[PATCH] posix-cpu-timers: prevent signal delivery starvation
[linux-2.6.git] / kernel / posix-cpu-timers.c
1 /*
2  * Implement CPU time clocks for the POSIX clock interface.
3  */
4
5 #include <linux/sched.h>
6 #include <linux/posix-timers.h>
7 #include <asm/uaccess.h>
8 #include <linux/errno.h>
9
10 static int check_clock(const clockid_t which_clock)
11 {
12         int error = 0;
13         struct task_struct *p;
14         const pid_t pid = CPUCLOCK_PID(which_clock);
15
16         if (CPUCLOCK_WHICH(which_clock) >= CPUCLOCK_MAX)
17                 return -EINVAL;
18
19         if (pid == 0)
20                 return 0;
21
22         read_lock(&tasklist_lock);
23         p = find_task_by_pid(pid);
24         if (!p || (CPUCLOCK_PERTHREAD(which_clock) ?
25                    p->tgid != current->tgid : p->tgid != pid)) {
26                 error = -EINVAL;
27         }
28         read_unlock(&tasklist_lock);
29
30         return error;
31 }
32
33 static inline union cpu_time_count
34 timespec_to_sample(const clockid_t which_clock, const struct timespec *tp)
35 {
36         union cpu_time_count ret;
37         ret.sched = 0;          /* high half always zero when .cpu used */
38         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
39                 ret.sched = (unsigned long long)tp->tv_sec * NSEC_PER_SEC + tp->tv_nsec;
40         } else {
41                 ret.cpu = timespec_to_cputime(tp);
42         }
43         return ret;
44 }
45
46 static void sample_to_timespec(const clockid_t which_clock,
47                                union cpu_time_count cpu,
48                                struct timespec *tp)
49 {
50         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
51                 tp->tv_sec = div_long_long_rem(cpu.sched,
52                                                NSEC_PER_SEC, &tp->tv_nsec);
53         } else {
54                 cputime_to_timespec(cpu.cpu, tp);
55         }
56 }
57
58 static inline int cpu_time_before(const clockid_t which_clock,
59                                   union cpu_time_count now,
60                                   union cpu_time_count then)
61 {
62         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
63                 return now.sched < then.sched;
64         }  else {
65                 return cputime_lt(now.cpu, then.cpu);
66         }
67 }
68 static inline void cpu_time_add(const clockid_t which_clock,
69                                 union cpu_time_count *acc,
70                                 union cpu_time_count val)
71 {
72         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
73                 acc->sched += val.sched;
74         }  else {
75                 acc->cpu = cputime_add(acc->cpu, val.cpu);
76         }
77 }
78 static inline union cpu_time_count cpu_time_sub(const clockid_t which_clock,
79                                                 union cpu_time_count a,
80                                                 union cpu_time_count b)
81 {
82         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
83                 a.sched -= b.sched;
84         }  else {
85                 a.cpu = cputime_sub(a.cpu, b.cpu);
86         }
87         return a;
88 }
89
90 /*
91  * Divide and limit the result to res >= 1
92  *
93  * This is necessary to prevent signal delivery starvation, when the result of
94  * the division would be rounded down to 0.
95  */
96 static inline cputime_t cputime_div_non_zero(cputime_t time, unsigned long div)
97 {
98         cputime_t res = cputime_div(time, div);
99
100         return max_t(cputime_t, res, 1);
101 }
102
103 /*
104  * Update expiry time from increment, and increase overrun count,
105  * given the current clock sample.
106  */
107 static void bump_cpu_timer(struct k_itimer *timer,
108                                   union cpu_time_count now)
109 {
110         int i;
111
112         if (timer->it.cpu.incr.sched == 0)
113                 return;
114
115         if (CPUCLOCK_WHICH(timer->it_clock) == CPUCLOCK_SCHED) {
116                 unsigned long long delta, incr;
117
118                 if (now.sched < timer->it.cpu.expires.sched)
119                         return;
120                 incr = timer->it.cpu.incr.sched;
121                 delta = now.sched + incr - timer->it.cpu.expires.sched;
122                 /* Don't use (incr*2 < delta), incr*2 might overflow. */
123                 for (i = 0; incr < delta - incr; i++)
124                         incr = incr << 1;
125                 for (; i >= 0; incr >>= 1, i--) {
126                         if (delta < incr)
127                                 continue;
128                         timer->it.cpu.expires.sched += incr;
129                         timer->it_overrun += 1 << i;
130                         delta -= incr;
131                 }
132         } else {
133                 cputime_t delta, incr;
134
135                 if (cputime_lt(now.cpu, timer->it.cpu.expires.cpu))
136                         return;
137                 incr = timer->it.cpu.incr.cpu;
138                 delta = cputime_sub(cputime_add(now.cpu, incr),
139                                     timer->it.cpu.expires.cpu);
140                 /* Don't use (incr*2 < delta), incr*2 might overflow. */
141                 for (i = 0; cputime_lt(incr, cputime_sub(delta, incr)); i++)
142                              incr = cputime_add(incr, incr);
143                 for (; i >= 0; incr = cputime_halve(incr), i--) {
144                         if (cputime_lt(delta, incr))
145                                 continue;
146                         timer->it.cpu.expires.cpu =
147                                 cputime_add(timer->it.cpu.expires.cpu, incr);
148                         timer->it_overrun += 1 << i;
149                         delta = cputime_sub(delta, incr);
150                 }
151         }
152 }
153
154 static inline cputime_t prof_ticks(struct task_struct *p)
155 {
156         return cputime_add(p->utime, p->stime);
157 }
158 static inline cputime_t virt_ticks(struct task_struct *p)
159 {
160         return p->utime;
161 }
162 static inline unsigned long long sched_ns(struct task_struct *p)
163 {
164         return (p == current) ? current_sched_time(p) : p->sched_time;
165 }
166
167 int posix_cpu_clock_getres(const clockid_t which_clock, struct timespec *tp)
168 {
169         int error = check_clock(which_clock);
170         if (!error) {
171                 tp->tv_sec = 0;
172                 tp->tv_nsec = ((NSEC_PER_SEC + HZ - 1) / HZ);
173                 if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
174                         /*
175                          * If sched_clock is using a cycle counter, we
176                          * don't have any idea of its true resolution
177                          * exported, but it is much more than 1s/HZ.
178                          */
179                         tp->tv_nsec = 1;
180                 }
181         }
182         return error;
183 }
184
185 int posix_cpu_clock_set(const clockid_t which_clock, const struct timespec *tp)
186 {
187         /*
188          * You can never reset a CPU clock, but we check for other errors
189          * in the call before failing with EPERM.
190          */
191         int error = check_clock(which_clock);
192         if (error == 0) {
193                 error = -EPERM;
194         }
195         return error;
196 }
197
198
199 /*
200  * Sample a per-thread clock for the given task.
201  */
202 static int cpu_clock_sample(const clockid_t which_clock, struct task_struct *p,
203                             union cpu_time_count *cpu)
204 {
205         switch (CPUCLOCK_WHICH(which_clock)) {
206         default:
207                 return -EINVAL;
208         case CPUCLOCK_PROF:
209                 cpu->cpu = prof_ticks(p);
210                 break;
211         case CPUCLOCK_VIRT:
212                 cpu->cpu = virt_ticks(p);
213                 break;
214         case CPUCLOCK_SCHED:
215                 cpu->sched = sched_ns(p);
216                 break;
217         }
218         return 0;
219 }
220
221 /*
222  * Sample a process (thread group) clock for the given group_leader task.
223  * Must be called with tasklist_lock held for reading.
224  * Must be called with tasklist_lock held for reading, and p->sighand->siglock.
225  */
226 static int cpu_clock_sample_group_locked(unsigned int clock_idx,
227                                          struct task_struct *p,
228                                          union cpu_time_count *cpu)
229 {
230         struct task_struct *t = p;
231         switch (clock_idx) {
232         default:
233                 return -EINVAL;
234         case CPUCLOCK_PROF:
235                 cpu->cpu = cputime_add(p->signal->utime, p->signal->stime);
236                 do {
237                         cpu->cpu = cputime_add(cpu->cpu, prof_ticks(t));
238                         t = next_thread(t);
239                 } while (t != p);
240                 break;
241         case CPUCLOCK_VIRT:
242                 cpu->cpu = p->signal->utime;
243                 do {
244                         cpu->cpu = cputime_add(cpu->cpu, virt_ticks(t));
245                         t = next_thread(t);
246                 } while (t != p);
247                 break;
248         case CPUCLOCK_SCHED:
249                 cpu->sched = p->signal->sched_time;
250                 /* Add in each other live thread.  */
251                 while ((t = next_thread(t)) != p) {
252                         cpu->sched += t->sched_time;
253                 }
254                 cpu->sched += sched_ns(p);
255                 break;
256         }
257         return 0;
258 }
259
260 /*
261  * Sample a process (thread group) clock for the given group_leader task.
262  * Must be called with tasklist_lock held for reading.
263  */
264 static int cpu_clock_sample_group(const clockid_t which_clock,
265                                   struct task_struct *p,
266                                   union cpu_time_count *cpu)
267 {
268         int ret;
269         unsigned long flags;
270         spin_lock_irqsave(&p->sighand->siglock, flags);
271         ret = cpu_clock_sample_group_locked(CPUCLOCK_WHICH(which_clock), p,
272                                             cpu);
273         spin_unlock_irqrestore(&p->sighand->siglock, flags);
274         return ret;
275 }
276
277
278 int posix_cpu_clock_get(const clockid_t which_clock, struct timespec *tp)
279 {
280         const pid_t pid = CPUCLOCK_PID(which_clock);
281         int error = -EINVAL;
282         union cpu_time_count rtn;
283
284         if (pid == 0) {
285                 /*
286                  * Special case constant value for our own clocks.
287                  * We don't have to do any lookup to find ourselves.
288                  */
289                 if (CPUCLOCK_PERTHREAD(which_clock)) {
290                         /*
291                          * Sampling just ourselves we can do with no locking.
292                          */
293                         error = cpu_clock_sample(which_clock,
294                                                  current, &rtn);
295                 } else {
296                         read_lock(&tasklist_lock);
297                         error = cpu_clock_sample_group(which_clock,
298                                                        current, &rtn);
299                         read_unlock(&tasklist_lock);
300                 }
301         } else {
302                 /*
303                  * Find the given PID, and validate that the caller
304                  * should be able to see it.
305                  */
306                 struct task_struct *p;
307                 read_lock(&tasklist_lock);
308                 p = find_task_by_pid(pid);
309                 if (p) {
310                         if (CPUCLOCK_PERTHREAD(which_clock)) {
311                                 if (p->tgid == current->tgid) {
312                                         error = cpu_clock_sample(which_clock,
313                                                                  p, &rtn);
314                                 }
315                         } else if (p->tgid == pid && p->signal) {
316                                 error = cpu_clock_sample_group(which_clock,
317                                                                p, &rtn);
318                         }
319                 }
320                 read_unlock(&tasklist_lock);
321         }
322
323         if (error)
324                 return error;
325         sample_to_timespec(which_clock, rtn, tp);
326         return 0;
327 }
328
329
330 /*
331  * Validate the clockid_t for a new CPU-clock timer, and initialize the timer.
332  * This is called from sys_timer_create with the new timer already locked.
333  */
334 int posix_cpu_timer_create(struct k_itimer *new_timer)
335 {
336         int ret = 0;
337         const pid_t pid = CPUCLOCK_PID(new_timer->it_clock);
338         struct task_struct *p;
339
340         if (CPUCLOCK_WHICH(new_timer->it_clock) >= CPUCLOCK_MAX)
341                 return -EINVAL;
342
343         INIT_LIST_HEAD(&new_timer->it.cpu.entry);
344         new_timer->it.cpu.incr.sched = 0;
345         new_timer->it.cpu.expires.sched = 0;
346
347         read_lock(&tasklist_lock);
348         if (CPUCLOCK_PERTHREAD(new_timer->it_clock)) {
349                 if (pid == 0) {
350                         p = current;
351                 } else {
352                         p = find_task_by_pid(pid);
353                         if (p && p->tgid != current->tgid)
354                                 p = NULL;
355                 }
356         } else {
357                 if (pid == 0) {
358                         p = current->group_leader;
359                 } else {
360                         p = find_task_by_pid(pid);
361                         if (p && p->tgid != pid)
362                                 p = NULL;
363                 }
364         }
365         new_timer->it.cpu.task = p;
366         if (p) {
367                 get_task_struct(p);
368         } else {
369                 ret = -EINVAL;
370         }
371         read_unlock(&tasklist_lock);
372
373         return ret;
374 }
375
376 /*
377  * Clean up a CPU-clock timer that is about to be destroyed.
378  * This is called from timer deletion with the timer already locked.
379  * If we return TIMER_RETRY, it's necessary to release the timer's lock
380  * and try again.  (This happens when the timer is in the middle of firing.)
381  */
382 int posix_cpu_timer_del(struct k_itimer *timer)
383 {
384         struct task_struct *p = timer->it.cpu.task;
385         int ret = 0;
386
387         if (likely(p != NULL)) {
388                 read_lock(&tasklist_lock);
389                 if (unlikely(p->signal == NULL)) {
390                         /*
391                          * We raced with the reaping of the task.
392                          * The deletion should have cleared us off the list.
393                          */
394                         BUG_ON(!list_empty(&timer->it.cpu.entry));
395                 } else {
396                         spin_lock(&p->sighand->siglock);
397                         if (timer->it.cpu.firing)
398                                 ret = TIMER_RETRY;
399                         else
400                                 list_del(&timer->it.cpu.entry);
401                         spin_unlock(&p->sighand->siglock);
402                 }
403                 read_unlock(&tasklist_lock);
404
405                 if (!ret)
406                         put_task_struct(p);
407         }
408
409         return ret;
410 }
411
412 /*
413  * Clean out CPU timers still ticking when a thread exited.  The task
414  * pointer is cleared, and the expiry time is replaced with the residual
415  * time for later timer_gettime calls to return.
416  * This must be called with the siglock held.
417  */
418 static void cleanup_timers(struct list_head *head,
419                            cputime_t utime, cputime_t stime,
420                            unsigned long long sched_time)
421 {
422         struct cpu_timer_list *timer, *next;
423         cputime_t ptime = cputime_add(utime, stime);
424
425         list_for_each_entry_safe(timer, next, head, entry) {
426                 list_del_init(&timer->entry);
427                 if (cputime_lt(timer->expires.cpu, ptime)) {
428                         timer->expires.cpu = cputime_zero;
429                 } else {
430                         timer->expires.cpu = cputime_sub(timer->expires.cpu,
431                                                          ptime);
432                 }
433         }
434
435         ++head;
436         list_for_each_entry_safe(timer, next, head, entry) {
437                 list_del_init(&timer->entry);
438                 if (cputime_lt(timer->expires.cpu, utime)) {
439                         timer->expires.cpu = cputime_zero;
440                 } else {
441                         timer->expires.cpu = cputime_sub(timer->expires.cpu,
442                                                          utime);
443                 }
444         }
445
446         ++head;
447         list_for_each_entry_safe(timer, next, head, entry) {
448                 list_del_init(&timer->entry);
449                 if (timer->expires.sched < sched_time) {
450                         timer->expires.sched = 0;
451                 } else {
452                         timer->expires.sched -= sched_time;
453                 }
454         }
455 }
456
457 /*
458  * These are both called with the siglock held, when the current thread
459  * is being reaped.  When the final (leader) thread in the group is reaped,
460  * posix_cpu_timers_exit_group will be called after posix_cpu_timers_exit.
461  */
462 void posix_cpu_timers_exit(struct task_struct *tsk)
463 {
464         cleanup_timers(tsk->cpu_timers,
465                        tsk->utime, tsk->stime, tsk->sched_time);
466
467 }
468 void posix_cpu_timers_exit_group(struct task_struct *tsk)
469 {
470         cleanup_timers(tsk->signal->cpu_timers,
471                        cputime_add(tsk->utime, tsk->signal->utime),
472                        cputime_add(tsk->stime, tsk->signal->stime),
473                        tsk->sched_time + tsk->signal->sched_time);
474 }
475
476
477 /*
478  * Set the expiry times of all the threads in the process so one of them
479  * will go off before the process cumulative expiry total is reached.
480  */
481 static void process_timer_rebalance(struct task_struct *p,
482                                     unsigned int clock_idx,
483                                     union cpu_time_count expires,
484                                     union cpu_time_count val)
485 {
486         cputime_t ticks, left;
487         unsigned long long ns, nsleft;
488         struct task_struct *t = p;
489         unsigned int nthreads = atomic_read(&p->signal->live);
490
491         if (!nthreads)
492                 return;
493
494         switch (clock_idx) {
495         default:
496                 BUG();
497                 break;
498         case CPUCLOCK_PROF:
499                 left = cputime_div_non_zero(cputime_sub(expires.cpu, val.cpu),
500                                        nthreads);
501                 do {
502                         if (likely(!(t->flags & PF_EXITING))) {
503                                 ticks = cputime_add(prof_ticks(t), left);
504                                 if (cputime_eq(t->it_prof_expires,
505                                                cputime_zero) ||
506                                     cputime_gt(t->it_prof_expires, ticks)) {
507                                         t->it_prof_expires = ticks;
508                                 }
509                         }
510                         t = next_thread(t);
511                 } while (t != p);
512                 break;
513         case CPUCLOCK_VIRT:
514                 left = cputime_div_non_zero(cputime_sub(expires.cpu, val.cpu),
515                                        nthreads);
516                 do {
517                         if (likely(!(t->flags & PF_EXITING))) {
518                                 ticks = cputime_add(virt_ticks(t), left);
519                                 if (cputime_eq(t->it_virt_expires,
520                                                cputime_zero) ||
521                                     cputime_gt(t->it_virt_expires, ticks)) {
522                                         t->it_virt_expires = ticks;
523                                 }
524                         }
525                         t = next_thread(t);
526                 } while (t != p);
527                 break;
528         case CPUCLOCK_SCHED:
529                 nsleft = expires.sched - val.sched;
530                 do_div(nsleft, nthreads);
531                 nsleft = max_t(unsigned long long, nsleft, 1);
532                 do {
533                         if (likely(!(t->flags & PF_EXITING))) {
534                                 ns = t->sched_time + nsleft;
535                                 if (t->it_sched_expires == 0 ||
536                                     t->it_sched_expires > ns) {
537                                         t->it_sched_expires = ns;
538                                 }
539                         }
540                         t = next_thread(t);
541                 } while (t != p);
542                 break;
543         }
544 }
545
546 static void clear_dead_task(struct k_itimer *timer, union cpu_time_count now)
547 {
548         /*
549          * That's all for this thread or process.
550          * We leave our residual in expires to be reported.
551          */
552         put_task_struct(timer->it.cpu.task);
553         timer->it.cpu.task = NULL;
554         timer->it.cpu.expires = cpu_time_sub(timer->it_clock,
555                                              timer->it.cpu.expires,
556                                              now);
557 }
558
559 /*
560  * Insert the timer on the appropriate list before any timers that
561  * expire later.  This must be called with the tasklist_lock held
562  * for reading, and interrupts disabled.
563  */
564 static void arm_timer(struct k_itimer *timer, union cpu_time_count now)
565 {
566         struct task_struct *p = timer->it.cpu.task;
567         struct list_head *head, *listpos;
568         struct cpu_timer_list *const nt = &timer->it.cpu;
569         struct cpu_timer_list *next;
570         unsigned long i;
571
572         head = (CPUCLOCK_PERTHREAD(timer->it_clock) ?
573                 p->cpu_timers : p->signal->cpu_timers);
574         head += CPUCLOCK_WHICH(timer->it_clock);
575
576         BUG_ON(!irqs_disabled());
577         spin_lock(&p->sighand->siglock);
578
579         listpos = head;
580         if (CPUCLOCK_WHICH(timer->it_clock) == CPUCLOCK_SCHED) {
581                 list_for_each_entry(next, head, entry) {
582                         if (next->expires.sched > nt->expires.sched)
583                                 break;
584                         listpos = &next->entry;
585                 }
586         } else {
587                 list_for_each_entry(next, head, entry) {
588                         if (cputime_gt(next->expires.cpu, nt->expires.cpu))
589                                 break;
590                         listpos = &next->entry;
591                 }
592         }
593         list_add(&nt->entry, listpos);
594
595         if (listpos == head) {
596                 /*
597                  * We are the new earliest-expiring timer.
598                  * If we are a thread timer, there can always
599                  * be a process timer telling us to stop earlier.
600                  */
601
602                 if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
603                         switch (CPUCLOCK_WHICH(timer->it_clock)) {
604                         default:
605                                 BUG();
606                         case CPUCLOCK_PROF:
607                                 if (cputime_eq(p->it_prof_expires,
608                                                cputime_zero) ||
609                                     cputime_gt(p->it_prof_expires,
610                                                nt->expires.cpu))
611                                         p->it_prof_expires = nt->expires.cpu;
612                                 break;
613                         case CPUCLOCK_VIRT:
614                                 if (cputime_eq(p->it_virt_expires,
615                                                cputime_zero) ||
616                                     cputime_gt(p->it_virt_expires,
617                                                nt->expires.cpu))
618                                         p->it_virt_expires = nt->expires.cpu;
619                                 break;
620                         case CPUCLOCK_SCHED:
621                                 if (p->it_sched_expires == 0 ||
622                                     p->it_sched_expires > nt->expires.sched)
623                                         p->it_sched_expires = nt->expires.sched;
624                                 break;
625                         }
626                 } else {
627                         /*
628                          * For a process timer, we must balance
629                          * all the live threads' expirations.
630                          */
631                         switch (CPUCLOCK_WHICH(timer->it_clock)) {
632                         default:
633                                 BUG();
634                         case CPUCLOCK_VIRT:
635                                 if (!cputime_eq(p->signal->it_virt_expires,
636                                                 cputime_zero) &&
637                                     cputime_lt(p->signal->it_virt_expires,
638                                                timer->it.cpu.expires.cpu))
639                                         break;
640                                 goto rebalance;
641                         case CPUCLOCK_PROF:
642                                 if (!cputime_eq(p->signal->it_prof_expires,
643                                                 cputime_zero) &&
644                                     cputime_lt(p->signal->it_prof_expires,
645                                                timer->it.cpu.expires.cpu))
646                                         break;
647                                 i = p->signal->rlim[RLIMIT_CPU].rlim_cur;
648                                 if (i != RLIM_INFINITY &&
649                                     i <= cputime_to_secs(timer->it.cpu.expires.cpu))
650                                         break;
651                                 goto rebalance;
652                         case CPUCLOCK_SCHED:
653                         rebalance:
654                                 process_timer_rebalance(
655                                         timer->it.cpu.task,
656                                         CPUCLOCK_WHICH(timer->it_clock),
657                                         timer->it.cpu.expires, now);
658                                 break;
659                         }
660                 }
661         }
662
663         spin_unlock(&p->sighand->siglock);
664 }
665
666 /*
667  * The timer is locked, fire it and arrange for its reload.
668  */
669 static void cpu_timer_fire(struct k_itimer *timer)
670 {
671         if (unlikely(timer->sigq == NULL)) {
672                 /*
673                  * This a special case for clock_nanosleep,
674                  * not a normal timer from sys_timer_create.
675                  */
676                 wake_up_process(timer->it_process);
677                 timer->it.cpu.expires.sched = 0;
678         } else if (timer->it.cpu.incr.sched == 0) {
679                 /*
680                  * One-shot timer.  Clear it as soon as it's fired.
681                  */
682                 posix_timer_event(timer, 0);
683                 timer->it.cpu.expires.sched = 0;
684         } else if (posix_timer_event(timer, ++timer->it_requeue_pending)) {
685                 /*
686                  * The signal did not get queued because the signal
687                  * was ignored, so we won't get any callback to
688                  * reload the timer.  But we need to keep it
689                  * ticking in case the signal is deliverable next time.
690                  */
691                 posix_cpu_timer_schedule(timer);
692         }
693 }
694
695 /*
696  * Guts of sys_timer_settime for CPU timers.
697  * This is called with the timer locked and interrupts disabled.
698  * If we return TIMER_RETRY, it's necessary to release the timer's lock
699  * and try again.  (This happens when the timer is in the middle of firing.)
700  */
701 int posix_cpu_timer_set(struct k_itimer *timer, int flags,
702                         struct itimerspec *new, struct itimerspec *old)
703 {
704         struct task_struct *p = timer->it.cpu.task;
705         union cpu_time_count old_expires, new_expires, val;
706         int ret;
707
708         if (unlikely(p == NULL)) {
709                 /*
710                  * Timer refers to a dead task's clock.
711                  */
712                 return -ESRCH;
713         }
714
715         new_expires = timespec_to_sample(timer->it_clock, &new->it_value);
716
717         read_lock(&tasklist_lock);
718         /*
719          * We need the tasklist_lock to protect against reaping that
720          * clears p->signal.  If p has just been reaped, we can no
721          * longer get any information about it at all.
722          */
723         if (unlikely(p->signal == NULL)) {
724                 read_unlock(&tasklist_lock);
725                 put_task_struct(p);
726                 timer->it.cpu.task = NULL;
727                 return -ESRCH;
728         }
729
730         /*
731          * Disarm any old timer after extracting its expiry time.
732          */
733         BUG_ON(!irqs_disabled());
734
735         ret = 0;
736         spin_lock(&p->sighand->siglock);
737         old_expires = timer->it.cpu.expires;
738         if (unlikely(timer->it.cpu.firing)) {
739                 timer->it.cpu.firing = -1;
740                 ret = TIMER_RETRY;
741         } else
742                 list_del_init(&timer->it.cpu.entry);
743         spin_unlock(&p->sighand->siglock);
744
745         /*
746          * We need to sample the current value to convert the new
747          * value from to relative and absolute, and to convert the
748          * old value from absolute to relative.  To set a process
749          * timer, we need a sample to balance the thread expiry
750          * times (in arm_timer).  With an absolute time, we must
751          * check if it's already passed.  In short, we need a sample.
752          */
753         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
754                 cpu_clock_sample(timer->it_clock, p, &val);
755         } else {
756                 cpu_clock_sample_group(timer->it_clock, p, &val);
757         }
758
759         if (old) {
760                 if (old_expires.sched == 0) {
761                         old->it_value.tv_sec = 0;
762                         old->it_value.tv_nsec = 0;
763                 } else {
764                         /*
765                          * Update the timer in case it has
766                          * overrun already.  If it has,
767                          * we'll report it as having overrun
768                          * and with the next reloaded timer
769                          * already ticking, though we are
770                          * swallowing that pending
771                          * notification here to install the
772                          * new setting.
773                          */
774                         bump_cpu_timer(timer, val);
775                         if (cpu_time_before(timer->it_clock, val,
776                                             timer->it.cpu.expires)) {
777                                 old_expires = cpu_time_sub(
778                                         timer->it_clock,
779                                         timer->it.cpu.expires, val);
780                                 sample_to_timespec(timer->it_clock,
781                                                    old_expires,
782                                                    &old->it_value);
783                         } else {
784                                 old->it_value.tv_nsec = 1;
785                                 old->it_value.tv_sec = 0;
786                         }
787                 }
788         }
789
790         if (unlikely(ret)) {
791                 /*
792                  * We are colliding with the timer actually firing.
793                  * Punt after filling in the timer's old value, and
794                  * disable this firing since we are already reporting
795                  * it as an overrun (thanks to bump_cpu_timer above).
796                  */
797                 read_unlock(&tasklist_lock);
798                 goto out;
799         }
800
801         if (new_expires.sched != 0 && !(flags & TIMER_ABSTIME)) {
802                 cpu_time_add(timer->it_clock, &new_expires, val);
803         }
804
805         /*
806          * Install the new expiry time (or zero).
807          * For a timer with no notification action, we don't actually
808          * arm the timer (we'll just fake it for timer_gettime).
809          */
810         timer->it.cpu.expires = new_expires;
811         if (new_expires.sched != 0 &&
812             (timer->it_sigev_notify & ~SIGEV_THREAD_ID) != SIGEV_NONE &&
813             cpu_time_before(timer->it_clock, val, new_expires)) {
814                 arm_timer(timer, val);
815         }
816
817         read_unlock(&tasklist_lock);
818
819         /*
820          * Install the new reload setting, and
821          * set up the signal and overrun bookkeeping.
822          */
823         timer->it.cpu.incr = timespec_to_sample(timer->it_clock,
824                                                 &new->it_interval);
825
826         /*
827          * This acts as a modification timestamp for the timer,
828          * so any automatic reload attempt will punt on seeing
829          * that we have reset the timer manually.
830          */
831         timer->it_requeue_pending = (timer->it_requeue_pending + 2) &
832                 ~REQUEUE_PENDING;
833         timer->it_overrun_last = 0;
834         timer->it_overrun = -1;
835
836         if (new_expires.sched != 0 &&
837             (timer->it_sigev_notify & ~SIGEV_THREAD_ID) != SIGEV_NONE &&
838             !cpu_time_before(timer->it_clock, val, new_expires)) {
839                 /*
840                  * The designated time already passed, so we notify
841                  * immediately, even if the thread never runs to
842                  * accumulate more time on this clock.
843                  */
844                 cpu_timer_fire(timer);
845         }
846
847         ret = 0;
848  out:
849         if (old) {
850                 sample_to_timespec(timer->it_clock,
851                                    timer->it.cpu.incr, &old->it_interval);
852         }
853         return ret;
854 }
855
856 void posix_cpu_timer_get(struct k_itimer *timer, struct itimerspec *itp)
857 {
858         union cpu_time_count now;
859         struct task_struct *p = timer->it.cpu.task;
860         int clear_dead;
861
862         /*
863          * Easy part: convert the reload time.
864          */
865         sample_to_timespec(timer->it_clock,
866                            timer->it.cpu.incr, &itp->it_interval);
867
868         if (timer->it.cpu.expires.sched == 0) { /* Timer not armed at all.  */
869                 itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
870                 return;
871         }
872
873         if (unlikely(p == NULL)) {
874                 /*
875                  * This task already died and the timer will never fire.
876                  * In this case, expires is actually the dead value.
877                  */
878         dead:
879                 sample_to_timespec(timer->it_clock, timer->it.cpu.expires,
880                                    &itp->it_value);
881                 return;
882         }
883
884         /*
885          * Sample the clock to take the difference with the expiry time.
886          */
887         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
888                 cpu_clock_sample(timer->it_clock, p, &now);
889                 clear_dead = p->exit_state;
890         } else {
891                 read_lock(&tasklist_lock);
892                 if (unlikely(p->signal == NULL)) {
893                         /*
894                          * The process has been reaped.
895                          * We can't even collect a sample any more.
896                          * Call the timer disarmed, nothing else to do.
897                          */
898                         put_task_struct(p);
899                         timer->it.cpu.task = NULL;
900                         timer->it.cpu.expires.sched = 0;
901                         read_unlock(&tasklist_lock);
902                         goto dead;
903                 } else {
904                         cpu_clock_sample_group(timer->it_clock, p, &now);
905                         clear_dead = (unlikely(p->exit_state) &&
906                                       thread_group_empty(p));
907                 }
908                 read_unlock(&tasklist_lock);
909         }
910
911         if ((timer->it_sigev_notify & ~SIGEV_THREAD_ID) == SIGEV_NONE) {
912                 if (timer->it.cpu.incr.sched == 0 &&
913                     cpu_time_before(timer->it_clock,
914                                     timer->it.cpu.expires, now)) {
915                         /*
916                          * Do-nothing timer expired and has no reload,
917                          * so it's as if it was never set.
918                          */
919                         timer->it.cpu.expires.sched = 0;
920                         itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
921                         return;
922                 }
923                 /*
924                  * Account for any expirations and reloads that should
925                  * have happened.
926                  */
927                 bump_cpu_timer(timer, now);
928         }
929
930         if (unlikely(clear_dead)) {
931                 /*
932                  * We've noticed that the thread is dead, but
933                  * not yet reaped.  Take this opportunity to
934                  * drop our task ref.
935                  */
936                 clear_dead_task(timer, now);
937                 goto dead;
938         }
939
940         if (cpu_time_before(timer->it_clock, now, timer->it.cpu.expires)) {
941                 sample_to_timespec(timer->it_clock,
942                                    cpu_time_sub(timer->it_clock,
943                                                 timer->it.cpu.expires, now),
944                                    &itp->it_value);
945         } else {
946                 /*
947                  * The timer should have expired already, but the firing
948                  * hasn't taken place yet.  Say it's just about to expire.
949                  */
950                 itp->it_value.tv_nsec = 1;
951                 itp->it_value.tv_sec = 0;
952         }
953 }
954
955 /*
956  * Check for any per-thread CPU timers that have fired and move them off
957  * the tsk->cpu_timers[N] list onto the firing list.  Here we update the
958  * tsk->it_*_expires values to reflect the remaining thread CPU timers.
959  */
960 static void check_thread_timers(struct task_struct *tsk,
961                                 struct list_head *firing)
962 {
963         int maxfire;
964         struct list_head *timers = tsk->cpu_timers;
965
966         maxfire = 20;
967         tsk->it_prof_expires = cputime_zero;
968         while (!list_empty(timers)) {
969                 struct cpu_timer_list *t = list_entry(timers->next,
970                                                       struct cpu_timer_list,
971                                                       entry);
972                 if (!--maxfire || cputime_lt(prof_ticks(tsk), t->expires.cpu)) {
973                         tsk->it_prof_expires = t->expires.cpu;
974                         break;
975                 }
976                 t->firing = 1;
977                 list_move_tail(&t->entry, firing);
978         }
979
980         ++timers;
981         maxfire = 20;
982         tsk->it_virt_expires = cputime_zero;
983         while (!list_empty(timers)) {
984                 struct cpu_timer_list *t = list_entry(timers->next,
985                                                       struct cpu_timer_list,
986                                                       entry);
987                 if (!--maxfire || cputime_lt(virt_ticks(tsk), t->expires.cpu)) {
988                         tsk->it_virt_expires = t->expires.cpu;
989                         break;
990                 }
991                 t->firing = 1;
992                 list_move_tail(&t->entry, firing);
993         }
994
995         ++timers;
996         maxfire = 20;
997         tsk->it_sched_expires = 0;
998         while (!list_empty(timers)) {
999                 struct cpu_timer_list *t = list_entry(timers->next,
1000                                                       struct cpu_timer_list,
1001                                                       entry);
1002                 if (!--maxfire || tsk->sched_time < t->expires.sched) {
1003                         tsk->it_sched_expires = t->expires.sched;
1004                         break;
1005                 }
1006                 t->firing = 1;
1007                 list_move_tail(&t->entry, firing);
1008         }
1009 }
1010
1011 /*
1012  * Check for any per-thread CPU timers that have fired and move them
1013  * off the tsk->*_timers list onto the firing list.  Per-thread timers
1014  * have already been taken off.
1015  */
1016 static void check_process_timers(struct task_struct *tsk,
1017                                  struct list_head *firing)
1018 {
1019         int maxfire;
1020         struct signal_struct *const sig = tsk->signal;
1021         cputime_t utime, stime, ptime, virt_expires, prof_expires;
1022         unsigned long long sched_time, sched_expires;
1023         struct task_struct *t;
1024         struct list_head *timers = sig->cpu_timers;
1025
1026         /*
1027          * Don't sample the current process CPU clocks if there are no timers.
1028          */
1029         if (list_empty(&timers[CPUCLOCK_PROF]) &&
1030             cputime_eq(sig->it_prof_expires, cputime_zero) &&
1031             sig->rlim[RLIMIT_CPU].rlim_cur == RLIM_INFINITY &&
1032             list_empty(&timers[CPUCLOCK_VIRT]) &&
1033             cputime_eq(sig->it_virt_expires, cputime_zero) &&
1034             list_empty(&timers[CPUCLOCK_SCHED]))
1035                 return;
1036
1037         /*
1038          * Collect the current process totals.
1039          */
1040         utime = sig->utime;
1041         stime = sig->stime;
1042         sched_time = sig->sched_time;
1043         t = tsk;
1044         do {
1045                 utime = cputime_add(utime, t->utime);
1046                 stime = cputime_add(stime, t->stime);
1047                 sched_time += t->sched_time;
1048                 t = next_thread(t);
1049         } while (t != tsk);
1050         ptime = cputime_add(utime, stime);
1051
1052         maxfire = 20;
1053         prof_expires = cputime_zero;
1054         while (!list_empty(timers)) {
1055                 struct cpu_timer_list *t = list_entry(timers->next,
1056                                                       struct cpu_timer_list,
1057                                                       entry);
1058                 if (!--maxfire || cputime_lt(ptime, t->expires.cpu)) {
1059                         prof_expires = t->expires.cpu;
1060                         break;
1061                 }
1062                 t->firing = 1;
1063                 list_move_tail(&t->entry, firing);
1064         }
1065
1066         ++timers;
1067         maxfire = 20;
1068         virt_expires = cputime_zero;
1069         while (!list_empty(timers)) {
1070                 struct cpu_timer_list *t = list_entry(timers->next,
1071                                                       struct cpu_timer_list,
1072                                                       entry);
1073                 if (!--maxfire || cputime_lt(utime, t->expires.cpu)) {
1074                         virt_expires = t->expires.cpu;
1075                         break;
1076                 }
1077                 t->firing = 1;
1078                 list_move_tail(&t->entry, firing);
1079         }
1080
1081         ++timers;
1082         maxfire = 20;
1083         sched_expires = 0;
1084         while (!list_empty(timers)) {
1085                 struct cpu_timer_list *t = list_entry(timers->next,
1086                                                       struct cpu_timer_list,
1087                                                       entry);
1088                 if (!--maxfire || sched_time < t->expires.sched) {
1089                         sched_expires = t->expires.sched;
1090                         break;
1091                 }
1092                 t->firing = 1;
1093                 list_move_tail(&t->entry, firing);
1094         }
1095
1096         /*
1097          * Check for the special case process timers.
1098          */
1099         if (!cputime_eq(sig->it_prof_expires, cputime_zero)) {
1100                 if (cputime_ge(ptime, sig->it_prof_expires)) {
1101                         /* ITIMER_PROF fires and reloads.  */
1102                         sig->it_prof_expires = sig->it_prof_incr;
1103                         if (!cputime_eq(sig->it_prof_expires, cputime_zero)) {
1104                                 sig->it_prof_expires = cputime_add(
1105                                         sig->it_prof_expires, ptime);
1106                         }
1107                         __group_send_sig_info(SIGPROF, SEND_SIG_PRIV, tsk);
1108                 }
1109                 if (!cputime_eq(sig->it_prof_expires, cputime_zero) &&
1110                     (cputime_eq(prof_expires, cputime_zero) ||
1111                      cputime_lt(sig->it_prof_expires, prof_expires))) {
1112                         prof_expires = sig->it_prof_expires;
1113                 }
1114         }
1115         if (!cputime_eq(sig->it_virt_expires, cputime_zero)) {
1116                 if (cputime_ge(utime, sig->it_virt_expires)) {
1117                         /* ITIMER_VIRTUAL fires and reloads.  */
1118                         sig->it_virt_expires = sig->it_virt_incr;
1119                         if (!cputime_eq(sig->it_virt_expires, cputime_zero)) {
1120                                 sig->it_virt_expires = cputime_add(
1121                                         sig->it_virt_expires, utime);
1122                         }
1123                         __group_send_sig_info(SIGVTALRM, SEND_SIG_PRIV, tsk);
1124                 }
1125                 if (!cputime_eq(sig->it_virt_expires, cputime_zero) &&
1126                     (cputime_eq(virt_expires, cputime_zero) ||
1127                      cputime_lt(sig->it_virt_expires, virt_expires))) {
1128                         virt_expires = sig->it_virt_expires;
1129                 }
1130         }
1131         if (sig->rlim[RLIMIT_CPU].rlim_cur != RLIM_INFINITY) {
1132                 unsigned long psecs = cputime_to_secs(ptime);
1133                 cputime_t x;
1134                 if (psecs >= sig->rlim[RLIMIT_CPU].rlim_max) {
1135                         /*
1136                          * At the hard limit, we just die.
1137                          * No need to calculate anything else now.
1138                          */
1139                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
1140                         return;
1141                 }
1142                 if (psecs >= sig->rlim[RLIMIT_CPU].rlim_cur) {
1143                         /*
1144                          * At the soft limit, send a SIGXCPU every second.
1145                          */
1146                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
1147                         if (sig->rlim[RLIMIT_CPU].rlim_cur
1148                             < sig->rlim[RLIMIT_CPU].rlim_max) {
1149                                 sig->rlim[RLIMIT_CPU].rlim_cur++;
1150                         }
1151                 }
1152                 x = secs_to_cputime(sig->rlim[RLIMIT_CPU].rlim_cur);
1153                 if (cputime_eq(prof_expires, cputime_zero) ||
1154                     cputime_lt(x, prof_expires)) {
1155                         prof_expires = x;
1156                 }
1157         }
1158
1159         if (!cputime_eq(prof_expires, cputime_zero) ||
1160             !cputime_eq(virt_expires, cputime_zero) ||
1161             sched_expires != 0) {
1162                 /*
1163                  * Rebalance the threads' expiry times for the remaining
1164                  * process CPU timers.
1165                  */
1166
1167                 cputime_t prof_left, virt_left, ticks;
1168                 unsigned long long sched_left, sched;
1169                 const unsigned int nthreads = atomic_read(&sig->live);
1170
1171                 if (!nthreads)
1172                         return;
1173
1174                 prof_left = cputime_sub(prof_expires, utime);
1175                 prof_left = cputime_sub(prof_left, stime);
1176                 prof_left = cputime_div_non_zero(prof_left, nthreads);
1177                 virt_left = cputime_sub(virt_expires, utime);
1178                 virt_left = cputime_div_non_zero(virt_left, nthreads);
1179                 if (sched_expires) {
1180                         sched_left = sched_expires - sched_time;
1181                         do_div(sched_left, nthreads);
1182                         sched_left = max_t(unsigned long long, sched_left, 1);
1183                 } else {
1184                         sched_left = 0;
1185                 }
1186                 t = tsk;
1187                 do {
1188                         if (unlikely(t->flags & PF_EXITING))
1189                                 continue;
1190
1191                         ticks = cputime_add(cputime_add(t->utime, t->stime),
1192                                             prof_left);
1193                         if (!cputime_eq(prof_expires, cputime_zero) &&
1194                             (cputime_eq(t->it_prof_expires, cputime_zero) ||
1195                              cputime_gt(t->it_prof_expires, ticks))) {
1196                                 t->it_prof_expires = ticks;
1197                         }
1198
1199                         ticks = cputime_add(t->utime, virt_left);
1200                         if (!cputime_eq(virt_expires, cputime_zero) &&
1201                             (cputime_eq(t->it_virt_expires, cputime_zero) ||
1202                              cputime_gt(t->it_virt_expires, ticks))) {
1203                                 t->it_virt_expires = ticks;
1204                         }
1205
1206                         sched = t->sched_time + sched_left;
1207                         if (sched_expires && (t->it_sched_expires == 0 ||
1208                                               t->it_sched_expires > sched)) {
1209                                 t->it_sched_expires = sched;
1210                         }
1211                 } while ((t = next_thread(t)) != tsk);
1212         }
1213 }
1214
1215 /*
1216  * This is called from the signal code (via do_schedule_next_timer)
1217  * when the last timer signal was delivered and we have to reload the timer.
1218  */
1219 void posix_cpu_timer_schedule(struct k_itimer *timer)
1220 {
1221         struct task_struct *p = timer->it.cpu.task;
1222         union cpu_time_count now;
1223
1224         if (unlikely(p == NULL))
1225                 /*
1226                  * The task was cleaned up already, no future firings.
1227                  */
1228                 goto out;
1229
1230         /*
1231          * Fetch the current sample and update the timer's expiry time.
1232          */
1233         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
1234                 cpu_clock_sample(timer->it_clock, p, &now);
1235                 bump_cpu_timer(timer, now);
1236                 if (unlikely(p->exit_state)) {
1237                         clear_dead_task(timer, now);
1238                         goto out;
1239                 }
1240                 read_lock(&tasklist_lock); /* arm_timer needs it.  */
1241         } else {
1242                 read_lock(&tasklist_lock);
1243                 if (unlikely(p->signal == NULL)) {
1244                         /*
1245                          * The process has been reaped.
1246                          * We can't even collect a sample any more.
1247                          */
1248                         put_task_struct(p);
1249                         timer->it.cpu.task = p = NULL;
1250                         timer->it.cpu.expires.sched = 0;
1251                         goto out_unlock;
1252                 } else if (unlikely(p->exit_state) && thread_group_empty(p)) {
1253                         /*
1254                          * We've noticed that the thread is dead, but
1255                          * not yet reaped.  Take this opportunity to
1256                          * drop our task ref.
1257                          */
1258                         clear_dead_task(timer, now);
1259                         goto out_unlock;
1260                 }
1261                 cpu_clock_sample_group(timer->it_clock, p, &now);
1262                 bump_cpu_timer(timer, now);
1263                 /* Leave the tasklist_lock locked for the call below.  */
1264         }
1265
1266         /*
1267          * Now re-arm for the new expiry time.
1268          */
1269         arm_timer(timer, now);
1270
1271 out_unlock:
1272         read_unlock(&tasklist_lock);
1273
1274 out:
1275         timer->it_overrun_last = timer->it_overrun;
1276         timer->it_overrun = -1;
1277         ++timer->it_requeue_pending;
1278 }
1279
1280 /*
1281  * This is called from the timer interrupt handler.  The irq handler has
1282  * already updated our counts.  We need to check if any timers fire now.
1283  * Interrupts are disabled.
1284  */
1285 void run_posix_cpu_timers(struct task_struct *tsk)
1286 {
1287         LIST_HEAD(firing);
1288         struct k_itimer *timer, *next;
1289
1290         BUG_ON(!irqs_disabled());
1291
1292 #define UNEXPIRED(clock) \
1293                 (cputime_eq(tsk->it_##clock##_expires, cputime_zero) || \
1294                  cputime_lt(clock##_ticks(tsk), tsk->it_##clock##_expires))
1295
1296         if (UNEXPIRED(prof) && UNEXPIRED(virt) &&
1297             (tsk->it_sched_expires == 0 ||
1298              tsk->sched_time < tsk->it_sched_expires))
1299                 return;
1300
1301 #undef  UNEXPIRED
1302
1303         /*
1304          * Double-check with locks held.
1305          */
1306         read_lock(&tasklist_lock);
1307         if (likely(tsk->signal != NULL)) {
1308                 spin_lock(&tsk->sighand->siglock);
1309
1310                 /*
1311                  * Here we take off tsk->cpu_timers[N] and tsk->signal->cpu_timers[N]
1312                  * all the timers that are firing, and put them on the firing list.
1313                  */
1314                 check_thread_timers(tsk, &firing);
1315                 check_process_timers(tsk, &firing);
1316
1317                 /*
1318                  * We must release these locks before taking any timer's lock.
1319                  * There is a potential race with timer deletion here, as the
1320                  * siglock now protects our private firing list.  We have set
1321                  * the firing flag in each timer, so that a deletion attempt
1322                  * that gets the timer lock before we do will give it up and
1323                  * spin until we've taken care of that timer below.
1324                  */
1325                 spin_unlock(&tsk->sighand->siglock);
1326         }
1327         read_unlock(&tasklist_lock);
1328
1329         /*
1330          * Now that all the timers on our list have the firing flag,
1331          * noone will touch their list entries but us.  We'll take
1332          * each timer's lock before clearing its firing flag, so no
1333          * timer call will interfere.
1334          */
1335         list_for_each_entry_safe(timer, next, &firing, it.cpu.entry) {
1336                 int firing;
1337                 spin_lock(&timer->it_lock);
1338                 list_del_init(&timer->it.cpu.entry);
1339                 firing = timer->it.cpu.firing;
1340                 timer->it.cpu.firing = 0;
1341                 /*
1342                  * The firing flag is -1 if we collided with a reset
1343                  * of the timer, which already reported this
1344                  * almost-firing as an overrun.  So don't generate an event.
1345                  */
1346                 if (likely(firing >= 0)) {
1347                         cpu_timer_fire(timer);
1348                 }
1349                 spin_unlock(&timer->it_lock);
1350         }
1351 }
1352
1353 /*
1354  * Set one of the process-wide special case CPU timers.
1355  * The tasklist_lock and tsk->sighand->siglock must be held by the caller.
1356  * The oldval argument is null for the RLIMIT_CPU timer, where *newval is
1357  * absolute; non-null for ITIMER_*, where *newval is relative and we update
1358  * it to be absolute, *oldval is absolute and we update it to be relative.
1359  */
1360 void set_process_cpu_timer(struct task_struct *tsk, unsigned int clock_idx,
1361                            cputime_t *newval, cputime_t *oldval)
1362 {
1363         union cpu_time_count now;
1364         struct list_head *head;
1365
1366         BUG_ON(clock_idx == CPUCLOCK_SCHED);
1367         cpu_clock_sample_group_locked(clock_idx, tsk, &now);
1368
1369         if (oldval) {
1370                 if (!cputime_eq(*oldval, cputime_zero)) {
1371                         if (cputime_le(*oldval, now.cpu)) {
1372                                 /* Just about to fire. */
1373                                 *oldval = jiffies_to_cputime(1);
1374                         } else {
1375                                 *oldval = cputime_sub(*oldval, now.cpu);
1376                         }
1377                 }
1378
1379                 if (cputime_eq(*newval, cputime_zero))
1380                         return;
1381                 *newval = cputime_add(*newval, now.cpu);
1382
1383                 /*
1384                  * If the RLIMIT_CPU timer will expire before the
1385                  * ITIMER_PROF timer, we have nothing else to do.
1386                  */
1387                 if (tsk->signal->rlim[RLIMIT_CPU].rlim_cur
1388                     < cputime_to_secs(*newval))
1389                         return;
1390         }
1391
1392         /*
1393          * Check whether there are any process timers already set to fire
1394          * before this one.  If so, we don't have anything more to do.
1395          */
1396         head = &tsk->signal->cpu_timers[clock_idx];
1397         if (list_empty(head) ||
1398             cputime_ge(list_entry(head->next,
1399                                   struct cpu_timer_list, entry)->expires.cpu,
1400                        *newval)) {
1401                 /*
1402                  * Rejigger each thread's expiry time so that one will
1403                  * notice before we hit the process-cumulative expiry time.
1404                  */
1405                 union cpu_time_count expires = { .sched = 0 };
1406                 expires.cpu = *newval;
1407                 process_timer_rebalance(tsk, clock_idx, expires, now);
1408         }
1409 }
1410
1411 static int do_cpu_nanosleep(const clockid_t which_clock, int flags,
1412                             struct timespec *rqtp, struct itimerspec *it)
1413 {
1414         struct k_itimer timer;
1415         int error;
1416
1417         /*
1418          * Set up a temporary timer and then wait for it to go off.
1419          */
1420         memset(&timer, 0, sizeof timer);
1421         spin_lock_init(&timer.it_lock);
1422         timer.it_clock = which_clock;
1423         timer.it_overrun = -1;
1424         error = posix_cpu_timer_create(&timer);
1425         timer.it_process = current;
1426         if (!error) {
1427                 static struct itimerspec zero_it;
1428
1429                 memset(it, 0, sizeof *it);
1430                 it->it_value = *rqtp;
1431
1432                 spin_lock_irq(&timer.it_lock);
1433                 error = posix_cpu_timer_set(&timer, flags, it, NULL);
1434                 if (error) {
1435                         spin_unlock_irq(&timer.it_lock);
1436                         return error;
1437                 }
1438
1439                 while (!signal_pending(current)) {
1440                         if (timer.it.cpu.expires.sched == 0) {
1441                                 /*
1442                                  * Our timer fired and was reset.
1443                                  */
1444                                 spin_unlock_irq(&timer.it_lock);
1445                                 return 0;
1446                         }
1447
1448                         /*
1449                          * Block until cpu_timer_fire (or a signal) wakes us.
1450                          */
1451                         __set_current_state(TASK_INTERRUPTIBLE);
1452                         spin_unlock_irq(&timer.it_lock);
1453                         schedule();
1454                         spin_lock_irq(&timer.it_lock);
1455                 }
1456
1457                 /*
1458                  * We were interrupted by a signal.
1459                  */
1460                 sample_to_timespec(which_clock, timer.it.cpu.expires, rqtp);
1461                 posix_cpu_timer_set(&timer, 0, &zero_it, it);
1462                 spin_unlock_irq(&timer.it_lock);
1463
1464                 if ((it->it_value.tv_sec | it->it_value.tv_nsec) == 0) {
1465                         /*
1466                          * It actually did fire already.
1467                          */
1468                         return 0;
1469                 }
1470
1471                 error = -ERESTART_RESTARTBLOCK;
1472         }
1473
1474         return error;
1475 }
1476
1477 int posix_cpu_nsleep(const clockid_t which_clock, int flags,
1478                      struct timespec *rqtp, struct timespec __user *rmtp)
1479 {
1480         struct restart_block *restart_block =
1481             &current_thread_info()->restart_block;
1482         struct itimerspec it;
1483         int error;
1484
1485         /*
1486          * Diagnose required errors first.
1487          */
1488         if (CPUCLOCK_PERTHREAD(which_clock) &&
1489             (CPUCLOCK_PID(which_clock) == 0 ||
1490              CPUCLOCK_PID(which_clock) == current->pid))
1491                 return -EINVAL;
1492
1493         error = do_cpu_nanosleep(which_clock, flags, rqtp, &it);
1494
1495         if (error == -ERESTART_RESTARTBLOCK) {
1496
1497                 if (flags & TIMER_ABSTIME)
1498                         return -ERESTARTNOHAND;
1499                 /*
1500                  * Report back to the user the time still remaining.
1501                  */
1502                 if (rmtp != NULL && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1503                         return -EFAULT;
1504
1505                 restart_block->fn = posix_cpu_nsleep_restart;
1506                 restart_block->arg0 = which_clock;
1507                 restart_block->arg1 = (unsigned long) rmtp;
1508                 restart_block->arg2 = rqtp->tv_sec;
1509                 restart_block->arg3 = rqtp->tv_nsec;
1510         }
1511         return error;
1512 }
1513
1514 long posix_cpu_nsleep_restart(struct restart_block *restart_block)
1515 {
1516         clockid_t which_clock = restart_block->arg0;
1517         struct timespec __user *rmtp;
1518         struct timespec t;
1519         struct itimerspec it;
1520         int error;
1521
1522         rmtp = (struct timespec __user *) restart_block->arg1;
1523         t.tv_sec = restart_block->arg2;
1524         t.tv_nsec = restart_block->arg3;
1525
1526         restart_block->fn = do_no_restart_syscall;
1527         error = do_cpu_nanosleep(which_clock, TIMER_ABSTIME, &t, &it);
1528
1529         if (error == -ERESTART_RESTARTBLOCK) {
1530                 /*
1531                  * Report back to the user the time still remaining.
1532                  */
1533                 if (rmtp != NULL && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1534                         return -EFAULT;
1535
1536                 restart_block->fn = posix_cpu_nsleep_restart;
1537                 restart_block->arg0 = which_clock;
1538                 restart_block->arg1 = (unsigned long) rmtp;
1539                 restart_block->arg2 = t.tv_sec;
1540                 restart_block->arg3 = t.tv_nsec;
1541         }
1542         return error;
1543
1544 }
1545
1546
1547 #define PROCESS_CLOCK   MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED)
1548 #define THREAD_CLOCK    MAKE_THREAD_CPUCLOCK(0, CPUCLOCK_SCHED)
1549
1550 static int process_cpu_clock_getres(const clockid_t which_clock,
1551                                     struct timespec *tp)
1552 {
1553         return posix_cpu_clock_getres(PROCESS_CLOCK, tp);
1554 }
1555 static int process_cpu_clock_get(const clockid_t which_clock,
1556                                  struct timespec *tp)
1557 {
1558         return posix_cpu_clock_get(PROCESS_CLOCK, tp);
1559 }
1560 static int process_cpu_timer_create(struct k_itimer *timer)
1561 {
1562         timer->it_clock = PROCESS_CLOCK;
1563         return posix_cpu_timer_create(timer);
1564 }
1565 static int process_cpu_nsleep(const clockid_t which_clock, int flags,
1566                               struct timespec *rqtp,
1567                               struct timespec __user *rmtp)
1568 {
1569         return posix_cpu_nsleep(PROCESS_CLOCK, flags, rqtp, rmtp);
1570 }
1571 static long process_cpu_nsleep_restart(struct restart_block *restart_block)
1572 {
1573         return -EINVAL;
1574 }
1575 static int thread_cpu_clock_getres(const clockid_t which_clock,
1576                                    struct timespec *tp)
1577 {
1578         return posix_cpu_clock_getres(THREAD_CLOCK, tp);
1579 }
1580 static int thread_cpu_clock_get(const clockid_t which_clock,
1581                                 struct timespec *tp)
1582 {
1583         return posix_cpu_clock_get(THREAD_CLOCK, tp);
1584 }
1585 static int thread_cpu_timer_create(struct k_itimer *timer)
1586 {
1587         timer->it_clock = THREAD_CLOCK;
1588         return posix_cpu_timer_create(timer);
1589 }
1590 static int thread_cpu_nsleep(const clockid_t which_clock, int flags,
1591                               struct timespec *rqtp, struct timespec __user *rmtp)
1592 {
1593         return -EINVAL;
1594 }
1595 static long thread_cpu_nsleep_restart(struct restart_block *restart_block)
1596 {
1597         return -EINVAL;
1598 }
1599
1600 static __init int init_posix_cpu_timers(void)
1601 {
1602         struct k_clock process = {
1603                 .clock_getres = process_cpu_clock_getres,
1604                 .clock_get = process_cpu_clock_get,
1605                 .clock_set = do_posix_clock_nosettime,
1606                 .timer_create = process_cpu_timer_create,
1607                 .nsleep = process_cpu_nsleep,
1608                 .nsleep_restart = process_cpu_nsleep_restart,
1609         };
1610         struct k_clock thread = {
1611                 .clock_getres = thread_cpu_clock_getres,
1612                 .clock_get = thread_cpu_clock_get,
1613                 .clock_set = do_posix_clock_nosettime,
1614                 .timer_create = thread_cpu_timer_create,
1615                 .nsleep = thread_cpu_nsleep,
1616                 .nsleep_restart = thread_cpu_nsleep_restart,
1617         };
1618
1619         register_posix_clock(CLOCK_PROCESS_CPUTIME_ID, &process);
1620         register_posix_clock(CLOCK_THREAD_CPUTIME_ID, &thread);
1621
1622         return 0;
1623 }
1624 __initcall(init_posix_cpu_timers);