blob: 25776fe8f7ac9c57e8ee738b4ec31889cd17ca4f [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 Tarreau8d388052017-11-05 13:34:20 +010020#include <eb32sctree.h>
Willy Tarreau45cb4fb2009-10-26 21:10:04 +010021#include <eb32tree.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020022
Willy Tarreaud825eef2007-05-12 22:35:00 +020023#include <proto/proxy.h>
Willy Tarreau87b09662015-04-03 00:22:06 +020024#include <proto/stream.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020025#include <proto/task.h>
26
Willy Tarreau9789f7b2008-06-24 08:17:16 +020027struct pool_head *pool2_task;
Willy Tarreau96bcfd72007-04-29 10:41:56 +020028
Thierry FOURNIERd6975962017-07-12 14:31:10 +020029/* This is the memory pool containing all the signal structs. These
30 * struct are used to store each requiered signal between two tasks.
31 */
32struct pool_head *pool2_notification;
33
Willy Tarreaua4613182009-03-21 18:13:21 +010034unsigned int nb_tasks = 0;
Christopher Faulet3911ee82017-11-14 10:26:53 +010035unsigned long active_tasks_mask = 0; /* Mask of threads with active tasks */
Christopher Faulet34c5cc92016-12-06 09:15:30 +010036unsigned int tasks_run_queue = 0;
37unsigned int tasks_run_queue_cur = 0; /* copy of the run queue size */
Willy Tarreauc7bdf092009-03-21 18:33:52 +010038unsigned int nb_tasks_cur = 0; /* copy of the tasks count */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010039unsigned int niced_tasks = 0; /* number of niced tasks in the run queue */
Emeric Brun01948972017-03-30 15:37:25 +020040
Christopher Faulet9dcf9b62017-11-13 10:34:01 +010041__decl_hathreads(HA_SPINLOCK_T rq_lock); /* spin lock related to run queue */
42__decl_hathreads(HA_SPINLOCK_T wq_lock); /* spin lock related to wait queue */
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++;
Christopher Faulet3911ee82017-11-14 10:26:53 +010059 active_tasks_mask |= t->thread_mask;
Willy Tarreau4726f532009-03-07 17:25:21 +010060 t->rq.key = ++rqueue_ticks;
Willy Tarreau91e99932008-06-30 07:51:00 +020061
62 if (likely(t->nice)) {
63 int offset;
64
65 niced_tasks++;
66 if (likely(t->nice > 0))
Christopher Faulet34c5cc92016-12-06 09:15:30 +010067 offset = (unsigned)((tasks_run_queue * (unsigned int)t->nice) / 32U);
Willy Tarreau91e99932008-06-30 07:51:00 +020068 else
Christopher Faulet34c5cc92016-12-06 09:15:30 +010069 offset = -(unsigned)((tasks_run_queue * (unsigned int)-t->nice) / 32U);
Willy Tarreau4726f532009-03-07 17:25:21 +010070 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +020071 }
72
Emeric Brun01948972017-03-30 15:37:25 +020073 /* reset flag to pending ones
74 * Note: __task_wakeup must not be called
75 * if task is running
76 */
77 t->state = t->pending_state;
Willy Tarreau8d388052017-11-05 13:34:20 +010078 eb32sc_insert(&rqueue, &t->rq, t->thread_mask);
Willy Tarreau58b458d2008-06-29 22:40:23 +020079 return t;
Willy Tarreaue33aece2007-04-30 13:15:14 +020080}
Willy Tarreaud825eef2007-05-12 22:35:00 +020081
Willy Tarreau96bcfd72007-04-29 10:41:56 +020082/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +010083 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +020084 *
85 * Inserts a task into the wait queue at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +010086 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +010087 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +010088 * Last, tasks must not be queued further than the end of the tree, which is
89 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +010090 *
91 * This function should not be used directly, it is meant to be called by the
92 * inline version of task_queue() which performs a few cheap preliminary tests
93 * before deciding to call __task_queue().
Willy Tarreau96bcfd72007-04-29 10:41:56 +020094 */
Willy Tarreau531cf0c2009-03-08 16:35:27 +010095void __task_queue(struct task *task)
Willy Tarreau964c9362007-01-07 00:38:00 +010096{
Willy Tarreau531cf0c2009-03-08 16:35:27 +010097 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +010098 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +010099
100 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100101 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200102#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100103 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +0200104 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +0100105 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200106#endif
Willy Tarreauce44f122008-07-05 18:16:19 +0200107
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100108 eb32_insert(&timers, &task->wq);
SaVaGe1d7a4202009-10-06 18:53:37 +0300109
Willy Tarreau4726f532009-03-07 17:25:21 +0100110 return;
Willy Tarreau964c9362007-01-07 00:38:00 +0100111}
112
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200113/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200114 * Extract all expired timers from the timer queue, and wakes up all
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100115 * associated tasks. Returns the date of next event (or eternity).
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200116 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100117int wake_expired_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200118{
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200119 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200120 struct eb32_node *eb;
Willy Tarreaub992ba12017-11-05 19:09:27 +0100121 int ret = TICK_ETERNITY;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200122
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100123 while (1) {
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100124 HA_SPIN_LOCK(TASK_WQ_LOCK, &wq_lock);
Emeric Brunc60def82017-09-27 14:59:38 +0200125 lookup_next:
Emeric Brun01948972017-03-30 15:37:25 +0200126 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200127 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100128 /* 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);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100133 if (likely(!eb))
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100134 break;
135 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200136
Willy Tarreaub992ba12017-11-05 19:09:27 +0100137 if (tick_is_lt(now_ms, eb->key)) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100138 /* timer not expired yet, revisit it later */
Willy Tarreaub992ba12017-11-05 19:09:27 +0100139 ret = eb->key;
140 break;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100141 }
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);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100145 __task_unlink_wq(task);
Willy Tarreau41365222009-03-08 07:46:27 +0100146
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100147 /* It is possible that this task was left at an earlier place in the
148 * tree because a recent call to task_queue() has not moved it. This
149 * happens when the new expiration date is later than the old one.
150 * Since it is very unlikely that we reach a timeout anyway, it's a
151 * lot cheaper to proceed like this because we almost never update
152 * the tree. We may also find disabled expiration dates there. Since
153 * we have detached the task from the tree, we simply call task_queue
Willy Tarreau814c9782009-07-14 23:48:55 +0200154 * to take care of this. Note that we might occasionally requeue it at
155 * the same place, before <eb>, so we have to check if this happens,
156 * and adjust <eb>, otherwise we may skip it which is not what we want.
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200157 * We may also not requeue the task (and not point eb at it) if its
158 * expiration time is not set.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100159 */
160 if (!tick_is_expired(task->expire, now_ms)) {
Willy Tarreaub992ba12017-11-05 19:09:27 +0100161 if (tick_isset(task->expire))
162 __task_queue(task);
Emeric Brunc60def82017-09-27 14:59:38 +0200163 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200164 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100165 task_wakeup(task, TASK_WOKEN_TIMER);
Willy Tarreau51753452017-11-23 18:36:50 +0100166 HA_SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100167 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200168
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100169 HA_SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100170 return ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200171}
172
Willy Tarreau58b458d2008-06-29 22:40:23 +0200173/* The run queue is chronologically sorted in a tree. An insertion counter is
174 * used to assign a position to each task. This counter may be combined with
175 * other variables (eg: nice value) to set the final position in the tree. The
176 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100177 * tasks processed to 200 in any case, so that general latency remains low and
178 * so that task positions have a chance to be considered.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200179 *
180 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200181 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100182void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200183{
Willy Tarreau964c9362007-01-07 00:38:00 +0100184 struct task *t;
Emeric Brun01948972017-03-30 15:37:25 +0200185 int i;
186 int max_processed;
Willy Tarreau8d388052017-11-05 13:34:20 +0100187 struct eb32sc_node *rq_next;
Emeric Brun01948972017-03-30 15:37:25 +0200188 struct task *local_tasks[16];
189 int local_tasks_count;
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100190 int final_tasks_count;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100191
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100192 tasks_run_queue_cur = tasks_run_queue; /* keep a copy for reporting */
Willy Tarreauc7bdf092009-03-21 18:33:52 +0100193 nb_tasks_cur = nb_tasks;
Christopher Faulet8a48f672017-11-14 10:38:36 +0100194 max_processed = 200;
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100195 if (unlikely(global.nbthread <= 1)) {
196 /* when no lock is needed, this loop is much faster */
Christopher Faulet8a48f672017-11-14 10:38:36 +0100197 if (!(active_tasks_mask & tid_bit))
198 return;
199
Christopher Faulet3911ee82017-11-14 10:26:53 +0100200 active_tasks_mask &= ~tid_bit;
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100201 rq_next = eb32sc_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
202 while (1) {
203 if (!rq_next) {
204 /* we might have reached the end of the tree, typically because
205 * <rqueue_ticks> is in the first half and we're first scanning
206 * the last half. Let's loop back to the beginning of the tree now.
207 */
208 rq_next = eb32sc_first(&rqueue, tid_bit);
209 if (!rq_next)
210 break;
211 }
212
213 t = eb32sc_entry(rq_next, struct task, rq);
214 rq_next = eb32sc_next(rq_next, tid_bit);
215 __task_unlink_rq(t);
216 t->state |= TASK_RUNNING;
217 t->pending_state = 0;
218
219 t->calls++;
220 /* This is an optimisation to help the processor's branch
221 * predictor take this most common call.
222 */
223 if (likely(t->process == process_stream))
224 t = process_stream(t);
225 else
226 t = t->process(t);
227
228 if (likely(t != NULL)) {
229 t->state &= ~TASK_RUNNING;
230 /* If there is a pending state
231 * we have to wake up the task
232 * immediatly, else we defer
233 * it into wait queue
234 */
235 if (t->pending_state)
236 __task_wakeup(t);
237 else
238 task_queue(t);
239 }
240
241 max_processed--;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100242 if (max_processed <= 0) {
243 active_tasks_mask |= tid_bit;
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100244 break;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100245 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100246 }
247 return;
248 }
249
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100250 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Christopher Faulet8a48f672017-11-14 10:38:36 +0100251 if (!(active_tasks_mask & tid_bit)) {
252 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
253 return;
254 }
255
Christopher Faulet3911ee82017-11-14 10:26:53 +0100256 active_tasks_mask &= ~tid_bit;
257 while (1) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100258 /* Note: this loop is one of the fastest code path in
259 * the whole program. It should not be re-arranged
260 * without a good reason.
261 */
Willy Tarreau9e45b332017-11-08 14:05:19 +0100262
263 /* we have to restart looking up after every batch */
264 rq_next = eb32sc_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreauce4e0aa2017-11-05 23:57:00 +0100265 for (local_tasks_count = 0; local_tasks_count < 16; local_tasks_count++) {
266 if (unlikely(!rq_next)) {
267 /* either we just started or we reached the end
268 * of the tree, typically because <rqueue_ticks>
269 * is in the first half and we're first scanning
270 * the last half. Let's loop back to the beginning
271 * of the tree now.
272 */
Christopher Faulet919b7392017-11-14 10:17:48 +0100273 rq_next = eb32sc_first(&rqueue, tid_bit);
274 if (!rq_next)
275 break;
Emeric Brun01948972017-03-30 15:37:25 +0200276 }
Emeric Brun01948972017-03-30 15:37:25 +0200277
Willy Tarreau8d388052017-11-05 13:34:20 +0100278 t = eb32sc_entry(rq_next, struct task, rq);
279 rq_next = eb32sc_next(rq_next, tid_bit);
Willy Tarreauce4e0aa2017-11-05 23:57:00 +0100280
281 /* detach the task from the queue */
282 __task_unlink_rq(t);
283 local_tasks[local_tasks_count] = t;
284 t->state |= TASK_RUNNING;
285 t->pending_state = 0;
286 t->calls++;
287 max_processed--;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100288 }
Willy Tarreau58b458d2008-06-29 22:40:23 +0200289
Emeric Brun01948972017-03-30 15:37:25 +0200290 if (!local_tasks_count)
291 break;
Willy Tarreau58b458d2008-06-29 22:40:23 +0200292
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100293 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Emeric Brun01948972017-03-30 15:37:25 +0200294
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100295 final_tasks_count = 0;
Emeric Brun01948972017-03-30 15:37:25 +0200296 for (i = 0; i < local_tasks_count ; i++) {
297 t = local_tasks[i];
298 /* This is an optimisation to help the processor's branch
299 * predictor take this most common call.
300 */
301 if (likely(t->process == process_stream))
302 t = process_stream(t);
303 else
304 t = t->process(t);
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100305 if (t)
306 local_tasks[final_tasks_count++] = t;
Emeric Brun01948972017-03-30 15:37:25 +0200307 }
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100308
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100309 for (i = 0; i < final_tasks_count ; i++) {
Emeric Brun01948972017-03-30 15:37:25 +0200310 t = local_tasks[i];
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100311 /* If there is a pending state
312 * we have to wake up the task
313 * immediatly, else we defer
314 * it into wait queue
315 */
Willy Tarreau51753452017-11-23 18:36:50 +0100316 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
317 t->state &= ~TASK_RUNNING;
318 if (t->pending_state) {
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100319 __task_wakeup(t);
Willy Tarreau51753452017-11-23 18:36:50 +0100320 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
321 }
322 else {
323 /* we must never hold the RQ lock before the WQ lock */
324 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100325 task_queue(t);
Willy Tarreau51753452017-11-23 18:36:50 +0100326 }
Willy Tarreau58b458d2008-06-29 22:40:23 +0200327 }
Willy Tarreauce4e0aa2017-11-05 23:57:00 +0100328
Willy Tarreau51753452017-11-23 18:36:50 +0100329 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Christopher Faulet3911ee82017-11-14 10:26:53 +0100330 if (max_processed <= 0) {
331 active_tasks_mask |= tid_bit;
332 break;
333 }
334 }
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100335 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200336}
337
Willy Tarreau4726f532009-03-07 17:25:21 +0100338/* perform minimal intializations, report 0 in case of error, 1 if OK. */
339int init_task()
340{
341 memset(&timers, 0, sizeof(timers));
342 memset(&rqueue, 0, sizeof(rqueue));
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100343 HA_SPIN_INIT(&wq_lock);
344 HA_SPIN_INIT(&rq_lock);
Willy Tarreau4726f532009-03-07 17:25:21 +0100345 pool2_task = create_pool("task", sizeof(struct task), MEM_F_SHARED);
Thierry FOURNIERd6975962017-07-12 14:31:10 +0200346 if (!pool2_task)
347 return 0;
348 pool2_notification = create_pool("notification", sizeof(struct notification), MEM_F_SHARED);
349 if (!pool2_notification)
350 return 0;
351 return 1;
Willy Tarreau4726f532009-03-07 17:25:21 +0100352}
353
Willy Tarreaubaaee002006-06-26 02:48:02 +0200354/*
355 * Local variables:
356 * c-indent-level: 8
357 * c-basic-offset: 8
358 * End:
359 */