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 15
   8
   9 static int show_schedstat(struct seq_file *seq, void *v)
  10 {
  11         int cpu;
  12         int mask_len = DIV_ROUND_UP(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 %llu %llu %lu",
  30                     cpu, rq->yld_count,
  31                     rq->sched_switch, rq->sched_count, rq->sched_goidle,
  32                     rq->ttwu_count, rq->ttwu_local,
  33                     rq->rq_cpu_time,
  34                     rq->rq_sched_info.run_delay, rq->rq_sched_info.pcount);
  35
  36                 seq_printf(seq, "\n");
  37
  38 #ifdef CONFIG_SMP
  39                 /* domain-specific stats */
  40                 rcu_read_lock();
  41                 for_each_domain(cpu, sd) {
  42                         enum cpu_idle_type itype;
  43
  44                         cpumask_scnprintf(mask_str, mask_len,
  45                                           sched_domain_span(sd));
  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                 rcu_read_unlock();
  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 static 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 static int __init proc_schedstat_init(void)
 101 {
 102         proc_create("schedstat", 0, NULL, &proc_schedstat_operations);
 103         return 0;
 104 }
 105 module_init(proc_schedstat_init);
 106
 107 /*
 108  * Expects runqueue lock to be held for atomicity of update
 109  */
 110 static inline void
 111 rq_sched_info_arrive(struct rq *rq, unsigned long long delta)
 112 {
 113         if (rq) {
 114                 rq->rq_sched_info.run_delay += delta;
 115                 rq->rq_sched_info.pcount++;
 116         }
 117 }
 118
 119 /*
 120  * Expects runqueue lock to be held for atomicity of update
 121  */
 122 static inline void
 123 rq_sched_info_depart(struct rq *rq, unsigned long long delta)
 124 {
 125         if (rq)
 126                 rq->rq_cpu_time += delta;
 127 }
 128
 129 static inline void
 130 rq_sched_info_dequeued(struct rq *rq, unsigned long long delta)
 131 {
 132         if (rq)
 133                 rq->rq_sched_info.run_delay += delta;
 134 }
 135 # define schedstat_inc(rq, field)       do { (rq)->field++; } while (0)
 136 # define schedstat_add(rq, field, amt)  do { (rq)->field += (amt); } while (0)
 137 # define schedstat_set(var, val)        do { var = (val); } while (0)
 138 #else /* !CONFIG_SCHEDSTATS */
 139 static inline void
 140 rq_sched_info_arrive(struct rq *rq, unsigned long long delta)
 141 {}
 142 static inline void
 143 rq_sched_info_dequeued(struct rq *rq, unsigned long long delta)
 144 {}
 145 static inline void
 146 rq_sched_info_depart(struct rq *rq, unsigned long long delta)
 147 {}
 148 # define schedstat_inc(rq, field)       do { } while (0)
 149 # define schedstat_add(rq, field, amt)  do { } while (0)
 150 # define schedstat_set(var, val)        do { } while (0)
 151 #endif
 152
 153 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
 154 static inline void sched_info_reset_dequeued(struct task_struct *t)
 155 {
 156         t->sched_info.last_queued = 0;
 157 }
 158
 159 /*
 160  * We are interested in knowing how long it was from the *first* time a
 161  * task was queued to the time that it finally hit a cpu, we call this routine
 162  * from dequeue_task() to account for possible rq->clock skew across cpus. The
 163  * delta taken on each cpu would annul the skew.
 164  */
 165 static inline void sched_info_dequeued(struct task_struct *t)
 166 {
 167         unsigned long long now = task_rq(t)->clock, delta = 0;
 168
 169         if (unlikely(sched_info_on()))
 170                 if (t->sched_info.last_queued)
 171                         delta = now - t->sched_info.last_queued;
 172         sched_info_reset_dequeued(t);
 173         t->sched_info.run_delay += delta;
 174
 175         rq_sched_info_dequeued(task_rq(t), delta);
 176 }
 177
 178 /*
 179  * Called when a task finally hits the cpu.  We can now calculate how
 180  * long it was waiting to run.  We also note when it began so that we
 181  * can keep stats on how long its timeslice is.
 182  */
 183 static void sched_info_arrive(struct task_struct *t)
 184 {
 185         unsigned long long now = task_rq(t)->clock, delta = 0;
 186
 187         if (t->sched_info.last_queued)
 188                 delta = now - t->sched_info.last_queued;
 189         sched_info_reset_dequeued(t);
 190         t->sched_info.run_delay += delta;
 191         t->sched_info.last_arrival = now;
 192         t->sched_info.pcount++;
 193
 194         rq_sched_info_arrive(task_rq(t), delta);
 195 }
 196
 197 /*
 198  * This function is only called from enqueue_task(), but also only updates
 199  * the timestamp if it is already not set.  It's assumed that
 200  * sched_info_dequeued() will clear that stamp when appropriate.
 201  */
 202 static inline void sched_info_queued(struct task_struct *t)
 203 {
 204         if (unlikely(sched_info_on()))
 205                 if (!t->sched_info.last_queued)
 206                         t->sched_info.last_queued = task_rq(t)->clock;
 207 }
 208
 209 /*
 210  * Called when a process ceases being the active-running process, either
 211  * voluntarily or involuntarily.  Now we can calculate how long we ran.
 212  * Also, if the process is still in the TASK_RUNNING state, call
 213  * sched_info_queued() to mark that it has now again started waiting on
 214  * the runqueue.
 215  */
 216 static inline void sched_info_depart(struct task_struct *t)
 217 {
 218         unsigned long long delta = task_rq(t)->clock -
 219                                         t->sched_info.last_arrival;
 220
 221         rq_sched_info_depart(task_rq(t), delta);
 222
 223         if (t->state == TASK_RUNNING)
 224                 sched_info_queued(t);
 225 }
 226
 227 /*
 228  * Called when tasks are switched involuntarily due, typically, to expiring
 229  * their time slice.  (This may also be called when switching to or from
 230  * the idle task.)  We are only called when prev != next.
 231  */
 232 static inline void
 233 __sched_info_switch(struct task_struct *prev, struct task_struct *next)
 234 {
 235         struct rq *rq = task_rq(prev);
 236
 237         /*
 238          * prev now departs the cpu.  It's not interesting to record
 239          * stats about how efficient we were at scheduling the idle
 240          * process, however.
 241          */
 242         if (prev != rq->idle)
 243                 sched_info_depart(prev);
 244
 245         if (next != rq->idle)
 246                 sched_info_arrive(next);
 247 }
 248 static inline void
 249 sched_info_switch(struct task_struct *prev, struct task_struct *next)
 250 {
 251         if (unlikely(sched_info_on()))
 252                 __sched_info_switch(prev, next);
 253 }
 254 #else
 255 #define sched_info_queued(t)                    do { } while (0)
 256 #define sched_info_reset_dequeued(t)    do { } while (0)
 257 #define sched_info_dequeued(t)                  do { } while (0)
 258 #define sched_info_switch(t, next)              do { } while (0)
 259 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
 260
 261 /*
 262  * The following are functions that support scheduler-internal time accounting.
 263  * These functions are generally called at the timer tick.  None of this depends
 264  * on CONFIG_SCHEDSTATS.
 265  */
 266
 267 /**
 268  * account_group_user_time - Maintain utime for a thread group.
 269  *
 270  * @tsk:        Pointer to task structure.
 271  * @cputime:    Time value by which to increment the utime field of the
 272  *              thread_group_cputime structure.
 273  *
 274  * If thread group time is being maintained, get the structure for the
 275  * running CPU and update the utime field there.
 276  */
 277 static inline void account_group_user_time(struct task_struct *tsk,
 278                                            cputime_t cputime)
 279 {
 280         struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
 281
 282         if (!cputimer->running)
 283                 return;
 284
 285         spin_lock(&cputimer->lock);
 286         cputimer->cputime.utime =
 287                 cputime_add(cputimer->cputime.utime, cputime);
 288         spin_unlock(&cputimer->lock);
 289 }
 290
 291 /**
 292  * account_group_system_time - Maintain stime for a thread group.
 293  *
 294  * @tsk:        Pointer to task structure.
 295  * @cputime:    Time value by which to increment the stime field of the
 296  *              thread_group_cputime structure.
 297  *
 298  * If thread group time is being maintained, get the structure for the
 299  * running CPU and update the stime field there.
 300  */
 301 static inline void account_group_system_time(struct task_struct *tsk,
 302                                              cputime_t cputime)
 303 {
 304         struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
 305
 306         if (!cputimer->running)
 307                 return;
 308
 309         spin_lock(&cputimer->lock);
 310         cputimer->cputime.stime =
 311                 cputime_add(cputimer->cputime.stime, cputime);
 312         spin_unlock(&cputimer->lock);
 313 }
 314
 315 /**
 316  * account_group_exec_runtime - Maintain exec runtime for a thread group.
 317  *
 318  * @tsk:        Pointer to task structure.
 319  * @ns:         Time value by which to increment the sum_exec_runtime field
 320  *              of the thread_group_cputime structure.
 321  *
 322  * If thread group time is being maintained, get the structure for the
 323  * running CPU and update the sum_exec_runtime field there.
 324  */
 325 static inline void account_group_exec_runtime(struct task_struct *tsk,
 326                                               unsigned long long ns)
 327 {
 328         struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
 329
 330         if (!cputimer->running)
 331                 return;
 332
 333         spin_lock(&cputimer->lock);
 334         cputimer->cputime.sum_exec_runtime += ns;
 335         spin_unlock(&cputimer->lock);
 336 }