blob: dbc67010058bc120d5b7551ee52e1ebca3c74159 [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 Tarreaub2551052020-06-09 09:07:15 +020015#include <import/eb32sctree.h>
16#include <import/eb32tree.h>
17
Willy Tarreau4c7e4b72020-05-27 12:58:42 +020018#include <haproxy/api.h>
Willy Tarreaue7723bd2020-06-24 11:11:02 +020019#include <haproxy/cfgparse.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020020#include <haproxy/fd.h>
21#include <haproxy/freq_ctr.h>
Willy Tarreau853b2972020-05-27 18:01:47 +020022#include <haproxy/list.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020023#include <haproxy/pool.h>
Willy Tarreaudfd3de82020-06-04 23:46:14 +020024#include <haproxy/stream.h>
Willy Tarreaucea0e1b2020-06-04 17:25:40 +020025#include <haproxy/task.h>
Willy Tarreau92b4f132020-06-01 11:05:15 +020026#include <haproxy/time.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020027#include <haproxy/tools.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020028
Willy Tarreaubaaee002006-06-26 02:48:02 +020029
Willy Tarreau8ceae722018-11-26 11:58:30 +010030DECLARE_POOL(pool_head_task, "task", sizeof(struct task));
31DECLARE_POOL(pool_head_tasklet, "tasklet", sizeof(struct tasklet));
Willy Tarreau96bcfd72007-04-29 10:41:56 +020032
Thierry FOURNIERd6975962017-07-12 14:31:10 +020033/* This is the memory pool containing all the signal structs. These
Joseph Herlantcf92b6d2018-11-15 14:19:23 -080034 * struct are used to store each required signal between two tasks.
Thierry FOURNIERd6975962017-07-12 14:31:10 +020035 */
Willy Tarreau8ceae722018-11-26 11:58:30 +010036DECLARE_POOL(pool_head_notification, "notification", sizeof(struct notification));
Thierry FOURNIERd6975962017-07-12 14:31:10 +020037
Olivier Houchardeba0c0b2018-07-26 18:53:28 +020038volatile unsigned long global_tasks_mask = 0; /* Mask of threads with tasks in the global runqueue */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010039unsigned int niced_tasks = 0; /* number of niced tasks in the run queue */
Emeric Brun01948972017-03-30 15:37:25 +020040
Willy Tarreaud022e9c2019-09-24 08:25:15 +020041THREAD_LOCAL struct task_per_thread *sched = &task_per_thread[0]; /* scheduler context for the current thread */
Willy Tarreau6d1222c2017-11-26 10:08:06 +010042
Willy Tarreau86abe442018-11-25 20:12:18 +010043__decl_aligned_spinlock(rq_lock); /* spin lock related to run queue */
Willy Tarreauef28dc12019-05-28 18:48:07 +020044__decl_aligned_rwlock(wq_lock); /* RW lock related to the wait queue */
Willy Tarreau964c9362007-01-07 00:38:00 +010045
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020046#ifdef USE_THREAD
Willy Tarreauc6ba9a02021-02-20 12:49:54 +010047struct eb_root timers; /* sorted timers tree, global, accessed under wq_lock */
48struct eb_root rqueue; /* tree constituting the global run queue, accessed under rq_lock */
Willy Tarreau9c7b8082021-02-24 15:10:07 +010049unsigned int grq_total; /* total number of entries in the global run queue, use grq_lock */
Willy Tarreauc6ba9a02021-02-20 12:49:54 +010050static unsigned int global_rqueue_ticks; /* insertion count in the grq, use rq_lock */
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020051#endif
Willy Tarreaub20aa9e2018-10-15 14:52:21 +020052
Willy Tarreau8d8747a2018-10-15 16:12:48 +020053
54struct task_per_thread task_per_thread[MAX_THREADS];
Willy Tarreau9789f7b2008-06-24 08:17:16 +020055
Willy Tarreaueb8c2c62020-06-30 11:48:48 +020056
57/* Flags the task <t> for immediate destruction and puts it into its first
58 * thread's shared tasklet list if not yet queued/running. This will bypass
59 * the priority scheduling and make the task show up as fast as possible in
60 * the other thread's queue. Note that this operation isn't idempotent and is
61 * not supposed to be run on the same task from multiple threads at once. It's
62 * the caller's responsibility to make sure it is the only one able to kill the
63 * task.
64 */
65void task_kill(struct task *t)
66{
67 unsigned short state = t->state;
68 unsigned int thr;
69
70 BUG_ON(state & TASK_KILLED);
71
72 while (1) {
73 while (state & (TASK_RUNNING | TASK_QUEUED)) {
74 /* task already in the queue and about to be executed,
75 * or even currently running. Just add the flag and be
76 * done with it, the process loop will detect it and kill
77 * it. The CAS will fail if we arrive too late.
78 */
79 if (_HA_ATOMIC_CAS(&t->state, &state, state | TASK_KILLED))
80 return;
81 }
82
83 /* We'll have to wake it up, but we must also secure it so that
84 * it doesn't vanish under us. TASK_QUEUED guarantees nobody will
85 * add past us.
86 */
87 if (_HA_ATOMIC_CAS(&t->state, &state, state | TASK_QUEUED | TASK_KILLED)) {
88 /* Bypass the tree and go directly into the shared tasklet list.
89 * Note: that's a task so it must be accounted for as such. Pick
90 * the task's first thread for the job.
91 */
92 thr = my_ffsl(t->thread_mask) - 1;
Willy Tarreau54d31172020-07-02 14:14:00 +020093
94 /* Beware: tasks that have never run don't have their ->list empty yet! */
Willy Tarreau950954f2020-07-10 08:32:10 +020095 MT_LIST_ADDQ(&task_per_thread[thr].shared_tasklet_list,
Willy Tarreau4f589262020-07-02 17:17:42 +020096 (struct mt_list *)&((struct tasklet *)t)->list);
Willy Tarreau9c7b8082021-02-24 15:10:07 +010097 _HA_ATOMIC_ADD(&task_per_thread[thr].rq_total, 1);
Willy Tarreau1f3b1412021-02-24 14:13:40 +010098 _HA_ATOMIC_ADD(&task_per_thread[thr].tasks_in_list, 1);
Willy Tarreau54d31172020-07-02 14:14:00 +020099 if (sleeping_thread_mask & (1UL << thr)) {
100 _HA_ATOMIC_AND(&sleeping_thread_mask, ~(1UL << thr));
101 wake_thread(thr);
Willy Tarreaueb8c2c62020-06-30 11:48:48 +0200102 }
Willy Tarreau54d31172020-07-02 14:14:00 +0200103 return;
Willy Tarreaueb8c2c62020-06-30 11:48:48 +0200104 }
105 }
106}
107
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100108/* Do not call this one, please use tasklet_wakeup_on() instead, as this one is
109 * the slow path of tasklet_wakeup_on() which performs some preliminary checks
110 * and sets TASK_IN_LIST before calling this one. A negative <thr> designates
111 * the current thread.
112 */
113void __tasklet_wakeup_on(struct tasklet *tl, int thr)
114{
115 if (likely(thr < 0)) {
116 /* this tasklet runs on the caller thread */
117 if (tl->state & TASK_SELF_WAKING) {
118 LIST_ADDQ(&sched->tasklets[TL_BULK], &tl->list);
119 sched->tl_class_mask |= 1 << TL_BULK;
120 }
121 else if ((struct task *)tl == sched->current) {
122 _HA_ATOMIC_OR(&tl->state, TASK_SELF_WAKING);
123 LIST_ADDQ(&sched->tasklets[TL_BULK], &tl->list);
124 sched->tl_class_mask |= 1 << TL_BULK;
125 }
126 else if (sched->current_queue < 0) {
127 LIST_ADDQ(&sched->tasklets[TL_URGENT], &tl->list);
128 sched->tl_class_mask |= 1 << TL_URGENT;
129 }
130 else {
131 LIST_ADDQ(&sched->tasklets[sched->current_queue], &tl->list);
132 sched->tl_class_mask |= 1 << sched->current_queue;
133 }
134 _HA_ATOMIC_ADD(&sched->rq_total, 1);
135 } else {
136 /* this tasklet runs on a specific thread. */
137 MT_LIST_ADDQ(&task_per_thread[thr].shared_tasklet_list, (struct mt_list *)&tl->list);
138 _HA_ATOMIC_ADD(&task_per_thread[thr].rq_total, 1);
139 if (sleeping_thread_mask & (1UL << thr)) {
140 _HA_ATOMIC_AND(&sleeping_thread_mask, ~(1UL << thr));
141 wake_thread(thr);
142 }
143 }
144}
145
Willy Tarreau4726f532009-03-07 17:25:21 +0100146/* Puts the task <t> in run queue at a position depending on t->nice. <t> is
147 * returned. The nice value assigns boosts in 32th of the run queue size. A
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100148 * nice value of -1024 sets the task to -tasks_run_queue*32, while a nice value
149 * of 1024 sets the task to tasks_run_queue*32. The state flags are cleared, so
150 * the caller will have to set its flags after this call.
Willy Tarreau4726f532009-03-07 17:25:21 +0100151 * The task must not already be in the run queue. If unsure, use the safer
152 * task_wakeup() function.
Willy Tarreau91e99932008-06-30 07:51:00 +0200153 */
Willy Tarreau018564e2021-02-24 16:41:11 +0100154void __task_wakeup(struct task *t)
Willy Tarreaue33aece2007-04-30 13:15:14 +0200155{
Willy Tarreau018564e2021-02-24 16:41:11 +0100156 struct eb_root *root = &sched->rqueue;
157
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200158#ifdef USE_THREAD
Willy Tarreau018564e2021-02-24 16:41:11 +0100159 if (t->thread_mask != tid_bit && global.nbthread != 1) {
160 root = &rqueue;
161
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200162 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreau9c7b8082021-02-24 15:10:07 +0100163
Olivier Houchardde82aea2019-04-17 19:10:22 +0200164 global_tasks_mask |= t->thread_mask;
Willy Tarreau9c7b8082021-02-24 15:10:07 +0100165 grq_total++;
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100166 t->rq.key = ++global_rqueue_ticks;
Olivier Houcharded1a6a02019-04-18 14:12:51 +0200167 __ha_barrier_store();
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100168 } else
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200169#endif
Willy Tarreau9c7b8082021-02-24 15:10:07 +0100170 {
171 _HA_ATOMIC_ADD(&sched->rq_total, 1);
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100172 t->rq.key = ++sched->rqueue_ticks;
Willy Tarreau9c7b8082021-02-24 15:10:07 +0100173 }
Willy Tarreau91e99932008-06-30 07:51:00 +0200174
175 if (likely(t->nice)) {
176 int offset;
177
Olivier Houchard4c2832852019-03-08 18:48:47 +0100178 _HA_ATOMIC_ADD(&niced_tasks, 1);
Willy Tarreau2d1fd0a2019-04-15 09:18:31 +0200179 offset = t->nice * (int)global.tune.runqueue_depth;
Willy Tarreau4726f532009-03-07 17:25:21 +0100180 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +0200181 }
182
Willy Tarreaud9add3a2019-04-25 08:57:41 +0200183 if (task_profiling_mask & tid_bit)
Willy Tarreau9efd7452018-05-31 14:48:54 +0200184 t->call_date = now_mono_time();
185
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200186 eb32sc_insert(root, &t->rq, t->thread_mask);
Willy Tarreau018564e2021-02-24 16:41:11 +0100187
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200188#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200189 if (root == &rqueue) {
Olivier Houchard4c2832852019-03-08 18:48:47 +0100190 _HA_ATOMIC_OR(&t->state, TASK_GLOBAL);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200191 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreau2c41d772021-02-24 16:13:03 +0100192
Willy Tarreaueeffb3d2021-02-24 16:44:51 +0100193 /* If all threads that are supposed to handle this task are sleeping,
194 * wake one.
195 */
196 if ((((t->thread_mask & all_threads_mask) & sleeping_thread_mask) ==
197 (t->thread_mask & all_threads_mask))) {
198 unsigned long m = (t->thread_mask & all_threads_mask) &~ tid_bit;
Olivier Houchard1b327902019-03-15 00:23:10 +0100199
Willy Tarreaueeffb3d2021-02-24 16:44:51 +0100200 m = (m & (m - 1)) ^ m; // keep lowest bit set
201 _HA_ATOMIC_AND(&sleeping_thread_mask, ~m);
202 wake_thread(my_ffsl(m) - 1);
203 }
Olivier Houchard1b327902019-03-15 00:23:10 +0100204 }
Willy Tarreau85d9b842018-07-27 17:14:41 +0200205#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200206 return;
Willy Tarreaue33aece2007-04-30 13:15:14 +0200207}
Willy Tarreaud825eef2007-05-12 22:35:00 +0200208
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200209/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100210 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200211 *
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200212 * Inserts a task into wait queue <wq> at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +0100213 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100214 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100215 * Last, tasks must not be queued further than the end of the tree, which is
216 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100217 *
218 * This function should not be used directly, it is meant to be called by the
219 * inline version of task_queue() which performs a few cheap preliminary tests
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200220 * before deciding to call __task_queue(). Moreover this function doesn't care
221 * at all about locking so the caller must be careful when deciding whether to
222 * lock or not around this call.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200223 */
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200224void __task_queue(struct task *task, struct eb_root *wq)
Willy Tarreau964c9362007-01-07 00:38:00 +0100225{
Willy Tarreaue5d79bc2020-07-22 14:29:42 +0200226#ifdef USE_THREAD
227 BUG_ON((wq == &timers && !(task->state & TASK_SHARED_WQ)) ||
228 (wq == &sched->timers && (task->state & TASK_SHARED_WQ)) ||
229 (wq != &timers && wq != &sched->timers));
230#endif
231
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100232 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +0100233 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +0100234
235 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100236 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200237#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100238 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +0200239 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +0100240 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200241#endif
Willy Tarreauce44f122008-07-05 18:16:19 +0200242
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200243 eb32_insert(wq, &task->wq);
Willy Tarreau964c9362007-01-07 00:38:00 +0100244}
245
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200246/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200247 * Extract all expired timers from the timer queue, and wakes up all
Willy Tarreauc49ba522019-12-11 08:12:23 +0100248 * associated tasks.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200249 */
Willy Tarreauc49ba522019-12-11 08:12:23 +0100250void wake_expired_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200251{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200252 struct task_per_thread * const tt = sched; // thread's tasks
Willy Tarreau3cfaa8d2020-10-16 09:26:22 +0200253 int max_processed = global.tune.runqueue_depth;
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200254 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200255 struct eb32_node *eb;
Willy Tarreauaf613e82020-06-05 08:40:51 +0200256 __decl_thread(int key);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200257
Willy Tarreau3cfaa8d2020-10-16 09:26:22 +0200258 while (max_processed-- > 0) {
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200259 lookup_next_local:
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200260 eb = eb32_lookup_ge(&tt->timers, now_ms - TIMER_LOOK_BACK);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200261 if (!eb) {
262 /* we might have reached the end of the tree, typically because
263 * <now_ms> is in the first half and we're first scanning the last
264 * half. Let's loop back to the beginning of the tree now.
265 */
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200266 eb = eb32_first(&tt->timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200267 if (likely(!eb))
268 break;
269 }
270
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200271 /* It is possible that this task was left at an earlier place in the
272 * tree because a recent call to task_queue() has not moved it. This
273 * happens when the new expiration date is later than the old one.
274 * Since it is very unlikely that we reach a timeout anyway, it's a
275 * lot cheaper to proceed like this because we almost never update
276 * the tree. We may also find disabled expiration dates there. Since
277 * we have detached the task from the tree, we simply call task_queue
278 * to take care of this. Note that we might occasionally requeue it at
279 * the same place, before <eb>, so we have to check if this happens,
280 * and adjust <eb>, otherwise we may skip it which is not what we want.
281 * We may also not requeue the task (and not point eb at it) if its
Willy Tarreau77015ab2020-06-19 11:50:27 +0200282 * expiration time is not set. We also make sure we leave the real
283 * expiration date for the next task in the queue so that when calling
284 * next_timer_expiry() we're guaranteed to see the next real date and
285 * not the next apparent date. This is in order to avoid useless
286 * wakeups.
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200287 */
Willy Tarreau77015ab2020-06-19 11:50:27 +0200288
289 task = eb32_entry(eb, struct task, wq);
290 if (tick_is_expired(task->expire, now_ms)) {
291 /* expired task, wake it up */
292 __task_unlink_wq(task);
293 task_wakeup(task, TASK_WOKEN_TIMER);
294 }
295 else if (task->expire != eb->key) {
296 /* task is not expired but its key doesn't match so let's
297 * update it and skip to next apparently expired task.
298 */
299 __task_unlink_wq(task);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200300 if (tick_isset(task->expire))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200301 __task_queue(task, &tt->timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200302 }
Willy Tarreau77015ab2020-06-19 11:50:27 +0200303 else {
304 /* task not expired and correctly placed */
305 break;
306 }
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200307 }
308
309#ifdef USE_THREAD
Willy Tarreau1e928c02019-05-28 18:57:25 +0200310 if (eb_is_empty(&timers))
311 goto leave;
312
313 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
314 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
315 if (!eb) {
316 eb = eb32_first(&timers);
317 if (likely(!eb)) {
318 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
319 goto leave;
320 }
321 }
322 key = eb->key;
Willy Tarreau1e928c02019-05-28 18:57:25 +0200323
Willy Tarreaud48ed662020-10-16 09:31:41 +0200324 if (tick_is_lt(now_ms, key)) {
325 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau1e928c02019-05-28 18:57:25 +0200326 goto leave;
Willy Tarreaud48ed662020-10-16 09:31:41 +0200327 }
Willy Tarreau1e928c02019-05-28 18:57:25 +0200328
329 /* There's really something of interest here, let's visit the queue */
330
Willy Tarreaud48ed662020-10-16 09:31:41 +0200331 if (HA_RWLOCK_TRYRDTOSK(TASK_WQ_LOCK, &wq_lock)) {
332 /* if we failed to grab the lock it means another thread is
333 * already doing the same here, so let it do the job.
334 */
335 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
336 goto leave;
337 }
338
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200339 while (1) {
Emeric Brunc60def82017-09-27 14:59:38 +0200340 lookup_next:
Willy Tarreau3cfaa8d2020-10-16 09:26:22 +0200341 if (max_processed-- <= 0)
342 break;
Emeric Brun01948972017-03-30 15:37:25 +0200343 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200344 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100345 /* we might have reached the end of the tree, typically because
346 * <now_ms> is in the first half and we're first scanning the last
347 * half. Let's loop back to the beginning of the tree now.
348 */
349 eb = eb32_first(&timers);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100350 if (likely(!eb))
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100351 break;
352 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200353
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100354 task = eb32_entry(eb, struct task, wq);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200355 if (tick_is_expired(task->expire, now_ms)) {
356 /* expired task, wake it up */
Willy Tarreaud48ed662020-10-16 09:31:41 +0200357 HA_RWLOCK_SKTOWR(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200358 __task_unlink_wq(task);
Willy Tarreaud48ed662020-10-16 09:31:41 +0200359 HA_RWLOCK_WRTOSK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200360 task_wakeup(task, TASK_WOKEN_TIMER);
361 }
362 else if (task->expire != eb->key) {
363 /* task is not expired but its key doesn't match so let's
364 * update it and skip to next apparently expired task.
365 */
Willy Tarreaud48ed662020-10-16 09:31:41 +0200366 HA_RWLOCK_SKTOWR(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200367 __task_unlink_wq(task);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100368 if (tick_isset(task->expire))
Willy Tarreau783afbe2020-07-22 14:12:45 +0200369 __task_queue(task, &timers);
Willy Tarreaud48ed662020-10-16 09:31:41 +0200370 HA_RWLOCK_WRTOSK(TASK_WQ_LOCK, &wq_lock);
Emeric Brunc60def82017-09-27 14:59:38 +0200371 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200372 }
Willy Tarreau77015ab2020-06-19 11:50:27 +0200373 else {
374 /* task not expired and correctly placed */
375 break;
376 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100377 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200378
Willy Tarreaud48ed662020-10-16 09:31:41 +0200379 HA_RWLOCK_SKUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200380#endif
Willy Tarreau1e928c02019-05-28 18:57:25 +0200381leave:
Willy Tarreauc49ba522019-12-11 08:12:23 +0100382 return;
383}
384
385/* Checks the next timer for the current thread by looking into its own timer
386 * list and the global one. It may return TICK_ETERNITY if no timer is present.
Ilya Shipitsin856aabc2020-04-16 23:51:34 +0500387 * Note that the next timer might very well be slightly in the past.
Willy Tarreauc49ba522019-12-11 08:12:23 +0100388 */
389int next_timer_expiry()
390{
391 struct task_per_thread * const tt = sched; // thread's tasks
392 struct eb32_node *eb;
393 int ret = TICK_ETERNITY;
Willy Tarreau6ce02322020-08-21 05:48:34 +0200394 __decl_thread(int key = TICK_ETERNITY);
Willy Tarreauc49ba522019-12-11 08:12:23 +0100395
396 /* first check in the thread-local timers */
397 eb = eb32_lookup_ge(&tt->timers, now_ms - TIMER_LOOK_BACK);
398 if (!eb) {
399 /* we might have reached the end of the tree, typically because
400 * <now_ms> is in the first half and we're first scanning the last
401 * half. Let's loop back to the beginning of the tree now.
402 */
403 eb = eb32_first(&tt->timers);
404 }
405
406 if (eb)
407 ret = eb->key;
408
409#ifdef USE_THREAD
410 if (!eb_is_empty(&timers)) {
411 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
412 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
413 if (!eb)
414 eb = eb32_first(&timers);
415 if (eb)
416 key = eb->key;
417 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
418 if (eb)
419 ret = tick_first(ret, key);
420 }
421#endif
Willy Tarreaub992ba12017-11-05 19:09:27 +0100422 return ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200423}
424
Willy Tarreau59153fe2020-06-24 10:17:29 +0200425/* Walks over tasklet lists sched->tasklets[0..TL_CLASSES-1] and run at most
426 * budget[TL_*] of them. Returns the number of entries effectively processed
427 * (tasks and tasklets merged). The count of tasks in the list for the current
428 * thread is adjusted.
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100429 */
Willy Tarreau59153fe2020-06-24 10:17:29 +0200430unsigned int run_tasks_from_lists(unsigned int budgets[])
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100431{
432 struct task *(*process)(struct task *t, void *ctx, unsigned short state);
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200433 struct list *tl_queues = sched->tasklets;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100434 struct task *t;
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200435 uint8_t budget_mask = (1 << TL_CLASSES) - 1;
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100436 struct sched_activity *profile_entry = NULL;
Willy Tarreau59153fe2020-06-24 10:17:29 +0200437 unsigned int done = 0;
438 unsigned int queue;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100439 unsigned short state;
440 void *ctx;
Willy Tarreau59153fe2020-06-24 10:17:29 +0200441
442 for (queue = 0; queue < TL_CLASSES;) {
443 sched->current_queue = queue;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100444
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200445 /* global.tune.sched.low-latency is set */
446 if (global.tune.options & GTUNE_SCHED_LOW_LATENCY) {
447 if (unlikely(sched->tl_class_mask & budget_mask & ((1 << queue) - 1))) {
448 /* a lower queue index has tasks again and still has a
449 * budget to run them. Let's switch to it now.
450 */
451 queue = (sched->tl_class_mask & 1) ? 0 :
452 (sched->tl_class_mask & 2) ? 1 : 2;
453 continue;
454 }
455
456 if (unlikely(queue > TL_URGENT &&
457 budget_mask & (1 << TL_URGENT) &&
458 !MT_LIST_ISEMPTY(&sched->shared_tasklet_list))) {
459 /* an urgent tasklet arrived from another thread */
460 break;
461 }
462
463 if (unlikely(queue > TL_NORMAL &&
464 budget_mask & (1 << TL_NORMAL) &&
Willy Tarreau2c41d772021-02-24 16:13:03 +0100465 (!eb_is_empty(&sched->rqueue) ||
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200466 (global_tasks_mask & tid_bit)))) {
467 /* a task was woken up by a bulk tasklet or another thread */
468 break;
469 }
470 }
471
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200472 if (LIST_ISEMPTY(&tl_queues[queue])) {
473 sched->tl_class_mask &= ~(1 << queue);
Willy Tarreau59153fe2020-06-24 10:17:29 +0200474 queue++;
475 continue;
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200476 }
477
Willy Tarreau59153fe2020-06-24 10:17:29 +0200478 if (!budgets[queue]) {
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200479 budget_mask &= ~(1 << queue);
Willy Tarreau59153fe2020-06-24 10:17:29 +0200480 queue++;
481 continue;
482 }
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200483
Willy Tarreau59153fe2020-06-24 10:17:29 +0200484 budgets[queue]--;
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200485 t = (struct task *)LIST_ELEM(tl_queues[queue].n, struct tasklet *, list);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200486 state = t->state & (TASK_SHARED_WQ|TASK_SELF_WAKING|TASK_KILLED);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100487
488 ti->flags &= ~TI_FL_STUCK; // this thread is still running
489 activity[tid].ctxsw++;
490 ctx = t->context;
491 process = t->process;
492 t->calls++;
Willy Tarreaud23d4132020-01-31 10:39:03 +0100493 sched->current = t;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100494
Willy Tarreau9c7b8082021-02-24 15:10:07 +0100495 _HA_ATOMIC_SUB(&sched->rq_total, 1);
Willy Tarreau2da4c312020-11-30 14:52:11 +0100496
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100497 if (TASK_IS_TASKLET(t)) {
Willy Tarreau4d6c5942020-11-30 14:58:53 +0100498 LIST_DEL_INIT(&((struct tasklet *)t)->list);
499 __ha_barrier_store();
Willy Tarreau952c2642020-01-31 16:39:30 +0100500 state = _HA_ATOMIC_XCHG(&t->state, state);
501 __ha_barrier_atomic_store();
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100502
503 if (unlikely(task_profiling_mask & tid_bit)) {
504 uint64_t before;
505
506 profile_entry = sched_activity_entry(sched_activity, t->process);
507 before = now_mono_time();
508 process(t, ctx, state);
509 HA_ATOMIC_ADD(&profile_entry->calls, 1);
510 HA_ATOMIC_ADD(&profile_entry->cpu_time, now_mono_time() - before);
511 } else {
512 process(t, ctx, state);
513 }
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100514 done++;
Willy Tarreaud23d4132020-01-31 10:39:03 +0100515 sched->current = NULL;
516 __ha_barrier_store();
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100517 continue;
518 }
519
Willy Tarreau4d6c5942020-11-30 14:58:53 +0100520 LIST_DEL_INIT(&((struct tasklet *)t)->list);
521 __ha_barrier_store();
Willy Tarreau952c2642020-01-31 16:39:30 +0100522 state = _HA_ATOMIC_XCHG(&t->state, state | TASK_RUNNING);
523 __ha_barrier_atomic_store();
Willy Tarreau952c2642020-01-31 16:39:30 +0100524
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100525 /* OK then this is a regular task */
526
Willy Tarreau1f3b1412021-02-24 14:13:40 +0100527 _HA_ATOMIC_SUB(&task_per_thread[tid].tasks_in_list, 1);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100528 if (unlikely(t->call_date)) {
529 uint64_t now_ns = now_mono_time();
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100530 uint64_t lat = now_ns - t->call_date;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100531
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100532 t->lat_time += lat;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100533 t->call_date = now_ns;
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100534 profile_entry = sched_activity_entry(sched_activity, t->process);
535 HA_ATOMIC_ADD(&profile_entry->lat_time, lat);
536 HA_ATOMIC_ADD(&profile_entry->calls, 1);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100537 }
538
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100539 __ha_barrier_store();
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200540
541 /* Note for below: if TASK_KILLED arrived before we've read the state, we
542 * directly free the task. Otherwise it will be seen after processing and
543 * it's freed on the exit path.
544 */
545 if (likely(!(state & TASK_KILLED) && process == process_stream))
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100546 t = process_stream(t, ctx, state);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200547 else if (!(state & TASK_KILLED) && process != NULL)
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100548 t = process(t, ctx, state);
549 else {
Willy Tarreau273aea42020-07-17 14:37:51 +0200550 task_unlink_wq(t);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100551 __task_free(t);
552 sched->current = NULL;
553 __ha_barrier_store();
554 /* We don't want max_processed to be decremented if
555 * we're just freeing a destroyed task, we should only
556 * do so if we really ran a task.
557 */
558 continue;
559 }
560 sched->current = NULL;
561 __ha_barrier_store();
562 /* If there is a pending state we have to wake up the task
563 * immediately, else we defer it into wait queue
564 */
565 if (t != NULL) {
566 if (unlikely(t->call_date)) {
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100567 uint64_t cpu = now_mono_time() - t->call_date;
568
569 t->cpu_time += cpu;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100570 t->call_date = 0;
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100571 HA_ATOMIC_ADD(&profile_entry->cpu_time, cpu);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100572 }
573
574 state = _HA_ATOMIC_AND(&t->state, ~TASK_RUNNING);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200575 if (unlikely(state & TASK_KILLED)) {
Willy Tarreau273aea42020-07-17 14:37:51 +0200576 task_unlink_wq(t);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200577 __task_free(t);
578 }
579 else if (state & TASK_WOKEN_ANY)
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100580 task_wakeup(t, 0);
581 else
582 task_queue(t);
583 }
584 done++;
585 }
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200586 sched->current_queue = -1;
Willy Tarreau116ef222020-06-23 16:35:38 +0200587
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100588 return done;
589}
590
Willy Tarreau58b458d2008-06-29 22:40:23 +0200591/* The run queue is chronologically sorted in a tree. An insertion counter is
592 * used to assign a position to each task. This counter may be combined with
593 * other variables (eg: nice value) to set the final position in the tree. The
594 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100595 * tasks processed to 200 in any case, so that general latency remains low and
Willy Tarreaucde79022019-04-12 18:03:41 +0200596 * so that task positions have a chance to be considered. The function scans
597 * both the global and local run queues and picks the most urgent task between
598 * the two. We need to grab the global runqueue lock to touch it so it's taken
599 * on the very first access to the global run queue and is released as soon as
600 * it reaches the end.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200601 *
602 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200603 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100604void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200605{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200606 struct task_per_thread * const tt = sched;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200607 struct eb32sc_node *lrq; // next local run queue entry
608 struct eb32sc_node *grq; // next global run queue entry
Willy Tarreau964c9362007-01-07 00:38:00 +0100609 struct task *t;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200610 const unsigned int default_weights[TL_CLASSES] = {
611 [TL_URGENT] = 64, // ~50% of CPU bandwidth for I/O
612 [TL_NORMAL] = 48, // ~37% of CPU bandwidth for tasks
613 [TL_BULK] = 16, // ~13% of CPU bandwidth for self-wakers
614 };
615 unsigned int max[TL_CLASSES]; // max to be run per class
616 unsigned int max_total; // sum of max above
Olivier Houchard06910462019-10-11 16:35:01 +0200617 struct mt_list *tmp_list;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200618 unsigned int queue;
619 int max_processed;
Willy Tarreauc309dbd2020-11-30 15:39:00 +0100620 int picked;
621 int budget;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100622
Willy Tarreaue6a02fa2019-05-22 07:06:44 +0200623 ti->flags &= ~TI_FL_STUCK; // this thread is still running
624
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200625 if (!thread_has_tasks()) {
626 activity[tid].empty_rq++;
627 return;
628 }
629
Willy Tarreau5c8be272020-06-19 12:17:55 +0200630 max_processed = global.tune.runqueue_depth;
631
632 if (likely(niced_tasks))
633 max_processed = (max_processed + 3) / 4;
634
Willy Tarreau5c8be272020-06-19 12:17:55 +0200635 not_done_yet:
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200636 max[TL_URGENT] = max[TL_NORMAL] = max[TL_BULK] = 0;
Willy Tarreaucde79022019-04-12 18:03:41 +0200637
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200638 /* urgent tasklets list gets a default weight of ~50% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200639 if ((tt->tl_class_mask & (1 << TL_URGENT)) ||
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200640 !MT_LIST_ISEMPTY(&tt->shared_tasklet_list))
641 max[TL_URGENT] = default_weights[TL_URGENT];
Willy Tarreaua62917b2020-01-30 18:37:28 +0100642
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200643 /* normal tasklets list gets a default weight of ~37% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200644 if ((tt->tl_class_mask & (1 << TL_NORMAL)) ||
Willy Tarreau2c41d772021-02-24 16:13:03 +0100645 !eb_is_empty(&sched->rqueue) || (global_tasks_mask & tid_bit))
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200646 max[TL_NORMAL] = default_weights[TL_NORMAL];
Willy Tarreaua62917b2020-01-30 18:37:28 +0100647
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200648 /* bulk tasklets list gets a default weight of ~13% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200649 if ((tt->tl_class_mask & (1 << TL_BULK)))
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200650 max[TL_BULK] = default_weights[TL_BULK];
651
652 /* Now compute a fair share of the weights. Total may slightly exceed
Willy Tarreau1553b662020-06-30 13:46:21 +0200653 * 100% due to rounding, this is not a problem. Note that while in
654 * theory the sum cannot be NULL as we cannot get there without tasklets
655 * to process, in practice it seldom happens when multiple writers
Willy Tarreaude4db172020-07-10 08:10:29 +0200656 * conflict and rollback on MT_LIST_TRY_ADDQ(shared_tasklet_list), causing
Willy Tarreau1553b662020-06-30 13:46:21 +0200657 * a first MT_LIST_ISEMPTY() to succeed for thread_has_task() and the
658 * one above to finally fail. This is extremely rare and not a problem.
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200659 */
660 max_total = max[TL_URGENT] + max[TL_NORMAL] + max[TL_BULK];
Willy Tarreau1553b662020-06-30 13:46:21 +0200661 if (!max_total)
662 return;
663
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200664 for (queue = 0; queue < TL_CLASSES; queue++)
665 max[queue] = ((unsigned)max_processed * max[queue] + max_total - 1) / max_total;
666
667 lrq = grq = NULL;
Christopher Faulet8a48f672017-11-14 10:38:36 +0100668
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200669 /* pick up to max[TL_NORMAL] regular tasks from prio-ordered run queues */
670 /* Note: the grq lock is always held when grq is not null */
Willy Tarreauc309dbd2020-11-30 15:39:00 +0100671 picked = 0;
Willy Tarreau1f3b1412021-02-24 14:13:40 +0100672 budget = max[TL_NORMAL] - tt->tasks_in_list;
Willy Tarreauc309dbd2020-11-30 15:39:00 +0100673 while (picked < budget) {
Willy Tarreaucde79022019-04-12 18:03:41 +0200674 if ((global_tasks_mask & tid_bit) && !grq) {
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200675#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200676 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100677 grq = eb32sc_lookup_ge(&rqueue, global_rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreaucde79022019-04-12 18:03:41 +0200678 if (unlikely(!grq)) {
679 grq = eb32sc_first(&rqueue, tid_bit);
680 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200681 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200682 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200683 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100684 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200685#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200686 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100687
Willy Tarreaucde79022019-04-12 18:03:41 +0200688 /* If a global task is available for this thread, it's in grq
689 * now and the global RQ is locked.
690 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200691
Willy Tarreaucde79022019-04-12 18:03:41 +0200692 if (!lrq) {
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100693 lrq = eb32sc_lookup_ge(&tt->rqueue, tt->rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreaucde79022019-04-12 18:03:41 +0200694 if (unlikely(!lrq))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200695 lrq = eb32sc_first(&tt->rqueue, tid_bit);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100696 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100697
Willy Tarreaucde79022019-04-12 18:03:41 +0200698 if (!lrq && !grq)
699 break;
700
701 if (likely(!grq || (lrq && (int)(lrq->key - grq->key) <= 0))) {
702 t = eb32sc_entry(lrq, struct task, rq);
703 lrq = eb32sc_next(lrq, tid_bit);
704 __task_unlink_rq(t);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200705 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200706#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200707 else {
708 t = eb32sc_entry(grq, struct task, rq);
709 grq = eb32sc_next(grq, tid_bit);
710 __task_unlink_rq(t);
711 if (unlikely(!grq)) {
712 grq = eb32sc_first(&rqueue, tid_bit);
713 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200714 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200715 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaucde79022019-04-12 18:03:41 +0200716 }
717 }
Emeric Brun01948972017-03-30 15:37:25 +0200718 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200719#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200720
Willy Tarreaua868c292020-11-30 15:30:22 +0100721 /* Add it to the local task list */
722 LIST_ADDQ(&tt->tasklets[TL_NORMAL], &((struct tasklet *)t)->list);
Willy Tarreauc309dbd2020-11-30 15:39:00 +0100723 picked++;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200724 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200725
726 /* release the rqueue lock */
727 if (grq) {
728 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
729 grq = NULL;
730 }
731
Willy Tarreauc309dbd2020-11-30 15:39:00 +0100732 if (picked) {
733 tt->tl_class_mask |= 1 << TL_NORMAL;
Willy Tarreau1f3b1412021-02-24 14:13:40 +0100734 _HA_ATOMIC_ADD(&tt->tasks_in_list, picked);
Willy Tarreau9c7b8082021-02-24 15:10:07 +0100735 _HA_ATOMIC_ADD(&tt->rq_total, picked);
Willy Tarreauc309dbd2020-11-30 15:39:00 +0100736 activity[tid].tasksw += picked;
737 }
738
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200739 /* Merge the list of tasklets waken up by other threads to the
740 * main list.
741 */
742 tmp_list = MT_LIST_BEHEAD(&tt->shared_tasklet_list);
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200743 if (tmp_list) {
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200744 LIST_SPLICE_END_DETACHED(&tt->tasklets[TL_URGENT], (struct list *)tmp_list);
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200745 if (!LIST_ISEMPTY(&tt->tasklets[TL_URGENT]))
746 tt->tl_class_mask |= 1 << TL_URGENT;
747 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200748
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200749 /* execute tasklets in each queue */
Willy Tarreau59153fe2020-06-24 10:17:29 +0200750 max_processed -= run_tasks_from_lists(max);
Willy Tarreaua62917b2020-01-30 18:37:28 +0100751
Willy Tarreau5c8be272020-06-19 12:17:55 +0200752 /* some tasks may have woken other ones up */
Willy Tarreau0c0c85e2020-06-23 11:32:35 +0200753 if (max_processed > 0 && thread_has_tasks())
Willy Tarreau5c8be272020-06-19 12:17:55 +0200754 goto not_done_yet;
755
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200756 if (tt->tl_class_mask)
Willy Tarreaucde79022019-04-12 18:03:41 +0200757 activity[tid].long_rq++;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200758}
759
Willy Tarreau64e60122019-07-12 08:31:17 +0200760/* create a work list array for <nbthread> threads, using tasks made of
761 * function <fct>. The context passed to the function will be the pointer to
762 * the thread's work list, which will contain a copy of argument <arg>. The
763 * wake up reason will be TASK_WOKEN_OTHER. The pointer to the work_list array
764 * is returned on success, otherwise NULL on failure.
765 */
766struct work_list *work_list_create(int nbthread,
767 struct task *(*fct)(struct task *, void *, unsigned short),
768 void *arg)
769{
770 struct work_list *wl;
771 int i;
772
773 wl = calloc(nbthread, sizeof(*wl));
774 if (!wl)
775 goto fail;
776
777 for (i = 0; i < nbthread; i++) {
Olivier Houchard859dc802019-08-08 15:47:21 +0200778 MT_LIST_INIT(&wl[i].head);
Willy Tarreau64e60122019-07-12 08:31:17 +0200779 wl[i].task = task_new(1UL << i);
780 if (!wl[i].task)
781 goto fail;
782 wl[i].task->process = fct;
783 wl[i].task->context = &wl[i];
784 wl[i].arg = arg;
785 }
786 return wl;
787
788 fail:
789 work_list_destroy(wl, nbthread);
790 return NULL;
791}
792
793/* destroy work list <work> */
794void work_list_destroy(struct work_list *work, int nbthread)
795{
796 int t;
797
798 if (!work)
799 return;
800 for (t = 0; t < nbthread; t++)
801 task_destroy(work[t].task);
802 free(work);
803}
804
William Lallemand27f3fa52018-12-06 14:05:20 +0100805/*
806 * Delete every tasks before running the master polling loop
807 */
808void mworker_cleantasks()
809{
810 struct task *t;
811 int i;
William Lallemandb5823392018-12-06 15:14:37 +0100812 struct eb32_node *tmp_wq = NULL;
813 struct eb32sc_node *tmp_rq = NULL;
William Lallemand27f3fa52018-12-06 14:05:20 +0100814
815#ifdef USE_THREAD
816 /* cleanup the global run queue */
William Lallemandb5823392018-12-06 15:14:37 +0100817 tmp_rq = eb32sc_first(&rqueue, MAX_THREADS_MASK);
818 while (tmp_rq) {
819 t = eb32sc_entry(tmp_rq, struct task, rq);
820 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200821 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100822 }
823 /* cleanup the timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100824 tmp_wq = eb32_first(&timers);
825 while (tmp_wq) {
826 t = eb32_entry(tmp_wq, struct task, wq);
827 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200828 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100829 }
830#endif
831 /* clean the per thread run queue */
832 for (i = 0; i < global.nbthread; i++) {
William Lallemandb5823392018-12-06 15:14:37 +0100833 tmp_rq = eb32sc_first(&task_per_thread[i].rqueue, MAX_THREADS_MASK);
834 while (tmp_rq) {
835 t = eb32sc_entry(tmp_rq, struct task, rq);
836 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200837 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100838 }
839 /* cleanup the per thread timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100840 tmp_wq = eb32_first(&task_per_thread[i].timers);
841 while (tmp_wq) {
842 t = eb32_entry(tmp_wq, struct task, wq);
843 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200844 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100845 }
846 }
847}
848
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100849/* perform minimal intializations */
850static void init_task()
Willy Tarreau4726f532009-03-07 17:25:21 +0100851{
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200852 int i;
853
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200854#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200855 memset(&timers, 0, sizeof(timers));
Willy Tarreau4726f532009-03-07 17:25:21 +0100856 memset(&rqueue, 0, sizeof(rqueue));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200857#endif
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200858 memset(&task_per_thread, 0, sizeof(task_per_thread));
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200859 for (i = 0; i < MAX_THREADS; i++) {
Willy Tarreaua62917b2020-01-30 18:37:28 +0100860 LIST_INIT(&task_per_thread[i].tasklets[TL_URGENT]);
861 LIST_INIT(&task_per_thread[i].tasklets[TL_NORMAL]);
862 LIST_INIT(&task_per_thread[i].tasklets[TL_BULK]);
Olivier Houchard06910462019-10-11 16:35:01 +0200863 MT_LIST_INIT(&task_per_thread[i].shared_tasklet_list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200864 }
Willy Tarreau4726f532009-03-07 17:25:21 +0100865}
866
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200867/* config parser for global "tune.sched.low-latency", accepts "on" or "off" */
868static int cfg_parse_tune_sched_low_latency(char **args, int section_type, struct proxy *curpx,
869 struct proxy *defpx, const char *file, int line,
870 char **err)
871{
872 if (too_many_args(1, args, err, NULL))
873 return -1;
874
875 if (strcmp(args[1], "on") == 0)
876 global.tune.options |= GTUNE_SCHED_LOW_LATENCY;
877 else if (strcmp(args[1], "off") == 0)
878 global.tune.options &= ~GTUNE_SCHED_LOW_LATENCY;
879 else {
880 memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
881 return -1;
882 }
883 return 0;
884}
885
886/* config keyword parsers */
887static struct cfg_kw_list cfg_kws = {ILH, {
888 { CFG_GLOBAL, "tune.sched.low-latency", cfg_parse_tune_sched_low_latency },
889 { 0, NULL, NULL }
890}};
891
892INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100893INITCALL0(STG_PREPARE, init_task);
894
Willy Tarreaubaaee002006-06-26 02:48:02 +0200895/*
896 * Local variables:
897 * c-indent-level: 8
898 * c-basic-offset: 8
899 * End:
900 */