kernel/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/sched/rt.h>
   7 #include <linux/posix-timers.h>
   8 #include <linux/errno.h>
   9 #include <linux/math64.h>
  10 #include <asm/uaccess.h>
  11 #include <linux/kernel_stat.h>
  12 #include <trace/events/timer.h>
  13 #include <linux/random.h>
  14 #include <linux/tick.h>
  15 #include <linux/workqueue.h>
  16
  17 /*
  18  * Called after updating RLIMIT_CPU to run cpu timer and update
  19  * tsk->signal->cputime_expires expiration cache if necessary. Needs
  20  * siglock protection since other code may update expiration cache as
  21  * well.
  22  */
  23 void update_rlimit_cpu(struct task_struct *task, unsigned long rlim_new)
  24 {
  25         cputime_t cputime = secs_to_cputime(rlim_new);
  26
  27         spin_lock_irq(&task->sighand->siglock);
  28         set_process_cpu_timer(task, CPUCLOCK_PROF, &cputime, NULL);
  29         spin_unlock_irq(&task->sighand->siglock);
  30 }
  31
  32 static int check_clock(const clockid_t which_clock)
  33 {
  34         int error = 0;
  35         struct task_struct *p;
  36         const pid_t pid = CPUCLOCK_PID(which_clock);
  37
  38         if (CPUCLOCK_WHICH(which_clock) >= CPUCLOCK_MAX)
  39                 return -EINVAL;
  40
  41         if (pid == 0)
  42                 return 0;
  43
  44         rcu_read_lock();
  45         p = find_task_by_vpid(pid);
  46         if (!p || !(CPUCLOCK_PERTHREAD(which_clock) ?
  47                    same_thread_group(p, current) : has_group_leader_pid(p))) {
  48                 error = -EINVAL;
  49         }
  50         rcu_read_unlock();
  51
  52         return error;
  53 }
  54
  55 static inline unsigned long long
  56 timespec_to_sample(const clockid_t which_clock, const struct timespec *tp)
  57 {
  58         unsigned long long ret;
  59
  60         ret = 0;                /* high half always zero when .cpu used */
  61         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  62                 ret = (unsigned long long)tp->tv_sec * NSEC_PER_SEC + tp->tv_nsec;
  63         } else {
  64                 ret = cputime_to_expires(timespec_to_cputime(tp));
  65         }
  66         return ret;
  67 }
  68
  69 static void sample_to_timespec(const clockid_t which_clock,
  70                                unsigned long long expires,
  71                                struct timespec *tp)
  72 {
  73         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED)
  74                 *tp = ns_to_timespec(expires);
  75         else
  76                 cputime_to_timespec((__force cputime_t)expires, tp);
  77 }
  78
  79 /*
  80  * Update expiry time from increment, and increase overrun count,
  81  * given the current clock sample.
  82  */
  83 static void bump_cpu_timer(struct k_itimer *timer,
  84                            unsigned long long now)
  85 {
  86         int i;
  87         unsigned long long delta, incr;
  88
  89         if (timer->it.cpu.incr == 0)
  90                 return;
  91
  92         if (now < timer->it.cpu.expires)
  93                 return;
  94
  95         incr = timer->it.cpu.incr;
  96         delta = now + incr - timer->it.cpu.expires;
  97
  98         /* Don't use (incr*2 < delta), incr*2 might overflow. */
  99         for (i = 0; incr < delta - incr; i++)
 100                 incr = incr << 1;
 101
 102         for (; i >= 0; incr >>= 1, i--) {
 103                 if (delta < incr)
 104                         continue;
 105
 106                 timer->it.cpu.expires += incr;
 107                 timer->it_overrun += 1 << i;
 108                 delta -= incr;
 109         }
 110 }
 111
 112 /**
 113  * task_cputime_zero - Check a task_cputime struct for all zero fields.
 114  *
 115  * @cputime:    The struct to compare.
 116  *
 117  * Checks @cputime to see if all fields are zero.  Returns true if all fields
 118  * are zero, false if any field is nonzero.
 119  */
 120 static inline int task_cputime_zero(const struct task_cputime *cputime)
 121 {
 122         if (!cputime->utime && !cputime->stime && !cputime->sum_exec_runtime)
 123                 return 1;
 124         return 0;
 125 }
 126
 127 static inline unsigned long long prof_ticks(struct task_struct *p)
 128 {
 129         cputime_t utime, stime;
 130
 131         task_cputime(p, &utime, &stime);
 132
 133         return cputime_to_expires(utime + stime);
 134 }
 135 static inline unsigned long long virt_ticks(struct task_struct *p)
 136 {
 137         cputime_t utime;
 138
 139         task_cputime(p, &utime, NULL);
 140
 141         return cputime_to_expires(utime);
 142 }
 143
 144 static int
 145 posix_cpu_clock_getres(const clockid_t which_clock, struct timespec *tp)
 146 {
 147         int error = check_clock(which_clock);
 148         if (!error) {
 149                 tp->tv_sec = 0;
 150                 tp->tv_nsec = ((NSEC_PER_SEC + HZ - 1) / HZ);
 151                 if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
 152                         /*
 153                          * If sched_clock is using a cycle counter, we
 154                          * don't have any idea of its true resolution
 155                          * exported, but it is much more than 1s/HZ.
 156                          */
 157                         tp->tv_nsec = 1;
 158                 }
 159         }
 160         return error;
 161 }
 162
 163 static int
 164 posix_cpu_clock_set(const clockid_t which_clock, const struct timespec *tp)
 165 {
 166         /*
 167          * You can never reset a CPU clock, but we check for other errors
 168          * in the call before failing with EPERM.
 169          */
 170         int error = check_clock(which_clock);
 171         if (error == 0) {
 172                 error = -EPERM;
 173         }
 174         return error;
 175 }
 176
 177
 178 /*
 179  * Sample a per-thread clock for the given task.
 180  */
 181 static int cpu_clock_sample(const clockid_t which_clock, struct task_struct *p,
 182                             unsigned long long *sample)
 183 {
 184         switch (CPUCLOCK_WHICH(which_clock)) {
 185         default:
 186                 return -EINVAL;
 187         case CPUCLOCK_PROF:
 188                 *sample = prof_ticks(p);
 189                 break;
 190         case CPUCLOCK_VIRT:
 191                 *sample = virt_ticks(p);
 192                 break;
 193         case CPUCLOCK_SCHED:
 194                 *sample = task_sched_runtime(p);
 195                 break;
 196         }
 197         return 0;
 198 }
 199
 200 static void update_gt_cputime(struct task_cputime *a, struct task_cputime *b)
 201 {
 202         if (b->utime > a->utime)
 203                 a->utime = b->utime;
 204
 205         if (b->stime > a->stime)
 206                 a->stime = b->stime;
 207
 208         if (b->sum_exec_runtime > a->sum_exec_runtime)
 209                 a->sum_exec_runtime = b->sum_exec_runtime;
 210 }
 211
 212 void thread_group_cputimer(struct task_struct *tsk, struct task_cputime *times)
 213 {
 214         struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
 215         struct task_cputime sum;
 216         unsigned long flags;
 217
 218         if (!cputimer->running) {
 219                 /*
 220                  * The POSIX timer interface allows for absolute time expiry
 221                  * values through the TIMER_ABSTIME flag, therefore we have
 222                  * to synchronize the timer to the clock every time we start
 223                  * it.
 224                  */
 225                 thread_group_cputime(tsk, &sum);
 226                 raw_spin_lock_irqsave(&cputimer->lock, flags);
 227                 cputimer->running = 1;
 228                 update_gt_cputime(&cputimer->cputime, &sum);
 229         } else
 230                 raw_spin_lock_irqsave(&cputimer->lock, flags);
 231         *times = cputimer->cputime;
 232         raw_spin_unlock_irqrestore(&cputimer->lock, flags);
 233 }
 234
 235 /*
 236  * Sample a process (thread group) clock for the given group_leader task.
 237  * Must be called with task sighand lock held for safe while_each_thread()
 238  * traversal.
 239  */
 240 static int cpu_clock_sample_group(const clockid_t which_clock,
 241                                   struct task_struct *p,
 242                                   unsigned long long *sample)
 243 {
 244         struct task_cputime cputime;
 245
 246         switch (CPUCLOCK_WHICH(which_clock)) {
 247         default:
 248                 return -EINVAL;
 249         case CPUCLOCK_PROF:
 250                 thread_group_cputime(p, &cputime);
 251                 *sample = cputime_to_expires(cputime.utime + cputime.stime);
 252                 break;
 253         case CPUCLOCK_VIRT:
 254                 thread_group_cputime(p, &cputime);
 255                 *sample = cputime_to_expires(cputime.utime);
 256                 break;
 257         case CPUCLOCK_SCHED:
 258                 thread_group_cputime(p, &cputime);
 259                 *sample = cputime.sum_exec_runtime;
 260                 break;
 261         }
 262         return 0;
 263 }
 264
 265 static int posix_cpu_clock_get_task(struct task_struct *tsk,
 266                                     const clockid_t which_clock,
 267                                     struct timespec *tp)
 268 {
 269         int err = -EINVAL;
 270         unsigned long long rtn;
 271
 272         if (CPUCLOCK_PERTHREAD(which_clock)) {
 273                 if (same_thread_group(tsk, current))
 274                         err = cpu_clock_sample(which_clock, tsk, &rtn);
 275         } else {
 276                 if (tsk == current || thread_group_leader(tsk))
 277                         err = cpu_clock_sample_group(which_clock, tsk, &rtn);
 278         }
 279
 280         if (!err)
 281                 sample_to_timespec(which_clock, rtn, tp);
 282
 283         return err;
 284 }
 285
 286
 287 static int posix_cpu_clock_get(const clockid_t which_clock, struct timespec *tp)
 288 {
 289         const pid_t pid = CPUCLOCK_PID(which_clock);
 290         int err = -EINVAL;
 291
 292         if (pid == 0) {
 293                 /*
 294                  * Special case constant value for our own clocks.
 295                  * We don't have to do any lookup to find ourselves.
 296                  */
 297                 err = posix_cpu_clock_get_task(current, which_clock, tp);
 298         } else {
 299                 /*
 300                  * Find the given PID, and validate that the caller
 301                  * should be able to see it.
 302                  */
 303                 struct task_struct *p;
 304                 rcu_read_lock();
 305                 p = find_task_by_vpid(pid);
 306                 if (p)
 307                         err = posix_cpu_clock_get_task(p, which_clock, tp);
 308                 rcu_read_unlock();
 309         }
 310
 311         return err;
 312 }
 313
 314
 315 /*
 316  * Validate the clockid_t for a new CPU-clock timer, and initialize the timer.
 317  * This is called from sys_timer_create() and do_cpu_nanosleep() with the
 318  * new timer already all-zeros initialized.
 319  */
 320 static int posix_cpu_timer_create(struct k_itimer *new_timer)
 321 {
 322         int ret = 0;
 323         const pid_t pid = CPUCLOCK_PID(new_timer->it_clock);
 324         struct task_struct *p;
 325
 326         if (CPUCLOCK_WHICH(new_timer->it_clock) >= CPUCLOCK_MAX)
 327                 return -EINVAL;
 328
 329         INIT_LIST_HEAD(&new_timer->it.cpu.entry);
 330
 331         rcu_read_lock();
 332         if (CPUCLOCK_PERTHREAD(new_timer->it_clock)) {
 333                 if (pid == 0) {
 334                         p = current;
 335                 } else {
 336                         p = find_task_by_vpid(pid);
 337                         if (p && !same_thread_group(p, current))
 338                                 p = NULL;
 339                 }
 340         } else {
 341                 if (pid == 0) {
 342                         p = current->group_leader;
 343                 } else {
 344                         p = find_task_by_vpid(pid);
 345                         if (p && !has_group_leader_pid(p))
 346                                 p = NULL;
 347                 }
 348         }
 349         new_timer->it.cpu.task = p;
 350         if (p) {
 351                 get_task_struct(p);
 352         } else {
 353                 ret = -EINVAL;
 354         }
 355         rcu_read_unlock();
 356
 357         return ret;
 358 }
 359
 360 /*
 361  * Clean up a CPU-clock timer that is about to be destroyed.
 362  * This is called from timer deletion with the timer already locked.
 363  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 364  * and try again.  (This happens when the timer is in the middle of firing.)
 365  */
 366 static int posix_cpu_timer_del(struct k_itimer *timer)
 367 {
 368         int ret = 0;
 369         unsigned long flags;
 370         struct sighand_struct *sighand;
 371         struct task_struct *p = timer->it.cpu.task;
 372
 373         WARN_ON_ONCE(p == NULL);
 374
 375         /*
 376          * Protect against sighand release/switch in exit/exec and process/
 377          * thread timer list entry concurrent read/writes.
 378          */
 379         sighand = lock_task_sighand(p, &flags);
 380         if (unlikely(sighand == NULL)) {
 381                 /*
 382                  * We raced with the reaping of the task.
 383                  * The deletion should have cleared us off the list.
 384                  */
 385                 WARN_ON_ONCE(!list_empty(&timer->it.cpu.entry));
 386         } else {
 387                 if (timer->it.cpu.firing)
 388                         ret = TIMER_RETRY;
 389                 else
 390                         list_del(&timer->it.cpu.entry);
 391
 392                 unlock_task_sighand(p, &flags);
 393         }
 394
 395         if (!ret)
 396                 put_task_struct(p);
 397
 398         return ret;
 399 }
 400
 401 static void cleanup_timers_list(struct list_head *head)
 402 {
 403         struct cpu_timer_list *timer, *next;
 404
 405         list_for_each_entry_safe(timer, next, head, entry)
 406                 list_del_init(&timer->entry);
 407 }
 408
 409 /*
 410  * Clean out CPU timers still ticking when a thread exited.  The task
 411  * pointer is cleared, and the expiry time is replaced with the residual
 412  * time for later timer_gettime calls to return.
 413  * This must be called with the siglock held.
 414  */
 415 static void cleanup_timers(struct list_head *head)
 416 {
 417         cleanup_timers_list(head);
 418         cleanup_timers_list(++head);
 419         cleanup_timers_list(++head);
 420 }
 421
 422 /*
 423  * These are both called with the siglock held, when the current thread
 424  * is being reaped.  When the final (leader) thread in the group is reaped,
 425  * posix_cpu_timers_exit_group will be called after posix_cpu_timers_exit.
 426  */
 427 void posix_cpu_timers_exit(struct task_struct *tsk)
 428 {
 429         add_device_randomness((const void*) &tsk->se.sum_exec_runtime,
 430                                                 sizeof(unsigned long long));
 431         cleanup_timers(tsk->cpu_timers);
 432
 433 }
 434 void posix_cpu_timers_exit_group(struct task_struct *tsk)
 435 {
 436         cleanup_timers(tsk->signal->cpu_timers);
 437 }
 438
 439 static inline int expires_gt(cputime_t expires, cputime_t new_exp)
 440 {
 441         return expires == 0 || expires > new_exp;
 442 }
 443
 444 /*
 445  * Insert the timer on the appropriate list before any timers that
 446  * expire later.  This must be called with the sighand lock held.
 447  */
 448 static void arm_timer(struct k_itimer *timer)
 449 {
 450         struct task_struct *p = timer->it.cpu.task;
 451         struct list_head *head, *listpos;
 452         struct task_cputime *cputime_expires;
 453         struct cpu_timer_list *const nt = &timer->it.cpu;
 454         struct cpu_timer_list *next;
 455
 456         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 457                 head = p->cpu_timers;
 458                 cputime_expires = &p->cputime_expires;
 459         } else {
 460                 head = p->signal->cpu_timers;
 461                 cputime_expires = &p->signal->cputime_expires;
 462         }
 463         head += CPUCLOCK_WHICH(timer->it_clock);
 464
 465         listpos = head;
 466         list_for_each_entry(next, head, entry) {
 467                 if (nt->expires < next->expires)
 468                         break;
 469                 listpos = &next->entry;
 470         }
 471         list_add(&nt->entry, listpos);
 472
 473         if (listpos == head) {
 474                 unsigned long long exp = nt->expires;
 475
 476                 /*
 477                  * We are the new earliest-expiring POSIX 1.b timer, hence
 478                  * need to update expiration cache. Take into account that
 479                  * for process timers we share expiration cache with itimers
 480                  * and RLIMIT_CPU and for thread timers with RLIMIT_RTTIME.
 481                  */
 482
 483                 switch (CPUCLOCK_WHICH(timer->it_clock)) {
 484                 case CPUCLOCK_PROF:
 485                         if (expires_gt(cputime_expires->prof_exp, expires_to_cputime(exp)))
 486                                 cputime_expires->prof_exp = expires_to_cputime(exp);
 487                         break;
 488                 case CPUCLOCK_VIRT:
 489                         if (expires_gt(cputime_expires->virt_exp, expires_to_cputime(exp)))
 490                                 cputime_expires->virt_exp = expires_to_cputime(exp);
 491                         break;
 492                 case CPUCLOCK_SCHED:
 493                         if (cputime_expires->sched_exp == 0 ||
 494                             cputime_expires->sched_exp > exp)
 495                                 cputime_expires->sched_exp = exp;
 496                         break;
 497                 }
 498         }
 499 }
 500
 501 /*
 502  * The timer is locked, fire it and arrange for its reload.
 503  */
 504 static void cpu_timer_fire(struct k_itimer *timer)
 505 {
 506         if ((timer->it_sigev_notify & ~SIGEV_THREAD_ID) == SIGEV_NONE) {
 507                 /*
 508                  * User don't want any signal.
 509                  */
 510                 timer->it.cpu.expires = 0;
 511         } else if (unlikely(timer->sigq == NULL)) {
 512                 /*
 513                  * This a special case for clock_nanosleep,
 514                  * not a normal timer from sys_timer_create.
 515                  */
 516                 wake_up_process(timer->it_process);
 517                 timer->it.cpu.expires = 0;
 518         } else if (timer->it.cpu.incr == 0) {
 519                 /*
 520                  * One-shot timer.  Clear it as soon as it's fired.
 521                  */
 522                 posix_timer_event(timer, 0);
 523                 timer->it.cpu.expires = 0;
 524         } else if (posix_timer_event(timer, ++timer->it_requeue_pending)) {
 525                 /*
 526                  * The signal did not get queued because the signal
 527                  * was ignored, so we won't get any callback to
 528                  * reload the timer.  But we need to keep it
 529                  * ticking in case the signal is deliverable next time.
 530                  */
 531                 posix_cpu_timer_schedule(timer);
 532         }
 533 }
 534
 535 /*
 536  * Sample a process (thread group) timer for the given group_leader task.
 537  * Must be called with task sighand lock held for safe while_each_thread()
 538  * traversal.
 539  */
 540 static int cpu_timer_sample_group(const clockid_t which_clock,
 541                                   struct task_struct *p,
 542                                   unsigned long long *sample)
 543 {
 544         struct task_cputime cputime;
 545
 546         thread_group_cputimer(p, &cputime);
 547         switch (CPUCLOCK_WHICH(which_clock)) {
 548         default:
 549                 return -EINVAL;
 550         case CPUCLOCK_PROF:
 551                 *sample = cputime_to_expires(cputime.utime + cputime.stime);
 552                 break;
 553         case CPUCLOCK_VIRT:
 554                 *sample = cputime_to_expires(cputime.utime);
 555                 break;
 556         case CPUCLOCK_SCHED:
 557                 *sample = cputime.sum_exec_runtime;
 558                 break;
 559         }
 560         return 0;
 561 }
 562
 563 #ifdef CONFIG_NO_HZ_FULL
 564 static void nohz_kick_work_fn(struct work_struct *work)
 565 {
 566         tick_nohz_full_kick_all();
 567 }
 568
 569 static DECLARE_WORK(nohz_kick_work, nohz_kick_work_fn);
 570
 571 /*
 572  * We need the IPIs to be sent from sane process context.
 573  * The posix cpu timers are always set with irqs disabled.
 574  */
 575 static void posix_cpu_timer_kick_nohz(void)
 576 {
 577         if (context_tracking_is_enabled())
 578                 schedule_work(&nohz_kick_work);
 579 }
 580
 581 bool posix_cpu_timers_can_stop_tick(struct task_struct *tsk)
 582 {
 583         if (!task_cputime_zero(&tsk->cputime_expires))
 584                 return false;
 585
 586         if (tsk->signal->cputimer.running)
 587                 return false;
 588
 589         return true;
 590 }
 591 #else
 592 static inline void posix_cpu_timer_kick_nohz(void) { }
 593 #endif
 594
 595 /*
 596  * Guts of sys_timer_settime for CPU timers.
 597  * This is called with the timer locked and interrupts disabled.
 598  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 599  * and try again.  (This happens when the timer is in the middle of firing.)
 600  */
 601 static int posix_cpu_timer_set(struct k_itimer *timer, int timer_flags,
 602                                struct itimerspec *new, struct itimerspec *old)
 603 {
 604         unsigned long flags;
 605         struct sighand_struct *sighand;
 606         struct task_struct *p = timer->it.cpu.task;
 607         unsigned long long old_expires, new_expires, old_incr, val;
 608         int ret;
 609
 610         WARN_ON_ONCE(p == NULL);
 611
 612         new_expires = timespec_to_sample(timer->it_clock, &new->it_value);
 613
 614         /*
 615          * Protect against sighand release/switch in exit/exec and p->cpu_timers
 616          * and p->signal->cpu_timers read/write in arm_timer()
 617          */
 618         sighand = lock_task_sighand(p, &flags);
 619         /*
 620          * If p has just been reaped, we can no
 621          * longer get any information about it at all.
 622          */
 623         if (unlikely(sighand == NULL)) {
 624                 return -ESRCH;
 625         }
 626
 627         /*
 628          * Disarm any old timer after extracting its expiry time.
 629          */
 630         WARN_ON_ONCE_NONRT(!irqs_disabled());
 631
 632         ret = 0;
 633         old_incr = timer->it.cpu.incr;
 634         old_expires = timer->it.cpu.expires;
 635         if (unlikely(timer->it.cpu.firing)) {
 636                 timer->it.cpu.firing = -1;
 637                 ret = TIMER_RETRY;
 638         } else
 639                 list_del_init(&timer->it.cpu.entry);
 640
 641         /*
 642          * We need to sample the current value to convert the new
 643          * value from to relative and absolute, and to convert the
 644          * old value from absolute to relative.  To set a process
 645          * timer, we need a sample to balance the thread expiry
 646          * times (in arm_timer).  With an absolute time, we must
 647          * check if it's already passed.  In short, we need a sample.
 648          */
 649         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 650                 cpu_clock_sample(timer->it_clock, p, &val);
 651         } else {
 652                 cpu_timer_sample_group(timer->it_clock, p, &val);
 653         }
 654
 655         if (old) {
 656                 if (old_expires == 0) {
 657                         old->it_value.tv_sec = 0;
 658                         old->it_value.tv_nsec = 0;
 659                 } else {
 660                         /*
 661                          * Update the timer in case it has
 662                          * overrun already.  If it has,
 663                          * we'll report it as having overrun
 664                          * and with the next reloaded timer
 665                          * already ticking, though we are
 666                          * swallowing that pending
 667                          * notification here to install the
 668                          * new setting.
 669                          */
 670                         bump_cpu_timer(timer, val);
 671                         if (val < timer->it.cpu.expires) {
 672                                 old_expires = timer->it.cpu.expires - val;
 673                                 sample_to_timespec(timer->it_clock,
 674                                                    old_expires,
 675                                                    &old->it_value);
 676                         } else {
 677                                 old->it_value.tv_nsec = 1;
 678                                 old->it_value.tv_sec = 0;
 679                         }
 680                 }
 681         }
 682
 683         if (unlikely(ret)) {
 684                 /*
 685                  * We are colliding with the timer actually firing.
 686                  * Punt after filling in the timer's old value, and
 687                  * disable this firing since we are already reporting
 688                  * it as an overrun (thanks to bump_cpu_timer above).
 689                  */
 690                 unlock_task_sighand(p, &flags);
 691                 goto out;
 692         }
 693
 694         if (new_expires != 0 && !(timer_flags & TIMER_ABSTIME)) {
 695                 new_expires += val;
 696         }
 697
 698         /*
 699          * Install the new expiry time (or zero).
 700          * For a timer with no notification action, we don't actually
 701          * arm the timer (we'll just fake it for timer_gettime).
 702          */
 703         timer->it.cpu.expires = new_expires;
 704         if (new_expires != 0 && val < new_expires) {
 705                 arm_timer(timer);
 706         }
 707
 708         unlock_task_sighand(p, &flags);
 709         /*
 710          * Install the new reload setting, and
 711          * set up the signal and overrun bookkeeping.
 712          */
 713         timer->it.cpu.incr = timespec_to_sample(timer->it_clock,
 714                                                 &new->it_interval);
 715
 716         /*
 717          * This acts as a modification timestamp for the timer,
 718          * so any automatic reload attempt will punt on seeing
 719          * that we have reset the timer manually.
 720          */
 721         timer->it_requeue_pending = (timer->it_requeue_pending + 2) &
 722                 ~REQUEUE_PENDING;
 723         timer->it_overrun_last = 0;
 724         timer->it_overrun = -1;
 725
 726         if (new_expires != 0 && !(val < new_expires)) {
 727                 /*
 728                  * The designated time already passed, so we notify
 729                  * immediately, even if the thread never runs to
 730                  * accumulate more time on this clock.
 731                  */
 732                 cpu_timer_fire(timer);
 733         }
 734
 735         ret = 0;
 736  out:
 737         if (old) {
 738                 sample_to_timespec(timer->it_clock,
 739                                    old_incr, &old->it_interval);
 740         }
 741         if (!ret)
 742                 posix_cpu_timer_kick_nohz();
 743         return ret;
 744 }
 745
 746 static void posix_cpu_timer_get(struct k_itimer *timer, struct itimerspec *itp)
 747 {
 748         unsigned long long now;
 749         struct task_struct *p = timer->it.cpu.task;
 750
 751         WARN_ON_ONCE(p == NULL);
 752
 753         /*
 754          * Easy part: convert the reload time.
 755          */
 756         sample_to_timespec(timer->it_clock,
 757                            timer->it.cpu.incr, &itp->it_interval);
 758
 759         if (timer->it.cpu.expires == 0) {       /* Timer not armed at all.  */
 760                 itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
 761                 return;
 762         }
 763
 764         /*
 765          * Sample the clock to take the difference with the expiry time.
 766          */
 767         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 768                 cpu_clock_sample(timer->it_clock, p, &now);
 769         } else {
 770                 struct sighand_struct *sighand;
 771                 unsigned long flags;
 772
 773                 /*
 774                  * Protect against sighand release/switch in exit/exec and
 775                  * also make timer sampling safe if it ends up calling
 776                  * thread_group_cputime().
 777                  */
 778                 sighand = lock_task_sighand(p, &flags);
 779                 if (unlikely(sighand == NULL)) {
 780                         /*
 781                          * The process has been reaped.
 782                          * We can't even collect a sample any more.
 783                          * Call the timer disarmed, nothing else to do.
 784                          */
 785                         timer->it.cpu.expires = 0;
 786                         sample_to_timespec(timer->it_clock, timer->it.cpu.expires,
 787                                            &itp->it_value);
 788                 } else {
 789                         cpu_timer_sample_group(timer->it_clock, p, &now);
 790                         unlock_task_sighand(p, &flags);
 791                 }
 792         }
 793
 794         if (now < timer->it.cpu.expires) {
 795                 sample_to_timespec(timer->it_clock,
 796                                    timer->it.cpu.expires - now,
 797                                    &itp->it_value);
 798         } else {
 799                 /*
 800                  * The timer should have expired already, but the firing
 801                  * hasn't taken place yet.  Say it's just about to expire.
 802                  */
 803                 itp->it_value.tv_nsec = 1;
 804                 itp->it_value.tv_sec = 0;
 805         }
 806 }
 807
 808 static unsigned long long
 809 check_timers_list(struct list_head *timers,
 810                   struct list_head *firing,
 811                   unsigned long long curr)
 812 {
 813         int maxfire = 20;
 814
 815         while (!list_empty(timers)) {
 816                 struct cpu_timer_list *t;
 817
 818                 t = list_first_entry(timers, struct cpu_timer_list, entry);
 819
 820                 if (!--maxfire || curr < t->expires)
 821                         return t->expires;
 822
 823                 t->firing = 1;
 824                 list_move_tail(&t->entry, firing);
 825         }
 826
 827         return 0;
 828 }
 829
 830 /*
 831  * Check for any per-thread CPU timers that have fired and move them off
 832  * the tsk->cpu_timers[N] list onto the firing list.  Here we update the
 833  * tsk->it_*_expires values to reflect the remaining thread CPU timers.
 834  */
 835 static void check_thread_timers(struct task_struct *tsk,
 836                                 struct list_head *firing)
 837 {
 838         struct list_head *timers = tsk->cpu_timers;
 839         struct signal_struct *const sig = tsk->signal;
 840         struct task_cputime *tsk_expires = &tsk->cputime_expires;
 841         unsigned long long expires;
 842         unsigned long soft;
 843
 844         expires = check_timers_list(timers, firing, prof_ticks(tsk));
 845         tsk_expires->prof_exp = expires_to_cputime(expires);
 846
 847         expires = check_timers_list(++timers, firing, virt_ticks(tsk));
 848         tsk_expires->virt_exp = expires_to_cputime(expires);
 849
 850         tsk_expires->sched_exp = check_timers_list(++timers, firing,
 851                                                    tsk->se.sum_exec_runtime);
 852
 853         /*
 854          * Check for the special case thread timers.
 855          */
 856         soft = ACCESS_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_cur);
 857         if (soft != RLIM_INFINITY) {
 858                 unsigned long hard =
 859                         ACCESS_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_max);
 860
 861                 if (hard != RLIM_INFINITY &&
 862                     tsk->rt.timeout > DIV_ROUND_UP(hard, USEC_PER_SEC/HZ)) {
 863                         /*
 864                          * At the hard limit, we just die.
 865                          * No need to calculate anything else now.
 866                          */
 867                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
 868                         return;
 869                 }
 870                 if (tsk->rt.timeout > DIV_ROUND_UP(soft, USEC_PER_SEC/HZ)) {
 871                         /*
 872                          * At the soft limit, send a SIGXCPU every second.
 873                          */
 874                         if (soft < hard) {
 875                                 soft += USEC_PER_SEC;
 876                                 sig->rlim[RLIMIT_RTTIME].rlim_cur = soft;
 877                         }
 878                         printk(KERN_INFO
 879                                 "RT Watchdog Timeout: %s[%d]\n",
 880                                 tsk->comm, task_pid_nr(tsk));
 881                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
 882                 }
 883         }
 884 }
 885
 886 static void stop_process_timers(struct signal_struct *sig)
 887 {
 888         struct thread_group_cputimer *cputimer = &sig->cputimer;
 889         unsigned long flags;
 890
 891         raw_spin_lock_irqsave(&cputimer->lock, flags);
 892         cputimer->running = 0;
 893         raw_spin_unlock_irqrestore(&cputimer->lock, flags);
 894 }
 895
 896 static u32 onecputick;
 897
 898 static void check_cpu_itimer(struct task_struct *tsk, struct cpu_itimer *it,
 899                              unsigned long long *expires,
 900                              unsigned long long cur_time, int signo)
 901 {
 902         if (!it->expires)
 903                 return;
 904
 905         if (cur_time >= it->expires) {
 906                 if (it->incr) {
 907                         it->expires += it->incr;
 908                         it->error += it->incr_error;
 909                         if (it->error >= onecputick) {
 910                                 it->expires -= cputime_one_jiffy;
 911                                 it->error -= onecputick;
 912                         }
 913                 } else {
 914                         it->expires = 0;
 915                 }
 916
 917                 trace_itimer_expire(signo == SIGPROF ?
 918                                     ITIMER_PROF : ITIMER_VIRTUAL,
 919                                     tsk->signal->leader_pid, cur_time);
 920                 __group_send_sig_info(signo, SEND_SIG_PRIV, tsk);
 921         }
 922
 923         if (it->expires && (!*expires || it->expires < *expires)) {
 924                 *expires = it->expires;
 925         }
 926 }
 927
 928 /*
 929  * Check for any per-thread CPU timers that have fired and move them
 930  * off the tsk->*_timers list onto the firing list.  Per-thread timers
 931  * have already been taken off.
 932  */
 933 static void check_process_timers(struct task_struct *tsk,
 934                                  struct list_head *firing)
 935 {
 936         struct signal_struct *const sig = tsk->signal;
 937         unsigned long long utime, ptime, virt_expires, prof_expires;
 938         unsigned long long sum_sched_runtime, sched_expires;
 939         struct list_head *timers = sig->cpu_timers;
 940         struct task_cputime cputime;
 941         unsigned long soft;
 942
 943         /*
 944          * Collect the current process totals.
 945          */
 946         thread_group_cputimer(tsk, &cputime);
 947         utime = cputime_to_expires(cputime.utime);
 948         ptime = utime + cputime_to_expires(cputime.stime);
 949         sum_sched_runtime = cputime.sum_exec_runtime;
 950
 951         prof_expires = check_timers_list(timers, firing, ptime);
 952         virt_expires = check_timers_list(++timers, firing, utime);
 953         sched_expires = check_timers_list(++timers, firing, sum_sched_runtime);
 954
 955         /*
 956          * Check for the special case process timers.
 957          */
 958         check_cpu_itimer(tsk, &sig->it[CPUCLOCK_PROF], &prof_expires, ptime,
 959                          SIGPROF);
 960         check_cpu_itimer(tsk, &sig->it[CPUCLOCK_VIRT], &virt_expires, utime,
 961                          SIGVTALRM);
 962         soft = ACCESS_ONCE(sig->rlim[RLIMIT_CPU].rlim_cur);
 963         if (soft != RLIM_INFINITY) {
 964                 unsigned long psecs = cputime_to_secs(ptime);
 965                 unsigned long hard =
 966                         ACCESS_ONCE(sig->rlim[RLIMIT_CPU].rlim_max);
 967                 cputime_t x;
 968                 if (psecs >= hard) {
 969                         /*
 970                          * At the hard limit, we just die.
 971                          * No need to calculate anything else now.
 972                          */
 973                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
 974                         return;
 975                 }
 976                 if (psecs >= soft) {
 977                         /*
 978                          * At the soft limit, send a SIGXCPU every second.
 979                          */
 980                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
 981                         if (soft < hard) {
 982                                 soft++;
 983                                 sig->rlim[RLIMIT_CPU].rlim_cur = soft;
 984                         }
 985                 }
 986                 x = secs_to_cputime(soft);
 987                 if (!prof_expires || x < prof_expires) {
 988                         prof_expires = x;
 989                 }
 990         }
 991
 992         sig->cputime_expires.prof_exp = expires_to_cputime(prof_expires);
 993         sig->cputime_expires.virt_exp = expires_to_cputime(virt_expires);
 994         sig->cputime_expires.sched_exp = sched_expires;
 995         if (task_cputime_zero(&sig->cputime_expires))
 996                 stop_process_timers(sig);
 997 }
 998
 999 /*
1000  * This is called from the signal code (via do_schedule_next_timer)
1001  * when the last timer signal was delivered and we have to reload the timer.
1002  */
1003 void posix_cpu_timer_schedule(struct k_itimer *timer)
1004 {
1005         struct sighand_struct *sighand;
1006         unsigned long flags;
1007         struct task_struct *p = timer->it.cpu.task;
1008         unsigned long long now;
1009
1010         WARN_ON_ONCE(p == NULL);
1011
1012         /*
1013          * Fetch the current sample and update the timer's expiry time.
1014          */
1015         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
1016                 cpu_clock_sample(timer->it_clock, p, &now);
1017                 bump_cpu_timer(timer, now);
1018                 if (unlikely(p->exit_state))
1019                         goto out;
1020
1021                 /* Protect timer list r/w in arm_timer() */
1022                 sighand = lock_task_sighand(p, &flags);
1023                 if (!sighand)
1024                         goto out;
1025         } else {
1026                 /*
1027                  * Protect arm_timer() and timer sampling in case of call to
1028                  * thread_group_cputime().
1029                  */
1030                 sighand = lock_task_sighand(p, &flags);
1031                 if (unlikely(sighand == NULL)) {
1032                         /*
1033                          * The process has been reaped.
1034                          * We can't even collect a sample any more.
1035                          */
1036                         timer->it.cpu.expires = 0;
1037                         goto out;
1038                 } else if (unlikely(p->exit_state) && thread_group_empty(p)) {
1039                         unlock_task_sighand(p, &flags);
1040                         /* Optimizations: if the process is dying, no need to rearm */
1041                         goto out;
1042                 }
1043                 cpu_timer_sample_group(timer->it_clock, p, &now);
1044                 bump_cpu_timer(timer, now);
1045                 /* Leave the sighand locked for the call below.  */
1046         }
1047
1048         /*
1049          * Now re-arm for the new expiry time.
1050          */
1051         WARN_ON_ONCE_NONRT(!irqs_disabled());
1052         arm_timer(timer);
1053         unlock_task_sighand(p, &flags);
1054
1055         /* Kick full dynticks CPUs in case they need to tick on the new timer */
1056         posix_cpu_timer_kick_nohz();
1057 out:
1058         timer->it_overrun_last = timer->it_overrun;
1059         timer->it_overrun = -1;
1060         ++timer->it_requeue_pending;
1061 }
1062
1063 /**
1064  * task_cputime_expired - Compare two task_cputime entities.
1065  *
1066  * @sample:     The task_cputime structure to be checked for expiration.
1067  * @expires:    Expiration times, against which @sample will be checked.
1068  *
1069  * Checks @sample against @expires to see if any field of @sample has expired.
1070  * Returns true if any field of the former is greater than the corresponding
1071  * field of the latter if the latter field is set.  Otherwise returns false.
1072  */
1073 static inline int task_cputime_expired(const struct task_cputime *sample,
1074                                         const struct task_cputime *expires)
1075 {
1076         if (expires->utime && sample->utime >= expires->utime)
1077                 return 1;
1078         if (expires->stime && sample->utime + sample->stime >= expires->stime)
1079                 return 1;
1080         if (expires->sum_exec_runtime != 0 &&
1081             sample->sum_exec_runtime >= expires->sum_exec_runtime)
1082                 return 1;
1083         return 0;
1084 }
1085
1086 /**
1087  * fastpath_timer_check - POSIX CPU timers fast path.
1088  *
1089  * @tsk:        The task (thread) being checked.
1090  *
1091  * Check the task and thread group timers.  If both are zero (there are no
1092  * timers set) return false.  Otherwise snapshot the task and thread group
1093  * timers and compare them with the corresponding expiration times.  Return
1094  * true if a timer has expired, else return false.
1095  */
1096 static inline int fastpath_timer_check(struct task_struct *tsk)
1097 {
1098         struct signal_struct *sig;
1099         cputime_t utime, stime;
1100
1101         task_cputime(tsk, &utime, &stime);
1102
1103         if (!task_cputime_zero(&tsk->cputime_expires)) {
1104                 struct task_cputime task_sample = {
1105                         .utime = utime,
1106                         .stime = stime,
1107                         .sum_exec_runtime = tsk->se.sum_exec_runtime
1108                 };
1109
1110                 if (task_cputime_expired(&task_sample, &tsk->cputime_expires))
1111                         return 1;
1112         }
1113
1114         sig = tsk->signal;
1115         if (sig->cputimer.running) {
1116                 struct task_cputime group_sample;
1117                 unsigned long flags;
1118
1119                 raw_spin_lock_irqsave(&sig->cputimer.lock, flags);
1120                 group_sample = sig->cputimer.cputime;
1121                 raw_spin_unlock_irqrestore(&sig->cputimer.lock, flags);
1122
1123                 if (task_cputime_expired(&group_sample, &sig->cputime_expires))
1124                         return 1;
1125         }
1126
1127         return 0;
1128 }
1129
1130 /*
1131  * This is called from the timer interrupt handler.  The irq handler has
1132  * already updated our counts.  We need to check if any timers fire now.
1133  * Interrupts are disabled.
1134  */
1135 static void __run_posix_cpu_timers(struct task_struct *tsk)
1136 {
1137         LIST_HEAD(firing);
1138         struct k_itimer *timer, *next;
1139         unsigned long flags;
1140
1141         WARN_ON_ONCE_NONRT(!irqs_disabled());
1142
1143         /*
1144          * The fast path checks that there are no expired thread or thread
1145          * group timers.  If that's so, just return.
1146          */
1147         if (!fastpath_timer_check(tsk))
1148                 return;
1149
1150         if (!lock_task_sighand(tsk, &flags))
1151                 return;
1152         /*
1153          * Here we take off tsk->signal->cpu_timers[N] and
1154          * tsk->cpu_timers[N] all the timers that are firing, and
1155          * put them on the firing list.
1156          */
1157         check_thread_timers(tsk, &firing);
1158         /*
1159          * If there are any active process wide timers (POSIX 1.b, itimers,
1160          * RLIMIT_CPU) cputimer must be running.
1161          */
1162         if (tsk->signal->cputimer.running)
1163                 check_process_timers(tsk, &firing);
1164
1165         /*
1166          * We must release these locks before taking any timer's lock.
1167          * There is a potential race with timer deletion here, as the
1168          * siglock now protects our private firing list.  We have set
1169          * the firing flag in each timer, so that a deletion attempt
1170          * that gets the timer lock before we do will give it up and
1171          * spin until we've taken care of that timer below.
1172          */
1173         unlock_task_sighand(tsk, &flags);
1174
1175         /*
1176          * Now that all the timers on our list have the firing flag,
1177          * no one will touch their list entries but us.  We'll take
1178          * each timer's lock before clearing its firing flag, so no
1179          * timer call will interfere.
1180          */
1181         list_for_each_entry_safe(timer, next, &firing, it.cpu.entry) {
1182                 int cpu_firing;
1183
1184                 spin_lock(&timer->it_lock);
1185                 list_del_init(&timer->it.cpu.entry);
1186                 cpu_firing = timer->it.cpu.firing;
1187                 timer->it.cpu.firing = 0;
1188                 /*
1189                  * The firing flag is -1 if we collided with a reset
1190                  * of the timer, which already reported this
1191                  * almost-firing as an overrun.  So don't generate an event.
1192                  */
1193                 if (likely(cpu_firing >= 0))
1194                         cpu_timer_fire(timer);
1195                 spin_unlock(&timer->it_lock);
1196         }
1197 }
1198
1199 #ifdef CONFIG_PREEMPT_RT_BASE
1200 #include <linux/kthread.h>
1201 #include <linux/cpu.h>
1202 DEFINE_PER_CPU(struct task_struct *, posix_timer_task);
1203 DEFINE_PER_CPU(struct task_struct *, posix_timer_tasklist);
1204
1205 static int posix_cpu_timers_thread(void *data)
1206 {
1207         int cpu = (long)data;
1208
1209         BUG_ON(per_cpu(posix_timer_task,cpu) != current);
1210
1211         while (!kthread_should_stop()) {
1212                 struct task_struct *tsk = NULL;
1213                 struct task_struct *next = NULL;
1214
1215                 if (cpu_is_offline(cpu))
1216                         goto wait_to_die;
1217
1218                 /* grab task list */
1219                 raw_local_irq_disable();
1220                 tsk = per_cpu(posix_timer_tasklist, cpu);
1221                 per_cpu(posix_timer_tasklist, cpu) = NULL;
1222                 raw_local_irq_enable();
1223
1224                 /* its possible the list is empty, just return */
1225                 if (!tsk) {
1226                         set_current_state(TASK_INTERRUPTIBLE);
1227                         schedule();
1228                         __set_current_state(TASK_RUNNING);
1229                         continue;
1230                 }
1231
1232                 /* Process task list */
1233                 while (1) {
1234                         /* save next */
1235                         next = tsk->posix_timer_list;
1236
1237                         /* run the task timers, clear its ptr and
1238                          * unreference it
1239                          */
1240                         __run_posix_cpu_timers(tsk);
1241                         tsk->posix_timer_list = NULL;
1242                         put_task_struct(tsk);
1243
1244                         /* check if this is the last on the list */
1245                         if (next == tsk)
1246                                 break;
1247                         tsk = next;
1248                 }
1249         }
1250         return 0;
1251
1252 wait_to_die:
1253         /* Wait for kthread_stop */
1254         set_current_state(TASK_INTERRUPTIBLE);
1255         while (!kthread_should_stop()) {
1256                 schedule();
1257                 set_current_state(TASK_INTERRUPTIBLE);
1258         }
1259         __set_current_state(TASK_RUNNING);
1260         return 0;
1261 }
1262
1263 static inline int __fastpath_timer_check(struct task_struct *tsk)
1264 {
1265         /* tsk == current, ensure it is safe to use ->signal/sighand */
1266         if (unlikely(tsk->exit_state))
1267                 return 0;
1268
1269         if (!task_cputime_zero(&tsk->cputime_expires))
1270                         return 1;
1271
1272         if (!task_cputime_zero(&tsk->signal->cputime_expires))
1273                         return 1;
1274
1275         return 0;
1276 }
1277
1278 void run_posix_cpu_timers(struct task_struct *tsk)
1279 {
1280         unsigned long cpu = smp_processor_id();
1281         struct task_struct *tasklist;
1282
1283         BUG_ON(!irqs_disabled());
1284         if(!per_cpu(posix_timer_task, cpu))
1285                 return;
1286         /* get per-cpu references */
1287         tasklist = per_cpu(posix_timer_tasklist, cpu);
1288
1289         /* check to see if we're already queued */
1290         if (!tsk->posix_timer_list && __fastpath_timer_check(tsk)) {
1291                 get_task_struct(tsk);
1292                 if (tasklist) {
1293                         tsk->posix_timer_list = tasklist;
1294                 } else {
1295                         /*
1296                          * The list is terminated by a self-pointing
1297                          * task_struct
1298                          */
1299                         tsk->posix_timer_list = tsk;
1300                 }
1301                 per_cpu(posix_timer_tasklist, cpu) = tsk;
1302
1303                 wake_up_process(per_cpu(posix_timer_task, cpu));
1304         }
1305 }
1306
1307 /*
1308  * posix_cpu_thread_call - callback that gets triggered when a CPU is added.
1309  * Here we can start up the necessary migration thread for the new CPU.
1310  */
1311 static int posix_cpu_thread_call(struct notifier_block *nfb,
1312                                  unsigned long action, void *hcpu)
1313 {
1314         int cpu = (long)hcpu;
1315         struct task_struct *p;
1316         struct sched_param param;
1317
1318         switch (action) {
1319         case CPU_UP_PREPARE:
1320                 p = kthread_create(posix_cpu_timers_thread, hcpu,
1321                                         "posixcputmr/%d",cpu);
1322                 if (IS_ERR(p))
1323                         return NOTIFY_BAD;
1324                 p->flags |= PF_NOFREEZE;
1325                 kthread_bind(p, cpu);
1326                 /* Must be high prio to avoid getting starved */
1327                 param.sched_priority = MAX_RT_PRIO-1;
1328                 sched_setscheduler(p, SCHED_FIFO, &param);
1329                 per_cpu(posix_timer_task,cpu) = p;
1330                 break;
1331         case CPU_ONLINE:
1332                 /* Strictly unneccessary, as first user will wake it. */
1333                 wake_up_process(per_cpu(posix_timer_task,cpu));
1334                 break;
1335 #ifdef CONFIG_HOTPLUG_CPU
1336         case CPU_UP_CANCELED:
1337                 /* Unbind it from offline cpu so it can run.  Fall thru. */
1338                 kthread_bind(per_cpu(posix_timer_task, cpu),
1339                              cpumask_any(cpu_online_mask));
1340                 kthread_stop(per_cpu(posix_timer_task,cpu));
1341                 per_cpu(posix_timer_task,cpu) = NULL;
1342                 break;
1343         case CPU_DEAD:
1344                 kthread_stop(per_cpu(posix_timer_task,cpu));
1345                 per_cpu(posix_timer_task,cpu) = NULL;
1346                 break;
1347 #endif
1348         }
1349         return NOTIFY_OK;
1350 }
1351
1352 /* Register at highest priority so that task migration (migrate_all_tasks)
1353  * happens before everything else.
1354  */
1355 static struct notifier_block posix_cpu_thread_notifier = {
1356         .notifier_call = posix_cpu_thread_call,
1357         .priority = 10
1358 };
1359
1360 static int __init posix_cpu_thread_init(void)
1361 {
1362         void *hcpu = (void *)(long)smp_processor_id();
1363         /* Start one for boot CPU. */
1364         unsigned long cpu;
1365
1366         /* init the per-cpu posix_timer_tasklets */
1367         for_each_possible_cpu(cpu)
1368                 per_cpu(posix_timer_tasklist, cpu) = NULL;
1369
1370         posix_cpu_thread_call(&posix_cpu_thread_notifier, CPU_UP_PREPARE, hcpu);
1371         posix_cpu_thread_call(&posix_cpu_thread_notifier, CPU_ONLINE, hcpu);
1372         register_cpu_notifier(&posix_cpu_thread_notifier);
1373         return 0;
1374 }
1375 early_initcall(posix_cpu_thread_init);
1376 #else /* CONFIG_PREEMPT_RT_BASE */
1377 void run_posix_cpu_timers(struct task_struct *tsk)
1378 {
1379         __run_posix_cpu_timers(tsk);
1380 }
1381 #endif /* CONFIG_PREEMPT_RT_BASE */
1382
1383 /*
1384  * Set one of the process-wide special case CPU timers or RLIMIT_CPU.
1385  * The tsk->sighand->siglock must be held by the caller.
1386  */
1387 void set_process_cpu_timer(struct task_struct *tsk, unsigned int clock_idx,
1388                            cputime_t *newval, cputime_t *oldval)
1389 {
1390         unsigned long long now;
1391
1392         WARN_ON_ONCE(clock_idx == CPUCLOCK_SCHED);
1393         cpu_timer_sample_group(clock_idx, tsk, &now);
1394
1395         if (oldval) {
1396                 /*
1397                  * We are setting itimer. The *oldval is absolute and we update
1398                  * it to be relative, *newval argument is relative and we update
1399                  * it to be absolute.
1400                  */
1401                 if (*oldval) {
1402                         if (*oldval <= now) {
1403                                 /* Just about to fire. */
1404                                 *oldval = cputime_one_jiffy;
1405                         } else {
1406                                 *oldval -= now;
1407                         }
1408                 }
1409
1410                 if (!*newval)
1411                         goto out;
1412                 *newval += now;
1413         }
1414
1415         /*
1416          * Update expiration cache if we are the earliest timer, or eventually
1417          * RLIMIT_CPU limit is earlier than prof_exp cpu timer expire.
1418          */
1419         switch (clock_idx) {
1420         case CPUCLOCK_PROF:
1421                 if (expires_gt(tsk->signal->cputime_expires.prof_exp, *newval))
1422                         tsk->signal->cputime_expires.prof_exp = *newval;
1423                 break;
1424         case CPUCLOCK_VIRT:
1425                 if (expires_gt(tsk->signal->cputime_expires.virt_exp, *newval))
1426                         tsk->signal->cputime_expires.virt_exp = *newval;
1427                 break;
1428         }
1429 out:
1430         posix_cpu_timer_kick_nohz();
1431 }
1432
1433 static int do_cpu_nanosleep(const clockid_t which_clock, int flags,
1434                             struct timespec *rqtp, struct itimerspec *it)
1435 {
1436         struct k_itimer timer;
1437         int error;
1438
1439         /*
1440          * Set up a temporary timer and then wait for it to go off.
1441          */
1442         memset(&timer, 0, sizeof timer);
1443         spin_lock_init(&timer.it_lock);
1444         timer.it_clock = which_clock;
1445         timer.it_overrun = -1;
1446         error = posix_cpu_timer_create(&timer);
1447         timer.it_process = current;
1448         if (!error) {
1449                 static struct itimerspec zero_it;
1450
1451                 memset(it, 0, sizeof *it);
1452                 it->it_value = *rqtp;
1453
1454                 spin_lock_irq(&timer.it_lock);
1455                 error = posix_cpu_timer_set(&timer, flags, it, NULL);
1456                 if (error) {
1457                         spin_unlock_irq(&timer.it_lock);
1458                         return error;
1459                 }
1460
1461                 while (!signal_pending(current)) {
1462                         if (timer.it.cpu.expires == 0) {
1463                                 /*
1464                                  * Our timer fired and was reset, below
1465                                  * deletion can not fail.
1466                                  */
1467                                 posix_cpu_timer_del(&timer);
1468                                 spin_unlock_irq(&timer.it_lock);
1469                                 return 0;
1470                         }
1471
1472                         /*
1473                          * Block until cpu_timer_fire (or a signal) wakes us.
1474                          */
1475                         __set_current_state(TASK_INTERRUPTIBLE);
1476                         spin_unlock_irq(&timer.it_lock);
1477                         schedule();
1478                         spin_lock_irq(&timer.it_lock);
1479                 }
1480
1481                 /*
1482                  * We were interrupted by a signal.
1483                  */
1484                 sample_to_timespec(which_clock, timer.it.cpu.expires, rqtp);
1485                 error = posix_cpu_timer_set(&timer, 0, &zero_it, it);
1486                 if (!error) {
1487                         /*
1488                          * Timer is now unarmed, deletion can not fail.
1489                          */
1490                         posix_cpu_timer_del(&timer);
1491                 }
1492                 spin_unlock_irq(&timer.it_lock);
1493
1494                 while (error == TIMER_RETRY) {
1495                         /*
1496                          * We need to handle case when timer was or is in the
1497                          * middle of firing. In other cases we already freed
1498                          * resources.
1499                          */
1500                         spin_lock_irq(&timer.it_lock);
1501                         error = posix_cpu_timer_del(&timer);
1502                         spin_unlock_irq(&timer.it_lock);
1503                 }
1504
1505                 if ((it->it_value.tv_sec | it->it_value.tv_nsec) == 0) {
1506                         /*
1507                          * It actually did fire already.
1508                          */
1509                         return 0;
1510                 }
1511
1512                 error = -ERESTART_RESTARTBLOCK;
1513         }
1514
1515         return error;
1516 }
1517
1518 static long posix_cpu_nsleep_restart(struct restart_block *restart_block);
1519
1520 static int posix_cpu_nsleep(const clockid_t which_clock, int flags,
1521                             struct timespec *rqtp, struct timespec __user *rmtp)
1522 {
1523         struct restart_block *restart_block = &current->restart_block;
1524         struct itimerspec it;
1525         int error;
1526
1527         /*
1528          * Diagnose required errors first.
1529          */
1530         if (CPUCLOCK_PERTHREAD(which_clock) &&
1531             (CPUCLOCK_PID(which_clock) == 0 ||
1532              CPUCLOCK_PID(which_clock) == current->pid))
1533                 return -EINVAL;
1534
1535         error = do_cpu_nanosleep(which_clock, flags, rqtp, &it);
1536
1537         if (error == -ERESTART_RESTARTBLOCK) {
1538
1539                 if (flags & TIMER_ABSTIME)
1540                         return -ERESTARTNOHAND;
1541                 /*
1542                  * Report back to the user the time still remaining.
1543                  */
1544                 if (rmtp && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1545                         return -EFAULT;
1546
1547                 restart_block->fn = posix_cpu_nsleep_restart;
1548                 restart_block->nanosleep.clockid = which_clock;
1549                 restart_block->nanosleep.rmtp = rmtp;
1550                 restart_block->nanosleep.expires = timespec_to_ns(rqtp);
1551         }
1552         return error;
1553 }
1554
1555 static long posix_cpu_nsleep_restart(struct restart_block *restart_block)
1556 {
1557         clockid_t which_clock = restart_block->nanosleep.clockid;
1558         struct timespec t;
1559         struct itimerspec it;
1560         int error;
1561
1562         t = ns_to_timespec(restart_block->nanosleep.expires);
1563
1564         error = do_cpu_nanosleep(which_clock, TIMER_ABSTIME, &t, &it);
1565
1566         if (error == -ERESTART_RESTARTBLOCK) {
1567                 struct timespec __user *rmtp = restart_block->nanosleep.rmtp;
1568                 /*
1569                  * Report back to the user the time still remaining.
1570                  */
1571                 if (rmtp && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1572                         return -EFAULT;
1573
1574                 restart_block->nanosleep.expires = timespec_to_ns(&t);
1575         }
1576         return error;
1577
1578 }
1579
1580 #define PROCESS_CLOCK   MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED)
1581 #define THREAD_CLOCK    MAKE_THREAD_CPUCLOCK(0, CPUCLOCK_SCHED)
1582
1583 static int process_cpu_clock_getres(const clockid_t which_clock,
1584                                     struct timespec *tp)
1585 {
1586         return posix_cpu_clock_getres(PROCESS_CLOCK, tp);
1587 }
1588 static int process_cpu_clock_get(const clockid_t which_clock,
1589                                  struct timespec *tp)
1590 {
1591         return posix_cpu_clock_get(PROCESS_CLOCK, tp);
1592 }
1593 static int process_cpu_timer_create(struct k_itimer *timer)
1594 {
1595         timer->it_clock = PROCESS_CLOCK;
1596         return posix_cpu_timer_create(timer);
1597 }
1598 static int process_cpu_nsleep(const clockid_t which_clock, int flags,
1599                               struct timespec *rqtp,
1600                               struct timespec __user *rmtp)
1601 {
1602         return posix_cpu_nsleep(PROCESS_CLOCK, flags, rqtp, rmtp);
1603 }
1604 static long process_cpu_nsleep_restart(struct restart_block *restart_block)
1605 {
1606         return -EINVAL;
1607 }
1608 static int thread_cpu_clock_getres(const clockid_t which_clock,
1609                                    struct timespec *tp)
1610 {
1611         return posix_cpu_clock_getres(THREAD_CLOCK, tp);
1612 }
1613 static int thread_cpu_clock_get(const clockid_t which_clock,
1614                                 struct timespec *tp)
1615 {
1616         return posix_cpu_clock_get(THREAD_CLOCK, tp);
1617 }
1618 static int thread_cpu_timer_create(struct k_itimer *timer)
1619 {
1620         timer->it_clock = THREAD_CLOCK;
1621         return posix_cpu_timer_create(timer);
1622 }
1623
1624 struct k_clock clock_posix_cpu = {
1625         .clock_getres   = posix_cpu_clock_getres,
1626         .clock_set      = posix_cpu_clock_set,
1627         .clock_get      = posix_cpu_clock_get,
1628         .timer_create   = posix_cpu_timer_create,
1629         .nsleep         = posix_cpu_nsleep,
1630         .nsleep_restart = posix_cpu_nsleep_restart,
1631         .timer_set      = posix_cpu_timer_set,
1632         .timer_del      = posix_cpu_timer_del,
1633         .timer_get      = posix_cpu_timer_get,
1634 };
1635
1636 static __init int init_posix_cpu_timers(void)
1637 {
1638         struct k_clock process = {
1639                 .clock_getres   = process_cpu_clock_getres,
1640                 .clock_get      = process_cpu_clock_get,
1641                 .timer_create   = process_cpu_timer_create,
1642                 .nsleep         = process_cpu_nsleep,
1643                 .nsleep_restart = process_cpu_nsleep_restart,
1644         };
1645         struct k_clock thread = {
1646                 .clock_getres   = thread_cpu_clock_getres,
1647                 .clock_get      = thread_cpu_clock_get,
1648                 .timer_create   = thread_cpu_timer_create,
1649         };
1650         struct timespec ts;
1651
1652         posix_timers_register_clock(CLOCK_PROCESS_CPUTIME_ID, &process);
1653         posix_timers_register_clock(CLOCK_THREAD_CPUTIME_ID, &thread);
1654
1655         cputime_to_timespec(cputime_one_jiffy, &ts);
1656         onecputick = ts.tv_nsec;
1657         WARN_ON(ts.tv_sec != 0);
1658
1659         return 0;
1660 }
1661 __initcall(init_posix_cpu_timers);