mm: introduce single zone pcplists drain
[deliverable/linux.git] / 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/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 int 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 return 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, (unsigned long long)ctx->ticks,
308 ctx->settime_flags,
309 (unsigned long long)t.it_value.tv_sec,
310 (unsigned long long)t.it_value.tv_nsec,
311 (unsigned long long)t.it_interval.tv_sec,
312 (unsigned long long)t.it_interval.tv_nsec);
313 }
314 #else
315 #define timerfd_show NULL
316 #endif
317
318 #ifdef CONFIG_CHECKPOINT_RESTORE
319 static long timerfd_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
320 {
321 struct timerfd_ctx *ctx = file->private_data;
322 int ret = 0;
323
324 switch (cmd) {
325 case TFD_IOC_SET_TICKS: {
326 u64 ticks;
327
328 if (copy_from_user(&ticks, (u64 __user *)arg, sizeof(ticks)))
329 return -EFAULT;
330 if (!ticks)
331 return -EINVAL;
332
333 spin_lock_irq(&ctx->wqh.lock);
334 if (!timerfd_canceled(ctx)) {
335 ctx->ticks = ticks;
336 wake_up_locked(&ctx->wqh);
337 } else
338 ret = -ECANCELED;
339 spin_unlock_irq(&ctx->wqh.lock);
340 break;
341 }
342 default:
343 ret = -ENOTTY;
344 break;
345 }
346
347 return ret;
348 }
349 #else
350 #define timerfd_ioctl NULL
351 #endif
352
353 static const struct file_operations timerfd_fops = {
354 .release = timerfd_release,
355 .poll = timerfd_poll,
356 .read = timerfd_read,
357 .llseek = noop_llseek,
358 .show_fdinfo = timerfd_show,
359 .unlocked_ioctl = timerfd_ioctl,
360 };
361
362 static int timerfd_fget(int fd, struct fd *p)
363 {
364 struct fd f = fdget(fd);
365 if (!f.file)
366 return -EBADF;
367 if (f.file->f_op != &timerfd_fops) {
368 fdput(f);
369 return -EINVAL;
370 }
371 *p = f;
372 return 0;
373 }
374
375 SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
376 {
377 int ufd;
378 struct timerfd_ctx *ctx;
379
380 /* Check the TFD_* constants for consistency. */
381 BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC);
382 BUILD_BUG_ON(TFD_NONBLOCK != O_NONBLOCK);
383
384 if ((flags & ~TFD_CREATE_FLAGS) ||
385 (clockid != CLOCK_MONOTONIC &&
386 clockid != CLOCK_REALTIME &&
387 clockid != CLOCK_REALTIME_ALARM &&
388 clockid != CLOCK_BOOTTIME &&
389 clockid != CLOCK_BOOTTIME_ALARM))
390 return -EINVAL;
391
392 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
393 if (!ctx)
394 return -ENOMEM;
395
396 init_waitqueue_head(&ctx->wqh);
397 ctx->clockid = clockid;
398
399 if (isalarm(ctx))
400 alarm_init(&ctx->t.alarm,
401 ctx->clockid == CLOCK_REALTIME_ALARM ?
402 ALARM_REALTIME : ALARM_BOOTTIME,
403 timerfd_alarmproc);
404 else
405 hrtimer_init(&ctx->t.tmr, clockid, HRTIMER_MODE_ABS);
406
407 ctx->moffs = ktime_mono_to_real((ktime_t){ .tv64 = 0 });
408
409 ufd = anon_inode_getfd("[timerfd]", &timerfd_fops, ctx,
410 O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
411 if (ufd < 0)
412 kfree(ctx);
413
414 return ufd;
415 }
416
417 static int do_timerfd_settime(int ufd, int flags,
418 const struct itimerspec *new,
419 struct itimerspec *old)
420 {
421 struct fd f;
422 struct timerfd_ctx *ctx;
423 int ret;
424
425 if ((flags & ~TFD_SETTIME_FLAGS) ||
426 !timespec_valid(&new->it_value) ||
427 !timespec_valid(&new->it_interval))
428 return -EINVAL;
429
430 ret = timerfd_fget(ufd, &f);
431 if (ret)
432 return ret;
433 ctx = f.file->private_data;
434
435 timerfd_setup_cancel(ctx, flags);
436
437 /*
438 * We need to stop the existing timer before reprogramming
439 * it to the new values.
440 */
441 for (;;) {
442 spin_lock_irq(&ctx->wqh.lock);
443
444 if (isalarm(ctx)) {
445 if (alarm_try_to_cancel(&ctx->t.alarm) >= 0)
446 break;
447 } else {
448 if (hrtimer_try_to_cancel(&ctx->t.tmr) >= 0)
449 break;
450 }
451 spin_unlock_irq(&ctx->wqh.lock);
452 cpu_relax();
453 }
454
455 /*
456 * If the timer is expired and it's periodic, we need to advance it
457 * because the caller may want to know the previous expiration time.
458 * We do not update "ticks" and "expired" since the timer will be
459 * re-programmed again in the following timerfd_setup() call.
460 */
461 if (ctx->expired && ctx->tintv.tv64) {
462 if (isalarm(ctx))
463 alarm_forward_now(&ctx->t.alarm, ctx->tintv);
464 else
465 hrtimer_forward_now(&ctx->t.tmr, ctx->tintv);
466 }
467
468 old->it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
469 old->it_interval = ktime_to_timespec(ctx->tintv);
470
471 /*
472 * Re-program the timer to the new value ...
473 */
474 ret = timerfd_setup(ctx, flags, new);
475
476 spin_unlock_irq(&ctx->wqh.lock);
477 fdput(f);
478 return ret;
479 }
480
481 static int do_timerfd_gettime(int ufd, struct itimerspec *t)
482 {
483 struct fd f;
484 struct timerfd_ctx *ctx;
485 int ret = timerfd_fget(ufd, &f);
486 if (ret)
487 return ret;
488 ctx = f.file->private_data;
489
490 spin_lock_irq(&ctx->wqh.lock);
491 if (ctx->expired && ctx->tintv.tv64) {
492 ctx->expired = 0;
493
494 if (isalarm(ctx)) {
495 ctx->ticks +=
496 alarm_forward_now(
497 &ctx->t.alarm, ctx->tintv) - 1;
498 alarm_restart(&ctx->t.alarm);
499 } else {
500 ctx->ticks +=
501 hrtimer_forward_now(&ctx->t.tmr, ctx->tintv)
502 - 1;
503 hrtimer_restart(&ctx->t.tmr);
504 }
505 }
506 t->it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
507 t->it_interval = ktime_to_timespec(ctx->tintv);
508 spin_unlock_irq(&ctx->wqh.lock);
509 fdput(f);
510 return 0;
511 }
512
513 SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
514 const struct itimerspec __user *, utmr,
515 struct itimerspec __user *, otmr)
516 {
517 struct itimerspec new, old;
518 int ret;
519
520 if (copy_from_user(&new, utmr, sizeof(new)))
521 return -EFAULT;
522 ret = do_timerfd_settime(ufd, flags, &new, &old);
523 if (ret)
524 return ret;
525 if (otmr && copy_to_user(otmr, &old, sizeof(old)))
526 return -EFAULT;
527
528 return ret;
529 }
530
531 SYSCALL_DEFINE2(timerfd_gettime, int, ufd, struct itimerspec __user *, otmr)
532 {
533 struct itimerspec kotmr;
534 int ret = do_timerfd_gettime(ufd, &kotmr);
535 if (ret)
536 return ret;
537 return copy_to_user(otmr, &kotmr, sizeof(kotmr)) ? -EFAULT: 0;
538 }
539
540 #ifdef CONFIG_COMPAT
541 COMPAT_SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
542 const struct compat_itimerspec __user *, utmr,
543 struct compat_itimerspec __user *, otmr)
544 {
545 struct itimerspec new, old;
546 int ret;
547
548 if (get_compat_itimerspec(&new, utmr))
549 return -EFAULT;
550 ret = do_timerfd_settime(ufd, flags, &new, &old);
551 if (ret)
552 return ret;
553 if (otmr && put_compat_itimerspec(otmr, &old))
554 return -EFAULT;
555 return ret;
556 }
557
558 COMPAT_SYSCALL_DEFINE2(timerfd_gettime, int, ufd,
559 struct compat_itimerspec __user *, otmr)
560 {
561 struct itimerspec kotmr;
562 int ret = do_timerfd_gettime(ufd, &kotmr);
563 if (ret)
564 return ret;
565 return put_compat_itimerspec(otmr, &kotmr) ? -EFAULT: 0;
566 }
567 #endif
This page took 0.042825 seconds and 5 git commands to generate.