blob: 0022bff302912c76e5a74f42cfc910bd5f329e51 [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 Tarreau87b09662015-04-03 00:22:06 +020023#include <proto/stream.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
Thierry FOURNIERd6975962017-07-12 14:31:10 +020028/* This is the memory pool containing all the signal structs. These
29 * struct are used to store each requiered signal between two tasks.
30 */
31struct pool_head *pool2_notification;
32
Willy Tarreaua4613182009-03-21 18:13:21 +010033unsigned int nb_tasks = 0;
Christopher Faulet34c5cc92016-12-06 09:15:30 +010034unsigned int tasks_run_queue = 0;
35unsigned int tasks_run_queue_cur = 0; /* copy of the run queue size */
Willy Tarreauc7bdf092009-03-21 18:33:52 +010036unsigned int nb_tasks_cur = 0; /* copy of the tasks count */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010037unsigned int niced_tasks = 0; /* number of niced tasks in the run queue */
Emeric Brun01948972017-03-30 15:37:25 +020038
Emeric Brunc60def82017-09-27 14:59:38 +020039#ifdef USE_THREAD
40HA_SPINLOCK_T rq_lock; /* spin lock related to run queue */
41HA_SPINLOCK_T wq_lock; /* spin lock related to wait queue */
42#endif
Willy Tarreau964c9362007-01-07 00:38:00 +010043
Willy Tarreaue35c94a2009-03-21 10:01:42 +010044static struct eb_root timers; /* sorted timers tree */
45static struct eb_root rqueue; /* tree constituting the run queue */
46static unsigned int rqueue_ticks; /* insertion count */
Willy Tarreau9789f7b2008-06-24 08:17:16 +020047
Willy Tarreau4726f532009-03-07 17:25:21 +010048/* Puts the task <t> in run queue at a position depending on t->nice. <t> is
49 * returned. The nice value assigns boosts in 32th of the run queue size. A
Christopher Faulet34c5cc92016-12-06 09:15:30 +010050 * nice value of -1024 sets the task to -tasks_run_queue*32, while a nice value
51 * of 1024 sets the task to tasks_run_queue*32. The state flags are cleared, so
52 * the caller will have to set its flags after this call.
Willy Tarreau4726f532009-03-07 17:25:21 +010053 * The task must not already be in the run queue. If unsure, use the safer
54 * task_wakeup() function.
Willy Tarreau91e99932008-06-30 07:51:00 +020055 */
Willy Tarreau4df82062008-08-29 15:26:14 +020056struct task *__task_wakeup(struct task *t)
Willy Tarreaue33aece2007-04-30 13:15:14 +020057{
Christopher Faulet34c5cc92016-12-06 09:15:30 +010058 tasks_run_queue++;
Willy Tarreau4726f532009-03-07 17:25:21 +010059 t->rq.key = ++rqueue_ticks;
Willy Tarreau91e99932008-06-30 07:51:00 +020060
61 if (likely(t->nice)) {
62 int offset;
63
64 niced_tasks++;
65 if (likely(t->nice > 0))
Christopher Faulet34c5cc92016-12-06 09:15:30 +010066 offset = (unsigned)((tasks_run_queue * (unsigned int)t->nice) / 32U);
Willy Tarreau91e99932008-06-30 07:51:00 +020067 else
Christopher Faulet34c5cc92016-12-06 09:15:30 +010068 offset = -(unsigned)((tasks_run_queue * (unsigned int)-t->nice) / 32U);
Willy Tarreau4726f532009-03-07 17:25:21 +010069 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +020070 }
71
Emeric Brun01948972017-03-30 15:37:25 +020072 /* reset flag to pending ones
73 * Note: __task_wakeup must not be called
74 * if task is running
75 */
76 t->state = t->pending_state;
Willy Tarreaue35c94a2009-03-21 10:01:42 +010077 eb32_insert(&rqueue, &t->rq);
Willy Tarreau58b458d2008-06-29 22:40:23 +020078 return t;
Willy Tarreaue33aece2007-04-30 13:15:14 +020079}
Willy Tarreaud825eef2007-05-12 22:35:00 +020080
Willy Tarreau96bcfd72007-04-29 10:41:56 +020081/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +010082 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +020083 *
84 * Inserts a task into the wait queue at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +010085 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +010086 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +010087 * Last, tasks must not be queued further than the end of the tree, which is
88 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +010089 *
90 * This function should not be used directly, it is meant to be called by the
91 * inline version of task_queue() which performs a few cheap preliminary tests
92 * before deciding to call __task_queue().
Willy Tarreau96bcfd72007-04-29 10:41:56 +020093 */
Willy Tarreau531cf0c2009-03-08 16:35:27 +010094void __task_queue(struct task *task)
Willy Tarreau964c9362007-01-07 00:38:00 +010095{
Willy Tarreau531cf0c2009-03-08 16:35:27 +010096 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +010097 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +010098
99 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100100 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200101#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100102 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +0200103 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +0100104 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200105#endif
Willy Tarreauce44f122008-07-05 18:16:19 +0200106
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100107 eb32_insert(&timers, &task->wq);
SaVaGe1d7a4202009-10-06 18:53:37 +0300108
Willy Tarreau4726f532009-03-07 17:25:21 +0100109 return;
Willy Tarreau964c9362007-01-07 00:38:00 +0100110}
111
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200112/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200113 * Extract all expired timers from the timer queue, and wakes up all
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100114 * associated tasks. Returns the date of next event (or eternity).
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200115 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100116int wake_expired_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200117{
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200118 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200119 struct eb32_node *eb;
Emeric Brunc60def82017-09-27 14:59:38 +0200120 int ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200121
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100122 while (1) {
Emeric Brunc60def82017-09-27 14:59:38 +0200123 SPIN_LOCK(TASK_WQ_LOCK, &wq_lock);
124 lookup_next:
Emeric Brun01948972017-03-30 15:37:25 +0200125 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200126 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100127 /* we might have reached the end of the tree, typically because
128 * <now_ms> is in the first half and we're first scanning the last
129 * half. Let's loop back to the beginning of the tree now.
130 */
131 eb = eb32_first(&timers);
Emeric Brunc60def82017-09-27 14:59:38 +0200132 if (likely(!eb)) {
133 SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100134 break;
Emeric Brunc60def82017-09-27 14:59:38 +0200135 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100136 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200137
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100138 if (likely(tick_is_lt(now_ms, eb->key))) {
Emeric Brunc60def82017-09-27 14:59:38 +0200139 ret = eb->key;
140 SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100141 /* timer not expired yet, revisit it later */
Emeric Brunc60def82017-09-27 14:59:38 +0200142 return ret;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100143 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200144
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100145 /* timer looks expired, detach it from the queue */
146 task = eb32_entry(eb, struct task, wq);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100147 __task_unlink_wq(task);
Willy Tarreau41365222009-03-08 07:46:27 +0100148
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100149 /* It is possible that this task was left at an earlier place in the
150 * tree because a recent call to task_queue() has not moved it. This
151 * happens when the new expiration date is later than the old one.
152 * Since it is very unlikely that we reach a timeout anyway, it's a
153 * lot cheaper to proceed like this because we almost never update
154 * the tree. We may also find disabled expiration dates there. Since
155 * we have detached the task from the tree, we simply call task_queue
Willy Tarreau814c9782009-07-14 23:48:55 +0200156 * to take care of this. Note that we might occasionally requeue it at
157 * the same place, before <eb>, so we have to check if this happens,
158 * and adjust <eb>, otherwise we may skip it which is not what we want.
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200159 * We may also not requeue the task (and not point eb at it) if its
160 * expiration time is not set.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100161 */
162 if (!tick_is_expired(task->expire, now_ms)) {
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200163 if (!tick_isset(task->expire))
Emeric Brunc60def82017-09-27 14:59:38 +0200164 goto lookup_next;
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200165 __task_queue(task);
Emeric Brunc60def82017-09-27 14:59:38 +0200166 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200167 }
Emeric Brunc60def82017-09-27 14:59:38 +0200168 SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100169 task_wakeup(task, TASK_WOKEN_TIMER);
170 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200171
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100172 /* No task is expired */
173 return TICK_ETERNITY;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200174}
175
Willy Tarreau58b458d2008-06-29 22:40:23 +0200176/* The run queue is chronologically sorted in a tree. An insertion counter is
177 * used to assign a position to each task. This counter may be combined with
178 * other variables (eg: nice value) to set the final position in the tree. The
179 * counter may wrap without a problem, of course. We then limit the number of
Willy Tarreau91e99932008-06-30 07:51:00 +0200180 * tasks processed at once to 1/4 of the number of tasks in the queue, and to
181 * 200 max in any case, so that general latency remains low and so that task
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100182 * positions have a chance to be considered.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200183 *
184 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200185 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100186void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200187{
Willy Tarreau964c9362007-01-07 00:38:00 +0100188 struct task *t;
Emeric Brun01948972017-03-30 15:37:25 +0200189 int i;
190 int max_processed;
191 struct eb32_node *rq_next;
192 int rewind;
193 struct task *local_tasks[16];
194 int local_tasks_count;
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100195 tasks_run_queue_cur = tasks_run_queue; /* keep a copy for reporting */
Willy Tarreauc7bdf092009-03-21 18:33:52 +0100196 nb_tasks_cur = nb_tasks;
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100197 max_processed = tasks_run_queue;
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100198 if (!tasks_run_queue)
Willy Tarreau98c61212011-09-10 20:08:49 +0200199 return;
200
Willy Tarreau91e99932008-06-30 07:51:00 +0200201 if (max_processed > 200)
202 max_processed = 200;
203
204 if (likely(niced_tasks))
Willy Tarreau218859a2009-03-21 11:53:09 +0100205 max_processed = (max_processed + 3) / 4;
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200206
Emeric Brunc60def82017-09-27 14:59:38 +0200207 SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Emeric Brun01948972017-03-30 15:37:25 +0200208 while (max_processed > 0) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100209 /* Note: this loop is one of the fastest code path in
210 * the whole program. It should not be re-arranged
211 * without a good reason.
212 */
Emeric Brun01948972017-03-30 15:37:25 +0200213
214 rewind = 0;
215 rq_next = eb32_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK);
216 if (!rq_next) {
217 /* we might have reached the end of the tree, typically because
218 * <rqueue_ticks> is in the first half and we're first scanning
219 * the last half. Let's loop back to the beginning of the tree now.
220 */
221 rq_next = eb32_first(&rqueue);
Willy Tarreau501260b2015-02-23 16:07:01 +0100222 if (!rq_next) {
Emeric Brun01948972017-03-30 15:37:25 +0200223 break;
224 }
225 rewind = 1;
226 }
227
228 local_tasks_count = 0;
229 while (local_tasks_count < 16) {
230 t = eb32_entry(rq_next, struct task, rq);
231 rq_next = eb32_next(rq_next);
Emeric Brunc60def82017-09-27 14:59:38 +0200232 if (t->process_mask & (1UL << tid)) {
233 /* detach the task from the queue */
234 __task_unlink_rq(t);
235 t->state |= TASK_RUNNING;
236 t->pending_state = 0;
237 t->calls++;
238 local_tasks[local_tasks_count++] = t;
239 }
Emeric Brun01948972017-03-30 15:37:25 +0200240 if (!rq_next) {
241 if (rewind || !(rq_next = eb32_first(&rqueue))) {
Willy Tarreau501260b2015-02-23 16:07:01 +0100242 break;
Emeric Brun01948972017-03-30 15:37:25 +0200243 }
244 rewind = 1;
Willy Tarreau501260b2015-02-23 16:07:01 +0100245 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100246 }
Willy Tarreau58b458d2008-06-29 22:40:23 +0200247
Emeric Brun01948972017-03-30 15:37:25 +0200248 if (!local_tasks_count)
249 break;
Willy Tarreau58b458d2008-06-29 22:40:23 +0200250
Emeric Brunc60def82017-09-27 14:59:38 +0200251 SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Emeric Brun01948972017-03-30 15:37:25 +0200252
253 for (i = 0; i < local_tasks_count ; i++) {
254 t = local_tasks[i];
255 /* This is an optimisation to help the processor's branch
256 * predictor take this most common call.
257 */
258 if (likely(t->process == process_stream))
259 t = process_stream(t);
260 else
261 t = t->process(t);
262 local_tasks[i] = t;
263 }
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100264
Emeric Brun01948972017-03-30 15:37:25 +0200265 max_processed -= local_tasks_count;
Emeric Brunc60def82017-09-27 14:59:38 +0200266 SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Emeric Brun01948972017-03-30 15:37:25 +0200267 for (i = 0; i < local_tasks_count ; i++) {
268 t = local_tasks[i];
269 if (likely(t != NULL)) {
270 t->state &= ~TASK_RUNNING;
271 /* If there is a pending state
272 * we have to wake up the task
273 * immediatly, else we defer
274 * it into wait queue
275 */
276 if (t->pending_state)
277 __task_wakeup(t);
278 else
279 task_queue(t);
280 }
Willy Tarreau58b458d2008-06-29 22:40:23 +0200281 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100282 }
Emeric Brunc60def82017-09-27 14:59:38 +0200283 SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200284}
285
Willy Tarreau4726f532009-03-07 17:25:21 +0100286/* perform minimal intializations, report 0 in case of error, 1 if OK. */
287int init_task()
288{
289 memset(&timers, 0, sizeof(timers));
290 memset(&rqueue, 0, sizeof(rqueue));
Emeric Brunc60def82017-09-27 14:59:38 +0200291 SPIN_INIT(&wq_lock);
292 SPIN_INIT(&rq_lock);
Willy Tarreau4726f532009-03-07 17:25:21 +0100293 pool2_task = create_pool("task", sizeof(struct task), MEM_F_SHARED);
Thierry FOURNIERd6975962017-07-12 14:31:10 +0200294 if (!pool2_task)
295 return 0;
296 pool2_notification = create_pool("notification", sizeof(struct notification), MEM_F_SHARED);
297 if (!pool2_notification)
298 return 0;
299 return 1;
Willy Tarreau4726f532009-03-07 17:25:21 +0100300}
301
Willy Tarreaubaaee002006-06-26 02:48:02 +0200302/*
303 * Local variables:
304 * c-indent-level: 8
305 * c-basic-offset: 8
306 * End:
307 */