blob: 82d76233804368e14f9f253266262508bad1d027 [file] [log] [blame]
Willy Tarreauf89c1872009-10-01 11:19:37 +02001/*
2 * Fast Weighted Round Robin 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
13#include <common/compat.h>
14#include <common/config.h>
15#include <common/debug.h>
Willy Tarreau45cb4fb2009-10-26 21:10:04 +010016#include <eb32tree.h>
Willy Tarreauf89c1872009-10-01 11:19:37 +020017
18#include <types/global.h>
19#include <types/server.h>
20
21#include <proto/backend.h>
22#include <proto/queue.h>
23
24static inline void fwrr_remove_from_tree(struct server *s);
25static inline void fwrr_queue_by_weight(struct eb_root *root, struct server *s);
26static inline void fwrr_dequeue_srv(struct server *s);
27static void fwrr_get_srv(struct server *s);
28static void fwrr_queue_srv(struct server *s);
29
30
31/* This function updates the server trees according to server <srv>'s new
32 * state. It should be called when server <srv>'s status changes to down.
33 * It is not important whether the server was already down or not. It is not
34 * important either that the new state is completely down (the caller may not
35 * know all the variables of a server's state).
Willy Tarreau1b877482018-08-21 19:44:53 +020036 *
37 * The server's lock must be held. The lbprm's lock will be used.
Willy Tarreauf89c1872009-10-01 11:19:37 +020038 */
39static void fwrr_set_server_status_down(struct server *srv)
40{
41 struct proxy *p = srv->proxy;
42 struct fwrr_group *grp;
43
Willy Tarreauc5150da2014-05-13 19:27:31 +020044 if (!srv_lb_status_changed(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +020045 return;
46
Emeric Brun52a91d32017-08-31 14:41:55 +020047 if (srv_willbe_usable(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +020048 goto out_update_state;
49
Willy Tarreau1b877482018-08-21 19:44:53 +020050 HA_SPIN_LOCK(LBPRM_LOCK, &p->lbprm.lock);
51
Emeric Brun52a91d32017-08-31 14:41:55 +020052 if (!srv_currently_usable(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +020053 /* server was already down */
54 goto out_update_backend;
55
Willy Tarreauc93cd162014-05-13 15:54:22 +020056 grp = (srv->flags & SRV_F_BACKUP) ? &p->lbprm.fwrr.bck : &p->lbprm.fwrr.act;
Emeric Brun52a91d32017-08-31 14:41:55 +020057 grp->next_weight -= srv->cur_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +020058
Willy Tarreauc93cd162014-05-13 15:54:22 +020059 if (srv->flags & SRV_F_BACKUP) {
Willy Tarreauf89c1872009-10-01 11:19:37 +020060 p->lbprm.tot_wbck = p->lbprm.fwrr.bck.next_weight;
61 p->srv_bck--;
62
63 if (srv == p->lbprm.fbck) {
64 /* we lost the first backup server in a single-backup
65 * configuration, we must search another one.
66 */
67 struct server *srv2 = p->lbprm.fbck;
68 do {
69 srv2 = srv2->next;
70 } while (srv2 &&
Willy Tarreauc93cd162014-05-13 15:54:22 +020071 !((srv2->flags & SRV_F_BACKUP) &&
Emeric Brun52a91d32017-08-31 14:41:55 +020072 srv_willbe_usable(srv2)));
Willy Tarreauf89c1872009-10-01 11:19:37 +020073 p->lbprm.fbck = srv2;
74 }
75 } else {
76 p->lbprm.tot_wact = p->lbprm.fwrr.act.next_weight;
77 p->srv_act--;
78 }
79
80 fwrr_dequeue_srv(srv);
81 fwrr_remove_from_tree(srv);
82
83out_update_backend:
84 /* check/update tot_used, tot_weight */
85 update_backend_weight(p);
Willy Tarreau1b877482018-08-21 19:44:53 +020086 HA_SPIN_UNLOCK(LBPRM_LOCK, &p->lbprm.lock);
87
Willy Tarreauf89c1872009-10-01 11:19:37 +020088 out_update_state:
Willy Tarreauc5150da2014-05-13 19:27:31 +020089 srv_lb_commit_status(srv);
Willy Tarreauf89c1872009-10-01 11:19:37 +020090}
91
92/* This function updates the server trees according to server <srv>'s new
93 * state. It should be called when server <srv>'s status changes to up.
94 * It is not important whether the server was already down or not. It is not
95 * important either that the new state is completely UP (the caller may not
96 * know all the variables of a server's state). This function will not change
97 * the weight of a server which was already up.
Willy Tarreau1b877482018-08-21 19:44:53 +020098 *
99 * The server's lock must be held. The lbprm's lock will be used.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200100 */
101static void fwrr_set_server_status_up(struct server *srv)
102{
103 struct proxy *p = srv->proxy;
104 struct fwrr_group *grp;
105
Willy Tarreauc5150da2014-05-13 19:27:31 +0200106 if (!srv_lb_status_changed(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +0200107 return;
108
Emeric Brun52a91d32017-08-31 14:41:55 +0200109 if (!srv_willbe_usable(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +0200110 goto out_update_state;
111
Willy Tarreau1b877482018-08-21 19:44:53 +0200112 HA_SPIN_LOCK(LBPRM_LOCK, &p->lbprm.lock);
113
Emeric Brun52a91d32017-08-31 14:41:55 +0200114 if (srv_currently_usable(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +0200115 /* server was already up */
116 goto out_update_backend;
117
Willy Tarreauc93cd162014-05-13 15:54:22 +0200118 grp = (srv->flags & SRV_F_BACKUP) ? &p->lbprm.fwrr.bck : &p->lbprm.fwrr.act;
Emeric Brun52a91d32017-08-31 14:41:55 +0200119 grp->next_weight += srv->next_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200120
Willy Tarreauc93cd162014-05-13 15:54:22 +0200121 if (srv->flags & SRV_F_BACKUP) {
Willy Tarreauf89c1872009-10-01 11:19:37 +0200122 p->lbprm.tot_wbck = p->lbprm.fwrr.bck.next_weight;
123 p->srv_bck++;
124
125 if (!(p->options & PR_O_USE_ALL_BK)) {
126 if (!p->lbprm.fbck) {
127 /* there was no backup server anymore */
128 p->lbprm.fbck = srv;
129 } else {
130 /* we may have restored a backup server prior to fbck,
131 * in which case it should replace it.
132 */
133 struct server *srv2 = srv;
134 do {
135 srv2 = srv2->next;
136 } while (srv2 && (srv2 != p->lbprm.fbck));
137 if (srv2)
138 p->lbprm.fbck = srv;
139 }
140 }
141 } else {
142 p->lbprm.tot_wact = p->lbprm.fwrr.act.next_weight;
143 p->srv_act++;
144 }
145
146 /* note that eweight cannot be 0 here */
147 fwrr_get_srv(srv);
Emeric Brun52a91d32017-08-31 14:41:55 +0200148 srv->npos = grp->curr_pos + (grp->next_weight + grp->curr_weight - grp->curr_pos) / srv->next_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200149 fwrr_queue_srv(srv);
150
151out_update_backend:
152 /* check/update tot_used, tot_weight */
153 update_backend_weight(p);
Willy Tarreau1b877482018-08-21 19:44:53 +0200154 HA_SPIN_UNLOCK(LBPRM_LOCK, &p->lbprm.lock);
155
Willy Tarreauf89c1872009-10-01 11:19:37 +0200156 out_update_state:
Willy Tarreauc5150da2014-05-13 19:27:31 +0200157 srv_lb_commit_status(srv);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200158}
159
160/* This function must be called after an update to server <srv>'s effective
161 * weight. It may be called after a state change too.
Willy Tarreau1b877482018-08-21 19:44:53 +0200162 *
163 * The server's lock must be held. The lbprm's lock will be used.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200164 */
165static void fwrr_update_server_weight(struct server *srv)
166{
167 int old_state, new_state;
168 struct proxy *p = srv->proxy;
169 struct fwrr_group *grp;
170
Willy Tarreauc5150da2014-05-13 19:27:31 +0200171 if (!srv_lb_status_changed(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +0200172 return;
173
174 /* If changing the server's weight changes its state, we simply apply
175 * the procedures we already have for status change. If the state
176 * remains down, the server is not in any tree, so it's as easy as
177 * updating its values. If the state remains up with different weights,
178 * there are some computations to perform to find a new place and
179 * possibly a new tree for this server.
180 */
181
Emeric Brun52a91d32017-08-31 14:41:55 +0200182 old_state = srv_currently_usable(srv);
183 new_state = srv_willbe_usable(srv);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200184
185 if (!old_state && !new_state) {
Willy Tarreauc5150da2014-05-13 19:27:31 +0200186 srv_lb_commit_status(srv);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200187 return;
188 }
189 else if (!old_state && new_state) {
190 fwrr_set_server_status_up(srv);
191 return;
192 }
193 else if (old_state && !new_state) {
194 fwrr_set_server_status_down(srv);
195 return;
196 }
197
Willy Tarreau1b877482018-08-21 19:44:53 +0200198 HA_SPIN_LOCK(LBPRM_LOCK, &p->lbprm.lock);
199
Willy Tarreauc93cd162014-05-13 15:54:22 +0200200 grp = (srv->flags & SRV_F_BACKUP) ? &p->lbprm.fwrr.bck : &p->lbprm.fwrr.act;
Emeric Brun52a91d32017-08-31 14:41:55 +0200201 grp->next_weight = grp->next_weight - srv->cur_eweight + srv->next_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200202
203 p->lbprm.tot_wact = p->lbprm.fwrr.act.next_weight;
204 p->lbprm.tot_wbck = p->lbprm.fwrr.bck.next_weight;
205
206 if (srv->lb_tree == grp->init) {
207 fwrr_dequeue_srv(srv);
208 fwrr_queue_by_weight(grp->init, srv);
209 }
210 else if (!srv->lb_tree) {
211 /* FIXME: server was down. This is not possible right now but
212 * may be needed soon for slowstart or graceful shutdown.
213 */
214 fwrr_dequeue_srv(srv);
215 fwrr_get_srv(srv);
Emeric Brun52a91d32017-08-31 14:41:55 +0200216 srv->npos = grp->curr_pos + (grp->next_weight + grp->curr_weight - grp->curr_pos) / srv->next_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200217 fwrr_queue_srv(srv);
218 } else {
219 /* The server is either active or in the next queue. If it's
220 * still in the active queue and it has not consumed all of its
221 * places, let's adjust its next position.
222 */
223 fwrr_get_srv(srv);
224
Emeric Brun52a91d32017-08-31 14:41:55 +0200225 if (srv->next_eweight > 0) {
Willy Tarreauf89c1872009-10-01 11:19:37 +0200226 int prev_next = srv->npos;
Emeric Brun52a91d32017-08-31 14:41:55 +0200227 int step = grp->next_weight / srv->next_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200228
229 srv->npos = srv->lpos + step;
230 srv->rweight = 0;
231
232 if (srv->npos > prev_next)
233 srv->npos = prev_next;
234 if (srv->npos < grp->curr_pos + 2)
235 srv->npos = grp->curr_pos + step;
236 } else {
237 /* push it into the next tree */
238 srv->npos = grp->curr_pos + grp->curr_weight;
239 }
240
241 fwrr_dequeue_srv(srv);
242 fwrr_queue_srv(srv);
243 }
244
245 update_backend_weight(p);
Willy Tarreau1b877482018-08-21 19:44:53 +0200246 HA_SPIN_UNLOCK(LBPRM_LOCK, &p->lbprm.lock);
247
Willy Tarreauc5150da2014-05-13 19:27:31 +0200248 srv_lb_commit_status(srv);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200249}
250
251/* Remove a server from a tree. It must have previously been dequeued. This
252 * function is meant to be called when a server is going down or has its
253 * weight disabled.
Willy Tarreau1b877482018-08-21 19:44:53 +0200254 *
255 * The server's lock and the lbprm's lock must be held.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200256 */
257static inline void fwrr_remove_from_tree(struct server *s)
258{
259 s->lb_tree = NULL;
260}
261
262/* Queue a server in the weight tree <root>, assuming the weight is >0.
263 * We want to sort them by inverted weights, because we need to place
264 * heavy servers first in order to get a smooth distribution.
Willy Tarreau1b877482018-08-21 19:44:53 +0200265 *
266 * The server's lock and the lbprm's lock must be held.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200267 */
268static inline void fwrr_queue_by_weight(struct eb_root *root, struct server *s)
269{
Emeric Brun52a91d32017-08-31 14:41:55 +0200270 s->lb_node.key = SRV_EWGHT_MAX - s->next_eweight;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200271 eb32_insert(root, &s->lb_node);
272 s->lb_tree = root;
273}
274
275/* This function is responsible for building the weight trees in case of fast
276 * weighted round-robin. It also sets p->lbprm.wdiv to the eweight to uweight
277 * ratio. Both active and backup groups are initialized.
278 */
279void fwrr_init_server_groups(struct proxy *p)
280{
281 struct server *srv;
282 struct eb_root init_head = EB_ROOT;
283
284 p->lbprm.set_server_status_up = fwrr_set_server_status_up;
285 p->lbprm.set_server_status_down = fwrr_set_server_status_down;
286 p->lbprm.update_server_eweight = fwrr_update_server_weight;
287
288 p->lbprm.wdiv = BE_WEIGHT_SCALE;
289 for (srv = p->srv; srv; srv = srv->next) {
Emeric Brun52a91d32017-08-31 14:41:55 +0200290 srv->next_eweight = (srv->uweight * p->lbprm.wdiv + p->lbprm.wmult - 1) / p->lbprm.wmult;
Willy Tarreauc5150da2014-05-13 19:27:31 +0200291 srv_lb_commit_status(srv);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200292 }
293
294 recount_servers(p);
295 update_backend_weight(p);
296
297 /* prepare the active servers group */
298 p->lbprm.fwrr.act.curr_pos = p->lbprm.fwrr.act.curr_weight =
299 p->lbprm.fwrr.act.next_weight = p->lbprm.tot_wact;
300 p->lbprm.fwrr.act.curr = p->lbprm.fwrr.act.t0 =
301 p->lbprm.fwrr.act.t1 = init_head;
302 p->lbprm.fwrr.act.init = &p->lbprm.fwrr.act.t0;
303 p->lbprm.fwrr.act.next = &p->lbprm.fwrr.act.t1;
304
305 /* prepare the backup servers group */
306 p->lbprm.fwrr.bck.curr_pos = p->lbprm.fwrr.bck.curr_weight =
307 p->lbprm.fwrr.bck.next_weight = p->lbprm.tot_wbck;
308 p->lbprm.fwrr.bck.curr = p->lbprm.fwrr.bck.t0 =
309 p->lbprm.fwrr.bck.t1 = init_head;
310 p->lbprm.fwrr.bck.init = &p->lbprm.fwrr.bck.t0;
311 p->lbprm.fwrr.bck.next = &p->lbprm.fwrr.bck.t1;
312
313 /* queue active and backup servers in two distinct groups */
314 for (srv = p->srv; srv; srv = srv->next) {
Emeric Brun52a91d32017-08-31 14:41:55 +0200315 if (!srv_currently_usable(srv))
Willy Tarreauf89c1872009-10-01 11:19:37 +0200316 continue;
Willy Tarreauc93cd162014-05-13 15:54:22 +0200317 fwrr_queue_by_weight((srv->flags & SRV_F_BACKUP) ?
Willy Tarreauf89c1872009-10-01 11:19:37 +0200318 p->lbprm.fwrr.bck.init :
319 p->lbprm.fwrr.act.init,
320 srv);
321 }
322}
323
Willy Tarreau1b877482018-08-21 19:44:53 +0200324/* simply removes a server from a weight tree.
325 *
326 * The server's lock and the lbprm's lock must be held.
327 */
Willy Tarreauf89c1872009-10-01 11:19:37 +0200328static inline void fwrr_dequeue_srv(struct server *s)
329{
330 eb32_delete(&s->lb_node);
331}
332
333/* queues a server into the appropriate group and tree depending on its
334 * backup status, and ->npos. If the server is disabled, simply assign
335 * it to the NULL tree.
Willy Tarreau1b877482018-08-21 19:44:53 +0200336 *
337 * The server's lock and the lbprm's lock must be held.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200338 */
339static void fwrr_queue_srv(struct server *s)
340{
341 struct proxy *p = s->proxy;
342 struct fwrr_group *grp;
343
Willy Tarreauc93cd162014-05-13 15:54:22 +0200344 grp = (s->flags & SRV_F_BACKUP) ? &p->lbprm.fwrr.bck : &p->lbprm.fwrr.act;
Christopher Faulet5b517552017-06-09 14:17:53 +0200345
Willy Tarreauf89c1872009-10-01 11:19:37 +0200346 /* Delay everything which does not fit into the window and everything
347 * which does not fit into the theorical new window.
348 */
Emeric Brun52a91d32017-08-31 14:41:55 +0200349 if (!srv_willbe_usable(s)) {
Willy Tarreauf89c1872009-10-01 11:19:37 +0200350 fwrr_remove_from_tree(s);
351 }
Emeric Brun52a91d32017-08-31 14:41:55 +0200352 else if (s->next_eweight <= 0 ||
Willy Tarreauf89c1872009-10-01 11:19:37 +0200353 s->npos >= 2 * grp->curr_weight ||
354 s->npos >= grp->curr_weight + grp->next_weight) {
355 /* put into next tree, and readjust npos in case we could
356 * finally take this back to current. */
Olivier Houchard36a8e6f2019-03-08 18:52:46 +0100357 _HA_ATOMIC_SUB(&s->npos, grp->curr_weight);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200358 fwrr_queue_by_weight(grp->next, s);
359 }
360 else {
361 /* The sorting key is stored in units of s->npos * user_weight
362 * in order to avoid overflows. As stated in backend.h, the
363 * lower the scale, the rougher the weights modulation, and the
364 * higher the scale, the lower the number of servers without
365 * overflow. With this formula, the result is always positive,
Godbacha34bdc02013-07-22 07:44:53 +0800366 * so we can use eb32_insert().
Willy Tarreauf89c1872009-10-01 11:19:37 +0200367 */
368 s->lb_node.key = SRV_UWGHT_RANGE * s->npos +
Emeric Brun52a91d32017-08-31 14:41:55 +0200369 (unsigned)(SRV_EWGHT_MAX + s->rweight - s->next_eweight) / BE_WEIGHT_SCALE;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200370
371 eb32_insert(&grp->curr, &s->lb_node);
372 s->lb_tree = &grp->curr;
373 }
374}
375
Willy Tarreau1b877482018-08-21 19:44:53 +0200376/* prepares a server when extracting it from the "init" tree.
377 *
378 * The server's lock and the lbprm's lock must be held.
379 */
Willy Tarreauf89c1872009-10-01 11:19:37 +0200380static inline void fwrr_get_srv_init(struct server *s)
381{
382 s->npos = s->rweight = 0;
383}
384
Willy Tarreau1b877482018-08-21 19:44:53 +0200385/* prepares a server when extracting it from the "next" tree.
386 *
387 * The server's lock and the lbprm's lock must be held.
388 */
Willy Tarreauf89c1872009-10-01 11:19:37 +0200389static inline void fwrr_get_srv_next(struct server *s)
390{
Willy Tarreauc93cd162014-05-13 15:54:22 +0200391 struct fwrr_group *grp = (s->flags & SRV_F_BACKUP) ?
Willy Tarreauf89c1872009-10-01 11:19:37 +0200392 &s->proxy->lbprm.fwrr.bck :
393 &s->proxy->lbprm.fwrr.act;
394
Olivier Houchard36a8e6f2019-03-08 18:52:46 +0100395 _HA_ATOMIC_ADD(&s->npos, grp->curr_weight);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200396}
397
Willy Tarreau1b877482018-08-21 19:44:53 +0200398/* prepares a server when it was marked down.
399 *
400 * The server's lock and the lbprm's lock must be held.
401 */
Willy Tarreauf89c1872009-10-01 11:19:37 +0200402static inline void fwrr_get_srv_down(struct server *s)
403{
Willy Tarreauc93cd162014-05-13 15:54:22 +0200404 struct fwrr_group *grp = (s->flags & SRV_F_BACKUP) ?
Willy Tarreauf89c1872009-10-01 11:19:37 +0200405 &s->proxy->lbprm.fwrr.bck :
406 &s->proxy->lbprm.fwrr.act;
407
408 s->npos = grp->curr_pos;
409}
410
Willy Tarreau1b877482018-08-21 19:44:53 +0200411/* prepares a server when extracting it from its tree.
412 *
413 * The server's lock and the lbprm's lock must be held.
414 */
Willy Tarreauf89c1872009-10-01 11:19:37 +0200415static void fwrr_get_srv(struct server *s)
416{
417 struct proxy *p = s->proxy;
Willy Tarreauc93cd162014-05-13 15:54:22 +0200418 struct fwrr_group *grp = (s->flags & SRV_F_BACKUP) ?
Willy Tarreauf89c1872009-10-01 11:19:37 +0200419 &p->lbprm.fwrr.bck :
420 &p->lbprm.fwrr.act;
421
422 if (s->lb_tree == grp->init) {
423 fwrr_get_srv_init(s);
424 }
425 else if (s->lb_tree == grp->next) {
426 fwrr_get_srv_next(s);
427 }
428 else if (s->lb_tree == NULL) {
429 fwrr_get_srv_down(s);
430 }
431}
432
433/* switches trees "init" and "next" for FWRR group <grp>. "init" should be empty
434 * when this happens, and "next" filled with servers sorted by weights.
Willy Tarreau1b877482018-08-21 19:44:53 +0200435 *
436 * The lbprm's lock must be held.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200437 */
438static inline void fwrr_switch_trees(struct fwrr_group *grp)
439{
440 struct eb_root *swap;
441 swap = grp->init;
442 grp->init = grp->next;
443 grp->next = swap;
444 grp->curr_weight = grp->next_weight;
445 grp->curr_pos = grp->curr_weight;
446}
447
448/* return next server from the current tree in FWRR group <grp>, or a server
449 * from the "init" tree if appropriate. If both trees are empty, return NULL.
Willy Tarreau1b877482018-08-21 19:44:53 +0200450 *
Willy Tarreau9df86f92019-04-16 11:21:14 +0200451 * The lbprm's lock must be held. The server's lock will be held on return if
452 * a non-null server is returned.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200453 */
454static struct server *fwrr_get_server_from_group(struct fwrr_group *grp)
455{
Willy Tarreau9df86f92019-04-16 11:21:14 +0200456 struct eb32_node *node1;
457 struct eb32_node *node2;
458 struct server *s1 = NULL;
459 struct server *s2 = NULL;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200460
Willy Tarreau9df86f92019-04-16 11:21:14 +0200461 node1 = eb32_first(&grp->curr);
462 if (node1) {
463 s1 = eb32_entry(node1, struct server, lb_node);
464 HA_SPIN_LOCK(SERVER_LOCK, &s1->lock);
465 if (s1->cur_eweight && s1->npos <= grp->curr_pos)
466 return s1;
467 }
Christopher Faulet5b517552017-06-09 14:17:53 +0200468
Willy Tarreau9df86f92019-04-16 11:21:14 +0200469 /* Either we have no server left, or we have a hole. We'll look in the
470 * init tree or a better proposal. At this point, if <s1> is non-null,
471 * it is locked.
472 */
473 node2 = eb32_first(grp->init);
474 if (node2) {
475 s2 = eb32_entry(node2, struct server, lb_node);
476 HA_SPIN_LOCK(SERVER_LOCK, &s2->lock);
477 if (s2->cur_eweight) {
478 if (s1)
479 HA_SPIN_UNLOCK(SERVER_LOCK, &s1->lock);
480 fwrr_get_srv_init(s2);
481 return s2;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200482 }
483 }
Willy Tarreau9df86f92019-04-16 11:21:14 +0200484 return s1;
Willy Tarreauf89c1872009-10-01 11:19:37 +0200485}
486
487/* Computes next position of server <s> in the group. It is mandatory for <s>
488 * to have a non-zero, positive eweight.
Willy Tarreau1b877482018-08-21 19:44:53 +0200489 *
490 * The server's lock and the lbprm's lock must be held.
491 */
Willy Tarreauf89c1872009-10-01 11:19:37 +0200492static inline void fwrr_update_position(struct fwrr_group *grp, struct server *s)
493{
494 if (!s->npos) {
495 /* first time ever for this server */
496 s->lpos = grp->curr_pos;
Emeric Brun52a91d32017-08-31 14:41:55 +0200497 s->npos = grp->curr_pos + grp->next_weight / s->cur_eweight;
Olivier Houchard36a8e6f2019-03-08 18:52:46 +0100498 _HA_ATOMIC_ADD(&s->rweight, (grp->next_weight % s->cur_eweight));
Willy Tarreauf89c1872009-10-01 11:19:37 +0200499
Emeric Brun52a91d32017-08-31 14:41:55 +0200500 if (s->rweight >= s->cur_eweight) {
Olivier Houchard36a8e6f2019-03-08 18:52:46 +0100501 _HA_ATOMIC_SUB(&s->rweight, s->cur_eweight);
502 _HA_ATOMIC_ADD(&s->npos, 1);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200503 }
504 } else {
505 s->lpos = s->npos;
Olivier Houchard36a8e6f2019-03-08 18:52:46 +0100506 _HA_ATOMIC_ADD(&s->npos, (grp->next_weight / s->cur_eweight));
507 _HA_ATOMIC_ADD(&s->rweight, (grp->next_weight % s->cur_eweight));
Willy Tarreauf89c1872009-10-01 11:19:37 +0200508
Emeric Brun52a91d32017-08-31 14:41:55 +0200509 if (s->rweight >= s->cur_eweight) {
Olivier Houchard36a8e6f2019-03-08 18:52:46 +0100510 _HA_ATOMIC_SUB(&s->rweight, s->cur_eweight);
511 _HA_ATOMIC_ADD(&s->npos, 1);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200512 }
513 }
514}
515
516/* Return next server from the current tree in backend <p>, or a server from
517 * the init tree if appropriate. If both trees are empty, return NULL.
518 * Saturated servers are skipped and requeued.
Willy Tarreau1b877482018-08-21 19:44:53 +0200519 *
Willy Tarreau9df86f92019-04-16 11:21:14 +0200520 * The lbprm's lock and the server's lock will be used.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200521 */
522struct server *fwrr_get_next_server(struct proxy *p, struct server *srvtoavoid)
523{
524 struct server *srv, *full, *avoided;
525 struct fwrr_group *grp;
526 int switched;
527
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100528 HA_SPIN_LOCK(LBPRM_LOCK, &p->lbprm.lock);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200529 if (p->srv_act)
530 grp = &p->lbprm.fwrr.act;
Christopher Faulet5b517552017-06-09 14:17:53 +0200531 else if (p->lbprm.fbck) {
532 srv = p->lbprm.fbck;
533 goto out;
534 }
Willy Tarreauf89c1872009-10-01 11:19:37 +0200535 else if (p->srv_bck)
536 grp = &p->lbprm.fwrr.bck;
Christopher Faulet5b517552017-06-09 14:17:53 +0200537 else {
538 srv = NULL;
539 goto out;
540 }
Willy Tarreauf89c1872009-10-01 11:19:37 +0200541
542 switched = 0;
543 avoided = NULL;
544 full = NULL; /* NULL-terminated list of saturated servers */
545 while (1) {
546 /* if we see an empty group, let's first try to collect weights
547 * which might have recently changed.
548 */
549 if (!grp->curr_weight)
550 grp->curr_pos = grp->curr_weight = grp->next_weight;
551
552 /* get first server from the "current" tree. When the end of
553 * the tree is reached, we may have to switch, but only once.
554 */
555 while (1) {
556 srv = fwrr_get_server_from_group(grp);
557 if (srv)
558 break;
559 if (switched) {
560 if (avoided) {
Willy Tarreau9df86f92019-04-16 11:21:14 +0200561 /* no need to lock it, it was already
562 * locked on first pass.
563 */
Willy Tarreauf89c1872009-10-01 11:19:37 +0200564 srv = avoided;
565 break;
566 }
567 goto requeue_servers;
568 }
569 switched = 1;
570 fwrr_switch_trees(grp);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200571 }
572
Willy Tarreau9df86f92019-04-16 11:21:14 +0200573 /* OK, we have a server and it is locked. However, it may be
574 * saturated, in which case we don't want to reconsider it for
575 * now. We'll update its position and dequeue it anyway, so
576 * that we can move it to a better place afterwards.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200577 */
578 fwrr_update_position(grp, srv);
579 fwrr_dequeue_srv(srv);
580 grp->curr_pos++;
581 if (!srv->maxconn || (!srv->nbpend && srv->served < srv_dynamic_maxconn(srv))) {
582 /* make sure it is not the server we are trying to exclude... */
583 if (srv != srvtoavoid || avoided)
584 break;
585
586 avoided = srv; /* ...but remember that is was selected yet avoided */
587 }
588
Willy Tarreau9df86f92019-04-16 11:21:14 +0200589 /* the server is saturated or avoided, let's chain it for later reinsertion.
590 * Note that servers chained this way are all locked.
591 */
Willy Tarreauf89c1872009-10-01 11:19:37 +0200592 srv->next_full = full;
593 full = srv;
594 }
595
596 /* OK, we got the best server, let's update it */
597 fwrr_queue_srv(srv);
598
Willy Tarreau9df86f92019-04-16 11:21:14 +0200599 /* we don't need to keep this lock anymore */
600 HA_SPIN_UNLOCK(SERVER_LOCK, &srv->lock);
601
Willy Tarreauf89c1872009-10-01 11:19:37 +0200602 requeue_servers:
603 /* Requeue all extracted servers. If full==srv then it was
Willy Tarreau9df86f92019-04-16 11:21:14 +0200604 * avoided (unsuccessfully) and chained, omit it now. The
605 * only way to get there is by having <avoided>==NULL or
606 * <avoided>==<srv>.
Willy Tarreauf89c1872009-10-01 11:19:37 +0200607 */
608 if (unlikely(full != NULL)) {
609 if (switched) {
610 /* the tree has switched, requeue all extracted servers
611 * into "init", because their place was lost, and only
612 * their weight matters.
613 */
614 do {
Willy Tarreau9df86f92019-04-16 11:21:14 +0200615 if (likely(full != srv)) {
Willy Tarreauf89c1872009-10-01 11:19:37 +0200616 fwrr_queue_by_weight(grp->init, full);
Willy Tarreau9df86f92019-04-16 11:21:14 +0200617 HA_SPIN_UNLOCK(SERVER_LOCK, &full->lock);
618 }
Willy Tarreauf89c1872009-10-01 11:19:37 +0200619 full = full->next_full;
620 } while (full);
621 } else {
622 /* requeue all extracted servers just as if they were consumed
623 * so that they regain their expected place.
624 */
625 do {
Willy Tarreau9df86f92019-04-16 11:21:14 +0200626 if (likely(full != srv)) {
Willy Tarreauf89c1872009-10-01 11:19:37 +0200627 fwrr_queue_srv(full);
Willy Tarreau9df86f92019-04-16 11:21:14 +0200628 HA_SPIN_UNLOCK(SERVER_LOCK, &full->lock);
629 }
Willy Tarreauf89c1872009-10-01 11:19:37 +0200630 full = full->next_full;
631 } while (full);
632 }
633 }
Christopher Faulet5b517552017-06-09 14:17:53 +0200634 out:
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100635 HA_SPIN_UNLOCK(LBPRM_LOCK, &p->lbprm.lock);
Willy Tarreauf89c1872009-10-01 11:19:37 +0200636 return srv;
637}
638
639/*
640 * Local variables:
641 * c-indent-level: 8
642 * c-basic-offset: 8
643 * End:
644 */