]>
Commit | Line | Data |
---|---|---|
f620b43a BP |
1 | /* |
2 | * Copyright (c) 2008, 2009, 2010, 2011 Nicira Networks. | |
3 | * | |
4 | * Licensed under the Apache License, Version 2.0 (the "License"); | |
5 | * you may not use this file except in compliance with the License. | |
6 | * You may obtain a copy of the License at: | |
7 | * | |
8 | * http://www.apache.org/licenses/LICENSE-2.0 | |
9 | * | |
10 | * Unless required by applicable law or agreed to in writing, software | |
11 | * distributed under the License is distributed on an "AS IS" BASIS, | |
12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |
13 | * See the License for the specific language governing permissions and | |
14 | * limitations under the License. | |
15 | */ | |
16 | ||
17 | #include <config.h> | |
18 | ||
19 | #include "bond.h" | |
20 | ||
21 | #include <limits.h> | |
22 | #include <stdint.h> | |
23 | #include <stdlib.h> | |
24 | ||
25 | #include "coverage.h" | |
26 | #include "dynamic-string.h" | |
27 | #include "flow.h" | |
28 | #include "hmap.h" | |
29 | #include "lacp.h" | |
30 | #include "list.h" | |
31 | #include "netdev.h" | |
32 | #include "odp-util.h" | |
33 | #include "ofpbuf.h" | |
34 | #include "packets.h" | |
35 | #include "poll-loop.h" | |
36 | #include "tag.h" | |
37 | #include "timeval.h" | |
38 | #include "unixctl.h" | |
39 | #include "vlog.h" | |
40 | ||
41 | VLOG_DEFINE_THIS_MODULE(bond); | |
42 | ||
43 | COVERAGE_DEFINE(bond_process_lacp); | |
44 | ||
45 | /* Bit-mask for hashing a flow down to a bucket. | |
46 | * There are (BOND_MASK + 1) buckets. */ | |
47 | #define BOND_MASK 0xff | |
48 | ||
49 | /* A hash bucket for mapping a flow to a slave. | |
50 | * "struct bond" has an array of (BOND_MASK + 1) of these. */ | |
51 | struct bond_entry { | |
52 | struct bond_slave *slave; /* Assigned slave, NULL if unassigned. */ | |
53 | uint64_t tx_bytes; /* Count of bytes recently transmitted. */ | |
54 | tag_type tag; /* Tag for entry<->slave association. */ | |
55 | struct list list_node; /* In bond_slave's 'entries' list. */ | |
56 | }; | |
57 | ||
58 | /* A bond slave, that is, one of the links comprising a bond. */ | |
59 | struct bond_slave { | |
60 | struct hmap_node hmap_node; /* In struct bond's slaves hmap. */ | |
61 | struct bond *bond; /* The bond that contains this slave. */ | |
62 | void *aux; /* Client-provided handle for this slave. */ | |
63 | ||
64 | struct netdev *netdev; /* Network device, owned by the client. */ | |
65 | char *name; /* Name (a copy of netdev_get_name(netdev)). */ | |
66 | ||
67 | /* Link status. */ | |
68 | long long delay_expires; /* Time after which 'enabled' may change. */ | |
69 | bool up; /* Last link status read from netdev. */ | |
70 | bool enabled; /* May be chosen for flows? */ | |
71 | tag_type tag; /* Tag associated with this slave. */ | |
72 | ||
73 | /* Rebalancing info. Used only by bond_rebalance(). */ | |
74 | struct list bal_node; /* In bond_rebalance()'s 'bals' list. */ | |
75 | struct list entries; /* 'struct bond_entry's assigned here. */ | |
76 | uint64_t tx_bytes; /* Sum across 'tx_bytes' of entries. */ | |
77 | }; | |
78 | ||
79 | /* A bond, that is, a set of network devices grouped to improve performance or | |
80 | * robustness. */ | |
81 | struct bond { | |
82 | struct hmap_node hmap_node; /* In 'all_bonds' hmap. */ | |
83 | char *name; /* Name provided by client. */ | |
84 | ||
85 | /* Slaves. */ | |
86 | struct hmap slaves; | |
87 | ||
88 | /* Bonding info. */ | |
89 | enum bond_mode balance; /* Balancing mode, one of BM_*. */ | |
90 | struct bond_slave *active_slave; | |
91 | tag_type no_slaves_tag; /* Tag for flows when all slaves disabled. */ | |
92 | int updelay, downdelay; /* Delay before slave goes up/down, in ms. */ | |
93 | ||
94 | /* SLB specific bonding info. */ | |
95 | struct bond_entry *hash; /* An array of (BOND_MASK + 1) elements. */ | |
96 | int rebalance_interval; /* Interval between rebalances, in ms. */ | |
97 | long long int next_rebalance; /* Next rebalancing time. */ | |
98 | bool send_learning_packets; | |
99 | ||
100 | /* LACP. */ | |
101 | struct lacp *lacp; /* LACP object. NULL if LACP is disabled. */ | |
102 | ||
103 | /* Monitoring. */ | |
104 | enum bond_detect_mode detect; /* Link status mode, one of BLSM_*. */ | |
105 | struct netdev_monitor *monitor; /* detect == BLSM_CARRIER only. */ | |
106 | long long int miimon_interval; /* Miimon status refresh interval. */ | |
107 | long long int miimon_next_update; /* Time of next miimon update. */ | |
108 | ||
109 | /* Legacy compatibility. */ | |
110 | long long int next_fake_iface_update; /* LLONG_MAX if disabled. */ | |
111 | ||
112 | /* Tag set saved for next bond_run(). This tag set is a kluge for cases | |
113 | * where we can't otherwise provide revalidation feedback to the client. | |
114 | * That's only unixctl commands now; I hope no other cases will arise. */ | |
115 | struct tag_set unixctl_tags; | |
116 | }; | |
117 | ||
118 | static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds); | |
119 | ||
120 | static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_); | |
121 | static bool bond_is_link_up(struct bond *, struct netdev *); | |
122 | static void bond_enable_slave(struct bond_slave *, bool enable, | |
123 | struct tag_set *); | |
124 | static void bond_link_status_update(struct bond_slave *, struct tag_set *); | |
125 | static void bond_choose_active_slave(struct bond *, struct tag_set *); | |
126 | static bool bond_is_tcp_hash(const struct bond *); | |
127 | static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], | |
128 | uint16_t vlan); | |
129 | static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan); | |
130 | static struct bond_entry *lookup_bond_entry(const struct bond *, | |
131 | const struct flow *, | |
132 | uint16_t vlan); | |
133 | static tag_type bond_get_active_slave_tag(const struct bond *); | |
134 | static struct bond_slave *choose_output_slave(const struct bond *, | |
135 | const struct flow *, | |
136 | uint16_t vlan); | |
137 | static void bond_update_fake_slave_stats(struct bond *); | |
138 | ||
139 | /* Attempts to parse 's' as the name of a bond balancing mode. If successful, | |
140 | * stores the mode in '*balance' and returns true. Otherwise returns false | |
141 | * without modifying '*balance'. */ | |
142 | bool | |
143 | bond_mode_from_string(enum bond_mode *balance, const char *s) | |
144 | { | |
145 | if (!strcmp(s, bond_mode_to_string(BM_TCP))) { | |
146 | *balance = BM_TCP; | |
147 | } else if (!strcmp(s, bond_mode_to_string(BM_SLB))) { | |
148 | *balance = BM_SLB; | |
149 | } else if (!strcmp(s, bond_mode_to_string(BM_AB))) { | |
150 | *balance = BM_AB; | |
151 | } else { | |
152 | return false; | |
153 | } | |
154 | return true; | |
155 | } | |
156 | ||
157 | /* Returns a string representing 'balance'. */ | |
158 | const char * | |
159 | bond_mode_to_string(enum bond_mode balance) { | |
160 | switch (balance) { | |
161 | case BM_TCP: | |
162 | return "balance-tcp"; | |
163 | case BM_SLB: | |
164 | return "balance-slb"; | |
165 | case BM_AB: | |
166 | return "active-backup"; | |
167 | } | |
168 | NOT_REACHED(); | |
169 | } | |
170 | ||
171 | /* Attempts to parse 's' as the name of a bond link status detection mode. If | |
172 | * successful, stores the mode in '*detect' and returns true. Otherwise | |
173 | * returns false without modifying '*detect'. */ | |
174 | bool | |
175 | bond_detect_mode_from_string(enum bond_detect_mode *detect, const char *s) | |
176 | { | |
177 | if (!strcmp(s, bond_detect_mode_to_string(BLSM_CARRIER))) { | |
178 | *detect = BLSM_CARRIER; | |
179 | } else if (!strcmp(s, bond_detect_mode_to_string(BLSM_MIIMON))) { | |
180 | *detect = BLSM_MIIMON; | |
181 | } else { | |
182 | return false; | |
183 | } | |
184 | return true; | |
185 | } | |
186 | ||
187 | /* Returns a string representing 'detect'. */ | |
188 | const char * | |
189 | bond_detect_mode_to_string(enum bond_detect_mode detect) | |
190 | { | |
191 | switch (detect) { | |
192 | case BLSM_CARRIER: | |
193 | return "carrier"; | |
194 | case BLSM_MIIMON: | |
195 | return "miimon"; | |
196 | } | |
197 | NOT_REACHED(); | |
198 | } | |
199 | \f | |
200 | /* Creates and returns a new bond whose configuration is initially taken from | |
201 | * 's'. | |
202 | * | |
203 | * The caller should register each slave on the new bond by calling | |
204 | * bond_slave_register(). */ | |
205 | struct bond * | |
206 | bond_create(const struct bond_settings *s) | |
207 | { | |
208 | struct bond *bond; | |
209 | ||
210 | bond = xzalloc(sizeof *bond); | |
211 | hmap_init(&bond->slaves); | |
212 | bond->no_slaves_tag = tag_create_random(); | |
213 | bond->miimon_next_update = LLONG_MAX; | |
214 | bond->next_fake_iface_update = LLONG_MAX; | |
215 | ||
216 | bond_reconfigure(bond, s); | |
217 | ||
218 | tag_set_init(&bond->unixctl_tags); | |
219 | ||
220 | return bond; | |
221 | } | |
222 | ||
223 | /* Frees 'bond'. */ | |
224 | void | |
225 | bond_destroy(struct bond *bond) | |
226 | { | |
227 | struct bond_slave *slave, *next_slave; | |
228 | ||
229 | if (!bond) { | |
230 | return; | |
231 | } | |
232 | ||
233 | hmap_remove(&all_bonds, &bond->hmap_node); | |
234 | ||
235 | HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) { | |
236 | hmap_remove(&bond->slaves, &slave->hmap_node); | |
237 | /* Client owns 'slave->netdev'. */ | |
238 | free(slave->name); | |
239 | free(slave); | |
240 | } | |
241 | hmap_destroy(&bond->slaves); | |
242 | ||
243 | free(bond->hash); | |
244 | ||
245 | lacp_destroy(bond->lacp); | |
246 | ||
247 | netdev_monitor_destroy(bond->monitor); | |
248 | ||
249 | free(bond->name); | |
250 | free(bond); | |
251 | } | |
252 | ||
253 | /* Updates 'bond''s overall configuration to 's'. | |
254 | * | |
255 | * The caller should register each slave on 'bond' by calling | |
256 | * bond_slave_register(). This is optional if none of the slaves' | |
257 | * configuration has changed, except that it is mandatory if 's' enables LACP | |
258 | * and 'bond' previously didn't have LACP enabled. In any case it can't | |
259 | * hurt. */ | |
260 | void | |
261 | bond_reconfigure(struct bond *bond, const struct bond_settings *s) | |
262 | { | |
263 | if (!bond->name || strcmp(bond->name, s->name)) { | |
264 | if (bond->name) { | |
265 | hmap_remove(&all_bonds, &bond->hmap_node); | |
266 | free(bond->name); | |
267 | } | |
268 | bond->name = xstrdup(s->name); | |
269 | hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0)); | |
270 | } | |
271 | ||
272 | bond->balance = s->balance; | |
273 | bond->detect = s->detect; | |
274 | bond->miimon_interval = s->miimon_interval; | |
275 | bond->updelay = s->up_delay; | |
276 | bond->downdelay = s->down_delay; | |
277 | bond->rebalance_interval = s->rebalance_interval; | |
278 | ||
279 | if (bond->balance != BM_AB) { | |
280 | if (!bond->hash) { | |
281 | bond->hash = xcalloc(BOND_MASK + 1, sizeof *bond->hash); | |
282 | bond->next_rebalance = time_msec() + bond->rebalance_interval; | |
283 | } | |
284 | } else { | |
285 | if (bond->hash) { | |
286 | free(bond->hash); | |
287 | bond->hash = NULL; | |
288 | } | |
289 | } | |
290 | ||
291 | if (bond->detect == BLSM_CARRIER) { | |
292 | struct bond_slave *slave; | |
293 | ||
294 | if (!bond->monitor) { | |
295 | bond->monitor = netdev_monitor_create(); | |
296 | } | |
297 | ||
298 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
299 | netdev_monitor_add(bond->monitor, slave->netdev); | |
300 | } | |
301 | } else { | |
302 | netdev_monitor_destroy(bond->monitor); | |
303 | bond->monitor = NULL; | |
304 | ||
305 | if (bond->miimon_next_update == LLONG_MAX) { | |
306 | bond->miimon_next_update = time_msec() + bond->miimon_interval; | |
307 | } | |
308 | } | |
309 | ||
310 | if (s->lacp) { | |
311 | if (!bond->lacp) { | |
312 | bond->lacp = lacp_create(); | |
313 | } | |
314 | lacp_configure(bond->lacp, s->lacp); | |
315 | } else { | |
316 | lacp_destroy(bond->lacp); | |
317 | bond->lacp = NULL; | |
318 | } | |
319 | ||
320 | if (s->fake_iface) { | |
321 | if (bond->next_fake_iface_update == LLONG_MAX) { | |
322 | bond->next_fake_iface_update = time_msec(); | |
323 | } | |
324 | } else { | |
325 | bond->next_fake_iface_update = LLONG_MAX; | |
326 | } | |
327 | } | |
328 | ||
329 | /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an | |
330 | * arbitrary client-provided pointer that uniquely identifies a slave within a | |
331 | * bond. If 'slave_' already exists within 'bond' then this function | |
332 | * reconfigures the existing slave. | |
333 | * | |
334 | * 'netdev' must be the network device that 'slave_' represents. It is owned | |
335 | * by the client, so the client must not close it before either unregistering | |
336 | * 'slave_' or destroying 'bond'. | |
337 | * | |
338 | * If 'bond' has a LACP configuration then 'lacp_settings' must point to LACP | |
339 | * settings for 'slave_'; otherwise 'lacp_settings' is ignored. | |
340 | */ | |
341 | void | |
342 | bond_slave_register(struct bond *bond, void *slave_, struct netdev *netdev, | |
343 | const struct lacp_slave_settings *lacp_settings) | |
344 | { | |
345 | struct bond_slave *slave = bond_slave_lookup(bond, slave_); | |
346 | ||
347 | if (!slave) { | |
348 | slave = xzalloc(sizeof *slave); | |
349 | ||
350 | hmap_insert(&bond->slaves, &slave->hmap_node, hash_pointer(slave_, 0)); | |
351 | slave->bond = bond; | |
352 | slave->aux = slave_; | |
353 | slave->delay_expires = LLONG_MAX; | |
354 | slave->up = bond_is_link_up(bond, netdev); | |
355 | slave->enabled = slave->up; | |
356 | } | |
357 | ||
358 | slave->netdev = netdev; | |
359 | free(slave->name); | |
360 | slave->name = xstrdup(netdev_get_name(netdev)); | |
361 | ||
362 | if (bond->lacp) { | |
363 | assert(lacp_settings != NULL); | |
364 | lacp_slave_register(bond->lacp, slave, lacp_settings); | |
365 | } | |
366 | } | |
367 | ||
368 | /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave | |
369 | * then this function has no effect. | |
370 | * | |
371 | * Unregistering a slave invalidates all flows. */ | |
372 | void | |
373 | bond_slave_unregister(struct bond *bond, const void *slave_) | |
374 | { | |
375 | struct bond_slave *slave = bond_slave_lookup(bond, slave_); | |
376 | bool del_active; | |
377 | ||
378 | if (!slave) { | |
379 | return; | |
380 | } | |
381 | ||
382 | del_active = bond->active_slave == slave; | |
383 | if (bond->hash) { | |
384 | struct bond_entry *e; | |
385 | for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) { | |
386 | if (e->slave == slave) { | |
387 | e->slave = NULL; | |
388 | } | |
389 | } | |
390 | } | |
391 | ||
392 | free(slave->name); | |
393 | ||
394 | hmap_remove(&bond->slaves, &slave->hmap_node); | |
395 | /* Client owns 'slave->netdev'. */ | |
396 | free(slave); | |
397 | ||
398 | if (del_active) { | |
399 | struct tag_set tags; | |
400 | ||
401 | tag_set_init(&tags); | |
402 | bond_choose_active_slave(bond, &tags); | |
403 | bond->send_learning_packets = true; | |
404 | } | |
405 | } | |
406 | ||
407 | /* Callback for lacp_run(). */ | |
408 | static void | |
409 | bond_send_pdu_cb(void *slave_, const struct lacp_pdu *pdu) | |
410 | { | |
411 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 10); | |
412 | struct bond_slave *slave = slave_; | |
413 | uint8_t ea[ETH_ADDR_LEN]; | |
414 | int error; | |
415 | ||
416 | error = netdev_get_etheraddr(slave->netdev, ea); | |
417 | if (!error) { | |
418 | struct lacp_pdu *packet_pdu; | |
419 | struct ofpbuf packet; | |
420 | ||
421 | ofpbuf_init(&packet, 0); | |
422 | packet_pdu = eth_compose(&packet, eth_addr_lacp, ea, ETH_TYPE_LACP, | |
423 | sizeof *packet_pdu); | |
424 | *packet_pdu = *pdu; | |
425 | error = netdev_send(slave->netdev, &packet); | |
426 | if (error) { | |
427 | VLOG_WARN_RL(&rl, "bond %s: sending LACP PDU on slave %s failed " | |
428 | "(%s)", | |
429 | slave->bond->name, slave->name, strerror(error)); | |
430 | } | |
431 | ofpbuf_uninit(&packet); | |
432 | } else { | |
433 | VLOG_ERR_RL(&rl, "bond %s: cannot obtain Ethernet address of slave " | |
434 | "%s (%s)", | |
435 | slave->bond->name, slave->name, strerror(error)); | |
436 | } | |
437 | } | |
438 | ||
439 | /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to | |
440 | * allow tagged flows to be invalidated. | |
441 | * | |
442 | * The caller should check bond_should_send_learning_packets() afterward. */ | |
443 | void | |
444 | bond_run(struct bond *bond, struct tag_set *tags) | |
445 | { | |
446 | struct bond_slave *slave; | |
447 | ||
448 | /* Update link status. */ | |
449 | if (bond->detect == BLSM_CARRIER | |
450 | || time_msec() >= bond->miimon_next_update) | |
451 | { | |
452 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
453 | slave->up = bond_is_link_up(bond, slave->netdev); | |
454 | } | |
455 | bond->miimon_next_update = time_msec() + bond->miimon_interval; | |
456 | } | |
457 | ||
458 | /* Update LACP. */ | |
459 | if (bond->lacp) { | |
460 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
461 | lacp_slave_enable(bond->lacp, slave, slave->enabled); | |
462 | } | |
463 | ||
464 | lacp_run(bond->lacp, bond_send_pdu_cb); | |
465 | } | |
466 | ||
467 | /* Enable slaves based on link status and LACP feedback. */ | |
468 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
469 | bond_link_status_update(slave, tags); | |
470 | } | |
471 | if (!bond->active_slave || !bond->active_slave->enabled) { | |
472 | bond_choose_active_slave(bond, tags); | |
473 | } | |
474 | ||
475 | /* Update fake bond interface stats. */ | |
476 | if (time_msec() >= bond->next_fake_iface_update) { | |
477 | bond_update_fake_slave_stats(bond); | |
478 | bond->next_fake_iface_update = time_msec() + 1000; | |
479 | } | |
480 | ||
481 | /* Invalidate any tags required by */ | |
482 | tag_set_union(tags, &bond->unixctl_tags); | |
483 | tag_set_init(&bond->unixctl_tags); | |
484 | } | |
485 | ||
486 | /* Causes poll_block() to wake up when 'bond' needs something to be done. */ | |
487 | void | |
488 | bond_wait(struct bond *bond) | |
489 | { | |
490 | struct bond_slave *slave; | |
491 | ||
492 | if (bond->detect == BLSM_CARRIER) { | |
493 | netdev_monitor_poll_wait(bond->monitor); | |
494 | } else { | |
495 | poll_timer_wait_until(bond->miimon_next_update); | |
496 | } | |
497 | ||
498 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
499 | if (slave->delay_expires != LLONG_MAX) { | |
500 | poll_timer_wait_until(slave->delay_expires); | |
501 | } | |
502 | } | |
503 | ||
504 | if (bond->next_fake_iface_update != LLONG_MAX) { | |
505 | poll_timer_wait_until(bond->next_fake_iface_update); | |
506 | } | |
507 | ||
508 | /* Ensure that any saved tags get revalidated right away. */ | |
509 | if (!tag_set_is_empty(&bond->unixctl_tags)) { | |
510 | poll_immediate_wake(); | |
511 | } | |
512 | ||
513 | /* We don't wait for bond->next_rebalance because rebalancing can only run | |
514 | * at a flow account checkpoint. ofproto does checkpointing on its own | |
515 | * schedule and bond_rebalance() gets called afterward, so we'd just be | |
516 | * waking up for no purpose. */ | |
517 | } | |
518 | \f | |
519 | /* MAC learning table interaction. */ | |
520 | ||
521 | static bool | |
522 | may_send_learning_packets(const struct bond *bond) | |
523 | { | |
524 | return !lacp_negotiated(bond->lacp) && bond->balance != BM_AB; | |
525 | } | |
526 | ||
527 | /* Returns true if 'bond' needs the client to send out packets to assist with | |
528 | * MAC learning on 'bond'. If this function returns true, then the client | |
529 | * should iterate through its MAC learning table for the bridge on which 'bond' | |
530 | * is located. For each MAC that has been learned on a port other than 'bond', | |
531 | * it should call bond_send_learning_packet(). | |
532 | * | |
533 | * This function will only return true if 'bond' is in SLB mode and LACP is not | |
534 | * negotiated. Otherwise sending learning packets isn't necessary. | |
535 | * | |
536 | * Calling this function resets the state that it checks. */ | |
537 | bool | |
538 | bond_should_send_learning_packets(struct bond *bond) | |
539 | { | |
540 | bool send = bond->send_learning_packets && may_send_learning_packets(bond); | |
541 | bond->send_learning_packets = false; | |
542 | return send; | |
543 | } | |
544 | ||
545 | /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'. | |
546 | * | |
547 | * See bond_should_send_learning_packets() for description of usage. */ | |
548 | int | |
549 | bond_send_learning_packet(struct bond *bond, | |
550 | const uint8_t eth_src[ETH_ADDR_LEN], | |
551 | uint16_t vlan) | |
552 | { | |
553 | struct bond_slave *slave; | |
554 | struct ofpbuf packet; | |
555 | struct flow flow; | |
556 | int error; | |
557 | ||
558 | assert(may_send_learning_packets(bond)); | |
559 | if (!bond->active_slave) { | |
560 | /* Nowhere to send the learning packet. */ | |
561 | return 0; | |
562 | } | |
563 | ||
564 | memset(&flow, 0, sizeof flow); | |
565 | memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN); | |
566 | slave = choose_output_slave(bond, &flow, vlan); | |
567 | ||
568 | ofpbuf_init(&packet, 0); | |
569 | compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177, | |
570 | eth_src); | |
571 | if (vlan) { | |
572 | eth_set_vlan_tci(&packet, htons(vlan)); | |
573 | } | |
574 | error = netdev_send(slave->netdev, &packet); | |
575 | ofpbuf_uninit(&packet); | |
576 | ||
577 | return error; | |
578 | } | |
579 | \f | |
580 | /* Checks whether a packet that arrived on 'slave_' within 'bond', with an | |
581 | * Ethernet destination address of 'eth_dst', should be admitted. | |
582 | * | |
583 | * The return value is one of the following: | |
584 | * | |
585 | * - BV_ACCEPT: Admit the packet. | |
586 | * | |
587 | * - BV_DROP: Drop the packet. | |
588 | * | |
589 | * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's | |
590 | * Ethernet source address and VLAN. If there is none, or if the packet | |
591 | * is on the learned port, then admit the packet. If a different port has | |
592 | * been learned, however, drop the packet (and do not use it for MAC | |
593 | * learning). | |
594 | */ | |
595 | enum bond_verdict | |
596 | bond_check_admissibility(struct bond *bond, const void *slave_, | |
597 | const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags) | |
598 | { | |
599 | /* Admit all packets if LACP has been negotiated, because that means that | |
600 | * the remote switch is aware of the bond and will "do the right thing". */ | |
601 | if (lacp_negotiated(bond->lacp)) { | |
602 | return BV_ACCEPT; | |
603 | } | |
604 | ||
605 | /* Drop all multicast packets on inactive slaves. */ | |
606 | if (eth_addr_is_multicast(eth_dst)) { | |
607 | *tags |= bond_get_active_slave_tag(bond); | |
608 | if (bond->active_slave != bond_slave_lookup(bond, slave_)) { | |
609 | return BV_DROP; | |
610 | } | |
611 | } | |
612 | ||
613 | /* Drop all packets for which we have learned a different input port, | |
614 | * because we probably sent the packet on one slave and got it back on the | |
615 | * other. Gratuitous ARP packets are an exception to this rule: the host | |
616 | * has moved to another switch. The exception to the exception is if we | |
617 | * locked the learning table to avoid reflections on bond slaves. */ | |
618 | return BV_DROP_IF_MOVED; | |
619 | } | |
620 | ||
621 | /* Returns the slave (registered on 'bond' by bond_slave_register()) to which | |
622 | * a packet with the given 'flow' and 'vlan' should be forwarded. Returns | |
623 | * NULL if the packet should be dropped because no slaves are enabled. | |
624 | * | |
625 | * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan' | |
626 | * should be a VID only (i.e. excluding the PCP bits). Second, | |
627 | * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be | |
628 | * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the | |
629 | * packet belongs to (so for an access port it will be the access port's VLAN). | |
630 | * | |
631 | * Adds a tag to '*tags' that associates the flow with the returned slave. | |
632 | */ | |
633 | void * | |
634 | bond_choose_output_slave(struct bond *bond, const struct flow *flow, | |
635 | uint16_t vlan, tag_type *tags) | |
636 | { | |
637 | struct bond_slave *slave = choose_output_slave(bond, flow, vlan); | |
638 | if (slave) { | |
639 | *tags |= slave->tag; | |
640 | return slave->aux; | |
641 | } else { | |
642 | *tags |= bond->no_slaves_tag; | |
643 | return NULL; | |
644 | } | |
645 | } | |
646 | ||
647 | /* Processes LACP packet 'packet', which was received on 'slave_' within | |
648 | * 'bond'. | |
649 | * | |
650 | * The client should use this function to pass along LACP messages received on | |
651 | * any of 'bond''s slaves. */ | |
652 | void | |
653 | bond_process_lacp(struct bond *bond, void *slave_, const struct ofpbuf *packet) | |
654 | { | |
655 | if (bond->lacp) { | |
656 | struct bond_slave *slave = bond_slave_lookup(bond, slave_); | |
657 | const struct lacp_pdu *pdu = parse_lacp_packet(packet); | |
658 | if (slave && pdu) { | |
659 | COVERAGE_INC(bond_process_lacp); | |
660 | lacp_process_pdu(bond->lacp, slave, pdu); | |
661 | } | |
662 | } | |
663 | } | |
664 | \f | |
665 | /* Rebalancing. */ | |
666 | ||
667 | /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */ | |
668 | void | |
669 | bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan, | |
670 | uint64_t n_bytes) | |
671 | { | |
672 | switch (bond->balance) { | |
673 | case BM_AB: | |
674 | /* Nothing to do. */ | |
675 | break; | |
676 | ||
677 | case BM_SLB: | |
678 | case BM_TCP: | |
679 | lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes; | |
680 | break; | |
681 | ||
682 | default: | |
683 | NOT_REACHED(); | |
684 | } | |
685 | } | |
686 | ||
687 | static struct bond_slave * | |
688 | bond_slave_from_bal_node(struct list *bal) | |
689 | { | |
690 | return CONTAINER_OF(bal, struct bond_slave, bal_node); | |
691 | } | |
692 | ||
693 | static void | |
694 | log_bals(struct bond *bond, const struct list *bals) | |
695 | { | |
696 | if (VLOG_IS_DBG_ENABLED()) { | |
697 | struct ds ds = DS_EMPTY_INITIALIZER; | |
698 | const struct bond_slave *slave; | |
699 | ||
700 | LIST_FOR_EACH (slave, bal_node, bals) { | |
701 | if (ds.length) { | |
702 | ds_put_char(&ds, ','); | |
703 | } | |
704 | ds_put_format(&ds, " %s %"PRIu64"kB", | |
705 | slave->name, slave->tx_bytes / 1024); | |
706 | ||
707 | if (!slave->enabled) { | |
708 | ds_put_cstr(&ds, " (disabled)"); | |
709 | } | |
710 | if (!list_is_empty(&slave->entries)) { | |
711 | struct bond_entry *e; | |
712 | ||
713 | ds_put_cstr(&ds, " ("); | |
714 | LIST_FOR_EACH (e, list_node, &slave->entries) { | |
715 | if (&e->list_node != list_front(&slave->entries)) { | |
716 | ds_put_cstr(&ds, " + "); | |
717 | } | |
718 | ds_put_format(&ds, "h%td: %"PRIu64"kB", | |
719 | e - bond->hash, e->tx_bytes / 1024); | |
720 | } | |
721 | ds_put_cstr(&ds, ")"); | |
722 | } | |
723 | } | |
724 | VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds)); | |
725 | ds_destroy(&ds); | |
726 | } | |
727 | } | |
728 | ||
729 | /* Shifts 'hash' from its current slave to 'to'. */ | |
730 | static void | |
731 | bond_shift_load(struct bond_entry *hash, struct bond_slave *to, | |
732 | struct tag_set *set) | |
733 | { | |
734 | struct bond_slave *from = hash->slave; | |
735 | struct bond *bond = from->bond; | |
736 | uint64_t delta = hash->tx_bytes; | |
737 | ||
738 | VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) " | |
739 | "from %s to %s (now carrying %"PRIu64"kB and " | |
740 | "%"PRIu64"kB load, respectively)", | |
741 | bond->name, delta / 1024, hash - bond->hash, | |
742 | from->name, to->name, | |
743 | (from->tx_bytes - delta) / 1024, | |
744 | (to->tx_bytes + delta) / 1024); | |
745 | ||
746 | /* Shift load away from 'from' to 'to'. */ | |
747 | from->tx_bytes -= delta; | |
748 | to->tx_bytes += delta; | |
749 | ||
750 | /* Arrange for flows to be revalidated. */ | |
751 | tag_set_add(set, hash->tag); | |
752 | hash->slave = to; | |
753 | hash->tag = tag_create_random(); | |
754 | } | |
755 | ||
756 | /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave), | |
757 | * given that doing so must decrease the ratio of the load on the two slaves by | |
758 | * at least 0.1. Returns NULL if there is no appropriate entry. | |
759 | * | |
760 | * The list of entries isn't sorted. I don't know of a reason to prefer to | |
761 | * shift away small hashes or large hashes. */ | |
762 | static struct bond_entry * | |
763 | choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes) | |
764 | { | |
765 | struct bond_entry *e; | |
766 | ||
767 | if (list_is_short(&from->entries)) { | |
768 | /* 'from' carries no more than one MAC hash, so shifting load away from | |
769 | * it would be pointless. */ | |
770 | return NULL; | |
771 | } | |
772 | ||
773 | LIST_FOR_EACH (e, list_node, &from->entries) { | |
774 | double old_ratio, new_ratio; | |
775 | uint64_t delta; | |
776 | ||
777 | if (to_tx_bytes == 0) { | |
778 | /* Nothing on the new slave, move it. */ | |
779 | return e; | |
780 | } | |
781 | ||
782 | delta = e->tx_bytes; | |
783 | old_ratio = (double)from->tx_bytes / to_tx_bytes; | |
784 | new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta); | |
785 | if (old_ratio - new_ratio > 0.1) { | |
786 | /* Would decrease the ratio, move it. */ | |
787 | return e; | |
788 | } | |
789 | } | |
790 | ||
791 | return NULL; | |
792 | } | |
793 | ||
794 | /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is | |
795 | * maintained. */ | |
796 | static void | |
797 | insert_bal(struct list *bals, struct bond_slave *slave) | |
798 | { | |
799 | struct bond_slave *pos; | |
800 | ||
801 | LIST_FOR_EACH (pos, bal_node, bals) { | |
802 | if (slave->tx_bytes > pos->tx_bytes) { | |
803 | break; | |
804 | } | |
805 | } | |
806 | list_insert(&pos->bal_node, &slave->bal_node); | |
807 | } | |
808 | ||
809 | /* Removes 'slave' from its current list and then inserts it into 'bals' so | |
810 | * that descending order of 'tx_bytes' is maintained. */ | |
811 | static void | |
812 | reinsert_bal(struct list *bals, struct bond_slave *slave) | |
813 | { | |
814 | list_remove(&slave->bal_node); | |
815 | insert_bal(bals, slave); | |
816 | } | |
817 | ||
818 | /* If 'bond' needs rebalancing, does so. | |
819 | * | |
820 | * The caller should have called bond_account() for each active flow, to ensure | |
821 | * that flow data is consistently accounted at this point. */ | |
822 | void | |
823 | bond_rebalance(struct bond *bond, struct tag_set *tags) | |
824 | { | |
825 | struct bond_slave *slave; | |
826 | struct bond_entry *e; | |
827 | struct list bals; | |
828 | ||
829 | if (bond->balance == BM_AB || time_msec() < bond->next_rebalance) { | |
830 | return; | |
831 | } | |
832 | bond->next_rebalance = time_msec() + bond->rebalance_interval; | |
833 | ||
834 | /* Add each bond_entry to its slave's 'entries' list. | |
835 | * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */ | |
836 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
837 | slave->tx_bytes = 0; | |
838 | list_init(&slave->entries); | |
839 | } | |
840 | for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) { | |
841 | if (e->slave && e->tx_bytes) { | |
842 | e->slave->tx_bytes += e->tx_bytes; | |
843 | list_push_back(&e->slave->entries, &e->list_node); | |
844 | } | |
845 | } | |
846 | ||
847 | /* Add enabled slaves to 'bals' in descending order of tx_bytes. | |
848 | * | |
849 | * XXX This is O(n**2) in the number of slaves but it could be O(n lg n) | |
850 | * with a proper list sort algorithm. */ | |
851 | list_init(&bals); | |
852 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
853 | if (slave->enabled) { | |
854 | insert_bal(&bals, slave); | |
855 | } | |
856 | } | |
857 | log_bals(bond, &bals); | |
858 | ||
859 | /* Shift load from the most-loaded slaves to the least-loaded slaves. */ | |
860 | while (!list_is_short(&bals)) { | |
861 | struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals)); | |
862 | struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals)); | |
863 | uint64_t overload; | |
864 | ||
865 | overload = from->tx_bytes - to->tx_bytes; | |
866 | if (overload < to->tx_bytes >> 5 || overload < 100000) { | |
867 | /* The extra load on 'from' (and all less-loaded slaves), compared | |
868 | * to that of 'to' (the least-loaded slave), is less than ~3%, or | |
869 | * it is less than ~1Mbps. No point in rebalancing. */ | |
870 | break; | |
871 | } | |
872 | ||
873 | /* 'from' is carrying significantly more load than 'to', and that load | |
874 | * is split across at least two different hashes. */ | |
875 | e = choose_entry_to_migrate(from, to->tx_bytes); | |
876 | if (e) { | |
877 | bond_shift_load(e, to, tags); | |
878 | ||
879 | /* Delete element from from->entries. | |
880 | * | |
881 | * We don't add the element to to->hashes. That would only allow | |
882 | * 'e' to be migrated to another slave in this rebalancing run, and | |
883 | * there is no point in doing that. */ | |
884 | list_remove(&e->list_node); | |
885 | ||
886 | /* Re-sort 'bals'. */ | |
887 | reinsert_bal(&bals, from); | |
888 | reinsert_bal(&bals, to); | |
889 | } else { | |
890 | /* Can't usefully migrate anything away from 'from'. | |
891 | * Don't reconsider it. */ | |
892 | list_remove(&from->bal_node); | |
893 | } | |
894 | } | |
895 | ||
896 | /* Implement exponentially weighted moving average. A weight of 1/2 causes | |
897 | * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes | |
898 | * take 20 rebalancing runs to decay to 0 and get deleted entirely. */ | |
899 | for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) { | |
900 | e->tx_bytes /= 2; | |
901 | if (!e->tx_bytes) { | |
902 | e->slave = NULL; | |
903 | } | |
904 | } | |
905 | } | |
906 | \f | |
907 | /* Bonding unixctl user interface functions. */ | |
908 | ||
909 | static struct bond * | |
910 | bond_find(const char *name) | |
911 | { | |
912 | struct bond *bond; | |
913 | ||
914 | HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0), | |
915 | &all_bonds) { | |
916 | if (!strcmp(bond->name, name)) { | |
917 | return bond; | |
918 | } | |
919 | } | |
920 | return NULL; | |
921 | } | |
922 | ||
923 | static struct bond_slave * | |
924 | bond_lookup_slave(struct bond *bond, const char *slave_name) | |
925 | { | |
926 | struct bond_slave *slave; | |
927 | ||
928 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
929 | if (!strcmp(slave->name, slave_name)) { | |
930 | return slave; | |
931 | } | |
932 | } | |
933 | return NULL; | |
934 | } | |
935 | ||
936 | static void | |
937 | bond_unixctl_list(struct unixctl_conn *conn, | |
938 | const char *args OVS_UNUSED, void *aux OVS_UNUSED) | |
939 | { | |
940 | struct ds ds = DS_EMPTY_INITIALIZER; | |
941 | const struct bond *bond; | |
942 | ||
943 | ds_put_cstr(&ds, "bond\ttype\tslaves\n"); | |
944 | ||
945 | HMAP_FOR_EACH (bond, hmap_node, &all_bonds) { | |
946 | const struct bond_slave *slave; | |
947 | size_t i; | |
948 | ||
949 | ds_put_format(&ds, "%s\t%s\t", | |
950 | bond->name, bond_mode_to_string(bond->balance)); | |
951 | ||
952 | i = 0; | |
953 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
954 | if (i++ > 0) { | |
955 | ds_put_cstr(&ds, ", "); | |
956 | } | |
957 | ds_put_cstr(&ds, slave->name); | |
958 | } | |
959 | ds_put_char(&ds, '\n'); | |
960 | } | |
961 | unixctl_command_reply(conn, 200, ds_cstr(&ds)); | |
962 | ds_destroy(&ds); | |
963 | } | |
964 | ||
965 | static void | |
966 | bond_unixctl_show(struct unixctl_conn *conn, | |
967 | const char *args, void *aux OVS_UNUSED) | |
968 | { | |
969 | struct ds ds = DS_EMPTY_INITIALIZER; | |
970 | const struct bond_slave *slave; | |
971 | const struct bond *bond; | |
972 | ||
973 | bond = bond_find(args); | |
974 | if (!bond) { | |
975 | unixctl_command_reply(conn, 501, "no such bond"); | |
976 | return; | |
977 | } | |
978 | ||
979 | ds_put_format(&ds, "bond_mode: %s\n", | |
980 | bond_mode_to_string(bond->balance)); | |
981 | ||
982 | if (bond->lacp) { | |
983 | ds_put_format(&ds, "lacp: %s\n", | |
984 | lacp_is_active(bond->lacp) ? "active" : "passive"); | |
985 | } else { | |
986 | ds_put_cstr(&ds, "lacp: off\n"); | |
987 | } | |
988 | ||
989 | if (bond->balance != BM_AB) { | |
990 | ds_put_format(&ds, "bond-hash-algorithm: %s\n", | |
991 | bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb"); | |
992 | } | |
993 | ||
994 | ds_put_format(&ds, "bond-detect-mode: %s\n", | |
995 | bond->monitor ? "carrier" : "miimon"); | |
996 | ||
997 | if (!bond->monitor) { | |
998 | ds_put_format(&ds, "bond-miimon-interval: %lld\n", | |
999 | bond->miimon_interval); | |
1000 | } | |
1001 | ||
1002 | ds_put_format(&ds, "updelay: %d ms\n", bond->updelay); | |
1003 | ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay); | |
1004 | ||
1005 | if (bond->balance != BM_AB) { | |
1006 | ds_put_format(&ds, "next rebalance: %lld ms\n", | |
1007 | bond->next_rebalance - time_msec()); | |
1008 | } | |
1009 | ||
1010 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
1011 | struct bond_entry *be; | |
1012 | struct flow flow; | |
1013 | ||
1014 | /* Basic info. */ | |
1015 | ds_put_format(&ds, "\nslave %s: %s\n", | |
1016 | slave->name, slave->enabled ? "enabled" : "disabled"); | |
1017 | if (slave == bond->active_slave) { | |
1018 | ds_put_cstr(&ds, "\tactive slave\n"); | |
1019 | } | |
1020 | if (slave->delay_expires != LLONG_MAX) { | |
1021 | ds_put_format(&ds, "\t%s expires in %lld ms\n", | |
1022 | slave->enabled ? "downdelay" : "updelay", | |
1023 | slave->delay_expires - time_msec()); | |
1024 | } | |
1025 | ||
1026 | if (bond->balance == BM_AB) { | |
1027 | continue; | |
1028 | } | |
1029 | ||
1030 | /* Hashes. */ | |
1031 | memset(&flow, 0, sizeof flow); | |
1032 | for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) { | |
1033 | int hash = be - bond->hash; | |
1034 | ||
1035 | if (be->slave != slave) { | |
1036 | continue; | |
1037 | } | |
1038 | ||
1039 | ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n", | |
1040 | hash, be->tx_bytes / 1024); | |
1041 | ||
1042 | if (bond->balance != BM_SLB) { | |
1043 | continue; | |
1044 | } | |
1045 | ||
1046 | /* XXX How can we list the MACs assigned to hashes? */ | |
1047 | } | |
1048 | } | |
1049 | unixctl_command_reply(conn, 200, ds_cstr(&ds)); | |
1050 | ds_destroy(&ds); | |
1051 | } | |
1052 | ||
1053 | static void | |
1054 | bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_, | |
1055 | void *aux OVS_UNUSED) | |
1056 | { | |
1057 | char *args = (char *) args_; | |
1058 | char *save_ptr = NULL; | |
1059 | char *bond_s, *hash_s, *slave_s; | |
1060 | struct bond *bond; | |
1061 | struct bond_slave *slave; | |
1062 | struct bond_entry *entry; | |
1063 | int hash; | |
1064 | ||
1065 | bond_s = strtok_r(args, " ", &save_ptr); | |
1066 | hash_s = strtok_r(NULL, " ", &save_ptr); | |
1067 | slave_s = strtok_r(NULL, " ", &save_ptr); | |
1068 | if (!slave_s) { | |
1069 | unixctl_command_reply(conn, 501, | |
1070 | "usage: bond/migrate BOND HASH SLAVE"); | |
1071 | return; | |
1072 | } | |
1073 | ||
1074 | bond = bond_find(bond_s); | |
1075 | if (!bond) { | |
1076 | unixctl_command_reply(conn, 501, "no such bond"); | |
1077 | return; | |
1078 | } | |
1079 | ||
1080 | if (bond->balance != BM_SLB) { | |
1081 | unixctl_command_reply(conn, 501, "not an SLB bond"); | |
1082 | return; | |
1083 | } | |
1084 | ||
1085 | if (strspn(hash_s, "0123456789") == strlen(hash_s)) { | |
1086 | hash = atoi(hash_s) & BOND_MASK; | |
1087 | } else { | |
1088 | unixctl_command_reply(conn, 501, "bad hash"); | |
1089 | return; | |
1090 | } | |
1091 | ||
1092 | slave = bond_lookup_slave(bond, slave_s); | |
1093 | if (!slave) { | |
1094 | unixctl_command_reply(conn, 501, "no such slave"); | |
1095 | return; | |
1096 | } | |
1097 | ||
1098 | if (!slave->enabled) { | |
1099 | unixctl_command_reply(conn, 501, "cannot migrate to disabled slave"); | |
1100 | return; | |
1101 | } | |
1102 | ||
1103 | entry = &bond->hash[hash]; | |
1104 | tag_set_add(&bond->unixctl_tags, entry->tag); | |
1105 | entry->slave = slave; | |
1106 | entry->tag = tag_create_random(); | |
1107 | unixctl_command_reply(conn, 200, "migrated"); | |
1108 | } | |
1109 | ||
1110 | static void | |
1111 | bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_, | |
1112 | void *aux OVS_UNUSED) | |
1113 | { | |
1114 | char *args = (char *) args_; | |
1115 | char *save_ptr = NULL; | |
1116 | char *bond_s, *slave_s; | |
1117 | struct bond *bond; | |
1118 | struct bond_slave *slave; | |
1119 | ||
1120 | bond_s = strtok_r(args, " ", &save_ptr); | |
1121 | slave_s = strtok_r(NULL, " ", &save_ptr); | |
1122 | if (!slave_s) { | |
1123 | unixctl_command_reply(conn, 501, | |
1124 | "usage: bond/set-active-slave BOND SLAVE"); | |
1125 | return; | |
1126 | } | |
1127 | ||
1128 | bond = bond_find(bond_s); | |
1129 | if (!bond) { | |
1130 | unixctl_command_reply(conn, 501, "no such bond"); | |
1131 | return; | |
1132 | } | |
1133 | ||
1134 | slave = bond_lookup_slave(bond, slave_s); | |
1135 | if (!slave) { | |
1136 | unixctl_command_reply(conn, 501, "no such slave"); | |
1137 | return; | |
1138 | } | |
1139 | ||
1140 | if (!slave->enabled) { | |
1141 | unixctl_command_reply(conn, 501, "cannot make disabled slave active"); | |
1142 | return; | |
1143 | } | |
1144 | ||
1145 | if (bond->active_slave != slave) { | |
1146 | tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond)); | |
1147 | bond->active_slave = slave; | |
1148 | bond->active_slave->tag = tag_create_random(); | |
1149 | VLOG_INFO("bond %s: active interface is now %s", | |
1150 | bond->name, slave->name); | |
1151 | bond->send_learning_packets = true; | |
1152 | unixctl_command_reply(conn, 200, "done"); | |
1153 | } else { | |
1154 | unixctl_command_reply(conn, 200, "no change"); | |
1155 | } | |
1156 | } | |
1157 | ||
1158 | static void | |
1159 | enable_slave(struct unixctl_conn *conn, const char *args_, bool enable) | |
1160 | { | |
1161 | char *args = (char *) args_; | |
1162 | char *save_ptr = NULL; | |
1163 | char *bond_s, *slave_s; | |
1164 | struct bond *bond; | |
1165 | struct bond_slave *slave; | |
1166 | ||
1167 | bond_s = strtok_r(args, " ", &save_ptr); | |
1168 | slave_s = strtok_r(NULL, " ", &save_ptr); | |
1169 | if (!slave_s) { | |
1170 | char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE", | |
1171 | enable ? "enable" : "disable"); | |
1172 | unixctl_command_reply(conn, 501, usage); | |
1173 | free(usage); | |
1174 | return; | |
1175 | } | |
1176 | ||
1177 | bond = bond_find(bond_s); | |
1178 | if (!bond) { | |
1179 | unixctl_command_reply(conn, 501, "no such bond"); | |
1180 | return; | |
1181 | } | |
1182 | ||
1183 | slave = bond_lookup_slave(bond, slave_s); | |
1184 | if (!slave) { | |
1185 | unixctl_command_reply(conn, 501, "no such slave"); | |
1186 | return; | |
1187 | } | |
1188 | ||
1189 | bond_enable_slave(slave, enable, &bond->unixctl_tags); | |
1190 | unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled"); | |
1191 | } | |
1192 | ||
1193 | static void | |
1194 | bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args, | |
1195 | void *aux OVS_UNUSED) | |
1196 | { | |
1197 | enable_slave(conn, args, true); | |
1198 | } | |
1199 | ||
1200 | static void | |
1201 | bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args, | |
1202 | void *aux OVS_UNUSED) | |
1203 | { | |
1204 | enable_slave(conn, args, false); | |
1205 | } | |
1206 | ||
1207 | static void | |
1208 | bond_unixctl_hash(struct unixctl_conn *conn, const char *args_, | |
1209 | void *aux OVS_UNUSED) | |
1210 | { | |
1211 | char *args = (char *) args_; | |
1212 | uint8_t mac[ETH_ADDR_LEN]; | |
1213 | uint8_t hash; | |
1214 | char *hash_cstr; | |
1215 | unsigned int vlan; | |
1216 | char *mac_s, *vlan_s; | |
1217 | char *save_ptr = NULL; | |
1218 | ||
1219 | mac_s = strtok_r(args, " ", &save_ptr); | |
1220 | vlan_s = strtok_r(NULL, " ", &save_ptr); | |
1221 | ||
1222 | if (vlan_s) { | |
1223 | if (sscanf(vlan_s, "%u", &vlan) != 1) { | |
1224 | unixctl_command_reply(conn, 501, "invalid vlan"); | |
1225 | return; | |
1226 | } | |
1227 | } else { | |
1228 | vlan = OFP_VLAN_NONE; | |
1229 | } | |
1230 | ||
1231 | if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac)) | |
1232 | == ETH_ADDR_SCAN_COUNT) { | |
1233 | hash = bond_hash_src(mac, vlan) & BOND_MASK; | |
1234 | ||
1235 | hash_cstr = xasprintf("%u", hash); | |
1236 | unixctl_command_reply(conn, 200, hash_cstr); | |
1237 | free(hash_cstr); | |
1238 | } else { | |
1239 | unixctl_command_reply(conn, 501, "invalid mac"); | |
1240 | } | |
1241 | } | |
1242 | ||
1243 | void | |
1244 | bond_init(void) | |
1245 | { | |
1246 | lacp_init(); | |
1247 | ||
1248 | unixctl_command_register("bond/list", bond_unixctl_list, NULL); | |
1249 | unixctl_command_register("bond/show", bond_unixctl_show, NULL); | |
1250 | unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL); | |
1251 | unixctl_command_register("bond/set-active-slave", | |
1252 | bond_unixctl_set_active_slave, NULL); | |
1253 | unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave, | |
1254 | NULL); | |
1255 | unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave, | |
1256 | NULL); | |
1257 | unixctl_command_register("bond/hash", bond_unixctl_hash, NULL); | |
1258 | } | |
1259 | \f | |
1260 | static struct bond_slave * | |
1261 | bond_slave_lookup(struct bond *bond, const void *slave_) | |
1262 | { | |
1263 | struct bond_slave *slave; | |
1264 | ||
1265 | HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0), | |
1266 | &bond->slaves) { | |
1267 | if (slave->aux == slave_) { | |
1268 | return slave; | |
1269 | } | |
1270 | } | |
1271 | ||
1272 | return NULL; | |
1273 | } | |
1274 | ||
1275 | static bool | |
1276 | bond_is_link_up(struct bond *bond, struct netdev *netdev) | |
1277 | { | |
1278 | return (bond->detect == BLSM_CARRIER | |
1279 | ? netdev_get_carrier(netdev) | |
1280 | : netdev_get_miimon(netdev)); | |
1281 | } | |
1282 | ||
1283 | static void | |
1284 | bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags) | |
1285 | { | |
1286 | slave->delay_expires = LLONG_MAX; | |
1287 | if (enable != slave->enabled) { | |
1288 | slave->enabled = enable; | |
1289 | if (!slave->enabled) { | |
1290 | VLOG_WARN("interface %s: disabled", slave->name); | |
1291 | tag_set_add(tags, slave->tag); | |
1292 | } else { | |
1293 | VLOG_WARN("interface %s: enabled", slave->name); | |
1294 | slave->tag = tag_create_random(); | |
1295 | } | |
1296 | } | |
1297 | } | |
1298 | ||
1299 | static void | |
1300 | bond_link_status_update(struct bond_slave *slave, struct tag_set *tags) | |
1301 | { | |
1302 | struct bond *bond = slave->bond; | |
1303 | bool up; | |
1304 | ||
1305 | up = slave->up && lacp_slave_may_enable(bond->lacp, slave); | |
1306 | if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) { | |
1307 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20); | |
1308 | VLOG_INFO_RL(&rl, "interface %s: link state %s", | |
1309 | slave->name, up ? "up" : "down"); | |
1310 | if (up == slave->enabled) { | |
1311 | slave->delay_expires = LLONG_MAX; | |
1312 | VLOG_INFO_RL(&rl, "interface %s: will not be %s", | |
1313 | slave->name, up ? "disabled" : "enabled"); | |
1314 | } else { | |
1315 | int delay = (lacp_negotiated(bond->lacp) ? 0 | |
1316 | : up ? bond->updelay : bond->downdelay); | |
1317 | slave->delay_expires = time_msec() + delay; | |
1318 | if (delay) { | |
1319 | VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s " | |
1320 | "for %d ms", | |
1321 | slave->name, | |
1322 | up ? "enabled" : "disabled", | |
1323 | up ? "up" : "down", | |
1324 | delay); | |
1325 | } | |
1326 | } | |
1327 | } | |
1328 | ||
1329 | if (time_msec() >= slave->delay_expires) { | |
1330 | bond_enable_slave(slave, up, tags); | |
1331 | } | |
1332 | } | |
1333 | ||
1334 | static bool | |
1335 | bond_is_tcp_hash(const struct bond *bond) | |
1336 | { | |
1337 | return bond->balance == BM_TCP && lacp_negotiated(bond->lacp); | |
1338 | } | |
1339 | ||
1340 | static unsigned int | |
1341 | bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan) | |
1342 | { | |
1343 | return hash_bytes(mac, ETH_ADDR_LEN, vlan); | |
1344 | } | |
1345 | ||
1346 | static unsigned int | |
1347 | bond_hash_tcp(const struct flow *flow, uint16_t vlan) | |
1348 | { | |
1349 | struct flow hash_flow = *flow; | |
1350 | hash_flow.vlan_tci = vlan; | |
1351 | ||
1352 | /* The symmetric quality of this hash function is not required, but | |
1353 | * flow_hash_symmetric_l4 already exists, and is sufficient for our | |
1354 | * purposes, so we use it out of convenience. */ | |
1355 | return flow_hash_symmetric_l4(&hash_flow, 0); | |
1356 | } | |
1357 | ||
1358 | static struct bond_entry * | |
1359 | lookup_bond_entry(const struct bond *bond, const struct flow *flow, | |
1360 | uint16_t vlan) | |
1361 | { | |
1362 | assert(bond->balance != BM_AB); | |
1363 | return &bond->hash[(bond_is_tcp_hash(bond) | |
1364 | ? bond_hash_tcp(flow, vlan) | |
1365 | : bond_hash_src(flow->dl_src, vlan)) & BOND_MASK]; | |
1366 | } | |
1367 | ||
1368 | static struct bond_slave * | |
1369 | choose_output_slave(const struct bond *bond, const struct flow *flow, | |
1370 | uint16_t vlan) | |
1371 | { | |
1372 | struct bond_entry *e; | |
1373 | ||
1374 | switch (bond->balance) { | |
1375 | case BM_AB: | |
1376 | return bond->active_slave; | |
1377 | ||
1378 | case BM_SLB: | |
1379 | case BM_TCP: | |
1380 | e = lookup_bond_entry(bond, flow, vlan); | |
1381 | if (!e->slave || !e->slave->enabled) { | |
1382 | /* XXX select interface properly. The current interface selection | |
1383 | * is only good for testing the rebalancing code. */ | |
1384 | e->slave = bond->active_slave; | |
1385 | e->tag = tag_create_random(); | |
1386 | } | |
1387 | return e->slave; | |
1388 | ||
1389 | default: | |
1390 | NOT_REACHED(); | |
1391 | } | |
1392 | } | |
1393 | ||
1394 | static struct bond_slave * | |
1395 | bond_choose_slave(const struct bond *bond) | |
1396 | { | |
1397 | struct bond_slave *slave, *best; | |
1398 | ||
1399 | /* Find an enabled slave. */ | |
1400 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
1401 | if (slave->enabled) { | |
1402 | return slave; | |
1403 | } | |
1404 | } | |
1405 | ||
1406 | /* All interfaces are disabled. Find an interface that will be enabled | |
1407 | * after its updelay expires. */ | |
1408 | best = NULL; | |
1409 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
1410 | if (slave->delay_expires != LLONG_MAX | |
1411 | && lacp_slave_may_enable(bond->lacp, slave) | |
1412 | && (!best || slave->delay_expires < best->delay_expires)) { | |
1413 | best = slave; | |
1414 | } | |
1415 | } | |
1416 | return best; | |
1417 | } | |
1418 | ||
1419 | static void | |
1420 | bond_choose_active_slave(struct bond *bond, struct tag_set *tags) | |
1421 | { | |
1422 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20); | |
1423 | struct bond_slave *old_active_slave = bond->active_slave; | |
1424 | ||
1425 | bond->active_slave = bond_choose_slave(bond); | |
1426 | if (bond->active_slave) { | |
1427 | if (bond->active_slave->enabled) { | |
1428 | VLOG_INFO_RL(&rl, "bond %s: active interface is now %s", | |
1429 | bond->name, bond->active_slave->name); | |
1430 | } else { | |
1431 | VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping " | |
1432 | "remaining %lld ms updelay (since no interface was " | |
1433 | "enabled)", bond->name, bond->active_slave->name, | |
1434 | bond->active_slave->delay_expires - time_msec()); | |
1435 | bond_enable_slave(bond->active_slave, true, tags); | |
1436 | } | |
1437 | ||
1438 | if (!old_active_slave) { | |
1439 | tag_set_add(tags, bond->no_slaves_tag); | |
1440 | } | |
1441 | ||
1442 | bond->send_learning_packets = true; | |
1443 | } else if (old_active_slave) { | |
1444 | VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name); | |
1445 | } | |
1446 | } | |
1447 | ||
1448 | /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if | |
1449 | * there is no active slave. */ | |
1450 | static tag_type | |
1451 | bond_get_active_slave_tag(const struct bond *bond) | |
1452 | { | |
1453 | return (bond->active_slave | |
1454 | ? bond->active_slave->tag | |
1455 | : bond->no_slaves_tag); | |
1456 | } | |
1457 | ||
1458 | /* Attempts to make the sum of the bond slaves' statistics appear on the fake | |
1459 | * bond interface. */ | |
1460 | static void | |
1461 | bond_update_fake_slave_stats(struct bond *bond) | |
1462 | { | |
1463 | struct netdev_stats bond_stats; | |
1464 | struct bond_slave *slave; | |
1465 | struct netdev *bond_dev; | |
1466 | ||
1467 | memset(&bond_stats, 0, sizeof bond_stats); | |
1468 | ||
1469 | HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { | |
1470 | struct netdev_stats slave_stats; | |
1471 | ||
1472 | if (!netdev_get_stats(slave->netdev, &slave_stats)) { | |
1473 | /* XXX: We swap the stats here because they are swapped back when | |
1474 | * reported by the internal device. The reason for this is | |
1475 | * internal devices normally represent packets going into the | |
1476 | * system but when used as fake bond device they represent packets | |
1477 | * leaving the system. We really should do this in the internal | |
1478 | * device itself because changing it here reverses the counts from | |
1479 | * the perspective of the switch. However, the internal device | |
1480 | * doesn't know what type of device it represents so we have to do | |
1481 | * it here for now. */ | |
1482 | bond_stats.tx_packets += slave_stats.rx_packets; | |
1483 | bond_stats.tx_bytes += slave_stats.rx_bytes; | |
1484 | bond_stats.rx_packets += slave_stats.tx_packets; | |
1485 | bond_stats.rx_bytes += slave_stats.tx_bytes; | |
1486 | } | |
1487 | } | |
1488 | ||
1489 | if (!netdev_open_default(bond->name, &bond_dev)) { | |
1490 | netdev_set_stats(bond_dev, &bond_stats); | |
1491 | netdev_close(bond_dev); | |
1492 | } | |
1493 | } |