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"
25 #include "lib/dpif-provider.h"
27 #include "openvswitch/dynamic-string.h"
28 #include "fail-open.h"
29 #include "guarded-list.h"
31 #include "openvswitch/list.h"
33 #include "openvswitch/ofpbuf.h"
34 #include "ofproto-dpif-ipfix.h"
35 #include "ofproto-dpif-sflow.h"
36 #include "ofproto-dpif-xlate.h"
37 #include "ofproto-dpif-xlate-cache.h"
38 #include "ofproto-dpif-trace.h"
41 #include "openvswitch/poll-loop.h"
45 #include "openvswitch/vlog.h"
46 #include "lib/netdev-provider.h"
48 #define UPCALL_MAX_BATCH 64
49 #define REVALIDATE_MAX_BATCH 50
51 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall
);
53 COVERAGE_DEFINE(dumped_duplicate_flow
);
54 COVERAGE_DEFINE(dumped_new_flow
);
55 COVERAGE_DEFINE(handler_duplicate_upcall
);
56 COVERAGE_DEFINE(upcall_ukey_contention
);
57 COVERAGE_DEFINE(upcall_ukey_replace
);
58 COVERAGE_DEFINE(revalidate_missed_dp_flow
);
60 /* A thread that reads upcalls from dpif, forwards each upcall's packet,
61 * and possibly sets up a kernel flow as a cache. */
63 struct udpif
*udpif
; /* Parent udpif. */
64 pthread_t thread
; /* Thread ID. */
65 uint32_t handler_id
; /* Handler id. */
68 /* In the absence of a multiple-writer multiple-reader datastructure for
69 * storing udpif_keys ("ukeys"), we use a large number of cmaps, each with its
70 * own lock for writing. */
71 #define N_UMAPS 512 /* per udpif. */
73 struct ovs_mutex mutex
; /* Take for writing to the following. */
74 struct cmap cmap
; /* Datapath flow keys. */
77 /* A thread that processes datapath flows, updates OpenFlow statistics, and
78 * updates or removes them if necessary.
80 * Revalidator threads operate in two phases: "dump" and "sweep". In between
81 * each phase, all revalidators sync up so that all revalidator threads are
82 * either in one phase or the other, but not a combination.
84 * During the dump phase, revalidators fetch flows from the datapath and
85 * attribute the statistics to OpenFlow rules. Each datapath flow has a
86 * corresponding ukey which caches the most recently seen statistics. If
87 * a flow needs to be deleted (for example, because it is unused over a
88 * period of time), revalidator threads may delete the flow during the
89 * dump phase. The datapath is not guaranteed to reliably dump all flows
90 * from the datapath, and there is no mapping between datapath flows to
91 * revalidators, so a particular flow may be handled by zero or more
92 * revalidators during a single dump phase. To avoid duplicate attribution
93 * of statistics, ukeys are never deleted during this phase.
95 * During the sweep phase, each revalidator takes ownership of a different
96 * slice of umaps and sweeps through all ukeys in those umaps to figure out
97 * whether they need to be deleted. During this phase, revalidators may
98 * fetch individual flows which were not dumped during the dump phase to
99 * validate them and attribute statistics.
102 struct udpif
*udpif
; /* Parent udpif. */
103 pthread_t thread
; /* Thread ID. */
104 unsigned int id
; /* ovsthread_id_self(). */
107 /* An upcall handler for ofproto_dpif.
109 * udpif keeps records of two kind of logically separate units:
114 * - An array of 'struct handler's for upcall handling and flow
120 * - Revalidation threads which read the datapath flow table and maintains
124 struct ovs_list list_node
; /* In all_udpifs list. */
126 struct dpif
*dpif
; /* Datapath handle. */
127 struct dpif_backer
*backer
; /* Opaque dpif_backer pointer. */
129 struct handler
*handlers
; /* Upcall handlers. */
132 struct revalidator
*revalidators
; /* Flow revalidators. */
133 size_t n_revalidators
;
135 struct latch exit_latch
; /* Tells child threads to exit. */
138 struct seq
*reval_seq
; /* Incremented to force revalidation. */
139 bool reval_exit
; /* Set by leader on 'exit_latch. */
140 struct ovs_barrier reval_barrier
; /* Barrier used by revalidators. */
141 struct dpif_flow_dump
*dump
; /* DPIF flow dump state. */
142 long long int dump_duration
; /* Duration of the last flow dump. */
143 struct seq
*dump_seq
; /* Increments each dump iteration. */
144 atomic_bool enable_ufid
; /* If true, skip dumping flow attrs. */
146 /* These variables provide a mechanism for the main thread to pause
147 * all revalidation without having to completely shut the threads down.
148 * 'pause_latch' is shared between the main thread and the lead
149 * revalidator thread, so when it is desirable to halt revalidation, the
150 * main thread will set the latch. 'pause' and 'pause_barrier' are shared
151 * by revalidator threads. The lead revalidator will set 'pause' when it
152 * observes the latch has been set, and this will cause all revalidator
153 * threads to wait on 'pause_barrier' at the beginning of the next
154 * revalidation round. */
155 bool pause
; /* Set by leader on 'pause_latch. */
156 struct latch pause_latch
; /* Set to force revalidators pause. */
157 struct ovs_barrier pause_barrier
; /* Barrier used to pause all */
158 /* revalidators by main thread. */
160 /* There are 'N_UMAPS' maps containing 'struct udpif_key' elements.
162 * During the flow dump phase, revalidators insert into these with a random
163 * distribution. During the garbage collection phase, each revalidator
164 * takes care of garbage collecting a slice of these maps. */
167 /* Datapath flow statistics. */
168 unsigned int max_n_flows
;
169 unsigned int avg_n_flows
;
171 /* Following fields are accessed and modified by different threads. */
172 atomic_uint flow_limit
; /* Datapath flow hard limit. */
174 /* n_flows_mutex prevents multiple threads updating these concurrently. */
175 atomic_uint n_flows
; /* Number of flows in the datapath. */
176 atomic_llong n_flows_timestamp
; /* Last time n_flows was updated. */
177 struct ovs_mutex n_flows_mutex
;
179 /* Following fields are accessed and modified only from the main thread. */
180 struct unixctl_conn
**conns
; /* Connections waiting on dump_seq. */
181 uint64_t conn_seq
; /* Corresponds to 'dump_seq' when
182 conns[n_conns-1] was stored. */
183 size_t n_conns
; /* Number of connections waiting. */
185 long long int offload_rebalance_time
; /* Time of last offload rebalance */
189 BAD_UPCALL
, /* Some kind of bug somewhere. */
190 MISS_UPCALL
, /* A flow miss. */
191 SLOW_PATH_UPCALL
, /* Slow path upcall. */
192 SFLOW_UPCALL
, /* sFlow sample. */
193 FLOW_SAMPLE_UPCALL
, /* Per-flow sampling. */
194 IPFIX_UPCALL
, /* Per-bridge sampling. */
195 CONTROLLER_UPCALL
/* Destined for the controller. */
205 struct ofproto_dpif
*ofproto
; /* Parent ofproto. */
206 const struct recirc_id_node
*recirc
; /* Recirculation context. */
207 bool have_recirc_ref
; /* Reference held on recirc ctx? */
209 /* The flow and packet are only required to be constant when using
210 * dpif-netdev. If a modification is absolutely necessary, a const cast
211 * may be used with other datapaths. */
212 const struct flow
*flow
; /* Parsed representation of the packet. */
213 enum odp_key_fitness fitness
; /* Fitness of 'flow' relative to ODP key. */
214 const ovs_u128
*ufid
; /* Unique identifier for 'flow'. */
215 unsigned pmd_id
; /* Datapath poll mode driver id. */
216 const struct dp_packet
*packet
; /* Packet associated with this upcall. */
217 ofp_port_t ofp_in_port
; /* OpenFlow in port, or OFPP_NONE. */
218 uint16_t mru
; /* If !0, Maximum receive unit of
219 fragmented IP packet */
221 enum upcall_type type
; /* Type of the upcall. */
222 const struct nlattr
*actions
; /* Flow actions in DPIF_UC_ACTION Upcalls. */
224 bool xout_initialized
; /* True if 'xout' must be uninitialized. */
225 struct xlate_out xout
; /* Result of xlate_actions(). */
226 struct ofpbuf odp_actions
; /* Datapath actions from xlate_actions(). */
227 struct flow_wildcards wc
; /* Dependencies that megaflow must match. */
228 struct ofpbuf put_actions
; /* Actions 'put' in the fastpath. */
230 struct dpif_ipfix
*ipfix
; /* IPFIX pointer or NULL. */
231 struct dpif_sflow
*sflow
; /* SFlow pointer or NULL. */
233 struct udpif_key
*ukey
; /* Revalidator flow cache. */
234 bool ukey_persists
; /* Set true to keep 'ukey' beyond the
235 lifetime of this upcall. */
237 uint64_t reval_seq
; /* udpif->reval_seq at translation time. */
239 /* Not used by the upcall callback interface. */
240 const struct nlattr
*key
; /* Datapath flow key. */
241 size_t key_len
; /* Datapath flow key length. */
242 const struct nlattr
*out_tun_key
; /* Datapath output tunnel key. */
244 struct user_action_cookie cookie
;
246 uint64_t odp_actions_stub
[1024 / 8]; /* Stub for odp_actions. */
249 /* Ukeys must transition through these states using transition_ukey(). */
252 UKEY_VISIBLE
, /* Ukey is in umap, datapath flow install is queued. */
253 UKEY_OPERATIONAL
, /* Ukey is in umap, datapath flow is installed. */
254 UKEY_EVICTING
, /* Ukey is in umap, datapath flow delete is queued. */
255 UKEY_EVICTED
, /* Ukey is in umap, datapath flow is deleted. */
256 UKEY_DELETED
, /* Ukey removed from umap, ukey free is deferred. */
258 #define N_UKEY_STATES (UKEY_DELETED + 1)
260 /* 'udpif_key's are responsible for tracking the little bit of state udpif
261 * needs to do flow expiration which can't be pulled directly from the
262 * datapath. They may be created by any handler or revalidator thread at any
263 * time, and read by any revalidator during the dump phase. They are however
264 * each owned by a single revalidator which takes care of destroying them
265 * during the garbage-collection phase.
267 * The mutex within the ukey protects some members of the ukey. The ukey
268 * itself is protected by RCU and is held within a umap in the parent udpif.
269 * Adding or removing a ukey from a umap is only safe when holding the
270 * corresponding umap lock. */
272 struct cmap_node cmap_node
; /* In parent revalidator 'ukeys' map. */
274 /* These elements are read only once created, and therefore aren't
275 * protected by a mutex. */
276 const struct nlattr
*key
; /* Datapath flow key. */
277 size_t key_len
; /* Length of 'key'. */
278 const struct nlattr
*mask
; /* Datapath flow mask. */
279 size_t mask_len
; /* Length of 'mask'. */
280 ovs_u128 ufid
; /* Unique flow identifier. */
281 bool ufid_present
; /* True if 'ufid' is in datapath. */
282 uint32_t hash
; /* Pre-computed hash for 'key'. */
283 unsigned pmd_id
; /* Datapath poll mode driver id. */
285 struct ovs_mutex mutex
; /* Guards the following. */
286 struct dpif_flow_stats stats OVS_GUARDED
; /* Last known stats.*/
287 long long int created OVS_GUARDED
; /* Estimate of creation time. */
288 uint64_t dump_seq OVS_GUARDED
; /* Tracks udpif->dump_seq. */
289 uint64_t reval_seq OVS_GUARDED
; /* Tracks udpif->reval_seq. */
290 enum ukey_state state OVS_GUARDED
; /* Tracks ukey lifetime. */
292 /* 'state' debug information. */
293 unsigned int state_thread OVS_GUARDED
; /* Thread that transitions. */
294 const char *state_where OVS_GUARDED
; /* transition_ukey() locator. */
296 /* Datapath flow actions as nlattrs. Protected by RCU. Read with
297 * ukey_get_actions(), and write with ukey_set_actions(). */
298 OVSRCU_TYPE(struct ofpbuf
*) actions
;
300 struct xlate_cache
*xcache OVS_GUARDED
; /* Cache for xlate entries that
301 * are affected by this ukey.
302 * Used for stats and learning.*/
304 struct odputil_keybuf buf
;
308 uint32_t key_recirc_id
; /* Non-zero if reference is held by the ukey. */
309 struct recirc_refs recircs
; /* Action recirc IDs with references held. */
311 #define OFFL_REBAL_INTVL_MSEC 3000 /* dynamic offload rebalance freq */
312 struct netdev
*in_netdev
; /* in_odp_port's netdev */
313 bool offloaded
; /* True if flow is offloaded */
314 uint64_t flow_pps_rate
; /* Packets-Per-Second rate */
315 long long int flow_time
; /* last pps update time */
316 uint64_t flow_packets
; /* #pkts seen in interval */
317 uint64_t flow_backlog_packets
; /* prev-mode #pkts (offl or kernel) */
320 /* Datapath operation with optional ukey attached. */
322 struct udpif_key
*ukey
;
323 struct dpif_flow_stats stats
; /* Stats for 'op'. */
324 struct dpif_op dop
; /* Flow operation. */
327 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
328 static struct ovs_list all_udpifs
= OVS_LIST_INITIALIZER(&all_udpifs
);
330 static size_t recv_upcalls(struct handler
*);
331 static int process_upcall(struct udpif
*, struct upcall
*,
332 struct ofpbuf
*odp_actions
, struct flow_wildcards
*);
333 static void handle_upcalls(struct udpif
*, struct upcall
*, size_t n_upcalls
);
334 static void udpif_stop_threads(struct udpif
*);
335 static void udpif_start_threads(struct udpif
*, size_t n_handlers
,
336 size_t n_revalidators
);
337 static void udpif_pause_revalidators(struct udpif
*);
338 static void udpif_resume_revalidators(struct udpif
*);
339 static void *udpif_upcall_handler(void *);
340 static void *udpif_revalidator(void *);
341 static unsigned long udpif_get_n_flows(struct udpif
*);
342 static void revalidate(struct revalidator
*);
343 static void revalidator_pause(struct revalidator
*);
344 static void revalidator_sweep(struct revalidator
*);
345 static void revalidator_purge(struct revalidator
*);
346 static void upcall_unixctl_show(struct unixctl_conn
*conn
, int argc
,
347 const char *argv
[], void *aux
);
348 static void upcall_unixctl_disable_megaflows(struct unixctl_conn
*, int argc
,
349 const char *argv
[], void *aux
);
350 static void upcall_unixctl_enable_megaflows(struct unixctl_conn
*, int argc
,
351 const char *argv
[], void *aux
);
352 static void upcall_unixctl_disable_ufid(struct unixctl_conn
*, int argc
,
353 const char *argv
[], void *aux
);
354 static void upcall_unixctl_enable_ufid(struct unixctl_conn
*, int argc
,
355 const char *argv
[], void *aux
);
356 static void upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
, int argc
,
357 const char *argv
[], void *aux
);
358 static void upcall_unixctl_dump_wait(struct unixctl_conn
*conn
, int argc
,
359 const char *argv
[], void *aux
);
360 static void upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc
,
361 const char *argv
[], void *aux
);
363 static struct udpif_key
*ukey_create_from_upcall(struct upcall
*,
364 struct flow_wildcards
*);
365 static int ukey_create_from_dpif_flow(const struct udpif
*,
366 const struct dpif_flow
*,
367 struct udpif_key
**);
368 static void ukey_get_actions(struct udpif_key
*, const struct nlattr
**actions
,
370 static bool ukey_install__(struct udpif
*, struct udpif_key
*ukey
)
371 OVS_TRY_LOCK(true, ukey
->mutex
);
372 static bool ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
);
373 static void transition_ukey_at(struct udpif_key
*ukey
, enum ukey_state dst
,
375 OVS_REQUIRES(ukey
->mutex
);
376 #define transition_ukey(UKEY, DST) \
377 transition_ukey_at(UKEY, DST, OVS_SOURCE_LOCATOR)
378 static struct udpif_key
*ukey_lookup(struct udpif
*udpif
,
379 const ovs_u128
*ufid
,
380 const unsigned pmd_id
);
381 static int ukey_acquire(struct udpif
*, const struct dpif_flow
*,
382 struct udpif_key
**result
, int *error
);
383 static void ukey_delete__(struct udpif_key
*);
384 static void ukey_delete(struct umap
*, struct udpif_key
*);
385 static enum upcall_type
classify_upcall(enum dpif_upcall_type type
,
386 const struct nlattr
*userdata
,
387 struct user_action_cookie
*cookie
);
389 static void put_op_init(struct ukey_op
*op
, struct udpif_key
*ukey
,
390 enum dpif_flow_put_flags flags
);
391 static void delete_op_init(struct udpif
*udpif
, struct ukey_op
*op
,
392 struct udpif_key
*ukey
);
394 static int upcall_receive(struct upcall
*, const struct dpif_backer
*,
395 const struct dp_packet
*packet
, enum dpif_upcall_type
,
396 const struct nlattr
*userdata
, const struct flow
*,
397 const unsigned int mru
,
398 const ovs_u128
*ufid
, const unsigned pmd_id
);
399 static void upcall_uninit(struct upcall
*);
401 static void udpif_flow_rebalance(struct udpif
*udpif
);
402 static int udpif_flow_program(struct udpif
*udpif
, struct udpif_key
*ukey
,
403 enum dpif_offload_type offload_type
);
404 static int udpif_flow_unprogram(struct udpif
*udpif
, struct udpif_key
*ukey
,
405 enum dpif_offload_type offload_type
);
407 static upcall_callback upcall_cb
;
408 static dp_purge_callback dp_purge_cb
;
410 static atomic_bool enable_megaflows
= ATOMIC_VAR_INIT(true);
411 static atomic_bool enable_ufid
= ATOMIC_VAR_INIT(true);
416 static struct ovsthread_once once
= OVSTHREAD_ONCE_INITIALIZER
;
417 if (ovsthread_once_start(&once
)) {
418 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show
,
420 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
421 upcall_unixctl_disable_megaflows
, NULL
);
422 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
423 upcall_unixctl_enable_megaflows
, NULL
);
424 unixctl_command_register("upcall/disable-ufid", "", 0, 0,
425 upcall_unixctl_disable_ufid
, NULL
);
426 unixctl_command_register("upcall/enable-ufid", "", 0, 0,
427 upcall_unixctl_enable_ufid
, NULL
);
428 unixctl_command_register("upcall/set-flow-limit", "flow-limit-number",
429 1, 1, upcall_unixctl_set_flow_limit
, NULL
);
430 unixctl_command_register("revalidator/wait", "", 0, 0,
431 upcall_unixctl_dump_wait
, NULL
);
432 unixctl_command_register("revalidator/purge", "", 0, 0,
433 upcall_unixctl_purge
, NULL
);
434 ovsthread_once_done(&once
);
439 udpif_create(struct dpif_backer
*backer
, struct dpif
*dpif
)
441 struct udpif
*udpif
= xzalloc(sizeof *udpif
);
444 udpif
->backer
= backer
;
445 atomic_init(&udpif
->flow_limit
, MIN(ofproto_flow_limit
, 10000));
446 udpif
->reval_seq
= seq_create();
447 udpif
->dump_seq
= seq_create();
448 latch_init(&udpif
->exit_latch
);
449 latch_init(&udpif
->pause_latch
);
450 ovs_list_push_back(&all_udpifs
, &udpif
->list_node
);
451 atomic_init(&udpif
->enable_ufid
, false);
452 atomic_init(&udpif
->n_flows
, 0);
453 atomic_init(&udpif
->n_flows_timestamp
, LLONG_MIN
);
454 ovs_mutex_init(&udpif
->n_flows_mutex
);
455 udpif
->ukeys
= xmalloc(N_UMAPS
* sizeof *udpif
->ukeys
);
456 for (int i
= 0; i
< N_UMAPS
; i
++) {
457 cmap_init(&udpif
->ukeys
[i
].cmap
);
458 ovs_mutex_init(&udpif
->ukeys
[i
].mutex
);
461 dpif_register_upcall_cb(dpif
, upcall_cb
, udpif
);
462 dpif_register_dp_purge_cb(dpif
, dp_purge_cb
, udpif
);
468 udpif_run(struct udpif
*udpif
)
470 if (udpif
->conns
&& udpif
->conn_seq
!= seq_read(udpif
->dump_seq
)) {
473 for (i
= 0; i
< udpif
->n_conns
; i
++) {
474 unixctl_command_reply(udpif
->conns
[i
], NULL
);
483 udpif_destroy(struct udpif
*udpif
)
485 udpif_stop_threads(udpif
);
487 dpif_register_dp_purge_cb(udpif
->dpif
, NULL
, udpif
);
488 dpif_register_upcall_cb(udpif
->dpif
, NULL
, udpif
);
490 for (int i
= 0; i
< N_UMAPS
; i
++) {
491 cmap_destroy(&udpif
->ukeys
[i
].cmap
);
492 ovs_mutex_destroy(&udpif
->ukeys
[i
].mutex
);
497 ovs_list_remove(&udpif
->list_node
);
498 latch_destroy(&udpif
->exit_latch
);
499 latch_destroy(&udpif
->pause_latch
);
500 seq_destroy(udpif
->reval_seq
);
501 seq_destroy(udpif
->dump_seq
);
502 ovs_mutex_destroy(&udpif
->n_flows_mutex
);
506 /* Stops the handler and revalidator threads. */
508 udpif_stop_threads(struct udpif
*udpif
)
510 if (udpif
&& (udpif
->n_handlers
!= 0 || udpif
->n_revalidators
!= 0)) {
513 /* Tell the threads to exit. */
514 latch_set(&udpif
->exit_latch
);
516 /* Wait for the threads to exit. Quiesce because this can take a long
518 ovsrcu_quiesce_start();
519 for (i
= 0; i
< udpif
->n_handlers
; i
++) {
520 xpthread_join(udpif
->handlers
[i
].thread
, NULL
);
522 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
523 xpthread_join(udpif
->revalidators
[i
].thread
, NULL
);
525 dpif_disable_upcall(udpif
->dpif
);
526 ovsrcu_quiesce_end();
528 /* Delete ukeys, and delete all flows from the datapath to prevent
529 * double-counting stats. */
530 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
531 revalidator_purge(&udpif
->revalidators
[i
]);
534 latch_poll(&udpif
->exit_latch
);
536 ovs_barrier_destroy(&udpif
->reval_barrier
);
537 ovs_barrier_destroy(&udpif
->pause_barrier
);
539 free(udpif
->revalidators
);
540 udpif
->revalidators
= NULL
;
541 udpif
->n_revalidators
= 0;
543 free(udpif
->handlers
);
544 udpif
->handlers
= NULL
;
545 udpif
->n_handlers
= 0;
549 /* Starts the handler and revalidator threads. */
551 udpif_start_threads(struct udpif
*udpif
, size_t n_handlers_
,
552 size_t n_revalidators_
)
554 if (udpif
&& n_handlers_
&& n_revalidators_
) {
555 /* Creating a thread can take a significant amount of time on some
556 * systems, even hundred of milliseconds, so quiesce around it. */
557 ovsrcu_quiesce_start();
559 udpif
->n_handlers
= n_handlers_
;
560 udpif
->n_revalidators
= n_revalidators_
;
562 udpif
->handlers
= xzalloc(udpif
->n_handlers
* sizeof *udpif
->handlers
);
563 for (size_t i
= 0; i
< udpif
->n_handlers
; i
++) {
564 struct handler
*handler
= &udpif
->handlers
[i
];
566 handler
->udpif
= udpif
;
567 handler
->handler_id
= i
;
568 handler
->thread
= ovs_thread_create(
569 "handler", udpif_upcall_handler
, handler
);
572 atomic_init(&udpif
->enable_ufid
, udpif
->backer
->rt_support
.ufid
);
573 dpif_enable_upcall(udpif
->dpif
);
575 ovs_barrier_init(&udpif
->reval_barrier
, udpif
->n_revalidators
);
576 ovs_barrier_init(&udpif
->pause_barrier
, udpif
->n_revalidators
+ 1);
577 udpif
->reval_exit
= false;
578 udpif
->pause
= false;
579 udpif
->offload_rebalance_time
= time_msec();
580 udpif
->revalidators
= xzalloc(udpif
->n_revalidators
581 * sizeof *udpif
->revalidators
);
582 for (size_t i
= 0; i
< udpif
->n_revalidators
; i
++) {
583 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
585 revalidator
->udpif
= udpif
;
586 revalidator
->thread
= ovs_thread_create(
587 "revalidator", udpif_revalidator
, revalidator
);
589 ovsrcu_quiesce_end();
593 /* Pauses all revalidators. Should only be called by the main thread.
594 * When function returns, all revalidators are paused and will proceed
595 * only after udpif_resume_revalidators() is called. */
597 udpif_pause_revalidators(struct udpif
*udpif
)
599 if (udpif
->backer
->recv_set_enable
) {
600 latch_set(&udpif
->pause_latch
);
601 ovs_barrier_block(&udpif
->pause_barrier
);
605 /* Resumes the pausing of revalidators. Should only be called by the
608 udpif_resume_revalidators(struct udpif
*udpif
)
610 if (udpif
->backer
->recv_set_enable
) {
611 latch_poll(&udpif
->pause_latch
);
612 ovs_barrier_block(&udpif
->pause_barrier
);
616 /* Tells 'udpif' how many threads it should use to handle upcalls.
617 * 'n_handlers_' and 'n_revalidators_' can never be zero. 'udpif''s
618 * datapath handle must have packet reception enabled before starting
621 udpif_set_threads(struct udpif
*udpif
, size_t n_handlers_
,
622 size_t n_revalidators_
)
625 ovs_assert(n_handlers_
&& n_revalidators_
);
627 if (udpif
->n_handlers
!= n_handlers_
628 || udpif
->n_revalidators
!= n_revalidators_
) {
629 udpif_stop_threads(udpif
);
632 if (!udpif
->handlers
&& !udpif
->revalidators
) {
635 error
= dpif_handlers_set(udpif
->dpif
, n_handlers_
);
637 VLOG_ERR("failed to configure handlers in dpif %s: %s",
638 dpif_name(udpif
->dpif
), ovs_strerror(error
));
642 udpif_start_threads(udpif
, n_handlers_
, n_revalidators_
);
646 /* Waits for all ongoing upcall translations to complete. This ensures that
647 * there are no transient references to any removed ofprotos (or other
648 * objects). In particular, this should be called after an ofproto is removed
649 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
651 udpif_synchronize(struct udpif
*udpif
)
653 /* This is stronger than necessary. It would be sufficient to ensure
654 * (somehow) that each handler and revalidator thread had passed through
655 * its main loop once. */
656 size_t n_handlers_
= udpif
->n_handlers
;
657 size_t n_revalidators_
= udpif
->n_revalidators
;
659 udpif_stop_threads(udpif
);
660 udpif_start_threads(udpif
, n_handlers_
, n_revalidators_
);
663 /* Notifies 'udpif' that something changed which may render previous
664 * xlate_actions() results invalid. */
666 udpif_revalidate(struct udpif
*udpif
)
668 seq_change(udpif
->reval_seq
);
671 /* Returns a seq which increments every time 'udpif' pulls stats from the
672 * datapath. Callers can use this to get a sense of when might be a good time
673 * to do periodic work which relies on relatively up to date statistics. */
675 udpif_dump_seq(struct udpif
*udpif
)
677 return udpif
->dump_seq
;
681 udpif_get_memory_usage(struct udpif
*udpif
, struct simap
*usage
)
685 simap_increase(usage
, "handlers", udpif
->n_handlers
);
687 simap_increase(usage
, "revalidators", udpif
->n_revalidators
);
688 for (i
= 0; i
< N_UMAPS
; i
++) {
689 simap_increase(usage
, "udpif keys", cmap_count(&udpif
->ukeys
[i
].cmap
));
693 /* Remove flows from a single datapath. */
695 udpif_flush(struct udpif
*udpif
)
697 size_t n_handlers_
= udpif
->n_handlers
;
698 size_t n_revalidators_
= udpif
->n_revalidators
;
700 udpif_stop_threads(udpif
);
701 dpif_flow_flush(udpif
->dpif
);
702 udpif_start_threads(udpif
, n_handlers_
, n_revalidators_
);
705 /* Removes all flows from all datapaths. */
707 udpif_flush_all_datapaths(void)
711 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
717 udpif_use_ufid(struct udpif
*udpif
)
721 atomic_read_relaxed(&enable_ufid
, &enable
);
722 return enable
&& udpif
->backer
->rt_support
.ufid
;
727 udpif_get_n_flows(struct udpif
*udpif
)
729 long long int time
, now
;
730 unsigned long flow_count
;
733 atomic_read_relaxed(&udpif
->n_flows_timestamp
, &time
);
734 if (time
< now
- 100 && !ovs_mutex_trylock(&udpif
->n_flows_mutex
)) {
735 struct dpif_dp_stats stats
;
737 atomic_store_relaxed(&udpif
->n_flows_timestamp
, now
);
738 dpif_get_dp_stats(udpif
->dpif
, &stats
);
739 flow_count
= stats
.n_flows
;
740 atomic_store_relaxed(&udpif
->n_flows
, flow_count
);
741 ovs_mutex_unlock(&udpif
->n_flows_mutex
);
743 atomic_read_relaxed(&udpif
->n_flows
, &flow_count
);
748 /* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH
749 * upcalls from dpif, processes the batch and installs corresponding flows
752 udpif_upcall_handler(void *arg
)
754 struct handler
*handler
= arg
;
755 struct udpif
*udpif
= handler
->udpif
;
757 while (!latch_is_set(&handler
->udpif
->exit_latch
)) {
758 if (recv_upcalls(handler
)) {
759 poll_immediate_wake();
761 dpif_recv_wait(udpif
->dpif
, handler
->handler_id
);
762 latch_wait(&udpif
->exit_latch
);
771 recv_upcalls(struct handler
*handler
)
773 struct udpif
*udpif
= handler
->udpif
;
774 uint64_t recv_stubs
[UPCALL_MAX_BATCH
][512 / 8];
775 struct ofpbuf recv_bufs
[UPCALL_MAX_BATCH
];
776 struct dpif_upcall dupcalls
[UPCALL_MAX_BATCH
];
777 struct upcall upcalls
[UPCALL_MAX_BATCH
];
778 struct flow flows
[UPCALL_MAX_BATCH
];
782 while (n_upcalls
< UPCALL_MAX_BATCH
) {
783 struct ofpbuf
*recv_buf
= &recv_bufs
[n_upcalls
];
784 struct dpif_upcall
*dupcall
= &dupcalls
[n_upcalls
];
785 struct upcall
*upcall
= &upcalls
[n_upcalls
];
786 struct flow
*flow
= &flows
[n_upcalls
];
790 ofpbuf_use_stub(recv_buf
, recv_stubs
[n_upcalls
],
791 sizeof recv_stubs
[n_upcalls
]);
792 if (dpif_recv(udpif
->dpif
, handler
->handler_id
, dupcall
, recv_buf
)) {
793 ofpbuf_uninit(recv_buf
);
797 upcall
->fitness
= odp_flow_key_to_flow(dupcall
->key
, dupcall
->key_len
,
799 if (upcall
->fitness
== ODP_FIT_ERROR
) {
804 mru
= nl_attr_get_u16(dupcall
->mru
);
809 error
= upcall_receive(upcall
, udpif
->backer
, &dupcall
->packet
,
810 dupcall
->type
, dupcall
->userdata
, flow
, mru
,
811 &dupcall
->ufid
, PMD_ID_NULL
);
813 if (error
== ENODEV
) {
814 /* Received packet on datapath port for which we couldn't
815 * associate an ofproto. This can happen if a port is removed
816 * while traffic is being received. Print a rate-limited
817 * message in case it happens frequently. */
818 dpif_flow_put(udpif
->dpif
, DPIF_FP_CREATE
, dupcall
->key
,
819 dupcall
->key_len
, NULL
, 0, NULL
, 0,
820 &dupcall
->ufid
, PMD_ID_NULL
, NULL
);
821 VLOG_INFO_RL(&rl
, "received packet on unassociated datapath "
822 "port %"PRIu32
, flow
->in_port
.odp_port
);
827 upcall
->key
= dupcall
->key
;
828 upcall
->key_len
= dupcall
->key_len
;
829 upcall
->ufid
= &dupcall
->ufid
;
831 upcall
->out_tun_key
= dupcall
->out_tun_key
;
832 upcall
->actions
= dupcall
->actions
;
834 pkt_metadata_from_flow(&dupcall
->packet
.md
, flow
);
835 flow_extract(&dupcall
->packet
, flow
);
837 error
= process_upcall(udpif
, upcall
,
838 &upcall
->odp_actions
, &upcall
->wc
);
847 upcall_uninit(upcall
);
849 dp_packet_uninit(&dupcall
->packet
);
850 ofpbuf_uninit(recv_buf
);
854 handle_upcalls(handler
->udpif
, upcalls
, n_upcalls
);
855 for (i
= 0; i
< n_upcalls
; i
++) {
856 dp_packet_uninit(&dupcalls
[i
].packet
);
857 ofpbuf_uninit(&recv_bufs
[i
]);
858 upcall_uninit(&upcalls
[i
]);
866 udpif_run_flow_rebalance(struct udpif
*udpif
)
868 long long int now
= 0;
870 /* Don't rebalance if OFFL_REBAL_INTVL_MSEC have not elapsed */
872 if (now
< udpif
->offload_rebalance_time
+ OFFL_REBAL_INTVL_MSEC
) {
876 if (!netdev_any_oor()) {
880 VLOG_DBG("Offload rebalance: Found OOR netdevs");
881 udpif
->offload_rebalance_time
= now
;
882 udpif_flow_rebalance(udpif
);
886 udpif_revalidator(void *arg
)
888 /* Used by all revalidators. */
889 struct revalidator
*revalidator
= arg
;
890 struct udpif
*udpif
= revalidator
->udpif
;
891 bool leader
= revalidator
== &udpif
->revalidators
[0];
893 /* Used only by the leader. */
894 long long int start_time
= 0;
895 uint64_t last_reval_seq
= 0;
898 revalidator
->id
= ovsthread_id_self();
903 recirc_run(); /* Recirculation cleanup. */
905 reval_seq
= seq_read(udpif
->reval_seq
);
906 last_reval_seq
= reval_seq
;
908 n_flows
= udpif_get_n_flows(udpif
);
909 udpif
->max_n_flows
= MAX(n_flows
, udpif
->max_n_flows
);
910 udpif
->avg_n_flows
= (udpif
->avg_n_flows
+ n_flows
) / 2;
912 /* Only the leader checks the pause latch to prevent a race where
913 * some threads think it's false and proceed to block on
914 * reval_barrier and others think it's true and block indefinitely
915 * on the pause_barrier */
916 udpif
->pause
= latch_is_set(&udpif
->pause_latch
);
918 /* Only the leader checks the exit latch to prevent a race where
919 * some threads think it's true and exit and others think it's
920 * false and block indefinitely on the reval_barrier */
921 udpif
->reval_exit
= latch_is_set(&udpif
->exit_latch
);
923 start_time
= time_msec();
924 if (!udpif
->reval_exit
) {
927 terse_dump
= udpif_use_ufid(udpif
);
928 udpif
->dump
= dpif_flow_dump_create(udpif
->dpif
, terse_dump
,
933 /* Wait for the leader to start the flow dump. */
934 ovs_barrier_block(&udpif
->reval_barrier
);
936 revalidator_pause(revalidator
);
939 if (udpif
->reval_exit
) {
942 revalidate(revalidator
);
944 /* Wait for all flows to have been dumped before we garbage collect. */
945 ovs_barrier_block(&udpif
->reval_barrier
);
946 revalidator_sweep(revalidator
);
948 /* Wait for all revalidators to finish garbage collection. */
949 ovs_barrier_block(&udpif
->reval_barrier
);
952 unsigned int flow_limit
;
953 long long int duration
;
955 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
957 dpif_flow_dump_destroy(udpif
->dump
);
958 seq_change(udpif
->dump_seq
);
959 if (netdev_is_offload_rebalance_policy_enabled()) {
960 udpif_run_flow_rebalance(udpif
);
963 duration
= MAX(time_msec() - start_time
, 1);
964 udpif
->dump_duration
= duration
;
965 if (duration
> 2000) {
966 flow_limit
/= duration
/ 1000;
967 } else if (duration
> 1300) {
968 flow_limit
= flow_limit
* 3 / 4;
969 } else if (duration
< 1000 &&
970 flow_limit
< n_flows
* 1000 / duration
) {
973 flow_limit
= MIN(ofproto_flow_limit
, MAX(flow_limit
, 1000));
974 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
976 if (duration
> 2000) {
977 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
981 poll_timer_wait_until(start_time
+ MIN(ofproto_max_idle
, 500));
982 seq_wait(udpif
->reval_seq
, last_reval_seq
);
983 latch_wait(&udpif
->exit_latch
);
984 latch_wait(&udpif
->pause_latch
);
987 if (!latch_is_set(&udpif
->pause_latch
) &&
988 !latch_is_set(&udpif
->exit_latch
)) {
989 long long int now
= time_msec();
990 /* Block again if we are woken up within 5ms of the last start
994 if (now
< start_time
) {
995 poll_timer_wait_until(start_time
);
996 latch_wait(&udpif
->exit_latch
);
997 latch_wait(&udpif
->pause_latch
);
1007 static enum upcall_type
1008 classify_upcall(enum dpif_upcall_type type
, const struct nlattr
*userdata
,
1009 struct user_action_cookie
*cookie
)
1011 /* First look at the upcall type. */
1013 case DPIF_UC_ACTION
:
1019 case DPIF_N_UC_TYPES
:
1021 VLOG_WARN_RL(&rl
, "upcall has unexpected type %"PRIu32
, type
);
1025 /* "action" upcalls need a closer look. */
1027 VLOG_WARN_RL(&rl
, "action upcall missing cookie");
1031 size_t userdata_len
= nl_attr_get_size(userdata
);
1032 if (userdata_len
!= sizeof *cookie
) {
1033 VLOG_WARN_RL(&rl
, "action upcall cookie has unexpected size %"PRIuSIZE
,
1037 memcpy(cookie
, nl_attr_get(userdata
), sizeof *cookie
);
1038 if (cookie
->type
== USER_ACTION_COOKIE_SFLOW
) {
1039 return SFLOW_UPCALL
;
1040 } else if (cookie
->type
== USER_ACTION_COOKIE_SLOW_PATH
) {
1041 return SLOW_PATH_UPCALL
;
1042 } else if (cookie
->type
== USER_ACTION_COOKIE_FLOW_SAMPLE
) {
1043 return FLOW_SAMPLE_UPCALL
;
1044 } else if (cookie
->type
== USER_ACTION_COOKIE_IPFIX
) {
1045 return IPFIX_UPCALL
;
1046 } else if (cookie
->type
== USER_ACTION_COOKIE_CONTROLLER
) {
1047 return CONTROLLER_UPCALL
;
1049 VLOG_WARN_RL(&rl
, "invalid user cookie of type %"PRIu16
1050 " and size %"PRIuSIZE
, cookie
->type
, userdata_len
);
1055 /* Calculates slow path actions for 'xout'. 'buf' must statically be
1056 * initialized with at least 128 bytes of space. */
1058 compose_slow_path(struct udpif
*udpif
, struct xlate_out
*xout
,
1059 odp_port_t odp_in_port
, ofp_port_t ofp_in_port
,
1060 struct ofpbuf
*buf
, uint32_t meter_id
,
1061 struct uuid
*ofproto_uuid
)
1063 struct user_action_cookie cookie
;
1067 cookie
.type
= USER_ACTION_COOKIE_SLOW_PATH
;
1068 cookie
.ofp_in_port
= ofp_in_port
;
1069 cookie
.ofproto_uuid
= *ofproto_uuid
;
1070 cookie
.slow_path
.reason
= xout
->slow
;
1072 port
= xout
->slow
& (SLOW_CFM
| SLOW_BFD
| SLOW_LACP
| SLOW_STP
)
1075 pid
= dpif_port_get_pid(udpif
->dpif
, port
);
1079 if (meter_id
!= UINT32_MAX
) {
1080 /* If slowpath meter is configured, generate clone(meter, userspace)
1082 offset
= nl_msg_start_nested(buf
, OVS_ACTION_ATTR_SAMPLE
);
1083 nl_msg_put_u32(buf
, OVS_SAMPLE_ATTR_PROBABILITY
, UINT32_MAX
);
1084 ac_offset
= nl_msg_start_nested(buf
, OVS_SAMPLE_ATTR_ACTIONS
);
1085 nl_msg_put_u32(buf
, OVS_ACTION_ATTR_METER
, meter_id
);
1088 odp_put_userspace_action(pid
, &cookie
, sizeof cookie
,
1089 ODPP_NONE
, false, buf
);
1091 if (meter_id
!= UINT32_MAX
) {
1092 nl_msg_end_nested(buf
, ac_offset
);
1093 nl_msg_end_nested(buf
, offset
);
1097 /* If there is no error, the upcall must be destroyed with upcall_uninit()
1098 * before quiescing, as the referred objects are guaranteed to exist only
1099 * until the calling thread quiesces. Otherwise, do not call upcall_uninit()
1100 * since the 'upcall->put_actions' remains uninitialized. */
1102 upcall_receive(struct upcall
*upcall
, const struct dpif_backer
*backer
,
1103 const struct dp_packet
*packet
, enum dpif_upcall_type type
,
1104 const struct nlattr
*userdata
, const struct flow
*flow
,
1105 const unsigned int mru
,
1106 const ovs_u128
*ufid
, const unsigned pmd_id
)
1110 upcall
->type
= classify_upcall(type
, userdata
, &upcall
->cookie
);
1111 if (upcall
->type
== BAD_UPCALL
) {
1113 } else if (upcall
->type
== MISS_UPCALL
) {
1114 error
= xlate_lookup(backer
, flow
, &upcall
->ofproto
, &upcall
->ipfix
,
1115 &upcall
->sflow
, NULL
, &upcall
->ofp_in_port
);
1120 struct ofproto_dpif
*ofproto
1121 = ofproto_dpif_lookup_by_uuid(&upcall
->cookie
.ofproto_uuid
);
1123 VLOG_INFO_RL(&rl
, "upcall could not find ofproto");
1126 upcall
->ofproto
= ofproto
;
1127 upcall
->ipfix
= ofproto
->ipfix
;
1128 upcall
->sflow
= ofproto
->sflow
;
1129 upcall
->ofp_in_port
= upcall
->cookie
.ofp_in_port
;
1132 upcall
->recirc
= NULL
;
1133 upcall
->have_recirc_ref
= false;
1134 upcall
->flow
= flow
;
1135 upcall
->packet
= packet
;
1136 upcall
->ufid
= ufid
;
1137 upcall
->pmd_id
= pmd_id
;
1138 ofpbuf_use_stub(&upcall
->odp_actions
, upcall
->odp_actions_stub
,
1139 sizeof upcall
->odp_actions_stub
);
1140 ofpbuf_init(&upcall
->put_actions
, 0);
1142 upcall
->xout_initialized
= false;
1143 upcall
->ukey_persists
= false;
1145 upcall
->ukey
= NULL
;
1147 upcall
->key_len
= 0;
1150 upcall
->out_tun_key
= NULL
;
1151 upcall
->actions
= NULL
;
1157 upcall_xlate(struct udpif
*udpif
, struct upcall
*upcall
,
1158 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1160 struct dpif_flow_stats stats
;
1161 enum xlate_error xerr
;
1162 struct xlate_in xin
;
1165 stats
.n_packets
= 1;
1166 stats
.n_bytes
= dp_packet_size(upcall
->packet
);
1167 stats
.used
= time_msec();
1168 stats
.tcp_flags
= ntohs(upcall
->flow
->tcp_flags
);
1170 xlate_in_init(&xin
, upcall
->ofproto
,
1171 ofproto_dpif_get_tables_version(upcall
->ofproto
),
1172 upcall
->flow
, upcall
->ofp_in_port
, NULL
,
1173 stats
.tcp_flags
, upcall
->packet
, wc
, odp_actions
);
1175 if (upcall
->type
== MISS_UPCALL
) {
1176 xin
.resubmit_stats
= &stats
;
1178 if (xin
.frozen_state
) {
1179 /* We may install a datapath flow only if we get a reference to the
1180 * recirculation context (otherwise we could have recirculation
1181 * upcalls using recirculation ID for which no context can be
1182 * found). We may still execute the flow's actions even if we
1183 * don't install the flow. */
1184 upcall
->recirc
= recirc_id_node_from_state(xin
.frozen_state
);
1185 upcall
->have_recirc_ref
= recirc_id_node_try_ref_rcu(upcall
->recirc
);
1188 /* For non-miss upcalls, we are either executing actions (one of which
1189 * is an userspace action) for an upcall, in which case the stats have
1190 * already been taken care of, or there's a flow in the datapath which
1191 * this packet was accounted to. Presumably the revalidators will deal
1192 * with pushing its stats eventually. */
1195 upcall
->reval_seq
= seq_read(udpif
->reval_seq
);
1197 xerr
= xlate_actions(&xin
, &upcall
->xout
);
1199 /* Translate again and log the ofproto trace for
1200 * these two error types. */
1201 if (xerr
== XLATE_RECURSION_TOO_DEEP
||
1202 xerr
== XLATE_TOO_MANY_RESUBMITS
) {
1203 static struct vlog_rate_limit rll
= VLOG_RATE_LIMIT_INIT(1, 1);
1205 /* This is a huge log, so be conservative. */
1206 if (!VLOG_DROP_WARN(&rll
)) {
1208 ofproto_trace(upcall
->ofproto
, upcall
->flow
,
1209 upcall
->packet
, NULL
, 0, NULL
, &output
);
1210 VLOG_WARN("%s", ds_cstr(&output
));
1211 ds_destroy(&output
);
1216 /* Convert the input port wildcard from OFP to ODP format. There's no
1217 * real way to do this for arbitrary bitmasks since the numbering spaces
1218 * aren't the same. However, flow translation always exact matches the
1219 * whole thing, so we can do the same here. */
1220 WC_MASK_FIELD(wc
, in_port
.odp_port
);
1223 upcall
->xout_initialized
= true;
1225 if (upcall
->fitness
== ODP_FIT_TOO_LITTLE
) {
1226 upcall
->xout
.slow
|= SLOW_MATCH
;
1228 if (!upcall
->xout
.slow
) {
1229 ofpbuf_use_const(&upcall
->put_actions
,
1230 odp_actions
->data
, odp_actions
->size
);
1232 /* upcall->put_actions already initialized by upcall_receive(). */
1233 compose_slow_path(udpif
, &upcall
->xout
,
1234 upcall
->flow
->in_port
.odp_port
, upcall
->ofp_in_port
,
1235 &upcall
->put_actions
,
1236 upcall
->ofproto
->up
.slowpath_meter_id
,
1237 &upcall
->ofproto
->uuid
);
1240 /* This function is also called for slow-pathed flows. As we are only
1241 * going to create new datapath flows for actual datapath misses, there is
1242 * no point in creating a ukey otherwise. */
1243 if (upcall
->type
== MISS_UPCALL
) {
1244 upcall
->ukey
= ukey_create_from_upcall(upcall
, wc
);
1249 upcall_uninit(struct upcall
*upcall
)
1252 if (upcall
->xout_initialized
) {
1253 xlate_out_uninit(&upcall
->xout
);
1255 ofpbuf_uninit(&upcall
->odp_actions
);
1256 ofpbuf_uninit(&upcall
->put_actions
);
1258 if (!upcall
->ukey_persists
) {
1259 ukey_delete__(upcall
->ukey
);
1261 } else if (upcall
->have_recirc_ref
) {
1262 /* The reference was transferred to the ukey if one was created. */
1263 recirc_id_node_unref(upcall
->recirc
);
1268 /* If there are less flows than the limit, and this is a miss upcall which
1270 * - Has no recirc_id, OR
1271 * - Has a recirc_id and we can get a reference on the recirc ctx,
1273 * Then we should install the flow (true). Otherwise, return false. */
1275 should_install_flow(struct udpif
*udpif
, struct upcall
*upcall
)
1277 unsigned int flow_limit
;
1279 if (upcall
->type
!= MISS_UPCALL
) {
1281 } else if (upcall
->recirc
&& !upcall
->have_recirc_ref
) {
1282 VLOG_DBG_RL(&rl
, "upcall: no reference for recirc flow");
1286 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
1287 if (udpif_get_n_flows(udpif
) >= flow_limit
) {
1288 VLOG_WARN_RL(&rl
, "upcall: datapath flow limit reached");
1296 upcall_cb(const struct dp_packet
*packet
, const struct flow
*flow
, ovs_u128
*ufid
,
1297 unsigned pmd_id
, enum dpif_upcall_type type
,
1298 const struct nlattr
*userdata
, struct ofpbuf
*actions
,
1299 struct flow_wildcards
*wc
, struct ofpbuf
*put_actions
, void *aux
)
1301 struct udpif
*udpif
= aux
;
1302 struct upcall upcall
;
1306 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1308 error
= upcall_receive(&upcall
, udpif
->backer
, packet
, type
, userdata
,
1309 flow
, 0, ufid
, pmd_id
);
1314 upcall
.fitness
= ODP_FIT_PERFECT
;
1315 error
= process_upcall(udpif
, &upcall
, actions
, wc
);
1320 if (upcall
.xout
.slow
&& put_actions
) {
1321 ofpbuf_put(put_actions
, upcall
.put_actions
.data
,
1322 upcall
.put_actions
.size
);
1325 if (OVS_UNLIKELY(!megaflow
&& wc
)) {
1326 flow_wildcards_init_for_packet(wc
, flow
);
1329 if (!should_install_flow(udpif
, &upcall
)) {
1334 if (upcall
.ukey
&& !ukey_install(udpif
, upcall
.ukey
)) {
1335 static struct vlog_rate_limit rll
= VLOG_RATE_LIMIT_INIT(1, 1);
1336 VLOG_WARN_RL(&rll
, "upcall_cb failure: ukey installation fails");
1341 upcall
.ukey_persists
= true;
1343 upcall_uninit(&upcall
);
1348 dpif_get_actions(struct udpif
*udpif
, struct upcall
*upcall
,
1349 const struct nlattr
**actions
)
1351 size_t actions_len
= 0;
1353 if (upcall
->actions
) {
1354 /* Actions were passed up from datapath. */
1355 *actions
= nl_attr_get(upcall
->actions
);
1356 actions_len
= nl_attr_get_size(upcall
->actions
);
1359 if (actions_len
== 0) {
1360 /* Lookup actions in userspace cache. */
1361 struct udpif_key
*ukey
= ukey_lookup(udpif
, upcall
->ufid
,
1364 ukey_get_actions(ukey
, actions
, &actions_len
);
1372 dpif_read_actions(struct udpif
*udpif
, struct upcall
*upcall
,
1373 const struct flow
*flow
, enum upcall_type type
,
1376 const struct nlattr
*actions
= NULL
;
1377 size_t actions_len
= dpif_get_actions(udpif
, upcall
, &actions
);
1379 if (!actions
|| !actions_len
) {
1385 dpif_sflow_read_actions(flow
, actions
, actions_len
, upcall_data
, true);
1387 case FLOW_SAMPLE_UPCALL
:
1389 dpif_ipfix_read_actions(flow
, actions
, actions_len
, upcall_data
);
1393 case SLOW_PATH_UPCALL
:
1394 case CONTROLLER_UPCALL
:
1403 process_upcall(struct udpif
*udpif
, struct upcall
*upcall
,
1404 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1406 const struct dp_packet
*packet
= upcall
->packet
;
1407 const struct flow
*flow
= upcall
->flow
;
1408 size_t actions_len
= 0;
1410 switch (upcall
->type
) {
1412 case SLOW_PATH_UPCALL
:
1413 upcall_xlate(udpif
, upcall
, odp_actions
, wc
);
1417 if (upcall
->sflow
) {
1418 struct dpif_sflow_actions sflow_actions
;
1420 memset(&sflow_actions
, 0, sizeof sflow_actions
);
1422 actions_len
= dpif_read_actions(udpif
, upcall
, flow
,
1423 upcall
->type
, &sflow_actions
);
1424 dpif_sflow_received(upcall
->sflow
, packet
, flow
,
1425 flow
->in_port
.odp_port
, &upcall
->cookie
,
1426 actions_len
> 0 ? &sflow_actions
: NULL
);
1431 case FLOW_SAMPLE_UPCALL
:
1432 if (upcall
->ipfix
) {
1433 struct flow_tnl output_tunnel_key
;
1434 struct dpif_ipfix_actions ipfix_actions
;
1436 memset(&ipfix_actions
, 0, sizeof ipfix_actions
);
1438 if (upcall
->out_tun_key
) {
1439 odp_tun_key_from_attr(upcall
->out_tun_key
, &output_tunnel_key
,
1443 actions_len
= dpif_read_actions(udpif
, upcall
, flow
,
1444 upcall
->type
, &ipfix_actions
);
1445 if (upcall
->type
== IPFIX_UPCALL
) {
1446 dpif_ipfix_bridge_sample(upcall
->ipfix
, packet
, flow
,
1447 flow
->in_port
.odp_port
,
1448 upcall
->cookie
.ipfix
.output_odp_port
,
1449 upcall
->out_tun_key
?
1450 &output_tunnel_key
: NULL
,
1452 &ipfix_actions
: NULL
);
1454 /* The flow reflects exactly the contents of the packet.
1455 * Sample the packet using it. */
1456 dpif_ipfix_flow_sample(upcall
->ipfix
, packet
, flow
,
1457 &upcall
->cookie
, flow
->in_port
.odp_port
,
1458 upcall
->out_tun_key
?
1459 &output_tunnel_key
: NULL
,
1460 actions_len
> 0 ? &ipfix_actions
: NULL
);
1465 case CONTROLLER_UPCALL
:
1467 struct user_action_cookie
*cookie
= &upcall
->cookie
;
1469 if (cookie
->controller
.dont_send
) {
1473 uint32_t recirc_id
= cookie
->controller
.recirc_id
;
1478 const struct recirc_id_node
*recirc_node
1479 = recirc_id_node_find(recirc_id
);
1484 const struct frozen_state
*state
= &recirc_node
->state
;
1486 struct ofproto_async_msg
*am
= xmalloc(sizeof *am
);
1487 *am
= (struct ofproto_async_msg
) {
1488 .controller_id
= cookie
->controller
.controller_id
,
1489 .oam
= OAM_PACKET_IN
,
1493 .packet
= xmemdup(dp_packet_data(packet
),
1494 dp_packet_size(packet
)),
1495 .packet_len
= dp_packet_size(packet
),
1496 .reason
= cookie
->controller
.reason
,
1497 .table_id
= state
->table_id
,
1498 .cookie
= get_32aligned_be64(
1499 &cookie
->controller
.rule_cookie
),
1500 .userdata
= (recirc_node
->state
.userdata_len
1501 ? xmemdup(recirc_node
->state
.userdata
,
1502 recirc_node
->state
.userdata_len
)
1504 .userdata_len
= recirc_node
->state
.userdata_len
,
1507 .max_len
= cookie
->controller
.max_len
,
1511 if (cookie
->controller
.continuation
) {
1512 am
->pin
.up
.stack
= (state
->stack_size
1513 ? xmemdup(state
->stack
, state
->stack_size
)
1515 am
->pin
.up
.stack_size
= state
->stack_size
,
1516 am
->pin
.up
.mirrors
= state
->mirrors
,
1517 am
->pin
.up
.conntracked
= state
->conntracked
,
1518 am
->pin
.up
.actions
= (state
->ofpacts_len
1519 ? xmemdup(state
->ofpacts
,
1520 state
->ofpacts_len
) : NULL
),
1521 am
->pin
.up
.actions_len
= state
->ofpacts_len
,
1522 am
->pin
.up
.action_set
= (state
->action_set_len
1523 ? xmemdup(state
->action_set
,
1524 state
->action_set_len
)
1526 am
->pin
.up
.action_set_len
= state
->action_set_len
,
1527 am
->pin
.up
.bridge
= upcall
->ofproto
->uuid
;
1528 am
->pin
.up
.odp_port
= upcall
->packet
->md
.in_port
.odp_port
;
1531 /* We don't want to use the upcall 'flow', since it may be
1532 * more specific than the point at which the "controller"
1533 * action was specified. */
1534 struct flow frozen_flow
;
1536 frozen_flow
= *flow
;
1537 if (!state
->conntracked
) {
1538 flow_clear_conntrack(&frozen_flow
);
1541 frozen_metadata_to_flow(&state
->metadata
, &frozen_flow
);
1542 flow_get_metadata(&frozen_flow
, &am
->pin
.up
.base
.flow_metadata
);
1544 ofproto_dpif_send_async_msg(upcall
->ofproto
, am
);
1556 handle_upcalls(struct udpif
*udpif
, struct upcall
*upcalls
,
1559 struct dpif_op
*opsp
[UPCALL_MAX_BATCH
* 2];
1560 struct ukey_op ops
[UPCALL_MAX_BATCH
* 2];
1561 size_t n_ops
, n_opsp
, i
;
1563 /* Handle the packets individually in order of arrival.
1565 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, SLOW_BFD, and SLOW_LLDP,
1566 * translation is what processes received packets for these
1569 * - For SLOW_ACTION, translation executes the actions directly.
1571 * The loop fills 'ops' with an array of operations to execute in the
1574 for (i
= 0; i
< n_upcalls
; i
++) {
1575 struct upcall
*upcall
= &upcalls
[i
];
1576 const struct dp_packet
*packet
= upcall
->packet
;
1579 if (should_install_flow(udpif
, upcall
)) {
1580 struct udpif_key
*ukey
= upcall
->ukey
;
1582 if (ukey_install(udpif
, ukey
)) {
1583 upcall
->ukey_persists
= true;
1584 put_op_init(&ops
[n_ops
++], ukey
, DPIF_FP_CREATE
);
1588 if (upcall
->odp_actions
.size
) {
1591 op
->dop
.type
= DPIF_OP_EXECUTE
;
1592 op
->dop
.execute
.packet
= CONST_CAST(struct dp_packet
*, packet
);
1593 op
->dop
.execute
.flow
= upcall
->flow
;
1594 odp_key_to_dp_packet(upcall
->key
, upcall
->key_len
,
1595 op
->dop
.execute
.packet
);
1596 op
->dop
.execute
.actions
= upcall
->odp_actions
.data
;
1597 op
->dop
.execute
.actions_len
= upcall
->odp_actions
.size
;
1598 op
->dop
.execute
.needs_help
= (upcall
->xout
.slow
& SLOW_ACTION
) != 0;
1599 op
->dop
.execute
.probe
= false;
1600 op
->dop
.execute
.mtu
= upcall
->mru
;
1604 /* Execute batch. */
1606 for (i
= 0; i
< n_ops
; i
++) {
1607 opsp
[n_opsp
++] = &ops
[i
].dop
;
1609 dpif_operate(udpif
->dpif
, opsp
, n_opsp
, DPIF_OFFLOAD_AUTO
);
1610 for (i
= 0; i
< n_ops
; i
++) {
1611 struct udpif_key
*ukey
= ops
[i
].ukey
;
1614 ovs_mutex_lock(&ukey
->mutex
);
1615 if (ops
[i
].dop
.error
) {
1616 transition_ukey(ukey
, UKEY_EVICTED
);
1617 } else if (ukey
->state
< UKEY_OPERATIONAL
) {
1618 transition_ukey(ukey
, UKEY_OPERATIONAL
);
1620 ovs_mutex_unlock(&ukey
->mutex
);
1626 get_ukey_hash(const ovs_u128
*ufid
, const unsigned pmd_id
)
1628 return hash_2words(ufid
->u32
[0], pmd_id
);
1631 static struct udpif_key
*
1632 ukey_lookup(struct udpif
*udpif
, const ovs_u128
*ufid
, const unsigned pmd_id
)
1634 struct udpif_key
*ukey
;
1635 int idx
= get_ukey_hash(ufid
, pmd_id
) % N_UMAPS
;
1636 struct cmap
*cmap
= &udpif
->ukeys
[idx
].cmap
;
1638 CMAP_FOR_EACH_WITH_HASH (ukey
, cmap_node
,
1639 get_ukey_hash(ufid
, pmd_id
), cmap
) {
1640 if (ovs_u128_equals(ukey
->ufid
, *ufid
)) {
1647 /* Provides safe lockless access of RCU protected 'ukey->actions'. Callers may
1648 * alternatively access the field directly if they take 'ukey->mutex'. */
1650 ukey_get_actions(struct udpif_key
*ukey
, const struct nlattr
**actions
, size_t *size
)
1652 const struct ofpbuf
*buf
= ovsrcu_get(struct ofpbuf
*, &ukey
->actions
);
1653 *actions
= buf
->data
;
1658 ukey_set_actions(struct udpif_key
*ukey
, const struct ofpbuf
*actions
)
1660 struct ofpbuf
*old_actions
= ovsrcu_get_protected(struct ofpbuf
*,
1664 ovsrcu_postpone(ofpbuf_delete
, old_actions
);
1667 ovsrcu_set(&ukey
->actions
, ofpbuf_clone(actions
));
1670 static struct udpif_key
*
1671 ukey_create__(const struct nlattr
*key
, size_t key_len
,
1672 const struct nlattr
*mask
, size_t mask_len
,
1673 bool ufid_present
, const ovs_u128
*ufid
,
1674 const unsigned pmd_id
, const struct ofpbuf
*actions
,
1675 uint64_t reval_seq
, long long int used
,
1676 uint32_t key_recirc_id
, struct xlate_out
*xout
)
1677 OVS_NO_THREAD_SAFETY_ANALYSIS
1679 struct udpif_key
*ukey
= xmalloc(sizeof *ukey
);
1681 memcpy(&ukey
->keybuf
, key
, key_len
);
1682 ukey
->key
= &ukey
->keybuf
.nla
;
1683 ukey
->key_len
= key_len
;
1684 memcpy(&ukey
->maskbuf
, mask
, mask_len
);
1685 ukey
->mask
= &ukey
->maskbuf
.nla
;
1686 ukey
->mask_len
= mask_len
;
1687 ukey
->ufid_present
= ufid_present
;
1689 ukey
->pmd_id
= pmd_id
;
1690 ukey
->hash
= get_ukey_hash(&ukey
->ufid
, pmd_id
);
1692 ovsrcu_init(&ukey
->actions
, NULL
);
1693 ukey_set_actions(ukey
, actions
);
1695 ovs_mutex_init(&ukey
->mutex
);
1696 ukey
->dump_seq
= 0; /* Not yet dumped */
1697 ukey
->reval_seq
= reval_seq
;
1698 ukey
->state
= UKEY_CREATED
;
1699 ukey
->state_thread
= ovsthread_id_self();
1700 ukey
->state_where
= OVS_SOURCE_LOCATOR
;
1701 ukey
->created
= ukey
->flow_time
= time_msec();
1702 memset(&ukey
->stats
, 0, sizeof ukey
->stats
);
1703 ukey
->stats
.used
= used
;
1704 ukey
->xcache
= NULL
;
1706 ukey
->offloaded
= false;
1707 ukey
->in_netdev
= NULL
;
1708 ukey
->flow_packets
= ukey
->flow_backlog_packets
= 0;
1710 ukey
->key_recirc_id
= key_recirc_id
;
1711 recirc_refs_init(&ukey
->recircs
);
1713 /* Take ownership of the action recirc id references. */
1714 recirc_refs_swap(&ukey
->recircs
, &xout
->recircs
);
1720 static struct udpif_key
*
1721 ukey_create_from_upcall(struct upcall
*upcall
, struct flow_wildcards
*wc
)
1723 struct odputil_keybuf keystub
, maskstub
;
1724 struct ofpbuf keybuf
, maskbuf
;
1726 struct odp_flow_key_parms odp_parms
= {
1727 .flow
= upcall
->flow
,
1728 .mask
= wc
? &wc
->masks
: NULL
,
1731 odp_parms
.support
= upcall
->ofproto
->backer
->rt_support
.odp
;
1732 if (upcall
->key_len
) {
1733 ofpbuf_use_const(&keybuf
, upcall
->key
, upcall
->key_len
);
1735 /* dpif-netdev doesn't provide a netlink-formatted flow key in the
1736 * upcall, so convert the upcall's flow here. */
1737 ofpbuf_use_stack(&keybuf
, &keystub
, sizeof keystub
);
1738 odp_flow_key_from_flow(&odp_parms
, &keybuf
);
1741 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1742 ofpbuf_use_stack(&maskbuf
, &maskstub
, sizeof maskstub
);
1743 if (megaflow
&& wc
) {
1744 odp_parms
.key_buf
= &keybuf
;
1745 odp_flow_key_from_mask(&odp_parms
, &maskbuf
);
1748 return ukey_create__(keybuf
.data
, keybuf
.size
, maskbuf
.data
, maskbuf
.size
,
1749 true, upcall
->ufid
, upcall
->pmd_id
,
1750 &upcall
->put_actions
, upcall
->reval_seq
, 0,
1751 upcall
->have_recirc_ref
? upcall
->recirc
->id
: 0,
1756 ukey_create_from_dpif_flow(const struct udpif
*udpif
,
1757 const struct dpif_flow
*flow
,
1758 struct udpif_key
**ukey
)
1760 struct dpif_flow full_flow
;
1761 struct ofpbuf actions
;
1763 uint64_t stub
[DPIF_FLOW_BUFSIZE
/ 8];
1764 const struct nlattr
*a
;
1767 if (!flow
->key_len
|| !flow
->actions_len
) {
1771 /* If the key or actions were not provided by the datapath, fetch the
1773 ofpbuf_use_stack(&buf
, &stub
, sizeof stub
);
1774 err
= dpif_flow_get(udpif
->dpif
, flow
->key
, flow
->key_len
,
1775 flow
->ufid_present
? &flow
->ufid
: NULL
,
1776 flow
->pmd_id
, &buf
, &full_flow
);
1783 /* Check the flow actions for recirculation action. As recirculation
1784 * relies on OVS userspace internal state, we need to delete all old
1785 * datapath flows with either a non-zero recirc_id in the key, or any
1786 * recirculation actions upon OVS restart. */
1787 NL_ATTR_FOR_EACH (a
, left
, flow
->key
, flow
->key_len
) {
1788 if (nl_attr_type(a
) == OVS_KEY_ATTR_RECIRC_ID
1789 && nl_attr_get_u32(a
) != 0) {
1793 NL_ATTR_FOR_EACH (a
, left
, flow
->actions
, flow
->actions_len
) {
1794 if (nl_attr_type(a
) == OVS_ACTION_ATTR_RECIRC
) {
1799 reval_seq
= seq_read(udpif
->reval_seq
) - 1; /* Ensure revalidation. */
1800 ofpbuf_use_const(&actions
, &flow
->actions
, flow
->actions_len
);
1801 *ukey
= ukey_create__(flow
->key
, flow
->key_len
,
1802 flow
->mask
, flow
->mask_len
, flow
->ufid_present
,
1803 &flow
->ufid
, flow
->pmd_id
, &actions
,
1804 reval_seq
, flow
->stats
.used
, 0, NULL
);
1810 try_ukey_replace(struct umap
*umap
, struct udpif_key
*old_ukey
,
1811 struct udpif_key
*new_ukey
)
1812 OVS_REQUIRES(umap
->mutex
)
1813 OVS_TRY_LOCK(true, new_ukey
->mutex
)
1815 bool replaced
= false;
1817 if (!ovs_mutex_trylock(&old_ukey
->mutex
)) {
1818 if (old_ukey
->state
== UKEY_EVICTED
) {
1819 /* The flow was deleted during the current revalidator dump,
1820 * but its ukey won't be fully cleaned up until the sweep phase.
1821 * In the mean time, we are receiving upcalls for this traffic.
1822 * Expedite the (new) flow install by replacing the ukey. */
1823 ovs_mutex_lock(&new_ukey
->mutex
);
1824 cmap_replace(&umap
->cmap
, &old_ukey
->cmap_node
,
1825 &new_ukey
->cmap_node
, new_ukey
->hash
);
1826 ovsrcu_postpone(ukey_delete__
, old_ukey
);
1827 transition_ukey(old_ukey
, UKEY_DELETED
);
1828 transition_ukey(new_ukey
, UKEY_VISIBLE
);
1831 ovs_mutex_unlock(&old_ukey
->mutex
);
1835 COVERAGE_INC(upcall_ukey_replace
);
1837 COVERAGE_INC(handler_duplicate_upcall
);
1842 /* Attempts to insert a ukey into the shared ukey maps.
1844 * On success, returns true, installs the ukey and returns it in a locked
1845 * state. Otherwise, returns false. */
1847 ukey_install__(struct udpif
*udpif
, struct udpif_key
*new_ukey
)
1848 OVS_TRY_LOCK(true, new_ukey
->mutex
)
1851 struct udpif_key
*old_ukey
;
1853 bool locked
= false;
1855 idx
= new_ukey
->hash
% N_UMAPS
;
1856 umap
= &udpif
->ukeys
[idx
];
1857 ovs_mutex_lock(&umap
->mutex
);
1858 old_ukey
= ukey_lookup(udpif
, &new_ukey
->ufid
, new_ukey
->pmd_id
);
1860 /* Uncommon case: A ukey is already installed with the same UFID. */
1861 if (old_ukey
->key_len
== new_ukey
->key_len
1862 && !memcmp(old_ukey
->key
, new_ukey
->key
, new_ukey
->key_len
)) {
1863 locked
= try_ukey_replace(umap
, old_ukey
, new_ukey
);
1865 struct ds ds
= DS_EMPTY_INITIALIZER
;
1867 odp_format_ufid(&old_ukey
->ufid
, &ds
);
1868 ds_put_cstr(&ds
, " ");
1869 odp_flow_key_format(old_ukey
->key
, old_ukey
->key_len
, &ds
);
1870 ds_put_cstr(&ds
, "\n");
1871 odp_format_ufid(&new_ukey
->ufid
, &ds
);
1872 ds_put_cstr(&ds
, " ");
1873 odp_flow_key_format(new_ukey
->key
, new_ukey
->key_len
, &ds
);
1875 VLOG_WARN_RL(&rl
, "Conflicting ukey for flows:\n%s", ds_cstr(&ds
));
1879 ovs_mutex_lock(&new_ukey
->mutex
);
1880 cmap_insert(&umap
->cmap
, &new_ukey
->cmap_node
, new_ukey
->hash
);
1881 transition_ukey(new_ukey
, UKEY_VISIBLE
);
1884 ovs_mutex_unlock(&umap
->mutex
);
1890 transition_ukey_at(struct udpif_key
*ukey
, enum ukey_state dst
,
1892 OVS_REQUIRES(ukey
->mutex
)
1894 if (dst
< ukey
->state
) {
1895 VLOG_ABORT("Invalid ukey transition %d->%d (last transitioned from "
1896 "thread %u at %s)", ukey
->state
, dst
, ukey
->state_thread
,
1899 if (ukey
->state
== dst
&& dst
== UKEY_OPERATIONAL
) {
1903 /* Valid state transitions:
1904 * UKEY_CREATED -> UKEY_VISIBLE
1905 * Ukey is now visible in the umap.
1906 * UKEY_VISIBLE -> UKEY_OPERATIONAL
1907 * A handler has installed the flow, and the flow is in the datapath.
1908 * UKEY_VISIBLE -> UKEY_EVICTING
1909 * A handler installs the flow, then revalidator sweeps the ukey before
1910 * the flow is dumped. Most likely the flow was installed; start trying
1912 * UKEY_VISIBLE -> UKEY_EVICTED
1913 * A handler attempts to install the flow, but the datapath rejects it.
1914 * Consider that the datapath has already destroyed it.
1915 * UKEY_OPERATIONAL -> UKEY_EVICTING
1916 * A revalidator decides to evict the datapath flow.
1917 * UKEY_EVICTING -> UKEY_EVICTED
1918 * A revalidator has evicted the datapath flow.
1919 * UKEY_EVICTED -> UKEY_DELETED
1920 * A revalidator has removed the ukey from the umap and is deleting it.
1922 if (ukey
->state
== dst
- 1 || (ukey
->state
== UKEY_VISIBLE
&&
1923 dst
< UKEY_DELETED
)) {
1926 struct ds ds
= DS_EMPTY_INITIALIZER
;
1928 odp_format_ufid(&ukey
->ufid
, &ds
);
1929 VLOG_WARN_RL(&rl
, "Invalid state transition for ukey %s: %d -> %d",
1930 ds_cstr(&ds
), ukey
->state
, dst
);
1933 ukey
->state_thread
= ovsthread_id_self();
1934 ukey
->state_where
= where
;
1938 ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
)
1942 installed
= ukey_install__(udpif
, ukey
);
1944 ovs_mutex_unlock(&ukey
->mutex
);
1950 /* Searches for a ukey in 'udpif->ukeys' that matches 'flow' and attempts to
1951 * lock the ukey. If the ukey does not exist, create it.
1953 * Returns 0 on success, setting *result to the matching ukey and returning it
1954 * in a locked state. Otherwise, returns an errno and clears *result. EBUSY
1955 * indicates that another thread is handling this flow. Other errors indicate
1956 * an unexpected condition creating a new ukey.
1958 * *error is an output parameter provided to appease the threadsafety analyser,
1959 * and its value matches the return value. */
1961 ukey_acquire(struct udpif
*udpif
, const struct dpif_flow
*flow
,
1962 struct udpif_key
**result
, int *error
)
1963 OVS_TRY_LOCK(0, (*result
)->mutex
)
1965 struct udpif_key
*ukey
;
1968 ukey
= ukey_lookup(udpif
, &flow
->ufid
, flow
->pmd_id
);
1970 retval
= ovs_mutex_trylock(&ukey
->mutex
);
1972 /* Usually we try to avoid installing flows from revalidator threads,
1973 * because locking on a umap may cause handler threads to block.
1974 * However there are certain cases, like when ovs-vswitchd is
1975 * restarted, where it is desirable to handle flows that exist in the
1976 * datapath gracefully (ie, don't just clear the datapath). */
1979 retval
= ukey_create_from_dpif_flow(udpif
, flow
, &ukey
);
1983 install
= ukey_install__(udpif
, ukey
);
1987 ukey_delete__(ukey
);
2003 ukey_delete__(struct udpif_key
*ukey
)
2004 OVS_NO_THREAD_SAFETY_ANALYSIS
2007 if (ukey
->key_recirc_id
) {
2008 recirc_free_id(ukey
->key_recirc_id
);
2010 recirc_refs_unref(&ukey
->recircs
);
2011 xlate_cache_delete(ukey
->xcache
);
2012 ofpbuf_delete(ovsrcu_get(struct ofpbuf
*, &ukey
->actions
));
2013 ovs_mutex_destroy(&ukey
->mutex
);
2019 ukey_delete(struct umap
*umap
, struct udpif_key
*ukey
)
2020 OVS_REQUIRES(umap
->mutex
)
2022 ovs_mutex_lock(&ukey
->mutex
);
2023 if (ukey
->state
< UKEY_DELETED
) {
2024 cmap_remove(&umap
->cmap
, &ukey
->cmap_node
, ukey
->hash
);
2025 ovsrcu_postpone(ukey_delete__
, ukey
);
2026 transition_ukey(ukey
, UKEY_DELETED
);
2028 ovs_mutex_unlock(&ukey
->mutex
);
2032 should_revalidate(const struct udpif
*udpif
, uint64_t packets
,
2035 long long int metric
, now
, duration
;
2038 /* Always revalidate the first time a flow is dumped. */
2042 if (udpif
->dump_duration
< 200) {
2043 /* We are likely to handle full revalidation for the flows. */
2047 /* Calculate the mean time between seeing these packets. If this
2048 * exceeds the threshold, then delete the flow rather than performing
2049 * costly revalidation for flows that aren't being hit frequently.
2051 * This is targeted at situations where the dump_duration is high (~1s),
2052 * and revalidation is triggered by a call to udpif_revalidate(). In
2053 * these situations, revalidation of all flows causes fluctuations in the
2054 * flow_limit due to the interaction with the dump_duration and max_idle.
2055 * This tends to result in deletion of low-throughput flows anyway, so
2056 * skip the revalidation and just delete those flows. */
2057 packets
= MAX(packets
, 1);
2058 now
= MAX(used
, time_msec());
2059 duration
= now
- used
;
2060 metric
= duration
/ packets
;
2063 /* The flow is receiving more than ~5pps, so keep it. */
2069 struct reval_context
{
2070 /* Optional output parameters */
2071 struct flow_wildcards
*wc
;
2072 struct ofpbuf
*odp_actions
;
2073 struct netflow
**netflow
;
2074 struct xlate_cache
*xcache
;
2076 /* Required output parameters */
2077 struct xlate_out xout
;
2081 /* Translates 'key' into a flow, populating 'ctx' as it goes along.
2083 * Returns 0 on success, otherwise a positive errno value.
2085 * The caller is responsible for uninitializing ctx->xout on success.
2088 xlate_key(struct udpif
*udpif
, const struct nlattr
*key
, unsigned int len
,
2089 const struct dpif_flow_stats
*push
, struct reval_context
*ctx
)
2091 struct ofproto_dpif
*ofproto
;
2092 ofp_port_t ofp_in_port
;
2093 enum odp_key_fitness fitness
;
2094 struct xlate_in xin
;
2097 fitness
= odp_flow_key_to_flow(key
, len
, &ctx
->flow
, NULL
);
2098 if (fitness
== ODP_FIT_ERROR
) {
2102 error
= xlate_lookup(udpif
->backer
, &ctx
->flow
, &ofproto
, NULL
, NULL
,
2103 ctx
->netflow
, &ofp_in_port
);
2108 xlate_in_init(&xin
, ofproto
, ofproto_dpif_get_tables_version(ofproto
),
2109 &ctx
->flow
, ofp_in_port
, NULL
, push
->tcp_flags
,
2110 NULL
, ctx
->wc
, ctx
->odp_actions
);
2111 if (push
->n_packets
) {
2112 xin
.resubmit_stats
= push
;
2113 xin
.allow_side_effects
= true;
2115 xin
.xcache
= ctx
->xcache
;
2116 xlate_actions(&xin
, &ctx
->xout
);
2117 if (fitness
== ODP_FIT_TOO_LITTLE
) {
2118 ctx
->xout
.slow
|= SLOW_MATCH
;
2125 xlate_ukey(struct udpif
*udpif
, const struct udpif_key
*ukey
,
2126 uint16_t tcp_flags
, struct reval_context
*ctx
)
2128 struct dpif_flow_stats push
= {
2129 .tcp_flags
= tcp_flags
,
2131 return xlate_key(udpif
, ukey
->key
, ukey
->key_len
, &push
, ctx
);
2135 populate_xcache(struct udpif
*udpif
, struct udpif_key
*ukey
,
2137 OVS_REQUIRES(ukey
->mutex
)
2139 struct reval_context ctx
= {
2140 .odp_actions
= NULL
,
2146 ovs_assert(!ukey
->xcache
);
2147 ukey
->xcache
= ctx
.xcache
= xlate_cache_new();
2148 error
= xlate_ukey(udpif
, ukey
, tcp_flags
, &ctx
);
2152 xlate_out_uninit(&ctx
.xout
);
2157 static enum reval_result
2158 revalidate_ukey__(struct udpif
*udpif
, const struct udpif_key
*ukey
,
2159 uint16_t tcp_flags
, struct ofpbuf
*odp_actions
,
2160 struct recirc_refs
*recircs
, struct xlate_cache
*xcache
)
2162 struct xlate_out
*xoutp
;
2163 struct netflow
*netflow
;
2164 struct flow_wildcards dp_mask
, wc
;
2165 enum reval_result result
;
2166 struct reval_context ctx
= {
2167 .odp_actions
= odp_actions
,
2168 .netflow
= &netflow
,
2173 result
= UKEY_DELETE
;
2177 if (xlate_ukey(udpif
, ukey
, tcp_flags
, &ctx
)) {
2182 if (xoutp
->avoid_caching
) {
2187 struct ofproto_dpif
*ofproto
;
2188 ofp_port_t ofp_in_port
;
2190 ofproto
= xlate_lookup_ofproto(udpif
->backer
, &ctx
.flow
, &ofp_in_port
,
2193 ofpbuf_clear(odp_actions
);
2199 compose_slow_path(udpif
, xoutp
, ctx
.flow
.in_port
.odp_port
,
2200 ofp_in_port
, odp_actions
,
2201 ofproto
->up
.slowpath_meter_id
, &ofproto
->uuid
);
2204 if (odp_flow_key_to_mask(ukey
->mask
, ukey
->mask_len
, &dp_mask
, &ctx
.flow
,
2210 /* Do not modify if any bit is wildcarded by the installed datapath flow,
2211 * but not the newly revalidated wildcard mask (wc), i.e., if revalidation
2212 * tells that the datapath flow is now too generic and must be narrowed
2213 * down. Note that we do not know if the datapath has ignored any of the
2214 * wildcarded bits, so we may be overly conservative here. */
2215 if (flow_wildcards_has_extra(&dp_mask
, ctx
.wc
)) {
2219 if (!ofpbuf_equal(odp_actions
,
2220 ovsrcu_get(struct ofpbuf
*, &ukey
->actions
))) {
2221 /* The datapath mask was OK, but the actions seem to have changed.
2222 * Let's modify it in place. */
2223 result
= UKEY_MODIFY
;
2224 /* Transfer recirc action ID references to the caller. */
2225 recirc_refs_swap(recircs
, &xoutp
->recircs
);
2232 if (netflow
&& result
== UKEY_DELETE
) {
2233 netflow_flow_clear(netflow
, &ctx
.flow
);
2235 xlate_out_uninit(xoutp
);
2239 /* Verifies that the datapath actions of 'ukey' are still correct, and pushes
2242 * Returns a recommended action for 'ukey', options include:
2243 * UKEY_DELETE The ukey should be deleted.
2244 * UKEY_KEEP The ukey is fine as is.
2245 * UKEY_MODIFY The ukey's actions should be changed but is otherwise
2246 * fine. Callers should change the actions to those found
2247 * in the caller supplied 'odp_actions' buffer. The
2248 * recirculation references can be found in 'recircs' and
2249 * must be handled by the caller.
2251 * If the result is UKEY_MODIFY, then references to all recirc_ids used by the
2252 * new flow will be held within 'recircs' (which may be none).
2254 * The caller is responsible for both initializing 'recircs' prior this call,
2255 * and ensuring any references are eventually freed.
2257 static enum reval_result
2258 revalidate_ukey(struct udpif
*udpif
, struct udpif_key
*ukey
,
2259 const struct dpif_flow_stats
*stats
,
2260 struct ofpbuf
*odp_actions
, uint64_t reval_seq
,
2261 struct recirc_refs
*recircs
)
2262 OVS_REQUIRES(ukey
->mutex
)
2264 bool need_revalidate
= ukey
->reval_seq
!= reval_seq
;
2265 enum reval_result result
= UKEY_DELETE
;
2266 struct dpif_flow_stats push
;
2268 ofpbuf_clear(odp_actions
);
2270 push
.used
= stats
->used
;
2271 push
.tcp_flags
= stats
->tcp_flags
;
2272 push
.n_packets
= (stats
->n_packets
> ukey
->stats
.n_packets
2273 ? stats
->n_packets
- ukey
->stats
.n_packets
2275 push
.n_bytes
= (stats
->n_bytes
> ukey
->stats
.n_bytes
2276 ? stats
->n_bytes
- ukey
->stats
.n_bytes
2279 if (need_revalidate
) {
2280 if (should_revalidate(udpif
, push
.n_packets
, ukey
->stats
.used
)) {
2281 if (!ukey
->xcache
) {
2282 ukey
->xcache
= xlate_cache_new();
2284 xlate_cache_clear(ukey
->xcache
);
2286 result
= revalidate_ukey__(udpif
, ukey
, push
.tcp_flags
,
2287 odp_actions
, recircs
, ukey
->xcache
);
2288 } /* else delete; too expensive to revalidate */
2289 } else if (!push
.n_packets
|| ukey
->xcache
2290 || !populate_xcache(udpif
, ukey
, push
.tcp_flags
)) {
2294 /* Stats for deleted flows will be attributed upon flow deletion. Skip. */
2295 if (result
!= UKEY_DELETE
) {
2296 xlate_push_stats(ukey
->xcache
, &push
);
2297 ukey
->stats
= *stats
;
2298 ukey
->reval_seq
= reval_seq
;
2305 delete_op_init__(struct udpif
*udpif
, struct ukey_op
*op
,
2306 const struct dpif_flow
*flow
)
2309 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
2310 op
->dop
.flow_del
.key
= flow
->key
;
2311 op
->dop
.flow_del
.key_len
= flow
->key_len
;
2312 op
->dop
.flow_del
.ufid
= flow
->ufid_present
? &flow
->ufid
: NULL
;
2313 op
->dop
.flow_del
.pmd_id
= flow
->pmd_id
;
2314 op
->dop
.flow_del
.stats
= &op
->stats
;
2315 op
->dop
.flow_del
.terse
= udpif_use_ufid(udpif
);
2319 delete_op_init(struct udpif
*udpif
, struct ukey_op
*op
, struct udpif_key
*ukey
)
2322 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
2323 op
->dop
.flow_del
.key
= ukey
->key
;
2324 op
->dop
.flow_del
.key_len
= ukey
->key_len
;
2325 op
->dop
.flow_del
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
2326 op
->dop
.flow_del
.pmd_id
= ukey
->pmd_id
;
2327 op
->dop
.flow_del
.stats
= &op
->stats
;
2328 op
->dop
.flow_del
.terse
= udpif_use_ufid(udpif
);
2332 put_op_init(struct ukey_op
*op
, struct udpif_key
*ukey
,
2333 enum dpif_flow_put_flags flags
)
2336 op
->dop
.type
= DPIF_OP_FLOW_PUT
;
2337 op
->dop
.flow_put
.flags
= flags
;
2338 op
->dop
.flow_put
.key
= ukey
->key
;
2339 op
->dop
.flow_put
.key_len
= ukey
->key_len
;
2340 op
->dop
.flow_put
.mask
= ukey
->mask
;
2341 op
->dop
.flow_put
.mask_len
= ukey
->mask_len
;
2342 op
->dop
.flow_put
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
2343 op
->dop
.flow_put
.pmd_id
= ukey
->pmd_id
;
2344 op
->dop
.flow_put
.stats
= NULL
;
2345 ukey_get_actions(ukey
, &op
->dop
.flow_put
.actions
,
2346 &op
->dop
.flow_put
.actions_len
);
2349 /* Executes datapath operations 'ops' and attributes stats retrieved from the
2350 * datapath as part of those operations. */
2352 push_dp_ops(struct udpif
*udpif
, struct ukey_op
*ops
, size_t n_ops
)
2354 struct dpif_op
*opsp
[REVALIDATE_MAX_BATCH
];
2357 ovs_assert(n_ops
<= REVALIDATE_MAX_BATCH
);
2358 for (i
= 0; i
< n_ops
; i
++) {
2359 opsp
[i
] = &ops
[i
].dop
;
2361 dpif_operate(udpif
->dpif
, opsp
, n_ops
, DPIF_OFFLOAD_AUTO
);
2363 for (i
= 0; i
< n_ops
; i
++) {
2364 struct ukey_op
*op
= &ops
[i
];
2365 struct dpif_flow_stats
*push
, *stats
, push_buf
;
2367 stats
= op
->dop
.flow_del
.stats
;
2370 if (op
->dop
.type
!= DPIF_OP_FLOW_DEL
) {
2371 /* Only deleted flows need their stats pushed. */
2375 if (op
->dop
.error
) {
2376 /* flow_del error, 'stats' is unusable. */
2378 ovs_mutex_lock(&op
->ukey
->mutex
);
2379 transition_ukey(op
->ukey
, UKEY_EVICTED
);
2380 ovs_mutex_unlock(&op
->ukey
->mutex
);
2386 ovs_mutex_lock(&op
->ukey
->mutex
);
2387 transition_ukey(op
->ukey
, UKEY_EVICTED
);
2388 push
->used
= MAX(stats
->used
, op
->ukey
->stats
.used
);
2389 push
->tcp_flags
= stats
->tcp_flags
| op
->ukey
->stats
.tcp_flags
;
2390 push
->n_packets
= stats
->n_packets
- op
->ukey
->stats
.n_packets
;
2391 push
->n_bytes
= stats
->n_bytes
- op
->ukey
->stats
.n_bytes
;
2392 ovs_mutex_unlock(&op
->ukey
->mutex
);
2397 if (push
->n_packets
|| netflow_exists()) {
2398 const struct nlattr
*key
= op
->dop
.flow_del
.key
;
2399 size_t key_len
= op
->dop
.flow_del
.key_len
;
2400 struct netflow
*netflow
;
2401 struct reval_context ctx
= {
2402 .netflow
= &netflow
,
2407 ovs_mutex_lock(&op
->ukey
->mutex
);
2408 if (op
->ukey
->xcache
) {
2409 xlate_push_stats(op
->ukey
->xcache
, push
);
2410 ovs_mutex_unlock(&op
->ukey
->mutex
);
2413 ovs_mutex_unlock(&op
->ukey
->mutex
);
2414 key
= op
->ukey
->key
;
2415 key_len
= op
->ukey
->key_len
;
2418 error
= xlate_key(udpif
, key
, key_len
, push
, &ctx
);
2420 static struct vlog_rate_limit rll
= VLOG_RATE_LIMIT_INIT(1, 5);
2421 VLOG_WARN_RL(&rll
, "xlate_key failed (%s)!",
2422 ovs_strerror(error
));
2424 xlate_out_uninit(&ctx
.xout
);
2426 netflow_flow_clear(netflow
, &ctx
.flow
);
2433 /* Executes datapath operations 'ops', attributes stats retrieved from the
2434 * datapath, and deletes ukeys corresponding to deleted flows. */
2436 push_ukey_ops(struct udpif
*udpif
, struct umap
*umap
,
2437 struct ukey_op
*ops
, size_t n_ops
)
2441 push_dp_ops(udpif
, ops
, n_ops
);
2442 ovs_mutex_lock(&umap
->mutex
);
2443 for (i
= 0; i
< n_ops
; i
++) {
2444 if (ops
[i
].dop
.type
== DPIF_OP_FLOW_DEL
) {
2445 ukey_delete(umap
, ops
[i
].ukey
);
2448 ovs_mutex_unlock(&umap
->mutex
);
2452 log_unexpected_flow(const struct dpif_flow
*flow
, int error
)
2454 struct ds ds
= DS_EMPTY_INITIALIZER
;
2456 ds_put_format(&ds
, "Failed to acquire udpif_key corresponding to "
2457 "unexpected flow (%s): ", ovs_strerror(error
));
2458 odp_format_ufid(&flow
->ufid
, &ds
);
2460 static struct vlog_rate_limit rll
= VLOG_RATE_LIMIT_INIT(10, 60);
2461 VLOG_WARN_RL(&rll
, "%s", ds_cstr(&ds
));
2467 reval_op_init(struct ukey_op
*op
, enum reval_result result
,
2468 struct udpif
*udpif
, struct udpif_key
*ukey
,
2469 struct recirc_refs
*recircs
, struct ofpbuf
*odp_actions
)
2470 OVS_REQUIRES(ukey
->mutex
)
2472 if (result
== UKEY_DELETE
) {
2473 delete_op_init(udpif
, op
, ukey
);
2474 transition_ukey(ukey
, UKEY_EVICTING
);
2475 } else if (result
== UKEY_MODIFY
) {
2476 /* Store the new recircs. */
2477 recirc_refs_swap(&ukey
->recircs
, recircs
);
2478 /* Release old recircs. */
2479 recirc_refs_unref(recircs
);
2480 /* ukey->key_recirc_id remains, as the key is the same as before. */
2482 ukey_set_actions(ukey
, odp_actions
);
2483 put_op_init(op
, ukey
, DPIF_FP_MODIFY
);
2488 ukey_netdev_unref(struct udpif_key
*ukey
)
2490 if (!ukey
->in_netdev
) {
2493 netdev_close(ukey
->in_netdev
);
2494 ukey
->in_netdev
= NULL
;
2498 * Given a udpif_key, get its input port (netdev) by parsing the flow keys
2499 * and actions. The flow may not contain flow attributes if it is a terse
2500 * dump; read its attributes from the ukey and then parse the flow to get
2501 * the port info. Save them in udpif_key.
2504 ukey_to_flow_netdev(struct udpif
*udpif
, struct udpif_key
*ukey
)
2506 const struct dpif
*dpif
= udpif
->dpif
;
2507 const struct dpif_class
*dpif_class
= dpif
->dpif_class
;
2508 const struct nlattr
*k
;
2511 /* Remove existing references to netdev */
2512 ukey_netdev_unref(ukey
);
2514 /* Find the input port and get a reference to its netdev */
2515 NL_ATTR_FOR_EACH (k
, left
, ukey
->key
, ukey
->key_len
) {
2516 enum ovs_key_attr type
= nl_attr_type(k
);
2518 if (type
== OVS_KEY_ATTR_IN_PORT
) {
2519 ukey
->in_netdev
= netdev_ports_get(nl_attr_get_odp_port(k
),
2521 } else if (type
== OVS_KEY_ATTR_TUNNEL
) {
2522 struct flow_tnl tnl
;
2523 enum odp_key_fitness res
;
2525 if (ukey
->in_netdev
) {
2526 netdev_close(ukey
->in_netdev
);
2527 ukey
->in_netdev
= NULL
;
2529 res
= odp_tun_key_from_attr(k
, &tnl
, NULL
);
2530 if (res
!= ODP_FIT_ERROR
) {
2531 ukey
->in_netdev
= flow_get_tunnel_netdev(&tnl
);
2539 udpif_flow_packet_delta(struct udpif_key
*ukey
, const struct dpif_flow
*f
)
2541 return f
->stats
.n_packets
+ ukey
->flow_backlog_packets
-
2545 static long long int
2546 udpif_flow_time_delta(struct udpif
*udpif
, struct udpif_key
*ukey
)
2548 return (udpif
->dpif
->current_ms
- ukey
->flow_time
) / 1000;
2552 * Save backlog packet count while switching modes
2553 * between offloaded and kernel datapaths.
2556 udpif_set_ukey_backlog_packets(struct udpif_key
*ukey
)
2558 ukey
->flow_backlog_packets
= ukey
->flow_packets
;
2561 /* Gather pps-rate for the given dpif_flow and save it in its ukey */
2563 udpif_update_flow_pps(struct udpif
*udpif
, struct udpif_key
*ukey
,
2564 const struct dpif_flow
*f
)
2568 /* Update pps-rate only when we are close to rebalance interval */
2569 if (udpif
->dpif
->current_ms
- ukey
->flow_time
< OFFL_REBAL_INTVL_MSEC
) {
2573 ukey
->offloaded
= f
->attrs
.offloaded
;
2574 pps
= udpif_flow_packet_delta(ukey
, f
) /
2575 udpif_flow_time_delta(udpif
, ukey
);
2576 ukey
->flow_pps_rate
= pps
;
2577 ukey
->flow_packets
= ukey
->flow_backlog_packets
+ f
->stats
.n_packets
;
2578 ukey
->flow_time
= udpif
->dpif
->current_ms
;
2582 revalidate(struct revalidator
*revalidator
)
2584 uint64_t odp_actions_stub
[1024 / 8];
2585 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2587 struct udpif
*udpif
= revalidator
->udpif
;
2588 struct dpif_flow_dump_thread
*dump_thread
;
2589 uint64_t dump_seq
, reval_seq
;
2590 unsigned int flow_limit
;
2592 dump_seq
= seq_read(udpif
->dump_seq
);
2593 reval_seq
= seq_read(udpif
->reval_seq
);
2594 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2595 dump_thread
= dpif_flow_dump_thread_create(udpif
->dump
);
2597 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2600 struct dpif_flow flows
[REVALIDATE_MAX_BATCH
];
2601 const struct dpif_flow
*f
;
2604 long long int max_idle
;
2609 n_dumped
= dpif_flow_dump_next(dump_thread
, flows
, ARRAY_SIZE(flows
));
2616 /* In normal operation we want to keep flows around until they have
2617 * been idle for 'ofproto_max_idle' milliseconds. However:
2619 * - If the number of datapath flows climbs above 'flow_limit',
2620 * drop that down to 100 ms to try to bring the flows down to
2623 * - If the number of datapath flows climbs above twice
2624 * 'flow_limit', delete all the datapath flows as an emergency
2625 * measure. (We reassess this condition for the next batch of
2626 * datapath flows, so we will recover before all the flows are
2628 n_dp_flows
= udpif_get_n_flows(udpif
);
2629 kill_them_all
= n_dp_flows
> flow_limit
* 2;
2630 max_idle
= n_dp_flows
> flow_limit
? 100 : ofproto_max_idle
;
2632 udpif
->dpif
->current_ms
= time_msec();
2633 for (f
= flows
; f
< &flows
[n_dumped
]; f
++) {
2634 long long int used
= f
->stats
.used
;
2635 struct recirc_refs recircs
= RECIRC_REFS_EMPTY_INITIALIZER
;
2636 enum reval_result result
;
2637 struct udpif_key
*ukey
;
2638 bool already_dumped
;
2641 if (ukey_acquire(udpif
, f
, &ukey
, &error
)) {
2642 if (error
== EBUSY
) {
2643 /* Another thread is processing this flow, so don't bother
2645 COVERAGE_INC(upcall_ukey_contention
);
2647 log_unexpected_flow(f
, error
);
2648 if (error
!= ENOENT
) {
2649 delete_op_init__(udpif
, &ops
[n_ops
++], f
);
2655 already_dumped
= ukey
->dump_seq
== dump_seq
;
2656 if (already_dumped
) {
2657 /* The flow has already been handled during this flow dump
2658 * operation. Skip it. */
2660 COVERAGE_INC(dumped_duplicate_flow
);
2662 COVERAGE_INC(dumped_new_flow
);
2664 ovs_mutex_unlock(&ukey
->mutex
);
2668 if (ukey
->state
<= UKEY_OPERATIONAL
) {
2669 /* The flow is now confirmed to be in the datapath. */
2670 transition_ukey(ukey
, UKEY_OPERATIONAL
);
2672 VLOG_INFO("Unexpected ukey transition from state %d "
2673 "(last transitioned from thread %u at %s)",
2674 ukey
->state
, ukey
->state_thread
, ukey
->state_where
);
2675 ovs_mutex_unlock(&ukey
->mutex
);
2680 used
= ukey
->created
;
2682 if (kill_them_all
|| (used
&& used
< now
- max_idle
)) {
2683 result
= UKEY_DELETE
;
2685 result
= revalidate_ukey(udpif
, ukey
, &f
->stats
, &odp_actions
,
2686 reval_seq
, &recircs
);
2688 ukey
->dump_seq
= dump_seq
;
2690 if (netdev_is_offload_rebalance_policy_enabled() &&
2691 result
!= UKEY_DELETE
) {
2692 udpif_update_flow_pps(udpif
, ukey
, f
);
2695 if (result
!= UKEY_KEEP
) {
2696 /* Takes ownership of 'recircs'. */
2697 reval_op_init(&ops
[n_ops
++], result
, udpif
, ukey
, &recircs
,
2700 ovs_mutex_unlock(&ukey
->mutex
);
2704 /* Push datapath ops but defer ukey deletion to 'sweep' phase. */
2705 push_dp_ops(udpif
, ops
, n_ops
);
2709 dpif_flow_dump_thread_destroy(dump_thread
);
2710 ofpbuf_uninit(&odp_actions
);
2713 /* Pauses the 'revalidator', can only proceed after main thread
2714 * calls udpif_resume_revalidators(). */
2716 revalidator_pause(struct revalidator
*revalidator
)
2718 /* The first block is for sync'ing the pause with main thread. */
2719 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2720 /* The second block is for pausing until main thread resumes. */
2721 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2725 revalidator_sweep__(struct revalidator
*revalidator
, bool purge
)
2727 struct udpif
*udpif
;
2728 uint64_t dump_seq
, reval_seq
;
2731 udpif
= revalidator
->udpif
;
2732 dump_seq
= seq_read(udpif
->dump_seq
);
2733 reval_seq
= seq_read(udpif
->reval_seq
);
2734 slice
= revalidator
- udpif
->revalidators
;
2735 ovs_assert(slice
< udpif
->n_revalidators
);
2737 for (int i
= slice
; i
< N_UMAPS
; i
+= udpif
->n_revalidators
) {
2738 uint64_t odp_actions_stub
[1024 / 8];
2739 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2741 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2742 struct udpif_key
*ukey
;
2743 struct umap
*umap
= &udpif
->ukeys
[i
];
2746 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2747 enum ukey_state ukey_state
;
2749 /* Handler threads could be holding a ukey lock while it installs a
2750 * new flow, so don't hang around waiting for access to it. */
2751 if (ovs_mutex_trylock(&ukey
->mutex
)) {
2754 ukey_state
= ukey
->state
;
2755 if (ukey_state
== UKEY_OPERATIONAL
2756 || (ukey_state
== UKEY_VISIBLE
&& purge
)) {
2757 struct recirc_refs recircs
= RECIRC_REFS_EMPTY_INITIALIZER
;
2758 bool seq_mismatch
= (ukey
->dump_seq
!= dump_seq
2759 && ukey
->reval_seq
!= reval_seq
);
2760 enum reval_result result
;
2763 result
= UKEY_DELETE
;
2764 } else if (!seq_mismatch
) {
2767 struct dpif_flow_stats stats
;
2768 COVERAGE_INC(revalidate_missed_dp_flow
);
2769 memset(&stats
, 0, sizeof stats
);
2770 result
= revalidate_ukey(udpif
, ukey
, &stats
, &odp_actions
,
2771 reval_seq
, &recircs
);
2773 if (result
!= UKEY_KEEP
) {
2774 /* Clears 'recircs' if filled by revalidate_ukey(). */
2775 reval_op_init(&ops
[n_ops
++], result
, udpif
, ukey
, &recircs
,
2779 ovs_mutex_unlock(&ukey
->mutex
);
2781 if (ukey_state
== UKEY_EVICTED
) {
2782 /* The common flow deletion case involves deletion of the flow
2783 * during the dump phase and ukey deletion here. */
2784 ovs_mutex_lock(&umap
->mutex
);
2785 ukey_delete(umap
, ukey
);
2786 ovs_mutex_unlock(&umap
->mutex
);
2789 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2790 /* Update/delete missed flows and clean up corresponding ukeys
2792 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2798 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2801 ofpbuf_uninit(&odp_actions
);
2807 revalidator_sweep(struct revalidator
*revalidator
)
2809 revalidator_sweep__(revalidator
, false);
2813 revalidator_purge(struct revalidator
*revalidator
)
2815 revalidator_sweep__(revalidator
, true);
2818 /* In reaction to dpif purge, purges all 'ukey's with same 'pmd_id'. */
2820 dp_purge_cb(void *aux
, unsigned pmd_id
)
2821 OVS_NO_THREAD_SAFETY_ANALYSIS
2823 struct udpif
*udpif
= aux
;
2826 udpif_pause_revalidators(udpif
);
2827 for (i
= 0; i
< N_UMAPS
; i
++) {
2828 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2829 struct udpif_key
*ukey
;
2830 struct umap
*umap
= &udpif
->ukeys
[i
];
2833 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2834 if (ukey
->pmd_id
== pmd_id
) {
2835 delete_op_init(udpif
, &ops
[n_ops
++], ukey
);
2836 transition_ukey(ukey
, UKEY_EVICTING
);
2838 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2839 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2846 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2851 udpif_resume_revalidators(udpif
);
2855 upcall_unixctl_show(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2856 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2858 struct ds ds
= DS_EMPTY_INITIALIZER
;
2859 struct udpif
*udpif
;
2861 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2862 unsigned int flow_limit
;
2866 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2867 ufid_enabled
= udpif_use_ufid(udpif
);
2869 ds_put_format(&ds
, "%s:\n", dpif_name(udpif
->dpif
));
2870 ds_put_format(&ds
, " flows : (current %lu)"
2871 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif
),
2872 udpif
->avg_n_flows
, udpif
->max_n_flows
, flow_limit
);
2873 ds_put_format(&ds
, " dump duration : %lldms\n", udpif
->dump_duration
);
2874 ds_put_format(&ds
, " ufid enabled : ");
2876 ds_put_format(&ds
, "true\n");
2878 ds_put_format(&ds
, "false\n");
2880 ds_put_char(&ds
, '\n');
2882 for (i
= 0; i
< n_revalidators
; i
++) {
2883 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
2884 int j
, elements
= 0;
2886 for (j
= i
; j
< N_UMAPS
; j
+= n_revalidators
) {
2887 elements
+= cmap_count(&udpif
->ukeys
[j
].cmap
);
2889 ds_put_format(&ds
, " %u: (keys %d)\n", revalidator
->id
, elements
);
2893 unixctl_command_reply(conn
, ds_cstr(&ds
));
2897 /* Disable using the megaflows.
2899 * This command is only needed for advanced debugging, so it's not
2900 * documented in the man page. */
2902 upcall_unixctl_disable_megaflows(struct unixctl_conn
*conn
,
2903 int argc OVS_UNUSED
,
2904 const char *argv
[] OVS_UNUSED
,
2905 void *aux OVS_UNUSED
)
2907 atomic_store_relaxed(&enable_megaflows
, false);
2908 udpif_flush_all_datapaths();
2909 unixctl_command_reply(conn
, "megaflows disabled");
2912 /* Re-enable using megaflows.
2914 * This command is only needed for advanced debugging, so it's not
2915 * documented in the man page. */
2917 upcall_unixctl_enable_megaflows(struct unixctl_conn
*conn
,
2918 int argc OVS_UNUSED
,
2919 const char *argv
[] OVS_UNUSED
,
2920 void *aux OVS_UNUSED
)
2922 atomic_store_relaxed(&enable_megaflows
, true);
2923 udpif_flush_all_datapaths();
2924 unixctl_command_reply(conn
, "megaflows enabled");
2927 /* Disable skipping flow attributes during flow dump.
2929 * This command is only needed for advanced debugging, so it's not
2930 * documented in the man page. */
2932 upcall_unixctl_disable_ufid(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2933 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2935 atomic_store_relaxed(&enable_ufid
, false);
2936 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID disabled");
2939 /* Re-enable skipping flow attributes during flow dump.
2941 * This command is only needed for advanced debugging, so it's not documented
2942 * in the man page. */
2944 upcall_unixctl_enable_ufid(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2945 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2947 atomic_store_relaxed(&enable_ufid
, true);
2948 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID enabled "
2949 "for supported datapaths");
2952 /* Set the flow limit.
2954 * This command is only needed for advanced debugging, so it's not
2955 * documented in the man page. */
2957 upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
,
2958 int argc OVS_UNUSED
,
2960 void *aux OVS_UNUSED
)
2962 struct ds ds
= DS_EMPTY_INITIALIZER
;
2963 struct udpif
*udpif
;
2964 unsigned int flow_limit
= atoi(argv
[1]);
2966 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2967 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
2969 ds_put_format(&ds
, "set flow_limit to %u\n", flow_limit
);
2970 unixctl_command_reply(conn
, ds_cstr(&ds
));
2975 upcall_unixctl_dump_wait(struct unixctl_conn
*conn
,
2976 int argc OVS_UNUSED
,
2977 const char *argv
[] OVS_UNUSED
,
2978 void *aux OVS_UNUSED
)
2980 if (ovs_list_is_singleton(&all_udpifs
)) {
2981 struct udpif
*udpif
= NULL
;
2984 udpif
= OBJECT_CONTAINING(ovs_list_front(&all_udpifs
), udpif
, list_node
);
2985 len
= (udpif
->n_conns
+ 1) * sizeof *udpif
->conns
;
2986 udpif
->conn_seq
= seq_read(udpif
->dump_seq
);
2987 udpif
->conns
= xrealloc(udpif
->conns
, len
);
2988 udpif
->conns
[udpif
->n_conns
++] = conn
;
2990 unixctl_command_reply_error(conn
, "can't wait on multiple udpifs.");
2995 upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2996 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2998 struct udpif
*udpif
;
3000 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
3003 for (n
= 0; n
< udpif
->n_revalidators
; n
++) {
3004 revalidator_purge(&udpif
->revalidators
[n
]);
3007 unixctl_command_reply(conn
, "");
3010 /* Flows are sorted in the following order:
3011 * netdev, flow state (offloaded/kernel path), flow_pps_rate.
3014 flow_compare_rebalance(const void *elem1
, const void *elem2
)
3016 const struct udpif_key
*f1
= *(struct udpif_key
**)elem1
;
3017 const struct udpif_key
*f2
= *(struct udpif_key
**)elem2
;
3020 if (f1
->in_netdev
< f2
->in_netdev
) {
3022 } else if (f1
->in_netdev
> f2
->in_netdev
) {
3026 if (f1
->offloaded
!= f2
->offloaded
) {
3027 return f2
->offloaded
- f1
->offloaded
;
3030 diff
= (f1
->offloaded
== true) ?
3031 f1
->flow_pps_rate
- f2
->flow_pps_rate
:
3032 f2
->flow_pps_rate
- f1
->flow_pps_rate
;
3034 return (diff
< 0) ? -1 : 1;
3037 /* Insert flows from pending array during rebalancing */
3039 rebalance_insert_pending(struct udpif
*udpif
, struct udpif_key
**pending_flows
,
3040 int pending_count
, int insert_count
,
3041 uint64_t rate_threshold
)
3045 for (int i
= 0; i
< pending_count
; i
++) {
3046 struct udpif_key
*flow
= pending_flows
[i
];
3049 /* Stop offloading pending flows if the insert count is
3050 * reached and the flow rate is less than the threshold
3052 if (count
>= insert_count
&& flow
->flow_pps_rate
< rate_threshold
) {
3056 /* Offload the flow to netdev */
3057 err
= udpif_flow_program(udpif
, flow
, DPIF_OFFLOAD_ALWAYS
);
3059 if (err
== ENOSPC
) {
3060 /* Stop if we are out of resources */
3068 /* Offload succeeded; delete it from the kernel datapath */
3069 udpif_flow_unprogram(udpif
, flow
, DPIF_OFFLOAD_NEVER
);
3071 /* Change the state of the flow, adjust dpif counters */
3072 flow
->offloaded
= true;
3074 udpif_set_ukey_backlog_packets(flow
);
3081 /* Remove flows from offloaded array during rebalancing */
3083 rebalance_remove_offloaded(struct udpif
*udpif
,
3084 struct udpif_key
**offloaded_flows
,
3087 for (int i
= 0; i
< offload_count
; i
++) {
3088 struct udpif_key
*flow
= offloaded_flows
[i
];
3091 /* Install the flow into kernel path first */
3092 err
= udpif_flow_program(udpif
, flow
, DPIF_OFFLOAD_NEVER
);
3097 /* Success; now remove offloaded flow from netdev */
3098 err
= udpif_flow_unprogram(udpif
, flow
, DPIF_OFFLOAD_ALWAYS
);
3100 udpif_flow_unprogram(udpif
, flow
, DPIF_OFFLOAD_NEVER
);
3103 udpif_set_ukey_backlog_packets(flow
);
3104 flow
->offloaded
= false;
3109 * Rebalance offloaded flows on a netdev that's in OOR state.
3111 * The rebalancing is done in two phases. In the first phase, we check if
3112 * the pending flows can be offloaded (if some resources became available
3113 * in the meantime) by trying to offload each pending flow. If all pending
3114 * flows get successfully offloaded, the OOR state is cleared on the netdev
3115 * and there's nothing to rebalance.
3117 * If some of the pending flows could not be offloaded, i.e, we still see
3118 * the OOR error, then we move to the second phase of rebalancing. In this
3119 * phase, the rebalancer compares pps-rate of an offloaded flow with the
3120 * least pps-rate with that of a pending flow with the highest pps-rate from
3121 * their respective sorted arrays. If pps-rate of the offloaded flow is less
3122 * than the pps-rate of the pending flow, then it deletes the offloaded flow
3123 * from the HW/netdev and adds it to kernel datapath and then offloads pending
3124 * to HW/netdev. This process is repeated for every pair of offloaded and
3125 * pending flows in the ordered list. The process stops when we encounter an
3126 * offloaded flow that has a higher pps-rate than the corresponding pending
3127 * flow. The entire rebalancing process is repeated in the next iteration.
3130 rebalance_device(struct udpif
*udpif
, struct udpif_key
**offloaded_flows
,
3131 int offload_count
, struct udpif_key
**pending_flows
,
3136 int num_inserted
= rebalance_insert_pending(udpif
, pending_flows
,
3137 pending_count
, pending_count
,
3140 VLOG_DBG("Offload rebalance: Phase1: inserted %d pending flows",
3144 /* Adjust pending array */
3145 pending_flows
= &pending_flows
[num_inserted
];
3146 pending_count
-= num_inserted
;
3148 if (!pending_count
) {
3150 * Successfully offloaded all pending flows. The device
3151 * is no longer in OOR state; done rebalancing this device.
3157 * Phase 2; determine how many offloaded flows to churn.
3159 #define OFFL_REBAL_MAX_CHURN 1024
3160 int churn_count
= 0;
3161 while (churn_count
< OFFL_REBAL_MAX_CHURN
&& churn_count
< offload_count
3162 && churn_count
< pending_count
) {
3163 if (pending_flows
[churn_count
]->flow_pps_rate
<=
3164 offloaded_flows
[churn_count
]->flow_pps_rate
)
3170 VLOG_DBG("Offload rebalance: Phase2: removing %d offloaded flows",
3174 /* Bail early if nothing to churn */
3179 /* Remove offloaded flows */
3180 rebalance_remove_offloaded(udpif
, offloaded_flows
, churn_count
);
3182 /* Adjust offloaded array */
3183 offloaded_flows
= &offloaded_flows
[churn_count
];
3184 offload_count
-= churn_count
;
3186 /* Replace offloaded flows with pending flows */
3187 num_inserted
= rebalance_insert_pending(udpif
, pending_flows
,
3188 pending_count
, churn_count
,
3190 offloaded_flows
[0]->flow_pps_rate
:
3193 VLOG_DBG("Offload rebalance: Phase2: inserted %d pending flows",
3200 static struct udpif_key
**
3201 udpif_add_oor_flows(struct udpif_key
**sort_flows
, size_t *total_flow_count
,
3202 size_t *alloc_flow_count
, struct udpif_key
*ukey
)
3204 if (*total_flow_count
>= *alloc_flow_count
) {
3205 sort_flows
= x2nrealloc(sort_flows
, alloc_flow_count
, sizeof ukey
);
3207 sort_flows
[(*total_flow_count
)++] = ukey
;
3212 * Build sort_flows[] initially with flows that
3213 * reference an 'OOR' netdev as their input port.
3215 static struct udpif_key
**
3216 udpif_build_oor_flows(struct udpif_key
**sort_flows
, size_t *total_flow_count
,
3217 size_t *alloc_flow_count
, struct udpif_key
*ukey
,
3218 int *oor_netdev_count
)
3220 struct netdev
*netdev
;
3223 /* Input netdev must be available for the flow */
3224 netdev
= ukey
->in_netdev
;
3229 /* Is the in-netdev for this flow in OOR state ? */
3230 if (!netdev_get_hw_info(netdev
, HW_INFO_TYPE_OOR
)) {
3231 ukey_netdev_unref(ukey
);
3235 /* Add the flow to sort_flows[] */
3236 sort_flows
= udpif_add_oor_flows(sort_flows
, total_flow_count
,
3237 alloc_flow_count
, ukey
);
3238 if (ukey
->offloaded
) {
3239 count
= netdev_get_hw_info(netdev
, HW_INFO_TYPE_OFFL_COUNT
);
3240 ovs_assert(count
>= 0);
3242 (*oor_netdev_count
)++;
3244 netdev_set_hw_info(netdev
, HW_INFO_TYPE_OFFL_COUNT
, count
);
3246 count
= netdev_get_hw_info(netdev
, HW_INFO_TYPE_PEND_COUNT
);
3247 ovs_assert(count
>= 0);
3248 netdev_set_hw_info(netdev
, HW_INFO_TYPE_PEND_COUNT
, ++count
);
3255 * Rebalance offloaded flows on HW netdevs that are in OOR state.
3258 udpif_flow_rebalance(struct udpif
*udpif
)
3260 struct udpif_key
**sort_flows
= NULL
;
3261 size_t alloc_flow_count
= 0;
3262 size_t total_flow_count
= 0;
3263 int oor_netdev_count
= 0;
3264 int offload_index
= 0;
3267 /* Collect flows (offloaded and pending) that reference OOR netdevs */
3268 for (size_t i
= 0; i
< N_UMAPS
; i
++) {
3269 struct udpif_key
*ukey
;
3270 struct umap
*umap
= &udpif
->ukeys
[i
];
3272 CMAP_FOR_EACH (ukey
, cmap_node
, &umap
->cmap
) {
3273 ukey_to_flow_netdev(udpif
, ukey
);
3274 sort_flows
= udpif_build_oor_flows(sort_flows
, &total_flow_count
,
3275 &alloc_flow_count
, ukey
,
3280 /* Sort flows by OOR netdevs, state (offloaded/pending) and pps-rate */
3281 qsort(sort_flows
, total_flow_count
, sizeof(struct udpif_key
*),
3282 flow_compare_rebalance
);
3285 * We now have flows referencing OOR netdevs, that are sorted. We also
3286 * have a count of offloaded and pending flows on each of the netdevs
3287 * that are in OOR state. Now rebalance each oor-netdev.
3289 while (oor_netdev_count
) {
3290 struct netdev
*netdev
;
3295 netdev
= sort_flows
[offload_index
]->in_netdev
;
3296 ovs_assert(netdev_get_hw_info(netdev
, HW_INFO_TYPE_OOR
) == true);
3297 VLOG_DBG("Offload rebalance: netdev: %s is OOR", netdev
->name
);
3299 offload_count
= netdev_get_hw_info(netdev
, HW_INFO_TYPE_OFFL_COUNT
);
3300 pending_count
= netdev_get_hw_info(netdev
, HW_INFO_TYPE_PEND_COUNT
);
3301 pending_index
= offload_index
+ offload_count
;
3303 oor
= rebalance_device(udpif
,
3304 &sort_flows
[offload_index
], offload_count
,
3305 &sort_flows
[pending_index
], pending_count
);
3306 netdev_set_hw_info(netdev
, HW_INFO_TYPE_OOR
, oor
);
3308 offload_index
= pending_index
+ pending_count
;
3309 netdev_set_hw_info(netdev
, HW_INFO_TYPE_OFFL_COUNT
, 0);
3310 netdev_set_hw_info(netdev
, HW_INFO_TYPE_PEND_COUNT
, 0);
3314 for (int i
= 0; i
< total_flow_count
; i
++) {
3315 struct udpif_key
*ukey
= sort_flows
[i
];
3316 ukey_netdev_unref(ukey
);
3322 udpif_flow_program(struct udpif
*udpif
, struct udpif_key
*ukey
,
3323 enum dpif_offload_type offload_type
)
3325 struct dpif_op
*opsp
;
3329 put_op_init(&uop
, ukey
, DPIF_FP_CREATE
);
3330 dpif_operate(udpif
->dpif
, &opsp
, 1, offload_type
);
3336 udpif_flow_unprogram(struct udpif
*udpif
, struct udpif_key
*ukey
,
3337 enum dpif_offload_type offload_type
)
3339 struct dpif_op
*opsp
;
3343 delete_op_init(udpif
, &uop
, ukey
);
3344 dpif_operate(udpif
->dpif
, &opsp
, 1, offload_type
);