blob: ce5b4f907e2c1442d3c416064d71b64167b659e7 [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
32 * struct are used to store each requiered signal between two tasks.
33 */
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
Olivier Houchardb0bdae72018-05-18 18:45:28 +020047struct list task_list[MAX_THREADS]; /* List of tasks to be run, mixing tasks and tasklets */
48int task_list_size[MAX_THREADS]; /* Number of tasks in the task_list */
49
Willy Tarreaua24d1d02017-11-26 10:19:16 +010050__decl_hathreads(HA_SPINLOCK_T __attribute__((aligned(64))) rq_lock); /* spin lock related to run queue */
51__decl_hathreads(HA_SPINLOCK_T __attribute__((aligned(64))) wq_lock); /* spin lock related to wait queue */
Willy Tarreau964c9362007-01-07 00:38:00 +010052
Willy Tarreaue35c94a2009-03-21 10:01:42 +010053static struct eb_root timers; /* sorted timers tree */
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020054#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020055struct eb_root rqueue; /* tree constituting the run queue */
Olivier Houchard77551ee2018-07-26 15:59:38 +020056int global_rqueue_size; /* Number of element sin the global runqueue */
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020057#endif
58struct eb_root rqueue_local[MAX_THREADS]; /* tree constituting the per-thread run queue */
Olivier Houchard77551ee2018-07-26 15:59:38 +020059int rqueue_size[MAX_THREADS]; /* Number of elements in the per-thread run queue */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010060static unsigned int rqueue_ticks; /* insertion count */
Willy Tarreau9789f7b2008-06-24 08:17:16 +020061
Willy Tarreau4726f532009-03-07 17:25:21 +010062/* Puts the task <t> in run queue at a position depending on t->nice. <t> is
63 * returned. The nice value assigns boosts in 32th of the run queue size. A
Christopher Faulet34c5cc92016-12-06 09:15:30 +010064 * nice value of -1024 sets the task to -tasks_run_queue*32, while a nice value
65 * of 1024 sets the task to tasks_run_queue*32. The state flags are cleared, so
66 * the caller will have to set its flags after this call.
Willy Tarreau4726f532009-03-07 17:25:21 +010067 * The task must not already be in the run queue. If unsure, use the safer
68 * task_wakeup() function.
Willy Tarreau91e99932008-06-30 07:51:00 +020069 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020070void __task_wakeup(struct task *t, struct eb_root *root)
Willy Tarreaue33aece2007-04-30 13:15:14 +020071{
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020072 void *expected = NULL;
73 int *rq_size;
Willy Tarreau85d9b842018-07-27 17:14:41 +020074 unsigned long __maybe_unused old_active_mask;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020075
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020076#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020077 if (root == &rqueue) {
78 rq_size = &global_rqueue_size;
79 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020080 } else
81#endif
82 {
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020083 int nb = root - &rqueue_local[0];
84 rq_size = &rqueue_size[nb];
85 }
86 /* Make sure if the task isn't in the runqueue, nobody inserts it
87 * in the meanwhile.
88 */
89redo:
David Carliercc0a9572018-06-05 10:41:03 +000090 if (unlikely(!HA_ATOMIC_CAS(&t->rq.node.leaf_p, &expected, (void *)0x1))) {
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020091#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020092 if (root == &rqueue)
93 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020094#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020095 return;
96 }
97 /* There's a small race condition, when running a task, the thread
98 * first sets TASK_RUNNING, and then unlink the task.
99 * If an another thread calls task_wakeup() for the same task,
100 * it may set t->state before TASK_RUNNING was set, and then try
101 * to set t->rq.nod.leaf_p after it was unlinked.
102 * To make sure it is not a problem, we check if TASK_RUNNING is set
103 * again. If it is, we unset t->rq.node.leaf_p.
104 * We then check for TASK_RUNNING a third time. If it is still there,
105 * then we can give up, the task will be re-queued later if it needs
106 * to be. If it's not there, and there is still something in t->state,
107 * then we have to requeue.
108 */
109 if (((volatile unsigned short)(t->state)) & TASK_RUNNING) {
110 unsigned short state;
111 t->rq.node.leaf_p = NULL;
112 __ha_barrier_store();
113
114 state = (volatile unsigned short)(t->state);
115 if (unlikely(state != 0 && !(state & TASK_RUNNING)))
116 goto redo;
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200117#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200118 if (root == &rqueue)
119 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200120#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200121 return;
122 }
123 HA_ATOMIC_ADD(&tasks_run_queue, 1);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200124#ifdef USE_THREAD
125 if (root == &rqueue) {
126 HA_ATOMIC_OR(&global_tasks_mask, t->thread_mask);
127 __ha_barrier_store();
128 }
129#endif
Olivier Houchard79321b92018-07-26 17:55:11 +0200130 old_active_mask = active_tasks_mask;
Willy Tarreau189ea852018-07-26 15:16:43 +0200131 HA_ATOMIC_OR(&active_tasks_mask, t->thread_mask);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200132 t->rq.key = HA_ATOMIC_ADD(&rqueue_ticks, 1);
Willy Tarreau91e99932008-06-30 07:51:00 +0200133
134 if (likely(t->nice)) {
135 int offset;
136
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200137 HA_ATOMIC_ADD(&niced_tasks, 1);
Willy Tarreau91e99932008-06-30 07:51:00 +0200138 if (likely(t->nice > 0))
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200139 offset = (unsigned)((*rq_size * (unsigned int)t->nice) / 32U);
Willy Tarreau91e99932008-06-30 07:51:00 +0200140 else
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200141 offset = -(unsigned)((*rq_size * (unsigned int)-t->nice) / 32U);
Willy Tarreau4726f532009-03-07 17:25:21 +0100142 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +0200143 }
144
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200145 eb32sc_insert(root, &t->rq, t->thread_mask);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200146#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200147 if (root == &rqueue) {
148 global_rqueue_size++;
Olivier Houchard76e45182018-07-26 16:19:58 +0200149 HA_ATOMIC_OR(&t->state, TASK_GLOBAL);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200150 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200151 } else
152#endif
153 {
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200154 int nb = root - &rqueue_local[0];
155
156 rqueue_size[nb]++;
157 }
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 *
173 * Inserts a task into the wait queue 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
181 * before deciding to call __task_queue().
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200182 */
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100183void __task_queue(struct task *task)
Willy Tarreau964c9362007-01-07 00:38:00 +0100184{
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100185 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +0100186 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +0100187
188 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100189 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200190#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100191 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +0200192 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +0100193 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200194#endif
Willy Tarreauce44f122008-07-05 18:16:19 +0200195
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100196 eb32_insert(&timers, &task->wq);
SaVaGe1d7a4202009-10-06 18:53:37 +0300197
Willy Tarreau4726f532009-03-07 17:25:21 +0100198 return;
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) {
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100212 HA_SPIN_LOCK(TASK_WQ_LOCK, &wq_lock);
Emeric Brunc60def82017-09-27 14:59:38 +0200213 lookup_next:
Emeric Brun01948972017-03-30 15:37:25 +0200214 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200215 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100216 /* we might have reached the end of the tree, typically because
217 * <now_ms> is in the first half and we're first scanning the last
218 * half. Let's loop back to the beginning of the tree now.
219 */
220 eb = eb32_first(&timers);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100221 if (likely(!eb))
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100222 break;
223 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200224
Willy Tarreaub992ba12017-11-05 19:09:27 +0100225 if (tick_is_lt(now_ms, eb->key)) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100226 /* timer not expired yet, revisit it later */
Willy Tarreaub992ba12017-11-05 19:09:27 +0100227 ret = eb->key;
228 break;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100229 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200230
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100231 /* timer looks expired, detach it from the queue */
232 task = eb32_entry(eb, struct task, wq);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100233 __task_unlink_wq(task);
Willy Tarreau41365222009-03-08 07:46:27 +0100234
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100235 /* It is possible that this task was left at an earlier place in the
236 * tree because a recent call to task_queue() has not moved it. This
237 * happens when the new expiration date is later than the old one.
238 * Since it is very unlikely that we reach a timeout anyway, it's a
239 * lot cheaper to proceed like this because we almost never update
240 * the tree. We may also find disabled expiration dates there. Since
241 * we have detached the task from the tree, we simply call task_queue
Willy Tarreau814c9782009-07-14 23:48:55 +0200242 * to take care of this. Note that we might occasionally requeue it at
243 * the same place, before <eb>, so we have to check if this happens,
244 * and adjust <eb>, otherwise we may skip it which is not what we want.
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200245 * We may also not requeue the task (and not point eb at it) if its
246 * expiration time is not set.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100247 */
248 if (!tick_is_expired(task->expire, now_ms)) {
Willy Tarreaub992ba12017-11-05 19:09:27 +0100249 if (tick_isset(task->expire))
250 __task_queue(task);
Emeric Brunc60def82017-09-27 14:59:38 +0200251 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200252 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100253 task_wakeup(task, TASK_WOKEN_TIMER);
Willy Tarreau51753452017-11-23 18:36:50 +0100254 HA_SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100255 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200256
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100257 HA_SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100258 return ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200259}
260
Willy Tarreau58b458d2008-06-29 22:40:23 +0200261/* The run queue is chronologically sorted in a tree. An insertion counter is
262 * used to assign a position to each task. This counter may be combined with
263 * other variables (eg: nice value) to set the final position in the tree. The
264 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100265 * tasks processed to 200 in any case, so that general latency remains low and
266 * so that task positions have a chance to be considered.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200267 *
268 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200269 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100270void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200271{
Willy Tarreau964c9362007-01-07 00:38:00 +0100272 struct task *t;
Emeric Brun01948972017-03-30 15:37:25 +0200273 int max_processed;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100274
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100275 tasks_run_queue_cur = tasks_run_queue; /* keep a copy for reporting */
Willy Tarreauc7bdf092009-03-21 18:33:52 +0100276 nb_tasks_cur = nb_tasks;
Olivier Houchard1599b802018-05-24 18:59:04 +0200277 max_processed = global.tune.runqueue_depth;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200278
279 if (likely(global.nbthread > 1)) {
280 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100281 if (!(active_tasks_mask & tid_bit)) {
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200282 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100283 activity[tid].empty_rq++;
Christopher Faulet8a48f672017-11-14 10:38:36 +0100284 return;
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100285 }
Christopher Faulet8a48f672017-11-14 10:38:36 +0100286
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200287#ifdef USE_THREAD
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200288 /* Get some elements from the global run queue and put it in the
289 * local run queue. To try to keep a bit of fairness, just get as
290 * much elements from the global list as to have a bigger local queue
291 * than the average.
292 */
Willy Tarreau9a771862018-07-26 13:36:01 +0200293 while ((task_list_size[tid] + rqueue_size[tid]) * global.nbthread <= tasks_run_queue) {
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200294 /* we have to restart looking up after every batch */
295 rq_next = eb32sc_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
296 if (unlikely(!rq_next)) {
297 /* either we just started or we reached the end
298 * of the tree, typically because <rqueue_ticks>
299 * is in the first half and we're first scanning
300 * the last half. Let's loop back to the beginning
301 * of the tree now.
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100302 */
303 rq_next = eb32sc_first(&rqueue, tid_bit);
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200304 if (!rq_next) {
305 HA_ATOMIC_AND(&global_tasks_mask, ~tid_bit);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100306 break;
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200307 }
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100308 }
309
310 t = eb32sc_entry(rq_next, struct task, rq);
311 rq_next = eb32sc_next(rq_next, tid_bit);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200312
313 /* detach the task from the queue */
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100314 __task_unlink_rq(t);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200315 __task_wakeup(t, &rqueue_local[tid]);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100316 }
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200317#endif
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100318
Christopher Faulet8a48f672017-11-14 10:38:36 +0100319 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200320 } else {
321 if (!(active_tasks_mask & tid_bit)) {
322 activity[tid].empty_rq++;
323 return;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200324 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200325 }
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200326 /* Get some tasks from the run queue, make sure we don't
327 * get too much in the task list, but put a bit more than
328 * the max that will be run, to give a bit more fairness
329 */
Olivier Houchard1599b802018-05-24 18:59:04 +0200330 while (max_processed + (max_processed / 10) > task_list_size[tid]) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100331 /* Note: this loop is one of the fastest code path in
332 * the whole program. It should not be re-arranged
333 * without a good reason.
334 */
Willy Tarreau9e45b332017-11-08 14:05:19 +0100335
336 /* we have to restart looking up after every batch */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200337 rq_next = eb32sc_lookup_ge(&rqueue_local[tid], rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
338 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.
Emeric Brun01948972017-03-30 15:37:25 +0200344 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200345 rq_next = eb32sc_first(&rqueue_local[tid], tid_bit);
346 if (!rq_next)
347 break;
Emeric Brun01948972017-03-30 15:37:25 +0200348 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200349 t = eb32sc_entry(rq_next, struct task, rq);
350 rq_next = eb32sc_next(rq_next, tid_bit);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200351 /* Make sure nobody re-adds the task in the runqueue */
352 HA_ATOMIC_OR(&t->state, TASK_RUNNING);
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100353
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200354 /* detach the task from the queue */
355 __task_unlink_rq(t);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200356 /* And add it to the local task list */
357 task_insert_into_tasklet_list(t);
358 }
Olivier Houchardc4aac9e2018-07-26 15:25:49 +0200359 if (!(global_tasks_mask & tid_bit) && rqueue_size[tid] == 0) {
360 HA_ATOMIC_AND(&active_tasks_mask, ~tid_bit);
361 __ha_barrier_load();
362 if (global_tasks_mask & tid_bit)
363 HA_ATOMIC_OR(&active_tasks_mask, tid_bit);
364 }
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200365 while (max_processed > 0 && !LIST_ISEMPTY(&task_list[tid])) {
366 struct task *t;
367 unsigned short state;
368 void *ctx;
369 struct task *(*process)(struct task *t, void *ctx, unsigned short state);
370
371 t = (struct task *)LIST_ELEM(task_list[tid].n, struct tasklet *, list);
372 state = HA_ATOMIC_XCHG(&t->state, TASK_RUNNING);
373 __ha_barrier_store();
374 task_remove_from_task_list(t);
375
376 ctx = t->context;
377 process = t->process;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200378 t->calls++;
379 curr_task = (struct task *)t;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200380 if (likely(process == process_stream))
381 t = process_stream(t, ctx, state);
382 else {
383 if (t->process != NULL)
Olivier Houchard9db0fed2018-06-14 15:40:47 +0200384 t = process(TASK_IS_TASKLET(t) ? NULL : t, ctx, state);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200385 else {
386 __task_free(t);
387 t = NULL;
388 }
389 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200390 curr_task = NULL;
391 /* If there is a pending state we have to wake up the task
392 * immediatly, else we defer it into wait queue
393 */
394 if (t != NULL) {
395 state = HA_ATOMIC_AND(&t->state, ~TASK_RUNNING);
396 if (state)
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200397#ifdef USE_THREAD
Olivier Houchard19bdf242018-08-17 13:36:08 +0200398 __task_wakeup(t, ((t->thread_mask & all_threads_mask) == tid_bit) ?
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200399 &rqueue_local[tid] : &rqueue);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200400#else
401 __task_wakeup(t, &rqueue_local[tid]);
402#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200403 else
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100404 task_queue(t);
Willy Tarreau58b458d2008-06-29 22:40:23 +0200405 }
Willy Tarreauce4e0aa2017-11-05 23:57:00 +0100406
Olivier Houchard736ea412018-05-28 14:54:49 +0200407 max_processed--;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100408 if (max_processed <= 0) {
Willy Tarreau189ea852018-07-26 15:16:43 +0200409 HA_ATOMIC_OR(&active_tasks_mask, tid_bit);
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100410 activity[tid].long_rq++;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100411 break;
412 }
413 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200414}
415
Willy Tarreau4726f532009-03-07 17:25:21 +0100416/* perform minimal intializations, report 0 in case of error, 1 if OK. */
417int init_task()
418{
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200419 int i;
420
Willy Tarreau4726f532009-03-07 17:25:21 +0100421 memset(&timers, 0, sizeof(timers));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200422#ifdef USE_THREAD
Willy Tarreau4726f532009-03-07 17:25:21 +0100423 memset(&rqueue, 0, sizeof(rqueue));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200424#endif
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100425 HA_SPIN_INIT(&wq_lock);
426 HA_SPIN_INIT(&rq_lock);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200427 for (i = 0; i < MAX_THREADS; i++) {
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200428 memset(&rqueue_local[i], 0, sizeof(rqueue_local[i]));
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200429 LIST_INIT(&task_list[i]);
430 task_list_size[i] = 0;
431 }
Willy Tarreaubafbe012017-11-24 17:34:44 +0100432 pool_head_task = create_pool("task", sizeof(struct task), MEM_F_SHARED);
433 if (!pool_head_task)
Thierry FOURNIERd6975962017-07-12 14:31:10 +0200434 return 0;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200435 pool_head_tasklet = create_pool("tasklet", sizeof(struct tasklet), MEM_F_SHARED);
436 if (!pool_head_tasklet)
437 return 0;
Willy Tarreaubafbe012017-11-24 17:34:44 +0100438 pool_head_notification = create_pool("notification", sizeof(struct notification), MEM_F_SHARED);
439 if (!pool_head_notification)
Thierry FOURNIERd6975962017-07-12 14:31:10 +0200440 return 0;
441 return 1;
Willy Tarreau4726f532009-03-07 17:25:21 +0100442}
443
Willy Tarreaubaaee002006-06-26 02:48:02 +0200444/*
445 * Local variables:
446 * c-indent-level: 8
447 * c-basic-offset: 8
448 * End:
449 */