| 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 /* !CONFIG_IRQ_TIME_ACCOUNTING: */ |
| 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, int flags) |
| 131 | { |
| 132 | int clear = 0, set = 0; |
| 133 | |
| 134 | if (static_branch_likely(&psi_disabled)) |
| 135 | return; |
| 136 | |
| 137 | /* Same runqueue, nothing changed for psi */ |
| 138 | if (flags & ENQUEUE_RESTORE) |
| 139 | return; |
| 140 | |
| 141 | /* psi_sched_switch() will handle the flags */ |
| 142 | if (task_on_cpu(task_rq(p), p)) |
| 143 | return; |
| 144 | |
| 145 | if (p->se.sched_delayed) { |
| 146 | /* CPU migration of "sleeping" task */ |
| 147 | WARN_ON_ONCE(!(flags & ENQUEUE_MIGRATED)); |
| 148 | if (p->in_memstall) |
| 149 | set |= TSK_MEMSTALL; |
| 150 | if (p->in_iowait) |
| 151 | set |= TSK_IOWAIT; |
| 152 | } else if (flags & ENQUEUE_MIGRATED) { |
| 153 | /* CPU migration of runnable task */ |
| 154 | set = TSK_RUNNING; |
| 155 | if (p->in_memstall) |
| 156 | set |= TSK_MEMSTALL | TSK_MEMSTALL_RUNNING; |
| 157 | } else { |
| 158 | /* Wakeup of new or sleeping task */ |
| 159 | if (p->in_iowait) |
| 160 | clear |= TSK_IOWAIT; |
| 161 | set = TSK_RUNNING; |
| 162 | if (p->in_memstall) |
| 163 | set |= TSK_MEMSTALL_RUNNING; |
| 164 | } |
| 165 | |
| 166 | psi_task_change(task: p, clear, set); |
| 167 | } |
| 168 | |
| 169 | static inline void psi_dequeue(struct task_struct *p, int flags) |
| 170 | { |
| 171 | if (static_branch_likely(&psi_disabled)) |
| 172 | return; |
| 173 | |
| 174 | /* Same runqueue, nothing changed for psi */ |
| 175 | if (flags & DEQUEUE_SAVE) |
| 176 | return; |
| 177 | |
| 178 | /* |
| 179 | * A voluntary sleep is a dequeue followed by a task switch. To |
| 180 | * avoid walking all ancestors twice, psi_task_switch() handles |
| 181 | * TSK_RUNNING and TSK_IOWAIT for us when it moves TSK_ONCPU. |
| 182 | * Do nothing here. |
| 183 | * |
| 184 | * In the SCHED_PROXY_EXECUTION case we may do sleeping |
| 185 | * dequeues that are not followed by a task switch, so check |
| 186 | * TSK_ONCPU is set to ensure the task switch is imminent. |
| 187 | * Otherwise clear the flags as usual. |
| 188 | */ |
| 189 | if ((flags & DEQUEUE_SLEEP) && (p->psi_flags & TSK_ONCPU)) |
| 190 | return; |
| 191 | |
| 192 | /* |
| 193 | * When migrating a task to another CPU, clear all psi |
| 194 | * state. The enqueue callback above will work it out. |
| 195 | */ |
| 196 | psi_task_change(task: p, clear: p->psi_flags, set: 0); |
| 197 | } |
| 198 | |
| 199 | static inline void psi_ttwu_dequeue(struct task_struct *p) |
| 200 | { |
| 201 | if (static_branch_likely(&psi_disabled)) |
| 202 | return; |
| 203 | /* |
| 204 | * Is the task being migrated during a wakeup? Make sure to |
| 205 | * deregister its sleep-persistent psi states from the old |
| 206 | * queue, and let psi_enqueue() know it has to requeue. |
| 207 | */ |
| 208 | if (unlikely(p->psi_flags)) { |
| 209 | struct rq_flags rf; |
| 210 | struct rq *rq; |
| 211 | |
| 212 | rq = __task_rq_lock(p, rf: &rf); |
| 213 | psi_task_change(task: p, clear: p->psi_flags, set: 0); |
| 214 | __task_rq_unlock(rq, p, rf: &rf); |
| 215 | } |
| 216 | } |
| 217 | |
| 218 | static inline void psi_sched_switch(struct task_struct *prev, |
| 219 | struct task_struct *next, |
| 220 | bool sleep) |
| 221 | { |
| 222 | if (static_branch_likely(&psi_disabled)) |
| 223 | return; |
| 224 | |
| 225 | psi_task_switch(prev, next, sleep); |
| 226 | } |
| 227 | |
| 228 | #else /* !CONFIG_PSI: */ |
| 229 | static inline void psi_enqueue(struct task_struct *p, bool migrate) {} |
| 230 | static inline void psi_dequeue(struct task_struct *p, bool migrate) {} |
| 231 | static inline void psi_ttwu_dequeue(struct task_struct *p) {} |
| 232 | static inline void psi_sched_switch(struct task_struct *prev, |
| 233 | struct task_struct *next, |
| 234 | bool sleep) {} |
| 235 | static inline void psi_account_irqtime(struct rq *rq, struct task_struct *curr, |
| 236 | struct task_struct *prev) {} |
| 237 | #endif /* !CONFIG_PSI */ |
| 238 | |
| 239 | #ifdef CONFIG_SCHED_INFO |
| 240 | /* |
| 241 | * We are interested in knowing how long it was from the *first* time a |
| 242 | * task was queued to the time that it finally hit a CPU, we call this routine |
| 243 | * from dequeue_task() to account for possible rq->clock skew across CPUs. The |
| 244 | * delta taken on each CPU would annul the skew. |
| 245 | */ |
| 246 | static inline void sched_info_dequeue(struct rq *rq, struct task_struct *t) |
| 247 | { |
| 248 | unsigned long long delta = 0; |
| 249 | |
| 250 | if (!t->sched_info.last_queued) |
| 251 | return; |
| 252 | |
| 253 | delta = rq_clock(rq) - t->sched_info.last_queued; |
| 254 | t->sched_info.last_queued = 0; |
| 255 | t->sched_info.run_delay += delta; |
| 256 | if (delta > t->sched_info.max_run_delay) |
| 257 | t->sched_info.max_run_delay = delta; |
| 258 | if (delta && (!t->sched_info.min_run_delay || delta < t->sched_info.min_run_delay)) |
| 259 | t->sched_info.min_run_delay = delta; |
| 260 | rq_sched_info_dequeue(rq, delta); |
| 261 | } |
| 262 | |
| 263 | /* |
| 264 | * Called when a task finally hits the CPU. We can now calculate how |
| 265 | * long it was waiting to run. We also note when it began so that we |
| 266 | * can keep stats on how long its time-slice is. |
| 267 | */ |
| 268 | static void sched_info_arrive(struct rq *rq, struct task_struct *t) |
| 269 | { |
| 270 | unsigned long long now, delta = 0; |
| 271 | |
| 272 | if (!t->sched_info.last_queued) |
| 273 | return; |
| 274 | |
| 275 | now = rq_clock(rq); |
| 276 | delta = now - t->sched_info.last_queued; |
| 277 | t->sched_info.last_queued = 0; |
| 278 | t->sched_info.run_delay += delta; |
| 279 | t->sched_info.last_arrival = now; |
| 280 | t->sched_info.pcount++; |
| 281 | if (delta > t->sched_info.max_run_delay) |
| 282 | t->sched_info.max_run_delay = delta; |
| 283 | if (delta && (!t->sched_info.min_run_delay || delta < t->sched_info.min_run_delay)) |
| 284 | t->sched_info.min_run_delay = delta; |
| 285 | |
| 286 | rq_sched_info_arrive(rq, delta); |
| 287 | } |
| 288 | |
| 289 | /* |
| 290 | * This function is only called from enqueue_task(), but also only updates |
| 291 | * the timestamp if it is already not set. It's assumed that |
| 292 | * sched_info_dequeue() will clear that stamp when appropriate. |
| 293 | */ |
| 294 | static inline void sched_info_enqueue(struct rq *rq, struct task_struct *t) |
| 295 | { |
| 296 | if (!t->sched_info.last_queued) |
| 297 | t->sched_info.last_queued = rq_clock(rq); |
| 298 | } |
| 299 | |
| 300 | /* |
| 301 | * Called when a process ceases being the active-running process involuntarily |
| 302 | * due, typically, to expiring its time slice (this may also be called when |
| 303 | * switching to the idle task). Now we can calculate how long we ran. |
| 304 | * Also, if the process is still in the TASK_RUNNING state, call |
| 305 | * sched_info_enqueue() to mark that it has now again started waiting on |
| 306 | * the runqueue. |
| 307 | */ |
| 308 | static inline void sched_info_depart(struct rq *rq, struct task_struct *t) |
| 309 | { |
| 310 | unsigned long long delta = rq_clock(rq) - t->sched_info.last_arrival; |
| 311 | |
| 312 | rq_sched_info_depart(rq, delta); |
| 313 | |
| 314 | if (task_is_running(t)) |
| 315 | sched_info_enqueue(rq, t); |
| 316 | } |
| 317 | |
| 318 | /* |
| 319 | * Called when tasks are switched involuntarily due, typically, to expiring |
| 320 | * their time slice. (This may also be called when switching to or from |
| 321 | * the idle task.) We are only called when prev != next. |
| 322 | */ |
| 323 | static inline void |
| 324 | sched_info_switch(struct rq *rq, struct task_struct *prev, struct task_struct *next) |
| 325 | { |
| 326 | /* |
| 327 | * prev now departs the CPU. It's not interesting to record |
| 328 | * stats about how efficient we were at scheduling the idle |
| 329 | * process, however. |
| 330 | */ |
| 331 | if (prev != rq->idle) |
| 332 | sched_info_depart(rq, t: prev); |
| 333 | |
| 334 | if (next != rq->idle) |
| 335 | sched_info_arrive(rq, t: next); |
| 336 | } |
| 337 | |
| 338 | #else /* !CONFIG_SCHED_INFO: */ |
| 339 | # define sched_info_enqueue(rq, t) do { } while (0) |
| 340 | # define sched_info_dequeue(rq, t) do { } while (0) |
| 341 | # define sched_info_switch(rq, t, next) do { } while (0) |
| 342 | #endif /* !CONFIG_SCHED_INFO */ |
| 343 | |
| 344 | #endif /* _KERNEL_STATS_H */ |
| 345 | |