Commit | Line | Data |
---|---|---|
73fbec60 FW |
1 | #include <linux/export.h> |
2 | #include <linux/sched.h> | |
3 | #include <linux/tsacct_kern.h> | |
4 | #include <linux/kernel_stat.h> | |
5 | #include <linux/static_key.h> | |
abf917cd | 6 | #include <linux/context_tracking.h> |
73fbec60 FW |
7 | #include "sched.h" |
8 | ||
9 | ||
10 | #ifdef CONFIG_IRQ_TIME_ACCOUNTING | |
11 | ||
12 | /* | |
13 | * There are no locks covering percpu hardirq/softirq time. | |
bf9fae9f | 14 | * They are only modified in vtime_account, on corresponding CPU |
73fbec60 FW |
15 | * with interrupts disabled. So, writes are safe. |
16 | * They are read and saved off onto struct rq in update_rq_clock(). | |
17 | * This may result in other CPU reading this CPU's irq time and can | |
bf9fae9f | 18 | * race with irq/vtime_account on this CPU. We would either get old |
73fbec60 FW |
19 | * or new value with a side effect of accounting a slice of irq time to wrong |
20 | * task when irq is in progress while we read rq->clock. That is a worthy | |
21 | * compromise in place of having locks on each irq in account_system_time. | |
22 | */ | |
23 | DEFINE_PER_CPU(u64, cpu_hardirq_time); | |
24 | DEFINE_PER_CPU(u64, cpu_softirq_time); | |
25 | ||
26 | static DEFINE_PER_CPU(u64, irq_start_time); | |
27 | static int sched_clock_irqtime; | |
28 | ||
29 | void enable_sched_clock_irqtime(void) | |
30 | { | |
31 | sched_clock_irqtime = 1; | |
32 | } | |
33 | ||
34 | void disable_sched_clock_irqtime(void) | |
35 | { | |
36 | sched_clock_irqtime = 0; | |
37 | } | |
38 | ||
39 | #ifndef CONFIG_64BIT | |
40 | DEFINE_PER_CPU(seqcount_t, irq_time_seq); | |
41 | #endif /* CONFIG_64BIT */ | |
42 | ||
43 | /* | |
44 | * Called before incrementing preempt_count on {soft,}irq_enter | |
45 | * and before decrementing preempt_count on {soft,}irq_exit. | |
46 | */ | |
3e1df4f5 | 47 | void irqtime_account_irq(struct task_struct *curr) |
73fbec60 FW |
48 | { |
49 | unsigned long flags; | |
50 | s64 delta; | |
51 | int cpu; | |
52 | ||
53 | if (!sched_clock_irqtime) | |
54 | return; | |
55 | ||
56 | local_irq_save(flags); | |
57 | ||
58 | cpu = smp_processor_id(); | |
59 | delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time); | |
60 | __this_cpu_add(irq_start_time, delta); | |
61 | ||
62 | irq_time_write_begin(); | |
63 | /* | |
64 | * We do not account for softirq time from ksoftirqd here. | |
65 | * We want to continue accounting softirq time to ksoftirqd thread | |
66 | * in that case, so as not to confuse scheduler with a special task | |
67 | * that do not consume any time, but still wants to run. | |
68 | */ | |
69 | if (hardirq_count()) | |
70 | __this_cpu_add(cpu_hardirq_time, delta); | |
71 | else if (in_serving_softirq() && curr != this_cpu_ksoftirqd()) | |
72 | __this_cpu_add(cpu_softirq_time, delta); | |
73 | ||
74 | irq_time_write_end(); | |
75 | local_irq_restore(flags); | |
76 | } | |
3e1df4f5 | 77 | EXPORT_SYMBOL_GPL(irqtime_account_irq); |
73fbec60 FW |
78 | |
79 | static int irqtime_account_hi_update(void) | |
80 | { | |
81 | u64 *cpustat = kcpustat_this_cpu->cpustat; | |
82 | unsigned long flags; | |
83 | u64 latest_ns; | |
84 | int ret = 0; | |
85 | ||
86 | local_irq_save(flags); | |
87 | latest_ns = this_cpu_read(cpu_hardirq_time); | |
88 | if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ]) | |
89 | ret = 1; | |
90 | local_irq_restore(flags); | |
91 | return ret; | |
92 | } | |
93 | ||
94 | static int irqtime_account_si_update(void) | |
95 | { | |
96 | u64 *cpustat = kcpustat_this_cpu->cpustat; | |
97 | unsigned long flags; | |
98 | u64 latest_ns; | |
99 | int ret = 0; | |
100 | ||
101 | local_irq_save(flags); | |
102 | latest_ns = this_cpu_read(cpu_softirq_time); | |
103 | if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ]) | |
104 | ret = 1; | |
105 | local_irq_restore(flags); | |
106 | return ret; | |
107 | } | |
108 | ||
109 | #else /* CONFIG_IRQ_TIME_ACCOUNTING */ | |
110 | ||
111 | #define sched_clock_irqtime (0) | |
112 | ||
113 | #endif /* !CONFIG_IRQ_TIME_ACCOUNTING */ | |
114 | ||
115 | static inline void task_group_account_field(struct task_struct *p, int index, | |
116 | u64 tmp) | |
117 | { | |
73fbec60 FW |
118 | /* |
119 | * Since all updates are sure to touch the root cgroup, we | |
120 | * get ourselves ahead and touch it first. If the root cgroup | |
121 | * is the only cgroup, then nothing else should be necessary. | |
122 | * | |
123 | */ | |
a4f61cc0 | 124 | __this_cpu_add(kernel_cpustat.cpustat[index], tmp); |
73fbec60 | 125 | |
1966aaf7 | 126 | cpuacct_account_field(p, index, tmp); |
73fbec60 FW |
127 | } |
128 | ||
129 | /* | |
130 | * Account user cpu time to a process. | |
131 | * @p: the process that the cpu time gets accounted to | |
132 | * @cputime: the cpu time spent in user space since the last update | |
133 | * @cputime_scaled: cputime scaled by cpu frequency | |
134 | */ | |
135 | void account_user_time(struct task_struct *p, cputime_t cputime, | |
136 | cputime_t cputime_scaled) | |
137 | { | |
138 | int index; | |
139 | ||
140 | /* Add user time to process. */ | |
141 | p->utime += cputime; | |
142 | p->utimescaled += cputime_scaled; | |
143 | account_group_user_time(p, cputime); | |
144 | ||
145 | index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER; | |
146 | ||
147 | /* Add user time to cpustat. */ | |
148 | task_group_account_field(p, index, (__force u64) cputime); | |
149 | ||
150 | /* Account for user time used */ | |
6fac4829 | 151 | acct_account_cputime(p); |
73fbec60 FW |
152 | } |
153 | ||
154 | /* | |
155 | * Account guest cpu time to a process. | |
156 | * @p: the process that the cpu time gets accounted to | |
157 | * @cputime: the cpu time spent in virtual machine since the last update | |
158 | * @cputime_scaled: cputime scaled by cpu frequency | |
159 | */ | |
160 | static void account_guest_time(struct task_struct *p, cputime_t cputime, | |
161 | cputime_t cputime_scaled) | |
162 | { | |
163 | u64 *cpustat = kcpustat_this_cpu->cpustat; | |
164 | ||
165 | /* Add guest time to process. */ | |
166 | p->utime += cputime; | |
167 | p->utimescaled += cputime_scaled; | |
168 | account_group_user_time(p, cputime); | |
169 | p->gtime += cputime; | |
170 | ||
171 | /* Add guest time to cpustat. */ | |
172 | if (TASK_NICE(p) > 0) { | |
173 | cpustat[CPUTIME_NICE] += (__force u64) cputime; | |
174 | cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime; | |
175 | } else { | |
176 | cpustat[CPUTIME_USER] += (__force u64) cputime; | |
177 | cpustat[CPUTIME_GUEST] += (__force u64) cputime; | |
178 | } | |
179 | } | |
180 | ||
181 | /* | |
182 | * Account system cpu time to a process and desired cpustat field | |
183 | * @p: the process that the cpu time gets accounted to | |
184 | * @cputime: the cpu time spent in kernel space since the last update | |
185 | * @cputime_scaled: cputime scaled by cpu frequency | |
186 | * @target_cputime64: pointer to cpustat field that has to be updated | |
187 | */ | |
188 | static inline | |
189 | void __account_system_time(struct task_struct *p, cputime_t cputime, | |
190 | cputime_t cputime_scaled, int index) | |
191 | { | |
192 | /* Add system time to process. */ | |
193 | p->stime += cputime; | |
194 | p->stimescaled += cputime_scaled; | |
195 | account_group_system_time(p, cputime); | |
196 | ||
197 | /* Add system time to cpustat. */ | |
198 | task_group_account_field(p, index, (__force u64) cputime); | |
199 | ||
200 | /* Account for system time used */ | |
6fac4829 | 201 | acct_account_cputime(p); |
73fbec60 FW |
202 | } |
203 | ||
204 | /* | |
205 | * Account system cpu time to a process. | |
206 | * @p: the process that the cpu time gets accounted to | |
207 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
208 | * @cputime: the cpu time spent in kernel space since the last update | |
209 | * @cputime_scaled: cputime scaled by cpu frequency | |
210 | */ | |
211 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
212 | cputime_t cputime, cputime_t cputime_scaled) | |
213 | { | |
214 | int index; | |
215 | ||
216 | if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { | |
217 | account_guest_time(p, cputime, cputime_scaled); | |
218 | return; | |
219 | } | |
220 | ||
221 | if (hardirq_count() - hardirq_offset) | |
222 | index = CPUTIME_IRQ; | |
223 | else if (in_serving_softirq()) | |
224 | index = CPUTIME_SOFTIRQ; | |
225 | else | |
226 | index = CPUTIME_SYSTEM; | |
227 | ||
228 | __account_system_time(p, cputime, cputime_scaled, index); | |
229 | } | |
230 | ||
231 | /* | |
232 | * Account for involuntary wait time. | |
233 | * @cputime: the cpu time spent in involuntary wait | |
234 | */ | |
235 | void account_steal_time(cputime_t cputime) | |
236 | { | |
237 | u64 *cpustat = kcpustat_this_cpu->cpustat; | |
238 | ||
239 | cpustat[CPUTIME_STEAL] += (__force u64) cputime; | |
240 | } | |
241 | ||
242 | /* | |
243 | * Account for idle time. | |
244 | * @cputime: the cpu time spent in idle wait | |
245 | */ | |
246 | void account_idle_time(cputime_t cputime) | |
247 | { | |
248 | u64 *cpustat = kcpustat_this_cpu->cpustat; | |
249 | struct rq *rq = this_rq(); | |
250 | ||
251 | if (atomic_read(&rq->nr_iowait) > 0) | |
252 | cpustat[CPUTIME_IOWAIT] += (__force u64) cputime; | |
253 | else | |
254 | cpustat[CPUTIME_IDLE] += (__force u64) cputime; | |
255 | } | |
256 | ||
257 | static __always_inline bool steal_account_process_tick(void) | |
258 | { | |
259 | #ifdef CONFIG_PARAVIRT | |
260 | if (static_key_false(¶virt_steal_enabled)) { | |
261 | u64 steal, st = 0; | |
262 | ||
263 | steal = paravirt_steal_clock(smp_processor_id()); | |
264 | steal -= this_rq()->prev_steal_time; | |
265 | ||
266 | st = steal_ticks(steal); | |
267 | this_rq()->prev_steal_time += st * TICK_NSEC; | |
268 | ||
269 | account_steal_time(st); | |
270 | return st; | |
271 | } | |
272 | #endif | |
273 | return false; | |
274 | } | |
275 | ||
a634f933 FW |
276 | /* |
277 | * Accumulate raw cputime values of dead tasks (sig->[us]time) and live | |
278 | * tasks (sum on group iteration) belonging to @tsk's group. | |
279 | */ | |
280 | void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times) | |
281 | { | |
282 | struct signal_struct *sig = tsk->signal; | |
6fac4829 | 283 | cputime_t utime, stime; |
a634f933 FW |
284 | struct task_struct *t; |
285 | ||
286 | times->utime = sig->utime; | |
287 | times->stime = sig->stime; | |
288 | times->sum_exec_runtime = sig->sum_sched_runtime; | |
289 | ||
290 | rcu_read_lock(); | |
291 | /* make sure we can trust tsk->thread_group list */ | |
292 | if (!likely(pid_alive(tsk))) | |
293 | goto out; | |
294 | ||
295 | t = tsk; | |
296 | do { | |
e614b333 | 297 | task_cputime(t, &utime, &stime); |
6fac4829 FW |
298 | times->utime += utime; |
299 | times->stime += stime; | |
a634f933 FW |
300 | times->sum_exec_runtime += task_sched_runtime(t); |
301 | } while_each_thread(tsk, t); | |
302 | out: | |
303 | rcu_read_unlock(); | |
304 | } | |
305 | ||
73fbec60 FW |
306 | #ifdef CONFIG_IRQ_TIME_ACCOUNTING |
307 | /* | |
308 | * Account a tick to a process and cpustat | |
309 | * @p: the process that the cpu time gets accounted to | |
310 | * @user_tick: is the tick from userspace | |
311 | * @rq: the pointer to rq | |
312 | * | |
313 | * Tick demultiplexing follows the order | |
314 | * - pending hardirq update | |
315 | * - pending softirq update | |
316 | * - user_time | |
317 | * - idle_time | |
318 | * - system time | |
319 | * - check for guest_time | |
320 | * - else account as system_time | |
321 | * | |
322 | * Check for hardirq is done both for system and user time as there is | |
323 | * no timer going off while we are on hardirq and hence we may never get an | |
324 | * opportunity to update it solely in system time. | |
325 | * p->stime and friends are only updated on system time and not on irq | |
326 | * softirq as those do not count in task exec_runtime any more. | |
327 | */ | |
328 | static void irqtime_account_process_tick(struct task_struct *p, int user_tick, | |
329 | struct rq *rq) | |
330 | { | |
331 | cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); | |
332 | u64 *cpustat = kcpustat_this_cpu->cpustat; | |
333 | ||
334 | if (steal_account_process_tick()) | |
335 | return; | |
336 | ||
337 | if (irqtime_account_hi_update()) { | |
338 | cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy; | |
339 | } else if (irqtime_account_si_update()) { | |
340 | cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy; | |
341 | } else if (this_cpu_ksoftirqd() == p) { | |
342 | /* | |
343 | * ksoftirqd time do not get accounted in cpu_softirq_time. | |
344 | * So, we have to handle it separately here. | |
345 | * Also, p->stime needs to be updated for ksoftirqd. | |
346 | */ | |
347 | __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, | |
348 | CPUTIME_SOFTIRQ); | |
349 | } else if (user_tick) { | |
350 | account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); | |
351 | } else if (p == rq->idle) { | |
352 | account_idle_time(cputime_one_jiffy); | |
353 | } else if (p->flags & PF_VCPU) { /* System time or guest time */ | |
354 | account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled); | |
355 | } else { | |
356 | __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, | |
357 | CPUTIME_SYSTEM); | |
358 | } | |
359 | } | |
360 | ||
361 | static void irqtime_account_idle_ticks(int ticks) | |
362 | { | |
363 | int i; | |
364 | struct rq *rq = this_rq(); | |
365 | ||
366 | for (i = 0; i < ticks; i++) | |
367 | irqtime_account_process_tick(current, 0, rq); | |
368 | } | |
369 | #else /* CONFIG_IRQ_TIME_ACCOUNTING */ | |
3f4724ea FW |
370 | static inline void irqtime_account_idle_ticks(int ticks) {} |
371 | static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick, | |
73fbec60 FW |
372 | struct rq *rq) {} |
373 | #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ | |
374 | ||
73fbec60 FW |
375 | /* |
376 | * Use precise platform statistics if available: | |
377 | */ | |
378 | #ifdef CONFIG_VIRT_CPU_ACCOUNTING | |
a7e1a9e3 | 379 | |
e3942ba0 | 380 | #ifndef __ARCH_HAS_VTIME_TASK_SWITCH |
b0493406 | 381 | void vtime_common_task_switch(struct task_struct *prev) |
e3942ba0 FW |
382 | { |
383 | if (is_idle_task(prev)) | |
384 | vtime_account_idle(prev); | |
385 | else | |
386 | vtime_account_system(prev); | |
387 | ||
abf917cd | 388 | #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE |
e3942ba0 | 389 | vtime_account_user(prev); |
abf917cd | 390 | #endif |
e3942ba0 FW |
391 | arch_vtime_task_switch(prev); |
392 | } | |
393 | #endif | |
11113334 | 394 | |
a7e1a9e3 FW |
395 | /* |
396 | * Archs that account the whole time spent in the idle task | |
397 | * (outside irq) as idle time can rely on this and just implement | |
fd25b4c2 | 398 | * vtime_account_system() and vtime_account_idle(). Archs that |
a7e1a9e3 FW |
399 | * have other meaning of the idle time (s390 only includes the |
400 | * time spent by the CPU when it's in low power mode) must override | |
401 | * vtime_account(). | |
402 | */ | |
403 | #ifndef __ARCH_HAS_VTIME_ACCOUNT | |
b0493406 | 404 | void vtime_common_account_irq_enter(struct task_struct *tsk) |
a7e1a9e3 | 405 | { |
abf917cd FW |
406 | if (!in_interrupt()) { |
407 | /* | |
408 | * If we interrupted user, context_tracking_in_user() | |
409 | * is 1 because the context tracking don't hook | |
410 | * on irq entry/exit. This way we know if | |
411 | * we need to flush user time on kernel entry. | |
412 | */ | |
413 | if (context_tracking_in_user()) { | |
414 | vtime_account_user(tsk); | |
415 | return; | |
416 | } | |
417 | ||
418 | if (is_idle_task(tsk)) { | |
419 | vtime_account_idle(tsk); | |
420 | return; | |
421 | } | |
422 | } | |
423 | vtime_account_system(tsk); | |
a7e1a9e3 | 424 | } |
b0493406 | 425 | EXPORT_SYMBOL_GPL(vtime_common_account_irq_enter); |
a7e1a9e3 | 426 | #endif /* __ARCH_HAS_VTIME_ACCOUNT */ |
9fbc42ea FW |
427 | #endif /* CONFIG_VIRT_CPU_ACCOUNTING */ |
428 | ||
429 | ||
430 | #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE | |
431 | void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st) | |
432 | { | |
433 | *ut = p->utime; | |
434 | *st = p->stime; | |
435 | } | |
a7e1a9e3 | 436 | |
9fbc42ea FW |
437 | void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st) |
438 | { | |
439 | struct task_cputime cputime; | |
73fbec60 | 440 | |
9fbc42ea FW |
441 | thread_group_cputime(p, &cputime); |
442 | ||
443 | *ut = cputime.utime; | |
444 | *st = cputime.stime; | |
445 | } | |
446 | #else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */ | |
447 | /* | |
448 | * Account a single tick of cpu time. | |
449 | * @p: the process that the cpu time gets accounted to | |
450 | * @user_tick: indicates if the tick is a user or a system tick | |
451 | */ | |
452 | void account_process_tick(struct task_struct *p, int user_tick) | |
73fbec60 | 453 | { |
9fbc42ea FW |
454 | cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); |
455 | struct rq *rq = this_rq(); | |
73fbec60 | 456 | |
9fbc42ea FW |
457 | if (vtime_accounting_enabled()) |
458 | return; | |
459 | ||
460 | if (sched_clock_irqtime) { | |
461 | irqtime_account_process_tick(p, user_tick, rq); | |
462 | return; | |
463 | } | |
464 | ||
465 | if (steal_account_process_tick()) | |
466 | return; | |
73fbec60 | 467 | |
9fbc42ea FW |
468 | if (user_tick) |
469 | account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); | |
470 | else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET)) | |
471 | account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy, | |
472 | one_jiffy_scaled); | |
73fbec60 | 473 | else |
9fbc42ea FW |
474 | account_idle_time(cputime_one_jiffy); |
475 | } | |
73fbec60 | 476 | |
9fbc42ea FW |
477 | /* |
478 | * Account multiple ticks of steal time. | |
479 | * @p: the process from which the cpu time has been stolen | |
480 | * @ticks: number of stolen ticks | |
481 | */ | |
482 | void account_steal_ticks(unsigned long ticks) | |
483 | { | |
484 | account_steal_time(jiffies_to_cputime(ticks)); | |
485 | } | |
486 | ||
487 | /* | |
488 | * Account multiple ticks of idle time. | |
489 | * @ticks: number of stolen ticks | |
490 | */ | |
491 | void account_idle_ticks(unsigned long ticks) | |
492 | { | |
493 | ||
494 | if (sched_clock_irqtime) { | |
495 | irqtime_account_idle_ticks(ticks); | |
496 | return; | |
497 | } | |
498 | ||
499 | account_idle_time(jiffies_to_cputime(ticks)); | |
500 | } | |
73fbec60 | 501 | |
d9a3c982 | 502 | /* |
55eaa7c1 SG |
503 | * Perform (stime * rtime) / total, but avoid multiplication overflow by |
504 | * loosing precision when the numbers are big. | |
d9a3c982 FW |
505 | */ |
506 | static cputime_t scale_stime(u64 stime, u64 rtime, u64 total) | |
73fbec60 | 507 | { |
55eaa7c1 | 508 | u64 scaled; |
73fbec60 | 509 | |
55eaa7c1 SG |
510 | for (;;) { |
511 | /* Make sure "rtime" is the bigger of stime/rtime */ | |
84f9f3a1 SG |
512 | if (stime > rtime) |
513 | swap(rtime, stime); | |
55eaa7c1 SG |
514 | |
515 | /* Make sure 'total' fits in 32 bits */ | |
516 | if (total >> 32) | |
517 | goto drop_precision; | |
518 | ||
519 | /* Does rtime (and thus stime) fit in 32 bits? */ | |
520 | if (!(rtime >> 32)) | |
521 | break; | |
522 | ||
523 | /* Can we just balance rtime/stime rather than dropping bits? */ | |
524 | if (stime >> 31) | |
525 | goto drop_precision; | |
526 | ||
527 | /* We can grow stime and shrink rtime and try to make them both fit */ | |
528 | stime <<= 1; | |
529 | rtime >>= 1; | |
530 | continue; | |
531 | ||
532 | drop_precision: | |
533 | /* We drop from rtime, it has more bits than stime */ | |
534 | rtime >>= 1; | |
535 | total >>= 1; | |
d9a3c982 | 536 | } |
73fbec60 | 537 | |
55eaa7c1 SG |
538 | /* |
539 | * Make sure gcc understands that this is a 32x32->64 multiply, | |
540 | * followed by a 64/32->64 divide. | |
541 | */ | |
542 | scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total); | |
d9a3c982 | 543 | return (__force cputime_t) scaled; |
73fbec60 FW |
544 | } |
545 | ||
fa092057 FW |
546 | /* |
547 | * Adjust tick based cputime random precision against scheduler | |
548 | * runtime accounting. | |
549 | */ | |
d37f761d FW |
550 | static void cputime_adjust(struct task_cputime *curr, |
551 | struct cputime *prev, | |
552 | cputime_t *ut, cputime_t *st) | |
73fbec60 | 553 | { |
5a8e01f8 | 554 | cputime_t rtime, stime, utime; |
fa092057 | 555 | |
73fbec60 | 556 | /* |
fa092057 FW |
557 | * Tick based cputime accounting depend on random scheduling |
558 | * timeslices of a task to be interrupted or not by the timer. | |
559 | * Depending on these circumstances, the number of these interrupts | |
560 | * may be over or under-optimistic, matching the real user and system | |
561 | * cputime with a variable precision. | |
562 | * | |
563 | * Fix this by scaling these tick based values against the total | |
564 | * runtime accounted by the CFS scheduler. | |
73fbec60 | 565 | */ |
d37f761d | 566 | rtime = nsecs_to_cputime(curr->sum_exec_runtime); |
73fbec60 | 567 | |
772c808a SG |
568 | /* |
569 | * Update userspace visible utime/stime values only if actual execution | |
570 | * time is bigger than already exported. Note that can happen, that we | |
571 | * provided bigger values due to scaling inaccuracy on big numbers. | |
572 | */ | |
573 | if (prev->stime + prev->utime >= rtime) | |
574 | goto out; | |
575 | ||
5a8e01f8 SG |
576 | stime = curr->stime; |
577 | utime = curr->utime; | |
578 | ||
579 | if (utime == 0) { | |
580 | stime = rtime; | |
581 | } else if (stime == 0) { | |
582 | utime = rtime; | |
583 | } else { | |
584 | cputime_t total = stime + utime; | |
585 | ||
d9a3c982 FW |
586 | stime = scale_stime((__force u64)stime, |
587 | (__force u64)rtime, (__force u64)total); | |
68aa8efc | 588 | utime = rtime - stime; |
d9a3c982 | 589 | } |
73fbec60 FW |
590 | |
591 | /* | |
fa092057 FW |
592 | * If the tick based count grows faster than the scheduler one, |
593 | * the result of the scaling may go backward. | |
594 | * Let's enforce monotonicity. | |
73fbec60 | 595 | */ |
62188451 | 596 | prev->stime = max(prev->stime, stime); |
68aa8efc | 597 | prev->utime = max(prev->utime, utime); |
d37f761d | 598 | |
772c808a | 599 | out: |
d37f761d FW |
600 | *ut = prev->utime; |
601 | *st = prev->stime; | |
602 | } | |
73fbec60 | 603 | |
d37f761d FW |
604 | void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st) |
605 | { | |
606 | struct task_cputime cputime = { | |
d37f761d FW |
607 | .sum_exec_runtime = p->se.sum_exec_runtime, |
608 | }; | |
609 | ||
6fac4829 | 610 | task_cputime(p, &cputime.utime, &cputime.stime); |
d37f761d | 611 | cputime_adjust(&cputime, &p->prev_cputime, ut, st); |
73fbec60 FW |
612 | } |
613 | ||
614 | /* | |
615 | * Must be called with siglock held. | |
616 | */ | |
e80d0a1a | 617 | void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st) |
73fbec60 | 618 | { |
73fbec60 | 619 | struct task_cputime cputime; |
73fbec60 FW |
620 | |
621 | thread_group_cputime(p, &cputime); | |
d37f761d | 622 | cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st); |
73fbec60 | 623 | } |
9fbc42ea | 624 | #endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */ |
abf917cd FW |
625 | |
626 | #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN | |
6a61671b FW |
627 | static unsigned long long vtime_delta(struct task_struct *tsk) |
628 | { | |
629 | unsigned long long clock; | |
630 | ||
7f6575f1 | 631 | clock = local_clock(); |
6a61671b FW |
632 | if (clock < tsk->vtime_snap) |
633 | return 0; | |
abf917cd | 634 | |
6a61671b FW |
635 | return clock - tsk->vtime_snap; |
636 | } | |
637 | ||
638 | static cputime_t get_vtime_delta(struct task_struct *tsk) | |
abf917cd | 639 | { |
6a61671b | 640 | unsigned long long delta = vtime_delta(tsk); |
abf917cd | 641 | |
6a61671b FW |
642 | WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_SLEEPING); |
643 | tsk->vtime_snap += delta; | |
abf917cd FW |
644 | |
645 | /* CHECKME: always safe to convert nsecs to cputime? */ | |
646 | return nsecs_to_cputime(delta); | |
647 | } | |
648 | ||
6a61671b FW |
649 | static void __vtime_account_system(struct task_struct *tsk) |
650 | { | |
651 | cputime_t delta_cpu = get_vtime_delta(tsk); | |
652 | ||
653 | account_system_time(tsk, irq_count(), delta_cpu, cputime_to_scaled(delta_cpu)); | |
654 | } | |
655 | ||
abf917cd FW |
656 | void vtime_account_system(struct task_struct *tsk) |
657 | { | |
6a61671b FW |
658 | write_seqlock(&tsk->vtime_seqlock); |
659 | __vtime_account_system(tsk); | |
660 | write_sequnlock(&tsk->vtime_seqlock); | |
661 | } | |
3f4724ea | 662 | |
b0493406 | 663 | void vtime_gen_account_irq_exit(struct task_struct *tsk) |
6a61671b | 664 | { |
6a61671b | 665 | write_seqlock(&tsk->vtime_seqlock); |
af2350bd | 666 | __vtime_account_system(tsk); |
6a61671b FW |
667 | if (context_tracking_in_user()) |
668 | tsk->vtime_snap_whence = VTIME_USER; | |
6a61671b | 669 | write_sequnlock(&tsk->vtime_seqlock); |
abf917cd FW |
670 | } |
671 | ||
672 | void vtime_account_user(struct task_struct *tsk) | |
673 | { | |
3f4724ea FW |
674 | cputime_t delta_cpu; |
675 | ||
6a61671b | 676 | write_seqlock(&tsk->vtime_seqlock); |
54461562 | 677 | delta_cpu = get_vtime_delta(tsk); |
6a61671b | 678 | tsk->vtime_snap_whence = VTIME_SYS; |
abf917cd | 679 | account_user_time(tsk, delta_cpu, cputime_to_scaled(delta_cpu)); |
6a61671b FW |
680 | write_sequnlock(&tsk->vtime_seqlock); |
681 | } | |
682 | ||
683 | void vtime_user_enter(struct task_struct *tsk) | |
684 | { | |
6a61671b | 685 | write_seqlock(&tsk->vtime_seqlock); |
6a61671b | 686 | __vtime_account_system(tsk); |
af2350bd | 687 | tsk->vtime_snap_whence = VTIME_USER; |
6a61671b FW |
688 | write_sequnlock(&tsk->vtime_seqlock); |
689 | } | |
690 | ||
691 | void vtime_guest_enter(struct task_struct *tsk) | |
692 | { | |
5b206d48 FW |
693 | /* |
694 | * The flags must be updated under the lock with | |
695 | * the vtime_snap flush and update. | |
696 | * That enforces a right ordering and update sequence | |
697 | * synchronization against the reader (task_gtime()) | |
698 | * that can thus safely catch up with a tickless delta. | |
699 | */ | |
6a61671b FW |
700 | write_seqlock(&tsk->vtime_seqlock); |
701 | __vtime_account_system(tsk); | |
702 | current->flags |= PF_VCPU; | |
703 | write_sequnlock(&tsk->vtime_seqlock); | |
704 | } | |
48d6a816 | 705 | EXPORT_SYMBOL_GPL(vtime_guest_enter); |
6a61671b FW |
706 | |
707 | void vtime_guest_exit(struct task_struct *tsk) | |
708 | { | |
709 | write_seqlock(&tsk->vtime_seqlock); | |
710 | __vtime_account_system(tsk); | |
711 | current->flags &= ~PF_VCPU; | |
712 | write_sequnlock(&tsk->vtime_seqlock); | |
abf917cd | 713 | } |
48d6a816 | 714 | EXPORT_SYMBOL_GPL(vtime_guest_exit); |
abf917cd FW |
715 | |
716 | void vtime_account_idle(struct task_struct *tsk) | |
717 | { | |
6a61671b | 718 | cputime_t delta_cpu = get_vtime_delta(tsk); |
abf917cd FW |
719 | |
720 | account_idle_time(delta_cpu); | |
721 | } | |
3f4724ea | 722 | |
6a61671b FW |
723 | void arch_vtime_task_switch(struct task_struct *prev) |
724 | { | |
725 | write_seqlock(&prev->vtime_seqlock); | |
726 | prev->vtime_snap_whence = VTIME_SLEEPING; | |
727 | write_sequnlock(&prev->vtime_seqlock); | |
728 | ||
729 | write_seqlock(¤t->vtime_seqlock); | |
730 | current->vtime_snap_whence = VTIME_SYS; | |
45eacc69 | 731 | current->vtime_snap = sched_clock_cpu(smp_processor_id()); |
6a61671b FW |
732 | write_sequnlock(¤t->vtime_seqlock); |
733 | } | |
734 | ||
45eacc69 | 735 | void vtime_init_idle(struct task_struct *t, int cpu) |
6a61671b FW |
736 | { |
737 | unsigned long flags; | |
738 | ||
739 | write_seqlock_irqsave(&t->vtime_seqlock, flags); | |
740 | t->vtime_snap_whence = VTIME_SYS; | |
45eacc69 | 741 | t->vtime_snap = sched_clock_cpu(cpu); |
6a61671b FW |
742 | write_sequnlock_irqrestore(&t->vtime_seqlock, flags); |
743 | } | |
744 | ||
745 | cputime_t task_gtime(struct task_struct *t) | |
746 | { | |
6a61671b FW |
747 | unsigned int seq; |
748 | cputime_t gtime; | |
749 | ||
750 | do { | |
cdc4e86b | 751 | seq = read_seqbegin(&t->vtime_seqlock); |
6a61671b FW |
752 | |
753 | gtime = t->gtime; | |
754 | if (t->flags & PF_VCPU) | |
755 | gtime += vtime_delta(t); | |
756 | ||
cdc4e86b | 757 | } while (read_seqretry(&t->vtime_seqlock, seq)); |
6a61671b FW |
758 | |
759 | return gtime; | |
760 | } | |
761 | ||
762 | /* | |
763 | * Fetch cputime raw values from fields of task_struct and | |
764 | * add up the pending nohz execution time since the last | |
765 | * cputime snapshot. | |
766 | */ | |
767 | static void | |
768 | fetch_task_cputime(struct task_struct *t, | |
769 | cputime_t *u_dst, cputime_t *s_dst, | |
770 | cputime_t *u_src, cputime_t *s_src, | |
771 | cputime_t *udelta, cputime_t *sdelta) | |
772 | { | |
6a61671b FW |
773 | unsigned int seq; |
774 | unsigned long long delta; | |
775 | ||
776 | do { | |
777 | *udelta = 0; | |
778 | *sdelta = 0; | |
779 | ||
cdc4e86b | 780 | seq = read_seqbegin(&t->vtime_seqlock); |
6a61671b FW |
781 | |
782 | if (u_dst) | |
783 | *u_dst = *u_src; | |
784 | if (s_dst) | |
785 | *s_dst = *s_src; | |
786 | ||
787 | /* Task is sleeping, nothing to add */ | |
788 | if (t->vtime_snap_whence == VTIME_SLEEPING || | |
789 | is_idle_task(t)) | |
790 | continue; | |
791 | ||
792 | delta = vtime_delta(t); | |
793 | ||
794 | /* | |
795 | * Task runs either in user or kernel space, add pending nohz time to | |
796 | * the right place. | |
797 | */ | |
798 | if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU) { | |
799 | *udelta = delta; | |
800 | } else { | |
801 | if (t->vtime_snap_whence == VTIME_SYS) | |
802 | *sdelta = delta; | |
803 | } | |
cdc4e86b | 804 | } while (read_seqretry(&t->vtime_seqlock, seq)); |
6a61671b FW |
805 | } |
806 | ||
807 | ||
808 | void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime) | |
809 | { | |
810 | cputime_t udelta, sdelta; | |
811 | ||
812 | fetch_task_cputime(t, utime, stime, &t->utime, | |
813 | &t->stime, &udelta, &sdelta); | |
814 | if (utime) | |
815 | *utime += udelta; | |
816 | if (stime) | |
817 | *stime += sdelta; | |
818 | } | |
819 | ||
820 | void task_cputime_scaled(struct task_struct *t, | |
821 | cputime_t *utimescaled, cputime_t *stimescaled) | |
822 | { | |
823 | cputime_t udelta, sdelta; | |
824 | ||
825 | fetch_task_cputime(t, utimescaled, stimescaled, | |
826 | &t->utimescaled, &t->stimescaled, &udelta, &sdelta); | |
827 | if (utimescaled) | |
828 | *utimescaled += cputime_to_scaled(udelta); | |
829 | if (stimescaled) | |
830 | *stimescaled += cputime_to_scaled(sdelta); | |
831 | } | |
abf917cd | 832 | #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */ |