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 <asm/uaccess.h>
   8 #include <linux/errno.h>
   9
  10 static int check_clock(clockid_t which_clock)
  11 {
  12         int error = 0;
  13         struct task_struct *p;
  14         const pid_t pid = CPUCLOCK_PID(which_clock);
  15
  16         if (CPUCLOCK_WHICH(which_clock) >= CPUCLOCK_MAX)
  17                 return -EINVAL;
  18
  19         if (pid == 0)
  20                 return 0;
  21
  22         read_lock(&tasklist_lock);
  23         p = find_task_by_pid(pid);
  24         if (!p || (CPUCLOCK_PERTHREAD(which_clock) ?
  25                    p->tgid != current->tgid : p->tgid != pid)) {
  26                 error = -EINVAL;
  27         }
  28         read_unlock(&tasklist_lock);
  29
  30         return error;
  31 }
  32
  33 static inline union cpu_time_count
  34 timespec_to_sample(clockid_t which_clock, const struct timespec *tp)
  35 {
  36         union cpu_time_count ret;
  37         ret.sched = 0;          /* high half always zero when .cpu used */
  38         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  39                 ret.sched = tp->tv_sec * NSEC_PER_SEC + tp->tv_nsec;
  40         } else {
  41                 ret.cpu = timespec_to_cputime(tp);
  42         }
  43         return ret;
  44 }
  45
  46 static void sample_to_timespec(clockid_t which_clock,
  47                                union cpu_time_count cpu,
  48                                struct timespec *tp)
  49 {
  50         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  51                 tp->tv_sec = div_long_long_rem(cpu.sched,
  52                                                NSEC_PER_SEC, &tp->tv_nsec);
  53         } else {
  54                 cputime_to_timespec(cpu.cpu, tp);
  55         }
  56 }
  57
  58 static inline int cpu_time_before(clockid_t which_clock,
  59                                   union cpu_time_count now,
  60                                   union cpu_time_count then)
  61 {
  62         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  63                 return now.sched < then.sched;
  64         }  else {
  65                 return cputime_lt(now.cpu, then.cpu);
  66         }
  67 }
  68 static inline void cpu_time_add(clockid_t which_clock,
  69                                 union cpu_time_count *acc,
  70                                 union cpu_time_count val)
  71 {
  72         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  73                 acc->sched += val.sched;
  74         }  else {
  75                 acc->cpu = cputime_add(acc->cpu, val.cpu);
  76         }
  77 }
  78 static inline union cpu_time_count cpu_time_sub(clockid_t which_clock,
  79                                                 union cpu_time_count a,
  80                                                 union cpu_time_count b)
  81 {
  82         if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
  83                 a.sched -= b.sched;
  84         }  else {
  85                 a.cpu = cputime_sub(a.cpu, b.cpu);
  86         }
  87         return a;
  88 }
  89
  90 /*
  91  * Update expiry time from increment, and increase overrun count,
  92  * given the current clock sample.
  93  */
  94 static inline void bump_cpu_timer(struct k_itimer *timer,
  95                                   union cpu_time_count now)
  96 {
  97         int i;
  98
  99         if (timer->it.cpu.incr.sched == 0)
 100                 return;
 101
 102         if (CPUCLOCK_WHICH(timer->it_clock) == CPUCLOCK_SCHED) {
 103                 unsigned long long delta, incr;
 104
 105                 if (now.sched < timer->it.cpu.expires.sched)
 106                         return;
 107                 incr = timer->it.cpu.incr.sched;
 108                 delta = now.sched + incr - timer->it.cpu.expires.sched;
 109                 /* Don't use (incr*2 < delta), incr*2 might overflow. */
 110                 for (i = 0; incr < delta - incr; i++)
 111                         incr = incr << 1;
 112                 for (; i >= 0; incr >>= 1, i--) {
 113                         if (delta <= incr)
 114                                 continue;
 115                         timer->it.cpu.expires.sched += incr;
 116                         timer->it_overrun += 1 << i;
 117                         delta -= incr;
 118                 }
 119         } else {
 120                 cputime_t delta, incr;
 121
 122                 if (cputime_lt(now.cpu, timer->it.cpu.expires.cpu))
 123                         return;
 124                 incr = timer->it.cpu.incr.cpu;
 125                 delta = cputime_sub(cputime_add(now.cpu, incr),
 126                                     timer->it.cpu.expires.cpu);
 127                 /* Don't use (incr*2 < delta), incr*2 might overflow. */
 128                 for (i = 0; cputime_lt(incr, cputime_sub(delta, incr)); i++)
 129                              incr = cputime_add(incr, incr);
 130                 for (; i >= 0; incr = cputime_halve(incr), i--) {
 131                         if (cputime_le(delta, incr))
 132                                 continue;
 133                         timer->it.cpu.expires.cpu =
 134                                 cputime_add(timer->it.cpu.expires.cpu, incr);
 135                         timer->it_overrun += 1 << i;
 136                         delta = cputime_sub(delta, incr);
 137                 }
 138         }
 139 }
 140
 141 static inline cputime_t prof_ticks(struct task_struct *p)
 142 {
 143         return cputime_add(p->utime, p->stime);
 144 }
 145 static inline cputime_t virt_ticks(struct task_struct *p)
 146 {
 147         return p->utime;
 148 }
 149 static inline unsigned long long sched_ns(struct task_struct *p)
 150 {
 151         return (p == current) ? current_sched_time(p) : p->sched_time;
 152 }
 153
 154 int posix_cpu_clock_getres(clockid_t which_clock, struct timespec *tp)
 155 {
 156         int error = check_clock(which_clock);
 157         if (!error) {
 158                 tp->tv_sec = 0;
 159                 tp->tv_nsec = ((NSEC_PER_SEC + HZ - 1) / HZ);
 160                 if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
 161                         /*
 162                          * If sched_clock is using a cycle counter, we
 163                          * don't have any idea of its true resolution
 164                          * exported, but it is much more than 1s/HZ.
 165                          */
 166                         tp->tv_nsec = 1;
 167                 }
 168         }
 169         return error;
 170 }
 171
 172 int posix_cpu_clock_set(clockid_t which_clock, const struct timespec *tp)
 173 {
 174         /*
 175          * You can never reset a CPU clock, but we check for other errors
 176          * in the call before failing with EPERM.
 177          */
 178         int error = check_clock(which_clock);
 179         if (error == 0) {
 180                 error = -EPERM;
 181         }
 182         return error;
 183 }
 184
 185
 186 /*
 187  * Sample a per-thread clock for the given task.
 188  */
 189 static int cpu_clock_sample(clockid_t which_clock, struct task_struct *p,
 190                             union cpu_time_count *cpu)
 191 {
 192         switch (CPUCLOCK_WHICH(which_clock)) {
 193         default:
 194                 return -EINVAL;
 195         case CPUCLOCK_PROF:
 196                 cpu->cpu = prof_ticks(p);
 197                 break;
 198         case CPUCLOCK_VIRT:
 199                 cpu->cpu = virt_ticks(p);
 200                 break;
 201         case CPUCLOCK_SCHED:
 202                 cpu->sched = sched_ns(p);
 203                 break;
 204         }
 205         return 0;
 206 }
 207
 208 /*
 209  * Sample a process (thread group) clock for the given group_leader task.
 210  * Must be called with tasklist_lock held for reading.
 211  * Must be called with tasklist_lock held for reading, and p->sighand->siglock.
 212  */
 213 static int cpu_clock_sample_group_locked(unsigned int clock_idx,
 214                                          struct task_struct *p,
 215                                          union cpu_time_count *cpu)
 216 {
 217         struct task_struct *t = p;
 218         switch (clock_idx) {
 219         default:
 220                 return -EINVAL;
 221         case CPUCLOCK_PROF:
 222                 cpu->cpu = cputime_add(p->signal->utime, p->signal->stime);
 223                 do {
 224                         cpu->cpu = cputime_add(cpu->cpu, prof_ticks(t));
 225                         t = next_thread(t);
 226                 } while (t != p);
 227                 break;
 228         case CPUCLOCK_VIRT:
 229                 cpu->cpu = p->signal->utime;
 230                 do {
 231                         cpu->cpu = cputime_add(cpu->cpu, virt_ticks(t));
 232                         t = next_thread(t);
 233                 } while (t != p);
 234                 break;
 235         case CPUCLOCK_SCHED:
 236                 cpu->sched = p->signal->sched_time;
 237                 /* Add in each other live thread.  */
 238                 while ((t = next_thread(t)) != p) {
 239                         cpu->sched += t->sched_time;
 240                 }
 241                 if (p->tgid == current->tgid) {
 242                         /*
 243                          * We're sampling ourselves, so include the
 244                          * cycles not yet banked.  We still omit
 245                          * other threads running on other CPUs,
 246                          * so the total can always be behind as
 247                          * much as max(nthreads-1,ncpus) * (NSEC_PER_SEC/HZ).
 248                          */
 249                         cpu->sched += current_sched_time(current);
 250                 } else {
 251                         cpu->sched += p->sched_time;
 252                 }
 253                 break;
 254         }
 255         return 0;
 256 }
 257
 258 /*
 259  * Sample a process (thread group) clock for the given group_leader task.
 260  * Must be called with tasklist_lock held for reading.
 261  */
 262 static int cpu_clock_sample_group(clockid_t which_clock,
 263                                   struct task_struct *p,
 264                                   union cpu_time_count *cpu)
 265 {
 266         int ret;
 267         unsigned long flags;
 268         spin_lock_irqsave(&p->sighand->siglock, flags);
 269         ret = cpu_clock_sample_group_locked(CPUCLOCK_WHICH(which_clock), p,
 270                                             cpu);
 271         spin_unlock_irqrestore(&p->sighand->siglock, flags);
 272         return ret;
 273 }
 274
 275
 276 int posix_cpu_clock_get(clockid_t which_clock, struct timespec *tp)
 277 {
 278         const pid_t pid = CPUCLOCK_PID(which_clock);
 279         int error = -EINVAL;
 280         union cpu_time_count rtn;
 281
 282         if (pid == 0) {
 283                 /*
 284                  * Special case constant value for our own clocks.
 285                  * We don't have to do any lookup to find ourselves.
 286                  */
 287                 if (CPUCLOCK_PERTHREAD(which_clock)) {
 288                         /*
 289                          * Sampling just ourselves we can do with no locking.
 290                          */
 291                         error = cpu_clock_sample(which_clock,
 292                                                  current, &rtn);
 293                 } else {
 294                         read_lock(&tasklist_lock);
 295                         error = cpu_clock_sample_group(which_clock,
 296                                                        current, &rtn);
 297                         read_unlock(&tasklist_lock);
 298                 }
 299         } else {
 300                 /*
 301                  * Find the given PID, and validate that the caller
 302                  * should be able to see it.
 303                  */
 304                 struct task_struct *p;
 305                 read_lock(&tasklist_lock);
 306                 p = find_task_by_pid(pid);
 307                 if (p) {
 308                         if (CPUCLOCK_PERTHREAD(which_clock)) {
 309                                 if (p->tgid == current->tgid) {
 310                                         error = cpu_clock_sample(which_clock,
 311                                                                  p, &rtn);
 312                                 }
 313                         } else if (p->tgid == pid && p->signal) {
 314                                 error = cpu_clock_sample_group(which_clock,
 315                                                                p, &rtn);
 316                         }
 317                 }
 318                 read_unlock(&tasklist_lock);
 319         }
 320
 321         if (error)
 322                 return error;
 323         sample_to_timespec(which_clock, rtn, tp);
 324         return 0;
 325 }
 326
 327
 328 /*
 329  * Validate the clockid_t for a new CPU-clock timer, and initialize the timer.
 330  * This is called from sys_timer_create with the new timer already locked.
 331  */
 332 int posix_cpu_timer_create(struct k_itimer *new_timer)
 333 {
 334         int ret = 0;
 335         const pid_t pid = CPUCLOCK_PID(new_timer->it_clock);
 336         struct task_struct *p;
 337
 338         if (CPUCLOCK_WHICH(new_timer->it_clock) >= CPUCLOCK_MAX)
 339                 return -EINVAL;
 340
 341         INIT_LIST_HEAD(&new_timer->it.cpu.entry);
 342         new_timer->it.cpu.incr.sched = 0;
 343         new_timer->it.cpu.expires.sched = 0;
 344
 345         read_lock(&tasklist_lock);
 346         if (CPUCLOCK_PERTHREAD(new_timer->it_clock)) {
 347                 if (pid == 0) {
 348                         p = current;
 349                 } else {
 350                         p = find_task_by_pid(pid);
 351                         if (p && p->tgid != current->tgid)
 352                                 p = NULL;
 353                 }
 354         } else {
 355                 if (pid == 0) {
 356                         p = current->group_leader;
 357                 } else {
 358                         p = find_task_by_pid(pid);
 359                         if (p && p->tgid != pid)
 360                                 p = NULL;
 361                 }
 362         }
 363         new_timer->it.cpu.task = p;
 364         if (p) {
 365                 get_task_struct(p);
 366         } else {
 367                 ret = -EINVAL;
 368         }
 369         read_unlock(&tasklist_lock);
 370
 371         return ret;
 372 }
 373
 374 /*
 375  * Clean up a CPU-clock timer that is about to be destroyed.
 376  * This is called from timer deletion with the timer already locked.
 377  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 378  * and try again.  (This happens when the timer is in the middle of firing.)
 379  */
 380 int posix_cpu_timer_del(struct k_itimer *timer)
 381 {
 382         struct task_struct *p = timer->it.cpu.task;
 383
 384         if (timer->it.cpu.firing)
 385                 return TIMER_RETRY;
 386
 387         if (unlikely(p == NULL))
 388                 return 0;
 389
 390         if (!list_empty(&timer->it.cpu.entry)) {
 391                 read_lock(&tasklist_lock);
 392                 if (unlikely(p->signal == NULL)) {
 393                         /*
 394                          * We raced with the reaping of the task.
 395                          * The deletion should have cleared us off the list.
 396                          */
 397                         BUG_ON(!list_empty(&timer->it.cpu.entry));
 398                 } else {
 399                         /*
 400                          * Take us off the task's timer list.
 401                          */
 402                         spin_lock(&p->sighand->siglock);
 403                         list_del(&timer->it.cpu.entry);
 404                         spin_unlock(&p->sighand->siglock);
 405                 }
 406                 read_unlock(&tasklist_lock);
 407         }
 408         put_task_struct(p);
 409
 410         return 0;
 411 }
 412
 413 /*
 414  * Clean out CPU timers still ticking when a thread exited.  The task
 415  * pointer is cleared, and the expiry time is replaced with the residual
 416  * time for later timer_gettime calls to return.
 417  * This must be called with the siglock held.
 418  */
 419 static void cleanup_timers(struct list_head *head,
 420                            cputime_t utime, cputime_t stime,
 421                            unsigned long long sched_time)
 422 {
 423         struct cpu_timer_list *timer, *next;
 424         cputime_t ptime = cputime_add(utime, stime);
 425
 426         list_for_each_entry_safe(timer, next, head, entry) {
 427                 timer->task = NULL;
 428                 list_del_init(&timer->entry);
 429                 if (cputime_lt(timer->expires.cpu, ptime)) {
 430                         timer->expires.cpu = cputime_zero;
 431                 } else {
 432                         timer->expires.cpu = cputime_sub(timer->expires.cpu,
 433                                                          ptime);
 434                 }
 435         }
 436
 437         ++head;
 438         list_for_each_entry_safe(timer, next, head, entry) {
 439                 timer->task = NULL;
 440                 list_del_init(&timer->entry);
 441                 if (cputime_lt(timer->expires.cpu, utime)) {
 442                         timer->expires.cpu = cputime_zero;
 443                 } else {
 444                         timer->expires.cpu = cputime_sub(timer->expires.cpu,
 445                                                          utime);
 446                 }
 447         }
 448
 449         ++head;
 450         list_for_each_entry_safe(timer, next, head, entry) {
 451                 timer->task = NULL;
 452                 list_del_init(&timer->entry);
 453                 if (timer->expires.sched < sched_time) {
 454                         timer->expires.sched = 0;
 455                 } else {
 456                         timer->expires.sched -= sched_time;
 457                 }
 458         }
 459 }
 460
 461 /*
 462  * These are both called with the siglock held, when the current thread
 463  * is being reaped.  When the final (leader) thread in the group is reaped,
 464  * posix_cpu_timers_exit_group will be called after posix_cpu_timers_exit.
 465  */
 466 void posix_cpu_timers_exit(struct task_struct *tsk)
 467 {
 468         cleanup_timers(tsk->cpu_timers,
 469                        tsk->utime, tsk->stime, tsk->sched_time);
 470
 471 }
 472 void posix_cpu_timers_exit_group(struct task_struct *tsk)
 473 {
 474         cleanup_timers(tsk->signal->cpu_timers,
 475                        cputime_add(tsk->utime, tsk->signal->utime),
 476                        cputime_add(tsk->stime, tsk->signal->stime),
 477                        tsk->sched_time + tsk->signal->sched_time);
 478 }
 479
 480
 481 /*
 482  * Set the expiry times of all the threads in the process so one of them
 483  * will go off before the process cumulative expiry total is reached.
 484  */
 485 static void process_timer_rebalance(struct task_struct *p,
 486                                     unsigned int clock_idx,
 487                                     union cpu_time_count expires,
 488                                     union cpu_time_count val)
 489 {
 490         cputime_t ticks, left;
 491         unsigned long long ns, nsleft;
 492         struct task_struct *t = p;
 493         unsigned int nthreads = atomic_read(&p->signal->live);
 494
 495         switch (clock_idx) {
 496         default:
 497                 BUG();
 498                 break;
 499         case CPUCLOCK_PROF:
 500                 left = cputime_div(cputime_sub(expires.cpu, val.cpu),
 501                                    nthreads);
 502                 do {
 503                         if (!unlikely(t->exit_state)) {
 504                                 ticks = cputime_add(prof_ticks(t), left);
 505                                 if (cputime_eq(t->it_prof_expires,
 506                                                cputime_zero) ||
 507                                     cputime_gt(t->it_prof_expires, ticks)) {
 508                                         t->it_prof_expires = ticks;
 509                                 }
 510                         }
 511                         t = next_thread(t);
 512                 } while (t != p);
 513                 break;
 514         case CPUCLOCK_VIRT:
 515                 left = cputime_div(cputime_sub(expires.cpu, val.cpu),
 516                                    nthreads);
 517                 do {
 518                         if (!unlikely(t->exit_state)) {
 519                                 ticks = cputime_add(virt_ticks(t), left);
 520                                 if (cputime_eq(t->it_virt_expires,
 521                                                cputime_zero) ||
 522                                     cputime_gt(t->it_virt_expires, ticks)) {
 523                                         t->it_virt_expires = ticks;
 524                                 }
 525                         }
 526                         t = next_thread(t);
 527                 } while (t != p);
 528                 break;
 529         case CPUCLOCK_SCHED:
 530                 nsleft = expires.sched - val.sched;
 531                 do_div(nsleft, nthreads);
 532                 do {
 533                         if (!unlikely(t->exit_state)) {
 534                                 ns = t->sched_time + nsleft;
 535                                 if (t->it_sched_expires == 0 ||
 536                                     t->it_sched_expires > ns) {
 537                                         t->it_sched_expires = ns;
 538                                 }
 539                         }
 540                         t = next_thread(t);
 541                 } while (t != p);
 542                 break;
 543         }
 544 }
 545
 546 static void clear_dead_task(struct k_itimer *timer, union cpu_time_count now)
 547 {
 548         /*
 549          * That's all for this thread or process.
 550          * We leave our residual in expires to be reported.
 551          */
 552         put_task_struct(timer->it.cpu.task);
 553         timer->it.cpu.task = NULL;
 554         timer->it.cpu.expires = cpu_time_sub(timer->it_clock,
 555                                              timer->it.cpu.expires,
 556                                              now);
 557 }
 558
 559 /*
 560  * Insert the timer on the appropriate list before any timers that
 561  * expire later.  This must be called with the tasklist_lock held
 562  * for reading, and interrupts disabled.
 563  */
 564 static void arm_timer(struct k_itimer *timer, union cpu_time_count now)
 565 {
 566         struct task_struct *p = timer->it.cpu.task;
 567         struct list_head *head, *listpos;
 568         struct cpu_timer_list *const nt = &timer->it.cpu;
 569         struct cpu_timer_list *next;
 570         unsigned long i;
 571
 572         head = (CPUCLOCK_PERTHREAD(timer->it_clock) ?
 573                 p->cpu_timers : p->signal->cpu_timers);
 574         head += CPUCLOCK_WHICH(timer->it_clock);
 575
 576         BUG_ON(!irqs_disabled());
 577         spin_lock(&p->sighand->siglock);
 578
 579         listpos = head;
 580         if (CPUCLOCK_WHICH(timer->it_clock) == CPUCLOCK_SCHED) {
 581                 list_for_each_entry(next, head, entry) {
 582                         if (next->expires.sched > nt->expires.sched) {
 583                                 listpos = &next->entry;
 584                                 break;
 585                         }
 586                 }
 587         } else {
 588                 list_for_each_entry(next, head, entry) {
 589                         if (cputime_gt(next->expires.cpu, nt->expires.cpu)) {
 590                                 listpos = &next->entry;
 591                                 break;
 592                         }
 593                 }
 594         }
 595         list_add(&nt->entry, listpos);
 596
 597         if (listpos == head) {
 598                 /*
 599                  * We are the new earliest-expiring timer.
 600                  * If we are a thread timer, there can always
 601                  * be a process timer telling us to stop earlier.
 602                  */
 603
 604                 if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 605                         switch (CPUCLOCK_WHICH(timer->it_clock)) {
 606                         default:
 607                                 BUG();
 608                         case CPUCLOCK_PROF:
 609                                 if (cputime_eq(p->it_prof_expires,
 610                                                cputime_zero) ||
 611                                     cputime_gt(p->it_prof_expires,
 612                                                nt->expires.cpu))
 613                                         p->it_prof_expires = nt->expires.cpu;
 614                                 break;
 615                         case CPUCLOCK_VIRT:
 616                                 if (cputime_eq(p->it_virt_expires,
 617                                                cputime_zero) ||
 618                                     cputime_gt(p->it_virt_expires,
 619                                                nt->expires.cpu))
 620                                         p->it_virt_expires = nt->expires.cpu;
 621                                 break;
 622                         case CPUCLOCK_SCHED:
 623                                 if (p->it_sched_expires == 0 ||
 624                                     p->it_sched_expires > nt->expires.sched)
 625                                         p->it_sched_expires = nt->expires.sched;
 626                                 break;
 627                         }
 628                 } else {
 629                         /*
 630                          * For a process timer, we must balance
 631                          * all the live threads' expirations.
 632                          */
 633                         switch (CPUCLOCK_WHICH(timer->it_clock)) {
 634                         default:
 635                                 BUG();
 636                         case CPUCLOCK_VIRT:
 637                                 if (!cputime_eq(p->signal->it_virt_expires,
 638                                                 cputime_zero) &&
 639                                     cputime_lt(p->signal->it_virt_expires,
 640                                                timer->it.cpu.expires.cpu))
 641                                         break;
 642                                 goto rebalance;
 643                         case CPUCLOCK_PROF:
 644                                 if (!cputime_eq(p->signal->it_prof_expires,
 645                                                 cputime_zero) &&
 646                                     cputime_lt(p->signal->it_prof_expires,
 647                                                timer->it.cpu.expires.cpu))
 648                                         break;
 649                                 i = p->signal->rlim[RLIMIT_CPU].rlim_cur;
 650                                 if (i != RLIM_INFINITY &&
 651                                     i <= cputime_to_secs(timer->it.cpu.expires.cpu))
 652                                         break;
 653                                 goto rebalance;
 654                         case CPUCLOCK_SCHED:
 655                         rebalance:
 656                                 process_timer_rebalance(
 657                                         timer->it.cpu.task,
 658                                         CPUCLOCK_WHICH(timer->it_clock),
 659                                         timer->it.cpu.expires, now);
 660                                 break;
 661                         }
 662                 }
 663         }
 664
 665         spin_unlock(&p->sighand->siglock);
 666 }
 667
 668 /*
 669  * The timer is locked, fire it and arrange for its reload.
 670  */
 671 static void cpu_timer_fire(struct k_itimer *timer)
 672 {
 673         if (unlikely(timer->sigq == NULL)) {
 674                 /*
 675                  * This a special case for clock_nanosleep,
 676                  * not a normal timer from sys_timer_create.
 677                  */
 678                 wake_up_process(timer->it_process);
 679                 timer->it.cpu.expires.sched = 0;
 680         } else if (timer->it.cpu.incr.sched == 0) {
 681                 /*
 682                  * One-shot timer.  Clear it as soon as it's fired.
 683                  */
 684                 posix_timer_event(timer, 0);
 685                 timer->it.cpu.expires.sched = 0;
 686         } else if (posix_timer_event(timer, ++timer->it_requeue_pending)) {
 687                 /*
 688                  * The signal did not get queued because the signal
 689                  * was ignored, so we won't get any callback to
 690                  * reload the timer.  But we need to keep it
 691                  * ticking in case the signal is deliverable next time.
 692                  */
 693                 posix_cpu_timer_schedule(timer);
 694         }
 695 }
 696
 697 /*
 698  * Guts of sys_timer_settime for CPU timers.
 699  * This is called with the timer locked and interrupts disabled.
 700  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 701  * and try again.  (This happens when the timer is in the middle of firing.)
 702  */
 703 int posix_cpu_timer_set(struct k_itimer *timer, int flags,
 704                         struct itimerspec *new, struct itimerspec *old)
 705 {
 706         struct task_struct *p = timer->it.cpu.task;
 707         union cpu_time_count old_expires, new_expires, val;
 708         int ret;
 709
 710         if (unlikely(p == NULL)) {
 711                 /*
 712                  * Timer refers to a dead task's clock.
 713                  */
 714                 return -ESRCH;
 715         }
 716
 717         new_expires = timespec_to_sample(timer->it_clock, &new->it_value);
 718
 719         read_lock(&tasklist_lock);
 720         /*
 721          * We need the tasklist_lock to protect against reaping that
 722          * clears p->signal.  If p has just been reaped, we can no
 723          * longer get any information about it at all.
 724          */
 725         if (unlikely(p->signal == NULL)) {
 726                 read_unlock(&tasklist_lock);
 727                 put_task_struct(p);
 728                 timer->it.cpu.task = NULL;
 729                 return -ESRCH;
 730         }
 731
 732         /*
 733          * Disarm any old timer after extracting its expiry time.
 734          */
 735         BUG_ON(!irqs_disabled());
 736         spin_lock(&p->sighand->siglock);
 737         old_expires = timer->it.cpu.expires;
 738         list_del_init(&timer->it.cpu.entry);
 739         spin_unlock(&p->sighand->siglock);
 740
 741         /*
 742          * We need to sample the current value to convert the new
 743          * value from to relative and absolute, and to convert the
 744          * old value from absolute to relative.  To set a process
 745          * timer, we need a sample to balance the thread expiry
 746          * times (in arm_timer).  With an absolute time, we must
 747          * check if it's already passed.  In short, we need a sample.
 748          */
 749         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 750                 cpu_clock_sample(timer->it_clock, p, &val);
 751         } else {
 752                 cpu_clock_sample_group(timer->it_clock, p, &val);
 753         }
 754
 755         if (old) {
 756                 if (old_expires.sched == 0) {
 757                         old->it_value.tv_sec = 0;
 758                         old->it_value.tv_nsec = 0;
 759                 } else {
 760                         /*
 761                          * Update the timer in case it has
 762                          * overrun already.  If it has,
 763                          * we'll report it as having overrun
 764                          * and with the next reloaded timer
 765                          * already ticking, though we are
 766                          * swallowing that pending
 767                          * notification here to install the
 768                          * new setting.
 769                          */
 770                         bump_cpu_timer(timer, val);
 771                         if (cpu_time_before(timer->it_clock, val,
 772                                             timer->it.cpu.expires)) {
 773                                 old_expires = cpu_time_sub(
 774                                         timer->it_clock,
 775                                         timer->it.cpu.expires, val);
 776                                 sample_to_timespec(timer->it_clock,
 777                                                    old_expires,
 778                                                    &old->it_value);
 779                         } else {
 780                                 old->it_value.tv_nsec = 1;
 781                                 old->it_value.tv_sec = 0;
 782                         }
 783                 }
 784         }
 785
 786         if (unlikely(timer->it.cpu.firing)) {
 787                 /*
 788                  * We are colliding with the timer actually firing.
 789                  * Punt after filling in the timer's old value, and
 790                  * disable this firing since we are already reporting
 791                  * it as an overrun (thanks to bump_cpu_timer above).
 792                  */
 793                 read_unlock(&tasklist_lock);
 794                 timer->it.cpu.firing = -1;
 795                 ret = TIMER_RETRY;
 796                 goto out;
 797         }
 798
 799         if (new_expires.sched != 0 && !(flags & TIMER_ABSTIME)) {
 800                 cpu_time_add(timer->it_clock, &new_expires, val);
 801         }
 802
 803         /*
 804          * Install the new expiry time (or zero).
 805          * For a timer with no notification action, we don't actually
 806          * arm the timer (we'll just fake it for timer_gettime).
 807          */
 808         timer->it.cpu.expires = new_expires;
 809         if (new_expires.sched != 0 &&
 810             (timer->it_sigev_notify & ~SIGEV_THREAD_ID) != SIGEV_NONE &&
 811             cpu_time_before(timer->it_clock, val, new_expires)) {
 812                 arm_timer(timer, val);
 813         }
 814
 815         read_unlock(&tasklist_lock);
 816
 817         /*
 818          * Install the new reload setting, and
 819          * set up the signal and overrun bookkeeping.
 820          */
 821         timer->it.cpu.incr = timespec_to_sample(timer->it_clock,
 822                                                 &new->it_interval);
 823
 824         /*
 825          * This acts as a modification timestamp for the timer,
 826          * so any automatic reload attempt will punt on seeing
 827          * that we have reset the timer manually.
 828          */
 829         timer->it_requeue_pending = (timer->it_requeue_pending + 2) &
 830                 ~REQUEUE_PENDING;
 831         timer->it_overrun_last = 0;
 832         timer->it_overrun = -1;
 833
 834         if (new_expires.sched != 0 &&
 835             (timer->it_sigev_notify & ~SIGEV_THREAD_ID) != SIGEV_NONE &&
 836             !cpu_time_before(timer->it_clock, val, new_expires)) {
 837                 /*
 838                  * The designated time already passed, so we notify
 839                  * immediately, even if the thread never runs to
 840                  * accumulate more time on this clock.
 841                  */
 842                 cpu_timer_fire(timer);
 843         }
 844
 845         ret = 0;
 846  out:
 847         if (old) {
 848                 sample_to_timespec(timer->it_clock,
 849                                    timer->it.cpu.incr, &old->it_interval);
 850         }
 851         return ret;
 852 }
 853
 854 void posix_cpu_timer_get(struct k_itimer *timer, struct itimerspec *itp)
 855 {
 856         union cpu_time_count now;
 857         struct task_struct *p = timer->it.cpu.task;
 858         int clear_dead;
 859
 860         /*
 861          * Easy part: convert the reload time.
 862          */
 863         sample_to_timespec(timer->it_clock,
 864                            timer->it.cpu.incr, &itp->it_interval);
 865
 866         if (timer->it.cpu.expires.sched == 0) { /* Timer not armed at all.  */
 867                 itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
 868                 return;
 869         }
 870
 871         if (unlikely(p == NULL)) {
 872                 /*
 873                  * This task already died and the timer will never fire.
 874                  * In this case, expires is actually the dead value.
 875                  */
 876         dead:
 877                 sample_to_timespec(timer->it_clock, timer->it.cpu.expires,
 878                                    &itp->it_value);
 879                 return;
 880         }
 881
 882         /*
 883          * Sample the clock to take the difference with the expiry time.
 884          */
 885         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 886                 cpu_clock_sample(timer->it_clock, p, &now);
 887                 clear_dead = p->exit_state;
 888         } else {
 889                 read_lock(&tasklist_lock);
 890                 if (unlikely(p->signal == NULL)) {
 891                         /*
 892                          * The process has been reaped.
 893                          * We can't even collect a sample any more.
 894                          * Call the timer disarmed, nothing else to do.
 895                          */
 896                         put_task_struct(p);
 897                         timer->it.cpu.task = NULL;
 898                         timer->it.cpu.expires.sched = 0;
 899                         read_unlock(&tasklist_lock);
 900                         goto dead;
 901                 } else {
 902                         cpu_clock_sample_group(timer->it_clock, p, &now);
 903                         clear_dead = (unlikely(p->exit_state) &&
 904                                       thread_group_empty(p));
 905                 }
 906                 read_unlock(&tasklist_lock);
 907         }
 908
 909         if ((timer->it_sigev_notify & ~SIGEV_THREAD_ID) == SIGEV_NONE) {
 910                 if (timer->it.cpu.incr.sched == 0 &&
 911                     cpu_time_before(timer->it_clock,
 912                                     timer->it.cpu.expires, now)) {
 913                         /*
 914                          * Do-nothing timer expired and has no reload,
 915                          * so it's as if it was never set.
 916                          */
 917                         timer->it.cpu.expires.sched = 0;
 918                         itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
 919                         return;
 920                 }
 921                 /*
 922                  * Account for any expirations and reloads that should
 923                  * have happened.
 924                  */
 925                 bump_cpu_timer(timer, now);
 926         }
 927
 928         if (unlikely(clear_dead)) {
 929                 /*
 930                  * We've noticed that the thread is dead, but
 931                  * not yet reaped.  Take this opportunity to
 932                  * drop our task ref.
 933                  */
 934                 clear_dead_task(timer, now);
 935                 goto dead;
 936         }
 937
 938         if (cpu_time_before(timer->it_clock, now, timer->it.cpu.expires)) {
 939                 sample_to_timespec(timer->it_clock,
 940                                    cpu_time_sub(timer->it_clock,
 941                                                 timer->it.cpu.expires, now),
 942                                    &itp->it_value);
 943         } else {
 944                 /*
 945                  * The timer should have expired already, but the firing
 946                  * hasn't taken place yet.  Say it's just about to expire.
 947                  */
 948                 itp->it_value.tv_nsec = 1;
 949                 itp->it_value.tv_sec = 0;
 950         }
 951 }
 952
 953 /*
 954  * Check for any per-thread CPU timers that have fired and move them off
 955  * the tsk->cpu_timers[N] list onto the firing list.  Here we update the
 956  * tsk->it_*_expires values to reflect the remaining thread CPU timers.
 957  */
 958 static void check_thread_timers(struct task_struct *tsk,
 959                                 struct list_head *firing)
 960 {
 961         struct list_head *timers = tsk->cpu_timers;
 962
 963         tsk->it_prof_expires = cputime_zero;
 964         while (!list_empty(timers)) {
 965                 struct cpu_timer_list *t = list_entry(timers->next,
 966                                                       struct cpu_timer_list,
 967                                                       entry);
 968                 if (cputime_lt(prof_ticks(tsk), t->expires.cpu)) {
 969                         tsk->it_prof_expires = t->expires.cpu;
 970                         break;
 971                 }
 972                 t->firing = 1;
 973                 list_move_tail(&t->entry, firing);
 974         }
 975
 976         ++timers;
 977         tsk->it_virt_expires = cputime_zero;
 978         while (!list_empty(timers)) {
 979                 struct cpu_timer_list *t = list_entry(timers->next,
 980                                                       struct cpu_timer_list,
 981                                                       entry);
 982                 if (cputime_lt(virt_ticks(tsk), t->expires.cpu)) {
 983                         tsk->it_virt_expires = t->expires.cpu;
 984                         break;
 985                 }
 986                 t->firing = 1;
 987                 list_move_tail(&t->entry, firing);
 988         }
 989
 990         ++timers;
 991         tsk->it_sched_expires = 0;
 992         while (!list_empty(timers)) {
 993                 struct cpu_timer_list *t = list_entry(timers->next,
 994                                                       struct cpu_timer_list,
 995                                                       entry);
 996                 if (tsk->sched_time < t->expires.sched) {
 997                         tsk->it_sched_expires = t->expires.sched;
 998                         break;
 999                 }
1000                 t->firing = 1;
1001                 list_move_tail(&t->entry, firing);
1002         }
1003 }
1004
1005 /*
1006  * Check for any per-thread CPU timers that have fired and move them
1007  * off the tsk->*_timers list onto the firing list.  Per-thread timers
1008  * have already been taken off.
1009  */
1010 static void check_process_timers(struct task_struct *tsk,
1011                                  struct list_head *firing)
1012 {
1013         struct signal_struct *const sig = tsk->signal;
1014         cputime_t utime, stime, ptime, virt_expires, prof_expires;
1015         unsigned long long sched_time, sched_expires;
1016         struct task_struct *t;
1017         struct list_head *timers = sig->cpu_timers;
1018
1019         /*
1020          * Don't sample the current process CPU clocks if there are no timers.
1021          */
1022         if (list_empty(&timers[CPUCLOCK_PROF]) &&
1023             cputime_eq(sig->it_prof_expires, cputime_zero) &&
1024             sig->rlim[RLIMIT_CPU].rlim_cur == RLIM_INFINITY &&
1025             list_empty(&timers[CPUCLOCK_VIRT]) &&
1026             cputime_eq(sig->it_virt_expires, cputime_zero) &&
1027             list_empty(&timers[CPUCLOCK_SCHED]))
1028                 return;
1029
1030         /*
1031          * Collect the current process totals.
1032          */
1033         utime = sig->utime;
1034         stime = sig->stime;
1035         sched_time = sig->sched_time;
1036         t = tsk;
1037         do {
1038                 utime = cputime_add(utime, t->utime);
1039                 stime = cputime_add(stime, t->stime);
1040                 sched_time += t->sched_time;
1041                 t = next_thread(t);
1042         } while (t != tsk);
1043         ptime = cputime_add(utime, stime);
1044
1045         prof_expires = cputime_zero;
1046         while (!list_empty(timers)) {
1047                 struct cpu_timer_list *t = list_entry(timers->next,
1048                                                       struct cpu_timer_list,
1049                                                       entry);
1050                 if (cputime_lt(ptime, t->expires.cpu)) {
1051                         prof_expires = t->expires.cpu;
1052                         break;
1053                 }
1054                 t->firing = 1;
1055                 list_move_tail(&t->entry, firing);
1056         }
1057
1058         ++timers;
1059         virt_expires = cputime_zero;
1060         while (!list_empty(timers)) {
1061                 struct cpu_timer_list *t = list_entry(timers->next,
1062                                                       struct cpu_timer_list,
1063                                                       entry);
1064                 if (cputime_lt(utime, t->expires.cpu)) {
1065                         virt_expires = t->expires.cpu;
1066                         break;
1067                 }
1068                 t->firing = 1;
1069                 list_move_tail(&t->entry, firing);
1070         }
1071
1072         ++timers;
1073         sched_expires = 0;
1074         while (!list_empty(timers)) {
1075                 struct cpu_timer_list *t = list_entry(timers->next,
1076                                                       struct cpu_timer_list,
1077                                                       entry);
1078                 if (sched_time < t->expires.sched) {
1079                         sched_expires = t->expires.sched;
1080                         break;
1081                 }
1082                 t->firing = 1;
1083                 list_move_tail(&t->entry, firing);
1084         }
1085
1086         /*
1087          * Check for the special case process timers.
1088          */
1089         if (!cputime_eq(sig->it_prof_expires, cputime_zero)) {
1090                 if (cputime_ge(ptime, sig->it_prof_expires)) {
1091                         /* ITIMER_PROF fires and reloads.  */
1092                         sig->it_prof_expires = sig->it_prof_incr;
1093                         if (!cputime_eq(sig->it_prof_expires, cputime_zero)) {
1094                                 sig->it_prof_expires = cputime_add(
1095                                         sig->it_prof_expires, ptime);
1096                         }
1097                         __group_send_sig_info(SIGPROF, SEND_SIG_PRIV, tsk);
1098                 }
1099                 if (!cputime_eq(sig->it_prof_expires, cputime_zero) &&
1100                     (cputime_eq(prof_expires, cputime_zero) ||
1101                      cputime_lt(sig->it_prof_expires, prof_expires))) {
1102                         prof_expires = sig->it_prof_expires;
1103                 }
1104         }
1105         if (!cputime_eq(sig->it_virt_expires, cputime_zero)) {
1106                 if (cputime_ge(utime, sig->it_virt_expires)) {
1107                         /* ITIMER_VIRTUAL fires and reloads.  */
1108                         sig->it_virt_expires = sig->it_virt_incr;
1109                         if (!cputime_eq(sig->it_virt_expires, cputime_zero)) {
1110                                 sig->it_virt_expires = cputime_add(
1111                                         sig->it_virt_expires, utime);
1112                         }
1113                         __group_send_sig_info(SIGVTALRM, SEND_SIG_PRIV, tsk);
1114                 }
1115                 if (!cputime_eq(sig->it_virt_expires, cputime_zero) &&
1116                     (cputime_eq(virt_expires, cputime_zero) ||
1117                      cputime_lt(sig->it_virt_expires, virt_expires))) {
1118                         virt_expires = sig->it_virt_expires;
1119                 }
1120         }
1121         if (sig->rlim[RLIMIT_CPU].rlim_cur != RLIM_INFINITY) {
1122                 unsigned long psecs = cputime_to_secs(ptime);
1123                 cputime_t x;
1124                 if (psecs >= sig->rlim[RLIMIT_CPU].rlim_max) {
1125                         /*
1126                          * At the hard limit, we just die.
1127                          * No need to calculate anything else now.
1128                          */
1129                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
1130                         return;
1131                 }
1132                 if (psecs >= sig->rlim[RLIMIT_CPU].rlim_cur) {
1133                         /*
1134                          * At the soft limit, send a SIGXCPU every second.
1135                          */
1136                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
1137                         if (sig->rlim[RLIMIT_CPU].rlim_cur
1138                             < sig->rlim[RLIMIT_CPU].rlim_max) {
1139                                 sig->rlim[RLIMIT_CPU].rlim_cur++;
1140                         }
1141                 }
1142                 x = secs_to_cputime(sig->rlim[RLIMIT_CPU].rlim_cur);
1143                 if (cputime_eq(prof_expires, cputime_zero) ||
1144                     cputime_lt(x, prof_expires)) {
1145                         prof_expires = x;
1146                 }
1147         }
1148
1149         if (!cputime_eq(prof_expires, cputime_zero) ||
1150             !cputime_eq(virt_expires, cputime_zero) ||
1151             sched_expires != 0) {
1152                 /*
1153                  * Rebalance the threads' expiry times for the remaining
1154                  * process CPU timers.
1155                  */
1156
1157                 cputime_t prof_left, virt_left, ticks;
1158                 unsigned long long sched_left, sched;
1159                 const unsigned int nthreads = atomic_read(&sig->live);
1160
1161                 prof_left = cputime_sub(prof_expires, utime);
1162                 prof_left = cputime_sub(prof_left, stime);
1163                 prof_left = cputime_div(prof_left, nthreads);
1164                 virt_left = cputime_sub(virt_expires, utime);
1165                 virt_left = cputime_div(virt_left, nthreads);
1166                 if (sched_expires) {
1167                         sched_left = sched_expires - sched_time;
1168                         do_div(sched_left, nthreads);
1169                 } else {
1170                         sched_left = 0;
1171                 }
1172                 t = tsk;
1173                 do {
1174                         ticks = cputime_add(cputime_add(t->utime, t->stime),
1175                                             prof_left);
1176                         if (!cputime_eq(prof_expires, cputime_zero) &&
1177                             (cputime_eq(t->it_prof_expires, cputime_zero) ||
1178                              cputime_gt(t->it_prof_expires, ticks))) {
1179                                 t->it_prof_expires = ticks;
1180                         }
1181
1182                         ticks = cputime_add(t->utime, virt_left);
1183                         if (!cputime_eq(virt_expires, cputime_zero) &&
1184                             (cputime_eq(t->it_virt_expires, cputime_zero) ||
1185                              cputime_gt(t->it_virt_expires, ticks))) {
1186                                 t->it_virt_expires = ticks;
1187                         }
1188
1189                         sched = t->sched_time + sched_left;
1190                         if (sched_expires && (t->it_sched_expires == 0 ||
1191                                               t->it_sched_expires > sched)) {
1192                                 t->it_sched_expires = sched;
1193                         }
1194
1195                         do {
1196                                 t = next_thread(t);
1197                         } while (unlikely(t->exit_state));
1198                 } while (t != tsk);
1199         }
1200 }
1201
1202 /*
1203  * This is called from the signal code (via do_schedule_next_timer)
1204  * when the last timer signal was delivered and we have to reload the timer.
1205  */
1206 void posix_cpu_timer_schedule(struct k_itimer *timer)
1207 {
1208         struct task_struct *p = timer->it.cpu.task;
1209         union cpu_time_count now;
1210
1211         if (unlikely(p == NULL))
1212                 /*
1213                  * The task was cleaned up already, no future firings.
1214                  */
1215                 return;
1216
1217         /*
1218          * Fetch the current sample and update the timer's expiry time.
1219          */
1220         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
1221                 cpu_clock_sample(timer->it_clock, p, &now);
1222                 bump_cpu_timer(timer, now);
1223                 if (unlikely(p->exit_state)) {
1224                         clear_dead_task(timer, now);
1225                         return;
1226                 }
1227                 read_lock(&tasklist_lock); /* arm_timer needs it.  */
1228         } else {
1229                 read_lock(&tasklist_lock);
1230                 if (unlikely(p->signal == NULL)) {
1231                         /*
1232                          * The process has been reaped.
1233                          * We can't even collect a sample any more.
1234                          */
1235                         put_task_struct(p);
1236                         timer->it.cpu.task = p = NULL;
1237                         timer->it.cpu.expires.sched = 0;
1238                         read_unlock(&tasklist_lock);
1239                         return;
1240                 } else if (unlikely(p->exit_state) && thread_group_empty(p)) {
1241                         /*
1242                          * We've noticed that the thread is dead, but
1243                          * not yet reaped.  Take this opportunity to
1244                          * drop our task ref.
1245                          */
1246                         clear_dead_task(timer, now);
1247                         read_unlock(&tasklist_lock);
1248                         return;
1249                 }
1250                 cpu_clock_sample_group(timer->it_clock, p, &now);
1251                 bump_cpu_timer(timer, now);
1252                 /* Leave the tasklist_lock locked for the call below.  */
1253         }
1254
1255         /*
1256          * Now re-arm for the new expiry time.
1257          */
1258         arm_timer(timer, now);
1259
1260         read_unlock(&tasklist_lock);
1261 }
1262
1263 /*
1264  * This is called from the timer interrupt handler.  The irq handler has
1265  * already updated our counts.  We need to check if any timers fire now.
1266  * Interrupts are disabled.
1267  */
1268 void run_posix_cpu_timers(struct task_struct *tsk)
1269 {
1270         LIST_HEAD(firing);
1271         struct k_itimer *timer, *next;
1272
1273         BUG_ON(!irqs_disabled());
1274
1275 #define UNEXPIRED(clock) \
1276                 (cputime_eq(tsk->it_##clock##_expires, cputime_zero) || \
1277                  cputime_lt(clock##_ticks(tsk), tsk->it_##clock##_expires))
1278
1279         if (UNEXPIRED(prof) && UNEXPIRED(virt) &&
1280             (tsk->it_sched_expires == 0 ||
1281              tsk->sched_time < tsk->it_sched_expires))
1282                 return;
1283
1284 #undef  UNEXPIRED
1285
1286         BUG_ON(tsk->exit_state);
1287
1288         /*
1289          * Double-check with locks held.
1290          */
1291         read_lock(&tasklist_lock);
1292         spin_lock(&tsk->sighand->siglock);
1293
1294         /*
1295          * Here we take off tsk->cpu_timers[N] and tsk->signal->cpu_timers[N]
1296          * all the timers that are firing, and put them on the firing list.
1297          */
1298         check_thread_timers(tsk, &firing);
1299         check_process_timers(tsk, &firing);
1300
1301         /*
1302          * We must release these locks before taking any timer's lock.
1303          * There is a potential race with timer deletion here, as the
1304          * siglock now protects our private firing list.  We have set
1305          * the firing flag in each timer, so that a deletion attempt
1306          * that gets the timer lock before we do will give it up and
1307          * spin until we've taken care of that timer below.
1308          */
1309         spin_unlock(&tsk->sighand->siglock);
1310         read_unlock(&tasklist_lock);
1311
1312         /*
1313          * Now that all the timers on our list have the firing flag,
1314          * noone will touch their list entries but us.  We'll take
1315          * each timer's lock before clearing its firing flag, so no
1316          * timer call will interfere.
1317          */
1318         list_for_each_entry_safe(timer, next, &firing, it.cpu.entry) {
1319                 int firing;
1320                 spin_lock(&timer->it_lock);
1321                 list_del_init(&timer->it.cpu.entry);
1322                 firing = timer->it.cpu.firing;
1323                 timer->it.cpu.firing = 0;
1324                 /*
1325                  * The firing flag is -1 if we collided with a reset
1326                  * of the timer, which already reported this
1327                  * almost-firing as an overrun.  So don't generate an event.
1328                  */
1329                 if (likely(firing >= 0)) {
1330                         cpu_timer_fire(timer);
1331                 }
1332                 spin_unlock(&timer->it_lock);
1333         }
1334 }
1335
1336 /*
1337  * Set one of the process-wide special case CPU timers.
1338  * The tasklist_lock and tsk->sighand->siglock must be held by the caller.
1339  * The oldval argument is null for the RLIMIT_CPU timer, where *newval is
1340  * absolute; non-null for ITIMER_*, where *newval is relative and we update
1341  * it to be absolute, *oldval is absolute and we update it to be relative.
1342  */
1343 void set_process_cpu_timer(struct task_struct *tsk, unsigned int clock_idx,
1344                            cputime_t *newval, cputime_t *oldval)
1345 {
1346         union cpu_time_count now;
1347         struct list_head *head;
1348
1349         BUG_ON(clock_idx == CPUCLOCK_SCHED);
1350         cpu_clock_sample_group_locked(clock_idx, tsk, &now);
1351
1352         if (oldval) {
1353                 if (!cputime_eq(*oldval, cputime_zero)) {
1354                         if (cputime_le(*oldval, now.cpu)) {
1355                                 /* Just about to fire. */
1356                                 *oldval = jiffies_to_cputime(1);
1357                         } else {
1358                                 *oldval = cputime_sub(*oldval, now.cpu);
1359                         }
1360                 }
1361
1362                 if (cputime_eq(*newval, cputime_zero))
1363                         return;
1364                 *newval = cputime_add(*newval, now.cpu);
1365
1366                 /*
1367                  * If the RLIMIT_CPU timer will expire before the
1368                  * ITIMER_PROF timer, we have nothing else to do.
1369                  */
1370                 if (tsk->signal->rlim[RLIMIT_CPU].rlim_cur
1371                     < cputime_to_secs(*newval))
1372                         return;
1373         }
1374
1375         /*
1376          * Check whether there are any process timers already set to fire
1377          * before this one.  If so, we don't have anything more to do.
1378          */
1379         head = &tsk->signal->cpu_timers[clock_idx];
1380         if (list_empty(head) ||
1381             cputime_ge(list_entry(head->next,
1382                                   struct cpu_timer_list, entry)->expires.cpu,
1383                        *newval)) {
1384                 /*
1385                  * Rejigger each thread's expiry time so that one will
1386                  * notice before we hit the process-cumulative expiry time.
1387                  */
1388                 union cpu_time_count expires = { .sched = 0 };
1389                 expires.cpu = *newval;
1390                 process_timer_rebalance(tsk, clock_idx, expires, now);
1391         }
1392 }
1393
1394 static long posix_cpu_clock_nanosleep_restart(struct restart_block *);
1395
1396 int posix_cpu_nsleep(clockid_t which_clock, int flags,
1397                      struct timespec *rqtp)
1398 {
1399         struct restart_block *restart_block =
1400             &current_thread_info()->restart_block;
1401         struct k_itimer timer;
1402         int error;
1403
1404         /*
1405          * Diagnose required errors first.
1406          */
1407         if (CPUCLOCK_PERTHREAD(which_clock) &&
1408             (CPUCLOCK_PID(which_clock) == 0 ||
1409              CPUCLOCK_PID(which_clock) == current->pid))
1410                 return -EINVAL;
1411
1412         /*
1413          * Set up a temporary timer and then wait for it to go off.
1414          */
1415         memset(&timer, 0, sizeof timer);
1416         spin_lock_init(&timer.it_lock);
1417         timer.it_clock = which_clock;
1418         timer.it_overrun = -1;
1419         error = posix_cpu_timer_create(&timer);
1420         timer.it_process = current;
1421         if (!error) {
1422                 struct timespec __user *rmtp;
1423                 static struct itimerspec zero_it;
1424                 struct itimerspec it = { .it_value = *rqtp,
1425                                          .it_interval = {} };
1426
1427                 spin_lock_irq(&timer.it_lock);
1428                 error = posix_cpu_timer_set(&timer, flags, &it, NULL);
1429                 if (error) {
1430                         spin_unlock_irq(&timer.it_lock);
1431                         return error;
1432                 }
1433
1434                 while (!signal_pending(current)) {
1435                         if (timer.it.cpu.expires.sched == 0) {
1436                                 /*
1437                                  * Our timer fired and was reset.
1438                                  */
1439                                 spin_unlock_irq(&timer.it_lock);
1440                                 return 0;
1441                         }
1442
1443                         /*
1444                          * Block until cpu_timer_fire (or a signal) wakes us.
1445                          */
1446                         __set_current_state(TASK_INTERRUPTIBLE);
1447                         spin_unlock_irq(&timer.it_lock);
1448                         schedule();
1449                         spin_lock_irq(&timer.it_lock);
1450                 }
1451
1452                 /*
1453                  * We were interrupted by a signal.
1454                  */
1455                 sample_to_timespec(which_clock, timer.it.cpu.expires, rqtp);
1456                 posix_cpu_timer_set(&timer, 0, &zero_it, &it);
1457                 spin_unlock_irq(&timer.it_lock);
1458
1459                 if ((it.it_value.tv_sec | it.it_value.tv_nsec) == 0) {
1460                         /*
1461                          * It actually did fire already.
1462                          */
1463                         return 0;
1464                 }
1465
1466                 /*
1467                  * Report back to the user the time still remaining.
1468                  */
1469                 rmtp = (struct timespec __user *) restart_block->arg1;
1470                 if (rmtp != NULL && !(flags & TIMER_ABSTIME) &&
1471                     copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1472                         return -EFAULT;
1473
1474                 restart_block->fn = posix_cpu_clock_nanosleep_restart;
1475                 /* Caller already set restart_block->arg1 */
1476                 restart_block->arg0 = which_clock;
1477                 restart_block->arg2 = rqtp->tv_sec;
1478                 restart_block->arg3 = rqtp->tv_nsec;
1479
1480                 error = -ERESTART_RESTARTBLOCK;
1481         }
1482
1483         return error;
1484 }
1485
1486 static long
1487 posix_cpu_clock_nanosleep_restart(struct restart_block *restart_block)
1488 {
1489         clockid_t which_clock = restart_block->arg0;
1490         struct timespec t = { .tv_sec = restart_block->arg2,
1491                               .tv_nsec = restart_block->arg3 };
1492         restart_block->fn = do_no_restart_syscall;
1493         return posix_cpu_nsleep(which_clock, TIMER_ABSTIME, &t);
1494 }
1495
1496
1497 #define PROCESS_CLOCK   MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED)
1498 #define THREAD_CLOCK    MAKE_THREAD_CPUCLOCK(0, CPUCLOCK_SCHED)
1499
1500 static int process_cpu_clock_getres(clockid_t which_clock, struct timespec *tp)
1501 {
1502         return posix_cpu_clock_getres(PROCESS_CLOCK, tp);
1503 }
1504 static int process_cpu_clock_get(clockid_t which_clock, struct timespec *tp)
1505 {
1506         return posix_cpu_clock_get(PROCESS_CLOCK, tp);
1507 }
1508 static int process_cpu_timer_create(struct k_itimer *timer)
1509 {
1510         timer->it_clock = PROCESS_CLOCK;
1511         return posix_cpu_timer_create(timer);
1512 }
1513 static int process_cpu_nsleep(clockid_t which_clock, int flags,
1514                               struct timespec *rqtp)
1515 {
1516         return posix_cpu_nsleep(PROCESS_CLOCK, flags, rqtp);
1517 }
1518 static int thread_cpu_clock_getres(clockid_t which_clock, struct timespec *tp)
1519 {
1520         return posix_cpu_clock_getres(THREAD_CLOCK, tp);
1521 }
1522 static int thread_cpu_clock_get(clockid_t which_clock, struct timespec *tp)
1523 {
1524         return posix_cpu_clock_get(THREAD_CLOCK, tp);
1525 }
1526 static int thread_cpu_timer_create(struct k_itimer *timer)
1527 {
1528         timer->it_clock = THREAD_CLOCK;
1529         return posix_cpu_timer_create(timer);
1530 }
1531 static int thread_cpu_nsleep(clockid_t which_clock, int flags,
1532                               struct timespec *rqtp)
1533 {
1534         return -EINVAL;
1535 }
1536
1537 static __init int init_posix_cpu_timers(void)
1538 {
1539         struct k_clock process = {
1540                 .clock_getres = process_cpu_clock_getres,
1541                 .clock_get = process_cpu_clock_get,
1542                 .clock_set = do_posix_clock_nosettime,
1543                 .timer_create = process_cpu_timer_create,
1544                 .nsleep = process_cpu_nsleep,
1545         };
1546         struct k_clock thread = {
1547                 .clock_getres = thread_cpu_clock_getres,
1548                 .clock_get = thread_cpu_clock_get,
1549                 .clock_set = do_posix_clock_nosettime,
1550                 .timer_create = thread_cpu_timer_create,
1551                 .nsleep = thread_cpu_nsleep,
1552         };
1553
1554         register_posix_clock(CLOCK_PROCESS_CPUTIME_ID, &process);
1555         register_posix_clock(CLOCK_THREAD_CPUTIME_ID, &thread);
1556
1557         return 0;
1558 }
1559 __initcall(init_posix_cpu_timers);