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 /*
 201  * Set cputime to sum_cputime if sum_cputime > cputime. Use cmpxchg
 202  * to avoid race conditions with concurrent updates to cputime.
 203  */
 204 static inline void __update_gt_cputime(atomic64_t *cputime, u64 sum_cputime)
 205 {
 206         u64 curr_cputime;
 207 retry:
 208         curr_cputime = atomic64_read(cputime);
 209         if (sum_cputime > curr_cputime) {
 210                 if (atomic64_cmpxchg(cputime, curr_cputime, sum_cputime) != curr_cputime)
 211                         goto retry;
 212         }
 213 }
 214
 215 static void update_gt_cputime(struct task_cputime_atomic *cputime_atomic, struct task_cputime *sum)
 216 {
 217         __update_gt_cputime(&cputime_atomic->utime, sum->utime);
 218         __update_gt_cputime(&cputime_atomic->stime, sum->stime);
 219         __update_gt_cputime(&cputime_atomic->sum_exec_runtime, sum->sum_exec_runtime);
 220 }
 221
 222 /* Sample task_cputime_atomic values in "atomic_timers", store results in "times". */
 223 static inline void sample_cputime_atomic(struct task_cputime *times,
 224                                          struct task_cputime_atomic *atomic_times)
 225 {
 226         times->utime = atomic64_read(&atomic_times->utime);
 227         times->stime = atomic64_read(&atomic_times->stime);
 228         times->sum_exec_runtime = atomic64_read(&atomic_times->sum_exec_runtime);
 229 }
 230
 231 void thread_group_cputimer(struct task_struct *tsk, struct task_cputime *times)
 232 {
 233         struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
 234         struct task_cputime sum;
 235
 236         /* Check if cputimer isn't running. This is accessed without locking. */
 237         if (!READ_ONCE(cputimer->running)) {
 238                 /*
 239                  * The POSIX timer interface allows for absolute time expiry
 240                  * values through the TIMER_ABSTIME flag, therefore we have
 241                  * to synchronize the timer to the clock every time we start it.
 242                  */
 243                 thread_group_cputime(tsk, &sum);
 244                 update_gt_cputime(&cputimer->cputime_atomic, &sum);
 245
 246                 /*
 247                  * We're setting cputimer->running without a lock. Ensure
 248                  * this only gets written to in one operation. We set
 249                  * running after update_gt_cputime() as a small optimization,
 250                  * but barriers are not required because update_gt_cputime()
 251                  * can handle concurrent updates.
 252                  */
 253                 WRITE_ONCE(cputimer->running, true);
 254         }
 255         sample_cputime_atomic(times, &cputimer->cputime_atomic);
 256 }
 257
 258 /*
 259  * Sample a process (thread group) clock for the given group_leader task.
 260  * Must be called with task sighand lock held for safe while_each_thread()
 261  * traversal.
 262  */
 263 static int cpu_clock_sample_group(const clockid_t which_clock,
 264                                   struct task_struct *p,
 265                                   unsigned long long *sample)
 266 {
 267         struct task_cputime cputime;
 268
 269         switch (CPUCLOCK_WHICH(which_clock)) {
 270         default:
 271                 return -EINVAL;
 272         case CPUCLOCK_PROF:
 273                 thread_group_cputime(p, &cputime);
 274                 *sample = cputime_to_expires(cputime.utime + cputime.stime);
 275                 break;
 276         case CPUCLOCK_VIRT:
 277                 thread_group_cputime(p, &cputime);
 278                 *sample = cputime_to_expires(cputime.utime);
 279                 break;
 280         case CPUCLOCK_SCHED:
 281                 thread_group_cputime(p, &cputime);
 282                 *sample = cputime.sum_exec_runtime;
 283                 break;
 284         }
 285         return 0;
 286 }
 287
 288 static int posix_cpu_clock_get_task(struct task_struct *tsk,
 289                                     const clockid_t which_clock,
 290                                     struct timespec *tp)
 291 {
 292         int err = -EINVAL;
 293         unsigned long long rtn;
 294
 295         if (CPUCLOCK_PERTHREAD(which_clock)) {
 296                 if (same_thread_group(tsk, current))
 297                         err = cpu_clock_sample(which_clock, tsk, &rtn);
 298         } else {
 299                 if (tsk == current || thread_group_leader(tsk))
 300                         err = cpu_clock_sample_group(which_clock, tsk, &rtn);
 301         }
 302
 303         if (!err)
 304                 sample_to_timespec(which_clock, rtn, tp);
 305
 306         return err;
 307 }
 308
 309
 310 static int posix_cpu_clock_get(const clockid_t which_clock, struct timespec *tp)
 311 {
 312         const pid_t pid = CPUCLOCK_PID(which_clock);
 313         int err = -EINVAL;
 314
 315         if (pid == 0) {
 316                 /*
 317                  * Special case constant value for our own clocks.
 318                  * We don't have to do any lookup to find ourselves.
 319                  */
 320                 err = posix_cpu_clock_get_task(current, which_clock, tp);
 321         } else {
 322                 /*
 323                  * Find the given PID, and validate that the caller
 324                  * should be able to see it.
 325                  */
 326                 struct task_struct *p;
 327                 rcu_read_lock();
 328                 p = find_task_by_vpid(pid);
 329                 if (p)
 330                         err = posix_cpu_clock_get_task(p, which_clock, tp);
 331                 rcu_read_unlock();
 332         }
 333
 334         return err;
 335 }
 336
 337
 338 /*
 339  * Validate the clockid_t for a new CPU-clock timer, and initialize the timer.
 340  * This is called from sys_timer_create() and do_cpu_nanosleep() with the
 341  * new timer already all-zeros initialized.
 342  */
 343 static int posix_cpu_timer_create(struct k_itimer *new_timer)
 344 {
 345         int ret = 0;
 346         const pid_t pid = CPUCLOCK_PID(new_timer->it_clock);
 347         struct task_struct *p;
 348
 349         if (CPUCLOCK_WHICH(new_timer->it_clock) >= CPUCLOCK_MAX)
 350                 return -EINVAL;
 351
 352         INIT_LIST_HEAD(&new_timer->it.cpu.entry);
 353
 354         rcu_read_lock();
 355         if (CPUCLOCK_PERTHREAD(new_timer->it_clock)) {
 356                 if (pid == 0) {
 357                         p = current;
 358                 } else {
 359                         p = find_task_by_vpid(pid);
 360                         if (p && !same_thread_group(p, current))
 361                                 p = NULL;
 362                 }
 363         } else {
 364                 if (pid == 0) {
 365                         p = current->group_leader;
 366                 } else {
 367                         p = find_task_by_vpid(pid);
 368                         if (p && !has_group_leader_pid(p))
 369                                 p = NULL;
 370                 }
 371         }
 372         new_timer->it.cpu.task = p;
 373         if (p) {
 374                 get_task_struct(p);
 375         } else {
 376                 ret = -EINVAL;
 377         }
 378         rcu_read_unlock();
 379
 380         return ret;
 381 }
 382
 383 /*
 384  * Clean up a CPU-clock timer that is about to be destroyed.
 385  * This is called from timer deletion with the timer already locked.
 386  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 387  * and try again.  (This happens when the timer is in the middle of firing.)
 388  */
 389 static int posix_cpu_timer_del(struct k_itimer *timer)
 390 {
 391         int ret = 0;
 392         unsigned long flags;
 393         struct sighand_struct *sighand;
 394         struct task_struct *p = timer->it.cpu.task;
 395
 396         WARN_ON_ONCE(p == NULL);
 397
 398         /*
 399          * Protect against sighand release/switch in exit/exec and process/
 400          * thread timer list entry concurrent read/writes.
 401          */
 402         sighand = lock_task_sighand(p, &flags);
 403         if (unlikely(sighand == NULL)) {
 404                 /*
 405                  * We raced with the reaping of the task.
 406                  * The deletion should have cleared us off the list.
 407                  */
 408                 WARN_ON_ONCE(!list_empty(&timer->it.cpu.entry));
 409         } else {
 410                 if (timer->it.cpu.firing)
 411                         ret = TIMER_RETRY;
 412                 else
 413                         list_del(&timer->it.cpu.entry);
 414
 415                 unlock_task_sighand(p, &flags);
 416         }
 417
 418         if (!ret)
 419                 put_task_struct(p);
 420
 421         return ret;
 422 }
 423
 424 static void cleanup_timers_list(struct list_head *head)
 425 {
 426         struct cpu_timer_list *timer, *next;
 427
 428         list_for_each_entry_safe(timer, next, head, entry)
 429                 list_del_init(&timer->entry);
 430 }
 431
 432 /*
 433  * Clean out CPU timers still ticking when a thread exited.  The task
 434  * pointer is cleared, and the expiry time is replaced with the residual
 435  * time for later timer_gettime calls to return.
 436  * This must be called with the siglock held.
 437  */
 438 static void cleanup_timers(struct list_head *head)
 439 {
 440         cleanup_timers_list(head);
 441         cleanup_timers_list(++head);
 442         cleanup_timers_list(++head);
 443 }
 444
 445 /*
 446  * These are both called with the siglock held, when the current thread
 447  * is being reaped.  When the final (leader) thread in the group is reaped,
 448  * posix_cpu_timers_exit_group will be called after posix_cpu_timers_exit.
 449  */
 450 void posix_cpu_timers_exit(struct task_struct *tsk)
 451 {
 452         add_device_randomness((const void*) &tsk->se.sum_exec_runtime,
 453                                                 sizeof(unsigned long long));
 454         cleanup_timers(tsk->cpu_timers);
 455
 456 }
 457 void posix_cpu_timers_exit_group(struct task_struct *tsk)
 458 {
 459         cleanup_timers(tsk->signal->cpu_timers);
 460 }
 461
 462 static inline int expires_gt(cputime_t expires, cputime_t new_exp)
 463 {
 464         return expires == 0 || expires > new_exp;
 465 }
 466
 467 /*
 468  * Insert the timer on the appropriate list before any timers that
 469  * expire later.  This must be called with the sighand lock held.
 470  */
 471 static void arm_timer(struct k_itimer *timer)
 472 {
 473         struct task_struct *p = timer->it.cpu.task;
 474         struct list_head *head, *listpos;
 475         struct task_cputime *cputime_expires;
 476         struct cpu_timer_list *const nt = &timer->it.cpu;
 477         struct cpu_timer_list *next;
 478
 479         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 480                 head = p->cpu_timers;
 481                 cputime_expires = &p->cputime_expires;
 482         } else {
 483                 head = p->signal->cpu_timers;
 484                 cputime_expires = &p->signal->cputime_expires;
 485         }
 486         head += CPUCLOCK_WHICH(timer->it_clock);
 487
 488         listpos = head;
 489         list_for_each_entry(next, head, entry) {
 490                 if (nt->expires < next->expires)
 491                         break;
 492                 listpos = &next->entry;
 493         }
 494         list_add(&nt->entry, listpos);
 495
 496         if (listpos == head) {
 497                 unsigned long long exp = nt->expires;
 498
 499                 /*
 500                  * We are the new earliest-expiring POSIX 1.b timer, hence
 501                  * need to update expiration cache. Take into account that
 502                  * for process timers we share expiration cache with itimers
 503                  * and RLIMIT_CPU and for thread timers with RLIMIT_RTTIME.
 504                  */
 505
 506                 switch (CPUCLOCK_WHICH(timer->it_clock)) {
 507                 case CPUCLOCK_PROF:
 508                         if (expires_gt(cputime_expires->prof_exp, expires_to_cputime(exp)))
 509                                 cputime_expires->prof_exp = expires_to_cputime(exp);
 510                         break;
 511                 case CPUCLOCK_VIRT:
 512                         if (expires_gt(cputime_expires->virt_exp, expires_to_cputime(exp)))
 513                                 cputime_expires->virt_exp = expires_to_cputime(exp);
 514                         break;
 515                 case CPUCLOCK_SCHED:
 516                         if (cputime_expires->sched_exp == 0 ||
 517                             cputime_expires->sched_exp > exp)
 518                                 cputime_expires->sched_exp = exp;
 519                         break;
 520                 }
 521         }
 522 }
 523
 524 /*
 525  * The timer is locked, fire it and arrange for its reload.
 526  */
 527 static void cpu_timer_fire(struct k_itimer *timer)
 528 {
 529         if ((timer->it_sigev_notify & ~SIGEV_THREAD_ID) == SIGEV_NONE) {
 530                 /*
 531                  * User don't want any signal.
 532                  */
 533                 timer->it.cpu.expires = 0;
 534         } else if (unlikely(timer->sigq == NULL)) {
 535                 /*
 536                  * This a special case for clock_nanosleep,
 537                  * not a normal timer from sys_timer_create.
 538                  */
 539                 wake_up_process(timer->it_process);
 540                 timer->it.cpu.expires = 0;
 541         } else if (timer->it.cpu.incr == 0) {
 542                 /*
 543                  * One-shot timer.  Clear it as soon as it's fired.
 544                  */
 545                 posix_timer_event(timer, 0);
 546                 timer->it.cpu.expires = 0;
 547         } else if (posix_timer_event(timer, ++timer->it_requeue_pending)) {
 548                 /*
 549                  * The signal did not get queued because the signal
 550                  * was ignored, so we won't get any callback to
 551                  * reload the timer.  But we need to keep it
 552                  * ticking in case the signal is deliverable next time.
 553                  */
 554                 posix_cpu_timer_schedule(timer);
 555         }
 556 }
 557
 558 /*
 559  * Sample a process (thread group) timer for the given group_leader task.
 560  * Must be called with task sighand lock held for safe while_each_thread()
 561  * traversal.
 562  */
 563 static int cpu_timer_sample_group(const clockid_t which_clock,
 564                                   struct task_struct *p,
 565                                   unsigned long long *sample)
 566 {
 567         struct task_cputime cputime;
 568
 569         thread_group_cputimer(p, &cputime);
 570         switch (CPUCLOCK_WHICH(which_clock)) {
 571         default:
 572                 return -EINVAL;
 573         case CPUCLOCK_PROF:
 574                 *sample = cputime_to_expires(cputime.utime + cputime.stime);
 575                 break;
 576         case CPUCLOCK_VIRT:
 577                 *sample = cputime_to_expires(cputime.utime);
 578                 break;
 579         case CPUCLOCK_SCHED:
 580                 *sample = cputime.sum_exec_runtime;
 581                 break;
 582         }
 583         return 0;
 584 }
 585
 586 #ifdef CONFIG_NO_HZ_FULL
 587 static void nohz_kick_work_fn(struct work_struct *work)
 588 {
 589         tick_nohz_full_kick_all();
 590 }
 591
 592 static DECLARE_WORK(nohz_kick_work, nohz_kick_work_fn);
 593
 594 /*
 595  * We need the IPIs to be sent from sane process context.
 596  * The posix cpu timers are always set with irqs disabled.
 597  */
 598 static void posix_cpu_timer_kick_nohz(void)
 599 {
 600         if (context_tracking_is_enabled())
 601                 schedule_work(&nohz_kick_work);
 602 }
 603
 604 bool posix_cpu_timers_can_stop_tick(struct task_struct *tsk)
 605 {
 606         if (!task_cputime_zero(&tsk->cputime_expires))
 607                 return false;
 608
 609         /* Check if cputimer is running. This is accessed without locking. */
 610         if (READ_ONCE(tsk->signal->cputimer.running))
 611                 return false;
 612
 613         return true;
 614 }
 615 #else
 616 static inline void posix_cpu_timer_kick_nohz(void) { }
 617 #endif
 618
 619 /*
 620  * Guts of sys_timer_settime for CPU timers.
 621  * This is called with the timer locked and interrupts disabled.
 622  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 623  * and try again.  (This happens when the timer is in the middle of firing.)
 624  */
 625 static int posix_cpu_timer_set(struct k_itimer *timer, int timer_flags,
 626                                struct itimerspec *new, struct itimerspec *old)
 627 {
 628         unsigned long flags;
 629         struct sighand_struct *sighand;
 630         struct task_struct *p = timer->it.cpu.task;
 631         unsigned long long old_expires, new_expires, old_incr, val;
 632         int ret;
 633
 634         WARN_ON_ONCE(p == NULL);
 635
 636         new_expires = timespec_to_sample(timer->it_clock, &new->it_value);
 637
 638         /*
 639          * Protect against sighand release/switch in exit/exec and p->cpu_timers
 640          * and p->signal->cpu_timers read/write in arm_timer()
 641          */
 642         sighand = lock_task_sighand(p, &flags);
 643         /*
 644          * If p has just been reaped, we can no
 645          * longer get any information about it at all.
 646          */
 647         if (unlikely(sighand == NULL)) {
 648                 return -ESRCH;
 649         }
 650
 651         /*
 652          * Disarm any old timer after extracting its expiry time.
 653          */
 654         WARN_ON_ONCE_NONRT(!irqs_disabled());
 655
 656         ret = 0;
 657         old_incr = timer->it.cpu.incr;
 658         old_expires = timer->it.cpu.expires;
 659         if (unlikely(timer->it.cpu.firing)) {
 660                 timer->it.cpu.firing = -1;
 661                 ret = TIMER_RETRY;
 662         } else
 663                 list_del_init(&timer->it.cpu.entry);
 664
 665         /*
 666          * We need to sample the current value to convert the new
 667          * value from to relative and absolute, and to convert the
 668          * old value from absolute to relative.  To set a process
 669          * timer, we need a sample to balance the thread expiry
 670          * times (in arm_timer).  With an absolute time, we must
 671          * check if it's already passed.  In short, we need a sample.
 672          */
 673         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 674                 cpu_clock_sample(timer->it_clock, p, &val);
 675         } else {
 676                 cpu_timer_sample_group(timer->it_clock, p, &val);
 677         }
 678
 679         if (old) {
 680                 if (old_expires == 0) {
 681                         old->it_value.tv_sec = 0;
 682                         old->it_value.tv_nsec = 0;
 683                 } else {
 684                         /*
 685                          * Update the timer in case it has
 686                          * overrun already.  If it has,
 687                          * we'll report it as having overrun
 688                          * and with the next reloaded timer
 689                          * already ticking, though we are
 690                          * swallowing that pending
 691                          * notification here to install the
 692                          * new setting.
 693                          */
 694                         bump_cpu_timer(timer, val);
 695                         if (val < timer->it.cpu.expires) {
 696                                 old_expires = timer->it.cpu.expires - val;
 697                                 sample_to_timespec(timer->it_clock,
 698                                                    old_expires,
 699                                                    &old->it_value);
 700                         } else {
 701                                 old->it_value.tv_nsec = 1;
 702                                 old->it_value.tv_sec = 0;
 703                         }
 704                 }
 705         }
 706
 707         if (unlikely(ret)) {
 708                 /*
 709                  * We are colliding with the timer actually firing.
 710                  * Punt after filling in the timer's old value, and
 711                  * disable this firing since we are already reporting
 712                  * it as an overrun (thanks to bump_cpu_timer above).
 713                  */
 714                 unlock_task_sighand(p, &flags);
 715                 goto out;
 716         }
 717
 718         if (new_expires != 0 && !(timer_flags & TIMER_ABSTIME)) {
 719                 new_expires += val;
 720         }
 721
 722         /*
 723          * Install the new expiry time (or zero).
 724          * For a timer with no notification action, we don't actually
 725          * arm the timer (we'll just fake it for timer_gettime).
 726          */
 727         timer->it.cpu.expires = new_expires;
 728         if (new_expires != 0 && val < new_expires) {
 729                 arm_timer(timer);
 730         }
 731
 732         unlock_task_sighand(p, &flags);
 733         /*
 734          * Install the new reload setting, and
 735          * set up the signal and overrun bookkeeping.
 736          */
 737         timer->it.cpu.incr = timespec_to_sample(timer->it_clock,
 738                                                 &new->it_interval);
 739
 740         /*
 741          * This acts as a modification timestamp for the timer,
 742          * so any automatic reload attempt will punt on seeing
 743          * that we have reset the timer manually.
 744          */
 745         timer->it_requeue_pending = (timer->it_requeue_pending + 2) &
 746                 ~REQUEUE_PENDING;
 747         timer->it_overrun_last = 0;
 748         timer->it_overrun = -1;
 749
 750         if (new_expires != 0 && !(val < new_expires)) {
 751                 /*
 752                  * The designated time already passed, so we notify
 753                  * immediately, even if the thread never runs to
 754                  * accumulate more time on this clock.
 755                  */
 756                 cpu_timer_fire(timer);
 757         }
 758
 759         ret = 0;
 760  out:
 761         if (old) {
 762                 sample_to_timespec(timer->it_clock,
 763                                    old_incr, &old->it_interval);
 764         }
 765         if (!ret)
 766                 posix_cpu_timer_kick_nohz();
 767         return ret;
 768 }
 769
 770 static void posix_cpu_timer_get(struct k_itimer *timer, struct itimerspec *itp)
 771 {
 772         unsigned long long now;
 773         struct task_struct *p = timer->it.cpu.task;
 774
 775         WARN_ON_ONCE(p == NULL);
 776
 777         /*
 778          * Easy part: convert the reload time.
 779          */
 780         sample_to_timespec(timer->it_clock,
 781                            timer->it.cpu.incr, &itp->it_interval);
 782
 783         if (timer->it.cpu.expires == 0) {       /* Timer not armed at all.  */
 784                 itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
 785                 return;
 786         }
 787
 788         /*
 789          * Sample the clock to take the difference with the expiry time.
 790          */
 791         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 792                 cpu_clock_sample(timer->it_clock, p, &now);
 793         } else {
 794                 struct sighand_struct *sighand;
 795                 unsigned long flags;
 796
 797                 /*
 798                  * Protect against sighand release/switch in exit/exec and
 799                  * also make timer sampling safe if it ends up calling
 800                  * thread_group_cputime().
 801                  */
 802                 sighand = lock_task_sighand(p, &flags);
 803                 if (unlikely(sighand == NULL)) {
 804                         /*
 805                          * The process has been reaped.
 806                          * We can't even collect a sample any more.
 807                          * Call the timer disarmed, nothing else to do.
 808                          */
 809                         timer->it.cpu.expires = 0;
 810                         sample_to_timespec(timer->it_clock, timer->it.cpu.expires,
 811                                            &itp->it_value);
 812                 } else {
 813                         cpu_timer_sample_group(timer->it_clock, p, &now);
 814                         unlock_task_sighand(p, &flags);
 815                 }
 816         }
 817
 818         if (now < timer->it.cpu.expires) {
 819                 sample_to_timespec(timer->it_clock,
 820                                    timer->it.cpu.expires - now,
 821                                    &itp->it_value);
 822         } else {
 823                 /*
 824                  * The timer should have expired already, but the firing
 825                  * hasn't taken place yet.  Say it's just about to expire.
 826                  */
 827                 itp->it_value.tv_nsec = 1;
 828                 itp->it_value.tv_sec = 0;
 829         }
 830 }
 831
 832 static unsigned long long
 833 check_timers_list(struct list_head *timers,
 834                   struct list_head *firing,
 835                   unsigned long long curr)
 836 {
 837         int maxfire = 20;
 838
 839         while (!list_empty(timers)) {
 840                 struct cpu_timer_list *t;
 841
 842                 t = list_first_entry(timers, struct cpu_timer_list, entry);
 843
 844                 if (!--maxfire || curr < t->expires)
 845                         return t->expires;
 846
 847                 t->firing = 1;
 848                 list_move_tail(&t->entry, firing);
 849         }
 850
 851         return 0;
 852 }
 853
 854 /*
 855  * Check for any per-thread CPU timers that have fired and move them off
 856  * the tsk->cpu_timers[N] list onto the firing list.  Here we update the
 857  * tsk->it_*_expires values to reflect the remaining thread CPU timers.
 858  */
 859 static void check_thread_timers(struct task_struct *tsk,
 860                                 struct list_head *firing)
 861 {
 862         struct list_head *timers = tsk->cpu_timers;
 863         struct signal_struct *const sig = tsk->signal;
 864         struct task_cputime *tsk_expires = &tsk->cputime_expires;
 865         unsigned long long expires;
 866         unsigned long soft;
 867
 868         /*
 869          * If cputime_expires is zero, then there are no active
 870          * per thread CPU timers.
 871          */
 872         if (task_cputime_zero(&tsk->cputime_expires))
 873                 return;
 874
 875         expires = check_timers_list(timers, firing, prof_ticks(tsk));
 876         tsk_expires->prof_exp = expires_to_cputime(expires);
 877
 878         expires = check_timers_list(++timers, firing, virt_ticks(tsk));
 879         tsk_expires->virt_exp = expires_to_cputime(expires);
 880
 881         tsk_expires->sched_exp = check_timers_list(++timers, firing,
 882                                                    tsk->se.sum_exec_runtime);
 883
 884         /*
 885          * Check for the special case thread timers.
 886          */
 887         soft = READ_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_cur);
 888         if (soft != RLIM_INFINITY) {
 889                 unsigned long hard =
 890                         READ_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_max);
 891
 892                 if (hard != RLIM_INFINITY &&
 893                     tsk->rt.timeout > DIV_ROUND_UP(hard, USEC_PER_SEC/HZ)) {
 894                         /*
 895                          * At the hard limit, we just die.
 896                          * No need to calculate anything else now.
 897                          */
 898                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
 899                         return;
 900                 }
 901                 if (tsk->rt.timeout > DIV_ROUND_UP(soft, USEC_PER_SEC/HZ)) {
 902                         /*
 903                          * At the soft limit, send a SIGXCPU every second.
 904                          */
 905                         if (soft < hard) {
 906                                 soft += USEC_PER_SEC;
 907                                 sig->rlim[RLIMIT_RTTIME].rlim_cur = soft;
 908                         }
 909                         printk(KERN_INFO
 910                                 "RT Watchdog Timeout: %s[%d]\n",
 911                                 tsk->comm, task_pid_nr(tsk));
 912                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
 913                 }
 914         }
 915 }
 916
 917 static inline void stop_process_timers(struct signal_struct *sig)
 918 {
 919         struct thread_group_cputimer *cputimer = &sig->cputimer;
 920
 921         /* Turn off cputimer->running. This is done without locking. */
 922         WRITE_ONCE(cputimer->running, false);
 923 }
 924
 925 static u32 onecputick;
 926
 927 static void check_cpu_itimer(struct task_struct *tsk, struct cpu_itimer *it,
 928                              unsigned long long *expires,
 929                              unsigned long long cur_time, int signo)
 930 {
 931         if (!it->expires)
 932                 return;
 933
 934         if (cur_time >= it->expires) {
 935                 if (it->incr) {
 936                         it->expires += it->incr;
 937                         it->error += it->incr_error;
 938                         if (it->error >= onecputick) {
 939                                 it->expires -= cputime_one_jiffy;
 940                                 it->error -= onecputick;
 941                         }
 942                 } else {
 943                         it->expires = 0;
 944                 }
 945
 946                 trace_itimer_expire(signo == SIGPROF ?
 947                                     ITIMER_PROF : ITIMER_VIRTUAL,
 948                                     tsk->signal->leader_pid, cur_time);
 949                 __group_send_sig_info(signo, SEND_SIG_PRIV, tsk);
 950         }
 951
 952         if (it->expires && (!*expires || it->expires < *expires)) {
 953                 *expires = it->expires;
 954         }
 955 }
 956
 957 /*
 958  * Check for any per-thread CPU timers that have fired and move them
 959  * off the tsk->*_timers list onto the firing list.  Per-thread timers
 960  * have already been taken off.
 961  */
 962 static void check_process_timers(struct task_struct *tsk,
 963                                  struct list_head *firing)
 964 {
 965         struct signal_struct *const sig = tsk->signal;
 966         unsigned long long utime, ptime, virt_expires, prof_expires;
 967         unsigned long long sum_sched_runtime, sched_expires;
 968         struct list_head *timers = sig->cpu_timers;
 969         struct task_cputime cputime;
 970         unsigned long soft;
 971
 972         /*
 973          * If cputimer is not running, then there are no active
 974          * process wide timers (POSIX 1.b, itimers, RLIMIT_CPU).
 975          */
 976         if (!READ_ONCE(tsk->signal->cputimer.running))
 977                 return;
 978
 979         /*
 980          * Signify that a thread is checking for process timers.
 981          * Write access to this field is protected by the sighand lock.
 982          */
 983         sig->cputimer.checking_timer = true;
 984
 985         /*
 986          * Collect the current process totals.
 987          */
 988         thread_group_cputimer(tsk, &cputime);
 989         utime = cputime_to_expires(cputime.utime);
 990         ptime = utime + cputime_to_expires(cputime.stime);
 991         sum_sched_runtime = cputime.sum_exec_runtime;
 992
 993         prof_expires = check_timers_list(timers, firing, ptime);
 994         virt_expires = check_timers_list(++timers, firing, utime);
 995         sched_expires = check_timers_list(++timers, firing, sum_sched_runtime);
 996
 997         /*
 998          * Check for the special case process timers.
 999          */
1000         check_cpu_itimer(tsk, &sig->it[CPUCLOCK_PROF], &prof_expires, ptime,
1001                          SIGPROF);
1002         check_cpu_itimer(tsk, &sig->it[CPUCLOCK_VIRT], &virt_expires, utime,
1003                          SIGVTALRM);
1004         soft = READ_ONCE(sig->rlim[RLIMIT_CPU].rlim_cur);
1005         if (soft != RLIM_INFINITY) {
1006                 unsigned long psecs = cputime_to_secs(ptime);
1007                 unsigned long hard =
1008                         READ_ONCE(sig->rlim[RLIMIT_CPU].rlim_max);
1009                 cputime_t x;
1010                 if (psecs >= hard) {
1011                         /*
1012                          * At the hard limit, we just die.
1013                          * No need to calculate anything else now.
1014                          */
1015                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
1016                         return;
1017                 }
1018                 if (psecs >= soft) {
1019                         /*
1020                          * At the soft limit, send a SIGXCPU every second.
1021                          */
1022                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
1023                         if (soft < hard) {
1024                                 soft++;
1025                                 sig->rlim[RLIMIT_CPU].rlim_cur = soft;
1026                         }
1027                 }
1028                 x = secs_to_cputime(soft);
1029                 if (!prof_expires || x < prof_expires) {
1030                         prof_expires = x;
1031                 }
1032         }
1033
1034         sig->cputime_expires.prof_exp = expires_to_cputime(prof_expires);
1035         sig->cputime_expires.virt_exp = expires_to_cputime(virt_expires);
1036         sig->cputime_expires.sched_exp = sched_expires;
1037         if (task_cputime_zero(&sig->cputime_expires))
1038                 stop_process_timers(sig);
1039
1040         sig->cputimer.checking_timer = false;
1041 }
1042
1043 /*
1044  * This is called from the signal code (via do_schedule_next_timer)
1045  * when the last timer signal was delivered and we have to reload the timer.
1046  */
1047 void posix_cpu_timer_schedule(struct k_itimer *timer)
1048 {
1049         struct sighand_struct *sighand;
1050         unsigned long flags;
1051         struct task_struct *p = timer->it.cpu.task;
1052         unsigned long long now;
1053
1054         WARN_ON_ONCE(p == NULL);
1055
1056         /*
1057          * Fetch the current sample and update the timer's expiry time.
1058          */
1059         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
1060                 cpu_clock_sample(timer->it_clock, p, &now);
1061                 bump_cpu_timer(timer, now);
1062                 if (unlikely(p->exit_state))
1063                         goto out;
1064
1065                 /* Protect timer list r/w in arm_timer() */
1066                 sighand = lock_task_sighand(p, &flags);
1067                 if (!sighand)
1068                         goto out;
1069         } else {
1070                 /*
1071                  * Protect arm_timer() and timer sampling in case of call to
1072                  * thread_group_cputime().
1073                  */
1074                 sighand = lock_task_sighand(p, &flags);
1075                 if (unlikely(sighand == NULL)) {
1076                         /*
1077                          * The process has been reaped.
1078                          * We can't even collect a sample any more.
1079                          */
1080                         timer->it.cpu.expires = 0;
1081                         goto out;
1082                 } else if (unlikely(p->exit_state) && thread_group_empty(p)) {
1083                         unlock_task_sighand(p, &flags);
1084                         /* Optimizations: if the process is dying, no need to rearm */
1085                         goto out;
1086                 }
1087                 cpu_timer_sample_group(timer->it_clock, p, &now);
1088                 bump_cpu_timer(timer, now);
1089                 /* Leave the sighand locked for the call below.  */
1090         }
1091
1092         /*
1093          * Now re-arm for the new expiry time.
1094          */
1095         WARN_ON_ONCE_NONRT(!irqs_disabled());
1096         arm_timer(timer);
1097         unlock_task_sighand(p, &flags);
1098
1099         /* Kick full dynticks CPUs in case they need to tick on the new timer */
1100         posix_cpu_timer_kick_nohz();
1101 out:
1102         timer->it_overrun_last = timer->it_overrun;
1103         timer->it_overrun = -1;
1104         ++timer->it_requeue_pending;
1105 }
1106
1107 /**
1108  * task_cputime_expired - Compare two task_cputime entities.
1109  *
1110  * @sample:     The task_cputime structure to be checked for expiration.
1111  * @expires:    Expiration times, against which @sample will be checked.
1112  *
1113  * Checks @sample against @expires to see if any field of @sample has expired.
1114  * Returns true if any field of the former is greater than the corresponding
1115  * field of the latter if the latter field is set.  Otherwise returns false.
1116  */
1117 static inline int task_cputime_expired(const struct task_cputime *sample,
1118                                         const struct task_cputime *expires)
1119 {
1120         if (expires->utime && sample->utime >= expires->utime)
1121                 return 1;
1122         if (expires->stime && sample->utime + sample->stime >= expires->stime)
1123                 return 1;
1124         if (expires->sum_exec_runtime != 0 &&
1125             sample->sum_exec_runtime >= expires->sum_exec_runtime)
1126                 return 1;
1127         return 0;
1128 }
1129
1130 /**
1131  * fastpath_timer_check - POSIX CPU timers fast path.
1132  *
1133  * @tsk:        The task (thread) being checked.
1134  *
1135  * Check the task and thread group timers.  If both are zero (there are no
1136  * timers set) return false.  Otherwise snapshot the task and thread group
1137  * timers and compare them with the corresponding expiration times.  Return
1138  * true if a timer has expired, else return false.
1139  */
1140 static inline int fastpath_timer_check(struct task_struct *tsk)
1141 {
1142         struct signal_struct *sig;
1143
1144         if (!task_cputime_zero(&tsk->cputime_expires)) {
1145                 struct task_cputime task_sample;
1146
1147                 task_cputime(tsk, &task_sample.utime, &task_sample.stime);
1148                 task_sample.sum_exec_runtime = tsk->se.sum_exec_runtime;
1149                 if (task_cputime_expired(&task_sample, &tsk->cputime_expires))
1150                         return 1;
1151         }
1152
1153         sig = tsk->signal;
1154         /*
1155          * Check if thread group timers expired when the cputimer is
1156          * running and no other thread in the group is already checking
1157          * for thread group cputimers. These fields are read without the
1158          * sighand lock. However, this is fine because this is meant to
1159          * be a fastpath heuristic to determine whether we should try to
1160          * acquire the sighand lock to check/handle timers.
1161          *
1162          * In the worst case scenario, if 'running' or 'checking_timer' gets
1163          * set but the current thread doesn't see the change yet, we'll wait
1164          * until the next thread in the group gets a scheduler interrupt to
1165          * handle the timer. This isn't an issue in practice because these
1166          * types of delays with signals actually getting sent are expected.
1167          */
1168         if (READ_ONCE(sig->cputimer.running) &&
1169             !READ_ONCE(sig->cputimer.checking_timer)) {
1170                 struct task_cputime group_sample;
1171
1172                 sample_cputime_atomic(&group_sample, &sig->cputimer.cputime_atomic);
1173
1174                 if (task_cputime_expired(&group_sample, &sig->cputime_expires))
1175                         return 1;
1176         }
1177
1178         return 0;
1179 }
1180
1181 /*
1182  * This is called from the timer interrupt handler.  The irq handler has
1183  * already updated our counts.  We need to check if any timers fire now.
1184  * Interrupts are disabled.
1185  */
1186 static void __run_posix_cpu_timers(struct task_struct *tsk)
1187 {
1188         LIST_HEAD(firing);
1189         struct k_itimer *timer, *next;
1190         unsigned long flags;
1191
1192         WARN_ON_ONCE_NONRT(!irqs_disabled());
1193
1194         /*
1195          * The fast path checks that there are no expired thread or thread
1196          * group timers.  If that's so, just return.
1197          */
1198         if (!fastpath_timer_check(tsk))
1199                 return;
1200
1201         if (!lock_task_sighand(tsk, &flags))
1202                 return;
1203         /*
1204          * Here we take off tsk->signal->cpu_timers[N] and
1205          * tsk->cpu_timers[N] all the timers that are firing, and
1206          * put them on the firing list.
1207          */
1208         check_thread_timers(tsk, &firing);
1209
1210         check_process_timers(tsk, &firing);
1211
1212         /*
1213          * We must release these locks before taking any timer's lock.
1214          * There is a potential race with timer deletion here, as the
1215          * siglock now protects our private firing list.  We have set
1216          * the firing flag in each timer, so that a deletion attempt
1217          * that gets the timer lock before we do will give it up and
1218          * spin until we've taken care of that timer below.
1219          */
1220         unlock_task_sighand(tsk, &flags);
1221
1222         /*
1223          * Now that all the timers on our list have the firing flag,
1224          * no one will touch their list entries but us.  We'll take
1225          * each timer's lock before clearing its firing flag, so no
1226          * timer call will interfere.
1227          */
1228         list_for_each_entry_safe(timer, next, &firing, it.cpu.entry) {
1229                 int cpu_firing;
1230
1231                 spin_lock(&timer->it_lock);
1232                 list_del_init(&timer->it.cpu.entry);
1233                 cpu_firing = timer->it.cpu.firing;
1234                 timer->it.cpu.firing = 0;
1235                 /*
1236                  * The firing flag is -1 if we collided with a reset
1237                  * of the timer, which already reported this
1238                  * almost-firing as an overrun.  So don't generate an event.
1239                  */
1240                 if (likely(cpu_firing >= 0))
1241                         cpu_timer_fire(timer);
1242                 spin_unlock(&timer->it_lock);
1243         }
1244 }
1245
1246 #ifdef CONFIG_PREEMPT_RT_BASE
1247 #include <linux/kthread.h>
1248 #include <linux/cpu.h>
1249 DEFINE_PER_CPU(struct task_struct *, posix_timer_task);
1250 DEFINE_PER_CPU(struct task_struct *, posix_timer_tasklist);
1251
1252 static int posix_cpu_timers_thread(void *data)
1253 {
1254         int cpu = (long)data;
1255
1256         BUG_ON(per_cpu(posix_timer_task,cpu) != current);
1257
1258         while (!kthread_should_stop()) {
1259                 struct task_struct *tsk = NULL;
1260                 struct task_struct *next = NULL;
1261
1262                 if (cpu_is_offline(cpu))
1263                         goto wait_to_die;
1264
1265                 /* grab task list */
1266                 raw_local_irq_disable();
1267                 tsk = per_cpu(posix_timer_tasklist, cpu);
1268                 per_cpu(posix_timer_tasklist, cpu) = NULL;
1269                 raw_local_irq_enable();
1270
1271                 /* its possible the list is empty, just return */
1272                 if (!tsk) {
1273                         set_current_state(TASK_INTERRUPTIBLE);
1274                         schedule();
1275                         __set_current_state(TASK_RUNNING);
1276                         continue;
1277                 }
1278
1279                 /* Process task list */
1280                 while (1) {
1281                         /* save next */
1282                         next = tsk->posix_timer_list;
1283
1284                         /* run the task timers, clear its ptr and
1285                          * unreference it
1286                          */
1287                         __run_posix_cpu_timers(tsk);
1288                         tsk->posix_timer_list = NULL;
1289                         put_task_struct(tsk);
1290
1291                         /* check if this is the last on the list */
1292                         if (next == tsk)
1293                                 break;
1294                         tsk = next;
1295                 }
1296         }
1297         return 0;
1298
1299 wait_to_die:
1300         /* Wait for kthread_stop */
1301         set_current_state(TASK_INTERRUPTIBLE);
1302         while (!kthread_should_stop()) {
1303                 schedule();
1304                 set_current_state(TASK_INTERRUPTIBLE);
1305         }
1306         __set_current_state(TASK_RUNNING);
1307         return 0;
1308 }
1309
1310 static inline int __fastpath_timer_check(struct task_struct *tsk)
1311 {
1312         /* tsk == current, ensure it is safe to use ->signal/sighand */
1313         if (unlikely(tsk->exit_state))
1314                 return 0;
1315
1316         if (!task_cputime_zero(&tsk->cputime_expires))
1317                         return 1;
1318
1319         if (!task_cputime_zero(&tsk->signal->cputime_expires))
1320                         return 1;
1321
1322         return 0;
1323 }
1324
1325 void run_posix_cpu_timers(struct task_struct *tsk)
1326 {
1327         unsigned long cpu = smp_processor_id();
1328         struct task_struct *tasklist;
1329
1330         BUG_ON(!irqs_disabled());
1331         if(!per_cpu(posix_timer_task, cpu))
1332                 return;
1333         /* get per-cpu references */
1334         tasklist = per_cpu(posix_timer_tasklist, cpu);
1335
1336         /* check to see if we're already queued */
1337         if (!tsk->posix_timer_list && __fastpath_timer_check(tsk)) {
1338                 get_task_struct(tsk);
1339                 if (tasklist) {
1340                         tsk->posix_timer_list = tasklist;
1341                 } else {
1342                         /*
1343                          * The list is terminated by a self-pointing
1344                          * task_struct
1345                          */
1346                         tsk->posix_timer_list = tsk;
1347                 }
1348                 per_cpu(posix_timer_tasklist, cpu) = tsk;
1349
1350                 wake_up_process(per_cpu(posix_timer_task, cpu));
1351         }
1352 }
1353
1354 /*
1355  * posix_cpu_thread_call - callback that gets triggered when a CPU is added.
1356  * Here we can start up the necessary migration thread for the new CPU.
1357  */
1358 static int posix_cpu_thread_call(struct notifier_block *nfb,
1359                                  unsigned long action, void *hcpu)
1360 {
1361         int cpu = (long)hcpu;
1362         struct task_struct *p;
1363         struct sched_param param;
1364
1365         switch (action) {
1366         case CPU_UP_PREPARE:
1367                 p = kthread_create(posix_cpu_timers_thread, hcpu,
1368                                         "posixcputmr/%d",cpu);
1369                 if (IS_ERR(p))
1370                         return NOTIFY_BAD;
1371                 p->flags |= PF_NOFREEZE;
1372                 kthread_bind(p, cpu);
1373                 /* Must be high prio to avoid getting starved */
1374                 param.sched_priority = MAX_RT_PRIO-1;
1375                 sched_setscheduler(p, SCHED_FIFO, &param);
1376                 per_cpu(posix_timer_task,cpu) = p;
1377                 break;
1378         case CPU_ONLINE:
1379                 /* Strictly unneccessary, as first user will wake it. */
1380                 wake_up_process(per_cpu(posix_timer_task,cpu));
1381                 break;
1382 #ifdef CONFIG_HOTPLUG_CPU
1383         case CPU_UP_CANCELED:
1384                 /* Unbind it from offline cpu so it can run.  Fall thru. */
1385                 kthread_bind(per_cpu(posix_timer_task, cpu),
1386                              cpumask_any(cpu_online_mask));
1387                 kthread_stop(per_cpu(posix_timer_task,cpu));
1388                 per_cpu(posix_timer_task,cpu) = NULL;
1389                 break;
1390         case CPU_DEAD:
1391                 kthread_stop(per_cpu(posix_timer_task,cpu));
1392                 per_cpu(posix_timer_task,cpu) = NULL;
1393                 break;
1394 #endif
1395         }
1396         return NOTIFY_OK;
1397 }
1398
1399 /* Register at highest priority so that task migration (migrate_all_tasks)
1400  * happens before everything else.
1401  */
1402 static struct notifier_block posix_cpu_thread_notifier = {
1403         .notifier_call = posix_cpu_thread_call,
1404         .priority = 10
1405 };
1406
1407 static int __init posix_cpu_thread_init(void)
1408 {
1409         void *hcpu = (void *)(long)smp_processor_id();
1410         /* Start one for boot CPU. */
1411         unsigned long cpu;
1412
1413         /* init the per-cpu posix_timer_tasklets */
1414         for_each_possible_cpu(cpu)
1415                 per_cpu(posix_timer_tasklist, cpu) = NULL;
1416
1417         posix_cpu_thread_call(&posix_cpu_thread_notifier, CPU_UP_PREPARE, hcpu);
1418         posix_cpu_thread_call(&posix_cpu_thread_notifier, CPU_ONLINE, hcpu);
1419         register_cpu_notifier(&posix_cpu_thread_notifier);
1420         return 0;
1421 }
1422 early_initcall(posix_cpu_thread_init);
1423 #else /* CONFIG_PREEMPT_RT_BASE */
1424 void run_posix_cpu_timers(struct task_struct *tsk)
1425 {
1426         __run_posix_cpu_timers(tsk);
1427 }
1428 #endif /* CONFIG_PREEMPT_RT_BASE */
1429
1430 /*
1431  * Set one of the process-wide special case CPU timers or RLIMIT_CPU.
1432  * The tsk->sighand->siglock must be held by the caller.
1433  */
1434 void set_process_cpu_timer(struct task_struct *tsk, unsigned int clock_idx,
1435                            cputime_t *newval, cputime_t *oldval)
1436 {
1437         unsigned long long now;
1438
1439         WARN_ON_ONCE(clock_idx == CPUCLOCK_SCHED);
1440         cpu_timer_sample_group(clock_idx, tsk, &now);
1441
1442         if (oldval) {
1443                 /*
1444                  * We are setting itimer. The *oldval is absolute and we update
1445                  * it to be relative, *newval argument is relative and we update
1446                  * it to be absolute.
1447                  */
1448                 if (*oldval) {
1449                         if (*oldval <= now) {
1450                                 /* Just about to fire. */
1451                                 *oldval = cputime_one_jiffy;
1452                         } else {
1453                                 *oldval -= now;
1454                         }
1455                 }
1456
1457                 if (!*newval)
1458                         goto out;
1459                 *newval += now;
1460         }
1461
1462         /*
1463          * Update expiration cache if we are the earliest timer, or eventually
1464          * RLIMIT_CPU limit is earlier than prof_exp cpu timer expire.
1465          */
1466         switch (clock_idx) {
1467         case CPUCLOCK_PROF:
1468                 if (expires_gt(tsk->signal->cputime_expires.prof_exp, *newval))
1469                         tsk->signal->cputime_expires.prof_exp = *newval;
1470                 break;
1471         case CPUCLOCK_VIRT:
1472                 if (expires_gt(tsk->signal->cputime_expires.virt_exp, *newval))
1473                         tsk->signal->cputime_expires.virt_exp = *newval;
1474                 break;
1475         }
1476 out:
1477         posix_cpu_timer_kick_nohz();
1478 }
1479
1480 static int do_cpu_nanosleep(const clockid_t which_clock, int flags,
1481                             struct timespec *rqtp, struct itimerspec *it)
1482 {
1483         struct k_itimer timer;
1484         int error;
1485
1486         /*
1487          * Set up a temporary timer and then wait for it to go off.
1488          */
1489         memset(&timer, 0, sizeof timer);
1490         spin_lock_init(&timer.it_lock);
1491         timer.it_clock = which_clock;
1492         timer.it_overrun = -1;
1493         error = posix_cpu_timer_create(&timer);
1494         timer.it_process = current;
1495         if (!error) {
1496                 static struct itimerspec zero_it;
1497
1498                 memset(it, 0, sizeof *it);
1499                 it->it_value = *rqtp;
1500
1501                 spin_lock_irq(&timer.it_lock);
1502                 error = posix_cpu_timer_set(&timer, flags, it, NULL);
1503                 if (error) {
1504                         spin_unlock_irq(&timer.it_lock);
1505                         return error;
1506                 }
1507
1508                 while (!signal_pending(current)) {
1509                         if (timer.it.cpu.expires == 0) {
1510                                 /*
1511                                  * Our timer fired and was reset, below
1512                                  * deletion can not fail.
1513                                  */
1514                                 posix_cpu_timer_del(&timer);
1515                                 spin_unlock_irq(&timer.it_lock);
1516                                 return 0;
1517                         }
1518
1519                         /*
1520                          * Block until cpu_timer_fire (or a signal) wakes us.
1521                          */
1522                         __set_current_state(TASK_INTERRUPTIBLE);
1523                         spin_unlock_irq(&timer.it_lock);
1524                         schedule();
1525                         spin_lock_irq(&timer.it_lock);
1526                 }
1527
1528                 /*
1529                  * We were interrupted by a signal.
1530                  */
1531                 sample_to_timespec(which_clock, timer.it.cpu.expires, rqtp);
1532                 error = posix_cpu_timer_set(&timer, 0, &zero_it, it);
1533                 if (!error) {
1534                         /*
1535                          * Timer is now unarmed, deletion can not fail.
1536                          */
1537                         posix_cpu_timer_del(&timer);
1538                 }
1539                 spin_unlock_irq(&timer.it_lock);
1540
1541                 while (error == TIMER_RETRY) {
1542                         /*
1543                          * We need to handle case when timer was or is in the
1544                          * middle of firing. In other cases we already freed
1545                          * resources.
1546                          */
1547                         spin_lock_irq(&timer.it_lock);
1548                         error = posix_cpu_timer_del(&timer);
1549                         spin_unlock_irq(&timer.it_lock);
1550                 }
1551
1552                 if ((it->it_value.tv_sec | it->it_value.tv_nsec) == 0) {
1553                         /*
1554                          * It actually did fire already.
1555                          */
1556                         return 0;
1557                 }
1558
1559                 error = -ERESTART_RESTARTBLOCK;
1560         }
1561
1562         return error;
1563 }
1564
1565 static long posix_cpu_nsleep_restart(struct restart_block *restart_block);
1566
1567 static int posix_cpu_nsleep(const clockid_t which_clock, int flags,
1568                             struct timespec *rqtp, struct timespec __user *rmtp)
1569 {
1570         struct restart_block *restart_block = &current->restart_block;
1571         struct itimerspec it;
1572         int error;
1573
1574         /*
1575          * Diagnose required errors first.
1576          */
1577         if (CPUCLOCK_PERTHREAD(which_clock) &&
1578             (CPUCLOCK_PID(which_clock) == 0 ||
1579              CPUCLOCK_PID(which_clock) == current->pid))
1580                 return -EINVAL;
1581
1582         error = do_cpu_nanosleep(which_clock, flags, rqtp, &it);
1583
1584         if (error == -ERESTART_RESTARTBLOCK) {
1585
1586                 if (flags & TIMER_ABSTIME)
1587                         return -ERESTARTNOHAND;
1588                 /*
1589                  * Report back to the user the time still remaining.
1590                  */
1591                 if (rmtp && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1592                         return -EFAULT;
1593
1594                 restart_block->fn = posix_cpu_nsleep_restart;
1595                 restart_block->nanosleep.clockid = which_clock;
1596                 restart_block->nanosleep.rmtp = rmtp;
1597                 restart_block->nanosleep.expires = timespec_to_ns(rqtp);
1598         }
1599         return error;
1600 }
1601
1602 static long posix_cpu_nsleep_restart(struct restart_block *restart_block)
1603 {
1604         clockid_t which_clock = restart_block->nanosleep.clockid;
1605         struct timespec t;
1606         struct itimerspec it;
1607         int error;
1608
1609         t = ns_to_timespec(restart_block->nanosleep.expires);
1610
1611         error = do_cpu_nanosleep(which_clock, TIMER_ABSTIME, &t, &it);
1612
1613         if (error == -ERESTART_RESTARTBLOCK) {
1614                 struct timespec __user *rmtp = restart_block->nanosleep.rmtp;
1615                 /*
1616                  * Report back to the user the time still remaining.
1617                  */
1618                 if (rmtp && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1619                         return -EFAULT;
1620
1621                 restart_block->nanosleep.expires = timespec_to_ns(&t);
1622         }
1623         return error;
1624
1625 }
1626
1627 #define PROCESS_CLOCK   MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED)
1628 #define THREAD_CLOCK    MAKE_THREAD_CPUCLOCK(0, CPUCLOCK_SCHED)
1629
1630 static int process_cpu_clock_getres(const clockid_t which_clock,
1631                                     struct timespec *tp)
1632 {
1633         return posix_cpu_clock_getres(PROCESS_CLOCK, tp);
1634 }
1635 static int process_cpu_clock_get(const clockid_t which_clock,
1636                                  struct timespec *tp)
1637 {
1638         return posix_cpu_clock_get(PROCESS_CLOCK, tp);
1639 }
1640 static int process_cpu_timer_create(struct k_itimer *timer)
1641 {
1642         timer->it_clock = PROCESS_CLOCK;
1643         return posix_cpu_timer_create(timer);
1644 }
1645 static int process_cpu_nsleep(const clockid_t which_clock, int flags,
1646                               struct timespec *rqtp,
1647                               struct timespec __user *rmtp)
1648 {
1649         return posix_cpu_nsleep(PROCESS_CLOCK, flags, rqtp, rmtp);
1650 }
1651 static long process_cpu_nsleep_restart(struct restart_block *restart_block)
1652 {
1653         return -EINVAL;
1654 }
1655 static int thread_cpu_clock_getres(const clockid_t which_clock,
1656                                    struct timespec *tp)
1657 {
1658         return posix_cpu_clock_getres(THREAD_CLOCK, tp);
1659 }
1660 static int thread_cpu_clock_get(const clockid_t which_clock,
1661                                 struct timespec *tp)
1662 {
1663         return posix_cpu_clock_get(THREAD_CLOCK, tp);
1664 }
1665 static int thread_cpu_timer_create(struct k_itimer *timer)
1666 {
1667         timer->it_clock = THREAD_CLOCK;
1668         return posix_cpu_timer_create(timer);
1669 }
1670
1671 struct k_clock clock_posix_cpu = {
1672         .clock_getres   = posix_cpu_clock_getres,
1673         .clock_set      = posix_cpu_clock_set,
1674         .clock_get      = posix_cpu_clock_get,
1675         .timer_create   = posix_cpu_timer_create,
1676         .nsleep         = posix_cpu_nsleep,
1677         .nsleep_restart = posix_cpu_nsleep_restart,
1678         .timer_set      = posix_cpu_timer_set,
1679         .timer_del      = posix_cpu_timer_del,
1680         .timer_get      = posix_cpu_timer_get,
1681 };
1682
1683 static __init int init_posix_cpu_timers(void)
1684 {
1685         struct k_clock process = {
1686                 .clock_getres   = process_cpu_clock_getres,
1687                 .clock_get      = process_cpu_clock_get,
1688                 .timer_create   = process_cpu_timer_create,
1689                 .nsleep         = process_cpu_nsleep,
1690                 .nsleep_restart = process_cpu_nsleep_restart,
1691         };
1692         struct k_clock thread = {
1693                 .clock_getres   = thread_cpu_clock_getres,
1694                 .clock_get      = thread_cpu_clock_get,
1695                 .timer_create   = thread_cpu_timer_create,
1696         };
1697         struct timespec ts;
1698
1699         posix_timers_register_clock(CLOCK_PROCESS_CPUTIME_ID, &process);
1700         posix_timers_register_clock(CLOCK_THREAD_CPUTIME_ID, &thread);
1701
1702         cputime_to_timespec(cputime_one_jiffy, &ts);
1703         onecputick = ts.tv_nsec;
1704         WARN_ON(ts.tv_sec != 0);
1705
1706         return 0;
1707 }
1708 __initcall(init_posix_cpu_timers);