kernel/posix-cpu-timers.c

   1 /*
   2  * Implement CPU time clocks for the POSIX clock interface.
   3  */
   4
   5 #include <linux/sched.h>
   6 #include <linux/posix-timers.h>
   7 #include <linux/errno.h>
   8 #include <linux/math64.h>
   9 #include <asm/uaccess.h>
  10 #include <linux/kernel_stat.h>
  11 #include <trace/events/timer.h>
  12 #include <linux/random.h>
  13 #include <linux/tick.h>
  14 #include <linux/workqueue.h>
  15
  16 /*
  17  * Called after updating RLIMIT_CPU to run cpu timer and update
  18  * tsk->signal->cputime_expires expiration cache if necessary. Needs
  19  * siglock protection since other code may update expiration cache as
  20  * well.
  21  */
  22 void update_rlimit_cpu(struct task_struct *task, unsigned long rlim_new)
  23 {
  24         cputime_t cputime = secs_to_cputime(rlim_new);
  25
  26         spin_lock_irq(&task->sighand->siglock);
  27         set_process_cpu_timer(task, CPUCLOCK_PROF, &cputime, NULL);
  28         spin_unlock_irq(&task->sighand->siglock);
  29 }
  30
  31 static int check_clock(const clockid_t which_clock)
  32 {
  33         int error = 0;
  34         struct task_struct *p;
  35         const pid_t pid = CPUCLOCK_PID(which_clock);
  36
  37         if (CPUCLOCK_WHICH(which_clock) >= CPUCLOCK_MAX)
  38                 return -EINVAL;
  39
  40         if (pid == 0)
  41                 return 0;
  42
  43         rcu_read_lock();
  44         p = find_task_by_vpid(pid);
  45         if (!p || !(CPUCLOCK_PERTHREAD(which_clock) ?
  46                    same_thread_group(p, current) : has_group_leader_pid(p))) {
  47                 error = -EINVAL;
  48         }
  49         rcu_read_unlock();
  50
  51         return error;
  52 }
  53
  54 static inline union cpu_time_count
  55 timespec_to_sample(const clockid_t which_clock, const struct timespec *tp)
  56 {
  57         union cpu_time_count ret;
  58         ret.sched = 0;          /* high half always zero when .cpu used */
  59         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  60                 ret.sched = (unsigned long long)tp->tv_sec * NSEC_PER_SEC + tp->tv_nsec;
  61         } else {
  62                 ret.cpu = timespec_to_cputime(tp);
  63         }
  64         return ret;
  65 }
  66
  67 static void sample_to_timespec(const clockid_t which_clock,
  68                                union cpu_time_count cpu,
  69                                struct timespec *tp)
  70 {
  71         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED)
  72                 *tp = ns_to_timespec(cpu.sched);
  73         else
  74                 cputime_to_timespec(cpu.cpu, tp);
  75 }
  76
  77 static inline int cpu_time_before(const clockid_t which_clock,
  78                                   union cpu_time_count now,
  79                                   union cpu_time_count then)
  80 {
  81         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  82                 return now.sched < then.sched;
  83         }  else {
  84                 return now.cpu < then.cpu;
  85         }
  86 }
  87 static inline void cpu_time_add(const clockid_t which_clock,
  88                                 union cpu_time_count *acc,
  89                                 union cpu_time_count val)
  90 {
  91         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  92                 acc->sched += val.sched;
  93         }  else {
  94                 acc->cpu += val.cpu;
  95         }
  96 }
  97 static inline union cpu_time_count cpu_time_sub(const clockid_t which_clock,
  98                                                 union cpu_time_count a,
  99                                                 union cpu_time_count b)
 100 {
 101         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
 102                 a.sched -= b.sched;
 103         }  else {
 104                 a.cpu -= b.cpu;
 105         }
 106         return a;
 107 }
 108
 109 /*
 110  * Update expiry time from increment, and increase overrun count,
 111  * given the current clock sample.
 112  */
 113 static void bump_cpu_timer(struct k_itimer *timer,
 114                                   union cpu_time_count now)
 115 {
 116         int i;
 117
 118         if (timer->it.cpu.incr.sched == 0)
 119                 return;
 120
 121         if (CPUCLOCK_WHICH(timer->it_clock) == CPUCLOCK_SCHED) {
 122                 unsigned long long delta, incr;
 123
 124                 if (now.sched < timer->it.cpu.expires.sched)
 125                         return;
 126                 incr = timer->it.cpu.incr.sched;
 127                 delta = now.sched + incr - timer->it.cpu.expires.sched;
 128                 /* Don't use (incr*2 < delta), incr*2 might overflow. */
 129                 for (i = 0; incr < delta - incr; i++)
 130                         incr = incr << 1;
 131                 for (; i >= 0; incr >>= 1, i--) {
 132                         if (delta < incr)
 133                                 continue;
 134                         timer->it.cpu.expires.sched += incr;
 135                         timer->it_overrun += 1 << i;
 136                         delta -= incr;
 137                 }
 138         } else {
 139                 cputime_t delta, incr;
 140
 141                 if (now.cpu < timer->it.cpu.expires.cpu)
 142                         return;
 143                 incr = timer->it.cpu.incr.cpu;
 144                 delta = now.cpu + incr - timer->it.cpu.expires.cpu;
 145                 /* Don't use (incr*2 < delta), incr*2 might overflow. */
 146                 for (i = 0; incr < delta - incr; i++)
 147                              incr += incr;
 148                 for (; i >= 0; incr = incr >> 1, i--) {
 149                         if (delta < incr)
 150                                 continue;
 151                         timer->it.cpu.expires.cpu += incr;
 152                         timer->it_overrun += 1 << i;
 153                         delta -= incr;
 154                 }
 155         }
 156 }
 157
 158 /**
 159  * task_cputime_zero - Check a task_cputime struct for all zero fields.
 160  *
 161  * @cputime:    The struct to compare.
 162  *
 163  * Checks @cputime to see if all fields are zero.  Returns true if all fields
 164  * are zero, false if any field is nonzero.
 165  */
 166 static inline int task_cputime_zero(const struct task_cputime *cputime)
 167 {
 168         if (!cputime->utime && !cputime->stime && !cputime->sum_exec_runtime)
 169                 return 1;
 170         return 0;
 171 }
 172
 173 static inline cputime_t prof_ticks(struct task_struct *p)
 174 {
 175         cputime_t utime, stime;
 176
 177         task_cputime(p, &utime, &stime);
 178
 179         return utime + stime;
 180 }
 181 static inline cputime_t virt_ticks(struct task_struct *p)
 182 {
 183         cputime_t utime;
 184
 185         task_cputime(p, &utime, NULL);
 186
 187         return utime;
 188 }
 189
 190 static int
 191 posix_cpu_clock_getres(const clockid_t which_clock, struct timespec *tp)
 192 {
 193         int error = check_clock(which_clock);
 194         if (!error) {
 195                 tp->tv_sec = 0;
 196                 tp->tv_nsec = ((NSEC_PER_SEC + HZ - 1) / HZ);
 197                 if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
 198                         /*
 199                          * If sched_clock is using a cycle counter, we
 200                          * don't have any idea of its true resolution
 201                          * exported, but it is much more than 1s/HZ.
 202                          */
 203                         tp->tv_nsec = 1;
 204                 }
 205         }
 206         return error;
 207 }
 208
 209 static int
 210 posix_cpu_clock_set(const clockid_t which_clock, const struct timespec *tp)
 211 {
 212         /*
 213          * You can never reset a CPU clock, but we check for other errors
 214          * in the call before failing with EPERM.
 215          */
 216         int error = check_clock(which_clock);
 217         if (error == 0) {
 218                 error = -EPERM;
 219         }
 220         return error;
 221 }
 222
 223
 224 /*
 225  * Sample a per-thread clock for the given task.
 226  */
 227 static int cpu_clock_sample(const clockid_t which_clock, struct task_struct *p,
 228                             union cpu_time_count *cpu)
 229 {
 230         switch (CPUCLOCK_WHICH(which_clock)) {
 231         default:
 232                 return -EINVAL;
 233         case CPUCLOCK_PROF:
 234                 cpu->cpu = prof_ticks(p);
 235                 break;
 236         case CPUCLOCK_VIRT:
 237                 cpu->cpu = virt_ticks(p);
 238                 break;
 239         case CPUCLOCK_SCHED:
 240                 cpu->sched = task_sched_runtime(p);
 241                 break;
 242         }
 243         return 0;
 244 }
 245
 246 static void update_gt_cputime(struct task_cputime *a, struct task_cputime *b)
 247 {
 248         if (b->utime > a->utime)
 249                 a->utime = b->utime;
 250
 251         if (b->stime > a->stime)
 252                 a->stime = b->stime;
 253
 254         if (b->sum_exec_runtime > a->sum_exec_runtime)
 255                 a->sum_exec_runtime = b->sum_exec_runtime;
 256 }
 257
 258 void thread_group_cputimer(struct task_struct *tsk, struct task_cputime *times)
 259 {
 260         struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
 261         struct task_cputime sum;
 262         unsigned long flags;
 263
 264         if (!cputimer->running) {
 265                 /*
 266                  * The POSIX timer interface allows for absolute time expiry
 267                  * values through the TIMER_ABSTIME flag, therefore we have
 268                  * to synchronize the timer to the clock every time we start
 269                  * it.
 270                  */
 271                 thread_group_cputime(tsk, &sum);
 272                 raw_spin_lock_irqsave(&cputimer->lock, flags);
 273                 cputimer->running = 1;
 274                 update_gt_cputime(&cputimer->cputime, &sum);
 275         } else
 276                 raw_spin_lock_irqsave(&cputimer->lock, flags);
 277         *times = cputimer->cputime;
 278         raw_spin_unlock_irqrestore(&cputimer->lock, flags);
 279 }
 280
 281 /*
 282  * Sample a process (thread group) clock for the given group_leader task.
 283  * Must be called with tasklist_lock held for reading.
 284  */
 285 static int cpu_clock_sample_group(const clockid_t which_clock,
 286                                   struct task_struct *p,
 287                                   union cpu_time_count *cpu)
 288 {
 289         struct task_cputime cputime;
 290
 291         switch (CPUCLOCK_WHICH(which_clock)) {
 292         default:
 293                 return -EINVAL;
 294         case CPUCLOCK_PROF:
 295                 thread_group_cputime(p, &cputime);
 296                 cpu->cpu = cputime.utime + cputime.stime;
 297                 break;
 298         case CPUCLOCK_VIRT:
 299                 thread_group_cputime(p, &cputime);
 300                 cpu->cpu = cputime.utime;
 301                 break;
 302         case CPUCLOCK_SCHED:
 303                 thread_group_cputime(p, &cputime);
 304                 cpu->sched = cputime.sum_exec_runtime;
 305                 break;
 306         }
 307         return 0;
 308 }
 309
 310
 311 static int posix_cpu_clock_get(const clockid_t which_clock, struct timespec *tp)
 312 {
 313         const pid_t pid = CPUCLOCK_PID(which_clock);
 314         int error = -EINVAL;
 315         union cpu_time_count rtn;
 316
 317         if (pid == 0) {
 318                 /*
 319                  * Special case constant value for our own clocks.
 320                  * We don't have to do any lookup to find ourselves.
 321                  */
 322                 if (CPUCLOCK_PERTHREAD(which_clock)) {
 323                         /*
 324                          * Sampling just ourselves we can do with no locking.
 325                          */
 326                         error = cpu_clock_sample(which_clock,
 327                                                  current, &rtn);
 328                 } else {
 329                         read_lock(&tasklist_lock);
 330                         error = cpu_clock_sample_group(which_clock,
 331                                                        current, &rtn);
 332                         read_unlock(&tasklist_lock);
 333                 }
 334         } else {
 335                 /*
 336                  * Find the given PID, and validate that the caller
 337                  * should be able to see it.
 338                  */
 339                 struct task_struct *p;
 340                 rcu_read_lock();
 341                 p = find_task_by_vpid(pid);
 342                 if (p) {
 343                         if (CPUCLOCK_PERTHREAD(which_clock)) {
 344                                 if (same_thread_group(p, current)) {
 345                                         error = cpu_clock_sample(which_clock,
 346                                                                  p, &rtn);
 347                                 }
 348                         } else {
 349                                 read_lock(&tasklist_lock);
 350                                 if (thread_group_leader(p) && p->sighand) {
 351                                         error =
 352                                             cpu_clock_sample_group(which_clock,
 353                                                                    p, &rtn);
 354                                 }
 355                                 read_unlock(&tasklist_lock);
 356                         }
 357                 }
 358                 rcu_read_unlock();
 359         }
 360
 361         if (error)
 362                 return error;
 363         sample_to_timespec(which_clock, rtn, tp);
 364         return 0;
 365 }
 366
 367
 368 /*
 369  * Validate the clockid_t for a new CPU-clock timer, and initialize the timer.
 370  * This is called from sys_timer_create() and do_cpu_nanosleep() with the
 371  * new timer already all-zeros initialized.
 372  */
 373 static int posix_cpu_timer_create(struct k_itimer *new_timer)
 374 {
 375         int ret = 0;
 376         const pid_t pid = CPUCLOCK_PID(new_timer->it_clock);
 377         struct task_struct *p;
 378
 379         if (CPUCLOCK_WHICH(new_timer->it_clock) >= CPUCLOCK_MAX)
 380                 return -EINVAL;
 381
 382         INIT_LIST_HEAD(&new_timer->it.cpu.entry);
 383
 384         rcu_read_lock();
 385         if (CPUCLOCK_PERTHREAD(new_timer->it_clock)) {
 386                 if (pid == 0) {
 387                         p = current;
 388                 } else {
 389                         p = find_task_by_vpid(pid);
 390                         if (p && !same_thread_group(p, current))
 391                                 p = NULL;
 392                 }
 393         } else {
 394                 if (pid == 0) {
 395                         p = current->group_leader;
 396                 } else {
 397                         p = find_task_by_vpid(pid);
 398                         if (p && !has_group_leader_pid(p))
 399                                 p = NULL;
 400                 }
 401         }
 402         new_timer->it.cpu.task = p;
 403         if (p) {
 404                 get_task_struct(p);
 405         } else {
 406                 ret = -EINVAL;
 407         }
 408         rcu_read_unlock();
 409
 410         return ret;
 411 }
 412
 413 /*
 414  * Clean up a CPU-clock timer that is about to be destroyed.
 415  * This is called from timer deletion with the timer already locked.
 416  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 417  * and try again.  (This happens when the timer is in the middle of firing.)
 418  */
 419 static int posix_cpu_timer_del(struct k_itimer *timer)
 420 {
 421         struct task_struct *p = timer->it.cpu.task;
 422         int ret = 0;
 423
 424         if (likely(p != NULL)) {
 425                 read_lock(&tasklist_lock);
 426                 if (unlikely(p->sighand == NULL)) {
 427                         /*
 428                          * We raced with the reaping of the task.
 429                          * The deletion should have cleared us off the list.
 430                          */
 431                         BUG_ON(!list_empty(&timer->it.cpu.entry));
 432                 } else {
 433                         spin_lock(&p->sighand->siglock);
 434                         if (timer->it.cpu.firing)
 435                                 ret = TIMER_RETRY;
 436                         else
 437                                 list_del(&timer->it.cpu.entry);
 438                         spin_unlock(&p->sighand->siglock);
 439                 }
 440                 read_unlock(&tasklist_lock);
 441
 442                 if (!ret)
 443                         put_task_struct(p);
 444         }
 445
 446         return ret;
 447 }
 448
 449 /*
 450  * Clean out CPU timers still ticking when a thread exited.  The task
 451  * pointer is cleared, and the expiry time is replaced with the residual
 452  * time for later timer_gettime calls to return.
 453  * This must be called with the siglock held.
 454  */
 455 static void cleanup_timers(struct list_head *head,
 456                            cputime_t utime, cputime_t stime,
 457                            unsigned long long sum_exec_runtime)
 458 {
 459         struct cpu_timer_list *timer, *next;
 460         cputime_t ptime = utime + stime;
 461
 462         list_for_each_entry_safe(timer, next, head, entry) {
 463                 list_del_init(&timer->entry);
 464                 if (timer->expires.cpu < ptime) {
 465                         timer->expires.cpu = 0;
 466                 } else {
 467                         timer->expires.cpu -= ptime;
 468                 }
 469         }
 470
 471         ++head;
 472         list_for_each_entry_safe(timer, next, head, entry) {
 473                 list_del_init(&timer->entry);
 474                 if (timer->expires.cpu < utime) {
 475                         timer->expires.cpu = 0;
 476                 } else {
 477                         timer->expires.cpu -= utime;
 478                 }
 479         }
 480
 481         ++head;
 482         list_for_each_entry_safe(timer, next, head, entry) {
 483                 list_del_init(&timer->entry);
 484                 if (timer->expires.sched < sum_exec_runtime) {
 485                         timer->expires.sched = 0;
 486                 } else {
 487                         timer->expires.sched -= sum_exec_runtime;
 488                 }
 489         }
 490 }
 491
 492 /*
 493  * These are both called with the siglock held, when the current thread
 494  * is being reaped.  When the final (leader) thread in the group is reaped,
 495  * posix_cpu_timers_exit_group will be called after posix_cpu_timers_exit.
 496  */
 497 void posix_cpu_timers_exit(struct task_struct *tsk)
 498 {
 499         cputime_t utime, stime;
 500
 501         add_device_randomness((const void*) &tsk->se.sum_exec_runtime,
 502                                                 sizeof(unsigned long long));
 503         task_cputime(tsk, &utime, &stime);
 504         cleanup_timers(tsk->cpu_timers,
 505                        utime, stime, tsk->se.sum_exec_runtime);
 506
 507 }
 508 void posix_cpu_timers_exit_group(struct task_struct *tsk)
 509 {
 510         struct signal_struct *const sig = tsk->signal;
 511         cputime_t utime, stime;
 512
 513         task_cputime(tsk, &utime, &stime);
 514         cleanup_timers(tsk->signal->cpu_timers,
 515                        utime + sig->utime, stime + sig->stime,
 516                        tsk->se.sum_exec_runtime + sig->sum_sched_runtime);
 517 }
 518
 519 static void clear_dead_task(struct k_itimer *timer, union cpu_time_count now)
 520 {
 521         /*
 522          * That's all for this thread or process.
 523          * We leave our residual in expires to be reported.
 524          */
 525         put_task_struct(timer->it.cpu.task);
 526         timer->it.cpu.task = NULL;
 527         timer->it.cpu.expires = cpu_time_sub(timer->it_clock,
 528                                              timer->it.cpu.expires,
 529                                              now);
 530 }
 531
 532 static inline int expires_gt(cputime_t expires, cputime_t new_exp)
 533 {
 534         return expires == 0 || expires > new_exp;
 535 }
 536
 537 /*
 538  * Insert the timer on the appropriate list before any timers that
 539  * expire later.  This must be called with the tasklist_lock held
 540  * for reading, interrupts disabled and p->sighand->siglock taken.
 541  */
 542 static void arm_timer(struct k_itimer *timer)
 543 {
 544         struct task_struct *p = timer->it.cpu.task;
 545         struct list_head *head, *listpos;
 546         struct task_cputime *cputime_expires;
 547         struct cpu_timer_list *const nt = &timer->it.cpu;
 548         struct cpu_timer_list *next;
 549
 550         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 551                 head = p->cpu_timers;
 552                 cputime_expires = &p->cputime_expires;
 553         } else {
 554                 head = p->signal->cpu_timers;
 555                 cputime_expires = &p->signal->cputime_expires;
 556         }
 557         head += CPUCLOCK_WHICH(timer->it_clock);
 558
 559         listpos = head;
 560         list_for_each_entry(next, head, entry) {
 561                 if (cpu_time_before(timer->it_clock, nt->expires, next->expires))
 562                         break;
 563                 listpos = &next->entry;
 564         }
 565         list_add(&nt->entry, listpos);
 566
 567         if (listpos == head) {
 568                 union cpu_time_count *exp = &nt->expires;
 569
 570                 /*
 571                  * We are the new earliest-expiring POSIX 1.b timer, hence
 572                  * need to update expiration cache. Take into account that
 573                  * for process timers we share expiration cache with itimers
 574                  * and RLIMIT_CPU and for thread timers with RLIMIT_RTTIME.
 575                  */
 576
 577                 switch (CPUCLOCK_WHICH(timer->it_clock)) {
 578                 case CPUCLOCK_PROF:
 579                         if (expires_gt(cputime_expires->prof_exp, exp->cpu))
 580                                 cputime_expires->prof_exp = exp->cpu;
 581                         break;
 582                 case CPUCLOCK_VIRT:
 583                         if (expires_gt(cputime_expires->virt_exp, exp->cpu))
 584                                 cputime_expires->virt_exp = exp->cpu;
 585                         break;
 586                 case CPUCLOCK_SCHED:
 587                         if (cputime_expires->sched_exp == 0 ||
 588                             cputime_expires->sched_exp > exp->sched)
 589                                 cputime_expires->sched_exp = exp->sched;
 590                         break;
 591                 }
 592         }
 593 }
 594
 595 /*
 596  * The timer is locked, fire it and arrange for its reload.
 597  */
 598 static void cpu_timer_fire(struct k_itimer *timer)
 599 {
 600         if ((timer->it_sigev_notify & ~SIGEV_THREAD_ID) == SIGEV_NONE) {
 601                 /*
 602                  * User don't want any signal.
 603                  */
 604                 timer->it.cpu.expires.sched = 0;
 605         } else if (unlikely(timer->sigq == NULL)) {
 606                 /*
 607                  * This a special case for clock_nanosleep,
 608                  * not a normal timer from sys_timer_create.
 609                  */
 610                 wake_up_process(timer->it_process);
 611                 timer->it.cpu.expires.sched = 0;
 612         } else if (timer->it.cpu.incr.sched == 0) {
 613                 /*
 614                  * One-shot timer.  Clear it as soon as it's fired.
 615                  */
 616                 posix_timer_event(timer, 0);
 617                 timer->it.cpu.expires.sched = 0;
 618         } else if (posix_timer_event(timer, ++timer->it_requeue_pending)) {
 619                 /*
 620                  * The signal did not get queued because the signal
 621                  * was ignored, so we won't get any callback to
 622                  * reload the timer.  But we need to keep it
 623                  * ticking in case the signal is deliverable next time.
 624                  */
 625                 posix_cpu_timer_schedule(timer);
 626         }
 627 }
 628
 629 /*
 630  * Sample a process (thread group) timer for the given group_leader task.
 631  * Must be called with tasklist_lock held for reading.
 632  */
 633 static int cpu_timer_sample_group(const clockid_t which_clock,
 634                                   struct task_struct *p,
 635                                   union cpu_time_count *cpu)
 636 {
 637         struct task_cputime cputime;
 638
 639         thread_group_cputimer(p, &cputime);
 640         switch (CPUCLOCK_WHICH(which_clock)) {
 641         default:
 642                 return -EINVAL;
 643         case CPUCLOCK_PROF:
 644                 cpu->cpu = cputime.utime + cputime.stime;
 645                 break;
 646         case CPUCLOCK_VIRT:
 647                 cpu->cpu = cputime.utime;
 648                 break;
 649         case CPUCLOCK_SCHED:
 650                 cpu->sched = cputime.sum_exec_runtime + task_delta_exec(p);
 651                 break;
 652         }
 653         return 0;
 654 }
 655
 656 #ifdef CONFIG_NO_HZ_FULL
 657 static void nohz_kick_work_fn(struct work_struct *work)
 658 {
 659         tick_nohz_full_kick_all();
 660 }
 661
 662 static DECLARE_WORK(nohz_kick_work, nohz_kick_work_fn);
 663
 664 /*
 665  * We need the IPIs to be sent from sane process context.
 666  * The posix cpu timers are always set with irqs disabled.
 667  */
 668 static void posix_cpu_timer_kick_nohz(void)
 669 {
 670         schedule_work(&nohz_kick_work);
 671 }
 672
 673 bool posix_cpu_timers_can_stop_tick(struct task_struct *tsk)
 674 {
 675         if (!task_cputime_zero(&tsk->cputime_expires))
 676                 return false;
 677
 678         if (tsk->signal->cputimer.running)
 679                 return false;
 680
 681         return true;
 682 }
 683 #else
 684 static inline void posix_cpu_timer_kick_nohz(void) { }
 685 #endif
 686
 687 /*
 688  * Guts of sys_timer_settime for CPU timers.
 689  * This is called with the timer locked and interrupts disabled.
 690  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 691  * and try again.  (This happens when the timer is in the middle of firing.)
 692  */
 693 static int posix_cpu_timer_set(struct k_itimer *timer, int flags,
 694                                struct itimerspec *new, struct itimerspec *old)
 695 {
 696         struct task_struct *p = timer->it.cpu.task;
 697         union cpu_time_count old_expires, new_expires, old_incr, val;
 698         int ret;
 699
 700         if (unlikely(p == NULL)) {
 701                 /*
 702                  * Timer refers to a dead task's clock.
 703                  */
 704                 return -ESRCH;
 705         }
 706
 707         new_expires = timespec_to_sample(timer->it_clock, &new->it_value);
 708
 709         read_lock(&tasklist_lock);
 710         /*
 711          * We need the tasklist_lock to protect against reaping that
 712          * clears p->sighand.  If p has just been reaped, we can no
 713          * longer get any information about it at all.
 714          */
 715         if (unlikely(p->sighand == NULL)) {
 716                 read_unlock(&tasklist_lock);
 717                 put_task_struct(p);
 718                 timer->it.cpu.task = NULL;
 719                 return -ESRCH;
 720         }
 721
 722         /*
 723          * Disarm any old timer after extracting its expiry time.
 724          */
 725         BUG_ON(!irqs_disabled());
 726
 727         ret = 0;
 728         old_incr = timer->it.cpu.incr;
 729         spin_lock(&p->sighand->siglock);
 730         old_expires = timer->it.cpu.expires;
 731         if (unlikely(timer->it.cpu.firing)) {
 732                 timer->it.cpu.firing = -1;
 733                 ret = TIMER_RETRY;
 734         } else
 735                 list_del_init(&timer->it.cpu.entry);
 736
 737         /*
 738          * We need to sample the current value to convert the new
 739          * value from to relative and absolute, and to convert the
 740          * old value from absolute to relative.  To set a process
 741          * timer, we need a sample to balance the thread expiry
 742          * times (in arm_timer).  With an absolute time, we must
 743          * check if it's already passed.  In short, we need a sample.
 744          */
 745         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 746                 cpu_clock_sample(timer->it_clock, p, &val);
 747         } else {
 748                 cpu_timer_sample_group(timer->it_clock, p, &val);
 749         }
 750
 751         if (old) {
 752                 if (old_expires.sched == 0) {
 753                         old->it_value.tv_sec = 0;
 754                         old->it_value.tv_nsec = 0;
 755                 } else {
 756                         /*
 757                          * Update the timer in case it has
 758                          * overrun already.  If it has,
 759                          * we'll report it as having overrun
 760                          * and with the next reloaded timer
 761                          * already ticking, though we are
 762                          * swallowing that pending
 763                          * notification here to install the
 764                          * new setting.
 765                          */
 766                         bump_cpu_timer(timer, val);
 767                         if (cpu_time_before(timer->it_clock, val,
 768                                             timer->it.cpu.expires)) {
 769                                 old_expires = cpu_time_sub(
 770                                         timer->it_clock,
 771                                         timer->it.cpu.expires, val);
 772                                 sample_to_timespec(timer->it_clock,
 773                                                    old_expires,
 774                                                    &old->it_value);
 775                         } else {
 776                                 old->it_value.tv_nsec = 1;
 777                                 old->it_value.tv_sec = 0;
 778                         }
 779                 }
 780         }
 781
 782         if (unlikely(ret)) {
 783                 /*
 784                  * We are colliding with the timer actually firing.
 785                  * Punt after filling in the timer's old value, and
 786                  * disable this firing since we are already reporting
 787                  * it as an overrun (thanks to bump_cpu_timer above).
 788                  */
 789                 spin_unlock(&p->sighand->siglock);
 790                 read_unlock(&tasklist_lock);
 791                 goto out;
 792         }
 793
 794         if (new_expires.sched != 0 && !(flags & TIMER_ABSTIME)) {
 795                 cpu_time_add(timer->it_clock, &new_expires, val);
 796         }
 797
 798         /*
 799          * Install the new expiry time (or zero).
 800          * For a timer with no notification action, we don't actually
 801          * arm the timer (we'll just fake it for timer_gettime).
 802          */
 803         timer->it.cpu.expires = new_expires;
 804         if (new_expires.sched != 0 &&
 805             cpu_time_before(timer->it_clock, val, new_expires)) {
 806                 arm_timer(timer);
 807         }
 808
 809         spin_unlock(&p->sighand->siglock);
 810         read_unlock(&tasklist_lock);
 811
 812         /*
 813          * Install the new reload setting, and
 814          * set up the signal and overrun bookkeeping.
 815          */
 816         timer->it.cpu.incr = timespec_to_sample(timer->it_clock,
 817                                                 &new->it_interval);
 818
 819         /*
 820          * This acts as a modification timestamp for the timer,
 821          * so any automatic reload attempt will punt on seeing
 822          * that we have reset the timer manually.
 823          */
 824         timer->it_requeue_pending = (timer->it_requeue_pending + 2) &
 825                 ~REQUEUE_PENDING;
 826         timer->it_overrun_last = 0;
 827         timer->it_overrun = -1;
 828
 829         if (new_expires.sched != 0 &&
 830             !cpu_time_before(timer->it_clock, val, new_expires)) {
 831                 /*
 832                  * The designated time already passed, so we notify
 833                  * immediately, even if the thread never runs to
 834                  * accumulate more time on this clock.
 835                  */
 836                 cpu_timer_fire(timer);
 837         }
 838
 839         ret = 0;
 840  out:
 841         if (old) {
 842                 sample_to_timespec(timer->it_clock,
 843                                    old_incr, &old->it_interval);
 844         }
 845         if (!ret)
 846                 posix_cpu_timer_kick_nohz();
 847         return ret;
 848 }
 849
 850 static void posix_cpu_timer_get(struct k_itimer *timer, struct itimerspec *itp)
 851 {
 852         union cpu_time_count now;
 853         struct task_struct *p = timer->it.cpu.task;
 854         int clear_dead;
 855
 856         /*
 857          * Easy part: convert the reload time.
 858          */
 859         sample_to_timespec(timer->it_clock,
 860                            timer->it.cpu.incr, &itp->it_interval);
 861
 862         if (timer->it.cpu.expires.sched == 0) { /* Timer not armed at all.  */
 863                 itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
 864                 return;
 865         }
 866
 867         if (unlikely(p == NULL)) {
 868                 /*
 869                  * This task already died and the timer will never fire.
 870                  * In this case, expires is actually the dead value.
 871                  */
 872         dead:
 873                 sample_to_timespec(timer->it_clock, timer->it.cpu.expires,
 874                                    &itp->it_value);
 875                 return;
 876         }
 877
 878         /*
 879          * Sample the clock to take the difference with the expiry time.
 880          */
 881         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 882                 cpu_clock_sample(timer->it_clock, p, &now);
 883                 clear_dead = p->exit_state;
 884         } else {
 885                 read_lock(&tasklist_lock);
 886                 if (unlikely(p->sighand == NULL)) {
 887                         /*
 888                          * The process has been reaped.
 889                          * We can't even collect a sample any more.
 890                          * Call the timer disarmed, nothing else to do.
 891                          */
 892                         put_task_struct(p);
 893                         timer->it.cpu.task = NULL;
 894                         timer->it.cpu.expires.sched = 0;
 895                         read_unlock(&tasklist_lock);
 896                         goto dead;
 897                 } else {
 898                         cpu_timer_sample_group(timer->it_clock, p, &now);
 899                         clear_dead = (unlikely(p->exit_state) &&
 900                                       thread_group_empty(p));
 901                 }
 902                 read_unlock(&tasklist_lock);
 903         }
 904
 905         if (unlikely(clear_dead)) {
 906                 /*
 907                  * We've noticed that the thread is dead, but
 908                  * not yet reaped.  Take this opportunity to
 909                  * drop our task ref.
 910                  */
 911                 clear_dead_task(timer, now);
 912                 goto dead;
 913         }
 914
 915         if (cpu_time_before(timer->it_clock, now, timer->it.cpu.expires)) {
 916                 sample_to_timespec(timer->it_clock,
 917                                    cpu_time_sub(timer->it_clock,
 918                                                 timer->it.cpu.expires, now),
 919                                    &itp->it_value);
 920         } else {
 921                 /*
 922                  * The timer should have expired already, but the firing
 923                  * hasn't taken place yet.  Say it's just about to expire.
 924                  */
 925                 itp->it_value.tv_nsec = 1;
 926                 itp->it_value.tv_sec = 0;
 927         }
 928 }
 929
 930 /*
 931  * Check for any per-thread CPU timers that have fired and move them off
 932  * the tsk->cpu_timers[N] list onto the firing list.  Here we update the
 933  * tsk->it_*_expires values to reflect the remaining thread CPU timers.
 934  */
 935 static void check_thread_timers(struct task_struct *tsk,
 936                                 struct list_head *firing)
 937 {
 938         int maxfire;
 939         struct list_head *timers = tsk->cpu_timers;
 940         struct signal_struct *const sig = tsk->signal;
 941         unsigned long soft;
 942
 943         maxfire = 20;
 944         tsk->cputime_expires.prof_exp = 0;
 945         while (!list_empty(timers)) {
 946                 struct cpu_timer_list *t = list_first_entry(timers,
 947                                                       struct cpu_timer_list,
 948                                                       entry);
 949                 if (!--maxfire || prof_ticks(tsk) < t->expires.cpu) {
 950                         tsk->cputime_expires.prof_exp = t->expires.cpu;
 951                         break;
 952                 }
 953                 t->firing = 1;
 954                 list_move_tail(&t->entry, firing);
 955         }
 956
 957         ++timers;
 958         maxfire = 20;
 959         tsk->cputime_expires.virt_exp = 0;
 960         while (!list_empty(timers)) {
 961                 struct cpu_timer_list *t = list_first_entry(timers,
 962                                                       struct cpu_timer_list,
 963                                                       entry);
 964                 if (!--maxfire || virt_ticks(tsk) < t->expires.cpu) {
 965                         tsk->cputime_expires.virt_exp = t->expires.cpu;
 966                         break;
 967                 }
 968                 t->firing = 1;
 969                 list_move_tail(&t->entry, firing);
 970         }
 971
 972         ++timers;
 973         maxfire = 20;
 974         tsk->cputime_expires.sched_exp = 0;
 975         while (!list_empty(timers)) {
 976                 struct cpu_timer_list *t = list_first_entry(timers,
 977                                                       struct cpu_timer_list,
 978                                                       entry);
 979                 if (!--maxfire || tsk->se.sum_exec_runtime < t->expires.sched) {
 980                         tsk->cputime_expires.sched_exp = t->expires.sched;
 981                         break;
 982                 }
 983                 t->firing = 1;
 984                 list_move_tail(&t->entry, firing);
 985         }
 986
 987         /*
 988          * Check for the special case thread timers.
 989          */
 990         soft = ACCESS_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_cur);
 991         if (soft != RLIM_INFINITY) {
 992                 unsigned long hard =
 993                         ACCESS_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_max);
 994
 995                 if (hard != RLIM_INFINITY &&
 996                     tsk->rt.timeout > DIV_ROUND_UP(hard, USEC_PER_SEC/HZ)) {
 997                         /*
 998                          * At the hard limit, we just die.
 999                          * No need to calculate anything else now.
1000                          */
1001                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
1002                         return;
1003                 }
1004                 if (tsk->rt.timeout > DIV_ROUND_UP(soft, USEC_PER_SEC/HZ)) {
1005                         /*
1006                          * At the soft limit, send a SIGXCPU every second.
1007                          */
1008                         if (soft < hard) {
1009                                 soft += USEC_PER_SEC;
1010                                 sig->rlim[RLIMIT_RTTIME].rlim_cur = soft;
1011                         }
1012                         printk(KERN_INFO
1013                                 "RT Watchdog Timeout: %s[%d]\n",
1014                                 tsk->comm, task_pid_nr(tsk));
1015                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
1016                 }
1017         }
1018 }
1019
1020 static void stop_process_timers(struct signal_struct *sig)
1021 {
1022         struct thread_group_cputimer *cputimer = &sig->cputimer;
1023         unsigned long flags;
1024
1025         raw_spin_lock_irqsave(&cputimer->lock, flags);
1026         cputimer->running = 0;
1027         raw_spin_unlock_irqrestore(&cputimer->lock, flags);
1028 }
1029
1030 static u32 onecputick;
1031
1032 static void check_cpu_itimer(struct task_struct *tsk, struct cpu_itimer *it,
1033                              cputime_t *expires, cputime_t cur_time, int signo)
1034 {
1035         if (!it->expires)
1036                 return;
1037
1038         if (cur_time >= it->expires) {
1039                 if (it->incr) {
1040                         it->expires += it->incr;
1041                         it->error += it->incr_error;
1042                         if (it->error >= onecputick) {
1043                                 it->expires -= cputime_one_jiffy;
1044                                 it->error -= onecputick;
1045                         }
1046                 } else {
1047                         it->expires = 0;
1048                 }
1049
1050                 trace_itimer_expire(signo == SIGPROF ?
1051                                     ITIMER_PROF : ITIMER_VIRTUAL,
1052                                     tsk->signal->leader_pid, cur_time);
1053                 __group_send_sig_info(signo, SEND_SIG_PRIV, tsk);
1054         }
1055
1056         if (it->expires && (!*expires || it->expires < *expires)) {
1057                 *expires = it->expires;
1058         }
1059 }
1060
1061 /*
1062  * Check for any per-thread CPU timers that have fired and move them
1063  * off the tsk->*_timers list onto the firing list.  Per-thread timers
1064  * have already been taken off.
1065  */
1066 static void check_process_timers(struct task_struct *tsk,
1067                                  struct list_head *firing)
1068 {
1069         int maxfire;
1070         struct signal_struct *const sig = tsk->signal;
1071         cputime_t utime, ptime, virt_expires, prof_expires;
1072         unsigned long long sum_sched_runtime, sched_expires;
1073         struct list_head *timers = sig->cpu_timers;
1074         struct task_cputime cputime;
1075         unsigned long soft;
1076
1077         /*
1078          * Collect the current process totals.
1079          */
1080         thread_group_cputimer(tsk, &cputime);
1081         utime = cputime.utime;
1082         ptime = utime + cputime.stime;
1083         sum_sched_runtime = cputime.sum_exec_runtime;
1084         maxfire = 20;
1085         prof_expires = 0;
1086         while (!list_empty(timers)) {
1087                 struct cpu_timer_list *tl = list_first_entry(timers,
1088                                                       struct cpu_timer_list,
1089                                                       entry);
1090                 if (!--maxfire || ptime < tl->expires.cpu) {
1091                         prof_expires = tl->expires.cpu;
1092                         break;
1093                 }
1094                 tl->firing = 1;
1095                 list_move_tail(&tl->entry, firing);
1096         }
1097
1098         ++timers;
1099         maxfire = 20;
1100         virt_expires = 0;
1101         while (!list_empty(timers)) {
1102                 struct cpu_timer_list *tl = list_first_entry(timers,
1103                                                       struct cpu_timer_list,
1104                                                       entry);
1105                 if (!--maxfire || utime < tl->expires.cpu) {
1106                         virt_expires = tl->expires.cpu;
1107                         break;
1108                 }
1109                 tl->firing = 1;
1110                 list_move_tail(&tl->entry, firing);
1111         }
1112
1113         ++timers;
1114         maxfire = 20;
1115         sched_expires = 0;
1116         while (!list_empty(timers)) {
1117                 struct cpu_timer_list *tl = list_first_entry(timers,
1118                                                       struct cpu_timer_list,
1119                                                       entry);
1120                 if (!--maxfire || sum_sched_runtime < tl->expires.sched) {
1121                         sched_expires = tl->expires.sched;
1122                         break;
1123                 }
1124                 tl->firing = 1;
1125                 list_move_tail(&tl->entry, firing);
1126         }
1127
1128         /*
1129          * Check for the special case process timers.
1130          */
1131         check_cpu_itimer(tsk, &sig->it[CPUCLOCK_PROF], &prof_expires, ptime,
1132                          SIGPROF);
1133         check_cpu_itimer(tsk, &sig->it[CPUCLOCK_VIRT], &virt_expires, utime,
1134                          SIGVTALRM);
1135         soft = ACCESS_ONCE(sig->rlim[RLIMIT_CPU].rlim_cur);
1136         if (soft != RLIM_INFINITY) {
1137                 unsigned long psecs = cputime_to_secs(ptime);
1138                 unsigned long hard =
1139                         ACCESS_ONCE(sig->rlim[RLIMIT_CPU].rlim_max);
1140                 cputime_t x;
1141                 if (psecs >= hard) {
1142                         /*
1143                          * At the hard limit, we just die.
1144                          * No need to calculate anything else now.
1145                          */
1146                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
1147                         return;
1148                 }
1149                 if (psecs >= soft) {
1150                         /*
1151                          * At the soft limit, send a SIGXCPU every second.
1152                          */
1153                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
1154                         if (soft < hard) {
1155                                 soft++;
1156                                 sig->rlim[RLIMIT_CPU].rlim_cur = soft;
1157                         }
1158                 }
1159                 x = secs_to_cputime(soft);
1160                 if (!prof_expires || x < prof_expires) {
1161                         prof_expires = x;
1162                 }
1163         }
1164
1165         sig->cputime_expires.prof_exp = prof_expires;
1166         sig->cputime_expires.virt_exp = virt_expires;
1167         sig->cputime_expires.sched_exp = sched_expires;
1168         if (task_cputime_zero(&sig->cputime_expires))
1169                 stop_process_timers(sig);
1170 }
1171
1172 /*
1173  * This is called from the signal code (via do_schedule_next_timer)
1174  * when the last timer signal was delivered and we have to reload the timer.
1175  */
1176 void posix_cpu_timer_schedule(struct k_itimer *timer)
1177 {
1178         struct task_struct *p = timer->it.cpu.task;
1179         union cpu_time_count now;
1180
1181         if (unlikely(p == NULL))
1182                 /*
1183                  * The task was cleaned up already, no future firings.
1184                  */
1185                 goto out;
1186
1187         /*
1188          * Fetch the current sample and update the timer's expiry time.
1189          */
1190         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
1191                 cpu_clock_sample(timer->it_clock, p, &now);
1192                 bump_cpu_timer(timer, now);
1193                 if (unlikely(p->exit_state)) {
1194                         clear_dead_task(timer, now);
1195                         goto out;
1196                 }
1197                 read_lock(&tasklist_lock); /* arm_timer needs it.  */
1198                 spin_lock(&p->sighand->siglock);
1199         } else {
1200                 read_lock(&tasklist_lock);
1201                 if (unlikely(p->sighand == NULL)) {
1202                         /*
1203                          * The process has been reaped.
1204                          * We can't even collect a sample any more.
1205                          */
1206                         put_task_struct(p);
1207                         timer->it.cpu.task = p = NULL;
1208                         timer->it.cpu.expires.sched = 0;
1209                         goto out_unlock;
1210                 } else if (unlikely(p->exit_state) && thread_group_empty(p)) {
1211                         /*
1212                          * We've noticed that the thread is dead, but
1213                          * not yet reaped.  Take this opportunity to
1214                          * drop our task ref.
1215                          */
1216                         clear_dead_task(timer, now);
1217                         goto out_unlock;
1218                 }
1219                 spin_lock(&p->sighand->siglock);
1220                 cpu_timer_sample_group(timer->it_clock, p, &now);
1221                 bump_cpu_timer(timer, now);
1222                 /* Leave the tasklist_lock locked for the call below.  */
1223         }
1224
1225         /*
1226          * Now re-arm for the new expiry time.
1227          */
1228         BUG_ON(!irqs_disabled());
1229         arm_timer(timer);
1230         spin_unlock(&p->sighand->siglock);
1231
1232 out_unlock:
1233         read_unlock(&tasklist_lock);
1234
1235 out:
1236         timer->it_overrun_last = timer->it_overrun;
1237         timer->it_overrun = -1;
1238         ++timer->it_requeue_pending;
1239 }
1240
1241 /**
1242  * task_cputime_expired - Compare two task_cputime entities.
1243  *
1244  * @sample:     The task_cputime structure to be checked for expiration.
1245  * @expires:    Expiration times, against which @sample will be checked.
1246  *
1247  * Checks @sample against @expires to see if any field of @sample has expired.
1248  * Returns true if any field of the former is greater than the corresponding
1249  * field of the latter if the latter field is set.  Otherwise returns false.
1250  */
1251 static inline int task_cputime_expired(const struct task_cputime *sample,
1252                                         const struct task_cputime *expires)
1253 {
1254         if (expires->utime && sample->utime >= expires->utime)
1255                 return 1;
1256         if (expires->stime && sample->utime + sample->stime >= expires->stime)
1257                 return 1;
1258         if (expires->sum_exec_runtime != 0 &&
1259             sample->sum_exec_runtime >= expires->sum_exec_runtime)
1260                 return 1;
1261         return 0;
1262 }
1263
1264 /**
1265  * fastpath_timer_check - POSIX CPU timers fast path.
1266  *
1267  * @tsk:        The task (thread) being checked.
1268  *
1269  * Check the task and thread group timers.  If both are zero (there are no
1270  * timers set) return false.  Otherwise snapshot the task and thread group
1271  * timers and compare them with the corresponding expiration times.  Return
1272  * true if a timer has expired, else return false.
1273  */
1274 static inline int fastpath_timer_check(struct task_struct *tsk)
1275 {
1276         struct signal_struct *sig;
1277         cputime_t utime, stime;
1278
1279         task_cputime(tsk, &utime, &stime);
1280
1281         if (!task_cputime_zero(&tsk->cputime_expires)) {
1282                 struct task_cputime task_sample = {
1283                         .utime = utime,
1284                         .stime = stime,
1285                         .sum_exec_runtime = tsk->se.sum_exec_runtime
1286                 };
1287
1288                 if (task_cputime_expired(&task_sample, &tsk->cputime_expires))
1289                         return 1;
1290         }
1291
1292         sig = tsk->signal;
1293         if (sig->cputimer.running) {
1294                 struct task_cputime group_sample;
1295
1296                 raw_spin_lock(&sig->cputimer.lock);
1297                 group_sample = sig->cputimer.cputime;
1298                 raw_spin_unlock(&sig->cputimer.lock);
1299
1300                 if (task_cputime_expired(&group_sample, &sig->cputime_expires))
1301                         return 1;
1302         }
1303
1304         return 0;
1305 }
1306
1307 /*
1308  * This is called from the timer interrupt handler.  The irq handler has
1309  * already updated our counts.  We need to check if any timers fire now.
1310  * Interrupts are disabled.
1311  */
1312 void run_posix_cpu_timers(struct task_struct *tsk)
1313 {
1314         LIST_HEAD(firing);
1315         struct k_itimer *timer, *next;
1316         unsigned long flags;
1317
1318         BUG_ON(!irqs_disabled());
1319
1320         /*
1321          * The fast path checks that there are no expired thread or thread
1322          * group timers.  If that's so, just return.
1323          */
1324         if (!fastpath_timer_check(tsk))
1325                 return;
1326
1327         if (!lock_task_sighand(tsk, &flags))
1328                 return;
1329         /*
1330          * Here we take off tsk->signal->cpu_timers[N] and
1331          * tsk->cpu_timers[N] all the timers that are firing, and
1332          * put them on the firing list.
1333          */
1334         check_thread_timers(tsk, &firing);
1335         /*
1336          * If there are any active process wide timers (POSIX 1.b, itimers,
1337          * RLIMIT_CPU) cputimer must be running.
1338          */
1339         if (tsk->signal->cputimer.running)
1340                 check_process_timers(tsk, &firing);
1341
1342         /*
1343          * We must release these locks before taking any timer's lock.
1344          * There is a potential race with timer deletion here, as the
1345          * siglock now protects our private firing list.  We have set
1346          * the firing flag in each timer, so that a deletion attempt
1347          * that gets the timer lock before we do will give it up and
1348          * spin until we've taken care of that timer below.
1349          */
1350         unlock_task_sighand(tsk, &flags);
1351
1352         /*
1353          * Now that all the timers on our list have the firing flag,
1354          * no one will touch their list entries but us.  We'll take
1355          * each timer's lock before clearing its firing flag, so no
1356          * timer call will interfere.
1357          */
1358         list_for_each_entry_safe(timer, next, &firing, it.cpu.entry) {
1359                 int cpu_firing;
1360
1361                 spin_lock(&timer->it_lock);
1362                 list_del_init(&timer->it.cpu.entry);
1363                 cpu_firing = timer->it.cpu.firing;
1364                 timer->it.cpu.firing = 0;
1365                 /*
1366                  * The firing flag is -1 if we collided with a reset
1367                  * of the timer, which already reported this
1368                  * almost-firing as an overrun.  So don't generate an event.
1369                  */
1370                 if (likely(cpu_firing >= 0))
1371                         cpu_timer_fire(timer);
1372                 spin_unlock(&timer->it_lock);
1373         }
1374
1375         /*
1376          * In case some timers were rescheduled after the queue got emptied,
1377          * wake up full dynticks CPUs.
1378          */
1379         if (tsk->signal->cputimer.running)
1380                 posix_cpu_timer_kick_nohz();
1381 }
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         union cpu_time_count now;
1391
1392         BUG_ON(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.cpu) {
1403                                 /* Just about to fire. */
1404                                 *oldval = cputime_one_jiffy;
1405                         } else {
1406                                 *oldval -= now.cpu;
1407                         }
1408                 }
1409
1410                 if (!*newval)
1411                         goto out;
1412                 *newval += now.cpu;
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.sched == 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 =
1524                 &current_thread_info()->restart_block;
1525         struct itimerspec it;
1526         int error;
1527
1528         /*
1529          * Diagnose required errors first.
1530          */
1531         if (CPUCLOCK_PERTHREAD(which_clock) &&
1532             (CPUCLOCK_PID(which_clock) == 0 ||
1533              CPUCLOCK_PID(which_clock) == current->pid))
1534                 return -EINVAL;
1535
1536         error = do_cpu_nanosleep(which_clock, flags, rqtp, &it);
1537
1538         if (error == -ERESTART_RESTARTBLOCK) {
1539
1540                 if (flags & TIMER_ABSTIME)
1541                         return -ERESTARTNOHAND;
1542                 /*
1543                  * Report back to the user the time still remaining.
1544                  */
1545                 if (rmtp && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1546                         return -EFAULT;
1547
1548                 restart_block->fn = posix_cpu_nsleep_restart;
1549                 restart_block->nanosleep.clockid = which_clock;
1550                 restart_block->nanosleep.rmtp = rmtp;
1551                 restart_block->nanosleep.expires = timespec_to_ns(rqtp);
1552         }
1553         return error;
1554 }
1555
1556 static long posix_cpu_nsleep_restart(struct restart_block *restart_block)
1557 {
1558         clockid_t which_clock = restart_block->nanosleep.clockid;
1559         struct timespec t;
1560         struct itimerspec it;
1561         int error;
1562
1563         t = ns_to_timespec(restart_block->nanosleep.expires);
1564
1565         error = do_cpu_nanosleep(which_clock, TIMER_ABSTIME, &t, &it);
1566
1567         if (error == -ERESTART_RESTARTBLOCK) {
1568                 struct timespec __user *rmtp = restart_block->nanosleep.rmtp;
1569                 /*
1570                  * Report back to the user the time still remaining.
1571                  */
1572                 if (rmtp && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1573                         return -EFAULT;
1574
1575                 restart_block->nanosleep.expires = timespec_to_ns(&t);
1576         }
1577         return error;
1578
1579 }
1580
1581 #define PROCESS_CLOCK   MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED)
1582 #define THREAD_CLOCK    MAKE_THREAD_CPUCLOCK(0, CPUCLOCK_SCHED)
1583
1584 static int process_cpu_clock_getres(const clockid_t which_clock,
1585                                     struct timespec *tp)
1586 {
1587         return posix_cpu_clock_getres(PROCESS_CLOCK, tp);
1588 }
1589 static int process_cpu_clock_get(const clockid_t which_clock,
1590                                  struct timespec *tp)
1591 {
1592         return posix_cpu_clock_get(PROCESS_CLOCK, tp);
1593 }
1594 static int process_cpu_timer_create(struct k_itimer *timer)
1595 {
1596         timer->it_clock = PROCESS_CLOCK;
1597         return posix_cpu_timer_create(timer);
1598 }
1599 static int process_cpu_nsleep(const clockid_t which_clock, int flags,
1600                               struct timespec *rqtp,
1601                               struct timespec __user *rmtp)
1602 {
1603         return posix_cpu_nsleep(PROCESS_CLOCK, flags, rqtp, rmtp);
1604 }
1605 static long process_cpu_nsleep_restart(struct restart_block *restart_block)
1606 {
1607         return -EINVAL;
1608 }
1609 static int thread_cpu_clock_getres(const clockid_t which_clock,
1610                                    struct timespec *tp)
1611 {
1612         return posix_cpu_clock_getres(THREAD_CLOCK, tp);
1613 }
1614 static int thread_cpu_clock_get(const clockid_t which_clock,
1615                                 struct timespec *tp)
1616 {
1617         return posix_cpu_clock_get(THREAD_CLOCK, tp);
1618 }
1619 static int thread_cpu_timer_create(struct k_itimer *timer)
1620 {
1621         timer->it_clock = THREAD_CLOCK;
1622         return posix_cpu_timer_create(timer);
1623 }
1624
1625 struct k_clock clock_posix_cpu = {
1626         .clock_getres   = posix_cpu_clock_getres,
1627         .clock_set      = posix_cpu_clock_set,
1628         .clock_get      = posix_cpu_clock_get,
1629         .timer_create   = posix_cpu_timer_create,
1630         .nsleep         = posix_cpu_nsleep,
1631         .nsleep_restart = posix_cpu_nsleep_restart,
1632         .timer_set      = posix_cpu_timer_set,
1633         .timer_del      = posix_cpu_timer_del,
1634         .timer_get      = posix_cpu_timer_get,
1635 };
1636
1637 static __init int init_posix_cpu_timers(void)
1638 {
1639         struct k_clock process = {
1640                 .clock_getres   = process_cpu_clock_getres,
1641                 .clock_get      = process_cpu_clock_get,
1642                 .timer_create   = process_cpu_timer_create,
1643                 .nsleep         = process_cpu_nsleep,
1644                 .nsleep_restart = process_cpu_nsleep_restart,
1645         };
1646         struct k_clock thread = {
1647                 .clock_getres   = thread_cpu_clock_getres,
1648                 .clock_get      = thread_cpu_clock_get,
1649                 .timer_create   = thread_cpu_timer_create,
1650         };
1651         struct timespec ts;
1652
1653         posix_timers_register_clock(CLOCK_PROCESS_CPUTIME_ID, &process);
1654         posix_timers_register_clock(CLOCK_THREAD_CPUTIME_ID, &thread);
1655
1656         cputime_to_timespec(cputime_one_jiffy, &ts);
1657         onecputick = ts.tv_nsec;
1658         WARN_ON(ts.tv_sec != 0);
1659
1660         return 0;
1661 }
1662 __initcall(init_posix_cpu_timers);