blob: 7dc0cd41cfa9db413a14dcf0f9d2b5d02ad91431 [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>
Willy Tarreau853b2972020-05-27 18:01:47 +020021#include <haproxy/list.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020022#include <haproxy/pool.h>
Willy Tarreaucea0e1b2020-06-04 17:25:40 +020023#include <haproxy/task.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020024#include <haproxy/tools.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020025
Willy Tarreaue08f4bf2021-05-08 20:10:13 +020026extern struct task *process_stream(struct task *t, void *context, unsigned int state);
Willy Tarreaubaaee002006-06-26 02:48:02 +020027
Willy Tarreau8ceae722018-11-26 11:58:30 +010028DECLARE_POOL(pool_head_task, "task", sizeof(struct task));
29DECLARE_POOL(pool_head_tasklet, "tasklet", sizeof(struct tasklet));
Willy Tarreau96bcfd72007-04-29 10:41:56 +020030
Thierry FOURNIERd6975962017-07-12 14:31:10 +020031/* This is the memory pool containing all the signal structs. These
Joseph Herlantcf92b6d2018-11-15 14:19:23 -080032 * struct are used to store each required signal between two tasks.
Thierry FOURNIERd6975962017-07-12 14:31:10 +020033 */
Willy Tarreau8ceae722018-11-26 11:58:30 +010034DECLARE_POOL(pool_head_notification, "notification", sizeof(struct notification));
Thierry FOURNIERd6975962017-07-12 14:31:10 +020035
Olivier Houchardeba0c0b2018-07-26 18:53:28 +020036volatile unsigned long global_tasks_mask = 0; /* Mask of threads with tasks in the global runqueue */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010037unsigned int niced_tasks = 0; /* number of niced tasks in the run queue */
Emeric Brun01948972017-03-30 15:37:25 +020038
Willy Tarreaud022e9c2019-09-24 08:25:15 +020039THREAD_LOCAL struct task_per_thread *sched = &task_per_thread[0]; /* scheduler context for the current thread */
Willy Tarreau6d1222c2017-11-26 10:08:06 +010040
Willy Tarreau86abe442018-11-25 20:12:18 +010041__decl_aligned_spinlock(rq_lock); /* spin lock related to run queue */
Willy Tarreauef28dc12019-05-28 18:48:07 +020042__decl_aligned_rwlock(wq_lock); /* RW lock related to the wait queue */
Willy Tarreau964c9362007-01-07 00:38:00 +010043
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020044#ifdef USE_THREAD
Willy Tarreauc6ba9a02021-02-20 12:49:54 +010045struct eb_root timers; /* sorted timers tree, global, accessed under wq_lock */
46struct eb_root rqueue; /* tree constituting the global run queue, accessed under rq_lock */
Willy Tarreau45499c52021-02-25 07:51:18 +010047unsigned int grq_total; /* total number of entries in the global run queue, atomic */
Willy Tarreauc6ba9a02021-02-20 12:49:54 +010048static unsigned int global_rqueue_ticks; /* insertion count in the grq, use rq_lock */
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020049#endif
Willy Tarreaub20aa9e2018-10-15 14:52:21 +020050
Willy Tarreau8d8747a2018-10-15 16:12:48 +020051
52struct task_per_thread task_per_thread[MAX_THREADS];
Willy Tarreau9789f7b2008-06-24 08:17:16 +020053
Willy Tarreaueb8c2c62020-06-30 11:48:48 +020054
55/* Flags the task <t> for immediate destruction and puts it into its first
56 * thread's shared tasklet list if not yet queued/running. This will bypass
57 * the priority scheduling and make the task show up as fast as possible in
58 * the other thread's queue. Note that this operation isn't idempotent and is
59 * not supposed to be run on the same task from multiple threads at once. It's
60 * the caller's responsibility to make sure it is the only one able to kill the
61 * task.
62 */
63void task_kill(struct task *t)
64{
Willy Tarreau144f84a2021-03-02 16:09:26 +010065 unsigned int state = t->state;
Willy Tarreaueb8c2c62020-06-30 11:48:48 +020066 unsigned int thr;
67
68 BUG_ON(state & TASK_KILLED);
69
70 while (1) {
71 while (state & (TASK_RUNNING | TASK_QUEUED)) {
72 /* task already in the queue and about to be executed,
73 * or even currently running. Just add the flag and be
74 * done with it, the process loop will detect it and kill
75 * it. The CAS will fail if we arrive too late.
76 */
77 if (_HA_ATOMIC_CAS(&t->state, &state, state | TASK_KILLED))
78 return;
79 }
80
81 /* We'll have to wake it up, but we must also secure it so that
82 * it doesn't vanish under us. TASK_QUEUED guarantees nobody will
83 * add past us.
84 */
85 if (_HA_ATOMIC_CAS(&t->state, &state, state | TASK_QUEUED | TASK_KILLED)) {
86 /* Bypass the tree and go directly into the shared tasklet list.
87 * Note: that's a task so it must be accounted for as such. Pick
88 * the task's first thread for the job.
89 */
90 thr = my_ffsl(t->thread_mask) - 1;
Willy Tarreau54d31172020-07-02 14:14:00 +020091
92 /* Beware: tasks that have never run don't have their ->list empty yet! */
Willy Tarreau2b718102021-04-21 07:32:39 +020093 MT_LIST_APPEND(&task_per_thread[thr].shared_tasklet_list,
Willy Tarreau4f589262020-07-02 17:17:42 +020094 (struct mt_list *)&((struct tasklet *)t)->list);
Willy Tarreau4781b152021-04-06 13:53:36 +020095 _HA_ATOMIC_INC(&task_per_thread[thr].rq_total);
96 _HA_ATOMIC_INC(&task_per_thread[thr].tasks_in_list);
Willy Tarreau54d31172020-07-02 14:14:00 +020097 if (sleeping_thread_mask & (1UL << thr)) {
98 _HA_ATOMIC_AND(&sleeping_thread_mask, ~(1UL << thr));
99 wake_thread(thr);
Willy Tarreaueb8c2c62020-06-30 11:48:48 +0200100 }
Willy Tarreau54d31172020-07-02 14:14:00 +0200101 return;
Willy Tarreaueb8c2c62020-06-30 11:48:48 +0200102 }
103 }
104}
105
Amaury Denoyelle7b368332021-07-28 16:12:57 +0200106/* Equivalent of task_kill for tasklets. Mark the tasklet <t> for destruction.
107 * It will be deleted on the next scheduler invocation. This function is
108 * thread-safe : a thread can kill a tasklet of another thread.
109 */
110void tasklet_kill(struct tasklet *t)
111{
112 unsigned int state = t->state;
113 unsigned int thr;
114
115 BUG_ON(state & TASK_KILLED);
116
117 while (1) {
118 while (state & (TASK_IN_LIST)) {
119 /* Tasklet already in the list ready to be executed. Add
120 * the killed flag and wait for the process loop to
121 * detect it.
122 */
123 if (_HA_ATOMIC_CAS(&t->state, &state, state | TASK_KILLED))
124 return;
125 }
126
127 /* Mark the tasklet as killed and wake the thread to process it
128 * as soon as possible.
129 */
130 if (_HA_ATOMIC_CAS(&t->state, &state, state | TASK_IN_LIST | TASK_KILLED)) {
131 thr = t->tid > 0 ? t->tid: tid;
132 MT_LIST_APPEND(&task_per_thread[thr].shared_tasklet_list,
133 (struct mt_list *)&t->list);
134 _HA_ATOMIC_INC(&task_per_thread[thr].rq_total);
135 if (sleeping_thread_mask & (1UL << thr)) {
136 _HA_ATOMIC_AND(&sleeping_thread_mask, ~(1UL << thr));
137 wake_thread(thr);
138 }
139 return;
140 }
141 }
142}
143
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100144/* Do not call this one, please use tasklet_wakeup_on() instead, as this one is
145 * the slow path of tasklet_wakeup_on() which performs some preliminary checks
146 * and sets TASK_IN_LIST before calling this one. A negative <thr> designates
147 * the current thread.
148 */
149void __tasklet_wakeup_on(struct tasklet *tl, int thr)
150{
151 if (likely(thr < 0)) {
152 /* this tasklet runs on the caller thread */
Willy Tarreau826fa872021-02-26 10:13:40 +0100153 if (tl->state & TASK_HEAVY) {
Willy Tarreau2b718102021-04-21 07:32:39 +0200154 LIST_APPEND(&sched->tasklets[TL_HEAVY], &tl->list);
Willy Tarreau826fa872021-02-26 10:13:40 +0100155 sched->tl_class_mask |= 1 << TL_HEAVY;
156 }
157 else if (tl->state & TASK_SELF_WAKING) {
Willy Tarreau2b718102021-04-21 07:32:39 +0200158 LIST_APPEND(&sched->tasklets[TL_BULK], &tl->list);
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100159 sched->tl_class_mask |= 1 << TL_BULK;
160 }
161 else if ((struct task *)tl == sched->current) {
162 _HA_ATOMIC_OR(&tl->state, TASK_SELF_WAKING);
Willy Tarreau2b718102021-04-21 07:32:39 +0200163 LIST_APPEND(&sched->tasklets[TL_BULK], &tl->list);
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100164 sched->tl_class_mask |= 1 << TL_BULK;
165 }
166 else if (sched->current_queue < 0) {
Willy Tarreau2b718102021-04-21 07:32:39 +0200167 LIST_APPEND(&sched->tasklets[TL_URGENT], &tl->list);
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100168 sched->tl_class_mask |= 1 << TL_URGENT;
169 }
170 else {
Willy Tarreau2b718102021-04-21 07:32:39 +0200171 LIST_APPEND(&sched->tasklets[sched->current_queue], &tl->list);
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100172 sched->tl_class_mask |= 1 << sched->current_queue;
173 }
Willy Tarreau4781b152021-04-06 13:53:36 +0200174 _HA_ATOMIC_INC(&sched->rq_total);
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100175 } else {
176 /* this tasklet runs on a specific thread. */
Willy Tarreau2b718102021-04-21 07:32:39 +0200177 MT_LIST_APPEND(&task_per_thread[thr].shared_tasklet_list, (struct mt_list *)&tl->list);
Willy Tarreau4781b152021-04-06 13:53:36 +0200178 _HA_ATOMIC_INC(&task_per_thread[thr].rq_total);
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100179 if (sleeping_thread_mask & (1UL << thr)) {
180 _HA_ATOMIC_AND(&sleeping_thread_mask, ~(1UL << thr));
181 wake_thread(thr);
182 }
183 }
184}
185
Willy Tarreau4726f532009-03-07 17:25:21 +0100186/* Puts the task <t> in run queue at a position depending on t->nice. <t> is
187 * returned. The nice value assigns boosts in 32th of the run queue size. A
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100188 * nice value of -1024 sets the task to -tasks_run_queue*32, while a nice value
189 * of 1024 sets the task to tasks_run_queue*32. The state flags are cleared, so
190 * the caller will have to set its flags after this call.
Willy Tarreau4726f532009-03-07 17:25:21 +0100191 * The task must not already be in the run queue. If unsure, use the safer
192 * task_wakeup() function.
Willy Tarreau91e99932008-06-30 07:51:00 +0200193 */
Willy Tarreau018564e2021-02-24 16:41:11 +0100194void __task_wakeup(struct task *t)
Willy Tarreaue33aece2007-04-30 13:15:14 +0200195{
Willy Tarreau018564e2021-02-24 16:41:11 +0100196 struct eb_root *root = &sched->rqueue;
197
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200198#ifdef USE_THREAD
Willy Tarreau018564e2021-02-24 16:41:11 +0100199 if (t->thread_mask != tid_bit && global.nbthread != 1) {
200 root = &rqueue;
201
Willy Tarreau4781b152021-04-06 13:53:36 +0200202 _HA_ATOMIC_INC(&grq_total);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200203 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreau9c7b8082021-02-24 15:10:07 +0100204
Olivier Houchardde82aea2019-04-17 19:10:22 +0200205 global_tasks_mask |= t->thread_mask;
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100206 t->rq.key = ++global_rqueue_ticks;
Olivier Houcharded1a6a02019-04-18 14:12:51 +0200207 __ha_barrier_store();
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100208 } else
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200209#endif
Willy Tarreau9c7b8082021-02-24 15:10:07 +0100210 {
Willy Tarreau4781b152021-04-06 13:53:36 +0200211 _HA_ATOMIC_INC(&sched->rq_total);
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100212 t->rq.key = ++sched->rqueue_ticks;
Willy Tarreau9c7b8082021-02-24 15:10:07 +0100213 }
Willy Tarreau91e99932008-06-30 07:51:00 +0200214
215 if (likely(t->nice)) {
216 int offset;
217
Willy Tarreau4781b152021-04-06 13:53:36 +0200218 _HA_ATOMIC_INC(&niced_tasks);
Willy Tarreau2d1fd0a2019-04-15 09:18:31 +0200219 offset = t->nice * (int)global.tune.runqueue_depth;
Willy Tarreau4726f532009-03-07 17:25:21 +0100220 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +0200221 }
222
Willy Tarreaud9add3a2019-04-25 08:57:41 +0200223 if (task_profiling_mask & tid_bit)
Willy Tarreau9efd7452018-05-31 14:48:54 +0200224 t->call_date = now_mono_time();
225
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200226 eb32sc_insert(root, &t->rq, t->thread_mask);
Willy Tarreau018564e2021-02-24 16:41:11 +0100227
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200228#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200229 if (root == &rqueue) {
Olivier Houchard4c2832852019-03-08 18:48:47 +0100230 _HA_ATOMIC_OR(&t->state, TASK_GLOBAL);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200231 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreau2c41d772021-02-24 16:13:03 +0100232
Willy Tarreaueeffb3d2021-02-24 16:44:51 +0100233 /* If all threads that are supposed to handle this task are sleeping,
234 * wake one.
235 */
236 if ((((t->thread_mask & all_threads_mask) & sleeping_thread_mask) ==
237 (t->thread_mask & all_threads_mask))) {
238 unsigned long m = (t->thread_mask & all_threads_mask) &~ tid_bit;
Olivier Houchard1b327902019-03-15 00:23:10 +0100239
Willy Tarreaueeffb3d2021-02-24 16:44:51 +0100240 m = (m & (m - 1)) ^ m; // keep lowest bit set
241 _HA_ATOMIC_AND(&sleeping_thread_mask, ~m);
242 wake_thread(my_ffsl(m) - 1);
243 }
Olivier Houchard1b327902019-03-15 00:23:10 +0100244 }
Willy Tarreau85d9b842018-07-27 17:14:41 +0200245#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200246 return;
Willy Tarreaue33aece2007-04-30 13:15:14 +0200247}
Willy Tarreaud825eef2007-05-12 22:35:00 +0200248
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200249/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100250 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200251 *
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200252 * Inserts a task into wait queue <wq> at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +0100253 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100254 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100255 * Last, tasks must not be queued further than the end of the tree, which is
256 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100257 *
258 * This function should not be used directly, it is meant to be called by the
259 * inline version of task_queue() which performs a few cheap preliminary tests
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200260 * before deciding to call __task_queue(). Moreover this function doesn't care
261 * at all about locking so the caller must be careful when deciding whether to
262 * lock or not around this call.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200263 */
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200264void __task_queue(struct task *task, struct eb_root *wq)
Willy Tarreau964c9362007-01-07 00:38:00 +0100265{
Willy Tarreaue5d79bc2020-07-22 14:29:42 +0200266#ifdef USE_THREAD
267 BUG_ON((wq == &timers && !(task->state & TASK_SHARED_WQ)) ||
268 (wq == &sched->timers && (task->state & TASK_SHARED_WQ)) ||
269 (wq != &timers && wq != &sched->timers));
270#endif
271
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100272 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +0100273 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +0100274
275 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100276 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200277#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100278 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +0200279 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +0100280 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200281#endif
Willy Tarreauce44f122008-07-05 18:16:19 +0200282
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200283 eb32_insert(wq, &task->wq);
Willy Tarreau964c9362007-01-07 00:38:00 +0100284}
285
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200286/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200287 * Extract all expired timers from the timer queue, and wakes up all
Willy Tarreauc49ba522019-12-11 08:12:23 +0100288 * associated tasks.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200289 */
Willy Tarreauc49ba522019-12-11 08:12:23 +0100290void wake_expired_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200291{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200292 struct task_per_thread * const tt = sched; // thread's tasks
Willy Tarreau3cfaa8d2020-10-16 09:26:22 +0200293 int max_processed = global.tune.runqueue_depth;
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200294 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200295 struct eb32_node *eb;
Willy Tarreauaf613e82020-06-05 08:40:51 +0200296 __decl_thread(int key);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200297
Willy Tarreau3cfaa8d2020-10-16 09:26:22 +0200298 while (max_processed-- > 0) {
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200299 lookup_next_local:
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200300 eb = eb32_lookup_ge(&tt->timers, now_ms - TIMER_LOOK_BACK);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200301 if (!eb) {
302 /* we might have reached the end of the tree, typically because
303 * <now_ms> is in the first half and we're first scanning the last
304 * half. Let's loop back to the beginning of the tree now.
305 */
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200306 eb = eb32_first(&tt->timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200307 if (likely(!eb))
308 break;
309 }
310
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200311 /* It is possible that this task was left at an earlier place in the
312 * tree because a recent call to task_queue() has not moved it. This
313 * happens when the new expiration date is later than the old one.
314 * Since it is very unlikely that we reach a timeout anyway, it's a
315 * lot cheaper to proceed like this because we almost never update
316 * the tree. We may also find disabled expiration dates there. Since
317 * we have detached the task from the tree, we simply call task_queue
318 * to take care of this. Note that we might occasionally requeue it at
319 * the same place, before <eb>, so we have to check if this happens,
320 * and adjust <eb>, otherwise we may skip it which is not what we want.
321 * We may also not requeue the task (and not point eb at it) if its
Willy Tarreau77015ab2020-06-19 11:50:27 +0200322 * expiration time is not set. We also make sure we leave the real
323 * expiration date for the next task in the queue so that when calling
324 * next_timer_expiry() we're guaranteed to see the next real date and
325 * not the next apparent date. This is in order to avoid useless
326 * wakeups.
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200327 */
Willy Tarreau77015ab2020-06-19 11:50:27 +0200328
329 task = eb32_entry(eb, struct task, wq);
330 if (tick_is_expired(task->expire, now_ms)) {
331 /* expired task, wake it up */
332 __task_unlink_wq(task);
333 task_wakeup(task, TASK_WOKEN_TIMER);
334 }
335 else if (task->expire != eb->key) {
336 /* task is not expired but its key doesn't match so let's
337 * update it and skip to next apparently expired task.
338 */
339 __task_unlink_wq(task);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200340 if (tick_isset(task->expire))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200341 __task_queue(task, &tt->timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200342 }
Willy Tarreau77015ab2020-06-19 11:50:27 +0200343 else {
344 /* task not expired and correctly placed */
345 break;
346 }
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200347 }
348
349#ifdef USE_THREAD
Willy Tarreau1e928c02019-05-28 18:57:25 +0200350 if (eb_is_empty(&timers))
351 goto leave;
352
353 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
354 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
355 if (!eb) {
356 eb = eb32_first(&timers);
357 if (likely(!eb)) {
358 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
359 goto leave;
360 }
361 }
362 key = eb->key;
Willy Tarreau1e928c02019-05-28 18:57:25 +0200363
Willy Tarreaud48ed662020-10-16 09:31:41 +0200364 if (tick_is_lt(now_ms, key)) {
365 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau1e928c02019-05-28 18:57:25 +0200366 goto leave;
Willy Tarreaud48ed662020-10-16 09:31:41 +0200367 }
Willy Tarreau1e928c02019-05-28 18:57:25 +0200368
369 /* There's really something of interest here, let's visit the queue */
370
Willy Tarreaud48ed662020-10-16 09:31:41 +0200371 if (HA_RWLOCK_TRYRDTOSK(TASK_WQ_LOCK, &wq_lock)) {
372 /* if we failed to grab the lock it means another thread is
373 * already doing the same here, so let it do the job.
374 */
375 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
376 goto leave;
377 }
378
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200379 while (1) {
Emeric Brunc60def82017-09-27 14:59:38 +0200380 lookup_next:
Willy Tarreau3cfaa8d2020-10-16 09:26:22 +0200381 if (max_processed-- <= 0)
382 break;
Emeric Brun01948972017-03-30 15:37:25 +0200383 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200384 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100385 /* we might have reached the end of the tree, typically because
386 * <now_ms> is in the first half and we're first scanning the last
387 * half. Let's loop back to the beginning of the tree now.
388 */
389 eb = eb32_first(&timers);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100390 if (likely(!eb))
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100391 break;
392 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200393
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100394 task = eb32_entry(eb, struct task, wq);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200395 if (tick_is_expired(task->expire, now_ms)) {
396 /* expired task, wake it up */
Willy Tarreaud48ed662020-10-16 09:31:41 +0200397 HA_RWLOCK_SKTOWR(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200398 __task_unlink_wq(task);
Willy Tarreaud48ed662020-10-16 09:31:41 +0200399 HA_RWLOCK_WRTOSK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200400 task_wakeup(task, TASK_WOKEN_TIMER);
401 }
402 else if (task->expire != eb->key) {
403 /* task is not expired but its key doesn't match so let's
404 * update it and skip to next apparently expired task.
405 */
Willy Tarreaud48ed662020-10-16 09:31:41 +0200406 HA_RWLOCK_SKTOWR(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200407 __task_unlink_wq(task);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100408 if (tick_isset(task->expire))
Willy Tarreau783afbe2020-07-22 14:12:45 +0200409 __task_queue(task, &timers);
Willy Tarreaud48ed662020-10-16 09:31:41 +0200410 HA_RWLOCK_WRTOSK(TASK_WQ_LOCK, &wq_lock);
Emeric Brunc60def82017-09-27 14:59:38 +0200411 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200412 }
Willy Tarreau77015ab2020-06-19 11:50:27 +0200413 else {
414 /* task not expired and correctly placed */
415 break;
416 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100417 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200418
Willy Tarreaud48ed662020-10-16 09:31:41 +0200419 HA_RWLOCK_SKUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200420#endif
Willy Tarreau1e928c02019-05-28 18:57:25 +0200421leave:
Willy Tarreauc49ba522019-12-11 08:12:23 +0100422 return;
423}
424
425/* Checks the next timer for the current thread by looking into its own timer
426 * list and the global one. It may return TICK_ETERNITY if no timer is present.
Ilya Shipitsin856aabc2020-04-16 23:51:34 +0500427 * Note that the next timer might very well be slightly in the past.
Willy Tarreauc49ba522019-12-11 08:12:23 +0100428 */
429int next_timer_expiry()
430{
431 struct task_per_thread * const tt = sched; // thread's tasks
432 struct eb32_node *eb;
433 int ret = TICK_ETERNITY;
Willy Tarreau6ce02322020-08-21 05:48:34 +0200434 __decl_thread(int key = TICK_ETERNITY);
Willy Tarreauc49ba522019-12-11 08:12:23 +0100435
436 /* first check in the thread-local timers */
437 eb = eb32_lookup_ge(&tt->timers, now_ms - TIMER_LOOK_BACK);
438 if (!eb) {
439 /* we might have reached the end of the tree, typically because
440 * <now_ms> is in the first half and we're first scanning the last
441 * half. Let's loop back to the beginning of the tree now.
442 */
443 eb = eb32_first(&tt->timers);
444 }
445
446 if (eb)
447 ret = eb->key;
448
449#ifdef USE_THREAD
450 if (!eb_is_empty(&timers)) {
451 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
452 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
453 if (!eb)
454 eb = eb32_first(&timers);
455 if (eb)
456 key = eb->key;
457 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
458 if (eb)
459 ret = tick_first(ret, key);
460 }
461#endif
Willy Tarreaub992ba12017-11-05 19:09:27 +0100462 return ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200463}
464
Willy Tarreau59153fe2020-06-24 10:17:29 +0200465/* Walks over tasklet lists sched->tasklets[0..TL_CLASSES-1] and run at most
466 * budget[TL_*] of them. Returns the number of entries effectively processed
467 * (tasks and tasklets merged). The count of tasks in the list for the current
468 * thread is adjusted.
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100469 */
Willy Tarreau59153fe2020-06-24 10:17:29 +0200470unsigned int run_tasks_from_lists(unsigned int budgets[])
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100471{
Willy Tarreau144f84a2021-03-02 16:09:26 +0100472 struct task *(*process)(struct task *t, void *ctx, unsigned int state);
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200473 struct list *tl_queues = sched->tasklets;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100474 struct task *t;
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200475 uint8_t budget_mask = (1 << TL_CLASSES) - 1;
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100476 struct sched_activity *profile_entry = NULL;
Willy Tarreau59153fe2020-06-24 10:17:29 +0200477 unsigned int done = 0;
478 unsigned int queue;
Willy Tarreau144f84a2021-03-02 16:09:26 +0100479 unsigned int state;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100480 void *ctx;
Willy Tarreau59153fe2020-06-24 10:17:29 +0200481
482 for (queue = 0; queue < TL_CLASSES;) {
483 sched->current_queue = queue;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100484
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200485 /* global.tune.sched.low-latency is set */
486 if (global.tune.options & GTUNE_SCHED_LOW_LATENCY) {
487 if (unlikely(sched->tl_class_mask & budget_mask & ((1 << queue) - 1))) {
488 /* a lower queue index has tasks again and still has a
489 * budget to run them. Let's switch to it now.
490 */
491 queue = (sched->tl_class_mask & 1) ? 0 :
492 (sched->tl_class_mask & 2) ? 1 : 2;
493 continue;
494 }
495
496 if (unlikely(queue > TL_URGENT &&
497 budget_mask & (1 << TL_URGENT) &&
498 !MT_LIST_ISEMPTY(&sched->shared_tasklet_list))) {
499 /* an urgent tasklet arrived from another thread */
500 break;
501 }
502
503 if (unlikely(queue > TL_NORMAL &&
504 budget_mask & (1 << TL_NORMAL) &&
Willy Tarreau2c41d772021-02-24 16:13:03 +0100505 (!eb_is_empty(&sched->rqueue) ||
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200506 (global_tasks_mask & tid_bit)))) {
507 /* a task was woken up by a bulk tasklet or another thread */
508 break;
509 }
510 }
511
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200512 if (LIST_ISEMPTY(&tl_queues[queue])) {
513 sched->tl_class_mask &= ~(1 << queue);
Willy Tarreau59153fe2020-06-24 10:17:29 +0200514 queue++;
515 continue;
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200516 }
517
Willy Tarreau59153fe2020-06-24 10:17:29 +0200518 if (!budgets[queue]) {
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200519 budget_mask &= ~(1 << queue);
Willy Tarreau59153fe2020-06-24 10:17:29 +0200520 queue++;
521 continue;
522 }
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200523
Willy Tarreau59153fe2020-06-24 10:17:29 +0200524 budgets[queue]--;
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200525 t = (struct task *)LIST_ELEM(tl_queues[queue].n, struct tasklet *, list);
Amaury Denoyelle7b368332021-07-28 16:12:57 +0200526 state = t->state & (TASK_SHARED_WQ|TASK_SELF_WAKING|TASK_HEAVY|TASK_F_TASKLET|TASK_KILLED|TASK_F_USR1|TASK_KILLED);
Willy Tarreau74dea8c2021-02-26 00:25:51 +0100527
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100528 ti->flags &= ~TI_FL_STUCK; // this thread is still running
529 activity[tid].ctxsw++;
530 ctx = t->context;
531 process = t->process;
532 t->calls++;
Willy Tarreaud23d4132020-01-31 10:39:03 +0100533 sched->current = t;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100534
Willy Tarreau4781b152021-04-06 13:53:36 +0200535 _HA_ATOMIC_DEC(&sched->rq_total);
Willy Tarreau2da4c312020-11-30 14:52:11 +0100536
Willy Tarreaudb4e2382021-03-02 15:54:11 +0100537 if (state & TASK_F_TASKLET) {
Willy Tarreau2a54ffb2021-02-25 09:32:58 +0100538 uint64_t before = 0;
539
Willy Tarreau4d6c5942020-11-30 14:58:53 +0100540 LIST_DEL_INIT(&((struct tasklet *)t)->list);
541 __ha_barrier_store();
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100542
543 if (unlikely(task_profiling_mask & tid_bit)) {
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100544 profile_entry = sched_activity_entry(sched_activity, t->process);
545 before = now_mono_time();
Willy Tarreaub2285de2021-02-25 08:39:07 +0100546#ifdef DEBUG_TASK
Willy Tarreau2a54ffb2021-02-25 09:32:58 +0100547 if (((struct tasklet *)t)->call_date) {
548 HA_ATOMIC_ADD(&profile_entry->lat_time, before - ((struct tasklet *)t)->call_date);
549 ((struct tasklet *)t)->call_date = 0;
550 }
Willy Tarreaub2285de2021-02-25 08:39:07 +0100551#endif
Willy Tarreau2a54ffb2021-02-25 09:32:58 +0100552 }
553
554 state = _HA_ATOMIC_XCHG(&t->state, state);
555 __ha_barrier_atomic_store();
556
Amaury Denoyelle7b368332021-07-28 16:12:57 +0200557 if (likely(!(state & TASK_KILLED))) {
558 process(t, ctx, state);
559 }
560 else {
561 done++;
562 sched->current = NULL;
563 pool_free(pool_head_tasklet, t);
564 __ha_barrier_store();
565 continue;
566 }
Willy Tarreau2a54ffb2021-02-25 09:32:58 +0100567
568 if (unlikely(task_profiling_mask & tid_bit)) {
Willy Tarreau4781b152021-04-06 13:53:36 +0200569 HA_ATOMIC_INC(&profile_entry->calls);
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100570 HA_ATOMIC_ADD(&profile_entry->cpu_time, now_mono_time() - before);
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100571 }
Willy Tarreau2a54ffb2021-02-25 09:32:58 +0100572
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100573 done++;
Willy Tarreaud23d4132020-01-31 10:39:03 +0100574 sched->current = NULL;
575 __ha_barrier_store();
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100576 continue;
577 }
578
Willy Tarreau4d6c5942020-11-30 14:58:53 +0100579 LIST_DEL_INIT(&((struct tasklet *)t)->list);
580 __ha_barrier_store();
Willy Tarreau6fa8bcd2021-03-02 16:26:05 +0100581 state = _HA_ATOMIC_XCHG(&t->state, state|TASK_RUNNING|TASK_F_USR1);
Willy Tarreau952c2642020-01-31 16:39:30 +0100582 __ha_barrier_atomic_store();
Willy Tarreau952c2642020-01-31 16:39:30 +0100583
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100584 /* OK then this is a regular task */
585
Willy Tarreau4781b152021-04-06 13:53:36 +0200586 _HA_ATOMIC_DEC(&task_per_thread[tid].tasks_in_list);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100587 if (unlikely(t->call_date)) {
588 uint64_t now_ns = now_mono_time();
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100589 uint64_t lat = now_ns - t->call_date;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100590
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100591 t->lat_time += lat;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100592 t->call_date = now_ns;
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100593 profile_entry = sched_activity_entry(sched_activity, t->process);
594 HA_ATOMIC_ADD(&profile_entry->lat_time, lat);
Willy Tarreau4781b152021-04-06 13:53:36 +0200595 HA_ATOMIC_INC(&profile_entry->calls);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100596 }
597
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100598 __ha_barrier_store();
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200599
600 /* Note for below: if TASK_KILLED arrived before we've read the state, we
601 * directly free the task. Otherwise it will be seen after processing and
602 * it's freed on the exit path.
603 */
604 if (likely(!(state & TASK_KILLED) && process == process_stream))
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100605 t = process_stream(t, ctx, state);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200606 else if (!(state & TASK_KILLED) && process != NULL)
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100607 t = process(t, ctx, state);
608 else {
Willy Tarreau273aea42020-07-17 14:37:51 +0200609 task_unlink_wq(t);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100610 __task_free(t);
611 sched->current = NULL;
612 __ha_barrier_store();
613 /* We don't want max_processed to be decremented if
614 * we're just freeing a destroyed task, we should only
615 * do so if we really ran a task.
616 */
617 continue;
618 }
619 sched->current = NULL;
620 __ha_barrier_store();
621 /* If there is a pending state we have to wake up the task
622 * immediately, else we defer it into wait queue
623 */
624 if (t != NULL) {
625 if (unlikely(t->call_date)) {
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100626 uint64_t cpu = now_mono_time() - t->call_date;
627
628 t->cpu_time += cpu;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100629 t->call_date = 0;
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100630 HA_ATOMIC_ADD(&profile_entry->cpu_time, cpu);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100631 }
632
Willy Tarreau1db42732021-04-06 11:44:07 +0200633 state = _HA_ATOMIC_AND_FETCH(&t->state, ~TASK_RUNNING);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200634 if (unlikely(state & TASK_KILLED)) {
Willy Tarreau273aea42020-07-17 14:37:51 +0200635 task_unlink_wq(t);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200636 __task_free(t);
637 }
638 else if (state & TASK_WOKEN_ANY)
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100639 task_wakeup(t, 0);
640 else
641 task_queue(t);
642 }
643 done++;
644 }
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200645 sched->current_queue = -1;
Willy Tarreau116ef222020-06-23 16:35:38 +0200646
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100647 return done;
648}
649
Willy Tarreau58b458d2008-06-29 22:40:23 +0200650/* The run queue is chronologically sorted in a tree. An insertion counter is
651 * used to assign a position to each task. This counter may be combined with
652 * other variables (eg: nice value) to set the final position in the tree. The
653 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100654 * tasks processed to 200 in any case, so that general latency remains low and
Willy Tarreaucde79022019-04-12 18:03:41 +0200655 * so that task positions have a chance to be considered. The function scans
656 * both the global and local run queues and picks the most urgent task between
657 * the two. We need to grab the global runqueue lock to touch it so it's taken
658 * on the very first access to the global run queue and is released as soon as
659 * it reaches the end.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200660 *
661 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200662 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100663void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200664{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200665 struct task_per_thread * const tt = sched;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200666 struct eb32sc_node *lrq; // next local run queue entry
667 struct eb32sc_node *grq; // next global run queue entry
Willy Tarreau964c9362007-01-07 00:38:00 +0100668 struct task *t;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200669 const unsigned int default_weights[TL_CLASSES] = {
670 [TL_URGENT] = 64, // ~50% of CPU bandwidth for I/O
671 [TL_NORMAL] = 48, // ~37% of CPU bandwidth for tasks
672 [TL_BULK] = 16, // ~13% of CPU bandwidth for self-wakers
Willy Tarreau401135c2021-02-26 09:16:22 +0100673 [TL_HEAVY] = 1, // never more than 1 heavy task at once
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200674 };
675 unsigned int max[TL_CLASSES]; // max to be run per class
676 unsigned int max_total; // sum of max above
Olivier Houchard06910462019-10-11 16:35:01 +0200677 struct mt_list *tmp_list;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200678 unsigned int queue;
679 int max_processed;
Willy Tarreaue7923c12021-02-25 07:09:08 +0100680 int lpicked, gpicked;
Willy Tarreau76390da2021-02-26 10:18:11 +0100681 int heavy_queued = 0;
Willy Tarreauc309dbd2020-11-30 15:39:00 +0100682 int budget;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100683
Willy Tarreaue6a02fa2019-05-22 07:06:44 +0200684 ti->flags &= ~TI_FL_STUCK; // this thread is still running
685
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200686 if (!thread_has_tasks()) {
687 activity[tid].empty_rq++;
688 return;
689 }
690
Willy Tarreau5c8be272020-06-19 12:17:55 +0200691 max_processed = global.tune.runqueue_depth;
692
693 if (likely(niced_tasks))
694 max_processed = (max_processed + 3) / 4;
695
Willy Tarreau1691ba32021-03-10 09:26:24 +0100696 if (max_processed < sched->rq_total && sched->rq_total <= 2*max_processed) {
697 /* If the run queue exceeds the budget by up to 50%, let's cut it
698 * into two identical halves to improve latency.
699 */
700 max_processed = sched->rq_total / 2;
701 }
702
Willy Tarreau5c8be272020-06-19 12:17:55 +0200703 not_done_yet:
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200704 max[TL_URGENT] = max[TL_NORMAL] = max[TL_BULK] = 0;
Willy Tarreaucde79022019-04-12 18:03:41 +0200705
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200706 /* urgent tasklets list gets a default weight of ~50% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200707 if ((tt->tl_class_mask & (1 << TL_URGENT)) ||
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200708 !MT_LIST_ISEMPTY(&tt->shared_tasklet_list))
709 max[TL_URGENT] = default_weights[TL_URGENT];
Willy Tarreaua62917b2020-01-30 18:37:28 +0100710
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200711 /* normal tasklets list gets a default weight of ~37% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200712 if ((tt->tl_class_mask & (1 << TL_NORMAL)) ||
Willy Tarreau2c41d772021-02-24 16:13:03 +0100713 !eb_is_empty(&sched->rqueue) || (global_tasks_mask & tid_bit))
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200714 max[TL_NORMAL] = default_weights[TL_NORMAL];
Willy Tarreaua62917b2020-01-30 18:37:28 +0100715
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200716 /* bulk tasklets list gets a default weight of ~13% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200717 if ((tt->tl_class_mask & (1 << TL_BULK)))
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200718 max[TL_BULK] = default_weights[TL_BULK];
719
Willy Tarreau401135c2021-02-26 09:16:22 +0100720 /* heavy tasks are processed only once and never refilled in a
Willy Tarreau76390da2021-02-26 10:18:11 +0100721 * call round. That budget is not lost either as we don't reset
722 * it unless consumed.
Willy Tarreau401135c2021-02-26 09:16:22 +0100723 */
Willy Tarreau76390da2021-02-26 10:18:11 +0100724 if (!heavy_queued) {
725 if ((tt->tl_class_mask & (1 << TL_HEAVY)))
726 max[TL_HEAVY] = default_weights[TL_HEAVY];
727 else
728 max[TL_HEAVY] = 0;
729 heavy_queued = 1;
730 }
Willy Tarreau401135c2021-02-26 09:16:22 +0100731
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200732 /* Now compute a fair share of the weights. Total may slightly exceed
Willy Tarreau1553b662020-06-30 13:46:21 +0200733 * 100% due to rounding, this is not a problem. Note that while in
734 * theory the sum cannot be NULL as we cannot get there without tasklets
735 * to process, in practice it seldom happens when multiple writers
Willy Tarreau2b718102021-04-21 07:32:39 +0200736 * conflict and rollback on MT_LIST_TRY_APPEND(shared_tasklet_list), causing
Willy Tarreau1553b662020-06-30 13:46:21 +0200737 * a first MT_LIST_ISEMPTY() to succeed for thread_has_task() and the
738 * one above to finally fail. This is extremely rare and not a problem.
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200739 */
Willy Tarreau401135c2021-02-26 09:16:22 +0100740 max_total = max[TL_URGENT] + max[TL_NORMAL] + max[TL_BULK] + max[TL_HEAVY];
Willy Tarreau1553b662020-06-30 13:46:21 +0200741 if (!max_total)
742 return;
743
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200744 for (queue = 0; queue < TL_CLASSES; queue++)
745 max[queue] = ((unsigned)max_processed * max[queue] + max_total - 1) / max_total;
746
Willy Tarreau76390da2021-02-26 10:18:11 +0100747 /* The heavy queue must never process more than one task at once
748 * anyway.
749 */
750 if (max[TL_HEAVY] > 1)
751 max[TL_HEAVY] = 1;
752
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200753 lrq = grq = NULL;
Christopher Faulet8a48f672017-11-14 10:38:36 +0100754
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200755 /* pick up to max[TL_NORMAL] regular tasks from prio-ordered run queues */
756 /* Note: the grq lock is always held when grq is not null */
Willy Tarreaue7923c12021-02-25 07:09:08 +0100757 lpicked = gpicked = 0;
Willy Tarreau1f3b1412021-02-24 14:13:40 +0100758 budget = max[TL_NORMAL] - tt->tasks_in_list;
Willy Tarreaue7923c12021-02-25 07:09:08 +0100759 while (lpicked + gpicked < budget) {
Willy Tarreaucde79022019-04-12 18:03:41 +0200760 if ((global_tasks_mask & tid_bit) && !grq) {
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200761#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200762 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100763 grq = eb32sc_lookup_ge(&rqueue, global_rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreaucde79022019-04-12 18:03:41 +0200764 if (unlikely(!grq)) {
765 grq = eb32sc_first(&rqueue, tid_bit);
766 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200767 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200768 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200769 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100770 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200771#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200772 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100773
Willy Tarreaucde79022019-04-12 18:03:41 +0200774 /* If a global task is available for this thread, it's in grq
775 * now and the global RQ is locked.
776 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200777
Willy Tarreaucde79022019-04-12 18:03:41 +0200778 if (!lrq) {
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100779 lrq = eb32sc_lookup_ge(&tt->rqueue, tt->rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreaucde79022019-04-12 18:03:41 +0200780 if (unlikely(!lrq))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200781 lrq = eb32sc_first(&tt->rqueue, tid_bit);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100782 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100783
Willy Tarreaucde79022019-04-12 18:03:41 +0200784 if (!lrq && !grq)
785 break;
786
787 if (likely(!grq || (lrq && (int)(lrq->key - grq->key) <= 0))) {
788 t = eb32sc_entry(lrq, struct task, rq);
789 lrq = eb32sc_next(lrq, tid_bit);
Willy Tarreau2b363ac2021-02-25 07:14:58 +0100790 eb32sc_delete(&t->rq);
Willy Tarreaue7923c12021-02-25 07:09:08 +0100791 lpicked++;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200792 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200793#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200794 else {
795 t = eb32sc_entry(grq, struct task, rq);
796 grq = eb32sc_next(grq, tid_bit);
Willy Tarreau2b363ac2021-02-25 07:14:58 +0100797 _HA_ATOMIC_AND(&t->state, ~TASK_GLOBAL);
798 eb32sc_delete(&t->rq);
799
Willy Tarreaucde79022019-04-12 18:03:41 +0200800 if (unlikely(!grq)) {
801 grq = eb32sc_first(&rqueue, tid_bit);
802 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200803 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200804 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaucde79022019-04-12 18:03:41 +0200805 }
806 }
Willy Tarreaue7923c12021-02-25 07:09:08 +0100807 gpicked++;
Emeric Brun01948972017-03-30 15:37:25 +0200808 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200809#endif
Willy Tarreau2b363ac2021-02-25 07:14:58 +0100810 if (t->nice)
Willy Tarreau4781b152021-04-06 13:53:36 +0200811 _HA_ATOMIC_DEC(&niced_tasks);
Willy Tarreaucde79022019-04-12 18:03:41 +0200812
Willy Tarreaua868c292020-11-30 15:30:22 +0100813 /* Add it to the local task list */
Willy Tarreau2b718102021-04-21 07:32:39 +0200814 LIST_APPEND(&tt->tasklets[TL_NORMAL], &((struct tasklet *)t)->list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200815 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200816
817 /* release the rqueue lock */
818 if (grq) {
819 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
820 grq = NULL;
821 }
822
Willy Tarreaue7923c12021-02-25 07:09:08 +0100823 if (lpicked + gpicked) {
Willy Tarreauc309dbd2020-11-30 15:39:00 +0100824 tt->tl_class_mask |= 1 << TL_NORMAL;
Willy Tarreaue7923c12021-02-25 07:09:08 +0100825 _HA_ATOMIC_ADD(&tt->tasks_in_list, lpicked + gpicked);
Willy Tarreaub7e0c632021-03-09 09:59:50 +0100826#ifdef USE_THREAD
Willy Tarreau45499c52021-02-25 07:51:18 +0100827 if (gpicked) {
828 _HA_ATOMIC_SUB(&grq_total, gpicked);
Willy Tarreauc9afbb12021-02-25 07:19:45 +0100829 _HA_ATOMIC_ADD(&tt->rq_total, gpicked);
Willy Tarreau45499c52021-02-25 07:51:18 +0100830 }
Willy Tarreaub7e0c632021-03-09 09:59:50 +0100831#endif
Willy Tarreaue7923c12021-02-25 07:09:08 +0100832 activity[tid].tasksw += lpicked + gpicked;
Willy Tarreauc309dbd2020-11-30 15:39:00 +0100833 }
834
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200835 /* Merge the list of tasklets waken up by other threads to the
836 * main list.
837 */
838 tmp_list = MT_LIST_BEHEAD(&tt->shared_tasklet_list);
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200839 if (tmp_list) {
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200840 LIST_SPLICE_END_DETACHED(&tt->tasklets[TL_URGENT], (struct list *)tmp_list);
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200841 if (!LIST_ISEMPTY(&tt->tasklets[TL_URGENT]))
842 tt->tl_class_mask |= 1 << TL_URGENT;
843 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200844
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200845 /* execute tasklets in each queue */
Willy Tarreau59153fe2020-06-24 10:17:29 +0200846 max_processed -= run_tasks_from_lists(max);
Willy Tarreaua62917b2020-01-30 18:37:28 +0100847
Willy Tarreau5c8be272020-06-19 12:17:55 +0200848 /* some tasks may have woken other ones up */
Willy Tarreau0c0c85e2020-06-23 11:32:35 +0200849 if (max_processed > 0 && thread_has_tasks())
Willy Tarreau5c8be272020-06-19 12:17:55 +0200850 goto not_done_yet;
851
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200852 if (tt->tl_class_mask)
Willy Tarreaucde79022019-04-12 18:03:41 +0200853 activity[tid].long_rq++;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200854}
855
Willy Tarreau64e60122019-07-12 08:31:17 +0200856/* create a work list array for <nbthread> threads, using tasks made of
857 * function <fct>. The context passed to the function will be the pointer to
858 * the thread's work list, which will contain a copy of argument <arg>. The
859 * wake up reason will be TASK_WOKEN_OTHER. The pointer to the work_list array
860 * is returned on success, otherwise NULL on failure.
861 */
862struct work_list *work_list_create(int nbthread,
Willy Tarreau144f84a2021-03-02 16:09:26 +0100863 struct task *(*fct)(struct task *, void *, unsigned int),
Willy Tarreau64e60122019-07-12 08:31:17 +0200864 void *arg)
865{
866 struct work_list *wl;
867 int i;
868
869 wl = calloc(nbthread, sizeof(*wl));
870 if (!wl)
871 goto fail;
872
873 for (i = 0; i < nbthread; i++) {
Olivier Houchard859dc802019-08-08 15:47:21 +0200874 MT_LIST_INIT(&wl[i].head);
Willy Tarreau64e60122019-07-12 08:31:17 +0200875 wl[i].task = task_new(1UL << i);
876 if (!wl[i].task)
877 goto fail;
878 wl[i].task->process = fct;
879 wl[i].task->context = &wl[i];
880 wl[i].arg = arg;
881 }
882 return wl;
883
884 fail:
885 work_list_destroy(wl, nbthread);
886 return NULL;
887}
888
889/* destroy work list <work> */
890void work_list_destroy(struct work_list *work, int nbthread)
891{
892 int t;
893
894 if (!work)
895 return;
896 for (t = 0; t < nbthread; t++)
897 task_destroy(work[t].task);
898 free(work);
899}
900
William Lallemand27f3fa52018-12-06 14:05:20 +0100901/*
902 * Delete every tasks before running the master polling loop
903 */
904void mworker_cleantasks()
905{
906 struct task *t;
907 int i;
William Lallemandb5823392018-12-06 15:14:37 +0100908 struct eb32_node *tmp_wq = NULL;
909 struct eb32sc_node *tmp_rq = NULL;
William Lallemand27f3fa52018-12-06 14:05:20 +0100910
911#ifdef USE_THREAD
912 /* cleanup the global run queue */
William Lallemandb5823392018-12-06 15:14:37 +0100913 tmp_rq = eb32sc_first(&rqueue, MAX_THREADS_MASK);
914 while (tmp_rq) {
915 t = eb32sc_entry(tmp_rq, struct task, rq);
916 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200917 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100918 }
919 /* cleanup the timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100920 tmp_wq = eb32_first(&timers);
921 while (tmp_wq) {
922 t = eb32_entry(tmp_wq, struct task, wq);
923 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200924 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100925 }
926#endif
927 /* clean the per thread run queue */
928 for (i = 0; i < global.nbthread; i++) {
William Lallemandb5823392018-12-06 15:14:37 +0100929 tmp_rq = eb32sc_first(&task_per_thread[i].rqueue, MAX_THREADS_MASK);
930 while (tmp_rq) {
931 t = eb32sc_entry(tmp_rq, struct task, rq);
932 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200933 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100934 }
935 /* cleanup the per thread timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100936 tmp_wq = eb32_first(&task_per_thread[i].timers);
937 while (tmp_wq) {
938 t = eb32_entry(tmp_wq, struct task, wq);
939 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200940 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100941 }
942 }
943}
944
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100945/* perform minimal intializations */
946static void init_task()
Willy Tarreau4726f532009-03-07 17:25:21 +0100947{
Willy Tarreau401135c2021-02-26 09:16:22 +0100948 int i, q;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200949
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200950#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200951 memset(&timers, 0, sizeof(timers));
Willy Tarreau4726f532009-03-07 17:25:21 +0100952 memset(&rqueue, 0, sizeof(rqueue));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200953#endif
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200954 memset(&task_per_thread, 0, sizeof(task_per_thread));
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200955 for (i = 0; i < MAX_THREADS; i++) {
Willy Tarreau401135c2021-02-26 09:16:22 +0100956 for (q = 0; q < TL_CLASSES; q++)
957 LIST_INIT(&task_per_thread[i].tasklets[q]);
Olivier Houchard06910462019-10-11 16:35:01 +0200958 MT_LIST_INIT(&task_per_thread[i].shared_tasklet_list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200959 }
Willy Tarreau4726f532009-03-07 17:25:21 +0100960}
961
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200962/* config parser for global "tune.sched.low-latency", accepts "on" or "off" */
963static int cfg_parse_tune_sched_low_latency(char **args, int section_type, struct proxy *curpx,
Willy Tarreau01825162021-03-09 09:53:46 +0100964 const struct proxy *defpx, const char *file, int line,
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200965 char **err)
966{
967 if (too_many_args(1, args, err, NULL))
968 return -1;
969
970 if (strcmp(args[1], "on") == 0)
971 global.tune.options |= GTUNE_SCHED_LOW_LATENCY;
972 else if (strcmp(args[1], "off") == 0)
973 global.tune.options &= ~GTUNE_SCHED_LOW_LATENCY;
974 else {
975 memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
976 return -1;
977 }
978 return 0;
979}
980
981/* config keyword parsers */
982static struct cfg_kw_list cfg_kws = {ILH, {
983 { CFG_GLOBAL, "tune.sched.low-latency", cfg_parse_tune_sched_low_latency },
984 { 0, NULL, NULL }
985}};
986
987INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100988INITCALL0(STG_PREPARE, init_task);
989
Willy Tarreaubaaee002006-06-26 02:48:02 +0200990/*
991 * Local variables:
992 * c-indent-level: 8
993 * c-basic-offset: 8
994 * End:
995 */