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/file.h>
  12 #include <linux/poll.h>
  13 #include <linux/init.h>
  14 #include <linux/fs.h>
  15 #include <linux/sched.h>
  16 #include <linux/kernel.h>
  17 #include <linux/slab.h>
  18 #include <linux/list.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/time.h>
  21 #include <linux/hrtimer.h>
  22 #include <linux/anon_inodes.h>
  23 #include <linux/timerfd.h>
  24 #include <linux/syscalls.h>
  25 #include <linux/compat.h>
  26 #include <linux/rcupdate.h>
  27
  28 struct timerfd_ctx {
  29         struct hrtimer tmr;
  30         ktime_t tintv;
  31         ktime_t moffs;
  32         wait_queue_head_t wqh;
  33         u64 ticks;
  34         int expired;
  35         int clockid;
  36         struct rcu_head rcu;
  37         struct list_head clist;
  38         bool might_cancel;
  39 };
  40
  41 static LIST_HEAD(cancel_list);
  42 static DEFINE_SPINLOCK(cancel_lock);
  43
  44 /*
  45  * This gets called when the timer event triggers. We set the "expired"
  46  * flag, but we do not re-arm the timer (in case it's necessary,
  47  * tintv.tv64 != 0) until the timer is accessed.
  48  */
  49 static enum hrtimer_restart timerfd_tmrproc(struct hrtimer *htmr)
  50 {
  51         struct timerfd_ctx *ctx = container_of(htmr, struct timerfd_ctx, tmr);
  52         unsigned long flags;
  53
  54         spin_lock_irqsave(&ctx->wqh.lock, flags);
  55         ctx->expired = 1;
  56         ctx->ticks++;
  57         wake_up_locked(&ctx->wqh);
  58         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
  59
  60         return HRTIMER_NORESTART;
  61 }
  62
  63 /*
  64  * Called when the clock was set to cancel the timers in the cancel
  65  * list. This will wake up processes waiting on these timers. The
  66  * wake-up requires ctx->ticks to be non zero, therefore we increment
  67  * it before calling wake_up_locked().
  68  */
  69 void timerfd_clock_was_set(void)
  70 {
  71         ktime_t moffs = ktime_get_monotonic_offset();
  72         struct timerfd_ctx *ctx;
  73         unsigned long flags;
  74
  75         rcu_read_lock();
  76         list_for_each_entry_rcu(ctx, &cancel_list, clist) {
  77                 if (!ctx->might_cancel)
  78                         continue;
  79                 spin_lock_irqsave(&ctx->wqh.lock, flags);
  80                 if (ctx->moffs.tv64 != moffs.tv64) {
  81                         ctx->moffs.tv64 = KTIME_MAX;
  82                         ctx->ticks++;
  83                         wake_up_locked(&ctx->wqh);
  84                 }
  85                 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
  86         }
  87         rcu_read_unlock();
  88 }
  89
  90 static void timerfd_remove_cancel(struct timerfd_ctx *ctx)
  91 {
  92         if (ctx->might_cancel) {
  93                 ctx->might_cancel = false;
  94                 spin_lock(&cancel_lock);
  95                 list_del_rcu(&ctx->clist);
  96                 spin_unlock(&cancel_lock);
  97         }
  98 }
  99
 100 static bool timerfd_canceled(struct timerfd_ctx *ctx)
 101 {
 102         if (!ctx->might_cancel || ctx->moffs.tv64 != KTIME_MAX)
 103                 return false;
 104         ctx->moffs = ktime_get_monotonic_offset();
 105         return true;
 106 }
 107
 108 static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags)
 109 {
 110         if (ctx->clockid == CLOCK_REALTIME && (flags & TFD_TIMER_ABSTIME) &&
 111             (flags & TFD_TIMER_CANCEL_ON_SET)) {
 112                 if (!ctx->might_cancel) {
 113                         ctx->might_cancel = true;
 114                         spin_lock(&cancel_lock);
 115                         list_add_rcu(&ctx->clist, &cancel_list);
 116                         spin_unlock(&cancel_lock);
 117                 }
 118         } else if (ctx->might_cancel) {
 119                 timerfd_remove_cancel(ctx);
 120         }
 121 }
 122
 123 static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx)
 124 {
 125         ktime_t remaining;
 126
 127         remaining = hrtimer_expires_remaining(&ctx->tmr);
 128         return remaining.tv64 < 0 ? ktime_set(0, 0): remaining;
 129 }
 130
 131 static int timerfd_setup(struct timerfd_ctx *ctx, int flags,
 132                          const struct itimerspec *ktmr)
 133 {
 134         enum hrtimer_mode htmode;
 135         ktime_t texp;
 136         int clockid = ctx->clockid;
 137
 138         htmode = (flags & TFD_TIMER_ABSTIME) ?
 139                 HRTIMER_MODE_ABS: HRTIMER_MODE_REL;
 140
 141         texp = timespec_to_ktime(ktmr->it_value);
 142         ctx->expired = 0;
 143         ctx->ticks = 0;
 144         ctx->tintv = timespec_to_ktime(ktmr->it_interval);
 145         hrtimer_init(&ctx->tmr, clockid, htmode);
 146         hrtimer_set_expires(&ctx->tmr, texp);
 147         ctx->tmr.function = timerfd_tmrproc;
 148         if (texp.tv64 != 0) {
 149                 hrtimer_start(&ctx->tmr, texp, htmode);
 150                 if (timerfd_canceled(ctx))
 151                         return -ECANCELED;
 152         }
 153         return 0;
 154 }
 155
 156 static int timerfd_release(struct inode *inode, struct file *file)
 157 {
 158         struct timerfd_ctx *ctx = file->private_data;
 159
 160         timerfd_remove_cancel(ctx);
 161         hrtimer_cancel(&ctx->tmr);
 162         kfree_rcu(ctx, rcu);
 163         return 0;
 164 }
 165
 166 static unsigned int timerfd_poll(struct file *file, poll_table *wait)
 167 {
 168         struct timerfd_ctx *ctx = file->private_data;
 169         unsigned int events = 0;
 170         unsigned long flags;
 171
 172         poll_wait(file, &ctx->wqh, wait);
 173
 174         spin_lock_irqsave(&ctx->wqh.lock, flags);
 175         if (ctx->ticks)
 176                 events |= POLLIN;
 177         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
 178
 179         return events;
 180 }
 181
 182 static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
 183                             loff_t *ppos)
 184 {
 185         struct timerfd_ctx *ctx = file->private_data;
 186         ssize_t res;
 187         u64 ticks = 0;
 188
 189         if (count < sizeof(ticks))
 190                 return -EINVAL;
 191         spin_lock_irq(&ctx->wqh.lock);
 192         if (file->f_flags & O_NONBLOCK)
 193                 res = -EAGAIN;
 194         else
 195                 res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks);
 196
 197         /*
 198          * If clock has changed, we do not care about the
 199          * ticks and we do not rearm the timer. Userspace must
 200          * reevaluate anyway.
 201          */
 202         if (timerfd_canceled(ctx)) {
 203                 ctx->ticks = 0;
 204                 ctx->expired = 0;
 205                 res = -ECANCELED;
 206         }
 207
 208         if (ctx->ticks) {
 209                 ticks = ctx->ticks;
 210
 211                 if (ctx->expired && ctx->tintv.tv64) {
 212                         /*
 213                          * If tintv.tv64 != 0, this is a periodic timer that
 214                          * needs to be re-armed. We avoid doing it in the timer
 215                          * callback to avoid DoS attacks specifying a very
 216                          * short timer period.
 217                          */
 218                         ticks += hrtimer_forward_now(&ctx->tmr,
 219                                                      ctx->tintv) - 1;
 220                         hrtimer_restart(&ctx->tmr);
 221                 }
 222                 ctx->expired = 0;
 223                 ctx->ticks = 0;
 224         }
 225         spin_unlock_irq(&ctx->wqh.lock);
 226         if (ticks)
 227                 res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks);
 228         return res;
 229 }
 230
 231 static const struct file_operations timerfd_fops = {
 232         .release        = timerfd_release,
 233         .poll           = timerfd_poll,
 234         .read           = timerfd_read,
 235         .llseek         = noop_llseek,
 236 };
 237
 238 static int timerfd_fget(int fd, struct fd *p)
 239 {
 240         struct fd f = fdget(fd);
 241         if (!f.file)
 242                 return -EBADF;
 243         if (f.file->f_op != &timerfd_fops) {
 244                 fdput(f);
 245                 return -EINVAL;
 246         }
 247         *p = f;
 248         return 0;
 249 }
 250
 251 SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
 252 {
 253         int ufd;
 254         struct timerfd_ctx *ctx;
 255
 256         /* Check the TFD_* constants for consistency.  */
 257         BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC);
 258         BUILD_BUG_ON(TFD_NONBLOCK != O_NONBLOCK);
 259
 260         if ((flags & ~TFD_CREATE_FLAGS) ||
 261             (clockid != CLOCK_MONOTONIC &&
 262              clockid != CLOCK_REALTIME))
 263                 return -EINVAL;
 264
 265         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 266         if (!ctx)
 267                 return -ENOMEM;
 268
 269         init_waitqueue_head(&ctx->wqh);
 270         ctx->clockid = clockid;
 271         hrtimer_init(&ctx->tmr, clockid, HRTIMER_MODE_ABS);
 272         ctx->moffs = ktime_get_monotonic_offset();
 273
 274         ufd = anon_inode_getfd("[timerfd]", &timerfd_fops, ctx,
 275                                O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
 276         if (ufd < 0)
 277                 kfree(ctx);
 278
 279         return ufd;
 280 }
 281
 282 static int do_timerfd_settime(int ufd, int flags,
 283                 const struct itimerspec *new,
 284                 struct itimerspec *old)
 285 {
 286         struct fd f;
 287         struct timerfd_ctx *ctx;
 288         int ret;
 289
 290         if ((flags & ~TFD_SETTIME_FLAGS) ||
 291             !timespec_valid(&new->it_value) ||
 292             !timespec_valid(&new->it_interval))
 293                 return -EINVAL;
 294
 295         ret = timerfd_fget(ufd, &f);
 296         if (ret)
 297                 return ret;
 298         ctx = f.file->private_data;
 299
 300         timerfd_setup_cancel(ctx, flags);
 301
 302         /*
 303          * We need to stop the existing timer before reprogramming
 304          * it to the new values.
 305          */
 306         for (;;) {
 307                 spin_lock_irq(&ctx->wqh.lock);
 308                 if (hrtimer_try_to_cancel(&ctx->tmr) >= 0)
 309                         break;
 310                 spin_unlock_irq(&ctx->wqh.lock);
 311                 cpu_relax();
 312         }
 313
 314         /*
 315          * If the timer is expired and it's periodic, we need to advance it
 316          * because the caller may want to know the previous expiration time.
 317          * We do not update "ticks" and "expired" since the timer will be
 318          * re-programmed again in the following timerfd_setup() call.
 319          */
 320         if (ctx->expired && ctx->tintv.tv64)
 321                 hrtimer_forward_now(&ctx->tmr, ctx->tintv);
 322
 323         old->it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
 324         old->it_interval = ktime_to_timespec(ctx->tintv);
 325
 326         /*
 327          * Re-program the timer to the new value ...
 328          */
 329         ret = timerfd_setup(ctx, flags, new);
 330
 331         spin_unlock_irq(&ctx->wqh.lock);
 332         fdput(f);
 333         return ret;
 334 }
 335
 336 static int do_timerfd_gettime(int ufd, struct itimerspec *t)
 337 {
 338         struct fd f;
 339         struct timerfd_ctx *ctx;
 340         int ret = timerfd_fget(ufd, &f);
 341         if (ret)
 342                 return ret;
 343         ctx = f.file->private_data;
 344
 345         spin_lock_irq(&ctx->wqh.lock);
 346         if (ctx->expired && ctx->tintv.tv64) {
 347                 ctx->expired = 0;
 348                 ctx->ticks +=
 349                         hrtimer_forward_now(&ctx->tmr, ctx->tintv) - 1;
 350                 hrtimer_restart(&ctx->tmr);
 351         }
 352         t->it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
 353         t->it_interval = ktime_to_timespec(ctx->tintv);
 354         spin_unlock_irq(&ctx->wqh.lock);
 355         fdput(f);
 356         return 0;
 357 }
 358
 359 SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
 360                 const struct itimerspec __user *, utmr,
 361                 struct itimerspec __user *, otmr)
 362 {
 363         struct itimerspec new, old;
 364         int ret;
 365
 366         if (copy_from_user(&new, utmr, sizeof(new)))
 367                 return -EFAULT;
 368         ret = do_timerfd_settime(ufd, flags, &new, &old);
 369         if (ret)
 370                 return ret;
 371         if (otmr && copy_to_user(otmr, &old, sizeof(old)))
 372                 return -EFAULT;
 373
 374         return ret;
 375 }
 376
 377 SYSCALL_DEFINE2(timerfd_gettime, int, ufd, struct itimerspec __user *, otmr)
 378 {
 379         struct itimerspec kotmr;
 380         int ret = do_timerfd_gettime(ufd, &kotmr);
 381         if (ret)
 382                 return ret;
 383         return copy_to_user(otmr, &kotmr, sizeof(kotmr)) ? -EFAULT: 0;
 384 }
 385
 386 #ifdef CONFIG_COMPAT
 387 COMPAT_SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
 388                 const struct compat_itimerspec __user *, utmr,
 389                 struct compat_itimerspec __user *, otmr)
 390 {
 391         struct itimerspec new, old;
 392         int ret;
 393
 394         if (get_compat_itimerspec(&new, utmr))
 395                 return -EFAULT;
 396         ret = do_timerfd_settime(ufd, flags, &new, &old);
 397         if (ret)
 398                 return ret;
 399         if (otmr && put_compat_itimerspec(otmr, &old))
 400                 return -EFAULT;
 401         return ret;
 402 }
 403
 404 COMPAT_SYSCALL_DEFINE2(timerfd_gettime, int, ufd,
 405                 struct compat_itimerspec __user *, otmr)
 406 {
 407         struct itimerspec kotmr;
 408         int ret = do_timerfd_gettime(ufd, &kotmr);
 409         if (ret)
 410                 return ret;
 411         return put_compat_itimerspec(otmr, &kotmr) ? -EFAULT: 0;
 412 }
 413 #endif