blob: 268f8d9e978394b99513396c3b66869d5c59a337 [file] [log] [blame]
Willy Tarreaubaaee002006-06-26 02:48:02 +02001/*
2 * Task management functions.
3 *
Willy Tarreau4726f532009-03-07 17:25:21 +01004 * 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 Tarreau87bed622009-03-08 22:25:28 +010013#include <string.h>
14
Willy Tarreau2dd0d472006-06-29 17:53:05 +020015#include <common/config.h>
Willy Tarreau9789f7b2008-06-24 08:17:16 +020016#include <common/eb32tree.h>
Willy Tarreauc6ca1a02007-05-13 19:43:47 +020017#include <common/memory.h>
Willy Tarreau2dd0d472006-06-29 17:53:05 +020018#include <common/mini-clist.h>
Willy Tarreau96bcfd72007-04-29 10:41:56 +020019#include <common/standard.h>
Willy Tarreaua6a6a932007-04-28 22:40:08 +020020#include <common/time.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020021
Willy Tarreaud825eef2007-05-12 22:35:00 +020022#include <proto/proxy.h>
Willy Tarreau531cf0c2009-03-08 16:35:27 +010023#include <proto/session.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020024#include <proto/task.h>
25
Willy Tarreau9789f7b2008-06-24 08:17:16 +020026struct pool_head *pool2_task;
Willy Tarreau96bcfd72007-04-29 10:41:56 +020027
Willy Tarreaua4613182009-03-21 18:13:21 +010028unsigned int nb_tasks = 0;
Willy Tarreau58b458d2008-06-29 22:40:23 +020029unsigned int run_queue = 0;
Willy Tarreauc7bdf092009-03-21 18:33:52 +010030unsigned int run_queue_cur = 0; /* copy of the run queue size */
31unsigned int nb_tasks_cur = 0; /* copy of the tasks count */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010032unsigned int niced_tasks = 0; /* number of niced tasks in the run queue */
Willy Tarreau26ca34e2009-03-21 12:51:40 +010033struct eb32_node *last_timer = NULL; /* optimization: last queued timer */
Willy Tarreau964c9362007-01-07 00:38:00 +010034
Willy Tarreaue35c94a2009-03-21 10:01:42 +010035static struct eb_root timers; /* sorted timers tree */
36static struct eb_root rqueue; /* tree constituting the run queue */
37static unsigned int rqueue_ticks; /* insertion count */
Willy Tarreau9789f7b2008-06-24 08:17:16 +020038
Willy Tarreau4726f532009-03-07 17:25:21 +010039/* Puts the task <t> in run queue at a position depending on t->nice. <t> is
40 * returned. The nice value assigns boosts in 32th of the run queue size. A
41 * nice value of -1024 sets the task to -run_queue*32, while a nice value of
42 * 1024 sets the task to run_queue*32. The state flags are cleared, so the
43 * caller will have to set its flags after this call.
44 * The task must not already be in the run queue. If unsure, use the safer
45 * task_wakeup() function.
Willy Tarreau91e99932008-06-30 07:51:00 +020046 */
Willy Tarreau4df82062008-08-29 15:26:14 +020047struct task *__task_wakeup(struct task *t)
Willy Tarreaue33aece2007-04-30 13:15:14 +020048{
Willy Tarreau58b458d2008-06-29 22:40:23 +020049 run_queue++;
Willy Tarreau4726f532009-03-07 17:25:21 +010050 t->rq.key = ++rqueue_ticks;
Willy Tarreau91e99932008-06-30 07:51:00 +020051
52 if (likely(t->nice)) {
53 int offset;
54
55 niced_tasks++;
56 if (likely(t->nice > 0))
57 offset = (unsigned)((run_queue * (unsigned int)t->nice) / 32U);
58 else
59 offset = -(unsigned)((run_queue * (unsigned int)-t->nice) / 32U);
Willy Tarreau4726f532009-03-07 17:25:21 +010060 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +020061 }
62
Willy Tarreaufdccded2008-08-29 18:19:04 +020063 /* clear state flags at the same time */
Willy Tarreau4726f532009-03-07 17:25:21 +010064 t->state &= ~TASK_WOKEN_ANY;
Willy Tarreau58b458d2008-06-29 22:40:23 +020065
Willy Tarreaue35c94a2009-03-21 10:01:42 +010066 eb32_insert(&rqueue, &t->rq);
Willy Tarreau58b458d2008-06-29 22:40:23 +020067 return t;
Willy Tarreaue33aece2007-04-30 13:15:14 +020068}
Willy Tarreaud825eef2007-05-12 22:35:00 +020069
Willy Tarreau96bcfd72007-04-29 10:41:56 +020070/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +010071 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +020072 *
73 * Inserts a task into the wait queue at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +010074 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +010075 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +010076 * Last, tasks must not be queued further than the end of the tree, which is
77 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +010078 *
79 * This function should not be used directly, it is meant to be called by the
80 * inline version of task_queue() which performs a few cheap preliminary tests
81 * before deciding to call __task_queue().
Willy Tarreau96bcfd72007-04-29 10:41:56 +020082 */
Willy Tarreau531cf0c2009-03-08 16:35:27 +010083void __task_queue(struct task *task)
Willy Tarreau964c9362007-01-07 00:38:00 +010084{
Willy Tarreau531cf0c2009-03-08 16:35:27 +010085 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +010086 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +010087
88 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010089 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +020090#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +010091 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +020092 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +010093 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +020094#endif
Willy Tarreauce44f122008-07-05 18:16:19 +020095
96 if (likely(last_timer &&
Willy Tarreau26ca34e2009-03-21 12:51:40 +010097 last_timer->node.bit < 0 &&
98 last_timer->key == task->wq.key &&
99 last_timer->node.node_p)) {
Willy Tarreauce44f122008-07-05 18:16:19 +0200100 /* Most often, last queued timer has the same expiration date, so
101 * if it's not queued at the root, let's queue a dup directly there.
Willy Tarreau26ca34e2009-03-21 12:51:40 +0100102 * Note that we can only use dups at the dup tree's root (most
103 * negative bit).
Willy Tarreauce44f122008-07-05 18:16:19 +0200104 */
Willy Tarreau26ca34e2009-03-21 12:51:40 +0100105 eb_insert_dup(&last_timer->node, &task->wq.node);
106 if (task->wq.node.bit < last_timer->node.bit)
107 last_timer = &task->wq;
Willy Tarreau4726f532009-03-07 17:25:21 +0100108 return;
Willy Tarreauce44f122008-07-05 18:16:19 +0200109 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100110 eb32_insert(&timers, &task->wq);
Willy Tarreau26ca34e2009-03-21 12:51:40 +0100111 if (!last_timer || (task->wq.node.bit < last_timer->node.bit))
112 last_timer = &task->wq;
Willy Tarreau4726f532009-03-07 17:25:21 +0100113 return;
Willy Tarreau964c9362007-01-07 00:38:00 +0100114}
115
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200116/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200117 * Extract all expired timers from the timer queue, and wakes up all
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100118 * associated tasks. Returns the date of next event (or eternity) in <next>.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200119 */
Willy Tarreau0c303ee2008-07-07 00:09:58 +0200120void wake_expired_tasks(int *next)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200121{
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200122 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200123 struct eb32_node *eb;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200124
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100125 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
126 while (1) {
127 if (unlikely(!eb)) {
128 /* we might have reached the end of the tree, typically because
129 * <now_ms> is in the first half and we're first scanning the last
130 * half. Let's loop back to the beginning of the tree now.
131 */
132 eb = eb32_first(&timers);
133 if (likely(!eb))
134 break;
135 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200136
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100137 if (likely(tick_is_lt(now_ms, eb->key))) {
138 /* timer not expired yet, revisit it later */
139 *next = eb->key;
140 return;
141 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200142
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100143 /* timer looks expired, detach it from the queue */
144 task = eb32_entry(eb, struct task, wq);
145 eb = eb32_next(eb);
146 __task_unlink_wq(task);
Willy Tarreau41365222009-03-08 07:46:27 +0100147
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100148 /* It is possible that this task was left at an earlier place in the
149 * tree because a recent call to task_queue() has not moved it. This
150 * happens when the new expiration date is later than the old one.
151 * Since it is very unlikely that we reach a timeout anyway, it's a
152 * lot cheaper to proceed like this because we almost never update
153 * the tree. We may also find disabled expiration dates there. Since
154 * we have detached the task from the tree, we simply call task_queue
Willy Tarreau814c9782009-07-14 23:48:55 +0200155 * to take care of this. Note that we might occasionally requeue it at
156 * the same place, before <eb>, so we have to check if this happens,
157 * and adjust <eb>, otherwise we may skip it which is not what we want.
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200158 * We may also not requeue the task (and not point eb at it) if its
159 * expiration time is not set.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100160 */
161 if (!tick_is_expired(task->expire, now_ms)) {
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200162 if (!tick_isset(task->expire))
163 continue;
164 __task_queue(task);
Willy Tarreau814c9782009-07-14 23:48:55 +0200165 if (!eb || eb->key > task->wq.key)
166 eb = &task->wq;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100167 continue;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200168 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100169 task_wakeup(task, TASK_WOKEN_TIMER);
170 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200171
Willy Tarreau28c41a42008-06-29 17:00:59 +0200172 /* We have found no task to expire in any tree */
Willy Tarreau0c303ee2008-07-07 00:09:58 +0200173 *next = TICK_ETERNITY;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200174 return;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200175}
176
Willy Tarreau58b458d2008-06-29 22:40:23 +0200177/* The run queue is chronologically sorted in a tree. An insertion counter is
178 * used to assign a position to each task. This counter may be combined with
179 * other variables (eg: nice value) to set the final position in the tree. The
180 * counter may wrap without a problem, of course. We then limit the number of
Willy Tarreau91e99932008-06-30 07:51:00 +0200181 * tasks processed at once to 1/4 of the number of tasks in the queue, and to
182 * 200 max in any case, so that general latency remains low and so that task
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100183 * positions have a chance to be considered.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200184 *
185 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200186 */
Willy Tarreau0c303ee2008-07-07 00:09:58 +0200187void process_runnable_tasks(int *next)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200188{
Willy Tarreau964c9362007-01-07 00:38:00 +0100189 struct task *t;
Willy Tarreau58b458d2008-06-29 22:40:23 +0200190 struct eb32_node *eb;
Willy Tarreau58b458d2008-06-29 22:40:23 +0200191 unsigned int max_processed;
Willy Tarreau26c25062009-03-08 09:38:41 +0100192 int expire;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200193
Willy Tarreauc7bdf092009-03-21 18:33:52 +0100194 run_queue_cur = run_queue; /* keep a copy for reporting */
195 nb_tasks_cur = nb_tasks;
Willy Tarreau91e99932008-06-30 07:51:00 +0200196 max_processed = run_queue;
197 if (max_processed > 200)
198 max_processed = 200;
199
200 if (likely(niced_tasks))
Willy Tarreau218859a2009-03-21 11:53:09 +0100201 max_processed = (max_processed + 3) / 4;
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200202
Willy Tarreau26c25062009-03-08 09:38:41 +0100203 expire = *next;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100204 eb = eb32_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK);
205 while (max_processed--) {
206 /* Note: this loop is one of the fastest code path in
207 * the whole program. It should not be re-arranged
208 * without a good reason.
209 */
210
211 if (unlikely(!eb)) {
212 /* we might have reached the end of the tree, typically because
213 * <rqueue_ticks> is in the first half and we're first scanning
214 * the last half. Let's loop back to the beginning of the tree now.
215 */
216 eb = eb32_first(&rqueue);
217 if (likely(!eb))
218 break;
219 }
Willy Tarreau58b458d2008-06-29 22:40:23 +0200220
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100221 /* detach the task from the queue */
222 t = eb32_entry(eb, struct task, rq);
223 eb = eb32_next(eb);
224 __task_unlink_rq(t);
Willy Tarreau58b458d2008-06-29 22:40:23 +0200225
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100226 t->state |= TASK_RUNNING;
227 /* This is an optimisation to help the processor's branch
228 * predictor take this most common call.
229 */
Willy Tarreau3884cba2009-03-28 17:54:35 +0100230 t->calls++;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100231 if (likely(t->process == process_session))
232 t = process_session(t);
233 else
234 t = t->process(t);
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100235
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100236 if (likely(t != NULL)) {
237 t->state &= ~TASK_RUNNING;
238 if (t->expire) {
239 task_queue(t);
240 expire = tick_first_2nz(expire, t->expire);
Willy Tarreau26c25062009-03-08 09:38:41 +0100241 }
Willy Tarreau135a1132009-03-21 13:26:05 +0100242
243 /* if the task has put itself back into the run queue, we want to ensure
244 * it will be served at the proper time, especially if it's reniced.
245 */
246 if (unlikely(task_in_rq(t)) && (!eb || tick_is_lt(t->rq.key, eb->key))) {
247 eb = eb32_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK);
248 }
Willy Tarreau58b458d2008-06-29 22:40:23 +0200249 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100250 }
Willy Tarreau26c25062009-03-08 09:38:41 +0100251 *next = expire;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200252}
253
Willy Tarreau4726f532009-03-07 17:25:21 +0100254/* perform minimal intializations, report 0 in case of error, 1 if OK. */
255int init_task()
256{
257 memset(&timers, 0, sizeof(timers));
258 memset(&rqueue, 0, sizeof(rqueue));
259 pool2_task = create_pool("task", sizeof(struct task), MEM_F_SHARED);
260 return pool2_task != NULL;
261}
262
Willy Tarreaubaaee002006-06-26 02:48:02 +0200263/*
264 * Local variables:
265 * c-indent-level: 8
266 * c-basic-offset: 8
267 * End:
268 */