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 int posix_cpu_clock_getres(const clockid_t which_clock, struct timespec *tp)
 180 {
 181         int error = check_clock(which_clock);
 182         if (!error) {
 183                 tp->tv_sec = 0;
 184                 tp->tv_nsec = ((NSEC_PER_SEC + HZ - 1) / HZ);
 185                 if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
 186                         /*
 187                          * If sched_clock is using a cycle counter, we
 188                          * don't have any idea of its true resolution
 189                          * exported, but it is much more than 1s/HZ.
 190                          */
 191                         tp->tv_nsec = 1;
 192                 }
 193         }
 194         return error;
 195 }
 196
 197 int posix_cpu_clock_set(const clockid_t which_clock, const struct timespec *tp)
 198 {
 199         /*
 200          * You can never reset a CPU clock, but we check for other errors
 201          * in the call before failing with EPERM.
 202          */
 203         int error = check_clock(which_clock);
 204         if (error == 0) {
 205                 error = -EPERM;
 206         }
 207         return error;
 208 }
 209
 210
 211 /*
 212  * Sample a per-thread clock for the given task.
 213  */
 214 static int cpu_clock_sample(const clockid_t which_clock, struct task_struct *p,
 215                             union cpu_time_count *cpu)
 216 {
 217         switch (CPUCLOCK_WHICH(which_clock)) {
 218         default:
 219                 return -EINVAL;
 220         case CPUCLOCK_PROF:
 221                 cpu->cpu = prof_ticks(p);
 222                 break;
 223         case CPUCLOCK_VIRT:
 224                 cpu->cpu = virt_ticks(p);
 225                 break;
 226         case CPUCLOCK_SCHED:
 227                 cpu->sched = task_sched_runtime(p);
 228                 break;
 229         }
 230         return 0;
 231 }
 232
 233 void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
 234 {
 235         struct signal_struct *sig = tsk->signal;
 236         struct task_struct *t;
 237
 238         times->utime = sig->utime;
 239         times->stime = sig->stime;
 240         times->sum_exec_runtime = sig->sum_sched_runtime;
 241
 242         rcu_read_lock();
 243         /* make sure we can trust tsk->thread_group list */
 244         if (!likely(pid_alive(tsk)))
 245                 goto out;
 246
 247         t = tsk;
 248         do {
 249                 times->utime = cputime_add(times->utime, t->utime);
 250                 times->stime = cputime_add(times->stime, t->stime);
 251                 times->sum_exec_runtime += t->se.sum_exec_runtime;
 252         } while_each_thread(tsk, t);
 253 out:
 254         rcu_read_unlock();
 255 }
 256
 257 static void update_gt_cputime(struct task_cputime *a, struct task_cputime *b)
 258 {
 259         if (cputime_gt(b->utime, a->utime))
 260                 a->utime = b->utime;
 261
 262         if (cputime_gt(b->stime, a->stime))
 263                 a->stime = b->stime;
 264
 265         if (b->sum_exec_runtime > a->sum_exec_runtime)
 266                 a->sum_exec_runtime = b->sum_exec_runtime;
 267 }
 268
 269 void thread_group_cputimer(struct task_struct *tsk, struct task_cputime *times)
 270 {
 271         struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
 272         struct task_cputime sum;
 273         unsigned long flags;
 274
 275         spin_lock_irqsave(&cputimer->lock, flags);
 276         if (!cputimer->running) {
 277                 cputimer->running = 1;
 278                 /*
 279                  * The POSIX timer interface allows for absolute time expiry
 280                  * values through the TIMER_ABSTIME flag, therefore we have
 281                  * to synchronize the timer to the clock every time we start
 282                  * it.
 283                  */
 284                 thread_group_cputime(tsk, &sum);
 285                 update_gt_cputime(&cputimer->cputime, &sum);
 286         }
 287         *times = cputimer->cputime;
 288         spin_unlock_irqrestore(&cputimer->lock, flags);
 289 }
 290
 291 /*
 292  * Sample a process (thread group) clock for the given group_leader task.
 293  * Must be called with tasklist_lock held for reading.
 294  */
 295 static int cpu_clock_sample_group(const clockid_t which_clock,
 296                                   struct task_struct *p,
 297                                   union cpu_time_count *cpu)
 298 {
 299         struct task_cputime cputime;
 300
 301         switch (CPUCLOCK_WHICH(which_clock)) {
 302         default:
 303                 return -EINVAL;
 304         case CPUCLOCK_PROF:
 305                 thread_group_cputime(p, &cputime);
 306                 cpu->cpu = cputime_add(cputime.utime, cputime.stime);
 307                 break;
 308         case CPUCLOCK_VIRT:
 309                 thread_group_cputime(p, &cputime);
 310                 cpu->cpu = cputime.utime;
 311                 break;
 312         case CPUCLOCK_SCHED:
 313                 cpu->sched = thread_group_sched_runtime(p);
 314                 break;
 315         }
 316         return 0;
 317 }
 318
 319
 320 int posix_cpu_clock_get(const clockid_t which_clock, struct timespec *tp)
 321 {
 322         const pid_t pid = CPUCLOCK_PID(which_clock);
 323         int error = -EINVAL;
 324         union cpu_time_count rtn;
 325
 326         if (pid == 0) {
 327                 /*
 328                  * Special case constant value for our own clocks.
 329                  * We don't have to do any lookup to find ourselves.
 330                  */
 331                 if (CPUCLOCK_PERTHREAD(which_clock)) {
 332                         /*
 333                          * Sampling just ourselves we can do with no locking.
 334                          */
 335                         error = cpu_clock_sample(which_clock,
 336                                                  current, &rtn);
 337                 } else {
 338                         read_lock(&tasklist_lock);
 339                         error = cpu_clock_sample_group(which_clock,
 340                                                        current, &rtn);
 341                         read_unlock(&tasklist_lock);
 342                 }
 343         } else {
 344                 /*
 345                  * Find the given PID, and validate that the caller
 346                  * should be able to see it.
 347                  */
 348                 struct task_struct *p;
 349                 rcu_read_lock();
 350                 p = find_task_by_vpid(pid);
 351                 if (p) {
 352                         if (CPUCLOCK_PERTHREAD(which_clock)) {
 353                                 if (same_thread_group(p, current)) {
 354                                         error = cpu_clock_sample(which_clock,
 355                                                                  p, &rtn);
 356                                 }
 357                         } else {
 358                                 read_lock(&tasklist_lock);
 359                                 if (thread_group_leader(p) && p->sighand) {
 360                                         error =
 361                                             cpu_clock_sample_group(which_clock,
 362                                                                    p, &rtn);
 363                                 }
 364                                 read_unlock(&tasklist_lock);
 365                         }
 366                 }
 367                 rcu_read_unlock();
 368         }
 369
 370         if (error)
 371                 return error;
 372         sample_to_timespec(which_clock, rtn, tp);
 373         return 0;
 374 }
 375
 376
 377 /*
 378  * Validate the clockid_t for a new CPU-clock timer, and initialize the timer.
 379  * This is called from sys_timer_create() and do_cpu_nanosleep() with the
 380  * new timer already all-zeros initialized.
 381  */
 382 int posix_cpu_timer_create(struct k_itimer *new_timer)
 383 {
 384         int ret = 0;
 385         const pid_t pid = CPUCLOCK_PID(new_timer->it_clock);
 386         struct task_struct *p;
 387
 388         if (CPUCLOCK_WHICH(new_timer->it_clock) >= CPUCLOCK_MAX)
 389                 return -EINVAL;
 390
 391         INIT_LIST_HEAD(&new_timer->it.cpu.entry);
 392
 393         rcu_read_lock();
 394         if (CPUCLOCK_PERTHREAD(new_timer->it_clock)) {
 395                 if (pid == 0) {
 396                         p = current;
 397                 } else {
 398                         p = find_task_by_vpid(pid);
 399                         if (p && !same_thread_group(p, current))
 400                                 p = NULL;
 401                 }
 402         } else {
 403                 if (pid == 0) {
 404                         p = current->group_leader;
 405                 } else {
 406                         p = find_task_by_vpid(pid);
 407                         if (p && !has_group_leader_pid(p))
 408                                 p = NULL;
 409                 }
 410         }
 411         new_timer->it.cpu.task = p;
 412         if (p) {
 413                 get_task_struct(p);
 414         } else {
 415                 ret = -EINVAL;
 416         }
 417         rcu_read_unlock();
 418
 419         return ret;
 420 }
 421
 422 /*
 423  * Clean up a CPU-clock timer that is about to be destroyed.
 424  * This is called from timer deletion with the timer already locked.
 425  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 426  * and try again.  (This happens when the timer is in the middle of firing.)
 427  */
 428 int posix_cpu_timer_del(struct k_itimer *timer)
 429 {
 430         struct task_struct *p = timer->it.cpu.task;
 431         int ret = 0;
 432
 433         if (likely(p != NULL)) {
 434                 read_lock(&tasklist_lock);
 435                 if (unlikely(p->sighand == NULL)) {
 436                         /*
 437                          * We raced with the reaping of the task.
 438                          * The deletion should have cleared us off the list.
 439                          */
 440                         BUG_ON(!list_empty(&timer->it.cpu.entry));
 441                 } else {
 442                         spin_lock(&p->sighand->siglock);
 443                         if (timer->it.cpu.firing)
 444                                 ret = TIMER_RETRY;
 445                         else
 446                                 list_del(&timer->it.cpu.entry);
 447                         spin_unlock(&p->sighand->siglock);
 448                 }
 449                 read_unlock(&tasklist_lock);
 450
 451                 if (!ret)
 452                         put_task_struct(p);
 453         }
 454
 455         return ret;
 456 }
 457
 458 /*
 459  * Clean out CPU timers still ticking when a thread exited.  The task
 460  * pointer is cleared, and the expiry time is replaced with the residual
 461  * time for later timer_gettime calls to return.
 462  * This must be called with the siglock held.
 463  */
 464 static void cleanup_timers(struct list_head *head,
 465                            cputime_t utime, cputime_t stime,
 466                            unsigned long long sum_exec_runtime)
 467 {
 468         struct cpu_timer_list *timer, *next;
 469         cputime_t ptime = cputime_add(utime, stime);
 470
 471         list_for_each_entry_safe(timer, next, head, entry) {
 472                 list_del_init(&timer->entry);
 473                 if (cputime_lt(timer->expires.cpu, ptime)) {
 474                         timer->expires.cpu = cputime_zero;
 475                 } else {
 476                         timer->expires.cpu = cputime_sub(timer->expires.cpu,
 477                                                          ptime);
 478                 }
 479         }
 480
 481         ++head;
 482         list_for_each_entry_safe(timer, next, head, entry) {
 483                 list_del_init(&timer->entry);
 484                 if (cputime_lt(timer->expires.cpu, utime)) {
 485                         timer->expires.cpu = cputime_zero;
 486                 } else {
 487                         timer->expires.cpu = cputime_sub(timer->expires.cpu,
 488                                                          utime);
 489                 }
 490         }
 491
 492         ++head;
 493         list_for_each_entry_safe(timer, next, head, entry) {
 494                 list_del_init(&timer->entry);
 495                 if (timer->expires.sched < sum_exec_runtime) {
 496                         timer->expires.sched = 0;
 497                 } else {
 498                         timer->expires.sched -= sum_exec_runtime;
 499                 }
 500         }
 501 }
 502
 503 /*
 504  * These are both called with the siglock held, when the current thread
 505  * is being reaped.  When the final (leader) thread in the group is reaped,
 506  * posix_cpu_timers_exit_group will be called after posix_cpu_timers_exit.
 507  */
 508 void posix_cpu_timers_exit(struct task_struct *tsk)
 509 {
 510         cleanup_timers(tsk->cpu_timers,
 511                        tsk->utime, tsk->stime, tsk->se.sum_exec_runtime);
 512
 513 }
 514 void posix_cpu_timers_exit_group(struct task_struct *tsk)
 515 {
 516         struct signal_struct *const sig = tsk->signal;
 517
 518         cleanup_timers(tsk->signal->cpu_timers,
 519                        cputime_add(tsk->utime, sig->utime),
 520                        cputime_add(tsk->stime, sig->stime),
 521                        tsk->se.sum_exec_runtime + sig->sum_sched_runtime);
 522 }
 523
 524 static void clear_dead_task(struct k_itimer *timer, union cpu_time_count now)
 525 {
 526         /*
 527          * That's all for this thread or process.
 528          * We leave our residual in expires to be reported.
 529          */
 530         put_task_struct(timer->it.cpu.task);
 531         timer->it.cpu.task = NULL;
 532         timer->it.cpu.expires = cpu_time_sub(timer->it_clock,
 533                                              timer->it.cpu.expires,
 534                                              now);
 535 }
 536
 537 static inline int expires_gt(cputime_t expires, cputime_t new_exp)
 538 {
 539         return cputime_eq(expires, cputime_zero) ||
 540                cputime_gt(expires, new_exp);
 541 }
 542
 543 /*
 544  * Insert the timer on the appropriate list before any timers that
 545  * expire later.  This must be called with the tasklist_lock held
 546  * for reading, interrupts disabled and p->sighand->siglock taken.
 547  */
 548 static void arm_timer(struct k_itimer *timer)
 549 {
 550         struct task_struct *p = timer->it.cpu.task;
 551         struct list_head *head, *listpos;
 552         struct task_cputime *cputime_expires;
 553         struct cpu_timer_list *const nt = &timer->it.cpu;
 554         struct cpu_timer_list *next;
 555
 556         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 557                 head = p->cpu_timers;
 558                 cputime_expires = &p->cputime_expires;
 559         } else {
 560                 head = p->signal->cpu_timers;
 561                 cputime_expires = &p->signal->cputime_expires;
 562         }
 563         head += CPUCLOCK_WHICH(timer->it_clock);
 564
 565         listpos = head;
 566         list_for_each_entry(next, head, entry) {
 567                 if (cpu_time_before(timer->it_clock, nt->expires, next->expires))
 568                         break;
 569                 listpos = &next->entry;
 570         }
 571         list_add(&nt->entry, listpos);
 572
 573         if (listpos == head) {
 574                 union cpu_time_count *exp = &nt->expires;
 575
 576                 /*
 577                  * We are the new earliest-expiring POSIX 1.b timer, hence
 578                  * need to update expiration cache. Take into account that
 579                  * for process timers we share expiration cache with itimers
 580                  * and RLIMIT_CPU and for thread timers with RLIMIT_RTTIME.
 581                  */
 582
 583                 switch (CPUCLOCK_WHICH(timer->it_clock)) {
 584                 case CPUCLOCK_PROF:
 585                         if (expires_gt(cputime_expires->prof_exp, exp->cpu))
 586                                 cputime_expires->prof_exp = exp->cpu;
 587                         break;
 588                 case CPUCLOCK_VIRT:
 589                         if (expires_gt(cputime_expires->virt_exp, exp->cpu))
 590                                 cputime_expires->virt_exp = exp->cpu;
 591                         break;
 592                 case CPUCLOCK_SCHED:
 593                         if (cputime_expires->sched_exp == 0 ||
 594                             cputime_expires->sched_exp > exp->sched)
 595                                 cputime_expires->sched_exp = exp->sched;
 596                         break;
 597                 }
 598         }
 599 }
 600
 601 /*
 602  * The timer is locked, fire it and arrange for its reload.
 603  */
 604 static void cpu_timer_fire(struct k_itimer *timer)
 605 {
 606         if ((timer->it_sigev_notify & ~SIGEV_THREAD_ID) == SIGEV_NONE) {
 607                 /*
 608                  * User don't want any signal.
 609                  */
 610                 timer->it.cpu.expires.sched = 0;
 611         } else if (unlikely(timer->sigq == NULL)) {
 612                 /*
 613                  * This a special case for clock_nanosleep,
 614                  * not a normal timer from sys_timer_create.
 615                  */
 616                 wake_up_process(timer->it_process);
 617                 timer->it.cpu.expires.sched = 0;
 618         } else if (timer->it.cpu.incr.sched == 0) {
 619                 /*
 620                  * One-shot timer.  Clear it as soon as it's fired.
 621                  */
 622                 posix_timer_event(timer, 0);
 623                 timer->it.cpu.expires.sched = 0;
 624         } else if (posix_timer_event(timer, ++timer->it_requeue_pending)) {
 625                 /*
 626                  * The signal did not get queued because the signal
 627                  * was ignored, so we won't get any callback to
 628                  * reload the timer.  But we need to keep it
 629                  * ticking in case the signal is deliverable next time.
 630                  */
 631                 posix_cpu_timer_schedule(timer);
 632         }
 633 }
 634
 635 /*
 636  * Sample a process (thread group) timer for the given group_leader task.
 637  * Must be called with tasklist_lock held for reading.
 638  */
 639 static int cpu_timer_sample_group(const clockid_t which_clock,
 640                                   struct task_struct *p,
 641                                   union cpu_time_count *cpu)
 642 {
 643         struct task_cputime cputime;
 644
 645         thread_group_cputimer(p, &cputime);
 646         switch (CPUCLOCK_WHICH(which_clock)) {
 647         default:
 648                 return -EINVAL;
 649         case CPUCLOCK_PROF:
 650                 cpu->cpu = cputime_add(cputime.utime, cputime.stime);
 651                 break;
 652         case CPUCLOCK_VIRT:
 653                 cpu->cpu = cputime.utime;
 654                 break;
 655         case CPUCLOCK_SCHED:
 656                 cpu->sched = cputime.sum_exec_runtime + task_delta_exec(p);
 657                 break;
 658         }
 659         return 0;
 660 }
 661
 662 /*
 663  * Guts of sys_timer_settime for CPU timers.
 664  * This is called with the timer locked and interrupts disabled.
 665  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 666  * and try again.  (This happens when the timer is in the middle of firing.)
 667  */
 668 int posix_cpu_timer_set(struct k_itimer *timer, int flags,
 669                         struct itimerspec *new, struct itimerspec *old)
 670 {
 671         struct task_struct *p = timer->it.cpu.task;
 672         union cpu_time_count old_expires, new_expires, old_incr, val;
 673         int ret;
 674
 675         if (unlikely(p == NULL)) {
 676                 /*
 677                  * Timer refers to a dead task's clock.
 678                  */
 679                 return -ESRCH;
 680         }
 681
 682         new_expires = timespec_to_sample(timer->it_clock, &new->it_value);
 683
 684         read_lock(&tasklist_lock);
 685         /*
 686          * We need the tasklist_lock to protect against reaping that
 687          * clears p->sighand.  If p has just been reaped, we can no
 688          * longer get any information about it at all.
 689          */
 690         if (unlikely(p->sighand == NULL)) {
 691                 read_unlock(&tasklist_lock);
 692                 put_task_struct(p);
 693                 timer->it.cpu.task = NULL;
 694                 return -ESRCH;
 695         }
 696
 697         /*
 698          * Disarm any old timer after extracting its expiry time.
 699          */
 700         BUG_ON(!irqs_disabled());
 701
 702         ret = 0;
 703         old_incr = timer->it.cpu.incr;
 704         spin_lock(&p->sighand->siglock);
 705         old_expires = timer->it.cpu.expires;
 706         if (unlikely(timer->it.cpu.firing)) {
 707                 timer->it.cpu.firing = -1;
 708                 ret = TIMER_RETRY;
 709         } else
 710                 list_del_init(&timer->it.cpu.entry);
 711
 712         /*
 713          * We need to sample the current value to convert the new
 714          * value from to relative and absolute, and to convert the
 715          * old value from absolute to relative.  To set a process
 716          * timer, we need a sample to balance the thread expiry
 717          * times (in arm_timer).  With an absolute time, we must
 718          * check if it's already passed.  In short, we need a sample.
 719          */
 720         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 721                 cpu_clock_sample(timer->it_clock, p, &val);
 722         } else {
 723                 cpu_timer_sample_group(timer->it_clock, p, &val);
 724         }
 725
 726         if (old) {
 727                 if (old_expires.sched == 0) {
 728                         old->it_value.tv_sec = 0;
 729                         old->it_value.tv_nsec = 0;
 730                 } else {
 731                         /*
 732                          * Update the timer in case it has
 733                          * overrun already.  If it has,
 734                          * we'll report it as having overrun
 735                          * and with the next reloaded timer
 736                          * already ticking, though we are
 737                          * swallowing that pending
 738                          * notification here to install the
 739                          * new setting.
 740                          */
 741                         bump_cpu_timer(timer, val);
 742                         if (cpu_time_before(timer->it_clock, val,
 743                                             timer->it.cpu.expires)) {
 744                                 old_expires = cpu_time_sub(
 745                                         timer->it_clock,
 746                                         timer->it.cpu.expires, val);
 747                                 sample_to_timespec(timer->it_clock,
 748                                                    old_expires,
 749                                                    &old->it_value);
 750                         } else {
 751                                 old->it_value.tv_nsec = 1;
 752                                 old->it_value.tv_sec = 0;
 753                         }
 754                 }
 755         }
 756
 757         if (unlikely(ret)) {
 758                 /*
 759                  * We are colliding with the timer actually firing.
 760                  * Punt after filling in the timer's old value, and
 761                  * disable this firing since we are already reporting
 762                  * it as an overrun (thanks to bump_cpu_timer above).
 763                  */
 764                 spin_unlock(&p->sighand->siglock);
 765                 read_unlock(&tasklist_lock);
 766                 goto out;
 767         }
 768
 769         if (new_expires.sched != 0 && !(flags & TIMER_ABSTIME)) {
 770                 cpu_time_add(timer->it_clock, &new_expires, val);
 771         }
 772
 773         /*
 774          * Install the new expiry time (or zero).
 775          * For a timer with no notification action, we don't actually
 776          * arm the timer (we'll just fake it for timer_gettime).
 777          */
 778         timer->it.cpu.expires = new_expires;
 779         if (new_expires.sched != 0 &&
 780             cpu_time_before(timer->it_clock, val, new_expires)) {
 781                 arm_timer(timer);
 782         }
 783
 784         spin_unlock(&p->sighand->siglock);
 785         read_unlock(&tasklist_lock);
 786
 787         /*
 788          * Install the new reload setting, and
 789          * set up the signal and overrun bookkeeping.
 790          */
 791         timer->it.cpu.incr = timespec_to_sample(timer->it_clock,
 792                                                 &new->it_interval);
 793
 794         /*
 795          * This acts as a modification timestamp for the timer,
 796          * so any automatic reload attempt will punt on seeing
 797          * that we have reset the timer manually.
 798          */
 799         timer->it_requeue_pending = (timer->it_requeue_pending + 2) &
 800                 ~REQUEUE_PENDING;
 801         timer->it_overrun_last = 0;
 802         timer->it_overrun = -1;
 803
 804         if (new_expires.sched != 0 &&
 805             !cpu_time_before(timer->it_clock, val, new_expires)) {
 806                 /*
 807                  * The designated time already passed, so we notify
 808                  * immediately, even if the thread never runs to
 809                  * accumulate more time on this clock.
 810                  */
 811                 cpu_timer_fire(timer);
 812         }
 813
 814         ret = 0;
 815  out:
 816         if (old) {
 817                 sample_to_timespec(timer->it_clock,
 818                                    old_incr, &old->it_interval);
 819         }
 820         return ret;
 821 }
 822
 823 void posix_cpu_timer_get(struct k_itimer *timer, struct itimerspec *itp)
 824 {
 825         union cpu_time_count now;
 826         struct task_struct *p = timer->it.cpu.task;
 827         int clear_dead;
 828
 829         /*
 830          * Easy part: convert the reload time.
 831          */
 832         sample_to_timespec(timer->it_clock,
 833                            timer->it.cpu.incr, &itp->it_interval);
 834
 835         if (timer->it.cpu.expires.sched == 0) { /* Timer not armed at all.  */
 836                 itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
 837                 return;
 838         }
 839
 840         if (unlikely(p == NULL)) {
 841                 /*
 842                  * This task already died and the timer will never fire.
 843                  * In this case, expires is actually the dead value.
 844                  */
 845         dead:
 846                 sample_to_timespec(timer->it_clock, timer->it.cpu.expires,
 847                                    &itp->it_value);
 848                 return;
 849         }
 850
 851         /*
 852          * Sample the clock to take the difference with the expiry time.
 853          */
 854         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 855                 cpu_clock_sample(timer->it_clock, p, &now);
 856                 clear_dead = p->exit_state;
 857         } else {
 858                 read_lock(&tasklist_lock);
 859                 if (unlikely(p->sighand == NULL)) {
 860                         /*
 861                          * The process has been reaped.
 862                          * We can't even collect a sample any more.
 863                          * Call the timer disarmed, nothing else to do.
 864                          */
 865                         put_task_struct(p);
 866                         timer->it.cpu.task = NULL;
 867                         timer->it.cpu.expires.sched = 0;
 868                         read_unlock(&tasklist_lock);
 869                         goto dead;
 870                 } else {
 871                         cpu_timer_sample_group(timer->it_clock, p, &now);
 872                         clear_dead = (unlikely(p->exit_state) &&
 873                                       thread_group_empty(p));
 874                 }
 875                 read_unlock(&tasklist_lock);
 876         }
 877
 878         if (unlikely(clear_dead)) {
 879                 /*
 880                  * We've noticed that the thread is dead, but
 881                  * not yet reaped.  Take this opportunity to
 882                  * drop our task ref.
 883                  */
 884                 clear_dead_task(timer, now);
 885                 goto dead;
 886         }
 887
 888         if (cpu_time_before(timer->it_clock, now, timer->it.cpu.expires)) {
 889                 sample_to_timespec(timer->it_clock,
 890                                    cpu_time_sub(timer->it_clock,
 891                                                 timer->it.cpu.expires, now),
 892                                    &itp->it_value);
 893         } else {
 894                 /*
 895                  * The timer should have expired already, but the firing
 896                  * hasn't taken place yet.  Say it's just about to expire.
 897                  */
 898                 itp->it_value.tv_nsec = 1;
 899                 itp->it_value.tv_sec = 0;
 900         }
 901 }
 902
 903 /*
 904  * Check for any per-thread CPU timers that have fired and move them off
 905  * the tsk->cpu_timers[N] list onto the firing list.  Here we update the
 906  * tsk->it_*_expires values to reflect the remaining thread CPU timers.
 907  */
 908 static void check_thread_timers(struct task_struct *tsk,
 909                                 struct list_head *firing)
 910 {
 911         int maxfire;
 912         struct list_head *timers = tsk->cpu_timers;
 913         struct signal_struct *const sig = tsk->signal;
 914         unsigned long soft;
 915
 916         maxfire = 20;
 917         tsk->cputime_expires.prof_exp = cputime_zero;
 918         while (!list_empty(timers)) {
 919                 struct cpu_timer_list *t = list_first_entry(timers,
 920                                                       struct cpu_timer_list,
 921                                                       entry);
 922                 if (!--maxfire || cputime_lt(prof_ticks(tsk), t->expires.cpu)) {
 923                         tsk->cputime_expires.prof_exp = t->expires.cpu;
 924                         break;
 925                 }
 926                 t->firing = 1;
 927                 list_move_tail(&t->entry, firing);
 928         }
 929
 930         ++timers;
 931         maxfire = 20;
 932         tsk->cputime_expires.virt_exp = cputime_zero;
 933         while (!list_empty(timers)) {
 934                 struct cpu_timer_list *t = list_first_entry(timers,
 935                                                       struct cpu_timer_list,
 936                                                       entry);
 937                 if (!--maxfire || cputime_lt(virt_ticks(tsk), t->expires.cpu)) {
 938                         tsk->cputime_expires.virt_exp = t->expires.cpu;
 939                         break;
 940                 }
 941                 t->firing = 1;
 942                 list_move_tail(&t->entry, firing);
 943         }
 944
 945         ++timers;
 946         maxfire = 20;
 947         tsk->cputime_expires.sched_exp = 0;
 948         while (!list_empty(timers)) {
 949                 struct cpu_timer_list *t = list_first_entry(timers,
 950                                                       struct cpu_timer_list,
 951                                                       entry);
 952                 if (!--maxfire || tsk->se.sum_exec_runtime < t->expires.sched) {
 953                         tsk->cputime_expires.sched_exp = t->expires.sched;
 954                         break;
 955                 }
 956                 t->firing = 1;
 957                 list_move_tail(&t->entry, firing);
 958         }
 959
 960         /*
 961          * Check for the special case thread timers.
 962          */
 963         soft = ACCESS_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_cur);
 964         if (soft != RLIM_INFINITY) {
 965                 unsigned long hard =
 966                         ACCESS_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_max);
 967
 968                 if (hard != RLIM_INFINITY &&
 969                     tsk->rt.timeout > DIV_ROUND_UP(hard, USEC_PER_SEC/HZ)) {
 970                         /*
 971                          * At the hard limit, we just die.
 972                          * No need to calculate anything else now.
 973                          */
 974                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
 975                         return;
 976                 }
 977                 if (tsk->rt.timeout > DIV_ROUND_UP(soft, USEC_PER_SEC/HZ)) {
 978                         /*
 979                          * At the soft limit, send a SIGXCPU every second.
 980                          */
 981                         if (soft < hard) {
 982                                 soft += USEC_PER_SEC;
 983                                 sig->rlim[RLIMIT_RTTIME].rlim_cur = soft;
 984                         }
 985                         printk(KERN_INFO
 986                                 "RT Watchdog Timeout: %s[%d]\n",
 987                                 tsk->comm, task_pid_nr(tsk));
 988                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
 989                 }
 990         }
 991 }
 992
 993 static void stop_process_timers(struct signal_struct *sig)
 994 {
 995         struct thread_group_cputimer *cputimer = &sig->cputimer;
 996         unsigned long flags;
 997
 998         spin_lock_irqsave(&cputimer->lock, flags);
 999         cputimer->running = 0;
1000         spin_unlock_irqrestore(&cputimer->lock, flags);
1001 }
1002
1003 static u32 onecputick;
1004
1005 static void check_cpu_itimer(struct task_struct *tsk, struct cpu_itimer *it,
1006                              cputime_t *expires, cputime_t cur_time, int signo)
1007 {
1008         if (cputime_eq(it->expires, cputime_zero))
1009                 return;
1010
1011         if (cputime_ge(cur_time, it->expires)) {
1012                 if (!cputime_eq(it->incr, cputime_zero)) {
1013                         it->expires = cputime_add(it->expires, it->incr);
1014                         it->error += it->incr_error;
1015                         if (it->error >= onecputick) {
1016                                 it->expires = cputime_sub(it->expires,
1017                                                           cputime_one_jiffy);
1018                                 it->error -= onecputick;
1019                         }
1020                 } else {
1021                         it->expires = cputime_zero;
1022                 }
1023
1024                 trace_itimer_expire(signo == SIGPROF ?
1025                                     ITIMER_PROF : ITIMER_VIRTUAL,
1026                                     tsk->signal->leader_pid, cur_time);
1027                 __group_send_sig_info(signo, SEND_SIG_PRIV, tsk);
1028         }
1029
1030         if (!cputime_eq(it->expires, cputime_zero) &&
1031             (cputime_eq(*expires, cputime_zero) ||
1032              cputime_lt(it->expires, *expires))) {
1033                 *expires = it->expires;
1034         }
1035 }
1036
1037 /**
1038  * task_cputime_zero - Check a task_cputime struct for all zero fields.
1039  *
1040  * @cputime:    The struct to compare.
1041  *
1042  * Checks @cputime to see if all fields are zero.  Returns true if all fields
1043  * are zero, false if any field is nonzero.
1044  */
1045 static inline int task_cputime_zero(const struct task_cputime *cputime)
1046 {
1047         if (cputime_eq(cputime->utime, cputime_zero) &&
1048             cputime_eq(cputime->stime, cputime_zero) &&
1049             cputime->sum_exec_runtime == 0)
1050                 return 1;
1051         return 0;
1052 }
1053
1054 /*
1055  * Check for any per-thread CPU timers that have fired and move them
1056  * off the tsk->*_timers list onto the firing list.  Per-thread timers
1057  * have already been taken off.
1058  */
1059 static void check_process_timers(struct task_struct *tsk,
1060                                  struct list_head *firing)
1061 {
1062         int maxfire;
1063         struct signal_struct *const sig = tsk->signal;
1064         cputime_t utime, ptime, virt_expires, prof_expires;
1065         unsigned long long sum_sched_runtime, sched_expires;
1066         struct list_head *timers = sig->cpu_timers;
1067         struct task_cputime cputime;
1068         unsigned long soft;
1069
1070         /*
1071          * Collect the current process totals.
1072          */
1073         thread_group_cputimer(tsk, &cputime);
1074         utime = cputime.utime;
1075         ptime = cputime_add(utime, cputime.stime);
1076         sum_sched_runtime = cputime.sum_exec_runtime;
1077         maxfire = 20;
1078         prof_expires = cputime_zero;
1079         while (!list_empty(timers)) {
1080                 struct cpu_timer_list *tl = list_first_entry(timers,
1081                                                       struct cpu_timer_list,
1082                                                       entry);
1083                 if (!--maxfire || cputime_lt(ptime, tl->expires.cpu)) {
1084                         prof_expires = tl->expires.cpu;
1085                         break;
1086                 }
1087                 tl->firing = 1;
1088                 list_move_tail(&tl->entry, firing);
1089         }
1090
1091         ++timers;
1092         maxfire = 20;
1093         virt_expires = cputime_zero;
1094         while (!list_empty(timers)) {
1095                 struct cpu_timer_list *tl = list_first_entry(timers,
1096                                                       struct cpu_timer_list,
1097                                                       entry);
1098                 if (!--maxfire || cputime_lt(utime, tl->expires.cpu)) {
1099                         virt_expires = tl->expires.cpu;
1100                         break;
1101                 }
1102                 tl->firing = 1;
1103                 list_move_tail(&tl->entry, firing);
1104         }
1105
1106         ++timers;
1107         maxfire = 20;
1108         sched_expires = 0;
1109         while (!list_empty(timers)) {
1110                 struct cpu_timer_list *tl = list_first_entry(timers,
1111                                                       struct cpu_timer_list,
1112                                                       entry);
1113                 if (!--maxfire || sum_sched_runtime < tl->expires.sched) {
1114                         sched_expires = tl->expires.sched;
1115                         break;
1116                 }
1117                 tl->firing = 1;
1118                 list_move_tail(&tl->entry, firing);
1119         }
1120
1121         /*
1122          * Check for the special case process timers.
1123          */
1124         check_cpu_itimer(tsk, &sig->it[CPUCLOCK_PROF], &prof_expires, ptime,
1125                          SIGPROF);
1126         check_cpu_itimer(tsk, &sig->it[CPUCLOCK_VIRT], &virt_expires, utime,
1127                          SIGVTALRM);
1128         soft = ACCESS_ONCE(sig->rlim[RLIMIT_CPU].rlim_cur);
1129         if (soft != RLIM_INFINITY) {
1130                 unsigned long psecs = cputime_to_secs(ptime);
1131                 unsigned long hard =
1132                         ACCESS_ONCE(sig->rlim[RLIMIT_CPU].rlim_max);
1133                 cputime_t x;
1134                 if (psecs >= hard) {
1135                         /*
1136                          * At the hard limit, we just die.
1137                          * No need to calculate anything else now.
1138                          */
1139                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
1140                         return;
1141                 }
1142                 if (psecs >= soft) {
1143                         /*
1144                          * At the soft limit, send a SIGXCPU every second.
1145                          */
1146                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
1147                         if (soft < hard) {
1148                                 soft++;
1149                                 sig->rlim[RLIMIT_CPU].rlim_cur = soft;
1150                         }
1151                 }
1152                 x = secs_to_cputime(soft);
1153                 if (cputime_eq(prof_expires, cputime_zero) ||
1154                     cputime_lt(x, prof_expires)) {
1155                         prof_expires = x;
1156                 }
1157         }
1158
1159         sig->cputime_expires.prof_exp = prof_expires;
1160         sig->cputime_expires.virt_exp = virt_expires;
1161         sig->cputime_expires.sched_exp = sched_expires;
1162         if (task_cputime_zero(&sig->cputime_expires))
1163                 stop_process_timers(sig);
1164 }
1165
1166 /*
1167  * This is called from the signal code (via do_schedule_next_timer)
1168  * when the last timer signal was delivered and we have to reload the timer.
1169  */
1170 void posix_cpu_timer_schedule(struct k_itimer *timer)
1171 {
1172         struct task_struct *p = timer->it.cpu.task;
1173         union cpu_time_count now;
1174
1175         if (unlikely(p == NULL))
1176                 /*
1177                  * The task was cleaned up already, no future firings.
1178                  */
1179                 goto out;
1180
1181         /*
1182          * Fetch the current sample and update the timer's expiry time.
1183          */
1184         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
1185                 cpu_clock_sample(timer->it_clock, p, &now);
1186                 bump_cpu_timer(timer, now);
1187                 if (unlikely(p->exit_state)) {
1188                         clear_dead_task(timer, now);
1189                         goto out;
1190                 }
1191                 read_lock(&tasklist_lock); /* arm_timer needs it.  */
1192                 spin_lock(&p->sighand->siglock);
1193         } else {
1194                 read_lock(&tasklist_lock);
1195                 if (unlikely(p->sighand == NULL)) {
1196                         /*
1197                          * The process has been reaped.
1198                          * We can't even collect a sample any more.
1199                          */
1200                         put_task_struct(p);
1201                         timer->it.cpu.task = p = NULL;
1202                         timer->it.cpu.expires.sched = 0;
1203                         goto out_unlock;
1204                 } else if (unlikely(p->exit_state) && thread_group_empty(p)) {
1205                         /*
1206                          * We've noticed that the thread is dead, but
1207                          * not yet reaped.  Take this opportunity to
1208                          * drop our task ref.
1209                          */
1210                         clear_dead_task(timer, now);
1211                         goto out_unlock;
1212                 }
1213                 spin_lock(&p->sighand->siglock);
1214                 cpu_timer_sample_group(timer->it_clock, p, &now);
1215                 bump_cpu_timer(timer, now);
1216                 /* Leave the tasklist_lock locked for the call below.  */
1217         }
1218
1219         /*
1220          * Now re-arm for the new expiry time.
1221          */
1222         BUG_ON(!irqs_disabled());
1223         arm_timer(timer);
1224         spin_unlock(&p->sighand->siglock);
1225
1226 out_unlock:
1227         read_unlock(&tasklist_lock);
1228
1229 out:
1230         timer->it_overrun_last = timer->it_overrun;
1231         timer->it_overrun = -1;
1232         ++timer->it_requeue_pending;
1233 }
1234
1235 /**
1236  * task_cputime_expired - Compare two task_cputime entities.
1237  *
1238  * @sample:     The task_cputime structure to be checked for expiration.
1239  * @expires:    Expiration times, against which @sample will be checked.
1240  *
1241  * Checks @sample against @expires to see if any field of @sample has expired.
1242  * Returns true if any field of the former is greater than the corresponding
1243  * field of the latter if the latter field is set.  Otherwise returns false.
1244  */
1245 static inline int task_cputime_expired(const struct task_cputime *sample,
1246                                         const struct task_cputime *expires)
1247 {
1248         if (!cputime_eq(expires->utime, cputime_zero) &&
1249             cputime_ge(sample->utime, expires->utime))
1250                 return 1;
1251         if (!cputime_eq(expires->stime, cputime_zero) &&
1252             cputime_ge(cputime_add(sample->utime, sample->stime),
1253                        expires->stime))
1254                 return 1;
1255         if (expires->sum_exec_runtime != 0 &&
1256             sample->sum_exec_runtime >= expires->sum_exec_runtime)
1257                 return 1;
1258         return 0;
1259 }
1260
1261 /**
1262  * fastpath_timer_check - POSIX CPU timers fast path.
1263  *
1264  * @tsk:        The task (thread) being checked.
1265  *
1266  * Check the task and thread group timers.  If both are zero (there are no
1267  * timers set) return false.  Otherwise snapshot the task and thread group
1268  * timers and compare them with the corresponding expiration times.  Return
1269  * true if a timer has expired, else return false.
1270  */
1271 static inline int fastpath_timer_check(struct task_struct *tsk)
1272 {
1273         struct signal_struct *sig;
1274
1275         if (!task_cputime_zero(&tsk->cputime_expires)) {
1276                 struct task_cputime task_sample = {
1277                         .utime = tsk->utime,
1278                         .stime = tsk->stime,
1279                         .sum_exec_runtime = tsk->se.sum_exec_runtime
1280                 };
1281
1282                 if (task_cputime_expired(&task_sample, &tsk->cputime_expires))
1283                         return 1;
1284         }
1285
1286         sig = tsk->signal;
1287         if (sig->cputimer.running) {
1288                 struct task_cputime group_sample;
1289
1290                 spin_lock(&sig->cputimer.lock);
1291                 group_sample = sig->cputimer.cputime;
1292                 spin_unlock(&sig->cputimer.lock);
1293
1294                 if (task_cputime_expired(&group_sample, &sig->cputime_expires))
1295                         return 1;
1296         }
1297
1298         return 0;
1299 }
1300
1301 /*
1302  * This is called from the timer interrupt handler.  The irq handler has
1303  * already updated our counts.  We need to check if any timers fire now.
1304  * Interrupts are disabled.
1305  */
1306 void run_posix_cpu_timers(struct task_struct *tsk)
1307 {
1308         LIST_HEAD(firing);
1309         struct k_itimer *timer, *next;
1310         unsigned long flags;
1311
1312         BUG_ON(!irqs_disabled());
1313
1314         /*
1315          * The fast path checks that there are no expired thread or thread
1316          * group timers.  If that's so, just return.
1317          */
1318         if (!fastpath_timer_check(tsk))
1319                 return;
1320
1321         if (!lock_task_sighand(tsk, &flags))
1322                 return;
1323         /*
1324          * Here we take off tsk->signal->cpu_timers[N] and
1325          * tsk->cpu_timers[N] all the timers that are firing, and
1326          * put them on the firing list.
1327          */
1328         check_thread_timers(tsk, &firing);
1329         /*
1330          * If there are any active process wide timers (POSIX 1.b, itimers,
1331          * RLIMIT_CPU) cputimer must be running.
1332          */
1333         if (tsk->signal->cputimer.running)
1334                 check_process_timers(tsk, &firing);
1335
1336         /*
1337          * We must release these locks before taking any timer's lock.
1338          * There is a potential race with timer deletion here, as the
1339          * siglock now protects our private firing list.  We have set
1340          * the firing flag in each timer, so that a deletion attempt
1341          * that gets the timer lock before we do will give it up and
1342          * spin until we've taken care of that timer below.
1343          */
1344         unlock_task_sighand(tsk, &flags);
1345
1346         /*
1347          * Now that all the timers on our list have the firing flag,
1348          * noone will touch their list entries but us.  We'll take
1349          * each timer's lock before clearing its firing flag, so no
1350          * timer call will interfere.
1351          */
1352         list_for_each_entry_safe(timer, next, &firing, it.cpu.entry) {
1353                 int cpu_firing;
1354
1355                 spin_lock(&timer->it_lock);
1356                 list_del_init(&timer->it.cpu.entry);
1357                 cpu_firing = timer->it.cpu.firing;
1358                 timer->it.cpu.firing = 0;
1359                 /*
1360                  * The firing flag is -1 if we collided with a reset
1361                  * of the timer, which already reported this
1362                  * almost-firing as an overrun.  So don't generate an event.
1363                  */
1364                 if (likely(cpu_firing >= 0))
1365                         cpu_timer_fire(timer);
1366                 spin_unlock(&timer->it_lock);
1367         }
1368 }
1369
1370 /*
1371  * Set one of the process-wide special case CPU timers or RLIMIT_CPU.
1372  * The tsk->sighand->siglock must be held by the caller.
1373  */
1374 void set_process_cpu_timer(struct task_struct *tsk, unsigned int clock_idx,
1375                            cputime_t *newval, cputime_t *oldval)
1376 {
1377         union cpu_time_count now;
1378
1379         BUG_ON(clock_idx == CPUCLOCK_SCHED);
1380         cpu_timer_sample_group(clock_idx, tsk, &now);
1381
1382         if (oldval) {
1383                 /*
1384                  * We are setting itimer. The *oldval is absolute and we update
1385                  * it to be relative, *newval argument is relative and we update
1386                  * it to be absolute.
1387                  */
1388                 if (!cputime_eq(*oldval, cputime_zero)) {
1389                         if (cputime_le(*oldval, now.cpu)) {
1390                                 /* Just about to fire. */
1391                                 *oldval = cputime_one_jiffy;
1392                         } else {
1393                                 *oldval = cputime_sub(*oldval, now.cpu);
1394                         }
1395                 }
1396
1397                 if (cputime_eq(*newval, cputime_zero))
1398                         return;
1399                 *newval = cputime_add(*newval, now.cpu);
1400         }
1401
1402         /*
1403          * Update expiration cache if we are the earliest timer, or eventually
1404          * RLIMIT_CPU limit is earlier than prof_exp cpu timer expire.
1405          */
1406         switch (clock_idx) {
1407         case CPUCLOCK_PROF:
1408                 if (expires_gt(tsk->signal->cputime_expires.prof_exp, *newval))
1409                         tsk->signal->cputime_expires.prof_exp = *newval;
1410                 break;
1411         case CPUCLOCK_VIRT:
1412                 if (expires_gt(tsk->signal->cputime_expires.virt_exp, *newval))
1413                         tsk->signal->cputime_expires.virt_exp = *newval;
1414                 break;
1415         }
1416 }
1417
1418 static int do_cpu_nanosleep(const clockid_t which_clock, int flags,
1419                             struct timespec *rqtp, struct itimerspec *it)
1420 {
1421         struct k_itimer timer;
1422         int error;
1423
1424         /*
1425          * Set up a temporary timer and then wait for it to go off.
1426          */
1427         memset(&timer, 0, sizeof timer);
1428         spin_lock_init(&timer.it_lock);
1429         timer.it_clock = which_clock;
1430         timer.it_overrun = -1;
1431         error = posix_cpu_timer_create(&timer);
1432         timer.it_process = current;
1433         if (!error) {
1434                 static struct itimerspec zero_it;
1435
1436                 memset(it, 0, sizeof *it);
1437                 it->it_value = *rqtp;
1438
1439                 spin_lock_irq(&timer.it_lock);
1440                 error = posix_cpu_timer_set(&timer, flags, it, NULL);
1441                 if (error) {
1442                         spin_unlock_irq(&timer.it_lock);
1443                         return error;
1444                 }
1445
1446                 while (!signal_pending(current)) {
1447                         if (timer.it.cpu.expires.sched == 0) {
1448                                 /*
1449                                  * Our timer fired and was reset.
1450                                  */
1451                                 spin_unlock_irq(&timer.it_lock);
1452                                 return 0;
1453                         }
1454
1455                         /*
1456                          * Block until cpu_timer_fire (or a signal) wakes us.
1457                          */
1458                         __set_current_state(TASK_INTERRUPTIBLE);
1459                         spin_unlock_irq(&timer.it_lock);
1460                         schedule();
1461                         spin_lock_irq(&timer.it_lock);
1462                 }
1463
1464                 /*
1465                  * We were interrupted by a signal.
1466                  */
1467                 sample_to_timespec(which_clock, timer.it.cpu.expires, rqtp);
1468                 posix_cpu_timer_set(&timer, 0, &zero_it, it);
1469                 spin_unlock_irq(&timer.it_lock);
1470
1471                 if ((it->it_value.tv_sec | it->it_value.tv_nsec) == 0) {
1472                         /*
1473                          * It actually did fire already.
1474                          */
1475                         return 0;
1476                 }
1477
1478                 error = -ERESTART_RESTARTBLOCK;
1479         }
1480
1481         return error;
1482 }
1483
1484 int posix_cpu_nsleep(const clockid_t which_clock, int flags,
1485                      struct timespec *rqtp, struct timespec __user *rmtp)
1486 {
1487         struct restart_block *restart_block =
1488             &current_thread_info()->restart_block;
1489         struct itimerspec it;
1490         int error;
1491
1492         /*
1493          * Diagnose required errors first.
1494          */
1495         if (CPUCLOCK_PERTHREAD(which_clock) &&
1496             (CPUCLOCK_PID(which_clock) == 0 ||
1497              CPUCLOCK_PID(which_clock) == current->pid))
1498                 return -EINVAL;
1499
1500         error = do_cpu_nanosleep(which_clock, flags, rqtp, &it);
1501
1502         if (error == -ERESTART_RESTARTBLOCK) {
1503
1504                 if (flags & TIMER_ABSTIME)
1505                         return -ERESTARTNOHAND;
1506                 /*
1507                  * Report back to the user the time still remaining.
1508                  */
1509                 if (rmtp != NULL && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1510                         return -EFAULT;
1511
1512                 restart_block->fn = posix_cpu_nsleep_restart;
1513                 restart_block->arg0 = which_clock;
1514                 restart_block->arg1 = (unsigned long) rmtp;
1515                 restart_block->arg2 = rqtp->tv_sec;
1516                 restart_block->arg3 = rqtp->tv_nsec;
1517         }
1518         return error;
1519 }
1520
1521 long posix_cpu_nsleep_restart(struct restart_block *restart_block)
1522 {
1523         clockid_t which_clock = restart_block->arg0;
1524         struct timespec __user *rmtp;
1525         struct timespec t;
1526         struct itimerspec it;
1527         int error;
1528
1529         rmtp = (struct timespec __user *) restart_block->arg1;
1530         t.tv_sec = restart_block->arg2;
1531         t.tv_nsec = restart_block->arg3;
1532
1533         restart_block->fn = do_no_restart_syscall;
1534         error = do_cpu_nanosleep(which_clock, TIMER_ABSTIME, &t, &it);
1535
1536         if (error == -ERESTART_RESTARTBLOCK) {
1537                 /*
1538                  * Report back to the user the time still remaining.
1539                  */
1540                 if (rmtp != NULL && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1541                         return -EFAULT;
1542
1543                 restart_block->fn = posix_cpu_nsleep_restart;
1544                 restart_block->arg0 = which_clock;
1545                 restart_block->arg1 = (unsigned long) rmtp;
1546                 restart_block->arg2 = t.tv_sec;
1547                 restart_block->arg3 = t.tv_nsec;
1548         }
1549         return error;
1550
1551 }
1552
1553
1554 #define PROCESS_CLOCK   MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED)
1555 #define THREAD_CLOCK    MAKE_THREAD_CPUCLOCK(0, CPUCLOCK_SCHED)
1556
1557 static int process_cpu_clock_getres(const clockid_t which_clock,
1558                                     struct timespec *tp)
1559 {
1560         return posix_cpu_clock_getres(PROCESS_CLOCK, tp);
1561 }
1562 static int process_cpu_clock_get(const clockid_t which_clock,
1563                                  struct timespec *tp)
1564 {
1565         return posix_cpu_clock_get(PROCESS_CLOCK, tp);
1566 }
1567 static int process_cpu_timer_create(struct k_itimer *timer)
1568 {
1569         timer->it_clock = PROCESS_CLOCK;
1570         return posix_cpu_timer_create(timer);
1571 }
1572 static int process_cpu_nsleep(const clockid_t which_clock, int flags,
1573                               struct timespec *rqtp,
1574                               struct timespec __user *rmtp)
1575 {
1576         return posix_cpu_nsleep(PROCESS_CLOCK, flags, rqtp, rmtp);
1577 }
1578 static long process_cpu_nsleep_restart(struct restart_block *restart_block)
1579 {
1580         return -EINVAL;
1581 }
1582 static int thread_cpu_clock_getres(const clockid_t which_clock,
1583                                    struct timespec *tp)
1584 {
1585         return posix_cpu_clock_getres(THREAD_CLOCK, tp);
1586 }
1587 static int thread_cpu_clock_get(const clockid_t which_clock,
1588                                 struct timespec *tp)
1589 {
1590         return posix_cpu_clock_get(THREAD_CLOCK, tp);
1591 }
1592 static int thread_cpu_timer_create(struct k_itimer *timer)
1593 {
1594         timer->it_clock = THREAD_CLOCK;
1595         return posix_cpu_timer_create(timer);
1596 }
1597 static int thread_cpu_nsleep(const clockid_t which_clock, int flags,
1598                               struct timespec *rqtp, struct timespec __user *rmtp)
1599 {
1600         return -EINVAL;
1601 }
1602 static long thread_cpu_nsleep_restart(struct restart_block *restart_block)
1603 {
1604         return -EINVAL;
1605 }
1606
1607 static __init int init_posix_cpu_timers(void)
1608 {
1609         struct k_clock process = {
1610                 .clock_getres = process_cpu_clock_getres,
1611                 .clock_get = process_cpu_clock_get,
1612                 .clock_set = do_posix_clock_nosettime,
1613                 .timer_create = process_cpu_timer_create,
1614                 .nsleep = process_cpu_nsleep,
1615                 .nsleep_restart = process_cpu_nsleep_restart,
1616         };
1617         struct k_clock thread = {
1618                 .clock_getres = thread_cpu_clock_getres,
1619                 .clock_get = thread_cpu_clock_get,
1620                 .clock_set = do_posix_clock_nosettime,
1621                 .timer_create = thread_cpu_timer_create,
1622                 .nsleep = thread_cpu_nsleep,
1623                 .nsleep_restart = thread_cpu_nsleep_restart,
1624         };
1625         struct timespec ts;
1626
1627         register_posix_clock(CLOCK_PROCESS_CPUTIME_ID, &process);
1628         register_posix_clock(CLOCK_THREAD_CPUTIME_ID, &thread);
1629
1630         cputime_to_timespec(cputime_one_jiffy, &ts);
1631         onecputick = ts.tv_nsec;
1632         WARN_ON(ts.tv_sec != 0);
1633
1634         return 0;
1635 }
1636 __initcall(init_posix_cpu_timers);