1 /*
2 * Task management functions.
3 *
4 * Copyright 2000-2009 Willy Tarreau <w@1wt.eu>
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
13 #include <string.h>
14
15 #include <common/config.h>
16 #include <common/memory.h>
17 #include <common/mini-clist.h>
18 #include <common/standard.h>
19 #include <common/time.h>
20 #include <eb32tree.h>
21
22 #include <proto/proxy.h>
23 #include <proto/stream.h>
24 #include <proto/task.h>
25
26 struct pool_head *pool2_task;
27
28 unsigned int nb_tasks = 0;
29 unsigned int tasks_run_queue = 0;
30 unsigned int tasks_run_queue_cur = 0; /* copy of the run queue size */
31 unsigned int nb_tasks_cur = 0; /* copy of the tasks count */
32 unsigned int niced_tasks = 0; /* number of niced tasks in the run queue */
33 struct eb32_node *last_timer = NULL; /* optimization: last queued timer */
34 struct eb32_node *rq_next = NULL; /* optimization: next task except if delete/insert */
35
36 static struct eb_root timers; /* sorted timers tree */
37 static struct eb_root rqueue; /* tree constituting the run queue */
38 static unsigned int rqueue_ticks; /* insertion count */
39
40 /* Puts the task <t> in run queue at a position depending on t->nice. <t> is
41 * returned. The nice value assigns boosts in 32th of the run queue size. A
42 * nice value of -1024 sets the task to -tasks_run_queue*32, while a nice value
43 * of 1024 sets the task to tasks_run_queue*32. The state flags are cleared, so
44 * the caller will have to set its flags after this call.
45 * The task must not already be in the run queue. If unsure, use the safer
46 * task_wakeup() function.
47 */
__task_wakeup(struct task * t)48 struct task *__task_wakeup(struct task *t)
49 {
50 tasks_run_queue++;
51 t->rq.key = ++rqueue_ticks;
52
53 if (likely(t->nice)) {
54 int offset;
55
56 niced_tasks++;
57 if (likely(t->nice > 0))
58 offset = (unsigned)((tasks_run_queue * (unsigned int)t->nice) / 32U);
59 else
60 offset = -(unsigned)((tasks_run_queue * (unsigned int)-t->nice) / 32U);
61 t->rq.key += offset;
62 }
63
64 /* clear state flags at the same time */
65 t->state &= ~TASK_WOKEN_ANY;
66
67 eb32_insert(&rqueue, &t->rq);
68 rq_next = NULL;
69 return t;
70 }
71
72 /*
73 * __task_queue()
74 *
75 * Inserts a task into the wait queue at the position given by its expiration
76 * date. It does not matter if the task was already in the wait queue or not,
77 * as it will be unlinked. The task must not have an infinite expiration timer.
78 * Last, tasks must not be queued further than the end of the tree, which is
79 * between <now_ms> and <now_ms> + 2^31 ms (now+24days in 32bit).
80 *
81 * This function should not be used directly, it is meant to be called by the
82 * inline version of task_queue() which performs a few cheap preliminary tests
83 * before deciding to call __task_queue().
84 */
__task_queue(struct task * task)85 void __task_queue(struct task *task)
86 {
87 if (likely(task_in_wq(task)))
88 __task_unlink_wq(task);
89
90 /* the task is not in the queue now */
91 task->wq.key = task->expire;
92 #ifdef DEBUG_CHECK_INVALID_EXPIRATION_DATES
93 if (tick_is_lt(task->wq.key, now_ms))
94 /* we're queuing too far away or in the past (most likely) */
95 return;
96 #endif
97
98 if (likely(last_timer &&
99 last_timer->node.bit < 0 &&
100 last_timer->key == task->wq.key &&
101 last_timer->node.node_p)) {
102 /* Most often, last queued timer has the same expiration date, so
103 * if it's not queued at the root, let's queue a dup directly there.
104 * Note that we can only use dups at the dup tree's root (most
105 * negative bit).
106 */
107 eb_insert_dup(&last_timer->node, &task->wq.node);
108 if (task->wq.node.bit < last_timer->node.bit)
109 last_timer = &task->wq;
110 return;
111 }
112 eb32_insert(&timers, &task->wq);
113
114 /* Make sure we don't assign the last_timer to a node-less entry */
115 if (task->wq.node.node_p && (!last_timer || (task->wq.node.bit < last_timer->node.bit)))
116 last_timer = &task->wq;
117 return;
118 }
119
120 /*
121 * Extract all expired timers from the timer queue, and wakes up all
122 * associated tasks. Returns the date of next event (or eternity).
123 */
wake_expired_tasks()124 int wake_expired_tasks()
125 {
126 struct task *task;
127 struct eb32_node *eb;
128
129 eb = eb32_lookup_ge(&timers, now_ms - TIMER_LOOK_BACK);
130 while (1) {
131 if (unlikely(!eb)) {
132 /* we might have reached the end of the tree, typically because
133 * <now_ms> is in the first half and we're first scanning the last
134 * half. Let's loop back to the beginning of the tree now.
135 */
136 eb = eb32_first(&timers);
137 if (likely(!eb))
138 break;
139 }
140
141 if (likely(tick_is_lt(now_ms, eb->key))) {
142 /* timer not expired yet, revisit it later */
143 return eb->key;
144 }
145
146 /* timer looks expired, detach it from the queue */
147 task = eb32_entry(eb, struct task, wq);
148 eb = eb32_next(eb);
149 __task_unlink_wq(task);
150
151 /* It is possible that this task was left at an earlier place in the
152 * tree because a recent call to task_queue() has not moved it. This
153 * happens when the new expiration date is later than the old one.
154 * Since it is very unlikely that we reach a timeout anyway, it's a
155 * lot cheaper to proceed like this because we almost never update
156 * the tree. We may also find disabled expiration dates there. Since
157 * we have detached the task from the tree, we simply call task_queue
158 * to take care of this. Note that we might occasionally requeue it at
159 * the same place, before <eb>, so we have to check if this happens,
160 * and adjust <eb>, otherwise we may skip it which is not what we want.
161 * We may also not requeue the task (and not point eb at it) if its
162 * expiration time is not set.
163 */
164 if (!tick_is_expired(task->expire, now_ms)) {
165 if (!tick_isset(task->expire))
166 continue;
167 __task_queue(task);
168 if (!eb || eb->key > task->wq.key)
169 eb = &task->wq;
170 continue;
171 }
172 task_wakeup(task, TASK_WOKEN_TIMER);
173 }
174
175 /* No task is expired */
176 return TICK_ETERNITY;
177 }
178
179 /* The run queue is chronologically sorted in a tree. An insertion counter is
180 * used to assign a position to each task. This counter may be combined with
181 * other variables (eg: nice value) to set the final position in the tree. The
182 * counter may wrap without a problem, of course. We then limit the number of
183 * tasks processed at once to 1/4 of the number of tasks in the queue, and to
184 * 200 max in any case, so that general latency remains low and so that task
185 * positions have a chance to be considered.
186 *
187 * The function adjusts <next> if a new event is closer.
188 */
process_runnable_tasks()189 void process_runnable_tasks()
190 {
191 struct task *t;
192 unsigned int max_processed;
193
194 tasks_run_queue_cur = tasks_run_queue; /* keep a copy for reporting */
195 nb_tasks_cur = nb_tasks;
196 max_processed = tasks_run_queue;
197
198 if (!tasks_run_queue)
199 return;
200
201 if (max_processed > 200)
202 max_processed = 200;
203
204 if (likely(niced_tasks))
205 max_processed = (max_processed + 3) / 4;
206
207 while (max_processed--) {
208 /* Note: this loop is one of the fastest code path in
209 * the whole program. It should not be re-arranged
210 * without a good reason.
211 */
212 if (unlikely(!rq_next)) {
213 rq_next = eb32_lookup_ge(&rqueue, rqueue_ticks - TIMER_LOOK_BACK);
214 if (!rq_next) {
215 /* we might have reached the end of the tree, typically because
216 * <rqueue_ticks> is in the first half and we're first scanning
217 * the last half. Let's loop back to the beginning of the tree now.
218 */
219 rq_next = eb32_first(&rqueue);
220 if (!rq_next)
221 break;
222 }
223 }
224
225 /* detach the task from the queue after updating the pointer to
226 * the next entry.
227 */
228 t = eb32_entry(rq_next, struct task, rq);
229 rq_next = eb32_next(rq_next);
230 __task_unlink_rq(t);
231
232 t->state |= TASK_RUNNING;
233 /* This is an optimisation to help the processor's branch
234 * predictor take this most common call.
235 */
236 t->calls++;
237 if (likely(t->process == process_stream))
238 t = process_stream(t);
239 else
240 t = t->process(t);
241
242 if (likely(t != NULL)) {
243 t->state &= ~TASK_RUNNING;
244 if (t->expire)
245 task_queue(t);
246 }
247 }
248 }
249
250 /* perform minimal intializations, report 0 in case of error, 1 if OK. */
init_task()251 int init_task()
252 {
253 memset(&timers, 0, sizeof(timers));
254 memset(&rqueue, 0, sizeof(rqueue));
255 pool2_task = create_pool("task", sizeof(struct task), MEM_F_SHARED);
256 return pool2_task != NULL;
257 }
258
259 /*
260 * Local variables:
261 * c-indent-level: 8
262 * c-basic-offset: 8
263 * End:
264 */
265