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