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(const 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(const 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 = (unsigned long long)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(const 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(const 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(const 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(const 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 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_lt(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(const 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(const 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(const 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                 cpu->sched += sched_ns(p);
 242                 break;
 243         }
 244         return 0;
 245 }
 246
 247 /*
 248  * Sample a process (thread group) clock for the given group_leader task.
 249  * Must be called with tasklist_lock held for reading.
 250  */
 251 static int cpu_clock_sample_group(const clockid_t which_clock,
 252                                   struct task_struct *p,
 253                                   union cpu_time_count *cpu)
 254 {
 255         int ret;
 256         unsigned long flags;
 257         spin_lock_irqsave(&p->sighand->siglock, flags);
 258         ret = cpu_clock_sample_group_locked(CPUCLOCK_WHICH(which_clock), p,
 259                                             cpu);
 260         spin_unlock_irqrestore(&p->sighand->siglock, flags);
 261         return ret;
 262 }
 263
 264
 265 int posix_cpu_clock_get(const clockid_t which_clock, struct timespec *tp)
 266 {
 267         const pid_t pid = CPUCLOCK_PID(which_clock);
 268         int error = -EINVAL;
 269         union cpu_time_count rtn;
 270
 271         if (pid == 0) {
 272                 /*
 273                  * Special case constant value for our own clocks.
 274                  * We don't have to do any lookup to find ourselves.
 275                  */
 276                 if (CPUCLOCK_PERTHREAD(which_clock)) {
 277                         /*
 278                          * Sampling just ourselves we can do with no locking.
 279                          */
 280                         error = cpu_clock_sample(which_clock,
 281                                                  current, &rtn);
 282                 } else {
 283                         read_lock(&tasklist_lock);
 284                         error = cpu_clock_sample_group(which_clock,
 285                                                        current, &rtn);
 286                         read_unlock(&tasklist_lock);
 287                 }
 288         } else {
 289                 /*
 290                  * Find the given PID, and validate that the caller
 291                  * should be able to see it.
 292                  */
 293                 struct task_struct *p;
 294                 read_lock(&tasklist_lock);
 295                 p = find_task_by_pid(pid);
 296                 if (p) {
 297                         if (CPUCLOCK_PERTHREAD(which_clock)) {
 298                                 if (p->tgid == current->tgid) {
 299                                         error = cpu_clock_sample(which_clock,
 300                                                                  p, &rtn);
 301                                 }
 302                         } else if (p->tgid == pid && p->signal) {
 303                                 error = cpu_clock_sample_group(which_clock,
 304                                                                p, &rtn);
 305                         }
 306                 }
 307                 read_unlock(&tasklist_lock);
 308         }
 309
 310         if (error)
 311                 return error;
 312         sample_to_timespec(which_clock, rtn, tp);
 313         return 0;
 314 }
 315
 316
 317 /*
 318  * Validate the clockid_t for a new CPU-clock timer, and initialize the timer.
 319  * This is called from sys_timer_create with the new timer already locked.
 320  */
 321 int posix_cpu_timer_create(struct k_itimer *new_timer)
 322 {
 323         int ret = 0;
 324         const pid_t pid = CPUCLOCK_PID(new_timer->it_clock);
 325         struct task_struct *p;
 326
 327         if (CPUCLOCK_WHICH(new_timer->it_clock) >= CPUCLOCK_MAX)
 328                 return -EINVAL;
 329
 330         INIT_LIST_HEAD(&new_timer->it.cpu.entry);
 331         new_timer->it.cpu.incr.sched = 0;
 332         new_timer->it.cpu.expires.sched = 0;
 333
 334         read_lock(&tasklist_lock);
 335         if (CPUCLOCK_PERTHREAD(new_timer->it_clock)) {
 336                 if (pid == 0) {
 337                         p = current;
 338                 } else {
 339                         p = find_task_by_pid(pid);
 340                         if (p && p->tgid != current->tgid)
 341                                 p = NULL;
 342                 }
 343         } else {
 344                 if (pid == 0) {
 345                         p = current->group_leader;
 346                 } else {
 347                         p = find_task_by_pid(pid);
 348                         if (p && p->tgid != pid)
 349                                 p = NULL;
 350                 }
 351         }
 352         new_timer->it.cpu.task = p;
 353         if (p) {
 354                 get_task_struct(p);
 355         } else {
 356                 ret = -EINVAL;
 357         }
 358         read_unlock(&tasklist_lock);
 359
 360         return ret;
 361 }
 362
 363 /*
 364  * Clean up a CPU-clock timer that is about to be destroyed.
 365  * This is called from timer deletion with the timer already locked.
 366  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 367  * and try again.  (This happens when the timer is in the middle of firing.)
 368  */
 369 int posix_cpu_timer_del(struct k_itimer *timer)
 370 {
 371         struct task_struct *p = timer->it.cpu.task;
 372         int ret = 0;
 373
 374         if (likely(p != NULL)) {
 375                 read_lock(&tasklist_lock);
 376                 if (unlikely(p->signal == NULL)) {
 377                         /*
 378                          * We raced with the reaping of the task.
 379                          * The deletion should have cleared us off the list.
 380                          */
 381                         BUG_ON(!list_empty(&timer->it.cpu.entry));
 382                 } else {
 383                         spin_lock(&p->sighand->siglock);
 384                         if (timer->it.cpu.firing)
 385                                 ret = TIMER_RETRY;
 386                         else
 387                                 list_del(&timer->it.cpu.entry);
 388                         spin_unlock(&p->sighand->siglock);
 389                 }
 390                 read_unlock(&tasklist_lock);
 391
 392                 if (!ret)
 393                         put_task_struct(p);
 394         }
 395
 396         return ret;
 397 }
 398
 399 /*
 400  * Clean out CPU timers still ticking when a thread exited.  The task
 401  * pointer is cleared, and the expiry time is replaced with the residual
 402  * time for later timer_gettime calls to return.
 403  * This must be called with the siglock held.
 404  */
 405 static void cleanup_timers(struct list_head *head,
 406                            cputime_t utime, cputime_t stime,
 407                            unsigned long long sched_time)
 408 {
 409         struct cpu_timer_list *timer, *next;
 410         cputime_t ptime = cputime_add(utime, stime);
 411
 412         list_for_each_entry_safe(timer, next, head, entry) {
 413                 list_del_init(&timer->entry);
 414                 if (cputime_lt(timer->expires.cpu, ptime)) {
 415                         timer->expires.cpu = cputime_zero;
 416                 } else {
 417                         timer->expires.cpu = cputime_sub(timer->expires.cpu,
 418                                                          ptime);
 419                 }
 420         }
 421
 422         ++head;
 423         list_for_each_entry_safe(timer, next, head, entry) {
 424                 list_del_init(&timer->entry);
 425                 if (cputime_lt(timer->expires.cpu, utime)) {
 426                         timer->expires.cpu = cputime_zero;
 427                 } else {
 428                         timer->expires.cpu = cputime_sub(timer->expires.cpu,
 429                                                          utime);
 430                 }
 431         }
 432
 433         ++head;
 434         list_for_each_entry_safe(timer, next, head, entry) {
 435                 list_del_init(&timer->entry);
 436                 if (timer->expires.sched < sched_time) {
 437                         timer->expires.sched = 0;
 438                 } else {
 439                         timer->expires.sched -= sched_time;
 440                 }
 441         }
 442 }
 443
 444 /*
 445  * These are both called with the siglock held, when the current thread
 446  * is being reaped.  When the final (leader) thread in the group is reaped,
 447  * posix_cpu_timers_exit_group will be called after posix_cpu_timers_exit.
 448  */
 449 void posix_cpu_timers_exit(struct task_struct *tsk)
 450 {
 451         cleanup_timers(tsk->cpu_timers,
 452                        tsk->utime, tsk->stime, tsk->sched_time);
 453
 454 }
 455 void posix_cpu_timers_exit_group(struct task_struct *tsk)
 456 {
 457         cleanup_timers(tsk->signal->cpu_timers,
 458                        cputime_add(tsk->utime, tsk->signal->utime),
 459                        cputime_add(tsk->stime, tsk->signal->stime),
 460                        tsk->sched_time + tsk->signal->sched_time);
 461 }
 462
 463
 464 /*
 465  * Set the expiry times of all the threads in the process so one of them
 466  * will go off before the process cumulative expiry total is reached.
 467  */
 468 static void process_timer_rebalance(struct task_struct *p,
 469                                     unsigned int clock_idx,
 470                                     union cpu_time_count expires,
 471                                     union cpu_time_count val)
 472 {
 473         cputime_t ticks, left;
 474         unsigned long long ns, nsleft;
 475         struct task_struct *t = p;
 476         unsigned int nthreads = atomic_read(&p->signal->live);
 477
 478         if (!nthreads)
 479                 return;
 480
 481         switch (clock_idx) {
 482         default:
 483                 BUG();
 484                 break;
 485         case CPUCLOCK_PROF:
 486                 left = cputime_div(cputime_sub(expires.cpu, val.cpu),
 487                                    nthreads);
 488                 do {
 489                         if (likely(!(t->flags & PF_EXITING))) {
 490                                 ticks = cputime_add(prof_ticks(t), left);
 491                                 if (cputime_eq(t->it_prof_expires,
 492                                                cputime_zero) ||
 493                                     cputime_gt(t->it_prof_expires, ticks)) {
 494                                         t->it_prof_expires = ticks;
 495                                 }
 496                         }
 497                         t = next_thread(t);
 498                 } while (t != p);
 499                 break;
 500         case CPUCLOCK_VIRT:
 501                 left = cputime_div(cputime_sub(expires.cpu, val.cpu),
 502                                    nthreads);
 503                 do {
 504                         if (likely(!(t->flags & PF_EXITING))) {
 505                                 ticks = cputime_add(virt_ticks(t), left);
 506                                 if (cputime_eq(t->it_virt_expires,
 507                                                cputime_zero) ||
 508                                     cputime_gt(t->it_virt_expires, ticks)) {
 509                                         t->it_virt_expires = ticks;
 510                                 }
 511                         }
 512                         t = next_thread(t);
 513                 } while (t != p);
 514                 break;
 515         case CPUCLOCK_SCHED:
 516                 nsleft = expires.sched - val.sched;
 517                 do_div(nsleft, nthreads);
 518                 do {
 519                         if (likely(!(t->flags & PF_EXITING))) {
 520                                 ns = t->sched_time + nsleft;
 521                                 if (t->it_sched_expires == 0 ||
 522                                     t->it_sched_expires > ns) {
 523                                         t->it_sched_expires = ns;
 524                                 }
 525                         }
 526                         t = next_thread(t);
 527                 } while (t != p);
 528                 break;
 529         }
 530 }
 531
 532 static void clear_dead_task(struct k_itimer *timer, union cpu_time_count now)
 533 {
 534         /*
 535          * That's all for this thread or process.
 536          * We leave our residual in expires to be reported.
 537          */
 538         put_task_struct(timer->it.cpu.task);
 539         timer->it.cpu.task = NULL;
 540         timer->it.cpu.expires = cpu_time_sub(timer->it_clock,
 541                                              timer->it.cpu.expires,
 542                                              now);
 543 }
 544
 545 /*
 546  * Insert the timer on the appropriate list before any timers that
 547  * expire later.  This must be called with the tasklist_lock held
 548  * for reading, and interrupts disabled.
 549  */
 550 static void arm_timer(struct k_itimer *timer, union cpu_time_count now)
 551 {
 552         struct task_struct *p = timer->it.cpu.task;
 553         struct list_head *head, *listpos;
 554         struct cpu_timer_list *const nt = &timer->it.cpu;
 555         struct cpu_timer_list *next;
 556         unsigned long i;
 557
 558         head = (CPUCLOCK_PERTHREAD(timer->it_clock) ?
 559                 p->cpu_timers : p->signal->cpu_timers);
 560         head += CPUCLOCK_WHICH(timer->it_clock);
 561
 562         BUG_ON(!irqs_disabled());
 563         spin_lock(&p->sighand->siglock);
 564
 565         listpos = head;
 566         if (CPUCLOCK_WHICH(timer->it_clock) == CPUCLOCK_SCHED) {
 567                 list_for_each_entry(next, head, entry) {
 568                         if (next->expires.sched > nt->expires.sched)
 569                                 break;
 570                         listpos = &next->entry;
 571                 }
 572         } else {
 573                 list_for_each_entry(next, head, entry) {
 574                         if (cputime_gt(next->expires.cpu, nt->expires.cpu))
 575                                 break;
 576                         listpos = &next->entry;
 577                 }
 578         }
 579         list_add(&nt->entry, listpos);
 580
 581         if (listpos == head) {
 582                 /*
 583                  * We are the new earliest-expiring timer.
 584                  * If we are a thread timer, there can always
 585                  * be a process timer telling us to stop earlier.
 586                  */
 587
 588                 if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 589                         switch (CPUCLOCK_WHICH(timer->it_clock)) {
 590                         default:
 591                                 BUG();
 592                         case CPUCLOCK_PROF:
 593                                 if (cputime_eq(p->it_prof_expires,
 594                                                cputime_zero) ||
 595                                     cputime_gt(p->it_prof_expires,
 596                                                nt->expires.cpu))
 597                                         p->it_prof_expires = nt->expires.cpu;
 598                                 break;
 599                         case CPUCLOCK_VIRT:
 600                                 if (cputime_eq(p->it_virt_expires,
 601                                                cputime_zero) ||
 602                                     cputime_gt(p->it_virt_expires,
 603                                                nt->expires.cpu))
 604                                         p->it_virt_expires = nt->expires.cpu;
 605                                 break;
 606                         case CPUCLOCK_SCHED:
 607                                 if (p->it_sched_expires == 0 ||
 608                                     p->it_sched_expires > nt->expires.sched)
 609                                         p->it_sched_expires = nt->expires.sched;
 610                                 break;
 611                         }
 612                 } else {
 613                         /*
 614                          * For a process timer, we must balance
 615                          * all the live threads' expirations.
 616                          */
 617                         switch (CPUCLOCK_WHICH(timer->it_clock)) {
 618                         default:
 619                                 BUG();
 620                         case CPUCLOCK_VIRT:
 621                                 if (!cputime_eq(p->signal->it_virt_expires,
 622                                                 cputime_zero) &&
 623                                     cputime_lt(p->signal->it_virt_expires,
 624                                                timer->it.cpu.expires.cpu))
 625                                         break;
 626                                 goto rebalance;
 627                         case CPUCLOCK_PROF:
 628                                 if (!cputime_eq(p->signal->it_prof_expires,
 629                                                 cputime_zero) &&
 630                                     cputime_lt(p->signal->it_prof_expires,
 631                                                timer->it.cpu.expires.cpu))
 632                                         break;
 633                                 i = p->signal->rlim[RLIMIT_CPU].rlim_cur;
 634                                 if (i != RLIM_INFINITY &&
 635                                     i <= cputime_to_secs(timer->it.cpu.expires.cpu))
 636                                         break;
 637                                 goto rebalance;
 638                         case CPUCLOCK_SCHED:
 639                         rebalance:
 640                                 process_timer_rebalance(
 641                                         timer->it.cpu.task,
 642                                         CPUCLOCK_WHICH(timer->it_clock),
 643                                         timer->it.cpu.expires, now);
 644                                 break;
 645                         }
 646                 }
 647         }
 648
 649         spin_unlock(&p->sighand->siglock);
 650 }
 651
 652 /*
 653  * The timer is locked, fire it and arrange for its reload.
 654  */
 655 static void cpu_timer_fire(struct k_itimer *timer)
 656 {
 657         if (unlikely(timer->sigq == NULL)) {
 658                 /*
 659                  * This a special case for clock_nanosleep,
 660                  * not a normal timer from sys_timer_create.
 661                  */
 662                 wake_up_process(timer->it_process);
 663                 timer->it.cpu.expires.sched = 0;
 664         } else if (timer->it.cpu.incr.sched == 0) {
 665                 /*
 666                  * One-shot timer.  Clear it as soon as it's fired.
 667                  */
 668                 posix_timer_event(timer, 0);
 669                 timer->it.cpu.expires.sched = 0;
 670         } else if (posix_timer_event(timer, ++timer->it_requeue_pending)) {
 671                 /*
 672                  * The signal did not get queued because the signal
 673                  * was ignored, so we won't get any callback to
 674                  * reload the timer.  But we need to keep it
 675                  * ticking in case the signal is deliverable next time.
 676                  */
 677                 posix_cpu_timer_schedule(timer);
 678         }
 679 }
 680
 681 /*
 682  * Guts of sys_timer_settime for CPU timers.
 683  * This is called with the timer locked and interrupts disabled.
 684  * If we return TIMER_RETRY, it's necessary to release the timer's lock
 685  * and try again.  (This happens when the timer is in the middle of firing.)
 686  */
 687 int posix_cpu_timer_set(struct k_itimer *timer, int flags,
 688                         struct itimerspec *new, struct itimerspec *old)
 689 {
 690         struct task_struct *p = timer->it.cpu.task;
 691         union cpu_time_count old_expires, new_expires, val;
 692         int ret;
 693
 694         if (unlikely(p == NULL)) {
 695                 /*
 696                  * Timer refers to a dead task's clock.
 697                  */
 698                 return -ESRCH;
 699         }
 700
 701         new_expires = timespec_to_sample(timer->it_clock, &new->it_value);
 702
 703         read_lock(&tasklist_lock);
 704         /*
 705          * We need the tasklist_lock to protect against reaping that
 706          * clears p->signal.  If p has just been reaped, we can no
 707          * longer get any information about it at all.
 708          */
 709         if (unlikely(p->signal == NULL)) {
 710                 read_unlock(&tasklist_lock);
 711                 put_task_struct(p);
 712                 timer->it.cpu.task = NULL;
 713                 return -ESRCH;
 714         }
 715
 716         /*
 717          * Disarm any old timer after extracting its expiry time.
 718          */
 719         BUG_ON(!irqs_disabled());
 720
 721         ret = 0;
 722         spin_lock(&p->sighand->siglock);
 723         old_expires = timer->it.cpu.expires;
 724         if (unlikely(timer->it.cpu.firing)) {
 725                 timer->it.cpu.firing = -1;
 726                 ret = TIMER_RETRY;
 727         } else
 728                 list_del_init(&timer->it.cpu.entry);
 729         spin_unlock(&p->sighand->siglock);
 730
 731         /*
 732          * We need to sample the current value to convert the new
 733          * value from to relative and absolute, and to convert the
 734          * old value from absolute to relative.  To set a process
 735          * timer, we need a sample to balance the thread expiry
 736          * times (in arm_timer).  With an absolute time, we must
 737          * check if it's already passed.  In short, we need a sample.
 738          */
 739         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 740                 cpu_clock_sample(timer->it_clock, p, &val);
 741         } else {
 742                 cpu_clock_sample_group(timer->it_clock, p, &val);
 743         }
 744
 745         if (old) {
 746                 if (old_expires.sched == 0) {
 747                         old->it_value.tv_sec = 0;
 748                         old->it_value.tv_nsec = 0;
 749                 } else {
 750                         /*
 751                          * Update the timer in case it has
 752                          * overrun already.  If it has,
 753                          * we'll report it as having overrun
 754                          * and with the next reloaded timer
 755                          * already ticking, though we are
 756                          * swallowing that pending
 757                          * notification here to install the
 758                          * new setting.
 759                          */
 760                         bump_cpu_timer(timer, val);
 761                         if (cpu_time_before(timer->it_clock, val,
 762                                             timer->it.cpu.expires)) {
 763                                 old_expires = cpu_time_sub(
 764                                         timer->it_clock,
 765                                         timer->it.cpu.expires, val);
 766                                 sample_to_timespec(timer->it_clock,
 767                                                    old_expires,
 768                                                    &old->it_value);
 769                         } else {
 770                                 old->it_value.tv_nsec = 1;
 771                                 old->it_value.tv_sec = 0;
 772                         }
 773                 }
 774         }
 775
 776         if (unlikely(ret)) {
 777                 /*
 778                  * We are colliding with the timer actually firing.
 779                  * Punt after filling in the timer's old value, and
 780                  * disable this firing since we are already reporting
 781                  * it as an overrun (thanks to bump_cpu_timer above).
 782                  */
 783                 read_unlock(&tasklist_lock);
 784                 goto out;
 785         }
 786
 787         if (new_expires.sched != 0 && !(flags & TIMER_ABSTIME)) {
 788                 cpu_time_add(timer->it_clock, &new_expires, val);
 789         }
 790
 791         /*
 792          * Install the new expiry time (or zero).
 793          * For a timer with no notification action, we don't actually
 794          * arm the timer (we'll just fake it for timer_gettime).
 795          */
 796         timer->it.cpu.expires = new_expires;
 797         if (new_expires.sched != 0 &&
 798             (timer->it_sigev_notify & ~SIGEV_THREAD_ID) != SIGEV_NONE &&
 799             cpu_time_before(timer->it_clock, val, new_expires)) {
 800                 arm_timer(timer, val);
 801         }
 802
 803         read_unlock(&tasklist_lock);
 804
 805         /*
 806          * Install the new reload setting, and
 807          * set up the signal and overrun bookkeeping.
 808          */
 809         timer->it.cpu.incr = timespec_to_sample(timer->it_clock,
 810                                                 &new->it_interval);
 811
 812         /*
 813          * This acts as a modification timestamp for the timer,
 814          * so any automatic reload attempt will punt on seeing
 815          * that we have reset the timer manually.
 816          */
 817         timer->it_requeue_pending = (timer->it_requeue_pending + 2) &
 818                 ~REQUEUE_PENDING;
 819         timer->it_overrun_last = 0;
 820         timer->it_overrun = -1;
 821
 822         if (new_expires.sched != 0 &&
 823             (timer->it_sigev_notify & ~SIGEV_THREAD_ID) != SIGEV_NONE &&
 824             !cpu_time_before(timer->it_clock, val, new_expires)) {
 825                 /*
 826                  * The designated time already passed, so we notify
 827                  * immediately, even if the thread never runs to
 828                  * accumulate more time on this clock.
 829                  */
 830                 cpu_timer_fire(timer);
 831         }
 832
 833         ret = 0;
 834  out:
 835         if (old) {
 836                 sample_to_timespec(timer->it_clock,
 837                                    timer->it.cpu.incr, &old->it_interval);
 838         }
 839         return ret;
 840 }
 841
 842 void posix_cpu_timer_get(struct k_itimer *timer, struct itimerspec *itp)
 843 {
 844         union cpu_time_count now;
 845         struct task_struct *p = timer->it.cpu.task;
 846         int clear_dead;
 847
 848         /*
 849          * Easy part: convert the reload time.
 850          */
 851         sample_to_timespec(timer->it_clock,
 852                            timer->it.cpu.incr, &itp->it_interval);
 853
 854         if (timer->it.cpu.expires.sched == 0) { /* Timer not armed at all.  */
 855                 itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
 856                 return;
 857         }
 858
 859         if (unlikely(p == NULL)) {
 860                 /*
 861                  * This task already died and the timer will never fire.
 862                  * In this case, expires is actually the dead value.
 863                  */
 864         dead:
 865                 sample_to_timespec(timer->it_clock, timer->it.cpu.expires,
 866                                    &itp->it_value);
 867                 return;
 868         }
 869
 870         /*
 871          * Sample the clock to take the difference with the expiry time.
 872          */
 873         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
 874                 cpu_clock_sample(timer->it_clock, p, &now);
 875                 clear_dead = p->exit_state;
 876         } else {
 877                 read_lock(&tasklist_lock);
 878                 if (unlikely(p->signal == NULL)) {
 879                         /*
 880                          * The process has been reaped.
 881                          * We can't even collect a sample any more.
 882                          * Call the timer disarmed, nothing else to do.
 883                          */
 884                         put_task_struct(p);
 885                         timer->it.cpu.task = NULL;
 886                         timer->it.cpu.expires.sched = 0;
 887                         read_unlock(&tasklist_lock);
 888                         goto dead;
 889                 } else {
 890                         cpu_clock_sample_group(timer->it_clock, p, &now);
 891                         clear_dead = (unlikely(p->exit_state) &&
 892                                       thread_group_empty(p));
 893                 }
 894                 read_unlock(&tasklist_lock);
 895         }
 896
 897         if ((timer->it_sigev_notify & ~SIGEV_THREAD_ID) == SIGEV_NONE) {
 898                 if (timer->it.cpu.incr.sched == 0 &&
 899                     cpu_time_before(timer->it_clock,
 900                                     timer->it.cpu.expires, now)) {
 901                         /*
 902                          * Do-nothing timer expired and has no reload,
 903                          * so it's as if it was never set.
 904                          */
 905                         timer->it.cpu.expires.sched = 0;
 906                         itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
 907                         return;
 908                 }
 909                 /*
 910                  * Account for any expirations and reloads that should
 911                  * have happened.
 912                  */
 913                 bump_cpu_timer(timer, now);
 914         }
 915
 916         if (unlikely(clear_dead)) {
 917                 /*
 918                  * We've noticed that the thread is dead, but
 919                  * not yet reaped.  Take this opportunity to
 920                  * drop our task ref.
 921                  */
 922                 clear_dead_task(timer, now);
 923                 goto dead;
 924         }
 925
 926         if (cpu_time_before(timer->it_clock, now, timer->it.cpu.expires)) {
 927                 sample_to_timespec(timer->it_clock,
 928                                    cpu_time_sub(timer->it_clock,
 929                                                 timer->it.cpu.expires, now),
 930                                    &itp->it_value);
 931         } else {
 932                 /*
 933                  * The timer should have expired already, but the firing
 934                  * hasn't taken place yet.  Say it's just about to expire.
 935                  */
 936                 itp->it_value.tv_nsec = 1;
 937                 itp->it_value.tv_sec = 0;
 938         }
 939 }
 940
 941 /*
 942  * Check for any per-thread CPU timers that have fired and move them off
 943  * the tsk->cpu_timers[N] list onto the firing list.  Here we update the
 944  * tsk->it_*_expires values to reflect the remaining thread CPU timers.
 945  */
 946 static void check_thread_timers(struct task_struct *tsk,
 947                                 struct list_head *firing)
 948 {
 949         int maxfire;
 950         struct list_head *timers = tsk->cpu_timers;
 951
 952         maxfire = 20;
 953         tsk->it_prof_expires = cputime_zero;
 954         while (!list_empty(timers)) {
 955                 struct cpu_timer_list *t = list_entry(timers->next,
 956                                                       struct cpu_timer_list,
 957                                                       entry);
 958                 if (!--maxfire || cputime_lt(prof_ticks(tsk), t->expires.cpu)) {
 959                         tsk->it_prof_expires = t->expires.cpu;
 960                         break;
 961                 }
 962                 t->firing = 1;
 963                 list_move_tail(&t->entry, firing);
 964         }
 965
 966         ++timers;
 967         maxfire = 20;
 968         tsk->it_virt_expires = cputime_zero;
 969         while (!list_empty(timers)) {
 970                 struct cpu_timer_list *t = list_entry(timers->next,
 971                                                       struct cpu_timer_list,
 972                                                       entry);
 973                 if (!--maxfire || cputime_lt(virt_ticks(tsk), t->expires.cpu)) {
 974                         tsk->it_virt_expires = t->expires.cpu;
 975                         break;
 976                 }
 977                 t->firing = 1;
 978                 list_move_tail(&t->entry, firing);
 979         }
 980
 981         ++timers;
 982         maxfire = 20;
 983         tsk->it_sched_expires = 0;
 984         while (!list_empty(timers)) {
 985                 struct cpu_timer_list *t = list_entry(timers->next,
 986                                                       struct cpu_timer_list,
 987                                                       entry);
 988                 if (!--maxfire || tsk->sched_time < t->expires.sched) {
 989                         tsk->it_sched_expires = t->expires.sched;
 990                         break;
 991                 }
 992                 t->firing = 1;
 993                 list_move_tail(&t->entry, firing);
 994         }
 995 }
 996
 997 /*
 998  * Check for any per-thread CPU timers that have fired and move them
 999  * off the tsk->*_timers list onto the firing list.  Per-thread timers
1000  * have already been taken off.
1001  */
1002 static void check_process_timers(struct task_struct *tsk,
1003                                  struct list_head *firing)
1004 {
1005         int maxfire;
1006         struct signal_struct *const sig = tsk->signal;
1007         cputime_t utime, stime, ptime, virt_expires, prof_expires;
1008         unsigned long long sched_time, sched_expires;
1009         struct task_struct *t;
1010         struct list_head *timers = sig->cpu_timers;
1011
1012         /*
1013          * Don't sample the current process CPU clocks if there are no timers.
1014          */
1015         if (list_empty(&timers[CPUCLOCK_PROF]) &&
1016             cputime_eq(sig->it_prof_expires, cputime_zero) &&
1017             sig->rlim[RLIMIT_CPU].rlim_cur == RLIM_INFINITY &&
1018             list_empty(&timers[CPUCLOCK_VIRT]) &&
1019             cputime_eq(sig->it_virt_expires, cputime_zero) &&
1020             list_empty(&timers[CPUCLOCK_SCHED]))
1021                 return;
1022
1023         /*
1024          * Collect the current process totals.
1025          */
1026         utime = sig->utime;
1027         stime = sig->stime;
1028         sched_time = sig->sched_time;
1029         t = tsk;
1030         do {
1031                 utime = cputime_add(utime, t->utime);
1032                 stime = cputime_add(stime, t->stime);
1033                 sched_time += t->sched_time;
1034                 t = next_thread(t);
1035         } while (t != tsk);
1036         ptime = cputime_add(utime, stime);
1037
1038         maxfire = 20;
1039         prof_expires = cputime_zero;
1040         while (!list_empty(timers)) {
1041                 struct cpu_timer_list *t = list_entry(timers->next,
1042                                                       struct cpu_timer_list,
1043                                                       entry);
1044                 if (!--maxfire || cputime_lt(ptime, t->expires.cpu)) {
1045                         prof_expires = t->expires.cpu;
1046                         break;
1047                 }
1048                 t->firing = 1;
1049                 list_move_tail(&t->entry, firing);
1050         }
1051
1052         ++timers;
1053         maxfire = 20;
1054         virt_expires = cputime_zero;
1055         while (!list_empty(timers)) {
1056                 struct cpu_timer_list *t = list_entry(timers->next,
1057                                                       struct cpu_timer_list,
1058                                                       entry);
1059                 if (!--maxfire || cputime_lt(utime, t->expires.cpu)) {
1060                         virt_expires = t->expires.cpu;
1061                         break;
1062                 }
1063                 t->firing = 1;
1064                 list_move_tail(&t->entry, firing);
1065         }
1066
1067         ++timers;
1068         maxfire = 20;
1069         sched_expires = 0;
1070         while (!list_empty(timers)) {
1071                 struct cpu_timer_list *t = list_entry(timers->next,
1072                                                       struct cpu_timer_list,
1073                                                       entry);
1074                 if (!--maxfire || sched_time < t->expires.sched) {
1075                         sched_expires = t->expires.sched;
1076                         break;
1077                 }
1078                 t->firing = 1;
1079                 list_move_tail(&t->entry, firing);
1080         }
1081
1082         /*
1083          * Check for the special case process timers.
1084          */
1085         if (!cputime_eq(sig->it_prof_expires, cputime_zero)) {
1086                 if (cputime_ge(ptime, sig->it_prof_expires)) {
1087                         /* ITIMER_PROF fires and reloads.  */
1088                         sig->it_prof_expires = sig->it_prof_incr;
1089                         if (!cputime_eq(sig->it_prof_expires, cputime_zero)) {
1090                                 sig->it_prof_expires = cputime_add(
1091                                         sig->it_prof_expires, ptime);
1092                         }
1093                         __group_send_sig_info(SIGPROF, SEND_SIG_PRIV, tsk);
1094                 }
1095                 if (!cputime_eq(sig->it_prof_expires, cputime_zero) &&
1096                     (cputime_eq(prof_expires, cputime_zero) ||
1097                      cputime_lt(sig->it_prof_expires, prof_expires))) {
1098                         prof_expires = sig->it_prof_expires;
1099                 }
1100         }
1101         if (!cputime_eq(sig->it_virt_expires, cputime_zero)) {
1102                 if (cputime_ge(utime, sig->it_virt_expires)) {
1103                         /* ITIMER_VIRTUAL fires and reloads.  */
1104                         sig->it_virt_expires = sig->it_virt_incr;
1105                         if (!cputime_eq(sig->it_virt_expires, cputime_zero)) {
1106                                 sig->it_virt_expires = cputime_add(
1107                                         sig->it_virt_expires, utime);
1108                         }
1109                         __group_send_sig_info(SIGVTALRM, SEND_SIG_PRIV, tsk);
1110                 }
1111                 if (!cputime_eq(sig->it_virt_expires, cputime_zero) &&
1112                     (cputime_eq(virt_expires, cputime_zero) ||
1113                      cputime_lt(sig->it_virt_expires, virt_expires))) {
1114                         virt_expires = sig->it_virt_expires;
1115                 }
1116         }
1117         if (sig->rlim[RLIMIT_CPU].rlim_cur != RLIM_INFINITY) {
1118                 unsigned long psecs = cputime_to_secs(ptime);
1119                 cputime_t x;
1120                 if (psecs >= sig->rlim[RLIMIT_CPU].rlim_max) {
1121                         /*
1122                          * At the hard limit, we just die.
1123                          * No need to calculate anything else now.
1124                          */
1125                         __group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
1126                         return;
1127                 }
1128                 if (psecs >= sig->rlim[RLIMIT_CPU].rlim_cur) {
1129                         /*
1130                          * At the soft limit, send a SIGXCPU every second.
1131                          */
1132                         __group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
1133                         if (sig->rlim[RLIMIT_CPU].rlim_cur
1134                             < sig->rlim[RLIMIT_CPU].rlim_max) {
1135                                 sig->rlim[RLIMIT_CPU].rlim_cur++;
1136                         }
1137                 }
1138                 x = secs_to_cputime(sig->rlim[RLIMIT_CPU].rlim_cur);
1139                 if (cputime_eq(prof_expires, cputime_zero) ||
1140                     cputime_lt(x, prof_expires)) {
1141                         prof_expires = x;
1142                 }
1143         }
1144
1145         if (!cputime_eq(prof_expires, cputime_zero) ||
1146             !cputime_eq(virt_expires, cputime_zero) ||
1147             sched_expires != 0) {
1148                 /*
1149                  * Rebalance the threads' expiry times for the remaining
1150                  * process CPU timers.
1151                  */
1152
1153                 cputime_t prof_left, virt_left, ticks;
1154                 unsigned long long sched_left, sched;
1155                 const unsigned int nthreads = atomic_read(&sig->live);
1156
1157                 if (!nthreads)
1158                         return;
1159
1160                 prof_left = cputime_sub(prof_expires, utime);
1161                 prof_left = cputime_sub(prof_left, stime);
1162                 prof_left = cputime_div(prof_left, nthreads);
1163                 virt_left = cputime_sub(virt_expires, utime);
1164                 virt_left = cputime_div(virt_left, nthreads);
1165                 if (sched_expires) {
1166                         sched_left = sched_expires - sched_time;
1167                         do_div(sched_left, nthreads);
1168                 } else {
1169                         sched_left = 0;
1170                 }
1171                 t = tsk;
1172                 do {
1173                         if (unlikely(t->flags & PF_EXITING))
1174                                 continue;
1175
1176                         ticks = cputime_add(cputime_add(t->utime, t->stime),
1177                                             prof_left);
1178                         if (!cputime_eq(prof_expires, cputime_zero) &&
1179                             (cputime_eq(t->it_prof_expires, cputime_zero) ||
1180                              cputime_gt(t->it_prof_expires, ticks))) {
1181                                 t->it_prof_expires = ticks;
1182                         }
1183
1184                         ticks = cputime_add(t->utime, virt_left);
1185                         if (!cputime_eq(virt_expires, cputime_zero) &&
1186                             (cputime_eq(t->it_virt_expires, cputime_zero) ||
1187                              cputime_gt(t->it_virt_expires, ticks))) {
1188                                 t->it_virt_expires = ticks;
1189                         }
1190
1191                         sched = t->sched_time + sched_left;
1192                         if (sched_expires && (t->it_sched_expires == 0 ||
1193                                               t->it_sched_expires > sched)) {
1194                                 t->it_sched_expires = sched;
1195                         }
1196                 } while ((t = next_thread(t)) != tsk);
1197         }
1198 }
1199
1200 /*
1201  * This is called from the signal code (via do_schedule_next_timer)
1202  * when the last timer signal was delivered and we have to reload the timer.
1203  */
1204 void posix_cpu_timer_schedule(struct k_itimer *timer)
1205 {
1206         struct task_struct *p = timer->it.cpu.task;
1207         union cpu_time_count now;
1208
1209         if (unlikely(p == NULL))
1210                 /*
1211                  * The task was cleaned up already, no future firings.
1212                  */
1213                 goto out;
1214
1215         /*
1216          * Fetch the current sample and update the timer's expiry time.
1217          */
1218         if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
1219                 cpu_clock_sample(timer->it_clock, p, &now);
1220                 bump_cpu_timer(timer, now);
1221                 if (unlikely(p->exit_state)) {
1222                         clear_dead_task(timer, now);
1223                         goto out;
1224                 }
1225                 read_lock(&tasklist_lock); /* arm_timer needs it.  */
1226         } else {
1227                 read_lock(&tasklist_lock);
1228                 if (unlikely(p->signal == NULL)) {
1229                         /*
1230                          * The process has been reaped.
1231                          * We can't even collect a sample any more.
1232                          */
1233                         put_task_struct(p);
1234                         timer->it.cpu.task = p = NULL;
1235                         timer->it.cpu.expires.sched = 0;
1236                         goto out_unlock;
1237                 } else if (unlikely(p->exit_state) && thread_group_empty(p)) {
1238                         /*
1239                          * We've noticed that the thread is dead, but
1240                          * not yet reaped.  Take this opportunity to
1241                          * drop our task ref.
1242                          */
1243                         clear_dead_task(timer, now);
1244                         goto out_unlock;
1245                 }
1246                 cpu_clock_sample_group(timer->it_clock, p, &now);
1247                 bump_cpu_timer(timer, now);
1248                 /* Leave the tasklist_lock locked for the call below.  */
1249         }
1250
1251         /*
1252          * Now re-arm for the new expiry time.
1253          */
1254         arm_timer(timer, now);
1255
1256 out_unlock:
1257         read_unlock(&tasklist_lock);
1258
1259 out:
1260         timer->it_overrun_last = timer->it_overrun;
1261         timer->it_overrun = -1;
1262         ++timer->it_requeue_pending;
1263 }
1264
1265 /*
1266  * This is called from the timer interrupt handler.  The irq handler has
1267  * already updated our counts.  We need to check if any timers fire now.
1268  * Interrupts are disabled.
1269  */
1270 void run_posix_cpu_timers(struct task_struct *tsk)
1271 {
1272         LIST_HEAD(firing);
1273         struct k_itimer *timer, *next;
1274
1275         BUG_ON(!irqs_disabled());
1276
1277 #define UNEXPIRED(clock) \
1278                 (cputime_eq(tsk->it_##clock##_expires, cputime_zero) || \
1279                  cputime_lt(clock##_ticks(tsk), tsk->it_##clock##_expires))
1280
1281         if (UNEXPIRED(prof) && UNEXPIRED(virt) &&
1282             (tsk->it_sched_expires == 0 ||
1283              tsk->sched_time < tsk->it_sched_expires))
1284                 return;
1285
1286 #undef  UNEXPIRED
1287
1288         /*
1289          * Double-check with locks held.
1290          */
1291         read_lock(&tasklist_lock);
1292         if (likely(tsk->signal != NULL)) {
1293                 spin_lock(&tsk->sighand->siglock);
1294
1295                 /*
1296                  * Here we take off tsk->cpu_timers[N] and tsk->signal->cpu_timers[N]
1297                  * all the timers that are firing, and put them on the firing list.
1298                  */
1299                 check_thread_timers(tsk, &firing);
1300                 check_process_timers(tsk, &firing);
1301
1302                 /*
1303                  * We must release these locks before taking any timer's lock.
1304                  * There is a potential race with timer deletion here, as the
1305                  * siglock now protects our private firing list.  We have set
1306                  * the firing flag in each timer, so that a deletion attempt
1307                  * that gets the timer lock before we do will give it up and
1308                  * spin until we've taken care of that timer below.
1309                  */
1310                 spin_unlock(&tsk->sighand->siglock);
1311         }
1312         read_unlock(&tasklist_lock);
1313
1314         /*
1315          * Now that all the timers on our list have the firing flag,
1316          * noone will touch their list entries but us.  We'll take
1317          * each timer's lock before clearing its firing flag, so no
1318          * timer call will interfere.
1319          */
1320         list_for_each_entry_safe(timer, next, &firing, it.cpu.entry) {
1321                 int firing;
1322                 spin_lock(&timer->it_lock);
1323                 list_del_init(&timer->it.cpu.entry);
1324                 firing = timer->it.cpu.firing;
1325                 timer->it.cpu.firing = 0;
1326                 /*
1327                  * The firing flag is -1 if we collided with a reset
1328                  * of the timer, which already reported this
1329                  * almost-firing as an overrun.  So don't generate an event.
1330                  */
1331                 if (likely(firing >= 0)) {
1332                         cpu_timer_fire(timer);
1333                 }
1334                 spin_unlock(&timer->it_lock);
1335         }
1336 }
1337
1338 /*
1339  * Set one of the process-wide special case CPU timers.
1340  * The tasklist_lock and tsk->sighand->siglock must be held by the caller.
1341  * The oldval argument is null for the RLIMIT_CPU timer, where *newval is
1342  * absolute; non-null for ITIMER_*, where *newval is relative and we update
1343  * it to be absolute, *oldval is absolute and we update it to be relative.
1344  */
1345 void set_process_cpu_timer(struct task_struct *tsk, unsigned int clock_idx,
1346                            cputime_t *newval, cputime_t *oldval)
1347 {
1348         union cpu_time_count now;
1349         struct list_head *head;
1350
1351         BUG_ON(clock_idx == CPUCLOCK_SCHED);
1352         cpu_clock_sample_group_locked(clock_idx, tsk, &now);
1353
1354         if (oldval) {
1355                 if (!cputime_eq(*oldval, cputime_zero)) {
1356                         if (cputime_le(*oldval, now.cpu)) {
1357                                 /* Just about to fire. */
1358                                 *oldval = jiffies_to_cputime(1);
1359                         } else {
1360                                 *oldval = cputime_sub(*oldval, now.cpu);
1361                         }
1362                 }
1363
1364                 if (cputime_eq(*newval, cputime_zero))
1365                         return;
1366                 *newval = cputime_add(*newval, now.cpu);
1367
1368                 /*
1369                  * If the RLIMIT_CPU timer will expire before the
1370                  * ITIMER_PROF timer, we have nothing else to do.
1371                  */
1372                 if (tsk->signal->rlim[RLIMIT_CPU].rlim_cur
1373                     < cputime_to_secs(*newval))
1374                         return;
1375         }
1376
1377         /*
1378          * Check whether there are any process timers already set to fire
1379          * before this one.  If so, we don't have anything more to do.
1380          */
1381         head = &tsk->signal->cpu_timers[clock_idx];
1382         if (list_empty(head) ||
1383             cputime_ge(list_entry(head->next,
1384                                   struct cpu_timer_list, entry)->expires.cpu,
1385                        *newval)) {
1386                 /*
1387                  * Rejigger each thread's expiry time so that one will
1388                  * notice before we hit the process-cumulative expiry time.
1389                  */
1390                 union cpu_time_count expires = { .sched = 0 };
1391                 expires.cpu = *newval;
1392                 process_timer_rebalance(tsk, clock_idx, expires, now);
1393         }
1394 }
1395
1396 static int do_cpu_nanosleep(const clockid_t which_clock, int flags,
1397                             struct timespec *rqtp, struct itimerspec *it)
1398 {
1399         struct k_itimer timer;
1400         int error;
1401
1402         /*
1403          * Set up a temporary timer and then wait for it to go off.
1404          */
1405         memset(&timer, 0, sizeof timer);
1406         spin_lock_init(&timer.it_lock);
1407         timer.it_clock = which_clock;
1408         timer.it_overrun = -1;
1409         error = posix_cpu_timer_create(&timer);
1410         timer.it_process = current;
1411         if (!error) {
1412                 static struct itimerspec zero_it;
1413
1414                 memset(it, 0, sizeof *it);
1415                 it->it_value = *rqtp;
1416
1417                 spin_lock_irq(&timer.it_lock);
1418                 error = posix_cpu_timer_set(&timer, flags, it, NULL);
1419                 if (error) {
1420                         spin_unlock_irq(&timer.it_lock);
1421                         return error;
1422                 }
1423
1424                 while (!signal_pending(current)) {
1425                         if (timer.it.cpu.expires.sched == 0) {
1426                                 /*
1427                                  * Our timer fired and was reset.
1428                                  */
1429                                 spin_unlock_irq(&timer.it_lock);
1430                                 return 0;
1431                         }
1432
1433                         /*
1434                          * Block until cpu_timer_fire (or a signal) wakes us.
1435                          */
1436                         __set_current_state(TASK_INTERRUPTIBLE);
1437                         spin_unlock_irq(&timer.it_lock);
1438                         schedule();
1439                         spin_lock_irq(&timer.it_lock);
1440                 }
1441
1442                 /*
1443                  * We were interrupted by a signal.
1444                  */
1445                 sample_to_timespec(which_clock, timer.it.cpu.expires, rqtp);
1446                 posix_cpu_timer_set(&timer, 0, &zero_it, it);
1447                 spin_unlock_irq(&timer.it_lock);
1448
1449                 if ((it->it_value.tv_sec | it->it_value.tv_nsec) == 0) {
1450                         /*
1451                          * It actually did fire already.
1452                          */
1453                         return 0;
1454                 }
1455
1456                 error = -ERESTART_RESTARTBLOCK;
1457         }
1458
1459         return error;
1460 }
1461
1462 int posix_cpu_nsleep(const clockid_t which_clock, int flags,
1463                      struct timespec *rqtp, struct timespec __user *rmtp)
1464 {
1465         struct restart_block *restart_block =
1466             &current_thread_info()->restart_block;
1467         struct itimerspec it;
1468         int error;
1469
1470         /*
1471          * Diagnose required errors first.
1472          */
1473         if (CPUCLOCK_PERTHREAD(which_clock) &&
1474             (CPUCLOCK_PID(which_clock) == 0 ||
1475              CPUCLOCK_PID(which_clock) == current->pid))
1476                 return -EINVAL;
1477
1478         error = do_cpu_nanosleep(which_clock, flags, rqtp, &it);
1479
1480         if (error == -ERESTART_RESTARTBLOCK) {
1481
1482                 if (flags & TIMER_ABSTIME)
1483                         return -ERESTARTNOHAND;
1484                 /*
1485                  * Report back to the user the time still remaining.
1486                  */
1487                 if (rmtp != NULL && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1488                         return -EFAULT;
1489
1490                 restart_block->fn = posix_cpu_nsleep_restart;
1491                 restart_block->arg0 = which_clock;
1492                 restart_block->arg1 = (unsigned long) rmtp;
1493                 restart_block->arg2 = rqtp->tv_sec;
1494                 restart_block->arg3 = rqtp->tv_nsec;
1495         }
1496         return error;
1497 }
1498
1499 long posix_cpu_nsleep_restart(struct restart_block *restart_block)
1500 {
1501         clockid_t which_clock = restart_block->arg0;
1502         struct timespec __user *rmtp;
1503         struct timespec t;
1504         struct itimerspec it;
1505         int error;
1506
1507         rmtp = (struct timespec __user *) restart_block->arg1;
1508         t.tv_sec = restart_block->arg2;
1509         t.tv_nsec = restart_block->arg3;
1510
1511         restart_block->fn = do_no_restart_syscall;
1512         error = do_cpu_nanosleep(which_clock, TIMER_ABSTIME, &t, &it);
1513
1514         if (error == -ERESTART_RESTARTBLOCK) {
1515                 /*
1516                  * Report back to the user the time still remaining.
1517                  */
1518                 if (rmtp != NULL && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
1519                         return -EFAULT;
1520
1521                 restart_block->fn = posix_cpu_nsleep_restart;
1522                 restart_block->arg0 = which_clock;
1523                 restart_block->arg1 = (unsigned long) rmtp;
1524                 restart_block->arg2 = t.tv_sec;
1525                 restart_block->arg3 = t.tv_nsec;
1526         }
1527         return error;
1528
1529 }
1530
1531
1532 #define PROCESS_CLOCK   MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED)
1533 #define THREAD_CLOCK    MAKE_THREAD_CPUCLOCK(0, CPUCLOCK_SCHED)
1534
1535 static int process_cpu_clock_getres(const clockid_t which_clock,
1536                                     struct timespec *tp)
1537 {
1538         return posix_cpu_clock_getres(PROCESS_CLOCK, tp);
1539 }
1540 static int process_cpu_clock_get(const clockid_t which_clock,
1541                                  struct timespec *tp)
1542 {
1543         return posix_cpu_clock_get(PROCESS_CLOCK, tp);
1544 }
1545 static int process_cpu_timer_create(struct k_itimer *timer)
1546 {
1547         timer->it_clock = PROCESS_CLOCK;
1548         return posix_cpu_timer_create(timer);
1549 }
1550 static int process_cpu_nsleep(const clockid_t which_clock, int flags,
1551                               struct timespec *rqtp,
1552                               struct timespec __user *rmtp)
1553 {
1554         return posix_cpu_nsleep(PROCESS_CLOCK, flags, rqtp, rmtp);
1555 }
1556 static long process_cpu_nsleep_restart(struct restart_block *restart_block)
1557 {
1558         return -EINVAL;
1559 }
1560 static int thread_cpu_clock_getres(const clockid_t which_clock,
1561                                    struct timespec *tp)
1562 {
1563         return posix_cpu_clock_getres(THREAD_CLOCK, tp);
1564 }
1565 static int thread_cpu_clock_get(const clockid_t which_clock,
1566                                 struct timespec *tp)
1567 {
1568         return posix_cpu_clock_get(THREAD_CLOCK, tp);
1569 }
1570 static int thread_cpu_timer_create(struct k_itimer *timer)
1571 {
1572         timer->it_clock = THREAD_CLOCK;
1573         return posix_cpu_timer_create(timer);
1574 }
1575 static int thread_cpu_nsleep(const clockid_t which_clock, int flags,
1576                               struct timespec *rqtp, struct timespec __user *rmtp)
1577 {
1578         return -EINVAL;
1579 }
1580 static long thread_cpu_nsleep_restart(struct restart_block *restart_block)
1581 {
1582         return -EINVAL;
1583 }
1584
1585 static __init int init_posix_cpu_timers(void)
1586 {
1587         struct k_clock process = {
1588                 .clock_getres = process_cpu_clock_getres,
1589                 .clock_get = process_cpu_clock_get,
1590                 .clock_set = do_posix_clock_nosettime,
1591                 .timer_create = process_cpu_timer_create,
1592                 .nsleep = process_cpu_nsleep,
1593                 .nsleep_restart = process_cpu_nsleep_restart,
1594         };
1595         struct k_clock thread = {
1596                 .clock_getres = thread_cpu_clock_getres,
1597                 .clock_get = thread_cpu_clock_get,
1598                 .clock_set = do_posix_clock_nosettime,
1599                 .timer_create = thread_cpu_timer_create,
1600                 .nsleep = thread_cpu_nsleep,
1601                 .nsleep_restart = thread_cpu_nsleep_restart,
1602         };
1603
1604         register_posix_clock(CLOCK_PROCESS_CPUTIME_ID, &process);
1605         register_posix_clock(CLOCK_THREAD_CPUTIME_ID, &thread);
1606
1607         return 0;
1608 }
1609 __initcall(init_posix_cpu_timers);