]>
Commit | Line | Data |
---|---|---|
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/alarmtimer.h> | |
12 | #include <linux/file.h> | |
13 | #include <linux/poll.h> | |
14 | #include <linux/init.h> | |
15 | #include <linux/fs.h> | |
16 | #include <linux/sched.h> | |
17 | #include <linux/kernel.h> | |
18 | #include <linux/slab.h> | |
19 | #include <linux/list.h> | |
20 | #include <linux/spinlock.h> | |
21 | #include <linux/time.h> | |
22 | #include <linux/hrtimer.h> | |
23 | #include <linux/anon_inodes.h> | |
24 | #include <linux/timerfd.h> | |
25 | #include <linux/syscalls.h> | |
26 | #include <linux/compat.h> | |
27 | #include <linux/rcupdate.h> | |
28 | ||
29 | struct timerfd_ctx { | |
30 | union { | |
31 | struct hrtimer tmr; | |
32 | struct alarm alarm; | |
33 | } t; | |
34 | ktime_t tintv; | |
35 | ktime_t moffs; | |
36 | wait_queue_head_t wqh; | |
37 | u64 ticks; | |
38 | int clockid; | |
39 | short unsigned expired; | |
40 | short unsigned settime_flags; /* to show in fdinfo */ | |
41 | struct rcu_head rcu; | |
42 | struct list_head clist; | |
43 | bool might_cancel; | |
44 | }; | |
45 | ||
46 | static LIST_HEAD(cancel_list); | |
47 | static DEFINE_SPINLOCK(cancel_lock); | |
48 | ||
49 | static inline bool isalarm(struct timerfd_ctx *ctx) | |
50 | { | |
51 | return ctx->clockid == CLOCK_REALTIME_ALARM || | |
52 | ctx->clockid == CLOCK_BOOTTIME_ALARM; | |
53 | } | |
54 | ||
55 | /* | |
56 | * This gets called when the timer event triggers. We set the "expired" | |
57 | * flag, but we do not re-arm the timer (in case it's necessary, | |
58 | * tintv.tv64 != 0) until the timer is accessed. | |
59 | */ | |
60 | static void timerfd_triggered(struct timerfd_ctx *ctx) | |
61 | { | |
62 | unsigned long flags; | |
63 | ||
64 | spin_lock_irqsave(&ctx->wqh.lock, flags); | |
65 | ctx->expired = 1; | |
66 | ctx->ticks++; | |
67 | wake_up_locked(&ctx->wqh); | |
68 | spin_unlock_irqrestore(&ctx->wqh.lock, flags); | |
69 | } | |
70 | ||
71 | static enum hrtimer_restart timerfd_tmrproc(struct hrtimer *htmr) | |
72 | { | |
73 | struct timerfd_ctx *ctx = container_of(htmr, struct timerfd_ctx, | |
74 | t.tmr); | |
75 | timerfd_triggered(ctx); | |
76 | return HRTIMER_NORESTART; | |
77 | } | |
78 | ||
79 | static enum alarmtimer_restart timerfd_alarmproc(struct alarm *alarm, | |
80 | ktime_t now) | |
81 | { | |
82 | struct timerfd_ctx *ctx = container_of(alarm, struct timerfd_ctx, | |
83 | t.alarm); | |
84 | timerfd_triggered(ctx); | |
85 | return ALARMTIMER_NORESTART; | |
86 | } | |
87 | ||
88 | /* | |
89 | * Called when the clock was set to cancel the timers in the cancel | |
90 | * list. This will wake up processes waiting on these timers. The | |
91 | * wake-up requires ctx->ticks to be non zero, therefore we increment | |
92 | * it before calling wake_up_locked(). | |
93 | */ | |
94 | void timerfd_clock_was_set(void) | |
95 | { | |
96 | ktime_t moffs = ktime_mono_to_real((ktime_t){ .tv64 = 0 }); | |
97 | struct timerfd_ctx *ctx; | |
98 | unsigned long flags; | |
99 | ||
100 | rcu_read_lock(); | |
101 | list_for_each_entry_rcu(ctx, &cancel_list, clist) { | |
102 | if (!ctx->might_cancel) | |
103 | continue; | |
104 | spin_lock_irqsave(&ctx->wqh.lock, flags); | |
105 | if (ctx->moffs.tv64 != moffs.tv64) { | |
106 | ctx->moffs.tv64 = KTIME_MAX; | |
107 | ctx->ticks++; | |
108 | wake_up_locked(&ctx->wqh); | |
109 | } | |
110 | spin_unlock_irqrestore(&ctx->wqh.lock, flags); | |
111 | } | |
112 | rcu_read_unlock(); | |
113 | } | |
114 | ||
115 | static void timerfd_remove_cancel(struct timerfd_ctx *ctx) | |
116 | { | |
117 | if (ctx->might_cancel) { | |
118 | ctx->might_cancel = false; | |
119 | spin_lock(&cancel_lock); | |
120 | list_del_rcu(&ctx->clist); | |
121 | spin_unlock(&cancel_lock); | |
122 | } | |
123 | } | |
124 | ||
125 | static bool timerfd_canceled(struct timerfd_ctx *ctx) | |
126 | { | |
127 | if (!ctx->might_cancel || ctx->moffs.tv64 != KTIME_MAX) | |
128 | return false; | |
129 | ctx->moffs = ktime_mono_to_real((ktime_t){ .tv64 = 0 }); | |
130 | return true; | |
131 | } | |
132 | ||
133 | static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags) | |
134 | { | |
135 | if ((ctx->clockid == CLOCK_REALTIME || | |
136 | ctx->clockid == CLOCK_REALTIME_ALARM) && | |
137 | (flags & TFD_TIMER_ABSTIME) && (flags & TFD_TIMER_CANCEL_ON_SET)) { | |
138 | if (!ctx->might_cancel) { | |
139 | ctx->might_cancel = true; | |
140 | spin_lock(&cancel_lock); | |
141 | list_add_rcu(&ctx->clist, &cancel_list); | |
142 | spin_unlock(&cancel_lock); | |
143 | } | |
144 | } else if (ctx->might_cancel) { | |
145 | timerfd_remove_cancel(ctx); | |
146 | } | |
147 | } | |
148 | ||
149 | static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx) | |
150 | { | |
151 | ktime_t remaining; | |
152 | ||
153 | if (isalarm(ctx)) | |
154 | remaining = alarm_expires_remaining(&ctx->t.alarm); | |
155 | else | |
156 | remaining = hrtimer_expires_remaining(&ctx->t.tmr); | |
157 | ||
158 | return remaining.tv64 < 0 ? ktime_set(0, 0): remaining; | |
159 | } | |
160 | ||
161 | static int timerfd_setup(struct timerfd_ctx *ctx, int flags, | |
162 | const struct itimerspec *ktmr) | |
163 | { | |
164 | enum hrtimer_mode htmode; | |
165 | ktime_t texp; | |
166 | int clockid = ctx->clockid; | |
167 | ||
168 | htmode = (flags & TFD_TIMER_ABSTIME) ? | |
169 | HRTIMER_MODE_ABS: HRTIMER_MODE_REL; | |
170 | ||
171 | texp = timespec_to_ktime(ktmr->it_value); | |
172 | ctx->expired = 0; | |
173 | ctx->ticks = 0; | |
174 | ctx->tintv = timespec_to_ktime(ktmr->it_interval); | |
175 | ||
176 | if (isalarm(ctx)) { | |
177 | alarm_init(&ctx->t.alarm, | |
178 | ctx->clockid == CLOCK_REALTIME_ALARM ? | |
179 | ALARM_REALTIME : ALARM_BOOTTIME, | |
180 | timerfd_alarmproc); | |
181 | } else { | |
182 | hrtimer_init(&ctx->t.tmr, clockid, htmode); | |
183 | hrtimer_set_expires(&ctx->t.tmr, texp); | |
184 | ctx->t.tmr.function = timerfd_tmrproc; | |
185 | } | |
186 | ||
187 | if (texp.tv64 != 0) { | |
188 | if (isalarm(ctx)) { | |
189 | if (flags & TFD_TIMER_ABSTIME) | |
190 | alarm_start(&ctx->t.alarm, texp); | |
191 | else | |
192 | alarm_start_relative(&ctx->t.alarm, texp); | |
193 | } else { | |
194 | hrtimer_start(&ctx->t.tmr, texp, htmode); | |
195 | } | |
196 | ||
197 | if (timerfd_canceled(ctx)) | |
198 | return -ECANCELED; | |
199 | } | |
200 | ||
201 | ctx->settime_flags = flags & TFD_SETTIME_FLAGS; | |
202 | return 0; | |
203 | } | |
204 | ||
205 | static int timerfd_release(struct inode *inode, struct file *file) | |
206 | { | |
207 | struct timerfd_ctx *ctx = file->private_data; | |
208 | ||
209 | timerfd_remove_cancel(ctx); | |
210 | ||
211 | if (isalarm(ctx)) | |
212 | alarm_cancel(&ctx->t.alarm); | |
213 | else | |
214 | hrtimer_cancel(&ctx->t.tmr); | |
215 | kfree_rcu(ctx, rcu); | |
216 | return 0; | |
217 | } | |
218 | ||
219 | static unsigned int timerfd_poll(struct file *file, poll_table *wait) | |
220 | { | |
221 | struct timerfd_ctx *ctx = file->private_data; | |
222 | unsigned int events = 0; | |
223 | unsigned long flags; | |
224 | ||
225 | poll_wait(file, &ctx->wqh, wait); | |
226 | ||
227 | spin_lock_irqsave(&ctx->wqh.lock, flags); | |
228 | if (ctx->ticks) | |
229 | events |= POLLIN; | |
230 | spin_unlock_irqrestore(&ctx->wqh.lock, flags); | |
231 | ||
232 | return events; | |
233 | } | |
234 | ||
235 | static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count, | |
236 | loff_t *ppos) | |
237 | { | |
238 | struct timerfd_ctx *ctx = file->private_data; | |
239 | ssize_t res; | |
240 | u64 ticks = 0; | |
241 | ||
242 | if (count < sizeof(ticks)) | |
243 | return -EINVAL; | |
244 | spin_lock_irq(&ctx->wqh.lock); | |
245 | if (file->f_flags & O_NONBLOCK) | |
246 | res = -EAGAIN; | |
247 | else | |
248 | res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks); | |
249 | ||
250 | /* | |
251 | * If clock has changed, we do not care about the | |
252 | * ticks and we do not rearm the timer. Userspace must | |
253 | * reevaluate anyway. | |
254 | */ | |
255 | if (timerfd_canceled(ctx)) { | |
256 | ctx->ticks = 0; | |
257 | ctx->expired = 0; | |
258 | res = -ECANCELED; | |
259 | } | |
260 | ||
261 | if (ctx->ticks) { | |
262 | ticks = ctx->ticks; | |
263 | ||
264 | if (ctx->expired && ctx->tintv.tv64) { | |
265 | /* | |
266 | * If tintv.tv64 != 0, this is a periodic timer that | |
267 | * needs to be re-armed. We avoid doing it in the timer | |
268 | * callback to avoid DoS attacks specifying a very | |
269 | * short timer period. | |
270 | */ | |
271 | if (isalarm(ctx)) { | |
272 | ticks += alarm_forward_now( | |
273 | &ctx->t.alarm, ctx->tintv) - 1; | |
274 | alarm_restart(&ctx->t.alarm); | |
275 | } else { | |
276 | ticks += hrtimer_forward_now(&ctx->t.tmr, | |
277 | ctx->tintv) - 1; | |
278 | hrtimer_restart(&ctx->t.tmr); | |
279 | } | |
280 | } | |
281 | ctx->expired = 0; | |
282 | ctx->ticks = 0; | |
283 | } | |
284 | spin_unlock_irq(&ctx->wqh.lock); | |
285 | if (ticks) | |
286 | res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks); | |
287 | return res; | |
288 | } | |
289 | ||
290 | #ifdef CONFIG_PROC_FS | |
291 | static void timerfd_show(struct seq_file *m, struct file *file) | |
292 | { | |
293 | struct timerfd_ctx *ctx = file->private_data; | |
294 | struct itimerspec t; | |
295 | ||
296 | spin_lock_irq(&ctx->wqh.lock); | |
297 | t.it_value = ktime_to_timespec(timerfd_get_remaining(ctx)); | |
298 | t.it_interval = ktime_to_timespec(ctx->tintv); | |
299 | spin_unlock_irq(&ctx->wqh.lock); | |
300 | ||
301 | seq_printf(m, | |
302 | "clockid: %d\n" | |
303 | "ticks: %llu\n" | |
304 | "settime flags: 0%o\n" | |
305 | "it_value: (%llu, %llu)\n" | |
306 | "it_interval: (%llu, %llu)\n", | |
307 | ctx->clockid, | |
308 | (unsigned long long)ctx->ticks, | |
309 | ctx->settime_flags, | |
310 | (unsigned long long)t.it_value.tv_sec, | |
311 | (unsigned long long)t.it_value.tv_nsec, | |
312 | (unsigned long long)t.it_interval.tv_sec, | |
313 | (unsigned long long)t.it_interval.tv_nsec); | |
314 | } | |
315 | #else | |
316 | #define timerfd_show NULL | |
317 | #endif | |
318 | ||
319 | #ifdef CONFIG_CHECKPOINT_RESTORE | |
320 | static long timerfd_ioctl(struct file *file, unsigned int cmd, unsigned long arg) | |
321 | { | |
322 | struct timerfd_ctx *ctx = file->private_data; | |
323 | int ret = 0; | |
324 | ||
325 | switch (cmd) { | |
326 | case TFD_IOC_SET_TICKS: { | |
327 | u64 ticks; | |
328 | ||
329 | if (copy_from_user(&ticks, (u64 __user *)arg, sizeof(ticks))) | |
330 | return -EFAULT; | |
331 | if (!ticks) | |
332 | return -EINVAL; | |
333 | ||
334 | spin_lock_irq(&ctx->wqh.lock); | |
335 | if (!timerfd_canceled(ctx)) { | |
336 | ctx->ticks = ticks; | |
337 | wake_up_locked(&ctx->wqh); | |
338 | } else | |
339 | ret = -ECANCELED; | |
340 | spin_unlock_irq(&ctx->wqh.lock); | |
341 | break; | |
342 | } | |
343 | default: | |
344 | ret = -ENOTTY; | |
345 | break; | |
346 | } | |
347 | ||
348 | return ret; | |
349 | } | |
350 | #else | |
351 | #define timerfd_ioctl NULL | |
352 | #endif | |
353 | ||
354 | static const struct file_operations timerfd_fops = { | |
355 | .release = timerfd_release, | |
356 | .poll = timerfd_poll, | |
357 | .read = timerfd_read, | |
358 | .llseek = noop_llseek, | |
359 | .show_fdinfo = timerfd_show, | |
360 | .unlocked_ioctl = timerfd_ioctl, | |
361 | }; | |
362 | ||
363 | static int timerfd_fget(int fd, struct fd *p) | |
364 | { | |
365 | struct fd f = fdget(fd); | |
366 | if (!f.file) | |
367 | return -EBADF; | |
368 | if (f.file->f_op != &timerfd_fops) { | |
369 | fdput(f); | |
370 | return -EINVAL; | |
371 | } | |
372 | *p = f; | |
373 | return 0; | |
374 | } | |
375 | ||
376 | SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags) | |
377 | { | |
378 | int ufd; | |
379 | struct timerfd_ctx *ctx; | |
380 | ||
381 | /* Check the TFD_* constants for consistency. */ | |
382 | BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC); | |
383 | BUILD_BUG_ON(TFD_NONBLOCK != O_NONBLOCK); | |
384 | ||
385 | if ((flags & ~TFD_CREATE_FLAGS) || | |
386 | (clockid != CLOCK_MONOTONIC && | |
387 | clockid != CLOCK_REALTIME && | |
388 | clockid != CLOCK_REALTIME_ALARM && | |
389 | clockid != CLOCK_BOOTTIME && | |
390 | clockid != CLOCK_BOOTTIME_ALARM)) | |
391 | return -EINVAL; | |
392 | ||
393 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
394 | if (!ctx) | |
395 | return -ENOMEM; | |
396 | ||
397 | init_waitqueue_head(&ctx->wqh); | |
398 | ctx->clockid = clockid; | |
399 | ||
400 | if (isalarm(ctx)) | |
401 | alarm_init(&ctx->t.alarm, | |
402 | ctx->clockid == CLOCK_REALTIME_ALARM ? | |
403 | ALARM_REALTIME : ALARM_BOOTTIME, | |
404 | timerfd_alarmproc); | |
405 | else | |
406 | hrtimer_init(&ctx->t.tmr, clockid, HRTIMER_MODE_ABS); | |
407 | ||
408 | ctx->moffs = ktime_mono_to_real((ktime_t){ .tv64 = 0 }); | |
409 | ||
410 | ufd = anon_inode_getfd("[timerfd]", &timerfd_fops, ctx, | |
411 | O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS)); | |
412 | if (ufd < 0) | |
413 | kfree(ctx); | |
414 | ||
415 | return ufd; | |
416 | } | |
417 | ||
418 | static int do_timerfd_settime(int ufd, int flags, | |
419 | const struct itimerspec *new, | |
420 | struct itimerspec *old) | |
421 | { | |
422 | struct fd f; | |
423 | struct timerfd_ctx *ctx; | |
424 | int ret; | |
425 | ||
426 | if ((flags & ~TFD_SETTIME_FLAGS) || | |
427 | !timespec_valid(&new->it_value) || | |
428 | !timespec_valid(&new->it_interval)) | |
429 | return -EINVAL; | |
430 | ||
431 | ret = timerfd_fget(ufd, &f); | |
432 | if (ret) | |
433 | return ret; | |
434 | ctx = f.file->private_data; | |
435 | ||
436 | timerfd_setup_cancel(ctx, flags); | |
437 | ||
438 | /* | |
439 | * We need to stop the existing timer before reprogramming | |
440 | * it to the new values. | |
441 | */ | |
442 | for (;;) { | |
443 | spin_lock_irq(&ctx->wqh.lock); | |
444 | ||
445 | if (isalarm(ctx)) { | |
446 | if (alarm_try_to_cancel(&ctx->t.alarm) >= 0) | |
447 | break; | |
448 | } else { | |
449 | if (hrtimer_try_to_cancel(&ctx->t.tmr) >= 0) | |
450 | break; | |
451 | } | |
452 | spin_unlock_irq(&ctx->wqh.lock); | |
453 | cpu_relax(); | |
454 | } | |
455 | ||
456 | /* | |
457 | * If the timer is expired and it's periodic, we need to advance it | |
458 | * because the caller may want to know the previous expiration time. | |
459 | * We do not update "ticks" and "expired" since the timer will be | |
460 | * re-programmed again in the following timerfd_setup() call. | |
461 | */ | |
462 | if (ctx->expired && ctx->tintv.tv64) { | |
463 | if (isalarm(ctx)) | |
464 | alarm_forward_now(&ctx->t.alarm, ctx->tintv); | |
465 | else | |
466 | hrtimer_forward_now(&ctx->t.tmr, ctx->tintv); | |
467 | } | |
468 | ||
469 | old->it_value = ktime_to_timespec(timerfd_get_remaining(ctx)); | |
470 | old->it_interval = ktime_to_timespec(ctx->tintv); | |
471 | ||
472 | /* | |
473 | * Re-program the timer to the new value ... | |
474 | */ | |
475 | ret = timerfd_setup(ctx, flags, new); | |
476 | ||
477 | spin_unlock_irq(&ctx->wqh.lock); | |
478 | fdput(f); | |
479 | return ret; | |
480 | } | |
481 | ||
482 | static int do_timerfd_gettime(int ufd, struct itimerspec *t) | |
483 | { | |
484 | struct fd f; | |
485 | struct timerfd_ctx *ctx; | |
486 | int ret = timerfd_fget(ufd, &f); | |
487 | if (ret) | |
488 | return ret; | |
489 | ctx = f.file->private_data; | |
490 | ||
491 | spin_lock_irq(&ctx->wqh.lock); | |
492 | if (ctx->expired && ctx->tintv.tv64) { | |
493 | ctx->expired = 0; | |
494 | ||
495 | if (isalarm(ctx)) { | |
496 | ctx->ticks += | |
497 | alarm_forward_now( | |
498 | &ctx->t.alarm, ctx->tintv) - 1; | |
499 | alarm_restart(&ctx->t.alarm); | |
500 | } else { | |
501 | ctx->ticks += | |
502 | hrtimer_forward_now(&ctx->t.tmr, ctx->tintv) | |
503 | - 1; | |
504 | hrtimer_restart(&ctx->t.tmr); | |
505 | } | |
506 | } | |
507 | t->it_value = ktime_to_timespec(timerfd_get_remaining(ctx)); | |
508 | t->it_interval = ktime_to_timespec(ctx->tintv); | |
509 | spin_unlock_irq(&ctx->wqh.lock); | |
510 | fdput(f); | |
511 | return 0; | |
512 | } | |
513 | ||
514 | SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags, | |
515 | const struct itimerspec __user *, utmr, | |
516 | struct itimerspec __user *, otmr) | |
517 | { | |
518 | struct itimerspec new, old; | |
519 | int ret; | |
520 | ||
521 | if (copy_from_user(&new, utmr, sizeof(new))) | |
522 | return -EFAULT; | |
523 | ret = do_timerfd_settime(ufd, flags, &new, &old); | |
524 | if (ret) | |
525 | return ret; | |
526 | if (otmr && copy_to_user(otmr, &old, sizeof(old))) | |
527 | return -EFAULT; | |
528 | ||
529 | return ret; | |
530 | } | |
531 | ||
532 | SYSCALL_DEFINE2(timerfd_gettime, int, ufd, struct itimerspec __user *, otmr) | |
533 | { | |
534 | struct itimerspec kotmr; | |
535 | int ret = do_timerfd_gettime(ufd, &kotmr); | |
536 | if (ret) | |
537 | return ret; | |
538 | return copy_to_user(otmr, &kotmr, sizeof(kotmr)) ? -EFAULT: 0; | |
539 | } | |
540 | ||
541 | #ifdef CONFIG_COMPAT | |
542 | COMPAT_SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags, | |
543 | const struct compat_itimerspec __user *, utmr, | |
544 | struct compat_itimerspec __user *, otmr) | |
545 | { | |
546 | struct itimerspec new, old; | |
547 | int ret; | |
548 | ||
549 | if (get_compat_itimerspec(&new, utmr)) | |
550 | return -EFAULT; | |
551 | ret = do_timerfd_settime(ufd, flags, &new, &old); | |
552 | if (ret) | |
553 | return ret; | |
554 | if (otmr && put_compat_itimerspec(otmr, &old)) | |
555 | return -EFAULT; | |
556 | return ret; | |
557 | } | |
558 | ||
559 | COMPAT_SYSCALL_DEFINE2(timerfd_gettime, int, ufd, | |
560 | struct compat_itimerspec __user *, otmr) | |
561 | { | |
562 | struct itimerspec kotmr; | |
563 | int ret = do_timerfd_gettime(ufd, &kotmr); | |
564 | if (ret) | |
565 | return ret; | |
566 | return put_compat_itimerspec(otmr, &kotmr) ? -EFAULT: 0; | |
567 | } | |
568 | #endif |