1 /*
2  * Consistent Hash implementation
3  * Please consult this very well detailed article for more information :
4  * http://www.spiteful.com/2008/03/17/programmers-toolbox-part-3-consistent-hashing/
5  *
6  * Our implementation has to support both weighted hashing and weighted round
7  * robin because we'll use it to replace the previous map-based implementation
8  * which offered both algorithms.
9  *
10  * Copyright 2000-2010 Willy Tarreau <w@1wt.eu>
11  *
12  * This program is free software; you can redistribute it and/or
13  * modify it under the terms of the GNU General Public License
14  * as published by the Free Software Foundation; either version
15  * 2 of the License, or (at your option) any later version.
16  *
17  */
18 
19 #include <common/compat.h>
20 #include <common/config.h>
21 #include <common/debug.h>
22 #include <common/standard.h>
23 #include <eb32tree.h>
24 
25 #include <types/global.h>
26 #include <types/server.h>
27 
28 #include <proto/backend.h>
29 #include <proto/queue.h>
30 
31 /* Return next tree node after <node> which must still be in the tree, or be
32  * NULL. Lookup wraps around the end to the beginning. If the next node is the
33  * same node, return NULL. This is designed to find a valid next node before
34  * deleting one from the tree.
35  */
chash_skip_node(struct eb_root * root,struct eb32_node * node)36 static inline struct eb32_node *chash_skip_node(struct eb_root *root, struct eb32_node *node)
37 {
38 	struct eb32_node *stop = node;
39 
40 	if (!node)
41 		return NULL;
42 	node = eb32_next(node);
43 	if (!node)
44 		node = eb32_first(root);
45 	if (node == stop)
46 		return NULL;
47 	return node;
48 }
49 
50 /* Remove all of a server's entries from its tree. This may be used when
51  * setting a server down.
52  */
chash_dequeue_srv(struct server * s)53 static inline void chash_dequeue_srv(struct server *s)
54 {
55 	while (s->lb_nodes_now > 0) {
56 		if (s->lb_nodes_now >= s->lb_nodes_tot) // should always be false anyway
57 			s->lb_nodes_now = s->lb_nodes_tot;
58 		s->lb_nodes_now--;
59 		if (s->proxy->lbprm.chash.last == &s->lb_nodes[s->lb_nodes_now].node)
60 			s->proxy->lbprm.chash.last = chash_skip_node(s->lb_tree, s->proxy->lbprm.chash.last);
61 		eb32_delete(&s->lb_nodes[s->lb_nodes_now].node);
62 	}
63 }
64 
65 /* Adjust the number of entries of a server in its tree. The server must appear
66  * as many times as its weight indicates it. If it's there too often, we remove
67  * the last occurrences. If it's not there enough, we add more occurrences. To
68  * remove a server from the tree, normally call this with eweight=0.
69  */
chash_queue_dequeue_srv(struct server * s)70 static inline void chash_queue_dequeue_srv(struct server *s)
71 {
72 	while (s->lb_nodes_now > s->eweight) {
73 		if (s->lb_nodes_now >= s->lb_nodes_tot) // should always be false anyway
74 			s->lb_nodes_now = s->lb_nodes_tot;
75 		s->lb_nodes_now--;
76 		if (s->proxy->lbprm.chash.last == &s->lb_nodes[s->lb_nodes_now].node)
77 			s->proxy->lbprm.chash.last = chash_skip_node(s->lb_tree, s->proxy->lbprm.chash.last);
78 		eb32_delete(&s->lb_nodes[s->lb_nodes_now].node);
79 	}
80 
81 	/* Attempt to increase the total number of nodes, if the user
82 	 * increased the weight beyond the original weight
83 	 */
84 	if (s->lb_nodes_tot < s->eweight) {
85 		struct tree_occ *new_nodes;
86 
87 		/* First we need to remove all server's entries from its tree
88 		 * because the realloc will change all nodes pointers */
89 		chash_dequeue_srv(s);
90 
91 		new_nodes = realloc(s->lb_nodes, s->eweight * sizeof(*new_nodes));
92 		if (new_nodes) {
93 			unsigned int j;
94 
95 			s->lb_nodes = new_nodes;
96 			memset(&s->lb_nodes[s->lb_nodes_tot], 0,
97 			    (s->eweight - s->lb_nodes_tot) * sizeof(*s->lb_nodes));
98 			for (j = s->lb_nodes_tot; j < s->eweight; j++) {
99 				s->lb_nodes[j].server = s;
100 				s->lb_nodes[j].node.key = full_hash(s->puid * SRV_EWGHT_RANGE + j);
101 			}
102 			s->lb_nodes_tot = s->eweight;
103 		}
104 	}
105 	while (s->lb_nodes_now < s->eweight) {
106 		if (s->lb_nodes_now >= s->lb_nodes_tot) // should always be false anyway
107 			break;
108 		if (s->proxy->lbprm.chash.last == &s->lb_nodes[s->lb_nodes_now].node)
109 			s->proxy->lbprm.chash.last = chash_skip_node(s->lb_tree, s->proxy->lbprm.chash.last);
110 		eb32_insert(s->lb_tree, &s->lb_nodes[s->lb_nodes_now].node);
111 		s->lb_nodes_now++;
112 	}
113 }
114 
115 /* This function updates the server trees according to server <srv>'s new
116  * state. It should be called when server <srv>'s status changes to down.
117  * It is not important whether the server was already down or not. It is not
118  * important either that the new state is completely down (the caller may not
119  * know all the variables of a server's state).
120  */
chash_set_server_status_down(struct server * srv)121 static void chash_set_server_status_down(struct server *srv)
122 {
123 	struct proxy *p = srv->proxy;
124 
125 	if (!srv_lb_status_changed(srv))
126 		return;
127 
128 	if (srv_is_usable(srv))
129 		goto out_update_state;
130 
131 	if (!srv_was_usable(srv))
132 		/* server was already down */
133 		goto out_update_backend;
134 
135 	if (srv->flags & SRV_F_BACKUP) {
136 		p->lbprm.tot_wbck -= srv->prev_eweight;
137 		p->srv_bck--;
138 
139 		if (srv == p->lbprm.fbck) {
140 			/* we lost the first backup server in a single-backup
141 			 * configuration, we must search another one.
142 			 */
143 			struct server *srv2 = p->lbprm.fbck;
144 			do {
145 				srv2 = srv2->next;
146 			} while (srv2 &&
147 				 !((srv2->flags & SRV_F_BACKUP) &&
148 				   srv_is_usable(srv2)));
149 			p->lbprm.fbck = srv2;
150 		}
151 	} else {
152 		p->lbprm.tot_wact -= srv->prev_eweight;
153 		p->srv_act--;
154 	}
155 
156 	chash_dequeue_srv(srv);
157 
158 out_update_backend:
159 	/* check/update tot_used, tot_weight */
160 	update_backend_weight(p);
161  out_update_state:
162 	srv_lb_commit_status(srv);
163 }
164 
165 /* This function updates the server trees according to server <srv>'s new
166  * state. It should be called when server <srv>'s status changes to up.
167  * It is not important whether the server was already down or not. It is not
168  * important either that the new state is completely UP (the caller may not
169  * know all the variables of a server's state). This function will not change
170  * the weight of a server which was already up.
171  */
chash_set_server_status_up(struct server * srv)172 static void chash_set_server_status_up(struct server *srv)
173 {
174 	struct proxy *p = srv->proxy;
175 
176 	if (!srv_lb_status_changed(srv))
177 		return;
178 
179 	if (!srv_is_usable(srv))
180 		goto out_update_state;
181 
182 	if (srv_was_usable(srv))
183 		/* server was already up */
184 		goto out_update_backend;
185 
186 	if (srv->flags & SRV_F_BACKUP) {
187 		p->lbprm.tot_wbck += srv->eweight;
188 		p->srv_bck++;
189 
190 		if (!(p->options & PR_O_USE_ALL_BK)) {
191 			if (!p->lbprm.fbck) {
192 				/* there was no backup server anymore */
193 				p->lbprm.fbck = srv;
194 			} else {
195 				/* we may have restored a backup server prior to fbck,
196 				 * in which case it should replace it.
197 				 */
198 				struct server *srv2 = srv;
199 				do {
200 					srv2 = srv2->next;
201 				} while (srv2 && (srv2 != p->lbprm.fbck));
202 				if (srv2)
203 					p->lbprm.fbck = srv;
204 			}
205 		}
206 	} else {
207 		p->lbprm.tot_wact += srv->eweight;
208 		p->srv_act++;
209 	}
210 
211 	/* note that eweight cannot be 0 here */
212 	chash_queue_dequeue_srv(srv);
213 
214  out_update_backend:
215 	/* check/update tot_used, tot_weight */
216 	update_backend_weight(p);
217  out_update_state:
218 	srv_lb_commit_status(srv);
219 }
220 
221 /* This function must be called after an update to server <srv>'s effective
222  * weight. It may be called after a state change too.
223  */
chash_update_server_weight(struct server * srv)224 static void chash_update_server_weight(struct server *srv)
225 {
226 	int old_state, new_state;
227 	struct proxy *p = srv->proxy;
228 
229 	if (!srv_lb_status_changed(srv))
230 		return;
231 
232 	/* If changing the server's weight changes its state, we simply apply
233 	 * the procedures we already have for status change. If the state
234 	 * remains down, the server is not in any tree, so it's as easy as
235 	 * updating its values. If the state remains up with different weights,
236 	 * there are some computations to perform to find a new place and
237 	 * possibly a new tree for this server.
238 	 */
239 
240 	old_state = srv_was_usable(srv);
241 	new_state = srv_is_usable(srv);
242 
243 	if (!old_state && !new_state) {
244 		srv_lb_commit_status(srv);
245 		return;
246 	}
247 	else if (!old_state && new_state) {
248 		chash_set_server_status_up(srv);
249 		return;
250 	}
251 	else if (old_state && !new_state) {
252 		chash_set_server_status_down(srv);
253 		return;
254 	}
255 
256 	/* only adjust the server's presence in the tree */
257 	chash_queue_dequeue_srv(srv);
258 
259 	if (srv->flags & SRV_F_BACKUP)
260 		p->lbprm.tot_wbck += srv->eweight - srv->prev_eweight;
261 	else
262 		p->lbprm.tot_wact += srv->eweight - srv->prev_eweight;
263 
264 	update_backend_weight(p);
265 	srv_lb_commit_status(srv);
266 }
267 
268 /*
269  * This function implements the "Consistent Hashing with Bounded Loads" algorithm
270  * of Mirrokni, Thorup, and Zadimoghaddam (arxiv:1608.01350), adapted for use with
271  * unequal server weights.
272  */
chash_server_is_eligible(struct server * s)273 int chash_server_is_eligible(struct server *s)
274 {
275 	/* The total number of slots to allocate is the total number of outstanding requests
276 	 * (including the one we're about to make) times the load-balance-factor, rounded up.
277 	 */
278 	unsigned tot_slots = ((s->proxy->served + 1) * s->proxy->lbprm.chash.balance_factor + 99) / 100;
279 	unsigned slots_per_weight = tot_slots / s->proxy->lbprm.tot_weight;
280 	unsigned remainder = tot_slots % s->proxy->lbprm.tot_weight;
281 
282 	/* Allocate a whole number of slots per weight unit... */
283 	unsigned slots = s->eweight * slots_per_weight;
284 
285 	/* And then distribute the rest among servers proportionally to their weight. */
286 	slots += ((s->cumulative_weight + s->eweight) * remainder) / s->proxy->lbprm.tot_weight
287 		- (s->cumulative_weight * remainder) / s->proxy->lbprm.tot_weight;
288 
289 	/* But never leave a server with 0. */
290 	if (slots == 0)
291 		slots = 1;
292 
293 	return s->served < slots;
294 }
295 
296 /*
297  * This function returns the running server from the CHASH tree, which is at
298  * the closest distance from the value of <hash>. Doing so ensures that even
299  * with a well imbalanced hash, if some servers are close to each other, they
300  * will still both receive traffic. If any server is found, it will be returned.
301  * If no valid server is found, NULL is returned.
302  */
chash_get_server_hash(struct proxy * p,unsigned int hash)303 struct server *chash_get_server_hash(struct proxy *p, unsigned int hash)
304 {
305 	struct eb32_node *next, *prev;
306 	struct server *nsrv, *psrv;
307 	struct eb_root *root;
308 	unsigned int dn, dp;
309 	int loop;
310 
311 	if (p->srv_act)
312 		root = &p->lbprm.chash.act;
313 	else if (p->lbprm.fbck)
314 		return p->lbprm.fbck;
315 	else if (p->srv_bck)
316 		root = &p->lbprm.chash.bck;
317 	else
318 		return NULL;
319 
320 	/* find the node after and the node before */
321 	next = eb32_lookup_ge(root, hash);
322 	if (!next)
323 		next = eb32_first(root);
324 	if (!next)
325 		return NULL; /* tree is empty */
326 
327 	prev = eb32_prev(next);
328 	if (!prev)
329 		prev = eb32_last(root);
330 
331 	nsrv = eb32_entry(next, struct tree_occ, node)->server;
332 	psrv = eb32_entry(prev, struct tree_occ, node)->server;
333 
334 	/* OK we're located between two servers, let's
335 	 * compare distances between hash and the two servers
336 	 * and select the closest server.
337 	 */
338 	dp = hash - prev->key;
339 	dn = next->key - hash;
340 
341 	if (dp <= dn) {
342 		next = prev;
343 		nsrv = psrv;
344 	}
345 
346 	loop = 0;
347 	while (p->lbprm.chash.balance_factor && !chash_server_is_eligible(nsrv)) {
348 		next = eb32_next(next);
349 		if (!next) {
350 			next = eb32_first(root);
351 			if (++loop > 1) // protection against accidental loop
352 				break;
353 		}
354 		nsrv = eb32_entry(next, struct tree_occ, node)->server;
355 	}
356 
357 	return nsrv;
358 }
359 
360 /* Return next server from the CHASH tree in backend <p>. If the tree is empty,
361  * return NULL. Saturated servers are skipped.
362  */
chash_get_next_server(struct proxy * p,struct server * srvtoavoid)363 struct server *chash_get_next_server(struct proxy *p, struct server *srvtoavoid)
364 {
365 	struct server *srv, *avoided;
366 	struct eb32_node *node, *stop, *avoided_node;
367 	struct eb_root *root;
368 
369 	srv = avoided = NULL;
370 	avoided_node = NULL;
371 
372 	if (p->srv_act)
373 		root = &p->lbprm.chash.act;
374 	else if (p->lbprm.fbck)
375 		return p->lbprm.fbck;
376 	else if (p->srv_bck)
377 		root = &p->lbprm.chash.bck;
378 	else
379 		return NULL;
380 
381 	stop = node = p->lbprm.chash.last;
382 	do {
383 		struct server *s;
384 
385 		if (node)
386 			node = eb32_next(node);
387 		if (!node)
388 			node = eb32_first(root);
389 
390 		p->lbprm.chash.last = node;
391 		if (!node)
392 			/* no node is available */
393 			return NULL;
394 
395 		/* Note: if we came here after a down/up cycle with no last
396 		 * pointer, and after a redispatch (srvtoavoid is set), we
397 		 * must set stop to non-null otherwise we can loop forever.
398 		 */
399 		if (!stop)
400 			stop = node;
401 
402 		/* OK, we have a server. However, it may be saturated, in which
403 		 * case we don't want to reconsider it for now, so we'll simply
404 		 * skip it. Same if it's the server we try to avoid, in which
405 		 * case we simply remember it for later use if needed.
406 		 */
407 		s = eb32_entry(node, struct tree_occ, node)->server;
408 		if (!s->maxconn || (!s->nbpend && s->served < srv_dynamic_maxconn(s))) {
409 			if (s != srvtoavoid) {
410 				srv = s;
411 				break;
412 			}
413 			avoided = s;
414 			avoided_node = node;
415 		}
416 	} while (node != stop);
417 
418 	if (!srv) {
419 		srv = avoided;
420 		p->lbprm.chash.last = avoided_node;
421 	}
422 
423 	return srv;
424 }
425 
426 /* This function is responsible for building the active and backup trees for
427  * constistent hashing. The servers receive an array of initialized nodes
428  * with their assigned keys. It also sets p->lbprm.wdiv to the eweight to
429  * uweight ratio.
430  */
chash_init_server_tree(struct proxy * p)431 void chash_init_server_tree(struct proxy *p)
432 {
433 	struct server *srv;
434 	struct eb_root init_head = EB_ROOT;
435 	int node;
436 
437 	p->lbprm.set_server_status_up   = chash_set_server_status_up;
438 	p->lbprm.set_server_status_down = chash_set_server_status_down;
439 	p->lbprm.update_server_eweight  = chash_update_server_weight;
440 	p->lbprm.server_take_conn = NULL;
441 	p->lbprm.server_drop_conn = NULL;
442 
443 	p->lbprm.wdiv = BE_WEIGHT_SCALE;
444 	for (srv = p->srv; srv; srv = srv->next) {
445 		srv->eweight = (srv->uweight * p->lbprm.wdiv + p->lbprm.wmult - 1) / p->lbprm.wmult;
446 		srv_lb_commit_status(srv);
447 	}
448 
449 	recount_servers(p);
450 	update_backend_weight(p);
451 
452 	p->lbprm.chash.act = init_head;
453 	p->lbprm.chash.bck = init_head;
454 	p->lbprm.chash.last = NULL;
455 
456 	/* queue active and backup servers in two distinct groups */
457 	for (srv = p->srv; srv; srv = srv->next) {
458 		srv->lb_tree = (srv->flags & SRV_F_BACKUP) ? &p->lbprm.chash.bck : &p->lbprm.chash.act;
459 		srv->lb_nodes_tot = srv->uweight * BE_WEIGHT_SCALE;
460 		srv->lb_nodes_now = 0;
461 		srv->lb_nodes = calloc(srv->lb_nodes_tot, sizeof(struct tree_occ));
462 		for (node = 0; node < srv->lb_nodes_tot; node++) {
463 			srv->lb_nodes[node].server = srv;
464 			srv->lb_nodes[node].node.key = full_hash(srv->puid * SRV_EWGHT_RANGE + node);
465 		}
466 
467 		if (srv_is_usable(srv))
468 			chash_queue_dequeue_srv(srv);
469 	}
470 }
471