blob: 1aa3fb9ac76259c88c07c44c9e3a5efa17ce97b4 [file] [log] [blame]
Willy Tarreaubaaee002006-06-26 02:48:02 +02001/*
2 * Queue management functions.
3 *
Willy Tarreauac68c5d2009-10-04 23:12:44 +02004 * Copyright 2000-2009 Willy Tarreau <w@1wt.eu>
Willy Tarreaubaaee002006-06-26 02:48:02 +02005 *
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 Tarreau6bdd05c2018-07-25 15:21:00 +020013/* Short explanation on the locking, which is far from being trivial : a
14 * pendconn is a list element which necessarily is associated with an existing
15 * stream. It has pendconn->strm always valid. A pendconn may only be in one of
16 * these three states :
17 * - unlinked : in this case it is an empty list head ;
18 * - linked into the server's queue ;
19 * - linked into the proxy's queue.
20 *
21 * A stream does not necessarily have such a pendconn. Thus the pendconn is
22 * designated by the stream->pend_pos pointer. This results in some properties :
23 * - pendconn->strm->pend_pos is never NULL for any valid pendconn
Patrick Hemmer0355dab2018-05-11 12:52:31 -040024 * - if p->node.node.leaf_p is NULL, the element is unlinked,
Willy Tarreau6bdd05c2018-07-25 15:21:00 +020025 * otherwise it necessarily belongs to one of the other lists ; this may
26 * not be atomically checked under threads though ;
27 * - pendconn->px is never NULL if pendconn->list is not empty
Willy Tarreau88930dd2018-07-26 07:38:54 +020028 * - pendconn->srv is never NULL if pendconn->list is in the server's queue,
Willy Tarreau6bdd05c2018-07-25 15:21:00 +020029 * and is always NULL if pendconn->list is in the backend's queue or empty.
Willy Tarreau88930dd2018-07-26 07:38:54 +020030 * - pendconn->target is NULL while the element is queued, and points to the
31 * assigned server when the pendconn is picked.
Willy Tarreau6bdd05c2018-07-25 15:21:00 +020032 *
33 * Threads complicate the design a little bit but rules remain simple :
Willy Tarreau6bdd05c2018-07-25 15:21:00 +020034 * - the server's queue lock must be held at least when manipulating the
35 * server's queue, which is when adding a pendconn to the queue and when
36 * removing a pendconn from the queue. It protects the queue's integrity.
37 *
38 * - the proxy's queue lock must be held at least when manipulating the
39 * proxy's queue, which is when adding a pendconn to the queue and when
40 * removing a pendconn from the queue. It protects the queue's integrity.
41 *
Willy Tarreau3201e4e2018-07-26 08:23:24 +020042 * - both locks are compatible and may be held at the same time.
Willy Tarreau6bdd05c2018-07-25 15:21:00 +020043 *
44 * - a pendconn_add() is only performed by the stream which will own the
45 * pendconn ; the pendconn is allocated at this moment and returned ; it is
46 * added to either the server or the proxy's queue while holding this
Willy Tarreau49801602020-06-04 22:50:02 +020047s * queue's lock.
Willy Tarreau6bdd05c2018-07-25 15:21:00 +020048 *
49 * - the pendconn is then met by a thread walking over the proxy or server's
50 * queue with the respective lock held. This lock is exclusive and the
51 * pendconn can only appear in one queue so by definition a single thread
52 * may find this pendconn at a time.
53 *
54 * - the pendconn is unlinked either by its own stream upon success/abort/
55 * free, or by another one offering it its server slot. This is achieved by
56 * pendconn_process_next_strm() under either the server or proxy's lock,
57 * pendconn_redistribute() under the server's lock, pendconn_grab_from_px()
58 * under the proxy's lock, or pendconn_unlink() under either the proxy's or
59 * the server's lock depending on the queue the pendconn is attached to.
60 *
61 * - no single operation except the pendconn initialisation prior to the
Willy Tarreau3201e4e2018-07-26 08:23:24 +020062 * insertion are performed without eithre a queue lock held or the element
63 * being unlinked and visible exclusively to its stream.
Willy Tarreau6bdd05c2018-07-25 15:21:00 +020064 *
Willy Tarreau88930dd2018-07-26 07:38:54 +020065 * - pendconn_grab_from_px() and pendconn_process_next_strm() assign ->target
66 * so that the stream knows what server to work with (via
67 * pendconn_dequeue() which sets it on strm->target).
Willy Tarreau6bdd05c2018-07-25 15:21:00 +020068 *
69 * - a pendconn doesn't switch between queues, it stays where it is.
Willy Tarreau6bdd05c2018-07-25 15:21:00 +020070 */
71
Willy Tarreaudfd3de82020-06-04 23:46:14 +020072#include <import/eb32tree.h>
Willy Tarreau4c7e4b72020-05-27 12:58:42 +020073#include <haproxy/api.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020074#include <haproxy/backend.h>
Willy Tarreauc761f842020-06-04 11:40:28 +020075#include <haproxy/http_rules.h>
Willy Tarreaud0ef4392020-06-02 09:38:52 +020076#include <haproxy/pool.h>
Willy Tarreaua55c4542020-06-04 22:59:39 +020077#include <haproxy/queue.h>
Willy Tarreaue6ce10b2020-06-04 15:33:47 +020078#include <haproxy/sample.h>
Willy Tarreau1e56f922020-06-04 23:20:13 +020079#include <haproxy/server-t.h>
Willy Tarreaudfd3de82020-06-04 23:46:14 +020080#include <haproxy/stream.h>
Willy Tarreau5e539c92020-06-04 20:45:39 +020081#include <haproxy/stream_interface.h>
Willy Tarreaucea0e1b2020-06-04 17:25:40 +020082#include <haproxy/task.h>
Willy Tarreau8b550af2020-06-04 17:42:48 +020083#include <haproxy/tcp_rules.h>
Willy Tarreau3f567e42020-05-28 15:29:19 +020084#include <haproxy/thread.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020085#include <haproxy/time.h>
Willy Tarreauc1a689f2021-05-08 13:59:05 +020086#include <haproxy/tools.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020087
88
Patrick Hemmer248cb4c2018-05-11 12:52:31 -040089#define NOW_OFFSET_BOUNDARY() ((now_ms - (TIMER_LOOK_BACK >> 12)) & 0xfffff)
90#define KEY_CLASS(key) ((u32)key & 0xfff00000)
91#define KEY_OFFSET(key) ((u32)key & 0x000fffff)
92#define KEY_CLASS_OFFSET_BOUNDARY(key) (KEY_CLASS(key) | NOW_OFFSET_BOUNDARY())
93#define MAKE_KEY(class, offset) (((u32)(class + 0x7ff) << 20) | ((u32)(now_ms + offset) & 0xfffff))
94
Willy Tarreau8ceae722018-11-26 11:58:30 +010095DECLARE_POOL(pool_head_pendconn, "pendconn", sizeof(struct pendconn));
Willy Tarreaubaaee002006-06-26 02:48:02 +020096
97/* returns the effective dynamic maxconn for a server, considering the minconn
Willy Tarreau86034312006-12-29 00:10:33 +010098 * and the proxy's usage relative to its dynamic connections limit. It is
Willy Tarreau9909fc12007-11-30 17:42:05 +010099 * expected that 0 < s->minconn <= s->maxconn when this is called. If the
100 * server is currently warming up, the slowstart is also applied to the
101 * resulting value, which can be lower than minconn in this case, but never
102 * less than 1.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200103 */
Willy Tarreaub17916e2006-10-15 15:17:57 +0200104unsigned int srv_dynamic_maxconn(const struct server *s)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200105{
Willy Tarreau9909fc12007-11-30 17:42:05 +0100106 unsigned int max;
107
Willy Tarreau86034312006-12-29 00:10:33 +0100108 if (s->proxy->beconn >= s->proxy->fullconn)
109 /* no fullconn or proxy is full */
Willy Tarreau9909fc12007-11-30 17:42:05 +0100110 max = s->maxconn;
111 else if (s->minconn == s->maxconn)
Willy Tarreau86034312006-12-29 00:10:33 +0100112 /* static limit */
Willy Tarreau9909fc12007-11-30 17:42:05 +0100113 max = s->maxconn;
114 else max = MAX(s->minconn,
115 s->proxy->beconn * s->maxconn / s->proxy->fullconn);
Willy Tarreau86034312006-12-29 00:10:33 +0100116
Emeric Brun52a91d32017-08-31 14:41:55 +0200117 if ((s->cur_state == SRV_ST_STARTING) &&
Willy Tarreau9909fc12007-11-30 17:42:05 +0100118 now.tv_sec < s->last_change + s->slowstart &&
119 now.tv_sec >= s->last_change) {
120 unsigned int ratio;
Willy Tarreau28a9e522008-09-14 17:43:27 +0200121 ratio = 100 * (now.tv_sec - s->last_change) / s->slowstart;
122 max = MAX(1, max * ratio / 100);
Willy Tarreau9909fc12007-11-30 17:42:05 +0100123 }
124 return max;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200125}
126
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200127/* Remove the pendconn from the server's queue. At this stage, the connection
Willy Tarreau96bca332020-10-21 12:01:28 +0200128 * is not really dequeued. It will be done during the process_stream. It is
129 * up to the caller to atomically decrement the pending counts.
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100130 *
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200131 * The caller must own the lock on the server queue. The pendconn must still be
132 * queued (p->node.leaf_p != NULL) and must be in a server (p->srv != NULL).
Christopher Fauletf3a55db2017-06-09 14:26:38 +0200133 */
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200134static void __pendconn_unlink_srv(struct pendconn *p)
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100135{
Willy Tarreau98c89102021-06-18 10:51:58 +0200136 p->strm->logs.srv_queue_pos += _HA_ATOMIC_LOAD(&p->srv->queue.idx) - p->queue_idx;
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200137 eb32_delete(&p->node);
138}
139
140/* Remove the pendconn from the proxy's queue. At this stage, the connection
Willy Tarreau96bca332020-10-21 12:01:28 +0200141 * is not really dequeued. It will be done during the process_stream. It is
142 * up to the caller to atomically decrement the pending counts.
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200143 *
144 * The caller must own the lock on the proxy queue. The pendconn must still be
145 * queued (p->node.leaf_p != NULL) and must be in the proxy (p->srv == NULL).
146 */
147static void __pendconn_unlink_prx(struct pendconn *p)
148{
Willy Tarreau98c89102021-06-18 10:51:58 +0200149 p->strm->logs.prx_queue_pos += _HA_ATOMIC_LOAD(&p->px->queue.idx) - p->queue_idx;
Patrick Hemmer0355dab2018-05-11 12:52:31 -0400150 eb32_delete(&p->node);
Christopher Fauletf3a55db2017-06-09 14:26:38 +0200151}
152
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200153/* Locks the queue the pendconn element belongs to. This relies on both p->px
154 * and p->srv to be properly initialized (which is always the case once the
155 * element has been added).
156 */
157static inline void pendconn_queue_lock(struct pendconn *p)
158{
159 if (p->srv)
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200160 HA_SPIN_LOCK(QUEUE_LOCK, &p->srv->queue.lock);
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200161 else
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200162 HA_SPIN_LOCK(QUEUE_LOCK, &p->px->queue.lock);
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200163}
164
165/* Unlocks the queue the pendconn element belongs to. This relies on both p->px
166 * and p->srv to be properly initialized (which is always the case once the
167 * element has been added).
168 */
169static inline void pendconn_queue_unlock(struct pendconn *p)
170{
171 if (p->srv)
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200172 HA_SPIN_UNLOCK(QUEUE_LOCK, &p->srv->queue.lock);
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200173 else
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200174 HA_SPIN_UNLOCK(QUEUE_LOCK, &p->px->queue.lock);
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200175}
176
Willy Tarreau9624fae2018-07-25 08:04:20 +0200177/* Removes the pendconn from the server/proxy queue. At this stage, the
178 * connection is not really dequeued. It will be done during process_stream().
Willy Tarreau9ada0302019-11-14 14:58:39 +0100179 * This function takes all the required locks for the operation. The pendconn
180 * must be valid, though it doesn't matter if it was already unlinked. Prefer
181 * pendconn_cond_unlink() to first check <p>. When the locks are already held,
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200182 * please use __pendconn_unlink_{srv,prx}() instead.
Willy Tarreau9624fae2018-07-25 08:04:20 +0200183 */
184void pendconn_unlink(struct pendconn *p)
185{
Willy Tarreau96bca332020-10-21 12:01:28 +0200186 int done = 0;
187
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200188 if (p->srv) {
189 /* queued in the server */
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200190 HA_SPIN_LOCK(QUEUE_LOCK, &p->srv->queue.lock);
Willy Tarreau96bca332020-10-21 12:01:28 +0200191 if (p->node.node.leaf_p) {
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200192 __pendconn_unlink_srv(p);
Willy Tarreau96bca332020-10-21 12:01:28 +0200193 done = 1;
194 }
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200195 HA_SPIN_UNLOCK(QUEUE_LOCK, &p->srv->queue.lock);
Willy Tarreau96bca332020-10-21 12:01:28 +0200196 if (done) {
Willy Tarreaua0570452021-06-18 09:30:30 +0200197 _HA_ATOMIC_DEC(&p->srv->queue.length);
Willy Tarreau4781b152021-04-06 13:53:36 +0200198 _HA_ATOMIC_DEC(&p->px->totpend);
Willy Tarreau96bca332020-10-21 12:01:28 +0200199 }
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200200 }
201 else {
202 /* queued in the proxy */
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200203 HA_SPIN_LOCK(QUEUE_LOCK, &p->px->queue.lock);
Willy Tarreau96bca332020-10-21 12:01:28 +0200204 if (p->node.node.leaf_p) {
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200205 __pendconn_unlink_prx(p);
Willy Tarreau96bca332020-10-21 12:01:28 +0200206 done = 1;
207 }
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200208 HA_SPIN_UNLOCK(QUEUE_LOCK, &p->px->queue.lock);
Willy Tarreau96bca332020-10-21 12:01:28 +0200209 if (done) {
Willy Tarreau7f3c1df2021-06-18 09:22:21 +0200210 _HA_ATOMIC_DEC(&p->px->queue.length);
Willy Tarreau4781b152021-04-06 13:53:36 +0200211 _HA_ATOMIC_DEC(&p->px->totpend);
Willy Tarreau96bca332020-10-21 12:01:28 +0200212 }
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200213 }
Willy Tarreau9624fae2018-07-25 08:04:20 +0200214}
215
Willy Tarreau2bf3f2c2021-06-24 07:20:26 +0200216/* Retrieve the first pendconn from tree <pendconns>. Classes are always
217 * considered first, then the time offset. The time does wrap, so the
218 * lookup is performed twice, one to retrieve the first class and a second
219 * time to retrieve the earliest time in this class.
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400220 */
Willy Tarreau2bf3f2c2021-06-24 07:20:26 +0200221static struct pendconn *pendconn_first(struct eb_root *pendconns)
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400222{
223 struct eb32_node *node, *node2 = NULL;
224 u32 key;
225
Willy Tarreau2bf3f2c2021-06-24 07:20:26 +0200226 node = eb32_first(pendconns);
227 if (!node)
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400228 return NULL;
229
230 key = KEY_CLASS_OFFSET_BOUNDARY(node->key);
Willy Tarreau2bf3f2c2021-06-24 07:20:26 +0200231 node2 = eb32_lookup_ge(pendconns, key);
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400232
233 if (!node2 ||
234 KEY_CLASS(node2->key) != KEY_CLASS(node->key)) {
235 /* no other key in the tree, or in this class */
236 return eb32_entry(node, struct pendconn, node);
237 }
238
239 /* found a better key */
240 return eb32_entry(node2, struct pendconn, node);
241}
242
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100243/* Process the next pending connection from either a server or a proxy, and
Christopher Fauletfd83f0b2018-03-19 15:22:09 +0100244 * returns a strictly positive value on success (see below). If no pending
245 * connection is found, 0 is returned. Note that neither <srv> nor <px> may be
246 * NULL. Priority is given to the oldest request in the queue if both <srv> and
247 * <px> have pending requests. This ensures that no request will be left
248 * unserved. The <px> queue is not considered if the server (or a tracked
249 * server) is not RUNNING, is disabled, or has a null weight (server going
250 * down). The <srv> queue is still considered in this case, because if some
251 * connections remain there, it means that some requests have been forced there
252 * after it was seen down (eg: due to option persist). The stream is
253 * immediately marked as "assigned", and both its <srv> and <srv_conn> are set
254 * to <srv>.
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100255 *
Willy Tarreaua0e9c552021-06-18 19:45:17 +0200256 * The proxy's queue will be consulted only if px_ok is non-zero.
257 *
Willy Tarreau5343d8e2021-06-24 07:22:03 +0200258 * This function must only be called if the server queue _AND_ the proxy queue
Willy Tarreaua0e9c552021-06-18 19:45:17 +0200259 * are locked (if px_ok is set). Today it is only called by process_srv_queue.
260 * When a pending connection is dequeued, this function returns 1 if a pendconn
261 * is dequeued, otherwise 0.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200262 */
Willy Tarreaua0e9c552021-06-18 19:45:17 +0200263static int pendconn_process_next_strm(struct server *srv, struct proxy *px, int px_ok)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200264{
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100265 struct pendconn *p = NULL;
Patrick Hemmerda282f42018-05-11 12:52:31 -0400266 struct pendconn *pp = NULL;
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400267 u32 pkey, ppkey;
Willy Tarreaud132f742010-08-06 10:08:23 +0200268
Willy Tarreau2bf3f2c2021-06-24 07:20:26 +0200269 p = NULL;
Willy Tarreau90a160a2021-06-24 07:21:59 +0200270 if (srv->queue.length)
Willy Tarreau2bf3f2c2021-06-24 07:20:26 +0200271 p = pendconn_first(&srv->queue.head);
Willy Tarreau2bf3f2c2021-06-24 07:20:26 +0200272
273 pp = NULL;
Willy Tarreaua0e9c552021-06-18 19:45:17 +0200274 if (px_ok && px->queue.length)
Willy Tarreau2bf3f2c2021-06-24 07:20:26 +0200275 pp = pendconn_first(&px->queue.head);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100276
Willy Tarreau5343d8e2021-06-24 07:22:03 +0200277 if (!p && !pp)
Willy Tarreaua48905b2021-06-24 07:27:01 +0200278 return 0;
Christopher Fauletcd7126b2021-02-11 11:13:33 +0100279 else if (!pp)
280 goto use_p; /* p != NULL */
281 else if (!p)
282 goto use_pp; /* pp != NULL */
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100283
Christopher Fauletcd7126b2021-02-11 11:13:33 +0100284 /* p != NULL && pp != NULL*/
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100285
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400286 if (KEY_CLASS(p->node.key) < KEY_CLASS(pp->node.key))
287 goto use_p;
288
289 if (KEY_CLASS(pp->node.key) < KEY_CLASS(p->node.key))
290 goto use_pp;
291
292 pkey = KEY_OFFSET(p->node.key);
293 ppkey = KEY_OFFSET(pp->node.key);
294
295 if (pkey < NOW_OFFSET_BOUNDARY())
296 pkey += 0x100000; // key in the future
297
298 if (ppkey < NOW_OFFSET_BOUNDARY())
299 ppkey += 0x100000; // key in the future
300
301 if (pkey <= ppkey)
302 goto use_p;
303
304 use_pp:
305 /* Let's switch from the server pendconn to the proxy pendconn */
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200306 __pendconn_unlink_prx(pp);
Willy Tarreau7f3c1df2021-06-18 09:22:21 +0200307 _HA_ATOMIC_DEC(&px->queue.length);
Willy Tarreau98c89102021-06-18 10:51:58 +0200308 _HA_ATOMIC_INC(&px->queue.idx);
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400309 p = pp;
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200310 goto unlinked;
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400311 use_p:
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200312 __pendconn_unlink_srv(p);
Willy Tarreaua0570452021-06-18 09:30:30 +0200313 _HA_ATOMIC_DEC(&srv->queue.length);
Willy Tarreau98c89102021-06-18 10:51:58 +0200314 _HA_ATOMIC_INC(&srv->queue.idx);
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200315 unlinked:
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100316 p->strm_flags |= SF_ASSIGNED;
Willy Tarreau88930dd2018-07-26 07:38:54 +0200317 p->target = srv;
Willy Tarreaua48905b2021-06-24 07:27:01 +0200318
Willy Tarreaua48905b2021-06-24 07:27:01 +0200319 stream_add_srv_conn(p->strm, srv);
320
321 task_wakeup(p->strm->task, TASK_WOKEN_RES);
322
323 return 1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200324}
325
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100326/* Manages a server's connection queue. This function will try to dequeue as
Willy Tarreau9ab78292021-06-22 18:47:51 +0200327 * many pending streams as possible, and wake them up.
Christopher Faulet87566c92017-06-06 10:34:51 +0200328 */
Willy Tarreau9ab78292021-06-22 18:47:51 +0200329void process_srv_queue(struct server *s)
Christopher Faulet87566c92017-06-06 10:34:51 +0200330{
Willy Tarreaua0e9c552021-06-18 19:45:17 +0200331 struct server *ref = s->track ? s->track : s;
Christopher Faulet87566c92017-06-06 10:34:51 +0200332 struct proxy *p = s->proxy;
Olivier Houchardecfe6732018-07-26 18:47:27 +0200333 int maxconn;
Willy Tarreau9cef43a2021-06-24 07:47:08 +0200334 int done = 0;
Willy Tarreaua0e9c552021-06-18 19:45:17 +0200335 int px_ok;
336
337 /* if a server is not usable or backup and must not be used
338 * to dequeue backend requests.
339 */
340 px_ok = srv_currently_usable(ref) &&
341 (!(s->flags & SRV_F_BACKUP) ||
342 (!p->srv_act &&
343 (s == p->lbprm.fbck || (p->options & PR_O_USE_ALL_BK))));
Christopher Faulet87566c92017-06-06 10:34:51 +0200344
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200345 HA_SPIN_LOCK(SERVER_LOCK, &s->queue.lock);
346 HA_SPIN_LOCK(PROXY_LOCK, &p->queue.lock);
Christopher Faulet87566c92017-06-06 10:34:51 +0200347 maxconn = srv_dynamic_maxconn(s);
348 while (s->served < maxconn) {
Willy Tarreaua0e9c552021-06-18 19:45:17 +0200349 int ret = pendconn_process_next_strm(s, p, px_ok);
Willy Tarreaua48905b2021-06-24 07:27:01 +0200350 if (!ret)
Christopher Faulet87566c92017-06-06 10:34:51 +0200351 break;
Willy Tarreau9cef43a2021-06-24 07:47:08 +0200352 _HA_ATOMIC_INC(&s->served);
353 done++;
Christopher Faulet87566c92017-06-06 10:34:51 +0200354 }
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200355 HA_SPIN_UNLOCK(PROXY_LOCK, &p->queue.lock);
356 HA_SPIN_UNLOCK(SERVER_LOCK, &s->queue.lock);
Willy Tarreau9cef43a2021-06-24 07:47:08 +0200357
358 if (done) {
359 _HA_ATOMIC_SUB(&p->totpend, done);
360 _HA_ATOMIC_ADD(&p->served, done);
361 __ha_barrier_atomic_store();
362 if (p->lbprm.server_take_conn)
363 p->lbprm.server_take_conn(s);
364 }
Christopher Faulet87566c92017-06-06 10:34:51 +0200365}
366
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400367/* Adds the stream <strm> to the pending connection queue of server <strm>->srv
Willy Tarreau87b09662015-04-03 00:22:06 +0200368 * or to the one of <strm>->proxy if srv is NULL. All counters and back pointers
Willy Tarreaubaaee002006-06-26 02:48:02 +0200369 * are updated accordingly. Returns NULL if no memory is available, otherwise the
Willy Tarreau87b09662015-04-03 00:22:06 +0200370 * pendconn itself. If the stream was already marked as served, its flag is
371 * cleared. It is illegal to call this function with a non-NULL strm->srv_conn.
Patrick Hemmerda282f42018-05-11 12:52:31 -0400372 * The stream's queue position is counted with an offset of -1 because we want
373 * to make sure that being at the first position in the queue reports 1.
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100374 *
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400375 * The queue is sorted by the composition of the priority_class, and the current
376 * timestamp offset by strm->priority_offset. The timestamp is in milliseconds
377 * and truncated to 20 bits, so will wrap every 17m28s575ms.
378 * The offset can be positive or negative, and an offset of 0 puts it in the
379 * middle of this range (~ 8 min). Note that this also means if the adjusted
380 * timestamp wraps around, the request will be misinterpreted as being of
Joseph Herlantd8499ec2018-11-25 11:26:48 -0800381 * the highest priority for that priority class.
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400382 *
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100383 * This function must be called by the stream itself, so in the context of
384 * process_stream.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200385 */
Willy Tarreau87b09662015-04-03 00:22:06 +0200386struct pendconn *pendconn_add(struct stream *strm)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200387{
388 struct pendconn *p;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100389 struct proxy *px;
390 struct server *srv;
Willy Tarreau12529c02021-06-18 10:21:20 +0200391 struct queue *q;
392 unsigned int *max_ptr;
393 unsigned int old_max, new_max;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200394
Willy Tarreaubafbe012017-11-24 17:34:44 +0100395 p = pool_alloc(pool_head_pendconn);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200396 if (!p)
397 return NULL;
398
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200399 if (strm->flags & SF_ASSIGNED)
400 srv = objt_server(strm->target);
401 else
402 srv = NULL;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100403
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200404 px = strm->be;
Willy Tarreau88930dd2018-07-26 07:38:54 +0200405 p->target = NULL;
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200406 p->srv = srv;
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400407 p->node.key = MAKE_KEY(strm->priority_class, strm->priority_offset);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100408 p->px = px;
409 p->strm = strm;
410 p->strm_flags = strm->flags;
Willy Tarreau901972e2021-06-18 10:33:47 +0200411 strm->pend_pos = p;
Willy Tarreau7c669d72008-06-20 15:04:11 +0200412
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200413 if (srv) {
Willy Tarreau12529c02021-06-18 10:21:20 +0200414 q = &srv->queue;
415 max_ptr = &srv->counters.nbpend_max;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100416 }
417 else {
Willy Tarreau12529c02021-06-18 10:21:20 +0200418 q = &px->queue;
419 max_ptr = &px->be_counters.nbpend_max;
420 }
Willy Tarreau3eecdb62021-06-18 10:21:20 +0200421
Willy Tarreau98c89102021-06-18 10:51:58 +0200422 p->queue_idx = _HA_ATOMIC_LOAD(&q->idx) - 1; // for logging only
Willy Tarreau12529c02021-06-18 10:21:20 +0200423 new_max = _HA_ATOMIC_ADD_FETCH(&q->length, 1);
424 old_max = _HA_ATOMIC_LOAD(max_ptr);
425 while (new_max > old_max) {
426 if (likely(_HA_ATOMIC_CAS(max_ptr, &old_max, new_max)))
427 break;
Willy Tarreau58f4dfb2021-06-24 07:22:15 +0200428 }
Willy Tarreau12529c02021-06-18 10:21:20 +0200429 __ha_barrier_atomic_store();
430
431 HA_SPIN_LOCK(QUEUE_LOCK, &q->lock);
Willy Tarreau12529c02021-06-18 10:21:20 +0200432 eb32_insert(&q->head, &p->node);
433 HA_SPIN_UNLOCK(QUEUE_LOCK, &q->lock);
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200434
Willy Tarreau4781b152021-04-06 13:53:36 +0200435 _HA_ATOMIC_INC(&px->totpend);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200436 return p;
437}
438
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200439/* Redistribute pending connections when a server goes down. The number of
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200440 * connections redistributed is returned. It will take the server queue lock
441 * and does not use nor depend on other locks.
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200442 */
443int pendconn_redistribute(struct server *s)
444{
Patrick Hemmer0355dab2018-05-11 12:52:31 -0400445 struct pendconn *p;
Willy Tarreaubff005a2019-05-27 08:10:11 +0200446 struct eb32_node *node, *nodeb;
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200447 int xferred = 0;
448
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100449 /* The REDISP option was specified. We will ignore cookie and force to
450 * balance or use the dispatcher. */
451 if ((s->proxy->options & (PR_O_REDISP|PR_O_PERSIST)) != PR_O_REDISP)
452 return 0;
453
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200454 HA_SPIN_LOCK(SERVER_LOCK, &s->queue.lock);
Willy Tarreaua0570452021-06-18 09:30:30 +0200455 for (node = eb32_first(&s->queue.head); node; node = nodeb) {
Willy Tarreaubff005a2019-05-27 08:10:11 +0200456 nodeb = eb32_next(node);
457
458 p = eb32_entry(node, struct pendconn, node);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100459 if (p->strm_flags & SF_FORCE_PRST)
460 continue;
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200461
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100462 /* it's left to the dispatcher to choose a server */
Willy Tarreau3e3ae252020-10-21 11:20:07 +0200463 __pendconn_unlink_srv(p);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100464 p->strm_flags &= ~(SF_DIRECT | SF_ASSIGNED | SF_ADDR_SET);
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200465
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100466 task_wakeup(p->strm->task, TASK_WOKEN_RES);
Willy Tarreauef71f012020-10-21 11:54:38 +0200467 xferred++;
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200468 }
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200469 HA_SPIN_UNLOCK(SERVER_LOCK, &s->queue.lock);
470
Willy Tarreau96bca332020-10-21 12:01:28 +0200471 if (xferred) {
Willy Tarreaua0570452021-06-18 09:30:30 +0200472 _HA_ATOMIC_SUB(&s->queue.length, xferred);
Willy Tarreau5472aa52020-10-24 12:57:41 +0200473 _HA_ATOMIC_SUB(&s->proxy->totpend, xferred);
Willy Tarreau96bca332020-10-21 12:01:28 +0200474 }
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200475 return xferred;
476}
477
478/* Check for pending connections at the backend, and assign some of them to
479 * the server coming up. The server's weight is checked before being assigned
480 * connections it may not be able to handle. The total number of transferred
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200481 * connections is returned. It will take the proxy's queue lock and will not
482 * use nor depend on other locks.
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200483 */
484int pendconn_grab_from_px(struct server *s)
485{
Patrick Hemmer0355dab2018-05-11 12:52:31 -0400486 struct pendconn *p;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100487 int maxconn, xferred = 0;
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200488
Emeric Brun52a91d32017-08-31 14:41:55 +0200489 if (!srv_currently_usable(s))
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200490 return 0;
491
Willy Tarreaua8694652018-08-07 10:44:58 +0200492 /* if this is a backup server and there are active servers or at
493 * least another backup server was elected, then this one must
494 * not dequeue requests from the proxy.
495 */
496 if ((s->flags & SRV_F_BACKUP) &&
497 (s->proxy->srv_act ||
498 ((s != s->proxy->lbprm.fbck) && !(s->proxy->options & PR_O_USE_ALL_BK))))
499 return 0;
500
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200501 HA_SPIN_LOCK(QUEUE_LOCK, &s->proxy->queue.lock);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100502 maxconn = srv_dynamic_maxconn(s);
Willy Tarreau2bf3f2c2021-06-24 07:20:26 +0200503 while ((p = pendconn_first(&s->proxy->queue.head))) {
504 if (s->maxconn && s->served + xferred >= maxconn)
505 break;
Willy Tarreau772e9682021-06-18 20:32:50 +0200506
Willy Tarreau2bf3f2c2021-06-24 07:20:26 +0200507 __pendconn_unlink_prx(p);
Willy Tarreau88930dd2018-07-26 07:38:54 +0200508 p->target = s;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100509
510 task_wakeup(p->strm->task, TASK_WOKEN_RES);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100511 xferred++;
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200512 }
Willy Tarreau16fbdda2021-06-18 09:45:27 +0200513 HA_SPIN_UNLOCK(QUEUE_LOCK, &s->proxy->queue.lock);
Willy Tarreau96bca332020-10-21 12:01:28 +0200514 if (xferred) {
Willy Tarreau7f3c1df2021-06-18 09:22:21 +0200515 _HA_ATOMIC_SUB(&s->proxy->queue.length, xferred);
Willy Tarreau5472aa52020-10-24 12:57:41 +0200516 _HA_ATOMIC_SUB(&s->proxy->totpend, xferred);
Willy Tarreau96bca332020-10-21 12:01:28 +0200517 }
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200518 return xferred;
519}
520
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100521/* Try to dequeue pending connection attached to the stream <strm>. It must
522 * always exists here. If the pendconn is still linked to the server or the
523 * proxy queue, nothing is done and the function returns 1. Otherwise,
524 * <strm>->flags and <strm>->target are updated, the pendconn is released and 0
525 * is returned.
526 *
527 * This function must be called by the stream itself, so in the context of
528 * process_stream.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200529 */
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100530int pendconn_dequeue(struct stream *strm)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200531{
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100532 struct pendconn *p;
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200533 int is_unlinked;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100534
535 if (unlikely(!strm->pend_pos)) {
536 /* unexpected case because it is called by the stream itself and
537 * only the stream can release a pendconn. So it is only
538 * possible if a pendconn is released by someone else or if the
539 * stream is supposed to be queued but without its associated
540 * pendconn. In both cases it is a bug! */
541 abort();
Christopher Faulet8ba59142017-06-27 15:43:53 +0200542 }
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100543 p = strm->pend_pos;
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200544
545 /* note below : we need to grab the queue's lock to check for emptiness
546 * because we don't want a partial _grab_from_px() or _redistribute()
547 * to be called in parallel and show an empty list without having the
548 * time to finish. With this we know that if we see the element
549 * unlinked, these functions were completely done.
550 */
551 pendconn_queue_lock(p);
Patrick Hemmer0355dab2018-05-11 12:52:31 -0400552 is_unlinked = !p->node.node.leaf_p;
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200553 pendconn_queue_unlock(p);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100554
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200555 if (!is_unlinked)
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100556 return 1;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100557
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200558 /* the pendconn is not queued anymore and will not be so we're safe
559 * to proceed.
560 */
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100561 strm->flags &= ~(SF_DIRECT | SF_ASSIGNED | SF_ADDR_SET);
562 strm->flags |= p->strm_flags & (SF_DIRECT | SF_ASSIGNED | SF_ADDR_SET);
Willy Tarreau7867ceb2021-06-16 08:42:23 +0200563
564 if (p->target) {
565 /* a server picked this pendconn, it must skip LB */
566 strm->target = &p->target->obj_type;
567 strm->flags |= SF_ASSIGNED;
568 }
569
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100570 strm->pend_pos = NULL;
Willy Tarreaubafbe012017-11-24 17:34:44 +0100571 pool_free(pool_head_pendconn, p);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100572 return 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200573}
574
Patrick Hemmer268a7072018-05-11 12:52:31 -0400575static enum act_return action_set_priority_class(struct act_rule *rule, struct proxy *px,
576 struct session *sess, struct stream *s, int flags)
577{
578 struct sample *smp;
579
580 smp = sample_fetch_as_type(px, sess, s, SMP_OPT_DIR_REQ|SMP_OPT_FINAL, rule->arg.expr, SMP_T_SINT);
581 if (!smp)
582 return ACT_RET_CONT;
583
584 s->priority_class = queue_limit_class(smp->data.u.sint);
585 return ACT_RET_CONT;
586}
587
588static enum act_return action_set_priority_offset(struct act_rule *rule, struct proxy *px,
589 struct session *sess, struct stream *s, int flags)
590{
591 struct sample *smp;
592
593 smp = sample_fetch_as_type(px, sess, s, SMP_OPT_DIR_REQ|SMP_OPT_FINAL, rule->arg.expr, SMP_T_SINT);
594 if (!smp)
595 return ACT_RET_CONT;
596
597 s->priority_offset = queue_limit_offset(smp->data.u.sint);
598
599 return ACT_RET_CONT;
600}
601
602static enum act_parse_ret parse_set_priority_class(const char **args, int *arg, struct proxy *px,
603 struct act_rule *rule, char **err)
604{
605 unsigned int where = 0;
606
607 rule->arg.expr = sample_parse_expr((char **)args, arg, px->conf.args.file,
Willy Tarreaue3b57bf2020-02-14 16:50:14 +0100608 px->conf.args.line, err, &px->conf.args, NULL);
Patrick Hemmer268a7072018-05-11 12:52:31 -0400609 if (!rule->arg.expr)
610 return ACT_RET_PRS_ERR;
611
612 if (px->cap & PR_CAP_FE)
613 where |= SMP_VAL_FE_HRQ_HDR;
614 if (px->cap & PR_CAP_BE)
615 where |= SMP_VAL_BE_HRQ_HDR;
616
617 if (!(rule->arg.expr->fetch->val & where)) {
618 memprintf(err,
619 "fetch method '%s' extracts information from '%s', none of which is available here",
620 args[0], sample_src_names(rule->arg.expr->fetch->use));
621 free(rule->arg.expr);
622 return ACT_RET_PRS_ERR;
623 }
624
625 rule->action = ACT_CUSTOM;
626 rule->action_ptr = action_set_priority_class;
627 return ACT_RET_PRS_OK;
628}
629
630static enum act_parse_ret parse_set_priority_offset(const char **args, int *arg, struct proxy *px,
631 struct act_rule *rule, char **err)
632{
633 unsigned int where = 0;
634
635 rule->arg.expr = sample_parse_expr((char **)args, arg, px->conf.args.file,
Willy Tarreaue3b57bf2020-02-14 16:50:14 +0100636 px->conf.args.line, err, &px->conf.args, NULL);
Patrick Hemmer268a7072018-05-11 12:52:31 -0400637 if (!rule->arg.expr)
638 return ACT_RET_PRS_ERR;
639
640 if (px->cap & PR_CAP_FE)
641 where |= SMP_VAL_FE_HRQ_HDR;
642 if (px->cap & PR_CAP_BE)
643 where |= SMP_VAL_BE_HRQ_HDR;
644
645 if (!(rule->arg.expr->fetch->val & where)) {
646 memprintf(err,
647 "fetch method '%s' extracts information from '%s', none of which is available here",
648 args[0], sample_src_names(rule->arg.expr->fetch->use));
649 free(rule->arg.expr);
650 return ACT_RET_PRS_ERR;
651 }
652
653 rule->action = ACT_CUSTOM;
654 rule->action_ptr = action_set_priority_offset;
655 return ACT_RET_PRS_OK;
656}
657
658static struct action_kw_list tcp_cont_kws = {ILH, {
659 { "set-priority-class", parse_set_priority_class },
660 { "set-priority-offset", parse_set_priority_offset },
661 { /* END */ }
662}};
663
Willy Tarreau0108d902018-11-25 19:14:37 +0100664INITCALL1(STG_REGISTER, tcp_req_cont_keywords_register, &tcp_cont_kws);
665
Patrick Hemmer268a7072018-05-11 12:52:31 -0400666static struct action_kw_list http_req_kws = {ILH, {
667 { "set-priority-class", parse_set_priority_class },
668 { "set-priority-offset", parse_set_priority_offset },
669 { /* END */ }
670}};
671
Willy Tarreau0108d902018-11-25 19:14:37 +0100672INITCALL1(STG_REGISTER, http_req_keywords_register, &http_req_kws);
673
Patrick Hemmer268a7072018-05-11 12:52:31 -0400674static int
675smp_fetch_priority_class(const struct arg *args, struct sample *smp, const char *kw, void *private)
676{
677 if (!smp->strm)
678 return 0;
679
680 smp->data.type = SMP_T_SINT;
681 smp->data.u.sint = smp->strm->priority_class;
682
683 return 1;
684}
685
686static int
687smp_fetch_priority_offset(const struct arg *args, struct sample *smp, const char *kw, void *private)
688{
689 if (!smp->strm)
690 return 0;
691
692 smp->data.type = SMP_T_SINT;
693 smp->data.u.sint = smp->strm->priority_offset;
694
695 return 1;
696}
697
698
699static struct sample_fetch_kw_list smp_kws = {ILH, {
700 { "prio_class", smp_fetch_priority_class, 0, NULL, SMP_T_SINT, SMP_USE_INTRN, },
701 { "prio_offset", smp_fetch_priority_offset, 0, NULL, SMP_T_SINT, SMP_USE_INTRN, },
702 { /* END */},
703}};
704
Willy Tarreau0108d902018-11-25 19:14:37 +0100705INITCALL1(STG_REGISTER, sample_register_fetches, &smp_kws);
Patrick Hemmer268a7072018-05-11 12:52:31 -0400706
Willy Tarreaubaaee002006-06-26 02:48:02 +0200707/*
708 * Local variables:
709 * c-indent-level: 8
710 * c-basic-offset: 8
711 * End:
712 */