Merge master.kernel.org:/pub/scm/linux/kernel/git/davej/cpufreq
[sfrench/cifs-2.6.git] / kernel / sched_rt.c
1 /*
2  * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
3  * policies)
4  */
5
6 /*
7  * Update the current task's runtime statistics. Skip current tasks that
8  * are not in our scheduling class.
9  */
10 static void update_curr_rt(struct rq *rq)
11 {
12         struct task_struct *curr = rq->curr;
13         u64 delta_exec;
14
15         if (!task_has_rt_policy(curr))
16                 return;
17
18         delta_exec = rq->clock - curr->se.exec_start;
19         if (unlikely((s64)delta_exec < 0))
20                 delta_exec = 0;
21
22         schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
23
24         curr->se.sum_exec_runtime += delta_exec;
25         curr->se.exec_start = rq->clock;
26 }
27
28 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
29 {
30         struct rt_prio_array *array = &rq->rt.active;
31
32         list_add_tail(&p->run_list, array->queue + p->prio);
33         __set_bit(p->prio, array->bitmap);
34 }
35
36 /*
37  * Adding/removing a task to/from a priority array:
38  */
39 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
40 {
41         struct rt_prio_array *array = &rq->rt.active;
42
43         update_curr_rt(rq);
44
45         list_del(&p->run_list);
46         if (list_empty(array->queue + p->prio))
47                 __clear_bit(p->prio, array->bitmap);
48 }
49
50 /*
51  * Put task to the end of the run list without the overhead of dequeue
52  * followed by enqueue.
53  */
54 static void requeue_task_rt(struct rq *rq, struct task_struct *p)
55 {
56         struct rt_prio_array *array = &rq->rt.active;
57
58         list_move_tail(&p->run_list, array->queue + p->prio);
59 }
60
61 static void
62 yield_task_rt(struct rq *rq)
63 {
64         requeue_task_rt(rq, rq->curr);
65 }
66
67 /*
68  * Preempt the current task with a newly woken task if needed:
69  */
70 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
71 {
72         if (p->prio < rq->curr->prio)
73                 resched_task(rq->curr);
74 }
75
76 static struct task_struct *pick_next_task_rt(struct rq *rq)
77 {
78         struct rt_prio_array *array = &rq->rt.active;
79         struct task_struct *next;
80         struct list_head *queue;
81         int idx;
82
83         idx = sched_find_first_bit(array->bitmap);
84         if (idx >= MAX_RT_PRIO)
85                 return NULL;
86
87         queue = array->queue + idx;
88         next = list_entry(queue->next, struct task_struct, run_list);
89
90         next->se.exec_start = rq->clock;
91
92         return next;
93 }
94
95 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
96 {
97         update_curr_rt(rq);
98         p->se.exec_start = 0;
99 }
100
101 #ifdef CONFIG_SMP
102 /*
103  * Load-balancing iterator. Note: while the runqueue stays locked
104  * during the whole iteration, the current task might be
105  * dequeued so the iterator has to be dequeue-safe. Here we
106  * achieve that by always pre-iterating before returning
107  * the current task:
108  */
109 static struct task_struct *load_balance_start_rt(void *arg)
110 {
111         struct rq *rq = arg;
112         struct rt_prio_array *array = &rq->rt.active;
113         struct list_head *head, *curr;
114         struct task_struct *p;
115         int idx;
116
117         idx = sched_find_first_bit(array->bitmap);
118         if (idx >= MAX_RT_PRIO)
119                 return NULL;
120
121         head = array->queue + idx;
122         curr = head->prev;
123
124         p = list_entry(curr, struct task_struct, run_list);
125
126         curr = curr->prev;
127
128         rq->rt.rt_load_balance_idx = idx;
129         rq->rt.rt_load_balance_head = head;
130         rq->rt.rt_load_balance_curr = curr;
131
132         return p;
133 }
134
135 static struct task_struct *load_balance_next_rt(void *arg)
136 {
137         struct rq *rq = arg;
138         struct rt_prio_array *array = &rq->rt.active;
139         struct list_head *head, *curr;
140         struct task_struct *p;
141         int idx;
142
143         idx = rq->rt.rt_load_balance_idx;
144         head = rq->rt.rt_load_balance_head;
145         curr = rq->rt.rt_load_balance_curr;
146
147         /*
148          * If we arrived back to the head again then
149          * iterate to the next queue (if any):
150          */
151         if (unlikely(head == curr)) {
152                 int next_idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
153
154                 if (next_idx >= MAX_RT_PRIO)
155                         return NULL;
156
157                 idx = next_idx;
158                 head = array->queue + idx;
159                 curr = head->prev;
160
161                 rq->rt.rt_load_balance_idx = idx;
162                 rq->rt.rt_load_balance_head = head;
163         }
164
165         p = list_entry(curr, struct task_struct, run_list);
166
167         curr = curr->prev;
168
169         rq->rt.rt_load_balance_curr = curr;
170
171         return p;
172 }
173
174 static unsigned long
175 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
176                 unsigned long max_load_move,
177                 struct sched_domain *sd, enum cpu_idle_type idle,
178                 int *all_pinned, int *this_best_prio)
179 {
180         struct rq_iterator rt_rq_iterator;
181
182         rt_rq_iterator.start = load_balance_start_rt;
183         rt_rq_iterator.next = load_balance_next_rt;
184         /* pass 'busiest' rq argument into
185          * load_balance_[start|next]_rt iterators
186          */
187         rt_rq_iterator.arg = busiest;
188
189         return balance_tasks(this_rq, this_cpu, busiest, max_load_move, sd,
190                              idle, all_pinned, this_best_prio, &rt_rq_iterator);
191 }
192
193 static int
194 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
195                  struct sched_domain *sd, enum cpu_idle_type idle)
196 {
197         struct rq_iterator rt_rq_iterator;
198
199         rt_rq_iterator.start = load_balance_start_rt;
200         rt_rq_iterator.next = load_balance_next_rt;
201         rt_rq_iterator.arg = busiest;
202
203         return iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
204                                   &rt_rq_iterator);
205 }
206 #endif
207
208 static void task_tick_rt(struct rq *rq, struct task_struct *p)
209 {
210         /*
211          * RR tasks need a special form of timeslice management.
212          * FIFO tasks have no timeslices.
213          */
214         if (p->policy != SCHED_RR)
215                 return;
216
217         if (--p->time_slice)
218                 return;
219
220         p->time_slice = DEF_TIMESLICE;
221
222         /*
223          * Requeue to the end of queue if we are not the only element
224          * on the queue:
225          */
226         if (p->run_list.prev != p->run_list.next) {
227                 requeue_task_rt(rq, p);
228                 set_tsk_need_resched(p);
229         }
230 }
231
232 static void set_curr_task_rt(struct rq *rq)
233 {
234         struct task_struct *p = rq->curr;
235
236         p->se.exec_start = rq->clock;
237 }
238
239 const struct sched_class rt_sched_class = {
240         .next                   = &fair_sched_class,
241         .enqueue_task           = enqueue_task_rt,
242         .dequeue_task           = dequeue_task_rt,
243         .yield_task             = yield_task_rt,
244
245         .check_preempt_curr     = check_preempt_curr_rt,
246
247         .pick_next_task         = pick_next_task_rt,
248         .put_prev_task          = put_prev_task_rt,
249
250 #ifdef CONFIG_SMP
251         .load_balance           = load_balance_rt,
252         .move_one_task          = move_one_task_rt,
253 #endif
254
255         .set_curr_task          = set_curr_task_rt,
256         .task_tick              = task_tick_rt,
257 };