blob: a2b78e8fce83427b4bbe22fdec7b2ab72afefd4a [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 Houchardb1ca58b2018-06-06 14:22:03 +020053#endif
Willy Tarreaub20aa9e2018-10-15 14:52:21 +020054
Willy Tarreaue35c94a2009-03-21 10:01:42 +010055static unsigned int rqueue_ticks; /* insertion count */
Willy Tarreau8d8747a2018-10-15 16:12:48 +020056
57struct task_per_thread task_per_thread[MAX_THREADS];
Willy Tarreau9789f7b2008-06-24 08:17:16 +020058
Willy Tarreaueb8c2c62020-06-30 11:48:48 +020059
60/* Flags the task <t> for immediate destruction and puts it into its first
61 * thread's shared tasklet list if not yet queued/running. This will bypass
62 * the priority scheduling and make the task show up as fast as possible in
63 * the other thread's queue. Note that this operation isn't idempotent and is
64 * not supposed to be run on the same task from multiple threads at once. It's
65 * the caller's responsibility to make sure it is the only one able to kill the
66 * task.
67 */
68void task_kill(struct task *t)
69{
70 unsigned short state = t->state;
71 unsigned int thr;
72
73 BUG_ON(state & TASK_KILLED);
74
75 while (1) {
76 while (state & (TASK_RUNNING | TASK_QUEUED)) {
77 /* task already in the queue and about to be executed,
78 * or even currently running. Just add the flag and be
79 * done with it, the process loop will detect it and kill
80 * it. The CAS will fail if we arrive too late.
81 */
82 if (_HA_ATOMIC_CAS(&t->state, &state, state | TASK_KILLED))
83 return;
84 }
85
86 /* We'll have to wake it up, but we must also secure it so that
87 * it doesn't vanish under us. TASK_QUEUED guarantees nobody will
88 * add past us.
89 */
90 if (_HA_ATOMIC_CAS(&t->state, &state, state | TASK_QUEUED | TASK_KILLED)) {
91 /* Bypass the tree and go directly into the shared tasklet list.
92 * Note: that's a task so it must be accounted for as such. Pick
93 * the task's first thread for the job.
94 */
95 thr = my_ffsl(t->thread_mask) - 1;
Willy Tarreau54d31172020-07-02 14:14:00 +020096
97 /* Beware: tasks that have never run don't have their ->list empty yet! */
Willy Tarreau950954f2020-07-10 08:32:10 +020098 MT_LIST_ADDQ(&task_per_thread[thr].shared_tasklet_list,
Willy Tarreau4f589262020-07-02 17:17:42 +020099 (struct mt_list *)&((struct tasklet *)t)->list);
Willy Tarreau54d31172020-07-02 14:14:00 +0200100 _HA_ATOMIC_ADD(&tasks_run_queue, 1);
101 _HA_ATOMIC_ADD(&task_per_thread[thr].task_list_size, 1);
102 if (sleeping_thread_mask & (1UL << thr)) {
103 _HA_ATOMIC_AND(&sleeping_thread_mask, ~(1UL << thr));
104 wake_thread(thr);
Willy Tarreaueb8c2c62020-06-30 11:48:48 +0200105 }
Willy Tarreau54d31172020-07-02 14:14:00 +0200106 return;
Willy Tarreaueb8c2c62020-06-30 11:48:48 +0200107 }
108 }
109}
110
Willy Tarreau4726f532009-03-07 17:25:21 +0100111/* Puts the task <t> in run queue at a position depending on t->nice. <t> is
112 * returned. The nice value assigns boosts in 32th of the run queue size. A
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100113 * nice value of -1024 sets the task to -tasks_run_queue*32, while a nice value
114 * of 1024 sets the task to tasks_run_queue*32. The state flags are cleared, so
115 * the caller will have to set its flags after this call.
Willy Tarreau4726f532009-03-07 17:25:21 +0100116 * The task must not already be in the run queue. If unsure, use the safer
117 * task_wakeup() function.
Willy Tarreau91e99932008-06-30 07:51:00 +0200118 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200119void __task_wakeup(struct task *t, struct eb_root *root)
Willy Tarreaue33aece2007-04-30 13:15:14 +0200120{
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200121#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200122 if (root == &rqueue) {
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200123 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200124 }
Willy Tarreau2d1fd0a2019-04-15 09:18:31 +0200125#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200126 /* Make sure if the task isn't in the runqueue, nobody inserts it
127 * in the meanwhile.
128 */
Olivier Houchard4c2832852019-03-08 18:48:47 +0100129 _HA_ATOMIC_ADD(&tasks_run_queue, 1);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200130#ifdef USE_THREAD
131 if (root == &rqueue) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200132 global_tasks_mask |= t->thread_mask;
Olivier Houcharded1a6a02019-04-18 14:12:51 +0200133 __ha_barrier_store();
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200134 }
135#endif
Olivier Houchard4c2832852019-03-08 18:48:47 +0100136 t->rq.key = _HA_ATOMIC_ADD(&rqueue_ticks, 1);
Willy Tarreau91e99932008-06-30 07:51:00 +0200137
138 if (likely(t->nice)) {
139 int offset;
140
Olivier Houchard4c2832852019-03-08 18:48:47 +0100141 _HA_ATOMIC_ADD(&niced_tasks, 1);
Willy Tarreau2d1fd0a2019-04-15 09:18:31 +0200142 offset = t->nice * (int)global.tune.runqueue_depth;
Willy Tarreau4726f532009-03-07 17:25:21 +0100143 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +0200144 }
145
Willy Tarreaud9add3a2019-04-25 08:57:41 +0200146 if (task_profiling_mask & tid_bit)
Willy Tarreau9efd7452018-05-31 14:48:54 +0200147 t->call_date = now_mono_time();
148
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200149 eb32sc_insert(root, &t->rq, t->thread_mask);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200150#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200151 if (root == &rqueue) {
Olivier Houchard4c2832852019-03-08 18:48:47 +0100152 _HA_ATOMIC_OR(&t->state, TASK_GLOBAL);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200153 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200154 } else
155#endif
156 {
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200157 int nb = ((void *)root - (void *)&task_per_thread[0].rqueue) / sizeof(task_per_thread[0]);
158 task_per_thread[nb].rqueue_size++;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200159 }
Willy Tarreau85d9b842018-07-27 17:14:41 +0200160#ifdef USE_THREAD
Olivier Houchard79321b92018-07-26 17:55:11 +0200161 /* If all threads that are supposed to handle this task are sleeping,
162 * wake one.
163 */
164 if ((((t->thread_mask & all_threads_mask) & sleeping_thread_mask) ==
Olivier Houchard1b327902019-03-15 00:23:10 +0100165 (t->thread_mask & all_threads_mask))) {
166 unsigned long m = (t->thread_mask & all_threads_mask) &~ tid_bit;
167
168 m = (m & (m - 1)) ^ m; // keep lowest bit set
169 _HA_ATOMIC_AND(&sleeping_thread_mask, ~m);
170 wake_thread(my_ffsl(m) - 1);
171 }
Willy Tarreau85d9b842018-07-27 17:14:41 +0200172#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200173 return;
Willy Tarreaue33aece2007-04-30 13:15:14 +0200174}
Willy Tarreaud825eef2007-05-12 22:35:00 +0200175
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200176/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100177 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200178 *
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200179 * Inserts a task into wait queue <wq> at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +0100180 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100181 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100182 * Last, tasks must not be queued further than the end of the tree, which is
183 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100184 *
185 * This function should not be used directly, it is meant to be called by the
186 * inline version of task_queue() which performs a few cheap preliminary tests
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200187 * before deciding to call __task_queue(). Moreover this function doesn't care
188 * at all about locking so the caller must be careful when deciding whether to
189 * lock or not around this call.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200190 */
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200191void __task_queue(struct task *task, struct eb_root *wq)
Willy Tarreau964c9362007-01-07 00:38:00 +0100192{
Willy Tarreaue5d79bc2020-07-22 14:29:42 +0200193#ifdef USE_THREAD
194 BUG_ON((wq == &timers && !(task->state & TASK_SHARED_WQ)) ||
195 (wq == &sched->timers && (task->state & TASK_SHARED_WQ)) ||
196 (wq != &timers && wq != &sched->timers));
197#endif
198
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100199 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +0100200 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +0100201
202 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100203 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200204#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100205 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +0200206 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +0100207 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200208#endif
Willy Tarreauce44f122008-07-05 18:16:19 +0200209
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200210 eb32_insert(wq, &task->wq);
Willy Tarreau964c9362007-01-07 00:38:00 +0100211}
212
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200213/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200214 * Extract all expired timers from the timer queue, and wakes up all
Willy Tarreauc49ba522019-12-11 08:12:23 +0100215 * associated tasks.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200216 */
Willy Tarreauc49ba522019-12-11 08:12:23 +0100217void wake_expired_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200218{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200219 struct task_per_thread * const tt = sched; // thread's tasks
Willy Tarreau3cfaa8d2020-10-16 09:26:22 +0200220 int max_processed = global.tune.runqueue_depth;
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200221 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200222 struct eb32_node *eb;
Willy Tarreauaf613e82020-06-05 08:40:51 +0200223 __decl_thread(int key);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200224
Willy Tarreau3cfaa8d2020-10-16 09:26:22 +0200225 while (max_processed-- > 0) {
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200226 lookup_next_local:
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200227 eb = eb32_lookup_ge(&tt->timers, now_ms - TIMER_LOOK_BACK);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200228 if (!eb) {
229 /* we might have reached the end of the tree, typically because
230 * <now_ms> is in the first half and we're first scanning the last
231 * half. Let's loop back to the beginning of the tree now.
232 */
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200233 eb = eb32_first(&tt->timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200234 if (likely(!eb))
235 break;
236 }
237
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200238 /* It is possible that this task was left at an earlier place in the
239 * tree because a recent call to task_queue() has not moved it. This
240 * happens when the new expiration date is later than the old one.
241 * Since it is very unlikely that we reach a timeout anyway, it's a
242 * lot cheaper to proceed like this because we almost never update
243 * the tree. We may also find disabled expiration dates there. Since
244 * we have detached the task from the tree, we simply call task_queue
245 * to take care of this. Note that we might occasionally requeue it at
246 * the same place, before <eb>, so we have to check if this happens,
247 * and adjust <eb>, otherwise we may skip it which is not what we want.
248 * We may also not requeue the task (and not point eb at it) if its
Willy Tarreau77015ab2020-06-19 11:50:27 +0200249 * expiration time is not set. We also make sure we leave the real
250 * expiration date for the next task in the queue so that when calling
251 * next_timer_expiry() we're guaranteed to see the next real date and
252 * not the next apparent date. This is in order to avoid useless
253 * wakeups.
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200254 */
Willy Tarreau77015ab2020-06-19 11:50:27 +0200255
256 task = eb32_entry(eb, struct task, wq);
257 if (tick_is_expired(task->expire, now_ms)) {
258 /* expired task, wake it up */
259 __task_unlink_wq(task);
260 task_wakeup(task, TASK_WOKEN_TIMER);
261 }
262 else if (task->expire != eb->key) {
263 /* task is not expired but its key doesn't match so let's
264 * update it and skip to next apparently expired task.
265 */
266 __task_unlink_wq(task);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200267 if (tick_isset(task->expire))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200268 __task_queue(task, &tt->timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200269 }
Willy Tarreau77015ab2020-06-19 11:50:27 +0200270 else {
271 /* task not expired and correctly placed */
272 break;
273 }
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200274 }
275
276#ifdef USE_THREAD
Willy Tarreau1e928c02019-05-28 18:57:25 +0200277 if (eb_is_empty(&timers))
278 goto leave;
279
280 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
281 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
282 if (!eb) {
283 eb = eb32_first(&timers);
284 if (likely(!eb)) {
285 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
286 goto leave;
287 }
288 }
289 key = eb->key;
Willy Tarreau1e928c02019-05-28 18:57:25 +0200290
Willy Tarreaud48ed662020-10-16 09:31:41 +0200291 if (tick_is_lt(now_ms, key)) {
292 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau1e928c02019-05-28 18:57:25 +0200293 goto leave;
Willy Tarreaud48ed662020-10-16 09:31:41 +0200294 }
Willy Tarreau1e928c02019-05-28 18:57:25 +0200295
296 /* There's really something of interest here, let's visit the queue */
297
Willy Tarreaud48ed662020-10-16 09:31:41 +0200298 if (HA_RWLOCK_TRYRDTOSK(TASK_WQ_LOCK, &wq_lock)) {
299 /* if we failed to grab the lock it means another thread is
300 * already doing the same here, so let it do the job.
301 */
302 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
303 goto leave;
304 }
305
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200306 while (1) {
Emeric Brunc60def82017-09-27 14:59:38 +0200307 lookup_next:
Willy Tarreau3cfaa8d2020-10-16 09:26:22 +0200308 if (max_processed-- <= 0)
309 break;
Emeric Brun01948972017-03-30 15:37:25 +0200310 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200311 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100312 /* we might have reached the end of the tree, typically because
313 * <now_ms> is in the first half and we're first scanning the last
314 * half. Let's loop back to the beginning of the tree now.
315 */
316 eb = eb32_first(&timers);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100317 if (likely(!eb))
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100318 break;
319 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200320
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100321 task = eb32_entry(eb, struct task, wq);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200322 if (tick_is_expired(task->expire, now_ms)) {
323 /* expired task, wake it up */
Willy Tarreaud48ed662020-10-16 09:31:41 +0200324 HA_RWLOCK_SKTOWR(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200325 __task_unlink_wq(task);
Willy Tarreaud48ed662020-10-16 09:31:41 +0200326 HA_RWLOCK_WRTOSK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200327 task_wakeup(task, TASK_WOKEN_TIMER);
328 }
329 else if (task->expire != eb->key) {
330 /* task is not expired but its key doesn't match so let's
331 * update it and skip to next apparently expired task.
332 */
Willy Tarreaud48ed662020-10-16 09:31:41 +0200333 HA_RWLOCK_SKTOWR(TASK_WQ_LOCK, &wq_lock);
Willy Tarreau77015ab2020-06-19 11:50:27 +0200334 __task_unlink_wq(task);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100335 if (tick_isset(task->expire))
Willy Tarreau783afbe2020-07-22 14:12:45 +0200336 __task_queue(task, &timers);
Willy Tarreaud48ed662020-10-16 09:31:41 +0200337 HA_RWLOCK_WRTOSK(TASK_WQ_LOCK, &wq_lock);
Emeric Brunc60def82017-09-27 14:59:38 +0200338 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200339 }
Willy Tarreau77015ab2020-06-19 11:50:27 +0200340 else {
341 /* task not expired and correctly placed */
342 break;
343 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100344 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200345
Willy Tarreaud48ed662020-10-16 09:31:41 +0200346 HA_RWLOCK_SKUNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200347#endif
Willy Tarreau1e928c02019-05-28 18:57:25 +0200348leave:
Willy Tarreauc49ba522019-12-11 08:12:23 +0100349 return;
350}
351
352/* Checks the next timer for the current thread by looking into its own timer
353 * list and the global one. It may return TICK_ETERNITY if no timer is present.
Ilya Shipitsin856aabc2020-04-16 23:51:34 +0500354 * Note that the next timer might very well be slightly in the past.
Willy Tarreauc49ba522019-12-11 08:12:23 +0100355 */
356int next_timer_expiry()
357{
358 struct task_per_thread * const tt = sched; // thread's tasks
359 struct eb32_node *eb;
360 int ret = TICK_ETERNITY;
Willy Tarreau6ce02322020-08-21 05:48:34 +0200361 __decl_thread(int key = TICK_ETERNITY);
Willy Tarreauc49ba522019-12-11 08:12:23 +0100362
363 /* first check in the thread-local timers */
364 eb = eb32_lookup_ge(&tt->timers, now_ms - TIMER_LOOK_BACK);
365 if (!eb) {
366 /* we might have reached the end of the tree, typically because
367 * <now_ms> is in the first half and we're first scanning the last
368 * half. Let's loop back to the beginning of the tree now.
369 */
370 eb = eb32_first(&tt->timers);
371 }
372
373 if (eb)
374 ret = eb->key;
375
376#ifdef USE_THREAD
377 if (!eb_is_empty(&timers)) {
378 HA_RWLOCK_RDLOCK(TASK_WQ_LOCK, &wq_lock);
379 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
380 if (!eb)
381 eb = eb32_first(&timers);
382 if (eb)
383 key = eb->key;
384 HA_RWLOCK_RDUNLOCK(TASK_WQ_LOCK, &wq_lock);
385 if (eb)
386 ret = tick_first(ret, key);
387 }
388#endif
Willy Tarreaub992ba12017-11-05 19:09:27 +0100389 return ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200390}
391
Willy Tarreau59153fe2020-06-24 10:17:29 +0200392/* Walks over tasklet lists sched->tasklets[0..TL_CLASSES-1] and run at most
393 * budget[TL_*] of them. Returns the number of entries effectively processed
394 * (tasks and tasklets merged). The count of tasks in the list for the current
395 * thread is adjusted.
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100396 */
Willy Tarreau59153fe2020-06-24 10:17:29 +0200397unsigned int run_tasks_from_lists(unsigned int budgets[])
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100398{
399 struct task *(*process)(struct task *t, void *ctx, unsigned short state);
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200400 struct list *tl_queues = sched->tasklets;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100401 struct task *t;
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200402 uint8_t budget_mask = (1 << TL_CLASSES) - 1;
Willy Tarreau59153fe2020-06-24 10:17:29 +0200403 unsigned int done = 0;
404 unsigned int queue;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100405 unsigned short state;
406 void *ctx;
Willy Tarreau59153fe2020-06-24 10:17:29 +0200407
408 for (queue = 0; queue < TL_CLASSES;) {
409 sched->current_queue = queue;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100410
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200411 /* global.tune.sched.low-latency is set */
412 if (global.tune.options & GTUNE_SCHED_LOW_LATENCY) {
413 if (unlikely(sched->tl_class_mask & budget_mask & ((1 << queue) - 1))) {
414 /* a lower queue index has tasks again and still has a
415 * budget to run them. Let's switch to it now.
416 */
417 queue = (sched->tl_class_mask & 1) ? 0 :
418 (sched->tl_class_mask & 2) ? 1 : 2;
419 continue;
420 }
421
422 if (unlikely(queue > TL_URGENT &&
423 budget_mask & (1 << TL_URGENT) &&
424 !MT_LIST_ISEMPTY(&sched->shared_tasklet_list))) {
425 /* an urgent tasklet arrived from another thread */
426 break;
427 }
428
429 if (unlikely(queue > TL_NORMAL &&
430 budget_mask & (1 << TL_NORMAL) &&
431 ((sched->rqueue_size > 0) ||
432 (global_tasks_mask & tid_bit)))) {
433 /* a task was woken up by a bulk tasklet or another thread */
434 break;
435 }
436 }
437
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200438 if (LIST_ISEMPTY(&tl_queues[queue])) {
439 sched->tl_class_mask &= ~(1 << queue);
Willy Tarreau59153fe2020-06-24 10:17:29 +0200440 queue++;
441 continue;
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200442 }
443
Willy Tarreau59153fe2020-06-24 10:17:29 +0200444 if (!budgets[queue]) {
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200445 budget_mask &= ~(1 << queue);
Willy Tarreau59153fe2020-06-24 10:17:29 +0200446 queue++;
447 continue;
448 }
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200449
Willy Tarreau59153fe2020-06-24 10:17:29 +0200450 budgets[queue]--;
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200451 t = (struct task *)LIST_ELEM(tl_queues[queue].n, struct tasklet *, list);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200452 state = t->state & (TASK_SHARED_WQ|TASK_SELF_WAKING|TASK_KILLED);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100453
454 ti->flags &= ~TI_FL_STUCK; // this thread is still running
455 activity[tid].ctxsw++;
456 ctx = t->context;
457 process = t->process;
458 t->calls++;
Willy Tarreaud23d4132020-01-31 10:39:03 +0100459 sched->current = t;
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100460
Willy Tarreau9a0374a2020-11-30 14:52:11 +0100461 _HA_ATOMIC_SUB(&tasks_run_queue, 1);
462
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100463 if (TASK_IS_TASKLET(t)) {
Willy Tarreau3027cdc2020-11-30 14:58:53 +0100464 LIST_DEL_INIT(&((struct tasklet *)t)->list);
465 __ha_barrier_store();
Willy Tarreau952c2642020-01-31 16:39:30 +0100466 state = _HA_ATOMIC_XCHG(&t->state, state);
467 __ha_barrier_atomic_store();
Olivier Houchardc62d9ab2020-03-17 18:15:04 +0100468 process(t, ctx, state);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100469 done++;
Willy Tarreaud23d4132020-01-31 10:39:03 +0100470 sched->current = NULL;
471 __ha_barrier_store();
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100472 continue;
473 }
474
Willy Tarreau3027cdc2020-11-30 14:58:53 +0100475 LIST_DEL_INIT(&((struct tasklet *)t)->list);
476 __ha_barrier_store();
Willy Tarreau952c2642020-01-31 16:39:30 +0100477 state = _HA_ATOMIC_XCHG(&t->state, state | TASK_RUNNING);
478 __ha_barrier_atomic_store();
Willy Tarreau952c2642020-01-31 16:39:30 +0100479
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100480 /* OK then this is a regular task */
481
Willy Tarreaud99177f2020-06-30 11:48:48 +0200482 _HA_ATOMIC_SUB(&task_per_thread[tid].task_list_size, 1);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100483 if (unlikely(t->call_date)) {
484 uint64_t now_ns = now_mono_time();
485
486 t->lat_time += now_ns - t->call_date;
487 t->call_date = now_ns;
488 }
489
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100490 __ha_barrier_store();
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200491
492 /* Note for below: if TASK_KILLED arrived before we've read the state, we
493 * directly free the task. Otherwise it will be seen after processing and
494 * it's freed on the exit path.
495 */
496 if (likely(!(state & TASK_KILLED) && process == process_stream))
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100497 t = process_stream(t, ctx, state);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200498 else if (!(state & TASK_KILLED) && process != NULL)
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100499 t = process(t, ctx, state);
500 else {
Willy Tarreau273aea42020-07-17 14:37:51 +0200501 task_unlink_wq(t);
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100502 __task_free(t);
503 sched->current = NULL;
504 __ha_barrier_store();
505 /* We don't want max_processed to be decremented if
506 * we're just freeing a destroyed task, we should only
507 * do so if we really ran a task.
508 */
509 continue;
510 }
511 sched->current = NULL;
512 __ha_barrier_store();
513 /* If there is a pending state we have to wake up the task
514 * immediately, else we defer it into wait queue
515 */
516 if (t != NULL) {
517 if (unlikely(t->call_date)) {
518 t->cpu_time += now_mono_time() - t->call_date;
519 t->call_date = 0;
520 }
521
522 state = _HA_ATOMIC_AND(&t->state, ~TASK_RUNNING);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200523 if (unlikely(state & TASK_KILLED)) {
Willy Tarreau273aea42020-07-17 14:37:51 +0200524 task_unlink_wq(t);
Willy Tarreau8a6049c2020-06-30 11:48:48 +0200525 __task_free(t);
526 }
527 else if (state & TASK_WOKEN_ANY)
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100528 task_wakeup(t, 0);
529 else
530 task_queue(t);
531 }
532 done++;
533 }
Willy Tarreauba48d5c2020-06-24 09:54:24 +0200534 sched->current_queue = -1;
Willy Tarreau116ef222020-06-23 16:35:38 +0200535
Willy Tarreau4ffa0b52020-01-30 18:13:13 +0100536 return done;
537}
538
Willy Tarreau58b458d2008-06-29 22:40:23 +0200539/* The run queue is chronologically sorted in a tree. An insertion counter is
540 * used to assign a position to each task. This counter may be combined with
541 * other variables (eg: nice value) to set the final position in the tree. The
542 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100543 * tasks processed to 200 in any case, so that general latency remains low and
Willy Tarreaucde79022019-04-12 18:03:41 +0200544 * so that task positions have a chance to be considered. The function scans
545 * both the global and local run queues and picks the most urgent task between
546 * the two. We need to grab the global runqueue lock to touch it so it's taken
547 * on the very first access to the global run queue and is released as soon as
548 * it reaches the end.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200549 *
550 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200551 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100552void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200553{
Willy Tarreaud022e9c2019-09-24 08:25:15 +0200554 struct task_per_thread * const tt = sched;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200555 struct eb32sc_node *lrq; // next local run queue entry
556 struct eb32sc_node *grq; // next global run queue entry
Willy Tarreau964c9362007-01-07 00:38:00 +0100557 struct task *t;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200558 const unsigned int default_weights[TL_CLASSES] = {
559 [TL_URGENT] = 64, // ~50% of CPU bandwidth for I/O
560 [TL_NORMAL] = 48, // ~37% of CPU bandwidth for tasks
561 [TL_BULK] = 16, // ~13% of CPU bandwidth for self-wakers
562 };
563 unsigned int max[TL_CLASSES]; // max to be run per class
564 unsigned int max_total; // sum of max above
Olivier Houchard06910462019-10-11 16:35:01 +0200565 struct mt_list *tmp_list;
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200566 unsigned int queue;
567 int max_processed;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100568
Willy Tarreaue6a02fa2019-05-22 07:06:44 +0200569 ti->flags &= ~TI_FL_STUCK; // this thread is still running
570
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200571 if (!thread_has_tasks()) {
572 activity[tid].empty_rq++;
573 return;
574 }
575
Willy Tarreau5c8be272020-06-19 12:17:55 +0200576 tasks_run_queue_cur = tasks_run_queue; /* keep a copy for reporting */
577 nb_tasks_cur = nb_tasks;
578 max_processed = global.tune.runqueue_depth;
579
580 if (likely(niced_tasks))
581 max_processed = (max_processed + 3) / 4;
582
Willy Tarreau5c8be272020-06-19 12:17:55 +0200583 not_done_yet:
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200584 max[TL_URGENT] = max[TL_NORMAL] = max[TL_BULK] = 0;
Willy Tarreaucde79022019-04-12 18:03:41 +0200585
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200586 /* urgent tasklets list gets a default weight of ~50% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200587 if ((tt->tl_class_mask & (1 << TL_URGENT)) ||
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200588 !MT_LIST_ISEMPTY(&tt->shared_tasklet_list))
589 max[TL_URGENT] = default_weights[TL_URGENT];
Willy Tarreaua62917b2020-01-30 18:37:28 +0100590
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200591 /* normal tasklets list gets a default weight of ~37% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200592 if ((tt->tl_class_mask & (1 << TL_NORMAL)) ||
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200593 (sched->rqueue_size > 0) || (global_tasks_mask & tid_bit))
594 max[TL_NORMAL] = default_weights[TL_NORMAL];
Willy Tarreaua62917b2020-01-30 18:37:28 +0100595
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200596 /* bulk tasklets list gets a default weight of ~13% */
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200597 if ((tt->tl_class_mask & (1 << TL_BULK)))
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200598 max[TL_BULK] = default_weights[TL_BULK];
599
600 /* Now compute a fair share of the weights. Total may slightly exceed
Willy Tarreau1553b662020-06-30 13:46:21 +0200601 * 100% due to rounding, this is not a problem. Note that while in
602 * theory the sum cannot be NULL as we cannot get there without tasklets
603 * to process, in practice it seldom happens when multiple writers
Willy Tarreaude4db172020-07-10 08:10:29 +0200604 * conflict and rollback on MT_LIST_TRY_ADDQ(shared_tasklet_list), causing
Willy Tarreau1553b662020-06-30 13:46:21 +0200605 * a first MT_LIST_ISEMPTY() to succeed for thread_has_task() and the
606 * one above to finally fail. This is extremely rare and not a problem.
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200607 */
608 max_total = max[TL_URGENT] + max[TL_NORMAL] + max[TL_BULK];
Willy Tarreau1553b662020-06-30 13:46:21 +0200609 if (!max_total)
610 return;
611
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200612 for (queue = 0; queue < TL_CLASSES; queue++)
613 max[queue] = ((unsigned)max_processed * max[queue] + max_total - 1) / max_total;
614
615 lrq = grq = NULL;
Christopher Faulet8a48f672017-11-14 10:38:36 +0100616
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200617 /* pick up to max[TL_NORMAL] regular tasks from prio-ordered run queues */
618 /* Note: the grq lock is always held when grq is not null */
619 while (tt->task_list_size < max[TL_NORMAL]) {
Willy Tarreaucde79022019-04-12 18:03:41 +0200620 if ((global_tasks_mask & tid_bit) && !grq) {
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200621#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200622 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
623 grq = eb32sc_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
624 if (unlikely(!grq)) {
625 grq = eb32sc_first(&rqueue, tid_bit);
626 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200627 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200628 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200629 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100630 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200631#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200632 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100633
Willy Tarreaucde79022019-04-12 18:03:41 +0200634 /* If a global task is available for this thread, it's in grq
635 * now and the global RQ is locked.
636 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200637
Willy Tarreaucde79022019-04-12 18:03:41 +0200638 if (!lrq) {
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200639 lrq = eb32sc_lookup_ge(&tt->rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreaucde79022019-04-12 18:03:41 +0200640 if (unlikely(!lrq))
Willy Tarreau4c1e1ad2019-09-24 07:19:08 +0200641 lrq = eb32sc_first(&tt->rqueue, tid_bit);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100642 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100643
Willy Tarreaucde79022019-04-12 18:03:41 +0200644 if (!lrq && !grq)
645 break;
646
647 if (likely(!grq || (lrq && (int)(lrq->key - grq->key) <= 0))) {
648 t = eb32sc_entry(lrq, struct task, rq);
649 lrq = eb32sc_next(lrq, tid_bit);
650 __task_unlink_rq(t);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200651 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200652#ifdef USE_THREAD
Willy Tarreaucde79022019-04-12 18:03:41 +0200653 else {
654 t = eb32sc_entry(grq, struct task, rq);
655 grq = eb32sc_next(grq, tid_bit);
656 __task_unlink_rq(t);
657 if (unlikely(!grq)) {
658 grq = eb32sc_first(&rqueue, tid_bit);
659 if (!grq) {
Olivier Houchardde82aea2019-04-17 19:10:22 +0200660 global_tasks_mask &= ~tid_bit;
Willy Tarreaucde79022019-04-12 18:03:41 +0200661 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaucde79022019-04-12 18:03:41 +0200662 }
663 }
Emeric Brun01948972017-03-30 15:37:25 +0200664 }
Willy Tarreau3466e3c2019-04-15 18:52:40 +0200665#endif
Willy Tarreaucde79022019-04-12 18:03:41 +0200666
Olivier Houchardff1e9f32019-09-20 17:18:35 +0200667 /* Make sure the entry doesn't appear to be in a list */
Olivier Houchard06910462019-10-11 16:35:01 +0200668 LIST_INIT(&((struct tasklet *)t)->list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200669 /* And add it to the local task list */
Willy Tarreaua62917b2020-01-30 18:37:28 +0100670 tasklet_insert_into_tasklet_list(&tt->tasklets[TL_NORMAL], (struct tasklet *)t);
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200671 tt->tl_class_mask |= 1 << TL_NORMAL;
Willy Tarreaud99177f2020-06-30 11:48:48 +0200672 _HA_ATOMIC_ADD(&tt->task_list_size, 1);
Willy Tarreaubc13bec2019-04-30 14:55:18 +0200673 activity[tid].tasksw++;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200674 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200675
676 /* release the rqueue lock */
677 if (grq) {
678 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
679 grq = NULL;
680 }
681
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200682 /* Merge the list of tasklets waken up by other threads to the
683 * main list.
684 */
685 tmp_list = MT_LIST_BEHEAD(&tt->shared_tasklet_list);
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200686 if (tmp_list) {
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200687 LIST_SPLICE_END_DETACHED(&tt->tasklets[TL_URGENT], (struct list *)tmp_list);
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200688 if (!LIST_ISEMPTY(&tt->tasklets[TL_URGENT]))
689 tt->tl_class_mask |= 1 << TL_URGENT;
690 }
Willy Tarreaucde79022019-04-12 18:03:41 +0200691
Willy Tarreau3ef7a192020-06-24 07:21:08 +0200692 /* execute tasklets in each queue */
Willy Tarreau59153fe2020-06-24 10:17:29 +0200693 max_processed -= run_tasks_from_lists(max);
Willy Tarreaua62917b2020-01-30 18:37:28 +0100694
Willy Tarreau5c8be272020-06-19 12:17:55 +0200695 /* some tasks may have woken other ones up */
Willy Tarreau0c0c85e2020-06-23 11:32:35 +0200696 if (max_processed > 0 && thread_has_tasks())
Willy Tarreau5c8be272020-06-19 12:17:55 +0200697 goto not_done_yet;
698
Willy Tarreau49f90bf2020-06-24 09:39:48 +0200699 if (tt->tl_class_mask)
Willy Tarreaucde79022019-04-12 18:03:41 +0200700 activity[tid].long_rq++;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200701}
702
Willy Tarreau64e60122019-07-12 08:31:17 +0200703/* create a work list array for <nbthread> threads, using tasks made of
704 * function <fct>. The context passed to the function will be the pointer to
705 * the thread's work list, which will contain a copy of argument <arg>. The
706 * wake up reason will be TASK_WOKEN_OTHER. The pointer to the work_list array
707 * is returned on success, otherwise NULL on failure.
708 */
709struct work_list *work_list_create(int nbthread,
710 struct task *(*fct)(struct task *, void *, unsigned short),
711 void *arg)
712{
713 struct work_list *wl;
714 int i;
715
716 wl = calloc(nbthread, sizeof(*wl));
717 if (!wl)
718 goto fail;
719
720 for (i = 0; i < nbthread; i++) {
Olivier Houchard859dc802019-08-08 15:47:21 +0200721 MT_LIST_INIT(&wl[i].head);
Willy Tarreau64e60122019-07-12 08:31:17 +0200722 wl[i].task = task_new(1UL << i);
723 if (!wl[i].task)
724 goto fail;
725 wl[i].task->process = fct;
726 wl[i].task->context = &wl[i];
727 wl[i].arg = arg;
728 }
729 return wl;
730
731 fail:
732 work_list_destroy(wl, nbthread);
733 return NULL;
734}
735
736/* destroy work list <work> */
737void work_list_destroy(struct work_list *work, int nbthread)
738{
739 int t;
740
741 if (!work)
742 return;
743 for (t = 0; t < nbthread; t++)
744 task_destroy(work[t].task);
745 free(work);
746}
747
William Lallemand27f3fa52018-12-06 14:05:20 +0100748/*
749 * Delete every tasks before running the master polling loop
750 */
751void mworker_cleantasks()
752{
753 struct task *t;
754 int i;
William Lallemandb5823392018-12-06 15:14:37 +0100755 struct eb32_node *tmp_wq = NULL;
756 struct eb32sc_node *tmp_rq = NULL;
William Lallemand27f3fa52018-12-06 14:05:20 +0100757
758#ifdef USE_THREAD
759 /* cleanup the global run queue */
William Lallemandb5823392018-12-06 15:14:37 +0100760 tmp_rq = eb32sc_first(&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 timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100767 tmp_wq = eb32_first(&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#endif
774 /* clean the per thread run queue */
775 for (i = 0; i < global.nbthread; i++) {
William Lallemandb5823392018-12-06 15:14:37 +0100776 tmp_rq = eb32sc_first(&task_per_thread[i].rqueue, MAX_THREADS_MASK);
777 while (tmp_rq) {
778 t = eb32sc_entry(tmp_rq, struct task, rq);
779 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200780 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100781 }
782 /* cleanup the per thread timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100783 tmp_wq = eb32_first(&task_per_thread[i].timers);
784 while (tmp_wq) {
785 t = eb32_entry(tmp_wq, struct task, wq);
786 tmp_wq = eb32_next(tmp_wq);
Olivier Houchard3f795f72019-04-17 22:51:06 +0200787 task_destroy(t);
William Lallemand27f3fa52018-12-06 14:05:20 +0100788 }
789 }
790}
791
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100792/* perform minimal intializations */
793static void init_task()
Willy Tarreau4726f532009-03-07 17:25:21 +0100794{
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200795 int i;
796
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200797#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200798 memset(&timers, 0, sizeof(timers));
Willy Tarreau4726f532009-03-07 17:25:21 +0100799 memset(&rqueue, 0, sizeof(rqueue));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200800#endif
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200801 memset(&task_per_thread, 0, sizeof(task_per_thread));
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200802 for (i = 0; i < MAX_THREADS; i++) {
Willy Tarreaua62917b2020-01-30 18:37:28 +0100803 LIST_INIT(&task_per_thread[i].tasklets[TL_URGENT]);
804 LIST_INIT(&task_per_thread[i].tasklets[TL_NORMAL]);
805 LIST_INIT(&task_per_thread[i].tasklets[TL_BULK]);
Olivier Houchard06910462019-10-11 16:35:01 +0200806 MT_LIST_INIT(&task_per_thread[i].shared_tasklet_list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200807 }
Willy Tarreau4726f532009-03-07 17:25:21 +0100808}
809
Willy Tarreaue7723bd2020-06-24 11:11:02 +0200810/* config parser for global "tune.sched.low-latency", accepts "on" or "off" */
811static int cfg_parse_tune_sched_low_latency(char **args, int section_type, struct proxy *curpx,
812 struct proxy *defpx, const char *file, int line,
813 char **err)
814{
815 if (too_many_args(1, args, err, NULL))
816 return -1;
817
818 if (strcmp(args[1], "on") == 0)
819 global.tune.options |= GTUNE_SCHED_LOW_LATENCY;
820 else if (strcmp(args[1], "off") == 0)
821 global.tune.options &= ~GTUNE_SCHED_LOW_LATENCY;
822 else {
823 memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
824 return -1;
825 }
826 return 0;
827}
828
829/* config keyword parsers */
830static struct cfg_kw_list cfg_kws = {ILH, {
831 { CFG_GLOBAL, "tune.sched.low-latency", cfg_parse_tune_sched_low_latency },
832 { 0, NULL, NULL }
833}};
834
835INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100836INITCALL0(STG_PREPARE, init_task);
837
Willy Tarreaubaaee002006-06-26 02:48:02 +0200838/*
839 * Local variables:
840 * c-indent-level: 8
841 * c-basic-offset: 8
842 * End:
843 */