct udpif
*udpif
, struct upcall
*upcall
,
1034 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1036 struct dpif_flow_stats stats
;
1037 struct xlate_in xin
;
1039 stats
.n_packets
= 1;
1040 stats
.n_bytes
= dp_packet_size(upcall
->packet
);
1041 stats
.used
= time_msec();
1042 stats
.tcp_flags
= ntohs(upcall
->flow
->tcp_flags
);
1044 xlate_in_init(&xin
, upcall
->ofproto
, upcall
->flow
, upcall
->in_port
, NULL
,
1045 stats
.tcp_flags
, upcall
->packet
, wc
, odp_actions
);
1047 if (upcall
->type
== DPIF_UC_MISS
) {
1048 xin
.resubmit_stats
= &stats
;
1051 /* We may install a datapath flow only if we get a reference to the
1052 * recirculation context (otherwise we could have recirculation
1053 * upcalls using recirculation ID for which no context can be
1054 * found). We may still execute the flow's actions even if we
1055 * don't install the flow. */
1056 upcall
->recirc
= xin
.recirc
;
1057 upcall
->have_recirc_ref
= recirc_id_node_try_ref_rcu(xin
.recirc
);
1060 /* For non-miss upcalls, we are either executing actions (one of which
1061 * is an userspace action) for an upcall, in which case the stats have
1062 * already been taken care of, or there's a flow in the datapath which
1063 * this packet was accounted to. Presumably the revalidators will deal
1064 * with pushing its stats eventually. */
1067 upcall
->dump_seq
= seq_read(udpif
->dump_seq
);
1068 upcall
->reval_seq
= seq_read(udpif
->reval_seq
);
1069 xlate_actions(&xin
, &upcall
->xout
);
1070 upcall
->xout_initialized
= true;
1072 /* Special case for fail-open mode.
1074 * If we are in fail-open mode, but we are connected to a controller too,
1075 * then we should send the packet up to the controller in the hope that it
1076 * will try to set up a flow and thereby allow us to exit fail-open.
1078 * See the top-level comment in fail-open.c for more information.
1080 * Copy packets before they are modified by execution. */
1081 if (upcall
->xout
.fail_open
) {
1082 const struct dp_packet
*packet
= upcall
->packet
;
1083 struct ofproto_packet_in
*pin
;
1085 pin
= xmalloc(sizeof *pin
);
1086 pin
->up
.packet
= xmemdup(dp_packet_data(packet
), dp_packet_size(packet
));
1087 pin
->up
.packet_len
= dp_packet_size(packet
);
1088 pin
->up
.reason
= OFPR_NO_MATCH
;
1089 pin
->up
.table_id
= 0;
1090 pin
->up
.cookie
= OVS_BE64_MAX
;
1091 flow_get_metadata(upcall
->flow
, &pin
->up
.flow_metadata
);
1092 pin
->send_len
= 0; /* Not used for flow table misses. */
1093 pin
->miss_type
= OFPROTO_PACKET_IN_NO_MISS
;
1094 ofproto_dpif_send_packet_in(upcall
->ofproto
, pin
);
1097 if (!upcall
->xout
.slow
) {
1098 ofpbuf_use_const(&upcall
->put_actions
,
1099 odp_actions
->data
, odp_actions
->size
);
1101 ofpbuf_init(&upcall
->put_actions
, 0);
1102 compose_slow_path(udpif
, &upcall
->xout
, upcall
->flow
,
1103 upcall
->flow
->in_port
.odp_port
,
1104 &upcall
->put_actions
);
1107 /* This function is also called for slow-pathed flows. As we are only
1108 * going to create new datapath flows for actual datapath misses, there is
1109 * no point in creating a ukey otherwise. */
1110 if (upcall
->type
== DPIF_UC_MISS
) {
1111 upcall
->ukey
= ukey_create_from_upcall(upcall
, wc
);
1116 upcall_uninit(struct upcall
*upcall
)
1119 if (upcall
->xout_initialized
) {
1120 xlate_out_uninit(&upcall
->xout
);
1122 ofpbuf_uninit(&upcall
->odp_actions
);
1123 ofpbuf_uninit(&upcall
->put_actions
);
1125 if (!upcall
->ukey_persists
) {
1126 ukey_delete__(upcall
->ukey
);
1128 } else if (upcall
->have_recirc_ref
) {
1129 /* The reference was transferred to the ukey if one was created. */
1130 recirc_id_node_unref(upcall
->recirc
);
1136 upcall_cb(const struct dp_packet
*packet
, const struct flow
*flow
, ovs_u128
*ufid
,
1137 unsigned pmd_id
, enum dpif_upcall_type type
,
1138 const struct nlattr
*userdata
, struct ofpbuf
*actions
,
1139 struct flow_wildcards
*wc
, struct ofpbuf
*put_actions
, void *aux
)
1141 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 1);
1142 struct udpif
*udpif
= aux
;
1143 unsigned int flow_limit
;
1144 struct upcall upcall
;
1148 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1149 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
1151 error
= upcall_receive(&upcall
, udpif
->backer
, packet
, type
, userdata
,
1152 flow
, 0, ufid
, pmd_id
);
1157 error
= process_upcall(udpif
, &upcall
, actions
, wc
);
1162 if (upcall
.xout
.slow
&& put_actions
) {
1163 ofpbuf_put(put_actions
, upcall
.put_actions
.data
,
1164 upcall
.put_actions
.size
);
1167 if (OVS_UNLIKELY(!megaflow
)) {
1168 flow_wildcards_init_for_packet(wc
, flow
);
1171 if (udpif_get_n_flows(udpif
) >= flow_limit
) {
1172 VLOG_WARN_RL(&rl
, "upcall_cb failure: datapath flow limit reached");
1177 /* Prevent miss flow installation if the key has recirculation ID but we
1178 * were not able to get a reference on it. */
1179 if (type
== DPIF_UC_MISS
&& upcall
.recirc
&& !upcall
.have_recirc_ref
) {
1180 VLOG_WARN_RL(&rl
, "upcall_cb failure: no reference for recirc flow");
1185 if (upcall
.ukey
&& !ukey_install(udpif
, upcall
.ukey
)) {
1186 VLOG_WARN_RL(&rl
, "upcall_cb failure: ukey installation fails");
1191 upcall
.ukey_persists
= true;
1193 upcall_uninit(&upcall
);
1198 process_upcall(struct udpif
*udpif
, struct upcall
*upcall
,
1199 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1201 const struct nlattr
*userdata
= upcall
->userdata
;
1202 const struct dp_packet
*packet
= upcall
->packet
;
1203 const struct flow
*flow
= upcall
->flow
;
1205 switch (classify_upcall(upcall
->type
, userdata
)) {
1207 upcall_xlate(udpif
, upcall
, odp_actions
, wc
);
1211 if (upcall
->sflow
) {
1212 union user_action_cookie cookie
;
1213 const struct nlattr
*actions
;
1214 size_t actions_len
= 0;
1215 struct dpif_sflow_actions sflow_actions
;
1216 memset(&sflow_actions
, 0, sizeof sflow_actions
);
1217 memset(&cookie
, 0, sizeof cookie
);
1218 memcpy(&cookie
, nl_attr_get(userdata
), sizeof cookie
.sflow
);
1219 if (upcall
->actions
) {
1220 /* Actions were passed up from datapath. */
1221 actions
= nl_attr_get(upcall
->actions
);
1222 actions_len
= nl_attr_get_size(upcall
->actions
);
1223 if (actions
&& actions_len
) {
1224 dpif_sflow_read_actions(flow
, actions
, actions_len
,
1228 if (actions_len
== 0) {
1229 /* Lookup actions in userspace cache. */
1230 struct udpif_key
*ukey
= ukey_lookup(udpif
, upcall
->ufid
);
1232 ukey_get_actions(ukey
, &actions
, &actions_len
);
1233 dpif_sflow_read_actions(flow
, actions
, actions_len
,
1237 dpif_sflow_received(upcall
->sflow
, packet
, flow
,
1238 flow
->in_port
.odp_port
, &cookie
,
1239 actions_len
> 0 ? &sflow_actions
: NULL
);
1244 if (upcall
->ipfix
) {
1245 union user_action_cookie cookie
;
1246 struct flow_tnl output_tunnel_key
;
1248 memset(&cookie
, 0, sizeof cookie
);
1249 memcpy(&cookie
, nl_attr_get(userdata
), sizeof cookie
.ipfix
);
1251 if (upcall
->out_tun_key
) {
1252 odp_tun_key_from_attr(upcall
->out_tun_key
, false,
1253 &output_tunnel_key
);
1255 dpif_ipfix_bridge_sample(upcall
->ipfix
, packet
, flow
,
1256 flow
->in_port
.odp_port
,
1257 cookie
.ipfix
.output_odp_port
,
1258 upcall
->out_tun_key
?
1259 &output_tunnel_key
: NULL
);
1263 case FLOW_SAMPLE_UPCALL
:
1264 if (upcall
->ipfix
) {
1265 union user_action_cookie cookie
;
1267 memset(&cookie
, 0, sizeof cookie
);
1268 memcpy(&cookie
, nl_attr_get(userdata
), sizeof cookie
.flow_sample
);
1270 /* The flow reflects exactly the contents of the packet.
1271 * Sample the packet using it. */
1272 dpif_ipfix_flow_sample(upcall
->ipfix
, packet
, flow
,
1273 cookie
.flow_sample
.collector_set_id
,
1274 cookie
.flow_sample
.probability
,
1275 cookie
.flow_sample
.obs_domain_id
,
1276 cookie
.flow_sample
.obs_point_id
);
1288 handle_upcalls(struct udpif
*udpif
, struct upcall
*upcalls
,
1291 struct dpif_op
*opsp
[UPCALL_MAX_BATCH
* 2];
1292 struct ukey_op ops
[UPCALL_MAX_BATCH
* 2];
1293 unsigned int flow_limit
;
1294 size_t n_ops
, n_opsp
, i
;
1297 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
1299 may_put
= udpif_get_n_flows(udpif
) < flow_limit
;
1301 /* Handle the packets individually in order of arrival.
1303 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what
1304 * processes received packets for these protocols.
1306 * - For SLOW_CONTROLLER, translation sends the packet to the OpenFlow
1309 * The loop fills 'ops' with an array of operations to execute in the
1312 for (i
= 0; i
< n_upcalls
; i
++) {
1313 struct upcall
*upcall
= &upcalls
[i
];
1314 const struct dp_packet
*packet
= upcall
->packet
;
1317 if (upcall
->vsp_adjusted
) {
1318 /* This packet was received on a VLAN splinter port. We added a
1319 * VLAN to the packet to make the packet resemble the flow, but the
1320 * actions were composed assuming that the packet contained no
1321 * VLAN. So, we must remove the VLAN header from the packet before
1322 * trying to execute the actions. */
1323 if (upcall
->odp_actions
.size
) {
1324 eth_pop_vlan(CONST_CAST(struct dp_packet
*, upcall
->packet
));
1327 /* Remove the flow vlan tags inserted by vlan splinter logic
1328 * to ensure megaflow masks generated match the data path flow. */
1329 CONST_CAST(struct flow
*, upcall
->flow
)->vlan_tci
= 0;
1332 /* Do not install a flow into the datapath if:
1334 * - The datapath already has too many flows.
1336 * - We received this packet via some flow installed in the kernel
1339 * - Upcall was a recirculation but we do not have a reference to
1340 * to the recirculation ID. */
1341 if (may_put
&& upcall
->type
== DPIF_UC_MISS
&&
1342 (!upcall
->recirc
|| upcall
->have_recirc_ref
)) {
1343 struct udpif_key
*ukey
= upcall
->ukey
;
1345 upcall
->ukey_persists
= true;
1349 op
->dop
.type
= DPIF_OP_FLOW_PUT
;
1350 op
->dop
.u
.flow_put
.flags
= DPIF_FP_CREATE
;
1351 op
->dop
.u
.flow_put
.key
= ukey
->key
;
1352 op
->dop
.u
.flow_put
.key_len
= ukey
->key_len
;
1353 op
->dop
.u
.flow_put
.mask
= ukey
->mask
;
1354 op
->dop
.u
.flow_put
.mask_len
= ukey
->mask_len
;
1355 op
->dop
.u
.flow_put
.ufid
= upcall
->ufid
;
1356 op
->dop
.u
.flow_put
.stats
= NULL
;
1357 ukey_get_actions(ukey
, &op
->dop
.u
.flow_put
.actions
,
1358 &op
->dop
.u
.flow_put
.actions_len
);
1361 if (upcall
->odp_actions
.size
) {
1364 op
->dop
.type
= DPIF_OP_EXECUTE
;
1365 op
->dop
.u
.execute
.packet
= CONST_CAST(struct dp_packet
*, packet
);
1366 odp_key_to_pkt_metadata(upcall
->key
, upcall
->key_len
,
1367 &op
->dop
.u
.execute
.packet
->md
);
1368 op
->dop
.u
.execute
.actions
= upcall
->odp_actions
.data
;
1369 op
->dop
.u
.execute
.actions_len
= upcall
->odp_actions
.size
;
1370 op
->dop
.u
.execute
.needs_help
= (upcall
->xout
.slow
& SLOW_ACTION
) != 0;
1371 op
->dop
.u
.execute
.probe
= false;
1372 op
->dop
.u
.execute
.mtu
= upcall
->mru
;
1378 * We install ukeys before installing the flows, locking them for exclusive
1379 * access by this thread for the period of installation. This ensures that
1380 * other threads won't attempt to delete the flows as we are creating them.
1383 for (i
= 0; i
< n_ops
; i
++) {
1384 struct udpif_key
*ukey
= ops
[i
].ukey
;
1387 /* If we can't install the ukey, don't install the flow. */
1388 if (!ukey_install_start(udpif
, ukey
)) {
1389 ukey_delete__(ukey
);
1394 opsp
[n_opsp
++] = &ops
[i
].dop
;
1396 dpif_operate(udpif
->dpif
, opsp
, n_opsp
);
1397 for (i
= 0; i
< n_ops
; i
++) {
1399 ukey_install_finish(ops
[i
].ukey
, ops
[i
].dop
.error
);
1405 get_ufid_hash(const ovs_u128
*ufid
)
1407 return ufid
->u32
[0];
1410 static struct udpif_key
*
1411 ukey_lookup(struct udpif
*udpif
, const ovs_u128
*ufid
)
1413 struct udpif_key
*ukey
;
1414 int idx
= get_ufid_hash(ufid
) % N_UMAPS
;
1415 struct cmap
*cmap
= &udpif
->ukeys
[idx
].cmap
;
1417 CMAP_FOR_EACH_WITH_HASH (ukey
, cmap_node
, get_ufid_hash(ufid
), cmap
) {
1418 if (ovs_u128_equals(&ukey
->ufid
, ufid
)) {
1425 /* Provides safe lockless access of RCU protected 'ukey->actions'. Callers may
1426 * alternatively access the field directly if they take 'ukey->mutex'. */
1428 ukey_get_actions(struct udpif_key
*ukey
, const struct nlattr
**actions
, size_t *size
)
1430 const struct ofpbuf
*buf
= ovsrcu_get(struct ofpbuf
*, &ukey
->actions
);
1431 *actions
= buf
->data
;
1436 ukey_set_actions(struct udpif_key
*ukey
, const struct ofpbuf
*actions
)
1438 ovsrcu_postpone(ofpbuf_delete
,
1439 ovsrcu_get_protected(struct ofpbuf
*, &ukey
->actions
));
1440 ovsrcu_set(&ukey
->actions
, ofpbuf_clone(actions
));
1443 static struct udpif_key
*
1444 ukey_create__(const struct nlattr
*key
, size_t key_len
,
1445 const struct nlattr
*mask
, size_t mask_len
,
1446 bool ufid_present
, const ovs_u128
*ufid
,
1447 const unsigned pmd_id
, const struct ofpbuf
*actions
,
1448 uint64_t dump_seq
, uint64_t reval_seq
, long long int used
,
1449 const struct recirc_id_node
*key_recirc
, struct xlate_out
*xout
)
1450 OVS_NO_THREAD_SAFETY_ANALYSIS
1452 unsigned n_recircs
= (key_recirc
? 1 : 0) + (xout
? xout
->n_recircs
: 0);
1453 struct udpif_key
*ukey
= xmalloc(sizeof *ukey
+
1454 n_recircs
* sizeof *ukey
->recircs
);
1456 memcpy(&ukey
->keybuf
, key
, key_len
);
1457 ukey
->key
= &ukey
->keybuf
.nla
;
1458 ukey
->key_len
= key_len
;
1459 memcpy(&ukey
->maskbuf
, mask
, mask_len
);
1460 ukey
->mask
= &ukey
->maskbuf
.nla
;
1461 ukey
->mask_len
= mask_len
;
1462 ukey
->ufid_present
= ufid_present
;
1464 ukey
->pmd_id
= pmd_id
;
1465 ukey
->hash
= get_ufid_hash(&ukey
->ufid
);
1467 ovsrcu_init(&ukey
->actions
, NULL
);
1468 ukey_set_actions(ukey
, actions
);
1470 ovs_mutex_init(&ukey
->mutex
);
1471 ukey
->dump_seq
= dump_seq
;
1472 ukey
->reval_seq
= reval_seq
;
1473 ukey
->flow_exists
= false;
1474 ukey
->created
= time_msec();
1475 memset(&ukey
->stats
, 0, sizeof ukey
->stats
);
1476 ukey
->stats
.used
= used
;
1477 ukey
->xcache
= NULL
;
1479 ukey
->n_recircs
= n_recircs
;
1481 ukey
->recircs
[0] = key_recirc
->id
;
1483 if (xout
&& xout
->n_recircs
) {
1484 const uint32_t *act_recircs
= xlate_out_get_recircs(xout
);
1486 memcpy(ukey
->recircs
+ (key_recirc
? 1 : 0), act_recircs
,
1487 xout
->n_recircs
* sizeof *ukey
->recircs
);
1488 xlate_out_take_recircs(xout
);
1493 static struct udpif_key
*
1494 ukey_create_from_upcall(struct upcall
*upcall
, struct flow_wildcards
*wc
)
1496 struct odputil_keybuf keystub
, maskstub
;
1497 struct ofpbuf keybuf
, maskbuf
;
1499 struct odp_flow_key_parms odp_parms
= {
1500 .flow
= upcall
->flow
,
1504 odp_parms
.support
= ofproto_dpif_get_support(upcall
->ofproto
)->odp
;
1505 if (upcall
->key_len
) {
1506 ofpbuf_use_const(&keybuf
, upcall
->key
, upcall
->key_len
);
1508 /* dpif-netdev doesn't provide a netlink-formatted flow key in the
1509 * upcall, so convert the upcall's flow here. */
1510 ofpbuf_use_stack(&keybuf
, &keystub
, sizeof keystub
);
1511 odp_parms
.odp_in_port
= upcall
->flow
->in_port
.odp_port
;
1512 odp_flow_key_from_flow(&odp_parms
, &keybuf
);
1515 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1516 ofpbuf_use_stack(&maskbuf
, &maskstub
, sizeof maskstub
);
1518 odp_parms
.odp_in_port
= ODPP_NONE
;
1519 odp_parms
.key_buf
= &keybuf
;
1521 odp_flow_key_from_mask(&odp_parms
, &maskbuf
);
1524 return ukey_create__(keybuf
.data
, keybuf
.size
, maskbuf
.data
, maskbuf
.size
,
1525 true, upcall
->ufid
, upcall
->pmd_id
,
1526 &upcall
->put_actions
, upcall
->dump_seq
,
1527 upcall
->reval_seq
, 0,
1528 upcall
->have_recirc_ref
? upcall
->recirc
: NULL
,
1533 ukey_create_from_dpif_flow(const struct udpif
*udpif
,
1534 const struct dpif_flow
*flow
,
1535 struct udpif_key
**ukey
)
1537 struct dpif_flow full_flow
;
1538 struct ofpbuf actions
;
1539 uint64_t dump_seq
, reval_seq
;
1540 uint64_t stub
[DPIF_FLOW_BUFSIZE
/ 8];
1541 const struct nlattr
*a
;
1544 if (!flow
->key_len
|| !flow
->actions_len
) {
1548 /* If the key or actions were not provided by the datapath, fetch the
1550 ofpbuf_use_stack(&buf
, &stub
, sizeof stub
);
1551 err
= dpif_flow_get(udpif
->dpif
, NULL
, 0, &flow
->ufid
,
1552 flow
->pmd_id
, &buf
, &full_flow
);
1559 /* Check the flow actions for recirculation action. As recirculation
1560 * relies on OVS userspace internal state, we need to delete all old
1561 * datapath flows with recirculation upon OVS restart. */
1562 NL_ATTR_FOR_EACH_UNSAFE (a
, left
, flow
->actions
, flow
->actions_len
) {
1563 if (nl_attr_type(a
) == OVS_ACTION_ATTR_RECIRC
) {
1568 dump_seq
= seq_read(udpif
->dump_seq
);
1569 reval_seq
= seq_read(udpif
->reval_seq
);
1570 ofpbuf_use_const(&actions
, &flow
->actions
, flow
->actions_len
);
1571 *ukey
= ukey_create__(flow
->key
, flow
->key_len
,
1572 flow
->mask
, flow
->mask_len
, flow
->ufid_present
,
1573 &flow
->ufid
, flow
->pmd_id
, &actions
, dump_seq
,
1574 reval_seq
, flow
->stats
.used
, NULL
, NULL
);
1579 /* Attempts to insert a ukey into the shared ukey maps.
1581 * On success, returns true, installs the ukey and returns it in a locked
1582 * state. Otherwise, returns false. */
1584 ukey_install_start(struct udpif
*udpif
, struct udpif_key
*new_ukey
)
1585 OVS_TRY_LOCK(true, new_ukey
->mutex
)
1588 struct udpif_key
*old_ukey
;
1590 bool locked
= false;
1592 idx
= new_ukey
->hash
% N_UMAPS
;
1593 umap
= &udpif
->ukeys
[idx
];
1594 ovs_mutex_lock(&umap
->mutex
);
1595 old_ukey
= ukey_lookup(udpif
, &new_ukey
->ufid
);
1597 /* Uncommon case: A ukey is already installed with the same UFID. */
1598 if (old_ukey
->key_len
== new_ukey
->key_len
1599 && !memcmp(old_ukey
->key
, new_ukey
->key
, new_ukey
->key_len
)) {
1600 COVERAGE_INC(handler_duplicate_upcall
);
1602 struct ds ds
= DS_EMPTY_INITIALIZER
;
1604 odp_format_ufid(&old_ukey
->ufid
, &ds
);
1605 ds_put_cstr(&ds
, " ");
1606 odp_flow_key_format(old_ukey
->key
, old_ukey
->key_len
, &ds
);
1607 ds_put_cstr(&ds
, "\n");
1608 odp_format_ufid(&new_ukey
->ufid
, &ds
);
1609 ds_put_cstr(&ds
, " ");
1610 odp_flow_key_format(new_ukey
->key
, new_ukey
->key_len
, &ds
);
1612 VLOG_WARN_RL(&rl
, "Conflicting ukey for flows:\n%s", ds_cstr(&ds
));
1616 ovs_mutex_lock(&new_ukey
->mutex
);
1617 cmap_insert(&umap
->cmap
, &new_ukey
->cmap_node
, new_ukey
->hash
);
1620 ovs_mutex_unlock(&umap
->mutex
);
1626 ukey_install_finish__(struct udpif_key
*ukey
) OVS_REQUIRES(ukey
->mutex
)
1628 ukey
->flow_exists
= true;
1632 ukey_install_finish(struct udpif_key
*ukey
, int error
)
1633 OVS_RELEASES(ukey
->mutex
)
1636 ukey_install_finish__(ukey
);
1638 ovs_mutex_unlock(&ukey
->mutex
);
1644 ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
)
1646 /* The usual way to keep 'ukey->flow_exists' in sync with the datapath is
1647 * to call ukey_install_start(), install the corresponding datapath flow,
1648 * then call ukey_install_finish(). The netdev interface using upcall_cb()
1649 * doesn't provide a function to separately finish the flow installation,
1650 * so we perform the operations together here.
1652 * This is fine currently, as revalidator threads will only delete this
1653 * ukey during revalidator_sweep() and only if the dump_seq is mismatched.
1654 * It is unlikely for a revalidator thread to advance dump_seq and reach
1655 * the next GC phase between ukey creation and flow installation. */
1656 return ukey_install_start(udpif
, ukey
) && ukey_install_finish(ukey
, 0);
1659 /* Searches for a ukey in 'udpif->ukeys' that matches 'flow' and attempts to
1660 * lock the ukey. If the ukey does not exist, create it.
1662 * Returns 0 on success, setting *result to the matching ukey and returning it
1663 * in a locked state. Otherwise, returns an errno and clears *result. EBUSY
1664 * indicates that another thread is handling this flow. Other errors indicate
1665 * an unexpected condition creating a new ukey.
1667 * *error is an output parameter provided to appease the threadsafety analyser,
1668 * and its value matches the return value. */
1670 ukey_acquire(struct udpif
*udpif
, const struct dpif_flow
*flow
,
1671 struct udpif_key
**result
, int *error
)
1672 OVS_TRY_LOCK(0, (*result
)->mutex
)
1674 struct udpif_key
*ukey
;
1677 ukey
= ukey_lookup(udpif
, &flow
->ufid
);
1679 retval
= ovs_mutex_trylock(&ukey
->mutex
);
1681 /* Usually we try to avoid installing flows from revalidator threads,
1682 * because locking on a umap may cause handler threads to block.
1683 * However there are certain cases, like when ovs-vswitchd is
1684 * restarted, where it is desirable to handle flows that exist in the
1685 * datapath gracefully (ie, don't just clear the datapath). */
1688 retval
= ukey_create_from_dpif_flow(udpif
, flow
, &ukey
);
1692 install
= ukey_install_start(udpif
, ukey
);
1694 ukey_install_finish__(ukey
);
1697 ukey_delete__(ukey
);
1713 ukey_delete__(struct udpif_key
*ukey
)
1714 OVS_NO_THREAD_SAFETY_ANALYSIS
1717 for (int i
= 0; i
< ukey
->n_recircs
; i
++) {
1718 recirc_free_id(ukey
->recircs
[i
]);
1720 xlate_cache_delete(ukey
->xcache
);
1721 ofpbuf_delete(ovsrcu_get(struct ofpbuf
*, &ukey
->actions
));
1722 ovs_mutex_destroy(&ukey
->mutex
);
1728 ukey_delete(struct umap
*umap
, struct udpif_key
*ukey
)
1729 OVS_REQUIRES(umap
->mutex
)
1731 cmap_remove(&umap
->cmap
, &ukey
->cmap_node
, ukey
->hash
);
1732 ovsrcu_postpone(ukey_delete__
, ukey
);
1736 should_revalidate(const struct udpif
*udpif
, uint64_t packets
,
1739 long long int metric
, now
, duration
;
1741 if (udpif
->dump_duration
< 200) {
1742 /* We are likely to handle full revalidation for the flows. */
1746 /* Calculate the mean time between seeing these packets. If this
1747 * exceeds the threshold, then delete the flow rather than performing
1748 * costly revalidation for flows that aren't being hit frequently.
1750 * This is targeted at situations where the dump_duration is high (~1s),
1751 * and revalidation is triggered by a call to udpif_revalidate(). In
1752 * these situations, revalidation of all flows causes fluctuations in the
1753 * flow_limit due to the interaction with the dump_duration and max_idle.
1754 * This tends to result in deletion of low-throughput flows anyway, so
1755 * skip the revalidation and just delete those flows. */
1756 packets
= MAX(packets
, 1);
1757 now
= MAX(used
, time_msec());
1758 duration
= now
- used
;
1759 metric
= duration
/ packets
;
1762 /* The flow is receiving more than ~5pps, so keep it. */
1768 /* Verifies that the datapath actions of 'ukey' are still correct, and pushes
1771 * Returns a recommended action for 'ukey', options include:
1772 * UKEY_DELETE The ukey should be deleted.
1773 * UKEY_KEEP The ukey is fine as is.
1774 * UKEY_MODIFY The ukey's actions should be changed but is otherwise
1775 * fine. Callers should change the actions to those found
1776 * in the caller supplied 'odp_actions' buffer. */
1777 static enum reval_result
1778 revalidate_ukey(struct udpif
*udpif
, struct udpif_key
*ukey
,
1779 const struct dpif_flow_stats
*stats
,
1780 struct ofpbuf
*odp_actions
, uint64_t reval_seq
)
1781 OVS_REQUIRES(ukey
->mutex
)
1783 struct xlate_out xout
, *xoutp
;
1784 struct netflow
*netflow
;
1785 struct ofproto_dpif
*ofproto
;
1786 struct dpif_flow_stats push
;
1788 struct flow_wildcards dp_mask
, wc
;
1789 enum reval_result result
;
1790 ofp_port_t ofp_in_port
;
1791 struct xlate_in xin
;
1792 long long int last_used
;
1794 bool need_revalidate
;
1796 result
= UKEY_DELETE
;
1800 ofpbuf_clear(odp_actions
);
1801 need_revalidate
= (ukey
->reval_seq
!= reval_seq
);
1802 last_used
= ukey
->stats
.used
;
1803 push
.used
= stats
->used
;
1804 push
.tcp_flags
= stats
->tcp_flags
;
1805 push
.n_packets
= (stats
->n_packets
> ukey
->stats
.n_packets
1806 ? stats
->n_packets
- ukey
->stats
.n_packets
1808 push
.n_bytes
= (stats
->n_bytes
> ukey
->stats
.n_bytes
1809 ? stats
->n_bytes
- ukey
->stats
.n_bytes
1812 if (need_revalidate
&& last_used
1813 && !should_revalidate(udpif
, push
.n_packets
, last_used
)) {
1817 /* We will push the stats, so update the ukey stats cache. */
1818 ukey
->stats
= *stats
;
1819 if (!push
.n_packets
&& !need_revalidate
) {
1824 if (ukey
->xcache
&& !need_revalidate
) {
1825 xlate_push_stats(ukey
->xcache
, &push
);
1830 if (odp_flow_key_to_flow(ukey
->key
, ukey
->key_len
, &flow
)
1835 error
= xlate_lookup(udpif
->backer
, &flow
, &ofproto
, NULL
, NULL
, &netflow
,
1841 if (need_revalidate
) {
1842 xlate_cache_clear(ukey
->xcache
);
1844 if (!ukey
->xcache
) {
1845 ukey
->xcache
= xlate_cache_new();
1848 xlate_in_init(&xin
, ofproto
, &flow
, ofp_in_port
, NULL
, push
.tcp_flags
,
1849 NULL
, need_revalidate
? &wc
: NULL
, odp_actions
);
1850 if (push
.n_packets
) {
1851 xin
.resubmit_stats
= &push
;
1852 xin
.may_learn
= true;
1854 xin
.xcache
= ukey
->xcache
;
1855 xlate_actions(&xin
, &xout
);
1858 if (!need_revalidate
) {
1864 ofpbuf_clear(odp_actions
);
1865 compose_slow_path(udpif
, &xout
, &flow
, flow
.in_port
.odp_port
,
1869 if (odp_flow_key_to_mask(ukey
->mask
, ukey
->mask_len
, ukey
->key
,
1870 ukey
->key_len
, &dp_mask
.masks
, &flow
)
1875 /* Do not modify if any bit is wildcarded by the installed datapath flow,
1876 * but not the newly revalidated wildcard mask (wc), i.e., if revalidation
1877 * tells that the datapath flow is now too generic and must be narrowed
1878 * down. Note that we do not know if the datapath has ignored any of the
1879 * wildcarded bits, so we may be overtly conservative here. */
1880 if (flow_wildcards_has_extra(&dp_mask
, &wc
)) {
1884 if (!ofpbuf_equal(odp_actions
,
1885 ovsrcu_get(struct ofpbuf
*, &ukey
->actions
))) {
1886 /* The datapath mask was OK, but the actions seem to have changed.
1887 * Let's modify it in place. */
1888 result
= UKEY_MODIFY
;
1895 if (result
!= UKEY_DELETE
) {
1896 ukey
->reval_seq
= reval_seq
;
1898 if (netflow
&& result
== UKEY_DELETE
) {
1899 netflow_flow_clear(netflow
, &flow
);
1901 xlate_out_uninit(xoutp
);
1906 delete_op_init__(struct udpif
*udpif
, struct ukey_op
*op
,
1907 const struct dpif_flow
*flow
)
1910 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
1911 op
->dop
.u
.flow_del
.key
= flow
->key
;
1912 op
->dop
.u
.flow_del
.key_len
= flow
->key_len
;
1913 op
->dop
.u
.flow_del
.ufid
= flow
->ufid_present
? &flow
->ufid
: NULL
;
1914 op
->dop
.u
.flow_del
.pmd_id
= flow
->pmd_id
;
1915 op
->dop
.u
.flow_del
.stats
= &op
->stats
;
1916 op
->dop
.u
.flow_del
.terse
= udpif_use_ufid(udpif
);
1920 delete_op_init(struct udpif
*udpif
, struct ukey_op
*op
, struct udpif_key
*ukey
)
1923 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
1924 op
->dop
.u
.flow_del
.key
= ukey
->key
;
1925 op
->dop
.u
.flow_del
.key_len
= ukey
->key_len
;
1926 op
->dop
.u
.flow_del
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
1927 op
->dop
.u
.flow_del
.pmd_id
= ukey
->pmd_id
;
1928 op
->dop
.u
.flow_del
.stats
= &op
->stats
;
1929 op
->dop
.u
.flow_del
.terse
= udpif_use_ufid(udpif
);
1933 modify_op_init(struct ukey_op
*op
, struct udpif_key
*ukey
)
1936 op
->dop
.type
= DPIF_OP_FLOW_PUT
;
1937 op
->dop
.u
.flow_put
.flags
= DPIF_FP_MODIFY
;
1938 op
->dop
.u
.flow_put
.key
= ukey
->key
;
1939 op
->dop
.u
.flow_put
.key_len
= ukey
->key_len
;
1940 op
->dop
.u
.flow_put
.mask
= ukey
->mask
;
1941 op
->dop
.u
.flow_put
.mask_len
= ukey
->mask_len
;
1942 op
->dop
.u
.flow_put
.ufid
= &ukey
->ufid
;
1943 op
->dop
.u
.flow_put
.pmd_id
= ukey
->pmd_id
;
1944 op
->dop
.u
.flow_put
.stats
= NULL
;
1945 ukey_get_actions(ukey
, &op
->dop
.u
.flow_put
.actions
,
1946 &op
->dop
.u
.flow_put
.actions_len
);
1950 push_ukey_ops__(struct udpif
*udpif
, struct ukey_op
*ops
, size_t n_ops
)
1952 struct dpif_op
*opsp
[REVALIDATE_MAX_BATCH
];
1955 ovs_assert(n_ops
<= REVALIDATE_MAX_BATCH
);
1956 for (i
= 0; i
< n_ops
; i
++) {
1957 opsp
[i
] = &ops
[i
].dop
;
1959 dpif_operate(udpif
->dpif
, opsp
, n_ops
);
1961 for (i
= 0; i
< n_ops
; i
++) {
1962 struct ukey_op
*op
= &ops
[i
];
1963 struct dpif_flow_stats
*push
, *stats
, push_buf
;
1965 stats
= op
->dop
.u
.flow_del
.stats
;
1968 if (op
->dop
.type
!= DPIF_OP_FLOW_DEL
) {
1969 /* Only deleted flows need their stats pushed. */
1973 if (op
->dop
.error
) {
1974 /* flow_del error, 'stats' is unusable. */
1979 ovs_mutex_lock(&op
->ukey
->mutex
);
1980 push
->used
= MAX(stats
->used
, op
->ukey
->stats
.used
);
1981 push
->tcp_flags
= stats
->tcp_flags
| op
->ukey
->stats
.tcp_flags
;
1982 push
->n_packets
= stats
->n_packets
- op
->ukey
->stats
.n_packets
;
1983 push
->n_bytes
= stats
->n_bytes
- op
->ukey
->stats
.n_bytes
;
1984 ovs_mutex_unlock(&op
->ukey
->mutex
);
1989 if (push
->n_packets
|| netflow_exists()) {
1990 const struct nlattr
*key
= op
->dop
.u
.flow_del
.key
;
1991 size_t key_len
= op
->dop
.u
.flow_del
.key_len
;
1992 struct ofproto_dpif
*ofproto
;
1993 struct netflow
*netflow
;
1994 ofp_port_t ofp_in_port
;
1999 ovs_mutex_lock(&op
->ukey
->mutex
);
2000 if (op
->ukey
->xcache
) {
2001 xlate_push_stats(op
->ukey
->xcache
, push
);
2002 ovs_mutex_unlock(&op
->ukey
->mutex
);
2005 ovs_mutex_unlock(&op
->ukey
->mutex
);
2006 key
= op
->ukey
->key
;
2007 key_len
= op
->ukey
->key_len
;
2010 if (odp_flow_key_to_flow(key
, key_len
, &flow
)
2015 error
= xlate_lookup(udpif
->backer
, &flow
, &ofproto
, NULL
, NULL
,
2016 &netflow
, &ofp_in_port
);
2018 struct xlate_in xin
;
2020 xlate_in_init(&xin
, ofproto
, &flow
, ofp_in_port
, NULL
,
2021 push
->tcp_flags
, NULL
, NULL
, NULL
);
2022 xin
.resubmit_stats
= push
->n_packets
? push
: NULL
;
2023 xin
.may_learn
= push
->n_packets
> 0;
2024 xlate_actions_for_side_effects(&xin
);
2027 netflow_flow_clear(netflow
, &flow
);
2035 push_ukey_ops(struct udpif
*udpif
, struct umap
*umap
,
2036 struct ukey_op
*ops
, size_t n_ops
)
2040 push_ukey_ops__(udpif
, ops
, n_ops
);
2041 ovs_mutex_lock(&umap
->mutex
);
2042 for (i
= 0; i
< n_ops
; i
++) {
2043 ukey_delete(umap
, ops
[i
].ukey
);
2045 ovs_mutex_unlock(&umap
->mutex
);
2049 log_unexpected_flow(const struct dpif_flow
*flow
, int error
)
2051 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(10, 60);
2052 struct ds ds
= DS_EMPTY_INITIALIZER
;
2054 ds_put_format(&ds
, "Failed to acquire udpif_key corresponding to "
2055 "unexpected flow (%s): ", ovs_strerror(error
));
2056 odp_format_ufid(&flow
->ufid
, &ds
);
2057 VLOG_WARN_RL(&rl
, "%s", ds_cstr(&ds
));
2061 revalidate(struct revalidator
*revalidator
)
2063 uint64_t odp_actions_stub
[1024 / 8];
2064 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2066 struct udpif
*udpif
= revalidator
->udpif
;
2067 struct dpif_flow_dump_thread
*dump_thread
;
2068 uint64_t dump_seq
, reval_seq
;
2069 unsigned int flow_limit
;
2071 dump_seq
= seq_read(udpif
->dump_seq
);
2072 reval_seq
= seq_read(udpif
->reval_seq
);
2073 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2074 dump_thread
= dpif_flow_dump_thread_create(udpif
->dump
);
2076 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2079 struct dpif_flow flows
[REVALIDATE_MAX_BATCH
];
2080 const struct dpif_flow
*f
;
2083 long long int max_idle
;
2088 n_dumped
= dpif_flow_dump_next(dump_thread
, flows
, ARRAY_SIZE(flows
));
2095 /* In normal operation we want to keep flows around until they have
2096 * been idle for 'ofproto_max_idle' milliseconds. However:
2098 * - If the number of datapath flows climbs above 'flow_limit',
2099 * drop that down to 100 ms to try to bring the flows down to
2102 * - If the number of datapath flows climbs above twice
2103 * 'flow_limit', delete all the datapath flows as an emergency
2104 * measure. (We reassess this condition for the next batch of
2105 * datapath flows, so we will recover before all the flows are
2107 n_dp_flows
= udpif_get_n_flows(udpif
);
2108 kill_them_all
= n_dp_flows
> flow_limit
* 2;
2109 max_idle
= n_dp_flows
> flow_limit
? 100 : ofproto_max_idle
;
2111 for (f
= flows
; f
< &flows
[n_dumped
]; f
++) {
2112 long long int used
= f
->stats
.used
;
2113 enum reval_result result
;
2114 struct udpif_key
*ukey
;
2115 bool already_dumped
;
2118 if (ukey_acquire(udpif
, f
, &ukey
, &error
)) {
2119 if (error
== EBUSY
) {
2120 /* Another thread is processing this flow, so don't bother
2122 COVERAGE_INC(upcall_ukey_contention
);
2124 log_unexpected_flow(f
, error
);
2125 if (error
!= ENOENT
) {
2126 delete_op_init__(udpif
, &ops
[n_ops
++], f
);
2132 already_dumped
= ukey
->dump_seq
== dump_seq
;
2133 if (already_dumped
) {
2134 /* The flow has already been handled during this flow dump
2135 * operation. Skip it. */
2137 COVERAGE_INC(dumped_duplicate_flow
);
2139 COVERAGE_INC(dumped_new_flow
);
2141 ovs_mutex_unlock(&ukey
->mutex
);
2146 used
= ukey
->created
;
2148 if (kill_them_all
|| (used
&& used
< now
- max_idle
)) {
2149 result
= UKEY_DELETE
;
2151 result
= revalidate_ukey(udpif
, ukey
, &f
->stats
, &odp_actions
,
2154 ukey
->dump_seq
= dump_seq
;
2155 ukey
->flow_exists
= result
!= UKEY_DELETE
;
2157 if (result
== UKEY_DELETE
) {
2158 delete_op_init(udpif
, &ops
[n_ops
++], ukey
);
2159 } else if (result
== UKEY_MODIFY
) {
2160 ukey_set_actions(ukey
, &odp_actions
);
2161 modify_op_init(&ops
[n_ops
++], ukey
);
2163 ovs_mutex_unlock(&ukey
->mutex
);
2167 push_ukey_ops__(udpif
, ops
, n_ops
);
2171 dpif_flow_dump_thread_destroy(dump_thread
);
2172 ofpbuf_uninit(&odp_actions
);
2175 /* Pauses the 'revalidator', can only proceed after main thread
2176 * calls udpif_resume_revalidators(). */
2178 revalidator_pause(struct revalidator
*revalidator
)
2180 /* The first block is for sync'ing the pause with main thread. */
2181 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2182 /* The second block is for pausing until main thread resumes. */
2183 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2187 revalidator_sweep__(struct revalidator
*revalidator
, bool purge
)
2189 struct udpif
*udpif
;
2190 uint64_t dump_seq
, reval_seq
;
2193 udpif
= revalidator
->udpif
;
2194 dump_seq
= seq_read(udpif
->dump_seq
);
2195 reval_seq
= seq_read(udpif
->reval_seq
);
2196 slice
= revalidator
- udpif
->revalidators
;
2197 ovs_assert(slice
< udpif
->n_revalidators
);
2199 for (int i
= slice
; i
< N_UMAPS
; i
+= udpif
->n_revalidators
) {
2200 uint64_t odp_actions_stub
[1024 / 8];
2201 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2203 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2204 struct udpif_key
*ukey
;
2205 struct umap
*umap
= &udpif
->ukeys
[i
];
2208 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2209 bool flow_exists
, seq_mismatch
;
2210 enum reval_result result
;
2212 /* Handler threads could be holding a ukey lock while it installs a
2213 * new flow, so don't hang around waiting for access to it. */
2214 if (ovs_mutex_trylock(&ukey
->mutex
)) {
2217 flow_exists
= ukey
->flow_exists
;
2218 seq_mismatch
= (ukey
->dump_seq
!= dump_seq
2219 && ukey
->reval_seq
!= reval_seq
);
2222 result
= UKEY_DELETE
;
2223 } else if (!seq_mismatch
) {
2226 struct dpif_flow_stats stats
;
2227 COVERAGE_INC(revalidate_missed_dp_flow
);
2228 memset(&stats
, 0, sizeof stats
);
2229 result
= revalidate_ukey(udpif
, ukey
, &stats
, &odp_actions
,
2232 ovs_mutex_unlock(&ukey
->mutex
);
2234 if (result
== UKEY_DELETE
) {
2235 delete_op_init(udpif
, &ops
[n_ops
++], ukey
);
2236 } else if (result
== UKEY_MODIFY
) {
2237 ukey_set_actions(ukey
, &odp_actions
);
2238 modify_op_init(&ops
[n_ops
++], ukey
);
2241 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2242 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2247 ovs_mutex_lock(&umap
->mutex
);
2248 ukey_delete(umap
, ukey
);
2249 ovs_mutex_unlock(&umap
->mutex
);
2254 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2257 ofpbuf_uninit(&odp_actions
);
2263 revalidator_sweep(struct revalidator
*revalidator
)
2265 revalidator_sweep__(revalidator
, false);
2269 revalidator_purge(struct revalidator
*revalidator
)
2271 revalidator_sweep__(revalidator
, true);
2274 /* In reaction to dpif purge, purges all 'ukey's with same 'pmd_id'. */
2276 dp_purge_cb(void *aux
, unsigned pmd_id
)
2278 struct udpif
*udpif
= aux
;
2281 udpif_pause_revalidators(udpif
);
2282 for (i
= 0; i
< N_UMAPS
; i
++) {
2283 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2284 struct udpif_key
*ukey
;
2285 struct umap
*umap
= &udpif
->ukeys
[i
];
2288 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2289 if (ukey
->pmd_id
== pmd_id
) {
2290 delete_op_init(udpif
, &ops
[n_ops
++], ukey
);
2291 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2292 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2299 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2304 udpif_resume_revalidators(udpif
);
2308 upcall_unixctl_show(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2309 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2311 struct ds ds
= DS_EMPTY_INITIALIZER
;
2312 struct udpif
*udpif
;
2314 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2315 unsigned int flow_limit
;
2319 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2320 ufid_enabled
= udpif_use_ufid(udpif
);
2322 ds_put_format(&ds
, "%s:\n", dpif_name(udpif
->dpif
));
2323 ds_put_format(&ds
, "\tflows : (current %lu)"
2324 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif
),
2325 udpif
->avg_n_flows
, udpif
->max_n_flows
, flow_limit
);
2326 ds_put_format(&ds
, "\tdump duration : %lldms\n", udpif
->dump_duration
);
2327 ds_put_format(&ds
, "\tufid enabled : ");
2329 ds_put_format(&ds
, "true\n");
2331 ds_put_format(&ds
, "false\n");
2333 ds_put_char(&ds
, '\n');
2335 for (i
= 0; i
< n_revalidators
; i
++) {
2336 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
2337 int j
, elements
= 0;
2339 for (j
= i
; j
< N_UMAPS
; j
+= n_revalidators
) {
2340 elements
+= cmap_count(&udpif
->ukeys
[j
].cmap
);
2342 ds_put_format(&ds
, "\t%u: (keys %d)\n", revalidator
->id
, elements
);
2346 unixctl_command_reply(conn
, ds_cstr(&ds
));
2350 /* Disable using the megaflows.
2352 * This command is only needed for advanced debugging, so it's not
2353 * documented in the man page. */
2355 upcall_unixctl_disable_megaflows(struct unixctl_conn
*conn
,
2356 int argc OVS_UNUSED
,
2357 const char *argv
[] OVS_UNUSED
,
2358 void *aux OVS_UNUSED
)
2360 atomic_store_relaxed(&enable_megaflows
, false);
2361 udpif_flush_all_datapaths();
2362 unixctl_command_reply(conn
, "megaflows disabled");
2365 /* Re-enable using megaflows.
2367 * This command is only needed for advanced debugging, so it's not
2368 * documented in the man page. */
2370 upcall_unixctl_enable_megaflows(struct unixctl_conn
*conn
,
2371 int argc OVS_UNUSED
,
2372 const char *argv
[] OVS_UNUSED
,
2373 void *aux OVS_UNUSED
)
2375 atomic_store_relaxed(&enable_megaflows
, true);
2376 udpif_flush_all_datapaths();
2377 unixctl_command_reply(conn
, "megaflows enabled");
2380 /* Disable skipping flow attributes during flow dump.
2382 * This command is only needed for advanced debugging, so it's not
2383 * documented in the man page. */
2385 upcall_unixctl_disable_ufid(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2386 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2388 atomic_store_relaxed(&enable_ufid
, false);
2389 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID disabled");
2392 /* Re-enable skipping flow attributes during flow dump.
2394 * This command is only needed for advanced debugging, so it's not documented
2395 * in the man page. */
2397 upcall_unixctl_enable_ufid(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2398 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2400 atomic_store_relaxed(&enable_ufid
, true);
2401 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID enabled "
2402 "for supported datapaths");
2405 /* Set the flow limit.
2407 * This command is only needed for advanced debugging, so it's not
2408 * documented in the man page. */
2410 upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
,
2411 int argc OVS_UNUSED
,
2412 const char *argv
[] OVS_UNUSED
,
2413 void *aux OVS_UNUSED
)
2415 struct ds ds
= DS_EMPTY_INITIALIZER
;
2416 struct udpif
*udpif
;
2417 unsigned int flow_limit
= atoi(argv
[1]);
2419 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2420 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
2422 ds_put_format(&ds
, "set flow_limit to %u\n", flow_limit
);
2423 unixctl_command_reply(conn
, ds_cstr(&ds
));
2428 upcall_unixctl_dump_wait(struct unixctl_conn
*conn
,
2429 int argc OVS_UNUSED
,
2430 const char *argv
[] OVS_UNUSED
,
2431 void *aux OVS_UNUSED
)
2433 if (list_is_singleton(&all_udpifs
)) {
2434 struct udpif
*udpif
= NULL
;
2437 udpif
= OBJECT_CONTAINING(list_front(&all_udpifs
), udpif
, list_node
);
2438 len
= (udpif
->n_conns
+ 1) * sizeof *udpif
->conns
;
2439 udpif
->conn_seq
= seq_read(udpif
->dump_seq
);
2440 udpif
->conns
= xrealloc(udpif
->conns
, len
);
2441 udpif
->conns
[udpif
->n_conns
++] = conn
;
2443 unixctl_command_reply_error(conn
, "can't wait on multiple udpifs.");
2448 upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2449 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2451 struct udpif
*udpif
;
2453 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2456 for (n
= 0; n
< udpif
->n_revalidators
; n
++) {
2457 revalidator_purge(&udpif
->revalidators
[n
]);
2460 unixctl_command_reply(conn
, "");