timers: fix itimer/many thread hang, v3
[linux-2.6.git] / kernel / sched_stats.h
1
2 #ifdef CONFIG_SCHEDSTATS
3 /*
4  * bump this up when changing the output format or the meaning of an existing
5  * format, so that tools can adapt (or abort)
6  */
7 #define SCHEDSTAT_VERSION 14
8
9 static int show_schedstat(struct seq_file *seq, void *v)
10 {
11         int cpu;
12         int mask_len = NR_CPUS/32 * 9;
13         char *mask_str = kmalloc(mask_len, GFP_KERNEL);
14
15         if (mask_str == NULL)
16                 return -ENOMEM;
17
18         seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
19         seq_printf(seq, "timestamp %lu\n", jiffies);
20         for_each_online_cpu(cpu) {
21                 struct rq *rq = cpu_rq(cpu);
22 #ifdef CONFIG_SMP
23                 struct sched_domain *sd;
24                 int dcount = 0;
25 #endif
26
27                 /* runqueue-specific stats */
28                 seq_printf(seq,
29                     "cpu%d %u %u %u %u %u %u %u %u %u %llu %llu %lu",
30                     cpu, rq->yld_both_empty,
31                     rq->yld_act_empty, rq->yld_exp_empty, rq->yld_count,
32                     rq->sched_switch, rq->sched_count, rq->sched_goidle,
33                     rq->ttwu_count, rq->ttwu_local,
34                     rq->rq_sched_info.cpu_time,
35                     rq->rq_sched_info.run_delay, rq->rq_sched_info.pcount);
36
37                 seq_printf(seq, "\n");
38
39 #ifdef CONFIG_SMP
40                 /* domain-specific stats */
41                 preempt_disable();
42                 for_each_domain(cpu, sd) {
43                         enum cpu_idle_type itype;
44
45                         cpumask_scnprintf(mask_str, mask_len, sd->span);
46                         seq_printf(seq, "domain%d %s", dcount++, mask_str);
47                         for (itype = CPU_IDLE; itype < CPU_MAX_IDLE_TYPES;
48                                         itype++) {
49                                 seq_printf(seq, " %u %u %u %u %u %u %u %u",
50                                     sd->lb_count[itype],
51                                     sd->lb_balanced[itype],
52                                     sd->lb_failed[itype],
53                                     sd->lb_imbalance[itype],
54                                     sd->lb_gained[itype],
55                                     sd->lb_hot_gained[itype],
56                                     sd->lb_nobusyq[itype],
57                                     sd->lb_nobusyg[itype]);
58                         }
59                         seq_printf(seq,
60                                    " %u %u %u %u %u %u %u %u %u %u %u %u\n",
61                             sd->alb_count, sd->alb_failed, sd->alb_pushed,
62                             sd->sbe_count, sd->sbe_balanced, sd->sbe_pushed,
63                             sd->sbf_count, sd->sbf_balanced, sd->sbf_pushed,
64                             sd->ttwu_wake_remote, sd->ttwu_move_affine,
65                             sd->ttwu_move_balance);
66                 }
67                 preempt_enable();
68 #endif
69         }
70         kfree(mask_str);
71         return 0;
72 }
73
74 static int schedstat_open(struct inode *inode, struct file *file)
75 {
76         unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
77         char *buf = kmalloc(size, GFP_KERNEL);
78         struct seq_file *m;
79         int res;
80
81         if (!buf)
82                 return -ENOMEM;
83         res = single_open(file, show_schedstat, NULL);
84         if (!res) {
85                 m = file->private_data;
86                 m->buf = buf;
87                 m->size = size;
88         } else
89                 kfree(buf);
90         return res;
91 }
92
93 const struct file_operations proc_schedstat_operations = {
94         .open    = schedstat_open,
95         .read    = seq_read,
96         .llseek  = seq_lseek,
97         .release = single_release,
98 };
99
100 /*
101  * Expects runqueue lock to be held for atomicity of update
102  */
103 static inline void
104 rq_sched_info_arrive(struct rq *rq, unsigned long long delta)
105 {
106         if (rq) {
107                 rq->rq_sched_info.run_delay += delta;
108                 rq->rq_sched_info.pcount++;
109         }
110 }
111
112 /*
113  * Expects runqueue lock to be held for atomicity of update
114  */
115 static inline void
116 rq_sched_info_depart(struct rq *rq, unsigned long long delta)
117 {
118         if (rq)
119                 rq->rq_sched_info.cpu_time += delta;
120 }
121
122 static inline void
123 rq_sched_info_dequeued(struct rq *rq, unsigned long long delta)
124 {
125         if (rq)
126                 rq->rq_sched_info.run_delay += delta;
127 }
128 # define schedstat_inc(rq, field)       do { (rq)->field++; } while (0)
129 # define schedstat_add(rq, field, amt)  do { (rq)->field += (amt); } while (0)
130 # define schedstat_set(var, val)        do { var = (val); } while (0)
131 #else /* !CONFIG_SCHEDSTATS */
132 static inline void
133 rq_sched_info_arrive(struct rq *rq, unsigned long long delta)
134 {}
135 static inline void
136 rq_sched_info_dequeued(struct rq *rq, unsigned long long delta)
137 {}
138 static inline void
139 rq_sched_info_depart(struct rq *rq, unsigned long long delta)
140 {}
141 # define schedstat_inc(rq, field)       do { } while (0)
142 # define schedstat_add(rq, field, amt)  do { } while (0)
143 # define schedstat_set(var, val)        do { } while (0)
144 #endif
145
146 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
147 static inline void sched_info_reset_dequeued(struct task_struct *t)
148 {
149         t->sched_info.last_queued = 0;
150 }
151
152 /*
153  * Called when a process is dequeued from the active array and given
154  * the cpu.  We should note that with the exception of interactive
155  * tasks, the expired queue will become the active queue after the active
156  * queue is empty, without explicitly dequeuing and requeuing tasks in the
157  * expired queue.  (Interactive tasks may be requeued directly to the
158  * active queue, thus delaying tasks in the expired queue from running;
159  * see scheduler_tick()).
160  *
161  * Though we are interested in knowing how long it was from the *first* time a
162  * task was queued to the time that it finally hit a cpu, we call this routine
163  * from dequeue_task() to account for possible rq->clock skew across cpus. The
164  * delta taken on each cpu would annul the skew.
165  */
166 static inline void sched_info_dequeued(struct task_struct *t)
167 {
168         unsigned long long now = task_rq(t)->clock, delta = 0;
169
170         if (unlikely(sched_info_on()))
171                 if (t->sched_info.last_queued)
172                         delta = now - t->sched_info.last_queued;
173         sched_info_reset_dequeued(t);
174         t->sched_info.run_delay += delta;
175
176         rq_sched_info_dequeued(task_rq(t), delta);
177 }
178
179 /*
180  * Called when a task finally hits the cpu.  We can now calculate how
181  * long it was waiting to run.  We also note when it began so that we
182  * can keep stats on how long its timeslice is.
183  */
184 static void sched_info_arrive(struct task_struct *t)
185 {
186         unsigned long long now = task_rq(t)->clock, delta = 0;
187
188         if (t->sched_info.last_queued)
189                 delta = now - t->sched_info.last_queued;
190         sched_info_reset_dequeued(t);
191         t->sched_info.run_delay += delta;
192         t->sched_info.last_arrival = now;
193         t->sched_info.pcount++;
194
195         rq_sched_info_arrive(task_rq(t), delta);
196 }
197
198 /*
199  * Called when a process is queued into either the active or expired
200  * array.  The time is noted and later used to determine how long we
201  * had to wait for us to reach the cpu.  Since the expired queue will
202  * become the active queue after active queue is empty, without dequeuing
203  * and requeuing any tasks, we are interested in queuing to either. It
204  * is unusual but not impossible for tasks to be dequeued and immediately
205  * requeued in the same or another array: this can happen in sched_yield(),
206  * set_user_nice(), and even load_balance() as it moves tasks from runqueue
207  * to runqueue.
208  *
209  * This function is only called from enqueue_task(), but also only updates
210  * the timestamp if it is already not set.  It's assumed that
211  * sched_info_dequeued() will clear that stamp when appropriate.
212  */
213 static inline void sched_info_queued(struct task_struct *t)
214 {
215         if (unlikely(sched_info_on()))
216                 if (!t->sched_info.last_queued)
217                         t->sched_info.last_queued = task_rq(t)->clock;
218 }
219
220 /*
221  * Called when a process ceases being the active-running process, either
222  * voluntarily or involuntarily.  Now we can calculate how long we ran.
223  * Also, if the process is still in the TASK_RUNNING state, call
224  * sched_info_queued() to mark that it has now again started waiting on
225  * the runqueue.
226  */
227 static inline void sched_info_depart(struct task_struct *t)
228 {
229         unsigned long long delta = task_rq(t)->clock -
230                                         t->sched_info.last_arrival;
231
232         t->sched_info.cpu_time += delta;
233         rq_sched_info_depart(task_rq(t), delta);
234
235         if (t->state == TASK_RUNNING)
236                 sched_info_queued(t);
237 }
238
239 /*
240  * Called when tasks are switched involuntarily due, typically, to expiring
241  * their time slice.  (This may also be called when switching to or from
242  * the idle task.)  We are only called when prev != next.
243  */
244 static inline void
245 __sched_info_switch(struct task_struct *prev, struct task_struct *next)
246 {
247         struct rq *rq = task_rq(prev);
248
249         /*
250          * prev now departs the cpu.  It's not interesting to record
251          * stats about how efficient we were at scheduling the idle
252          * process, however.
253          */
254         if (prev != rq->idle)
255                 sched_info_depart(prev);
256
257         if (next != rq->idle)
258                 sched_info_arrive(next);
259 }
260 static inline void
261 sched_info_switch(struct task_struct *prev, struct task_struct *next)
262 {
263         if (unlikely(sched_info_on()))
264                 __sched_info_switch(prev, next);
265 }
266 #else
267 #define sched_info_queued(t)                    do { } while (0)
268 #define sched_info_reset_dequeued(t)    do { } while (0)
269 #define sched_info_dequeued(t)                  do { } while (0)
270 #define sched_info_switch(t, next)              do { } while (0)
271 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
272
273 /*
274  * The following are functions that support scheduler-internal time accounting.
275  * These functions are generally called at the timer tick.  None of this depends
276  * on CONFIG_SCHEDSTATS.
277  */
278
279 /**
280  * account_group_user_time - Maintain utime for a thread group.
281  *
282  * @tsk:        Pointer to task structure.
283  * @cputime:    Time value by which to increment the utime field of the
284  *              thread_group_cputime structure.
285  *
286  * If thread group time is being maintained, get the structure for the
287  * running CPU and update the utime field there.
288  */
289 static inline void account_group_user_time(struct task_struct *tsk,
290                                            cputime_t cputime)
291 {
292         struct signal_struct *sig;
293
294         sig = tsk->signal;
295         if (unlikely(!sig))
296                 return;
297         if (sig->cputime.totals) {
298                 struct task_cputime *times;
299
300                 times = per_cpu_ptr(sig->cputime.totals, get_cpu());
301                 times->utime = cputime_add(times->utime, cputime);
302                 put_cpu_no_resched();
303         }
304 }
305
306 /**
307  * account_group_system_time - Maintain stime for a thread group.
308  *
309  * @tsk:        Pointer to task structure.
310  * @cputime:    Time value by which to increment the stime field of the
311  *              thread_group_cputime structure.
312  *
313  * If thread group time is being maintained, get the structure for the
314  * running CPU and update the stime field there.
315  */
316 static inline void account_group_system_time(struct task_struct *tsk,
317                                              cputime_t cputime)
318 {
319         struct signal_struct *sig;
320
321         sig = tsk->signal;
322         if (unlikely(!sig))
323                 return;
324         if (sig->cputime.totals) {
325                 struct task_cputime *times;
326
327                 times = per_cpu_ptr(sig->cputime.totals, get_cpu());
328                 times->stime = cputime_add(times->stime, cputime);
329                 put_cpu_no_resched();
330         }
331 }
332
333 /**
334  * account_group_exec_runtime - Maintain exec runtime for a thread group.
335  *
336  * @tsk:        Pointer to task structure.
337  * @ns:         Time value by which to increment the sum_exec_runtime field
338  *              of the thread_group_cputime structure.
339  *
340  * If thread group time is being maintained, get the structure for the
341  * running CPU and update the sum_exec_runtime field there.
342  */
343 static inline void account_group_exec_runtime(struct task_struct *tsk,
344                                               unsigned long long ns)
345 {
346         struct signal_struct *sig;
347
348         sig = tsk->signal;
349         if (unlikely(!sig))
350                 return;
351         if (sig->cputime.totals) {
352                 struct task_cputime *times;
353
354                 times = per_cpu_ptr(sig->cputime.totals, get_cpu());
355                 times->sum_exec_runtime += ns;
356                 put_cpu_no_resched();
357         }
358 }