blob: 7bbeb621b7fecad9e2592a38c8945e0b9bbd554a [file] [log] [blame]
Willy Tarreauf89c1872009-10-01 11:19:37 +02001/*
2 * Fast Weighted Least Connection load balancing algorithm.
3 *
4 * Copyright 2000-2009 Willy Tarreau <w@1wt.eu>
5 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
10 *
11 */
12
Willy Tarreau1e56f922020-06-04 23:20:13 +020013#include <import/eb32tree.h>
Willy Tarreau4c7e4b72020-05-27 12:58:42 +020014#include <haproxy/api.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020015#include <haproxy/backend.h>
Willy Tarreaua55c4542020-06-04 22:59:39 +020016#include <haproxy/queue.h>
Willy Tarreau1e56f922020-06-04 23:20:13 +020017#include <haproxy/server-t.h>
Willy Tarreauf89c1872009-10-01 11:19:37 +020018
19
20/* Remove a server from a tree. It must have previously been dequeued. This
21 * function is meant to be called when a server is going down or has its
22 * weight disabled.
Willy Tarreau1b877482018-08-21 19:44:53 +020023 *
24 * The server's lock and the lbprm's lock must be held.
Willy Tarreauf89c1872009-10-01 11:19:37 +020025 */
26static inline void fwlc_remove_from_tree(struct server *s)
27{
28 s->lb_tree = NULL;
29}
30
Willy Tarreau1b877482018-08-21 19:44:53 +020031/* simply removes a server from a tree.
32 *
Willy Tarreau17807ed2021-02-17 16:14:00 +010033 * The lbprm's lock must be held.
Willy Tarreau1b877482018-08-21 19:44:53 +020034 */
Willy Tarreauf89c1872009-10-01 11:19:37 +020035static inline void fwlc_dequeue_srv(struct server *s)
36{
37 eb32_delete(&s->lb_node);
38}
39
Christopher Faulet08818422020-12-11 15:36:01 +010040/* Queue a server in its associated tree, assuming the <eweight> is >0.
Willy Tarreau1eb6c552018-12-14 08:33:28 +010041 * Servers are sorted by (#conns+1)/weight. To ensure maximum accuracy,
42 * we use (#conns+1)*SRV_EWGHT_MAX/eweight as the sorting key. The reason
43 * for using #conns+1 is to sort by weights in case the server is picked
44 * and not before it is picked. This provides a better load accuracy for
45 * low connection counts when weights differ and makes sure the round-robin
Willy Tarreaued5ac9c2019-09-06 17:04:04 +020046 * applies between servers of highest weight first. However servers with no
47 * connection are always picked first so that under low loads, it's not
48 * always the single server with the highest weight that gets picked.
Willy Tarreau1b877482018-08-21 19:44:53 +020049 *
Christopher Faulet08818422020-12-11 15:36:01 +010050 * NOTE: Depending on the calling context, we use s->next_eweight or
51 * s->cur_eweight. The next value is used when the server state is updated
52 * (because the weight changed for instance). During this step, the server
53 * state is not yet committed. The current value is used to reposition the
54 * server in the tree. This happens when the server is used.
55 *
Willy Tarreau17807ed2021-02-17 16:14:00 +010056 * The lbprm's lock must be held.
Willy Tarreauf89c1872009-10-01 11:19:37 +020057 */
Christopher Faulet08818422020-12-11 15:36:01 +010058static inline void fwlc_queue_srv(struct server *s, unsigned int eweight)
Willy Tarreauf89c1872009-10-01 11:19:37 +020059{
Willy Tarreau17807ed2021-02-17 16:14:00 +010060 unsigned int inflight = _HA_ATOMIC_LOAD(&s->served) + _HA_ATOMIC_LOAD(&s->nbpend);
Willy Tarreau8c855f62020-10-22 17:41:45 +020061
Christopher Faulet08818422020-12-11 15:36:01 +010062 s->lb_node.key = inflight ? (inflight + 1) * SRV_EWGHT_MAX / eweight : 0;
Willy Tarreauf89c1872009-10-01 11:19:37 +020063 eb32_insert(s->lb_tree, &s->lb_node);
64}
65
66/* Re-position the server in the FWLC tree after it has been assigned one
67 * connection or after it has released one. Note that it is possible that
68 * the server has been moved out of the tree due to failed health-checks.
Willy Tarreau1b877482018-08-21 19:44:53 +020069 *
Willy Tarreau2b437002021-02-17 16:01:37 +010070 * <locked> must reflect the server's lock ownership. The lbprm's lock will
71 * be used.
Willy Tarreauf89c1872009-10-01 11:19:37 +020072 */
Willy Tarreau2b437002021-02-17 16:01:37 +010073static void fwlc_srv_reposition(struct server *s, int locked)
Willy Tarreauf89c1872009-10-01 11:19:37 +020074{
Willy Tarreaucd10def2020-10-17 18:48:47 +020075 HA_RWLOCK_WRLOCK(LBPRM_LOCK, &s->proxy->lbprm.lock);
Christopher Faulet1ae2a882019-06-19 10:50:38 +020076 if (s->lb_tree) {
77 fwlc_dequeue_srv(s);
Christopher Faulet08818422020-12-11 15:36:01 +010078 fwlc_queue_srv(s, s->cur_eweight);
Christopher Faulet1ae2a882019-06-19 10:50:38 +020079 }
Willy Tarreaucd10def2020-10-17 18:48:47 +020080 HA_RWLOCK_WRUNLOCK(LBPRM_LOCK, &s->proxy->lbprm.lock);
Willy Tarreauf89c1872009-10-01 11:19:37 +020081}
82
83/* This function updates the server trees according to server <srv>'s new
84 * state. It should be called when server <srv>'s status changes to down.
85 * It is not important whether the server was already down or not. It is not
86 * important either that the new state is completely down (the caller may not
87 * know all the variables of a server's state).
Willy Tarreau1b877482018-08-21 19:44:53 +020088 *
89 * The server's lock must be held. The lbprm's lock will be used.
Willy Tarreauf89c1872009-10-01 11:19:37 +020090 */
91static void fwlc_set_server_status_down(struct server *srv)
92{
93 struct proxy *p = srv->proxy;
94
Willy Tarreauc5150da2014-05-13 19:27:31 +020095 if (!srv_lb_status_changed(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +020096 return;
97
Emeric Brun52a91d32017-08-31 14:41:55 +020098 if (srv_willbe_usable(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +020099 goto out_update_state;
Willy Tarreaucd10def2020-10-17 18:48:47 +0200100 HA_RWLOCK_WRLOCK(LBPRM_LOCK, &p->lbprm.lock);
Willy Tarreau1b877482018-08-21 19:44:53 +0200101
Willy Tarreauf89c1872009-10-01 11:19:37 +0200102
Emeric Brun52a91d32017-08-31 14:41:55 +0200103 if (!srv_currently_usable(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +0200104 /* server was already down */
105 goto out_update_backend;
106
Willy Tarreauc93cd162014-05-13 15:54:22 +0200107 if (srv->flags & SRV_F_BACKUP) {
Emeric Brun52a91d32017-08-31 14:41:55 +0200108 p->lbprm.tot_wbck -= srv->cur_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200109 p->srv_bck--;
110
111 if (srv == p->lbprm.fbck) {
112 /* we lost the first backup server in a single-backup
113 * configuration, we must search another one.
114 */
115 struct server *srv2 = p->lbprm.fbck;
116 do {
117 srv2 = srv2->next;
118 } while (srv2 &&
Willy Tarreauc93cd162014-05-13 15:54:22 +0200119 !((srv2->flags & SRV_F_BACKUP) &&
Emeric Brun52a91d32017-08-31 14:41:55 +0200120 srv_willbe_usable(srv2)));
Willy Tarreauf89c1872009-10-01 11:19:37 +0200121 p->lbprm.fbck = srv2;
122 }
123 } else {
Emeric Brun52a91d32017-08-31 14:41:55 +0200124 p->lbprm.tot_wact -= srv->cur_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200125 p->srv_act--;
126 }
127
128 fwlc_dequeue_srv(srv);
129 fwlc_remove_from_tree(srv);
130
131out_update_backend:
132 /* check/update tot_used, tot_weight */
133 update_backend_weight(p);
Willy Tarreaucd10def2020-10-17 18:48:47 +0200134 HA_RWLOCK_WRUNLOCK(LBPRM_LOCK, &p->lbprm.lock);
Willy Tarreau1b877482018-08-21 19:44:53 +0200135
Willy Tarreauf89c1872009-10-01 11:19:37 +0200136 out_update_state:
Willy Tarreauc5150da2014-05-13 19:27:31 +0200137 srv_lb_commit_status(srv);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200138}
139
140/* This function updates the server trees according to server <srv>'s new
141 * state. It should be called when server <srv>'s status changes to up.
142 * It is not important whether the server was already down or not. It is not
143 * important either that the new state is completely UP (the caller may not
144 * know all the variables of a server's state). This function will not change
145 * the weight of a server which was already up.
Willy Tarreau1b877482018-08-21 19:44:53 +0200146 *
147 * The server's lock must be held. The lbprm's lock will be used.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200148 */
149static void fwlc_set_server_status_up(struct server *srv)
150{
151 struct proxy *p = srv->proxy;
152
Willy Tarreauc5150da2014-05-13 19:27:31 +0200153 if (!srv_lb_status_changed(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +0200154 return;
155
Emeric Brun52a91d32017-08-31 14:41:55 +0200156 if (!srv_willbe_usable(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +0200157 goto out_update_state;
158
Willy Tarreaucd10def2020-10-17 18:48:47 +0200159 HA_RWLOCK_WRLOCK(LBPRM_LOCK, &p->lbprm.lock);
Willy Tarreau1b877482018-08-21 19:44:53 +0200160
Emeric Brun52a91d32017-08-31 14:41:55 +0200161 if (srv_currently_usable(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +0200162 /* server was already up */
163 goto out_update_backend;
164
Willy Tarreauc93cd162014-05-13 15:54:22 +0200165 if (srv->flags & SRV_F_BACKUP) {
Willy Tarreauf89c1872009-10-01 11:19:37 +0200166 srv->lb_tree = &p->lbprm.fwlc.bck;
Emeric Brun52a91d32017-08-31 14:41:55 +0200167 p->lbprm.tot_wbck += srv->next_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200168 p->srv_bck++;
169
170 if (!(p->options & PR_O_USE_ALL_BK)) {
171 if (!p->lbprm.fbck) {
172 /* there was no backup server anymore */
173 p->lbprm.fbck = srv;
174 } else {
175 /* we may have restored a backup server prior to fbck,
176 * in which case it should replace it.
177 */
178 struct server *srv2 = srv;
179 do {
180 srv2 = srv2->next;
181 } while (srv2 && (srv2 != p->lbprm.fbck));
182 if (srv2)
183 p->lbprm.fbck = srv;
184 }
185 }
186 } else {
187 srv->lb_tree = &p->lbprm.fwlc.act;
Emeric Brun52a91d32017-08-31 14:41:55 +0200188 p->lbprm.tot_wact += srv->next_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200189 p->srv_act++;
190 }
191
192 /* note that eweight cannot be 0 here */
Christopher Faulet08818422020-12-11 15:36:01 +0100193 fwlc_queue_srv(srv, srv->next_eweight);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200194
195 out_update_backend:
196 /* check/update tot_used, tot_weight */
197 update_backend_weight(p);
Willy Tarreaucd10def2020-10-17 18:48:47 +0200198 HA_RWLOCK_WRUNLOCK(LBPRM_LOCK, &p->lbprm.lock);
Willy Tarreau1b877482018-08-21 19:44:53 +0200199
Willy Tarreauf89c1872009-10-01 11:19:37 +0200200 out_update_state:
Willy Tarreauc5150da2014-05-13 19:27:31 +0200201 srv_lb_commit_status(srv);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200202}
203
204/* This function must be called after an update to server <srv>'s effective
205 * weight. It may be called after a state change too.
Willy Tarreau1b877482018-08-21 19:44:53 +0200206 *
207 * The server's lock must be held. The lbprm's lock will be used.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200208 */
209static void fwlc_update_server_weight(struct server *srv)
210{
211 int old_state, new_state;
212 struct proxy *p = srv->proxy;
213
Willy Tarreauc5150da2014-05-13 19:27:31 +0200214 if (!srv_lb_status_changed(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +0200215 return;
216
217 /* If changing the server's weight changes its state, we simply apply
218 * the procedures we already have for status change. If the state
219 * remains down, the server is not in any tree, so it's as easy as
220 * updating its values. If the state remains up with different weights,
221 * there are some computations to perform to find a new place and
222 * possibly a new tree for this server.
223 */
224
Emeric Brun52a91d32017-08-31 14:41:55 +0200225 old_state = srv_currently_usable(srv);
226 new_state = srv_willbe_usable(srv);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200227
228 if (!old_state && !new_state) {
Willy Tarreauc5150da2014-05-13 19:27:31 +0200229 srv_lb_commit_status(srv);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200230 return;
231 }
232 else if (!old_state && new_state) {
233 fwlc_set_server_status_up(srv);
234 return;
235 }
236 else if (old_state && !new_state) {
237 fwlc_set_server_status_down(srv);
238 return;
239 }
240
Willy Tarreaucd10def2020-10-17 18:48:47 +0200241 HA_RWLOCK_WRLOCK(LBPRM_LOCK, &p->lbprm.lock);
Willy Tarreau1b877482018-08-21 19:44:53 +0200242
Willy Tarreauf89c1872009-10-01 11:19:37 +0200243 if (srv->lb_tree)
244 fwlc_dequeue_srv(srv);
245
Willy Tarreauc93cd162014-05-13 15:54:22 +0200246 if (srv->flags & SRV_F_BACKUP) {
Emeric Brun52a91d32017-08-31 14:41:55 +0200247 p->lbprm.tot_wbck += srv->next_eweight - srv->cur_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200248 srv->lb_tree = &p->lbprm.fwlc.bck;
249 } else {
Emeric Brun52a91d32017-08-31 14:41:55 +0200250 p->lbprm.tot_wact += srv->next_eweight - srv->cur_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200251 srv->lb_tree = &p->lbprm.fwlc.act;
252 }
253
Christopher Faulet08818422020-12-11 15:36:01 +0100254 fwlc_queue_srv(srv, srv->next_eweight);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200255
256 update_backend_weight(p);
Willy Tarreaucd10def2020-10-17 18:48:47 +0200257 HA_RWLOCK_WRUNLOCK(LBPRM_LOCK, &p->lbprm.lock);
Willy Tarreau1b877482018-08-21 19:44:53 +0200258
Willy Tarreauc5150da2014-05-13 19:27:31 +0200259 srv_lb_commit_status(srv);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200260}
261
262/* This function is responsible for building the trees in case of fast
263 * weighted least-conns. It also sets p->lbprm.wdiv to the eweight to
264 * uweight ratio. Both active and backup groups are initialized.
265 */
266void fwlc_init_server_tree(struct proxy *p)
267{
268 struct server *srv;
269 struct eb_root init_head = EB_ROOT;
270
271 p->lbprm.set_server_status_up = fwlc_set_server_status_up;
272 p->lbprm.set_server_status_down = fwlc_set_server_status_down;
273 p->lbprm.update_server_eweight = fwlc_update_server_weight;
274 p->lbprm.server_take_conn = fwlc_srv_reposition;
275 p->lbprm.server_drop_conn = fwlc_srv_reposition;
276
277 p->lbprm.wdiv = BE_WEIGHT_SCALE;
278 for (srv = p->srv; srv; srv = srv->next) {
Emeric Brun52a91d32017-08-31 14:41:55 +0200279 srv->next_eweight = (srv->uweight * p->lbprm.wdiv + p->lbprm.wmult - 1) / p->lbprm.wmult;
Willy Tarreauc5150da2014-05-13 19:27:31 +0200280 srv_lb_commit_status(srv);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200281 }
282
283 recount_servers(p);
284 update_backend_weight(p);
285
286 p->lbprm.fwlc.act = init_head;
287 p->lbprm.fwlc.bck = init_head;
288
289 /* queue active and backup servers in two distinct groups */
290 for (srv = p->srv; srv; srv = srv->next) {
Emeric Brun52a91d32017-08-31 14:41:55 +0200291 if (!srv_currently_usable(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +0200292 continue;
Willy Tarreauc93cd162014-05-13 15:54:22 +0200293 srv->lb_tree = (srv->flags & SRV_F_BACKUP) ? &p->lbprm.fwlc.bck : &p->lbprm.fwlc.act;
Christopher Faulet08818422020-12-11 15:36:01 +0100294 fwlc_queue_srv(srv, srv->next_eweight);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200295 }
296}
297
298/* Return next server from the FWLC tree in backend <p>. If the tree is empty,
299 * return NULL. Saturated servers are skipped.
Willy Tarreau1b877482018-08-21 19:44:53 +0200300 *
301 * The server's lock must be held. The lbprm's lock will be used.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200302 */
303struct server *fwlc_get_next_server(struct proxy *p, struct server *srvtoavoid)
304{
305 struct server *srv, *avoided;
306 struct eb32_node *node;
307
308 srv = avoided = NULL;
309
Willy Tarreau58bc9c12020-10-17 19:32:09 +0200310 HA_RWLOCK_RDLOCK(LBPRM_LOCK, &p->lbprm.lock);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200311 if (p->srv_act)
312 node = eb32_first(&p->lbprm.fwlc.act);
Christopher Faulet5b517552017-06-09 14:17:53 +0200313 else if (p->lbprm.fbck) {
314 srv = p->lbprm.fbck;
315 goto out;
316 }
Willy Tarreauf89c1872009-10-01 11:19:37 +0200317 else if (p->srv_bck)
318 node = eb32_first(&p->lbprm.fwlc.bck);
Christopher Faulet5b517552017-06-09 14:17:53 +0200319 else {
320 srv = NULL;
321 goto out;
322 }
Willy Tarreauf89c1872009-10-01 11:19:37 +0200323
324 while (node) {
325 /* OK, we have a server. However, it may be saturated, in which
326 * case we don't want to reconsider it for now, so we'll simply
327 * skip it. Same if it's the server we try to avoid, in which
328 * case we simply remember it for later use if needed.
329 */
330 struct server *s;
331
332 s = eb32_entry(node, struct server, lb_node);
Willy Tarreau8ae8c482020-10-22 17:19:07 +0200333 if (!s->maxconn || s->served + s->nbpend < srv_dynamic_maxconn(s) + s->maxqueue) {
Willy Tarreauf89c1872009-10-01 11:19:37 +0200334 if (s != srvtoavoid) {
335 srv = s;
336 break;
337 }
338 avoided = s;
339 }
340 node = eb32_next(node);
341 }
342
343 if (!srv)
344 srv = avoided;
Christopher Faulet5b517552017-06-09 14:17:53 +0200345 out:
Willy Tarreau58bc9c12020-10-17 19:32:09 +0200346 HA_RWLOCK_RDUNLOCK(LBPRM_LOCK, &p->lbprm.lock);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200347 return srv;
348}
349
350
351/*
352 * Local variables:
353 * c-indent-level: 8
354 * c-basic-offset: 8
355 * End:
356 */