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_resume_work(struct work_struct *work)
 119 {
 120         timerfd_clock_was_set();
 121 }
 122
 123 static DECLARE_WORK(timerfd_work, timerfd_resume_work);
 124
 125 /*
 126  * Invoked from timekeeping_resume(). Defer the actual update to work so
 127  * timerfd_clock_was_set() runs in task context.
 128  */
 129 void timerfd_resume(void)
 130 {
 131         schedule_work(&timerfd_work);
 132 }
 133
 134 static void __timerfd_remove_cancel(struct timerfd_ctx *ctx)
 135 {
 136         if (ctx->might_cancel) {
 137                 ctx->might_cancel = false;
 138                 spin_lock(&cancel_lock);
 139                 list_del_rcu(&ctx->clist);
 140                 spin_unlock(&cancel_lock);
 141         }
 142 }
 143
 144 static void timerfd_remove_cancel(struct timerfd_ctx *ctx)
 145 {
 146         spin_lock(&ctx->cancel_lock);
 147         __timerfd_remove_cancel(ctx);
 148         spin_unlock(&ctx->cancel_lock);
 149 }
 150
 151 static bool timerfd_canceled(struct timerfd_ctx *ctx)
 152 {
 153         if (!ctx->might_cancel || ctx->moffs != KTIME_MAX)
 154                 return false;
 155         ctx->moffs = ktime_mono_to_real(0);
 156         return true;
 157 }
 158
 159 static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags)
 160 {
 161         spin_lock(&ctx->cancel_lock);
 162         if ((ctx->clockid == CLOCK_REALTIME ||
 163              ctx->clockid == CLOCK_REALTIME_ALARM) &&
 164             (flags & TFD_TIMER_ABSTIME) && (flags & TFD_TIMER_CANCEL_ON_SET)) {
 165                 if (!ctx->might_cancel) {
 166                         ctx->might_cancel = true;
 167                         spin_lock(&cancel_lock);
 168                         list_add_rcu(&ctx->clist, &cancel_list);
 169                         spin_unlock(&cancel_lock);
 170                 }
 171         } else {
 172                 __timerfd_remove_cancel(ctx);
 173         }
 174         spin_unlock(&ctx->cancel_lock);
 175 }
 176
 177 static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx)
 178 {
 179         ktime_t remaining;
 180
 181         if (isalarm(ctx))
 182                 remaining = alarm_expires_remaining(&ctx->t.alarm);
 183         else
 184                 remaining = hrtimer_expires_remaining_adjusted(&ctx->t.tmr);
 185
 186         return remaining < 0 ? 0: remaining;
 187 }
 188
 189 static int timerfd_setup(struct timerfd_ctx *ctx, int flags,
 190                          const struct itimerspec64 *ktmr)
 191 {
 192         enum hrtimer_mode htmode;
 193         ktime_t texp;
 194         int clockid = ctx->clockid;
 195
 196         htmode = (flags & TFD_TIMER_ABSTIME) ?
 197                 HRTIMER_MODE_ABS: HRTIMER_MODE_REL;
 198
 199         texp = timespec64_to_ktime(ktmr->it_value);
 200         ctx->expired = 0;
 201         ctx->ticks = 0;
 202         ctx->tintv = timespec64_to_ktime(ktmr->it_interval);
 203
 204         if (isalarm(ctx)) {
 205                 alarm_init(&ctx->t.alarm,
 206                            ctx->clockid == CLOCK_REALTIME_ALARM ?
 207                            ALARM_REALTIME : ALARM_BOOTTIME,
 208                            timerfd_alarmproc);
 209         } else {
 210                 hrtimer_init(&ctx->t.tmr, clockid, htmode);
 211                 hrtimer_set_expires(&ctx->t.tmr, texp);
 212                 ctx->t.tmr.function = timerfd_tmrproc;
 213         }
 214
 215         if (texp != 0) {
 216                 if (flags & TFD_TIMER_ABSTIME)
 217                         texp = timens_ktime_to_host(clockid, texp);
 218                 if (isalarm(ctx)) {
 219                         if (flags & TFD_TIMER_ABSTIME)
 220                                 alarm_start(&ctx->t.alarm, texp);
 221                         else
 222                                 alarm_start_relative(&ctx->t.alarm, texp);
 223                 } else {
 224                         hrtimer_start(&ctx->t.tmr, texp, htmode);
 225                 }
 226
 227                 if (timerfd_canceled(ctx))
 228                         return -ECANCELED;
 229         }
 230
 231         ctx->settime_flags = flags & TFD_SETTIME_FLAGS;
 232         return 0;
 233 }
 234
 235 static int timerfd_release(struct inode *inode, struct file *file)
 236 {
 237         struct timerfd_ctx *ctx = file->private_data;
 238
 239         timerfd_remove_cancel(ctx);
 240
 241         if (isalarm(ctx))
 242                 alarm_cancel(&ctx->t.alarm);
 243         else
 244                 hrtimer_cancel(&ctx->t.tmr);
 245         kfree_rcu(ctx, rcu);
 246         return 0;
 247 }
 248
 249 static __poll_t timerfd_poll(struct file *file, poll_table *wait)
 250 {
 251         struct timerfd_ctx *ctx = file->private_data;
 252         __poll_t events = 0;
 253         unsigned long flags;
 254
 255         poll_wait(file, &ctx->wqh, wait);
 256
 257         spin_lock_irqsave(&ctx->wqh.lock, flags);
 258         if (ctx->ticks)
 259                 events |= EPOLLIN;
 260         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
 261
 262         return events;
 263 }
 264
 265 static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
 266                             loff_t *ppos)
 267 {
 268         struct timerfd_ctx *ctx = file->private_data;
 269         ssize_t res;
 270         u64 ticks = 0;
 271
 272         if (count < sizeof(ticks))
 273                 return -EINVAL;
 274         spin_lock_irq(&ctx->wqh.lock);
 275         if (file->f_flags & O_NONBLOCK)
 276                 res = -EAGAIN;
 277         else
 278                 res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks);
 279
 280         /*
 281          * If clock has changed, we do not care about the
 282          * ticks and we do not rearm the timer. Userspace must
 283          * reevaluate anyway.
 284          */
 285         if (timerfd_canceled(ctx)) {
 286                 ctx->ticks = 0;
 287                 ctx->expired = 0;
 288                 res = -ECANCELED;
 289         }
 290
 291         if (ctx->ticks) {
 292                 ticks = ctx->ticks;
 293
 294                 if (ctx->expired && ctx->tintv) {
 295                         /*
 296                          * If tintv != 0, this is a periodic timer that
 297                          * needs to be re-armed. We avoid doing it in the timer
 298                          * callback to avoid DoS attacks specifying a very
 299                          * short timer period.
 300                          */
 301                         if (isalarm(ctx)) {
 302                                 ticks += alarm_forward_now(
 303                                         &ctx->t.alarm, ctx->tintv) - 1;
 304                                 alarm_restart(&ctx->t.alarm);
 305                         } else {
 306                                 ticks += hrtimer_forward_now(&ctx->t.tmr,
 307                                                              ctx->tintv) - 1;
 308                                 hrtimer_restart(&ctx->t.tmr);
 309                         }
 310                 }
 311                 ctx->expired = 0;
 312                 ctx->ticks = 0;
 313         }
 314         spin_unlock_irq(&ctx->wqh.lock);
 315         if (ticks)
 316                 res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks);
 317         return res;
 318 }
 319
 320 #ifdef CONFIG_PROC_FS
 321 static void timerfd_show(struct seq_file *m, struct file *file)
 322 {
 323         struct timerfd_ctx *ctx = file->private_data;
 324         struct timespec64 value, interval;
 325
 326         spin_lock_irq(&ctx->wqh.lock);
 327         value = ktime_to_timespec64(timerfd_get_remaining(ctx));
 328         interval = ktime_to_timespec64(ctx->tintv);
 329         spin_unlock_irq(&ctx->wqh.lock);
 330
 331         seq_printf(m,
 332                    "clockid: %d\n"
 333                    "ticks: %llu\n"
 334                    "settime flags: 0%o\n"
 335                    "it_value: (%llu, %llu)\n"
 336                    "it_interval: (%llu, %llu)\n",
 337                    ctx->clockid,
 338                    (unsigned long long)ctx->ticks,
 339                    ctx->settime_flags,
 340                    (unsigned long long)value.tv_sec,
 341                    (unsigned long long)value.tv_nsec,
 342                    (unsigned long long)interval.tv_sec,
 343                    (unsigned long long)interval.tv_nsec);
 344 }
 345 #else
 346 #define timerfd_show NULL
 347 #endif
 348
 349 #ifdef CONFIG_CHECKPOINT_RESTORE
 350 static long timerfd_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 351 {
 352         struct timerfd_ctx *ctx = file->private_data;
 353         int ret = 0;
 354
 355         switch (cmd) {
 356         case TFD_IOC_SET_TICKS: {
 357                 u64 ticks;
 358
 359                 if (copy_from_user(&ticks, (u64 __user *)arg, sizeof(ticks)))
 360                         return -EFAULT;
 361                 if (!ticks)
 362                         return -EINVAL;
 363
 364                 spin_lock_irq(&ctx->wqh.lock);
 365                 if (!timerfd_canceled(ctx)) {
 366                         ctx->ticks = ticks;
 367                         wake_up_locked_poll(&ctx->wqh, EPOLLIN);
 368                 } else
 369                         ret = -ECANCELED;
 370                 spin_unlock_irq(&ctx->wqh.lock);
 371                 break;
 372         }
 373         default:
 374                 ret = -ENOTTY;
 375                 break;
 376         }
 377
 378         return ret;
 379 }
 380 #else
 381 #define timerfd_ioctl NULL
 382 #endif
 383
 384 static const struct file_operations timerfd_fops = {
 385         .release        = timerfd_release,
 386         .poll           = timerfd_poll,
 387         .read           = timerfd_read,
 388         .llseek         = noop_llseek,
 389         .show_fdinfo    = timerfd_show,
 390         .unlocked_ioctl = timerfd_ioctl,
 391 };
 392
 393 static int timerfd_fget(int fd, struct fd *p)
 394 {
 395         struct fd f = fdget(fd);
 396         if (!f.file)
 397                 return -EBADF;
 398         if (f.file->f_op != &timerfd_fops) {
 399                 fdput(f);
 400                 return -EINVAL;
 401         }
 402         *p = f;
 403         return 0;
 404 }
 405
 406 SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
 407 {
 408         int ufd;
 409         struct timerfd_ctx *ctx;
 410
 411         /* Check the TFD_* constants for consistency.  */
 412         BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC);
 413         BUILD_BUG_ON(TFD_NONBLOCK != O_NONBLOCK);
 414
 415         if ((flags & ~TFD_CREATE_FLAGS) ||
 416             (clockid != CLOCK_MONOTONIC &&
 417              clockid != CLOCK_REALTIME &&
 418              clockid != CLOCK_REALTIME_ALARM &&
 419              clockid != CLOCK_BOOTTIME &&
 420              clockid != CLOCK_BOOTTIME_ALARM))
 421                 return -EINVAL;
 422
 423         if ((clockid == CLOCK_REALTIME_ALARM ||
 424              clockid == CLOCK_BOOTTIME_ALARM) &&
 425             !capable(CAP_WAKE_ALARM))
 426                 return -EPERM;
 427
 428         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 429         if (!ctx)
 430                 return -ENOMEM;
 431
 432         init_waitqueue_head(&ctx->wqh);
 433         spin_lock_init(&ctx->cancel_lock);
 434         ctx->clockid = clockid;
 435
 436         if (isalarm(ctx))
 437                 alarm_init(&ctx->t.alarm,
 438                            ctx->clockid == CLOCK_REALTIME_ALARM ?
 439                            ALARM_REALTIME : ALARM_BOOTTIME,
 440                            timerfd_alarmproc);
 441         else
 442                 hrtimer_init(&ctx->t.tmr, clockid, HRTIMER_MODE_ABS);
 443
 444         ctx->moffs = ktime_mono_to_real(0);
 445
 446         ufd = anon_inode_getfd("[timerfd]", &timerfd_fops, ctx,
 447                                O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
 448         if (ufd < 0)
 449                 kfree(ctx);
 450
 451         return ufd;
 452 }
 453
 454 static int do_timerfd_settime(int ufd, int flags,
 455                 const struct itimerspec64 *new,
 456                 struct itimerspec64 *old)
 457 {
 458         struct fd f;
 459         struct timerfd_ctx *ctx;
 460         int ret;
 461
 462         if ((flags & ~TFD_SETTIME_FLAGS) ||
 463                  !itimerspec64_valid(new))
 464                 return -EINVAL;
 465
 466         ret = timerfd_fget(ufd, &f);
 467         if (ret)
 468                 return ret;
 469         ctx = f.file->private_data;
 470
 471         if (isalarm(ctx) && !capable(CAP_WAKE_ALARM)) {
 472                 fdput(f);
 473                 return -EPERM;
 474         }
 475
 476         timerfd_setup_cancel(ctx, flags);
 477
 478         /*
 479          * We need to stop the existing timer before reprogramming
 480          * it to the new values.
 481          */
 482         for (;;) {
 483                 spin_lock_irq(&ctx->wqh.lock);
 484
 485                 if (isalarm(ctx)) {
 486                         if (alarm_try_to_cancel(&ctx->t.alarm) >= 0)
 487                                 break;
 488                 } else {
 489                         if (hrtimer_try_to_cancel(&ctx->t.tmr) >= 0)
 490                                 break;
 491                 }
 492                 spin_unlock_irq(&ctx->wqh.lock);
 493
 494                 if (isalarm(ctx))
 495                         hrtimer_cancel_wait_running(&ctx->t.alarm.timer);
 496                 else
 497                         hrtimer_cancel_wait_running(&ctx->t.tmr);
 498         }
 499
 500         /*
 501          * If the timer is expired and it's periodic, we need to advance it
 502          * because the caller may want to know the previous expiration time.
 503          * We do not update "ticks" and "expired" since the timer will be
 504          * re-programmed again in the following timerfd_setup() call.
 505          */
 506         if (ctx->expired && ctx->tintv) {
 507                 if (isalarm(ctx))
 508                         alarm_forward_now(&ctx->t.alarm, ctx->tintv);
 509                 else
 510                         hrtimer_forward_now(&ctx->t.tmr, ctx->tintv);
 511         }
 512
 513         old->it_value = ktime_to_timespec64(timerfd_get_remaining(ctx));
 514         old->it_interval = ktime_to_timespec64(ctx->tintv);
 515
 516         /*
 517          * Re-program the timer to the new value ...
 518          */
 519         ret = timerfd_setup(ctx, flags, new);
 520
 521         spin_unlock_irq(&ctx->wqh.lock);
 522         fdput(f);
 523         return ret;
 524 }
 525
 526 static int do_timerfd_gettime(int ufd, struct itimerspec64 *t)
 527 {
 528         struct fd f;
 529         struct timerfd_ctx *ctx;
 530         int ret = timerfd_fget(ufd, &f);
 531         if (ret)
 532                 return ret;
 533         ctx = f.file->private_data;
 534
 535         spin_lock_irq(&ctx->wqh.lock);
 536         if (ctx->expired && ctx->tintv) {
 537                 ctx->expired = 0;
 538
 539                 if (isalarm(ctx)) {
 540                         ctx->ticks +=
 541                                 alarm_forward_now(
 542                                         &ctx->t.alarm, ctx->tintv) - 1;
 543                         alarm_restart(&ctx->t.alarm);
 544                 } else {
 545                         ctx->ticks +=
 546                                 hrtimer_forward_now(&ctx->t.tmr, ctx->tintv)
 547                                 - 1;
 548                         hrtimer_restart(&ctx->t.tmr);
 549                 }
 550         }
 551         t->it_value = ktime_to_timespec64(timerfd_get_remaining(ctx));
 552         t->it_interval = ktime_to_timespec64(ctx->tintv);
 553         spin_unlock_irq(&ctx->wqh.lock);
 554         fdput(f);
 555         return 0;
 556 }
 557
 558 SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
 559                 const struct __kernel_itimerspec __user *, utmr,
 560                 struct __kernel_itimerspec __user *, otmr)
 561 {
 562         struct itimerspec64 new, old;
 563         int ret;
 564
 565         if (get_itimerspec64(&new, utmr))
 566                 return -EFAULT;
 567         ret = do_timerfd_settime(ufd, flags, &new, &old);
 568         if (ret)
 569                 return ret;
 570         if (otmr && put_itimerspec64(&old, otmr))
 571                 return -EFAULT;
 572
 573         return ret;
 574 }
 575
 576 SYSCALL_DEFINE2(timerfd_gettime, int, ufd, struct __kernel_itimerspec __user *, otmr)
 577 {
 578         struct itimerspec64 kotmr;
 579         int ret = do_timerfd_gettime(ufd, &kotmr);
 580         if (ret)
 581                 return ret;
 582         return put_itimerspec64(&kotmr, otmr) ? -EFAULT : 0;
 583 }
 584
 585 #ifdef CONFIG_COMPAT_32BIT_TIME
 586 SYSCALL_DEFINE4(timerfd_settime32, int, ufd, int, flags,
 587                 const struct old_itimerspec32 __user *, utmr,
 588                 struct old_itimerspec32 __user *, otmr)
 589 {
 590         struct itimerspec64 new, old;
 591         int ret;
 592
 593         if (get_old_itimerspec32(&new, utmr))
 594                 return -EFAULT;
 595         ret = do_timerfd_settime(ufd, flags, &new, &old);
 596         if (ret)
 597                 return ret;
 598         if (otmr && put_old_itimerspec32(&old, otmr))
 599                 return -EFAULT;
 600         return ret;
 601 }
 602
 603 SYSCALL_DEFINE2(timerfd_gettime32, int, ufd,
 604                 struct old_itimerspec32 __user *, otmr)
 605 {
 606         struct itimerspec64 kotmr;
 607         int ret = do_timerfd_gettime(ufd, &kotmr);
 608         if (ret)
 609                 return ret;
 610         return put_old_itimerspec32(&kotmr, otmr) ? -EFAULT : 0;
 611 }
 612 #endif