fs/timerfd.c

   1 /*
   2  *  fs/timerfd.c
   3  *
   4  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
   5  *
   6  *
   7  *  Thanks to Thomas Gleixner for code reviews and useful comments.
   8  *
   9  */
  10
  11 #include <linux/alarmtimer.h>
  12 #include <linux/file.h>
  13 #include <linux/poll.h>
  14 #include <linux/init.h>
  15 #include <linux/fs.h>
  16 #include <linux/sched.h>
  17 #include <linux/kernel.h>
  18 #include <linux/slab.h>
  19 #include <linux/list.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/time.h>
  22 #include <linux/hrtimer.h>
  23 #include <linux/anon_inodes.h>
  24 #include <linux/timerfd.h>
  25 #include <linux/syscalls.h>
  26 #include <linux/compat.h>
  27 #include <linux/rcupdate.h>
  28
  29 struct timerfd_ctx {
  30         union {
  31                 struct hrtimer tmr;
  32                 struct alarm alarm;
  33         } t;
  34         ktime_t tintv;
  35         ktime_t moffs;
  36         wait_queue_head_t wqh;
  37         u64 ticks;
  38         int clockid;
  39         short unsigned expired;
  40         short unsigned settime_flags;   /* to show in fdinfo */
  41         struct rcu_head rcu;
  42         struct list_head clist;
  43         spinlock_t cancel_lock;
  44         bool might_cancel;
  45 };
  46
  47 static LIST_HEAD(cancel_list);
  48 static DEFINE_SPINLOCK(cancel_lock);
  49
  50 static inline bool isalarm(struct timerfd_ctx *ctx)
  51 {
  52         return ctx->clockid == CLOCK_REALTIME_ALARM ||
  53                 ctx->clockid == CLOCK_BOOTTIME_ALARM;
  54 }
  55
  56 /*
  57  * This gets called when the timer event triggers. We set the "expired"
  58  * flag, but we do not re-arm the timer (in case it's necessary,
  59  * tintv.tv64 != 0) until the timer is accessed.
  60  */
  61 static void timerfd_triggered(struct timerfd_ctx *ctx)
  62 {
  63         unsigned long flags;
  64
  65         spin_lock_irqsave(&ctx->wqh.lock, flags);
  66         ctx->expired = 1;
  67         ctx->ticks++;
  68         wake_up_locked(&ctx->wqh);
  69         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
  70 }
  71
  72 static enum hrtimer_restart timerfd_tmrproc(struct hrtimer *htmr)
  73 {
  74         struct timerfd_ctx *ctx = container_of(htmr, struct timerfd_ctx,
  75                                                t.tmr);
  76         timerfd_triggered(ctx);
  77         return HRTIMER_NORESTART;
  78 }
  79
  80 static enum alarmtimer_restart timerfd_alarmproc(struct alarm *alarm,
  81         ktime_t now)
  82 {
  83         struct timerfd_ctx *ctx = container_of(alarm, struct timerfd_ctx,
  84                                                t.alarm);
  85         timerfd_triggered(ctx);
  86         return ALARMTIMER_NORESTART;
  87 }
  88
  89 /*
  90  * Called when the clock was set to cancel the timers in the cancel
  91  * list. This will wake up processes waiting on these timers. The
  92  * wake-up requires ctx->ticks to be non zero, therefore we increment
  93  * it before calling wake_up_locked().
  94  */
  95 void timerfd_clock_was_set(void)
  96 {
  97         ktime_t moffs = ktime_mono_to_real((ktime_t){ .tv64 = 0 });
  98         struct timerfd_ctx *ctx;
  99         unsigned long flags;
 100
 101         rcu_read_lock();
 102         list_for_each_entry_rcu(ctx, &cancel_list, clist) {
 103                 if (!ctx->might_cancel)
 104                         continue;
 105                 spin_lock_irqsave(&ctx->wqh.lock, flags);
 106                 if (ctx->moffs.tv64 != moffs.tv64) {
 107                         ctx->moffs.tv64 = KTIME_MAX;
 108                         ctx->ticks++;
 109                         wake_up_locked(&ctx->wqh);
 110                 }
 111                 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
 112         }
 113         rcu_read_unlock();
 114 }
 115
 116 static void __timerfd_remove_cancel(struct timerfd_ctx *ctx)
 117 {
 118         if (ctx->might_cancel) {
 119                 ctx->might_cancel = false;
 120                 spin_lock(&cancel_lock);
 121                 list_del_rcu(&ctx->clist);
 122                 spin_unlock(&cancel_lock);
 123         }
 124 }
 125
 126 static void timerfd_remove_cancel(struct timerfd_ctx *ctx)
 127 {
 128         spin_lock(&ctx->cancel_lock);
 129         __timerfd_remove_cancel(ctx);
 130         spin_unlock(&ctx->cancel_lock);
 131 }
 132
 133 static bool timerfd_canceled(struct timerfd_ctx *ctx)
 134 {
 135         if (!ctx->might_cancel || ctx->moffs.tv64 != KTIME_MAX)
 136                 return false;
 137         ctx->moffs = ktime_mono_to_real((ktime_t){ .tv64 = 0 });
 138         return true;
 139 }
 140
 141 static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags)
 142 {
 143         spin_lock(&ctx->cancel_lock);
 144         if ((ctx->clockid == CLOCK_REALTIME ||
 145              ctx->clockid == CLOCK_REALTIME_ALARM) &&
 146             (flags & TFD_TIMER_ABSTIME) && (flags & TFD_TIMER_CANCEL_ON_SET)) {
 147                 if (!ctx->might_cancel) {
 148                         ctx->might_cancel = true;
 149                         spin_lock(&cancel_lock);
 150                         list_add_rcu(&ctx->clist, &cancel_list);
 151                         spin_unlock(&cancel_lock);
 152                 }
 153         } else {
 154                 __timerfd_remove_cancel(ctx);
 155         }
 156         spin_unlock(&ctx->cancel_lock);
 157 }
 158
 159 static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx)
 160 {
 161         ktime_t remaining;
 162
 163         if (isalarm(ctx))
 164                 remaining = alarm_expires_remaining(&ctx->t.alarm);
 165         else
 166                 remaining = hrtimer_expires_remaining(&ctx->t.tmr);
 167
 168         return remaining.tv64 < 0 ? ktime_set(0, 0): remaining;
 169 }
 170
 171 static int timerfd_setup(struct timerfd_ctx *ctx, int flags,
 172                          const struct itimerspec *ktmr)
 173 {
 174         enum hrtimer_mode htmode;
 175         ktime_t texp;
 176         int clockid = ctx->clockid;
 177
 178         htmode = (flags & TFD_TIMER_ABSTIME) ?
 179                 HRTIMER_MODE_ABS: HRTIMER_MODE_REL;
 180
 181         texp = timespec_to_ktime(ktmr->it_value);
 182         ctx->expired = 0;
 183         ctx->ticks = 0;
 184         ctx->tintv = timespec_to_ktime(ktmr->it_interval);
 185
 186         if (isalarm(ctx)) {
 187                 alarm_init(&ctx->t.alarm,
 188                            ctx->clockid == CLOCK_REALTIME_ALARM ?
 189                            ALARM_REALTIME : ALARM_BOOTTIME,
 190                            timerfd_alarmproc);
 191         } else {
 192                 hrtimer_init(&ctx->t.tmr, clockid, htmode);
 193                 hrtimer_set_expires(&ctx->t.tmr, texp);
 194                 ctx->t.tmr.function = timerfd_tmrproc;
 195         }
 196
 197         if (texp.tv64 != 0) {
 198                 if (isalarm(ctx)) {
 199                         if (flags & TFD_TIMER_ABSTIME)
 200                                 alarm_start(&ctx->t.alarm, texp);
 201                         else
 202                                 alarm_start_relative(&ctx->t.alarm, texp);
 203                 } else {
 204                         hrtimer_start(&ctx->t.tmr, texp, htmode);
 205                 }
 206
 207                 if (timerfd_canceled(ctx))
 208                         return -ECANCELED;
 209         }
 210
 211         ctx->settime_flags = flags & TFD_SETTIME_FLAGS;
 212         return 0;
 213 }
 214
 215 static int timerfd_release(struct inode *inode, struct file *file)
 216 {
 217         struct timerfd_ctx *ctx = file->private_data;
 218
 219         timerfd_remove_cancel(ctx);
 220
 221         if (isalarm(ctx))
 222                 alarm_cancel(&ctx->t.alarm);
 223         else
 224                 hrtimer_cancel(&ctx->t.tmr);
 225         kfree_rcu(ctx, rcu);
 226         return 0;
 227 }
 228
 229 static unsigned int timerfd_poll(struct file *file, poll_table *wait)
 230 {
 231         struct timerfd_ctx *ctx = file->private_data;
 232         unsigned int events = 0;
 233         unsigned long flags;
 234
 235         poll_wait(file, &ctx->wqh, wait);
 236
 237         spin_lock_irqsave(&ctx->wqh.lock, flags);
 238         if (ctx->ticks)
 239                 events |= POLLIN;
 240         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
 241
 242         return events;
 243 }
 244
 245 static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
 246                             loff_t *ppos)
 247 {
 248         struct timerfd_ctx *ctx = file->private_data;
 249         ssize_t res;
 250         u64 ticks = 0;
 251
 252         if (count < sizeof(ticks))
 253                 return -EINVAL;
 254         spin_lock_irq(&ctx->wqh.lock);
 255         if (file->f_flags & O_NONBLOCK)
 256                 res = -EAGAIN;
 257         else
 258                 res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks);
 259
 260         /*
 261          * If clock has changed, we do not care about the
 262          * ticks and we do not rearm the timer. Userspace must
 263          * reevaluate anyway.
 264          */
 265         if (timerfd_canceled(ctx)) {
 266                 ctx->ticks = 0;
 267                 ctx->expired = 0;
 268                 res = -ECANCELED;
 269         }
 270
 271         if (ctx->ticks) {
 272                 ticks = ctx->ticks;
 273
 274                 if (ctx->expired && ctx->tintv.tv64) {
 275                         /*
 276                          * If tintv.tv64 != 0, this is a periodic timer that
 277                          * needs to be re-armed. We avoid doing it in the timer
 278                          * callback to avoid DoS attacks specifying a very
 279                          * short timer period.
 280                          */
 281                         if (isalarm(ctx)) {
 282                                 ticks += alarm_forward_now(
 283                                         &ctx->t.alarm, ctx->tintv) - 1;
 284                                 alarm_restart(&ctx->t.alarm);
 285                         } else {
 286                                 ticks += hrtimer_forward_now(&ctx->t.tmr,
 287                                                              ctx->tintv) - 1;
 288                                 hrtimer_restart(&ctx->t.tmr);
 289                         }
 290                 }
 291                 ctx->expired = 0;
 292                 ctx->ticks = 0;
 293         }
 294         spin_unlock_irq(&ctx->wqh.lock);
 295         if (ticks)
 296                 res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks);
 297         return res;
 298 }
 299
 300 #ifdef CONFIG_PROC_FS
 301 static int timerfd_show(struct seq_file *m, struct file *file)
 302 {
 303         struct timerfd_ctx *ctx = file->private_data;
 304         struct itimerspec t;
 305
 306         spin_lock_irq(&ctx->wqh.lock);
 307         t.it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
 308         t.it_interval = ktime_to_timespec(ctx->tintv);
 309         spin_unlock_irq(&ctx->wqh.lock);
 310
 311         return seq_printf(m,
 312                           "clockid: %d\n"
 313                           "ticks: %llu\n"
 314                           "settime flags: 0%o\n"
 315                           "it_value: (%llu, %llu)\n"
 316                           "it_interval: (%llu, %llu)\n",
 317                           ctx->clockid, (unsigned long long)ctx->ticks,
 318                           ctx->settime_flags,
 319                           (unsigned long long)t.it_value.tv_sec,
 320                           (unsigned long long)t.it_value.tv_nsec,
 321                           (unsigned long long)t.it_interval.tv_sec,
 322                           (unsigned long long)t.it_interval.tv_nsec);
 323 }
 324 #else
 325 #define timerfd_show NULL
 326 #endif
 327
 328 #ifdef CONFIG_CHECKPOINT_RESTORE
 329 static long timerfd_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 330 {
 331         struct timerfd_ctx *ctx = file->private_data;
 332         int ret = 0;
 333
 334         switch (cmd) {
 335         case TFD_IOC_SET_TICKS: {
 336                 u64 ticks;
 337
 338                 if (copy_from_user(&ticks, (u64 __user *)arg, sizeof(ticks)))
 339                         return -EFAULT;
 340                 if (!ticks)
 341                         return -EINVAL;
 342
 343                 spin_lock_irq(&ctx->wqh.lock);
 344                 if (!timerfd_canceled(ctx)) {
 345                         ctx->ticks = ticks;
 346                         wake_up_locked(&ctx->wqh);
 347                 } else
 348                         ret = -ECANCELED;
 349                 spin_unlock_irq(&ctx->wqh.lock);
 350                 break;
 351         }
 352         default:
 353                 ret = -ENOTTY;
 354                 break;
 355         }
 356
 357         return ret;
 358 }
 359 #else
 360 #define timerfd_ioctl NULL
 361 #endif
 362
 363 static const struct file_operations timerfd_fops = {
 364         .release        = timerfd_release,
 365         .poll           = timerfd_poll,
 366         .read           = timerfd_read,
 367         .llseek         = noop_llseek,
 368         .show_fdinfo    = timerfd_show,
 369         .unlocked_ioctl = timerfd_ioctl,
 370 };
 371
 372 static int timerfd_fget(int fd, struct fd *p)
 373 {
 374         struct fd f = fdget(fd);
 375         if (!f.file)
 376                 return -EBADF;
 377         if (f.file->f_op != &timerfd_fops) {
 378                 fdput(f);
 379                 return -EINVAL;
 380         }
 381         *p = f;
 382         return 0;
 383 }
 384
 385 SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
 386 {
 387         int ufd;
 388         struct timerfd_ctx *ctx;
 389
 390         /* Check the TFD_* constants for consistency.  */
 391         BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC);
 392         BUILD_BUG_ON(TFD_NONBLOCK != O_NONBLOCK);
 393
 394         if ((flags & ~TFD_CREATE_FLAGS) ||
 395             (clockid != CLOCK_MONOTONIC &&
 396              clockid != CLOCK_REALTIME &&
 397              clockid != CLOCK_REALTIME_ALARM &&
 398              clockid != CLOCK_BOOTTIME &&
 399              clockid != CLOCK_BOOTTIME_ALARM))
 400                 return -EINVAL;
 401
 402         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 403         if (!ctx)
 404                 return -ENOMEM;
 405
 406         init_waitqueue_head(&ctx->wqh);
 407         spin_lock_init(&ctx->cancel_lock);
 408         ctx->clockid = clockid;
 409
 410         if (isalarm(ctx))
 411                 alarm_init(&ctx->t.alarm,
 412                            ctx->clockid == CLOCK_REALTIME_ALARM ?
 413                            ALARM_REALTIME : ALARM_BOOTTIME,
 414                            timerfd_alarmproc);
 415         else
 416                 hrtimer_init(&ctx->t.tmr, clockid, HRTIMER_MODE_ABS);
 417
 418         ctx->moffs = ktime_mono_to_real((ktime_t){ .tv64 = 0 });
 419
 420         ufd = anon_inode_getfd("[timerfd]", &timerfd_fops, ctx,
 421                                O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
 422         if (ufd < 0)
 423                 kfree(ctx);
 424
 425         return ufd;
 426 }
 427
 428 static int do_timerfd_settime(int ufd, int flags,
 429                 const struct itimerspec *new,
 430                 struct itimerspec *old)
 431 {
 432         struct fd f;
 433         struct timerfd_ctx *ctx;
 434         int ret;
 435
 436         if ((flags & ~TFD_SETTIME_FLAGS) ||
 437             !timespec_valid(&new->it_value) ||
 438             !timespec_valid(&new->it_interval))
 439                 return -EINVAL;
 440
 441         ret = timerfd_fget(ufd, &f);
 442         if (ret)
 443                 return ret;
 444         ctx = f.file->private_data;
 445
 446         timerfd_setup_cancel(ctx, flags);
 447
 448         /*
 449          * We need to stop the existing timer before reprogramming
 450          * it to the new values.
 451          */
 452         for (;;) {
 453                 spin_lock_irq(&ctx->wqh.lock);
 454
 455                 if (isalarm(ctx)) {
 456                         if (alarm_try_to_cancel(&ctx->t.alarm) >= 0)
 457                                 break;
 458                 } else {
 459                         if (hrtimer_try_to_cancel(&ctx->t.tmr) >= 0)
 460                                 break;
 461                 }
 462                 spin_unlock_irq(&ctx->wqh.lock);
 463                 cpu_relax();
 464         }
 465
 466         /*
 467          * If the timer is expired and it's periodic, we need to advance it
 468          * because the caller may want to know the previous expiration time.
 469          * We do not update "ticks" and "expired" since the timer will be
 470          * re-programmed again in the following timerfd_setup() call.
 471          */
 472         if (ctx->expired && ctx->tintv.tv64) {
 473                 if (isalarm(ctx))
 474                         alarm_forward_now(&ctx->t.alarm, ctx->tintv);
 475                 else
 476                         hrtimer_forward_now(&ctx->t.tmr, ctx->tintv);
 477         }
 478
 479         old->it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
 480         old->it_interval = ktime_to_timespec(ctx->tintv);
 481
 482         /*
 483          * Re-program the timer to the new value ...
 484          */
 485         ret = timerfd_setup(ctx, flags, new);
 486
 487         spin_unlock_irq(&ctx->wqh.lock);
 488         fdput(f);
 489         return ret;
 490 }
 491
 492 static int do_timerfd_gettime(int ufd, struct itimerspec *t)
 493 {
 494         struct fd f;
 495         struct timerfd_ctx *ctx;
 496         int ret = timerfd_fget(ufd, &f);
 497         if (ret)
 498                 return ret;
 499         ctx = f.file->private_data;
 500
 501         spin_lock_irq(&ctx->wqh.lock);
 502         if (ctx->expired && ctx->tintv.tv64) {
 503                 ctx->expired = 0;
 504
 505                 if (isalarm(ctx)) {
 506                         ctx->ticks +=
 507                                 alarm_forward_now(
 508                                         &ctx->t.alarm, ctx->tintv) - 1;
 509                         alarm_restart(&ctx->t.alarm);
 510                 } else {
 511                         ctx->ticks +=
 512                                 hrtimer_forward_now(&ctx->t.tmr, ctx->tintv)
 513                                 - 1;
 514                         hrtimer_restart(&ctx->t.tmr);
 515                 }
 516         }
 517         t->it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
 518         t->it_interval = ktime_to_timespec(ctx->tintv);
 519         spin_unlock_irq(&ctx->wqh.lock);
 520         fdput(f);
 521         return 0;
 522 }
 523
 524 SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
 525                 const struct itimerspec __user *, utmr,
 526                 struct itimerspec __user *, otmr)
 527 {
 528         struct itimerspec new, old;
 529         int ret;
 530
 531         if (copy_from_user(&new, utmr, sizeof(new)))
 532                 return -EFAULT;
 533         ret = do_timerfd_settime(ufd, flags, &new, &old);
 534         if (ret)
 535                 return ret;
 536         if (otmr && copy_to_user(otmr, &old, sizeof(old)))
 537                 return -EFAULT;
 538
 539         return ret;
 540 }
 541
 542 SYSCALL_DEFINE2(timerfd_gettime, int, ufd, struct itimerspec __user *, otmr)
 543 {
 544         struct itimerspec kotmr;
 545         int ret = do_timerfd_gettime(ufd, &kotmr);
 546         if (ret)
 547                 return ret;
 548         return copy_to_user(otmr, &kotmr, sizeof(kotmr)) ? -EFAULT: 0;
 549 }
 550
 551 #ifdef CONFIG_COMPAT
 552 COMPAT_SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
 553                 const struct compat_itimerspec __user *, utmr,
 554                 struct compat_itimerspec __user *, otmr)
 555 {
 556         struct itimerspec new, old;
 557         int ret;
 558
 559         if (get_compat_itimerspec(&new, utmr))
 560                 return -EFAULT;
 561         ret = do_timerfd_settime(ufd, flags, &new, &old);
 562         if (ret)
 563                 return ret;
 564         if (otmr && put_compat_itimerspec(otmr, &old))
 565                 return -EFAULT;
 566         return ret;
 567 }
 568
 569 COMPAT_SYSCALL_DEFINE2(timerfd_gettime, int, ufd,
 570                 struct compat_itimerspec __user *, otmr)
 571 {
 572         struct itimerspec kotmr;
 573         int ret = do_timerfd_gettime(ufd, &kotmr);
 574         if (ret)
 575                 return ret;
 576         return put_compat_itimerspec(otmr, &kotmr) ? -EFAULT: 0;
 577 }
 578 #endif