kernel/sched/stats.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef _KERNEL_STATS_H
   3 #define _KERNEL_STATS_H
   4
   5 #ifdef CONFIG_SCHEDSTATS
   6
   7 extern struct static_key_false sched_schedstats;
   8
   9 /*
  10  * Expects runqueue lock to be held for atomicity of update
  11  */
  12 static inline void
  13 rq_sched_info_arrive(struct rq *rq, unsigned long long delta)
  14 {
  15         if (rq) {
  16                 rq->rq_sched_info.run_delay += delta;
  17                 rq->rq_sched_info.pcount++;
  18         }
  19 }
  20
  21 /*
  22  * Expects runqueue lock to be held for atomicity of update
  23  */
  24 static inline void
  25 rq_sched_info_depart(struct rq *rq, unsigned long long delta)
  26 {
  27         if (rq)
  28                 rq->rq_cpu_time += delta;
  29 }
  30
  31 static inline void
  32 rq_sched_info_dequeue(struct rq *rq, unsigned long long delta)
  33 {
  34         if (rq)
  35                 rq->rq_sched_info.run_delay += delta;
  36 }
  37 #define   schedstat_enabled()           static_branch_unlikely(&sched_schedstats)
  38 #define __schedstat_inc(var)            do { var++; } while (0)
  39 #define   schedstat_inc(var)            do { if (schedstat_enabled()) { var++; } } while (0)
  40 #define __schedstat_add(var, amt)       do { var += (amt); } while (0)
  41 #define   schedstat_add(var, amt)       do { if (schedstat_enabled()) { var += (amt); } } while (0)
  42 #define __schedstat_set(var, val)       do { var = (val); } while (0)
  43 #define   schedstat_set(var, val)       do { if (schedstat_enabled()) { var = (val); } } while (0)
  44 #define   schedstat_val(var)            (var)
  45 #define   schedstat_val_or_zero(var)    ((schedstat_enabled()) ? (var) : 0)
  46
  47 void __update_stats_wait_start(struct rq *rq, struct task_struct *p,
  48                                struct sched_statistics *stats);
  49
  50 void __update_stats_wait_end(struct rq *rq, struct task_struct *p,
  51                              struct sched_statistics *stats);
  52 void __update_stats_enqueue_sleeper(struct rq *rq, struct task_struct *p,
  53                                     struct sched_statistics *stats);
  54
  55 static inline void
  56 check_schedstat_required(void)
  57 {
  58         if (schedstat_enabled())
  59                 return;
  60
  61         /* Force schedstat enabled if a dependent tracepoint is active */
  62         if (trace_sched_stat_wait_enabled()    ||
  63             trace_sched_stat_sleep_enabled()   ||
  64             trace_sched_stat_iowait_enabled()  ||
  65             trace_sched_stat_blocked_enabled() ||
  66             trace_sched_stat_runtime_enabled())
  67                 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, stat_blocked and stat_runtime require the kernel parameter schedstats=enable or kernel.sched_schedstats=1\n");
  68 }
  69
  70 #else /* !CONFIG_SCHEDSTATS: */
  71
  72 static inline void rq_sched_info_arrive  (struct rq *rq, unsigned long long delta) { }
  73 static inline void rq_sched_info_dequeue(struct rq *rq, unsigned long long delta) { }
  74 static inline void rq_sched_info_depart  (struct rq *rq, unsigned long long delta) { }
  75 # define   schedstat_enabled()          0
  76 # define __schedstat_inc(var)           do { } while (0)
  77 # define   schedstat_inc(var)           do { } while (0)
  78 # define __schedstat_add(var, amt)      do { } while (0)
  79 # define   schedstat_add(var, amt)      do { } while (0)
  80 # define __schedstat_set(var, val)      do { } while (0)
  81 # define   schedstat_set(var, val)      do { } while (0)
  82 # define   schedstat_val(var)           0
  83 # define   schedstat_val_or_zero(var)   0
  84
  85 # define __update_stats_wait_start(rq, p, stats)       do { } while (0)
  86 # define __update_stats_wait_end(rq, p, stats)         do { } while (0)
  87 # define __update_stats_enqueue_sleeper(rq, p, stats)  do { } while (0)
  88 # define check_schedstat_required()                    do { } while (0)
  89
  90 #endif /* CONFIG_SCHEDSTATS */
  91
  92 #ifdef CONFIG_FAIR_GROUP_SCHED
  93 struct sched_entity_stats {
  94         struct sched_entity     se;
  95         struct sched_statistics stats;
  96 } __no_randomize_layout;
  97 #endif
  98
  99 static inline struct sched_statistics *
 100 __schedstats_from_se(struct sched_entity *se)
 101 {
 102 #ifdef CONFIG_FAIR_GROUP_SCHED
 103         if (!entity_is_task(se))
 104                 return &container_of(se, struct sched_entity_stats, se)->stats;
 105 #endif
 106         return &task_of(se)->stats;
 107 }
 108
 109 #ifdef CONFIG_PSI
 110 void psi_task_change(struct task_struct *task, int clear, int set);
 111 void psi_task_switch(struct task_struct *prev, struct task_struct *next,
 112                      bool sleep);
 113 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
 114 void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_struct *prev);
 115 #else
 116 static inline void psi_account_irqtime(struct rq *rq, struct task_struct *curr,
 117                                        struct task_struct *prev) {}
 118 #endif /*CONFIG_IRQ_TIME_ACCOUNTING */
 119 /*
 120  * PSI tracks state that persists across sleeps, such as iowaits and
 121  * memory stalls. As a result, it has to distinguish between sleeps,
 122  * where a task's runnable state changes, and migrations, where a task
 123  * and its runnable state are being moved between CPUs and runqueues.
 124  *
 125  * A notable case is a task whose dequeue is delayed. PSI considers
 126  * those sleeping, but because they are still on the runqueue they can
 127  * go through migration requeues. In this case, *sleeping* states need
 128  * to be transferred.
 129  */
 130 static inline void psi_enqueue(struct task_struct *p, bool migrate)
 131 {
 132         int clear = 0, set = 0;
 133
 134         if (static_branch_likely(&psi_disabled))
 135                 return;
 136
 137         if (p->se.sched_delayed) {
 138                 /* CPU migration of "sleeping" task */
 139                 SCHED_WARN_ON(!migrate);
 140                 if (p->in_memstall)
 141                         set |= TSK_MEMSTALL;
 142                 if (p->in_iowait)
 143                         set |= TSK_IOWAIT;
 144         } else if (migrate) {
 145                 /* CPU migration of runnable task */
 146                 set = TSK_RUNNING;
 147                 if (p->in_memstall)
 148                         set |= TSK_MEMSTALL | TSK_MEMSTALL_RUNNING;
 149         } else {
 150                 /* Wakeup of new or sleeping task */
 151                 if (p->in_iowait)
 152                         clear |= TSK_IOWAIT;
 153                 set = TSK_RUNNING;
 154                 if (p->in_memstall)
 155                         set |= TSK_MEMSTALL_RUNNING;
 156         }
 157
 158         psi_task_change(p, clear, set);
 159 }
 160
 161 static inline void psi_dequeue(struct task_struct *p, bool migrate)
 162 {
 163         if (static_branch_likely(&psi_disabled))
 164                 return;
 165
 166         /*
 167          * When migrating a task to another CPU, clear all psi
 168          * state. The enqueue callback above will work it out.
 169          */
 170         if (migrate)
 171                 psi_task_change(p, p->psi_flags, 0);
 172
 173         /*
 174          * A voluntary sleep is a dequeue followed by a task switch. To
 175          * avoid walking all ancestors twice, psi_task_switch() handles
 176          * TSK_RUNNING and TSK_IOWAIT for us when it moves TSK_ONCPU.
 177          * Do nothing here.
 178          */
 179 }
 180
 181 static inline void psi_ttwu_dequeue(struct task_struct *p)
 182 {
 183         if (static_branch_likely(&psi_disabled))
 184                 return;
 185         /*
 186          * Is the task being migrated during a wakeup? Make sure to
 187          * deregister its sleep-persistent psi states from the old
 188          * queue, and let psi_enqueue() know it has to requeue.
 189          */
 190         if (unlikely(p->psi_flags)) {
 191                 struct rq_flags rf;
 192                 struct rq *rq;
 193
 194                 rq = __task_rq_lock(p, &rf);
 195                 psi_task_change(p, p->psi_flags, 0);
 196                 __task_rq_unlock(rq, &rf);
 197         }
 198 }
 199
 200 static inline void psi_sched_switch(struct task_struct *prev,
 201                                     struct task_struct *next,
 202                                     bool sleep)
 203 {
 204         if (static_branch_likely(&psi_disabled))
 205                 return;
 206
 207         psi_task_switch(prev, next, sleep);
 208 }
 209
 210 #else /* CONFIG_PSI */
 211 static inline void psi_enqueue(struct task_struct *p, bool migrate) {}
 212 static inline void psi_dequeue(struct task_struct *p, bool migrate) {}
 213 static inline void psi_ttwu_dequeue(struct task_struct *p) {}
 214 static inline void psi_sched_switch(struct task_struct *prev,
 215                                     struct task_struct *next,
 216                                     bool sleep) {}
 217 static inline void psi_account_irqtime(struct rq *rq, struct task_struct *curr,
 218                                        struct task_struct *prev) {}
 219 #endif /* CONFIG_PSI */
 220
 221 #ifdef CONFIG_SCHED_INFO
 222 /*
 223  * We are interested in knowing how long it was from the *first* time a
 224  * task was queued to the time that it finally hit a CPU, we call this routine
 225  * from dequeue_task() to account for possible rq->clock skew across CPUs. The
 226  * delta taken on each CPU would annul the skew.
 227  */
 228 static inline void sched_info_dequeue(struct rq *rq, struct task_struct *t)
 229 {
 230         unsigned long long delta = 0;
 231
 232         if (!t->sched_info.last_queued)
 233                 return;
 234
 235         delta = rq_clock(rq) - t->sched_info.last_queued;
 236         t->sched_info.last_queued = 0;
 237         t->sched_info.run_delay += delta;
 238
 239         rq_sched_info_dequeue(rq, delta);
 240 }
 241
 242 /*
 243  * Called when a task finally hits the CPU.  We can now calculate how
 244  * long it was waiting to run.  We also note when it began so that we
 245  * can keep stats on how long its time-slice is.
 246  */
 247 static void sched_info_arrive(struct rq *rq, struct task_struct *t)
 248 {
 249         unsigned long long now, delta = 0;
 250
 251         if (!t->sched_info.last_queued)
 252                 return;
 253
 254         now = rq_clock(rq);
 255         delta = now - t->sched_info.last_queued;
 256         t->sched_info.last_queued = 0;
 257         t->sched_info.run_delay += delta;
 258         t->sched_info.last_arrival = now;
 259         t->sched_info.pcount++;
 260
 261         rq_sched_info_arrive(rq, delta);
 262 }
 263
 264 /*
 265  * This function is only called from enqueue_task(), but also only updates
 266  * the timestamp if it is already not set.  It's assumed that
 267  * sched_info_dequeue() will clear that stamp when appropriate.
 268  */
 269 static inline void sched_info_enqueue(struct rq *rq, struct task_struct *t)
 270 {
 271         if (!t->sched_info.last_queued)
 272                 t->sched_info.last_queued = rq_clock(rq);
 273 }
 274
 275 /*
 276  * Called when a process ceases being the active-running process involuntarily
 277  * due, typically, to expiring its time slice (this may also be called when
 278  * switching to the idle task).  Now we can calculate how long we ran.
 279  * Also, if the process is still in the TASK_RUNNING state, call
 280  * sched_info_enqueue() to mark that it has now again started waiting on
 281  * the runqueue.
 282  */
 283 static inline void sched_info_depart(struct rq *rq, struct task_struct *t)
 284 {
 285         unsigned long long delta = rq_clock(rq) - t->sched_info.last_arrival;
 286
 287         rq_sched_info_depart(rq, delta);
 288
 289         if (task_is_running(t))
 290                 sched_info_enqueue(rq, t);
 291 }
 292
 293 /*
 294  * Called when tasks are switched involuntarily due, typically, to expiring
 295  * their time slice.  (This may also be called when switching to or from
 296  * the idle task.)  We are only called when prev != next.
 297  */
 298 static inline void
 299 sched_info_switch(struct rq *rq, struct task_struct *prev, struct task_struct *next)
 300 {
 301         /*
 302          * prev now departs the CPU.  It's not interesting to record
 303          * stats about how efficient we were at scheduling the idle
 304          * process, however.
 305          */
 306         if (prev != rq->idle)
 307                 sched_info_depart(rq, prev);
 308
 309         if (next != rq->idle)
 310                 sched_info_arrive(rq, next);
 311 }
 312
 313 #else /* !CONFIG_SCHED_INFO: */
 314 # define sched_info_enqueue(rq, t)      do { } while (0)
 315 # define sched_info_dequeue(rq, t)      do { } while (0)
 316 # define sched_info_switch(rq, t, next) do { } while (0)
 317 #endif /* CONFIG_SCHED_INFO */
 318
 319 #endif /* _KERNEL_STATS_H */