4 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
7 * Thanks to Thomas Gleixner for code reviews and useful comments.
11 #include <linux/file.h>
12 #include <linux/poll.h>
13 #include <linux/init.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>
31 wait_queue_head_t wqh;
36 struct list_head clist;
37 spinlock_t cancel_lock;
41 static LIST_HEAD(cancel_list);
42 static DEFINE_SPINLOCK(cancel_lock);
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.
49 static enum hrtimer_restart timerfd_tmrproc(struct hrtimer *htmr)
51 struct timerfd_ctx *ctx = container_of(htmr, struct timerfd_ctx, tmr);
54 spin_lock_irqsave(&ctx->wqh.lock, flags);
57 wake_up_locked(&ctx->wqh);
58 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
60 return HRTIMER_NORESTART;
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().
69 void timerfd_clock_was_set(void)
71 ktime_t moffs = ktime_get_monotonic_offset();
72 struct timerfd_ctx *ctx;
76 list_for_each_entry_rcu(ctx, &cancel_list, clist) {
77 if (!ctx->might_cancel)
79 spin_lock_irqsave(&ctx->wqh.lock, flags);
80 if (ctx->moffs.tv64 != moffs.tv64) {
81 ctx->moffs.tv64 = KTIME_MAX;
83 wake_up_locked(&ctx->wqh);
85 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
90 static void __timerfd_remove_cancel(struct timerfd_ctx *ctx)
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);
100 static void timerfd_remove_cancel(struct timerfd_ctx *ctx)
102 spin_lock(&ctx->cancel_lock);
103 __timerfd_remove_cancel(ctx);
104 spin_unlock(&ctx->cancel_lock);
107 static bool timerfd_canceled(struct timerfd_ctx *ctx)
109 if (!ctx->might_cancel || ctx->moffs.tv64 != KTIME_MAX)
111 ctx->moffs = ktime_get_monotonic_offset();
115 static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags)
117 spin_lock(&ctx->cancel_lock);
118 if (ctx->clockid == CLOCK_REALTIME && (flags & TFD_TIMER_ABSTIME) &&
119 (flags & TFD_TIMER_CANCEL_ON_SET)) {
120 if (!ctx->might_cancel) {
121 ctx->might_cancel = true;
122 spin_lock(&cancel_lock);
123 list_add_rcu(&ctx->clist, &cancel_list);
124 spin_unlock(&cancel_lock);
127 __timerfd_remove_cancel(ctx);
129 spin_unlock(&ctx->cancel_lock);
132 static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx)
136 remaining = hrtimer_expires_remaining_adjusted(&ctx->tmr);
137 return remaining.tv64 < 0 ? ktime_set(0, 0): remaining;
140 static int timerfd_setup(struct timerfd_ctx *ctx, int flags,
141 const struct itimerspec *ktmr)
143 enum hrtimer_mode htmode;
145 int clockid = ctx->clockid;
147 htmode = (flags & TFD_TIMER_ABSTIME) ?
148 HRTIMER_MODE_ABS: HRTIMER_MODE_REL;
150 texp = timespec_to_ktime(ktmr->it_value);
153 ctx->tintv = timespec_to_ktime(ktmr->it_interval);
154 hrtimer_init(&ctx->tmr, clockid, htmode);
155 hrtimer_set_expires(&ctx->tmr, texp);
156 ctx->tmr.function = timerfd_tmrproc;
157 if (texp.tv64 != 0) {
158 hrtimer_start(&ctx->tmr, texp, htmode);
159 if (timerfd_canceled(ctx))
165 static int timerfd_release(struct inode *inode, struct file *file)
167 struct timerfd_ctx *ctx = file->private_data;
169 timerfd_remove_cancel(ctx);
170 hrtimer_cancel(&ctx->tmr);
175 static unsigned int timerfd_poll(struct file *file, poll_table *wait)
177 struct timerfd_ctx *ctx = file->private_data;
178 unsigned int events = 0;
181 poll_wait(file, &ctx->wqh, wait);
183 spin_lock_irqsave(&ctx->wqh.lock, flags);
186 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
191 static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
194 struct timerfd_ctx *ctx = file->private_data;
198 if (count < sizeof(ticks))
200 spin_lock_irq(&ctx->wqh.lock);
201 if (file->f_flags & O_NONBLOCK)
204 res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks);
207 * If clock has changed, we do not care about the
208 * ticks and we do not rearm the timer. Userspace must
211 if (timerfd_canceled(ctx)) {
220 if (ctx->expired && ctx->tintv.tv64) {
222 * If tintv.tv64 != 0, this is a periodic timer that
223 * needs to be re-armed. We avoid doing it in the timer
224 * callback to avoid DoS attacks specifying a very
225 * short timer period.
227 ticks += hrtimer_forward_now(&ctx->tmr,
229 hrtimer_restart(&ctx->tmr);
234 spin_unlock_irq(&ctx->wqh.lock);
236 res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks);
240 static const struct file_operations timerfd_fops = {
241 .release = timerfd_release,
242 .poll = timerfd_poll,
243 .read = timerfd_read,
244 .llseek = noop_llseek,
247 static struct file *timerfd_fget(int fd)
253 return ERR_PTR(-EBADF);
254 if (file->f_op != &timerfd_fops) {
256 return ERR_PTR(-EINVAL);
262 SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
265 struct timerfd_ctx *ctx;
267 /* Check the TFD_* constants for consistency. */
268 BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC);
269 BUILD_BUG_ON(TFD_NONBLOCK != O_NONBLOCK);
271 if ((flags & ~TFD_CREATE_FLAGS) ||
272 (clockid != CLOCK_MONOTONIC &&
273 clockid != CLOCK_REALTIME))
276 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
280 init_waitqueue_head(&ctx->wqh);
281 spin_lock_init(&ctx->cancel_lock);
282 ctx->clockid = clockid;
283 hrtimer_init(&ctx->tmr, clockid, HRTIMER_MODE_ABS);
284 ctx->moffs = ktime_get_monotonic_offset();
286 ufd = anon_inode_getfd("[timerfd]", &timerfd_fops, ctx,
287 O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
294 SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
295 const struct itimerspec __user *, utmr,
296 struct itimerspec __user *, otmr)
299 struct timerfd_ctx *ctx;
300 struct itimerspec ktmr, kotmr;
303 if (copy_from_user(&ktmr, utmr, sizeof(ktmr)))
306 if ((flags & ~TFD_SETTIME_FLAGS) ||
307 !timespec_valid(&ktmr.it_value) ||
308 !timespec_valid(&ktmr.it_interval))
311 file = timerfd_fget(ufd);
313 return PTR_ERR(file);
314 ctx = file->private_data;
316 timerfd_setup_cancel(ctx, flags);
319 * We need to stop the existing timer before reprogramming
320 * it to the new values.
323 spin_lock_irq(&ctx->wqh.lock);
324 if (hrtimer_try_to_cancel(&ctx->tmr) >= 0)
326 spin_unlock_irq(&ctx->wqh.lock);
331 * If the timer is expired and it's periodic, we need to advance it
332 * because the caller may want to know the previous expiration time.
333 * We do not update "ticks" and "expired" since the timer will be
334 * re-programmed again in the following timerfd_setup() call.
336 if (ctx->expired && ctx->tintv.tv64)
337 hrtimer_forward_now(&ctx->tmr, ctx->tintv);
339 kotmr.it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
340 kotmr.it_interval = ktime_to_timespec(ctx->tintv);
343 * Re-program the timer to the new value ...
345 ret = timerfd_setup(ctx, flags, &ktmr);
347 spin_unlock_irq(&ctx->wqh.lock);
349 if (otmr && copy_to_user(otmr, &kotmr, sizeof(kotmr)))
355 SYSCALL_DEFINE2(timerfd_gettime, int, ufd, struct itimerspec __user *, otmr)
358 struct timerfd_ctx *ctx;
359 struct itimerspec kotmr;
361 file = timerfd_fget(ufd);
363 return PTR_ERR(file);
364 ctx = file->private_data;
366 spin_lock_irq(&ctx->wqh.lock);
367 if (ctx->expired && ctx->tintv.tv64) {
370 hrtimer_forward_now(&ctx->tmr, ctx->tintv) - 1;
371 hrtimer_restart(&ctx->tmr);
373 kotmr.it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
374 kotmr.it_interval = ktime_to_timespec(ctx->tintv);
375 spin_unlock_irq(&ctx->wqh.lock);
378 return copy_to_user(otmr, &kotmr, sizeof(kotmr)) ? -EFAULT: 0;