blob: cabfcbe1ec8823c03046448f48a49ae44a9f08a9 [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
Willy Tarreaua4613182009-03-21 18:13:21 +010038unsigned int nb_tasks = 0;
Olivier Houchardeba0c0b2018-07-26 18:53:28 +020039volatile unsigned long global_tasks_mask = 0; /* Mask of threads with tasks in the global runqueue */
Christopher Faulet34c5cc92016-12-06 09:15:30 +010040unsigned int tasks_run_queue = 0;
41unsigned int tasks_run_queue_cur = 0; /* copy of the run queue size */
Willy Tarreauc7bdf092009-03-21 18:33:52 +010042unsigned int nb_tasks_cur = 0; /* copy of the tasks count */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010043unsigned int niced_tasks = 0; /* number of niced tasks in the run queue */
Emeric Brun01948972017-03-30 15:37:25 +020044
Willy Tarreaud022e9c2019-09-24 08:25:15 +020045THREAD_LOCAL struct task_per_thread *sched = &task_per_thread[0]; /* scheduler context for the current thread */
Willy Tarreau6d1222c2017-11-26 10:08:06 +010046
Willy Tarreau86abe442018-11-25 20:12:18 +010047__decl_aligned_spinlock(rq_lock); /* spin lock related to run queue */
Willy Tarreauef28dc12019-05-28 18:48:07 +020048__decl_aligned_rwlock(wq_lock); /* RW lock related to the wait queue */
Willy Tarreau964c9362007-01-07 00:38:00 +010049
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020050#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +020051struct eb_root timers; /* sorted timers tree, global */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020052struct eb_root rqueue; /* tree constituting the run queue */
Olivier Houchard77551ee2018-07-26 15:59:38 +020053int global_rqueue_size; /* Number of element sin the global runqueue */
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020054#endif
Willy Tarreaub20aa9e2018-10-15 14:52:21 +020055
Willy Tarreaue35c94a2009-03-21 10:01:42 +010056static unsigned int rqueue_ticks; /* insertion count */
Willy Tarreau8d8747a2018-10-15 16:12:48 +020057
58struct task_per_thread task_per_thread[MAX_THREADS];
Willy Tarreau9789f7b2008-06-24 08:17:16 +020059
Willy Tarreaueb8c2c62020-06-30 11:48:48 +020060
61/* Flags the task <t> for immediate destruction and puts it into its first
62 * thread's shared tasklet list if not yet queued/running. This will bypass
63 * the priority scheduling and make the task show up as fast as possible in
64 * the other thread's queue. Note that this operation isn't idempotent and is
65 * not supposed to be run on the same task from multiple threads at once. It's
66 * the caller's responsibility to make sure it is the only one able to kill the
67 * task.
68 */
69void task_kill(struct task *t)
70{
71 unsigned short state = t->state;
72 unsigned int thr;
73
74 BUG_ON(state & TASK_KILLED);
75
76 while (1) {
77 while (state & (TASK_RUNNING | TASK_QUEUED)) {
78 /* task already in the queue and about to be executed,
79 * or even currently running. Just add the flag and be
80 * done with it, the process loop will detect it and kill
81 * it. The CAS will fail if we arrive too late.
82 */
83 if (_HA_ATOMIC_CAS(&t->state, &state, state | TASK_KILLED))
84 return;
85 }
86
87 /* We'll have to wake it up, but we must also secure it so that
88 * it doesn't vanish under us. TASK_QUEUED guarantees nobody will
89 * add past us.
90 */
91 if (_HA_ATOMIC_CAS(&t->state, &state, state | TASK_QUEUED | TASK_KILLED)) {
92 /* Bypass the tree and go directly into the shared tasklet list.
93 * Note: that's a task so it must be accounted for as such. Pick
94 * the task's first thread for the job.
95 */
96 thr = my_ffsl(t->thread_mask) - 1;
Willy Tarreau54d31172020-07-02 14:14:00 +020097
98 /* Beware: tasks that have never run don't have their ->list empty yet! */
Willy Tarreau950954f2020-07-10 08:32:10 +020099 MT_LIST_ADDQ(&task_per_thread[thr].shared_tasklet_list,
Willy Tarreau4f589262020-07-02 17:17:42 +0200100 (struct mt_list *)&((struct tasklet *)t)->list);
Willy Tarreau54d31172020-07-02 14:14:00 +0200101 _HA_ATOMIC_ADD(&tasks_run_queue, 1);
102 _HA_ATOMIC_ADD(&task_per_thread[thr].task_list_size, 1);
103 if (sleeping_thread_mask & (1UL << thr)) {
104 _HA_ATOMIC_AND(&sleeping_thread_mask, ~(1UL << thr));
105 wake_thread(thr);
Willy Tarreaueb8c2c62020-06-30 11:48:48 +0200106 }
Willy Tarreau54d31172020-07-02 14:14:00 +0200107 return;
Willy Tarreaueb8c2c62020-06-30 11:48:48 +0200108 }
109 }
110}
111
Willy Tarreau4726f532009-03-07 17:25:21 +0100112/* Puts the task <t> in run queue at a position depending on t->nice. <t> is
113 * returned. The nice value assigns boosts in 32th of the run queue size. A
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100114 * nice value of -1024 sets the task to -tasks_run_queue*32, while a nice value
115 * of 1024 sets the task to tasks_run_queue*32. The state flags are cleared, so
116 * the caller will have to set its flags after this call.
Willy Tarreau4726f532009-03-07 17:25:21 +0100117 * The task must not already be in the run queue. If unsure, use the safer
118 * task_wakeup() function.
Willy Tarreau91e99932008-06-30 07:51:00 +0200119 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200120void __task_wakeup(struct task *t, struct eb_root *root)
Willy Tarreaue33aece2007-04-30 13:15:14 +0200121{
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200122#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200123 if (root == &rqueue) {
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200124 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200125 }
Willy Tarreau2d1fd0a2019-04-15 09:18:31 +0200126#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200127 /* Make sure if the task isn't in the runqueue, nobody inserts it
128 * in the meanwhile.
129 */
Olivier Houchard4c2832852019-03-08 18:48:47 +0100130 _HA_ATOMIC_ADD(&tasks_run_queue, 1);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200131#ifdef USE_THREAD
132 if (root == &rqueue) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200133 global_tasks_mask |= t->thread_mask;
Olivier Houcharded1a6a02019-04-18 14:12:51 +0200134 __ha_barrier_store();
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200135 }
136#endif
Olivier Houchard4c2832852019-03-08 18:48:47 +0100137 t->rq.key = _HA_ATOMIC_ADD(&rqueue_ticks, 1);
Willy Tarreau91e99932008-06-30 07:51:00 +0200138
139 if (likely(t->nice)) {
140 int offset;
141
Olivier Houchard4c2832852019-03-08 18:48:47 +0100142 _HA_ATOMIC_ADD(&niced_tasks, 1);
Willy Tarreau2d1fd0a2019-04-15 09:18:31 +0200143 offset = t->nice * (int)global.tune.runqueue_depth;
Willy Tarreau4726f532009-03-07 17:25:21 +0100144 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +0200145 }
146
Willy Tarreaud9add3a2019-04-25 08:57:41 +0200147 if (task_profiling_mask & tid_bit)
Willy Tarreau9efd7452018-05-31 14:48:54 +0200148 t->call_date = now_mono_time();
149
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200150 eb32sc_insert(root, &t->rq, t->thread_mask);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200151#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200152 if (root == &rqueue) {
153 global_rqueue_size++;
Olivier Houchard4c2832852019-03-08 18:48:47 +0100154 _HA_ATOMIC_OR(&t->state, TASK_GLOBAL);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200155 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200156 } else
157#endif
158 {
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200159 int nb = ((void *)root - (void *)&task_per_thread[0].rqueue) / sizeof(task_per_thread[0]);
160 task_per_thread[nb].rqueue_size++;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200161 }
Willy Tarreau85d9b842018-07-27 17:14:41 +0200162#ifdef USE_THREAD
Olivier Houchard79321b92018-07-26 17:55:11 +0200163 /* If all threads that are supposed to handle this task are sleeping,
164 * wake one.
165 */
166 if ((((t->thread_mask & all_threads_mask) & sleeping_thread_mask) ==
Olivier Houchard1b327902019-03-15 00:23:10 +0100167 (t->thread_mask & all_threads_mask))) {
168 unsigned long m = (t->thread_mask & all_threads_mask) &~ tid_bit;
169
170 m = (m & (m - 1)) ^ m; // keep lowest bit set
171 _HA_ATOMIC_AND(&sleeping_thread_mask, ~m);
172 wake_thread(my_ffsl(m) - 1);
173 }
Willy Tarreau85d9b842018-07-27 17:14:41 +0200174#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200175 return;
Willy Tarreaue33aece2007-04-30 13:15:14 +0200176}
Willy Tarreaud825eef2007-05-12 22:35:00 +0200177
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200178/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100179 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200180 *
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200181 * Inserts a task into wait queue <wq> at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +0100182 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100183 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100184 * Last, tasks must not be queued further than the end of the tree, which is
185 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100186 *
187 * This function should not be used directly, it is meant to be called by the
188 * inline version of task_queue() which performs a few cheap preliminary tests
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200189 * before deciding to call __task_queue(). Moreover this function doesn't care
190 * at all about locking so the caller must be careful when deciding whether to
191 * lock or not around this call.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200192 */
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200193void __task_queue(struct task *task, struct eb_root *wq)
Willy Tarreau964c9362007-01-07 00:38:00 +0100194{
Willy Tarreaue5d79bc2020-07-22 14:29:42 +0200195#ifdef USE_THREAD
196 BUG_ON((wq == &timers && !(task->state & TASK_SHARED_WQ)) ||
197 (wq == &sched->timers && (task->state & TASK_SHARED_WQ)) ||
198 (wq != &timers && wq != &sched->timers));
199#endif
200
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100201 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +0100202 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +0100203
204 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100205 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200206#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100207 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +0200208 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +0100209 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200210#endif
Willy Tarreauce44f122008-07-05 18:16:19 +0200211
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200212 eb32_insert(wq, &task->wq);
Willy Tarreau964c9362007-01-07 00:38:00 +0100213}
214
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200215/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200216 * Extract all expired timers from the timer queue, and wakes up all
Willy Tarreauc49ba522019-12-11 08:12:23 +0100217 * associated tasks.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200218 */
Willy Tarreauc49ba522019-12-11 08:12:23 +0100219void wake_expired_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200220{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200221 struct task_per_thread * const tt = sched; // thread's tasks
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200222 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200223 struct eb32_node *eb;
Willy Tarreauaf613e82020-06-05 08:40:51 +0200224 __decl_thread(int key);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200225
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100226 while (1) {
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200227 lookup_next_local:
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200228 eb = eb32_lookup_ge(&tt->timers, now_ms - TIMER_LOOK_BACK);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200229 if (!eb) {
230 /* we might have reached the end of the tree, typically because
231 * <now_ms> is in the first half and we're first scanning the last
232 * half. Let's loop back to the beginning of the tree now.
233 */
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200234 eb = eb32_first(&tt->timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200235 if (likely(!eb))
236 break;
237 }
238
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200239 /* It is possible that this task was left at an earlier place in the
240 * tree because a recent call to task_queue() has not moved it. This
241 * happens when the new expiration date is later than the old one.
242 * Since it is very unlikely that we reach a timeout anyway, it's a
243 * lot cheaper to proceed like this because we almost never update
244 * the tree. We may also find disabled expiration dates there. Since
245 * we have detached the task from the tree, we simply call task_queue
246 * to take care of this. Note that we might occasionally requeue it at
247 * the same place, before <eb>, so we have to check if this happens,
248 * and adjust <eb>, otherwise we may skip it which is not what we want.
249 * We may also not requeue the task (and not point eb at it) if its
Willy Tarreau77015ab2020-06-19 11:50:27 +0200250 * expiration time is not set. We also make sure we leave the real
251 * expiration date for the next task in the queue so that when calling
252 * next_timer_expiry() we're guaranteed to see the next real date and
253 * not the next apparent date. This is in order to avoid useless
254 * wakeups.
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200255 */
Willy Tarreau77015ab2020-06-19 11:50:27 +0200256
257 task = eb32_entry(eb, struct task, wq);
258 if (tick_is_expired(task->expire, now_ms)) {
259 /* expired task, wake it up */
260 __task_unlink_wq(task);
261 task_wakeup(task, TASK_WOKEN_TIMER);
262 }
263 else if (task->expire != eb->key) {
264 /* task is not expired but its key doesn't match so let's
265 * update it and skip to next apparently expired task.
266 */
267 __task_unlink_wq(task);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200268 if (tick_isset(task->expire))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200269 __task_queue(task, &tt->timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200270 }
Willy Tarreau77015ab2020-06-19 11:50:27 +0200271 else {
272 /* task not expired and correctly placed */
273 break;
274 }
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200275 }
276
277#ifdef USE_THREAD
Willy Tarreau1e928c02019-05-28 18:57:25 +0200278 if (eb_is_empty(&timers))
279 goto leave;
280
281 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
282 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
283 if (!eb) {
284 eb = eb32_first(&timers);
285 if (likely(!eb)) {
286 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
287 goto leave;
288 }
289 }
290 key = eb->key;
291 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
292
Willy Tarreauc49ba522019-12-11 08:12:23 +0100293 if (tick_is_lt(now_ms, key))
Willy Tarreau1e928c02019-05-28 18:57:25 +0200294 goto leave;
Willy Tarreau1e928c02019-05-28 18:57:25 +0200295
296 /* There's really something of interest here, let's visit the queue */
297
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200298 while (1) {
Willy Tarreauef28dc12019-05-28 18:48:07 +0200299 HA_RWLOCK_WRLOCK(TASK_WQ_LOCK, &wq_lock);
Emeric Brunc60def82017-09-27 14:59:38 +0200300 lookup_next:
Emeric Brun01948972017-03-30 15:37:25 +0200301 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200302 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100303 /* we might have reached the end of the tree, typically because
304 * <now_ms> is in the first half and we're first scanning the last
305 * half. Let's loop back to the beginning of the tree now.
306 */
307 eb = eb32_first(&timers);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100308 if (likely(!eb))
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100309 break;
310 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200311
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100312 task = eb32_entry(eb, struct task, wq);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200313 if (tick_is_expired(task->expire, now_ms)) {
314 /* expired task, wake it up */
315 __task_unlink_wq(task);
316 task_wakeup(task, TASK_WOKEN_TIMER);
317 }
318 else if (task->expire != eb->key) {
319 /* task is not expired but its key doesn't match so let's
320 * update it and skip to next apparently expired task.
321 */
322 __task_unlink_wq(task);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100323 if (tick_isset(task->expire))
Willy Tarreau783afbe2020-07-22 14:12:45 +0200324 __task_queue(task, &timers);
Emeric Brunc60def82017-09-27 14:59:38 +0200325 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200326 }
Willy Tarreau77015ab2020-06-19 11:50:27 +0200327 else {
328 /* task not expired and correctly placed */
329 break;
330 }
Willy Tarreauef28dc12019-05-28 18:48:07 +0200331 HA_RWLOCK_WRUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100332 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200333
Willy Tarreauef28dc12019-05-28 18:48:07 +0200334 HA_RWLOCK_WRUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200335#endif
Willy Tarreau1e928c02019-05-28 18:57:25 +0200336leave:
Willy Tarreauc49ba522019-12-11 08:12:23 +0100337 return;
338}
339
340/* Checks the next timer for the current thread by looking into its own timer
341 * list and the global one. It may return TICK_ETERNITY if no timer is present.
Ilya Shipitsin856aabc2020-04-16 23:51:34 +0500342 * Note that the next timer might very well be slightly in the past.
Willy Tarreauc49ba522019-12-11 08:12:23 +0100343 */
344int next_timer_expiry()
345{
346 struct task_per_thread * const tt = sched; // thread's tasks
347 struct eb32_node *eb;
348 int ret = TICK_ETERNITY;
Willy Tarreau6ce02322020-08-21 05:48:34 +0200349 __decl_thread(int key = TICK_ETERNITY);
Willy Tarreauc49ba522019-12-11 08:12:23 +0100350
351 /* first check in the thread-local timers */
352 eb = eb32_lookup_ge(&tt->timers, now_ms - TIMER_LOOK_BACK);
353 if (!eb) {
354 /* we might have reached the end of the tree, typically because
355 * <now_ms> is in the first half and we're first scanning the last
356 * half. Let's loop back to the beginning of the tree now.
357 */
358 eb = eb32_first(&tt->timers);
359 }
360
361 if (eb)
362 ret = eb->key;
363
364#ifdef USE_THREAD
365 if (!eb_is_empty(&timers)) {
366 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
367 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
368 if (!eb)
369 eb = eb32_first(&timers);
370 if (eb)
371 key = eb->key;
372 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
373 if (eb)
374 ret = tick_first(ret, key);
375 }
376#endif
Willy Tarreaub992ba12017-11-05 19:09:27 +0100377 return ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200378}
379
Willy Tarreau59153fe2020-06-24 10:17:29 +0200380/* Walks over tasklet lists sched->tasklets[0..TL_CLASSES-1] and run at most
381 * budget[TL_*] of them. Returns the number of entries effectively processed
382 * (tasks and tasklets merged). The count of tasks in the list for the current
383 * thread is adjusted.
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100384 */
Willy Tarreau59153fe2020-06-24 10:17:29 +0200385unsigned int run_tasks_from_lists(unsigned int budgets[])
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100386{
387 struct task *(*process)(struct task *t, void *ctx, unsigned short state);
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200388 struct list *tl_queues = sched->tasklets;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100389 struct task *t;
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200390 uint8_t budget_mask = (1 << TL_CLASSES) - 1;
Willy Tarreau59153fe2020-06-24 10:17:29 +0200391 unsigned int done = 0;
392 unsigned int queue;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100393 unsigned short state;
394 void *ctx;
Willy Tarreau59153fe2020-06-24 10:17:29 +0200395
396 for (queue = 0; queue < TL_CLASSES;) {
397 sched->current_queue = queue;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100398
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200399 /* global.tune.sched.low-latency is set */
400 if (global.tune.options & GTUNE_SCHED_LOW_LATENCY) {
401 if (unlikely(sched->tl_class_mask & budget_mask & ((1 << queue) - 1))) {
402 /* a lower queue index has tasks again and still has a
403 * budget to run them. Let's switch to it now.
404 */
405 queue = (sched->tl_class_mask & 1) ? 0 :
406 (sched->tl_class_mask & 2) ? 1 : 2;
407 continue;
408 }
409
410 if (unlikely(queue > TL_URGENT &&
411 budget_mask & (1 << TL_URGENT) &&
412 !MT_LIST_ISEMPTY(&sched->shared_tasklet_list))) {
413 /* an urgent tasklet arrived from another thread */
414 break;
415 }
416
417 if (unlikely(queue > TL_NORMAL &&
418 budget_mask & (1 << TL_NORMAL) &&
419 ((sched->rqueue_size > 0) ||
420 (global_tasks_mask & tid_bit)))) {
421 /* a task was woken up by a bulk tasklet or another thread */
422 break;
423 }
424 }
425
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200426 if (LIST_ISEMPTY(&tl_queues[queue])) {
427 sched->tl_class_mask &= ~(1 << queue);
Willy Tarreau59153fe2020-06-24 10:17:29 +0200428 queue++;
429 continue;
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200430 }
431
Willy Tarreau59153fe2020-06-24 10:17:29 +0200432 if (!budgets[queue]) {
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200433 budget_mask &= ~(1 << queue);
Willy Tarreau59153fe2020-06-24 10:17:29 +0200434 queue++;
435 continue;
436 }
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200437
Willy Tarreau59153fe2020-06-24 10:17:29 +0200438 budgets[queue]--;
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200439 t = (struct task *)LIST_ELEM(tl_queues[queue].n, struct tasklet *, list);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200440 state = t->state & (TASK_SHARED_WQ|TASK_SELF_WAKING|TASK_KILLED);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100441
442 ti->flags &= ~TI_FL_STUCK; // this thread is still running
443 activity[tid].ctxsw++;
444 ctx = t->context;
445 process = t->process;
446 t->calls++;
Willy Tarreaud23d4132020-01-31 10:39:03 +0100447 sched->current = t;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100448
449 if (TASK_IS_TASKLET(t)) {
Willy Tarreau952c2642020-01-31 16:39:30 +0100450 state = _HA_ATOMIC_XCHG(&t->state, state);
451 __ha_barrier_atomic_store();
452 __tasklet_remove_from_tasklet_list((struct tasklet *)t);
Olivier Houchardc62d9ab2020-03-17 18:15:04 +0100453 process(t, ctx, state);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100454 done++;
Willy Tarreaud23d4132020-01-31 10:39:03 +0100455 sched->current = NULL;
456 __ha_barrier_store();
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100457 continue;
458 }
459
Willy Tarreau952c2642020-01-31 16:39:30 +0100460 state = _HA_ATOMIC_XCHG(&t->state, state | TASK_RUNNING);
461 __ha_barrier_atomic_store();
462 __tasklet_remove_from_tasklet_list((struct tasklet *)t);
463
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100464 /* OK then this is a regular task */
465
Willy Tarreaud99177f2020-06-30 11:48:48 +0200466 _HA_ATOMIC_SUB(&task_per_thread[tid].task_list_size, 1);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100467 if (unlikely(t->call_date)) {
468 uint64_t now_ns = now_mono_time();
469
470 t->lat_time += now_ns - t->call_date;
471 t->call_date = now_ns;
472 }
473
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100474 __ha_barrier_store();
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200475
476 /* Note for below: if TASK_KILLED arrived before we've read the state, we
477 * directly free the task. Otherwise it will be seen after processing and
478 * it's freed on the exit path.
479 */
480 if (likely(!(state & TASK_KILLED) && process == process_stream))
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100481 t = process_stream(t, ctx, state);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200482 else if (!(state & TASK_KILLED) && process != NULL)
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100483 t = process(t, ctx, state);
484 else {
Willy Tarreau273aea42020-07-17 14:37:51 +0200485 task_unlink_wq(t);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100486 __task_free(t);
487 sched->current = NULL;
488 __ha_barrier_store();
489 /* We don't want max_processed to be decremented if
490 * we're just freeing a destroyed task, we should only
491 * do so if we really ran a task.
492 */
493 continue;
494 }
495 sched->current = NULL;
496 __ha_barrier_store();
497 /* If there is a pending state we have to wake up the task
498 * immediately, else we defer it into wait queue
499 */
500 if (t != NULL) {
501 if (unlikely(t->call_date)) {
502 t->cpu_time += now_mono_time() - t->call_date;
503 t->call_date = 0;
504 }
505
506 state = _HA_ATOMIC_AND(&t->state, ~TASK_RUNNING);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200507 if (unlikely(state & TASK_KILLED)) {
Willy Tarreau273aea42020-07-17 14:37:51 +0200508 task_unlink_wq(t);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200509 __task_free(t);
510 }
511 else if (state & TASK_WOKEN_ANY)
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100512 task_wakeup(t, 0);
513 else
514 task_queue(t);
515 }
516 done++;
517 }
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200518 sched->current_queue = -1;
Willy Tarreau116ef222020-06-23 16:35:38 +0200519
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100520 return done;
521}
522
Willy Tarreau58b458d2008-06-29 22:40:23 +0200523/* The run queue is chronologically sorted in a tree. An insertion counter is
524 * used to assign a position to each task. This counter may be combined with
525 * other variables (eg: nice value) to set the final position in the tree. The
526 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100527 * tasks processed to 200 in any case, so that general latency remains low and
Willy Tarreaucde79022019-04-12 18:03:41 +0200528 * so that task positions have a chance to be considered. The function scans
529 * both the global and local run queues and picks the most urgent task between
530 * the two. We need to grab the global runqueue lock to touch it so it's taken
531 * on the very first access to the global run queue and is released as soon as
532 * it reaches the end.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200533 *
534 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200535 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100536void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200537{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200538 struct task_per_thread * const tt = sched;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200539 struct eb32sc_node *lrq; // next local run queue entry
540 struct eb32sc_node *grq; // next global run queue entry
Willy Tarreau964c9362007-01-07 00:38:00 +0100541 struct task *t;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200542 const unsigned int default_weights[TL_CLASSES] = {
543 [TL_URGENT] = 64, // ~50% of CPU bandwidth for I/O
544 [TL_NORMAL] = 48, // ~37% of CPU bandwidth for tasks
545 [TL_BULK] = 16, // ~13% of CPU bandwidth for self-wakers
546 };
547 unsigned int max[TL_CLASSES]; // max to be run per class
548 unsigned int max_total; // sum of max above
Olivier Houchard06910462019-10-11 16:35:01 +0200549 struct mt_list *tmp_list;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200550 unsigned int queue;
551 int max_processed;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100552
Willy Tarreaue6a02fa2019-05-22 07:06:44 +0200553 ti->flags &= ~TI_FL_STUCK; // this thread is still running
554
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200555 if (!thread_has_tasks()) {
556 activity[tid].empty_rq++;
557 return;
558 }
559
Willy Tarreau5c8be272020-06-19 12:17:55 +0200560 tasks_run_queue_cur = tasks_run_queue; /* keep a copy for reporting */
561 nb_tasks_cur = nb_tasks;
562 max_processed = global.tune.runqueue_depth;
563
564 if (likely(niced_tasks))
565 max_processed = (max_processed + 3) / 4;
566
Willy Tarreau5c8be272020-06-19 12:17:55 +0200567 not_done_yet:
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200568 max[TL_URGENT] = max[TL_NORMAL] = max[TL_BULK] = 0;
Willy Tarreaucde79022019-04-12 18:03:41 +0200569
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200570 /* urgent tasklets list gets a default weight of ~50% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200571 if ((tt->tl_class_mask & (1 << TL_URGENT)) ||
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200572 !MT_LIST_ISEMPTY(&tt->shared_tasklet_list))
573 max[TL_URGENT] = default_weights[TL_URGENT];
Willy Tarreaua62917b2020-01-30 18:37:28 +0100574
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200575 /* normal tasklets list gets a default weight of ~37% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200576 if ((tt->tl_class_mask & (1 << TL_NORMAL)) ||
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200577 (sched->rqueue_size > 0) || (global_tasks_mask & tid_bit))
578 max[TL_NORMAL] = default_weights[TL_NORMAL];
Willy Tarreaua62917b2020-01-30 18:37:28 +0100579
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200580 /* bulk tasklets list gets a default weight of ~13% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200581 if ((tt->tl_class_mask & (1 << TL_BULK)))
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200582 max[TL_BULK] = default_weights[TL_BULK];
583
584 /* Now compute a fair share of the weights. Total may slightly exceed
Willy Tarreau1553b662020-06-30 13:46:21 +0200585 * 100% due to rounding, this is not a problem. Note that while in
586 * theory the sum cannot be NULL as we cannot get there without tasklets
587 * to process, in practice it seldom happens when multiple writers
Willy Tarreaude4db172020-07-10 08:10:29 +0200588 * conflict and rollback on MT_LIST_TRY_ADDQ(shared_tasklet_list), causing
Willy Tarreau1553b662020-06-30 13:46:21 +0200589 * a first MT_LIST_ISEMPTY() to succeed for thread_has_task() and the
590 * one above to finally fail. This is extremely rare and not a problem.
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200591 */
592 max_total = max[TL_URGENT] + max[TL_NORMAL] + max[TL_BULK];
Willy Tarreau1553b662020-06-30 13:46:21 +0200593 if (!max_total)
594 return;
595
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200596 for (queue = 0; queue < TL_CLASSES; queue++)
597 max[queue] = ((unsigned)max_processed * max[queue] + max_total - 1) / max_total;
598
599 lrq = grq = NULL;
Christopher Faulet8a48f672017-11-14 10:38:36 +0100600
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200601 /* pick up to max[TL_NORMAL] regular tasks from prio-ordered run queues */
602 /* Note: the grq lock is always held when grq is not null */
603 while (tt->task_list_size < max[TL_NORMAL]) {
Willy Tarreaucde79022019-04-12 18:03:41 +0200604 if ((global_tasks_mask & tid_bit) && !grq) {
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200605#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200606 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
607 grq = eb32sc_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
608 if (unlikely(!grq)) {
609 grq = eb32sc_first(&rqueue, tid_bit);
610 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200611 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200612 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200613 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100614 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200615#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200616 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100617
Willy Tarreaucde79022019-04-12 18:03:41 +0200618 /* If a global task is available for this thread, it's in grq
619 * now and the global RQ is locked.
620 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200621
Willy Tarreaucde79022019-04-12 18:03:41 +0200622 if (!lrq) {
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200623 lrq = eb32sc_lookup_ge(&tt->rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreaucde79022019-04-12 18:03:41 +0200624 if (unlikely(!lrq))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200625 lrq = eb32sc_first(&tt->rqueue, tid_bit);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100626 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100627
Willy Tarreaucde79022019-04-12 18:03:41 +0200628 if (!lrq && !grq)
629 break;
630
631 if (likely(!grq || (lrq && (int)(lrq->key - grq->key) <= 0))) {
632 t = eb32sc_entry(lrq, struct task, rq);
633 lrq = eb32sc_next(lrq, tid_bit);
634 __task_unlink_rq(t);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200635 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200636#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200637 else {
638 t = eb32sc_entry(grq, struct task, rq);
639 grq = eb32sc_next(grq, tid_bit);
640 __task_unlink_rq(t);
641 if (unlikely(!grq)) {
642 grq = eb32sc_first(&rqueue, tid_bit);
643 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200644 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200645 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaucde79022019-04-12 18:03:41 +0200646 }
647 }
Emeric Brun01948972017-03-30 15:37:25 +0200648 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200649#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200650
Olivier Houchardff1e9f32019-09-20 17:18:35 +0200651 /* Make sure the entry doesn't appear to be in a list */
Olivier Houchard06910462019-10-11 16:35:01 +0200652 LIST_INIT(&((struct tasklet *)t)->list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200653 /* And add it to the local task list */
Willy Tarreaua62917b2020-01-30 18:37:28 +0100654 tasklet_insert_into_tasklet_list(&tt->tasklets[TL_NORMAL], (struct tasklet *)t);
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200655 tt->tl_class_mask |= 1 << TL_NORMAL;
Willy Tarreaud99177f2020-06-30 11:48:48 +0200656 _HA_ATOMIC_ADD(&tt->task_list_size, 1);
Willy Tarreaubc13bec2019-04-30 14:55:18 +0200657 activity[tid].tasksw++;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200658 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200659
660 /* release the rqueue lock */
661 if (grq) {
662 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
663 grq = NULL;
664 }
665
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200666 /* Merge the list of tasklets waken up by other threads to the
667 * main list.
668 */
669 tmp_list = MT_LIST_BEHEAD(&tt->shared_tasklet_list);
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200670 if (tmp_list) {
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200671 LIST_SPLICE_END_DETACHED(&tt->tasklets[TL_URGENT], (struct list *)tmp_list);
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200672 if (!LIST_ISEMPTY(&tt->tasklets[TL_URGENT]))
673 tt->tl_class_mask |= 1 << TL_URGENT;
674 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200675
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200676 /* execute tasklets in each queue */
Willy Tarreau59153fe2020-06-24 10:17:29 +0200677 max_processed -= run_tasks_from_lists(max);
Willy Tarreaua62917b2020-01-30 18:37:28 +0100678
Willy Tarreau5c8be272020-06-19 12:17:55 +0200679 /* some tasks may have woken other ones up */
Willy Tarreau0c0c85e2020-06-23 11:32:35 +0200680 if (max_processed > 0 && thread_has_tasks())
Willy Tarreau5c8be272020-06-19 12:17:55 +0200681 goto not_done_yet;
682
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200683 if (tt->tl_class_mask)
Willy Tarreaucde79022019-04-12 18:03:41 +0200684 activity[tid].long_rq++;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200685}
686
Willy Tarreau64e60122019-07-12 08:31:17 +0200687/* create a work list array for <nbthread> threads, using tasks made of
688 * function <fct>. The context passed to the function will be the pointer to
689 * the thread's work list, which will contain a copy of argument <arg>. The
690 * wake up reason will be TASK_WOKEN_OTHER. The pointer to the work_list array
691 * is returned on success, otherwise NULL on failure.
692 */
693struct work_list *work_list_create(int nbthread,
694 struct task *(*fct)(struct task *, void *, unsigned short),
695 void *arg)
696{
697 struct work_list *wl;
698 int i;
699
700 wl = calloc(nbthread, sizeof(*wl));
701 if (!wl)
702 goto fail;
703
704 for (i = 0; i < nbthread; i++) {
Olivier Houchard859dc802019-08-08 15:47:21 +0200705 MT_LIST_INIT(&wl[i].head);
Willy Tarreau64e60122019-07-12 08:31:17 +0200706 wl[i].task = task_new(1UL << i);
707 if (!wl[i].task)
708 goto fail;
709 wl[i].task->process = fct;
710 wl[i].task->context = &wl[i];
711 wl[i].arg = arg;
712 }
713 return wl;
714
715 fail:
716 work_list_destroy(wl, nbthread);
717 return NULL;
718}
719
720/* destroy work list <work> */
721void work_list_destroy(struct work_list *work, int nbthread)
722{
723 int t;
724
725 if (!work)
726 return;
727 for (t = 0; t < nbthread; t++)
728 task_destroy(work[t].task);
729 free(work);
730}
731
William Lallemand27f3fa52018-12-06 14:05:20 +0100732/*
733 * Delete every tasks before running the master polling loop
734 */
735void mworker_cleantasks()
736{
737 struct task *t;
738 int i;
William Lallemandb5823392018-12-06 15:14:37 +0100739 struct eb32_node *tmp_wq = NULL;
740 struct eb32sc_node *tmp_rq = NULL;
William Lallemand27f3fa52018-12-06 14:05:20 +0100741
742#ifdef USE_THREAD
743 /* cleanup the global run queue */
William Lallemandb5823392018-12-06 15:14:37 +0100744 tmp_rq = eb32sc_first(&rqueue, MAX_THREADS_MASK);
745 while (tmp_rq) {
746 t = eb32sc_entry(tmp_rq, struct task, rq);
747 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200748 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100749 }
750 /* cleanup the timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100751 tmp_wq = eb32_first(&timers);
752 while (tmp_wq) {
753 t = eb32_entry(tmp_wq, struct task, wq);
754 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200755 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100756 }
757#endif
758 /* clean the per thread run queue */
759 for (i = 0; i < global.nbthread; i++) {
William Lallemandb5823392018-12-06 15:14:37 +0100760 tmp_rq = eb32sc_first(&task_per_thread[i].rqueue, MAX_THREADS_MASK);
761 while (tmp_rq) {
762 t = eb32sc_entry(tmp_rq, struct task, rq);
763 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200764 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100765 }
766 /* cleanup the per thread timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100767 tmp_wq = eb32_first(&task_per_thread[i].timers);
768 while (tmp_wq) {
769 t = eb32_entry(tmp_wq, struct task, wq);
770 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200771 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100772 }
773 }
774}
775
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100776/* perform minimal intializations */
777static void init_task()
Willy Tarreau4726f532009-03-07 17:25:21 +0100778{
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200779 int i;
780
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200781#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200782 memset(&timers, 0, sizeof(timers));
Willy Tarreau4726f532009-03-07 17:25:21 +0100783 memset(&rqueue, 0, sizeof(rqueue));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200784#endif
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200785 memset(&task_per_thread, 0, sizeof(task_per_thread));
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200786 for (i = 0; i < MAX_THREADS; i++) {
Willy Tarreaua62917b2020-01-30 18:37:28 +0100787 LIST_INIT(&task_per_thread[i].tasklets[TL_URGENT]);
788 LIST_INIT(&task_per_thread[i].tasklets[TL_NORMAL]);
789 LIST_INIT(&task_per_thread[i].tasklets[TL_BULK]);
Olivier Houchard06910462019-10-11 16:35:01 +0200790 MT_LIST_INIT(&task_per_thread[i].shared_tasklet_list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200791 }
Willy Tarreau4726f532009-03-07 17:25:21 +0100792}
793
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200794/* config parser for global "tune.sched.low-latency", accepts "on" or "off" */
795static int cfg_parse_tune_sched_low_latency(char **args, int section_type, struct proxy *curpx,
796 struct proxy *defpx, const char *file, int line,
797 char **err)
798{
799 if (too_many_args(1, args, err, NULL))
800 return -1;
801
802 if (strcmp(args[1], "on") == 0)
803 global.tune.options |= GTUNE_SCHED_LOW_LATENCY;
804 else if (strcmp(args[1], "off") == 0)
805 global.tune.options &= ~GTUNE_SCHED_LOW_LATENCY;
806 else {
807 memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
808 return -1;
809 }
810 return 0;
811}
812
813/* config keyword parsers */
814static struct cfg_kw_list cfg_kws = {ILH, {
815 { CFG_GLOBAL, "tune.sched.low-latency", cfg_parse_tune_sched_low_latency },
816 { 0, NULL, NULL }
817}};
818
819INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100820INITCALL0(STG_PREPARE, init_task);
821
Willy Tarreaubaaee002006-06-26 02:48:02 +0200822/*
823 * Local variables:
824 * c-indent-level: 8
825 * c-basic-offset: 8
826 * End:
827 */