blob: 2176a91905942e40e557ebe3925107f1fc5d92a2 [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 Tarreau6d1222c2017-11-26 10:08:06 +010044THREAD_LOCAL struct task *curr_task = NULL; /* task currently running or NULL */
45
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
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100158 * associated tasks. Returns the date of next event (or eternity).
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200159 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100160int wake_expired_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200161{
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200162 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200163 struct eb32_node *eb;
Willy Tarreaub992ba12017-11-05 19:09:27 +0100164 int ret = TICK_ETERNITY;
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 Tarreau8d8747a2018-10-15 16:12:48 +0200169 eb = eb32_lookup_ge(&task_per_thread[tid].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 Tarreau8d8747a2018-10-15 16:12:48 +0200175 eb = eb32_first(&task_per_thread[tid].timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200176 if (likely(!eb))
177 break;
178 }
179
180 if (tick_is_lt(now_ms, eb->key)) {
181 /* timer not expired yet, revisit it later */
182 ret = eb->key;
183 break;
184 }
185
186 /* timer looks expired, detach it from the queue */
187 task = eb32_entry(eb, struct task, wq);
188 __task_unlink_wq(task);
189
190 /* It is possible that this task was left at an earlier place in the
191 * tree because a recent call to task_queue() has not moved it. This
192 * happens when the new expiration date is later than the old one.
193 * Since it is very unlikely that we reach a timeout anyway, it's a
194 * lot cheaper to proceed like this because we almost never update
195 * the tree. We may also find disabled expiration dates there. Since
196 * we have detached the task from the tree, we simply call task_queue
197 * to take care of this. Note that we might occasionally requeue it at
198 * the same place, before <eb>, so we have to check if this happens,
199 * and adjust <eb>, otherwise we may skip it which is not what we want.
200 * We may also not requeue the task (and not point eb at it) if its
201 * expiration time is not set.
202 */
203 if (!tick_is_expired(task->expire, now_ms)) {
204 if (tick_isset(task->expire))
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200205 __task_queue(task, &task_per_thread[tid].timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200206 goto lookup_next_local;
207 }
208 task_wakeup(task, TASK_WOKEN_TIMER);
209 }
210
211#ifdef USE_THREAD
Willy Tarreau1e928c02019-05-28 18:57:25 +0200212 if (eb_is_empty(&timers))
213 goto leave;
214
215 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
216 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
217 if (!eb) {
218 eb = eb32_first(&timers);
219 if (likely(!eb)) {
220 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
221 goto leave;
222 }
223 }
224 key = eb->key;
225 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
226
227 if (tick_is_lt(now_ms, key)) {
228 /* timer not expired yet, revisit it later */
229 ret = tick_first(ret, key);
230 goto leave;
231 }
232
233 /* There's really something of interest here, let's visit the queue */
234
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200235 while (1) {
Willy Tarreauef28dc12019-05-28 18:48:07 +0200236 HA_RWLOCK_WRLOCK(TASK_WQ_LOCK, &wq_lock);
Emeric Brunc60def82017-09-27 14:59:38 +0200237 lookup_next:
Emeric Brun01948972017-03-30 15:37:25 +0200238 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200239 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100240 /* we might have reached the end of the tree, typically because
241 * <now_ms> is in the first half and we're first scanning the last
242 * half. Let's loop back to the beginning of the tree now.
243 */
244 eb = eb32_first(&timers);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100245 if (likely(!eb))
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100246 break;
247 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200248
Willy Tarreaub992ba12017-11-05 19:09:27 +0100249 if (tick_is_lt(now_ms, eb->key)) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100250 /* timer not expired yet, revisit it later */
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200251 ret = tick_first(ret, eb->key);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100252 break;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100253 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200254
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100255 /* timer looks expired, detach it from the queue */
256 task = eb32_entry(eb, struct task, wq);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100257 __task_unlink_wq(task);
Willy Tarreau41365222009-03-08 07:46:27 +0100258
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100259 /* It is possible that this task was left at an earlier place in the
260 * tree because a recent call to task_queue() has not moved it. This
261 * happens when the new expiration date is later than the old one.
262 * Since it is very unlikely that we reach a timeout anyway, it's a
263 * lot cheaper to proceed like this because we almost never update
264 * the tree. We may also find disabled expiration dates there. Since
265 * we have detached the task from the tree, we simply call task_queue
Willy Tarreau814c9782009-07-14 23:48:55 +0200266 * to take care of this. Note that we might occasionally requeue it at
267 * the same place, before <eb>, so we have to check if this happens,
268 * and adjust <eb>, otherwise we may skip it which is not what we want.
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200269 * We may also not requeue the task (and not point eb at it) if its
270 * expiration time is not set.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100271 */
272 if (!tick_is_expired(task->expire, now_ms)) {
Willy Tarreaub992ba12017-11-05 19:09:27 +0100273 if (tick_isset(task->expire))
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200274 __task_queue(task, &timers);
Emeric Brunc60def82017-09-27 14:59:38 +0200275 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200276 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100277 task_wakeup(task, TASK_WOKEN_TIMER);
Willy Tarreauef28dc12019-05-28 18:48:07 +0200278 HA_RWLOCK_WRUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100279 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200280
Willy Tarreauef28dc12019-05-28 18:48:07 +0200281 HA_RWLOCK_WRUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200282#endif
Willy Tarreau1e928c02019-05-28 18:57:25 +0200283leave:
Willy Tarreaub992ba12017-11-05 19:09:27 +0100284 return ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200285}
286
Willy Tarreau58b458d2008-06-29 22:40:23 +0200287/* The run queue is chronologically sorted in a tree. An insertion counter is
288 * used to assign a position to each task. This counter may be combined with
289 * other variables (eg: nice value) to set the final position in the tree. The
290 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100291 * tasks processed to 200 in any case, so that general latency remains low and
Willy Tarreaucde79022019-04-12 18:03:41 +0200292 * so that task positions have a chance to be considered. The function scans
293 * both the global and local run queues and picks the most urgent task between
294 * the two. We need to grab the global runqueue lock to touch it so it's taken
295 * on the very first access to the global run queue and is released as soon as
296 * it reaches the end.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200297 *
298 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200299 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100300void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200301{
Willy Tarreaucde79022019-04-12 18:03:41 +0200302 struct eb32sc_node *lrq = NULL; // next local run queue entry
303 struct eb32sc_node *grq = NULL; // next global run queue entry
Willy Tarreau964c9362007-01-07 00:38:00 +0100304 struct task *t;
Emeric Brun01948972017-03-30 15:37:25 +0200305 int max_processed;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100306
Willy Tarreaue6a02fa2019-05-22 07:06:44 +0200307 ti->flags &= ~TI_FL_STUCK; // this thread is still running
308
Olivier Houchardcfbb3e62019-05-29 19:22:43 +0200309 if (!thread_has_tasks()) {
Willy Tarreaucde79022019-04-12 18:03:41 +0200310 activity[tid].empty_rq++;
311 return;
312 }
313
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100314 tasks_run_queue_cur = tasks_run_queue; /* keep a copy for reporting */
Willy Tarreauc7bdf092009-03-21 18:33:52 +0100315 nb_tasks_cur = nb_tasks;
Olivier Houchard1599b802018-05-24 18:59:04 +0200316 max_processed = global.tune.runqueue_depth;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200317
Willy Tarreauc8da0442019-04-15 09:33:42 +0200318 if (likely(niced_tasks))
319 max_processed = (max_processed + 3) / 4;
320
Willy Tarreaucde79022019-04-12 18:03:41 +0200321 /* Note: the grq lock is always held when grq is not null */
Christopher Faulet8a48f672017-11-14 10:38:36 +0100322
Willy Tarreaucde79022019-04-12 18:03:41 +0200323 while (task_per_thread[tid].task_list_size < max_processed) {
324 if ((global_tasks_mask & tid_bit) && !grq) {
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200325#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200326 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
327 grq = eb32sc_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
328 if (unlikely(!grq)) {
329 grq = eb32sc_first(&rqueue, tid_bit);
330 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200331 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200332 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200333 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100334 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200335#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200336 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100337
Willy Tarreaucde79022019-04-12 18:03:41 +0200338 /* If a global task is available for this thread, it's in grq
339 * now and the global RQ is locked.
340 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200341
Willy Tarreaucde79022019-04-12 18:03:41 +0200342 if (!lrq) {
343 lrq = eb32sc_lookup_ge(&task_per_thread[tid].rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
344 if (unlikely(!lrq))
345 lrq = eb32sc_first(&task_per_thread[tid].rqueue, tid_bit);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100346 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100347
Willy Tarreaucde79022019-04-12 18:03:41 +0200348 if (!lrq && !grq)
349 break;
350
351 if (likely(!grq || (lrq && (int)(lrq->key - grq->key) <= 0))) {
352 t = eb32sc_entry(lrq, struct task, rq);
353 lrq = eb32sc_next(lrq, tid_bit);
354 __task_unlink_rq(t);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200355 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200356#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200357 else {
358 t = eb32sc_entry(grq, struct task, rq);
359 grq = eb32sc_next(grq, tid_bit);
360 __task_unlink_rq(t);
361 if (unlikely(!grq)) {
362 grq = eb32sc_first(&rqueue, tid_bit);
363 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200364 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200365 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaucde79022019-04-12 18:03:41 +0200366 }
367 }
Emeric Brun01948972017-03-30 15:37:25 +0200368 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200369#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200370
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200371 /* And add it to the local task list */
Willy Tarreaubd20a9d2019-06-14 18:05:54 +0200372 tasklet_insert_into_tasklet_list((struct tasklet *)t);
373 task_per_thread[tid].task_list_size++;
Willy Tarreaubc13bec2019-04-30 14:55:18 +0200374 activity[tid].tasksw++;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200375 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200376
377 /* release the rqueue lock */
378 if (grq) {
379 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
380 grq = NULL;
381 }
382
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200383 while (max_processed > 0 && !LIST_ISEMPTY(&task_per_thread[tid].task_list)) {
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200384 struct task *t;
385 unsigned short state;
386 void *ctx;
387 struct task *(*process)(struct task *t, void *ctx, unsigned short state);
388
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200389 t = (struct task *)LIST_ELEM(task_per_thread[tid].task_list.n, struct tasklet *, list);
Olivier Houchard4c2832852019-03-08 18:48:47 +0100390 state = _HA_ATOMIC_XCHG(&t->state, TASK_RUNNING);
Olivier Houchardd2b5d162019-03-08 13:47:21 +0100391 __ha_barrier_atomic_store();
Willy Tarreau86eded62019-06-14 14:47:49 +0200392 __tasklet_remove_from_tasklet_list((struct tasklet *)t);
Willy Tarreaubd20a9d2019-06-14 18:05:54 +0200393 if (!TASK_IS_TASKLET(t))
394 task_per_thread[tid].task_list_size--;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200395
Willy Tarreaue6a02fa2019-05-22 07:06:44 +0200396 ti->flags &= ~TI_FL_STUCK; // this thread is still running
Willy Tarreaubc13bec2019-04-30 14:55:18 +0200397 activity[tid].ctxsw++;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200398 ctx = t->context;
399 process = t->process;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200400 t->calls++;
Willy Tarreau9efd7452018-05-31 14:48:54 +0200401
402 if (unlikely(!TASK_IS_TASKLET(t) && t->call_date)) {
403 uint64_t now_ns = now_mono_time();
404
405 t->lat_time += now_ns - t->call_date;
406 t->call_date = now_ns;
407 }
408
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200409 curr_task = (struct task *)t;
Willy Tarreau01f34892019-05-17 11:46:04 +0200410 __ha_barrier_store();
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200411 if (likely(process == process_stream))
412 t = process_stream(t, ctx, state);
Willy Tarreau03dd0292019-04-17 22:32:27 +0200413 else if (process != NULL)
414 t = process(TASK_IS_TASKLET(t) ? NULL : t, ctx, state);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200415 else {
Willy Tarreau03dd0292019-04-17 22:32:27 +0200416 __task_free(t);
Olivier Houchard1cfac372019-04-17 22:53:41 +0200417 curr_task = NULL;
Willy Tarreau01f34892019-05-17 11:46:04 +0200418 __ha_barrier_store();
Olivier Houchard1cfac372019-04-17 22:53:41 +0200419 /* We don't want max_processed to be decremented if
420 * we're just freeing a destroyed task, we should only
421 * do so if we really ran a task.
422 */
423 continue;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200424 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200425 curr_task = NULL;
Willy Tarreau01f34892019-05-17 11:46:04 +0200426 __ha_barrier_store();
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200427 /* If there is a pending state we have to wake up the task
Joseph Herlantcf92b6d2018-11-15 14:19:23 -0800428 * immediately, else we defer it into wait queue
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200429 */
430 if (t != NULL) {
Willy Tarreau9efd7452018-05-31 14:48:54 +0200431 if (unlikely(!TASK_IS_TASKLET(t) && t->call_date)) {
432 t->cpu_time += now_mono_time() - t->call_date;
433 t->call_date = 0;
434 }
435
Olivier Houchard4c2832852019-03-08 18:48:47 +0100436 state = _HA_ATOMIC_AND(&t->state, ~TASK_RUNNING);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200437 if (state)
Willy Tarreaub0380072019-04-17 11:47:18 +0200438 task_wakeup(t, 0);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200439 else
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100440 task_queue(t);
Willy Tarreau58b458d2008-06-29 22:40:23 +0200441 }
Willy Tarreauce4e0aa2017-11-05 23:57:00 +0100442
Olivier Houchard736ea412018-05-28 14:54:49 +0200443 max_processed--;
Willy Tarreaucde79022019-04-12 18:03:41 +0200444 }
445
Olivier Houchardcfbb3e62019-05-29 19:22:43 +0200446 if (!LIST_ISEMPTY(&task_per_thread[tid].task_list))
Willy Tarreaucde79022019-04-12 18:03:41 +0200447 activity[tid].long_rq++;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200448}
449
William Lallemand27f3fa52018-12-06 14:05:20 +0100450/*
451 * Delete every tasks before running the master polling loop
452 */
453void mworker_cleantasks()
454{
455 struct task *t;
456 int i;
William Lallemandb5823392018-12-06 15:14:37 +0100457 struct eb32_node *tmp_wq = NULL;
458 struct eb32sc_node *tmp_rq = NULL;
William Lallemand27f3fa52018-12-06 14:05:20 +0100459
460#ifdef USE_THREAD
461 /* cleanup the global run queue */
William Lallemandb5823392018-12-06 15:14:37 +0100462 tmp_rq = eb32sc_first(&rqueue, MAX_THREADS_MASK);
463 while (tmp_rq) {
464 t = eb32sc_entry(tmp_rq, struct task, rq);
465 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200466 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100467 }
468 /* cleanup the timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100469 tmp_wq = eb32_first(&timers);
470 while (tmp_wq) {
471 t = eb32_entry(tmp_wq, struct task, wq);
472 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200473 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100474 }
475#endif
476 /* clean the per thread run queue */
477 for (i = 0; i < global.nbthread; i++) {
William Lallemandb5823392018-12-06 15:14:37 +0100478 tmp_rq = eb32sc_first(&task_per_thread[i].rqueue, MAX_THREADS_MASK);
479 while (tmp_rq) {
480 t = eb32sc_entry(tmp_rq, struct task, rq);
481 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200482 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100483 }
484 /* cleanup the per thread timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100485 tmp_wq = eb32_first(&task_per_thread[i].timers);
486 while (tmp_wq) {
487 t = eb32_entry(tmp_wq, struct task, wq);
488 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200489 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100490 }
491 }
492}
493
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100494/* perform minimal intializations */
495static void init_task()
Willy Tarreau4726f532009-03-07 17:25:21 +0100496{
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200497 int i;
498
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200499#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200500 memset(&timers, 0, sizeof(timers));
Willy Tarreau4726f532009-03-07 17:25:21 +0100501 memset(&rqueue, 0, sizeof(rqueue));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200502#endif
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200503 memset(&task_per_thread, 0, sizeof(task_per_thread));
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200504 for (i = 0; i < MAX_THREADS; i++) {
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200505 LIST_INIT(&task_per_thread[i].task_list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200506 }
Willy Tarreau4726f532009-03-07 17:25:21 +0100507}
508
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100509INITCALL0(STG_PREPARE, init_task);
510
Willy Tarreaubaaee002006-06-26 02:48:02 +0200511/*
512 * Local variables:
513 * c-indent-level: 8
514 * c-basic-offset: 8
515 * End:
516 */