blob: 74ae1c9a77f64f3e3d62dbcc530f0b3f49f56547 [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 Tarreauc6ca1a02007-05-13 19:43:47 +020016#include <common/memory.h>
Willy Tarreau2dd0d472006-06-29 17:53:05 +020017#include <common/mini-clist.h>
Willy Tarreau96bcfd72007-04-29 10:41:56 +020018#include <common/standard.h>
Willy Tarreaua6a6a932007-04-28 22:40:08 +020019#include <common/time.h>
Willy Tarreau45cb4fb2009-10-26 21:10:04 +010020#include <eb32tree.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 Tarreau501260b2015-02-23 16:07:01 +010034struct eb32_node *rq_next = NULL; /* optimization: next task except if delete/insert */
Willy Tarreau964c9362007-01-07 00:38:00 +010035
Willy Tarreaue35c94a2009-03-21 10:01:42 +010036static struct eb_root timers; /* sorted timers tree */
37static struct eb_root rqueue; /* tree constituting the run queue */
38static unsigned int rqueue_ticks; /* insertion count */
Willy Tarreau9789f7b2008-06-24 08:17:16 +020039
Willy Tarreau4726f532009-03-07 17:25:21 +010040/* Puts the task <t> in run queue at a position depending on t->nice. <t> is
41 * returned. The nice value assigns boosts in 32th of the run queue size. A
42 * nice value of -1024 sets the task to -run_queue*32, while a nice value of
43 * 1024 sets the task to run_queue*32. The state flags are cleared, so the
44 * caller will have to set its flags after this call.
45 * The task must not already be in the run queue. If unsure, use the safer
46 * task_wakeup() function.
Willy Tarreau91e99932008-06-30 07:51:00 +020047 */
Willy Tarreau4df82062008-08-29 15:26:14 +020048struct task *__task_wakeup(struct task *t)
Willy Tarreaue33aece2007-04-30 13:15:14 +020049{
Willy Tarreau58b458d2008-06-29 22:40:23 +020050 run_queue++;
Willy Tarreau4726f532009-03-07 17:25:21 +010051 t->rq.key = ++rqueue_ticks;
Willy Tarreau91e99932008-06-30 07:51:00 +020052
53 if (likely(t->nice)) {
54 int offset;
55
56 niced_tasks++;
57 if (likely(t->nice > 0))
58 offset = (unsigned)((run_queue * (unsigned int)t->nice) / 32U);
59 else
60 offset = -(unsigned)((run_queue * (unsigned int)-t->nice) / 32U);
Willy Tarreau4726f532009-03-07 17:25:21 +010061 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +020062 }
63
Willy Tarreaufdccded2008-08-29 18:19:04 +020064 /* clear state flags at the same time */
Willy Tarreau4726f532009-03-07 17:25:21 +010065 t->state &= ~TASK_WOKEN_ANY;
Willy Tarreau58b458d2008-06-29 22:40:23 +020066
Willy Tarreaue35c94a2009-03-21 10:01:42 +010067 eb32_insert(&rqueue, &t->rq);
Willy Tarreau58b458d2008-06-29 22:40:23 +020068 return t;
Willy Tarreaue33aece2007-04-30 13:15:14 +020069}
Willy Tarreaud825eef2007-05-12 22:35:00 +020070
Willy Tarreau96bcfd72007-04-29 10:41:56 +020071/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +010072 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +020073 *
74 * Inserts a task into the wait queue at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +010075 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +010076 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +010077 * Last, tasks must not be queued further than the end of the tree, which is
78 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +010079 *
80 * This function should not be used directly, it is meant to be called by the
81 * inline version of task_queue() which performs a few cheap preliminary tests
82 * before deciding to call __task_queue().
Willy Tarreau96bcfd72007-04-29 10:41:56 +020083 */
Willy Tarreau531cf0c2009-03-08 16:35:27 +010084void __task_queue(struct task *task)
Willy Tarreau964c9362007-01-07 00:38:00 +010085{
Willy Tarreau531cf0c2009-03-08 16:35:27 +010086 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +010087 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +010088
89 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010090 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +020091#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +010092 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +020093 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +010094 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +020095#endif
Willy Tarreauce44f122008-07-05 18:16:19 +020096
97 if (likely(last_timer &&
Willy Tarreau26ca34e2009-03-21 12:51:40 +010098 last_timer->node.bit < 0 &&
99 last_timer->key == task->wq.key &&
100 last_timer->node.node_p)) {
Willy Tarreauce44f122008-07-05 18:16:19 +0200101 /* Most often, last queued timer has the same expiration date, so
102 * if it's not queued at the root, let's queue a dup directly there.
Willy Tarreau26ca34e2009-03-21 12:51:40 +0100103 * Note that we can only use dups at the dup tree's root (most
104 * negative bit).
Willy Tarreauce44f122008-07-05 18:16:19 +0200105 */
Willy Tarreau26ca34e2009-03-21 12:51:40 +0100106 eb_insert_dup(&last_timer->node, &task->wq.node);
107 if (task->wq.node.bit < last_timer->node.bit)
108 last_timer = &task->wq;
Willy Tarreau4726f532009-03-07 17:25:21 +0100109 return;
Willy Tarreauce44f122008-07-05 18:16:19 +0200110 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100111 eb32_insert(&timers, &task->wq);
SaVaGe1d7a4202009-10-06 18:53:37 +0300112
113 /* Make sure we don't assign the last_timer to a node-less entry */
114 if (task->wq.node.node_p && (!last_timer || (task->wq.node.bit < last_timer->node.bit)))
Willy Tarreau26ca34e2009-03-21 12:51:40 +0100115 last_timer = &task->wq;
Willy Tarreau4726f532009-03-07 17:25:21 +0100116 return;
Willy Tarreau964c9362007-01-07 00:38:00 +0100117}
118
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200119/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200120 * Extract all expired timers from the timer queue, and wakes up all
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100121 * associated tasks. Returns the date of next event (or eternity) in <next>.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200122 */
Willy Tarreau0c303ee2008-07-07 00:09:58 +0200123void wake_expired_tasks(int *next)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200124{
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200125 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200126 struct eb32_node *eb;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200127
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100128 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
129 while (1) {
130 if (unlikely(!eb)) {
131 /* we might have reached the end of the tree, typically because
132 * <now_ms> is in the first half and we're first scanning the last
133 * half. Let's loop back to the beginning of the tree now.
134 */
135 eb = eb32_first(&timers);
136 if (likely(!eb))
137 break;
138 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200139
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100140 if (likely(tick_is_lt(now_ms, eb->key))) {
141 /* timer not expired yet, revisit it later */
142 *next = eb->key;
143 return;
144 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200145
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100146 /* timer looks expired, detach it from the queue */
147 task = eb32_entry(eb, struct task, wq);
148 eb = eb32_next(eb);
149 __task_unlink_wq(task);
Willy Tarreau41365222009-03-08 07:46:27 +0100150
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100151 /* It is possible that this task was left at an earlier place in the
152 * tree because a recent call to task_queue() has not moved it. This
153 * happens when the new expiration date is later than the old one.
154 * Since it is very unlikely that we reach a timeout anyway, it's a
155 * lot cheaper to proceed like this because we almost never update
156 * the tree. We may also find disabled expiration dates there. Since
157 * we have detached the task from the tree, we simply call task_queue
Willy Tarreau814c9782009-07-14 23:48:55 +0200158 * to take care of this. Note that we might occasionally requeue it at
159 * the same place, before <eb>, so we have to check if this happens,
160 * and adjust <eb>, otherwise we may skip it which is not what we want.
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200161 * We may also not requeue the task (and not point eb at it) if its
162 * expiration time is not set.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100163 */
164 if (!tick_is_expired(task->expire, now_ms)) {
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200165 if (!tick_isset(task->expire))
166 continue;
167 __task_queue(task);
Willy Tarreau814c9782009-07-14 23:48:55 +0200168 if (!eb || eb->key > task->wq.key)
169 eb = &task->wq;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100170 continue;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200171 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100172 task_wakeup(task, TASK_WOKEN_TIMER);
173 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200174
Willy Tarreau28c41a42008-06-29 17:00:59 +0200175 /* We have found no task to expire in any tree */
Willy Tarreau0c303ee2008-07-07 00:09:58 +0200176 *next = TICK_ETERNITY;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200177 return;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200178}
179
Willy Tarreau58b458d2008-06-29 22:40:23 +0200180/* The run queue is chronologically sorted in a tree. An insertion counter is
181 * used to assign a position to each task. This counter may be combined with
182 * other variables (eg: nice value) to set the final position in the tree. The
183 * counter may wrap without a problem, of course. We then limit the number of
Willy Tarreau91e99932008-06-30 07:51:00 +0200184 * tasks processed at once to 1/4 of the number of tasks in the queue, and to
185 * 200 max in any case, so that general latency remains low and so that task
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100186 * positions have a chance to be considered.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200187 *
188 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200189 */
Willy Tarreau0c303ee2008-07-07 00:09:58 +0200190void process_runnable_tasks(int *next)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200191{
Willy Tarreau964c9362007-01-07 00:38:00 +0100192 struct task *t;
Willy Tarreau58b458d2008-06-29 22:40:23 +0200193 unsigned int max_processed;
Willy Tarreau26c25062009-03-08 09:38:41 +0100194 int expire;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200195
Willy Tarreauc7bdf092009-03-21 18:33:52 +0100196 run_queue_cur = run_queue; /* keep a copy for reporting */
197 nb_tasks_cur = nb_tasks;
Willy Tarreau91e99932008-06-30 07:51:00 +0200198 max_processed = run_queue;
Willy Tarreau98c61212011-09-10 20:08:49 +0200199
200 if (!run_queue)
201 return;
202
Willy Tarreau91e99932008-06-30 07:51:00 +0200203 if (max_processed > 200)
204 max_processed = 200;
205
206 if (likely(niced_tasks))
Willy Tarreau218859a2009-03-21 11:53:09 +0100207 max_processed = (max_processed + 3) / 4;
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200208
Willy Tarreau26c25062009-03-08 09:38:41 +0100209 expire = *next;
Willy Tarreau501260b2015-02-23 16:07:01 +0100210
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100211 while (max_processed--) {
212 /* Note: this loop is one of the fastest code path in
213 * the whole program. It should not be re-arranged
214 * without a good reason.
215 */
Willy Tarreau501260b2015-02-23 16:07:01 +0100216 if (unlikely(!rq_next)) {
217 rq_next = eb32_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK);
218 if (!rq_next) {
219 /* we might have reached the end of the tree, typically because
220 * <rqueue_ticks> is in the first half and we're first scanning
221 * the last half. Let's loop back to the beginning of the tree now.
222 */
223 rq_next = eb32_first(&rqueue);
224 if (!rq_next)
225 break;
226 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100227 }
Willy Tarreau58b458d2008-06-29 22:40:23 +0200228
Willy Tarreau501260b2015-02-23 16:07:01 +0100229 /* detach the task from the queue after updating the pointer to
230 * the next entry.
231 */
232 t = eb32_entry(rq_next, struct task, rq);
233 rq_next = eb32_next(rq_next);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100234 __task_unlink_rq(t);
Willy Tarreau58b458d2008-06-29 22:40:23 +0200235
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100236 t->state |= TASK_RUNNING;
237 /* This is an optimisation to help the processor's branch
238 * predictor take this most common call.
239 */
Willy Tarreau3884cba2009-03-28 17:54:35 +0100240 t->calls++;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100241 if (likely(t->process == process_session))
242 t = process_session(t);
243 else
244 t = t->process(t);
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100245
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100246 if (likely(t != NULL)) {
247 t->state &= ~TASK_RUNNING;
248 if (t->expire) {
249 task_queue(t);
250 expire = tick_first_2nz(expire, t->expire);
Willy Tarreau26c25062009-03-08 09:38:41 +0100251 }
Willy Tarreau58b458d2008-06-29 22:40:23 +0200252 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100253 }
Willy Tarreau26c25062009-03-08 09:38:41 +0100254 *next = expire;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200255}
256
Willy Tarreau4726f532009-03-07 17:25:21 +0100257/* perform minimal intializations, report 0 in case of error, 1 if OK. */
258int init_task()
259{
260 memset(&timers, 0, sizeof(timers));
261 memset(&rqueue, 0, sizeof(rqueue));
262 pool2_task = create_pool("task", sizeof(struct task), MEM_F_SHARED);
263 return pool2_task != NULL;
264}
265
Willy Tarreaubaaee002006-06-26 02:48:02 +0200266/*
267 * Local variables:
268 * c-indent-level: 8
269 * c-basic-offset: 8
270 * End:
271 */