blob: 67633275dde2fa1d7ab7498edd16c888d424d500 [file] [log] [blame]
Willy Tarreaubaaee002006-06-26 02:48:02 +02001/*
2 * Task management functions.
3 *
Willy Tarreau4726f532009-03-07 17:25:21 +01004 * Copyright 2000-2009 Willy Tarreau <w@1wt.eu>
Willy Tarreaubaaee002006-06-26 02:48:02 +02005 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
10 *
11 */
12
Willy Tarreau87bed622009-03-08 22:25:28 +010013#include <string.h>
14
Willy Tarreaub2551052020-06-09 09:07:15 +020015#include <import/eb32sctree.h>
16#include <import/eb32tree.h>
17
Willy Tarreau4c7e4b72020-05-27 12:58:42 +020018#include <haproxy/api.h>
Willy Tarreaue7723bd2020-06-24 11:11:02 +020019#include <haproxy/cfgparse.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020020#include <haproxy/fd.h>
21#include <haproxy/freq_ctr.h>
Willy Tarreau853b2972020-05-27 18:01:47 +020022#include <haproxy/list.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020023#include <haproxy/pool.h>
Willy Tarreaudfd3de82020-06-04 23:46:14 +020024#include <haproxy/stream.h>
Willy Tarreaucea0e1b2020-06-04 17:25:40 +020025#include <haproxy/task.h>
Willy Tarreau92b4f132020-06-01 11:05:15 +020026#include <haproxy/time.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020027#include <haproxy/tools.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020028
Willy Tarreaubaaee002006-06-26 02:48:02 +020029
Willy Tarreau8ceae722018-11-26 11:58:30 +010030DECLARE_POOL(pool_head_task, "task", sizeof(struct task));
31DECLARE_POOL(pool_head_tasklet, "tasklet", sizeof(struct tasklet));
Willy Tarreau96bcfd72007-04-29 10:41:56 +020032
Thierry FOURNIERd6975962017-07-12 14:31:10 +020033/* This is the memory pool containing all the signal structs. These
Joseph Herlantcf92b6d2018-11-15 14:19:23 -080034 * struct are used to store each required signal between two tasks.
Thierry FOURNIERd6975962017-07-12 14:31:10 +020035 */
Willy Tarreau8ceae722018-11-26 11:58:30 +010036DECLARE_POOL(pool_head_notification, "notification", sizeof(struct notification));
Thierry FOURNIERd6975962017-07-12 14:31:10 +020037
Olivier Houchardeba0c0b2018-07-26 18:53:28 +020038volatile unsigned long global_tasks_mask = 0; /* Mask of threads with tasks in the global runqueue */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010039unsigned int niced_tasks = 0; /* number of niced tasks in the run queue */
Emeric Brun01948972017-03-30 15:37:25 +020040
Willy Tarreaud022e9c2019-09-24 08:25:15 +020041THREAD_LOCAL struct task_per_thread *sched = &task_per_thread[0]; /* scheduler context for the current thread */
Willy Tarreau6d1222c2017-11-26 10:08:06 +010042
Willy Tarreau86abe442018-11-25 20:12:18 +010043__decl_aligned_spinlock(rq_lock); /* spin lock related to run queue */
Willy Tarreauef28dc12019-05-28 18:48:07 +020044__decl_aligned_rwlock(wq_lock); /* RW lock related to the wait queue */
Willy Tarreau964c9362007-01-07 00:38:00 +010045
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020046#ifdef USE_THREAD
Willy Tarreauc6ba9a02021-02-20 12:49:54 +010047struct eb_root timers; /* sorted timers tree, global, accessed under wq_lock */
48struct eb_root rqueue; /* tree constituting the global run queue, accessed under rq_lock */
Willy Tarreau45499c52021-02-25 07:51:18 +010049unsigned int grq_total; /* total number of entries in the global run queue, atomic */
Willy Tarreauc6ba9a02021-02-20 12:49:54 +010050static unsigned int global_rqueue_ticks; /* insertion count in the grq, use rq_lock */
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020051#endif
Willy Tarreaub20aa9e2018-10-15 14:52:21 +020052
Willy Tarreau8d8747a2018-10-15 16:12:48 +020053
54struct task_per_thread task_per_thread[MAX_THREADS];
Willy Tarreau9789f7b2008-06-24 08:17:16 +020055
Willy Tarreaueb8c2c62020-06-30 11:48:48 +020056
57/* Flags the task <t> for immediate destruction and puts it into its first
58 * thread's shared tasklet list if not yet queued/running. This will bypass
59 * the priority scheduling and make the task show up as fast as possible in
60 * the other thread's queue. Note that this operation isn't idempotent and is
61 * not supposed to be run on the same task from multiple threads at once. It's
62 * the caller's responsibility to make sure it is the only one able to kill the
63 * task.
64 */
65void task_kill(struct task *t)
66{
Willy Tarreau144f84a2021-03-02 16:09:26 +010067 unsigned int state = t->state;
Willy Tarreaueb8c2c62020-06-30 11:48:48 +020068 unsigned int thr;
69
70 BUG_ON(state & TASK_KILLED);
71
72 while (1) {
73 while (state & (TASK_RUNNING | TASK_QUEUED)) {
74 /* task already in the queue and about to be executed,
75 * or even currently running. Just add the flag and be
76 * done with it, the process loop will detect it and kill
77 * it. The CAS will fail if we arrive too late.
78 */
79 if (_HA_ATOMIC_CAS(&t->state, &state, state | TASK_KILLED))
80 return;
81 }
82
83 /* We'll have to wake it up, but we must also secure it so that
84 * it doesn't vanish under us. TASK_QUEUED guarantees nobody will
85 * add past us.
86 */
87 if (_HA_ATOMIC_CAS(&t->state, &state, state | TASK_QUEUED | TASK_KILLED)) {
88 /* Bypass the tree and go directly into the shared tasklet list.
89 * Note: that's a task so it must be accounted for as such. Pick
90 * the task's first thread for the job.
91 */
92 thr = my_ffsl(t->thread_mask) - 1;
Willy Tarreau54d31172020-07-02 14:14:00 +020093
94 /* Beware: tasks that have never run don't have their ->list empty yet! */
Willy Tarreau2b718102021-04-21 07:32:39 +020095 MT_LIST_APPEND(&task_per_thread[thr].shared_tasklet_list,
Willy Tarreau4f589262020-07-02 17:17:42 +020096 (struct mt_list *)&((struct tasklet *)t)->list);
Willy Tarreau4781b152021-04-06 13:53:36 +020097 _HA_ATOMIC_INC(&task_per_thread[thr].rq_total);
98 _HA_ATOMIC_INC(&task_per_thread[thr].tasks_in_list);
Willy Tarreau54d31172020-07-02 14:14:00 +020099 if (sleeping_thread_mask & (1UL << thr)) {
100 _HA_ATOMIC_AND(&sleeping_thread_mask, ~(1UL << thr));
101 wake_thread(thr);
Willy Tarreaueb8c2c62020-06-30 11:48:48 +0200102 }
Willy Tarreau54d31172020-07-02 14:14:00 +0200103 return;
Willy Tarreaueb8c2c62020-06-30 11:48:48 +0200104 }
105 }
106}
107
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100108/* Do not call this one, please use tasklet_wakeup_on() instead, as this one is
109 * the slow path of tasklet_wakeup_on() which performs some preliminary checks
110 * and sets TASK_IN_LIST before calling this one. A negative <thr> designates
111 * the current thread.
112 */
113void __tasklet_wakeup_on(struct tasklet *tl, int thr)
114{
115 if (likely(thr < 0)) {
116 /* this tasklet runs on the caller thread */
Willy Tarreau826fa872021-02-26 10:13:40 +0100117 if (tl->state & TASK_HEAVY) {
Willy Tarreau2b718102021-04-21 07:32:39 +0200118 LIST_APPEND(&sched->tasklets[TL_HEAVY], &tl->list);
Willy Tarreau826fa872021-02-26 10:13:40 +0100119 sched->tl_class_mask |= 1 << TL_HEAVY;
120 }
121 else if (tl->state & TASK_SELF_WAKING) {
Willy Tarreau2b718102021-04-21 07:32:39 +0200122 LIST_APPEND(&sched->tasklets[TL_BULK], &tl->list);
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100123 sched->tl_class_mask |= 1 << TL_BULK;
124 }
125 else if ((struct task *)tl == sched->current) {
126 _HA_ATOMIC_OR(&tl->state, TASK_SELF_WAKING);
Willy Tarreau2b718102021-04-21 07:32:39 +0200127 LIST_APPEND(&sched->tasklets[TL_BULK], &tl->list);
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100128 sched->tl_class_mask |= 1 << TL_BULK;
129 }
130 else if (sched->current_queue < 0) {
Willy Tarreau2b718102021-04-21 07:32:39 +0200131 LIST_APPEND(&sched->tasklets[TL_URGENT], &tl->list);
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100132 sched->tl_class_mask |= 1 << TL_URGENT;
133 }
134 else {
Willy Tarreau2b718102021-04-21 07:32:39 +0200135 LIST_APPEND(&sched->tasklets[sched->current_queue], &tl->list);
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100136 sched->tl_class_mask |= 1 << sched->current_queue;
137 }
Willy Tarreau4781b152021-04-06 13:53:36 +0200138 _HA_ATOMIC_INC(&sched->rq_total);
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100139 } else {
140 /* this tasklet runs on a specific thread. */
Willy Tarreau2b718102021-04-21 07:32:39 +0200141 MT_LIST_APPEND(&task_per_thread[thr].shared_tasklet_list, (struct mt_list *)&tl->list);
Willy Tarreau4781b152021-04-06 13:53:36 +0200142 _HA_ATOMIC_INC(&task_per_thread[thr].rq_total);
Willy Tarreau9c6dbf02021-02-24 17:51:38 +0100143 if (sleeping_thread_mask & (1UL << thr)) {
144 _HA_ATOMIC_AND(&sleeping_thread_mask, ~(1UL << thr));
145 wake_thread(thr);
146 }
147 }
148}
149
Willy Tarreau4726f532009-03-07 17:25:21 +0100150/* Puts the task <t> in run queue at a position depending on t->nice. <t> is
151 * returned. The nice value assigns boosts in 32th of the run queue size. A
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100152 * nice value of -1024 sets the task to -tasks_run_queue*32, while a nice value
153 * of 1024 sets the task to tasks_run_queue*32. The state flags are cleared, so
154 * the caller will have to set its flags after this call.
Willy Tarreau4726f532009-03-07 17:25:21 +0100155 * The task must not already be in the run queue. If unsure, use the safer
156 * task_wakeup() function.
Willy Tarreau91e99932008-06-30 07:51:00 +0200157 */
Willy Tarreau018564e2021-02-24 16:41:11 +0100158void __task_wakeup(struct task *t)
Willy Tarreaue33aece2007-04-30 13:15:14 +0200159{
Willy Tarreau018564e2021-02-24 16:41:11 +0100160 struct eb_root *root = &sched->rqueue;
161
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200162#ifdef USE_THREAD
Willy Tarreau018564e2021-02-24 16:41:11 +0100163 if (t->thread_mask != tid_bit && global.nbthread != 1) {
164 root = &rqueue;
165
Willy Tarreau4781b152021-04-06 13:53:36 +0200166 _HA_ATOMIC_INC(&grq_total);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200167 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreau9c7b8082021-02-24 15:10:07 +0100168
Olivier Houchardde82aea2019-04-17 19:10:22 +0200169 global_tasks_mask |= t->thread_mask;
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100170 t->rq.key = ++global_rqueue_ticks;
Olivier Houcharded1a6a02019-04-18 14:12:51 +0200171 __ha_barrier_store();
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100172 } else
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200173#endif
Willy Tarreau9c7b8082021-02-24 15:10:07 +0100174 {
Willy Tarreau4781b152021-04-06 13:53:36 +0200175 _HA_ATOMIC_INC(&sched->rq_total);
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100176 t->rq.key = ++sched->rqueue_ticks;
Willy Tarreau9c7b8082021-02-24 15:10:07 +0100177 }
Willy Tarreau91e99932008-06-30 07:51:00 +0200178
179 if (likely(t->nice)) {
180 int offset;
181
Willy Tarreau4781b152021-04-06 13:53:36 +0200182 _HA_ATOMIC_INC(&niced_tasks);
Willy Tarreau2d1fd0a2019-04-15 09:18:31 +0200183 offset = t->nice * (int)global.tune.runqueue_depth;
Willy Tarreau4726f532009-03-07 17:25:21 +0100184 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +0200185 }
186
Willy Tarreaud9add3a2019-04-25 08:57:41 +0200187 if (task_profiling_mask & tid_bit)
Willy Tarreau9efd7452018-05-31 14:48:54 +0200188 t->call_date = now_mono_time();
189
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200190 eb32sc_insert(root, &t->rq, t->thread_mask);
Willy Tarreau018564e2021-02-24 16:41:11 +0100191
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200192#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200193 if (root == &rqueue) {
Olivier Houchard4c2832852019-03-08 18:48:47 +0100194 _HA_ATOMIC_OR(&t->state, TASK_GLOBAL);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200195 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreau2c41d772021-02-24 16:13:03 +0100196
Willy Tarreaueeffb3d2021-02-24 16:44:51 +0100197 /* If all threads that are supposed to handle this task are sleeping,
198 * wake one.
199 */
200 if ((((t->thread_mask & all_threads_mask) & sleeping_thread_mask) ==
201 (t->thread_mask & all_threads_mask))) {
202 unsigned long m = (t->thread_mask & all_threads_mask) &~ tid_bit;
Olivier Houchard1b327902019-03-15 00:23:10 +0100203
Willy Tarreaueeffb3d2021-02-24 16:44:51 +0100204 m = (m & (m - 1)) ^ m; // keep lowest bit set
205 _HA_ATOMIC_AND(&sleeping_thread_mask, ~m);
206 wake_thread(my_ffsl(m) - 1);
207 }
Olivier Houchard1b327902019-03-15 00:23:10 +0100208 }
Willy Tarreau85d9b842018-07-27 17:14:41 +0200209#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200210 return;
Willy Tarreaue33aece2007-04-30 13:15:14 +0200211}
Willy Tarreaud825eef2007-05-12 22:35:00 +0200212
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200213/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100214 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200215 *
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200216 * Inserts a task into wait queue <wq> at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +0100217 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100218 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100219 * Last, tasks must not be queued further than the end of the tree, which is
220 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100221 *
222 * This function should not be used directly, it is meant to be called by the
223 * inline version of task_queue() which performs a few cheap preliminary tests
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200224 * before deciding to call __task_queue(). Moreover this function doesn't care
225 * at all about locking so the caller must be careful when deciding whether to
226 * lock or not around this call.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200227 */
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200228void __task_queue(struct task *task, struct eb_root *wq)
Willy Tarreau964c9362007-01-07 00:38:00 +0100229{
Willy Tarreaue5d79bc2020-07-22 14:29:42 +0200230#ifdef USE_THREAD
231 BUG_ON((wq == &timers && !(task->state & TASK_SHARED_WQ)) ||
232 (wq == &sched->timers && (task->state & TASK_SHARED_WQ)) ||
233 (wq != &timers && wq != &sched->timers));
234#endif
235
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100236 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +0100237 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +0100238
239 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100240 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200241#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100242 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +0200243 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +0100244 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200245#endif
Willy Tarreauce44f122008-07-05 18:16:19 +0200246
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200247 eb32_insert(wq, &task->wq);
Willy Tarreau964c9362007-01-07 00:38:00 +0100248}
249
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200250/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200251 * Extract all expired timers from the timer queue, and wakes up all
Willy Tarreauc49ba522019-12-11 08:12:23 +0100252 * associated tasks.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200253 */
Willy Tarreauc49ba522019-12-11 08:12:23 +0100254void wake_expired_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200255{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200256 struct task_per_thread * const tt = sched; // thread's tasks
Willy Tarreau3cfaa8d2020-10-16 09:26:22 +0200257 int max_processed = global.tune.runqueue_depth;
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200258 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200259 struct eb32_node *eb;
Willy Tarreauaf613e82020-06-05 08:40:51 +0200260 __decl_thread(int key);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200261
Willy Tarreau3cfaa8d2020-10-16 09:26:22 +0200262 while (max_processed-- > 0) {
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200263 lookup_next_local:
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200264 eb = eb32_lookup_ge(&tt->timers, now_ms - TIMER_LOOK_BACK);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200265 if (!eb) {
266 /* we might have reached the end of the tree, typically because
267 * <now_ms> is in the first half and we're first scanning the last
268 * half. Let's loop back to the beginning of the tree now.
269 */
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200270 eb = eb32_first(&tt->timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200271 if (likely(!eb))
272 break;
273 }
274
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200275 /* It is possible that this task was left at an earlier place in the
276 * tree because a recent call to task_queue() has not moved it. This
277 * happens when the new expiration date is later than the old one.
278 * Since it is very unlikely that we reach a timeout anyway, it's a
279 * lot cheaper to proceed like this because we almost never update
280 * the tree. We may also find disabled expiration dates there. Since
281 * we have detached the task from the tree, we simply call task_queue
282 * to take care of this. Note that we might occasionally requeue it at
283 * the same place, before <eb>, so we have to check if this happens,
284 * and adjust <eb>, otherwise we may skip it which is not what we want.
285 * We may also not requeue the task (and not point eb at it) if its
Willy Tarreau77015ab2020-06-19 11:50:27 +0200286 * expiration time is not set. We also make sure we leave the real
287 * expiration date for the next task in the queue so that when calling
288 * next_timer_expiry() we're guaranteed to see the next real date and
289 * not the next apparent date. This is in order to avoid useless
290 * wakeups.
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200291 */
Willy Tarreau77015ab2020-06-19 11:50:27 +0200292
293 task = eb32_entry(eb, struct task, wq);
294 if (tick_is_expired(task->expire, now_ms)) {
295 /* expired task, wake it up */
296 __task_unlink_wq(task);
297 task_wakeup(task, TASK_WOKEN_TIMER);
298 }
299 else if (task->expire != eb->key) {
300 /* task is not expired but its key doesn't match so let's
301 * update it and skip to next apparently expired task.
302 */
303 __task_unlink_wq(task);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200304 if (tick_isset(task->expire))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200305 __task_queue(task, &tt->timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200306 }
Willy Tarreau77015ab2020-06-19 11:50:27 +0200307 else {
308 /* task not expired and correctly placed */
309 break;
310 }
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200311 }
312
313#ifdef USE_THREAD
Willy Tarreau1e928c02019-05-28 18:57:25 +0200314 if (eb_is_empty(&timers))
315 goto leave;
316
317 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
318 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
319 if (!eb) {
320 eb = eb32_first(&timers);
321 if (likely(!eb)) {
322 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
323 goto leave;
324 }
325 }
326 key = eb->key;
Willy Tarreau1e928c02019-05-28 18:57:25 +0200327
Willy Tarreaud48ed662020-10-16 09:31:41 +0200328 if (tick_is_lt(now_ms, key)) {
329 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau1e928c02019-05-28 18:57:25 +0200330 goto leave;
Willy Tarreaud48ed662020-10-16 09:31:41 +0200331 }
Willy Tarreau1e928c02019-05-28 18:57:25 +0200332
333 /* There's really something of interest here, let's visit the queue */
334
Willy Tarreaud48ed662020-10-16 09:31:41 +0200335 if (HA_RWLOCK_TRYRDTOSK(TASK_WQ_LOCK, &wq_lock)) {
336 /* if we failed to grab the lock it means another thread is
337 * already doing the same here, so let it do the job.
338 */
339 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
340 goto leave;
341 }
342
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200343 while (1) {
Emeric Brunc60def82017-09-27 14:59:38 +0200344 lookup_next:
Willy Tarreau3cfaa8d2020-10-16 09:26:22 +0200345 if (max_processed-- <= 0)
346 break;
Emeric Brun01948972017-03-30 15:37:25 +0200347 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200348 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100349 /* we might have reached the end of the tree, typically because
350 * <now_ms> is in the first half and we're first scanning the last
351 * half. Let's loop back to the beginning of the tree now.
352 */
353 eb = eb32_first(&timers);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100354 if (likely(!eb))
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100355 break;
356 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200357
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100358 task = eb32_entry(eb, struct task, wq);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200359 if (tick_is_expired(task->expire, now_ms)) {
360 /* expired task, wake it up */
Willy Tarreaud48ed662020-10-16 09:31:41 +0200361 HA_RWLOCK_SKTOWR(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200362 __task_unlink_wq(task);
Willy Tarreaud48ed662020-10-16 09:31:41 +0200363 HA_RWLOCK_WRTOSK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200364 task_wakeup(task, TASK_WOKEN_TIMER);
365 }
366 else if (task->expire != eb->key) {
367 /* task is not expired but its key doesn't match so let's
368 * update it and skip to next apparently expired task.
369 */
Willy Tarreaud48ed662020-10-16 09:31:41 +0200370 HA_RWLOCK_SKTOWR(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200371 __task_unlink_wq(task);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100372 if (tick_isset(task->expire))
Willy Tarreau783afbe2020-07-22 14:12:45 +0200373 __task_queue(task, &timers);
Willy Tarreaud48ed662020-10-16 09:31:41 +0200374 HA_RWLOCK_WRTOSK(TASK_WQ_LOCK, &wq_lock);
Emeric Brunc60def82017-09-27 14:59:38 +0200375 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200376 }
Willy Tarreau77015ab2020-06-19 11:50:27 +0200377 else {
378 /* task not expired and correctly placed */
379 break;
380 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100381 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200382
Willy Tarreaud48ed662020-10-16 09:31:41 +0200383 HA_RWLOCK_SKUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200384#endif
Willy Tarreau1e928c02019-05-28 18:57:25 +0200385leave:
Willy Tarreauc49ba522019-12-11 08:12:23 +0100386 return;
387}
388
389/* Checks the next timer for the current thread by looking into its own timer
390 * list and the global one. It may return TICK_ETERNITY if no timer is present.
Ilya Shipitsin856aabc2020-04-16 23:51:34 +0500391 * Note that the next timer might very well be slightly in the past.
Willy Tarreauc49ba522019-12-11 08:12:23 +0100392 */
393int next_timer_expiry()
394{
395 struct task_per_thread * const tt = sched; // thread's tasks
396 struct eb32_node *eb;
397 int ret = TICK_ETERNITY;
Willy Tarreau6ce02322020-08-21 05:48:34 +0200398 __decl_thread(int key = TICK_ETERNITY);
Willy Tarreauc49ba522019-12-11 08:12:23 +0100399
400 /* first check in the thread-local timers */
401 eb = eb32_lookup_ge(&tt->timers, now_ms - TIMER_LOOK_BACK);
402 if (!eb) {
403 /* we might have reached the end of the tree, typically because
404 * <now_ms> is in the first half and we're first scanning the last
405 * half. Let's loop back to the beginning of the tree now.
406 */
407 eb = eb32_first(&tt->timers);
408 }
409
410 if (eb)
411 ret = eb->key;
412
413#ifdef USE_THREAD
414 if (!eb_is_empty(&timers)) {
415 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
416 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
417 if (!eb)
418 eb = eb32_first(&timers);
419 if (eb)
420 key = eb->key;
421 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
422 if (eb)
423 ret = tick_first(ret, key);
424 }
425#endif
Willy Tarreaub992ba12017-11-05 19:09:27 +0100426 return ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200427}
428
Willy Tarreau59153fe2020-06-24 10:17:29 +0200429/* Walks over tasklet lists sched->tasklets[0..TL_CLASSES-1] and run at most
430 * budget[TL_*] of them. Returns the number of entries effectively processed
431 * (tasks and tasklets merged). The count of tasks in the list for the current
432 * thread is adjusted.
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100433 */
Willy Tarreau59153fe2020-06-24 10:17:29 +0200434unsigned int run_tasks_from_lists(unsigned int budgets[])
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100435{
Willy Tarreau144f84a2021-03-02 16:09:26 +0100436 struct task *(*process)(struct task *t, void *ctx, unsigned int state);
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200437 struct list *tl_queues = sched->tasklets;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100438 struct task *t;
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200439 uint8_t budget_mask = (1 << TL_CLASSES) - 1;
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100440 struct sched_activity *profile_entry = NULL;
Willy Tarreau59153fe2020-06-24 10:17:29 +0200441 unsigned int done = 0;
442 unsigned int queue;
Willy Tarreau144f84a2021-03-02 16:09:26 +0100443 unsigned int state;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100444 void *ctx;
Willy Tarreau59153fe2020-06-24 10:17:29 +0200445
446 for (queue = 0; queue < TL_CLASSES;) {
447 sched->current_queue = queue;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100448
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200449 /* global.tune.sched.low-latency is set */
450 if (global.tune.options & GTUNE_SCHED_LOW_LATENCY) {
451 if (unlikely(sched->tl_class_mask & budget_mask & ((1 << queue) - 1))) {
452 /* a lower queue index has tasks again and still has a
453 * budget to run them. Let's switch to it now.
454 */
455 queue = (sched->tl_class_mask & 1) ? 0 :
456 (sched->tl_class_mask & 2) ? 1 : 2;
457 continue;
458 }
459
460 if (unlikely(queue > TL_URGENT &&
461 budget_mask & (1 << TL_URGENT) &&
462 !MT_LIST_ISEMPTY(&sched->shared_tasklet_list))) {
463 /* an urgent tasklet arrived from another thread */
464 break;
465 }
466
467 if (unlikely(queue > TL_NORMAL &&
468 budget_mask & (1 << TL_NORMAL) &&
Willy Tarreau2c41d772021-02-24 16:13:03 +0100469 (!eb_is_empty(&sched->rqueue) ||
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200470 (global_tasks_mask & tid_bit)))) {
471 /* a task was woken up by a bulk tasklet or another thread */
472 break;
473 }
474 }
475
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200476 if (LIST_ISEMPTY(&tl_queues[queue])) {
477 sched->tl_class_mask &= ~(1 << queue);
Willy Tarreau59153fe2020-06-24 10:17:29 +0200478 queue++;
479 continue;
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200480 }
481
Willy Tarreau59153fe2020-06-24 10:17:29 +0200482 if (!budgets[queue]) {
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200483 budget_mask &= ~(1 << queue);
Willy Tarreau59153fe2020-06-24 10:17:29 +0200484 queue++;
485 continue;
486 }
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200487
Willy Tarreau59153fe2020-06-24 10:17:29 +0200488 budgets[queue]--;
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200489 t = (struct task *)LIST_ELEM(tl_queues[queue].n, struct tasklet *, list);
Willy Tarreau6fa8bcd2021-03-02 16:26:05 +0100490 state = t->state & (TASK_SHARED_WQ|TASK_SELF_WAKING|TASK_HEAVY|TASK_F_TASKLET|TASK_KILLED|TASK_F_USR1);
Willy Tarreau74dea8c2021-02-26 00:25:51 +0100491
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100492 ti->flags &= ~TI_FL_STUCK; // this thread is still running
493 activity[tid].ctxsw++;
494 ctx = t->context;
495 process = t->process;
496 t->calls++;
Willy Tarreaud23d4132020-01-31 10:39:03 +0100497 sched->current = t;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100498
Willy Tarreau4781b152021-04-06 13:53:36 +0200499 _HA_ATOMIC_DEC(&sched->rq_total);
Willy Tarreau2da4c312020-11-30 14:52:11 +0100500
Willy Tarreaudb4e2382021-03-02 15:54:11 +0100501 if (state & TASK_F_TASKLET) {
Willy Tarreau2a54ffb2021-02-25 09:32:58 +0100502 uint64_t before = 0;
503
Willy Tarreau4d6c5942020-11-30 14:58:53 +0100504 LIST_DEL_INIT(&((struct tasklet *)t)->list);
505 __ha_barrier_store();
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100506
507 if (unlikely(task_profiling_mask & tid_bit)) {
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100508 profile_entry = sched_activity_entry(sched_activity, t->process);
509 before = now_mono_time();
Willy Tarreaub2285de2021-02-25 08:39:07 +0100510#ifdef DEBUG_TASK
Willy Tarreau2a54ffb2021-02-25 09:32:58 +0100511 if (((struct tasklet *)t)->call_date) {
512 HA_ATOMIC_ADD(&profile_entry->lat_time, before - ((struct tasklet *)t)->call_date);
513 ((struct tasklet *)t)->call_date = 0;
514 }
Willy Tarreaub2285de2021-02-25 08:39:07 +0100515#endif
Willy Tarreau2a54ffb2021-02-25 09:32:58 +0100516 }
517
518 state = _HA_ATOMIC_XCHG(&t->state, state);
519 __ha_barrier_atomic_store();
520
521 process(t, ctx, state);
522
523 if (unlikely(task_profiling_mask & tid_bit)) {
Willy Tarreau4781b152021-04-06 13:53:36 +0200524 HA_ATOMIC_INC(&profile_entry->calls);
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100525 HA_ATOMIC_ADD(&profile_entry->cpu_time, now_mono_time() - before);
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100526 }
Willy Tarreau2a54ffb2021-02-25 09:32:58 +0100527
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100528 done++;
Willy Tarreaud23d4132020-01-31 10:39:03 +0100529 sched->current = NULL;
530 __ha_barrier_store();
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100531 continue;
532 }
533
Willy Tarreau4d6c5942020-11-30 14:58:53 +0100534 LIST_DEL_INIT(&((struct tasklet *)t)->list);
535 __ha_barrier_store();
Willy Tarreau6fa8bcd2021-03-02 16:26:05 +0100536 state = _HA_ATOMIC_XCHG(&t->state, state|TASK_RUNNING|TASK_F_USR1);
Willy Tarreau952c2642020-01-31 16:39:30 +0100537 __ha_barrier_atomic_store();
Willy Tarreau952c2642020-01-31 16:39:30 +0100538
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100539 /* OK then this is a regular task */
540
Willy Tarreau4781b152021-04-06 13:53:36 +0200541 _HA_ATOMIC_DEC(&task_per_thread[tid].tasks_in_list);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100542 if (unlikely(t->call_date)) {
543 uint64_t now_ns = now_mono_time();
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100544 uint64_t lat = now_ns - t->call_date;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100545
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100546 t->lat_time += lat;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100547 t->call_date = now_ns;
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100548 profile_entry = sched_activity_entry(sched_activity, t->process);
549 HA_ATOMIC_ADD(&profile_entry->lat_time, lat);
Willy Tarreau4781b152021-04-06 13:53:36 +0200550 HA_ATOMIC_INC(&profile_entry->calls);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100551 }
552
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100553 __ha_barrier_store();
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200554
555 /* Note for below: if TASK_KILLED arrived before we've read the state, we
556 * directly free the task. Otherwise it will be seen after processing and
557 * it's freed on the exit path.
558 */
559 if (likely(!(state & TASK_KILLED) && process == process_stream))
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100560 t = process_stream(t, ctx, state);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200561 else if (!(state & TASK_KILLED) && process != NULL)
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100562 t = process(t, ctx, state);
563 else {
Willy Tarreau273aea42020-07-17 14:37:51 +0200564 task_unlink_wq(t);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100565 __task_free(t);
566 sched->current = NULL;
567 __ha_barrier_store();
568 /* We don't want max_processed to be decremented if
569 * we're just freeing a destroyed task, we should only
570 * do so if we really ran a task.
571 */
572 continue;
573 }
574 sched->current = NULL;
575 __ha_barrier_store();
576 /* If there is a pending state we have to wake up the task
577 * immediately, else we defer it into wait queue
578 */
579 if (t != NULL) {
580 if (unlikely(t->call_date)) {
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100581 uint64_t cpu = now_mono_time() - t->call_date;
582
583 t->cpu_time += cpu;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100584 t->call_date = 0;
Willy Tarreau4e2282f2021-01-29 00:07:40 +0100585 HA_ATOMIC_ADD(&profile_entry->cpu_time, cpu);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100586 }
587
Willy Tarreau1db42732021-04-06 11:44:07 +0200588 state = _HA_ATOMIC_AND_FETCH(&t->state, ~TASK_RUNNING);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200589 if (unlikely(state & TASK_KILLED)) {
Willy Tarreau273aea42020-07-17 14:37:51 +0200590 task_unlink_wq(t);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200591 __task_free(t);
592 }
593 else if (state & TASK_WOKEN_ANY)
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100594 task_wakeup(t, 0);
595 else
596 task_queue(t);
597 }
598 done++;
599 }
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200600 sched->current_queue = -1;
Willy Tarreau116ef222020-06-23 16:35:38 +0200601
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100602 return done;
603}
604
Willy Tarreau58b458d2008-06-29 22:40:23 +0200605/* The run queue is chronologically sorted in a tree. An insertion counter is
606 * used to assign a position to each task. This counter may be combined with
607 * other variables (eg: nice value) to set the final position in the tree. The
608 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100609 * tasks processed to 200 in any case, so that general latency remains low and
Willy Tarreaucde79022019-04-12 18:03:41 +0200610 * so that task positions have a chance to be considered. The function scans
611 * both the global and local run queues and picks the most urgent task between
612 * the two. We need to grab the global runqueue lock to touch it so it's taken
613 * on the very first access to the global run queue and is released as soon as
614 * it reaches the end.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200615 *
616 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200617 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100618void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200619{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200620 struct task_per_thread * const tt = sched;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200621 struct eb32sc_node *lrq; // next local run queue entry
622 struct eb32sc_node *grq; // next global run queue entry
Willy Tarreau964c9362007-01-07 00:38:00 +0100623 struct task *t;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200624 const unsigned int default_weights[TL_CLASSES] = {
625 [TL_URGENT] = 64, // ~50% of CPU bandwidth for I/O
626 [TL_NORMAL] = 48, // ~37% of CPU bandwidth for tasks
627 [TL_BULK] = 16, // ~13% of CPU bandwidth for self-wakers
Willy Tarreau401135c2021-02-26 09:16:22 +0100628 [TL_HEAVY] = 1, // never more than 1 heavy task at once
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200629 };
630 unsigned int max[TL_CLASSES]; // max to be run per class
631 unsigned int max_total; // sum of max above
Olivier Houchard06910462019-10-11 16:35:01 +0200632 struct mt_list *tmp_list;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200633 unsigned int queue;
634 int max_processed;
Willy Tarreaue7923c12021-02-25 07:09:08 +0100635 int lpicked, gpicked;
Willy Tarreau76390da2021-02-26 10:18:11 +0100636 int heavy_queued = 0;
Willy Tarreauc309dbd2020-11-30 15:39:00 +0100637 int budget;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100638
Willy Tarreaue6a02fa2019-05-22 07:06:44 +0200639 ti->flags &= ~TI_FL_STUCK; // this thread is still running
640
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200641 if (!thread_has_tasks()) {
642 activity[tid].empty_rq++;
643 return;
644 }
645
Willy Tarreau5c8be272020-06-19 12:17:55 +0200646 max_processed = global.tune.runqueue_depth;
647
648 if (likely(niced_tasks))
649 max_processed = (max_processed + 3) / 4;
650
Willy Tarreau1691ba32021-03-10 09:26:24 +0100651 if (max_processed < sched->rq_total && sched->rq_total <= 2*max_processed) {
652 /* If the run queue exceeds the budget by up to 50%, let's cut it
653 * into two identical halves to improve latency.
654 */
655 max_processed = sched->rq_total / 2;
656 }
657
Willy Tarreau5c8be272020-06-19 12:17:55 +0200658 not_done_yet:
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200659 max[TL_URGENT] = max[TL_NORMAL] = max[TL_BULK] = 0;
Willy Tarreaucde79022019-04-12 18:03:41 +0200660
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200661 /* urgent tasklets list gets a default weight of ~50% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200662 if ((tt->tl_class_mask & (1 << TL_URGENT)) ||
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200663 !MT_LIST_ISEMPTY(&tt->shared_tasklet_list))
664 max[TL_URGENT] = default_weights[TL_URGENT];
Willy Tarreaua62917b2020-01-30 18:37:28 +0100665
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200666 /* normal tasklets list gets a default weight of ~37% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200667 if ((tt->tl_class_mask & (1 << TL_NORMAL)) ||
Willy Tarreau2c41d772021-02-24 16:13:03 +0100668 !eb_is_empty(&sched->rqueue) || (global_tasks_mask & tid_bit))
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200669 max[TL_NORMAL] = default_weights[TL_NORMAL];
Willy Tarreaua62917b2020-01-30 18:37:28 +0100670
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200671 /* bulk tasklets list gets a default weight of ~13% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200672 if ((tt->tl_class_mask & (1 << TL_BULK)))
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200673 max[TL_BULK] = default_weights[TL_BULK];
674
Willy Tarreau401135c2021-02-26 09:16:22 +0100675 /* heavy tasks are processed only once and never refilled in a
Willy Tarreau76390da2021-02-26 10:18:11 +0100676 * call round. That budget is not lost either as we don't reset
677 * it unless consumed.
Willy Tarreau401135c2021-02-26 09:16:22 +0100678 */
Willy Tarreau76390da2021-02-26 10:18:11 +0100679 if (!heavy_queued) {
680 if ((tt->tl_class_mask & (1 << TL_HEAVY)))
681 max[TL_HEAVY] = default_weights[TL_HEAVY];
682 else
683 max[TL_HEAVY] = 0;
684 heavy_queued = 1;
685 }
Willy Tarreau401135c2021-02-26 09:16:22 +0100686
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200687 /* Now compute a fair share of the weights. Total may slightly exceed
Willy Tarreau1553b662020-06-30 13:46:21 +0200688 * 100% due to rounding, this is not a problem. Note that while in
689 * theory the sum cannot be NULL as we cannot get there without tasklets
690 * to process, in practice it seldom happens when multiple writers
Willy Tarreau2b718102021-04-21 07:32:39 +0200691 * conflict and rollback on MT_LIST_TRY_APPEND(shared_tasklet_list), causing
Willy Tarreau1553b662020-06-30 13:46:21 +0200692 * a first MT_LIST_ISEMPTY() to succeed for thread_has_task() and the
693 * one above to finally fail. This is extremely rare and not a problem.
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200694 */
Willy Tarreau401135c2021-02-26 09:16:22 +0100695 max_total = max[TL_URGENT] + max[TL_NORMAL] + max[TL_BULK] + max[TL_HEAVY];
Willy Tarreau1553b662020-06-30 13:46:21 +0200696 if (!max_total)
697 return;
698
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200699 for (queue = 0; queue < TL_CLASSES; queue++)
700 max[queue] = ((unsigned)max_processed * max[queue] + max_total - 1) / max_total;
701
Willy Tarreau76390da2021-02-26 10:18:11 +0100702 /* The heavy queue must never process more than one task at once
703 * anyway.
704 */
705 if (max[TL_HEAVY] > 1)
706 max[TL_HEAVY] = 1;
707
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200708 lrq = grq = NULL;
Christopher Faulet8a48f672017-11-14 10:38:36 +0100709
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200710 /* pick up to max[TL_NORMAL] regular tasks from prio-ordered run queues */
711 /* Note: the grq lock is always held when grq is not null */
Willy Tarreaue7923c12021-02-25 07:09:08 +0100712 lpicked = gpicked = 0;
Willy Tarreau1f3b1412021-02-24 14:13:40 +0100713 budget = max[TL_NORMAL] - tt->tasks_in_list;
Willy Tarreaue7923c12021-02-25 07:09:08 +0100714 while (lpicked + gpicked < budget) {
Willy Tarreaucde79022019-04-12 18:03:41 +0200715 if ((global_tasks_mask & tid_bit) && !grq) {
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200716#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200717 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100718 grq = eb32sc_lookup_ge(&rqueue, global_rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreaucde79022019-04-12 18:03:41 +0200719 if (unlikely(!grq)) {
720 grq = eb32sc_first(&rqueue, tid_bit);
721 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200722 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200723 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200724 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100725 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200726#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200727 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100728
Willy Tarreaucde79022019-04-12 18:03:41 +0200729 /* If a global task is available for this thread, it's in grq
730 * now and the global RQ is locked.
731 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200732
Willy Tarreaucde79022019-04-12 18:03:41 +0200733 if (!lrq) {
Willy Tarreauc6ba9a02021-02-20 12:49:54 +0100734 lrq = eb32sc_lookup_ge(&tt->rqueue, tt->rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreaucde79022019-04-12 18:03:41 +0200735 if (unlikely(!lrq))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200736 lrq = eb32sc_first(&tt->rqueue, tid_bit);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100737 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100738
Willy Tarreaucde79022019-04-12 18:03:41 +0200739 if (!lrq && !grq)
740 break;
741
742 if (likely(!grq || (lrq && (int)(lrq->key - grq->key) <= 0))) {
743 t = eb32sc_entry(lrq, struct task, rq);
744 lrq = eb32sc_next(lrq, tid_bit);
Willy Tarreau2b363ac2021-02-25 07:14:58 +0100745 eb32sc_delete(&t->rq);
Willy Tarreaue7923c12021-02-25 07:09:08 +0100746 lpicked++;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200747 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200748#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200749 else {
750 t = eb32sc_entry(grq, struct task, rq);
751 grq = eb32sc_next(grq, tid_bit);
Willy Tarreau2b363ac2021-02-25 07:14:58 +0100752 _HA_ATOMIC_AND(&t->state, ~TASK_GLOBAL);
753 eb32sc_delete(&t->rq);
754
Willy Tarreaucde79022019-04-12 18:03:41 +0200755 if (unlikely(!grq)) {
756 grq = eb32sc_first(&rqueue, tid_bit);
757 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200758 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200759 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaucde79022019-04-12 18:03:41 +0200760 }
761 }
Willy Tarreaue7923c12021-02-25 07:09:08 +0100762 gpicked++;
Emeric Brun01948972017-03-30 15:37:25 +0200763 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200764#endif
Willy Tarreau2b363ac2021-02-25 07:14:58 +0100765 if (t->nice)
Willy Tarreau4781b152021-04-06 13:53:36 +0200766 _HA_ATOMIC_DEC(&niced_tasks);
Willy Tarreaucde79022019-04-12 18:03:41 +0200767
Willy Tarreaua868c292020-11-30 15:30:22 +0100768 /* Add it to the local task list */
Willy Tarreau2b718102021-04-21 07:32:39 +0200769 LIST_APPEND(&tt->tasklets[TL_NORMAL], &((struct tasklet *)t)->list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200770 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200771
772 /* release the rqueue lock */
773 if (grq) {
774 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
775 grq = NULL;
776 }
777
Willy Tarreaue7923c12021-02-25 07:09:08 +0100778 if (lpicked + gpicked) {
Willy Tarreauc309dbd2020-11-30 15:39:00 +0100779 tt->tl_class_mask |= 1 << TL_NORMAL;
Willy Tarreaue7923c12021-02-25 07:09:08 +0100780 _HA_ATOMIC_ADD(&tt->tasks_in_list, lpicked + gpicked);
Willy Tarreaub7e0c632021-03-09 09:59:50 +0100781#ifdef USE_THREAD
Willy Tarreau45499c52021-02-25 07:51:18 +0100782 if (gpicked) {
783 _HA_ATOMIC_SUB(&grq_total, gpicked);
Willy Tarreauc9afbb12021-02-25 07:19:45 +0100784 _HA_ATOMIC_ADD(&tt->rq_total, gpicked);
Willy Tarreau45499c52021-02-25 07:51:18 +0100785 }
Willy Tarreaub7e0c632021-03-09 09:59:50 +0100786#endif
Willy Tarreaue7923c12021-02-25 07:09:08 +0100787 activity[tid].tasksw += lpicked + gpicked;
Willy Tarreauc309dbd2020-11-30 15:39:00 +0100788 }
789
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200790 /* Merge the list of tasklets waken up by other threads to the
791 * main list.
792 */
793 tmp_list = MT_LIST_BEHEAD(&tt->shared_tasklet_list);
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200794 if (tmp_list) {
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200795 LIST_SPLICE_END_DETACHED(&tt->tasklets[TL_URGENT], (struct list *)tmp_list);
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200796 if (!LIST_ISEMPTY(&tt->tasklets[TL_URGENT]))
797 tt->tl_class_mask |= 1 << TL_URGENT;
798 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200799
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200800 /* execute tasklets in each queue */
Willy Tarreau59153fe2020-06-24 10:17:29 +0200801 max_processed -= run_tasks_from_lists(max);
Willy Tarreaua62917b2020-01-30 18:37:28 +0100802
Willy Tarreau5c8be272020-06-19 12:17:55 +0200803 /* some tasks may have woken other ones up */
Willy Tarreau0c0c85e2020-06-23 11:32:35 +0200804 if (max_processed > 0 && thread_has_tasks())
Willy Tarreau5c8be272020-06-19 12:17:55 +0200805 goto not_done_yet;
806
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200807 if (tt->tl_class_mask)
Willy Tarreaucde79022019-04-12 18:03:41 +0200808 activity[tid].long_rq++;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200809}
810
Willy Tarreau64e60122019-07-12 08:31:17 +0200811/* create a work list array for <nbthread> threads, using tasks made of
812 * function <fct>. The context passed to the function will be the pointer to
813 * the thread's work list, which will contain a copy of argument <arg>. The
814 * wake up reason will be TASK_WOKEN_OTHER. The pointer to the work_list array
815 * is returned on success, otherwise NULL on failure.
816 */
817struct work_list *work_list_create(int nbthread,
Willy Tarreau144f84a2021-03-02 16:09:26 +0100818 struct task *(*fct)(struct task *, void *, unsigned int),
Willy Tarreau64e60122019-07-12 08:31:17 +0200819 void *arg)
820{
821 struct work_list *wl;
822 int i;
823
824 wl = calloc(nbthread, sizeof(*wl));
825 if (!wl)
826 goto fail;
827
828 for (i = 0; i < nbthread; i++) {
Olivier Houchard859dc802019-08-08 15:47:21 +0200829 MT_LIST_INIT(&wl[i].head);
Willy Tarreau64e60122019-07-12 08:31:17 +0200830 wl[i].task = task_new(1UL << i);
831 if (!wl[i].task)
832 goto fail;
833 wl[i].task->process = fct;
834 wl[i].task->context = &wl[i];
835 wl[i].arg = arg;
836 }
837 return wl;
838
839 fail:
840 work_list_destroy(wl, nbthread);
841 return NULL;
842}
843
844/* destroy work list <work> */
845void work_list_destroy(struct work_list *work, int nbthread)
846{
847 int t;
848
849 if (!work)
850 return;
851 for (t = 0; t < nbthread; t++)
852 task_destroy(work[t].task);
853 free(work);
854}
855
William Lallemand27f3fa52018-12-06 14:05:20 +0100856/*
857 * Delete every tasks before running the master polling loop
858 */
859void mworker_cleantasks()
860{
861 struct task *t;
862 int i;
William Lallemandb5823392018-12-06 15:14:37 +0100863 struct eb32_node *tmp_wq = NULL;
864 struct eb32sc_node *tmp_rq = NULL;
William Lallemand27f3fa52018-12-06 14:05:20 +0100865
866#ifdef USE_THREAD
867 /* cleanup the global run queue */
William Lallemandb5823392018-12-06 15:14:37 +0100868 tmp_rq = eb32sc_first(&rqueue, MAX_THREADS_MASK);
869 while (tmp_rq) {
870 t = eb32sc_entry(tmp_rq, struct task, rq);
871 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200872 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100873 }
874 /* cleanup the timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100875 tmp_wq = eb32_first(&timers);
876 while (tmp_wq) {
877 t = eb32_entry(tmp_wq, struct task, wq);
878 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200879 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100880 }
881#endif
882 /* clean the per thread run queue */
883 for (i = 0; i < global.nbthread; i++) {
William Lallemandb5823392018-12-06 15:14:37 +0100884 tmp_rq = eb32sc_first(&task_per_thread[i].rqueue, MAX_THREADS_MASK);
885 while (tmp_rq) {
886 t = eb32sc_entry(tmp_rq, struct task, rq);
887 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200888 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100889 }
890 /* cleanup the per thread timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100891 tmp_wq = eb32_first(&task_per_thread[i].timers);
892 while (tmp_wq) {
893 t = eb32_entry(tmp_wq, struct task, wq);
894 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200895 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100896 }
897 }
898}
899
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100900/* perform minimal intializations */
901static void init_task()
Willy Tarreau4726f532009-03-07 17:25:21 +0100902{
Willy Tarreau401135c2021-02-26 09:16:22 +0100903 int i, q;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200904
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200905#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200906 memset(&timers, 0, sizeof(timers));
Willy Tarreau4726f532009-03-07 17:25:21 +0100907 memset(&rqueue, 0, sizeof(rqueue));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200908#endif
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200909 memset(&task_per_thread, 0, sizeof(task_per_thread));
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200910 for (i = 0; i < MAX_THREADS; i++) {
Willy Tarreau401135c2021-02-26 09:16:22 +0100911 for (q = 0; q < TL_CLASSES; q++)
912 LIST_INIT(&task_per_thread[i].tasklets[q]);
Olivier Houchard06910462019-10-11 16:35:01 +0200913 MT_LIST_INIT(&task_per_thread[i].shared_tasklet_list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200914 }
Willy Tarreau4726f532009-03-07 17:25:21 +0100915}
916
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200917/* config parser for global "tune.sched.low-latency", accepts "on" or "off" */
918static int cfg_parse_tune_sched_low_latency(char **args, int section_type, struct proxy *curpx,
Willy Tarreau01825162021-03-09 09:53:46 +0100919 const struct proxy *defpx, const char *file, int line,
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200920 char **err)
921{
922 if (too_many_args(1, args, err, NULL))
923 return -1;
924
925 if (strcmp(args[1], "on") == 0)
926 global.tune.options |= GTUNE_SCHED_LOW_LATENCY;
927 else if (strcmp(args[1], "off") == 0)
928 global.tune.options &= ~GTUNE_SCHED_LOW_LATENCY;
929 else {
930 memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
931 return -1;
932 }
933 return 0;
934}
935
936/* config keyword parsers */
937static struct cfg_kw_list cfg_kws = {ILH, {
938 { CFG_GLOBAL, "tune.sched.low-latency", cfg_parse_tune_sched_low_latency },
939 { 0, NULL, NULL }
940}};
941
942INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100943INITCALL0(STG_PREPARE, init_task);
944
Willy Tarreaubaaee002006-06-26 02:48:02 +0200945/*
946 * Local variables:
947 * c-indent-level: 8
948 * c-basic-offset: 8
949 * End:
950 */