blob: 0c26063bc71234198c016333255031f18210eea0 [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 Tarreau0212fad2019-04-24 08:10:57 +020023#include <proto/fd.h>
24#include <proto/freq_ctr.h>
Willy Tarreaud825eef2007-05-12 22:35:00 +020025#include <proto/proxy.h>
Willy Tarreau87b09662015-04-03 00:22:06 +020026#include <proto/stream.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020027#include <proto/task.h>
28
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 Tarreau5598d172019-06-04 17:16:29 +0200165 __decl_hathreads(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.
280 * Note that the next timer might very well be slighly in the past.
281 */
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;
287 __decl_hathreads(int key);
288
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 Tarreau58b458d2008-06-29 22:40:23 +0200318/* The run queue is chronologically sorted in a tree. An insertion counter is
319 * used to assign a position to each task. This counter may be combined with
320 * other variables (eg: nice value) to set the final position in the tree. The
321 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100322 * tasks processed to 200 in any case, so that general latency remains low and
Willy Tarreaucde79022019-04-12 18:03:41 +0200323 * so that task positions have a chance to be considered. The function scans
324 * both the global and local run queues and picks the most urgent task between
325 * the two. We need to grab the global runqueue lock to touch it so it's taken
326 * on the very first access to the global run queue and is released as soon as
327 * it reaches the end.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200328 *
329 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200330 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100331void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200332{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200333 struct task_per_thread * const tt = sched;
Willy Tarreaucde79022019-04-12 18:03:41 +0200334 struct eb32sc_node *lrq = NULL; // next local run queue entry
335 struct eb32sc_node *grq = NULL; // next global run queue entry
Willy Tarreau964c9362007-01-07 00:38:00 +0100336 struct task *t;
Emeric Brun01948972017-03-30 15:37:25 +0200337 int max_processed;
Olivier Houchard06910462019-10-11 16:35:01 +0200338 struct mt_list *tmp_list;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100339
Willy Tarreaue6a02fa2019-05-22 07:06:44 +0200340 ti->flags &= ~TI_FL_STUCK; // this thread is still running
341
Olivier Houchardcfbb3e62019-05-29 19:22:43 +0200342 if (!thread_has_tasks()) {
Willy Tarreaucde79022019-04-12 18:03:41 +0200343 activity[tid].empty_rq++;
344 return;
345 }
Olivier Houchard06910462019-10-11 16:35:01 +0200346 /* Merge the list of tasklets waken up by other threads to the
347 * main list.
348 */
349 tmp_list = MT_LIST_BEHEAD(&sched->shared_tasklet_list);
350 if (tmp_list)
351 LIST_SPLICE_END_DETACHED(&sched->task_list, (struct list *)tmp_list);
Willy Tarreaucde79022019-04-12 18:03:41 +0200352
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100353 tasks_run_queue_cur = tasks_run_queue; /* keep a copy for reporting */
Willy Tarreauc7bdf092009-03-21 18:33:52 +0100354 nb_tasks_cur = nb_tasks;
Olivier Houchard1599b802018-05-24 18:59:04 +0200355 max_processed = global.tune.runqueue_depth;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200356
Willy Tarreauc8da0442019-04-15 09:33:42 +0200357 if (likely(niced_tasks))
358 max_processed = (max_processed + 3) / 4;
359
Willy Tarreaucde79022019-04-12 18:03:41 +0200360 /* Note: the grq lock is always held when grq is not null */
Christopher Faulet8a48f672017-11-14 10:38:36 +0100361
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200362 while (tt->task_list_size < max_processed) {
Willy Tarreaucde79022019-04-12 18:03:41 +0200363 if ((global_tasks_mask & tid_bit) && !grq) {
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200364#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200365 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
366 grq = eb32sc_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
367 if (unlikely(!grq)) {
368 grq = eb32sc_first(&rqueue, tid_bit);
369 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200370 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200371 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200372 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100373 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200374#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200375 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100376
Willy Tarreaucde79022019-04-12 18:03:41 +0200377 /* If a global task is available for this thread, it's in grq
378 * now and the global RQ is locked.
379 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200380
Willy Tarreaucde79022019-04-12 18:03:41 +0200381 if (!lrq) {
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200382 lrq = eb32sc_lookup_ge(&tt->rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreaucde79022019-04-12 18:03:41 +0200383 if (unlikely(!lrq))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200384 lrq = eb32sc_first(&tt->rqueue, tid_bit);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100385 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100386
Willy Tarreaucde79022019-04-12 18:03:41 +0200387 if (!lrq && !grq)
388 break;
389
390 if (likely(!grq || (lrq && (int)(lrq->key - grq->key) <= 0))) {
391 t = eb32sc_entry(lrq, struct task, rq);
392 lrq = eb32sc_next(lrq, tid_bit);
393 __task_unlink_rq(t);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200394 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200395#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200396 else {
397 t = eb32sc_entry(grq, struct task, rq);
398 grq = eb32sc_next(grq, tid_bit);
399 __task_unlink_rq(t);
400 if (unlikely(!grq)) {
401 grq = eb32sc_first(&rqueue, tid_bit);
402 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200403 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200404 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaucde79022019-04-12 18:03:41 +0200405 }
406 }
Emeric Brun01948972017-03-30 15:37:25 +0200407 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200408#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200409
Olivier Houchardff1e9f32019-09-20 17:18:35 +0200410 /* Make sure the entry doesn't appear to be in a list */
Olivier Houchard06910462019-10-11 16:35:01 +0200411 LIST_INIT(&((struct tasklet *)t)->list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200412 /* And add it to the local task list */
Willy Tarreaubd20a9d2019-06-14 18:05:54 +0200413 tasklet_insert_into_tasklet_list((struct tasklet *)t);
Olivier Houchard06910462019-10-11 16:35:01 +0200414 tt->task_list_size++;
Willy Tarreaubc13bec2019-04-30 14:55:18 +0200415 activity[tid].tasksw++;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200416 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200417
418 /* release the rqueue lock */
419 if (grq) {
420 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
421 grq = NULL;
422 }
423
Olivier Houchard06910462019-10-11 16:35:01 +0200424 while (max_processed > 0 && !LIST_ISEMPTY(&tt->task_list)) {
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200425 struct task *t;
426 unsigned short state;
427 void *ctx;
428 struct task *(*process)(struct task *t, void *ctx, unsigned short state);
429
Olivier Houchard06910462019-10-11 16:35:01 +0200430 t = (struct task *)LIST_ELEM(task_per_thread[tid].task_list.n, struct tasklet *, list);
Willy Tarreaudd0e89a2019-12-19 07:39:06 +0100431 state = (t->state & TASK_SHARED_WQ) | TASK_RUNNING;
432 state = _HA_ATOMIC_XCHG(&t->state, state);
Olivier Houchardd2b5d162019-03-08 13:47:21 +0100433 __ha_barrier_atomic_store();
Olivier Houchard06910462019-10-11 16:35:01 +0200434 __tasklet_remove_from_tasklet_list((struct tasklet *)t);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200435
Willy Tarreaue6a02fa2019-05-22 07:06:44 +0200436 ti->flags &= ~TI_FL_STUCK; // this thread is still running
Willy Tarreaubc13bec2019-04-30 14:55:18 +0200437 activity[tid].ctxsw++;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200438 ctx = t->context;
439 process = t->process;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200440 t->calls++;
Willy Tarreau9efd7452018-05-31 14:48:54 +0200441
Willy Tarreaud66d7562019-09-24 07:34:09 +0200442 if (TASK_IS_TASKLET(t)) {
443 process(NULL, ctx, state);
444 max_processed--;
445 continue;
446 }
447
448 /* OK then this is a regular task */
449
Olivier Houchard06910462019-10-11 16:35:01 +0200450 tt->task_list_size--;
Willy Tarreaud66d7562019-09-24 07:34:09 +0200451 if (unlikely(t->call_date)) {
Willy Tarreau9efd7452018-05-31 14:48:54 +0200452 uint64_t now_ns = now_mono_time();
453
454 t->lat_time += now_ns - t->call_date;
455 t->call_date = now_ns;
456 }
457
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200458 sched->current = t;
Willy Tarreau01f34892019-05-17 11:46:04 +0200459 __ha_barrier_store();
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200460 if (likely(process == process_stream))
461 t = process_stream(t, ctx, state);
Willy Tarreau03dd0292019-04-17 22:32:27 +0200462 else if (process != NULL)
Willy Tarreaud66d7562019-09-24 07:34:09 +0200463 t = process(t, ctx, state);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200464 else {
Willy Tarreau03dd0292019-04-17 22:32:27 +0200465 __task_free(t);
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200466 sched->current = NULL;
Willy Tarreau01f34892019-05-17 11:46:04 +0200467 __ha_barrier_store();
Olivier Houchard1cfac372019-04-17 22:53:41 +0200468 /* We don't want max_processed to be decremented if
469 * we're just freeing a destroyed task, we should only
470 * do so if we really ran a task.
471 */
472 continue;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200473 }
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200474 sched->current = NULL;
Willy Tarreau01f34892019-05-17 11:46:04 +0200475 __ha_barrier_store();
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200476 /* If there is a pending state we have to wake up the task
Joseph Herlantcf92b6d2018-11-15 14:19:23 -0800477 * immediately, else we defer it into wait queue
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200478 */
479 if (t != NULL) {
Willy Tarreau9b48c622019-09-24 07:02:40 +0200480 if (unlikely(t->call_date)) {
Willy Tarreau9efd7452018-05-31 14:48:54 +0200481 t->cpu_time += now_mono_time() - t->call_date;
482 t->call_date = 0;
483 }
484
Olivier Houchard4c2832852019-03-08 18:48:47 +0100485 state = _HA_ATOMIC_AND(&t->state, ~TASK_RUNNING);
Willy Tarreau8fe42532019-12-19 07:34:20 +0100486 if (state & TASK_WOKEN_ANY)
Willy Tarreaub0380072019-04-17 11:47:18 +0200487 task_wakeup(t, 0);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200488 else
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100489 task_queue(t);
Willy Tarreau58b458d2008-06-29 22:40:23 +0200490 }
Willy Tarreauce4e0aa2017-11-05 23:57:00 +0100491
Olivier Houchard736ea412018-05-28 14:54:49 +0200492 max_processed--;
Willy Tarreaucde79022019-04-12 18:03:41 +0200493 }
494
Olivier Houchard06910462019-10-11 16:35:01 +0200495 if (!LIST_ISEMPTY(&tt->task_list))
Willy Tarreaucde79022019-04-12 18:03:41 +0200496 activity[tid].long_rq++;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200497}
498
Willy Tarreau64e60122019-07-12 08:31:17 +0200499/* create a work list array for <nbthread> threads, using tasks made of
500 * function <fct>. The context passed to the function will be the pointer to
501 * the thread's work list, which will contain a copy of argument <arg>. The
502 * wake up reason will be TASK_WOKEN_OTHER. The pointer to the work_list array
503 * is returned on success, otherwise NULL on failure.
504 */
505struct work_list *work_list_create(int nbthread,
506 struct task *(*fct)(struct task *, void *, unsigned short),
507 void *arg)
508{
509 struct work_list *wl;
510 int i;
511
512 wl = calloc(nbthread, sizeof(*wl));
513 if (!wl)
514 goto fail;
515
516 for (i = 0; i < nbthread; i++) {
Olivier Houchard859dc802019-08-08 15:47:21 +0200517 MT_LIST_INIT(&wl[i].head);
Willy Tarreau64e60122019-07-12 08:31:17 +0200518 wl[i].task = task_new(1UL << i);
519 if (!wl[i].task)
520 goto fail;
521 wl[i].task->process = fct;
522 wl[i].task->context = &wl[i];
523 wl[i].arg = arg;
524 }
525 return wl;
526
527 fail:
528 work_list_destroy(wl, nbthread);
529 return NULL;
530}
531
532/* destroy work list <work> */
533void work_list_destroy(struct work_list *work, int nbthread)
534{
535 int t;
536
537 if (!work)
538 return;
539 for (t = 0; t < nbthread; t++)
540 task_destroy(work[t].task);
541 free(work);
542}
543
William Lallemand27f3fa52018-12-06 14:05:20 +0100544/*
545 * Delete every tasks before running the master polling loop
546 */
547void mworker_cleantasks()
548{
549 struct task *t;
550 int i;
William Lallemandb5823392018-12-06 15:14:37 +0100551 struct eb32_node *tmp_wq = NULL;
552 struct eb32sc_node *tmp_rq = NULL;
William Lallemand27f3fa52018-12-06 14:05:20 +0100553
554#ifdef USE_THREAD
555 /* cleanup the global run queue */
William Lallemandb5823392018-12-06 15:14:37 +0100556 tmp_rq = eb32sc_first(&rqueue, MAX_THREADS_MASK);
557 while (tmp_rq) {
558 t = eb32sc_entry(tmp_rq, struct task, rq);
559 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200560 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100561 }
562 /* cleanup the timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100563 tmp_wq = eb32_first(&timers);
564 while (tmp_wq) {
565 t = eb32_entry(tmp_wq, struct task, wq);
566 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200567 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100568 }
569#endif
570 /* clean the per thread run queue */
571 for (i = 0; i < global.nbthread; i++) {
William Lallemandb5823392018-12-06 15:14:37 +0100572 tmp_rq = eb32sc_first(&task_per_thread[i].rqueue, MAX_THREADS_MASK);
573 while (tmp_rq) {
574 t = eb32sc_entry(tmp_rq, struct task, rq);
575 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200576 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100577 }
578 /* cleanup the per thread timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100579 tmp_wq = eb32_first(&task_per_thread[i].timers);
580 while (tmp_wq) {
581 t = eb32_entry(tmp_wq, struct task, wq);
582 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200583 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100584 }
585 }
586}
587
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100588/* perform minimal intializations */
589static void init_task()
Willy Tarreau4726f532009-03-07 17:25:21 +0100590{
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200591 int i;
592
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200593#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200594 memset(&timers, 0, sizeof(timers));
Willy Tarreau4726f532009-03-07 17:25:21 +0100595 memset(&rqueue, 0, sizeof(rqueue));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200596#endif
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200597 memset(&task_per_thread, 0, sizeof(task_per_thread));
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200598 for (i = 0; i < MAX_THREADS; i++) {
Olivier Houchard06910462019-10-11 16:35:01 +0200599 LIST_INIT(&task_per_thread[i].task_list);
600 MT_LIST_INIT(&task_per_thread[i].shared_tasklet_list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200601 }
Willy Tarreau4726f532009-03-07 17:25:21 +0100602}
603
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100604INITCALL0(STG_PREPARE, init_task);
605
Willy Tarreaubaaee002006-06-26 02:48:02 +0200606/*
607 * Local variables:
608 * c-indent-level: 8
609 * c-basic-offset: 8
610 * End:
611 */