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
);
59 COVERAGE_DEFINE(upcall_flow_limit_hit
);
61 /* A thread that reads upcalls from dpif, forwards each upcall's packet,
62 * and possibly sets up a kernel flow as a cache. */
64 struct udpif
*udpif
; /* Parent udpif. */
65 pthread_t thread
; /* Thread ID. */
66 uint32_t handler_id
; /* Handler id. */
69 /* In the absence of a multiple-writer multiple-reader datastructure for
70 * storing udpif_keys ("ukeys"), we use a large number of cmaps, each with its
71 * own lock for writing. */
72 #define N_UMAPS 512 /* per udpif. */
74 struct ovs_mutex mutex
; /* Take for writing to the following. */
75 struct cmap cmap
; /* Datapath flow keys. */
78 /* A thread that processes datapath flows, updates OpenFlow statistics, and
79 * updates or removes them if necessary.
81 * Revalidator threads operate in two phases: "dump" and "sweep". In between
82 * each phase, all revalidators sync up so that all revalidator threads are
83 * either in one phase or the other, but not a combination.
85 * During the dump phase, revalidators fetch flows from the datapath and
86 * attribute the statistics to OpenFlow rules. Each datapath flow has a
87 * corresponding ukey which caches the most recently seen statistics. If
88 * a flow needs to be deleted (for example, because it is unused over a
89 * period of time), revalidator threads may delete the flow during the
90 * dump phase. The datapath is not guaranteed to reliably dump all flows
91 * from the datapath, and there is no mapping between datapath flows to
92 * revalidators, so a particular flow may be handled by zero or more
93 * revalidators during a single dump phase. To avoid duplicate attribution
94 * of statistics, ukeys are never deleted during this phase.
96 * During the sweep phase, each revalidator takes ownership of a different
97 * slice of umaps and sweeps through all ukeys in those umaps to figure out
98 * whether they need to be deleted. During this phase, revalidators may
99 * fetch individual flows which were not dumped during the dump phase to
100 * validate them and attribute statistics.
103 struct udpif
*udpif
; /* Parent udpif. */
104 pthread_t thread
; /* Thread ID. */
105 unsigned int id
; /* ovsthread_id_self(). */
108 /* An upcall handler for ofproto_dpif.
110 * udpif keeps records of two kind of logically separate units:
115 * - An array of 'struct handler's for upcall handling and flow
121 * - Revalidation threads which read the datapath flow table and maintains
125 struct ovs_list list_node
; /* In all_udpifs list. */
127 struct dpif
*dpif
; /* Datapath handle. */
128 struct dpif_backer
*backer
; /* Opaque dpif_backer pointer. */
130 struct handler
*handlers
; /* Upcall handlers. */
133 struct revalidator
*revalidators
; /* Flow revalidators. */
134 size_t n_revalidators
;
136 struct latch exit_latch
; /* Tells child threads to exit. */
139 struct seq
*reval_seq
; /* Incremented to force revalidation. */
140 bool reval_exit
; /* Set by leader on 'exit_latch. */
141 struct ovs_barrier reval_barrier
; /* Barrier used by revalidators. */
142 struct dpif_flow_dump
*dump
; /* DPIF flow dump state. */
143 long long int dump_duration
; /* Duration of the last flow dump. */
144 struct seq
*dump_seq
; /* Increments each dump iteration. */
145 atomic_bool enable_ufid
; /* If true, skip dumping flow attrs. */
147 /* These variables provide a mechanism for the main thread to pause
148 * all revalidation without having to completely shut the threads down.
149 * 'pause_latch' is shared between the main thread and the lead
150 * revalidator thread, so when it is desirable to halt revalidation, the
151 * main thread will set the latch. 'pause' and 'pause_barrier' are shared
152 * by revalidator threads. The lead revalidator will set 'pause' when it
153 * observes the latch has been set, and this will cause all revalidator
154 * threads to wait on 'pause_barrier' at the beginning of the next
155 * revalidation round. */
156 bool pause
; /* Set by leader on 'pause_latch. */
157 struct latch pause_latch
; /* Set to force revalidators pause. */
158 struct ovs_barrier pause_barrier
; /* Barrier used to pause all */
159 /* revalidators by main thread. */
161 /* There are 'N_UMAPS' maps containing 'struct udpif_key' elements.
163 * During the flow dump phase, revalidators insert into these with a random
164 * distribution. During the garbage collection phase, each revalidator
165 * takes care of garbage collecting a slice of these maps. */
168 /* Datapath flow statistics. */
169 unsigned int max_n_flows
;
170 unsigned int avg_n_flows
;
172 /* Following fields are accessed and modified by different threads. */
173 atomic_uint flow_limit
; /* Datapath flow hard limit. */
175 /* n_flows_mutex prevents multiple threads updating these concurrently. */
176 atomic_uint n_flows
; /* Number of flows in the datapath. */
177 atomic_llong n_flows_timestamp
; /* Last time n_flows was updated. */
178 struct ovs_mutex n_flows_mutex
;
180 /* Following fields are accessed and modified only from the main thread. */
181 struct unixctl_conn
**conns
; /* Connections waiting on dump_seq. */
182 uint64_t conn_seq
; /* Corresponds to 'dump_seq' when
183 conns[n_conns-1] was stored. */
184 size_t n_conns
; /* Number of connections waiting. */
186 long long int offload_rebalance_time
; /* Time of last offload rebalance */
190 BAD_UPCALL
, /* Some kind of bug somewhere. */
191 MISS_UPCALL
, /* A flow miss. */
192 SLOW_PATH_UPCALL
, /* Slow path upcall. */
193 SFLOW_UPCALL
, /* sFlow sample. */
194 FLOW_SAMPLE_UPCALL
, /* Per-flow sampling. */
195 IPFIX_UPCALL
, /* Per-bridge sampling. */
196 CONTROLLER_UPCALL
/* Destined for the controller. */
206 struct ofproto_dpif
*ofproto
; /* Parent ofproto. */
207 const struct recirc_id_node
*recirc
; /* Recirculation context. */
208 bool have_recirc_ref
; /* Reference held on recirc ctx? */
210 /* The flow and packet are only required to be constant when using
211 * dpif-netdev. If a modification is absolutely necessary, a const cast
212 * may be used with other datapaths. */
213 const struct flow
*flow
; /* Parsed representation of the packet. */
214 enum odp_key_fitness fitness
; /* Fitness of 'flow' relative to ODP key. */
215 const ovs_u128
*ufid
; /* Unique identifier for 'flow'. */
216 unsigned pmd_id
; /* Datapath poll mode driver id. */
217 const struct dp_packet
*packet
; /* Packet associated with this upcall. */
218 ofp_port_t ofp_in_port
; /* OpenFlow in port, or OFPP_NONE. */
219 uint16_t mru
; /* If !0, Maximum receive unit of
220 fragmented IP packet */
223 enum upcall_type type
; /* Type of the upcall. */
224 const struct nlattr
*actions
; /* Flow actions in DPIF_UC_ACTION Upcalls. */
226 bool xout_initialized
; /* True if 'xout' must be uninitialized. */
227 struct xlate_out xout
; /* Result of xlate_actions(). */
228 struct ofpbuf odp_actions
; /* Datapath actions from xlate_actions(). */
229 struct flow_wildcards wc
; /* Dependencies that megaflow must match. */
230 struct ofpbuf put_actions
; /* Actions 'put' in the fastpath. */
232 struct dpif_ipfix
*ipfix
; /* IPFIX pointer or NULL. */
233 struct dpif_sflow
*sflow
; /* SFlow pointer or NULL. */
235 struct udpif_key
*ukey
; /* Revalidator flow cache. */
236 bool ukey_persists
; /* Set true to keep 'ukey' beyond the
237 lifetime of this upcall. */
239 uint64_t reval_seq
; /* udpif->reval_seq at translation time. */
241 /* Not used by the upcall callback interface. */
242 const struct nlattr
*key
; /* Datapath flow key. */
243 size_t key_len
; /* Datapath flow key length. */
244 const struct nlattr
*out_tun_key
; /* Datapath output tunnel key. */
246 struct user_action_cookie cookie
;
248 uint64_t odp_actions_stub
[1024 / 8]; /* Stub for odp_actions. */
251 /* Ukeys must transition through these states using transition_ukey(). */
254 UKEY_VISIBLE
, /* Ukey is in umap, datapath flow install is queued. */
255 UKEY_OPERATIONAL
, /* Ukey is in umap, datapath flow is installed. */
256 UKEY_EVICTING
, /* Ukey is in umap, datapath flow delete is queued. */
257 UKEY_EVICTED
, /* Ukey is in umap, datapath flow is deleted. */
258 UKEY_DELETED
, /* Ukey removed from umap, ukey free is deferred. */
260 #define N_UKEY_STATES (UKEY_DELETED + 1)
262 /* 'udpif_key's are responsible for tracking the little bit of state udpif
263 * needs to do flow expiration which can't be pulled directly from the
264 * datapath. They may be created by any handler or revalidator thread at any
265 * time, and read by any revalidator during the dump phase. They are however
266 * each owned by a single revalidator which takes care of destroying them
267 * during the garbage-collection phase.
269 * The mutex within the ukey protects some members of the ukey. The ukey
270 * itself is protected by RCU and is held within a umap in the parent udpif.
271 * Adding or removing a ukey from a umap is only safe when holding the
272 * corresponding umap lock. */
274 struct cmap_node cmap_node
; /* In parent revalidator 'ukeys' map. */
276 /* These elements are read only once created, and therefore aren't
277 * protected by a mutex. */
278 const struct nlattr
*key
; /* Datapath flow key. */
279 size_t key_len
; /* Length of 'key'. */
280 const struct nlattr
*mask
; /* Datapath flow mask. */
281 size_t mask_len
; /* Length of 'mask'. */
282 ovs_u128 ufid
; /* Unique flow identifier. */
283 bool ufid_present
; /* True if 'ufid' is in datapath. */
284 uint32_t hash
; /* Pre-computed hash for 'key'. */
285 unsigned pmd_id
; /* Datapath poll mode driver id. */
287 struct ovs_mutex mutex
; /* Guards the following. */
288 struct dpif_flow_stats stats OVS_GUARDED
; /* Last known stats.*/
289 long long int created OVS_GUARDED
; /* Estimate of creation time. */
290 uint64_t dump_seq OVS_GUARDED
; /* Tracks udpif->dump_seq. */
291 uint64_t reval_seq OVS_GUARDED
; /* Tracks udpif->reval_seq. */
292 enum ukey_state state OVS_GUARDED
; /* Tracks ukey lifetime. */
294 /* 'state' debug information. */
295 unsigned int state_thread OVS_GUARDED
; /* Thread that transitions. */
296 const char *state_where OVS_GUARDED
; /* transition_ukey() locator. */
298 /* Datapath flow actions as nlattrs. Protected by RCU. Read with
299 * ukey_get_actions(), and write with ukey_set_actions(). */
300 OVSRCU_TYPE(struct ofpbuf
*) actions
;
302 struct xlate_cache
*xcache OVS_GUARDED
; /* Cache for xlate entries that
303 * are affected by this ukey.
304 * Used for stats and learning.*/
306 struct odputil_keybuf buf
;
310 uint32_t key_recirc_id
; /* Non-zero if reference is held by the ukey. */
311 struct recirc_refs recircs
; /* Action recirc IDs with references held. */
313 #define OFFL_REBAL_INTVL_MSEC 3000 /* dynamic offload rebalance freq */
314 struct netdev
*in_netdev
; /* in_odp_port's netdev */
315 bool offloaded
; /* True if flow is offloaded */
316 uint64_t flow_pps_rate
; /* Packets-Per-Second rate */
317 long long int flow_time
; /* last pps update time */
318 uint64_t flow_packets
; /* #pkts seen in interval */
319 uint64_t flow_backlog_packets
; /* prev-mode #pkts (offl or kernel) */
322 /* Datapath operation with optional ukey attached. */
324 struct udpif_key
*ukey
;
325 struct dpif_flow_stats stats
; /* Stats for 'op'. */
326 struct dpif_op dop
; /* Flow operation. */
329 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
330 static struct ovs_list all_udpifs
= OVS_LIST_INITIALIZER(&all_udpifs
);
332 static size_t recv_upcalls(struct handler
*);
333 static int process_upcall(struct udpif
*, struct upcall
*,
334 struct ofpbuf
*odp_actions
, struct flow_wildcards
*);
335 static void handle_upcalls(struct udpif
*, struct upcall
*, size_t n_upcalls
);
336 static void udpif_stop_threads(struct udpif
*, bool delete_flows
);
337 static void udpif_start_threads(struct udpif
*, size_t n_handlers
,
338 size_t n_revalidators
);
339 static void udpif_pause_revalidators(struct udpif
*);
340 static void udpif_resume_revalidators(struct udpif
*);
341 static void *udpif_upcall_handler(void *);
342 static void *udpif_revalidator(void *);
343 static unsigned long udpif_get_n_flows(struct udpif
*);
344 static void revalidate(struct revalidator
*);
345 static void revalidator_pause(struct revalidator
*);
346 static void revalidator_sweep(struct revalidator
*);
347 static void revalidator_purge(struct revalidator
*);
348 static void upcall_unixctl_show(struct unixctl_conn
*conn
, int argc
,
349 const char *argv
[], void *aux
);
350 static void upcall_unixctl_disable_megaflows(struct unixctl_conn
*, int argc
,
351 const char *argv
[], void *aux
);
352 static void upcall_unixctl_enable_megaflows(struct unixctl_conn
*, int argc
,
353 const char *argv
[], void *aux
);
354 static void upcall_unixctl_disable_ufid(struct unixctl_conn
*, int argc
,
355 const char *argv
[], void *aux
);
356 static void upcall_unixctl_enable_ufid(struct unixctl_conn
*, int argc
,
357 const char *argv
[], void *aux
);
358 static void upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
, int argc
,
359 const char *argv
[], void *aux
);
360 static void upcall_unixctl_dump_wait(struct unixctl_conn
*conn
, int argc
,
361 const char *argv
[], void *aux
);
362 static void upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc
,
363 const char *argv
[], void *aux
);
365 static struct udpif_key
*ukey_create_from_upcall(struct upcall
*,
366 struct flow_wildcards
*);
367 static int ukey_create_from_dpif_flow(const struct udpif
*,
368 const struct dpif_flow
*,
369 struct udpif_key
**);
370 static void ukey_get_actions(struct udpif_key
*, const struct nlattr
**actions
,
372 static bool ukey_install__(struct udpif
*, struct udpif_key
*ukey
)
373 OVS_TRY_LOCK(true, ukey
->mutex
);
374 static bool ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
);
375 static void transition_ukey_at(struct udpif_key
*ukey
, enum ukey_state dst
,
377 OVS_REQUIRES(ukey
->mutex
);
378 #define transition_ukey(UKEY, DST) \
379 transition_ukey_at(UKEY, DST, OVS_SOURCE_LOCATOR)
380 static struct udpif_key
*ukey_lookup(struct udpif
*udpif
,
381 const ovs_u128
*ufid
,
382 const unsigned pmd_id
);
383 static int ukey_acquire(struct udpif
*, const struct dpif_flow
*,
384 struct udpif_key
**result
, int *error
);
385 static void ukey_delete__(struct udpif_key
*);
386 static void ukey_delete(struct umap
*, struct udpif_key
*);
387 static enum upcall_type
classify_upcall(enum dpif_upcall_type type
,
388 const struct nlattr
*userdata
,
389 struct user_action_cookie
*cookie
);
391 static void put_op_init(struct ukey_op
*op
, struct udpif_key
*ukey
,
392 enum dpif_flow_put_flags flags
);
393 static void delete_op_init(struct udpif
*udpif
, struct ukey_op
*op
,
394 struct udpif_key
*ukey
);
396 static int upcall_receive(struct upcall
*, const struct dpif_backer
*,
397 const struct dp_packet
*packet
, enum dpif_upcall_type
,
398 const struct nlattr
*userdata
, const struct flow
*,
399 const unsigned int mru
,
400 const ovs_u128
*ufid
, const unsigned pmd_id
);
401 static void upcall_uninit(struct upcall
*);
403 static void udpif_flow_rebalance(struct udpif
*udpif
);
404 static int udpif_flow_program(struct udpif
*udpif
, struct udpif_key
*ukey
,
405 enum dpif_offload_type offload_type
);
406 static int udpif_flow_unprogram(struct udpif
*udpif
, struct udpif_key
*ukey
,
407 enum dpif_offload_type offload_type
);
409 static upcall_callback upcall_cb
;
410 static dp_purge_callback dp_purge_cb
;
412 static atomic_bool enable_megaflows
= ATOMIC_VAR_INIT(true);
413 static atomic_bool enable_ufid
= ATOMIC_VAR_INIT(true);
418 static struct ovsthread_once once
= OVSTHREAD_ONCE_INITIALIZER
;
419 if (ovsthread_once_start(&once
)) {
420 unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show
,
422 unixctl_command_register("upcall/disable-megaflows", "", 0, 0,
423 upcall_unixctl_disable_megaflows
, NULL
);
424 unixctl_command_register("upcall/enable-megaflows", "", 0, 0,
425 upcall_unixctl_enable_megaflows
, NULL
);
426 unixctl_command_register("upcall/disable-ufid", "", 0, 0,
427 upcall_unixctl_disable_ufid
, NULL
);
428 unixctl_command_register("upcall/enable-ufid", "", 0, 0,
429 upcall_unixctl_enable_ufid
, NULL
);
430 unixctl_command_register("upcall/set-flow-limit", "flow-limit-number",
431 1, 1, upcall_unixctl_set_flow_limit
, NULL
);
432 unixctl_command_register("revalidator/wait", "", 0, 0,
433 upcall_unixctl_dump_wait
, NULL
);
434 unixctl_command_register("revalidator/purge", "", 0, 0,
435 upcall_unixctl_purge
, NULL
);
436 ovsthread_once_done(&once
);
441 udpif_create(struct dpif_backer
*backer
, struct dpif
*dpif
)
443 struct udpif
*udpif
= xzalloc(sizeof *udpif
);
446 udpif
->backer
= backer
;
447 atomic_init(&udpif
->flow_limit
, MIN(ofproto_flow_limit
, 10000));
448 udpif
->reval_seq
= seq_create();
449 udpif
->dump_seq
= seq_create();
450 latch_init(&udpif
->exit_latch
);
451 latch_init(&udpif
->pause_latch
);
452 ovs_list_push_back(&all_udpifs
, &udpif
->list_node
);
453 atomic_init(&udpif
->enable_ufid
, false);
454 atomic_init(&udpif
->n_flows
, 0);
455 atomic_init(&udpif
->n_flows_timestamp
, LLONG_MIN
);
456 ovs_mutex_init(&udpif
->n_flows_mutex
);
457 udpif
->ukeys
= xmalloc(N_UMAPS
* sizeof *udpif
->ukeys
);
458 for (int i
= 0; i
< N_UMAPS
; i
++) {
459 cmap_init(&udpif
->ukeys
[i
].cmap
);
460 ovs_mutex_init(&udpif
->ukeys
[i
].mutex
);
463 dpif_register_upcall_cb(dpif
, upcall_cb
, udpif
);
464 dpif_register_dp_purge_cb(dpif
, dp_purge_cb
, udpif
);
470 udpif_run(struct udpif
*udpif
)
472 if (udpif
->conns
&& udpif
->conn_seq
!= seq_read(udpif
->dump_seq
)) {
475 for (i
= 0; i
< udpif
->n_conns
; i
++) {
476 unixctl_command_reply(udpif
->conns
[i
], NULL
);
485 udpif_destroy(struct udpif
*udpif
)
487 udpif_stop_threads(udpif
, false);
489 dpif_register_dp_purge_cb(udpif
->dpif
, NULL
, udpif
);
490 dpif_register_upcall_cb(udpif
->dpif
, NULL
, udpif
);
492 for (int i
= 0; i
< N_UMAPS
; i
++) {
493 cmap_destroy(&udpif
->ukeys
[i
].cmap
);
494 ovs_mutex_destroy(&udpif
->ukeys
[i
].mutex
);
499 ovs_list_remove(&udpif
->list_node
);
500 latch_destroy(&udpif
->exit_latch
);
501 latch_destroy(&udpif
->pause_latch
);
502 seq_destroy(udpif
->reval_seq
);
503 seq_destroy(udpif
->dump_seq
);
504 ovs_mutex_destroy(&udpif
->n_flows_mutex
);
508 /* Stops the handler and revalidator threads.
510 * If 'delete_flows' is true, we delete ukeys and delete all flows from the
511 * datapath. Otherwise, we end up double-counting stats for flows that remain
512 * in the datapath. If 'delete_flows' is false, we skip this step. This is
513 * appropriate if OVS is about to exit anyway and it is desirable to let
514 * existing network connections continue being forwarded afterward. */
516 udpif_stop_threads(struct udpif
*udpif
, bool delete_flows
)
518 if (udpif
&& (udpif
->n_handlers
!= 0 || udpif
->n_revalidators
!= 0)) {
521 /* Tell the threads to exit. */
522 latch_set(&udpif
->exit_latch
);
524 /* Wait for the threads to exit. Quiesce because this can take a long
526 ovsrcu_quiesce_start();
527 for (i
= 0; i
< udpif
->n_handlers
; i
++) {
528 xpthread_join(udpif
->handlers
[i
].thread
, NULL
);
530 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
531 xpthread_join(udpif
->revalidators
[i
].thread
, NULL
);
533 dpif_disable_upcall(udpif
->dpif
);
534 ovsrcu_quiesce_end();
537 for (i
= 0; i
< udpif
->n_revalidators
; i
++) {
538 revalidator_purge(&udpif
->revalidators
[i
]);
542 latch_poll(&udpif
->exit_latch
);
544 ovs_barrier_destroy(&udpif
->reval_barrier
);
545 ovs_barrier_destroy(&udpif
->pause_barrier
);
547 free(udpif
->revalidators
);
548 udpif
->revalidators
= NULL
;
549 udpif
->n_revalidators
= 0;
551 free(udpif
->handlers
);
552 udpif
->handlers
= NULL
;
553 udpif
->n_handlers
= 0;
557 /* Starts the handler and revalidator threads. */
559 udpif_start_threads(struct udpif
*udpif
, size_t n_handlers_
,
560 size_t n_revalidators_
)
562 if (udpif
&& n_handlers_
&& n_revalidators_
) {
563 /* Creating a thread can take a significant amount of time on some
564 * systems, even hundred of milliseconds, so quiesce around it. */
565 ovsrcu_quiesce_start();
567 udpif
->n_handlers
= n_handlers_
;
568 udpif
->n_revalidators
= n_revalidators_
;
570 udpif
->handlers
= xzalloc(udpif
->n_handlers
* sizeof *udpif
->handlers
);
571 for (size_t i
= 0; i
< udpif
->n_handlers
; i
++) {
572 struct handler
*handler
= &udpif
->handlers
[i
];
574 handler
->udpif
= udpif
;
575 handler
->handler_id
= i
;
576 handler
->thread
= ovs_thread_create(
577 "handler", udpif_upcall_handler
, handler
);
580 atomic_init(&udpif
->enable_ufid
, udpif
->backer
->rt_support
.ufid
);
581 dpif_enable_upcall(udpif
->dpif
);
583 ovs_barrier_init(&udpif
->reval_barrier
, udpif
->n_revalidators
);
584 ovs_barrier_init(&udpif
->pause_barrier
, udpif
->n_revalidators
+ 1);
585 udpif
->reval_exit
= false;
586 udpif
->pause
= false;
587 udpif
->offload_rebalance_time
= time_msec();
588 udpif
->revalidators
= xzalloc(udpif
->n_revalidators
589 * sizeof *udpif
->revalidators
);
590 for (size_t i
= 0; i
< udpif
->n_revalidators
; i
++) {
591 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
593 revalidator
->udpif
= udpif
;
594 revalidator
->thread
= ovs_thread_create(
595 "revalidator", udpif_revalidator
, revalidator
);
597 ovsrcu_quiesce_end();
601 /* Pauses all revalidators. Should only be called by the main thread.
602 * When function returns, all revalidators are paused and will proceed
603 * only after udpif_resume_revalidators() is called. */
605 udpif_pause_revalidators(struct udpif
*udpif
)
607 if (udpif
->backer
->recv_set_enable
) {
608 latch_set(&udpif
->pause_latch
);
609 ovs_barrier_block(&udpif
->pause_barrier
);
613 /* Resumes the pausing of revalidators. Should only be called by the
616 udpif_resume_revalidators(struct udpif
*udpif
)
618 if (udpif
->backer
->recv_set_enable
) {
619 latch_poll(&udpif
->pause_latch
);
620 ovs_barrier_block(&udpif
->pause_barrier
);
624 /* Tells 'udpif' how many threads it should use to handle upcalls.
625 * 'n_handlers_' and 'n_revalidators_' can never be zero. 'udpif''s
626 * datapath handle must have packet reception enabled before starting
629 udpif_set_threads(struct udpif
*udpif
, size_t n_handlers_
,
630 size_t n_revalidators_
)
633 ovs_assert(n_handlers_
&& n_revalidators_
);
635 if (udpif
->n_handlers
!= n_handlers_
636 || udpif
->n_revalidators
!= n_revalidators_
) {
637 udpif_stop_threads(udpif
, true);
640 if (!udpif
->handlers
&& !udpif
->revalidators
) {
643 error
= dpif_handlers_set(udpif
->dpif
, n_handlers_
);
645 VLOG_ERR("failed to configure handlers in dpif %s: %s",
646 dpif_name(udpif
->dpif
), ovs_strerror(error
));
650 udpif_start_threads(udpif
, n_handlers_
, n_revalidators_
);
654 /* Notifies 'udpif' that something changed which may render previous
655 * xlate_actions() results invalid. */
657 udpif_revalidate(struct udpif
*udpif
)
659 seq_change(udpif
->reval_seq
);
662 /* Returns a seq which increments every time 'udpif' pulls stats from the
663 * datapath. Callers can use this to get a sense of when might be a good time
664 * to do periodic work which relies on relatively up to date statistics. */
666 udpif_dump_seq(struct udpif
*udpif
)
668 return udpif
->dump_seq
;
672 udpif_get_memory_usage(struct udpif
*udpif
, struct simap
*usage
)
676 simap_increase(usage
, "handlers", udpif
->n_handlers
);
678 simap_increase(usage
, "revalidators", udpif
->n_revalidators
);
679 for (i
= 0; i
< N_UMAPS
; i
++) {
680 simap_increase(usage
, "udpif keys", cmap_count(&udpif
->ukeys
[i
].cmap
));
684 /* Remove flows from a single datapath. */
686 udpif_flush(struct udpif
*udpif
)
688 size_t n_handlers_
= udpif
->n_handlers
;
689 size_t n_revalidators_
= udpif
->n_revalidators
;
691 udpif_stop_threads(udpif
, true);
692 dpif_flow_flush(udpif
->dpif
);
693 udpif_start_threads(udpif
, n_handlers_
, n_revalidators_
);
696 /* Removes all flows from all datapaths. */
698 udpif_flush_all_datapaths(void)
702 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
708 udpif_use_ufid(struct udpif
*udpif
)
712 atomic_read_relaxed(&enable_ufid
, &enable
);
713 return enable
&& udpif
->backer
->rt_support
.ufid
;
718 udpif_get_n_flows(struct udpif
*udpif
)
720 long long int time
, now
;
721 unsigned long flow_count
;
724 atomic_read_relaxed(&udpif
->n_flows_timestamp
, &time
);
725 if (time
< now
- 100 && !ovs_mutex_trylock(&udpif
->n_flows_mutex
)) {
726 struct dpif_dp_stats stats
;
728 atomic_store_relaxed(&udpif
->n_flows_timestamp
, now
);
729 dpif_get_dp_stats(udpif
->dpif
, &stats
);
730 flow_count
= stats
.n_flows
;
731 atomic_store_relaxed(&udpif
->n_flows
, flow_count
);
732 ovs_mutex_unlock(&udpif
->n_flows_mutex
);
734 atomic_read_relaxed(&udpif
->n_flows
, &flow_count
);
739 /* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH
740 * upcalls from dpif, processes the batch and installs corresponding flows
743 udpif_upcall_handler(void *arg
)
745 struct handler
*handler
= arg
;
746 struct udpif
*udpif
= handler
->udpif
;
748 while (!latch_is_set(&handler
->udpif
->exit_latch
)) {
749 if (recv_upcalls(handler
)) {
750 poll_immediate_wake();
752 dpif_recv_wait(udpif
->dpif
, handler
->handler_id
);
753 latch_wait(&udpif
->exit_latch
);
762 recv_upcalls(struct handler
*handler
)
764 struct udpif
*udpif
= handler
->udpif
;
765 uint64_t recv_stubs
[UPCALL_MAX_BATCH
][512 / 8];
766 struct ofpbuf recv_bufs
[UPCALL_MAX_BATCH
];
767 struct dpif_upcall dupcalls
[UPCALL_MAX_BATCH
];
768 struct upcall upcalls
[UPCALL_MAX_BATCH
];
769 struct flow flows
[UPCALL_MAX_BATCH
];
773 while (n_upcalls
< UPCALL_MAX_BATCH
) {
774 struct ofpbuf
*recv_buf
= &recv_bufs
[n_upcalls
];
775 struct dpif_upcall
*dupcall
= &dupcalls
[n_upcalls
];
776 struct upcall
*upcall
= &upcalls
[n_upcalls
];
777 struct flow
*flow
= &flows
[n_upcalls
];
778 unsigned int mru
= 0;
782 ofpbuf_use_stub(recv_buf
, recv_stubs
[n_upcalls
],
783 sizeof recv_stubs
[n_upcalls
]);
784 if (dpif_recv(udpif
->dpif
, handler
->handler_id
, dupcall
, recv_buf
)) {
785 ofpbuf_uninit(recv_buf
);
789 upcall
->fitness
= odp_flow_key_to_flow(dupcall
->key
, dupcall
->key_len
,
791 if (upcall
->fitness
== ODP_FIT_ERROR
) {
796 mru
= nl_attr_get_u16(dupcall
->mru
);
800 hash
= nl_attr_get_u64(dupcall
->hash
);
803 error
= upcall_receive(upcall
, udpif
->backer
, &dupcall
->packet
,
804 dupcall
->type
, dupcall
->userdata
, flow
, mru
,
805 &dupcall
->ufid
, PMD_ID_NULL
);
807 if (error
== ENODEV
) {
808 /* Received packet on datapath port for which we couldn't
809 * associate an ofproto. This can happen if a port is removed
810 * while traffic is being received. Print a rate-limited
811 * message in case it happens frequently. */
812 dpif_flow_put(udpif
->dpif
, DPIF_FP_CREATE
, dupcall
->key
,
813 dupcall
->key_len
, NULL
, 0, NULL
, 0,
814 &dupcall
->ufid
, PMD_ID_NULL
, NULL
);
815 VLOG_INFO_RL(&rl
, "received packet on unassociated datapath "
816 "port %"PRIu32
, flow
->in_port
.odp_port
);
821 upcall
->key
= dupcall
->key
;
822 upcall
->key_len
= dupcall
->key_len
;
823 upcall
->ufid
= &dupcall
->ufid
;
826 upcall
->out_tun_key
= dupcall
->out_tun_key
;
827 upcall
->actions
= dupcall
->actions
;
829 pkt_metadata_from_flow(&dupcall
->packet
.md
, flow
);
830 flow_extract(&dupcall
->packet
, flow
);
832 error
= process_upcall(udpif
, upcall
,
833 &upcall
->odp_actions
, &upcall
->wc
);
842 upcall_uninit(upcall
);
844 dp_packet_uninit(&dupcall
->packet
);
845 ofpbuf_uninit(recv_buf
);
849 handle_upcalls(handler
->udpif
, upcalls
, n_upcalls
);
850 for (i
= 0; i
< n_upcalls
; i
++) {
851 dp_packet_uninit(&dupcalls
[i
].packet
);
852 ofpbuf_uninit(&recv_bufs
[i
]);
853 upcall_uninit(&upcalls
[i
]);
861 udpif_run_flow_rebalance(struct udpif
*udpif
)
863 long long int now
= 0;
865 /* Don't rebalance if OFFL_REBAL_INTVL_MSEC have not elapsed */
867 if (now
< udpif
->offload_rebalance_time
+ OFFL_REBAL_INTVL_MSEC
) {
871 if (!netdev_any_oor()) {
875 VLOG_DBG("Offload rebalance: Found OOR netdevs");
876 udpif
->offload_rebalance_time
= now
;
877 udpif_flow_rebalance(udpif
);
881 udpif_revalidator(void *arg
)
883 /* Used by all revalidators. */
884 struct revalidator
*revalidator
= arg
;
885 struct udpif
*udpif
= revalidator
->udpif
;
886 bool leader
= revalidator
== &udpif
->revalidators
[0];
888 /* Used only by the leader. */
889 long long int start_time
= 0;
890 uint64_t last_reval_seq
= 0;
893 revalidator
->id
= ovsthread_id_self();
898 recirc_run(); /* Recirculation cleanup. */
900 reval_seq
= seq_read(udpif
->reval_seq
);
901 last_reval_seq
= reval_seq
;
903 n_flows
= udpif_get_n_flows(udpif
);
904 udpif
->max_n_flows
= MAX(n_flows
, udpif
->max_n_flows
);
905 udpif
->avg_n_flows
= (udpif
->avg_n_flows
+ n_flows
) / 2;
907 /* Only the leader checks the pause latch to prevent a race where
908 * some threads think it's false and proceed to block on
909 * reval_barrier and others think it's true and block indefinitely
910 * on the pause_barrier */
911 udpif
->pause
= latch_is_set(&udpif
->pause_latch
);
913 /* Only the leader checks the exit latch to prevent a race where
914 * some threads think it's true and exit and others think it's
915 * false and block indefinitely on the reval_barrier */
916 udpif
->reval_exit
= latch_is_set(&udpif
->exit_latch
);
918 start_time
= time_msec();
919 if (!udpif
->reval_exit
) {
922 terse_dump
= udpif_use_ufid(udpif
);
923 udpif
->dump
= dpif_flow_dump_create(udpif
->dpif
, terse_dump
,
928 /* Wait for the leader to start the flow dump. */
929 ovs_barrier_block(&udpif
->reval_barrier
);
931 revalidator_pause(revalidator
);
934 if (udpif
->reval_exit
) {
937 revalidate(revalidator
);
939 /* Wait for all flows to have been dumped before we garbage collect. */
940 ovs_barrier_block(&udpif
->reval_barrier
);
941 revalidator_sweep(revalidator
);
943 /* Wait for all revalidators to finish garbage collection. */
944 ovs_barrier_block(&udpif
->reval_barrier
);
947 unsigned int flow_limit
;
948 long long int duration
;
950 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
952 dpif_flow_dump_destroy(udpif
->dump
);
953 seq_change(udpif
->dump_seq
);
954 if (netdev_is_offload_rebalance_policy_enabled()) {
955 udpif_run_flow_rebalance(udpif
);
958 duration
= MAX(time_msec() - start_time
, 1);
959 udpif
->dump_duration
= duration
;
960 if (duration
> 2000) {
961 flow_limit
/= duration
/ 1000;
962 } else if (duration
> 1300) {
963 flow_limit
= flow_limit
* 3 / 4;
964 } else if (duration
< 1000 &&
965 flow_limit
< n_flows
* 1000 / duration
) {
968 flow_limit
= MIN(ofproto_flow_limit
, MAX(flow_limit
, 1000));
969 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
971 if (duration
> 2000) {
972 VLOG_INFO("Spent an unreasonably long %lldms dumping flows",
976 poll_timer_wait_until(start_time
+ MIN(ofproto_max_idle
,
977 ofproto_max_revalidator
));
978 seq_wait(udpif
->reval_seq
, last_reval_seq
);
979 latch_wait(&udpif
->exit_latch
);
980 latch_wait(&udpif
->pause_latch
);
983 if (!latch_is_set(&udpif
->pause_latch
) &&
984 !latch_is_set(&udpif
->exit_latch
)) {
985 long long int now
= time_msec();
986 /* Block again if we are woken up within 5ms of the last start
990 if (now
< start_time
) {
991 poll_timer_wait_until(start_time
);
992 latch_wait(&udpif
->exit_latch
);
993 latch_wait(&udpif
->pause_latch
);
1003 static enum upcall_type
1004 classify_upcall(enum dpif_upcall_type type
, const struct nlattr
*userdata
,
1005 struct user_action_cookie
*cookie
)
1007 /* First look at the upcall type. */
1009 case DPIF_UC_ACTION
:
1015 case DPIF_N_UC_TYPES
:
1017 VLOG_WARN_RL(&rl
, "upcall has unexpected type %"PRIu32
, type
);
1021 /* "action" upcalls need a closer look. */
1023 VLOG_WARN_RL(&rl
, "action upcall missing cookie");
1027 size_t userdata_len
= nl_attr_get_size(userdata
);
1028 if (userdata_len
!= sizeof *cookie
) {
1029 VLOG_WARN_RL(&rl
, "action upcall cookie has unexpected size %"PRIuSIZE
,
1033 memcpy(cookie
, nl_attr_get(userdata
), sizeof *cookie
);
1034 if (cookie
->type
== USER_ACTION_COOKIE_SFLOW
) {
1035 return SFLOW_UPCALL
;
1036 } else if (cookie
->type
== USER_ACTION_COOKIE_SLOW_PATH
) {
1037 return SLOW_PATH_UPCALL
;
1038 } else if (cookie
->type
== USER_ACTION_COOKIE_FLOW_SAMPLE
) {
1039 return FLOW_SAMPLE_UPCALL
;
1040 } else if (cookie
->type
== USER_ACTION_COOKIE_IPFIX
) {
1041 return IPFIX_UPCALL
;
1042 } else if (cookie
->type
== USER_ACTION_COOKIE_CONTROLLER
) {
1043 return CONTROLLER_UPCALL
;
1045 VLOG_WARN_RL(&rl
, "invalid user cookie of type %"PRIu16
1046 " and size %"PRIuSIZE
, cookie
->type
, userdata_len
);
1051 /* Calculates slow path actions for 'xout'. 'buf' must statically be
1052 * initialized with at least 128 bytes of space. */
1054 compose_slow_path(struct udpif
*udpif
, struct xlate_out
*xout
,
1055 odp_port_t odp_in_port
, ofp_port_t ofp_in_port
,
1056 struct ofpbuf
*buf
, uint32_t meter_id
,
1057 struct uuid
*ofproto_uuid
)
1059 struct user_action_cookie cookie
;
1063 memset(&cookie
, 0, sizeof cookie
);
1064 cookie
.type
= USER_ACTION_COOKIE_SLOW_PATH
;
1065 cookie
.ofp_in_port
= ofp_in_port
;
1066 cookie
.ofproto_uuid
= *ofproto_uuid
;
1067 cookie
.slow_path
.reason
= xout
->slow
;
1069 port
= xout
->slow
& (SLOW_CFM
| SLOW_BFD
| SLOW_LACP
| SLOW_STP
)
1072 pid
= dpif_port_get_pid(udpif
->dpif
, port
);
1076 if (meter_id
!= UINT32_MAX
) {
1077 /* If slowpath meter is configured, generate clone(meter, userspace)
1079 offset
= nl_msg_start_nested(buf
, OVS_ACTION_ATTR_SAMPLE
);
1080 nl_msg_put_u32(buf
, OVS_SAMPLE_ATTR_PROBABILITY
, UINT32_MAX
);
1081 ac_offset
= nl_msg_start_nested(buf
, OVS_SAMPLE_ATTR_ACTIONS
);
1082 nl_msg_put_u32(buf
, OVS_ACTION_ATTR_METER
, meter_id
);
1085 odp_put_userspace_action(pid
, &cookie
, sizeof cookie
,
1086 ODPP_NONE
, false, buf
);
1088 if (meter_id
!= UINT32_MAX
) {
1089 nl_msg_end_nested(buf
, ac_offset
);
1090 nl_msg_end_nested(buf
, offset
);
1094 /* If there is no error, the upcall must be destroyed with upcall_uninit()
1095 * before quiescing, as the referred objects are guaranteed to exist only
1096 * until the calling thread quiesces. Otherwise, do not call upcall_uninit()
1097 * since the 'upcall->put_actions' remains uninitialized. */
1099 upcall_receive(struct upcall
*upcall
, const struct dpif_backer
*backer
,
1100 const struct dp_packet
*packet
, enum dpif_upcall_type type
,
1101 const struct nlattr
*userdata
, const struct flow
*flow
,
1102 const unsigned int mru
,
1103 const ovs_u128
*ufid
, const unsigned pmd_id
)
1107 upcall
->type
= classify_upcall(type
, userdata
, &upcall
->cookie
);
1108 if (upcall
->type
== BAD_UPCALL
) {
1110 } else if (upcall
->type
== MISS_UPCALL
) {
1111 error
= xlate_lookup(backer
, flow
, &upcall
->ofproto
, &upcall
->ipfix
,
1112 &upcall
->sflow
, NULL
, &upcall
->ofp_in_port
);
1117 struct ofproto_dpif
*ofproto
1118 = ofproto_dpif_lookup_by_uuid(&upcall
->cookie
.ofproto_uuid
);
1120 VLOG_INFO_RL(&rl
, "upcall could not find ofproto");
1123 upcall
->ofproto
= ofproto
;
1124 upcall
->ipfix
= ofproto
->ipfix
;
1125 upcall
->sflow
= ofproto
->sflow
;
1126 upcall
->ofp_in_port
= upcall
->cookie
.ofp_in_port
;
1129 upcall
->recirc
= NULL
;
1130 upcall
->have_recirc_ref
= false;
1131 upcall
->flow
= flow
;
1132 upcall
->packet
= packet
;
1133 upcall
->ufid
= ufid
;
1134 upcall
->pmd_id
= pmd_id
;
1135 ofpbuf_use_stub(&upcall
->odp_actions
, upcall
->odp_actions_stub
,
1136 sizeof upcall
->odp_actions_stub
);
1137 ofpbuf_init(&upcall
->put_actions
, 0);
1139 upcall
->xout_initialized
= false;
1140 upcall
->ukey_persists
= false;
1142 upcall
->ukey
= NULL
;
1144 upcall
->key_len
= 0;
1147 upcall
->out_tun_key
= NULL
;
1148 upcall
->actions
= NULL
;
1154 upcall_xlate(struct udpif
*udpif
, struct upcall
*upcall
,
1155 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1157 struct dpif_flow_stats stats
;
1158 enum xlate_error xerr
;
1159 struct xlate_in xin
;
1162 stats
.n_packets
= 1;
1163 stats
.n_bytes
= dp_packet_size(upcall
->packet
);
1164 stats
.used
= time_msec();
1165 stats
.tcp_flags
= ntohs(upcall
->flow
->tcp_flags
);
1167 xlate_in_init(&xin
, upcall
->ofproto
,
1168 ofproto_dpif_get_tables_version(upcall
->ofproto
),
1169 upcall
->flow
, upcall
->ofp_in_port
, NULL
,
1170 stats
.tcp_flags
, upcall
->packet
, wc
, odp_actions
);
1172 if (upcall
->type
== MISS_UPCALL
) {
1173 xin
.resubmit_stats
= &stats
;
1175 if (xin
.frozen_state
) {
1176 /* We may install a datapath flow only if we get a reference to the
1177 * recirculation context (otherwise we could have recirculation
1178 * upcalls using recirculation ID for which no context can be
1179 * found). We may still execute the flow's actions even if we
1180 * don't install the flow. */
1181 upcall
->recirc
= recirc_id_node_from_state(xin
.frozen_state
);
1182 upcall
->have_recirc_ref
= recirc_id_node_try_ref_rcu(upcall
->recirc
);
1185 /* For non-miss upcalls, we are either executing actions (one of which
1186 * is an userspace action) for an upcall, in which case the stats have
1187 * already been taken care of, or there's a flow in the datapath which
1188 * this packet was accounted to. Presumably the revalidators will deal
1189 * with pushing its stats eventually. */
1192 upcall
->reval_seq
= seq_read(udpif
->reval_seq
);
1194 xerr
= xlate_actions(&xin
, &upcall
->xout
);
1196 /* Translate again and log the ofproto trace for
1197 * these two error types. */
1198 if (xerr
== XLATE_RECURSION_TOO_DEEP
||
1199 xerr
== XLATE_TOO_MANY_RESUBMITS
) {
1200 static struct vlog_rate_limit rll
= VLOG_RATE_LIMIT_INIT(1, 1);
1202 /* This is a huge log, so be conservative. */
1203 if (!VLOG_DROP_WARN(&rll
)) {
1205 ofproto_trace(upcall
->ofproto
, upcall
->flow
,
1206 upcall
->packet
, NULL
, 0, NULL
, &output
);
1207 VLOG_WARN("%s", ds_cstr(&output
));
1208 ds_destroy(&output
);
1213 /* Convert the input port wildcard from OFP to ODP format. There's no
1214 * real way to do this for arbitrary bitmasks since the numbering spaces
1215 * aren't the same. However, flow translation always exact matches the
1216 * whole thing, so we can do the same here. */
1217 WC_MASK_FIELD(wc
, in_port
.odp_port
);
1220 upcall
->xout_initialized
= true;
1222 if (upcall
->fitness
== ODP_FIT_TOO_LITTLE
) {
1223 upcall
->xout
.slow
|= SLOW_MATCH
;
1225 if (!upcall
->xout
.slow
) {
1226 ofpbuf_use_const(&upcall
->put_actions
,
1227 odp_actions
->data
, odp_actions
->size
);
1229 /* upcall->put_actions already initialized by upcall_receive(). */
1230 compose_slow_path(udpif
, &upcall
->xout
,
1231 upcall
->flow
->in_port
.odp_port
, upcall
->ofp_in_port
,
1232 &upcall
->put_actions
,
1233 upcall
->ofproto
->up
.slowpath_meter_id
,
1234 &upcall
->ofproto
->uuid
);
1237 /* This function is also called for slow-pathed flows. As we are only
1238 * going to create new datapath flows for actual datapath misses, there is
1239 * no point in creating a ukey otherwise. */
1240 if (upcall
->type
== MISS_UPCALL
) {
1241 upcall
->ukey
= ukey_create_from_upcall(upcall
, wc
);
1246 upcall_uninit(struct upcall
*upcall
)
1249 if (upcall
->xout_initialized
) {
1250 xlate_out_uninit(&upcall
->xout
);
1252 ofpbuf_uninit(&upcall
->odp_actions
);
1253 ofpbuf_uninit(&upcall
->put_actions
);
1255 if (!upcall
->ukey_persists
) {
1256 ukey_delete__(upcall
->ukey
);
1258 } else if (upcall
->have_recirc_ref
) {
1259 /* The reference was transferred to the ukey if one was created. */
1260 recirc_id_node_unref(upcall
->recirc
);
1265 /* If there are less flows than the limit, and this is a miss upcall which
1267 * - Has no recirc_id, OR
1268 * - Has a recirc_id and we can get a reference on the recirc ctx,
1270 * Then we should install the flow (true). Otherwise, return false. */
1272 should_install_flow(struct udpif
*udpif
, struct upcall
*upcall
)
1274 unsigned int flow_limit
;
1276 if (upcall
->type
!= MISS_UPCALL
) {
1278 } else if (upcall
->recirc
&& !upcall
->have_recirc_ref
) {
1279 VLOG_DBG_RL(&rl
, "upcall: no reference for recirc flow");
1283 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
1284 if (udpif_get_n_flows(udpif
) >= flow_limit
) {
1285 COVERAGE_INC(upcall_flow_limit_hit
);
1286 VLOG_WARN_RL(&rl
, "upcall: datapath flow limit reached");
1294 upcall_cb(const struct dp_packet
*packet
, const struct flow
*flow
, ovs_u128
*ufid
,
1295 unsigned pmd_id
, enum dpif_upcall_type type
,
1296 const struct nlattr
*userdata
, struct ofpbuf
*actions
,
1297 struct flow_wildcards
*wc
, struct ofpbuf
*put_actions
, void *aux
)
1299 struct udpif
*udpif
= aux
;
1300 struct upcall upcall
;
1304 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1306 error
= upcall_receive(&upcall
, udpif
->backer
, packet
, type
, userdata
,
1307 flow
, 0, ufid
, pmd_id
);
1312 upcall
.fitness
= ODP_FIT_PERFECT
;
1313 error
= process_upcall(udpif
, &upcall
, actions
, wc
);
1318 if (upcall
.xout
.slow
&& put_actions
) {
1319 ofpbuf_put(put_actions
, upcall
.put_actions
.data
,
1320 upcall
.put_actions
.size
);
1323 if (OVS_UNLIKELY(!megaflow
&& wc
)) {
1324 flow_wildcards_init_for_packet(wc
, flow
);
1327 if (!should_install_flow(udpif
, &upcall
)) {
1332 if (upcall
.ukey
&& !ukey_install(udpif
, upcall
.ukey
)) {
1333 static struct vlog_rate_limit rll
= VLOG_RATE_LIMIT_INIT(1, 1);
1334 VLOG_WARN_RL(&rll
, "upcall_cb failure: ukey installation fails");
1339 upcall
.ukey_persists
= true;
1341 upcall_uninit(&upcall
);
1346 dpif_get_actions(struct udpif
*udpif
, struct upcall
*upcall
,
1347 const struct nlattr
**actions
)
1349 size_t actions_len
= 0;
1351 if (upcall
->actions
) {
1352 /* Actions were passed up from datapath. */
1353 *actions
= nl_attr_get(upcall
->actions
);
1354 actions_len
= nl_attr_get_size(upcall
->actions
);
1357 if (actions_len
== 0) {
1358 /* Lookup actions in userspace cache. */
1359 struct udpif_key
*ukey
= ukey_lookup(udpif
, upcall
->ufid
,
1362 ukey_get_actions(ukey
, actions
, &actions_len
);
1370 dpif_read_actions(struct udpif
*udpif
, struct upcall
*upcall
,
1371 const struct flow
*flow
, enum upcall_type type
,
1374 const struct nlattr
*actions
= NULL
;
1375 size_t actions_len
= dpif_get_actions(udpif
, upcall
, &actions
);
1377 if (!actions
|| !actions_len
) {
1383 dpif_sflow_read_actions(flow
, actions
, actions_len
, upcall_data
, true);
1385 case FLOW_SAMPLE_UPCALL
:
1387 dpif_ipfix_read_actions(flow
, actions
, actions_len
, upcall_data
);
1391 case SLOW_PATH_UPCALL
:
1392 case CONTROLLER_UPCALL
:
1401 process_upcall(struct udpif
*udpif
, struct upcall
*upcall
,
1402 struct ofpbuf
*odp_actions
, struct flow_wildcards
*wc
)
1404 const struct dp_packet
*packet
= upcall
->packet
;
1405 const struct flow
*flow
= upcall
->flow
;
1406 size_t actions_len
= 0;
1408 switch (upcall
->type
) {
1410 case SLOW_PATH_UPCALL
:
1411 upcall_xlate(udpif
, upcall
, odp_actions
, wc
);
1415 if (upcall
->sflow
) {
1416 struct dpif_sflow_actions sflow_actions
;
1418 memset(&sflow_actions
, 0, sizeof sflow_actions
);
1420 actions_len
= dpif_read_actions(udpif
, upcall
, flow
,
1421 upcall
->type
, &sflow_actions
);
1422 dpif_sflow_received(upcall
->sflow
, packet
, flow
,
1423 flow
->in_port
.odp_port
, &upcall
->cookie
,
1424 actions_len
> 0 ? &sflow_actions
: NULL
);
1429 case FLOW_SAMPLE_UPCALL
:
1430 if (upcall
->ipfix
) {
1431 struct flow_tnl output_tunnel_key
;
1432 struct dpif_ipfix_actions ipfix_actions
;
1434 memset(&ipfix_actions
, 0, sizeof ipfix_actions
);
1436 if (upcall
->out_tun_key
) {
1437 odp_tun_key_from_attr(upcall
->out_tun_key
, &output_tunnel_key
,
1441 actions_len
= dpif_read_actions(udpif
, upcall
, flow
,
1442 upcall
->type
, &ipfix_actions
);
1443 if (upcall
->type
== IPFIX_UPCALL
) {
1444 dpif_ipfix_bridge_sample(upcall
->ipfix
, packet
, flow
,
1445 flow
->in_port
.odp_port
,
1446 upcall
->cookie
.ipfix
.output_odp_port
,
1447 upcall
->out_tun_key
?
1448 &output_tunnel_key
: NULL
,
1450 &ipfix_actions
: NULL
);
1452 /* The flow reflects exactly the contents of the packet.
1453 * Sample the packet using it. */
1454 dpif_ipfix_flow_sample(upcall
->ipfix
, packet
, flow
,
1455 &upcall
->cookie
, flow
->in_port
.odp_port
,
1456 upcall
->out_tun_key
?
1457 &output_tunnel_key
: NULL
,
1458 actions_len
> 0 ? &ipfix_actions
: NULL
);
1463 case CONTROLLER_UPCALL
:
1465 struct user_action_cookie
*cookie
= &upcall
->cookie
;
1467 if (cookie
->controller
.dont_send
) {
1471 uint32_t recirc_id
= cookie
->controller
.recirc_id
;
1476 const struct recirc_id_node
*recirc_node
1477 = recirc_id_node_find(recirc_id
);
1482 const struct frozen_state
*state
= &recirc_node
->state
;
1484 struct ofproto_async_msg
*am
= xmalloc(sizeof *am
);
1485 *am
= (struct ofproto_async_msg
) {
1486 .controller_id
= cookie
->controller
.controller_id
,
1487 .oam
= OAM_PACKET_IN
,
1491 .packet
= xmemdup(dp_packet_data(packet
),
1492 dp_packet_size(packet
)),
1493 .packet_len
= dp_packet_size(packet
),
1494 .reason
= cookie
->controller
.reason
,
1495 .table_id
= state
->table_id
,
1496 .cookie
= get_32aligned_be64(
1497 &cookie
->controller
.rule_cookie
),
1498 .userdata
= (recirc_node
->state
.userdata_len
1499 ? xmemdup(recirc_node
->state
.userdata
,
1500 recirc_node
->state
.userdata_len
)
1502 .userdata_len
= recirc_node
->state
.userdata_len
,
1505 .max_len
= cookie
->controller
.max_len
,
1509 if (cookie
->controller
.continuation
) {
1510 am
->pin
.up
.stack
= (state
->stack_size
1511 ? xmemdup(state
->stack
, state
->stack_size
)
1513 am
->pin
.up
.stack_size
= state
->stack_size
,
1514 am
->pin
.up
.mirrors
= state
->mirrors
,
1515 am
->pin
.up
.conntracked
= state
->conntracked
,
1516 am
->pin
.up
.actions
= (state
->ofpacts_len
1517 ? xmemdup(state
->ofpacts
,
1518 state
->ofpacts_len
) : NULL
),
1519 am
->pin
.up
.actions_len
= state
->ofpacts_len
,
1520 am
->pin
.up
.action_set
= (state
->action_set_len
1521 ? xmemdup(state
->action_set
,
1522 state
->action_set_len
)
1524 am
->pin
.up
.action_set_len
= state
->action_set_len
,
1525 am
->pin
.up
.bridge
= upcall
->ofproto
->uuid
;
1526 am
->pin
.up
.odp_port
= upcall
->packet
->md
.in_port
.odp_port
;
1529 /* We don't want to use the upcall 'flow', since it may be
1530 * more specific than the point at which the "controller"
1531 * action was specified. */
1532 struct flow frozen_flow
;
1534 frozen_flow
= *flow
;
1535 if (!state
->conntracked
) {
1536 flow_clear_conntrack(&frozen_flow
);
1539 frozen_metadata_to_flow(&upcall
->ofproto
->up
, &state
->metadata
,
1541 flow_get_metadata(&frozen_flow
, &am
->pin
.up
.base
.flow_metadata
);
1543 ofproto_dpif_send_async_msg(upcall
->ofproto
, am
);
1555 handle_upcalls(struct udpif
*udpif
, struct upcall
*upcalls
,
1558 struct dpif_op
*opsp
[UPCALL_MAX_BATCH
* 2];
1559 struct ukey_op ops
[UPCALL_MAX_BATCH
* 2];
1560 size_t n_ops
, n_opsp
, i
;
1562 /* Handle the packets individually in order of arrival.
1564 * - For SLOW_CFM, SLOW_LACP, SLOW_STP, SLOW_BFD, and SLOW_LLDP,
1565 * translation is what processes received packets for these
1568 * - For SLOW_ACTION, translation executes the actions directly.
1570 * The loop fills 'ops' with an array of operations to execute in the
1573 for (i
= 0; i
< n_upcalls
; i
++) {
1574 struct upcall
*upcall
= &upcalls
[i
];
1575 const struct dp_packet
*packet
= upcall
->packet
;
1578 if (should_install_flow(udpif
, upcall
)) {
1579 struct udpif_key
*ukey
= upcall
->ukey
;
1581 if (ukey_install(udpif
, ukey
)) {
1582 upcall
->ukey_persists
= true;
1583 put_op_init(&ops
[n_ops
++], ukey
, DPIF_FP_CREATE
);
1587 if (upcall
->odp_actions
.size
) {
1590 op
->dop
.type
= DPIF_OP_EXECUTE
;
1591 op
->dop
.execute
.packet
= CONST_CAST(struct dp_packet
*, packet
);
1592 op
->dop
.execute
.flow
= upcall
->flow
;
1593 odp_key_to_dp_packet(upcall
->key
, upcall
->key_len
,
1594 op
->dop
.execute
.packet
);
1595 op
->dop
.execute
.actions
= upcall
->odp_actions
.data
;
1596 op
->dop
.execute
.actions_len
= upcall
->odp_actions
.size
;
1597 op
->dop
.execute
.needs_help
= (upcall
->xout
.slow
& SLOW_ACTION
) != 0;
1598 op
->dop
.execute
.probe
= false;
1599 op
->dop
.execute
.mtu
= upcall
->mru
;
1600 op
->dop
.execute
.hash
= upcall
->hash
;
1604 /* Execute batch. */
1606 for (i
= 0; i
< n_ops
; i
++) {
1607 opsp
[n_opsp
++] = &ops
[i
].dop
;
1609 dpif_operate(udpif
->dpif
, opsp
, n_opsp
, DPIF_OFFLOAD_AUTO
);
1610 for (i
= 0; i
< n_ops
; i
++) {
1611 struct udpif_key
*ukey
= ops
[i
].ukey
;
1614 ovs_mutex_lock(&ukey
->mutex
);
1615 if (ops
[i
].dop
.error
) {
1616 transition_ukey(ukey
, UKEY_EVICTED
);
1617 } else if (ukey
->state
< UKEY_OPERATIONAL
) {
1618 transition_ukey(ukey
, UKEY_OPERATIONAL
);
1620 ovs_mutex_unlock(&ukey
->mutex
);
1626 get_ukey_hash(const ovs_u128
*ufid
, const unsigned pmd_id
)
1628 return hash_2words(ufid
->u32
[0], pmd_id
);
1631 static struct udpif_key
*
1632 ukey_lookup(struct udpif
*udpif
, const ovs_u128
*ufid
, const unsigned pmd_id
)
1634 struct udpif_key
*ukey
;
1635 int idx
= get_ukey_hash(ufid
, pmd_id
) % N_UMAPS
;
1636 struct cmap
*cmap
= &udpif
->ukeys
[idx
].cmap
;
1638 CMAP_FOR_EACH_WITH_HASH (ukey
, cmap_node
,
1639 get_ukey_hash(ufid
, pmd_id
), cmap
) {
1640 if (ovs_u128_equals(ukey
->ufid
, *ufid
)) {
1647 /* Provides safe lockless access of RCU protected 'ukey->actions'. Callers may
1648 * alternatively access the field directly if they take 'ukey->mutex'. */
1650 ukey_get_actions(struct udpif_key
*ukey
, const struct nlattr
**actions
, size_t *size
)
1652 const struct ofpbuf
*buf
= ovsrcu_get(struct ofpbuf
*, &ukey
->actions
);
1653 *actions
= buf
->data
;
1658 ukey_set_actions(struct udpif_key
*ukey
, const struct ofpbuf
*actions
)
1660 struct ofpbuf
*old_actions
= ovsrcu_get_protected(struct ofpbuf
*,
1664 ovsrcu_postpone(ofpbuf_delete
, old_actions
);
1667 ovsrcu_set(&ukey
->actions
, ofpbuf_clone(actions
));
1670 static struct udpif_key
*
1671 ukey_create__(const struct nlattr
*key
, size_t key_len
,
1672 const struct nlattr
*mask
, size_t mask_len
,
1673 bool ufid_present
, const ovs_u128
*ufid
,
1674 const unsigned pmd_id
, const struct ofpbuf
*actions
,
1675 uint64_t reval_seq
, long long int used
,
1676 uint32_t key_recirc_id
, struct xlate_out
*xout
)
1677 OVS_NO_THREAD_SAFETY_ANALYSIS
1679 struct udpif_key
*ukey
= xmalloc(sizeof *ukey
);
1681 memcpy(&ukey
->keybuf
, key
, key_len
);
1682 ukey
->key
= &ukey
->keybuf
.nla
;
1683 ukey
->key_len
= key_len
;
1684 memcpy(&ukey
->maskbuf
, mask
, mask_len
);
1685 ukey
->mask
= &ukey
->maskbuf
.nla
;
1686 ukey
->mask_len
= mask_len
;
1687 ukey
->ufid_present
= ufid_present
;
1689 ukey
->pmd_id
= pmd_id
;
1690 ukey
->hash
= get_ukey_hash(&ukey
->ufid
, pmd_id
);
1692 ovsrcu_init(&ukey
->actions
, NULL
);
1693 ukey_set_actions(ukey
, actions
);
1695 ovs_mutex_init(&ukey
->mutex
);
1696 ukey
->dump_seq
= 0; /* Not yet dumped */
1697 ukey
->reval_seq
= reval_seq
;
1698 ukey
->state
= UKEY_CREATED
;
1699 ukey
->state_thread
= ovsthread_id_self();
1700 ukey
->state_where
= OVS_SOURCE_LOCATOR
;
1701 ukey
->created
= ukey
->flow_time
= time_msec();
1702 memset(&ukey
->stats
, 0, sizeof ukey
->stats
);
1703 ukey
->stats
.used
= used
;
1704 ukey
->xcache
= NULL
;
1706 ukey
->offloaded
= false;
1707 ukey
->in_netdev
= NULL
;
1708 ukey
->flow_packets
= ukey
->flow_backlog_packets
= 0;
1710 ukey
->key_recirc_id
= key_recirc_id
;
1711 recirc_refs_init(&ukey
->recircs
);
1713 /* Take ownership of the action recirc id references. */
1714 recirc_refs_swap(&ukey
->recircs
, &xout
->recircs
);
1720 static struct udpif_key
*
1721 ukey_create_from_upcall(struct upcall
*upcall
, struct flow_wildcards
*wc
)
1723 struct odputil_keybuf keystub
, maskstub
;
1724 struct ofpbuf keybuf
, maskbuf
;
1726 struct odp_flow_key_parms odp_parms
= {
1727 .flow
= upcall
->flow
,
1728 .mask
= wc
? &wc
->masks
: NULL
,
1731 odp_parms
.support
= upcall
->ofproto
->backer
->rt_support
.odp
;
1732 if (upcall
->key_len
) {
1733 ofpbuf_use_const(&keybuf
, upcall
->key
, upcall
->key_len
);
1735 /* dpif-netdev doesn't provide a netlink-formatted flow key in the
1736 * upcall, so convert the upcall's flow here. */
1737 ofpbuf_use_stack(&keybuf
, &keystub
, sizeof keystub
);
1738 odp_flow_key_from_flow(&odp_parms
, &keybuf
);
1741 atomic_read_relaxed(&enable_megaflows
, &megaflow
);
1742 ofpbuf_use_stack(&maskbuf
, &maskstub
, sizeof maskstub
);
1743 if (megaflow
&& wc
) {
1744 odp_parms
.key_buf
= &keybuf
;
1745 odp_flow_key_from_mask(&odp_parms
, &maskbuf
);
1748 return ukey_create__(keybuf
.data
, keybuf
.size
, maskbuf
.data
, maskbuf
.size
,
1749 true, upcall
->ufid
, upcall
->pmd_id
,
1750 &upcall
->put_actions
, upcall
->reval_seq
, 0,
1751 upcall
->have_recirc_ref
? upcall
->recirc
->id
: 0,
1756 ukey_create_from_dpif_flow(const struct udpif
*udpif
,
1757 const struct dpif_flow
*flow
,
1758 struct udpif_key
**ukey
)
1760 struct dpif_flow full_flow
;
1761 struct ofpbuf actions
;
1763 uint64_t stub
[DPIF_FLOW_BUFSIZE
/ 8];
1764 const struct nlattr
*a
;
1767 if (!flow
->key_len
|| !flow
->actions_len
) {
1771 /* If the key or actions were not provided by the datapath, fetch the
1773 ofpbuf_use_stack(&buf
, &stub
, sizeof stub
);
1774 err
= dpif_flow_get(udpif
->dpif
, flow
->key
, flow
->key_len
,
1775 flow
->ufid_present
? &flow
->ufid
: NULL
,
1776 flow
->pmd_id
, &buf
, &full_flow
);
1783 /* Check the flow actions for recirculation action. As recirculation
1784 * relies on OVS userspace internal state, we need to delete all old
1785 * datapath flows with either a non-zero recirc_id in the key, or any
1786 * recirculation actions upon OVS restart. */
1787 NL_ATTR_FOR_EACH (a
, left
, flow
->key
, flow
->key_len
) {
1788 if (nl_attr_type(a
) == OVS_KEY_ATTR_RECIRC_ID
1789 && nl_attr_get_u32(a
) != 0) {
1793 NL_ATTR_FOR_EACH (a
, left
, flow
->actions
, flow
->actions_len
) {
1794 if (nl_attr_type(a
) == OVS_ACTION_ATTR_RECIRC
) {
1799 reval_seq
= seq_read(udpif
->reval_seq
) - 1; /* Ensure revalidation. */
1800 ofpbuf_use_const(&actions
, flow
->actions
, flow
->actions_len
);
1801 *ukey
= ukey_create__(flow
->key
, flow
->key_len
,
1802 flow
->mask
, flow
->mask_len
, flow
->ufid_present
,
1803 &flow
->ufid
, flow
->pmd_id
, &actions
,
1804 reval_seq
, flow
->stats
.used
, 0, NULL
);
1810 try_ukey_replace(struct umap
*umap
, struct udpif_key
*old_ukey
,
1811 struct udpif_key
*new_ukey
)
1812 OVS_REQUIRES(umap
->mutex
)
1813 OVS_TRY_LOCK(true, new_ukey
->mutex
)
1815 bool replaced
= false;
1817 if (!ovs_mutex_trylock(&old_ukey
->mutex
)) {
1818 if (old_ukey
->state
== UKEY_EVICTED
) {
1819 /* The flow was deleted during the current revalidator dump,
1820 * but its ukey won't be fully cleaned up until the sweep phase.
1821 * In the mean time, we are receiving upcalls for this traffic.
1822 * Expedite the (new) flow install by replacing the ukey. */
1823 ovs_mutex_lock(&new_ukey
->mutex
);
1824 cmap_replace(&umap
->cmap
, &old_ukey
->cmap_node
,
1825 &new_ukey
->cmap_node
, new_ukey
->hash
);
1826 ovsrcu_postpone(ukey_delete__
, old_ukey
);
1827 transition_ukey(old_ukey
, UKEY_DELETED
);
1828 transition_ukey(new_ukey
, UKEY_VISIBLE
);
1831 ovs_mutex_unlock(&old_ukey
->mutex
);
1835 COVERAGE_INC(upcall_ukey_replace
);
1837 COVERAGE_INC(handler_duplicate_upcall
);
1842 /* Attempts to insert a ukey into the shared ukey maps.
1844 * On success, returns true, installs the ukey and returns it in a locked
1845 * state. Otherwise, returns false. */
1847 ukey_install__(struct udpif
*udpif
, struct udpif_key
*new_ukey
)
1848 OVS_TRY_LOCK(true, new_ukey
->mutex
)
1851 struct udpif_key
*old_ukey
;
1853 bool locked
= false;
1855 idx
= new_ukey
->hash
% N_UMAPS
;
1856 umap
= &udpif
->ukeys
[idx
];
1857 ovs_mutex_lock(&umap
->mutex
);
1858 old_ukey
= ukey_lookup(udpif
, &new_ukey
->ufid
, new_ukey
->pmd_id
);
1860 /* Uncommon case: A ukey is already installed with the same UFID. */
1861 if (old_ukey
->key_len
== new_ukey
->key_len
1862 && !memcmp(old_ukey
->key
, new_ukey
->key
, new_ukey
->key_len
)) {
1863 locked
= try_ukey_replace(umap
, old_ukey
, new_ukey
);
1865 struct ds ds
= DS_EMPTY_INITIALIZER
;
1867 odp_format_ufid(&old_ukey
->ufid
, &ds
);
1868 ds_put_cstr(&ds
, " ");
1869 odp_flow_key_format(old_ukey
->key
, old_ukey
->key_len
, &ds
);
1870 ds_put_cstr(&ds
, "\n");
1871 odp_format_ufid(&new_ukey
->ufid
, &ds
);
1872 ds_put_cstr(&ds
, " ");
1873 odp_flow_key_format(new_ukey
->key
, new_ukey
->key_len
, &ds
);
1875 VLOG_WARN_RL(&rl
, "Conflicting ukey for flows:\n%s", ds_cstr(&ds
));
1879 ovs_mutex_lock(&new_ukey
->mutex
);
1880 cmap_insert(&umap
->cmap
, &new_ukey
->cmap_node
, new_ukey
->hash
);
1881 transition_ukey(new_ukey
, UKEY_VISIBLE
);
1884 ovs_mutex_unlock(&umap
->mutex
);
1890 transition_ukey_at(struct udpif_key
*ukey
, enum ukey_state dst
,
1892 OVS_REQUIRES(ukey
->mutex
)
1894 if (dst
< ukey
->state
) {
1895 VLOG_ABORT("Invalid ukey transition %d->%d (last transitioned from "
1896 "thread %u at %s)", ukey
->state
, dst
, ukey
->state_thread
,
1899 if (ukey
->state
== dst
&& dst
== UKEY_OPERATIONAL
) {
1903 /* Valid state transitions:
1904 * UKEY_CREATED -> UKEY_VISIBLE
1905 * Ukey is now visible in the umap.
1906 * UKEY_VISIBLE -> UKEY_OPERATIONAL
1907 * A handler has installed the flow, and the flow is in the datapath.
1908 * UKEY_VISIBLE -> UKEY_EVICTING
1909 * A handler installs the flow, then revalidator sweeps the ukey before
1910 * the flow is dumped. Most likely the flow was installed; start trying
1912 * UKEY_VISIBLE -> UKEY_EVICTED
1913 * A handler attempts to install the flow, but the datapath rejects it.
1914 * Consider that the datapath has already destroyed it.
1915 * UKEY_OPERATIONAL -> UKEY_EVICTING
1916 * A revalidator decides to evict the datapath flow.
1917 * UKEY_EVICTING -> UKEY_EVICTED
1918 * A revalidator has evicted the datapath flow.
1919 * UKEY_EVICTED -> UKEY_DELETED
1920 * A revalidator has removed the ukey from the umap and is deleting it.
1922 if (ukey
->state
== dst
- 1 || (ukey
->state
== UKEY_VISIBLE
&&
1923 dst
< UKEY_DELETED
)) {
1926 struct ds ds
= DS_EMPTY_INITIALIZER
;
1928 odp_format_ufid(&ukey
->ufid
, &ds
);
1929 VLOG_WARN_RL(&rl
, "Invalid state transition for ukey %s: %d -> %d",
1930 ds_cstr(&ds
), ukey
->state
, dst
);
1933 ukey
->state_thread
= ovsthread_id_self();
1934 ukey
->state_where
= where
;
1938 ukey_install(struct udpif
*udpif
, struct udpif_key
*ukey
)
1942 installed
= ukey_install__(udpif
, ukey
);
1944 ovs_mutex_unlock(&ukey
->mutex
);
1950 /* Searches for a ukey in 'udpif->ukeys' that matches 'flow' and attempts to
1951 * lock the ukey. If the ukey does not exist, create it.
1953 * Returns 0 on success, setting *result to the matching ukey and returning it
1954 * in a locked state. Otherwise, returns an errno and clears *result. EBUSY
1955 * indicates that another thread is handling this flow. Other errors indicate
1956 * an unexpected condition creating a new ukey.
1958 * *error is an output parameter provided to appease the threadsafety analyser,
1959 * and its value matches the return value. */
1961 ukey_acquire(struct udpif
*udpif
, const struct dpif_flow
*flow
,
1962 struct udpif_key
**result
, int *error
)
1963 OVS_TRY_LOCK(0, (*result
)->mutex
)
1965 struct udpif_key
*ukey
;
1968 ukey
= ukey_lookup(udpif
, &flow
->ufid
, flow
->pmd_id
);
1970 retval
= ovs_mutex_trylock(&ukey
->mutex
);
1972 /* Usually we try to avoid installing flows from revalidator threads,
1973 * because locking on a umap may cause handler threads to block.
1974 * However there are certain cases, like when ovs-vswitchd is
1975 * restarted, where it is desirable to handle flows that exist in the
1976 * datapath gracefully (ie, don't just clear the datapath). */
1979 retval
= ukey_create_from_dpif_flow(udpif
, flow
, &ukey
);
1983 install
= ukey_install__(udpif
, ukey
);
1987 ukey_delete__(ukey
);
2003 ukey_delete__(struct udpif_key
*ukey
)
2004 OVS_NO_THREAD_SAFETY_ANALYSIS
2007 if (ukey
->key_recirc_id
) {
2008 recirc_free_id(ukey
->key_recirc_id
);
2010 recirc_refs_unref(&ukey
->recircs
);
2011 xlate_cache_delete(ukey
->xcache
);
2012 ofpbuf_delete(ovsrcu_get(struct ofpbuf
*, &ukey
->actions
));
2013 ovs_mutex_destroy(&ukey
->mutex
);
2019 ukey_delete(struct umap
*umap
, struct udpif_key
*ukey
)
2020 OVS_REQUIRES(umap
->mutex
)
2022 ovs_mutex_lock(&ukey
->mutex
);
2023 if (ukey
->state
< UKEY_DELETED
) {
2024 cmap_remove(&umap
->cmap
, &ukey
->cmap_node
, ukey
->hash
);
2025 ovsrcu_postpone(ukey_delete__
, ukey
);
2026 transition_ukey(ukey
, UKEY_DELETED
);
2028 ovs_mutex_unlock(&ukey
->mutex
);
2032 should_revalidate(const struct udpif
*udpif
, uint64_t packets
,
2035 long long int metric
, now
, duration
;
2038 /* Always revalidate the first time a flow is dumped. */
2042 if (udpif
->dump_duration
< ofproto_max_revalidator
/ 2) {
2043 /* We are likely to handle full revalidation for the flows. */
2047 /* Calculate the mean time between seeing these packets. If this
2048 * exceeds the threshold, then delete the flow rather than performing
2049 * costly revalidation for flows that aren't being hit frequently.
2051 * This is targeted at situations where the dump_duration is high (~1s),
2052 * and revalidation is triggered by a call to udpif_revalidate(). In
2053 * these situations, revalidation of all flows causes fluctuations in the
2054 * flow_limit due to the interaction with the dump_duration and max_idle.
2055 * This tends to result in deletion of low-throughput flows anyway, so
2056 * skip the revalidation and just delete those flows. */
2057 packets
= MAX(packets
, 1);
2058 now
= MAX(used
, time_msec());
2059 duration
= now
- used
;
2060 metric
= duration
/ packets
;
2062 if (metric
< 1000 / ofproto_min_revalidate_pps
) {
2063 /* The flow is receiving more than min-revalidate-pps, so keep it. */
2069 struct reval_context
{
2070 /* Optional output parameters */
2071 struct flow_wildcards
*wc
;
2072 struct ofpbuf
*odp_actions
;
2073 struct netflow
**netflow
;
2074 struct xlate_cache
*xcache
;
2076 /* Required output parameters */
2077 struct xlate_out xout
;
2081 /* Translates 'key' into a flow, populating 'ctx' as it goes along.
2083 * Returns 0 on success, otherwise a positive errno value.
2085 * The caller is responsible for uninitializing ctx->xout on success.
2088 xlate_key(struct udpif
*udpif
, const struct nlattr
*key
, unsigned int len
,
2089 const struct dpif_flow_stats
*push
, struct reval_context
*ctx
)
2091 struct ofproto_dpif
*ofproto
;
2092 ofp_port_t ofp_in_port
;
2093 enum odp_key_fitness fitness
;
2094 struct xlate_in xin
;
2097 fitness
= odp_flow_key_to_flow(key
, len
, &ctx
->flow
, NULL
);
2098 if (fitness
== ODP_FIT_ERROR
) {
2102 error
= xlate_lookup(udpif
->backer
, &ctx
->flow
, &ofproto
, NULL
, NULL
,
2103 ctx
->netflow
, &ofp_in_port
);
2108 xlate_in_init(&xin
, ofproto
, ofproto_dpif_get_tables_version(ofproto
),
2109 &ctx
->flow
, ofp_in_port
, NULL
, push
->tcp_flags
,
2110 NULL
, ctx
->wc
, ctx
->odp_actions
);
2111 if (push
->n_packets
) {
2112 xin
.resubmit_stats
= push
;
2113 xin
.allow_side_effects
= true;
2115 xin
.xcache
= ctx
->xcache
;
2116 xlate_actions(&xin
, &ctx
->xout
);
2117 if (fitness
== ODP_FIT_TOO_LITTLE
) {
2118 ctx
->xout
.slow
|= SLOW_MATCH
;
2125 xlate_ukey(struct udpif
*udpif
, const struct udpif_key
*ukey
,
2126 uint16_t tcp_flags
, struct reval_context
*ctx
)
2128 struct dpif_flow_stats push
= {
2129 .tcp_flags
= tcp_flags
,
2131 return xlate_key(udpif
, ukey
->key
, ukey
->key_len
, &push
, ctx
);
2135 populate_xcache(struct udpif
*udpif
, struct udpif_key
*ukey
,
2137 OVS_REQUIRES(ukey
->mutex
)
2139 struct reval_context ctx
= {
2140 .odp_actions
= NULL
,
2146 ovs_assert(!ukey
->xcache
);
2147 ukey
->xcache
= ctx
.xcache
= xlate_cache_new();
2148 error
= xlate_ukey(udpif
, ukey
, tcp_flags
, &ctx
);
2152 xlate_out_uninit(&ctx
.xout
);
2157 static enum reval_result
2158 revalidate_ukey__(struct udpif
*udpif
, const struct udpif_key
*ukey
,
2159 uint16_t tcp_flags
, struct ofpbuf
*odp_actions
,
2160 struct recirc_refs
*recircs
, struct xlate_cache
*xcache
)
2162 struct xlate_out
*xoutp
;
2163 struct netflow
*netflow
;
2164 struct flow_wildcards dp_mask
, wc
;
2165 enum reval_result result
;
2166 struct reval_context ctx
= {
2167 .odp_actions
= odp_actions
,
2168 .netflow
= &netflow
,
2173 result
= UKEY_DELETE
;
2177 if (xlate_ukey(udpif
, ukey
, tcp_flags
, &ctx
)) {
2182 if (xoutp
->avoid_caching
) {
2187 struct ofproto_dpif
*ofproto
;
2188 ofp_port_t ofp_in_port
;
2190 ofproto
= xlate_lookup_ofproto(udpif
->backer
, &ctx
.flow
, &ofp_in_port
,
2193 ofpbuf_clear(odp_actions
);
2199 compose_slow_path(udpif
, xoutp
, ctx
.flow
.in_port
.odp_port
,
2200 ofp_in_port
, odp_actions
,
2201 ofproto
->up
.slowpath_meter_id
, &ofproto
->uuid
);
2204 if (odp_flow_key_to_mask(ukey
->mask
, ukey
->mask_len
, &dp_mask
, &ctx
.flow
,
2210 /* Do not modify if any bit is wildcarded by the installed datapath flow,
2211 * but not the newly revalidated wildcard mask (wc), i.e., if revalidation
2212 * tells that the datapath flow is now too generic and must be narrowed
2213 * down. Note that we do not know if the datapath has ignored any of the
2214 * wildcarded bits, so we may be overly conservative here. */
2215 if (flow_wildcards_has_extra(&dp_mask
, ctx
.wc
)) {
2219 if (!ofpbuf_equal(odp_actions
,
2220 ovsrcu_get(struct ofpbuf
*, &ukey
->actions
))) {
2221 /* The datapath mask was OK, but the actions seem to have changed.
2222 * Let's modify it in place. */
2223 result
= UKEY_MODIFY
;
2224 /* Transfer recirc action ID references to the caller. */
2225 recirc_refs_swap(recircs
, &xoutp
->recircs
);
2232 if (netflow
&& result
== UKEY_DELETE
) {
2233 netflow_flow_clear(netflow
, &ctx
.flow
);
2235 xlate_out_uninit(xoutp
);
2239 /* Verifies that the datapath actions of 'ukey' are still correct, and pushes
2242 * Returns a recommended action for 'ukey', options include:
2243 * UKEY_DELETE The ukey should be deleted.
2244 * UKEY_KEEP The ukey is fine as is.
2245 * UKEY_MODIFY The ukey's actions should be changed but is otherwise
2246 * fine. Callers should change the actions to those found
2247 * in the caller supplied 'odp_actions' buffer. The
2248 * recirculation references can be found in 'recircs' and
2249 * must be handled by the caller.
2251 * If the result is UKEY_MODIFY, then references to all recirc_ids used by the
2252 * new flow will be held within 'recircs' (which may be none).
2254 * The caller is responsible for both initializing 'recircs' prior this call,
2255 * and ensuring any references are eventually freed.
2257 static enum reval_result
2258 revalidate_ukey(struct udpif
*udpif
, struct udpif_key
*ukey
,
2259 const struct dpif_flow_stats
*stats
,
2260 struct ofpbuf
*odp_actions
, uint64_t reval_seq
,
2261 struct recirc_refs
*recircs
, bool offloaded
)
2262 OVS_REQUIRES(ukey
->mutex
)
2264 bool need_revalidate
= ukey
->reval_seq
!= reval_seq
;
2265 enum reval_result result
= UKEY_DELETE
;
2266 struct dpif_flow_stats push
;
2268 ofpbuf_clear(odp_actions
);
2270 push
.used
= stats
->used
;
2271 push
.tcp_flags
= stats
->tcp_flags
;
2272 push
.n_packets
= (stats
->n_packets
> ukey
->stats
.n_packets
2273 ? stats
->n_packets
- ukey
->stats
.n_packets
2275 push
.n_bytes
= (stats
->n_bytes
> ukey
->stats
.n_bytes
2276 ? stats
->n_bytes
- ukey
->stats
.n_bytes
2279 if (need_revalidate
) {
2280 if (should_revalidate(udpif
, push
.n_packets
, ukey
->stats
.used
)) {
2281 if (!ukey
->xcache
) {
2282 ukey
->xcache
= xlate_cache_new();
2284 xlate_cache_clear(ukey
->xcache
);
2286 result
= revalidate_ukey__(udpif
, ukey
, push
.tcp_flags
,
2287 odp_actions
, recircs
, ukey
->xcache
);
2288 } /* else delete; too expensive to revalidate */
2289 } else if (!push
.n_packets
|| ukey
->xcache
2290 || !populate_xcache(udpif
, ukey
, push
.tcp_flags
)) {
2294 /* Stats for deleted flows will be attributed upon flow deletion. Skip. */
2295 if (result
!= UKEY_DELETE
) {
2296 xlate_push_stats(ukey
->xcache
, &push
, offloaded
);
2297 ukey
->stats
= *stats
;
2298 ukey
->reval_seq
= reval_seq
;
2305 delete_op_init__(struct udpif
*udpif
, struct ukey_op
*op
,
2306 const struct dpif_flow
*flow
)
2309 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
2310 op
->dop
.flow_del
.key
= flow
->key
;
2311 op
->dop
.flow_del
.key_len
= flow
->key_len
;
2312 op
->dop
.flow_del
.ufid
= flow
->ufid_present
? &flow
->ufid
: NULL
;
2313 op
->dop
.flow_del
.pmd_id
= flow
->pmd_id
;
2314 op
->dop
.flow_del
.stats
= &op
->stats
;
2315 op
->dop
.flow_del
.terse
= udpif_use_ufid(udpif
);
2319 delete_op_init(struct udpif
*udpif
, struct ukey_op
*op
, struct udpif_key
*ukey
)
2322 op
->dop
.type
= DPIF_OP_FLOW_DEL
;
2323 op
->dop
.flow_del
.key
= ukey
->key
;
2324 op
->dop
.flow_del
.key_len
= ukey
->key_len
;
2325 op
->dop
.flow_del
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
2326 op
->dop
.flow_del
.pmd_id
= ukey
->pmd_id
;
2327 op
->dop
.flow_del
.stats
= &op
->stats
;
2328 op
->dop
.flow_del
.terse
= udpif_use_ufid(udpif
);
2332 put_op_init(struct ukey_op
*op
, struct udpif_key
*ukey
,
2333 enum dpif_flow_put_flags flags
)
2336 op
->dop
.type
= DPIF_OP_FLOW_PUT
;
2337 op
->dop
.flow_put
.flags
= flags
;
2338 op
->dop
.flow_put
.key
= ukey
->key
;
2339 op
->dop
.flow_put
.key_len
= ukey
->key_len
;
2340 op
->dop
.flow_put
.mask
= ukey
->mask
;
2341 op
->dop
.flow_put
.mask_len
= ukey
->mask_len
;
2342 op
->dop
.flow_put
.ufid
= ukey
->ufid_present
? &ukey
->ufid
: NULL
;
2343 op
->dop
.flow_put
.pmd_id
= ukey
->pmd_id
;
2344 op
->dop
.flow_put
.stats
= NULL
;
2345 ukey_get_actions(ukey
, &op
->dop
.flow_put
.actions
,
2346 &op
->dop
.flow_put
.actions_len
);
2349 /* Executes datapath operations 'ops' and attributes stats retrieved from the
2350 * datapath as part of those operations. */
2352 push_dp_ops(struct udpif
*udpif
, struct ukey_op
*ops
, size_t n_ops
)
2354 struct dpif_op
*opsp
[REVALIDATE_MAX_BATCH
];
2357 ovs_assert(n_ops
<= REVALIDATE_MAX_BATCH
);
2358 for (i
= 0; i
< n_ops
; i
++) {
2359 opsp
[i
] = &ops
[i
].dop
;
2361 dpif_operate(udpif
->dpif
, opsp
, n_ops
, DPIF_OFFLOAD_AUTO
);
2363 for (i
= 0; i
< n_ops
; i
++) {
2364 struct ukey_op
*op
= &ops
[i
];
2365 struct dpif_flow_stats
*push
, *stats
, push_buf
;
2367 stats
= op
->dop
.flow_del
.stats
;
2370 if (op
->dop
.type
!= DPIF_OP_FLOW_DEL
) {
2371 /* Only deleted flows need their stats pushed. */
2375 if (op
->dop
.error
) {
2376 /* flow_del error, 'stats' is unusable. */
2378 ovs_mutex_lock(&op
->ukey
->mutex
);
2379 transition_ukey(op
->ukey
, UKEY_EVICTED
);
2380 ovs_mutex_unlock(&op
->ukey
->mutex
);
2386 ovs_mutex_lock(&op
->ukey
->mutex
);
2387 transition_ukey(op
->ukey
, UKEY_EVICTED
);
2388 push
->used
= MAX(stats
->used
, op
->ukey
->stats
.used
);
2389 push
->tcp_flags
= stats
->tcp_flags
| op
->ukey
->stats
.tcp_flags
;
2390 push
->n_packets
= stats
->n_packets
- op
->ukey
->stats
.n_packets
;
2391 push
->n_bytes
= stats
->n_bytes
- op
->ukey
->stats
.n_bytes
;
2392 ovs_mutex_unlock(&op
->ukey
->mutex
);
2397 if (push
->n_packets
|| netflow_exists()) {
2398 const struct nlattr
*key
= op
->dop
.flow_del
.key
;
2399 size_t key_len
= op
->dop
.flow_del
.key_len
;
2400 struct netflow
*netflow
;
2401 struct reval_context ctx
= {
2402 .netflow
= &netflow
,
2407 ovs_mutex_lock(&op
->ukey
->mutex
);
2408 if (op
->ukey
->xcache
) {
2409 xlate_push_stats(op
->ukey
->xcache
, push
, false);
2410 ovs_mutex_unlock(&op
->ukey
->mutex
);
2413 ovs_mutex_unlock(&op
->ukey
->mutex
);
2414 key
= op
->ukey
->key
;
2415 key_len
= op
->ukey
->key_len
;
2418 error
= xlate_key(udpif
, key
, key_len
, push
, &ctx
);
2420 static struct vlog_rate_limit rll
= VLOG_RATE_LIMIT_INIT(1, 5);
2421 VLOG_WARN_RL(&rll
, "xlate_key failed (%s)!",
2422 ovs_strerror(error
));
2424 xlate_out_uninit(&ctx
.xout
);
2426 netflow_flow_clear(netflow
, &ctx
.flow
);
2433 /* Executes datapath operations 'ops', attributes stats retrieved from the
2434 * datapath, and deletes ukeys corresponding to deleted flows. */
2436 push_ukey_ops(struct udpif
*udpif
, struct umap
*umap
,
2437 struct ukey_op
*ops
, size_t n_ops
)
2441 push_dp_ops(udpif
, ops
, n_ops
);
2442 ovs_mutex_lock(&umap
->mutex
);
2443 for (i
= 0; i
< n_ops
; i
++) {
2444 if (ops
[i
].dop
.type
== DPIF_OP_FLOW_DEL
) {
2445 ukey_delete(umap
, ops
[i
].ukey
);
2448 ovs_mutex_unlock(&umap
->mutex
);
2452 log_unexpected_flow(const struct dpif_flow
*flow
, int error
)
2454 struct ds ds
= DS_EMPTY_INITIALIZER
;
2456 ds_put_format(&ds
, "Failed to acquire udpif_key corresponding to "
2457 "unexpected flow (%s): ", ovs_strerror(error
));
2458 odp_format_ufid(&flow
->ufid
, &ds
);
2460 static struct vlog_rate_limit rll
= VLOG_RATE_LIMIT_INIT(10, 60);
2461 VLOG_WARN_RL(&rll
, "%s", ds_cstr(&ds
));
2467 reval_op_init(struct ukey_op
*op
, enum reval_result result
,
2468 struct udpif
*udpif
, struct udpif_key
*ukey
,
2469 struct recirc_refs
*recircs
, struct ofpbuf
*odp_actions
)
2470 OVS_REQUIRES(ukey
->mutex
)
2472 if (result
== UKEY_DELETE
) {
2473 delete_op_init(udpif
, op
, ukey
);
2474 transition_ukey(ukey
, UKEY_EVICTING
);
2475 } else if (result
== UKEY_MODIFY
) {
2476 /* Store the new recircs. */
2477 recirc_refs_swap(&ukey
->recircs
, recircs
);
2478 /* Release old recircs. */
2479 recirc_refs_unref(recircs
);
2480 /* ukey->key_recirc_id remains, as the key is the same as before. */
2482 ukey_set_actions(ukey
, odp_actions
);
2483 put_op_init(op
, ukey
, DPIF_FP_MODIFY
);
2488 ukey_netdev_unref(struct udpif_key
*ukey
)
2490 if (!ukey
->in_netdev
) {
2493 netdev_close(ukey
->in_netdev
);
2494 ukey
->in_netdev
= NULL
;
2498 * Given a udpif_key, get its input port (netdev) by parsing the flow keys
2499 * and actions. The flow may not contain flow attributes if it is a terse
2500 * dump; read its attributes from the ukey and then parse the flow to get
2501 * the port info. Save them in udpif_key.
2504 ukey_to_flow_netdev(struct udpif
*udpif
, struct udpif_key
*ukey
)
2506 const char *dpif_type_str
= dpif_normalize_type(dpif_type(udpif
->dpif
));
2507 const struct nlattr
*k
;
2510 /* Remove existing references to netdev */
2511 ukey_netdev_unref(ukey
);
2513 /* Find the input port and get a reference to its netdev */
2514 NL_ATTR_FOR_EACH (k
, left
, ukey
->key
, ukey
->key_len
) {
2515 enum ovs_key_attr type
= nl_attr_type(k
);
2517 if (type
== OVS_KEY_ATTR_IN_PORT
) {
2518 ukey
->in_netdev
= netdev_ports_get(nl_attr_get_odp_port(k
),
2520 } else if (type
== OVS_KEY_ATTR_TUNNEL
) {
2521 struct flow_tnl tnl
;
2522 enum odp_key_fitness res
;
2524 if (ukey
->in_netdev
) {
2525 netdev_close(ukey
->in_netdev
);
2526 ukey
->in_netdev
= NULL
;
2528 res
= odp_tun_key_from_attr(k
, &tnl
, NULL
);
2529 if (res
!= ODP_FIT_ERROR
) {
2530 ukey
->in_netdev
= flow_get_tunnel_netdev(&tnl
);
2538 udpif_flow_packet_delta(struct udpif_key
*ukey
, const struct dpif_flow
*f
)
2540 return f
->stats
.n_packets
+ ukey
->flow_backlog_packets
-
2544 static long long int
2545 udpif_flow_time_delta(struct udpif
*udpif
, struct udpif_key
*ukey
)
2547 return (udpif
->dpif
->current_ms
- ukey
->flow_time
) / 1000;
2551 * Save backlog packet count while switching modes
2552 * between offloaded and kernel datapaths.
2555 udpif_set_ukey_backlog_packets(struct udpif_key
*ukey
)
2557 ukey
->flow_backlog_packets
= ukey
->flow_packets
;
2560 /* Gather pps-rate for the given dpif_flow and save it in its ukey */
2562 udpif_update_flow_pps(struct udpif
*udpif
, struct udpif_key
*ukey
,
2563 const struct dpif_flow
*f
)
2567 /* Update pps-rate only when we are close to rebalance interval */
2568 if (udpif
->dpif
->current_ms
- ukey
->flow_time
< OFFL_REBAL_INTVL_MSEC
) {
2572 ukey
->offloaded
= f
->attrs
.offloaded
;
2573 pps
= udpif_flow_packet_delta(ukey
, f
) /
2574 udpif_flow_time_delta(udpif
, ukey
);
2575 ukey
->flow_pps_rate
= pps
;
2576 ukey
->flow_packets
= ukey
->flow_backlog_packets
+ f
->stats
.n_packets
;
2577 ukey
->flow_time
= udpif
->dpif
->current_ms
;
2580 static long long int
2581 udpif_update_used(struct udpif
*udpif
, struct udpif_key
*ukey
,
2582 struct dpif_flow_stats
*stats
)
2583 OVS_REQUIRES(ukey
->mutex
)
2585 if (!udpif
->dump
->terse
) {
2586 return ukey
->created
;
2589 if (stats
->n_packets
> ukey
->stats
.n_packets
) {
2590 stats
->used
= udpif
->dpif
->current_ms
;
2591 } else if (ukey
->stats
.used
) {
2592 stats
->used
= ukey
->stats
.used
;
2594 stats
->used
= ukey
->created
;
2600 revalidate(struct revalidator
*revalidator
)
2602 uint64_t odp_actions_stub
[1024 / 8];
2603 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2605 struct udpif
*udpif
= revalidator
->udpif
;
2606 struct dpif_flow_dump_thread
*dump_thread
;
2607 uint64_t dump_seq
, reval_seq
;
2608 unsigned int flow_limit
;
2610 dump_seq
= seq_read(udpif
->dump_seq
);
2611 reval_seq
= seq_read(udpif
->reval_seq
);
2612 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2613 dump_thread
= dpif_flow_dump_thread_create(udpif
->dump
);
2615 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2618 struct dpif_flow flows
[REVALIDATE_MAX_BATCH
];
2619 const struct dpif_flow
*f
;
2622 long long int max_idle
;
2627 n_dumped
= dpif_flow_dump_next(dump_thread
, flows
, ARRAY_SIZE(flows
));
2634 /* In normal operation we want to keep flows around until they have
2635 * been idle for 'ofproto_max_idle' milliseconds. However:
2637 * - If the number of datapath flows climbs above 'flow_limit',
2638 * drop that down to 100 ms to try to bring the flows down to
2641 * - If the number of datapath flows climbs above twice
2642 * 'flow_limit', delete all the datapath flows as an emergency
2643 * measure. (We reassess this condition for the next batch of
2644 * datapath flows, so we will recover before all the flows are
2646 n_dp_flows
= udpif_get_n_flows(udpif
);
2647 if (n_dp_flows
>= flow_limit
) {
2648 COVERAGE_INC(upcall_flow_limit_hit
);
2651 kill_them_all
= n_dp_flows
> flow_limit
* 2;
2652 max_idle
= n_dp_flows
> flow_limit
? 100 : ofproto_max_idle
;
2654 udpif
->dpif
->current_ms
= time_msec();
2655 for (f
= flows
; f
< &flows
[n_dumped
]; f
++) {
2656 long long int used
= f
->stats
.used
;
2657 struct recirc_refs recircs
= RECIRC_REFS_EMPTY_INITIALIZER
;
2658 struct dpif_flow_stats stats
= f
->stats
;
2659 enum reval_result result
;
2660 struct udpif_key
*ukey
;
2661 bool already_dumped
;
2664 if (ukey_acquire(udpif
, f
, &ukey
, &error
)) {
2665 if (error
== EBUSY
) {
2666 /* Another thread is processing this flow, so don't bother
2668 COVERAGE_INC(upcall_ukey_contention
);
2670 log_unexpected_flow(f
, error
);
2671 if (error
!= ENOENT
) {
2672 delete_op_init__(udpif
, &ops
[n_ops
++], f
);
2678 already_dumped
= ukey
->dump_seq
== dump_seq
;
2679 if (already_dumped
) {
2680 /* The flow has already been handled during this flow dump
2681 * operation. Skip it. */
2683 COVERAGE_INC(dumped_duplicate_flow
);
2685 COVERAGE_INC(dumped_new_flow
);
2687 ovs_mutex_unlock(&ukey
->mutex
);
2691 if (ukey
->state
<= UKEY_OPERATIONAL
) {
2692 /* The flow is now confirmed to be in the datapath. */
2693 transition_ukey(ukey
, UKEY_OPERATIONAL
);
2695 VLOG_INFO("Unexpected ukey transition from state %d "
2696 "(last transitioned from thread %u at %s)",
2697 ukey
->state
, ukey
->state_thread
, ukey
->state_where
);
2698 ovs_mutex_unlock(&ukey
->mutex
);
2703 used
= udpif_update_used(udpif
, ukey
, &stats
);
2705 if (kill_them_all
|| (used
&& used
< now
- max_idle
)) {
2706 result
= UKEY_DELETE
;
2708 result
= revalidate_ukey(udpif
, ukey
, &stats
, &odp_actions
,
2709 reval_seq
, &recircs
,
2710 f
->attrs
.offloaded
);
2712 ukey
->dump_seq
= dump_seq
;
2714 if (netdev_is_offload_rebalance_policy_enabled() &&
2715 result
!= UKEY_DELETE
) {
2716 udpif_update_flow_pps(udpif
, ukey
, f
);
2719 if (result
!= UKEY_KEEP
) {
2720 /* Takes ownership of 'recircs'. */
2721 reval_op_init(&ops
[n_ops
++], result
, udpif
, ukey
, &recircs
,
2724 ovs_mutex_unlock(&ukey
->mutex
);
2728 /* Push datapath ops but defer ukey deletion to 'sweep' phase. */
2729 push_dp_ops(udpif
, ops
, n_ops
);
2733 dpif_flow_dump_thread_destroy(dump_thread
);
2734 ofpbuf_uninit(&odp_actions
);
2737 /* Pauses the 'revalidator', can only proceed after main thread
2738 * calls udpif_resume_revalidators(). */
2740 revalidator_pause(struct revalidator
*revalidator
)
2742 /* The first block is for sync'ing the pause with main thread. */
2743 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2744 /* The second block is for pausing until main thread resumes. */
2745 ovs_barrier_block(&revalidator
->udpif
->pause_barrier
);
2749 revalidator_sweep__(struct revalidator
*revalidator
, bool purge
)
2751 struct udpif
*udpif
;
2752 uint64_t dump_seq
, reval_seq
;
2755 udpif
= revalidator
->udpif
;
2756 dump_seq
= seq_read(udpif
->dump_seq
);
2757 reval_seq
= seq_read(udpif
->reval_seq
);
2758 slice
= revalidator
- udpif
->revalidators
;
2759 ovs_assert(slice
< udpif
->n_revalidators
);
2761 for (int i
= slice
; i
< N_UMAPS
; i
+= udpif
->n_revalidators
) {
2762 uint64_t odp_actions_stub
[1024 / 8];
2763 struct ofpbuf odp_actions
= OFPBUF_STUB_INITIALIZER(odp_actions_stub
);
2765 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2766 struct udpif_key
*ukey
;
2767 struct umap
*umap
= &udpif
->ukeys
[i
];
2770 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2771 enum ukey_state ukey_state
;
2773 /* Handler threads could be holding a ukey lock while it installs a
2774 * new flow, so don't hang around waiting for access to it. */
2775 if (ovs_mutex_trylock(&ukey
->mutex
)) {
2778 ukey_state
= ukey
->state
;
2779 if (ukey_state
== UKEY_OPERATIONAL
2780 || (ukey_state
== UKEY_VISIBLE
&& purge
)) {
2781 struct recirc_refs recircs
= RECIRC_REFS_EMPTY_INITIALIZER
;
2782 bool seq_mismatch
= (ukey
->dump_seq
!= dump_seq
2783 && ukey
->reval_seq
!= reval_seq
);
2784 enum reval_result result
;
2787 result
= UKEY_DELETE
;
2788 } else if (!seq_mismatch
) {
2791 struct dpif_flow_stats stats
;
2792 COVERAGE_INC(revalidate_missed_dp_flow
);
2793 memset(&stats
, 0, sizeof stats
);
2794 result
= revalidate_ukey(udpif
, ukey
, &stats
, &odp_actions
,
2795 reval_seq
, &recircs
, false);
2797 if (result
!= UKEY_KEEP
) {
2798 /* Clears 'recircs' if filled by revalidate_ukey(). */
2799 reval_op_init(&ops
[n_ops
++], result
, udpif
, ukey
, &recircs
,
2803 ovs_mutex_unlock(&ukey
->mutex
);
2805 if (ukey_state
== UKEY_EVICTED
) {
2806 /* The common flow deletion case involves deletion of the flow
2807 * during the dump phase and ukey deletion here. */
2808 ovs_mutex_lock(&umap
->mutex
);
2809 ukey_delete(umap
, ukey
);
2810 ovs_mutex_unlock(&umap
->mutex
);
2813 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2814 /* Update/delete missed flows and clean up corresponding ukeys
2816 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2822 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2825 ofpbuf_uninit(&odp_actions
);
2831 revalidator_sweep(struct revalidator
*revalidator
)
2833 revalidator_sweep__(revalidator
, false);
2837 revalidator_purge(struct revalidator
*revalidator
)
2839 revalidator_sweep__(revalidator
, true);
2842 /* In reaction to dpif purge, purges all 'ukey's with same 'pmd_id'. */
2844 dp_purge_cb(void *aux
, unsigned pmd_id
)
2845 OVS_NO_THREAD_SAFETY_ANALYSIS
2847 struct udpif
*udpif
= aux
;
2850 udpif_pause_revalidators(udpif
);
2851 for (i
= 0; i
< N_UMAPS
; i
++) {
2852 struct ukey_op ops
[REVALIDATE_MAX_BATCH
];
2853 struct udpif_key
*ukey
;
2854 struct umap
*umap
= &udpif
->ukeys
[i
];
2857 CMAP_FOR_EACH(ukey
, cmap_node
, &umap
->cmap
) {
2858 if (ukey
->pmd_id
== pmd_id
) {
2859 delete_op_init(udpif
, &ops
[n_ops
++], ukey
);
2860 transition_ukey(ukey
, UKEY_EVICTING
);
2862 if (n_ops
== REVALIDATE_MAX_BATCH
) {
2863 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2870 push_ukey_ops(udpif
, umap
, ops
, n_ops
);
2875 udpif_resume_revalidators(udpif
);
2879 upcall_unixctl_show(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2880 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2882 struct ds ds
= DS_EMPTY_INITIALIZER
;
2883 struct udpif
*udpif
;
2885 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2886 unsigned int flow_limit
;
2890 atomic_read_relaxed(&udpif
->flow_limit
, &flow_limit
);
2891 ufid_enabled
= udpif_use_ufid(udpif
);
2893 ds_put_format(&ds
, "%s:\n", dpif_name(udpif
->dpif
));
2894 ds_put_format(&ds
, " flows : (current %lu)"
2895 " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif
),
2896 udpif
->avg_n_flows
, udpif
->max_n_flows
, flow_limit
);
2897 ds_put_format(&ds
, " dump duration : %lldms\n", udpif
->dump_duration
);
2898 ds_put_format(&ds
, " ufid enabled : ");
2900 ds_put_format(&ds
, "true\n");
2902 ds_put_format(&ds
, "false\n");
2904 ds_put_char(&ds
, '\n');
2906 for (i
= 0; i
< n_revalidators
; i
++) {
2907 struct revalidator
*revalidator
= &udpif
->revalidators
[i
];
2908 int j
, elements
= 0;
2910 for (j
= i
; j
< N_UMAPS
; j
+= n_revalidators
) {
2911 elements
+= cmap_count(&udpif
->ukeys
[j
].cmap
);
2913 ds_put_format(&ds
, " %u: (keys %d)\n", revalidator
->id
, elements
);
2917 unixctl_command_reply(conn
, ds_cstr(&ds
));
2921 /* Disable using the megaflows.
2923 * This command is only needed for advanced debugging, so it's not
2924 * documented in the man page. */
2926 upcall_unixctl_disable_megaflows(struct unixctl_conn
*conn
,
2927 int argc OVS_UNUSED
,
2928 const char *argv
[] OVS_UNUSED
,
2929 void *aux OVS_UNUSED
)
2931 atomic_store_relaxed(&enable_megaflows
, false);
2932 udpif_flush_all_datapaths();
2933 unixctl_command_reply(conn
, "megaflows disabled");
2936 /* Re-enable using megaflows.
2938 * This command is only needed for advanced debugging, so it's not
2939 * documented in the man page. */
2941 upcall_unixctl_enable_megaflows(struct unixctl_conn
*conn
,
2942 int argc OVS_UNUSED
,
2943 const char *argv
[] OVS_UNUSED
,
2944 void *aux OVS_UNUSED
)
2946 atomic_store_relaxed(&enable_megaflows
, true);
2947 udpif_flush_all_datapaths();
2948 unixctl_command_reply(conn
, "megaflows enabled");
2951 /* Disable skipping flow attributes during flow dump.
2953 * This command is only needed for advanced debugging, so it's not
2954 * documented in the man page. */
2956 upcall_unixctl_disable_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
, false);
2960 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID disabled");
2963 /* Re-enable skipping flow attributes during flow dump.
2965 * This command is only needed for advanced debugging, so it's not documented
2966 * in the man page. */
2968 upcall_unixctl_enable_ufid(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
2969 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
2971 atomic_store_relaxed(&enable_ufid
, true);
2972 unixctl_command_reply(conn
, "Datapath dumping tersely using UFID enabled "
2973 "for supported datapaths");
2976 /* Set the flow limit.
2978 * This command is only needed for advanced debugging, so it's not
2979 * documented in the man page. */
2981 upcall_unixctl_set_flow_limit(struct unixctl_conn
*conn
,
2982 int argc OVS_UNUSED
,
2984 void *aux OVS_UNUSED
)
2986 struct ds ds
= DS_EMPTY_INITIALIZER
;
2987 struct udpif
*udpif
;
2988 unsigned int flow_limit
= atoi(argv
[1]);
2990 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
2991 atomic_store_relaxed(&udpif
->flow_limit
, flow_limit
);
2993 ds_put_format(&ds
, "set flow_limit to %u\n", flow_limit
);
2994 unixctl_command_reply(conn
, ds_cstr(&ds
));
2999 upcall_unixctl_dump_wait(struct unixctl_conn
*conn
,
3000 int argc OVS_UNUSED
,
3001 const char *argv
[] OVS_UNUSED
,
3002 void *aux OVS_UNUSED
)
3004 if (ovs_list_is_singleton(&all_udpifs
)) {
3005 struct udpif
*udpif
= NULL
;
3008 udpif
= OBJECT_CONTAINING(ovs_list_front(&all_udpifs
), udpif
, list_node
);
3009 len
= (udpif
->n_conns
+ 1) * sizeof *udpif
->conns
;
3010 udpif
->conn_seq
= seq_read(udpif
->dump_seq
);
3011 udpif
->conns
= xrealloc(udpif
->conns
, len
);
3012 udpif
->conns
[udpif
->n_conns
++] = conn
;
3014 unixctl_command_reply_error(conn
, "can't wait on multiple udpifs.");
3019 upcall_unixctl_purge(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
3020 const char *argv
[] OVS_UNUSED
, void *aux OVS_UNUSED
)
3022 struct udpif
*udpif
;
3024 LIST_FOR_EACH (udpif
, list_node
, &all_udpifs
) {
3027 for (n
= 0; n
< udpif
->n_revalidators
; n
++) {
3028 revalidator_purge(&udpif
->revalidators
[n
]);
3031 unixctl_command_reply(conn
, "");
3034 /* Flows are sorted in the following order:
3035 * netdev, flow state (offloaded/kernel path), flow_pps_rate.
3038 flow_compare_rebalance(const void *elem1
, const void *elem2
)
3040 const struct udpif_key
*f1
= *(struct udpif_key
**)elem1
;
3041 const struct udpif_key
*f2
= *(struct udpif_key
**)elem2
;
3044 if (f1
->in_netdev
< f2
->in_netdev
) {
3046 } else if (f1
->in_netdev
> f2
->in_netdev
) {
3050 if (f1
->offloaded
!= f2
->offloaded
) {
3051 return f2
->offloaded
- f1
->offloaded
;
3054 diff
= (f1
->offloaded
== true) ?
3055 f1
->flow_pps_rate
- f2
->flow_pps_rate
:
3056 f2
->flow_pps_rate
- f1
->flow_pps_rate
;
3058 return (diff
< 0) ? -1 : 1;
3061 /* Insert flows from pending array during rebalancing */
3063 rebalance_insert_pending(struct udpif
*udpif
, struct udpif_key
**pending_flows
,
3064 int pending_count
, int insert_count
,
3065 uint64_t rate_threshold
)
3069 for (int i
= 0; i
< pending_count
; i
++) {
3070 struct udpif_key
*flow
= pending_flows
[i
];
3073 /* Stop offloading pending flows if the insert count is
3074 * reached and the flow rate is less than the threshold
3076 if (count
>= insert_count
&& flow
->flow_pps_rate
< rate_threshold
) {
3080 /* Offload the flow to netdev */
3081 err
= udpif_flow_program(udpif
, flow
, DPIF_OFFLOAD_ALWAYS
);
3083 if (err
== ENOSPC
) {
3084 /* Stop if we are out of resources */
3092 /* Offload succeeded; delete it from the kernel datapath */
3093 udpif_flow_unprogram(udpif
, flow
, DPIF_OFFLOAD_NEVER
);
3095 /* Change the state of the flow, adjust dpif counters */
3096 flow
->offloaded
= true;
3098 udpif_set_ukey_backlog_packets(flow
);
3105 /* Remove flows from offloaded array during rebalancing */
3107 rebalance_remove_offloaded(struct udpif
*udpif
,
3108 struct udpif_key
**offloaded_flows
,
3111 for (int i
= 0; i
< offload_count
; i
++) {
3112 struct udpif_key
*flow
= offloaded_flows
[i
];
3115 /* Install the flow into kernel path first */
3116 err
= udpif_flow_program(udpif
, flow
, DPIF_OFFLOAD_NEVER
);
3121 /* Success; now remove offloaded flow from netdev */
3122 err
= udpif_flow_unprogram(udpif
, flow
, DPIF_OFFLOAD_ALWAYS
);
3124 udpif_flow_unprogram(udpif
, flow
, DPIF_OFFLOAD_NEVER
);
3127 udpif_set_ukey_backlog_packets(flow
);
3128 flow
->offloaded
= false;
3133 * Rebalance offloaded flows on a netdev that's in OOR state.
3135 * The rebalancing is done in two phases. In the first phase, we check if
3136 * the pending flows can be offloaded (if some resources became available
3137 * in the meantime) by trying to offload each pending flow. If all pending
3138 * flows get successfully offloaded, the OOR state is cleared on the netdev
3139 * and there's nothing to rebalance.
3141 * If some of the pending flows could not be offloaded, i.e, we still see
3142 * the OOR error, then we move to the second phase of rebalancing. In this
3143 * phase, the rebalancer compares pps-rate of an offloaded flow with the
3144 * least pps-rate with that of a pending flow with the highest pps-rate from
3145 * their respective sorted arrays. If pps-rate of the offloaded flow is less
3146 * than the pps-rate of the pending flow, then it deletes the offloaded flow
3147 * from the HW/netdev and adds it to kernel datapath and then offloads pending
3148 * to HW/netdev. This process is repeated for every pair of offloaded and
3149 * pending flows in the ordered list. The process stops when we encounter an
3150 * offloaded flow that has a higher pps-rate than the corresponding pending
3151 * flow. The entire rebalancing process is repeated in the next iteration.
3154 rebalance_device(struct udpif
*udpif
, struct udpif_key
**offloaded_flows
,
3155 int offload_count
, struct udpif_key
**pending_flows
,
3160 int num_inserted
= rebalance_insert_pending(udpif
, pending_flows
,
3161 pending_count
, pending_count
,
3164 VLOG_DBG("Offload rebalance: Phase1: inserted %d pending flows",
3168 /* Adjust pending array */
3169 pending_flows
= &pending_flows
[num_inserted
];
3170 pending_count
-= num_inserted
;
3172 if (!pending_count
) {
3174 * Successfully offloaded all pending flows. The device
3175 * is no longer in OOR state; done rebalancing this device.
3181 * Phase 2; determine how many offloaded flows to churn.
3183 #define OFFL_REBAL_MAX_CHURN 1024
3184 int churn_count
= 0;
3185 while (churn_count
< OFFL_REBAL_MAX_CHURN
&& churn_count
< offload_count
3186 && churn_count
< pending_count
) {
3187 if (pending_flows
[churn_count
]->flow_pps_rate
<=
3188 offloaded_flows
[churn_count
]->flow_pps_rate
)
3194 VLOG_DBG("Offload rebalance: Phase2: removing %d offloaded flows",
3198 /* Bail early if nothing to churn */
3203 /* Remove offloaded flows */
3204 rebalance_remove_offloaded(udpif
, offloaded_flows
, churn_count
);
3206 /* Adjust offloaded array */
3207 offloaded_flows
= &offloaded_flows
[churn_count
];
3208 offload_count
-= churn_count
;
3210 /* Replace offloaded flows with pending flows */
3211 num_inserted
= rebalance_insert_pending(udpif
, pending_flows
,
3212 pending_count
, churn_count
,
3214 offloaded_flows
[0]->flow_pps_rate
:
3217 VLOG_DBG("Offload rebalance: Phase2: inserted %d pending flows",
3224 static struct udpif_key
**
3225 udpif_add_oor_flows(struct udpif_key
**sort_flows
, size_t *total_flow_count
,
3226 size_t *alloc_flow_count
, struct udpif_key
*ukey
)
3228 if (*total_flow_count
>= *alloc_flow_count
) {
3229 sort_flows
= x2nrealloc(sort_flows
, alloc_flow_count
, sizeof ukey
);
3231 sort_flows
[(*total_flow_count
)++] = ukey
;
3236 * Build sort_flows[] initially with flows that
3237 * reference an 'OOR' netdev as their input port.
3239 static struct udpif_key
**
3240 udpif_build_oor_flows(struct udpif_key
**sort_flows
, size_t *total_flow_count
,
3241 size_t *alloc_flow_count
, struct udpif_key
*ukey
,
3242 int *oor_netdev_count
)
3244 struct netdev
*netdev
;
3247 /* Input netdev must be available for the flow */
3248 netdev
= ukey
->in_netdev
;
3253 /* Is the in-netdev for this flow in OOR state ? */
3254 if (!netdev_get_hw_info(netdev
, HW_INFO_TYPE_OOR
)) {
3255 ukey_netdev_unref(ukey
);
3259 /* Add the flow to sort_flows[] */
3260 sort_flows
= udpif_add_oor_flows(sort_flows
, total_flow_count
,
3261 alloc_flow_count
, ukey
);
3262 if (ukey
->offloaded
) {
3263 count
= netdev_get_hw_info(netdev
, HW_INFO_TYPE_OFFL_COUNT
);
3264 ovs_assert(count
>= 0);
3266 (*oor_netdev_count
)++;
3268 netdev_set_hw_info(netdev
, HW_INFO_TYPE_OFFL_COUNT
, count
);
3270 count
= netdev_get_hw_info(netdev
, HW_INFO_TYPE_PEND_COUNT
);
3271 ovs_assert(count
>= 0);
3272 netdev_set_hw_info(netdev
, HW_INFO_TYPE_PEND_COUNT
, ++count
);
3279 * Rebalance offloaded flows on HW netdevs that are in OOR state.
3282 udpif_flow_rebalance(struct udpif
*udpif
)
3284 struct udpif_key
**sort_flows
= NULL
;
3285 size_t alloc_flow_count
= 0;
3286 size_t total_flow_count
= 0;
3287 int oor_netdev_count
= 0;
3288 int offload_index
= 0;
3291 /* Collect flows (offloaded and pending) that reference OOR netdevs */
3292 for (size_t i
= 0; i
< N_UMAPS
; i
++) {
3293 struct udpif_key
*ukey
;
3294 struct umap
*umap
= &udpif
->ukeys
[i
];
3296 CMAP_FOR_EACH (ukey
, cmap_node
, &umap
->cmap
) {
3297 ukey_to_flow_netdev(udpif
, ukey
);
3298 sort_flows
= udpif_build_oor_flows(sort_flows
, &total_flow_count
,
3299 &alloc_flow_count
, ukey
,
3304 /* Sort flows by OOR netdevs, state (offloaded/pending) and pps-rate */
3305 qsort(sort_flows
, total_flow_count
, sizeof(struct udpif_key
*),
3306 flow_compare_rebalance
);
3309 * We now have flows referencing OOR netdevs, that are sorted. We also
3310 * have a count of offloaded and pending flows on each of the netdevs
3311 * that are in OOR state. Now rebalance each oor-netdev.
3313 while (oor_netdev_count
) {
3314 struct netdev
*netdev
;
3319 netdev
= sort_flows
[offload_index
]->in_netdev
;
3320 ovs_assert(netdev_get_hw_info(netdev
, HW_INFO_TYPE_OOR
) == true);
3321 VLOG_DBG("Offload rebalance: netdev: %s is OOR", netdev
->name
);
3323 offload_count
= netdev_get_hw_info(netdev
, HW_INFO_TYPE_OFFL_COUNT
);
3324 pending_count
= netdev_get_hw_info(netdev
, HW_INFO_TYPE_PEND_COUNT
);
3325 pending_index
= offload_index
+ offload_count
;
3327 oor
= rebalance_device(udpif
,
3328 &sort_flows
[offload_index
], offload_count
,
3329 &sort_flows
[pending_index
], pending_count
);
3330 netdev_set_hw_info(netdev
, HW_INFO_TYPE_OOR
, oor
);
3332 offload_index
= pending_index
+ pending_count
;
3333 netdev_set_hw_info(netdev
, HW_INFO_TYPE_OFFL_COUNT
, 0);
3334 netdev_set_hw_info(netdev
, HW_INFO_TYPE_PEND_COUNT
, 0);
3338 for (int i
= 0; i
< total_flow_count
; i
++) {
3339 struct udpif_key
*ukey
= sort_flows
[i
];
3340 ukey_netdev_unref(ukey
);
3346 udpif_flow_program(struct udpif
*udpif
, struct udpif_key
*ukey
,
3347 enum dpif_offload_type offload_type
)
3349 struct dpif_op
*opsp
;
3353 put_op_init(&uop
, ukey
, DPIF_FP_CREATE
);
3354 dpif_operate(udpif
->dpif
, &opsp
, 1, offload_type
);
3360 udpif_flow_unprogram(struct udpif
*udpif
, struct udpif_key
*ukey
,
3361 enum dpif_offload_type offload_type
)
3363 struct dpif_op
*opsp
;
3367 delete_op_init(udpif
, &uop
, ukey
);
3368 dpif_operate(udpif
->dpif
, &opsp
, 1, offload_type
);