blob: a3765139dc002e0591648c8ecdb83f6aed689f59 [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 Tarreau2dd0d472006-06-29 17:53:05 +020015#include <common/config.h>
Willy Tarreauc6ca1a02007-05-13 19:43:47 +020016#include <common/memory.h>
Willy Tarreau2dd0d472006-06-29 17:53:05 +020017#include <common/mini-clist.h>
Willy Tarreau96bcfd72007-04-29 10:41:56 +020018#include <common/standard.h>
Willy Tarreaua6a6a932007-04-28 22:40:08 +020019#include <common/time.h>
Willy Tarreau8d388052017-11-05 13:34:20 +010020#include <eb32sctree.h>
Willy Tarreau45cb4fb2009-10-26 21:10:04 +010021#include <eb32tree.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020022
Willy Tarreaud825eef2007-05-12 22:35:00 +020023#include <proto/proxy.h>
Willy Tarreau87b09662015-04-03 00:22:06 +020024#include <proto/stream.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020025#include <proto/task.h>
Olivier Houchard79321b92018-07-26 17:55:11 +020026#include <proto/fd.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020027
Willy Tarreau8ceae722018-11-26 11:58:30 +010028DECLARE_POOL(pool_head_task, "task", sizeof(struct task));
29DECLARE_POOL(pool_head_tasklet, "tasklet", sizeof(struct tasklet));
Willy Tarreau96bcfd72007-04-29 10:41:56 +020030
Thierry FOURNIERd6975962017-07-12 14:31:10 +020031/* This is the memory pool containing all the signal structs. These
Joseph Herlantcf92b6d2018-11-15 14:19:23 -080032 * struct are used to store each required signal between two tasks.
Thierry FOURNIERd6975962017-07-12 14:31:10 +020033 */
Willy Tarreau8ceae722018-11-26 11:58:30 +010034DECLARE_POOL(pool_head_notification, "notification", sizeof(struct notification));
Thierry FOURNIERd6975962017-07-12 14:31:10 +020035
Willy Tarreaua4613182009-03-21 18:13:21 +010036unsigned int nb_tasks = 0;
Olivier Houchard9b03c0c2018-07-26 18:45:22 +020037volatile unsigned long active_tasks_mask = 0; /* Mask of threads with active tasks */
Olivier Houchardeba0c0b2018-07-26 18:53:28 +020038volatile unsigned long global_tasks_mask = 0; /* Mask of threads with tasks in the global runqueue */
Christopher Faulet34c5cc92016-12-06 09:15:30 +010039unsigned int tasks_run_queue = 0;
40unsigned int tasks_run_queue_cur = 0; /* copy of the run queue size */
Willy Tarreauc7bdf092009-03-21 18:33:52 +010041unsigned int nb_tasks_cur = 0; /* copy of the tasks count */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010042unsigned int niced_tasks = 0; /* number of niced tasks in the run queue */
Emeric Brun01948972017-03-30 15:37:25 +020043
Willy Tarreau6d1222c2017-11-26 10:08:06 +010044THREAD_LOCAL struct task *curr_task = NULL; /* task currently running or NULL */
Olivier Houchard9b36cb42018-05-04 15:46:16 +020045THREAD_LOCAL struct eb32sc_node *rq_next = NULL; /* Next task to be potentially run */
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 */
48__decl_aligned_spinlock(wq_lock); /* spin lock related to wait queue */
Willy Tarreau964c9362007-01-07 00:38:00 +010049
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020050#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +020051struct eb_root timers; /* sorted timers tree, global */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020052struct eb_root rqueue; /* tree constituting the run queue */
Olivier Houchard77551ee2018-07-26 15:59:38 +020053int global_rqueue_size; /* Number of element sin the global runqueue */
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020054#endif
Willy Tarreaub20aa9e2018-10-15 14:52:21 +020055
Willy Tarreaue35c94a2009-03-21 10:01:42 +010056static unsigned int rqueue_ticks; /* insertion count */
Willy Tarreau8d8747a2018-10-15 16:12:48 +020057
58struct task_per_thread task_per_thread[MAX_THREADS];
Willy Tarreau9789f7b2008-06-24 08:17:16 +020059
Willy Tarreau4726f532009-03-07 17:25:21 +010060/* Puts the task <t> in run queue at a position depending on t->nice. <t> is
61 * returned. The nice value assigns boosts in 32th of the run queue size. A
Christopher Faulet34c5cc92016-12-06 09:15:30 +010062 * nice value of -1024 sets the task to -tasks_run_queue*32, while a nice value
63 * of 1024 sets the task to tasks_run_queue*32. The state flags are cleared, so
64 * the caller will have to set its flags after this call.
Willy Tarreau4726f532009-03-07 17:25:21 +010065 * The task must not already be in the run queue. If unsure, use the safer
66 * task_wakeup() function.
Willy Tarreau91e99932008-06-30 07:51:00 +020067 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020068void __task_wakeup(struct task *t, struct eb_root *root)
Willy Tarreaue33aece2007-04-30 13:15:14 +020069{
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020070 void *expected = NULL;
71 int *rq_size;
Willy Tarreau85d9b842018-07-27 17:14:41 +020072 unsigned long __maybe_unused old_active_mask;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020073
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020074#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020075 if (root == &rqueue) {
76 rq_size = &global_rqueue_size;
77 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020078 } else
79#endif
80 {
Willy Tarreau8d8747a2018-10-15 16:12:48 +020081 int nb = ((void *)root - (void *)&task_per_thread[0].rqueue) / sizeof(task_per_thread[0]);
82 rq_size = &task_per_thread[nb].rqueue_size;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020083 }
84 /* Make sure if the task isn't in the runqueue, nobody inserts it
85 * in the meanwhile.
86 */
87redo:
Olivier Houchard4c2832852019-03-08 18:48:47 +010088 if (unlikely(!_HA_ATOMIC_CAS(&t->rq.node.leaf_p, &expected, (void *)0x1))) {
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020089#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020090 if (root == &rqueue)
91 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020092#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020093 return;
94 }
95 /* There's a small race condition, when running a task, the thread
96 * first sets TASK_RUNNING, and then unlink the task.
97 * If an another thread calls task_wakeup() for the same task,
98 * it may set t->state before TASK_RUNNING was set, and then try
99 * to set t->rq.nod.leaf_p after it was unlinked.
100 * To make sure it is not a problem, we check if TASK_RUNNING is set
101 * again. If it is, we unset t->rq.node.leaf_p.
102 * We then check for TASK_RUNNING a third time. If it is still there,
103 * then we can give up, the task will be re-queued later if it needs
104 * to be. If it's not there, and there is still something in t->state,
105 * then we have to requeue.
106 */
107 if (((volatile unsigned short)(t->state)) & TASK_RUNNING) {
108 unsigned short state;
109 t->rq.node.leaf_p = NULL;
Willy Tarreau155acff2019-02-04 10:26:53 +0100110 __ha_barrier_full();
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200111
112 state = (volatile unsigned short)(t->state);
113 if (unlikely(state != 0 && !(state & TASK_RUNNING)))
114 goto redo;
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200115#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200116 if (root == &rqueue)
117 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200118#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200119 return;
120 }
Olivier Houchard4c2832852019-03-08 18:48:47 +0100121 _HA_ATOMIC_ADD(&tasks_run_queue, 1);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200122#ifdef USE_THREAD
123 if (root == &rqueue) {
Olivier Houchard4c2832852019-03-08 18:48:47 +0100124 _HA_ATOMIC_OR(&global_tasks_mask, t->thread_mask);
Olivier Houchardd2b5d162019-03-08 13:47:21 +0100125 __ha_barrier_atomic_store();
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200126 }
127#endif
Olivier Houchard79321b92018-07-26 17:55:11 +0200128 old_active_mask = active_tasks_mask;
Olivier Houchard4c2832852019-03-08 18:48:47 +0100129 _HA_ATOMIC_OR(&active_tasks_mask, t->thread_mask);
130 t->rq.key = _HA_ATOMIC_ADD(&rqueue_ticks, 1);
Willy Tarreau91e99932008-06-30 07:51:00 +0200131
132 if (likely(t->nice)) {
133 int offset;
134
Olivier Houchard4c2832852019-03-08 18:48:47 +0100135 _HA_ATOMIC_ADD(&niced_tasks, 1);
Willy Tarreau91e99932008-06-30 07:51:00 +0200136 if (likely(t->nice > 0))
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200137 offset = (unsigned)((*rq_size * (unsigned int)t->nice) / 32U);
Willy Tarreau91e99932008-06-30 07:51:00 +0200138 else
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200139 offset = -(unsigned)((*rq_size * (unsigned int)-t->nice) / 32U);
Willy Tarreau4726f532009-03-07 17:25:21 +0100140 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +0200141 }
142
Willy Tarreau9efd7452018-05-31 14:48:54 +0200143 if (profiling & HA_PROF_TASKS)
144 t->call_date = now_mono_time();
145
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200146 eb32sc_insert(root, &t->rq, t->thread_mask);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200147#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200148 if (root == &rqueue) {
149 global_rqueue_size++;
Olivier Houchard4c2832852019-03-08 18:48:47 +0100150 _HA_ATOMIC_OR(&t->state, TASK_GLOBAL);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200151 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200152 } else
153#endif
154 {
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200155 int nb = ((void *)root - (void *)&task_per_thread[0].rqueue) / sizeof(task_per_thread[0]);
156 task_per_thread[nb].rqueue_size++;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200157 }
Willy Tarreau85d9b842018-07-27 17:14:41 +0200158#ifdef USE_THREAD
Olivier Houchard79321b92018-07-26 17:55:11 +0200159 /* If all threads that are supposed to handle this task are sleeping,
160 * wake one.
161 */
162 if ((((t->thread_mask & all_threads_mask) & sleeping_thread_mask) ==
163 (t->thread_mask & all_threads_mask)) &&
164 !(t->thread_mask & old_active_mask))
165 wake_thread(my_ffsl((t->thread_mask & all_threads_mask) &~ tid_bit) - 1);
Willy Tarreau85d9b842018-07-27 17:14:41 +0200166#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200167 return;
Willy Tarreaue33aece2007-04-30 13:15:14 +0200168}
Willy Tarreaud825eef2007-05-12 22:35:00 +0200169
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200170/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100171 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200172 *
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200173 * Inserts a task into wait queue <wq> at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +0100174 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100175 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100176 * Last, tasks must not be queued further than the end of the tree, which is
177 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100178 *
179 * This function should not be used directly, it is meant to be called by the
180 * inline version of task_queue() which performs a few cheap preliminary tests
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200181 * before deciding to call __task_queue(). Moreover this function doesn't care
182 * at all about locking so the caller must be careful when deciding whether to
183 * lock or not around this call.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200184 */
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200185void __task_queue(struct task *task, struct eb_root *wq)
Willy Tarreau964c9362007-01-07 00:38:00 +0100186{
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100187 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +0100188 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +0100189
190 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100191 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200192#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100193 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +0200194 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +0100195 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200196#endif
Willy Tarreauce44f122008-07-05 18:16:19 +0200197
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200198 eb32_insert(wq, &task->wq);
Willy Tarreau964c9362007-01-07 00:38:00 +0100199}
200
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200201/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200202 * Extract all expired timers from the timer queue, and wakes up all
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100203 * associated tasks. Returns the date of next event (or eternity).
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200204 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100205int wake_expired_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200206{
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200207 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200208 struct eb32_node *eb;
Willy Tarreaub992ba12017-11-05 19:09:27 +0100209 int ret = TICK_ETERNITY;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200210
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100211 while (1) {
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200212 lookup_next_local:
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200213 eb = eb32_lookup_ge(&task_per_thread[tid].timers, now_ms - TIMER_LOOK_BACK);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200214 if (!eb) {
215 /* we might have reached the end of the tree, typically because
216 * <now_ms> is in the first half and we're first scanning the last
217 * half. Let's loop back to the beginning of the tree now.
218 */
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200219 eb = eb32_first(&task_per_thread[tid].timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200220 if (likely(!eb))
221 break;
222 }
223
224 if (tick_is_lt(now_ms, eb->key)) {
225 /* timer not expired yet, revisit it later */
226 ret = eb->key;
227 break;
228 }
229
230 /* timer looks expired, detach it from the queue */
231 task = eb32_entry(eb, struct task, wq);
232 __task_unlink_wq(task);
233
234 /* It is possible that this task was left at an earlier place in the
235 * tree because a recent call to task_queue() has not moved it. This
236 * happens when the new expiration date is later than the old one.
237 * Since it is very unlikely that we reach a timeout anyway, it's a
238 * lot cheaper to proceed like this because we almost never update
239 * the tree. We may also find disabled expiration dates there. Since
240 * we have detached the task from the tree, we simply call task_queue
241 * to take care of this. Note that we might occasionally requeue it at
242 * the same place, before <eb>, so we have to check if this happens,
243 * and adjust <eb>, otherwise we may skip it which is not what we want.
244 * We may also not requeue the task (and not point eb at it) if its
245 * expiration time is not set.
246 */
247 if (!tick_is_expired(task->expire, now_ms)) {
248 if (tick_isset(task->expire))
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200249 __task_queue(task, &task_per_thread[tid].timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200250 goto lookup_next_local;
251 }
252 task_wakeup(task, TASK_WOKEN_TIMER);
253 }
254
255#ifdef USE_THREAD
256 while (1) {
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100257 HA_SPIN_LOCK(TASK_WQ_LOCK, &wq_lock);
Emeric Brunc60def82017-09-27 14:59:38 +0200258 lookup_next:
Emeric Brun01948972017-03-30 15:37:25 +0200259 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200260 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100261 /* we might have reached the end of the tree, typically because
262 * <now_ms> is in the first half and we're first scanning the last
263 * half. Let's loop back to the beginning of the tree now.
264 */
265 eb = eb32_first(&timers);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100266 if (likely(!eb))
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100267 break;
268 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200269
Willy Tarreaub992ba12017-11-05 19:09:27 +0100270 if (tick_is_lt(now_ms, eb->key)) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100271 /* timer not expired yet, revisit it later */
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200272 ret = tick_first(ret, eb->key);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100273 break;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100274 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200275
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100276 /* timer looks expired, detach it from the queue */
277 task = eb32_entry(eb, struct task, wq);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100278 __task_unlink_wq(task);
Willy Tarreau41365222009-03-08 07:46:27 +0100279
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100280 /* It is possible that this task was left at an earlier place in the
281 * tree because a recent call to task_queue() has not moved it. This
282 * happens when the new expiration date is later than the old one.
283 * Since it is very unlikely that we reach a timeout anyway, it's a
284 * lot cheaper to proceed like this because we almost never update
285 * the tree. We may also find disabled expiration dates there. Since
286 * we have detached the task from the tree, we simply call task_queue
Willy Tarreau814c9782009-07-14 23:48:55 +0200287 * to take care of this. Note that we might occasionally requeue it at
288 * the same place, before <eb>, so we have to check if this happens,
289 * and adjust <eb>, otherwise we may skip it which is not what we want.
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200290 * We may also not requeue the task (and not point eb at it) if its
291 * expiration time is not set.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100292 */
293 if (!tick_is_expired(task->expire, now_ms)) {
Willy Tarreaub992ba12017-11-05 19:09:27 +0100294 if (tick_isset(task->expire))
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200295 __task_queue(task, &timers);
Emeric Brunc60def82017-09-27 14:59:38 +0200296 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200297 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100298 task_wakeup(task, TASK_WOKEN_TIMER);
Willy Tarreau51753452017-11-23 18:36:50 +0100299 HA_SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100300 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200301
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100302 HA_SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200303#endif
Willy Tarreaub992ba12017-11-05 19:09:27 +0100304 return ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200305}
306
Willy Tarreau58b458d2008-06-29 22:40:23 +0200307/* The run queue is chronologically sorted in a tree. An insertion counter is
308 * used to assign a position to each task. This counter may be combined with
309 * other variables (eg: nice value) to set the final position in the tree. The
310 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100311 * tasks processed to 200 in any case, so that general latency remains low and
312 * so that task positions have a chance to be considered.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200313 *
314 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200315 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100316void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200317{
Willy Tarreau964c9362007-01-07 00:38:00 +0100318 struct task *t;
Emeric Brun01948972017-03-30 15:37:25 +0200319 int max_processed;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100320
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100321 tasks_run_queue_cur = tasks_run_queue; /* keep a copy for reporting */
Willy Tarreauc7bdf092009-03-21 18:33:52 +0100322 nb_tasks_cur = nb_tasks;
Olivier Houchard1599b802018-05-24 18:59:04 +0200323 max_processed = global.tune.runqueue_depth;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200324
Willy Tarreau1ee55fd2018-12-14 15:49:45 +0100325 if (likely(global_tasks_mask & tid_bit)) {
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200326 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100327 if (!(active_tasks_mask & tid_bit)) {
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200328 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100329 activity[tid].empty_rq++;
Christopher Faulet8a48f672017-11-14 10:38:36 +0100330 return;
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100331 }
Christopher Faulet8a48f672017-11-14 10:38:36 +0100332
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200333#ifdef USE_THREAD
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200334 /* Get some elements from the global run queue and put it in the
335 * local run queue. To try to keep a bit of fairness, just get as
336 * much elements from the global list as to have a bigger local queue
337 * than the average.
338 */
Willy Tarreau0b25d5e2018-10-10 16:39:22 +0200339 rq_next = eb32sc_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200340 while ((task_per_thread[tid].task_list_size + task_per_thread[tid].rqueue_size) * global.nbthread <= tasks_run_queue) {
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200341 if (unlikely(!rq_next)) {
342 /* either we just started or we reached the end
343 * of the tree, typically because <rqueue_ticks>
344 * is in the first half and we're first scanning
345 * the last half. Let's loop back to the beginning
346 * of the tree now.
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100347 */
348 rq_next = eb32sc_first(&rqueue, tid_bit);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200349 if (!rq_next) {
Olivier Houchard4c2832852019-03-08 18:48:47 +0100350 _HA_ATOMIC_AND(&global_tasks_mask, ~tid_bit);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100351 break;
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200352 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100353 }
354
355 t = eb32sc_entry(rq_next, struct task, rq);
356 rq_next = eb32sc_next(rq_next, tid_bit);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200357
358 /* detach the task from the queue */
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100359 __task_unlink_rq(t);
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200360 __task_wakeup(t, &task_per_thread[tid].rqueue);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100361 }
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200362#endif
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100363
Christopher Faulet8a48f672017-11-14 10:38:36 +0100364 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200365 } else {
366 if (!(active_tasks_mask & tid_bit)) {
367 activity[tid].empty_rq++;
368 return;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200369 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200370 }
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200371 /* Get some tasks from the run queue, make sure we don't
372 * get too much in the task list, but put a bit more than
373 * the max that will be run, to give a bit more fairness
374 */
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200375 rq_next = eb32sc_lookup_ge(&task_per_thread[tid].rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
376 while (max_processed + (max_processed / 10) > task_per_thread[tid].task_list_size) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100377 /* Note: this loop is one of the fastest code path in
378 * the whole program. It should not be re-arranged
379 * without a good reason.
380 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200381 if (unlikely(!rq_next)) {
382 /* either we just started or we reached the end
383 * of the tree, typically because <rqueue_ticks>
384 * is in the first half and we're first scanning
385 * the last half. Let's loop back to the beginning
386 * of the tree now.
Emeric Brun01948972017-03-30 15:37:25 +0200387 */
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200388 rq_next = eb32sc_first(&task_per_thread[tid].rqueue, tid_bit);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200389 if (!rq_next)
390 break;
Emeric Brun01948972017-03-30 15:37:25 +0200391 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200392 t = eb32sc_entry(rq_next, struct task, rq);
393 rq_next = eb32sc_next(rq_next, tid_bit);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200394 /* Make sure nobody re-adds the task in the runqueue */
Olivier Houchard4c2832852019-03-08 18:48:47 +0100395 _HA_ATOMIC_OR(&t->state, TASK_RUNNING);
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100396
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200397 /* detach the task from the queue */
398 __task_unlink_rq(t);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200399 /* And add it to the local task list */
400 task_insert_into_tasklet_list(t);
401 }
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200402 if (!(global_tasks_mask & tid_bit) && task_per_thread[tid].rqueue_size == 0) {
Olivier Houchard4c2832852019-03-08 18:48:47 +0100403 _HA_ATOMIC_AND(&active_tasks_mask, ~tid_bit);
Olivier Houchardd2b5d162019-03-08 13:47:21 +0100404 __ha_barrier_atomic_load();
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200405 if (global_tasks_mask & tid_bit)
Olivier Houchard4c2832852019-03-08 18:48:47 +0100406 _HA_ATOMIC_OR(&active_tasks_mask, tid_bit);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200407 }
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200408 while (max_processed > 0 && !LIST_ISEMPTY(&task_per_thread[tid].task_list)) {
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200409 struct task *t;
410 unsigned short state;
411 void *ctx;
412 struct task *(*process)(struct task *t, void *ctx, unsigned short state);
413
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200414 t = (struct task *)LIST_ELEM(task_per_thread[tid].task_list.n, struct tasklet *, list);
Olivier Houchard4c2832852019-03-08 18:48:47 +0100415 state = _HA_ATOMIC_XCHG(&t->state, TASK_RUNNING);
Olivier Houchardd2b5d162019-03-08 13:47:21 +0100416 __ha_barrier_atomic_store();
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200417 task_remove_from_task_list(t);
418
419 ctx = t->context;
420 process = t->process;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200421 t->calls++;
Willy Tarreau9efd7452018-05-31 14:48:54 +0200422
423 if (unlikely(!TASK_IS_TASKLET(t) && t->call_date)) {
424 uint64_t now_ns = now_mono_time();
425
426 t->lat_time += now_ns - t->call_date;
427 t->call_date = now_ns;
428 }
429
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200430 curr_task = (struct task *)t;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200431 if (likely(process == process_stream))
432 t = process_stream(t, ctx, state);
433 else {
434 if (t->process != NULL)
Olivier Houchard9db0fed2018-06-14 15:40:47 +0200435 t = process(TASK_IS_TASKLET(t) ? NULL : t, ctx, state);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200436 else {
437 __task_free(t);
438 t = NULL;
439 }
440 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200441 curr_task = NULL;
442 /* If there is a pending state we have to wake up the task
Joseph Herlantcf92b6d2018-11-15 14:19:23 -0800443 * immediately, else we defer it into wait queue
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200444 */
445 if (t != NULL) {
Willy Tarreau9efd7452018-05-31 14:48:54 +0200446 if (unlikely(!TASK_IS_TASKLET(t) && t->call_date)) {
447 t->cpu_time += now_mono_time() - t->call_date;
448 t->call_date = 0;
449 }
450
Olivier Houchard4c2832852019-03-08 18:48:47 +0100451 state = _HA_ATOMIC_AND(&t->state, ~TASK_RUNNING);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200452 if (state)
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200453#ifdef USE_THREAD
Olivier Houchard19bdf242018-08-17 13:36:08 +0200454 __task_wakeup(t, ((t->thread_mask & all_threads_mask) == tid_bit) ?
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200455 &task_per_thread[tid].rqueue : &rqueue);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200456#else
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200457 __task_wakeup(t, &task_per_thread[tid].rqueue);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200458#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200459 else
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100460 task_queue(t);
Willy Tarreau58b458d2008-06-29 22:40:23 +0200461 }
Willy Tarreauce4e0aa2017-11-05 23:57:00 +0100462
Olivier Houchard736ea412018-05-28 14:54:49 +0200463 max_processed--;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100464 if (max_processed <= 0) {
Olivier Houchard4c2832852019-03-08 18:48:47 +0100465 _HA_ATOMIC_OR(&active_tasks_mask, tid_bit);
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100466 activity[tid].long_rq++;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100467 break;
468 }
469 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200470}
471
William Lallemand27f3fa52018-12-06 14:05:20 +0100472/*
473 * Delete every tasks before running the master polling loop
474 */
475void mworker_cleantasks()
476{
477 struct task *t;
478 int i;
William Lallemandb5823392018-12-06 15:14:37 +0100479 struct eb32_node *tmp_wq = NULL;
480 struct eb32sc_node *tmp_rq = NULL;
William Lallemand27f3fa52018-12-06 14:05:20 +0100481
482#ifdef USE_THREAD
483 /* cleanup the global run queue */
William Lallemandb5823392018-12-06 15:14:37 +0100484 tmp_rq = eb32sc_first(&rqueue, MAX_THREADS_MASK);
485 while (tmp_rq) {
486 t = eb32sc_entry(tmp_rq, struct task, rq);
487 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
William Lallemand27f3fa52018-12-06 14:05:20 +0100488 task_delete(t);
489 task_free(t);
490 }
491 /* cleanup the timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100492 tmp_wq = eb32_first(&timers);
493 while (tmp_wq) {
494 t = eb32_entry(tmp_wq, struct task, wq);
495 tmp_wq = eb32_next(tmp_wq);
William Lallemand27f3fa52018-12-06 14:05:20 +0100496 task_delete(t);
497 task_free(t);
498 }
499#endif
500 /* clean the per thread run queue */
501 for (i = 0; i < global.nbthread; i++) {
William Lallemandb5823392018-12-06 15:14:37 +0100502 tmp_rq = eb32sc_first(&task_per_thread[i].rqueue, MAX_THREADS_MASK);
503 while (tmp_rq) {
504 t = eb32sc_entry(tmp_rq, struct task, rq);
505 tmp_rq = eb32sc_next(tmp_rq, MAX_THREADS_MASK);
William Lallemand27f3fa52018-12-06 14:05:20 +0100506 task_delete(t);
507 task_free(t);
508 }
509 /* cleanup the per thread timers queue */
William Lallemandb5823392018-12-06 15:14:37 +0100510 tmp_wq = eb32_first(&task_per_thread[i].timers);
511 while (tmp_wq) {
512 t = eb32_entry(tmp_wq, struct task, wq);
513 tmp_wq = eb32_next(tmp_wq);
William Lallemand27f3fa52018-12-06 14:05:20 +0100514 task_delete(t);
515 task_free(t);
516 }
517 }
518}
519
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100520/* perform minimal intializations */
521static void init_task()
Willy Tarreau4726f532009-03-07 17:25:21 +0100522{
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200523 int i;
524
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200525#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200526 memset(&timers, 0, sizeof(timers));
Willy Tarreau4726f532009-03-07 17:25:21 +0100527 memset(&rqueue, 0, sizeof(rqueue));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200528#endif
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200529 memset(&task_per_thread, 0, sizeof(task_per_thread));
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200530 for (i = 0; i < MAX_THREADS; i++) {
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200531 LIST_INIT(&task_per_thread[i].task_list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200532 }
Willy Tarreau4726f532009-03-07 17:25:21 +0100533}
534
Willy Tarreaub6b3df32018-11-26 16:31:20 +0100535INITCALL0(STG_PREPARE, init_task);
536
Willy Tarreaubaaee002006-06-26 02:48:02 +0200537/*
538 * Local variables:
539 * c-indent-level: 8
540 * c-basic-offset: 8
541 * End:
542 */