blob: b863d645af96287d48131db885bb1ed57ff1abed [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 Tarreau3201e4e2018-07-26 08:23:24 +020047 * 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 Tarreau4c7e4b72020-05-27 12:58:42 +020072#include <haproxy/api.h>
Willy Tarreauc761f842020-06-04 11:40:28 +020073#include <haproxy/http_rules.h>
Willy Tarreaud0ef4392020-06-02 09:38:52 +020074#include <haproxy/pool.h>
Willy Tarreaue6ce10b2020-06-04 15:33:47 +020075#include <haproxy/sample.h>
Willy Tarreaucea0e1b2020-06-04 17:25:40 +020076#include <haproxy/task.h>
Willy Tarreau8b550af2020-06-04 17:42:48 +020077#include <haproxy/tcp_rules.h>
Willy Tarreau92b4f132020-06-01 11:05:15 +020078#include <haproxy/time.h>
Willy Tarreau3f567e42020-05-28 15:29:19 +020079#include <haproxy/thread.h>
Willy Tarreau8d2b7772020-05-27 10:58:19 +020080#include <import/eb32tree.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020081
Christopher Fauletfc9cfe42019-07-16 14:54:53 +020082#include <proto/http_ana.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020083#include <proto/queue.h>
84#include <proto/server.h>
Willy Tarreau87b09662015-04-03 00:22:06 +020085#include <proto/stream.h>
Willy Tarreau9e000c62011-03-10 14:03:36 +010086#include <proto/stream_interface.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
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100127/* Remove the pendconn from the server/proxy queue. At this stage, the
128 * connection is not really dequeued. It will be done during the
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200129 * process_stream. It also decreases the pending count.
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100130 *
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200131 * The caller must own the lock on the queue containing the pendconn. The
132 * pendconn must still be queued.
Christopher Fauletf3a55db2017-06-09 14:26:38 +0200133 */
Willy Tarreau9624fae2018-07-25 08:04:20 +0200134static void __pendconn_unlink(struct pendconn *p)
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100135{
Patrick Hemmerda282f42018-05-11 12:52:31 -0400136 if (p->srv) {
137 p->strm->logs.srv_queue_pos += p->srv->queue_idx - p->queue_idx;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100138 p->srv->nbpend--;
Patrick Hemmerda282f42018-05-11 12:52:31 -0400139 } else {
140 p->strm->logs.prx_queue_pos += p->px->queue_idx - p->queue_idx;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100141 p->px->nbpend--;
Patrick Hemmerda282f42018-05-11 12:52:31 -0400142 }
Olivier Houchardb4df4922019-03-08 18:54:16 +0100143 _HA_ATOMIC_SUB(&p->px->totpend, 1);
Patrick Hemmer0355dab2018-05-11 12:52:31 -0400144 eb32_delete(&p->node);
Christopher Fauletf3a55db2017-06-09 14:26:38 +0200145}
146
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200147/* Locks the queue the pendconn element belongs to. This relies on both p->px
148 * and p->srv to be properly initialized (which is always the case once the
149 * element has been added).
150 */
151static inline void pendconn_queue_lock(struct pendconn *p)
152{
153 if (p->srv)
154 HA_SPIN_LOCK(SERVER_LOCK, &p->srv->lock);
155 else
156 HA_SPIN_LOCK(PROXY_LOCK, &p->px->lock);
157}
158
159/* Unlocks the queue the pendconn element belongs to. This relies on both p->px
160 * and p->srv to be properly initialized (which is always the case once the
161 * element has been added).
162 */
163static inline void pendconn_queue_unlock(struct pendconn *p)
164{
165 if (p->srv)
166 HA_SPIN_UNLOCK(SERVER_LOCK, &p->srv->lock);
167 else
168 HA_SPIN_UNLOCK(PROXY_LOCK, &p->px->lock);
169}
170
Willy Tarreau9624fae2018-07-25 08:04:20 +0200171/* Removes the pendconn from the server/proxy queue. At this stage, the
172 * connection is not really dequeued. It will be done during process_stream().
Willy Tarreau9ada0302019-11-14 14:58:39 +0100173 * This function takes all the required locks for the operation. The pendconn
174 * must be valid, though it doesn't matter if it was already unlinked. Prefer
175 * pendconn_cond_unlink() to first check <p>. When the locks are already held,
176 * please use __pendconn_unlink() instead.
Willy Tarreau9624fae2018-07-25 08:04:20 +0200177 */
178void pendconn_unlink(struct pendconn *p)
179{
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200180 pendconn_queue_lock(p);
Willy Tarreau9624fae2018-07-25 08:04:20 +0200181
Willy Tarreau9ada0302019-11-14 14:58:39 +0100182 if (p->node.node.leaf_p)
183 __pendconn_unlink(p);
Willy Tarreau9624fae2018-07-25 08:04:20 +0200184
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200185 pendconn_queue_unlock(p);
Willy Tarreau9624fae2018-07-25 08:04:20 +0200186}
187
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400188/* Retrieve the first pendconn from tree <pendconns>. Classes are always
189 * considered first, then the time offset. The time does wrap, so the
190 * lookup is performed twice, one to retrieve the first class and a second
191 * time to retrieve the earliest time in this class.
192 */
193static struct pendconn *pendconn_first(struct eb_root *pendconns)
194{
195 struct eb32_node *node, *node2 = NULL;
196 u32 key;
197
198 node = eb32_first(pendconns);
199 if (!node)
200 return NULL;
201
202 key = KEY_CLASS_OFFSET_BOUNDARY(node->key);
203 node2 = eb32_lookup_ge(pendconns, key);
204
205 if (!node2 ||
206 KEY_CLASS(node2->key) != KEY_CLASS(node->key)) {
207 /* no other key in the tree, or in this class */
208 return eb32_entry(node, struct pendconn, node);
209 }
210
211 /* found a better key */
212 return eb32_entry(node2, struct pendconn, node);
213}
214
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100215/* Process the next pending connection from either a server or a proxy, and
Christopher Fauletfd83f0b2018-03-19 15:22:09 +0100216 * returns a strictly positive value on success (see below). If no pending
217 * connection is found, 0 is returned. Note that neither <srv> nor <px> may be
218 * NULL. Priority is given to the oldest request in the queue if both <srv> and
219 * <px> have pending requests. This ensures that no request will be left
220 * unserved. The <px> queue is not considered if the server (or a tracked
221 * server) is not RUNNING, is disabled, or has a null weight (server going
222 * down). The <srv> queue is still considered in this case, because if some
223 * connections remain there, it means that some requests have been forced there
224 * after it was seen down (eg: due to option persist). The stream is
225 * immediately marked as "assigned", and both its <srv> and <srv_conn> are set
226 * to <srv>.
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100227 *
228 * This function must only be called if the server queue _AND_ the proxy queue
Christopher Fauletfd83f0b2018-03-19 15:22:09 +0100229 * are locked. Today it is only called by process_srv_queue. When a pending
230 * connection is dequeued, this function returns 1 if the pending connection can
231 * be handled by the current thread, else it returns 2.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200232 */
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100233static int pendconn_process_next_strm(struct server *srv, struct proxy *px)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200234{
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100235 struct pendconn *p = NULL;
Patrick Hemmerda282f42018-05-11 12:52:31 -0400236 struct pendconn *pp = NULL;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100237 struct server *rsrv;
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400238 u32 pkey, ppkey;
Willy Tarreaud132f742010-08-06 10:08:23 +0200239
Willy Tarreau44267702011-10-28 15:35:33 +0200240 rsrv = srv->track;
Willy Tarreaud132f742010-08-06 10:08:23 +0200241 if (!rsrv)
242 rsrv = srv;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200243
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200244 p = NULL;
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400245 if (srv->nbpend)
246 p = pendconn_first(&srv->pendconns);
Willy Tarreau7c669d72008-06-20 15:04:11 +0200247
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400248 pp = NULL;
Willy Tarreaua8694652018-08-07 10:44:58 +0200249 if (srv_currently_usable(rsrv) && px->nbpend &&
250 (!(srv->flags & SRV_F_BACKUP) ||
251 (!px->srv_act &&
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400252 (srv == px->lbprm.fbck || (px->options & PR_O_USE_ALL_BK)))))
253 pp = pendconn_first(&px->pendconns);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100254
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400255 if (!p && !pp)
256 return 0;
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200257
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400258 if (p && !pp)
259 goto use_p;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100260
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400261 if (pp && !p)
262 goto use_pp;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100263
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400264 if (KEY_CLASS(p->node.key) < KEY_CLASS(pp->node.key))
265 goto use_p;
266
267 if (KEY_CLASS(pp->node.key) < KEY_CLASS(p->node.key))
268 goto use_pp;
269
270 pkey = KEY_OFFSET(p->node.key);
271 ppkey = KEY_OFFSET(pp->node.key);
272
273 if (pkey < NOW_OFFSET_BOUNDARY())
274 pkey += 0x100000; // key in the future
275
276 if (ppkey < NOW_OFFSET_BOUNDARY())
277 ppkey += 0x100000; // key in the future
278
279 if (pkey <= ppkey)
280 goto use_p;
281
282 use_pp:
283 /* Let's switch from the server pendconn to the proxy pendconn */
284 p = pp;
285 use_p:
Willy Tarreau9624fae2018-07-25 08:04:20 +0200286 __pendconn_unlink(p);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100287 p->strm_flags |= SF_ASSIGNED;
Willy Tarreau88930dd2018-07-26 07:38:54 +0200288 p->target = srv;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100289
Patrick Hemmerda282f42018-05-11 12:52:31 -0400290 if (p != pp)
291 srv->queue_idx++;
292 else
293 px->queue_idx++;
294
Olivier Houchardb4df4922019-03-08 18:54:16 +0100295 _HA_ATOMIC_ADD(&srv->served, 1);
296 _HA_ATOMIC_ADD(&srv->proxy->served, 1);
297 __ha_barrier_atomic_store();
Willy Tarreau7c669d72008-06-20 15:04:11 +0200298 if (px->lbprm.server_take_conn)
299 px->lbprm.server_take_conn(srv);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100300 __stream_add_srv_conn(p->strm, srv);
Willy Tarreau7c669d72008-06-20 15:04:11 +0200301
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100302 task_wakeup(p->strm->task, TASK_WOKEN_RES);
Christopher Fauletfd83f0b2018-03-19 15:22:09 +0100303
Olivier Houchardecfe6732018-07-26 18:47:27 +0200304 return 1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200305}
306
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100307/* Manages a server's connection queue. This function will try to dequeue as
Christopher Faulet87566c92017-06-06 10:34:51 +0200308 * many pending streams as possible, and wake them up.
309 */
310void process_srv_queue(struct server *s)
311{
312 struct proxy *p = s->proxy;
Olivier Houchardecfe6732018-07-26 18:47:27 +0200313 int maxconn;
Christopher Faulet87566c92017-06-06 10:34:51 +0200314
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100315 HA_SPIN_LOCK(SERVER_LOCK, &s->lock);
Willy Tarreau5e83d992019-07-30 11:59:34 +0200316 HA_SPIN_LOCK(PROXY_LOCK, &p->lock);
Christopher Faulet87566c92017-06-06 10:34:51 +0200317 maxconn = srv_dynamic_maxconn(s);
318 while (s->served < maxconn) {
Christopher Fauletfd83f0b2018-03-19 15:22:09 +0100319 int ret = pendconn_process_next_strm(s, p);
320 if (!ret)
Christopher Faulet87566c92017-06-06 10:34:51 +0200321 break;
Christopher Faulet87566c92017-06-06 10:34:51 +0200322 }
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100323 HA_SPIN_UNLOCK(PROXY_LOCK, &p->lock);
Willy Tarreau5e83d992019-07-30 11:59:34 +0200324 HA_SPIN_UNLOCK(SERVER_LOCK, &s->lock);
Christopher Faulet87566c92017-06-06 10:34:51 +0200325}
326
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400327/* Adds the stream <strm> to the pending connection queue of server <strm>->srv
Willy Tarreau87b09662015-04-03 00:22:06 +0200328 * or to the one of <strm>->proxy if srv is NULL. All counters and back pointers
Willy Tarreaubaaee002006-06-26 02:48:02 +0200329 * are updated accordingly. Returns NULL if no memory is available, otherwise the
Willy Tarreau87b09662015-04-03 00:22:06 +0200330 * pendconn itself. If the stream was already marked as served, its flag is
331 * cleared. It is illegal to call this function with a non-NULL strm->srv_conn.
Patrick Hemmerda282f42018-05-11 12:52:31 -0400332 * The stream's queue position is counted with an offset of -1 because we want
333 * to make sure that being at the first position in the queue reports 1.
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100334 *
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400335 * The queue is sorted by the composition of the priority_class, and the current
336 * timestamp offset by strm->priority_offset. The timestamp is in milliseconds
337 * and truncated to 20 bits, so will wrap every 17m28s575ms.
338 * The offset can be positive or negative, and an offset of 0 puts it in the
339 * middle of this range (~ 8 min). Note that this also means if the adjusted
340 * timestamp wraps around, the request will be misinterpreted as being of
Joseph Herlantd8499ec2018-11-25 11:26:48 -0800341 * the highest priority for that priority class.
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400342 *
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100343 * This function must be called by the stream itself, so in the context of
344 * process_stream.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200345 */
Willy Tarreau87b09662015-04-03 00:22:06 +0200346struct pendconn *pendconn_add(struct stream *strm)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200347{
348 struct pendconn *p;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100349 struct proxy *px;
350 struct server *srv;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200351
Willy Tarreaubafbe012017-11-24 17:34:44 +0100352 p = pool_alloc(pool_head_pendconn);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200353 if (!p)
354 return NULL;
355
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200356 if (strm->flags & SF_ASSIGNED)
357 srv = objt_server(strm->target);
358 else
359 srv = NULL;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100360
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200361 px = strm->be;
Willy Tarreau88930dd2018-07-26 07:38:54 +0200362 p->target = NULL;
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200363 p->srv = srv;
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400364 p->node.key = MAKE_KEY(strm->priority_class, strm->priority_offset);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100365 p->px = px;
366 p->strm = strm;
367 p->strm_flags = strm->flags;
Willy Tarreau7c669d72008-06-20 15:04:11 +0200368
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200369 pendconn_queue_lock(p);
370
371 if (srv) {
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100372 srv->nbpend++;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100373 if (srv->nbpend > srv->counters.nbpend_max)
374 srv->counters.nbpend_max = srv->nbpend;
Patrick Hemmerda282f42018-05-11 12:52:31 -0400375 p->queue_idx = srv->queue_idx - 1; // for increment
Patrick Hemmer0355dab2018-05-11 12:52:31 -0400376 eb32_insert(&srv->pendconns, &p->node);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100377 }
378 else {
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100379 px->nbpend++;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100380 if (px->nbpend > px->be_counters.nbpend_max)
381 px->be_counters.nbpend_max = px->nbpend;
Patrick Hemmerda282f42018-05-11 12:52:31 -0400382 p->queue_idx = px->queue_idx - 1; // for increment
Patrick Hemmer0355dab2018-05-11 12:52:31 -0400383 eb32_insert(&px->pendconns, &p->node);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200384 }
Willy Tarreau7c6f8a22018-07-26 08:03:14 +0200385 strm->pend_pos = p;
386
387 pendconn_queue_unlock(p);
388
Olivier Houchardb4df4922019-03-08 18:54:16 +0100389 _HA_ATOMIC_ADD(&px->totpend, 1);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200390 return p;
391}
392
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200393/* Redistribute pending connections when a server goes down. The number of
Willy Tarreaudeca26c2018-08-21 18:11:03 +0200394 * connections redistributed is returned. It must be called with the server
395 * lock held.
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200396 */
397int pendconn_redistribute(struct server *s)
398{
Patrick Hemmer0355dab2018-05-11 12:52:31 -0400399 struct pendconn *p;
Willy Tarreaubff005a2019-05-27 08:10:11 +0200400 struct eb32_node *node, *nodeb;
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200401 int xferred = 0;
402
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100403 /* The REDISP option was specified. We will ignore cookie and force to
404 * balance or use the dispatcher. */
405 if ((s->proxy->options & (PR_O_REDISP|PR_O_PERSIST)) != PR_O_REDISP)
406 return 0;
407
Willy Tarreaubff005a2019-05-27 08:10:11 +0200408 for (node = eb32_first(&s->pendconns); node; node = nodeb) {
409 nodeb = eb32_next(node);
410
411 p = eb32_entry(node, struct pendconn, node);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100412 if (p->strm_flags & SF_FORCE_PRST)
413 continue;
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200414
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100415 /* it's left to the dispatcher to choose a server */
Willy Tarreau9624fae2018-07-25 08:04:20 +0200416 __pendconn_unlink(p);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100417 p->strm_flags &= ~(SF_DIRECT | SF_ASSIGNED | SF_ADDR_SET);
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200418
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100419 task_wakeup(p->strm->task, TASK_WOKEN_RES);
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200420 }
421 return xferred;
422}
423
424/* Check for pending connections at the backend, and assign some of them to
425 * the server coming up. The server's weight is checked before being assigned
426 * connections it may not be able to handle. The total number of transferred
Willy Tarreau5e83d992019-07-30 11:59:34 +0200427 * connections is returned. It must be called with the server lock held, and
428 * will take the proxy's lock.
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200429 */
430int pendconn_grab_from_px(struct server *s)
431{
Patrick Hemmer0355dab2018-05-11 12:52:31 -0400432 struct pendconn *p;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100433 int maxconn, xferred = 0;
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200434
Emeric Brun52a91d32017-08-31 14:41:55 +0200435 if (!srv_currently_usable(s))
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200436 return 0;
437
Willy Tarreaua8694652018-08-07 10:44:58 +0200438 /* if this is a backup server and there are active servers or at
439 * least another backup server was elected, then this one must
440 * not dequeue requests from the proxy.
441 */
442 if ((s->flags & SRV_F_BACKUP) &&
443 (s->proxy->srv_act ||
444 ((s != s->proxy->lbprm.fbck) && !(s->proxy->options & PR_O_USE_ALL_BK))))
445 return 0;
446
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100447 HA_SPIN_LOCK(PROXY_LOCK, &s->proxy->lock);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100448 maxconn = srv_dynamic_maxconn(s);
Patrick Hemmer248cb4c2018-05-11 12:52:31 -0400449 while ((p = pendconn_first(&s->proxy->pendconns))) {
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100450 if (s->maxconn && s->served + xferred >= maxconn)
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200451 break;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100452
Willy Tarreau9624fae2018-07-25 08:04:20 +0200453 __pendconn_unlink(p);
Willy Tarreau88930dd2018-07-26 07:38:54 +0200454 p->target = s;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100455
456 task_wakeup(p->strm->task, TASK_WOKEN_RES);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100457 xferred++;
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200458 }
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100459 HA_SPIN_UNLOCK(PROXY_LOCK, &s->proxy->lock);
Willy Tarreau4aac7db2014-05-16 11:48:10 +0200460 return xferred;
461}
462
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100463/* Try to dequeue pending connection attached to the stream <strm>. It must
464 * always exists here. If the pendconn is still linked to the server or the
465 * proxy queue, nothing is done and the function returns 1. Otherwise,
466 * <strm>->flags and <strm>->target are updated, the pendconn is released and 0
467 * is returned.
468 *
469 * This function must be called by the stream itself, so in the context of
470 * process_stream.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200471 */
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100472int pendconn_dequeue(struct stream *strm)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200473{
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100474 struct pendconn *p;
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200475 int is_unlinked;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100476
477 if (unlikely(!strm->pend_pos)) {
478 /* unexpected case because it is called by the stream itself and
479 * only the stream can release a pendconn. So it is only
480 * possible if a pendconn is released by someone else or if the
481 * stream is supposed to be queued but without its associated
482 * pendconn. In both cases it is a bug! */
483 abort();
Christopher Faulet8ba59142017-06-27 15:43:53 +0200484 }
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100485 p = strm->pend_pos;
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200486
487 /* note below : we need to grab the queue's lock to check for emptiness
488 * because we don't want a partial _grab_from_px() or _redistribute()
489 * to be called in parallel and show an empty list without having the
490 * time to finish. With this we know that if we see the element
491 * unlinked, these functions were completely done.
492 */
493 pendconn_queue_lock(p);
Patrick Hemmer0355dab2018-05-11 12:52:31 -0400494 is_unlinked = !p->node.node.leaf_p;
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200495 pendconn_queue_unlock(p);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100496
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200497 if (!is_unlinked)
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100498 return 1;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100499
Willy Tarreau3201e4e2018-07-26 08:23:24 +0200500 /* the pendconn is not queued anymore and will not be so we're safe
501 * to proceed.
502 */
Willy Tarreau88930dd2018-07-26 07:38:54 +0200503 if (p->target)
504 strm->target = &p->target->obj_type;
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100505
506 strm->flags &= ~(SF_DIRECT | SF_ASSIGNED | SF_ADDR_SET);
507 strm->flags |= p->strm_flags & (SF_DIRECT | SF_ASSIGNED | SF_ADDR_SET);
508 strm->pend_pos = NULL;
Willy Tarreaubafbe012017-11-24 17:34:44 +0100509 pool_free(pool_head_pendconn, p);
Christopher Faulet5cd4bbd2018-03-14 16:18:06 +0100510 return 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200511}
512
Patrick Hemmer268a7072018-05-11 12:52:31 -0400513static enum act_return action_set_priority_class(struct act_rule *rule, struct proxy *px,
514 struct session *sess, struct stream *s, int flags)
515{
516 struct sample *smp;
517
518 smp = sample_fetch_as_type(px, sess, s, SMP_OPT_DIR_REQ|SMP_OPT_FINAL, rule->arg.expr, SMP_T_SINT);
519 if (!smp)
520 return ACT_RET_CONT;
521
522 s->priority_class = queue_limit_class(smp->data.u.sint);
523 return ACT_RET_CONT;
524}
525
526static enum act_return action_set_priority_offset(struct act_rule *rule, struct proxy *px,
527 struct session *sess, struct stream *s, int flags)
528{
529 struct sample *smp;
530
531 smp = sample_fetch_as_type(px, sess, s, SMP_OPT_DIR_REQ|SMP_OPT_FINAL, rule->arg.expr, SMP_T_SINT);
532 if (!smp)
533 return ACT_RET_CONT;
534
535 s->priority_offset = queue_limit_offset(smp->data.u.sint);
536
537 return ACT_RET_CONT;
538}
539
540static enum act_parse_ret parse_set_priority_class(const char **args, int *arg, struct proxy *px,
541 struct act_rule *rule, char **err)
542{
543 unsigned int where = 0;
544
545 rule->arg.expr = sample_parse_expr((char **)args, arg, px->conf.args.file,
Willy Tarreaue3b57bf2020-02-14 16:50:14 +0100546 px->conf.args.line, err, &px->conf.args, NULL);
Patrick Hemmer268a7072018-05-11 12:52:31 -0400547 if (!rule->arg.expr)
548 return ACT_RET_PRS_ERR;
549
550 if (px->cap & PR_CAP_FE)
551 where |= SMP_VAL_FE_HRQ_HDR;
552 if (px->cap & PR_CAP_BE)
553 where |= SMP_VAL_BE_HRQ_HDR;
554
555 if (!(rule->arg.expr->fetch->val & where)) {
556 memprintf(err,
557 "fetch method '%s' extracts information from '%s', none of which is available here",
558 args[0], sample_src_names(rule->arg.expr->fetch->use));
559 free(rule->arg.expr);
560 return ACT_RET_PRS_ERR;
561 }
562
563 rule->action = ACT_CUSTOM;
564 rule->action_ptr = action_set_priority_class;
565 return ACT_RET_PRS_OK;
566}
567
568static enum act_parse_ret parse_set_priority_offset(const char **args, int *arg, struct proxy *px,
569 struct act_rule *rule, char **err)
570{
571 unsigned int where = 0;
572
573 rule->arg.expr = sample_parse_expr((char **)args, arg, px->conf.args.file,
Willy Tarreaue3b57bf2020-02-14 16:50:14 +0100574 px->conf.args.line, err, &px->conf.args, NULL);
Patrick Hemmer268a7072018-05-11 12:52:31 -0400575 if (!rule->arg.expr)
576 return ACT_RET_PRS_ERR;
577
578 if (px->cap & PR_CAP_FE)
579 where |= SMP_VAL_FE_HRQ_HDR;
580 if (px->cap & PR_CAP_BE)
581 where |= SMP_VAL_BE_HRQ_HDR;
582
583 if (!(rule->arg.expr->fetch->val & where)) {
584 memprintf(err,
585 "fetch method '%s' extracts information from '%s', none of which is available here",
586 args[0], sample_src_names(rule->arg.expr->fetch->use));
587 free(rule->arg.expr);
588 return ACT_RET_PRS_ERR;
589 }
590
591 rule->action = ACT_CUSTOM;
592 rule->action_ptr = action_set_priority_offset;
593 return ACT_RET_PRS_OK;
594}
595
596static struct action_kw_list tcp_cont_kws = {ILH, {
597 { "set-priority-class", parse_set_priority_class },
598 { "set-priority-offset", parse_set_priority_offset },
599 { /* END */ }
600}};
601
Willy Tarreau0108d902018-11-25 19:14:37 +0100602INITCALL1(STG_REGISTER, tcp_req_cont_keywords_register, &tcp_cont_kws);
603
Patrick Hemmer268a7072018-05-11 12:52:31 -0400604static struct action_kw_list http_req_kws = {ILH, {
605 { "set-priority-class", parse_set_priority_class },
606 { "set-priority-offset", parse_set_priority_offset },
607 { /* END */ }
608}};
609
Willy Tarreau0108d902018-11-25 19:14:37 +0100610INITCALL1(STG_REGISTER, http_req_keywords_register, &http_req_kws);
611
Patrick Hemmer268a7072018-05-11 12:52:31 -0400612static int
613smp_fetch_priority_class(const struct arg *args, struct sample *smp, const char *kw, void *private)
614{
615 if (!smp->strm)
616 return 0;
617
618 smp->data.type = SMP_T_SINT;
619 smp->data.u.sint = smp->strm->priority_class;
620
621 return 1;
622}
623
624static int
625smp_fetch_priority_offset(const struct arg *args, struct sample *smp, const char *kw, void *private)
626{
627 if (!smp->strm)
628 return 0;
629
630 smp->data.type = SMP_T_SINT;
631 smp->data.u.sint = smp->strm->priority_offset;
632
633 return 1;
634}
635
636
637static struct sample_fetch_kw_list smp_kws = {ILH, {
638 { "prio_class", smp_fetch_priority_class, 0, NULL, SMP_T_SINT, SMP_USE_INTRN, },
639 { "prio_offset", smp_fetch_priority_offset, 0, NULL, SMP_T_SINT, SMP_USE_INTRN, },
640 { /* END */},
641}};
642
Willy Tarreau0108d902018-11-25 19:14:37 +0100643INITCALL1(STG_REGISTER, sample_register_fetches, &smp_kws);
Patrick Hemmer268a7072018-05-11 12:52:31 -0400644
Willy Tarreaubaaee002006-06-26 02:48:02 +0200645/*
646 * Local variables:
647 * c-indent-level: 8
648 * c-basic-offset: 8
649 * End:
650 */