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