fs/timerfd.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  fs/timerfd.c
   4  *
   5  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
   6  *
   7  *
   8  *  Thanks to Thomas Gleixner for code reviews and useful comments.
   9  *
  10  */
  11
  12 #include <linux/alarmtimer.h>
  13 #include <linux/file.h>
  14 #include <linux/poll.h>
  15 #include <linux/init.h>
  16 #include <linux/fs.h>
  17 #include <linux/sched.h>
  18 #include <linux/kernel.h>
  19 #include <linux/slab.h>
  20 #include <linux/list.h>
  21 #include <linux/spinlock.h>
  22 #include <linux/time.h>
  23 #include <linux/hrtimer.h>
  24 #include <linux/anon_inodes.h>
  25 #include <linux/timerfd.h>
  26 #include <linux/syscalls.h>
  27 #include <linux/compat.h>
  28 #include <linux/rcupdate.h>
  29 #include <linux/time_namespace.h>
  30
  31 struct timerfd_ctx {
  32         union {
  33                 struct hrtimer tmr;
  34                 struct alarm alarm;
  35         } t;
  36         ktime_t tintv;
  37         ktime_t moffs;
  38         wait_queue_head_t wqh;
  39         u64 ticks;
  40         int clockid;
  41         short unsigned expired;
  42         short unsigned settime_flags;   /* to show in fdinfo */
  43         struct rcu_head rcu;
  44         struct list_head clist;
  45         spinlock_t cancel_lock;
  46         bool might_cancel;
  47 };
  48
  49 static LIST_HEAD(cancel_list);
  50 static DEFINE_SPINLOCK(cancel_lock);
  51
  52 static inline bool isalarm(struct timerfd_ctx *ctx)
  53 {
  54         return ctx->clockid == CLOCK_REALTIME_ALARM ||
  55                 ctx->clockid == CLOCK_BOOTTIME_ALARM;
  56 }
  57
  58 /*
  59  * This gets called when the timer event triggers. We set the "expired"
  60  * flag, but we do not re-arm the timer (in case it's necessary,
  61  * tintv != 0) until the timer is accessed.
  62  */
  63 static void timerfd_triggered(struct timerfd_ctx *ctx)
  64 {
  65         unsigned long flags;
  66
  67         spin_lock_irqsave(&ctx->wqh.lock, flags);
  68         ctx->expired = 1;
  69         ctx->ticks++;
  70         wake_up_locked_poll(&ctx->wqh, EPOLLIN);
  71         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
  72 }
  73
  74 static enum hrtimer_restart timerfd_tmrproc(struct hrtimer *htmr)
  75 {
  76         struct timerfd_ctx *ctx = container_of(htmr, struct timerfd_ctx,
  77                                                t.tmr);
  78         timerfd_triggered(ctx);
  79         return HRTIMER_NORESTART;
  80 }
  81
  82 static enum alarmtimer_restart timerfd_alarmproc(struct alarm *alarm,
  83         ktime_t now)
  84 {
  85         struct timerfd_ctx *ctx = container_of(alarm, struct timerfd_ctx,
  86                                                t.alarm);
  87         timerfd_triggered(ctx);
  88         return ALARMTIMER_NORESTART;
  89 }
  90
  91 /*
  92  * Called when the clock was set to cancel the timers in the cancel
  93  * list. This will wake up processes waiting on these timers. The
  94  * wake-up requires ctx->ticks to be non zero, therefore we increment
  95  * it before calling wake_up_locked().
  96  */
  97 void timerfd_clock_was_set(void)
  98 {
  99         ktime_t moffs = ktime_mono_to_real(0);
 100         struct timerfd_ctx *ctx;
 101         unsigned long flags;
 102
 103         rcu_read_lock();
 104         list_for_each_entry_rcu(ctx, &cancel_list, clist) {
 105                 if (!ctx->might_cancel)
 106                         continue;
 107                 spin_lock_irqsave(&ctx->wqh.lock, flags);
 108                 if (ctx->moffs != moffs) {
 109                         ctx->moffs = KTIME_MAX;
 110                         ctx->ticks++;
 111                         wake_up_locked_poll(&ctx->wqh, EPOLLIN);
 112                 }
 113                 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
 114         }
 115         rcu_read_unlock();
 116 }
 117
 118 static void __timerfd_remove_cancel(struct timerfd_ctx *ctx)
 119 {
 120         if (ctx->might_cancel) {
 121                 ctx->might_cancel = false;
 122                 spin_lock(&cancel_lock);
 123                 list_del_rcu(&ctx->clist);
 124                 spin_unlock(&cancel_lock);
 125         }
 126 }
 127
 128 static void timerfd_remove_cancel(struct timerfd_ctx *ctx)
 129 {
 130         spin_lock(&ctx->cancel_lock);
 131         __timerfd_remove_cancel(ctx);
 132         spin_unlock(&ctx->cancel_lock);
 133 }
 134
 135 static bool timerfd_canceled(struct timerfd_ctx *ctx)
 136 {
 137         if (!ctx->might_cancel || ctx->moffs != KTIME_MAX)
 138                 return false;
 139         ctx->moffs = ktime_mono_to_real(0);
 140         return true;
 141 }
 142
 143 static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags)
 144 {
 145         spin_lock(&ctx->cancel_lock);
 146         if ((ctx->clockid == CLOCK_REALTIME ||
 147              ctx->clockid == CLOCK_REALTIME_ALARM) &&
 148             (flags & TFD_TIMER_ABSTIME) && (flags & TFD_TIMER_CANCEL_ON_SET)) {
 149                 if (!ctx->might_cancel) {
 150                         ctx->might_cancel = true;
 151                         spin_lock(&cancel_lock);
 152                         list_add_rcu(&ctx->clist, &cancel_list);
 153                         spin_unlock(&cancel_lock);
 154                 }
 155         } else {
 156                 __timerfd_remove_cancel(ctx);
 157         }
 158         spin_unlock(&ctx->cancel_lock);
 159 }
 160
 161 static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx)
 162 {
 163         ktime_t remaining;
 164
 165         if (isalarm(ctx))
 166                 remaining = alarm_expires_remaining(&ctx->t.alarm);
 167         else
 168                 remaining = hrtimer_expires_remaining_adjusted(&ctx->t.tmr);
 169
 170         return remaining < 0 ? 0: remaining;
 171 }
 172
 173 static int timerfd_setup(struct timerfd_ctx *ctx, int flags,
 174                          const struct itimerspec64 *ktmr)
 175 {
 176         enum hrtimer_mode htmode;
 177         ktime_t texp;
 178         int clockid = ctx->clockid;
 179
 180         htmode = (flags & TFD_TIMER_ABSTIME) ?
 181                 HRTIMER_MODE_ABS: HRTIMER_MODE_REL;
 182
 183         texp = timespec64_to_ktime(ktmr->it_value);
 184         ctx->expired = 0;
 185         ctx->ticks = 0;
 186         ctx->tintv = timespec64_to_ktime(ktmr->it_interval);
 187
 188         if (isalarm(ctx)) {
 189                 alarm_init(&ctx->t.alarm,
 190                            ctx->clockid == CLOCK_REALTIME_ALARM ?
 191                            ALARM_REALTIME : ALARM_BOOTTIME,
 192                            timerfd_alarmproc);
 193         } else {
 194                 hrtimer_init(&ctx->t.tmr, clockid, htmode);
 195                 hrtimer_set_expires(&ctx->t.tmr, texp);
 196                 ctx->t.tmr.function = timerfd_tmrproc;
 197         }
 198
 199         if (texp != 0) {
 200                 if (flags & TFD_TIMER_ABSTIME)
 201                         texp = timens_ktime_to_host(clockid, texp);
 202                 if (isalarm(ctx)) {
 203                         if (flags & TFD_TIMER_ABSTIME)
 204                                 alarm_start(&ctx->t.alarm, texp);
 205                         else
 206                                 alarm_start_relative(&ctx->t.alarm, texp);
 207                 } else {
 208                         hrtimer_start(&ctx->t.tmr, texp, htmode);
 209                 }
 210
 211                 if (timerfd_canceled(ctx))
 212                         return -ECANCELED;
 213         }
 214
 215         ctx->settime_flags = flags & TFD_SETTIME_FLAGS;
 216         return 0;
 217 }
 218
 219 static int timerfd_release(struct inode *inode, struct file *file)
 220 {
 221         struct timerfd_ctx *ctx = file->private_data;
 222
 223         timerfd_remove_cancel(ctx);
 224
 225         if (isalarm(ctx))
 226                 alarm_cancel(&ctx->t.alarm);
 227         else
 228                 hrtimer_cancel(&ctx->t.tmr);
 229         kfree_rcu(ctx, rcu);
 230         return 0;
 231 }
 232
 233 static __poll_t timerfd_poll(struct file *file, poll_table *wait)
 234 {
 235         struct timerfd_ctx *ctx = file->private_data;
 236         __poll_t events = 0;
 237         unsigned long flags;
 238
 239         poll_wait(file, &ctx->wqh, wait);
 240
 241         spin_lock_irqsave(&ctx->wqh.lock, flags);
 242         if (ctx->ticks)
 243                 events |= EPOLLIN;
 244         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
 245
 246         return events;
 247 }
 248
 249 static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
 250                             loff_t *ppos)
 251 {
 252         struct timerfd_ctx *ctx = file->private_data;
 253         ssize_t res;
 254         u64 ticks = 0;
 255
 256         if (count < sizeof(ticks))
 257                 return -EINVAL;
 258         spin_lock_irq(&ctx->wqh.lock);
 259         if (file->f_flags & O_NONBLOCK)
 260                 res = -EAGAIN;
 261         else
 262                 res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks);
 263
 264         /*
 265          * If clock has changed, we do not care about the
 266          * ticks and we do not rearm the timer. Userspace must
 267          * reevaluate anyway.
 268          */
 269         if (timerfd_canceled(ctx)) {
 270                 ctx->ticks = 0;
 271                 ctx->expired = 0;
 272                 res = -ECANCELED;
 273         }
 274
 275         if (ctx->ticks) {
 276                 ticks = ctx->ticks;
 277
 278                 if (ctx->expired && ctx->tintv) {
 279                         /*
 280                          * If tintv != 0, this is a periodic timer that
 281                          * needs to be re-armed. We avoid doing it in the timer
 282                          * callback to avoid DoS attacks specifying a very
 283                          * short timer period.
 284                          */
 285                         if (isalarm(ctx)) {
 286                                 ticks += alarm_forward_now(
 287                                         &ctx->t.alarm, ctx->tintv) - 1;
 288                                 alarm_restart(&ctx->t.alarm);
 289                         } else {
 290                                 ticks += hrtimer_forward_now(&ctx->t.tmr,
 291                                                              ctx->tintv) - 1;
 292                                 hrtimer_restart(&ctx->t.tmr);
 293                         }
 294                 }
 295                 ctx->expired = 0;
 296                 ctx->ticks = 0;
 297         }
 298         spin_unlock_irq(&ctx->wqh.lock);
 299         if (ticks)
 300                 res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks);
 301         return res;
 302 }
 303
 304 #ifdef CONFIG_PROC_FS
 305 static void timerfd_show(struct seq_file *m, struct file *file)
 306 {
 307         struct timerfd_ctx *ctx = file->private_data;
 308         struct timespec64 value, interval;
 309
 310         spin_lock_irq(&ctx->wqh.lock);
 311         value = ktime_to_timespec64(timerfd_get_remaining(ctx));
 312         interval = ktime_to_timespec64(ctx->tintv);
 313         spin_unlock_irq(&ctx->wqh.lock);
 314
 315         seq_printf(m,
 316                    "clockid: %d\n"
 317                    "ticks: %llu\n"
 318                    "settime flags: 0%o\n"
 319                    "it_value: (%llu, %llu)\n"
 320                    "it_interval: (%llu, %llu)\n",
 321                    ctx->clockid,
 322                    (unsigned long long)ctx->ticks,
 323                    ctx->settime_flags,
 324                    (unsigned long long)value.tv_sec,
 325                    (unsigned long long)value.tv_nsec,
 326                    (unsigned long long)interval.tv_sec,
 327                    (unsigned long long)interval.tv_nsec);
 328 }
 329 #else
 330 #define timerfd_show NULL
 331 #endif
 332
 333 #ifdef CONFIG_CHECKPOINT_RESTORE
 334 static long timerfd_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 335 {
 336         struct timerfd_ctx *ctx = file->private_data;
 337         int ret = 0;
 338
 339         switch (cmd) {
 340         case TFD_IOC_SET_TICKS: {
 341                 u64 ticks;
 342
 343                 if (copy_from_user(&ticks, (u64 __user *)arg, sizeof(ticks)))
 344                         return -EFAULT;
 345                 if (!ticks)
 346                         return -EINVAL;
 347
 348                 spin_lock_irq(&ctx->wqh.lock);
 349                 if (!timerfd_canceled(ctx)) {
 350                         ctx->ticks = ticks;
 351                         wake_up_locked_poll(&ctx->wqh, EPOLLIN);
 352                 } else
 353                         ret = -ECANCELED;
 354                 spin_unlock_irq(&ctx->wqh.lock);
 355                 break;
 356         }
 357         default:
 358                 ret = -ENOTTY;
 359                 break;
 360         }
 361
 362         return ret;
 363 }
 364 #else
 365 #define timerfd_ioctl NULL
 366 #endif
 367
 368 static const struct file_operations timerfd_fops = {
 369         .release        = timerfd_release,
 370         .poll           = timerfd_poll,
 371         .read           = timerfd_read,
 372         .llseek         = noop_llseek,
 373         .show_fdinfo    = timerfd_show,
 374         .unlocked_ioctl = timerfd_ioctl,
 375 };
 376
 377 static int timerfd_fget(int fd, struct fd *p)
 378 {
 379         struct fd f = fdget(fd);
 380         if (!f.file)
 381                 return -EBADF;
 382         if (f.file->f_op != &timerfd_fops) {
 383                 fdput(f);
 384                 return -EINVAL;
 385         }
 386         *p = f;
 387         return 0;
 388 }
 389
 390 SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
 391 {
 392         int ufd;
 393         struct timerfd_ctx *ctx;
 394
 395         /* Check the TFD_* constants for consistency.  */
 396         BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC);
 397         BUILD_BUG_ON(TFD_NONBLOCK != O_NONBLOCK);
 398
 399         if ((flags & ~TFD_CREATE_FLAGS) ||
 400             (clockid != CLOCK_MONOTONIC &&
 401              clockid != CLOCK_REALTIME &&
 402              clockid != CLOCK_REALTIME_ALARM &&
 403              clockid != CLOCK_BOOTTIME &&
 404              clockid != CLOCK_BOOTTIME_ALARM))
 405                 return -EINVAL;
 406
 407         if ((clockid == CLOCK_REALTIME_ALARM ||
 408              clockid == CLOCK_BOOTTIME_ALARM) &&
 409             !capable(CAP_WAKE_ALARM))
 410                 return -EPERM;
 411
 412         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 413         if (!ctx)
 414                 return -ENOMEM;
 415
 416         init_waitqueue_head(&ctx->wqh);
 417         spin_lock_init(&ctx->cancel_lock);
 418         ctx->clockid = clockid;
 419
 420         if (isalarm(ctx))
 421                 alarm_init(&ctx->t.alarm,
 422                            ctx->clockid == CLOCK_REALTIME_ALARM ?
 423                            ALARM_REALTIME : ALARM_BOOTTIME,
 424                            timerfd_alarmproc);
 425         else
 426                 hrtimer_init(&ctx->t.tmr, clockid, HRTIMER_MODE_ABS);
 427
 428         ctx->moffs = ktime_mono_to_real(0);
 429
 430         ufd = anon_inode_getfd("[timerfd]", &timerfd_fops, ctx,
 431                                O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
 432         if (ufd < 0)
 433                 kfree(ctx);
 434
 435         return ufd;
 436 }
 437
 438 static int do_timerfd_settime(int ufd, int flags,
 439                 const struct itimerspec64 *new,
 440                 struct itimerspec64 *old)
 441 {
 442         struct fd f;
 443         struct timerfd_ctx *ctx;
 444         int ret;
 445
 446         if ((flags & ~TFD_SETTIME_FLAGS) ||
 447                  !itimerspec64_valid(new))
 448                 return -EINVAL;
 449
 450         ret = timerfd_fget(ufd, &f);
 451         if (ret)
 452                 return ret;
 453         ctx = f.file->private_data;
 454
 455         if (isalarm(ctx) && !capable(CAP_WAKE_ALARM)) {
 456                 fdput(f);
 457                 return -EPERM;
 458         }
 459
 460         timerfd_setup_cancel(ctx, flags);
 461
 462         /*
 463          * We need to stop the existing timer before reprogramming
 464          * it to the new values.
 465          */
 466         for (;;) {
 467                 spin_lock_irq(&ctx->wqh.lock);
 468
 469                 if (isalarm(ctx)) {
 470                         if (alarm_try_to_cancel(&ctx->t.alarm) >= 0)
 471                                 break;
 472                 } else {
 473                         if (hrtimer_try_to_cancel(&ctx->t.tmr) >= 0)
 474                                 break;
 475                 }
 476                 spin_unlock_irq(&ctx->wqh.lock);
 477
 478                 if (isalarm(ctx))
 479                         hrtimer_cancel_wait_running(&ctx->t.alarm.timer);
 480                 else
 481                         hrtimer_cancel_wait_running(&ctx->t.tmr);
 482         }
 483
 484         /*
 485          * If the timer is expired and it's periodic, we need to advance it
 486          * because the caller may want to know the previous expiration time.
 487          * We do not update "ticks" and "expired" since the timer will be
 488          * re-programmed again in the following timerfd_setup() call.
 489          */
 490         if (ctx->expired && ctx->tintv) {
 491                 if (isalarm(ctx))
 492                         alarm_forward_now(&ctx->t.alarm, ctx->tintv);
 493                 else
 494                         hrtimer_forward_now(&ctx->t.tmr, ctx->tintv);
 495         }
 496
 497         old->it_value = ktime_to_timespec64(timerfd_get_remaining(ctx));
 498         old->it_interval = ktime_to_timespec64(ctx->tintv);
 499
 500         /*
 501          * Re-program the timer to the new value ...
 502          */
 503         ret = timerfd_setup(ctx, flags, new);
 504
 505         spin_unlock_irq(&ctx->wqh.lock);
 506         fdput(f);
 507         return ret;
 508 }
 509
 510 static int do_timerfd_gettime(int ufd, struct itimerspec64 *t)
 511 {
 512         struct fd f;
 513         struct timerfd_ctx *ctx;
 514         int ret = timerfd_fget(ufd, &f);
 515         if (ret)
 516                 return ret;
 517         ctx = f.file->private_data;
 518
 519         spin_lock_irq(&ctx->wqh.lock);
 520         if (ctx->expired && ctx->tintv) {
 521                 ctx->expired = 0;
 522
 523                 if (isalarm(ctx)) {
 524                         ctx->ticks +=
 525                                 alarm_forward_now(
 526                                         &ctx->t.alarm, ctx->tintv) - 1;
 527                         alarm_restart(&ctx->t.alarm);
 528                 } else {
 529                         ctx->ticks +=
 530                                 hrtimer_forward_now(&ctx->t.tmr, ctx->tintv)
 531                                 - 1;
 532                         hrtimer_restart(&ctx->t.tmr);
 533                 }
 534         }
 535         t->it_value = ktime_to_timespec64(timerfd_get_remaining(ctx));
 536         t->it_interval = ktime_to_timespec64(ctx->tintv);
 537         spin_unlock_irq(&ctx->wqh.lock);
 538         fdput(f);
 539         return 0;
 540 }
 541
 542 SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
 543                 const struct __kernel_itimerspec __user *, utmr,
 544                 struct __kernel_itimerspec __user *, otmr)
 545 {
 546         struct itimerspec64 new, old;
 547         int ret;
 548
 549         if (get_itimerspec64(&new, utmr))
 550                 return -EFAULT;
 551         ret = do_timerfd_settime(ufd, flags, &new, &old);
 552         if (ret)
 553                 return ret;
 554         if (otmr && put_itimerspec64(&old, otmr))
 555                 return -EFAULT;
 556
 557         return ret;
 558 }
 559
 560 SYSCALL_DEFINE2(timerfd_gettime, int, ufd, struct __kernel_itimerspec __user *, otmr)
 561 {
 562         struct itimerspec64 kotmr;
 563         int ret = do_timerfd_gettime(ufd, &kotmr);
 564         if (ret)
 565                 return ret;
 566         return put_itimerspec64(&kotmr, otmr) ? -EFAULT : 0;
 567 }
 568
 569 #ifdef CONFIG_COMPAT_32BIT_TIME
 570 SYSCALL_DEFINE4(timerfd_settime32, int, ufd, int, flags,
 571                 const struct old_itimerspec32 __user *, utmr,
 572                 struct old_itimerspec32 __user *, otmr)
 573 {
 574         struct itimerspec64 new, old;
 575         int ret;
 576
 577         if (get_old_itimerspec32(&new, utmr))
 578                 return -EFAULT;
 579         ret = do_timerfd_settime(ufd, flags, &new, &old);
 580         if (ret)
 581                 return ret;
 582         if (otmr && put_old_itimerspec32(&old, otmr))
 583                 return -EFAULT;
 584         return ret;
 585 }
 586
 587 SYSCALL_DEFINE2(timerfd_gettime32, int, ufd,
 588                 struct old_itimerspec32 __user *, otmr)
 589 {
 590         struct itimerspec64 kotmr;
 591         int ret = do_timerfd_gettime(ufd, &kotmr);
 592         if (ret)
 593                 return ret;
 594         return put_old_itimerspec32(&kotmr, otmr) ? -EFAULT : 0;
 595 }
 596 #endif