Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 1 | /* |
| 2 | * Task management functions. |
| 3 | * |
Willy Tarreau | 4726f53 | 2009-03-07 17:25:21 +0100 | [diff] [blame] | 4 | * Copyright 2000-2009 Willy Tarreau <w@1wt.eu> |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 5 | * |
| 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 Tarreau | 87bed62 | 2009-03-08 22:25:28 +0100 | [diff] [blame] | 13 | #include <string.h> |
| 14 | |
Willy Tarreau | 2dd0d47 | 2006-06-29 17:53:05 +0200 | [diff] [blame] | 15 | #include <common/config.h> |
Willy Tarreau | 9789f7b | 2008-06-24 08:17:16 +0200 | [diff] [blame] | 16 | #include <common/eb32tree.h> |
Willy Tarreau | c6ca1a0 | 2007-05-13 19:43:47 +0200 | [diff] [blame] | 17 | #include <common/memory.h> |
Willy Tarreau | 2dd0d47 | 2006-06-29 17:53:05 +0200 | [diff] [blame] | 18 | #include <common/mini-clist.h> |
Willy Tarreau | 96bcfd7 | 2007-04-29 10:41:56 +0200 | [diff] [blame] | 19 | #include <common/standard.h> |
Willy Tarreau | a6a6a93 | 2007-04-28 22:40:08 +0200 | [diff] [blame] | 20 | #include <common/time.h> |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 21 | |
Willy Tarreau | d825eef | 2007-05-12 22:35:00 +0200 | [diff] [blame] | 22 | #include <proto/proxy.h> |
Willy Tarreau | 531cf0c | 2009-03-08 16:35:27 +0100 | [diff] [blame] | 23 | #include <proto/session.h> |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 24 | #include <proto/task.h> |
| 25 | |
Willy Tarreau | 9789f7b | 2008-06-24 08:17:16 +0200 | [diff] [blame] | 26 | struct pool_head *pool2_task; |
Willy Tarreau | 96bcfd7 | 2007-04-29 10:41:56 +0200 | [diff] [blame] | 27 | |
Willy Tarreau | 58b458d | 2008-06-29 22:40:23 +0200 | [diff] [blame] | 28 | unsigned int run_queue = 0; |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 29 | unsigned int niced_tasks = 0; /* number of niced tasks in the run queue */ |
Willy Tarreau | 26ca34e | 2009-03-21 12:51:40 +0100 | [diff] [blame^] | 30 | struct eb32_node *last_timer = NULL; /* optimization: last queued timer */ |
Willy Tarreau | 964c936 | 2007-01-07 00:38:00 +0100 | [diff] [blame] | 31 | |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 32 | static struct eb_root timers; /* sorted timers tree */ |
| 33 | static struct eb_root rqueue; /* tree constituting the run queue */ |
| 34 | static unsigned int rqueue_ticks; /* insertion count */ |
Willy Tarreau | 9789f7b | 2008-06-24 08:17:16 +0200 | [diff] [blame] | 35 | |
Willy Tarreau | 4726f53 | 2009-03-07 17:25:21 +0100 | [diff] [blame] | 36 | /* Puts the task <t> in run queue at a position depending on t->nice. <t> is |
| 37 | * returned. The nice value assigns boosts in 32th of the run queue size. A |
| 38 | * nice value of -1024 sets the task to -run_queue*32, while a nice value of |
| 39 | * 1024 sets the task to run_queue*32. The state flags are cleared, so the |
| 40 | * caller will have to set its flags after this call. |
| 41 | * The task must not already be in the run queue. If unsure, use the safer |
| 42 | * task_wakeup() function. |
Willy Tarreau | 91e9993 | 2008-06-30 07:51:00 +0200 | [diff] [blame] | 43 | */ |
Willy Tarreau | 4df8206 | 2008-08-29 15:26:14 +0200 | [diff] [blame] | 44 | struct task *__task_wakeup(struct task *t) |
Willy Tarreau | e33aece | 2007-04-30 13:15:14 +0200 | [diff] [blame] | 45 | { |
Willy Tarreau | 58b458d | 2008-06-29 22:40:23 +0200 | [diff] [blame] | 46 | run_queue++; |
Willy Tarreau | 4726f53 | 2009-03-07 17:25:21 +0100 | [diff] [blame] | 47 | t->rq.key = ++rqueue_ticks; |
Willy Tarreau | 91e9993 | 2008-06-30 07:51:00 +0200 | [diff] [blame] | 48 | |
| 49 | if (likely(t->nice)) { |
| 50 | int offset; |
| 51 | |
| 52 | niced_tasks++; |
| 53 | if (likely(t->nice > 0)) |
| 54 | offset = (unsigned)((run_queue * (unsigned int)t->nice) / 32U); |
| 55 | else |
| 56 | offset = -(unsigned)((run_queue * (unsigned int)-t->nice) / 32U); |
Willy Tarreau | 4726f53 | 2009-03-07 17:25:21 +0100 | [diff] [blame] | 57 | t->rq.key += offset; |
Willy Tarreau | 91e9993 | 2008-06-30 07:51:00 +0200 | [diff] [blame] | 58 | } |
| 59 | |
Willy Tarreau | fdccded | 2008-08-29 18:19:04 +0200 | [diff] [blame] | 60 | /* clear state flags at the same time */ |
Willy Tarreau | 4726f53 | 2009-03-07 17:25:21 +0100 | [diff] [blame] | 61 | t->state &= ~TASK_WOKEN_ANY; |
Willy Tarreau | 58b458d | 2008-06-29 22:40:23 +0200 | [diff] [blame] | 62 | |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 63 | eb32_insert(&rqueue, &t->rq); |
Willy Tarreau | 58b458d | 2008-06-29 22:40:23 +0200 | [diff] [blame] | 64 | return t; |
Willy Tarreau | e33aece | 2007-04-30 13:15:14 +0200 | [diff] [blame] | 65 | } |
Willy Tarreau | d825eef | 2007-05-12 22:35:00 +0200 | [diff] [blame] | 66 | |
Willy Tarreau | 96bcfd7 | 2007-04-29 10:41:56 +0200 | [diff] [blame] | 67 | /* |
Willy Tarreau | 531cf0c | 2009-03-08 16:35:27 +0100 | [diff] [blame] | 68 | * __task_queue() |
Willy Tarreau | 96bcfd7 | 2007-04-29 10:41:56 +0200 | [diff] [blame] | 69 | * |
| 70 | * Inserts a task into the wait queue at the position given by its expiration |
Willy Tarreau | 4726f53 | 2009-03-07 17:25:21 +0100 | [diff] [blame] | 71 | * date. It does not matter if the task was already in the wait queue or not, |
Willy Tarreau | 531cf0c | 2009-03-08 16:35:27 +0100 | [diff] [blame] | 72 | * as it will be unlinked. The task must not have an infinite expiration timer. |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 73 | * Last, tasks must not be queued further than the end of the tree, which is |
| 74 | * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit). |
Willy Tarreau | 531cf0c | 2009-03-08 16:35:27 +0100 | [diff] [blame] | 75 | * |
| 76 | * This function should not be used directly, it is meant to be called by the |
| 77 | * inline version of task_queue() which performs a few cheap preliminary tests |
| 78 | * before deciding to call __task_queue(). |
Willy Tarreau | 96bcfd7 | 2007-04-29 10:41:56 +0200 | [diff] [blame] | 79 | */ |
Willy Tarreau | 531cf0c | 2009-03-08 16:35:27 +0100 | [diff] [blame] | 80 | void __task_queue(struct task *task) |
Willy Tarreau | 964c936 | 2007-01-07 00:38:00 +0100 | [diff] [blame] | 81 | { |
Willy Tarreau | 531cf0c | 2009-03-08 16:35:27 +0100 | [diff] [blame] | 82 | if (likely(task_in_wq(task))) |
Willy Tarreau | 4726f53 | 2009-03-07 17:25:21 +0100 | [diff] [blame] | 83 | __task_unlink_wq(task); |
Willy Tarreau | 4726f53 | 2009-03-07 17:25:21 +0100 | [diff] [blame] | 84 | |
| 85 | /* the task is not in the queue now */ |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 86 | task->wq.key = task->expire; |
Willy Tarreau | 28c41a4 | 2008-06-29 17:00:59 +0200 | [diff] [blame] | 87 | #ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 88 | if (tick_is_lt(task->wq.key, now_ms)) |
Willy Tarreau | 28c41a4 | 2008-06-29 17:00:59 +0200 | [diff] [blame] | 89 | /* we're queuing too far away or in the past (most likely) */ |
Willy Tarreau | 4726f53 | 2009-03-07 17:25:21 +0100 | [diff] [blame] | 90 | return; |
Willy Tarreau | 28c41a4 | 2008-06-29 17:00:59 +0200 | [diff] [blame] | 91 | #endif |
Willy Tarreau | ce44f12 | 2008-07-05 18:16:19 +0200 | [diff] [blame] | 92 | |
| 93 | if (likely(last_timer && |
Willy Tarreau | 26ca34e | 2009-03-21 12:51:40 +0100 | [diff] [blame^] | 94 | last_timer->node.bit < 0 && |
| 95 | last_timer->key == task->wq.key && |
| 96 | last_timer->node.node_p)) { |
Willy Tarreau | ce44f12 | 2008-07-05 18:16:19 +0200 | [diff] [blame] | 97 | /* Most often, last queued timer has the same expiration date, so |
| 98 | * if it's not queued at the root, let's queue a dup directly there. |
Willy Tarreau | 26ca34e | 2009-03-21 12:51:40 +0100 | [diff] [blame^] | 99 | * Note that we can only use dups at the dup tree's root (most |
| 100 | * negative bit). |
Willy Tarreau | ce44f12 | 2008-07-05 18:16:19 +0200 | [diff] [blame] | 101 | */ |
Willy Tarreau | 26ca34e | 2009-03-21 12:51:40 +0100 | [diff] [blame^] | 102 | eb_insert_dup(&last_timer->node, &task->wq.node); |
| 103 | if (task->wq.node.bit < last_timer->node.bit) |
| 104 | last_timer = &task->wq; |
Willy Tarreau | 4726f53 | 2009-03-07 17:25:21 +0100 | [diff] [blame] | 105 | return; |
Willy Tarreau | ce44f12 | 2008-07-05 18:16:19 +0200 | [diff] [blame] | 106 | } |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 107 | eb32_insert(&timers, &task->wq); |
Willy Tarreau | 26ca34e | 2009-03-21 12:51:40 +0100 | [diff] [blame^] | 108 | if (!last_timer || (task->wq.node.bit < last_timer->node.bit)) |
| 109 | last_timer = &task->wq; |
Willy Tarreau | 4726f53 | 2009-03-07 17:25:21 +0100 | [diff] [blame] | 110 | return; |
Willy Tarreau | 964c936 | 2007-01-07 00:38:00 +0100 | [diff] [blame] | 111 | } |
| 112 | |
Willy Tarreau | 96bcfd7 | 2007-04-29 10:41:56 +0200 | [diff] [blame] | 113 | /* |
Willy Tarreau | 9789f7b | 2008-06-24 08:17:16 +0200 | [diff] [blame] | 114 | * Extract all expired timers from the timer queue, and wakes up all |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 115 | * associated tasks. Returns the date of next event (or eternity) in <next>. |
Willy Tarreau | 96bcfd7 | 2007-04-29 10:41:56 +0200 | [diff] [blame] | 116 | */ |
Willy Tarreau | 0c303ee | 2008-07-07 00:09:58 +0200 | [diff] [blame] | 117 | void wake_expired_tasks(int *next) |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 118 | { |
Willy Tarreau | 96bcfd7 | 2007-04-29 10:41:56 +0200 | [diff] [blame] | 119 | struct task *task; |
Willy Tarreau | 9789f7b | 2008-06-24 08:17:16 +0200 | [diff] [blame] | 120 | struct eb32_node *eb; |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 121 | |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 122 | eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK); |
| 123 | while (1) { |
| 124 | if (unlikely(!eb)) { |
| 125 | /* we might have reached the end of the tree, typically because |
| 126 | * <now_ms> is in the first half and we're first scanning the last |
| 127 | * half. Let's loop back to the beginning of the tree now. |
| 128 | */ |
| 129 | eb = eb32_first(&timers); |
| 130 | if (likely(!eb)) |
| 131 | break; |
| 132 | } |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 133 | |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 134 | if (likely(tick_is_lt(now_ms, eb->key))) { |
| 135 | /* timer not expired yet, revisit it later */ |
| 136 | *next = eb->key; |
| 137 | return; |
| 138 | } |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 139 | |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 140 | /* timer looks expired, detach it from the queue */ |
| 141 | task = eb32_entry(eb, struct task, wq); |
| 142 | eb = eb32_next(eb); |
| 143 | __task_unlink_wq(task); |
Willy Tarreau | 4136522 | 2009-03-08 07:46:27 +0100 | [diff] [blame] | 144 | |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 145 | /* It is possible that this task was left at an earlier place in the |
| 146 | * tree because a recent call to task_queue() has not moved it. This |
| 147 | * happens when the new expiration date is later than the old one. |
| 148 | * Since it is very unlikely that we reach a timeout anyway, it's a |
| 149 | * lot cheaper to proceed like this because we almost never update |
| 150 | * the tree. We may also find disabled expiration dates there. Since |
| 151 | * we have detached the task from the tree, we simply call task_queue |
| 152 | * to take care of this. |
| 153 | */ |
| 154 | if (!tick_is_expired(task->expire, now_ms)) { |
| 155 | task_queue(task); |
| 156 | continue; |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 157 | } |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 158 | task_wakeup(task, TASK_WOKEN_TIMER); |
| 159 | } |
Willy Tarreau | 9789f7b | 2008-06-24 08:17:16 +0200 | [diff] [blame] | 160 | |
Willy Tarreau | 28c41a4 | 2008-06-29 17:00:59 +0200 | [diff] [blame] | 161 | /* We have found no task to expire in any tree */ |
Willy Tarreau | 0c303ee | 2008-07-07 00:09:58 +0200 | [diff] [blame] | 162 | *next = TICK_ETERNITY; |
Willy Tarreau | 28c41a4 | 2008-06-29 17:00:59 +0200 | [diff] [blame] | 163 | return; |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 164 | } |
| 165 | |
Willy Tarreau | 58b458d | 2008-06-29 22:40:23 +0200 | [diff] [blame] | 166 | /* The run queue is chronologically sorted in a tree. An insertion counter is |
| 167 | * used to assign a position to each task. This counter may be combined with |
| 168 | * other variables (eg: nice value) to set the final position in the tree. The |
| 169 | * counter may wrap without a problem, of course. We then limit the number of |
Willy Tarreau | 91e9993 | 2008-06-30 07:51:00 +0200 | [diff] [blame] | 170 | * tasks processed at once to 1/4 of the number of tasks in the queue, and to |
| 171 | * 200 max in any case, so that general latency remains low and so that task |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 172 | * positions have a chance to be considered. |
Willy Tarreau | 58b458d | 2008-06-29 22:40:23 +0200 | [diff] [blame] | 173 | * |
| 174 | * The function adjusts <next> if a new event is closer. |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 175 | */ |
Willy Tarreau | 0c303ee | 2008-07-07 00:09:58 +0200 | [diff] [blame] | 176 | void process_runnable_tasks(int *next) |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 177 | { |
Willy Tarreau | 964c936 | 2007-01-07 00:38:00 +0100 | [diff] [blame] | 178 | struct task *t; |
Willy Tarreau | 58b458d | 2008-06-29 22:40:23 +0200 | [diff] [blame] | 179 | struct eb32_node *eb; |
Willy Tarreau | 58b458d | 2008-06-29 22:40:23 +0200 | [diff] [blame] | 180 | unsigned int max_processed; |
Willy Tarreau | 26c2506 | 2009-03-08 09:38:41 +0100 | [diff] [blame] | 181 | int expire; |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 182 | |
Willy Tarreau | 91e9993 | 2008-06-30 07:51:00 +0200 | [diff] [blame] | 183 | max_processed = run_queue; |
| 184 | if (max_processed > 200) |
| 185 | max_processed = 200; |
| 186 | |
| 187 | if (likely(niced_tasks)) |
Willy Tarreau | 218859a | 2009-03-21 11:53:09 +0100 | [diff] [blame] | 188 | max_processed = (max_processed + 3) / 4; |
Willy Tarreau | 96bcfd7 | 2007-04-29 10:41:56 +0200 | [diff] [blame] | 189 | |
Willy Tarreau | 26c2506 | 2009-03-08 09:38:41 +0100 | [diff] [blame] | 190 | expire = *next; |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 191 | eb = eb32_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK); |
| 192 | while (max_processed--) { |
| 193 | /* Note: this loop is one of the fastest code path in |
| 194 | * the whole program. It should not be re-arranged |
| 195 | * without a good reason. |
| 196 | */ |
| 197 | |
| 198 | if (unlikely(!eb)) { |
| 199 | /* we might have reached the end of the tree, typically because |
| 200 | * <rqueue_ticks> is in the first half and we're first scanning |
| 201 | * the last half. Let's loop back to the beginning of the tree now. |
| 202 | */ |
| 203 | eb = eb32_first(&rqueue); |
| 204 | if (likely(!eb)) |
| 205 | break; |
| 206 | } |
Willy Tarreau | 58b458d | 2008-06-29 22:40:23 +0200 | [diff] [blame] | 207 | |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 208 | /* detach the task from the queue */ |
| 209 | t = eb32_entry(eb, struct task, rq); |
| 210 | eb = eb32_next(eb); |
| 211 | __task_unlink_rq(t); |
Willy Tarreau | 58b458d | 2008-06-29 22:40:23 +0200 | [diff] [blame] | 212 | |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 213 | t->state |= TASK_RUNNING; |
| 214 | /* This is an optimisation to help the processor's branch |
| 215 | * predictor take this most common call. |
| 216 | */ |
| 217 | if (likely(t->process == process_session)) |
| 218 | t = process_session(t); |
| 219 | else |
| 220 | t = t->process(t); |
Willy Tarreau | 531cf0c | 2009-03-08 16:35:27 +0100 | [diff] [blame] | 221 | |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 222 | if (likely(t != NULL)) { |
| 223 | t->state &= ~TASK_RUNNING; |
| 224 | if (t->expire) { |
| 225 | task_queue(t); |
| 226 | expire = tick_first_2nz(expire, t->expire); |
Willy Tarreau | 26c2506 | 2009-03-08 09:38:41 +0100 | [diff] [blame] | 227 | } |
Willy Tarreau | 58b458d | 2008-06-29 22:40:23 +0200 | [diff] [blame] | 228 | } |
Willy Tarreau | e35c94a | 2009-03-21 10:01:42 +0100 | [diff] [blame] | 229 | } |
Willy Tarreau | 26c2506 | 2009-03-08 09:38:41 +0100 | [diff] [blame] | 230 | *next = expire; |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 231 | } |
| 232 | |
Willy Tarreau | 4726f53 | 2009-03-07 17:25:21 +0100 | [diff] [blame] | 233 | /* perform minimal intializations, report 0 in case of error, 1 if OK. */ |
| 234 | int init_task() |
| 235 | { |
| 236 | memset(&timers, 0, sizeof(timers)); |
| 237 | memset(&rqueue, 0, sizeof(rqueue)); |
| 238 | pool2_task = create_pool("task", sizeof(struct task), MEM_F_SHARED); |
| 239 | return pool2_task != NULL; |
| 240 | } |
| 241 | |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 242 | /* |
| 243 | * Local variables: |
| 244 | * c-indent-level: 8 |
| 245 | * c-basic-offset: 8 |
| 246 | * End: |
| 247 | */ |