blob: b0f9b498b54920fef2bec19771f3fe17ad305206 [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>
26
Willy Tarreaubafbe012017-11-24 17:34:44 +010027struct pool_head *pool_head_task;
Olivier Houchardb0bdae72018-05-18 18:45:28 +020028struct pool_head *pool_head_tasklet;
Willy Tarreau96bcfd72007-04-29 10:41:56 +020029
Thierry FOURNIERd6975962017-07-12 14:31:10 +020030/* This is the memory pool containing all the signal structs. These
31 * struct are used to store each requiered signal between two tasks.
32 */
Willy Tarreaubafbe012017-11-24 17:34:44 +010033struct pool_head *pool_head_notification;
Thierry FOURNIERd6975962017-07-12 14:31:10 +020034
Willy Tarreaua4613182009-03-21 18:13:21 +010035unsigned int nb_tasks = 0;
Christopher Faulet3911ee82017-11-14 10:26:53 +010036unsigned long active_tasks_mask = 0; /* Mask of threads with active tasks */
Christopher Faulet34c5cc92016-12-06 09:15:30 +010037unsigned int tasks_run_queue = 0;
38unsigned int tasks_run_queue_cur = 0; /* copy of the run queue size */
Willy Tarreauc7bdf092009-03-21 18:33:52 +010039unsigned int nb_tasks_cur = 0; /* copy of the tasks count */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010040unsigned int niced_tasks = 0; /* number of niced tasks in the run queue */
Emeric Brun01948972017-03-30 15:37:25 +020041
Willy Tarreau6d1222c2017-11-26 10:08:06 +010042THREAD_LOCAL struct task *curr_task = NULL; /* task currently running or NULL */
Olivier Houchard9b36cb42018-05-04 15:46:16 +020043THREAD_LOCAL struct eb32sc_node *rq_next = NULL; /* Next task to be potentially run */
Willy Tarreau6d1222c2017-11-26 10:08:06 +010044
Olivier Houchardb0bdae72018-05-18 18:45:28 +020045struct list task_list[MAX_THREADS]; /* List of tasks to be run, mixing tasks and tasklets */
46int task_list_size[MAX_THREADS]; /* Number of tasks in the task_list */
47
Willy Tarreaua24d1d02017-11-26 10:19:16 +010048__decl_hathreads(HA_SPINLOCK_T __attribute__((aligned(64))) rq_lock); /* spin lock related to run queue */
49__decl_hathreads(HA_SPINLOCK_T __attribute__((aligned(64))) wq_lock); /* spin lock related to wait queue */
Willy Tarreau964c9362007-01-07 00:38:00 +010050
Willy Tarreaue35c94a2009-03-21 10:01:42 +010051static struct eb_root timers; /* sorted timers tree */
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020052#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020053struct eb_root rqueue; /* tree constituting the run queue */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020054static int global_rqueue_size; /* Number of element sin the global runqueue */
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020055#endif
56struct eb_root rqueue_local[MAX_THREADS]; /* tree constituting the per-thread run queue */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020057static int rqueue_size[MAX_THREADS]; /* Number of elements in the per-thread run queue */
Willy Tarreaue35c94a2009-03-21 10:01:42 +010058static unsigned int rqueue_ticks; /* insertion count */
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;
72
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020073#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020074 if (root == &rqueue) {
75 rq_size = &global_rqueue_size;
76 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020077 } else
78#endif
79 {
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020080 int nb = root - &rqueue_local[0];
81 rq_size = &rqueue_size[nb];
82 }
83 /* Make sure if the task isn't in the runqueue, nobody inserts it
84 * in the meanwhile.
85 */
86redo:
David Carliercc0a9572018-06-05 10:41:03 +000087 if (unlikely(!HA_ATOMIC_CAS(&t->rq.node.leaf_p, &expected, (void *)0x1))) {
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020088#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020089 if (root == &rqueue)
90 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +020091#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +020092 return;
93 }
94 /* There's a small race condition, when running a task, the thread
95 * first sets TASK_RUNNING, and then unlink the task.
96 * If an another thread calls task_wakeup() for the same task,
97 * it may set t->state before TASK_RUNNING was set, and then try
98 * to set t->rq.nod.leaf_p after it was unlinked.
99 * To make sure it is not a problem, we check if TASK_RUNNING is set
100 * again. If it is, we unset t->rq.node.leaf_p.
101 * We then check for TASK_RUNNING a third time. If it is still there,
102 * then we can give up, the task will be re-queued later if it needs
103 * to be. If it's not there, and there is still something in t->state,
104 * then we have to requeue.
105 */
106 if (((volatile unsigned short)(t->state)) & TASK_RUNNING) {
107 unsigned short state;
108 t->rq.node.leaf_p = NULL;
109 __ha_barrier_store();
110
111 state = (volatile unsigned short)(t->state);
112 if (unlikely(state != 0 && !(state & TASK_RUNNING)))
113 goto redo;
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200114#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200115 if (root == &rqueue)
116 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200117#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200118 return;
119 }
120 HA_ATOMIC_ADD(&tasks_run_queue, 1);
Christopher Faulet3911ee82017-11-14 10:26:53 +0100121 active_tasks_mask |= t->thread_mask;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200122 t->rq.key = HA_ATOMIC_ADD(&rqueue_ticks, 1);
Willy Tarreau91e99932008-06-30 07:51:00 +0200123
124 if (likely(t->nice)) {
125 int offset;
126
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200127 HA_ATOMIC_ADD(&niced_tasks, 1);
Willy Tarreau91e99932008-06-30 07:51:00 +0200128 if (likely(t->nice > 0))
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200129 offset = (unsigned)((*rq_size * (unsigned int)t->nice) / 32U);
Willy Tarreau91e99932008-06-30 07:51:00 +0200130 else
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200131 offset = -(unsigned)((*rq_size * (unsigned int)-t->nice) / 32U);
Willy Tarreau4726f532009-03-07 17:25:21 +0100132 t->rq.key += offset;
Willy Tarreau91e99932008-06-30 07:51:00 +0200133 }
134
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200135 eb32sc_insert(root, &t->rq, t->thread_mask);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200136#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200137 if (root == &rqueue) {
138 global_rqueue_size++;
139 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200140 } else
141#endif
142 {
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200143 int nb = root - &rqueue_local[0];
144
145 rqueue_size[nb]++;
146 }
147 return;
Willy Tarreaue33aece2007-04-30 13:15:14 +0200148}
Willy Tarreaud825eef2007-05-12 22:35:00 +0200149
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200150/*
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100151 * __task_queue()
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200152 *
153 * Inserts a task into the wait queue at the position given by its expiration
Willy Tarreau4726f532009-03-07 17:25:21 +0100154 * date. It does not matter if the task was already in the wait queue or not,
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100155 * as it will be unlinked. The task must not have an infinite expiration timer.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100156 * Last, tasks must not be queued further than the end of the tree, which is
157 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100158 *
159 * This function should not be used directly, it is meant to be called by the
160 * inline version of task_queue() which performs a few cheap preliminary tests
161 * before deciding to call __task_queue().
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200162 */
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100163void __task_queue(struct task *task)
Willy Tarreau964c9362007-01-07 00:38:00 +0100164{
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100165 if (likely(task_in_wq(task)))
Willy Tarreau4726f532009-03-07 17:25:21 +0100166 __task_unlink_wq(task);
Willy Tarreau4726f532009-03-07 17:25:21 +0100167
168 /* the task is not in the queue now */
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100169 task->wq.key = task->expire;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200170#ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100171 if (tick_is_lt(task->wq.key, now_ms))
Willy Tarreau28c41a42008-06-29 17:00:59 +0200172 /* we're queuing too far away or in the past (most likely) */
Willy Tarreau4726f532009-03-07 17:25:21 +0100173 return;
Willy Tarreau28c41a42008-06-29 17:00:59 +0200174#endif
Willy Tarreauce44f122008-07-05 18:16:19 +0200175
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100176 eb32_insert(&timers, &task->wq);
SaVaGe1d7a4202009-10-06 18:53:37 +0300177
Willy Tarreau4726f532009-03-07 17:25:21 +0100178 return;
Willy Tarreau964c9362007-01-07 00:38:00 +0100179}
180
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200181/*
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200182 * Extract all expired timers from the timer queue, and wakes up all
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100183 * associated tasks. Returns the date of next event (or eternity).
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200184 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100185int wake_expired_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200186{
Willy Tarreau96bcfd72007-04-29 10:41:56 +0200187 struct task *task;
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200188 struct eb32_node *eb;
Willy Tarreaub992ba12017-11-05 19:09:27 +0100189 int ret = TICK_ETERNITY;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200190
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100191 while (1) {
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100192 HA_SPIN_LOCK(TASK_WQ_LOCK, &wq_lock);
Emeric Brunc60def82017-09-27 14:59:38 +0200193 lookup_next:
Emeric Brun01948972017-03-30 15:37:25 +0200194 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
Emeric Brunc60def82017-09-27 14:59:38 +0200195 if (!eb) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100196 /* we might have reached the end of the tree, typically because
197 * <now_ms> is in the first half and we're first scanning the last
198 * half. Let's loop back to the beginning of the tree now.
199 */
200 eb = eb32_first(&timers);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100201 if (likely(!eb))
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100202 break;
203 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200204
Willy Tarreaub992ba12017-11-05 19:09:27 +0100205 if (tick_is_lt(now_ms, eb->key)) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100206 /* timer not expired yet, revisit it later */
Willy Tarreaub992ba12017-11-05 19:09:27 +0100207 ret = eb->key;
208 break;
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100209 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200210
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100211 /* timer looks expired, detach it from the queue */
212 task = eb32_entry(eb, struct task, wq);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100213 __task_unlink_wq(task);
Willy Tarreau41365222009-03-08 07:46:27 +0100214
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100215 /* It is possible that this task was left at an earlier place in the
216 * tree because a recent call to task_queue() has not moved it. This
217 * happens when the new expiration date is later than the old one.
218 * Since it is very unlikely that we reach a timeout anyway, it's a
219 * lot cheaper to proceed like this because we almost never update
220 * the tree. We may also find disabled expiration dates there. Since
221 * we have detached the task from the tree, we simply call task_queue
Willy Tarreau814c9782009-07-14 23:48:55 +0200222 * to take care of this. Note that we might occasionally requeue it at
223 * the same place, before <eb>, so we have to check if this happens,
224 * and adjust <eb>, otherwise we may skip it which is not what we want.
Willy Tarreau34e98ea2009-08-09 09:09:54 +0200225 * We may also not requeue the task (and not point eb at it) if its
226 * expiration time is not set.
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100227 */
228 if (!tick_is_expired(task->expire, now_ms)) {
Willy Tarreaub992ba12017-11-05 19:09:27 +0100229 if (tick_isset(task->expire))
230 __task_queue(task);
Emeric Brunc60def82017-09-27 14:59:38 +0200231 goto lookup_next;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200232 }
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100233 task_wakeup(task, TASK_WOKEN_TIMER);
Willy Tarreau51753452017-11-23 18:36:50 +0100234 HA_SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100235 }
Willy Tarreau9789f7b2008-06-24 08:17:16 +0200236
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100237 HA_SPIN_UNLOCK(TASK_WQ_LOCK, &wq_lock);
Willy Tarreaub992ba12017-11-05 19:09:27 +0100238 return ret;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200239}
240
Willy Tarreau58b458d2008-06-29 22:40:23 +0200241/* The run queue is chronologically sorted in a tree. An insertion counter is
242 * used to assign a position to each task. This counter may be combined with
243 * other variables (eg: nice value) to set the final position in the tree. The
244 * counter may wrap without a problem, of course. We then limit the number of
Christopher Faulet8a48f672017-11-14 10:38:36 +0100245 * tasks processed to 200 in any case, so that general latency remains low and
246 * so that task positions have a chance to be considered.
Willy Tarreau58b458d2008-06-29 22:40:23 +0200247 *
248 * The function adjusts <next> if a new event is closer.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200249 */
Thierry FOURNIER9cf7c4b2014-12-15 13:26:01 +0100250void process_runnable_tasks()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200251{
Willy Tarreau964c9362007-01-07 00:38:00 +0100252 struct task *t;
Emeric Brun01948972017-03-30 15:37:25 +0200253 int max_processed;
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200254#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200255 uint64_t average = 0;
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200256#endif
Christopher Faulet3911ee82017-11-14 10:26:53 +0100257
Christopher Faulet34c5cc92016-12-06 09:15:30 +0100258 tasks_run_queue_cur = tasks_run_queue; /* keep a copy for reporting */
Willy Tarreauc7bdf092009-03-21 18:33:52 +0100259 nb_tasks_cur = nb_tasks;
Olivier Houchard1599b802018-05-24 18:59:04 +0200260 max_processed = global.tune.runqueue_depth;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200261
262 if (likely(global.nbthread > 1)) {
263 HA_SPIN_LOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100264 if (!(active_tasks_mask & tid_bit)) {
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200265 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100266 activity[tid].empty_rq++;
Christopher Faulet8a48f672017-11-14 10:38:36 +0100267 return;
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100268 }
Christopher Faulet8a48f672017-11-14 10:38:36 +0100269
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200270#ifdef USE_THREAD
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200271 average = tasks_run_queue / global.nbthread;
272
273 /* Get some elements from the global run queue and put it in the
274 * local run queue. To try to keep a bit of fairness, just get as
275 * much elements from the global list as to have a bigger local queue
276 * than the average.
277 */
Olivier Houchard082627a2018-05-28 14:53:08 +0200278 while ((task_list_size[tid] + rqueue_size[tid]) <= average) {
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200279
280 /* we have to restart looking up after every batch */
281 rq_next = eb32sc_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
282 if (unlikely(!rq_next)) {
283 /* either we just started or we reached the end
284 * of the tree, typically because <rqueue_ticks>
285 * is in the first half and we're first scanning
286 * the last half. Let's loop back to the beginning
287 * of the tree now.
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100288 */
289 rq_next = eb32sc_first(&rqueue, tid_bit);
290 if (!rq_next)
291 break;
292 }
293
294 t = eb32sc_entry(rq_next, struct task, rq);
295 rq_next = eb32sc_next(rq_next, tid_bit);
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200296 global_rqueue_size--;
297
298 /* detach the task from the queue */
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100299 __task_unlink_rq(t);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200300 __task_wakeup(t, &rqueue_local[tid]);
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100301 }
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200302#endif
Willy Tarreauf0c531a2017-11-05 16:35:59 +0100303
Christopher Faulet8a48f672017-11-14 10:38:36 +0100304 HA_SPIN_UNLOCK(TASK_RQ_LOCK, &rq_lock);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200305 } else {
306 if (!(active_tasks_mask & tid_bit)) {
307 activity[tid].empty_rq++;
308 return;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200309 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200310 }
Christopher Faulet3911ee82017-11-14 10:26:53 +0100311 active_tasks_mask &= ~tid_bit;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200312 /* Get some tasks from the run queue, make sure we don't
313 * get too much in the task list, but put a bit more than
314 * the max that will be run, to give a bit more fairness
315 */
Olivier Houchard1599b802018-05-24 18:59:04 +0200316 while (max_processed + (max_processed / 10) > task_list_size[tid]) {
Willy Tarreaue35c94a2009-03-21 10:01:42 +0100317 /* Note: this loop is one of the fastest code path in
318 * the whole program. It should not be re-arranged
319 * without a good reason.
320 */
Willy Tarreau9e45b332017-11-08 14:05:19 +0100321
322 /* we have to restart looking up after every batch */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200323 rq_next = eb32sc_lookup_ge(&rqueue_local[tid], rqueue_ticks - TIMER_LOOK_BACK, tid_bit);
324 if (unlikely(!rq_next)) {
325 /* either we just started or we reached the end
326 * of the tree, typically because <rqueue_ticks>
327 * is in the first half and we're first scanning
328 * the last half. Let's loop back to the beginning
329 * of the tree now.
Emeric Brun01948972017-03-30 15:37:25 +0200330 */
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200331 rq_next = eb32sc_first(&rqueue_local[tid], tid_bit);
332 if (!rq_next)
333 break;
Emeric Brun01948972017-03-30 15:37:25 +0200334 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200335 t = eb32sc_entry(rq_next, struct task, rq);
336 rq_next = eb32sc_next(rq_next, tid_bit);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200337 /* Make sure nobody re-adds the task in the runqueue */
338 HA_ATOMIC_OR(&t->state, TASK_RUNNING);
Willy Tarreau531cf0c2009-03-08 16:35:27 +0100339
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200340 /* detach the task from the queue */
341 __task_unlink_rq(t);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200342 /* And add it to the local task list */
343 task_insert_into_tasklet_list(t);
344 }
345 while (max_processed > 0 && !LIST_ISEMPTY(&task_list[tid])) {
346 struct task *t;
347 unsigned short state;
348 void *ctx;
349 struct task *(*process)(struct task *t, void *ctx, unsigned short state);
350
351 t = (struct task *)LIST_ELEM(task_list[tid].n, struct tasklet *, list);
352 state = HA_ATOMIC_XCHG(&t->state, TASK_RUNNING);
353 __ha_barrier_store();
354 task_remove_from_task_list(t);
355
356 ctx = t->context;
357 process = t->process;
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200358 rqueue_size[tid]--;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200359 t->calls++;
360 curr_task = (struct task *)t;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200361 if (likely(process == process_stream))
362 t = process_stream(t, ctx, state);
363 else {
364 if (t->process != NULL)
Olivier Houchard9db0fed2018-06-14 15:40:47 +0200365 t = process(TASK_IS_TASKLET(t) ? NULL : t, ctx, state);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200366 else {
367 __task_free(t);
368 t = NULL;
369 }
370 }
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200371 curr_task = NULL;
372 /* If there is a pending state we have to wake up the task
373 * immediatly, else we defer it into wait queue
374 */
375 if (t != NULL) {
376 state = HA_ATOMIC_AND(&t->state, ~TASK_RUNNING);
377 if (state)
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200378#ifdef USE_THREAD
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200379 __task_wakeup(t, (t->thread_mask == tid_bit) ?
380 &rqueue_local[tid] : &rqueue);
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200381#else
382 __task_wakeup(t, &rqueue_local[tid]);
383#endif
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200384 else
Willy Tarreau9d4b56b2017-11-06 08:36:53 +0100385 task_queue(t);
Willy Tarreau58b458d2008-06-29 22:40:23 +0200386 }
Willy Tarreauce4e0aa2017-11-05 23:57:00 +0100387
Olivier Houchard736ea412018-05-28 14:54:49 +0200388 max_processed--;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100389 if (max_processed <= 0) {
390 active_tasks_mask |= tid_bit;
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100391 activity[tid].long_rq++;
Christopher Faulet3911ee82017-11-14 10:26:53 +0100392 break;
393 }
394 }
Willy Tarreaubaaee002006-06-26 02:48:02 +0200395}
396
Willy Tarreau4726f532009-03-07 17:25:21 +0100397/* perform minimal intializations, report 0 in case of error, 1 if OK. */
398int init_task()
399{
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200400 int i;
401
Willy Tarreau4726f532009-03-07 17:25:21 +0100402 memset(&timers, 0, sizeof(timers));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200403#ifdef USE_THREAD
Willy Tarreau4726f532009-03-07 17:25:21 +0100404 memset(&rqueue, 0, sizeof(rqueue));
Olivier Houchardb1ca58b2018-06-06 14:22:03 +0200405#endif
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100406 HA_SPIN_INIT(&wq_lock);
407 HA_SPIN_INIT(&rq_lock);
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200408 for (i = 0; i < MAX_THREADS; i++) {
Olivier Houchardf6e6dc12018-05-18 18:38:23 +0200409 memset(&rqueue_local[i], 0, sizeof(rqueue_local[i]));
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200410 LIST_INIT(&task_list[i]);
411 task_list_size[i] = 0;
412 }
Willy Tarreaubafbe012017-11-24 17:34:44 +0100413 pool_head_task = create_pool("task", sizeof(struct task), MEM_F_SHARED);
414 if (!pool_head_task)
Thierry FOURNIERd6975962017-07-12 14:31:10 +0200415 return 0;
Olivier Houchardb0bdae72018-05-18 18:45:28 +0200416 pool_head_tasklet = create_pool("tasklet", sizeof(struct tasklet), MEM_F_SHARED);
417 if (!pool_head_tasklet)
418 return 0;
Willy Tarreaubafbe012017-11-24 17:34:44 +0100419 pool_head_notification = create_pool("notification", sizeof(struct notification), MEM_F_SHARED);
420 if (!pool_head_notification)
Thierry FOURNIERd6975962017-07-12 14:31:10 +0200421 return 0;
422 return 1;
Willy Tarreau4726f532009-03-07 17:25:21 +0100423}
424
Willy Tarreaubaaee002006-06-26 02:48:02 +0200425/*
426 * Local variables:
427 * c-indent-level: 8
428 * c-basic-offset: 8
429 * End:
430 */