blob: f055ebb2b10697b75cd22636ffa8a64408c0a370 [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 Tarreau4c7e4b72020-05-27 12:58:42 +020015#include <haproxy/api.h>
Willy Tarreaud0ef4392020-06-02 09:38:52 +020016#include <haproxy/pool.h>
Willy Tarreau853b2972020-05-27 18:01:47 +020017#include <haproxy/list.h>
Willy Tarreaucea0e1b2020-06-04 17:25:40 +020018#include <haproxy/task.h>
Willy Tarreau48fbcae2020-06-03 18:09:46 +020019#include <haproxy/tools.h>
Willy Tarreau92b4f132020-06-01 11:05:15 +020020#include <haproxy/time.h>
Willy Tarreau8d2b7772020-05-27 10:58:19 +020021#include <import/eb32sctree.h>
22#include <import/eb32tree.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020023
Willy Tarreau0f6ffd62020-06-03 19:33:00 +020024#include <haproxy/fd.h>
Willy Tarreau66347942020-06-01 12:18:08 +020025#include <haproxy/freq_ctr.h>
Willy Tarreaud825eef2007-05-12 22:35:00 +020026#include <proto/proxy.h>
Willy Tarreau87b09662015-04-03 00:22:06 +020027#include <proto/stream.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020028
Willy Tarreau8ceae722018-11-26 11:58:30 +010029DECLARE_POOL(pool_head_task, "task", sizeof(struct task));
30DECLARE_POOL(pool_head_tasklet, "tasklet", sizeof(struct tasklet));
Willy Tarreau96bcfd72007-04-29 10:41:56 +020031
Thierry FOURNIERd6975962017-07-12 14:31:10 +020032/* This is the memory pool containing all the signal structs. These
Joseph Herlantcf92b6d2018-11-15 14:19:23 -080033 * struct are used to store each required signal between two tasks.
Thierry FOURNIERd6975962017-07-12 14:31:10 +020034 */
Willy Tarreau8ceae722018-11-26 11:58:30 +010035DECLARE_POOL(pool_head_notification, "notification", sizeof(struct notification));
Thierry FOURNIERd6975962017-07-12 14:31:10 +020036
Willy Tarreaua4613182009-03-21 18:13:21 +010037unsigned int nb_tasks = 0;
Olivier Houchardeba0c0b2018-07-26 18:53:28 +020038volatile unsigned long global_tasks_mask = 0; /* Mask of threads with tasks in the global runqueue */
Christopher Faulet34c5cc92016-12-06 09:15:30 +010039unsigned int tasks_run_queue = 0;
40unsigned int tasks_run_queue_cur = 0; /* copy of the run queue size */
Willy Tarreauc7bdf092009-03-21 18:33:52 +010041unsigned int nb_tasks_cur = 0; /* copy of the tasks count */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010042unsigned int niced_tasks = 0; /* number of niced tasks in the run queue */
Emeric Brun01948972017-03-30 15:37:25 +020043
Willy Tarreaud022e9c2019-09-24 08:25:15 +020044THREAD_LOCAL struct task_per_thread *sched = &task_per_thread[0]; /* scheduler context for the current thread */
Willy Tarreau6d1222c2017-11-26 10:08:06 +010045
Willy Tarreau86abe442018-11-25 20:12:18 +010046__decl_aligned_spinlock(rq_lock); /* spin lock related to run queue */
Willy Tarreauef28dc12019-05-28 18:48:07 +020047__decl_aligned_rwlock(wq_lock); /* RW lock related to the wait queue */
Willy Tarreau964c9362007-01-07 00:38:00 +010048
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020049#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +020050struct eb_root timers; /* sorted timers tree, global */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020051struct eb_root rqueue; /* tree constituting the run queue */
Olivier Houchard77551ee2018-07-26 15:59:38 +020052int global_rqueue_size; /* Number of element sin the global runqueue */
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020053#endif
Willy Tarreaub20aa9e2018-10-15 14:52:21 +020054
Willy Tarreaue35c94a2009-03-21 10:01:42 +010055static unsigned int rqueue_ticks; /* insertion count */
Willy Tarreau8d8747a2018-10-15 16:12:48 +020056
57struct task_per_thread task_per_thread[MAX_THREADS];
Willy Tarreau9789f7b2008-06-24 08:17:16 +020058
Willy Tarreau4726f532009-03-07 17:25:21 +010059/* Puts the task <t> in run queue at a position depending on t->nice. <t> is
60 * returned. The nice value assigns boosts in 32th of the run queue size. A
Christopher Faulet34c5cc92016-12-06 09:15:30 +010061 * nice value of -1024 sets the task to -tasks_run_queue*32, while a nice value
62 * of 1024 sets the task to tasks_run_queue*32. The state flags are cleared, so
63 * the caller will have to set its flags after this call.
Willy Tarreau4726f532009-03-07 17:25:21 +010064 * The task must not already be in the run queue. If unsure, use the safer
65 * task_wakeup() function.
Willy Tarreau91e99932008-06-30 07:51:00 +020066 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020067void __task_wakeup(struct task *t, struct eb_root *root)
Willy Tarreaue33aece2007-04-30 13:15:14 +020068{
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020069#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020070 if (root == &rqueue) {
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020071 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020072 }
Willy Tarreau2d1fd0a2019-04-15 09:18:31 +020073#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020074 /* Make sure if the task isn't in the runqueue, nobody inserts it
75 * in the meanwhile.
76 */
Olivier Houchard4c2832852019-03-08 18:48:47 +010077 _HA_ATOMIC_ADD(&tasks_run_queue, 1);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +020078#ifdef USE_THREAD
79 if (root == &rqueue) {
Olivier Houchardde82aea2019-04-17 19:10:22 +020080 global_tasks_mask |= t->thread_mask;
Olivier Houcharded1a6a02019-04-18 14:12:51 +020081 __ha_barrier_store();
Olivier Houchardc4aac9e2018-07-26 15:25:49 +020082 }
83#endif
Olivier Houchard4c2832852019-03-08 18:48:47 +010084 t->rq.key = _HA_ATOMIC_ADD(&rqueue_ticks, 1);
Willy Tarreau91e99932008-06-30 07:51:00 +020085
86 if (likely(t->nice)) {
87 int offset;
88
Olivier Houchard4c2832852019-03-08 18:48:47 +010089 _HA_ATOMIC_ADD(&niced_tasks, 1);
Willy Tarreau2d1fd0a2019-04-15 09:18:31 +020090 offset = t->nice * (int)global.tune.runqueue_depth;
Willy Tarreau4726f532009-03-07 17:25:21 +010091 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +020092 }
93
Willy Tarreaud9add3a2019-04-25 08:57:41 +020094 if (task_profiling_mask & tid_bit)
Willy Tarreau9efd7452018-05-31 14:48:54 +020095 t->call_date = now_mono_time();
96
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020097 eb32sc_insert(root, &t->rq, t->thread_mask);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020098#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020099 if (root == &rqueue) {
100 global_rqueue_size++;
Olivier Houchard4c2832852019-03-08 18:48:47 +0100101 _HA_ATOMIC_OR(&t->state, TASK_GLOBAL);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200102 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200103 } else
104#endif
105 {
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200106 int nb = ((void *)root - (void *)&task_per_thread[0].rqueue) / sizeof(task_per_thread[0]);
107 task_per_thread[nb].rqueue_size++;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200108 }
Willy Tarreau85d9b842018-07-27 17:14:41 +0200109#ifdef USE_THREAD
Olivier Houchard79321b92018-07-26 17:55:11 +0200110 /* If all threads that are supposed to handle this task are sleeping,
111 * wake one.
112 */
113 if ((((t->thread_mask & all_threads_mask) & sleeping_thread_mask) ==
Olivier Houchard1b327902019-03-15 00:23:10 +0100114 (t->thread_mask & all_threads_mask))) {
115 unsigned long m = (t->thread_mask & all_threads_mask) &~ tid_bit;
116
117 m = (m & (m - 1)) ^ m; // keep lowest bit set
118 _HA_ATOMIC_AND(&sleeping_thread_mask, ~m);
119 wake_thread(my_ffsl(m) - 1);
120 }
Willy Tarreau85d9b842018-07-27 17:14:41 +0200121#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200122 return;
Willy Tarreaue33aece2007-04-30 13:15:14 +0200123}
Willy Tarreaud825eef2007-05-12 22:35:00 +0200124
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200125/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100126 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200127 *
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200128 * Inserts a task into wait queue <wq> at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +0100129 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100130 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100131 * Last, tasks must not be queued further than the end of the tree, which is
132 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100133 *
134 * This function should not be used directly, it is meant to be called by the
135 * inline version of task_queue() which performs a few cheap preliminary tests
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200136 * before deciding to call __task_queue(). Moreover this function doesn't care
137 * at all about locking so the caller must be careful when deciding whether to
138 * lock or not around this call.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200139 */
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200140void __task_queue(struct task *task, struct eb_root *wq)
Willy Tarreau964c9362007-01-07 00:38:00 +0100141{
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100142 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +0100143 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +0100144
145 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100146 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200147#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100148 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +0200149 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +0100150 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200151#endif
Willy Tarreauce44f122008-07-05 18:16:19 +0200152
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200153 eb32_insert(wq, &task->wq);
Willy Tarreau964c9362007-01-07 00:38:00 +0100154}
155
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200156/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200157 * Extract all expired timers from the timer queue, and wakes up all
Willy Tarreauc49ba522019-12-11 08:12:23 +0100158 * associated tasks.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200159 */
Willy Tarreauc49ba522019-12-11 08:12:23 +0100160void wake_expired_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200161{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200162 struct task_per_thread * const tt = sched; // thread's tasks
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200163 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200164 struct eb32_node *eb;
Willy Tarreauaf613e82020-06-05 08:40:51 +0200165 __decl_thread(int key);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200166
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100167 while (1) {
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200168 lookup_next_local:
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200169 eb = eb32_lookup_ge(&tt->timers, now_ms - TIMER_LOOK_BACK);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200170 if (!eb) {
171 /* we might have reached the end of the tree, typically because
172 * <now_ms> is in the first half and we're first scanning the last
173 * half. Let's loop back to the beginning of the tree now.
174 */
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200175 eb = eb32_first(&tt->timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200176 if (likely(!eb))
177 break;
178 }
179
Willy Tarreauc49ba522019-12-11 08:12:23 +0100180 if (tick_is_lt(now_ms, eb->key))
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200181 break;
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200182
183 /* timer looks expired, detach it from the queue */
184 task = eb32_entry(eb, struct task, wq);
185 __task_unlink_wq(task);
186
187 /* It is possible that this task was left at an earlier place in the
188 * tree because a recent call to task_queue() has not moved it. This
189 * happens when the new expiration date is later than the old one.
190 * Since it is very unlikely that we reach a timeout anyway, it's a
191 * lot cheaper to proceed like this because we almost never update
192 * the tree. We may also find disabled expiration dates there. Since
193 * we have detached the task from the tree, we simply call task_queue
194 * to take care of this. Note that we might occasionally requeue it at
195 * the same place, before <eb>, so we have to check if this happens,
196 * and adjust <eb>, otherwise we may skip it which is not what we want.
197 * We may also not requeue the task (and not point eb at it) if its
198 * expiration time is not set.
199 */
200 if (!tick_is_expired(task->expire, now_ms)) {
201 if (tick_isset(task->expire))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200202 __task_queue(task, &tt->timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200203 goto lookup_next_local;
204 }
205 task_wakeup(task, TASK_WOKEN_TIMER);
206 }
207
208#ifdef USE_THREAD
Willy Tarreau1e928c02019-05-28 18:57:25 +0200209 if (eb_is_empty(&timers))
210 goto leave;
211
212 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
213 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
214 if (!eb) {
215 eb = eb32_first(&timers);
216 if (likely(!eb)) {
217 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
218 goto leave;
219 }
220 }
221 key = eb->key;
222 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
223
Willy Tarreauc49ba522019-12-11 08:12:23 +0100224 if (tick_is_lt(now_ms, key))
Willy Tarreau1e928c02019-05-28 18:57:25 +0200225 goto leave;
Willy Tarreau1e928c02019-05-28 18:57:25 +0200226
227 /* There's really something of interest here, let's visit the queue */
228
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200229 while (1) {
Willy Tarreauef28dc12019-05-28 18:48:07 +0200230 HA_RWLOCK_WRLOCK(TASK_WQ_LOCK, &wq_lock);
Emeric Brunc60def82017-09-27 14:59:38 +0200231 lookup_next:
Emeric Brun01948972017-03-30 15:37:25 +0200232 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200233 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100234 /* we might have reached the end of the tree, typically because
235 * <now_ms> is in the first half and we're first scanning the last
236 * half. Let's loop back to the beginning of the tree now.
237 */
238 eb = eb32_first(&timers);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100239 if (likely(!eb))
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100240 break;
241 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200242
Willy Tarreauc49ba522019-12-11 08:12:23 +0100243 if (tick_is_lt(now_ms, eb->key))
Willy Tarreaub992ba12017-11-05 19:09:27 +0100244 break;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200245
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100246 /* timer looks expired, detach it from the queue */
247 task = eb32_entry(eb, struct task, wq);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100248 __task_unlink_wq(task);
Willy Tarreau41365222009-03-08 07:46:27 +0100249
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100250 /* It is possible that this task was left at an earlier place in the
251 * tree because a recent call to task_queue() has not moved it. This
252 * happens when the new expiration date is later than the old one.
253 * Since it is very unlikely that we reach a timeout anyway, it's a
254 * lot cheaper to proceed like this because we almost never update
255 * the tree. We may also find disabled expiration dates there. Since
256 * we have detached the task from the tree, we simply call task_queue
Willy Tarreau814c9782009-07-14 23:48:55 +0200257 * to take care of this. Note that we might occasionally requeue it at
258 * the same place, before <eb>, so we have to check if this happens,
259 * and adjust <eb>, otherwise we may skip it which is not what we want.
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200260 * We may also not requeue the task (and not point eb at it) if its
261 * expiration time is not set.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100262 */
263 if (!tick_is_expired(task->expire, now_ms)) {
Willy Tarreaub992ba12017-11-05 19:09:27 +0100264 if (tick_isset(task->expire))
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200265 __task_queue(task, &timers);
Emeric Brunc60def82017-09-27 14:59:38 +0200266 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200267 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100268 task_wakeup(task, TASK_WOKEN_TIMER);
Willy Tarreauef28dc12019-05-28 18:48:07 +0200269 HA_RWLOCK_WRUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100270 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200271
Willy Tarreauef28dc12019-05-28 18:48:07 +0200272 HA_RWLOCK_WRUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200273#endif
Willy Tarreau1e928c02019-05-28 18:57:25 +0200274leave:
Willy Tarreauc49ba522019-12-11 08:12:23 +0100275 return;
276}
277
278/* Checks the next timer for the current thread by looking into its own timer
279 * list and the global one. It may return TICK_ETERNITY if no timer is present.
Ilya Shipitsin856aabc2020-04-16 23:51:34 +0500280 * Note that the next timer might very well be slightly in the past.
Willy Tarreauc49ba522019-12-11 08:12:23 +0100281 */
282int next_timer_expiry()
283{
284 struct task_per_thread * const tt = sched; // thread's tasks
285 struct eb32_node *eb;
286 int ret = TICK_ETERNITY;
Willy Tarreauaf613e82020-06-05 08:40:51 +0200287 __decl_thread(int key);
Willy Tarreauc49ba522019-12-11 08:12:23 +0100288
289 /* first check in the thread-local timers */
290 eb = eb32_lookup_ge(&tt->timers, now_ms - TIMER_LOOK_BACK);
291 if (!eb) {
292 /* we might have reached the end of the tree, typically because
293 * <now_ms> is in the first half and we're first scanning the last
294 * half. Let's loop back to the beginning of the tree now.
295 */
296 eb = eb32_first(&tt->timers);
297 }
298
299 if (eb)
300 ret = eb->key;
301
302#ifdef USE_THREAD
303 if (!eb_is_empty(&timers)) {
304 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
305 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
306 if (!eb)
307 eb = eb32_first(&timers);
308 if (eb)
309 key = eb->key;
310 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
311 if (eb)
312 ret = tick_first(ret, key);
313 }
314#endif
Willy Tarreaub992ba12017-11-05 19:09:27 +0100315 return ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200316}
317
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100318/* Walks over tasklet list <list> and run at most <max> of them. Returns
319 * the number of entries effectively processed (tasks and tasklets merged).
320 * The count of tasks in the list for the current thread is adjusted.
321 */
Willy Tarreau27d00c02020-03-03 14:59:28 +0100322int run_tasks_from_list(struct list *list, int max)
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100323{
324 struct task *(*process)(struct task *t, void *ctx, unsigned short state);
325 struct task *t;
326 unsigned short state;
327 void *ctx;
328 int done = 0;
329
330 while (done < max && !LIST_ISEMPTY(list)) {
331 t = (struct task *)LIST_ELEM(list->n, struct tasklet *, list);
Willy Tarreau952c2642020-01-31 16:39:30 +0100332 state = (t->state & (TASK_SHARED_WQ|TASK_SELF_WAKING));
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100333
334 ti->flags &= ~TI_FL_STUCK; // this thread is still running
335 activity[tid].ctxsw++;
336 ctx = t->context;
337 process = t->process;
338 t->calls++;
Willy Tarreaud23d4132020-01-31 10:39:03 +0100339 sched->current = t;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100340
341 if (TASK_IS_TASKLET(t)) {
Willy Tarreau952c2642020-01-31 16:39:30 +0100342 state = _HA_ATOMIC_XCHG(&t->state, state);
343 __ha_barrier_atomic_store();
344 __tasklet_remove_from_tasklet_list((struct tasklet *)t);
Olivier Houchardc62d9ab2020-03-17 18:15:04 +0100345 process(t, ctx, state);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100346 done++;
Willy Tarreaud23d4132020-01-31 10:39:03 +0100347 sched->current = NULL;
348 __ha_barrier_store();
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100349 continue;
350 }
351
Willy Tarreau952c2642020-01-31 16:39:30 +0100352 state = _HA_ATOMIC_XCHG(&t->state, state | TASK_RUNNING);
353 __ha_barrier_atomic_store();
354 __tasklet_remove_from_tasklet_list((struct tasklet *)t);
355
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100356 /* OK then this is a regular task */
357
358 task_per_thread[tid].task_list_size--;
359 if (unlikely(t->call_date)) {
360 uint64_t now_ns = now_mono_time();
361
362 t->lat_time += now_ns - t->call_date;
363 t->call_date = now_ns;
364 }
365
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100366 __ha_barrier_store();
367 if (likely(process == process_stream))
368 t = process_stream(t, ctx, state);
369 else if (process != NULL)
370 t = process(t, ctx, state);
371 else {
372 __task_free(t);
373 sched->current = NULL;
374 __ha_barrier_store();
375 /* We don't want max_processed to be decremented if
376 * we're just freeing a destroyed task, we should only
377 * do so if we really ran a task.
378 */
379 continue;
380 }
381 sched->current = NULL;
382 __ha_barrier_store();
383 /* If there is a pending state we have to wake up the task
384 * immediately, else we defer it into wait queue
385 */
386 if (t != NULL) {
387 if (unlikely(t->call_date)) {
388 t->cpu_time += now_mono_time() - t->call_date;
389 t->call_date = 0;
390 }
391
392 state = _HA_ATOMIC_AND(&t->state, ~TASK_RUNNING);
393 if (state & TASK_WOKEN_ANY)
394 task_wakeup(t, 0);
395 else
396 task_queue(t);
397 }
398 done++;
399 }
400 return done;
401}
402
Willy Tarreau58b458d2008-06-29 22:40:23 +0200403/* The run queue is chronologically sorted in a tree. An insertion counter is
404 * used to assign a position to each task. This counter may be combined with
405 * other variables (eg: nice value) to set the final position in the tree. The
406 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100407 * tasks processed to 200 in any case, so that general latency remains low and
Willy Tarreaucde79022019-04-12 18:03:41 +0200408 * so that task positions have a chance to be considered. The function scans
409 * both the global and local run queues and picks the most urgent task between
410 * the two. We need to grab the global runqueue lock to touch it so it's taken
411 * on the very first access to the global run queue and is released as soon as
412 * it reaches the end.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200413 *
414 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200415 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100416void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200417{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200418 struct task_per_thread * const tt = sched;
Willy Tarreaucde79022019-04-12 18:03:41 +0200419 struct eb32sc_node *lrq = NULL; // next local run queue entry
420 struct eb32sc_node *grq = NULL; // next global run queue entry
Willy Tarreau964c9362007-01-07 00:38:00 +0100421 struct task *t;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100422 int max_processed, done;
Olivier Houchard06910462019-10-11 16:35:01 +0200423 struct mt_list *tmp_list;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100424
Willy Tarreaue6a02fa2019-05-22 07:06:44 +0200425 ti->flags &= ~TI_FL_STUCK; // this thread is still running
426
Olivier Houchardcfbb3e62019-05-29 19:22:43 +0200427 if (!thread_has_tasks()) {
Willy Tarreaucde79022019-04-12 18:03:41 +0200428 activity[tid].empty_rq++;
429 return;
430 }
Olivier Houchard06910462019-10-11 16:35:01 +0200431 /* Merge the list of tasklets waken up by other threads to the
432 * main list.
433 */
434 tmp_list = MT_LIST_BEHEAD(&sched->shared_tasklet_list);
435 if (tmp_list)
Willy Tarreaua62917b2020-01-30 18:37:28 +0100436 LIST_SPLICE_END_DETACHED(&sched->tasklets[TL_URGENT], (struct list *)tmp_list);
Willy Tarreaucde79022019-04-12 18:03:41 +0200437
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100438 tasks_run_queue_cur = tasks_run_queue; /* keep a copy for reporting */
Willy Tarreauc7bdf092009-03-21 18:33:52 +0100439 nb_tasks_cur = nb_tasks;
Olivier Houchard1599b802018-05-24 18:59:04 +0200440 max_processed = global.tune.runqueue_depth;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200441
Willy Tarreauc8da0442019-04-15 09:33:42 +0200442 if (likely(niced_tasks))
443 max_processed = (max_processed + 3) / 4;
444
Willy Tarreau1dfc9bb2020-01-31 17:55:09 +0100445 /* run up to max_processed/3 urgent tasklets */
446 done = run_tasks_from_list(&tt->tasklets[TL_URGENT], (max_processed + 2) / 3);
Willy Tarreaua62917b2020-01-30 18:37:28 +0100447 max_processed -= done;
448
Willy Tarreau1dfc9bb2020-01-31 17:55:09 +0100449 /* pick up to max_processed/2 (~=3/4*(max_processed-done)) regular tasks from prio-ordered run queues */
Willy Tarreaua62917b2020-01-30 18:37:28 +0100450
Willy Tarreaucde79022019-04-12 18:03:41 +0200451 /* Note: the grq lock is always held when grq is not null */
Christopher Faulet8a48f672017-11-14 10:38:36 +0100452
Willy Tarreau1dfc9bb2020-01-31 17:55:09 +0100453 while (tt->task_list_size < (3 * max_processed + 3) / 4) {
Willy Tarreaucde79022019-04-12 18:03:41 +0200454 if ((global_tasks_mask & tid_bit) && !grq) {
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200455#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200456 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
457 grq = eb32sc_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
458 if (unlikely(!grq)) {
459 grq = eb32sc_first(&rqueue, tid_bit);
460 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200461 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200462 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200463 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100464 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200465#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200466 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100467
Willy Tarreaucde79022019-04-12 18:03:41 +0200468 /* If a global task is available for this thread, it's in grq
469 * now and the global RQ is locked.
470 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200471
Willy Tarreaucde79022019-04-12 18:03:41 +0200472 if (!lrq) {
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200473 lrq = eb32sc_lookup_ge(&tt->rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreaucde79022019-04-12 18:03:41 +0200474 if (unlikely(!lrq))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200475 lrq = eb32sc_first(&tt->rqueue, tid_bit);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100476 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100477
Willy Tarreaucde79022019-04-12 18:03:41 +0200478 if (!lrq && !grq)
479 break;
480
481 if (likely(!grq || (lrq && (int)(lrq->key - grq->key) <= 0))) {
482 t = eb32sc_entry(lrq, struct task, rq);
483 lrq = eb32sc_next(lrq, tid_bit);
484 __task_unlink_rq(t);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200485 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200486#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200487 else {
488 t = eb32sc_entry(grq, struct task, rq);
489 grq = eb32sc_next(grq, tid_bit);
490 __task_unlink_rq(t);
491 if (unlikely(!grq)) {
492 grq = eb32sc_first(&rqueue, tid_bit);
493 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200494 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200495 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaucde79022019-04-12 18:03:41 +0200496 }
497 }
Emeric Brun01948972017-03-30 15:37:25 +0200498 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200499#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200500
Olivier Houchardff1e9f32019-09-20 17:18:35 +0200501 /* Make sure the entry doesn't appear to be in a list */
Olivier Houchard06910462019-10-11 16:35:01 +0200502 LIST_INIT(&((struct tasklet *)t)->list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200503 /* And add it to the local task list */
Willy Tarreaua62917b2020-01-30 18:37:28 +0100504 tasklet_insert_into_tasklet_list(&tt->tasklets[TL_NORMAL], (struct tasklet *)t);
Olivier Houchard06910462019-10-11 16:35:01 +0200505 tt->task_list_size++;
Willy Tarreaubc13bec2019-04-30 14:55:18 +0200506 activity[tid].tasksw++;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200507 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200508
509 /* release the rqueue lock */
510 if (grq) {
511 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
512 grq = NULL;
513 }
514
Willy Tarreau1dfc9bb2020-01-31 17:55:09 +0100515 /* run between 0.4*max_processed and max_processed/2 regular tasks */
516 done = run_tasks_from_list(&tt->tasklets[TL_NORMAL], (3 * max_processed + 3) / 4);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100517 max_processed -= done;
Willy Tarreaucde79022019-04-12 18:03:41 +0200518
Willy Tarreau1dfc9bb2020-01-31 17:55:09 +0100519 /* run between max_processed/4 and max_processed bulk tasklets */
Willy Tarreaua62917b2020-01-30 18:37:28 +0100520 done = run_tasks_from_list(&tt->tasklets[TL_BULK], max_processed);
521 max_processed -= done;
522
523 if (!LIST_ISEMPTY(&sched->tasklets[TL_URGENT]) |
524 !LIST_ISEMPTY(&sched->tasklets[TL_NORMAL]) |
525 !LIST_ISEMPTY(&sched->tasklets[TL_BULK]))
Willy Tarreaucde79022019-04-12 18:03:41 +0200526 activity[tid].long_rq++;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200527}
528
Willy Tarreau64e60122019-07-12 08:31:17 +0200529/* create a work list array for <nbthread> threads, using tasks made of
530 * function <fct>. The context passed to the function will be the pointer to
531 * the thread's work list, which will contain a copy of argument <arg>. The
532 * wake up reason will be TASK_WOKEN_OTHER. The pointer to the work_list array
533 * is returned on success, otherwise NULL on failure.
534 */
535struct work_list *work_list_create(int nbthread,
536 struct task *(*fct)(struct task *, void *, unsigned short),
537 void *arg)
538{
539 struct work_list *wl;
540 int i;
541
542 wl = calloc(nbthread, sizeof(*wl));
543 if (!wl)
544 goto fail;
545
546 for (i = 0; i < nbthread; i++) {
Olivier Houchard859dc802019-08-08 15:47:21 +0200547 MT_LIST_INIT(&wl[i].head);
Willy Tarreau64e60122019-07-12 08:31:17 +0200548 wl[i].task = task_new(1UL << i);
549 if (!wl[i].task)
550 goto fail;
551 wl[i].task->process = fct;
552 wl[i].task->context = &wl[i];
553 wl[i].arg = arg;
554 }
555 return wl;
556
557 fail:
558 work_list_destroy(wl, nbthread);
559 return NULL;
560}
561
562/* destroy work list <work> */
563void work_list_destroy(struct work_list *work, int nbthread)
564{
565 int t;
566
567 if (!work)
568 return;
569 for (t = 0; t < nbthread; t++)
570 task_destroy(work[t].task);
571 free(work);
572}
573
William Lallemand27f3fa52018-12-06 14:05:20 +0100574/*
575 * Delete every tasks before running the master polling loop
576 */
577void mworker_cleantasks()
578{
579 struct task *t;
580 int i;
William Lallemandb5823392018-12-06 15:14:37 +0100581 struct eb32_node *tmp_wq = NULL;
582 struct eb32sc_node *tmp_rq = NULL;
William Lallemand27f3fa52018-12-06 14:05:20 +0100583
584#ifdef USE_THREAD
585 /* cleanup the global run queue */
William Lallemandb5823392018-12-06 15:14:37 +0100586 tmp_rq = eb32sc_first(&rqueue, MAX_THREADS_MASK);
587 while (tmp_rq) {
588 t = eb32sc_entry(tmp_rq, struct task, rq);
589 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200590 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100591 }
592 /* cleanup the timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100593 tmp_wq = eb32_first(&timers);
594 while (tmp_wq) {
595 t = eb32_entry(tmp_wq, struct task, wq);
596 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200597 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100598 }
599#endif
600 /* clean the per thread run queue */
601 for (i = 0; i < global.nbthread; i++) {
William Lallemandb5823392018-12-06 15:14:37 +0100602 tmp_rq = eb32sc_first(&task_per_thread[i].rqueue, MAX_THREADS_MASK);
603 while (tmp_rq) {
604 t = eb32sc_entry(tmp_rq, struct task, rq);
605 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200606 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100607 }
608 /* cleanup the per thread timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100609 tmp_wq = eb32_first(&task_per_thread[i].timers);
610 while (tmp_wq) {
611 t = eb32_entry(tmp_wq, struct task, wq);
612 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200613 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100614 }
615 }
616}
617
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100618/* perform minimal intializations */
619static void init_task()
Willy Tarreau4726f532009-03-07 17:25:21 +0100620{
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200621 int i;
622
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200623#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200624 memset(&timers, 0, sizeof(timers));
Willy Tarreau4726f532009-03-07 17:25:21 +0100625 memset(&rqueue, 0, sizeof(rqueue));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200626#endif
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200627 memset(&task_per_thread, 0, sizeof(task_per_thread));
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200628 for (i = 0; i < MAX_THREADS; i++) {
Willy Tarreaua62917b2020-01-30 18:37:28 +0100629 LIST_INIT(&task_per_thread[i].tasklets[TL_URGENT]);
630 LIST_INIT(&task_per_thread[i].tasklets[TL_NORMAL]);
631 LIST_INIT(&task_per_thread[i].tasklets[TL_BULK]);
Olivier Houchard06910462019-10-11 16:35:01 +0200632 MT_LIST_INIT(&task_per_thread[i].shared_tasklet_list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200633 }
Willy Tarreau4726f532009-03-07 17:25:21 +0100634}
635
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100636INITCALL0(STG_PREPARE, init_task);
637
Willy Tarreaubaaee002006-06-26 02:48:02 +0200638/*
639 * Local variables:
640 * c-indent-level: 8
641 * c-basic-offset: 8
642 * End:
643 */