kernel/time/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 #include <linux/random.h>
  13 #include <linux/tick.h>
  14 #include <linux/workqueue.h>
  15
  16 /*
  17  * Called after updating RLIMIT_CPU to run cpu timer and update
  18  * tsk->signal->cputime_expires expiration cache if necessary. Needs
  19  * siglock protection since other code may update expiration cache as
  20  * well.
  21  */
  22 void update_rlimit_cpu(struct task_struct *task, unsigned long rlim_new)
  23 {
  24         cputime_t cputime = secs_to_cputime(rlim_new);
  25
  26         spin_lock_irq(&task->sighand->siglock);
  27         set_process_cpu_timer(task, CPUCLOCK_PROF, &cputime, NULL);
  28         spin_unlock_irq(&task->sighand->siglock);
  29 }
  30
  31 static int check_clock(const clockid_t which_clock)
  32 {
  33         int error = 0;
  34         struct task_struct *p;
  35         const pid_t pid = CPUCLOCK_PID(which_clock);
  36
  37         if (CPUCLOCK_WHICH(which_clock) >= CPUCLOCK_MAX)
  38                 return -EINVAL;
  39
  40         if (pid == 0)
  41                 return 0;
  42
  43         rcu_read_lock();
  44         p = find_task_by_vpid(pid);
  45         if (!p || !(CPUCLOCK_PERTHREAD(which_clock) ?
  46                    same_thread_group(p, current) : has_group_leader_pid(p))) {
  47                 error = -EINVAL;
  48         }
  49         rcu_read_unlock();
  50
  51         return error;
  52 }
  53
  54 static inline unsigned long long
  55 timespec_to_sample(const clockid_t which_clock, const struct timespec *tp)
  56 {
  57         unsigned long long ret;
  58
  59         ret = 0;                /* high half always zero when .cpu used */
  60         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  61                 ret = (unsigned long long)tp->tv_sec * NSEC_PER_SEC + tp->tv_nsec;
  62         } else {
  63                 ret = cputime_to_expires(timespec_to_cputime(tp));
  64         }
  65         return ret;
  66 }
  67
  68 static void sample_to_timespec(const clockid_t which_clock,
  69                                unsigned long long expires,
  70                                struct timespec *tp)
  71 {
  72         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED)
  73                 *tp = ns_to_timespec(expires);
  74         else
  75                 cputime_to_timespec((__force cputime_t)expires, tp);
  76 }
  77
  78 /*
  79  * Update expiry time from increment, and increase overrun count,
  80  * given the current clock sample.
  81  */
  82 static void bump_cpu_timer(struct k_itimer *timer,
  83                            unsigned long long now)
  84 {
  85         int i;
  86         unsigned long long delta, incr;
  87
  88         if (timer->it.cpu.incr == 0)
  89                 return;
  90
  91         if (now < timer->it.cpu.expires)
  92                 return;
  93
  94         incr = timer->it.cpu.incr;
  95         delta = now + incr - timer->it.cpu.expires;
  96
  97         /* Don't use (incr*2 < delta), incr*2 might overflow. */
  98         for (i = 0; incr < delta - incr; i++)
  99                 incr = incr << 1;
 100
 101         for (; i >= 0; incr >>= 1, i--) {
 102                 if (delta < incr)
 103                         continue;
 104
 105                 timer->it.cpu.expires += incr;
 106                 timer->it_overrun += 1 << i;
 107                 delta -= incr;
 108         }
 109 }
 110
 111 /**
 112  * task_cputime_zero - Check a task_cputime struct for all zero fields.
 113  *
 114  * @cputime:    The struct to compare.
 115  *
 116  * Checks @cputime to see if all fields are zero.  Returns true if all fields
 117  * are zero, false if any field is nonzero.
 118  */
 119 static inline int task_cputime_zero(const struct task_cputime *cputime)
 120 {
 121         if (!cputime->utime && !cputime->stime && !cputime->sum_exec_runtime)
 122                 return 1;
 123         return 0;
 124 }
 125
 126 static inline unsigned long long prof_ticks(struct task_struct *p)
 127 {
 128         cputime_t utime, stime;
 129
 130         task_cputime(p, &utime, &stime);
 131
 132         return cputime_to_expires(utime + stime);
 133 }
 134 static inline unsigned long long virt_ticks(struct task_struct *p)
 135 {
 136         cputime_t utime;
 137
 138         task_cputime(p, &utime, NULL);
 139
 140         return cputime_to_expires(utime);
 141 }
 142
 143 static int
 144 posix_cpu_clock_getres(const clockid_t which_clock, struct timespec *tp)
 145 {
 146         int error = check_clock(which_clock);
 147         if (!error) {
 148                 tp->tv_sec = 0;
 149                 tp->tv_nsec = ((NSEC_PER_SEC + HZ - 1) / HZ);
 150                 if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
 151                         /*
 152                          * If sched_clock is using a cycle counter, we
 153                          * don't have any idea of its true resolution
 154                          * exported, but it is much more than 1s/HZ.
 155                          */
 156                         tp->tv_nsec = 1;
 157                 }
 158         }
 159         return error;
 160 }
 161
 162 static int
 163 posix_cpu_clock_set(const clockid_t which_clock, const struct timespec *tp)
 164 {
 165         /*
 166          * You can never reset a CPU clock, but we check for other errors
 167          * in the call before failing with EPERM.
 168          */
 169         int error = check_clock(which_clock);
 170         if (error == 0) {
 171                 error = -EPERM;
 172         }
 173         return error;
 174 }
 175
 176
 177 /*
 178  * Sample a per-thread clock for the given task.
 179  */
 180 static int cpu_clock_sample(const clockid_t which_clock, struct task_struct *p,
 181                             unsigned long long *sample)
 182 {
 183         switch (CPUCLOCK_WHICH(which_clock)) {
 184         default:
 185                 return -EINVAL;
 186         case CPUCLOCK_PROF:
 187                 *sample = prof_ticks(p);
 188                 break;
 189         case CPUCLOCK_VIRT:
 190                 *sample = virt_ticks(p);
 191                 break;
 192         case CPUCLOCK_SCHED:
 193                 *sample = task_sched_runtime(p);
 194                 break;
 195         }
 196         return 0;
 197 }
 198
 199 /*
 200  * Set cputime to sum_cputime if sum_cputime > cputime. Use cmpxchg
 201  * to avoid race conditions with concurrent updates to cputime.
 202  */
 203 static inline void __update_gt_cputime(atomic64_t *cputime, u64 sum_cputime)
 204 {
 205         u64 curr_cputime;
 206 retry:
 207         curr_cputime = atomic64_read(cputime);
 208         if (sum_cputime > curr_cputime) {
 209                 if (atomic64_cmpxchg(cputime, curr_cputime, sum_cputime) != curr_cputime)
 210                         goto retry;
 211         }
 212 }
 213
 214 static void update_gt_cputime(struct task_cputime_atomic *cputime_atomic, struct task_cputime *sum)
 215 {
 216         __update_gt_cputime(&cputime_atomic->utime, sum->utime);
 217         __update_gt_cputime(&cputime_atomic->stime, sum->stime);
 218         __update_gt_cputime(&cputime_atomic->sum_exec_runtime, sum->sum_exec_runtime);
 219 }
 220
 221 /* Sample task_cputime_atomic values in "atomic_timers", store results in "times". */
 222 static inline void sample_cputime_atomic(struct task_cputime *times,
 223                                          struct task_cputime_atomic *atomic_times)
 224 {
 225         times->utime = atomic64_read(&atomic_times->utime);
 226         times->stime = atomic64_read(&atomic_times->stime);
 227         times->sum_exec_runtime = atomic64_read(&atomic_times->sum_exec_runtime);
 228 }
 229
 230 void thread_group_cputimer(struct task_struct *tsk, struct task_cputime *times)
 231 {
 232         struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
 233         struct task_cputime sum;
 234
 235         /* Check if cputimer isn't running. This is accessed without locking. */
 236         if (!READ_ONCE(cputimer->running)) {
 237                 /*
 238                  * The POSIX timer interface allows for absolute time expiry
 239                  * values through the TIMER_ABSTIME flag, therefore we have
 240                  * to synchronize the timer to the clock every time we start it.
 241                  */
 242                 thread_group_cputime(tsk, &sum);
 243                 update_gt_cputime(&cputimer->cputime_atomic, &sum);
 244
 245                 /*
 246                  * We're setting cputimer->running without a lock. Ensure
 247                  * this only gets written to in one operation. We set
 248                  * running after update_gt_cputime() as a small optimization,
 249                  * but barriers are not required because update_gt_cputime()
 250                  * can handle concurrent updates.
 251                  */
 252                 WRITE_ONCE(cputimer->running, true);
 253         }
 254         sample_cputime_atomic(times, &cputimer->cputime_atomic);
 255 }
 256
 257 /*
 258  * Sample a process (thread group) clock for the given group_leader task.
 259  * Must be called with task sighand lock held for safe while_each_thread()
 260  * traversal.
 261  */
 262 static int cpu_clock_sample_group(const clockid_t which_clock,
 263                                   struct task_struct *p,
 264                                   unsigned long long *sample)
 265 {
 266         struct task_cputime cputime;
 267
 268         switch (CPUCLOCK_WHICH(which_clock)) {
 269         default:
 270                 return -EINVAL;
 271         case CPUCLOCK_PROF:
 272                 thread_group_cputime(p, &cputime);
 273                 *sample = cputime_to_expires(cputime.utime + cputime.stime);
 274                 break;
 275         case CPUCLOCK_VIRT:
 276                 thread_group_cputime(p, &cputime);
 277                 *sample = cputime_to_expires(cputime.utime);
 278                 break;
 279         case CPUCLOCK_SCHED:
 280                 thread_group_cputime(p, &cputime);
 281                 *sample = cputime.sum_exec_runtime;
 282                 break;
 283         }
 284         return 0;
 285 }
 286
 287 static int posix_cpu_clock_get_task(struct task_struct *tsk,
 288                                     const clockid_t which_clock,
 289                                     struct timespec *tp)
 290 {
 291         int err = -EINVAL;
 292         unsigned long long rtn;
 293
 294         if (CPUCLOCK_PERTHREAD(which_clock)) {
 295                 if (same_thread_group(tsk, current))
 296                         err = cpu_clock_sample(which_clock, tsk, &rtn);
 297         } else {
 298                 if (tsk == current || thread_group_leader(tsk))
 299                         err = cpu_clock_sample_group(which_clock, tsk, &rtn);
 300         }
 301
 302         if (!err)
 303                 sample_to_timespec(which_clock, rtn, tp);
 304
 305         return err;
 306 }
 307
 308
 309 static int posix_cpu_clock_get(const clockid_t which_clock, struct timespec *tp)
 310 {
 311         const pid_t pid = CPUCLOCK_PID(which_clock);
 312         int err = -EINVAL;
 313
 314         if (pid == 0) {
 315                 /*
 316                  * Special case constant value for our own clocks.
 317                  * We don't have to do any lookup to find ourselves.
 318                  */
 319                 err = posix_cpu_clock_get_task(current, which_clock, tp);
 320         } else {
 321                 /*
 322                  * Find the given PID, and validate that the caller
 323                  * should be able to see it.
 324                  */
 325                 struct task_struct *p;
 326                 rcu_read_lock();
 327                 p = find_task_by_vpid(pid);
 328                 if (p)
 329                         err = posix_cpu_clock_get_task(p, which_clock, tp);
 330                 rcu_read_unlock();
 331         }
 332
 333         return err;
 334 }
 335
 336
 337 /*
 338  * Validate the clockid_t for a new CPU-clock timer, and initialize the timer.
 339  * This is called from sys_timer_create() and do_cpu_nanosleep() with the
 340  * new timer already all-zeros initialized.
 341  */
 342 static int posix_cpu_timer_create(struct k_itimer *new_timer)
 343 {
 344         int ret = 0;
 345         const pid_t pid = CPUCLOCK_PID(new_timer->it_clock);
 346         struct task_struct *p;
 347
 348         if (CPUCLOCK_WHICH(new_timer->it_clock) >= CPUCLOCK_MAX)
 349                 return -EINVAL;
 350
 351         INIT_LIST_HEAD(&new_timer->it.cpu.entry);
 352
 353         rcu_read_lock();
 354         if (CPUCLOCK_PERTHREAD(new_timer->it_clock)) {
 355                 if (pid == 0) {
 356                         p = current;
 357                 } else {
 358                         p = find_task_by_vpid(pid);
 359                         if (p && !same_thread_group(p, current))
 360                                 p = NULL;
 361                 }
 362         } else {
 363                 if (pid == 0) {
 364                         p = current->group_leader;
 365                 } else {
 366                         p = find_task_by_vpid(pid);
 367                         if (p && !has_group_leader_pid(p))
 368                                 p = NULL;
 369                 }
 370         }
 371         new_timer->it.cpu.task = p;
 372         if (p) {
 373                 get_task_struct(p);
 374         } else {
 375                 ret = -EINVAL;
 376         }
 377         rcu_read_unlock();
 378
 379         return ret;
 380 }
 381
 382 /*
 383  * Clean up a CPU-clock timer that is about to be destroyed.
 384  * This is called from timer deletion with the timer already locked.
 385  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 386  * and try again.  (This happens when the timer is in the middle of firing.)
 387  */
 388 static int posix_cpu_timer_del(struct k_itimer *timer)
 389 {
 390         int ret = 0;
 391         unsigned long flags;
 392         struct sighand_struct *sighand;
 393         struct task_struct *p = timer->it.cpu.task;
 394
 395         WARN_ON_ONCE(p == NULL);
 396
 397         /*
 398          * Protect against sighand release/switch in exit/exec and process/
 399          * thread timer list entry concurrent read/writes.
 400          */
 401         sighand = lock_task_sighand(p, &flags);
 402         if (unlikely(sighand == NULL)) {
 403                 /*
 404                  * We raced with the reaping of the task.
 405                  * The deletion should have cleared us off the list.
 406                  */
 407                 WARN_ON_ONCE(!list_empty(&timer->it.cpu.entry));
 408         } else {
 409                 if (timer->it.cpu.firing)
 410                         ret = TIMER_RETRY;
 411                 else
 412                         list_del(&timer->it.cpu.entry);
 413
 414                 unlock_task_sighand(p, &flags);
 415         }
 416
 417         if (!ret)
 418                 put_task_struct(p);
 419
 420         return ret;
 421 }
 422
 423 static void cleanup_timers_list(struct list_head *head)
 424 {
 425         struct cpu_timer_list *timer, *next;
 426
 427         list_for_each_entry_safe(timer, next, head, entry)
 428                 list_del_init(&timer->entry);
 429 }
 430
 431 /*
 432  * Clean out CPU timers still ticking when a thread exited.  The task
 433  * pointer is cleared, and the expiry time is replaced with the residual
 434  * time for later timer_gettime calls to return.
 435  * This must be called with the siglock held.
 436  */
 437 static void cleanup_timers(struct list_head *head)
 438 {
 439         cleanup_timers_list(head);
 440         cleanup_timers_list(++head);
 441         cleanup_timers_list(++head);
 442 }
 443
 444 /*
 445  * These are both called with the siglock held, when the current thread
 446  * is being reaped.  When the final (leader) thread in the group is reaped,
 447  * posix_cpu_timers_exit_group will be called after posix_cpu_timers_exit.
 448  */
 449 void posix_cpu_timers_exit(struct task_struct *tsk)
 450 {
 451         add_device_randomness((const void*) &tsk->se.sum_exec_runtime,
 452                                                 sizeof(unsigned long long));
 453         cleanup_timers(tsk->cpu_timers);
 454
 455 }
 456 void posix_cpu_timers_exit_group(struct task_struct *tsk)
 457 {
 458         cleanup_timers(tsk->signal->cpu_timers);
 459 }
 460
 461 static inline int expires_gt(cputime_t expires, cputime_t new_exp)
 462 {
 463         return expires == 0 || expires > new_exp;
 464 }
 465
 466 /*
 467  * Insert the timer on the appropriate list before any timers that
 468  * expire later.  This must be called with the sighand lock held.
 469  */
 470 static void arm_timer(struct k_itimer *timer)
 471 {
 472         struct task_struct *p = timer->it.cpu.task;
 473         struct list_head *head, *listpos;
 474         struct task_cputime *cputime_expires;
 475         struct cpu_timer_list *const nt = &timer->it.cpu;
 476         struct cpu_timer_list *next;
 477
 478         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 479                 head = p->cpu_timers;
 480                 cputime_expires = &p->cputime_expires;
 481         } else {
 482                 head = p->signal->cpu_timers;
 483                 cputime_expires = &p->signal->cputime_expires;
 484         }
 485         head += CPUCLOCK_WHICH(timer->it_clock);
 486
 487         listpos = head;
 488         list_for_each_entry(next, head, entry) {
 489                 if (nt->expires < next->expires)
 490                         break;
 491                 listpos = &next->entry;
 492         }
 493         list_add(&nt->entry, listpos);
 494
 495         if (listpos == head) {
 496                 unsigned long long exp = nt->expires;
 497
 498                 /*
 499                  * We are the new earliest-expiring POSIX 1.b timer, hence
 500                  * need to update expiration cache. Take into account that
 501                  * for process timers we share expiration cache with itimers
 502                  * and RLIMIT_CPU and for thread timers with RLIMIT_RTTIME.
 503                  */
 504
 505                 switch (CPUCLOCK_WHICH(timer->it_clock)) {
 506                 case CPUCLOCK_PROF:
 507                         if (expires_gt(cputime_expires->prof_exp, expires_to_cputime(exp)))
 508                                 cputime_expires->prof_exp = expires_to_cputime(exp);
 509                         break;
 510                 case CPUCLOCK_VIRT:
 511                         if (expires_gt(cputime_expires->virt_exp, expires_to_cputime(exp)))
 512                                 cputime_expires->virt_exp = expires_to_cputime(exp);
 513                         break;
 514                 case CPUCLOCK_SCHED:
 515                         if (cputime_expires->sched_exp == 0 ||
 516                             cputime_expires->sched_exp > exp)
 517                                 cputime_expires->sched_exp = exp;
 518                         break;
 519                 }
 520         }
 521 }
 522
 523 /*
 524  * The timer is locked, fire it and arrange for its reload.
 525  */
 526 static void cpu_timer_fire(struct k_itimer *timer)
 527 {
 528         if ((timer->it_sigev_notify & ~SIGEV_THREAD_ID) == SIGEV_NONE) {
 529                 /*
 530                  * User don't want any signal.
 531                  */
 532                 timer->it.cpu.expires = 0;
 533         } else if (unlikely(timer->sigq == NULL)) {
 534                 /*
 535                  * This a special case for clock_nanosleep,
 536                  * not a normal timer from sys_timer_create.
 537                  */
 538                 wake_up_process(timer->it_process);
 539                 timer->it.cpu.expires = 0;
 540         } else if (timer->it.cpu.incr == 0) {
 541                 /*
 542                  * One-shot timer.  Clear it as soon as it's fired.
 543                  */
 544                 posix_timer_event(timer, 0);
 545                 timer->it.cpu.expires = 0;
 546         } else if (posix_timer_event(timer, ++timer->it_requeue_pending)) {
 547                 /*
 548                  * The signal did not get queued because the signal
 549                  * was ignored, so we won't get any callback to
 550                  * reload the timer.  But we need to keep it
 551                  * ticking in case the signal is deliverable next time.
 552                  */
 553                 posix_cpu_timer_schedule(timer);
 554         }
 555 }
 556
 557 /*
 558  * Sample a process (thread group) timer for the given group_leader task.
 559  * Must be called with task sighand lock held for safe while_each_thread()
 560  * traversal.
 561  */
 562 static int cpu_timer_sample_group(const clockid_t which_clock,
 563                                   struct task_struct *p,
 564                                   unsigned long long *sample)
 565 {
 566         struct task_cputime cputime;
 567
 568         thread_group_cputimer(p, &cputime);
 569         switch (CPUCLOCK_WHICH(which_clock)) {
 570         default:
 571                 return -EINVAL;
 572         case CPUCLOCK_PROF:
 573                 *sample = cputime_to_expires(cputime.utime + cputime.stime);
 574                 break;
 575         case CPUCLOCK_VIRT:
 576                 *sample = cputime_to_expires(cputime.utime);
 577                 break;
 578         case CPUCLOCK_SCHED:
 579                 *sample = cputime.sum_exec_runtime;
 580                 break;
 581         }
 582         return 0;
 583 }
 584
 585 #ifdef CONFIG_NO_HZ_FULL
 586 static void nohz_kick_work_fn(struct work_struct *work)
 587 {
 588         tick_nohz_full_kick_all();
 589 }
 590
 591 static DECLARE_WORK(nohz_kick_work, nohz_kick_work_fn);
 592
 593 /*
 594  * We need the IPIs to be sent from sane process context.
 595  * The posix cpu timers are always set with irqs disabled.
 596  */
 597 static void posix_cpu_timer_kick_nohz(void)
 598 {
 599         if (context_tracking_is_enabled())
 600                 schedule_work(&nohz_kick_work);
 601 }
 602
 603 bool posix_cpu_timers_can_stop_tick(struct task_struct *tsk)
 604 {
 605         if (!task_cputime_zero(&tsk->cputime_expires))
 606                 return false;
 607
 608         /* Check if cputimer is running. This is accessed without locking. */
 609         if (READ_ONCE(tsk->signal->cputimer.running))
 610                 return false;
 611
 612         return true;
 613 }
 614 #else
 615 static inline void posix_cpu_timer_kick_nohz(void) { }
 616 #endif
 617
 618 /*
 619  * Guts of sys_timer_settime for CPU timers.
 620  * This is called with the timer locked and interrupts disabled.
 621  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 622  * and try again.  (This happens when the timer is in the middle of firing.)
 623  */
 624 static int posix_cpu_timer_set(struct k_itimer *timer, int timer_flags,
 625                                struct itimerspec *new, struct itimerspec *old)
 626 {
 627         unsigned long flags;
 628         struct sighand_struct *sighand;
 629         struct task_struct *p = timer->it.cpu.task;
 630         unsigned long long old_expires, new_expires, old_incr, val;
 631         int ret;
 632
 633         WARN_ON_ONCE(p == NULL);
 634
 635         new_expires = timespec_to_sample(timer->it_clock, &new->it_value);
 636
 637         /*
 638          * Protect against sighand release/switch in exit/exec and p->cpu_timers
 639          * and p->signal->cpu_timers read/write in arm_timer()
 640          */
 641         sighand = lock_task_sighand(p, &flags);
 642         /*
 643          * If p has just been reaped, we can no
 644          * longer get any information about it at all.
 645          */
 646         if (unlikely(sighand == NULL)) {
 647                 return -ESRCH;
 648         }
 649
 650         /*
 651          * Disarm any old timer after extracting its expiry time.
 652          */
 653         WARN_ON_ONCE(!irqs_disabled());
 654
 655         ret = 0;
 656         old_incr = timer->it.cpu.incr;
 657         old_expires = timer->it.cpu.expires;
 658         if (unlikely(timer->it.cpu.firing)) {
 659                 timer->it.cpu.firing = -1;
 660                 ret = TIMER_RETRY;
 661         } else
 662                 list_del_init(&timer->it.cpu.entry);
 663
 664         /*
 665          * We need to sample the current value to convert the new
 666          * value from to relative and absolute, and to convert the
 667          * old value from absolute to relative.  To set a process
 668          * timer, we need a sample to balance the thread expiry
 669          * times (in arm_timer).  With an absolute time, we must
 670          * check if it's already passed.  In short, we need a sample.
 671          */
 672         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 673                 cpu_clock_sample(timer->it_clock, p, &val);
 674         } else {
 675                 cpu_timer_sample_group(timer->it_clock, p, &val);
 676         }
 677
 678         if (old) {
 679                 if (old_expires == 0) {
 680                         old->it_value.tv_sec = 0;
 681                         old->it_value.tv_nsec = 0;
 682                 } else {
 683                         /*
 684                          * Update the timer in case it has
 685                          * overrun already.  If it has,
 686                          * we'll report it as having overrun
 687                          * and with the next reloaded timer
 688                          * already ticking, though we are
 689                          * swallowing that pending
 690                          * notification here to install the
 691                          * new setting.
 692                          */
 693                         bump_cpu_timer(timer, val);
 694                         if (val < timer->it.cpu.expires) {
 695                                 old_expires = timer->it.cpu.expires - val;
 696                                 sample_to_timespec(timer->it_clock,
 697                                                    old_expires,
 698                                                    &old->it_value);
 699                         } else {
 700                                 old->it_value.tv_nsec = 1;
 701                                 old->it_value.tv_sec = 0;
 702                         }
 703                 }
 704         }
 705
 706         if (unlikely(ret)) {
 707                 /*
 708                  * We are colliding with the timer actually firing.
 709                  * Punt after filling in the timer's old value, and
 710                  * disable this firing since we are already reporting
 711                  * it as an overrun (thanks to bump_cpu_timer above).
 712                  */
 713                 unlock_task_sighand(p, &flags);
 714                 goto out;
 715         }
 716
 717         if (new_expires != 0 && !(timer_flags & TIMER_ABSTIME)) {
 718                 new_expires += val;
 719         }
 720
 721         /*
 722          * Install the new expiry time (or zero).
 723          * For a timer with no notification action, we don't actually
 724          * arm the timer (we'll just fake it for timer_gettime).
 725          */
 726         timer->it.cpu.expires = new_expires;
 727         if (new_expires != 0 && val < new_expires) {
 728                 arm_timer(timer);
 729         }
 730
 731         unlock_task_sighand(p, &flags);
 732         /*
 733          * Install the new reload setting, and
 734          * set up the signal and overrun bookkeeping.
 735          */
 736         timer->it.cpu.incr = timespec_to_sample(timer->it_clock,
 737                                                 &new->it_interval);
 738
 739         /*
 740          * This acts as a modification timestamp for the timer,
 741          * so any automatic reload attempt will punt on seeing
 742          * that we have reset the timer manually.
 743          */
 744         timer->it_requeue_pending = (timer->it_requeue_pending + 2) &
 745                 ~REQUEUE_PENDING;
 746         timer->it_overrun_last = 0;
 747         timer->it_overrun = -1;
 748
 749         if (new_expires != 0 && !(val < new_expires)) {
 750                 /*
 751                  * The designated time already passed, so we notify
 752                  * immediately, even if the thread never runs to
 753                  * accumulate more time on this clock.
 754                  */
 755                 cpu_timer_fire(timer);
 756         }
 757
 758         ret = 0;
 759  out:
 760         if (old) {
 761                 sample_to_timespec(timer->it_clock,
 762                                    old_incr, &old->it_interval);
 763         }
 764         if (!ret)
 765                 posix_cpu_timer_kick_nohz();
 766         return ret;
 767 }
 768
 769 static void posix_cpu_timer_get(struct k_itimer *timer, struct itimerspec *itp)
 770 {
 771         unsigned long long now;
 772         struct task_struct *p = timer->it.cpu.task;
 773
 774         WARN_ON_ONCE(p == NULL);
 775
 776         /*
 777          * Easy part: convert the reload time.
 778          */
 779         sample_to_timespec(timer->it_clock,
 780                            timer->it.cpu.incr, &itp->it_interval);
 781
 782         if (timer->it.cpu.expires == 0) {       /* Timer not armed at all.  */
 783                 itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
 784                 return;
 785         }
 786
 787         /*
 788          * Sample the clock to take the difference with the expiry time.
 789          */
 790         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 791                 cpu_clock_sample(timer->it_clock, p, &now);
 792         } else {
 793                 struct sighand_struct *sighand;
 794                 unsigned long flags;
 795
 796                 /*
 797                  * Protect against sighand release/switch in exit/exec and
 798                  * also make timer sampling safe if it ends up calling
 799                  * thread_group_cputime().
 800                  */
 801                 sighand = lock_task_sighand(p, &flags);
 802                 if (unlikely(sighand == NULL)) {
 803                         /*
 804                          * The process has been reaped.
 805                          * We can't even collect a sample any more.
 806                          * Call the timer disarmed, nothing else to do.
 807                          */
 808                         timer->it.cpu.expires = 0;
 809                         sample_to_timespec(timer->it_clock, timer->it.cpu.expires,
 810                                            &itp->it_value);
 811                 } else {
 812                         cpu_timer_sample_group(timer->it_clock, p, &now);
 813                         unlock_task_sighand(p, &flags);
 814                 }
 815         }
 816
 817         if (now < timer->it.cpu.expires) {
 818                 sample_to_timespec(timer->it_clock,
 819                                    timer->it.cpu.expires - now,
 820                                    &itp->it_value);
 821         } else {
 822                 /*
 823                  * The timer should have expired already, but the firing
 824                  * hasn't taken place yet.  Say it's just about to expire.
 825                  */
 826                 itp->it_value.tv_nsec = 1;
 827                 itp->it_value.tv_sec = 0;
 828         }
 829 }
 830
 831 static unsigned long long
 832 check_timers_list(struct list_head *timers,
 833                   struct list_head *firing,
 834                   unsigned long long curr)
 835 {
 836         int maxfire = 20;
 837
 838         while (!list_empty(timers)) {
 839                 struct cpu_timer_list *t;
 840
 841                 t = list_first_entry(timers, struct cpu_timer_list, entry);
 842
 843                 if (!--maxfire || curr < t->expires)
 844                         return t->expires;
 845
 846                 t->firing = 1;
 847                 list_move_tail(&t->entry, firing);
 848         }
 849
 850         return 0;
 851 }
 852
 853 /*
 854  * Check for any per-thread CPU timers that have fired and move them off
 855  * the tsk->cpu_timers[N] list onto the firing list.  Here we update the
 856  * tsk->it_*_expires values to reflect the remaining thread CPU timers.
 857  */
 858 static void check_thread_timers(struct task_struct *tsk,
 859                                 struct list_head *firing)
 860 {
 861         struct list_head *timers = tsk->cpu_timers;
 862         struct signal_struct *const sig = tsk->signal;
 863         struct task_cputime *tsk_expires = &tsk->cputime_expires;
 864         unsigned long long expires;
 865         unsigned long soft;
 866
 867         /*
 868          * If cputime_expires is zero, then there are no active
 869          * per thread CPU timers.
 870          */
 871         if (task_cputime_zero(&tsk->cputime_expires))
 872                 return;
 873
 874         expires = check_timers_list(timers, firing, prof_ticks(tsk));
 875         tsk_expires->prof_exp = expires_to_cputime(expires);
 876
 877         expires = check_timers_list(++timers, firing, virt_ticks(tsk));
 878         tsk_expires->virt_exp = expires_to_cputime(expires);
 879
 880         tsk_expires->sched_exp = check_timers_list(++timers, firing,
 881                                                    tsk->se.sum_exec_runtime);
 882
 883         /*
 884          * Check for the special case thread timers.
 885          */
 886         soft = READ_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_cur);
 887         if (soft != RLIM_INFINITY) {
 888                 unsigned long hard =
 889                         READ_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_max);
 890
 891                 if (hard != RLIM_INFINITY &&
 892                     tsk->rt.timeout > DIV_ROUND_UP(hard, USEC_PER_SEC/HZ)) {
 893                         /*
 894                          * At the hard limit, we just die.
 895                          * No need to calculate anything else now.
 896                          */
 897                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
 898                         return;
 899                 }
 900                 if (tsk->rt.timeout > DIV_ROUND_UP(soft, USEC_PER_SEC/HZ)) {
 901                         /*
 902                          * At the soft limit, send a SIGXCPU every second.
 903                          */
 904                         if (soft < hard) {
 905                                 soft += USEC_PER_SEC;
 906                                 sig->rlim[RLIMIT_RTTIME].rlim_cur = soft;
 907                         }
 908                         printk(KERN_INFO
 909                                 "RT Watchdog Timeout: %s[%d]\n",
 910                                 tsk->comm, task_pid_nr(tsk));
 911                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
 912                 }
 913         }
 914 }
 915
 916 static inline void stop_process_timers(struct signal_struct *sig)
 917 {
 918         struct thread_group_cputimer *cputimer = &sig->cputimer;
 919
 920         /* Turn off cputimer->running. This is done without locking. */
 921         WRITE_ONCE(cputimer->running, false);
 922 }
 923
 924 static u32 onecputick;
 925
 926 static void check_cpu_itimer(struct task_struct *tsk, struct cpu_itimer *it,
 927                              unsigned long long *expires,
 928                              unsigned long long cur_time, int signo)
 929 {
 930         if (!it->expires)
 931                 return;
 932
 933         if (cur_time >= it->expires) {
 934                 if (it->incr) {
 935                         it->expires += it->incr;
 936                         it->error += it->incr_error;
 937                         if (it->error >= onecputick) {
 938                                 it->expires -= cputime_one_jiffy;
 939                                 it->error -= onecputick;
 940                         }
 941                 } else {
 942                         it->expires = 0;
 943                 }
 944
 945                 trace_itimer_expire(signo == SIGPROF ?
 946                                     ITIMER_PROF : ITIMER_VIRTUAL,
 947                                     tsk->signal->leader_pid, cur_time);
 948                 __group_send_sig_info(signo, SEND_SIG_PRIV, tsk);
 949         }
 950
 951         if (it->expires && (!*expires || it->expires < *expires)) {
 952                 *expires = it->expires;
 953         }
 954 }
 955
 956 /*
 957  * Check for any per-thread CPU timers that have fired and move them
 958  * off the tsk->*_timers list onto the firing list.  Per-thread timers
 959  * have already been taken off.
 960  */
 961 static void check_process_timers(struct task_struct *tsk,
 962                                  struct list_head *firing)
 963 {
 964         struct signal_struct *const sig = tsk->signal;
 965         unsigned long long utime, ptime, virt_expires, prof_expires;
 966         unsigned long long sum_sched_runtime, sched_expires;
 967         struct list_head *timers = sig->cpu_timers;
 968         struct task_cputime cputime;
 969         unsigned long soft;
 970
 971         /*
 972          * If cputimer is not running, then there are no active
 973          * process wide timers (POSIX 1.b, itimers, RLIMIT_CPU).
 974          */
 975         if (!READ_ONCE(tsk->signal->cputimer.running))
 976                 return;
 977
 978         /*
 979          * Signify that a thread is checking for process timers.
 980          * Write access to this field is protected by the sighand lock.
 981          */
 982         sig->cputimer.checking_timer = true;
 983
 984         /*
 985          * Collect the current process totals.
 986          */
 987         thread_group_cputimer(tsk, &cputime);
 988         utime = cputime_to_expires(cputime.utime);
 989         ptime = utime + cputime_to_expires(cputime.stime);
 990         sum_sched_runtime = cputime.sum_exec_runtime;
 991
 992         prof_expires = check_timers_list(timers, firing, ptime);
 993         virt_expires = check_timers_list(++timers, firing, utime);
 994         sched_expires = check_timers_list(++timers, firing, sum_sched_runtime);
 995
 996         /*
 997          * Check for the special case process timers.
 998          */
 999         check_cpu_itimer(tsk, &sig->it[CPUCLOCK_PROF], &prof_expires, ptime,
1000                          SIGPROF);
1001         check_cpu_itimer(tsk, &sig->it[CPUCLOCK_VIRT], &virt_expires, utime,
1002                          SIGVTALRM);
1003         soft = READ_ONCE(sig->rlim[RLIMIT_CPU].rlim_cur);
1004         if (soft != RLIM_INFINITY) {
1005                 unsigned long psecs = cputime_to_secs(ptime);
1006                 unsigned long hard =
1007                         READ_ONCE(sig->rlim[RLIMIT_CPU].rlim_max);
1008                 cputime_t x;
1009                 if (psecs >= hard) {
1010                         /*
1011                          * At the hard limit, we just die.
1012                          * No need to calculate anything else now.
1013                          */
1014                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
1015                         return;
1016                 }
1017                 if (psecs >= soft) {
1018                         /*
1019                          * At the soft limit, send a SIGXCPU every second.
1020                          */
1021                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
1022                         if (soft < hard) {
1023                                 soft++;
1024                                 sig->rlim[RLIMIT_CPU].rlim_cur = soft;
1025                         }
1026                 }
1027                 x = secs_to_cputime(soft);
1028                 if (!prof_expires || x < prof_expires) {
1029                         prof_expires = x;
1030                 }
1031         }
1032
1033         sig->cputime_expires.prof_exp = expires_to_cputime(prof_expires);
1034         sig->cputime_expires.virt_exp = expires_to_cputime(virt_expires);
1035         sig->cputime_expires.sched_exp = sched_expires;
1036         if (task_cputime_zero(&sig->cputime_expires))
1037                 stop_process_timers(sig);
1038
1039         sig->cputimer.checking_timer = false;
1040 }
1041
1042 /*
1043  * This is called from the signal code (via do_schedule_next_timer)
1044  * when the last timer signal was delivered and we have to reload the timer.
1045  */
1046 void posix_cpu_timer_schedule(struct k_itimer *timer)
1047 {
1048         struct sighand_struct *sighand;
1049         unsigned long flags;
1050         struct task_struct *p = timer->it.cpu.task;
1051         unsigned long long now;
1052
1053         WARN_ON_ONCE(p == NULL);
1054
1055         /*
1056          * Fetch the current sample and update the timer's expiry time.
1057          */
1058         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
1059                 cpu_clock_sample(timer->it_clock, p, &now);
1060                 bump_cpu_timer(timer, now);
1061                 if (unlikely(p->exit_state))
1062                         goto out;
1063
1064                 /* Protect timer list r/w in arm_timer() */
1065                 sighand = lock_task_sighand(p, &flags);
1066                 if (!sighand)
1067                         goto out;
1068         } else {
1069                 /*
1070                  * Protect arm_timer() and timer sampling in case of call to
1071                  * thread_group_cputime().
1072                  */
1073                 sighand = lock_task_sighand(p, &flags);
1074                 if (unlikely(sighand == NULL)) {
1075                         /*
1076                          * The process has been reaped.
1077                          * We can't even collect a sample any more.
1078                          */
1079                         timer->it.cpu.expires = 0;
1080                         goto out;
1081                 } else if (unlikely(p->exit_state) && thread_group_empty(p)) {
1082                         unlock_task_sighand(p, &flags);
1083                         /* Optimizations: if the process is dying, no need to rearm */
1084                         goto out;
1085                 }
1086                 cpu_timer_sample_group(timer->it_clock, p, &now);
1087                 bump_cpu_timer(timer, now);
1088                 /* Leave the sighand locked for the call below.  */
1089         }
1090
1091         /*
1092          * Now re-arm for the new expiry time.
1093          */
1094         WARN_ON_ONCE(!irqs_disabled());
1095         arm_timer(timer);
1096         unlock_task_sighand(p, &flags);
1097
1098         /* Kick full dynticks CPUs in case they need to tick on the new timer */
1099         posix_cpu_timer_kick_nohz();
1100 out:
1101         timer->it_overrun_last = timer->it_overrun;
1102         timer->it_overrun = -1;
1103         ++timer->it_requeue_pending;
1104 }
1105
1106 /**
1107  * task_cputime_expired - Compare two task_cputime entities.
1108  *
1109  * @sample:     The task_cputime structure to be checked for expiration.
1110  * @expires:    Expiration times, against which @sample will be checked.
1111  *
1112  * Checks @sample against @expires to see if any field of @sample has expired.
1113  * Returns true if any field of the former is greater than the corresponding
1114  * field of the latter if the latter field is set.  Otherwise returns false.
1115  */
1116 static inline int task_cputime_expired(const struct task_cputime *sample,
1117                                         const struct task_cputime *expires)
1118 {
1119         if (expires->utime && sample->utime >= expires->utime)
1120                 return 1;
1121         if (expires->stime && sample->utime + sample->stime >= expires->stime)
1122                 return 1;
1123         if (expires->sum_exec_runtime != 0 &&
1124             sample->sum_exec_runtime >= expires->sum_exec_runtime)
1125                 return 1;
1126         return 0;
1127 }
1128
1129 /**
1130  * fastpath_timer_check - POSIX CPU timers fast path.
1131  *
1132  * @tsk:        The task (thread) being checked.
1133  *
1134  * Check the task and thread group timers.  If both are zero (there are no
1135  * timers set) return false.  Otherwise snapshot the task and thread group
1136  * timers and compare them with the corresponding expiration times.  Return
1137  * true if a timer has expired, else return false.
1138  */
1139 static inline int fastpath_timer_check(struct task_struct *tsk)
1140 {
1141         struct signal_struct *sig;
1142
1143         if (!task_cputime_zero(&tsk->cputime_expires)) {
1144                 struct task_cputime task_sample;
1145
1146                 task_cputime(tsk, &task_sample.utime, &task_sample.stime);
1147                 task_sample.sum_exec_runtime = tsk->se.sum_exec_runtime;
1148                 if (task_cputime_expired(&task_sample, &tsk->cputime_expires))
1149                         return 1;
1150         }
1151
1152         sig = tsk->signal;
1153         /*
1154          * Check if thread group timers expired when the cputimer is
1155          * running and no other thread in the group is already checking
1156          * for thread group cputimers. These fields are read without the
1157          * sighand lock. However, this is fine because this is meant to
1158          * be a fastpath heuristic to determine whether we should try to
1159          * acquire the sighand lock to check/handle timers.
1160          *
1161          * In the worst case scenario, if 'running' or 'checking_timer' gets
1162          * set but the current thread doesn't see the change yet, we'll wait
1163          * until the next thread in the group gets a scheduler interrupt to
1164          * handle the timer. This isn't an issue in practice because these
1165          * types of delays with signals actually getting sent are expected.
1166          */
1167         if (READ_ONCE(sig->cputimer.running) &&
1168             !READ_ONCE(sig->cputimer.checking_timer)) {
1169                 struct task_cputime group_sample;
1170
1171                 sample_cputime_atomic(&group_sample, &sig->cputimer.cputime_atomic);
1172
1173                 if (task_cputime_expired(&group_sample, &sig->cputime_expires))
1174                         return 1;
1175         }
1176
1177         return 0;
1178 }
1179
1180 /*
1181  * This is called from the timer interrupt handler.  The irq handler has
1182  * already updated our counts.  We need to check if any timers fire now.
1183  * Interrupts are disabled.
1184  */
1185 void run_posix_cpu_timers(struct task_struct *tsk)
1186 {
1187         LIST_HEAD(firing);
1188         struct k_itimer *timer, *next;
1189         unsigned long flags;
1190
1191         WARN_ON_ONCE(!irqs_disabled());
1192
1193         /*
1194          * The fast path checks that there are no expired thread or thread
1195          * group timers.  If that's so, just return.
1196          */
1197         if (!fastpath_timer_check(tsk))
1198                 return;
1199
1200         if (!lock_task_sighand(tsk, &flags))
1201                 return;
1202         /*
1203          * Here we take off tsk->signal->cpu_timers[N] and
1204          * tsk->cpu_timers[N] all the timers that are firing, and
1205          * put them on the firing list.
1206          */
1207         check_thread_timers(tsk, &firing);
1208
1209         check_process_timers(tsk, &firing);
1210
1211         /*
1212          * We must release these locks before taking any timer's lock.
1213          * There is a potential race with timer deletion here, as the
1214          * siglock now protects our private firing list.  We have set
1215          * the firing flag in each timer, so that a deletion attempt
1216          * that gets the timer lock before we do will give it up and
1217          * spin until we've taken care of that timer below.
1218          */
1219         unlock_task_sighand(tsk, &flags);
1220
1221         /*
1222          * Now that all the timers on our list have the firing flag,
1223          * no one will touch their list entries but us.  We'll take
1224          * each timer's lock before clearing its firing flag, so no
1225          * timer call will interfere.
1226          */
1227         list_for_each_entry_safe(timer, next, &firing, it.cpu.entry) {
1228                 int cpu_firing;
1229
1230                 spin_lock(&timer->it_lock);
1231                 list_del_init(&timer->it.cpu.entry);
1232                 cpu_firing = timer->it.cpu.firing;
1233                 timer->it.cpu.firing = 0;
1234                 /*
1235                  * The firing flag is -1 if we collided with a reset
1236                  * of the timer, which already reported this
1237                  * almost-firing as an overrun.  So don't generate an event.
1238                  */
1239                 if (likely(cpu_firing >= 0))
1240                         cpu_timer_fire(timer);
1241                 spin_unlock(&timer->it_lock);
1242         }
1243 }
1244
1245 /*
1246  * Set one of the process-wide special case CPU timers or RLIMIT_CPU.
1247  * The tsk->sighand->siglock must be held by the caller.
1248  */
1249 void set_process_cpu_timer(struct task_struct *tsk, unsigned int clock_idx,
1250                            cputime_t *newval, cputime_t *oldval)
1251 {
1252         unsigned long long now;
1253
1254         WARN_ON_ONCE(clock_idx == CPUCLOCK_SCHED);
1255         cpu_timer_sample_group(clock_idx, tsk, &now);
1256
1257         if (oldval) {
1258                 /*
1259                  * We are setting itimer. The *oldval is absolute and we update
1260                  * it to be relative, *newval argument is relative and we update
1261                  * it to be absolute.
1262                  */
1263                 if (*oldval) {
1264                         if (*oldval <= now) {
1265                                 /* Just about to fire. */
1266                                 *oldval = cputime_one_jiffy;
1267                         } else {
1268                                 *oldval -= now;
1269                         }
1270                 }
1271
1272                 if (!*newval)
1273                         goto out;
1274                 *newval += now;
1275         }
1276
1277         /*
1278          * Update expiration cache if we are the earliest timer, or eventually
1279          * RLIMIT_CPU limit is earlier than prof_exp cpu timer expire.
1280          */
1281         switch (clock_idx) {
1282         case CPUCLOCK_PROF:
1283                 if (expires_gt(tsk->signal->cputime_expires.prof_exp, *newval))
1284                         tsk->signal->cputime_expires.prof_exp = *newval;
1285                 break;
1286         case CPUCLOCK_VIRT:
1287                 if (expires_gt(tsk->signal->cputime_expires.virt_exp, *newval))
1288                         tsk->signal->cputime_expires.virt_exp = *newval;
1289                 break;
1290         }
1291 out:
1292         posix_cpu_timer_kick_nohz();
1293 }
1294
1295 static int do_cpu_nanosleep(const clockid_t which_clock, int flags,
1296                             struct timespec *rqtp, struct itimerspec *it)
1297 {
1298         struct k_itimer timer;
1299         int error;
1300
1301         /*
1302          * Set up a temporary timer and then wait for it to go off.
1303          */
1304         memset(&timer, 0, sizeof timer);
1305         spin_lock_init(&timer.it_lock);
1306         timer.it_clock = which_clock;
1307         timer.it_overrun = -1;
1308         error = posix_cpu_timer_create(&timer);
1309         timer.it_process = current;
1310         if (!error) {
1311                 static struct itimerspec zero_it;
1312
1313                 memset(it, 0, sizeof *it);
1314                 it->it_value = *rqtp;
1315
1316                 spin_lock_irq(&timer.it_lock);
1317                 error = posix_cpu_timer_set(&timer, flags, it, NULL);
1318                 if (error) {
1319                         spin_unlock_irq(&timer.it_lock);
1320                         return error;
1321                 }
1322
1323                 while (!signal_pending(current)) {
1324                         if (timer.it.cpu.expires == 0) {
1325                                 /*
1326                                  * Our timer fired and was reset, below
1327                                  * deletion can not fail.
1328                                  */
1329                                 posix_cpu_timer_del(&timer);
1330                                 spin_unlock_irq(&timer.it_lock);
1331                                 return 0;
1332                         }
1333
1334                         /*
1335                          * Block until cpu_timer_fire (or a signal) wakes us.
1336                          */
1337                         __set_current_state(TASK_INTERRUPTIBLE);
1338                         spin_unlock_irq(&timer.it_lock);
1339                         schedule();
1340                         spin_lock_irq(&timer.it_lock);
1341                 }
1342
1343                 /*
1344                  * We were interrupted by a signal.
1345                  */
1346                 sample_to_timespec(which_clock, timer.it.cpu.expires, rqtp);
1347                 error = posix_cpu_timer_set(&timer, 0, &zero_it, it);
1348                 if (!error) {
1349                         /*
1350                          * Timer is now unarmed, deletion can not fail.
1351                          */
1352                         posix_cpu_timer_del(&timer);
1353                 }
1354                 spin_unlock_irq(&timer.it_lock);
1355
1356                 while (error == TIMER_RETRY) {
1357                         /*
1358                          * We need to handle case when timer was or is in the
1359                          * middle of firing. In other cases we already freed
1360                          * resources.
1361                          */
1362                         spin_lock_irq(&timer.it_lock);
1363                         error = posix_cpu_timer_del(&timer);
1364                         spin_unlock_irq(&timer.it_lock);
1365                 }
1366
1367                 if ((it->it_value.tv_sec | it->it_value.tv_nsec) == 0) {
1368                         /*
1369                          * It actually did fire already.
1370                          */
1371                         return 0;
1372                 }
1373
1374                 error = -ERESTART_RESTARTBLOCK;
1375         }
1376
1377         return error;
1378 }
1379
1380 static long posix_cpu_nsleep_restart(struct restart_block *restart_block);
1381
1382 static int posix_cpu_nsleep(const clockid_t which_clock, int flags,
1383                             struct timespec *rqtp, struct timespec __user *rmtp)
1384 {
1385         struct restart_block *restart_block = &current->restart_block;
1386         struct itimerspec it;
1387         int error;
1388
1389         /*
1390          * Diagnose required errors first.
1391          */
1392         if (CPUCLOCK_PERTHREAD(which_clock) &&
1393             (CPUCLOCK_PID(which_clock) == 0 ||
1394              CPUCLOCK_PID(which_clock) == current->pid))
1395                 return -EINVAL;
1396
1397         error = do_cpu_nanosleep(which_clock, flags, rqtp, &it);
1398
1399         if (error == -ERESTART_RESTARTBLOCK) {
1400
1401                 if (flags & TIMER_ABSTIME)
1402                         return -ERESTARTNOHAND;
1403                 /*
1404                  * Report back to the user the time still remaining.
1405                  */
1406                 if (rmtp && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1407                         return -EFAULT;
1408
1409                 restart_block->fn = posix_cpu_nsleep_restart;
1410                 restart_block->nanosleep.clockid = which_clock;
1411                 restart_block->nanosleep.rmtp = rmtp;
1412                 restart_block->nanosleep.expires = timespec_to_ns(rqtp);
1413         }
1414         return error;
1415 }
1416
1417 static long posix_cpu_nsleep_restart(struct restart_block *restart_block)
1418 {
1419         clockid_t which_clock = restart_block->nanosleep.clockid;
1420         struct timespec t;
1421         struct itimerspec it;
1422         int error;
1423
1424         t = ns_to_timespec(restart_block->nanosleep.expires);
1425
1426         error = do_cpu_nanosleep(which_clock, TIMER_ABSTIME, &t, &it);
1427
1428         if (error == -ERESTART_RESTARTBLOCK) {
1429                 struct timespec __user *rmtp = restart_block->nanosleep.rmtp;
1430                 /*
1431                  * Report back to the user the time still remaining.
1432                  */
1433                 if (rmtp && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1434                         return -EFAULT;
1435
1436                 restart_block->nanosleep.expires = timespec_to_ns(&t);
1437         }
1438         return error;
1439
1440 }
1441
1442 #define PROCESS_CLOCK   MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED)
1443 #define THREAD_CLOCK    MAKE_THREAD_CPUCLOCK(0, CPUCLOCK_SCHED)
1444
1445 static int process_cpu_clock_getres(const clockid_t which_clock,
1446                                     struct timespec *tp)
1447 {
1448         return posix_cpu_clock_getres(PROCESS_CLOCK, tp);
1449 }
1450 static int process_cpu_clock_get(const clockid_t which_clock,
1451                                  struct timespec *tp)
1452 {
1453         return posix_cpu_clock_get(PROCESS_CLOCK, tp);
1454 }
1455 static int process_cpu_timer_create(struct k_itimer *timer)
1456 {
1457         timer->it_clock = PROCESS_CLOCK;
1458         return posix_cpu_timer_create(timer);
1459 }
1460 static int process_cpu_nsleep(const clockid_t which_clock, int flags,
1461                               struct timespec *rqtp,
1462                               struct timespec __user *rmtp)
1463 {
1464         return posix_cpu_nsleep(PROCESS_CLOCK, flags, rqtp, rmtp);
1465 }
1466 static long process_cpu_nsleep_restart(struct restart_block *restart_block)
1467 {
1468         return -EINVAL;
1469 }
1470 static int thread_cpu_clock_getres(const clockid_t which_clock,
1471                                    struct timespec *tp)
1472 {
1473         return posix_cpu_clock_getres(THREAD_CLOCK, tp);
1474 }
1475 static int thread_cpu_clock_get(const clockid_t which_clock,
1476                                 struct timespec *tp)
1477 {
1478         return posix_cpu_clock_get(THREAD_CLOCK, tp);
1479 }
1480 static int thread_cpu_timer_create(struct k_itimer *timer)
1481 {
1482         timer->it_clock = THREAD_CLOCK;
1483         return posix_cpu_timer_create(timer);
1484 }
1485
1486 struct k_clock clock_posix_cpu = {
1487         .clock_getres   = posix_cpu_clock_getres,
1488         .clock_set      = posix_cpu_clock_set,
1489         .clock_get      = posix_cpu_clock_get,
1490         .timer_create   = posix_cpu_timer_create,
1491         .nsleep         = posix_cpu_nsleep,
1492         .nsleep_restart = posix_cpu_nsleep_restart,
1493         .timer_set      = posix_cpu_timer_set,
1494         .timer_del      = posix_cpu_timer_del,
1495         .timer_get      = posix_cpu_timer_get,
1496 };
1497
1498 static __init int init_posix_cpu_timers(void)
1499 {
1500         struct k_clock process = {
1501                 .clock_getres   = process_cpu_clock_getres,
1502                 .clock_get      = process_cpu_clock_get,
1503                 .timer_create   = process_cpu_timer_create,
1504                 .nsleep         = process_cpu_nsleep,
1505                 .nsleep_restart = process_cpu_nsleep_restart,
1506         };
1507         struct k_clock thread = {
1508                 .clock_getres   = thread_cpu_clock_getres,
1509                 .clock_get      = thread_cpu_clock_get,
1510                 .timer_create   = thread_cpu_timer_create,
1511         };
1512         struct timespec ts;
1513
1514         posix_timers_register_clock(CLOCK_PROCESS_CPUTIME_ID, &process);
1515         posix_timers_register_clock(CLOCK_THREAD_CPUTIME_ID, &thread);
1516
1517         cputime_to_timespec(cputime_one_jiffy, &ts);
1518         onecputick = ts.tv_nsec;
1519         WARN_ON(ts.tv_sec != 0);
1520
1521         return 0;
1522 }
1523 __initcall(init_posix_cpu_timers);