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 <linux/errno.h>
   8 #include <linux/math64.h>
   9 #include <asm/uaccess.h>
  10 #include <linux/kernel_stat.h>
  11 #include <trace/events/timer.h>
  12
  13 /*
  14  * Called after updating RLIMIT_CPU to run cpu timer and update
  15  * tsk->signal->cputime_expires expiration cache if necessary. Needs
  16  * siglock protection since other code may update expiration cache as
  17  * well.
  18  */
  19 void update_rlimit_cpu(struct task_struct *task, unsigned long rlim_new)
  20 {
  21         cputime_t cputime = secs_to_cputime(rlim_new);
  22
  23         spin_lock_irq(&task->sighand->siglock);
  24         set_process_cpu_timer(task, CPUCLOCK_PROF, &cputime, NULL);
  25         spin_unlock_irq(&task->sighand->siglock);
  26 }
  27
  28 static int check_clock(const clockid_t which_clock)
  29 {
  30         int error = 0;
  31         struct task_struct *p;
  32         const pid_t pid = CPUCLOCK_PID(which_clock);
  33
  34         if (CPUCLOCK_WHICH(which_clock) >= CPUCLOCK_MAX)
  35                 return -EINVAL;
  36
  37         if (pid == 0)
  38                 return 0;
  39
  40         rcu_read_lock();
  41         p = find_task_by_vpid(pid);
  42         if (!p || !(CPUCLOCK_PERTHREAD(which_clock) ?
  43                    same_thread_group(p, current) : has_group_leader_pid(p))) {
  44                 error = -EINVAL;
  45         }
  46         rcu_read_unlock();
  47
  48         return error;
  49 }
  50
  51 static inline union cpu_time_count
  52 timespec_to_sample(const clockid_t which_clock, const struct timespec *tp)
  53 {
  54         union cpu_time_count ret;
  55         ret.sched = 0;          /* high half always zero when .cpu used */
  56         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  57                 ret.sched = (unsigned long long)tp->tv_sec * NSEC_PER_SEC + tp->tv_nsec;
  58         } else {
  59                 ret.cpu = timespec_to_cputime(tp);
  60         }
  61         return ret;
  62 }
  63
  64 static void sample_to_timespec(const clockid_t which_clock,
  65                                union cpu_time_count cpu,
  66                                struct timespec *tp)
  67 {
  68         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED)
  69                 *tp = ns_to_timespec(cpu.sched);
  70         else
  71                 cputime_to_timespec(cpu.cpu, tp);
  72 }
  73
  74 static inline int cpu_time_before(const clockid_t which_clock,
  75                                   union cpu_time_count now,
  76                                   union cpu_time_count then)
  77 {
  78         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  79                 return now.sched < then.sched;
  80         }  else {
  81                 return cputime_lt(now.cpu, then.cpu);
  82         }
  83 }
  84 static inline void cpu_time_add(const clockid_t which_clock,
  85                                 union cpu_time_count *acc,
  86                                 union cpu_time_count val)
  87 {
  88         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  89                 acc->sched += val.sched;
  90         }  else {
  91                 acc->cpu = cputime_add(acc->cpu, val.cpu);
  92         }
  93 }
  94 static inline union cpu_time_count cpu_time_sub(const clockid_t which_clock,
  95                                                 union cpu_time_count a,
  96                                                 union cpu_time_count b)
  97 {
  98         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  99                 a.sched -= b.sched;
 100         }  else {
 101                 a.cpu = cputime_sub(a.cpu, b.cpu);
 102         }
 103         return a;
 104 }
 105
 106 /*
 107  * Divide and limit the result to res >= 1
 108  *
 109  * This is necessary to prevent signal delivery starvation, when the result of
 110  * the division would be rounded down to 0.
 111  */
 112 static inline cputime_t cputime_div_non_zero(cputime_t time, unsigned long div)
 113 {
 114         cputime_t res = cputime_div(time, div);
 115
 116         return max_t(cputime_t, res, 1);
 117 }
 118
 119 /*
 120  * Update expiry time from increment, and increase overrun count,
 121  * given the current clock sample.
 122  */
 123 static void bump_cpu_timer(struct k_itimer *timer,
 124                                   union cpu_time_count now)
 125 {
 126         int i;
 127
 128         if (timer->it.cpu.incr.sched == 0)
 129                 return;
 130
 131         if (CPUCLOCK_WHICH(timer->it_clock) == CPUCLOCK_SCHED) {
 132                 unsigned long long delta, incr;
 133
 134                 if (now.sched < timer->it.cpu.expires.sched)
 135                         return;
 136                 incr = timer->it.cpu.incr.sched;
 137                 delta = now.sched + incr - timer->it.cpu.expires.sched;
 138                 /* Don't use (incr*2 < delta), incr*2 might overflow. */
 139                 for (i = 0; incr < delta - incr; i++)
 140                         incr = incr << 1;
 141                 for (; i >= 0; incr >>= 1, i--) {
 142                         if (delta < incr)
 143                                 continue;
 144                         timer->it.cpu.expires.sched += incr;
 145                         timer->it_overrun += 1 << i;
 146                         delta -= incr;
 147                 }
 148         } else {
 149                 cputime_t delta, incr;
 150
 151                 if (cputime_lt(now.cpu, timer->it.cpu.expires.cpu))
 152                         return;
 153                 incr = timer->it.cpu.incr.cpu;
 154                 delta = cputime_sub(cputime_add(now.cpu, incr),
 155                                     timer->it.cpu.expires.cpu);
 156                 /* Don't use (incr*2 < delta), incr*2 might overflow. */
 157                 for (i = 0; cputime_lt(incr, cputime_sub(delta, incr)); i++)
 158                              incr = cputime_add(incr, incr);
 159                 for (; i >= 0; incr = cputime_halve(incr), i--) {
 160                         if (cputime_lt(delta, incr))
 161                                 continue;
 162                         timer->it.cpu.expires.cpu =
 163                                 cputime_add(timer->it.cpu.expires.cpu, incr);
 164                         timer->it_overrun += 1 << i;
 165                         delta = cputime_sub(delta, incr);
 166                 }
 167         }
 168 }
 169
 170 static inline cputime_t prof_ticks(struct task_struct *p)
 171 {
 172         return cputime_add(p->utime, p->stime);
 173 }
 174 static inline cputime_t virt_ticks(struct task_struct *p)
 175 {
 176         return p->utime;
 177 }
 178
 179 static int
 180 posix_cpu_clock_getres(const clockid_t which_clock, struct timespec *tp)
 181 {
 182         int error = check_clock(which_clock);
 183         if (!error) {
 184                 tp->tv_sec = 0;
 185                 tp->tv_nsec = ((NSEC_PER_SEC + HZ - 1) / HZ);
 186                 if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
 187                         /*
 188                          * If sched_clock is using a cycle counter, we
 189                          * don't have any idea of its true resolution
 190                          * exported, but it is much more than 1s/HZ.
 191                          */
 192                         tp->tv_nsec = 1;
 193                 }
 194         }
 195         return error;
 196 }
 197
 198 static int
 199 posix_cpu_clock_set(const clockid_t which_clock, const struct timespec *tp)
 200 {
 201         /*
 202          * You can never reset a CPU clock, but we check for other errors
 203          * in the call before failing with EPERM.
 204          */
 205         int error = check_clock(which_clock);
 206         if (error == 0) {
 207                 error = -EPERM;
 208         }
 209         return error;
 210 }
 211
 212
 213 /*
 214  * Sample a per-thread clock for the given task.
 215  */
 216 static int cpu_clock_sample(const clockid_t which_clock, struct task_struct *p,
 217                             union cpu_time_count *cpu)
 218 {
 219         switch (CPUCLOCK_WHICH(which_clock)) {
 220         default:
 221                 return -EINVAL;
 222         case CPUCLOCK_PROF:
 223                 cpu->cpu = prof_ticks(p);
 224                 break;
 225         case CPUCLOCK_VIRT:
 226                 cpu->cpu = virt_ticks(p);
 227                 break;
 228         case CPUCLOCK_SCHED:
 229                 cpu->sched = task_sched_runtime(p);
 230                 break;
 231         }
 232         return 0;
 233 }
 234
 235 void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
 236 {
 237         struct signal_struct *sig = tsk->signal;
 238         struct task_struct *t;
 239
 240         times->utime = sig->utime;
 241         times->stime = sig->stime;
 242         times->sum_exec_runtime = sig->sum_sched_runtime;
 243
 244         rcu_read_lock();
 245         /* make sure we can trust tsk->thread_group list */
 246         if (!likely(pid_alive(tsk)))
 247                 goto out;
 248
 249         t = tsk;
 250         do {
 251                 times->utime = cputime_add(times->utime, t->utime);
 252                 times->stime = cputime_add(times->stime, t->stime);
 253                 times->sum_exec_runtime += t->se.sum_exec_runtime;
 254         } while_each_thread(tsk, t);
 255 out:
 256         rcu_read_unlock();
 257 }
 258
 259 static void update_gt_cputime(struct task_cputime *a, struct task_cputime *b)
 260 {
 261         if (cputime_gt(b->utime, a->utime))
 262                 a->utime = b->utime;
 263
 264         if (cputime_gt(b->stime, a->stime))
 265                 a->stime = b->stime;
 266
 267         if (b->sum_exec_runtime > a->sum_exec_runtime)
 268                 a->sum_exec_runtime = b->sum_exec_runtime;
 269 }
 270
 271 void thread_group_cputimer(struct task_struct *tsk, struct task_cputime *times)
 272 {
 273         struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
 274         struct task_cputime sum;
 275         unsigned long flags;
 276
 277         spin_lock_irqsave(&cputimer->lock, flags);
 278         if (!cputimer->running) {
 279                 cputimer->running = 1;
 280                 /*
 281                  * The POSIX timer interface allows for absolute time expiry
 282                  * values through the TIMER_ABSTIME flag, therefore we have
 283                  * to synchronize the timer to the clock every time we start
 284                  * it.
 285                  */
 286                 thread_group_cputime(tsk, &sum);
 287                 update_gt_cputime(&cputimer->cputime, &sum);
 288         }
 289         *times = cputimer->cputime;
 290         spin_unlock_irqrestore(&cputimer->lock, flags);
 291 }
 292
 293 /*
 294  * Sample a process (thread group) clock for the given group_leader task.
 295  * Must be called with tasklist_lock held for reading.
 296  */
 297 static int cpu_clock_sample_group(const clockid_t which_clock,
 298                                   struct task_struct *p,
 299                                   union cpu_time_count *cpu)
 300 {
 301         struct task_cputime cputime;
 302
 303         switch (CPUCLOCK_WHICH(which_clock)) {
 304         default:
 305                 return -EINVAL;
 306         case CPUCLOCK_PROF:
 307                 thread_group_cputime(p, &cputime);
 308                 cpu->cpu = cputime_add(cputime.utime, cputime.stime);
 309                 break;
 310         case CPUCLOCK_VIRT:
 311                 thread_group_cputime(p, &cputime);
 312                 cpu->cpu = cputime.utime;
 313                 break;
 314         case CPUCLOCK_SCHED:
 315                 cpu->sched = thread_group_sched_runtime(p);
 316                 break;
 317         }
 318         return 0;
 319 }
 320
 321
 322 static int posix_cpu_clock_get(const clockid_t which_clock, struct timespec *tp)
 323 {
 324         const pid_t pid = CPUCLOCK_PID(which_clock);
 325         int error = -EINVAL;
 326         union cpu_time_count rtn;
 327
 328         if (pid == 0) {
 329                 /*
 330                  * Special case constant value for our own clocks.
 331                  * We don't have to do any lookup to find ourselves.
 332                  */
 333                 if (CPUCLOCK_PERTHREAD(which_clock)) {
 334                         /*
 335                          * Sampling just ourselves we can do with no locking.
 336                          */
 337                         error = cpu_clock_sample(which_clock,
 338                                                  current, &rtn);
 339                 } else {
 340                         read_lock(&tasklist_lock);
 341                         error = cpu_clock_sample_group(which_clock,
 342                                                        current, &rtn);
 343                         read_unlock(&tasklist_lock);
 344                 }
 345         } else {
 346                 /*
 347                  * Find the given PID, and validate that the caller
 348                  * should be able to see it.
 349                  */
 350                 struct task_struct *p;
 351                 rcu_read_lock();
 352                 p = find_task_by_vpid(pid);
 353                 if (p) {
 354                         if (CPUCLOCK_PERTHREAD(which_clock)) {
 355                                 if (same_thread_group(p, current)) {
 356                                         error = cpu_clock_sample(which_clock,
 357                                                                  p, &rtn);
 358                                 }
 359                         } else {
 360                                 read_lock(&tasklist_lock);
 361                                 if (thread_group_leader(p) && p->sighand) {
 362                                         error =
 363                                             cpu_clock_sample_group(which_clock,
 364                                                                    p, &rtn);
 365                                 }
 366                                 read_unlock(&tasklist_lock);
 367                         }
 368                 }
 369                 rcu_read_unlock();
 370         }
 371
 372         if (error)
 373                 return error;
 374         sample_to_timespec(which_clock, rtn, tp);
 375         return 0;
 376 }
 377
 378
 379 /*
 380  * Validate the clockid_t for a new CPU-clock timer, and initialize the timer.
 381  * This is called from sys_timer_create() and do_cpu_nanosleep() with the
 382  * new timer already all-zeros initialized.
 383  */
 384 static int posix_cpu_timer_create(struct k_itimer *new_timer)
 385 {
 386         int ret = 0;
 387         const pid_t pid = CPUCLOCK_PID(new_timer->it_clock);
 388         struct task_struct *p;
 389
 390         if (CPUCLOCK_WHICH(new_timer->it_clock) >= CPUCLOCK_MAX)
 391                 return -EINVAL;
 392
 393         INIT_LIST_HEAD(&new_timer->it.cpu.entry);
 394
 395         rcu_read_lock();
 396         if (CPUCLOCK_PERTHREAD(new_timer->it_clock)) {
 397                 if (pid == 0) {
 398                         p = current;
 399                 } else {
 400                         p = find_task_by_vpid(pid);
 401                         if (p && !same_thread_group(p, current))
 402                                 p = NULL;
 403                 }
 404         } else {
 405                 if (pid == 0) {
 406                         p = current->group_leader;
 407                 } else {
 408                         p = find_task_by_vpid(pid);
 409                         if (p && !has_group_leader_pid(p))
 410                                 p = NULL;
 411                 }
 412         }
 413         new_timer->it.cpu.task = p;
 414         if (p) {
 415                 get_task_struct(p);
 416         } else {
 417                 ret = -EINVAL;
 418         }
 419         rcu_read_unlock();
 420
 421         return ret;
 422 }
 423
 424 /*
 425  * Clean up a CPU-clock timer that is about to be destroyed.
 426  * This is called from timer deletion with the timer already locked.
 427  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 428  * and try again.  (This happens when the timer is in the middle of firing.)
 429  */
 430 static int posix_cpu_timer_del(struct k_itimer *timer)
 431 {
 432         struct task_struct *p = timer->it.cpu.task;
 433         int ret = 0;
 434
 435         if (likely(p != NULL)) {
 436                 read_lock(&tasklist_lock);
 437                 if (unlikely(p->sighand == NULL)) {
 438                         /*
 439                          * We raced with the reaping of the task.
 440                          * The deletion should have cleared us off the list.
 441                          */
 442                         BUG_ON(!list_empty(&timer->it.cpu.entry));
 443                 } else {
 444                         spin_lock(&p->sighand->siglock);
 445                         if (timer->it.cpu.firing)
 446                                 ret = TIMER_RETRY;
 447                         else
 448                                 list_del(&timer->it.cpu.entry);
 449                         spin_unlock(&p->sighand->siglock);
 450                 }
 451                 read_unlock(&tasklist_lock);
 452
 453                 if (!ret)
 454                         put_task_struct(p);
 455         }
 456
 457         return ret;
 458 }
 459
 460 /*
 461  * Clean out CPU timers still ticking when a thread exited.  The task
 462  * pointer is cleared, and the expiry time is replaced with the residual
 463  * time for later timer_gettime calls to return.
 464  * This must be called with the siglock held.
 465  */
 466 static void cleanup_timers(struct list_head *head,
 467                            cputime_t utime, cputime_t stime,
 468                            unsigned long long sum_exec_runtime)
 469 {
 470         struct cpu_timer_list *timer, *next;
 471         cputime_t ptime = cputime_add(utime, stime);
 472
 473         list_for_each_entry_safe(timer, next, head, entry) {
 474                 list_del_init(&timer->entry);
 475                 if (cputime_lt(timer->expires.cpu, ptime)) {
 476                         timer->expires.cpu = cputime_zero;
 477                 } else {
 478                         timer->expires.cpu = cputime_sub(timer->expires.cpu,
 479                                                          ptime);
 480                 }
 481         }
 482
 483         ++head;
 484         list_for_each_entry_safe(timer, next, head, entry) {
 485                 list_del_init(&timer->entry);
 486                 if (cputime_lt(timer->expires.cpu, utime)) {
 487                         timer->expires.cpu = cputime_zero;
 488                 } else {
 489                         timer->expires.cpu = cputime_sub(timer->expires.cpu,
 490                                                          utime);
 491                 }
 492         }
 493
 494         ++head;
 495         list_for_each_entry_safe(timer, next, head, entry) {
 496                 list_del_init(&timer->entry);
 497                 if (timer->expires.sched < sum_exec_runtime) {
 498                         timer->expires.sched = 0;
 499                 } else {
 500                         timer->expires.sched -= sum_exec_runtime;
 501                 }
 502         }
 503 }
 504
 505 /*
 506  * These are both called with the siglock held, when the current thread
 507  * is being reaped.  When the final (leader) thread in the group is reaped,
 508  * posix_cpu_timers_exit_group will be called after posix_cpu_timers_exit.
 509  */
 510 void posix_cpu_timers_exit(struct task_struct *tsk)
 511 {
 512         cleanup_timers(tsk->cpu_timers,
 513                        tsk->utime, tsk->stime, tsk->se.sum_exec_runtime);
 514
 515 }
 516 void posix_cpu_timers_exit_group(struct task_struct *tsk)
 517 {
 518         struct signal_struct *const sig = tsk->signal;
 519
 520         cleanup_timers(tsk->signal->cpu_timers,
 521                        cputime_add(tsk->utime, sig->utime),
 522                        cputime_add(tsk->stime, sig->stime),
 523                        tsk->se.sum_exec_runtime + sig->sum_sched_runtime);
 524 }
 525
 526 static void clear_dead_task(struct k_itimer *timer, union cpu_time_count now)
 527 {
 528         /*
 529          * That's all for this thread or process.
 530          * We leave our residual in expires to be reported.
 531          */
 532         put_task_struct(timer->it.cpu.task);
 533         timer->it.cpu.task = NULL;
 534         timer->it.cpu.expires = cpu_time_sub(timer->it_clock,
 535                                              timer->it.cpu.expires,
 536                                              now);
 537 }
 538
 539 static inline int expires_gt(cputime_t expires, cputime_t new_exp)
 540 {
 541         return cputime_eq(expires, cputime_zero) ||
 542                cputime_gt(expires, new_exp);
 543 }
 544
 545 /*
 546  * Insert the timer on the appropriate list before any timers that
 547  * expire later.  This must be called with the tasklist_lock held
 548  * for reading, interrupts disabled and p->sighand->siglock taken.
 549  */
 550 static void arm_timer(struct k_itimer *timer)
 551 {
 552         struct task_struct *p = timer->it.cpu.task;
 553         struct list_head *head, *listpos;
 554         struct task_cputime *cputime_expires;
 555         struct cpu_timer_list *const nt = &timer->it.cpu;
 556         struct cpu_timer_list *next;
 557
 558         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 559                 head = p->cpu_timers;
 560                 cputime_expires = &p->cputime_expires;
 561         } else {
 562                 head = p->signal->cpu_timers;
 563                 cputime_expires = &p->signal->cputime_expires;
 564         }
 565         head += CPUCLOCK_WHICH(timer->it_clock);
 566
 567         listpos = head;
 568         list_for_each_entry(next, head, entry) {
 569                 if (cpu_time_before(timer->it_clock, nt->expires, next->expires))
 570                         break;
 571                 listpos = &next->entry;
 572         }
 573         list_add(&nt->entry, listpos);
 574
 575         if (listpos == head) {
 576                 union cpu_time_count *exp = &nt->expires;
 577
 578                 /*
 579                  * We are the new earliest-expiring POSIX 1.b timer, hence
 580                  * need to update expiration cache. Take into account that
 581                  * for process timers we share expiration cache with itimers
 582                  * and RLIMIT_CPU and for thread timers with RLIMIT_RTTIME.
 583                  */
 584
 585                 switch (CPUCLOCK_WHICH(timer->it_clock)) {
 586                 case CPUCLOCK_PROF:
 587                         if (expires_gt(cputime_expires->prof_exp, exp->cpu))
 588                                 cputime_expires->prof_exp = exp->cpu;
 589                         break;
 590                 case CPUCLOCK_VIRT:
 591                         if (expires_gt(cputime_expires->virt_exp, exp->cpu))
 592                                 cputime_expires->virt_exp = exp->cpu;
 593                         break;
 594                 case CPUCLOCK_SCHED:
 595                         if (cputime_expires->sched_exp == 0 ||
 596                             cputime_expires->sched_exp > exp->sched)
 597                                 cputime_expires->sched_exp = exp->sched;
 598                         break;
 599                 }
 600         }
 601 }
 602
 603 /*
 604  * The timer is locked, fire it and arrange for its reload.
 605  */
 606 static void cpu_timer_fire(struct k_itimer *timer)
 607 {
 608         if ((timer->it_sigev_notify & ~SIGEV_THREAD_ID) == SIGEV_NONE) {
 609                 /*
 610                  * User don't want any signal.
 611                  */
 612                 timer->it.cpu.expires.sched = 0;
 613         } else if (unlikely(timer->sigq == NULL)) {
 614                 /*
 615                  * This a special case for clock_nanosleep,
 616                  * not a normal timer from sys_timer_create.
 617                  */
 618                 wake_up_process(timer->it_process);
 619                 timer->it.cpu.expires.sched = 0;
 620         } else if (timer->it.cpu.incr.sched == 0) {
 621                 /*
 622                  * One-shot timer.  Clear it as soon as it's fired.
 623                  */
 624                 posix_timer_event(timer, 0);
 625                 timer->it.cpu.expires.sched = 0;
 626         } else if (posix_timer_event(timer, ++timer->it_requeue_pending)) {
 627                 /*
 628                  * The signal did not get queued because the signal
 629                  * was ignored, so we won't get any callback to
 630                  * reload the timer.  But we need to keep it
 631                  * ticking in case the signal is deliverable next time.
 632                  */
 633                 posix_cpu_timer_schedule(timer);
 634         }
 635 }
 636
 637 /*
 638  * Sample a process (thread group) timer for the given group_leader task.
 639  * Must be called with tasklist_lock held for reading.
 640  */
 641 static int cpu_timer_sample_group(const clockid_t which_clock,
 642                                   struct task_struct *p,
 643                                   union cpu_time_count *cpu)
 644 {
 645         struct task_cputime cputime;
 646
 647         thread_group_cputimer(p, &cputime);
 648         switch (CPUCLOCK_WHICH(which_clock)) {
 649         default:
 650                 return -EINVAL;
 651         case CPUCLOCK_PROF:
 652                 cpu->cpu = cputime_add(cputime.utime, cputime.stime);
 653                 break;
 654         case CPUCLOCK_VIRT:
 655                 cpu->cpu = cputime.utime;
 656                 break;
 657         case CPUCLOCK_SCHED:
 658                 cpu->sched = cputime.sum_exec_runtime + task_delta_exec(p);
 659                 break;
 660         }
 661         return 0;
 662 }
 663
 664 /*
 665  * Guts of sys_timer_settime for CPU timers.
 666  * This is called with the timer locked and interrupts disabled.
 667  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 668  * and try again.  (This happens when the timer is in the middle of firing.)
 669  */
 670 static int posix_cpu_timer_set(struct k_itimer *timer, int flags,
 671                                struct itimerspec *new, struct itimerspec *old)
 672 {
 673         struct task_struct *p = timer->it.cpu.task;
 674         union cpu_time_count old_expires, new_expires, old_incr, val;
 675         int ret;
 676
 677         if (unlikely(p == NULL)) {
 678                 /*
 679                  * Timer refers to a dead task's clock.
 680                  */
 681                 return -ESRCH;
 682         }
 683
 684         new_expires = timespec_to_sample(timer->it_clock, &new->it_value);
 685
 686         read_lock(&tasklist_lock);
 687         /*
 688          * We need the tasklist_lock to protect against reaping that
 689          * clears p->sighand.  If p has just been reaped, we can no
 690          * longer get any information about it at all.
 691          */
 692         if (unlikely(p->sighand == NULL)) {
 693                 read_unlock(&tasklist_lock);
 694                 put_task_struct(p);
 695                 timer->it.cpu.task = NULL;
 696                 return -ESRCH;
 697         }
 698
 699         /*
 700          * Disarm any old timer after extracting its expiry time.
 701          */
 702         BUG_ON(!irqs_disabled());
 703
 704         ret = 0;
 705         old_incr = timer->it.cpu.incr;
 706         spin_lock(&p->sighand->siglock);
 707         old_expires = timer->it.cpu.expires;
 708         if (unlikely(timer->it.cpu.firing)) {
 709                 timer->it.cpu.firing = -1;
 710                 ret = TIMER_RETRY;
 711         } else
 712                 list_del_init(&timer->it.cpu.entry);
 713
 714         /*
 715          * We need to sample the current value to convert the new
 716          * value from to relative and absolute, and to convert the
 717          * old value from absolute to relative.  To set a process
 718          * timer, we need a sample to balance the thread expiry
 719          * times (in arm_timer).  With an absolute time, we must
 720          * check if it's already passed.  In short, we need a sample.
 721          */
 722         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 723                 cpu_clock_sample(timer->it_clock, p, &val);
 724         } else {
 725                 cpu_timer_sample_group(timer->it_clock, p, &val);
 726         }
 727
 728         if (old) {
 729                 if (old_expires.sched == 0) {
 730                         old->it_value.tv_sec = 0;
 731                         old->it_value.tv_nsec = 0;
 732                 } else {
 733                         /*
 734                          * Update the timer in case it has
 735                          * overrun already.  If it has,
 736                          * we'll report it as having overrun
 737                          * and with the next reloaded timer
 738                          * already ticking, though we are
 739                          * swallowing that pending
 740                          * notification here to install the
 741                          * new setting.
 742                          */
 743                         bump_cpu_timer(timer, val);
 744                         if (cpu_time_before(timer->it_clock, val,
 745                                             timer->it.cpu.expires)) {
 746                                 old_expires = cpu_time_sub(
 747                                         timer->it_clock,
 748                                         timer->it.cpu.expires, val);
 749                                 sample_to_timespec(timer->it_clock,
 750                                                    old_expires,
 751                                                    &old->it_value);
 752                         } else {
 753                                 old->it_value.tv_nsec = 1;
 754                                 old->it_value.tv_sec = 0;
 755                         }
 756                 }
 757         }
 758
 759         if (unlikely(ret)) {
 760                 /*
 761                  * We are colliding with the timer actually firing.
 762                  * Punt after filling in the timer's old value, and
 763                  * disable this firing since we are already reporting
 764                  * it as an overrun (thanks to bump_cpu_timer above).
 765                  */
 766                 spin_unlock(&p->sighand->siglock);
 767                 read_unlock(&tasklist_lock);
 768                 goto out;
 769         }
 770
 771         if (new_expires.sched != 0 && !(flags & TIMER_ABSTIME)) {
 772                 cpu_time_add(timer->it_clock, &new_expires, val);
 773         }
 774
 775         /*
 776          * Install the new expiry time (or zero).
 777          * For a timer with no notification action, we don't actually
 778          * arm the timer (we'll just fake it for timer_gettime).
 779          */
 780         timer->it.cpu.expires = new_expires;
 781         if (new_expires.sched != 0 &&
 782             cpu_time_before(timer->it_clock, val, new_expires)) {
 783                 arm_timer(timer);
 784         }
 785
 786         spin_unlock(&p->sighand->siglock);
 787         read_unlock(&tasklist_lock);
 788
 789         /*
 790          * Install the new reload setting, and
 791          * set up the signal and overrun bookkeeping.
 792          */
 793         timer->it.cpu.incr = timespec_to_sample(timer->it_clock,
 794                                                 &new->it_interval);
 795
 796         /*
 797          * This acts as a modification timestamp for the timer,
 798          * so any automatic reload attempt will punt on seeing
 799          * that we have reset the timer manually.
 800          */
 801         timer->it_requeue_pending = (timer->it_requeue_pending + 2) &
 802                 ~REQUEUE_PENDING;
 803         timer->it_overrun_last = 0;
 804         timer->it_overrun = -1;
 805
 806         if (new_expires.sched != 0 &&
 807             !cpu_time_before(timer->it_clock, val, new_expires)) {
 808                 /*
 809                  * The designated time already passed, so we notify
 810                  * immediately, even if the thread never runs to
 811                  * accumulate more time on this clock.
 812                  */
 813                 cpu_timer_fire(timer);
 814         }
 815
 816         ret = 0;
 817  out:
 818         if (old) {
 819                 sample_to_timespec(timer->it_clock,
 820                                    old_incr, &old->it_interval);
 821         }
 822         return ret;
 823 }
 824
 825 static void posix_cpu_timer_get(struct k_itimer *timer, struct itimerspec *itp)
 826 {
 827         union cpu_time_count now;
 828         struct task_struct *p = timer->it.cpu.task;
 829         int clear_dead;
 830
 831         /*
 832          * Easy part: convert the reload time.
 833          */
 834         sample_to_timespec(timer->it_clock,
 835                            timer->it.cpu.incr, &itp->it_interval);
 836
 837         if (timer->it.cpu.expires.sched == 0) { /* Timer not armed at all.  */
 838                 itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
 839                 return;
 840         }
 841
 842         if (unlikely(p == NULL)) {
 843                 /*
 844                  * This task already died and the timer will never fire.
 845                  * In this case, expires is actually the dead value.
 846                  */
 847         dead:
 848                 sample_to_timespec(timer->it_clock, timer->it.cpu.expires,
 849                                    &itp->it_value);
 850                 return;
 851         }
 852
 853         /*
 854          * Sample the clock to take the difference with the expiry time.
 855          */
 856         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 857                 cpu_clock_sample(timer->it_clock, p, &now);
 858                 clear_dead = p->exit_state;
 859         } else {
 860                 read_lock(&tasklist_lock);
 861                 if (unlikely(p->sighand == NULL)) {
 862                         /*
 863                          * The process has been reaped.
 864                          * We can't even collect a sample any more.
 865                          * Call the timer disarmed, nothing else to do.
 866                          */
 867                         put_task_struct(p);
 868                         timer->it.cpu.task = NULL;
 869                         timer->it.cpu.expires.sched = 0;
 870                         read_unlock(&tasklist_lock);
 871                         goto dead;
 872                 } else {
 873                         cpu_timer_sample_group(timer->it_clock, p, &now);
 874                         clear_dead = (unlikely(p->exit_state) &&
 875                                       thread_group_empty(p));
 876                 }
 877                 read_unlock(&tasklist_lock);
 878         }
 879
 880         if (unlikely(clear_dead)) {
 881                 /*
 882                  * We've noticed that the thread is dead, but
 883                  * not yet reaped.  Take this opportunity to
 884                  * drop our task ref.
 885                  */
 886                 clear_dead_task(timer, now);
 887                 goto dead;
 888         }
 889
 890         if (cpu_time_before(timer->it_clock, now, timer->it.cpu.expires)) {
 891                 sample_to_timespec(timer->it_clock,
 892                                    cpu_time_sub(timer->it_clock,
 893                                                 timer->it.cpu.expires, now),
 894                                    &itp->it_value);
 895         } else {
 896                 /*
 897                  * The timer should have expired already, but the firing
 898                  * hasn't taken place yet.  Say it's just about to expire.
 899                  */
 900                 itp->it_value.tv_nsec = 1;
 901                 itp->it_value.tv_sec = 0;
 902         }
 903 }
 904
 905 /*
 906  * Check for any per-thread CPU timers that have fired and move them off
 907  * the tsk->cpu_timers[N] list onto the firing list.  Here we update the
 908  * tsk->it_*_expires values to reflect the remaining thread CPU timers.
 909  */
 910 static void check_thread_timers(struct task_struct *tsk,
 911                                 struct list_head *firing)
 912 {
 913         int maxfire;
 914         struct list_head *timers = tsk->cpu_timers;
 915         struct signal_struct *const sig = tsk->signal;
 916         unsigned long soft;
 917
 918         maxfire = 20;
 919         tsk->cputime_expires.prof_exp = cputime_zero;
 920         while (!list_empty(timers)) {
 921                 struct cpu_timer_list *t = list_first_entry(timers,
 922                                                       struct cpu_timer_list,
 923                                                       entry);
 924                 if (!--maxfire || cputime_lt(prof_ticks(tsk), t->expires.cpu)) {
 925                         tsk->cputime_expires.prof_exp = t->expires.cpu;
 926                         break;
 927                 }
 928                 t->firing = 1;
 929                 list_move_tail(&t->entry, firing);
 930         }
 931
 932         ++timers;
 933         maxfire = 20;
 934         tsk->cputime_expires.virt_exp = cputime_zero;
 935         while (!list_empty(timers)) {
 936                 struct cpu_timer_list *t = list_first_entry(timers,
 937                                                       struct cpu_timer_list,
 938                                                       entry);
 939                 if (!--maxfire || cputime_lt(virt_ticks(tsk), t->expires.cpu)) {
 940                         tsk->cputime_expires.virt_exp = t->expires.cpu;
 941                         break;
 942                 }
 943                 t->firing = 1;
 944                 list_move_tail(&t->entry, firing);
 945         }
 946
 947         ++timers;
 948         maxfire = 20;
 949         tsk->cputime_expires.sched_exp = 0;
 950         while (!list_empty(timers)) {
 951                 struct cpu_timer_list *t = list_first_entry(timers,
 952                                                       struct cpu_timer_list,
 953                                                       entry);
 954                 if (!--maxfire || tsk->se.sum_exec_runtime < t->expires.sched) {
 955                         tsk->cputime_expires.sched_exp = t->expires.sched;
 956                         break;
 957                 }
 958                 t->firing = 1;
 959                 list_move_tail(&t->entry, firing);
 960         }
 961
 962         /*
 963          * Check for the special case thread timers.
 964          */
 965         soft = ACCESS_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_cur);
 966         if (soft != RLIM_INFINITY) {
 967                 unsigned long hard =
 968                         ACCESS_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_max);
 969
 970                 if (hard != RLIM_INFINITY &&
 971                     tsk->rt.timeout > DIV_ROUND_UP(hard, USEC_PER_SEC/HZ)) {
 972                         /*
 973                          * At the hard limit, we just die.
 974                          * No need to calculate anything else now.
 975                          */
 976                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
 977                         return;
 978                 }
 979                 if (tsk->rt.timeout > DIV_ROUND_UP(soft, USEC_PER_SEC/HZ)) {
 980                         /*
 981                          * At the soft limit, send a SIGXCPU every second.
 982                          */
 983                         if (soft < hard) {
 984                                 soft += USEC_PER_SEC;
 985                                 sig->rlim[RLIMIT_RTTIME].rlim_cur = soft;
 986                         }
 987                         printk(KERN_INFO
 988                                 "RT Watchdog Timeout: %s[%d]\n",
 989                                 tsk->comm, task_pid_nr(tsk));
 990                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
 991                 }
 992         }
 993 }
 994
 995 static void stop_process_timers(struct signal_struct *sig)
 996 {
 997         struct thread_group_cputimer *cputimer = &sig->cputimer;
 998         unsigned long flags;
 999
1000         spin_lock_irqsave(&cputimer->lock, flags);
1001         cputimer->running = 0;
1002         spin_unlock_irqrestore(&cputimer->lock, flags);
1003 }
1004
1005 static u32 onecputick;
1006
1007 static void check_cpu_itimer(struct task_struct *tsk, struct cpu_itimer *it,
1008                              cputime_t *expires, cputime_t cur_time, int signo)
1009 {
1010         if (cputime_eq(it->expires, cputime_zero))
1011                 return;
1012
1013         if (cputime_ge(cur_time, it->expires)) {
1014                 if (!cputime_eq(it->incr, cputime_zero)) {
1015                         it->expires = cputime_add(it->expires, it->incr);
1016                         it->error += it->incr_error;
1017                         if (it->error >= onecputick) {
1018                                 it->expires = cputime_sub(it->expires,
1019                                                           cputime_one_jiffy);
1020                                 it->error -= onecputick;
1021                         }
1022                 } else {
1023                         it->expires = cputime_zero;
1024                 }
1025
1026                 trace_itimer_expire(signo == SIGPROF ?
1027                                     ITIMER_PROF : ITIMER_VIRTUAL,
1028                                     tsk->signal->leader_pid, cur_time);
1029                 __group_send_sig_info(signo, SEND_SIG_PRIV, tsk);
1030         }
1031
1032         if (!cputime_eq(it->expires, cputime_zero) &&
1033             (cputime_eq(*expires, cputime_zero) ||
1034              cputime_lt(it->expires, *expires))) {
1035                 *expires = it->expires;
1036         }
1037 }
1038
1039 /**
1040  * task_cputime_zero - Check a task_cputime struct for all zero fields.
1041  *
1042  * @cputime:    The struct to compare.
1043  *
1044  * Checks @cputime to see if all fields are zero.  Returns true if all fields
1045  * are zero, false if any field is nonzero.
1046  */
1047 static inline int task_cputime_zero(const struct task_cputime *cputime)
1048 {
1049         if (cputime_eq(cputime->utime, cputime_zero) &&
1050             cputime_eq(cputime->stime, cputime_zero) &&
1051             cputime->sum_exec_runtime == 0)
1052                 return 1;
1053         return 0;
1054 }
1055
1056 /*
1057  * Check for any per-thread CPU timers that have fired and move them
1058  * off the tsk->*_timers list onto the firing list.  Per-thread timers
1059  * have already been taken off.
1060  */
1061 static void check_process_timers(struct task_struct *tsk,
1062                                  struct list_head *firing)
1063 {
1064         int maxfire;
1065         struct signal_struct *const sig = tsk->signal;
1066         cputime_t utime, ptime, virt_expires, prof_expires;
1067         unsigned long long sum_sched_runtime, sched_expires;
1068         struct list_head *timers = sig->cpu_timers;
1069         struct task_cputime cputime;
1070         unsigned long soft;
1071
1072         /*
1073          * Collect the current process totals.
1074          */
1075         thread_group_cputimer(tsk, &cputime);
1076         utime = cputime.utime;
1077         ptime = cputime_add(utime, cputime.stime);
1078         sum_sched_runtime = cputime.sum_exec_runtime;
1079         maxfire = 20;
1080         prof_expires = cputime_zero;
1081         while (!list_empty(timers)) {
1082                 struct cpu_timer_list *tl = list_first_entry(timers,
1083                                                       struct cpu_timer_list,
1084                                                       entry);
1085                 if (!--maxfire || cputime_lt(ptime, tl->expires.cpu)) {
1086                         prof_expires = tl->expires.cpu;
1087                         break;
1088                 }
1089                 tl->firing = 1;
1090                 list_move_tail(&tl->entry, firing);
1091         }
1092
1093         ++timers;
1094         maxfire = 20;
1095         virt_expires = cputime_zero;
1096         while (!list_empty(timers)) {
1097                 struct cpu_timer_list *tl = list_first_entry(timers,
1098                                                       struct cpu_timer_list,
1099                                                       entry);
1100                 if (!--maxfire || cputime_lt(utime, tl->expires.cpu)) {
1101                         virt_expires = tl->expires.cpu;
1102                         break;
1103                 }
1104                 tl->firing = 1;
1105                 list_move_tail(&tl->entry, firing);
1106         }
1107
1108         ++timers;
1109         maxfire = 20;
1110         sched_expires = 0;
1111         while (!list_empty(timers)) {
1112                 struct cpu_timer_list *tl = list_first_entry(timers,
1113                                                       struct cpu_timer_list,
1114                                                       entry);
1115                 if (!--maxfire || sum_sched_runtime < tl->expires.sched) {
1116                         sched_expires = tl->expires.sched;
1117                         break;
1118                 }
1119                 tl->firing = 1;
1120                 list_move_tail(&tl->entry, firing);
1121         }
1122
1123         /*
1124          * Check for the special case process timers.
1125          */
1126         check_cpu_itimer(tsk, &sig->it[CPUCLOCK_PROF], &prof_expires, ptime,
1127                          SIGPROF);
1128         check_cpu_itimer(tsk, &sig->it[CPUCLOCK_VIRT], &virt_expires, utime,
1129                          SIGVTALRM);
1130         soft = ACCESS_ONCE(sig->rlim[RLIMIT_CPU].rlim_cur);
1131         if (soft != RLIM_INFINITY) {
1132                 unsigned long psecs = cputime_to_secs(ptime);
1133                 unsigned long hard =
1134                         ACCESS_ONCE(sig->rlim[RLIMIT_CPU].rlim_max);
1135                 cputime_t x;
1136                 if (psecs >= hard) {
1137                         /*
1138                          * At the hard limit, we just die.
1139                          * No need to calculate anything else now.
1140                          */
1141                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
1142                         return;
1143                 }
1144                 if (psecs >= soft) {
1145                         /*
1146                          * At the soft limit, send a SIGXCPU every second.
1147                          */
1148                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
1149                         if (soft < hard) {
1150                                 soft++;
1151                                 sig->rlim[RLIMIT_CPU].rlim_cur = soft;
1152                         }
1153                 }
1154                 x = secs_to_cputime(soft);
1155                 if (cputime_eq(prof_expires, cputime_zero) ||
1156                     cputime_lt(x, prof_expires)) {
1157                         prof_expires = x;
1158                 }
1159         }
1160
1161         sig->cputime_expires.prof_exp = prof_expires;
1162         sig->cputime_expires.virt_exp = virt_expires;
1163         sig->cputime_expires.sched_exp = sched_expires;
1164         if (task_cputime_zero(&sig->cputime_expires))
1165                 stop_process_timers(sig);
1166 }
1167
1168 /*
1169  * This is called from the signal code (via do_schedule_next_timer)
1170  * when the last timer signal was delivered and we have to reload the timer.
1171  */
1172 void posix_cpu_timer_schedule(struct k_itimer *timer)
1173 {
1174         struct task_struct *p = timer->it.cpu.task;
1175         union cpu_time_count now;
1176
1177         if (unlikely(p == NULL))
1178                 /*
1179                  * The task was cleaned up already, no future firings.
1180                  */
1181                 goto out;
1182
1183         /*
1184          * Fetch the current sample and update the timer's expiry time.
1185          */
1186         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
1187                 cpu_clock_sample(timer->it_clock, p, &now);
1188                 bump_cpu_timer(timer, now);
1189                 if (unlikely(p->exit_state)) {
1190                         clear_dead_task(timer, now);
1191                         goto out;
1192                 }
1193                 read_lock(&tasklist_lock); /* arm_timer needs it.  */
1194                 spin_lock(&p->sighand->siglock);
1195         } else {
1196                 read_lock(&tasklist_lock);
1197                 if (unlikely(p->sighand == NULL)) {
1198                         /*
1199                          * The process has been reaped.
1200                          * We can't even collect a sample any more.
1201                          */
1202                         put_task_struct(p);
1203                         timer->it.cpu.task = p = NULL;
1204                         timer->it.cpu.expires.sched = 0;
1205                         goto out_unlock;
1206                 } else if (unlikely(p->exit_state) && thread_group_empty(p)) {
1207                         /*
1208                          * We've noticed that the thread is dead, but
1209                          * not yet reaped.  Take this opportunity to
1210                          * drop our task ref.
1211                          */
1212                         clear_dead_task(timer, now);
1213                         goto out_unlock;
1214                 }
1215                 spin_lock(&p->sighand->siglock);
1216                 cpu_timer_sample_group(timer->it_clock, p, &now);
1217                 bump_cpu_timer(timer, now);
1218                 /* Leave the tasklist_lock locked for the call below.  */
1219         }
1220
1221         /*
1222          * Now re-arm for the new expiry time.
1223          */
1224         BUG_ON(!irqs_disabled());
1225         arm_timer(timer);
1226         spin_unlock(&p->sighand->siglock);
1227
1228 out_unlock:
1229         read_unlock(&tasklist_lock);
1230
1231 out:
1232         timer->it_overrun_last = timer->it_overrun;
1233         timer->it_overrun = -1;
1234         ++timer->it_requeue_pending;
1235 }
1236
1237 /**
1238  * task_cputime_expired - Compare two task_cputime entities.
1239  *
1240  * @sample:     The task_cputime structure to be checked for expiration.
1241  * @expires:    Expiration times, against which @sample will be checked.
1242  *
1243  * Checks @sample against @expires to see if any field of @sample has expired.
1244  * Returns true if any field of the former is greater than the corresponding
1245  * field of the latter if the latter field is set.  Otherwise returns false.
1246  */
1247 static inline int task_cputime_expired(const struct task_cputime *sample,
1248                                         const struct task_cputime *expires)
1249 {
1250         if (!cputime_eq(expires->utime, cputime_zero) &&
1251             cputime_ge(sample->utime, expires->utime))
1252                 return 1;
1253         if (!cputime_eq(expires->stime, cputime_zero) &&
1254             cputime_ge(cputime_add(sample->utime, sample->stime),
1255                        expires->stime))
1256                 return 1;
1257         if (expires->sum_exec_runtime != 0 &&
1258             sample->sum_exec_runtime >= expires->sum_exec_runtime)
1259                 return 1;
1260         return 0;
1261 }
1262
1263 /**
1264  * fastpath_timer_check - POSIX CPU timers fast path.
1265  *
1266  * @tsk:        The task (thread) being checked.
1267  *
1268  * Check the task and thread group timers.  If both are zero (there are no
1269  * timers set) return false.  Otherwise snapshot the task and thread group
1270  * timers and compare them with the corresponding expiration times.  Return
1271  * true if a timer has expired, else return false.
1272  */
1273 static inline int fastpath_timer_check(struct task_struct *tsk)
1274 {
1275         struct signal_struct *sig;
1276
1277         if (!task_cputime_zero(&tsk->cputime_expires)) {
1278                 struct task_cputime task_sample = {
1279                         .utime = tsk->utime,
1280                         .stime = tsk->stime,
1281                         .sum_exec_runtime = tsk->se.sum_exec_runtime
1282                 };
1283
1284                 if (task_cputime_expired(&task_sample, &tsk->cputime_expires))
1285                         return 1;
1286         }
1287
1288         sig = tsk->signal;
1289         if (sig->cputimer.running) {
1290                 struct task_cputime group_sample;
1291
1292                 spin_lock(&sig->cputimer.lock);
1293                 group_sample = sig->cputimer.cputime;
1294                 spin_unlock(&sig->cputimer.lock);
1295
1296                 if (task_cputime_expired(&group_sample, &sig->cputime_expires))
1297                         return 1;
1298         }
1299
1300         return 0;
1301 }
1302
1303 /*
1304  * This is called from the timer interrupt handler.  The irq handler has
1305  * already updated our counts.  We need to check if any timers fire now.
1306  * Interrupts are disabled.
1307  */
1308 void run_posix_cpu_timers(struct task_struct *tsk)
1309 {
1310         LIST_HEAD(firing);
1311         struct k_itimer *timer, *next;
1312         unsigned long flags;
1313
1314         BUG_ON(!irqs_disabled());
1315
1316         /*
1317          * The fast path checks that there are no expired thread or thread
1318          * group timers.  If that's so, just return.
1319          */
1320         if (!fastpath_timer_check(tsk))
1321                 return;
1322
1323         if (!lock_task_sighand(tsk, &flags))
1324                 return;
1325         /*
1326          * Here we take off tsk->signal->cpu_timers[N] and
1327          * tsk->cpu_timers[N] all the timers that are firing, and
1328          * put them on the firing list.
1329          */
1330         check_thread_timers(tsk, &firing);
1331         /*
1332          * If there are any active process wide timers (POSIX 1.b, itimers,
1333          * RLIMIT_CPU) cputimer must be running.
1334          */
1335         if (tsk->signal->cputimer.running)
1336                 check_process_timers(tsk, &firing);
1337
1338         /*
1339          * We must release these locks before taking any timer's lock.
1340          * There is a potential race with timer deletion here, as the
1341          * siglock now protects our private firing list.  We have set
1342          * the firing flag in each timer, so that a deletion attempt
1343          * that gets the timer lock before we do will give it up and
1344          * spin until we've taken care of that timer below.
1345          */
1346         unlock_task_sighand(tsk, &flags);
1347
1348         /*
1349          * Now that all the timers on our list have the firing flag,
1350          * no one will touch their list entries but us.  We'll take
1351          * each timer's lock before clearing its firing flag, so no
1352          * timer call will interfere.
1353          */
1354         list_for_each_entry_safe(timer, next, &firing, it.cpu.entry) {
1355                 int cpu_firing;
1356
1357                 spin_lock(&timer->it_lock);
1358                 list_del_init(&timer->it.cpu.entry);
1359                 cpu_firing = timer->it.cpu.firing;
1360                 timer->it.cpu.firing = 0;
1361                 /*
1362                  * The firing flag is -1 if we collided with a reset
1363                  * of the timer, which already reported this
1364                  * almost-firing as an overrun.  So don't generate an event.
1365                  */
1366                 if (likely(cpu_firing >= 0))
1367                         cpu_timer_fire(timer);
1368                 spin_unlock(&timer->it_lock);
1369         }
1370 }
1371
1372 /*
1373  * Set one of the process-wide special case CPU timers or RLIMIT_CPU.
1374  * The tsk->sighand->siglock must be held by the caller.
1375  */
1376 void set_process_cpu_timer(struct task_struct *tsk, unsigned int clock_idx,
1377                            cputime_t *newval, cputime_t *oldval)
1378 {
1379         union cpu_time_count now;
1380
1381         BUG_ON(clock_idx == CPUCLOCK_SCHED);
1382         cpu_timer_sample_group(clock_idx, tsk, &now);
1383
1384         if (oldval) {
1385                 /*
1386                  * We are setting itimer. The *oldval is absolute and we update
1387                  * it to be relative, *newval argument is relative and we update
1388                  * it to be absolute.
1389                  */
1390                 if (!cputime_eq(*oldval, cputime_zero)) {
1391                         if (cputime_le(*oldval, now.cpu)) {
1392                                 /* Just about to fire. */
1393                                 *oldval = cputime_one_jiffy;
1394                         } else {
1395                                 *oldval = cputime_sub(*oldval, now.cpu);
1396                         }
1397                 }
1398
1399                 if (cputime_eq(*newval, cputime_zero))
1400                         return;
1401                 *newval = cputime_add(*newval, now.cpu);
1402         }
1403
1404         /*
1405          * Update expiration cache if we are the earliest timer, or eventually
1406          * RLIMIT_CPU limit is earlier than prof_exp cpu timer expire.
1407          */
1408         switch (clock_idx) {
1409         case CPUCLOCK_PROF:
1410                 if (expires_gt(tsk->signal->cputime_expires.prof_exp, *newval))
1411                         tsk->signal->cputime_expires.prof_exp = *newval;
1412                 break;
1413         case CPUCLOCK_VIRT:
1414                 if (expires_gt(tsk->signal->cputime_expires.virt_exp, *newval))
1415                         tsk->signal->cputime_expires.virt_exp = *newval;
1416                 break;
1417         }
1418 }
1419
1420 static int do_cpu_nanosleep(const clockid_t which_clock, int flags,
1421                             struct timespec *rqtp, struct itimerspec *it)
1422 {
1423         struct k_itimer timer;
1424         int error;
1425
1426         /*
1427          * Set up a temporary timer and then wait for it to go off.
1428          */
1429         memset(&timer, 0, sizeof timer);
1430         spin_lock_init(&timer.it_lock);
1431         timer.it_clock = which_clock;
1432         timer.it_overrun = -1;
1433         error = posix_cpu_timer_create(&timer);
1434         timer.it_process = current;
1435         if (!error) {
1436                 static struct itimerspec zero_it;
1437
1438                 memset(it, 0, sizeof *it);
1439                 it->it_value = *rqtp;
1440
1441                 spin_lock_irq(&timer.it_lock);
1442                 error = posix_cpu_timer_set(&timer, flags, it, NULL);
1443                 if (error) {
1444                         spin_unlock_irq(&timer.it_lock);
1445                         return error;
1446                 }
1447
1448                 while (!signal_pending(current)) {
1449                         if (timer.it.cpu.expires.sched == 0) {
1450                                 /*
1451                                  * Our timer fired and was reset.
1452                                  */
1453                                 spin_unlock_irq(&timer.it_lock);
1454                                 return 0;
1455                         }
1456
1457                         /*
1458                          * Block until cpu_timer_fire (or a signal) wakes us.
1459                          */
1460                         __set_current_state(TASK_INTERRUPTIBLE);
1461                         spin_unlock_irq(&timer.it_lock);
1462                         schedule();
1463                         spin_lock_irq(&timer.it_lock);
1464                 }
1465
1466                 /*
1467                  * We were interrupted by a signal.
1468                  */
1469                 sample_to_timespec(which_clock, timer.it.cpu.expires, rqtp);
1470                 posix_cpu_timer_set(&timer, 0, &zero_it, it);
1471                 spin_unlock_irq(&timer.it_lock);
1472
1473                 if ((it->it_value.tv_sec | it->it_value.tv_nsec) == 0) {
1474                         /*
1475                          * It actually did fire already.
1476                          */
1477                         return 0;
1478                 }
1479
1480                 error = -ERESTART_RESTARTBLOCK;
1481         }
1482
1483         return error;
1484 }
1485
1486 static long posix_cpu_nsleep_restart(struct restart_block *restart_block);
1487
1488 static int posix_cpu_nsleep(const clockid_t which_clock, int flags,
1489                             struct timespec *rqtp, struct timespec __user *rmtp)
1490 {
1491         struct restart_block *restart_block =
1492                 &current_thread_info()->restart_block;
1493         struct itimerspec it;
1494         int error;
1495
1496         /*
1497          * Diagnose required errors first.
1498          */
1499         if (CPUCLOCK_PERTHREAD(which_clock) &&
1500             (CPUCLOCK_PID(which_clock) == 0 ||
1501              CPUCLOCK_PID(which_clock) == current->pid))
1502                 return -EINVAL;
1503
1504         error = do_cpu_nanosleep(which_clock, flags, rqtp, &it);
1505
1506         if (error == -ERESTART_RESTARTBLOCK) {
1507
1508                 if (flags & TIMER_ABSTIME)
1509                         return -ERESTARTNOHAND;
1510                 /*
1511                  * Report back to the user the time still remaining.
1512                  */
1513                 if (rmtp && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1514                         return -EFAULT;
1515
1516                 restart_block->fn = posix_cpu_nsleep_restart;
1517                 restart_block->nanosleep.clockid = which_clock;
1518                 restart_block->nanosleep.rmtp = rmtp;
1519                 restart_block->nanosleep.expires = timespec_to_ns(rqtp);
1520         }
1521         return error;
1522 }
1523
1524 static long posix_cpu_nsleep_restart(struct restart_block *restart_block)
1525 {
1526         clockid_t which_clock = restart_block->nanosleep.clockid;
1527         struct timespec t;
1528         struct itimerspec it;
1529         int error;
1530
1531         t = ns_to_timespec(restart_block->nanosleep.expires);
1532
1533         error = do_cpu_nanosleep(which_clock, TIMER_ABSTIME, &t, &it);
1534
1535         if (error == -ERESTART_RESTARTBLOCK) {
1536                 struct timespec __user *rmtp = restart_block->nanosleep.rmtp;
1537                 /*
1538                  * Report back to the user the time still remaining.
1539                  */
1540                 if (rmtp && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1541                         return -EFAULT;
1542
1543                 restart_block->nanosleep.expires = timespec_to_ns(&t);
1544         }
1545         return error;
1546
1547 }
1548
1549 #define PROCESS_CLOCK   MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED)
1550 #define THREAD_CLOCK    MAKE_THREAD_CPUCLOCK(0, CPUCLOCK_SCHED)
1551
1552 static int process_cpu_clock_getres(const clockid_t which_clock,
1553                                     struct timespec *tp)
1554 {
1555         return posix_cpu_clock_getres(PROCESS_CLOCK, tp);
1556 }
1557 static int process_cpu_clock_get(const clockid_t which_clock,
1558                                  struct timespec *tp)
1559 {
1560         return posix_cpu_clock_get(PROCESS_CLOCK, tp);
1561 }
1562 static int process_cpu_timer_create(struct k_itimer *timer)
1563 {
1564         timer->it_clock = PROCESS_CLOCK;
1565         return posix_cpu_timer_create(timer);
1566 }
1567 static int process_cpu_nsleep(const clockid_t which_clock, int flags,
1568                               struct timespec *rqtp,
1569                               struct timespec __user *rmtp)
1570 {
1571         return posix_cpu_nsleep(PROCESS_CLOCK, flags, rqtp, rmtp);
1572 }
1573 static long process_cpu_nsleep_restart(struct restart_block *restart_block)
1574 {
1575         return -EINVAL;
1576 }
1577 static int thread_cpu_clock_getres(const clockid_t which_clock,
1578                                    struct timespec *tp)
1579 {
1580         return posix_cpu_clock_getres(THREAD_CLOCK, tp);
1581 }
1582 static int thread_cpu_clock_get(const clockid_t which_clock,
1583                                 struct timespec *tp)
1584 {
1585         return posix_cpu_clock_get(THREAD_CLOCK, tp);
1586 }
1587 static int thread_cpu_timer_create(struct k_itimer *timer)
1588 {
1589         timer->it_clock = THREAD_CLOCK;
1590         return posix_cpu_timer_create(timer);
1591 }
1592
1593 struct k_clock clock_posix_cpu = {
1594         .clock_getres   = posix_cpu_clock_getres,
1595         .clock_set      = posix_cpu_clock_set,
1596         .clock_get      = posix_cpu_clock_get,
1597         .timer_create   = posix_cpu_timer_create,
1598         .nsleep         = posix_cpu_nsleep,
1599         .nsleep_restart = posix_cpu_nsleep_restart,
1600         .timer_set      = posix_cpu_timer_set,
1601         .timer_del      = posix_cpu_timer_del,
1602         .timer_get      = posix_cpu_timer_get,
1603 };
1604
1605 static __init int init_posix_cpu_timers(void)
1606 {
1607         struct k_clock process = {
1608                 .clock_getres   = process_cpu_clock_getres,
1609                 .clock_get      = process_cpu_clock_get,
1610                 .timer_create   = process_cpu_timer_create,
1611                 .nsleep         = process_cpu_nsleep,
1612                 .nsleep_restart = process_cpu_nsleep_restart,
1613         };
1614         struct k_clock thread = {
1615                 .clock_getres   = thread_cpu_clock_getres,
1616                 .clock_get      = thread_cpu_clock_get,
1617                 .timer_create   = thread_cpu_timer_create,
1618         };
1619         struct timespec ts;
1620
1621         posix_timers_register_clock(CLOCK_PROCESS_CPUTIME_ID, &process);
1622         posix_timers_register_clock(CLOCK_THREAD_CPUTIME_ID, &thread);
1623
1624         cputime_to_timespec(cputime_one_jiffy, &ts);
1625         onecputick = ts.tv_nsec;
1626         WARN_ON(ts.tv_sec != 0);
1627
1628         return 0;
1629 }
1630 __initcall(init_posix_cpu_timers);