blob: 779dbf5d8890006adc416c09670e1d668971eaff [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 Tarreaubafbe012017-11-24 17:34:44 +010028struct pool_head *pool_head_task;
Olivier Houchardb0bdae72018-05-18 18:45:28 +020029struct pool_head *pool_head_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 Tarreaubafbe012017-11-24 17:34:44 +010034struct pool_head *pool_head_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 Tarreaua24d1d02017-11-26 10:19:16 +010047__decl_hathreads(HA_SPINLOCK_T __attribute__((aligned(64))) rq_lock); /* spin lock related to run queue */
48__decl_hathreads(HA_SPINLOCK_T __attribute__((aligned(64))) 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:
David Carliercc0a9572018-06-05 10:41:03 +000088 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;
110 __ha_barrier_store();
111
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 }
121 HA_ATOMIC_ADD(&tasks_run_queue, 1);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200122#ifdef USE_THREAD
123 if (root == &rqueue) {
124 HA_ATOMIC_OR(&global_tasks_mask, t->thread_mask);
125 __ha_barrier_store();
126 }
127#endif
Olivier Houchard79321b92018-07-26 17:55:11 +0200128 old_active_mask = active_tasks_mask;
Willy Tarreau189ea852018-07-26 15:16:43 +0200129 HA_ATOMIC_OR(&active_tasks_mask, t->thread_mask);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200130 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 Houchardf6e6dc12018-05-18 18:38:23 +0200135 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
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200143 eb32sc_insert(root, &t->rq, t->thread_mask);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200144#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200145 if (root == &rqueue) {
146 global_rqueue_size++;
Olivier Houchard76e45182018-07-26 16:19:58 +0200147 HA_ATOMIC_OR(&t->state, TASK_GLOBAL);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200148 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200149 } else
150#endif
151 {
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200152 int nb = ((void *)root - (void *)&task_per_thread[0].rqueue) / sizeof(task_per_thread[0]);
153 task_per_thread[nb].rqueue_size++;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200154 }
Willy Tarreau85d9b842018-07-27 17:14:41 +0200155#ifdef USE_THREAD
Olivier Houchard79321b92018-07-26 17:55:11 +0200156 /* If all threads that are supposed to handle this task are sleeping,
157 * wake one.
158 */
159 if ((((t->thread_mask & all_threads_mask) & sleeping_thread_mask) ==
160 (t->thread_mask & all_threads_mask)) &&
161 !(t->thread_mask & old_active_mask))
162 wake_thread(my_ffsl((t->thread_mask & all_threads_mask) &~ tid_bit) - 1);
Willy Tarreau85d9b842018-07-27 17:14:41 +0200163#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200164 return;
Willy Tarreaue33aece2007-04-30 13:15:14 +0200165}
Willy Tarreaud825eef2007-05-12 22:35:00 +0200166
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200167/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100168 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200169 *
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200170 * Inserts a task into wait queue <wq> at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +0100171 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100172 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100173 * Last, tasks must not be queued further than the end of the tree, which is
174 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100175 *
176 * This function should not be used directly, it is meant to be called by the
177 * inline version of task_queue() which performs a few cheap preliminary tests
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200178 * before deciding to call __task_queue(). Moreover this function doesn't care
179 * at all about locking so the caller must be careful when deciding whether to
180 * lock or not around this call.
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200181 */
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200182void __task_queue(struct task *task, struct eb_root *wq)
Willy Tarreau964c9362007-01-07 00:38:00 +0100183{
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100184 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +0100185 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +0100186
187 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100188 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200189#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100190 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +0200191 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +0100192 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200193#endif
Willy Tarreauce44f122008-07-05 18:16:19 +0200194
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200195 eb32_insert(wq, &task->wq);
Willy Tarreau964c9362007-01-07 00:38:00 +0100196}
197
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200198/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200199 * Extract all expired timers from the timer queue, and wakes up all
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100200 * associated tasks. Returns the date of next event (or eternity).
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200201 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100202int wake_expired_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200203{
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200204 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200205 struct eb32_node *eb;
Willy Tarreaub992ba12017-11-05 19:09:27 +0100206 int ret = TICK_ETERNITY;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200207
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100208 while (1) {
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200209 lookup_next_local:
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200210 eb = eb32_lookup_ge(&task_per_thread[tid].timers, now_ms - TIMER_LOOK_BACK);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200211 if (!eb) {
212 /* we might have reached the end of the tree, typically because
213 * <now_ms> is in the first half and we're first scanning the last
214 * half. Let's loop back to the beginning of the tree now.
215 */
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200216 eb = eb32_first(&task_per_thread[tid].timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200217 if (likely(!eb))
218 break;
219 }
220
221 if (tick_is_lt(now_ms, eb->key)) {
222 /* timer not expired yet, revisit it later */
223 ret = eb->key;
224 break;
225 }
226
227 /* timer looks expired, detach it from the queue */
228 task = eb32_entry(eb, struct task, wq);
229 __task_unlink_wq(task);
230
231 /* It is possible that this task was left at an earlier place in the
232 * tree because a recent call to task_queue() has not moved it. This
233 * happens when the new expiration date is later than the old one.
234 * Since it is very unlikely that we reach a timeout anyway, it's a
235 * lot cheaper to proceed like this because we almost never update
236 * the tree. We may also find disabled expiration dates there. Since
237 * we have detached the task from the tree, we simply call task_queue
238 * to take care of this. Note that we might occasionally requeue it at
239 * the same place, before <eb>, so we have to check if this happens,
240 * and adjust <eb>, otherwise we may skip it which is not what we want.
241 * We may also not requeue the task (and not point eb at it) if its
242 * expiration time is not set.
243 */
244 if (!tick_is_expired(task->expire, now_ms)) {
245 if (tick_isset(task->expire))
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200246 __task_queue(task, &task_per_thread[tid].timers);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200247 goto lookup_next_local;
248 }
249 task_wakeup(task, TASK_WOKEN_TIMER);
250 }
251
252#ifdef USE_THREAD
253 while (1) {
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100254 HA_SPIN_LOCK(TASK_WQ_LOCK, &wq_lock);
Emeric Brunc60def82017-09-27 14:59:38 +0200255 lookup_next:
Emeric Brun01948972017-03-30 15:37:25 +0200256 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200257 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100258 /* we might have reached the end of the tree, typically because
259 * <now_ms> is in the first half and we're first scanning the last
260 * half. Let's loop back to the beginning of the tree now.
261 */
262 eb = eb32_first(&timers);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100263 if (likely(!eb))
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100264 break;
265 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200266
Willy Tarreaub992ba12017-11-05 19:09:27 +0100267 if (tick_is_lt(now_ms, eb->key)) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100268 /* timer not expired yet, revisit it later */
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200269 ret = tick_first(ret, eb->key);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100270 break;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100271 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200272
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100273 /* timer looks expired, detach it from the queue */
274 task = eb32_entry(eb, struct task, wq);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100275 __task_unlink_wq(task);
Willy Tarreau41365222009-03-08 07:46:27 +0100276
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100277 /* It is possible that this task was left at an earlier place in the
278 * tree because a recent call to task_queue() has not moved it. This
279 * happens when the new expiration date is later than the old one.
280 * Since it is very unlikely that we reach a timeout anyway, it's a
281 * lot cheaper to proceed like this because we almost never update
282 * the tree. We may also find disabled expiration dates there. Since
283 * we have detached the task from the tree, we simply call task_queue
Willy Tarreau814c9782009-07-14 23:48:55 +0200284 * to take care of this. Note that we might occasionally requeue it at
285 * the same place, before <eb>, so we have to check if this happens,
286 * and adjust <eb>, otherwise we may skip it which is not what we want.
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200287 * We may also not requeue the task (and not point eb at it) if its
288 * expiration time is not set.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100289 */
290 if (!tick_is_expired(task->expire, now_ms)) {
Willy Tarreaub992ba12017-11-05 19:09:27 +0100291 if (tick_isset(task->expire))
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200292 __task_queue(task, &timers);
Emeric Brunc60def82017-09-27 14:59:38 +0200293 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200294 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100295 task_wakeup(task, TASK_WOKEN_TIMER);
Willy Tarreau51753452017-11-23 18:36:50 +0100296 HA_SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100297 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200298
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100299 HA_SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200300#endif
Willy Tarreaub992ba12017-11-05 19:09:27 +0100301 return ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200302}
303
Willy Tarreau58b458d2008-06-29 22:40:23 +0200304/* The run queue is chronologically sorted in a tree. An insertion counter is
305 * used to assign a position to each task. This counter may be combined with
306 * other variables (eg: nice value) to set the final position in the tree. The
307 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100308 * tasks processed to 200 in any case, so that general latency remains low and
309 * so that task positions have a chance to be considered.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200310 *
311 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200312 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100313void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200314{
Willy Tarreau964c9362007-01-07 00:38:00 +0100315 struct task *t;
Emeric Brun01948972017-03-30 15:37:25 +0200316 int max_processed;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100317
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100318 tasks_run_queue_cur = tasks_run_queue; /* keep a copy for reporting */
Willy Tarreauc7bdf092009-03-21 18:33:52 +0100319 nb_tasks_cur = nb_tasks;
Olivier Houchard1599b802018-05-24 18:59:04 +0200320 max_processed = global.tune.runqueue_depth;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200321
322 if (likely(global.nbthread > 1)) {
323 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100324 if (!(active_tasks_mask & tid_bit)) {
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200325 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100326 activity[tid].empty_rq++;
Christopher Faulet8a48f672017-11-14 10:38:36 +0100327 return;
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100328 }
Christopher Faulet8a48f672017-11-14 10:38:36 +0100329
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200330#ifdef USE_THREAD
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200331 /* Get some elements from the global run queue and put it in the
332 * local run queue. To try to keep a bit of fairness, just get as
333 * much elements from the global list as to have a bigger local queue
334 * than the average.
335 */
Willy Tarreau0b25d5e2018-10-10 16:39:22 +0200336 rq_next = eb32sc_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200337 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 +0200338 if (unlikely(!rq_next)) {
339 /* either we just started or we reached the end
340 * of the tree, typically because <rqueue_ticks>
341 * is in the first half and we're first scanning
342 * the last half. Let's loop back to the beginning
343 * of the tree now.
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100344 */
345 rq_next = eb32sc_first(&rqueue, tid_bit);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200346 if (!rq_next) {
347 HA_ATOMIC_AND(&global_tasks_mask, ~tid_bit);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100348 break;
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200349 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100350 }
351
352 t = eb32sc_entry(rq_next, struct task, rq);
353 rq_next = eb32sc_next(rq_next, tid_bit);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200354
355 /* detach the task from the queue */
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100356 __task_unlink_rq(t);
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200357 __task_wakeup(t, &task_per_thread[tid].rqueue);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100358 }
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200359#endif
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100360
Christopher Faulet8a48f672017-11-14 10:38:36 +0100361 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200362 } else {
363 if (!(active_tasks_mask & tid_bit)) {
364 activity[tid].empty_rq++;
365 return;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200366 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200367 }
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200368 /* Get some tasks from the run queue, make sure we don't
369 * get too much in the task list, but put a bit more than
370 * the max that will be run, to give a bit more fairness
371 */
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200372 rq_next = eb32sc_lookup_ge(&task_per_thread[tid].rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
373 while (max_processed + (max_processed / 10) > task_per_thread[tid].task_list_size) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100374 /* Note: this loop is one of the fastest code path in
375 * the whole program. It should not be re-arranged
376 * without a good reason.
377 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200378 if (unlikely(!rq_next)) {
379 /* either we just started or we reached the end
380 * of the tree, typically because <rqueue_ticks>
381 * is in the first half and we're first scanning
382 * the last half. Let's loop back to the beginning
383 * of the tree now.
Emeric Brun01948972017-03-30 15:37:25 +0200384 */
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200385 rq_next = eb32sc_first(&task_per_thread[tid].rqueue, tid_bit);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200386 if (!rq_next)
387 break;
Emeric Brun01948972017-03-30 15:37:25 +0200388 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200389 t = eb32sc_entry(rq_next, struct task, rq);
390 rq_next = eb32sc_next(rq_next, tid_bit);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200391 /* Make sure nobody re-adds the task in the runqueue */
392 HA_ATOMIC_OR(&t->state, TASK_RUNNING);
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100393
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200394 /* detach the task from the queue */
395 __task_unlink_rq(t);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200396 /* And add it to the local task list */
397 task_insert_into_tasklet_list(t);
398 }
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200399 if (!(global_tasks_mask & tid_bit) && task_per_thread[tid].rqueue_size == 0) {
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200400 HA_ATOMIC_AND(&active_tasks_mask, ~tid_bit);
401 __ha_barrier_load();
402 if (global_tasks_mask & tid_bit)
403 HA_ATOMIC_OR(&active_tasks_mask, tid_bit);
404 }
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200405 while (max_processed > 0 && !LIST_ISEMPTY(&task_per_thread[tid].task_list)) {
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200406 struct task *t;
407 unsigned short state;
408 void *ctx;
409 struct task *(*process)(struct task *t, void *ctx, unsigned short state);
410
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200411 t = (struct task *)LIST_ELEM(task_per_thread[tid].task_list.n, struct tasklet *, list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200412 state = HA_ATOMIC_XCHG(&t->state, TASK_RUNNING);
413 __ha_barrier_store();
414 task_remove_from_task_list(t);
415
416 ctx = t->context;
417 process = t->process;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200418 t->calls++;
419 curr_task = (struct task *)t;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200420 if (likely(process == process_stream))
421 t = process_stream(t, ctx, state);
422 else {
423 if (t->process != NULL)
Olivier Houchard9db0fed2018-06-14 15:40:47 +0200424 t = process(TASK_IS_TASKLET(t) ? NULL : t, ctx, state);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200425 else {
426 __task_free(t);
427 t = NULL;
428 }
429 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200430 curr_task = NULL;
431 /* If there is a pending state we have to wake up the task
Joseph Herlantcf92b6d2018-11-15 14:19:23 -0800432 * immediately, else we defer it into wait queue
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200433 */
434 if (t != NULL) {
435 state = HA_ATOMIC_AND(&t->state, ~TASK_RUNNING);
436 if (state)
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200437#ifdef USE_THREAD
Olivier Houchard19bdf242018-08-17 13:36:08 +0200438 __task_wakeup(t, ((t->thread_mask & all_threads_mask) == tid_bit) ?
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200439 &task_per_thread[tid].rqueue : &rqueue);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200440#else
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200441 __task_wakeup(t, &task_per_thread[tid].rqueue);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200442#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200443 else
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100444 task_queue(t);
Willy Tarreau58b458d2008-06-29 22:40:23 +0200445 }
Willy Tarreauce4e0aa2017-11-05 23:57:00 +0100446
Olivier Houchard736ea412018-05-28 14:54:49 +0200447 max_processed--;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100448 if (max_processed <= 0) {
Willy Tarreau189ea852018-07-26 15:16:43 +0200449 HA_ATOMIC_OR(&active_tasks_mask, tid_bit);
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100450 activity[tid].long_rq++;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100451 break;
452 }
453 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200454}
455
Willy Tarreau4726f532009-03-07 17:25:21 +0100456/* perform minimal intializations, report 0 in case of error, 1 if OK. */
457int init_task()
458{
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200459 int i;
460
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200461#ifdef USE_THREAD
Willy Tarreaub20aa9e2018-10-15 14:52:21 +0200462 memset(&timers, 0, sizeof(timers));
Willy Tarreau4726f532009-03-07 17:25:21 +0100463 memset(&rqueue, 0, sizeof(rqueue));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200464#endif
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100465 HA_SPIN_INIT(&wq_lock);
466 HA_SPIN_INIT(&rq_lock);
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200467 memset(&task_per_thread, 0, sizeof(task_per_thread));
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200468 for (i = 0; i < MAX_THREADS; i++) {
Willy Tarreau8d8747a2018-10-15 16:12:48 +0200469 LIST_INIT(&task_per_thread[i].task_list);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200470 }
Willy Tarreaubafbe012017-11-24 17:34:44 +0100471 pool_head_task = create_pool("task", sizeof(struct task), MEM_F_SHARED);
472 if (!pool_head_task)
Thierry FOURNIERd6975962017-07-12 14:31:10 +0200473 return 0;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200474 pool_head_tasklet = create_pool("tasklet", sizeof(struct tasklet), MEM_F_SHARED);
475 if (!pool_head_tasklet)
476 return 0;
Willy Tarreaubafbe012017-11-24 17:34:44 +0100477 pool_head_notification = create_pool("notification", sizeof(struct notification), MEM_F_SHARED);
478 if (!pool_head_notification)
Thierry FOURNIERd6975962017-07-12 14:31:10 +0200479 return 0;
480 return 1;
Willy Tarreau4726f532009-03-07 17:25:21 +0100481}
482
Willy Tarreaubaaee002006-06-26 02:48:02 +0200483/*
484 * Local variables:
485 * c-indent-level: 8
486 * c-basic-offset: 8
487 * End:
488 */