xref: /linux/kernel/sched/stats.h (revision 26fbb4c8c7c3ee9a4c3b4de555a8587b5a19154e)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 
3 #ifdef CONFIG_SCHEDSTATS
4 
5 /*
6  * Expects runqueue lock to be held for atomicity of update
7  */
8 static inline void
9 rq_sched_info_arrive(struct rq *rq, unsigned long long delta)
10 {
11 	if (rq) {
12 		rq->rq_sched_info.run_delay += delta;
13 		rq->rq_sched_info.pcount++;
14 	}
15 }
16 
17 /*
18  * Expects runqueue lock to be held for atomicity of update
19  */
20 static inline void
21 rq_sched_info_depart(struct rq *rq, unsigned long long delta)
22 {
23 	if (rq)
24 		rq->rq_cpu_time += delta;
25 }
26 
27 static inline void
28 rq_sched_info_dequeued(struct rq *rq, unsigned long long delta)
29 {
30 	if (rq)
31 		rq->rq_sched_info.run_delay += delta;
32 }
33 #define   schedstat_enabled()		static_branch_unlikely(&sched_schedstats)
34 #define __schedstat_inc(var)		do { var++; } while (0)
35 #define   schedstat_inc(var)		do { if (schedstat_enabled()) { var++; } } while (0)
36 #define __schedstat_add(var, amt)	do { var += (amt); } while (0)
37 #define   schedstat_add(var, amt)	do { if (schedstat_enabled()) { var += (amt); } } while (0)
38 #define __schedstat_set(var, val)	do { var = (val); } while (0)
39 #define   schedstat_set(var, val)	do { if (schedstat_enabled()) { var = (val); } } while (0)
40 #define   schedstat_val(var)		(var)
41 #define   schedstat_val_or_zero(var)	((schedstat_enabled()) ? (var) : 0)
42 
43 #else /* !CONFIG_SCHEDSTATS: */
44 static inline void rq_sched_info_arrive  (struct rq *rq, unsigned long long delta) { }
45 static inline void rq_sched_info_dequeued(struct rq *rq, unsigned long long delta) { }
46 static inline void rq_sched_info_depart  (struct rq *rq, unsigned long long delta) { }
47 # define   schedstat_enabled()		0
48 # define __schedstat_inc(var)		do { } while (0)
49 # define   schedstat_inc(var)		do { } while (0)
50 # define __schedstat_add(var, amt)	do { } while (0)
51 # define   schedstat_add(var, amt)	do { } while (0)
52 # define __schedstat_set(var, val)	do { } while (0)
53 # define   schedstat_set(var, val)	do { } while (0)
54 # define   schedstat_val(var)		0
55 # define   schedstat_val_or_zero(var)	0
56 #endif /* CONFIG_SCHEDSTATS */
57 
58 #ifdef CONFIG_PSI
59 /*
60  * PSI tracks state that persists across sleeps, such as iowaits and
61  * memory stalls. As a result, it has to distinguish between sleeps,
62  * where a task's runnable state changes, and requeues, where a task
63  * and its state are being moved between CPUs and runqueues.
64  */
65 static inline void psi_enqueue(struct task_struct *p, bool wakeup)
66 {
67 	int clear = 0, set = TSK_RUNNING;
68 
69 	if (static_branch_likely(&psi_disabled))
70 		return;
71 
72 	if (!wakeup || p->sched_psi_wake_requeue) {
73 		if (p->in_memstall)
74 			set |= TSK_MEMSTALL;
75 		if (p->sched_psi_wake_requeue)
76 			p->sched_psi_wake_requeue = 0;
77 	} else {
78 		if (p->in_iowait)
79 			clear |= TSK_IOWAIT;
80 	}
81 
82 	psi_task_change(p, clear, set);
83 }
84 
85 static inline void psi_dequeue(struct task_struct *p, bool sleep)
86 {
87 	int clear = TSK_RUNNING, set = 0;
88 
89 	if (static_branch_likely(&psi_disabled))
90 		return;
91 
92 	if (!sleep) {
93 		if (p->in_memstall)
94 			clear |= TSK_MEMSTALL;
95 	} else {
96 		/*
97 		 * When a task sleeps, schedule() dequeues it before
98 		 * switching to the next one. Merge the clearing of
99 		 * TSK_RUNNING and TSK_ONCPU to save an unnecessary
100 		 * psi_task_change() call in psi_sched_switch().
101 		 */
102 		clear |= TSK_ONCPU;
103 
104 		if (p->in_iowait)
105 			set |= TSK_IOWAIT;
106 	}
107 
108 	psi_task_change(p, clear, set);
109 }
110 
111 static inline void psi_ttwu_dequeue(struct task_struct *p)
112 {
113 	if (static_branch_likely(&psi_disabled))
114 		return;
115 	/*
116 	 * Is the task being migrated during a wakeup? Make sure to
117 	 * deregister its sleep-persistent psi states from the old
118 	 * queue, and let psi_enqueue() know it has to requeue.
119 	 */
120 	if (unlikely(p->in_iowait || p->in_memstall)) {
121 		struct rq_flags rf;
122 		struct rq *rq;
123 		int clear = 0;
124 
125 		if (p->in_iowait)
126 			clear |= TSK_IOWAIT;
127 		if (p->in_memstall)
128 			clear |= TSK_MEMSTALL;
129 
130 		rq = __task_rq_lock(p, &rf);
131 		psi_task_change(p, clear, 0);
132 		p->sched_psi_wake_requeue = 1;
133 		__task_rq_unlock(rq, &rf);
134 	}
135 }
136 
137 static inline void psi_sched_switch(struct task_struct *prev,
138 				    struct task_struct *next,
139 				    bool sleep)
140 {
141 	if (static_branch_likely(&psi_disabled))
142 		return;
143 
144 	psi_task_switch(prev, next, sleep);
145 }
146 
147 static inline void psi_task_tick(struct rq *rq)
148 {
149 	if (static_branch_likely(&psi_disabled))
150 		return;
151 
152 	if (unlikely(rq->curr->in_memstall))
153 		psi_memstall_tick(rq->curr, cpu_of(rq));
154 }
155 #else /* CONFIG_PSI */
156 static inline void psi_enqueue(struct task_struct *p, bool wakeup) {}
157 static inline void psi_dequeue(struct task_struct *p, bool sleep) {}
158 static inline void psi_ttwu_dequeue(struct task_struct *p) {}
159 static inline void psi_sched_switch(struct task_struct *prev,
160 				    struct task_struct *next,
161 				    bool sleep) {}
162 static inline void psi_task_tick(struct rq *rq) {}
163 #endif /* CONFIG_PSI */
164 
165 #ifdef CONFIG_SCHED_INFO
166 static inline void sched_info_reset_dequeued(struct task_struct *t)
167 {
168 	t->sched_info.last_queued = 0;
169 }
170 
171 /*
172  * We are interested in knowing how long it was from the *first* time a
173  * task was queued to the time that it finally hit a CPU, we call this routine
174  * from dequeue_task() to account for possible rq->clock skew across CPUs. The
175  * delta taken on each CPU would annul the skew.
176  */
177 static inline void sched_info_dequeued(struct rq *rq, struct task_struct *t)
178 {
179 	unsigned long long now = rq_clock(rq), delta = 0;
180 
181 	if (sched_info_on()) {
182 		if (t->sched_info.last_queued)
183 			delta = now - t->sched_info.last_queued;
184 	}
185 	sched_info_reset_dequeued(t);
186 	t->sched_info.run_delay += delta;
187 
188 	rq_sched_info_dequeued(rq, delta);
189 }
190 
191 /*
192  * Called when a task finally hits the CPU.  We can now calculate how
193  * long it was waiting to run.  We also note when it began so that we
194  * can keep stats on how long its timeslice is.
195  */
196 static void sched_info_arrive(struct rq *rq, struct task_struct *t)
197 {
198 	unsigned long long now = rq_clock(rq), delta = 0;
199 
200 	if (t->sched_info.last_queued)
201 		delta = now - t->sched_info.last_queued;
202 	sched_info_reset_dequeued(t);
203 	t->sched_info.run_delay += delta;
204 	t->sched_info.last_arrival = now;
205 	t->sched_info.pcount++;
206 
207 	rq_sched_info_arrive(rq, delta);
208 }
209 
210 /*
211  * This function is only called from enqueue_task(), but also only updates
212  * the timestamp if it is already not set.  It's assumed that
213  * sched_info_dequeued() will clear that stamp when appropriate.
214  */
215 static inline void sched_info_queued(struct rq *rq, struct task_struct *t)
216 {
217 	if (sched_info_on()) {
218 		if (!t->sched_info.last_queued)
219 			t->sched_info.last_queued = rq_clock(rq);
220 	}
221 }
222 
223 /*
224  * Called when a process ceases being the active-running process involuntarily
225  * due, typically, to expiring its time slice (this may also be called when
226  * switching to the idle task).  Now we can calculate how long we ran.
227  * Also, if the process is still in the TASK_RUNNING state, call
228  * sched_info_queued() to mark that it has now again started waiting on
229  * the runqueue.
230  */
231 static inline void sched_info_depart(struct rq *rq, struct task_struct *t)
232 {
233 	unsigned long long delta = rq_clock(rq) - t->sched_info.last_arrival;
234 
235 	rq_sched_info_depart(rq, delta);
236 
237 	if (t->state == TASK_RUNNING)
238 		sched_info_queued(rq, t);
239 }
240 
241 /*
242  * Called when tasks are switched involuntarily due, typically, to expiring
243  * their time slice.  (This may also be called when switching to or from
244  * the idle task.)  We are only called when prev != next.
245  */
246 static inline void
247 __sched_info_switch(struct rq *rq, struct task_struct *prev, struct task_struct *next)
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(rq, prev);
256 
257 	if (next != rq->idle)
258 		sched_info_arrive(rq, next);
259 }
260 
261 static inline void
262 sched_info_switch(struct rq *rq, struct task_struct *prev, struct task_struct *next)
263 {
264 	if (sched_info_on())
265 		__sched_info_switch(rq, prev, next);
266 }
267 
268 #else /* !CONFIG_SCHED_INFO: */
269 # define sched_info_queued(rq, t)	do { } while (0)
270 # define sched_info_reset_dequeued(t)	do { } while (0)
271 # define sched_info_dequeued(rq, t)	do { } while (0)
272 # define sched_info_depart(rq, t)	do { } while (0)
273 # define sched_info_arrive(rq, next)	do { } while (0)
274 # define sched_info_switch(rq, t, next)	do { } while (0)
275 #endif /* CONFIG_SCHED_INFO */
276