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 */
222 enum upcall_type type
; /* Type of the upcall. */
223 const struct nlattr
*actions
; /* Flow actions in DPIF_UC_ACTION Upcalls. */
225 bool xout_initialized
; /* True if 'xout' must be uninitialized. */
226 struct xlate_out xout
; /* Result of xlate_actions(). */
227 struct ofpbuf odp_actions
; /* Datapath actions from xlate_actions(). */
228 struct flow_wildcards wc
; /* Dependencies that megaflow must match. */
229 struct ofpbuf put_actions
; /* Actions 'put' in the fastpath. */
231 struct dpif_ipfix
*ipfix
; /* IPFIX pointer or NULL. */
232 struct dpif_sflow
*sflow
; /* SFlow pointer or NULL. */
234 struct udpif_key
*ukey
; /* Revalidator flow cache. */
235 bool ukey_persists
; /* Set true to keep 'ukey' beyond the
236 lifetime of this upcall. */
238 uint64_t reval_seq
; /* udpif->reval_seq at translation time. */
240 /* Not used by the upcall callback interface. */
241 const struct nlattr
*key
; /* Datapath flow key. */
242 size_t key_len
; /* Datapath flow key length. */
243 const struct nlattr
*out_tun_key
; /* Datapath output tunnel key. */
245 struct user_action_cookie cookie
;
247 uint64_t odp_actions_stub
[1024 / 8]; /* Stub for odp_actions. */
250 /* Ukeys must transition through these states using transition_ukey(). */
253 UKEY_VISIBLE
, /* Ukey is in umap, datapath flow install is queued. */
254 UKEY_OPERATIONAL
, /* Ukey is in umap, datapath flow is installed. */
255 UKEY_EVICTING
, /* Ukey is in umap, datapath flow delete is queued. */
256 UKEY_EVICTED
, /* Ukey is in umap, datapath flow is deleted. */
257 UKEY_DELETED
, /* Ukey removed from umap, ukey free is deferred. */
259 #define N_UKEY_STATES (UKEY_DELETED + 1)
261 /* 'udpif_key's are responsible for tracking the little bit of state udpif
262 * needs to do flow expiration which can't be pulled directly from the
263 * datapath. They may be created by any handler or revalidator thread at any
264 * time, and read by any revalidator during the dump phase. They are however
265 * each owned by a single revalidator which takes care of destroying them
266 * during the garbage-collection phase.
268 * The mutex within the ukey protects some members of the ukey. The ukey
269 * itself is protected by RCU and is held within a umap in the parent udpif.
270 * Adding or removing a ukey from a umap is only safe when holding the
271 * corresponding umap lock. */
273 struct cmap_node cmap_node
; /* In parent revalidator 'ukeys' map. */
275 /* These elements are read only once created, and therefore aren't
276 * protected by a mutex. */
277 const struct nlattr
*key
; /* Datapath flow key. */
278 size_t key_len
; /* Length of 'key'. */
279 const struct nlattr
*mask
; /* Datapath flow mask. */
280 size_t mask_len
; /* Length of 'mask'. */
281 ovs_u128 ufid
; /* Unique flow identifier. */
282 bool ufid_present
; /* True if 'ufid' is in datapath. */
283 uint32_t hash
; /* Pre-computed hash for 'key'. */
284 unsigned pmd_id
; /* Datapath poll mode driver id. */
286 struct ovs_mutex mutex
; /* Guards the following. */
287 struct dpif_flow_stats stats OVS_GUARDED
; /* Last known stats.*/
288 long long int created OVS_GUARDED
; /* Estimate of creation time. */
289 uint64_t dump_seq OVS_GUARDED
; /* Tracks udpif->dump_seq. */
290 uint64_t reval_seq OVS_GUARDED
; /* Tracks udpif->reval_seq. */
291 enum ukey_state state OVS_GUARDED
; /* Tracks ukey lifetime. */
293 /* 'state' debug information. */
294 unsigned int state_thread OVS_GUARDED
; /* Thread that transitions. */
295 const char *state_where OVS_GUARDED
; /* transition_ukey() locator. */
297 /* Datapath flow actions as nlattrs. Protected by RCU. Read with
298 * ukey_get_actions(), and write with ukey_set_actions(). */
299 OVSRCU_TYPE(struct ofpbuf
*) actions
;
301 struct xlate_cache
*xcache OVS_GUARDED
; /* Cache for xlate entries that
302 * are affected by this ukey.
303 * Used for stats and learning.*/
305 struct odputil_keybuf buf
;
309 uint32_t key_recirc_id
; /* Non-zero if reference is held by the ukey. */
310 struct recirc_refs recircs
; /* Action recirc IDs with references held. */
312 #define OFFL_REBAL_INTVL_MSEC 3000 /* dynamic offload rebalance freq */
313 struct netdev
*in_netdev
; /* in_odp_port's netdev */
314 bool offloaded
; /* True if flow is offloaded */
315 uint64_t flow_pps_rate
; /* Packets-Per-Second rate */
316 long long int flow_time
; /* last pps update time */
317 uint64_t flow_packets
; /* #pkts seen in interval */
318 uint64_t flow_backlog_packets
; /* prev-mode #pkts (offl or kernel) */
321 /* Datapath operation with optional ukey attached. */
323 struct udpif_key
*ukey
;
324 struct dpif_flow_stats stats
; /* Stats for 'op'. */
325 struct dpif_op dop
; /* Flow operation. */
328 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
329 static struct ovs_list all_udpifs
= OVS_LIST_INITIALIZER(&all_udpifs
);
331 static size_t recv_upcalls(struct handler
*);
332 static int process_upcall(struct udpif
*, struct upcall
*,
333 struct ofpbuf
*odp_actions
, struct flow_wildcards
*);
334 static void handle_upcalls(struct udpif
*, struct upcall
*, size_t n_upcalls
);
335 static void udpif_stop_threads(struct udpif
*);
336 static void udpif_start_threads(struct udpif
*, size_t n_handlers
,
337 size_t n_revalidators
);
338 static void udpif_pause_revalidators(struct udpif
*);
339 static void udpif_resume_revalidators(struct udpif
*);
340 static void *udpif_upcall_handler(void *);
341 static void *udpif_revalidator(void *);
342 static unsigned long udpif_get_n_flows(struct udpif
*);
343 static void revalidate(struct revalidator
*);
344 static void revalidator_pause(struct revalidator
*);
345 static void revalidator_sweep(struct revalidator
*);
346 static void revalidator_purge(struct revalidator
*);
347 static void upcall_unixctl_show(struct unixctl_conn
*conn
, int argc
,
348 const char *argv
[], void *aux
);
349 static void upcall_unixctl_disable_megaflows(struct unixctl_conn
*, int argc
,
350 const char *argv
[], void *aux
);
351 static void upcall_unixctl_enable_megaflows(struct unixctl_conn
*, int argc
,
352 const char *argv
[], void *aux
);
353 static void upcall_unixctl_disable_ufid(struct unixctl_conn
*, int argc
,
354 const char *argv
[], void *aux
);
355 static void upcall_unixctl_enable_ufid(struct unixctl_conn
*, int argc
,
356 const char *argv
[], void *aux
);
357 static void upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
, int argc
,
358 const char *argv
[], void *aux
);
359 static void upcall_unixctl_dump_wait(struct unixctl_conn
*conn
, int argc
,
360 const char *argv
[], void *aux
);
361 static void upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc
,
362 const char *argv
[], void *aux
);
364 static struct udpif_key
*ukey_create_from_upcall(struct upcall
*,
365 struct flow_wildcards
*);
366 static int ukey_create_from_dpif_flow(const struct udpif
*,
367 const struct dpif_flow
*,
368 struct udpif_key
**);
369 static void ukey_get_actions(struct udpif_key
*, const struct nlattr
**actions
,
371 static bool ukey_install__(struct udpif
*, struct udpif_key
*ukey
)
372 OVS_TRY_LOCK(true, ukey
->mutex
);
373 static bool ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
);
374 static void transition_ukey_at(struct udpif_key
*ukey
, enum ukey_state dst
,
376 OVS_REQUIRES(ukey
->mutex
);
377 #define transition_ukey(UKEY, DST) \
378 transition_ukey_at(UKEY, DST, OVS_SOURCE_LOCATOR)
379 static struct udpif_key
*ukey_lookup(struct udpif
*udpif
,
380 const ovs_u128
*ufid
,
381 const unsigned pmd_id
);
382 static int ukey_acquire(struct udpif
*, const struct dpif_flow
*,
383 struct udpif_key
**result
, int *error
);
384 static void ukey_delete__(struct udpif_key
*);
385 static void ukey_delete(struct umap
*, struct udpif_key
*);
386 static enum upcall_type
classify_upcall(enum dpif_upcall_type type
,
387 const struct nlattr
*userdata
,
388 struct user_action_cookie
*cookie
);
390 static void put_op_init(struct ukey_op
*op
, struct udpif_key
*ukey
,
391 enum dpif_flow_put_flags flags
);
392 static void delete_op_init(struct udpif
*udpif
, struct ukey_op
*op
,
393 struct udpif_key
*ukey
);
395 static int upcall_receive(struct upcall
*, const struct dpif_backer
*,
396 const struct dp_packet
*packet
, enum dpif_upcall_type
,
397 const struct nlattr
*userdata
, const struct flow
*,
398 const unsigned int mru
,
399 const ovs_u128
*ufid
, const unsigned pmd_id
);
400 static void upcall_uninit(struct upcall
*);
402 static void udpif_flow_rebalance(struct udpif
*udpif
);
403 static int udpif_flow_program(struct udpif
*udpif
, struct udpif_key
*ukey
,
404 enum dpif_offload_type offload_type
);
405 static int udpif_flow_unprogram(struct udpif
*udpif
, struct udpif_key
*ukey
,
406 enum dpif_offload_type offload_type
);
408 static upcall_callback upcall_cb
;
409 static dp_purge_callback dp_purge_cb
;
411 static atomic_bool enable_megaflows
= ATOMIC_VAR_INIT(true);
412 static atomic_bool enable_ufid
= ATOMIC_VAR_INIT(true);
417 static struct ovsthread_once once
= OVSTHREAD_ONCE_INITIALIZER
;
418 if (ovsthread_once_start(&once
)) {
419 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show
,
421 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
422 upcall_unixctl_disable_megaflows
, NULL
);
423 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
424 upcall_unixctl_enable_megaflows
, NULL
);
425 unixctl_command_register("upcall/disable-ufid", "", 0, 0,
426 upcall_unixctl_disable_ufid
, NULL
);
427 unixctl_command_register("upcall/enable-ufid", "", 0, 0,
428 upcall_unixctl_enable_ufid
, NULL
);
429 unixctl_command_register("upcall/set-flow-limit", "flow-limit-number",
430 1, 1, upcall_unixctl_set_flow_limit
, NULL
);
431 unixctl_command_register("revalidator/wait", "", 0, 0,
432 upcall_unixctl_dump_wait
, NULL
);
433 unixctl_command_register("revalidator/purge", "", 0, 0,
434 upcall_unixctl_purge
, NULL
);
435 ovsthread_once_done(&once
);
440 udpif_create(struct dpif_backer
*backer
, struct dpif
*dpif
)
442 struct udpif
*udpif
= xzalloc(sizeof *udpif
);
445 udpif
->backer
= backer
;
446 atomic_init(&udpif
->flow_limit
, MIN(ofproto_flow_limit
, 10000));
447 udpif
->reval_seq
= seq_create();
448 udpif
->dump_seq
= seq_create();
449 latch_init(&udpif
->exit_latch
);
450 latch_init(&udpif
->pause_latch
);
451 ovs_list_push_back(&all_udpifs
, &udpif
->list_node
);
452 atomic_init(&udpif
->enable_ufid
, false);
453 atomic_init(&udpif
->n_flows
, 0);
454 atomic_init(&udpif
->n_flows_timestamp
, LLONG_MIN
);
455 ovs_mutex_init(&udpif
->n_flows_mutex
);
456 udpif
->ukeys
= xmalloc(N_UMAPS
* sizeof *udpif
->ukeys
);
457 for (int i
= 0; i
< N_UMAPS
; i
++) {
458 cmap_init(&udpif
->ukeys
[i
].cmap
);
459 ovs_mutex_init(&udpif
->ukeys
[i
].mutex
);
462 dpif_register_upcall_cb(dpif
, upcall_cb
, udpif
);
463 dpif_register_dp_purge_cb(dpif
, dp_purge_cb
, udpif
);
469 udpif_run(struct udpif
*udpif
)
471 if (udpif
->conns
&& udpif
->conn_seq
!= seq_read(udpif
->dump_seq
)) {
474 for (i
= 0; i
< udpif
->n_conns
; i
++) {
475 unixctl_command_reply(udpif
->conns
[i
], NULL
);
484 udpif_destroy(struct udpif
*udpif
)
486 udpif_stop_threads(udpif
);
488 dpif_register_dp_purge_cb(udpif
->dpif
, NULL
, udpif
);
489 dpif_register_upcall_cb(udpif
->dpif
, NULL
, udpif
);
491 for (int i
= 0; i
< N_UMAPS
; i
++) {
492 cmap_destroy(&udpif
->ukeys
[i
].cmap
);
493 ovs_mutex_destroy(&udpif
->ukeys
[i
].mutex
);
498 ovs_list_remove(&udpif
->list_node
);
499 latch_destroy(&udpif
->exit_latch
);
500 latch_destroy(&udpif
->pause_latch
);
501 seq_destroy(udpif
->reval_seq
);
502 seq_destroy(udpif
->dump_seq
);
503 ovs_mutex_destroy(&udpif
->n_flows_mutex
);
507 /* Stops the handler and revalidator threads. */
509 udpif_stop_threads(struct udpif
*udpif
)
511 if (udpif
&& (udpif
->n_handlers
!= 0 || udpif
->n_revalidators
!= 0)) {
514 /* Tell the threads to exit. */
515 latch_set(&udpif
->exit_latch
);
517 /* Wait for the threads to exit. Quiesce because this can take a long
519 ovsrcu_quiesce_start();
520 for (i
= 0; i
< udpif
->n_handlers
; i
++) {
521 xpthread_join(udpif
->handlers
[i
].thread
, NULL
);
523 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
524 xpthread_join(udpif
->revalidators
[i
].thread
, NULL
);
526 dpif_disable_upcall(udpif
->dpif
);
527 ovsrcu_quiesce_end();
529 /* Delete ukeys, and delete all flows from the datapath to prevent
530 * double-counting stats. */
531 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
532 revalidator_purge(&udpif
->revalidators
[i
]);
535 latch_poll(&udpif
->exit_latch
);
537 ovs_barrier_destroy(&udpif
->reval_barrier
);
538 ovs_barrier_destroy(&udpif
->pause_barrier
);
540 free(udpif
->revalidators
);
541 udpif
->revalidators
= NULL
;
542 udpif
->n_revalidators
= 0;
544 free(udpif
->handlers
);
545 udpif
->handlers
= NULL
;
546 udpif
->n_handlers
= 0;
550 /* Starts the handler and revalidator threads. */
552 udpif_start_threads(struct udpif
*udpif
, size_t n_handlers_
,
553 size_t n_revalidators_
)
555 if (udpif
&& n_handlers_
&& n_revalidators_
) {
556 /* Creating a thread can take a significant amount of time on some
557 * systems, even hundred of milliseconds, so quiesce around it. */
558 ovsrcu_quiesce_start();
560 udpif
->n_handlers
= n_handlers_
;
561 udpif
->n_revalidators
= n_revalidators_
;
563 udpif
->handlers
= xzalloc(udpif
->n_handlers
* sizeof *udpif
->handlers
);
564 for (size_t i
= 0; i
< udpif
->n_handlers
; i
++) {
565 struct handler
*handler
= &udpif
->handlers
[i
];
567 handler
->udpif
= udpif
;
568 handler
->handler_id
= i
;
569 handler
->thread
= ovs_thread_create(
570 "handler", udpif_upcall_handler
, handler
);
573 atomic_init(&udpif
->enable_ufid
, udpif
->backer
->rt_support
.ufid
);
574 dpif_enable_upcall(udpif
->dpif
);
576 ovs_barrier_init(&udpif
->reval_barrier
, udpif
->n_revalidators
);
577 ovs_barrier_init(&udpif
->pause_barrier
, udpif
->n_revalidators
+ 1);
578 udpif
->reval_exit
= false;
579 udpif
->pause
= false;
580 udpif
->offload_rebalance_time
= time_msec();
581 udpif
->revalidators
= xzalloc(udpif
->n_revalidators
582 * sizeof *udpif
->revalidators
);
583 for (size_t i
= 0; i
< udpif
->n_revalidators
; i
++) {
584 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
586 revalidator
->udpif
= udpif
;
587 revalidator
->thread
= ovs_thread_create(
588 "revalidator", udpif_revalidator
, revalidator
);
590 ovsrcu_quiesce_end();
594 /* Pauses all revalidators. Should only be called by the main thread.
595 * When function returns, all revalidators are paused and will proceed
596 * only after udpif_resume_revalidators() is called. */
598 udpif_pause_revalidators(struct udpif
*udpif
)
600 if (udpif
->backer
->recv_set_enable
) {
601 latch_set(&udpif
->pause_latch
);
602 ovs_barrier_block(&udpif
->pause_barrier
);
606 /* Resumes the pausing of revalidators. Should only be called by the
609 udpif_resume_revalidators(struct udpif
*udpif
)
611 if (udpif
->backer
->recv_set_enable
) {
612 latch_poll(&udpif
->pause_latch
);
613 ovs_barrier_block(&udpif
->pause_barrier
);
617 /* Tells 'udpif' how many threads it should use to handle upcalls.
618 * 'n_handlers_' and 'n_revalidators_' can never be zero. 'udpif''s
619 * datapath handle must have packet reception enabled before starting
622 udpif_set_threads(struct udpif
*udpif
, size_t n_handlers_
,
623 size_t n_revalidators_
)
626 ovs_assert(n_handlers_
&& n_revalidators_
);
628 if (udpif
->n_handlers
!= n_handlers_
629 || udpif
->n_revalidators
!= n_revalidators_
) {
630 udpif_stop_threads(udpif
);
633 if (!udpif
->handlers
&& !udpif
->revalidators
) {
636 error
= dpif_handlers_set(udpif
->dpif
, n_handlers_
);
638 VLOG_ERR("failed to configure handlers in dpif %s: %s",
639 dpif_name(udpif
->dpif
), ovs_strerror(error
));
643 udpif_start_threads(udpif
, n_handlers_
, n_revalidators_
);
647 /* Waits for all ongoing upcall translations to complete. This ensures that
648 * there are no transient references to any removed ofprotos (or other
649 * objects). In particular, this should be called after an ofproto is removed
650 * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */
652 udpif_synchronize(struct udpif
*udpif
)
654 /* This is stronger than necessary. It would be sufficient to ensure
655 * (somehow) that each handler and revalidator thread had passed through
656 * its main loop once. */
657 size_t n_handlers_
= udpif
->n_handlers
;
658 size_t n_revalidators_
= udpif
->n_revalidators
;
660 udpif_stop_threads(udpif
);
661 udpif_start_threads(udpif
, n_handlers_
, n_revalidators_
);
664 /* Notifies 'udpif' that something changed which may render previous
665 * xlate_actions() results invalid. */
667 udpif_revalidate(struct udpif
*udpif
)
669 seq_change(udpif
->reval_seq
);
672 /* Returns a seq which increments every time 'udpif' pulls stats from the
673 * datapath. Callers can use this to get a sense of when might be a good time
674 * to do periodic work which relies on relatively up to date statistics. */
676 udpif_dump_seq(struct udpif
*udpif
)
678 return udpif
->dump_seq
;
682 udpif_get_memory_usage(struct udpif
*udpif
, struct simap
*usage
)
686 simap_increase(usage
, "handlers", udpif
->n_handlers
);
688 simap_increase(usage
, "revalidators", udpif
->n_revalidators
);
689 for (i
= 0; i
< N_UMAPS
; i
++) {
690 simap_increase(usage
, "udpif keys", cmap_count(&udpif
->ukeys
[i
].cmap
));
694 /* Remove flows from a single datapath. */
696 udpif_flush(struct udpif
*udpif
)
698 size_t n_handlers_
= udpif
->n_handlers
;
699 size_t n_revalidators_
= udpif
->n_revalidators
;
701 udpif_stop_threads(udpif
);
702 dpif_flow_flush(udpif
->dpif
);
703 udpif_start_threads(udpif
, n_handlers_
, n_revalidators_
);
706 /* Removes all flows from all datapaths. */
708 udpif_flush_all_datapaths(void)
712 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
718 udpif_use_ufid(struct udpif
*udpif
)
722 atomic_read_relaxed(&enable_ufid
, &enable
);
723 return enable
&& udpif
->backer
->rt_support
.ufid
;
728 udpif_get_n_flows(struct udpif
*udpif
)
730 long long int time
, now
;
731 unsigned long flow_count
;
734 atomic_read_relaxed(&udpif
->n_flows_timestamp
, &time
);
735 if (time
< now
- 100 && !ovs_mutex_trylock(&udpif
->n_flows_mutex
)) {
736 struct dpif_dp_stats stats
;
738 atomic_store_relaxed(&udpif
->n_flows_timestamp
, now
);
739 dpif_get_dp_stats(udpif
->dpif
, &stats
);
740 flow_count
= stats
.n_flows
;
741 atomic_store_relaxed(&udpif
->n_flows
, flow_count
);
742 ovs_mutex_unlock(&udpif
->n_flows_mutex
);
744 atomic_read_relaxed(&udpif
->n_flows
, &flow_count
);
749 /* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH
750 * upcalls from dpif, processes the batch and installs corresponding flows
753 udpif_upcall_handler(void *arg
)
755 struct handler
*handler
= arg
;
756 struct udpif
*udpif
= handler
->udpif
;
758 while (!latch_is_set(&handler
->udpif
->exit_latch
)) {
759 if (recv_upcalls(handler
)) {
760 poll_immediate_wake();
762 dpif_recv_wait(udpif
->dpif
, handler
->handler_id
);
763 latch_wait(&udpif
->exit_latch
);
772 recv_upcalls(struct handler
*handler
)
774 struct udpif
*udpif
= handler
->udpif
;
775 uint64_t recv_stubs
[UPCALL_MAX_BATCH
][512 / 8];
776 struct ofpbuf recv_bufs
[UPCALL_MAX_BATCH
];
777 struct dpif_upcall dupcalls
[UPCALL_MAX_BATCH
];
778 struct upcall upcalls
[UPCALL_MAX_BATCH
];
779 struct flow flows
[UPCALL_MAX_BATCH
];
783 while (n_upcalls
< UPCALL_MAX_BATCH
) {
784 struct ofpbuf
*recv_buf
= &recv_bufs
[n_upcalls
];
785 struct dpif_upcall
*dupcall
= &dupcalls
[n_upcalls
];
786 struct upcall
*upcall
= &upcalls
[n_upcalls
];
787 struct flow
*flow
= &flows
[n_upcalls
];
788 unsigned int mru
= 0;
792 ofpbuf_use_stub(recv_buf
, recv_stubs
[n_upcalls
],
793 sizeof recv_stubs
[n_upcalls
]);
794 if (dpif_recv(udpif
->dpif
, handler
->handler_id
, dupcall
, recv_buf
)) {
795 ofpbuf_uninit(recv_buf
);
799 upcall
->fitness
= odp_flow_key_to_flow(dupcall
->key
, dupcall
->key_len
,
801 if (upcall
->fitness
== ODP_FIT_ERROR
) {
806 mru
= nl_attr_get_u16(dupcall
->mru
);
810 hash
= nl_attr_get_u64(dupcall
->hash
);
813 error
= upcall_receive(upcall
, udpif
->backer
, &dupcall
->packet
,
814 dupcall
->type
, dupcall
->userdata
, flow
, mru
,
815 &dupcall
->ufid
, PMD_ID_NULL
);
817 if (error
== ENODEV
) {
818 /* Received packet on datapath port for which we couldn't
819 * associate an ofproto. This can happen if a port is removed
820 * while traffic is being received. Print a rate-limited
821 * message in case it happens frequently. */
822 dpif_flow_put(udpif
->dpif
, DPIF_FP_CREATE
, dupcall
->key
,
823 dupcall
->key_len
, NULL
, 0, NULL
, 0,
824 &dupcall
->ufid
, PMD_ID_NULL
, NULL
);
825 VLOG_INFO_RL(&rl
, "received packet on unassociated datapath "
826 "port %"PRIu32
, flow
->in_port
.odp_port
);
831 upcall
->key
= dupcall
->key
;
832 upcall
->key_len
= dupcall
->key_len
;
833 upcall
->ufid
= &dupcall
->ufid
;
836 upcall
->out_tun_key
= dupcall
->out_tun_key
;
837 upcall
->actions
= dupcall
->actions
;
839 pkt_metadata_from_flow(&dupcall
->packet
.md
, flow
);
840 flow_extract(&dupcall
->packet
, flow
);
842 error
= process_upcall(udpif
, upcall
,
843 &upcall
->odp_actions
, &upcall
->wc
);
852 upcall_uninit(upcall
);
854 dp_packet_uninit(&dupcall
->packet
);
855 ofpbuf_uninit(recv_buf
);
859 handle_upcalls(handler
->udpif
, upcalls
, n_upcalls
);
860 for (i
= 0; i
< n_upcalls
; i
++) {
861 dp_packet_uninit(&dupcalls
[i
].packet
);
862 ofpbuf_uninit(&recv_bufs
[i
]);
863 upcall_uninit(&upcalls
[i
]);
871 udpif_run_flow_rebalance(struct udpif
*udpif
)
873 long long int now
= 0;
875 /* Don't rebalance if OFFL_REBAL_INTVL_MSEC have not elapsed */
877 if (now
< udpif
->offload_rebalance_time
+ OFFL_REBAL_INTVL_MSEC
) {
881 if (!netdev_any_oor()) {
885 VLOG_DBG("Offload rebalance: Found OOR netdevs");
886 udpif
->offload_rebalance_time
= now
;
887 udpif_flow_rebalance(udpif
);
891 udpif_revalidator(void *arg
)
893 /* Used by all revalidators. */
894 struct revalidator
*revalidator
= arg
;
895 struct udpif
*udpif
= revalidator
->udpif
;
896 bool leader
= revalidator
== &udpif
->revalidators
[0];
898 /* Used only by the leader. */
899 long long int start_time
= 0;
900 uint64_t last_reval_seq
= 0;
903 revalidator
->id
= ovsthread_id_self();
908 recirc_run(); /* Recirculation cleanup. */
910 reval_seq
= seq_read(udpif
->reval_seq
);
911 last_reval_seq
= reval_seq
;
913 n_flows
= udpif_get_n_flows(udpif
);
914 udpif
->max_n_flows
= MAX(n_flows
, udpif
->max_n_flows
);
915 udpif
->avg_n_flows
= (udpif
->avg_n_flows
+ n_flows
) / 2;
917 /* Only the leader checks the pause latch to prevent a race where
918 * some threads think it's false and proceed to block on
919 * reval_barrier and others think it's true and block indefinitely
920 * on the pause_barrier */
921 udpif
->pause
= latch_is_set(&udpif
->pause_latch
);
923 /* Only the leader checks the exit latch to prevent a race where
924 * some threads think it's true and exit and others think it's
925 * false and block indefinitely on the reval_barrier */
926 udpif
->reval_exit
= latch_is_set(&udpif
->exit_latch
);
928 start_time
= time_msec();
929 if (!udpif
->reval_exit
) {
932 terse_dump
= udpif_use_ufid(udpif
);
933 udpif
->dump
= dpif_flow_dump_create(udpif
->dpif
, terse_dump
,
938 /* Wait for the leader to start the flow dump. */
939 ovs_barrier_block(&udpif
->reval_barrier
);
941 revalidator_pause(revalidator
);
944 if (udpif
->reval_exit
) {
947 revalidate(revalidator
);
949 /* Wait for all flows to have been dumped before we garbage collect. */
950 ovs_barrier_block(&udpif
->reval_barrier
);
951 revalidator_sweep(revalidator
);
953 /* Wait for all revalidators to finish garbage collection. */
954 ovs_barrier_block(&udpif
->reval_barrier
);
957 unsigned int flow_limit
;
958 long long int duration
;
960 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
962 dpif_flow_dump_destroy(udpif
->dump
);
963 seq_change(udpif
->dump_seq
);
964 if (netdev_is_offload_rebalance_policy_enabled()) {
965 udpif_run_flow_rebalance(udpif
);
968 duration
= MAX(time_msec() - start_time
, 1);
969 udpif
->dump_duration
= duration
;
970 if (duration
> 2000) {
971 flow_limit
/= duration
/ 1000;
972 } else if (duration
> 1300) {
973 flow_limit
= flow_limit
* 3 / 4;
974 } else if (duration
< 1000 &&
975 flow_limit
< n_flows
* 1000 / duration
) {
978 flow_limit
= MIN(ofproto_flow_limit
, MAX(flow_limit
, 1000));
979 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
981 if (duration
> 2000) {
982 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
986 poll_timer_wait_until(start_time
+ MIN(ofproto_max_idle
,
987 ofproto_max_revalidator
));
988 seq_wait(udpif
->reval_seq
, last_reval_seq
);
989 latch_wait(&udpif
->exit_latch
);
990 latch_wait(&udpif
->pause_latch
);
993 if (!latch_is_set(&udpif
->pause_latch
) &&
994 !latch_is_set(&udpif
->exit_latch
)) {
995 long long int now
= time_msec();
996 /* Block again if we are woken up within 5ms of the last start
1000 if (now
< start_time
) {
1001 poll_timer_wait_until(start_time
);
1002 latch_wait(&udpif
->exit_latch
);
1003 latch_wait(&udpif
->pause_latch
);
1013 static enum upcall_type
1014 classify_upcall(enum dpif_upcall_type type
, const struct nlattr
*userdata
,
1015 struct user_action_cookie
*cookie
)
1017 /* First look at the upcall type. */
1019 case DPIF_UC_ACTION
:
1025 case DPIF_N_UC_TYPES
:
1027 VLOG_WARN_RL(&rl
, "upcall has unexpected type %"PRIu32
, type
);
1031 /* "action" upcalls need a closer look. */
1033 VLOG_WARN_RL(&rl
, "action upcall missing cookie");
1037 size_t userdata_len
= nl_attr_get_size(userdata
);
1038 if (userdata_len
!= sizeof *cookie
) {
1039 VLOG_WARN_RL(&rl
, "action upcall cookie has unexpected size %"PRIuSIZE
,
1043 memcpy(cookie
, nl_attr_get(userdata
), sizeof *cookie
);
1044 if (cookie
->type
== USER_ACTION_COOKIE_SFLOW
) {
1045 return SFLOW_UPCALL
;
1046 } else if (cookie
->type
== USER_ACTION_COOKIE_SLOW_PATH
) {
1047 return SLOW_PATH_UPCALL
;
1048 } else if (cookie
->type
== USER_ACTION_COOKIE_FLOW_SAMPLE
) {
1049 return FLOW_SAMPLE_UPCALL
;
1050 } else if (cookie
->type
== USER_ACTION_COOKIE_IPFIX
) {
1051 return IPFIX_UPCALL
;
1052 } else if (cookie
->type
== USER_ACTION_COOKIE_CONTROLLER
) {
1053 return CONTROLLER_UPCALL
;
1055 VLOG_WARN_RL(&rl
, "invalid user cookie of type %"PRIu16
1056 " and size %"PRIuSIZE
, cookie
->type
, userdata_len
);
1061 /* Calculates slow path actions for 'xout'. 'buf' must statically be
1062 * initialized with at least 128 bytes of space. */
1064 compose_slow_path(struct udpif
*udpif
, struct xlate_out
*xout
,
1065 odp_port_t odp_in_port
, ofp_port_t ofp_in_port
,
1066 struct ofpbuf
*buf
, uint32_t meter_id
,
1067 struct uuid
*ofproto_uuid
)
1069 struct user_action_cookie cookie
;
1073 memset(&cookie
, 0, sizeof cookie
);
1074 cookie
.type
= USER_ACTION_COOKIE_SLOW_PATH
;
1075 cookie
.ofp_in_port
= ofp_in_port
;
1076 cookie
.ofproto_uuid
= *ofproto_uuid
;
1077 cookie
.slow_path
.reason
= xout
->slow
;
1079 port
= xout
->slow
& (SLOW_CFM
| SLOW_BFD
| SLOW_LACP
| SLOW_STP
)
1082 pid
= dpif_port_get_pid(udpif
->dpif
, port
);
1086 if (meter_id
!= UINT32_MAX
) {
1087 /* If slowpath meter is configured, generate clone(meter, userspace)
1089 offset
= nl_msg_start_nested(buf
, OVS_ACTION_ATTR_SAMPLE
);
1090 nl_msg_put_u32(buf
, OVS_SAMPLE_ATTR_PROBABILITY
, UINT32_MAX
);
1091 ac_offset
= nl_msg_start_nested(buf
, OVS_SAMPLE_ATTR_ACTIONS
);
1092 nl_msg_put_u32(buf
, OVS_ACTION_ATTR_METER
, meter_id
);
1095 odp_put_userspace_action(pid
, &cookie
, sizeof cookie
,
1096 ODPP_NONE
, false, buf
);
1098 if (meter_id
!= UINT32_MAX
) {
1099 nl_msg_end_nested(buf
, ac_offset
);
1100 nl_msg_end_nested(buf
, offset
);
1104 /* If there is no error, the upcall must be destroyed with upcall_uninit()
1105 * before quiescing, as the referred objects are guaranteed to exist only
1106 * until the calling thread quiesces. Otherwise, do not call upcall_uninit()
1107 * since the 'upcall->put_actions' remains uninitialized. */
1109 upcall_receive(struct upcall
*upcall
, const struct dpif_backer
*backer
,
1110 const struct dp_packet
*packet
, enum dpif_upcall_type type
,
1111 const struct nlattr
*userdata
, const struct flow
*flow
,
1112 const unsigned int mru
,
1113 const ovs_u128
*ufid
, const unsigned pmd_id
)
1117 upcall
->type
= classify_upcall(type
, userdata
, &upcall
->cookie
);
1118 if (upcall
->type
== BAD_UPCALL
) {
1120 } else if (upcall
->type
== MISS_UPCALL
) {
1121 error
= xlate_lookup(backer
, flow
, &upcall
->ofproto
, &upcall
->ipfix
,
1122 &upcall
->sflow
, NULL
, &upcall
->ofp_in_port
);
1127 struct ofproto_dpif
*ofproto
1128 = ofproto_dpif_lookup_by_uuid(&upcall
->cookie
.ofproto_uuid
);
1130 VLOG_INFO_RL(&rl
, "upcall could not find ofproto");
1133 upcall
->ofproto
= ofproto
;
1134 upcall
->ipfix
= ofproto
->ipfix
;
1135 upcall
->sflow
= ofproto
->sflow
;
1136 upcall
->ofp_in_port
= upcall
->cookie
.ofp_in_port
;
1139 upcall
->recirc
= NULL
;
1140 upcall
->have_recirc_ref
= false;
1141 upcall
->flow
= flow
;
1142 upcall
->packet
= packet
;
1143 upcall
->ufid
= ufid
;
1144 upcall
->pmd_id
= pmd_id
;
1145 ofpbuf_use_stub(&upcall
->odp_actions
, upcall
->odp_actions_stub
,
1146 sizeof upcall
->odp_actions_stub
);
1147 ofpbuf_init(&upcall
->put_actions
, 0);
1149 upcall
->xout_initialized
= false;
1150 upcall
->ukey_persists
= false;
1152 upcall
->ukey
= NULL
;
1154 upcall
->key_len
= 0;
1157 upcall
->out_tun_key
= NULL
;
1158 upcall
->actions
= NULL
;
1164 upcall_xlate(struct udpif
*udpif
, struct upcall
*upcall
,
1165 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1167 struct dpif_flow_stats stats
;
1168 enum xlate_error xerr
;
1169 struct xlate_in xin
;
1172 stats
.n_packets
= 1;
1173 stats
.n_bytes
= dp_packet_size(upcall
->packet
);
1174 stats
.used
= time_msec();
1175 stats
.tcp_flags
= ntohs(upcall
->flow
->tcp_flags
);
1177 xlate_in_init(&xin
, upcall
->ofproto
,
1178 ofproto_dpif_get_tables_version(upcall
->ofproto
),
1179 upcall
->flow
, upcall
->ofp_in_port
, NULL
,
1180 stats
.tcp_flags
, upcall
->packet
, wc
, odp_actions
);
1182 if (upcall
->type
== MISS_UPCALL
) {
1183 xin
.resubmit_stats
= &stats
;
1185 if (xin
.frozen_state
) {
1186 /* We may install a datapath flow only if we get a reference to the
1187 * recirculation context (otherwise we could have recirculation
1188 * upcalls using recirculation ID for which no context can be
1189 * found). We may still execute the flow's actions even if we
1190 * don't install the flow. */
1191 upcall
->recirc
= recirc_id_node_from_state(xin
.frozen_state
);
1192 upcall
->have_recirc_ref
= recirc_id_node_try_ref_rcu(upcall
->recirc
);
1195 /* For non-miss upcalls, we are either executing actions (one of which
1196 * is an userspace action) for an upcall, in which case the stats have
1197 * already been taken care of, or there's a flow in the datapath which
1198 * this packet was accounted to. Presumably the revalidators will deal
1199 * with pushing its stats eventually. */
1202 upcall
->reval_seq
= seq_read(udpif
->reval_seq
);
1204 xerr
= xlate_actions(&xin
, &upcall
->xout
);
1206 /* Translate again and log the ofproto trace for
1207 * these two error types. */
1208 if (xerr
== XLATE_RECURSION_TOO_DEEP
||
1209 xerr
== XLATE_TOO_MANY_RESUBMITS
) {
1210 static struct vlog_rate_limit rll
= VLOG_RATE_LIMIT_INIT(1, 1);
1212 /* This is a huge log, so be conservative. */
1213 if (!VLOG_DROP_WARN(&rll
)) {
1215 ofproto_trace(upcall
->ofproto
, upcall
->flow
,
1216 upcall
->packet
, NULL
, 0, NULL
, &output
);
1217 VLOG_WARN("%s", ds_cstr(&output
));
1218 ds_destroy(&output
);
1223 /* Convert the input port wildcard from OFP to ODP format. There's no
1224 * real way to do this for arbitrary bitmasks since the numbering spaces
1225 * aren't the same. However, flow translation always exact matches the
1226 * whole thing, so we can do the same here. */
1227 WC_MASK_FIELD(wc
, in_port
.odp_port
);
1230 upcall
->xout_initialized
= true;
1232 if (upcall
->fitness
== ODP_FIT_TOO_LITTLE
) {
1233 upcall
->xout
.slow
|= SLOW_MATCH
;
1235 if (!upcall
->xout
.slow
) {
1236 ofpbuf_use_const(&upcall
->put_actions
,
1237 odp_actions
->data
, odp_actions
->size
);
1239 /* upcall->put_actions already initialized by upcall_receive(). */
1240 compose_slow_path(udpif
, &upcall
->xout
,
1241 upcall
->flow
->in_port
.odp_port
, upcall
->ofp_in_port
,
1242 &upcall
->put_actions
,
1243 upcall
->ofproto
->up
.slowpath_meter_id
,
1244 &upcall
->ofproto
->uuid
);
1247 /* This function is also called for slow-pathed flows. As we are only
1248 * going to create new datapath flows for actual datapath misses, there is
1249 * no point in creating a ukey otherwise. */
1250 if (upcall
->type
== MISS_UPCALL
) {
1251 upcall
->ukey
= ukey_create_from_upcall(upcall
, wc
);
1256 upcall_uninit(struct upcall
*upcall
)
1259 if (upcall
->xout_initialized
) {
1260 xlate_out_uninit(&upcall
->xout
);
1262 ofpbuf_uninit(&upcall
->odp_actions
);
1263 ofpbuf_uninit(&upcall
->put_actions
);
1265 if (!upcall
->ukey_persists
) {
1266 ukey_delete__(upcall
->ukey
);
1268 } else if (upcall
->have_recirc_ref
) {
1269 /* The reference was transferred to the ukey if one was created. */
1270 recirc_id_node_unref(upcall
->recirc
);
1275 /* If there are less flows than the limit, and this is a miss upcall which
1277 * - Has no recirc_id, OR
1278 * - Has a recirc_id and we can get a reference on the recirc ctx,
1280 * Then we should install the flow (true). Otherwise, return false. */
1282 should_install_flow(struct udpif
*udpif
, struct upcall
*upcall
)
1284 unsigned int flow_limit
;
1286 if (upcall
->type
!= MISS_UPCALL
) {
1288 } else if (upcall
->recirc
&& !upcall
->have_recirc_ref
) {
1289 VLOG_DBG_RL(&rl
, "upcall: no reference for recirc flow");
1293 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
1294 if (udpif_get_n_flows(udpif
) >= flow_limit
) {
1295 VLOG_WARN_RL(&rl
, "upcall: datapath flow limit reached");
1303 upcall_cb(const struct dp_packet
*packet
, const struct flow
*flow
, ovs_u128
*ufid
,
1304 unsigned pmd_id
, enum dpif_upcall_type type
,
1305 const struct nlattr
*userdata
, struct ofpbuf
*actions
,
1306 struct flow_wildcards
*wc
, struct ofpbuf
*put_actions
, void *aux
)
1308 struct udpif
*udpif
= aux
;
1309 struct upcall upcall
;
1313 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1315 error
= upcall_receive(&upcall
, udpif
->backer
, packet
, type
, userdata
,
1316 flow
, 0, ufid
, pmd_id
);
1321 upcall
.fitness
= ODP_FIT_PERFECT
;
1322 error
= process_upcall(udpif
, &upcall
, actions
, wc
);
1327 if (upcall
.xout
.slow
&& put_actions
) {
1328 ofpbuf_put(put_actions
, upcall
.put_actions
.data
,
1329 upcall
.put_actions
.size
);
1332 if (OVS_UNLIKELY(!megaflow
&& wc
)) {
1333 flow_wildcards_init_for_packet(wc
, flow
);
1336 if (!should_install_flow(udpif
, &upcall
)) {
1341 if (upcall
.ukey
&& !ukey_install(udpif
, upcall
.ukey
)) {
1342 static struct vlog_rate_limit rll
= VLOG_RATE_LIMIT_INIT(1, 1);
1343 VLOG_WARN_RL(&rll
, "upcall_cb failure: ukey installation fails");
1348 upcall
.ukey_persists
= true;
1350 upcall_uninit(&upcall
);
1355 dpif_get_actions(struct udpif
*udpif
, struct upcall
*upcall
,
1356 const struct nlattr
**actions
)
1358 size_t actions_len
= 0;
1360 if (upcall
->actions
) {
1361 /* Actions were passed up from datapath. */
1362 *actions
= nl_attr_get(upcall
->actions
);
1363 actions_len
= nl_attr_get_size(upcall
->actions
);
1366 if (actions_len
== 0) {
1367 /* Lookup actions in userspace cache. */
1368 struct udpif_key
*ukey
= ukey_lookup(udpif
, upcall
->ufid
,
1371 ukey_get_actions(ukey
, actions
, &actions_len
);
1379 dpif_read_actions(struct udpif
*udpif
, struct upcall
*upcall
,
1380 const struct flow
*flow
, enum upcall_type type
,
1383 const struct nlattr
*actions
= NULL
;
1384 size_t actions_len
= dpif_get_actions(udpif
, upcall
, &actions
);
1386 if (!actions
|| !actions_len
) {
1392 dpif_sflow_read_actions(flow
, actions
, actions_len
, upcall_data
, true);
1394 case FLOW_SAMPLE_UPCALL
:
1396 dpif_ipfix_read_actions(flow
, actions
, actions_len
, upcall_data
);
1400 case SLOW_PATH_UPCALL
:
1401 case CONTROLLER_UPCALL
:
1410 process_upcall(struct udpif
*udpif
, struct upcall
*upcall
,
1411 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1413 const struct dp_packet
*packet
= upcall
->packet
;
1414 const struct flow
*flow
= upcall
->flow
;
1415 size_t actions_len
= 0;
1417 switch (upcall
->type
) {
1419 case SLOW_PATH_UPCALL
:
1420 upcall_xlate(udpif
, upcall
, odp_actions
, wc
);
1424 if (upcall
->sflow
) {
1425 struct dpif_sflow_actions sflow_actions
;
1427 memset(&sflow_actions
, 0, sizeof sflow_actions
);
1429 actions_len
= dpif_read_actions(udpif
, upcall
, flow
,
1430 upcall
->type
, &sflow_actions
);
1431 dpif_sflow_received(upcall
->sflow
, packet
, flow
,
1432 flow
->in_port
.odp_port
, &upcall
->cookie
,
1433 actions_len
> 0 ? &sflow_actions
: NULL
);
1438 case FLOW_SAMPLE_UPCALL
:
1439 if (upcall
->ipfix
) {
1440 struct flow_tnl output_tunnel_key
;
1441 struct dpif_ipfix_actions ipfix_actions
;
1443 memset(&ipfix_actions
, 0, sizeof ipfix_actions
);
1445 if (upcall
->out_tun_key
) {
1446 odp_tun_key_from_attr(upcall
->out_tun_key
, &output_tunnel_key
,
1450 actions_len
= dpif_read_actions(udpif
, upcall
, flow
,
1451 upcall
->type
, &ipfix_actions
);
1452 if (upcall
->type
== IPFIX_UPCALL
) {
1453 dpif_ipfix_bridge_sample(upcall
->ipfix
, packet
, flow
,
1454 flow
->in_port
.odp_port
,
1455 upcall
->cookie
.ipfix
.output_odp_port
,
1456 upcall
->out_tun_key
?
1457 &output_tunnel_key
: NULL
,
1459 &ipfix_actions
: NULL
);
1461 /* The flow reflects exactly the contents of the packet.
1462 * Sample the packet using it. */
1463 dpif_ipfix_flow_sample(upcall
->ipfix
, packet
, flow
,
1464 &upcall
->cookie
, flow
->in_port
.odp_port
,
1465 upcall
->out_tun_key
?
1466 &output_tunnel_key
: NULL
,
1467 actions_len
> 0 ? &ipfix_actions
: NULL
);
1472 case CONTROLLER_UPCALL
:
1474 struct user_action_cookie
*cookie
= &upcall
->cookie
;
1476 if (cookie
->controller
.dont_send
) {
1480 uint32_t recirc_id
= cookie
->controller
.recirc_id
;
1485 const struct recirc_id_node
*recirc_node
1486 = recirc_id_node_find(recirc_id
);
1491 const struct frozen_state
*state
= &recirc_node
->state
;
1493 struct ofproto_async_msg
*am
= xmalloc(sizeof *am
);
1494 *am
= (struct ofproto_async_msg
) {
1495 .controller_id
= cookie
->controller
.controller_id
,
1496 .oam
= OAM_PACKET_IN
,
1500 .packet
= xmemdup(dp_packet_data(packet
),
1501 dp_packet_size(packet
)),
1502 .packet_len
= dp_packet_size(packet
),
1503 .reason
= cookie
->controller
.reason
,
1504 .table_id
= state
->table_id
,
1505 .cookie
= get_32aligned_be64(
1506 &cookie
->controller
.rule_cookie
),
1507 .userdata
= (recirc_node
->state
.userdata_len
1508 ? xmemdup(recirc_node
->state
.userdata
,
1509 recirc_node
->state
.userdata_len
)
1511 .userdata_len
= recirc_node
->state
.userdata_len
,
1514 .max_len
= cookie
->controller
.max_len
,
1518 if (cookie
->controller
.continuation
) {
1519 am
->pin
.up
.stack
= (state
->stack_size
1520 ? xmemdup(state
->stack
, state
->stack_size
)
1522 am
->pin
.up
.stack_size
= state
->stack_size
,
1523 am
->pin
.up
.mirrors
= state
->mirrors
,
1524 am
->pin
.up
.conntracked
= state
->conntracked
,
1525 am
->pin
.up
.actions
= (state
->ofpacts_len
1526 ? xmemdup(state
->ofpacts
,
1527 state
->ofpacts_len
) : NULL
),
1528 am
->pin
.up
.actions_len
= state
->ofpacts_len
,
1529 am
->pin
.up
.action_set
= (state
->action_set_len
1530 ? xmemdup(state
->action_set
,
1531 state
->action_set_len
)
1533 am
->pin
.up
.action_set_len
= state
->action_set_len
,
1534 am
->pin
.up
.bridge
= upcall
->ofproto
->uuid
;
1535 am
->pin
.up
.odp_port
= upcall
->packet
->md
.in_port
.odp_port
;
1538 /* We don't want to use the upcall 'flow', since it may be
1539 * more specific than the point at which the "controller"
1540 * action was specified. */
1541 struct flow frozen_flow
;
1543 frozen_flow
= *flow
;
1544 if (!state
->conntracked
) {
1545 flow_clear_conntrack(&frozen_flow
);
1548 frozen_metadata_to_flow(&state
->metadata
, &frozen_flow
);
1549 flow_get_metadata(&frozen_flow
, &am
->pin
.up
.base
.flow_metadata
);
1551 ofproto_dpif_send_async_msg(upcall
->ofproto
, am
);
1563 handle_upcalls(struct udpif
*udpif
, struct upcall
*upcalls
,
1566 struct dpif_op
*opsp
[UPCALL_MAX_BATCH
* 2];
1567 struct ukey_op ops
[UPCALL_MAX_BATCH
* 2];
1568 size_t n_ops
, n_opsp
, i
;
1570 /* Handle the packets individually in order of arrival.
1572 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, SLOW_BFD, and SLOW_LLDP,
1573 * translation is what processes received packets for these
1576 * - For SLOW_ACTION, translation executes the actions directly.
1578 * The loop fills 'ops' with an array of operations to execute in the
1581 for (i
= 0; i
< n_upcalls
; i
++) {
1582 struct upcall
*upcall
= &upcalls
[i
];
1583 const struct dp_packet
*packet
= upcall
->packet
;
1586 if (should_install_flow(udpif
, upcall
)) {
1587 struct udpif_key
*ukey
= upcall
->ukey
;
1589 if (ukey_install(udpif
, ukey
)) {
1590 upcall
->ukey_persists
= true;
1591 put_op_init(&ops
[n_ops
++], ukey
, DPIF_FP_CREATE
);
1595 if (upcall
->odp_actions
.size
) {
1598 op
->dop
.type
= DPIF_OP_EXECUTE
;
1599 op
->dop
.execute
.packet
= CONST_CAST(struct dp_packet
*, packet
);
1600 op
->dop
.execute
.flow
= upcall
->flow
;
1601 odp_key_to_dp_packet(upcall
->key
, upcall
->key_len
,
1602 op
->dop
.execute
.packet
);
1603 op
->dop
.execute
.actions
= upcall
->odp_actions
.data
;
1604 op
->dop
.execute
.actions_len
= upcall
->odp_actions
.size
;
1605 op
->dop
.execute
.needs_help
= (upcall
->xout
.slow
& SLOW_ACTION
) != 0;
1606 op
->dop
.execute
.probe
= false;
1607 op
->dop
.execute
.mtu
= upcall
->mru
;
1608 op
->dop
.execute
.hash
= upcall
->hash
;
1612 /* Execute batch. */
1614 for (i
= 0; i
< n_ops
; i
++) {
1615 opsp
[n_opsp
++] = &ops
[i
].dop
;
1617 dpif_operate(udpif
->dpif
, opsp
, n_opsp
, DPIF_OFFLOAD_AUTO
);
1618 for (i
= 0; i
< n_ops
; i
++) {
1619 struct udpif_key
*ukey
= ops
[i
].ukey
;
1622 ovs_mutex_lock(&ukey
->mutex
);
1623 if (ops
[i
].dop
.error
) {
1624 transition_ukey(ukey
, UKEY_EVICTED
);
1625 } else if (ukey
->state
< UKEY_OPERATIONAL
) {
1626 transition_ukey(ukey
, UKEY_OPERATIONAL
);
1628 ovs_mutex_unlock(&ukey
->mutex
);
1634 get_ukey_hash(const ovs_u128
*ufid
, const unsigned pmd_id
)
1636 return hash_2words(ufid
->u32
[0], pmd_id
);
1639 static struct udpif_key
*
1640 ukey_lookup(struct udpif
*udpif
, const ovs_u128
*ufid
, const unsigned pmd_id
)
1642 struct udpif_key
*ukey
;
1643 int idx
= get_ukey_hash(ufid
, pmd_id
) % N_UMAPS
;
1644 struct cmap
*cmap
= &udpif
->ukeys
[idx
].cmap
;
1646 CMAP_FOR_EACH_WITH_HASH (ukey
, cmap_node
,
1647 get_ukey_hash(ufid
, pmd_id
), cmap
) {
1648 if (ovs_u128_equals(ukey
->ufid
, *ufid
)) {
1655 /* Provides safe lockless access of RCU protected 'ukey->actions'. Callers may
1656 * alternatively access the field directly if they take 'ukey->mutex'. */
1658 ukey_get_actions(struct udpif_key
*ukey
, const struct nlattr
**actions
, size_t *size
)
1660 const struct ofpbuf
*buf
= ovsrcu_get(struct ofpbuf
*, &ukey
->actions
);
1661 *actions
= buf
->data
;
1666 ukey_set_actions(struct udpif_key
*ukey
, const struct ofpbuf
*actions
)
1668 struct ofpbuf
*old_actions
= ovsrcu_get_protected(struct ofpbuf
*,
1672 ovsrcu_postpone(ofpbuf_delete
, old_actions
);
1675 ovsrcu_set(&ukey
->actions
, ofpbuf_clone(actions
));
1678 static struct udpif_key
*
1679 ukey_create__(const struct nlattr
*key
, size_t key_len
,
1680 const struct nlattr
*mask
, size_t mask_len
,
1681 bool ufid_present
, const ovs_u128
*ufid
,
1682 const unsigned pmd_id
, const struct ofpbuf
*actions
,
1683 uint64_t reval_seq
, long long int used
,
1684 uint32_t key_recirc_id
, struct xlate_out
*xout
)
1685 OVS_NO_THREAD_SAFETY_ANALYSIS
1687 struct udpif_key
*ukey
= xmalloc(sizeof *ukey
);
1689 memcpy(&ukey
->keybuf
, key
, key_len
);
1690 ukey
->key
= &ukey
->keybuf
.nla
;
1691 ukey
->key_len
= key_len
;
1692 memcpy(&ukey
->maskbuf
, mask
, mask_len
);
1693 ukey
->mask
= &ukey
->maskbuf
.nla
;
1694 ukey
->mask_len
= mask_len
;
1695 ukey
->ufid_present
= ufid_present
;
1697 ukey
->pmd_id
= pmd_id
;
1698 ukey
->hash
= get_ukey_hash(&ukey
->ufid
, pmd_id
);
1700 ovsrcu_init(&ukey
->actions
, NULL
);
1701 ukey_set_actions(ukey
, actions
);
1703 ovs_mutex_init(&ukey
->mutex
);
1704 ukey
->dump_seq
= 0; /* Not yet dumped */
1705 ukey
->reval_seq
= reval_seq
;
1706 ukey
->state
= UKEY_CREATED
;
1707 ukey
->state_thread
= ovsthread_id_self();
1708 ukey
->state_where
= OVS_SOURCE_LOCATOR
;
1709 ukey
->created
= ukey
->flow_time
= time_msec();
1710 memset(&ukey
->stats
, 0, sizeof ukey
->stats
);
1711 ukey
->stats
.used
= used
;
1712 ukey
->xcache
= NULL
;
1714 ukey
->offloaded
= false;
1715 ukey
->in_netdev
= NULL
;
1716 ukey
->flow_packets
= ukey
->flow_backlog_packets
= 0;
1718 ukey
->key_recirc_id
= key_recirc_id
;
1719 recirc_refs_init(&ukey
->recircs
);
1721 /* Take ownership of the action recirc id references. */
1722 recirc_refs_swap(&ukey
->recircs
, &xout
->recircs
);
1728 static struct udpif_key
*
1729 ukey_create_from_upcall(struct upcall
*upcall
, struct flow_wildcards
*wc
)
1731 struct odputil_keybuf keystub
, maskstub
;
1732 struct ofpbuf keybuf
, maskbuf
;
1734 struct odp_flow_key_parms odp_parms
= {
1735 .flow
= upcall
->flow
,
1736 .mask
= wc
? &wc
->masks
: NULL
,
1739 odp_parms
.support
= upcall
->ofproto
->backer
->rt_support
.odp
;
1740 if (upcall
->key_len
) {
1741 ofpbuf_use_const(&keybuf
, upcall
->key
, upcall
->key_len
);
1743 /* dpif-netdev doesn't provide a netlink-formatted flow key in the
1744 * upcall, so convert the upcall's flow here. */
1745 ofpbuf_use_stack(&keybuf
, &keystub
, sizeof keystub
);
1746 odp_flow_key_from_flow(&odp_parms
, &keybuf
);
1749 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1750 ofpbuf_use_stack(&maskbuf
, &maskstub
, sizeof maskstub
);
1751 if (megaflow
&& wc
) {
1752 odp_parms
.key_buf
= &keybuf
;
1753 odp_flow_key_from_mask(&odp_parms
, &maskbuf
);
1756 return ukey_create__(keybuf
.data
, keybuf
.size
, maskbuf
.data
, maskbuf
.size
,
1757 true, upcall
->ufid
, upcall
->pmd_id
,
1758 &upcall
->put_actions
, upcall
->reval_seq
, 0,
1759 upcall
->have_recirc_ref
? upcall
->recirc
->id
: 0,
1764 ukey_create_from_dpif_flow(const struct udpif
*udpif
,
1765 const struct dpif_flow
*flow
,
1766 struct udpif_key
**ukey
)
1768 struct dpif_flow full_flow
;
1769 struct ofpbuf actions
;
1771 uint64_t stub
[DPIF_FLOW_BUFSIZE
/ 8];
1772 const struct nlattr
*a
;
1775 if (!flow
->key_len
|| !flow
->actions_len
) {
1779 /* If the key or actions were not provided by the datapath, fetch the
1781 ofpbuf_use_stack(&buf
, &stub
, sizeof stub
);
1782 err
= dpif_flow_get(udpif
->dpif
, flow
->key
, flow
->key_len
,
1783 flow
->ufid_present
? &flow
->ufid
: NULL
,
1784 flow
->pmd_id
, &buf
, &full_flow
);
1791 /* Check the flow actions for recirculation action. As recirculation
1792 * relies on OVS userspace internal state, we need to delete all old
1793 * datapath flows with either a non-zero recirc_id in the key, or any
1794 * recirculation actions upon OVS restart. */
1795 NL_ATTR_FOR_EACH (a
, left
, flow
->key
, flow
->key_len
) {
1796 if (nl_attr_type(a
) == OVS_KEY_ATTR_RECIRC_ID
1797 && nl_attr_get_u32(a
) != 0) {
1801 NL_ATTR_FOR_EACH (a
, left
, flow
->actions
, flow
->actions_len
) {
1802 if (nl_attr_type(a
) == OVS_ACTION_ATTR_RECIRC
) {
1807 reval_seq
= seq_read(udpif
->reval_seq
) - 1; /* Ensure revalidation. */
1808 ofpbuf_use_const(&actions
, flow
->actions
, flow
->actions_len
);
1809 *ukey
= ukey_create__(flow
->key
, flow
->key_len
,
1810 flow
->mask
, flow
->mask_len
, flow
->ufid_present
,
1811 &flow
->ufid
, flow
->pmd_id
, &actions
,
1812 reval_seq
, flow
->stats
.used
, 0, NULL
);
1818 try_ukey_replace(struct umap
*umap
, struct udpif_key
*old_ukey
,
1819 struct udpif_key
*new_ukey
)
1820 OVS_REQUIRES(umap
->mutex
)
1821 OVS_TRY_LOCK(true, new_ukey
->mutex
)
1823 bool replaced
= false;
1825 if (!ovs_mutex_trylock(&old_ukey
->mutex
)) {
1826 if (old_ukey
->state
== UKEY_EVICTED
) {
1827 /* The flow was deleted during the current revalidator dump,
1828 * but its ukey won't be fully cleaned up until the sweep phase.
1829 * In the mean time, we are receiving upcalls for this traffic.
1830 * Expedite the (new) flow install by replacing the ukey. */
1831 ovs_mutex_lock(&new_ukey
->mutex
);
1832 cmap_replace(&umap
->cmap
, &old_ukey
->cmap_node
,
1833 &new_ukey
->cmap_node
, new_ukey
->hash
);
1834 ovsrcu_postpone(ukey_delete__
, old_ukey
);
1835 transition_ukey(old_ukey
, UKEY_DELETED
);
1836 transition_ukey(new_ukey
, UKEY_VISIBLE
);
1839 ovs_mutex_unlock(&old_ukey
->mutex
);
1843 COVERAGE_INC(upcall_ukey_replace
);
1845 COVERAGE_INC(handler_duplicate_upcall
);
1850 /* Attempts to insert a ukey into the shared ukey maps.
1852 * On success, returns true, installs the ukey and returns it in a locked
1853 * state. Otherwise, returns false. */
1855 ukey_install__(struct udpif
*udpif
, struct udpif_key
*new_ukey
)
1856 OVS_TRY_LOCK(true, new_ukey
->mutex
)
1859 struct udpif_key
*old_ukey
;
1861 bool locked
= false;
1863 idx
= new_ukey
->hash
% N_UMAPS
;
1864 umap
= &udpif
->ukeys
[idx
];
1865 ovs_mutex_lock(&umap
->mutex
);
1866 old_ukey
= ukey_lookup(udpif
, &new_ukey
->ufid
, new_ukey
->pmd_id
);
1868 /* Uncommon case: A ukey is already installed with the same UFID. */
1869 if (old_ukey
->key_len
== new_ukey
->key_len
1870 && !memcmp(old_ukey
->key
, new_ukey
->key
, new_ukey
->key_len
)) {
1871 locked
= try_ukey_replace(umap
, old_ukey
, new_ukey
);
1873 struct ds ds
= DS_EMPTY_INITIALIZER
;
1875 odp_format_ufid(&old_ukey
->ufid
, &ds
);
1876 ds_put_cstr(&ds
, " ");
1877 odp_flow_key_format(old_ukey
->key
, old_ukey
->key_len
, &ds
);
1878 ds_put_cstr(&ds
, "\n");
1879 odp_format_ufid(&new_ukey
->ufid
, &ds
);
1880 ds_put_cstr(&ds
, " ");
1881 odp_flow_key_format(new_ukey
->key
, new_ukey
->key_len
, &ds
);
1883 VLOG_WARN_RL(&rl
, "Conflicting ukey for flows:\n%s", ds_cstr(&ds
));
1887 ovs_mutex_lock(&new_ukey
->mutex
);
1888 cmap_insert(&umap
->cmap
, &new_ukey
->cmap_node
, new_ukey
->hash
);
1889 transition_ukey(new_ukey
, UKEY_VISIBLE
);
1892 ovs_mutex_unlock(&umap
->mutex
);
1898 transition_ukey_at(struct udpif_key
*ukey
, enum ukey_state dst
,
1900 OVS_REQUIRES(ukey
->mutex
)
1902 if (dst
< ukey
->state
) {
1903 VLOG_ABORT("Invalid ukey transition %d->%d (last transitioned from "
1904 "thread %u at %s)", ukey
->state
, dst
, ukey
->state_thread
,
1907 if (ukey
->state
== dst
&& dst
== UKEY_OPERATIONAL
) {
1911 /* Valid state transitions:
1912 * UKEY_CREATED -> UKEY_VISIBLE
1913 * Ukey is now visible in the umap.
1914 * UKEY_VISIBLE -> UKEY_OPERATIONAL
1915 * A handler has installed the flow, and the flow is in the datapath.
1916 * UKEY_VISIBLE -> UKEY_EVICTING
1917 * A handler installs the flow, then revalidator sweeps the ukey before
1918 * the flow is dumped. Most likely the flow was installed; start trying
1920 * UKEY_VISIBLE -> UKEY_EVICTED
1921 * A handler attempts to install the flow, but the datapath rejects it.
1922 * Consider that the datapath has already destroyed it.
1923 * UKEY_OPERATIONAL -> UKEY_EVICTING
1924 * A revalidator decides to evict the datapath flow.
1925 * UKEY_EVICTING -> UKEY_EVICTED
1926 * A revalidator has evicted the datapath flow.
1927 * UKEY_EVICTED -> UKEY_DELETED
1928 * A revalidator has removed the ukey from the umap and is deleting it.
1930 if (ukey
->state
== dst
- 1 || (ukey
->state
== UKEY_VISIBLE
&&
1931 dst
< UKEY_DELETED
)) {
1934 struct ds ds
= DS_EMPTY_INITIALIZER
;
1936 odp_format_ufid(&ukey
->ufid
, &ds
);
1937 VLOG_WARN_RL(&rl
, "Invalid state transition for ukey %s: %d -> %d",
1938 ds_cstr(&ds
), ukey
->state
, dst
);
1941 ukey
->state_thread
= ovsthread_id_self();
1942 ukey
->state_where
= where
;
1946 ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
)
1950 installed
= ukey_install__(udpif
, ukey
);
1952 ovs_mutex_unlock(&ukey
->mutex
);
1958 /* Searches for a ukey in 'udpif->ukeys' that matches 'flow' and attempts to
1959 * lock the ukey. If the ukey does not exist, create it.
1961 * Returns 0 on success, setting *result to the matching ukey and returning it
1962 * in a locked state. Otherwise, returns an errno and clears *result. EBUSY
1963 * indicates that another thread is handling this flow. Other errors indicate
1964 * an unexpected condition creating a new ukey.
1966 * *error is an output parameter provided to appease the threadsafety analyser,
1967 * and its value matches the return value. */
1969 ukey_acquire(struct udpif
*udpif
, const struct dpif_flow
*flow
,
1970 struct udpif_key
**result
, int *error
)
1971 OVS_TRY_LOCK(0, (*result
)->mutex
)
1973 struct udpif_key
*ukey
;
1976 ukey
= ukey_lookup(udpif
, &flow
->ufid
, flow
->pmd_id
);
1978 retval
= ovs_mutex_trylock(&ukey
->mutex
);
1980 /* Usually we try to avoid installing flows from revalidator threads,
1981 * because locking on a umap may cause handler threads to block.
1982 * However there are certain cases, like when ovs-vswitchd is
1983 * restarted, where it is desirable to handle flows that exist in the
1984 * datapath gracefully (ie, don't just clear the datapath). */
1987 retval
= ukey_create_from_dpif_flow(udpif
, flow
, &ukey
);
1991 install
= ukey_install__(udpif
, ukey
);
1995 ukey_delete__(ukey
);
2011 ukey_delete__(struct udpif_key
*ukey
)
2012 OVS_NO_THREAD_SAFETY_ANALYSIS
2015 if (ukey
->key_recirc_id
) {
2016 recirc_free_id(ukey
->key_recirc_id
);
2018 recirc_refs_unref(&ukey
->recircs
);
2019 xlate_cache_delete(ukey
->xcache
);
2020 ofpbuf_delete(ovsrcu_get(struct ofpbuf
*, &ukey
->actions
));
2021 ovs_mutex_destroy(&ukey
->mutex
);
2027 ukey_delete(struct umap
*umap
, struct udpif_key
*ukey
)
2028 OVS_REQUIRES(umap
->mutex
)
2030 ovs_mutex_lock(&ukey
->mutex
);
2031 if (ukey
->state
< UKEY_DELETED
) {
2032 cmap_remove(&umap
->cmap
, &ukey
->cmap_node
, ukey
->hash
);
2033 ovsrcu_postpone(ukey_delete__
, ukey
);
2034 transition_ukey(ukey
, UKEY_DELETED
);
2036 ovs_mutex_unlock(&ukey
->mutex
);
2040 should_revalidate(const struct udpif
*udpif
, uint64_t packets
,
2043 long long int metric
, now
, duration
;
2046 /* Always revalidate the first time a flow is dumped. */
2050 if (udpif
->dump_duration
< ofproto_max_revalidator
/ 2) {
2051 /* We are likely to handle full revalidation for the flows. */
2055 /* Calculate the mean time between seeing these packets. If this
2056 * exceeds the threshold, then delete the flow rather than performing
2057 * costly revalidation for flows that aren't being hit frequently.
2059 * This is targeted at situations where the dump_duration is high (~1s),
2060 * and revalidation is triggered by a call to udpif_revalidate(). In
2061 * these situations, revalidation of all flows causes fluctuations in the
2062 * flow_limit due to the interaction with the dump_duration and max_idle.
2063 * This tends to result in deletion of low-throughput flows anyway, so
2064 * skip the revalidation and just delete those flows. */
2065 packets
= MAX(packets
, 1);
2066 now
= MAX(used
, time_msec());
2067 duration
= now
- used
;
2068 metric
= duration
/ packets
;
2070 if (metric
< 1000 / ofproto_min_revalidate_pps
) {
2071 /* The flow is receiving more than min-revalidate-pps, so keep it. */
2077 struct reval_context
{
2078 /* Optional output parameters */
2079 struct flow_wildcards
*wc
;
2080 struct ofpbuf
*odp_actions
;
2081 struct netflow
**netflow
;
2082 struct xlate_cache
*xcache
;
2084 /* Required output parameters */
2085 struct xlate_out xout
;
2089 /* Translates 'key' into a flow, populating 'ctx' as it goes along.
2091 * Returns 0 on success, otherwise a positive errno value.
2093 * The caller is responsible for uninitializing ctx->xout on success.
2096 xlate_key(struct udpif
*udpif
, const struct nlattr
*key
, unsigned int len
,
2097 const struct dpif_flow_stats
*push
, struct reval_context
*ctx
)
2099 struct ofproto_dpif
*ofproto
;
2100 ofp_port_t ofp_in_port
;
2101 enum odp_key_fitness fitness
;
2102 struct xlate_in xin
;
2105 fitness
= odp_flow_key_to_flow(key
, len
, &ctx
->flow
, NULL
);
2106 if (fitness
== ODP_FIT_ERROR
) {
2110 error
= xlate_lookup(udpif
->backer
, &ctx
->flow
, &ofproto
, NULL
, NULL
,
2111 ctx
->netflow
, &ofp_in_port
);
2116 xlate_in_init(&xin
, ofproto
, ofproto_dpif_get_tables_version(ofproto
),
2117 &ctx
->flow
, ofp_in_port
, NULL
, push
->tcp_flags
,
2118 NULL
, ctx
->wc
, ctx
->odp_actions
);
2119 if (push
->n_packets
) {
2120 xin
.resubmit_stats
= push
;
2121 xin
.allow_side_effects
= true;
2123 xin
.xcache
= ctx
->xcache
;
2124 xlate_actions(&xin
, &ctx
->xout
);
2125 if (fitness
== ODP_FIT_TOO_LITTLE
) {
2126 ctx
->xout
.slow
|= SLOW_MATCH
;
2133 xlate_ukey(struct udpif
*udpif
, const struct udpif_key
*ukey
,
2134 uint16_t tcp_flags
, struct reval_context
*ctx
)
2136 struct dpif_flow_stats push
= {
2137 .tcp_flags
= tcp_flags
,
2139 return xlate_key(udpif
, ukey
->key
, ukey
->key_len
, &push
, ctx
);
2143 populate_xcache(struct udpif
*udpif
, struct udpif_key
*ukey
,
2145 OVS_REQUIRES(ukey
->mutex
)
2147 struct reval_context ctx
= {
2148 .odp_actions
= NULL
,
2154 ovs_assert(!ukey
->xcache
);
2155 ukey
->xcache
= ctx
.xcache
= xlate_cache_new();
2156 error
= xlate_ukey(udpif
, ukey
, tcp_flags
, &ctx
);
2160 xlate_out_uninit(&ctx
.xout
);
2165 static enum reval_result
2166 revalidate_ukey__(struct udpif
*udpif
, const struct udpif_key
*ukey
,
2167 uint16_t tcp_flags
, struct ofpbuf
*odp_actions
,
2168 struct recirc_refs
*recircs
, struct xlate_cache
*xcache
)
2170 struct xlate_out
*xoutp
;
2171 struct netflow
*netflow
;
2172 struct flow_wildcards dp_mask
, wc
;
2173 enum reval_result result
;
2174 struct reval_context ctx
= {
2175 .odp_actions
= odp_actions
,
2176 .netflow
= &netflow
,
2181 result
= UKEY_DELETE
;
2185 if (xlate_ukey(udpif
, ukey
, tcp_flags
, &ctx
)) {
2190 if (xoutp
->avoid_caching
) {
2195 struct ofproto_dpif
*ofproto
;
2196 ofp_port_t ofp_in_port
;
2198 ofproto
= xlate_lookup_ofproto(udpif
->backer
, &ctx
.flow
, &ofp_in_port
,
2201 ofpbuf_clear(odp_actions
);
2207 compose_slow_path(udpif
, xoutp
, ctx
.flow
.in_port
.odp_port
,
2208 ofp_in_port
, odp_actions
,
2209 ofproto
->up
.slowpath_meter_id
, &ofproto
->uuid
);
2212 if (odp_flow_key_to_mask(ukey
->mask
, ukey
->mask_len
, &dp_mask
, &ctx
.flow
,
2218 /* Do not modify if any bit is wildcarded by the installed datapath flow,
2219 * but not the newly revalidated wildcard mask (wc), i.e., if revalidation
2220 * tells that the datapath flow is now too generic and must be narrowed
2221 * down. Note that we do not know if the datapath has ignored any of the
2222 * wildcarded bits, so we may be overly conservative here. */
2223 if (flow_wildcards_has_extra(&dp_mask
, ctx
.wc
)) {
2227 if (!ofpbuf_equal(odp_actions
,
2228 ovsrcu_get(struct ofpbuf
*, &ukey
->actions
))) {
2229 /* The datapath mask was OK, but the actions seem to have changed.
2230 * Let's modify it in place. */
2231 result
= UKEY_MODIFY
;
2232 /* Transfer recirc action ID references to the caller. */
2233 recirc_refs_swap(recircs
, &xoutp
->recircs
);
2240 if (netflow
&& result
== UKEY_DELETE
) {
2241 netflow_flow_clear(netflow
, &ctx
.flow
);
2243 xlate_out_uninit(xoutp
);
2247 /* Verifies that the datapath actions of 'ukey' are still correct, and pushes
2250 * Returns a recommended action for 'ukey', options include:
2251 * UKEY_DELETE The ukey should be deleted.
2252 * UKEY_KEEP The ukey is fine as is.
2253 * UKEY_MODIFY The ukey's actions should be changed but is otherwise
2254 * fine. Callers should change the actions to those found
2255 * in the caller supplied 'odp_actions' buffer. The
2256 * recirculation references can be found in 'recircs' and
2257 * must be handled by the caller.
2259 * If the result is UKEY_MODIFY, then references to all recirc_ids used by the
2260 * new flow will be held within 'recircs' (which may be none).
2262 * The caller is responsible for both initializing 'recircs' prior this call,
2263 * and ensuring any references are eventually freed.
2265 static enum reval_result
2266 revalidate_ukey(struct udpif
*udpif
, struct udpif_key
*ukey
,
2267 const struct dpif_flow_stats
*stats
,
2268 struct ofpbuf
*odp_actions
, uint64_t reval_seq
,
2269 struct recirc_refs
*recircs
, bool offloaded
)
2270 OVS_REQUIRES(ukey
->mutex
)
2272 bool need_revalidate
= ukey
->reval_seq
!= reval_seq
;
2273 enum reval_result result
= UKEY_DELETE
;
2274 struct dpif_flow_stats push
;
2276 ofpbuf_clear(odp_actions
);
2278 push
.used
= stats
->used
;
2279 push
.tcp_flags
= stats
->tcp_flags
;
2280 push
.n_packets
= (stats
->n_packets
> ukey
->stats
.n_packets
2281 ? stats
->n_packets
- ukey
->stats
.n_packets
2283 push
.n_bytes
= (stats
->n_bytes
> ukey
->stats
.n_bytes
2284 ? stats
->n_bytes
- ukey
->stats
.n_bytes
2287 if (need_revalidate
) {
2288 if (should_revalidate(udpif
, push
.n_packets
, ukey
->stats
.used
)) {
2289 if (!ukey
->xcache
) {
2290 ukey
->xcache
= xlate_cache_new();
2292 xlate_cache_clear(ukey
->xcache
);
2294 result
= revalidate_ukey__(udpif
, ukey
, push
.tcp_flags
,
2295 odp_actions
, recircs
, ukey
->xcache
);
2296 } /* else delete; too expensive to revalidate */
2297 } else if (!push
.n_packets
|| ukey
->xcache
2298 || !populate_xcache(udpif
, ukey
, push
.tcp_flags
)) {
2302 /* Stats for deleted flows will be attributed upon flow deletion. Skip. */
2303 if (result
!= UKEY_DELETE
) {
2304 xlate_push_stats(ukey
->xcache
, &push
, offloaded
);
2305 ukey
->stats
= *stats
;
2306 ukey
->reval_seq
= reval_seq
;
2313 delete_op_init__(struct udpif
*udpif
, struct ukey_op
*op
,
2314 const struct dpif_flow
*flow
)
2317 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
2318 op
->dop
.flow_del
.key
= flow
->key
;
2319 op
->dop
.flow_del
.key_len
= flow
->key_len
;
2320 op
->dop
.flow_del
.ufid
= flow
->ufid_present
? &flow
->ufid
: NULL
;
2321 op
->dop
.flow_del
.pmd_id
= flow
->pmd_id
;
2322 op
->dop
.flow_del
.stats
= &op
->stats
;
2323 op
->dop
.flow_del
.terse
= udpif_use_ufid(udpif
);
2327 delete_op_init(struct udpif
*udpif
, struct ukey_op
*op
, struct udpif_key
*ukey
)
2330 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
2331 op
->dop
.flow_del
.key
= ukey
->key
;
2332 op
->dop
.flow_del
.key_len
= ukey
->key_len
;
2333 op
->dop
.flow_del
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
2334 op
->dop
.flow_del
.pmd_id
= ukey
->pmd_id
;
2335 op
->dop
.flow_del
.stats
= &op
->stats
;
2336 op
->dop
.flow_del
.terse
= udpif_use_ufid(udpif
);
2340 put_op_init(struct ukey_op
*op
, struct udpif_key
*ukey
,
2341 enum dpif_flow_put_flags flags
)
2344 op
->dop
.type
= DPIF_OP_FLOW_PUT
;
2345 op
->dop
.flow_put
.flags
= flags
;
2346 op
->dop
.flow_put
.key
= ukey
->key
;
2347 op
->dop
.flow_put
.key_len
= ukey
->key_len
;
2348 op
->dop
.flow_put
.mask
= ukey
->mask
;
2349 op
->dop
.flow_put
.mask_len
= ukey
->mask_len
;
2350 op
->dop
.flow_put
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
2351 op
->dop
.flow_put
.pmd_id
= ukey
->pmd_id
;
2352 op
->dop
.flow_put
.stats
= NULL
;
2353 ukey_get_actions(ukey
, &op
->dop
.flow_put
.actions
,
2354 &op
->dop
.flow_put
.actions_len
);
2357 /* Executes datapath operations 'ops' and attributes stats retrieved from the
2358 * datapath as part of those operations. */
2360 push_dp_ops(struct udpif
*udpif
, struct ukey_op
*ops
, size_t n_ops
)
2362 struct dpif_op
*opsp
[REVALIDATE_MAX_BATCH
];
2365 ovs_assert(n_ops
<= REVALIDATE_MAX_BATCH
);
2366 for (i
= 0; i
< n_ops
; i
++) {
2367 opsp
[i
] = &ops
[i
].dop
;
2369 dpif_operate(udpif
->dpif
, opsp
, n_ops
, DPIF_OFFLOAD_AUTO
);
2371 for (i
= 0; i
< n_ops
; i
++) {
2372 struct ukey_op
*op
= &ops
[i
];
2373 struct dpif_flow_stats
*push
, *stats
, push_buf
;
2375 stats
= op
->dop
.flow_del
.stats
;
2378 if (op
->dop
.type
!= DPIF_OP_FLOW_DEL
) {
2379 /* Only deleted flows need their stats pushed. */
2383 if (op
->dop
.error
) {
2384 /* flow_del error, 'stats' is unusable. */
2386 ovs_mutex_lock(&op
->ukey
->mutex
);
2387 transition_ukey(op
->ukey
, UKEY_EVICTED
);
2388 ovs_mutex_unlock(&op
->ukey
->mutex
);
2394 ovs_mutex_lock(&op
->ukey
->mutex
);
2395 transition_ukey(op
->ukey
, UKEY_EVICTED
);
2396 push
->used
= MAX(stats
->used
, op
->ukey
->stats
.used
);
2397 push
->tcp_flags
= stats
->tcp_flags
| op
->ukey
->stats
.tcp_flags
;
2398 push
->n_packets
= stats
->n_packets
- op
->ukey
->stats
.n_packets
;
2399 push
->n_bytes
= stats
->n_bytes
- op
->ukey
->stats
.n_bytes
;
2400 ovs_mutex_unlock(&op
->ukey
->mutex
);
2405 if (push
->n_packets
|| netflow_exists()) {
2406 const struct nlattr
*key
= op
->dop
.flow_del
.key
;
2407 size_t key_len
= op
->dop
.flow_del
.key_len
;
2408 struct netflow
*netflow
;
2409 struct reval_context ctx
= {
2410 .netflow
= &netflow
,
2415 ovs_mutex_lock(&op
->ukey
->mutex
);
2416 if (op
->ukey
->xcache
) {
2417 xlate_push_stats(op
->ukey
->xcache
, push
, false);
2418 ovs_mutex_unlock(&op
->ukey
->mutex
);
2421 ovs_mutex_unlock(&op
->ukey
->mutex
);
2422 key
= op
->ukey
->key
;
2423 key_len
= op
->ukey
->key_len
;
2426 error
= xlate_key(udpif
, key
, key_len
, push
, &ctx
);
2428 static struct vlog_rate_limit rll
= VLOG_RATE_LIMIT_INIT(1, 5);
2429 VLOG_WARN_RL(&rll
, "xlate_key failed (%s)!",
2430 ovs_strerror(error
));
2432 xlate_out_uninit(&ctx
.xout
);
2434 netflow_flow_clear(netflow
, &ctx
.flow
);
2441 /* Executes datapath operations 'ops', attributes stats retrieved from the
2442 * datapath, and deletes ukeys corresponding to deleted flows. */
2444 push_ukey_ops(struct udpif
*udpif
, struct umap
*umap
,
2445 struct ukey_op
*ops
, size_t n_ops
)
2449 push_dp_ops(udpif
, ops
, n_ops
);
2450 ovs_mutex_lock(&umap
->mutex
);
2451 for (i
= 0; i
< n_ops
; i
++) {
2452 if (ops
[i
].dop
.type
== DPIF_OP_FLOW_DEL
) {
2453 ukey_delete(umap
, ops
[i
].ukey
);
2456 ovs_mutex_unlock(&umap
->mutex
);
2460 log_unexpected_flow(const struct dpif_flow
*flow
, int error
)
2462 struct ds ds
= DS_EMPTY_INITIALIZER
;
2464 ds_put_format(&ds
, "Failed to acquire udpif_key corresponding to "
2465 "unexpected flow (%s): ", ovs_strerror(error
));
2466 odp_format_ufid(&flow
->ufid
, &ds
);
2468 static struct vlog_rate_limit rll
= VLOG_RATE_LIMIT_INIT(10, 60);
2469 VLOG_WARN_RL(&rll
, "%s", ds_cstr(&ds
));
2475 reval_op_init(struct ukey_op
*op
, enum reval_result result
,
2476 struct udpif
*udpif
, struct udpif_key
*ukey
,
2477 struct recirc_refs
*recircs
, struct ofpbuf
*odp_actions
)
2478 OVS_REQUIRES(ukey
->mutex
)
2480 if (result
== UKEY_DELETE
) {
2481 delete_op_init(udpif
, op
, ukey
);
2482 transition_ukey(ukey
, UKEY_EVICTING
);
2483 } else if (result
== UKEY_MODIFY
) {
2484 /* Store the new recircs. */
2485 recirc_refs_swap(&ukey
->recircs
, recircs
);
2486 /* Release old recircs. */
2487 recirc_refs_unref(recircs
);
2488 /* ukey->key_recirc_id remains, as the key is the same as before. */
2490 ukey_set_actions(ukey
, odp_actions
);
2491 put_op_init(op
, ukey
, DPIF_FP_MODIFY
);
2496 ukey_netdev_unref(struct udpif_key
*ukey
)
2498 if (!ukey
->in_netdev
) {
2501 netdev_close(ukey
->in_netdev
);
2502 ukey
->in_netdev
= NULL
;
2506 * Given a udpif_key, get its input port (netdev) by parsing the flow keys
2507 * and actions. The flow may not contain flow attributes if it is a terse
2508 * dump; read its attributes from the ukey and then parse the flow to get
2509 * the port info. Save them in udpif_key.
2512 ukey_to_flow_netdev(struct udpif
*udpif
, struct udpif_key
*ukey
)
2514 const struct dpif
*dpif
= udpif
->dpif
;
2515 const struct dpif_class
*dpif_class
= dpif
->dpif_class
;
2516 const struct nlattr
*k
;
2519 /* Remove existing references to netdev */
2520 ukey_netdev_unref(ukey
);
2522 /* Find the input port and get a reference to its netdev */
2523 NL_ATTR_FOR_EACH (k
, left
, ukey
->key
, ukey
->key_len
) {
2524 enum ovs_key_attr type
= nl_attr_type(k
);
2526 if (type
== OVS_KEY_ATTR_IN_PORT
) {
2527 ukey
->in_netdev
= netdev_ports_get(nl_attr_get_odp_port(k
),
2529 } else if (type
== OVS_KEY_ATTR_TUNNEL
) {
2530 struct flow_tnl tnl
;
2531 enum odp_key_fitness res
;
2533 if (ukey
->in_netdev
) {
2534 netdev_close(ukey
->in_netdev
);
2535 ukey
->in_netdev
= NULL
;
2537 res
= odp_tun_key_from_attr(k
, &tnl
, NULL
);
2538 if (res
!= ODP_FIT_ERROR
) {
2539 ukey
->in_netdev
= flow_get_tunnel_netdev(&tnl
);
2547 udpif_flow_packet_delta(struct udpif_key
*ukey
, const struct dpif_flow
*f
)
2549 return f
->stats
.n_packets
+ ukey
->flow_backlog_packets
-
2553 static long long int
2554 udpif_flow_time_delta(struct udpif
*udpif
, struct udpif_key
*ukey
)
2556 return (udpif
->dpif
->current_ms
- ukey
->flow_time
) / 1000;
2560 * Save backlog packet count while switching modes
2561 * between offloaded and kernel datapaths.
2564 udpif_set_ukey_backlog_packets(struct udpif_key
*ukey
)
2566 ukey
->flow_backlog_packets
= ukey
->flow_packets
;
2569 /* Gather pps-rate for the given dpif_flow and save it in its ukey */
2571 udpif_update_flow_pps(struct udpif
*udpif
, struct udpif_key
*ukey
,
2572 const struct dpif_flow
*f
)
2576 /* Update pps-rate only when we are close to rebalance interval */
2577 if (udpif
->dpif
->current_ms
- ukey
->flow_time
< OFFL_REBAL_INTVL_MSEC
) {
2581 ukey
->offloaded
= f
->attrs
.offloaded
;
2582 pps
= udpif_flow_packet_delta(ukey
, f
) /
2583 udpif_flow_time_delta(udpif
, ukey
);
2584 ukey
->flow_pps_rate
= pps
;
2585 ukey
->flow_packets
= ukey
->flow_backlog_packets
+ f
->stats
.n_packets
;
2586 ukey
->flow_time
= udpif
->dpif
->current_ms
;
2590 revalidate(struct revalidator
*revalidator
)
2592 uint64_t odp_actions_stub
[1024 / 8];
2593 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2595 struct udpif
*udpif
= revalidator
->udpif
;
2596 struct dpif_flow_dump_thread
*dump_thread
;
2597 uint64_t dump_seq
, reval_seq
;
2598 unsigned int flow_limit
;
2600 dump_seq
= seq_read(udpif
->dump_seq
);
2601 reval_seq
= seq_read(udpif
->reval_seq
);
2602 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2603 dump_thread
= dpif_flow_dump_thread_create(udpif
->dump
);
2605 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2608 struct dpif_flow flows
[REVALIDATE_MAX_BATCH
];
2609 const struct dpif_flow
*f
;
2612 long long int max_idle
;
2617 n_dumped
= dpif_flow_dump_next(dump_thread
, flows
, ARRAY_SIZE(flows
));
2624 /* In normal operation we want to keep flows around until they have
2625 * been idle for 'ofproto_max_idle' milliseconds. However:
2627 * - If the number of datapath flows climbs above 'flow_limit',
2628 * drop that down to 100 ms to try to bring the flows down to
2631 * - If the number of datapath flows climbs above twice
2632 * 'flow_limit', delete all the datapath flows as an emergency
2633 * measure. (We reassess this condition for the next batch of
2634 * datapath flows, so we will recover before all the flows are
2636 n_dp_flows
= udpif_get_n_flows(udpif
);
2637 kill_them_all
= n_dp_flows
> flow_limit
* 2;
2638 max_idle
= n_dp_flows
> flow_limit
? 100 : ofproto_max_idle
;
2640 udpif
->dpif
->current_ms
= time_msec();
2641 for (f
= flows
; f
< &flows
[n_dumped
]; f
++) {
2642 long long int used
= f
->stats
.used
;
2643 struct recirc_refs recircs
= RECIRC_REFS_EMPTY_INITIALIZER
;
2644 enum reval_result result
;
2645 struct udpif_key
*ukey
;
2646 bool already_dumped
;
2649 /* We don't need an extra information. */
2650 free(f
->attrs
.dp_extra_info
);
2652 if (ukey_acquire(udpif
, f
, &ukey
, &error
)) {
2653 if (error
== EBUSY
) {
2654 /* Another thread is processing this flow, so don't bother
2656 COVERAGE_INC(upcall_ukey_contention
);
2658 log_unexpected_flow(f
, error
);
2659 if (error
!= ENOENT
) {
2660 delete_op_init__(udpif
, &ops
[n_ops
++], f
);
2666 already_dumped
= ukey
->dump_seq
== dump_seq
;
2667 if (already_dumped
) {
2668 /* The flow has already been handled during this flow dump
2669 * operation. Skip it. */
2671 COVERAGE_INC(dumped_duplicate_flow
);
2673 COVERAGE_INC(dumped_new_flow
);
2675 ovs_mutex_unlock(&ukey
->mutex
);
2679 if (ukey
->state
<= UKEY_OPERATIONAL
) {
2680 /* The flow is now confirmed to be in the datapath. */
2681 transition_ukey(ukey
, UKEY_OPERATIONAL
);
2683 VLOG_INFO("Unexpected ukey transition from state %d "
2684 "(last transitioned from thread %u at %s)",
2685 ukey
->state
, ukey
->state_thread
, ukey
->state_where
);
2686 ovs_mutex_unlock(&ukey
->mutex
);
2691 used
= ukey
->created
;
2693 if (kill_them_all
|| (used
&& used
< now
- max_idle
)) {
2694 result
= UKEY_DELETE
;
2696 result
= revalidate_ukey(udpif
, ukey
, &f
->stats
, &odp_actions
,
2697 reval_seq
, &recircs
,
2698 f
->attrs
.offloaded
);
2700 ukey
->dump_seq
= dump_seq
;
2702 if (netdev_is_offload_rebalance_policy_enabled() &&
2703 result
!= UKEY_DELETE
) {
2704 udpif_update_flow_pps(udpif
, ukey
, f
);
2707 if (result
!= UKEY_KEEP
) {
2708 /* Takes ownership of 'recircs'. */
2709 reval_op_init(&ops
[n_ops
++], result
, udpif
, ukey
, &recircs
,
2712 ovs_mutex_unlock(&ukey
->mutex
);
2716 /* Push datapath ops but defer ukey deletion to 'sweep' phase. */
2717 push_dp_ops(udpif
, ops
, n_ops
);
2721 dpif_flow_dump_thread_destroy(dump_thread
);
2722 ofpbuf_uninit(&odp_actions
);
2725 /* Pauses the 'revalidator', can only proceed after main thread
2726 * calls udpif_resume_revalidators(). */
2728 revalidator_pause(struct revalidator
*revalidator
)
2730 /* The first block is for sync'ing the pause with main thread. */
2731 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2732 /* The second block is for pausing until main thread resumes. */
2733 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2737 revalidator_sweep__(struct revalidator
*revalidator
, bool purge
)
2739 struct udpif
*udpif
;
2740 uint64_t dump_seq
, reval_seq
;
2743 udpif
= revalidator
->udpif
;
2744 dump_seq
= seq_read(udpif
->dump_seq
);
2745 reval_seq
= seq_read(udpif
->reval_seq
);
2746 slice
= revalidator
- udpif
->revalidators
;
2747 ovs_assert(slice
< udpif
->n_revalidators
);
2749 for (int i
= slice
; i
< N_UMAPS
; i
+= udpif
->n_revalidators
) {
2750 uint64_t odp_actions_stub
[1024 / 8];
2751 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2753 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2754 struct udpif_key
*ukey
;
2755 struct umap
*umap
= &udpif
->ukeys
[i
];
2758 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2759 enum ukey_state ukey_state
;
2761 /* Handler threads could be holding a ukey lock while it installs a
2762 * new flow, so don't hang around waiting for access to it. */
2763 if (ovs_mutex_trylock(&ukey
->mutex
)) {
2766 ukey_state
= ukey
->state
;
2767 if (ukey_state
== UKEY_OPERATIONAL
2768 || (ukey_state
== UKEY_VISIBLE
&& purge
)) {
2769 struct recirc_refs recircs
= RECIRC_REFS_EMPTY_INITIALIZER
;
2770 bool seq_mismatch
= (ukey
->dump_seq
!= dump_seq
2771 && ukey
->reval_seq
!= reval_seq
);
2772 enum reval_result result
;
2775 result
= UKEY_DELETE
;
2776 } else if (!seq_mismatch
) {
2779 struct dpif_flow_stats stats
;
2780 COVERAGE_INC(revalidate_missed_dp_flow
);
2781 memset(&stats
, 0, sizeof stats
);
2782 result
= revalidate_ukey(udpif
, ukey
, &stats
, &odp_actions
,
2783 reval_seq
, &recircs
, false);
2785 if (result
!= UKEY_KEEP
) {
2786 /* Clears 'recircs' if filled by revalidate_ukey(). */
2787 reval_op_init(&ops
[n_ops
++], result
, udpif
, ukey
, &recircs
,
2791 ovs_mutex_unlock(&ukey
->mutex
);
2793 if (ukey_state
== UKEY_EVICTED
) {
2794 /* The common flow deletion case involves deletion of the flow
2795 * during the dump phase and ukey deletion here. */
2796 ovs_mutex_lock(&umap
->mutex
);
2797 ukey_delete(umap
, ukey
);
2798 ovs_mutex_unlock(&umap
->mutex
);
2801 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2802 /* Update/delete missed flows and clean up corresponding ukeys
2804 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2810 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2813 ofpbuf_uninit(&odp_actions
);
2819 revalidator_sweep(struct revalidator
*revalidator
)
2821 revalidator_sweep__(revalidator
, false);
2825 revalidator_purge(struct revalidator
*revalidator
)
2827 revalidator_sweep__(revalidator
, true);
2830 /* In reaction to dpif purge, purges all 'ukey's with same 'pmd_id'. */
2832 dp_purge_cb(void *aux
, unsigned pmd_id
)
2833 OVS_NO_THREAD_SAFETY_ANALYSIS
2835 struct udpif
*udpif
= aux
;
2838 udpif_pause_revalidators(udpif
);
2839 for (i
= 0; i
< N_UMAPS
; i
++) {
2840 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2841 struct udpif_key
*ukey
;
2842 struct umap
*umap
= &udpif
->ukeys
[i
];
2845 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2846 if (ukey
->pmd_id
== pmd_id
) {
2847 delete_op_init(udpif
, &ops
[n_ops
++], ukey
);
2848 transition_ukey(ukey
, UKEY_EVICTING
);
2850 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2851 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2858 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2863 udpif_resume_revalidators(udpif
);
2867 upcall_unixctl_show(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2868 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2870 struct ds ds
= DS_EMPTY_INITIALIZER
;
2871 struct udpif
*udpif
;
2873 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2874 unsigned int flow_limit
;
2878 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2879 ufid_enabled
= udpif_use_ufid(udpif
);
2881 ds_put_format(&ds
, "%s:\n", dpif_name(udpif
->dpif
));
2882 ds_put_format(&ds
, " flows : (current %lu)"
2883 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif
),
2884 udpif
->avg_n_flows
, udpif
->max_n_flows
, flow_limit
);
2885 ds_put_format(&ds
, " dump duration : %lldms\n", udpif
->dump_duration
);
2886 ds_put_format(&ds
, " ufid enabled : ");
2888 ds_put_format(&ds
, "true\n");
2890 ds_put_format(&ds
, "false\n");
2892 ds_put_char(&ds
, '\n');
2894 for (i
= 0; i
< n_revalidators
; i
++) {
2895 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
2896 int j
, elements
= 0;
2898 for (j
= i
; j
< N_UMAPS
; j
+= n_revalidators
) {
2899 elements
+= cmap_count(&udpif
->ukeys
[j
].cmap
);
2901 ds_put_format(&ds
, " %u: (keys %d)\n", revalidator
->id
, elements
);
2905 unixctl_command_reply(conn
, ds_cstr(&ds
));
2909 /* Disable using the megaflows.
2911 * This command is only needed for advanced debugging, so it's not
2912 * documented in the man page. */
2914 upcall_unixctl_disable_megaflows(struct unixctl_conn
*conn
,
2915 int argc OVS_UNUSED
,
2916 const char *argv
[] OVS_UNUSED
,
2917 void *aux OVS_UNUSED
)
2919 atomic_store_relaxed(&enable_megaflows
, false);
2920 udpif_flush_all_datapaths();
2921 unixctl_command_reply(conn
, "megaflows disabled");
2924 /* Re-enable using megaflows.
2926 * This command is only needed for advanced debugging, so it's not
2927 * documented in the man page. */
2929 upcall_unixctl_enable_megaflows(struct unixctl_conn
*conn
,
2930 int argc OVS_UNUSED
,
2931 const char *argv
[] OVS_UNUSED
,
2932 void *aux OVS_UNUSED
)
2934 atomic_store_relaxed(&enable_megaflows
, true);
2935 udpif_flush_all_datapaths();
2936 unixctl_command_reply(conn
, "megaflows enabled");
2939 /* Disable skipping flow attributes during flow dump.
2941 * This command is only needed for advanced debugging, so it's not
2942 * documented in the man page. */
2944 upcall_unixctl_disable_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
, false);
2948 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID disabled");
2951 /* Re-enable skipping flow attributes during flow dump.
2953 * This command is only needed for advanced debugging, so it's not documented
2954 * in the man page. */
2956 upcall_unixctl_enable_ufid(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2957 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2959 atomic_store_relaxed(&enable_ufid
, true);
2960 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID enabled "
2961 "for supported datapaths");
2964 /* Set the flow limit.
2966 * This command is only needed for advanced debugging, so it's not
2967 * documented in the man page. */
2969 upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
,
2970 int argc OVS_UNUSED
,
2972 void *aux OVS_UNUSED
)
2974 struct ds ds
= DS_EMPTY_INITIALIZER
;
2975 struct udpif
*udpif
;
2976 unsigned int flow_limit
= atoi(argv
[1]);
2978 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2979 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
2981 ds_put_format(&ds
, "set flow_limit to %u\n", flow_limit
);
2982 unixctl_command_reply(conn
, ds_cstr(&ds
));
2987 upcall_unixctl_dump_wait(struct unixctl_conn
*conn
,
2988 int argc OVS_UNUSED
,
2989 const char *argv
[] OVS_UNUSED
,
2990 void *aux OVS_UNUSED
)
2992 if (ovs_list_is_singleton(&all_udpifs
)) {
2993 struct udpif
*udpif
= NULL
;
2996 udpif
= OBJECT_CONTAINING(ovs_list_front(&all_udpifs
), udpif
, list_node
);
2997 len
= (udpif
->n_conns
+ 1) * sizeof *udpif
->conns
;
2998 udpif
->conn_seq
= seq_read(udpif
->dump_seq
);
2999 udpif
->conns
= xrealloc(udpif
->conns
, len
);
3000 udpif
->conns
[udpif
->n_conns
++] = conn
;
3002 unixctl_command_reply_error(conn
, "can't wait on multiple udpifs.");
3007 upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
3008 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
3010 struct udpif
*udpif
;
3012 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
3015 for (n
= 0; n
< udpif
->n_revalidators
; n
++) {
3016 revalidator_purge(&udpif
->revalidators
[n
]);
3019 unixctl_command_reply(conn
, "");
3022 /* Flows are sorted in the following order:
3023 * netdev, flow state (offloaded/kernel path), flow_pps_rate.
3026 flow_compare_rebalance(const void *elem1
, const void *elem2
)
3028 const struct udpif_key
*f1
= *(struct udpif_key
**)elem1
;
3029 const struct udpif_key
*f2
= *(struct udpif_key
**)elem2
;
3032 if (f1
->in_netdev
< f2
->in_netdev
) {
3034 } else if (f1
->in_netdev
> f2
->in_netdev
) {
3038 if (f1
->offloaded
!= f2
->offloaded
) {
3039 return f2
->offloaded
- f1
->offloaded
;
3042 diff
= (f1
->offloaded
== true) ?
3043 f1
->flow_pps_rate
- f2
->flow_pps_rate
:
3044 f2
->flow_pps_rate
- f1
->flow_pps_rate
;
3046 return (diff
< 0) ? -1 : 1;
3049 /* Insert flows from pending array during rebalancing */
3051 rebalance_insert_pending(struct udpif
*udpif
, struct udpif_key
**pending_flows
,
3052 int pending_count
, int insert_count
,
3053 uint64_t rate_threshold
)
3057 for (int i
= 0; i
< pending_count
; i
++) {
3058 struct udpif_key
*flow
= pending_flows
[i
];
3061 /* Stop offloading pending flows if the insert count is
3062 * reached and the flow rate is less than the threshold
3064 if (count
>= insert_count
&& flow
->flow_pps_rate
< rate_threshold
) {
3068 /* Offload the flow to netdev */
3069 err
= udpif_flow_program(udpif
, flow
, DPIF_OFFLOAD_ALWAYS
);
3071 if (err
== ENOSPC
) {
3072 /* Stop if we are out of resources */
3080 /* Offload succeeded; delete it from the kernel datapath */
3081 udpif_flow_unprogram(udpif
, flow
, DPIF_OFFLOAD_NEVER
);
3083 /* Change the state of the flow, adjust dpif counters */
3084 flow
->offloaded
= true;
3086 udpif_set_ukey_backlog_packets(flow
);
3093 /* Remove flows from offloaded array during rebalancing */
3095 rebalance_remove_offloaded(struct udpif
*udpif
,
3096 struct udpif_key
**offloaded_flows
,
3099 for (int i
= 0; i
< offload_count
; i
++) {
3100 struct udpif_key
*flow
= offloaded_flows
[i
];
3103 /* Install the flow into kernel path first */
3104 err
= udpif_flow_program(udpif
, flow
, DPIF_OFFLOAD_NEVER
);
3109 /* Success; now remove offloaded flow from netdev */
3110 err
= udpif_flow_unprogram(udpif
, flow
, DPIF_OFFLOAD_ALWAYS
);
3112 udpif_flow_unprogram(udpif
, flow
, DPIF_OFFLOAD_NEVER
);
3115 udpif_set_ukey_backlog_packets(flow
);
3116 flow
->offloaded
= false;
3121 * Rebalance offloaded flows on a netdev that's in OOR state.
3123 * The rebalancing is done in two phases. In the first phase, we check if
3124 * the pending flows can be offloaded (if some resources became available
3125 * in the meantime) by trying to offload each pending flow. If all pending
3126 * flows get successfully offloaded, the OOR state is cleared on the netdev
3127 * and there's nothing to rebalance.
3129 * If some of the pending flows could not be offloaded, i.e, we still see
3130 * the OOR error, then we move to the second phase of rebalancing. In this
3131 * phase, the rebalancer compares pps-rate of an offloaded flow with the
3132 * least pps-rate with that of a pending flow with the highest pps-rate from
3133 * their respective sorted arrays. If pps-rate of the offloaded flow is less
3134 * than the pps-rate of the pending flow, then it deletes the offloaded flow
3135 * from the HW/netdev and adds it to kernel datapath and then offloads pending
3136 * to HW/netdev. This process is repeated for every pair of offloaded and
3137 * pending flows in the ordered list. The process stops when we encounter an
3138 * offloaded flow that has a higher pps-rate than the corresponding pending
3139 * flow. The entire rebalancing process is repeated in the next iteration.
3142 rebalance_device(struct udpif
*udpif
, struct udpif_key
**offloaded_flows
,
3143 int offload_count
, struct udpif_key
**pending_flows
,
3148 int num_inserted
= rebalance_insert_pending(udpif
, pending_flows
,
3149 pending_count
, pending_count
,
3152 VLOG_DBG("Offload rebalance: Phase1: inserted %d pending flows",
3156 /* Adjust pending array */
3157 pending_flows
= &pending_flows
[num_inserted
];
3158 pending_count
-= num_inserted
;
3160 if (!pending_count
) {
3162 * Successfully offloaded all pending flows. The device
3163 * is no longer in OOR state; done rebalancing this device.
3169 * Phase 2; determine how many offloaded flows to churn.
3171 #define OFFL_REBAL_MAX_CHURN 1024
3172 int churn_count
= 0;
3173 while (churn_count
< OFFL_REBAL_MAX_CHURN
&& churn_count
< offload_count
3174 && churn_count
< pending_count
) {
3175 if (pending_flows
[churn_count
]->flow_pps_rate
<=
3176 offloaded_flows
[churn_count
]->flow_pps_rate
)
3182 VLOG_DBG("Offload rebalance: Phase2: removing %d offloaded flows",
3186 /* Bail early if nothing to churn */
3191 /* Remove offloaded flows */
3192 rebalance_remove_offloaded(udpif
, offloaded_flows
, churn_count
);
3194 /* Adjust offloaded array */
3195 offloaded_flows
= &offloaded_flows
[churn_count
];
3196 offload_count
-= churn_count
;
3198 /* Replace offloaded flows with pending flows */
3199 num_inserted
= rebalance_insert_pending(udpif
, pending_flows
,
3200 pending_count
, churn_count
,
3202 offloaded_flows
[0]->flow_pps_rate
:
3205 VLOG_DBG("Offload rebalance: Phase2: inserted %d pending flows",
3212 static struct udpif_key
**
3213 udpif_add_oor_flows(struct udpif_key
**sort_flows
, size_t *total_flow_count
,
3214 size_t *alloc_flow_count
, struct udpif_key
*ukey
)
3216 if (*total_flow_count
>= *alloc_flow_count
) {
3217 sort_flows
= x2nrealloc(sort_flows
, alloc_flow_count
, sizeof ukey
);
3219 sort_flows
[(*total_flow_count
)++] = ukey
;
3224 * Build sort_flows[] initially with flows that
3225 * reference an 'OOR' netdev as their input port.
3227 static struct udpif_key
**
3228 udpif_build_oor_flows(struct udpif_key
**sort_flows
, size_t *total_flow_count
,
3229 size_t *alloc_flow_count
, struct udpif_key
*ukey
,
3230 int *oor_netdev_count
)
3232 struct netdev
*netdev
;
3235 /* Input netdev must be available for the flow */
3236 netdev
= ukey
->in_netdev
;
3241 /* Is the in-netdev for this flow in OOR state ? */
3242 if (!netdev_get_hw_info(netdev
, HW_INFO_TYPE_OOR
)) {
3243 ukey_netdev_unref(ukey
);
3247 /* Add the flow to sort_flows[] */
3248 sort_flows
= udpif_add_oor_flows(sort_flows
, total_flow_count
,
3249 alloc_flow_count
, ukey
);
3250 if (ukey
->offloaded
) {
3251 count
= netdev_get_hw_info(netdev
, HW_INFO_TYPE_OFFL_COUNT
);
3252 ovs_assert(count
>= 0);
3254 (*oor_netdev_count
)++;
3256 netdev_set_hw_info(netdev
, HW_INFO_TYPE_OFFL_COUNT
, count
);
3258 count
= netdev_get_hw_info(netdev
, HW_INFO_TYPE_PEND_COUNT
);
3259 ovs_assert(count
>= 0);
3260 netdev_set_hw_info(netdev
, HW_INFO_TYPE_PEND_COUNT
, ++count
);
3267 * Rebalance offloaded flows on HW netdevs that are in OOR state.
3270 udpif_flow_rebalance(struct udpif
*udpif
)
3272 struct udpif_key
**sort_flows
= NULL
;
3273 size_t alloc_flow_count
= 0;
3274 size_t total_flow_count
= 0;
3275 int oor_netdev_count
= 0;
3276 int offload_index
= 0;
3279 /* Collect flows (offloaded and pending) that reference OOR netdevs */
3280 for (size_t i
= 0; i
< N_UMAPS
; i
++) {
3281 struct udpif_key
*ukey
;
3282 struct umap
*umap
= &udpif
->ukeys
[i
];
3284 CMAP_FOR_EACH (ukey
, cmap_node
, &umap
->cmap
) {
3285 ukey_to_flow_netdev(udpif
, ukey
);
3286 sort_flows
= udpif_build_oor_flows(sort_flows
, &total_flow_count
,
3287 &alloc_flow_count
, ukey
,
3292 /* Sort flows by OOR netdevs, state (offloaded/pending) and pps-rate */
3293 qsort(sort_flows
, total_flow_count
, sizeof(struct udpif_key
*),
3294 flow_compare_rebalance
);
3297 * We now have flows referencing OOR netdevs, that are sorted. We also
3298 * have a count of offloaded and pending flows on each of the netdevs
3299 * that are in OOR state. Now rebalance each oor-netdev.
3301 while (oor_netdev_count
) {
3302 struct netdev
*netdev
;
3307 netdev
= sort_flows
[offload_index
]->in_netdev
;
3308 ovs_assert(netdev_get_hw_info(netdev
, HW_INFO_TYPE_OOR
) == true);
3309 VLOG_DBG("Offload rebalance: netdev: %s is OOR", netdev
->name
);
3311 offload_count
= netdev_get_hw_info(netdev
, HW_INFO_TYPE_OFFL_COUNT
);
3312 pending_count
= netdev_get_hw_info(netdev
, HW_INFO_TYPE_PEND_COUNT
);
3313 pending_index
= offload_index
+ offload_count
;
3315 oor
= rebalance_device(udpif
,
3316 &sort_flows
[offload_index
], offload_count
,
3317 &sort_flows
[pending_index
], pending_count
);
3318 netdev_set_hw_info(netdev
, HW_INFO_TYPE_OOR
, oor
);
3320 offload_index
= pending_index
+ pending_count
;
3321 netdev_set_hw_info(netdev
, HW_INFO_TYPE_OFFL_COUNT
, 0);
3322 netdev_set_hw_info(netdev
, HW_INFO_TYPE_PEND_COUNT
, 0);
3326 for (int i
= 0; i
< total_flow_count
; i
++) {
3327 struct udpif_key
*ukey
= sort_flows
[i
];
3328 ukey_netdev_unref(ukey
);
3334 udpif_flow_program(struct udpif
*udpif
, struct udpif_key
*ukey
,
3335 enum dpif_offload_type offload_type
)
3337 struct dpif_op
*opsp
;
3341 put_op_init(&uop
, ukey
, DPIF_FP_CREATE
);
3342 dpif_operate(udpif
->dpif
, &opsp
, 1, offload_type
);
3348 udpif_flow_unprogram(struct udpif
*udpif
, struct udpif_key
*ukey
,
3349 enum dpif_offload_type offload_type
)
3351 struct dpif_op
*opsp
;
3355 delete_op_init(udpif
, &uop
, ukey
);
3356 dpif_operate(udpif
->dpif
, &opsp
, 1, offload_type
);