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