projects/linux·blob v6.12 · refs

628 lines · 14.3 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * fs/timerfd.c
4 *
5 * Copyright (C) 2007 Davide Libenzi <[email protected]>
6 *
7 *
8 * Thanks to Thomas Gleixner for code reviews and useful comments.
9 *
10 */
11
12#include <linux/alarmtimer.h>
13#include <linux/file.h>
14#include <linux/poll.h>
15#include <linux/init.h>
16#include <linux/fs.h>
17#include <linux/sched.h>
18#include <linux/kernel.h>
19#include <linux/slab.h>
20#include <linux/list.h>
21#include <linux/spinlock.h>
22#include <linux/time.h>
23#include <linux/hrtimer.h>
24#include <linux/anon_inodes.h>
25#include <linux/timerfd.h>
26#include <linux/syscalls.h>
27#include <linux/compat.h>
28#include <linux/rcupdate.h>
29#include <linux/time_namespace.h>
30
31struct timerfd_ctx {
32 union {
33 struct hrtimer tmr;
34 struct alarm alarm;
35 } t;
36 ktime_t tintv;
37 ktime_t moffs;
38 wait_queue_head_t wqh;
39 u64 ticks;
40 int clockid;
41 short unsigned expired;
42 short unsigned settime_flags; /* to show in fdinfo */
43 struct rcu_head rcu;
44 struct list_head clist;
45 spinlock_t cancel_lock;
46 bool might_cancel;
47};
48
49static LIST_HEAD(cancel_list);
50static DEFINE_SPINLOCK(cancel_lock);
51
52static inline bool isalarm(struct timerfd_ctx *ctx)
53{
54 return ctx->clockid == CLOCK_REALTIME_ALARM ||
55 ctx->clockid == CLOCK_BOOTTIME_ALARM;
56}
57
58/*
59 * This gets called when the timer event triggers. We set the "expired"
60 * flag, but we do not re-arm the timer (in case it's necessary,
61 * tintv != 0) until the timer is accessed.
62 */
63static void timerfd_triggered(struct timerfd_ctx *ctx)
64{
65 unsigned long flags;
66
67 spin_lock_irqsave(&ctx->wqh.lock, flags);
68 ctx->expired = 1;
69 ctx->ticks++;
70 wake_up_locked_poll(&ctx->wqh, EPOLLIN);
71 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
72}
73
74static enum hrtimer_restart timerfd_tmrproc(struct hrtimer *htmr)
75{
76 struct timerfd_ctx *ctx = container_of(htmr, struct timerfd_ctx,
77 t.tmr);
78 timerfd_triggered(ctx);
79 return HRTIMER_NORESTART;
80}
81
82static enum alarmtimer_restart timerfd_alarmproc(struct alarm *alarm,
83 ktime_t now)
84{
85 struct timerfd_ctx *ctx = container_of(alarm, struct timerfd_ctx,
86 t.alarm);
87 timerfd_triggered(ctx);
88 return ALARMTIMER_NORESTART;
89}
90
91/*
92 * Called when the clock was set to cancel the timers in the cancel
93 * list. This will wake up processes waiting on these timers. The
94 * wake-up requires ctx->ticks to be non zero, therefore we increment
95 * it before calling wake_up_locked().
96 */
97void timerfd_clock_was_set(void)
98{
99 ktime_t moffs = ktime_mono_to_real(0);
100 struct timerfd_ctx *ctx;
101 unsigned long flags;
102
103 rcu_read_lock();
104 list_for_each_entry_rcu(ctx, &cancel_list, clist) {
105 if (!ctx->might_cancel)
106 continue;
107 spin_lock_irqsave(&ctx->wqh.lock, flags);
108 if (ctx->moffs != moffs) {
109 ctx->moffs = KTIME_MAX;
110 ctx->ticks++;
111 wake_up_locked_poll(&ctx->wqh, EPOLLIN);
112 }
113 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
114 }
115 rcu_read_unlock();
116}
117
118static void timerfd_resume_work(struct work_struct *work)
119{
120 timerfd_clock_was_set();
121}
122
123static DECLARE_WORK(timerfd_work, timerfd_resume_work);
124
125/*
126 * Invoked from timekeeping_resume(). Defer the actual update to work so
127 * timerfd_clock_was_set() runs in task context.
128 */
129void timerfd_resume(void)
130{
131 schedule_work(&timerfd_work);
132}
133
134static void __timerfd_remove_cancel(struct timerfd_ctx *ctx)
135{
136 if (ctx->might_cancel) {
137 ctx->might_cancel = false;
138 spin_lock(&cancel_lock);
139 list_del_rcu(&ctx->clist);
140 spin_unlock(&cancel_lock);
141 }
142}
143
144static void timerfd_remove_cancel(struct timerfd_ctx *ctx)
145{
146 spin_lock(&ctx->cancel_lock);
147 __timerfd_remove_cancel(ctx);
148 spin_unlock(&ctx->cancel_lock);
149}
150
151static bool timerfd_canceled(struct timerfd_ctx *ctx)
152{
153 if (!ctx->might_cancel || ctx->moffs != KTIME_MAX)
154 return false;
155 ctx->moffs = ktime_mono_to_real(0);
156 return true;
157}
158
159static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags)
160{
161 spin_lock(&ctx->cancel_lock);
162 if ((ctx->clockid == CLOCK_REALTIME ||
163 ctx->clockid == CLOCK_REALTIME_ALARM) &&
164 (flags & TFD_TIMER_ABSTIME) && (flags & TFD_TIMER_CANCEL_ON_SET)) {
165 if (!ctx->might_cancel) {
166 ctx->might_cancel = true;
167 spin_lock(&cancel_lock);
168 list_add_rcu(&ctx->clist, &cancel_list);
169 spin_unlock(&cancel_lock);
170 }
171 } else {
172 __timerfd_remove_cancel(ctx);
173 }
174 spin_unlock(&ctx->cancel_lock);
175}
176
177static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx)
178{
179 ktime_t remaining;
180
181 if (isalarm(ctx))
182 remaining = alarm_expires_remaining(&ctx->t.alarm);
183 else
184 remaining = hrtimer_expires_remaining_adjusted(&ctx->t.tmr);
185
186 return remaining < 0 ? 0: remaining;
187}
188
189static int timerfd_setup(struct timerfd_ctx *ctx, int flags,
190 const struct itimerspec64 *ktmr)
191{
192 enum hrtimer_mode htmode;
193 ktime_t texp;
194 int clockid = ctx->clockid;
195
196 htmode = (flags & TFD_TIMER_ABSTIME) ?
197 HRTIMER_MODE_ABS: HRTIMER_MODE_REL;
198
199 texp = timespec64_to_ktime(ktmr->it_value);
200 ctx->expired = 0;
201 ctx->ticks = 0;
202 ctx->tintv = timespec64_to_ktime(ktmr->it_interval);
203
204 if (isalarm(ctx)) {
205 alarm_init(&ctx->t.alarm,
206 ctx->clockid == CLOCK_REALTIME_ALARM ?
207 ALARM_REALTIME : ALARM_BOOTTIME,
208 timerfd_alarmproc);
209 } else {
210 hrtimer_init(&ctx->t.tmr, clockid, htmode);
211 hrtimer_set_expires(&ctx->t.tmr, texp);
212 ctx->t.tmr.function = timerfd_tmrproc;
213 }
214
215 if (texp != 0) {
216 if (flags & TFD_TIMER_ABSTIME)
217 texp = timens_ktime_to_host(clockid, texp);
218 if (isalarm(ctx)) {
219 if (flags & TFD_TIMER_ABSTIME)
220 alarm_start(&ctx->t.alarm, texp);
221 else
222 alarm_start_relative(&ctx->t.alarm, texp);
223 } else {
224 hrtimer_start(&ctx->t.tmr, texp, htmode);
225 }
226
227 if (timerfd_canceled(ctx))
228 return -ECANCELED;
229 }
230
231 ctx->settime_flags = flags & TFD_SETTIME_FLAGS;
232 return 0;
233}
234
235static int timerfd_release(struct inode *inode, struct file *file)
236{
237 struct timerfd_ctx *ctx = file->private_data;
238
239 timerfd_remove_cancel(ctx);
240
241 if (isalarm(ctx))
242 alarm_cancel(&ctx->t.alarm);
243 else
244 hrtimer_cancel(&ctx->t.tmr);
245 kfree_rcu(ctx, rcu);
246 return 0;
247}
248
249static __poll_t timerfd_poll(struct file *file, poll_table *wait)
250{
251 struct timerfd_ctx *ctx = file->private_data;
252 __poll_t events = 0;
253 unsigned long flags;
254
255 poll_wait(file, &ctx->wqh, wait);
256
257 spin_lock_irqsave(&ctx->wqh.lock, flags);
258 if (ctx->ticks)
259 events |= EPOLLIN;
260 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
261
262 return events;
263}
264
265static ssize_t timerfd_read_iter(struct kiocb *iocb, struct iov_iter *to)
266{
267 struct file *file = iocb->ki_filp;
268 struct timerfd_ctx *ctx = file->private_data;
269 ssize_t res;
270 u64 ticks = 0;
271
272 if (iov_iter_count(to) < sizeof(ticks))
273 return -EINVAL;
274
275 spin_lock_irq(&ctx->wqh.lock);
276 if (file->f_flags & O_NONBLOCK || iocb->ki_flags & IOCB_NOWAIT)
277 res = -EAGAIN;
278 else
279 res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks);
280
281 /*
282 * If clock has changed, we do not care about the
283 * ticks and we do not rearm the timer. Userspace must
284 * reevaluate anyway.
285 */
286 if (timerfd_canceled(ctx)) {
287 ctx->ticks = 0;
288 ctx->expired = 0;
289 res = -ECANCELED;
290 }
291
292 if (ctx->ticks) {
293 ticks = ctx->ticks;
294
295 if (ctx->expired && ctx->tintv) {
296 /*
297 * If tintv != 0, this is a periodic timer that
298 * needs to be re-armed. We avoid doing it in the timer
299 * callback to avoid DoS attacks specifying a very
300 * short timer period.
301 */
302 if (isalarm(ctx)) {
303 ticks += alarm_forward_now(
304 &ctx->t.alarm, ctx->tintv) - 1;
305 alarm_restart(&ctx->t.alarm);
306 } else {
307 ticks += hrtimer_forward_now(&ctx->t.tmr,
308 ctx->tintv) - 1;
309 hrtimer_restart(&ctx->t.tmr);
310 }
311 }
312 ctx->expired = 0;
313 ctx->ticks = 0;
314 }
315 spin_unlock_irq(&ctx->wqh.lock);
316 if (ticks) {
317 res = copy_to_iter(&ticks, sizeof(ticks), to);
318 if (!res)
319 res = -EFAULT;
320 }
321 return res;
322}
323
324#ifdef CONFIG_PROC_FS
325static void timerfd_show(struct seq_file *m, struct file *file)
326{
327 struct timerfd_ctx *ctx = file->private_data;
328 struct timespec64 value, interval;
329
330 spin_lock_irq(&ctx->wqh.lock);
331 value = ktime_to_timespec64(timerfd_get_remaining(ctx));
332 interval = ktime_to_timespec64(ctx->tintv);
333 spin_unlock_irq(&ctx->wqh.lock);
334
335 seq_printf(m,
336 "clockid: %d\n"
337 "ticks: %llu\n"
338 "settime flags: 0%o\n"
339 "it_value: (%llu, %llu)\n"
340 "it_interval: (%llu, %llu)\n",
341 ctx->clockid,
342 (unsigned long long)ctx->ticks,
343 ctx->settime_flags,
344 (unsigned long long)value.tv_sec,
345 (unsigned long long)value.tv_nsec,
346 (unsigned long long)interval.tv_sec,
347 (unsigned long long)interval.tv_nsec);
348}
349#else
350#define timerfd_show NULL
351#endif
352
353#ifdef CONFIG_CHECKPOINT_RESTORE
354static long timerfd_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
355{
356 struct timerfd_ctx *ctx = file->private_data;
357 int ret = 0;
358
359 switch (cmd) {
360 case TFD_IOC_SET_TICKS: {
361 u64 ticks;
362
363 if (copy_from_user(&ticks, (u64 __user *)arg, sizeof(ticks)))
364 return -EFAULT;
365 if (!ticks)
366 return -EINVAL;
367
368 spin_lock_irq(&ctx->wqh.lock);
369 if (!timerfd_canceled(ctx)) {
370 ctx->ticks = ticks;
371 wake_up_locked_poll(&ctx->wqh, EPOLLIN);
372 } else
373 ret = -ECANCELED;
374 spin_unlock_irq(&ctx->wqh.lock);
375 break;
376 }
377 default:
378 ret = -ENOTTY;
379 break;
380 }
381
382 return ret;
383}
384#else
385#define timerfd_ioctl NULL
386#endif
387
388static const struct file_operations timerfd_fops = {
389 .release = timerfd_release,
390 .poll = timerfd_poll,
391 .read_iter = timerfd_read_iter,
392 .llseek = noop_llseek,
393 .show_fdinfo = timerfd_show,
394 .unlocked_ioctl = timerfd_ioctl,
395};
396
397static int timerfd_fget(int fd, struct fd *p)
398{
399 struct fd f = fdget(fd);
400 if (!fd_file(f))
401 return -EBADF;
402 if (fd_file(f)->f_op != &timerfd_fops) {
403 fdput(f);
404 return -EINVAL;
405 }
406 *p = f;
407 return 0;
408}
409
410SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
411{
412 int ufd;
413 struct timerfd_ctx *ctx;
414 struct file *file;
415
416 /* Check the TFD_* constants for consistency. */
417 BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC);
418 BUILD_BUG_ON(TFD_NONBLOCK != O_NONBLOCK);
419
420 if ((flags & ~TFD_CREATE_FLAGS) ||
421 (clockid != CLOCK_MONOTONIC &&
422 clockid != CLOCK_REALTIME &&
423 clockid != CLOCK_REALTIME_ALARM &&
424 clockid != CLOCK_BOOTTIME &&
425 clockid != CLOCK_BOOTTIME_ALARM))
426 return -EINVAL;
427
428 if ((clockid == CLOCK_REALTIME_ALARM ||
429 clockid == CLOCK_BOOTTIME_ALARM) &&
430 !capable(CAP_WAKE_ALARM))
431 return -EPERM;
432
433 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
434 if (!ctx)
435 return -ENOMEM;
436
437 init_waitqueue_head(&ctx->wqh);
438 spin_lock_init(&ctx->cancel_lock);
439 ctx->clockid = clockid;
440
441 if (isalarm(ctx))
442 alarm_init(&ctx->t.alarm,
443 ctx->clockid == CLOCK_REALTIME_ALARM ?
444 ALARM_REALTIME : ALARM_BOOTTIME,
445 timerfd_alarmproc);
446 else
447 hrtimer_init(&ctx->t.tmr, clockid, HRTIMER_MODE_ABS);
448
449 ctx->moffs = ktime_mono_to_real(0);
450
451 ufd = get_unused_fd_flags(flags & TFD_SHARED_FCNTL_FLAGS);
452 if (ufd < 0) {
453 kfree(ctx);
454 return ufd;
455 }
456
457 file = anon_inode_getfile("[timerfd]", &timerfd_fops, ctx,
458 O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
459 if (IS_ERR(file)) {
460 put_unused_fd(ufd);
461 kfree(ctx);
462 return PTR_ERR(file);
463 }
464
465 file->f_mode |= FMODE_NOWAIT;
466 fd_install(ufd, file);
467 return ufd;
468}
469
470static int do_timerfd_settime(int ufd, int flags,
471 const struct itimerspec64 *new,
472 struct itimerspec64 *old)
473{
474 struct fd f;
475 struct timerfd_ctx *ctx;
476 int ret;
477
478 if ((flags & ~TFD_SETTIME_FLAGS) ||
479 !itimerspec64_valid(new))
480 return -EINVAL;
481
482 ret = timerfd_fget(ufd, &f);
483 if (ret)
484 return ret;
485 ctx = fd_file(f)->private_data;
486
487 if (isalarm(ctx) && !capable(CAP_WAKE_ALARM)) {
488 fdput(f);
489 return -EPERM;
490 }
491
492 timerfd_setup_cancel(ctx, flags);
493
494 /*
495 * We need to stop the existing timer before reprogramming
496 * it to the new values.
497 */
498 for (;;) {
499 spin_lock_irq(&ctx->wqh.lock);
500
501 if (isalarm(ctx)) {
502 if (alarm_try_to_cancel(&ctx->t.alarm) >= 0)
503 break;
504 } else {
505 if (hrtimer_try_to_cancel(&ctx->t.tmr) >= 0)
506 break;
507 }
508 spin_unlock_irq(&ctx->wqh.lock);
509
510 if (isalarm(ctx))
511 hrtimer_cancel_wait_running(&ctx->t.alarm.timer);
512 else
513 hrtimer_cancel_wait_running(&ctx->t.tmr);
514 }
515
516 /*
517 * If the timer is expired and it's periodic, we need to advance it
518 * because the caller may want to know the previous expiration time.
519 * We do not update "ticks" and "expired" since the timer will be
520 * re-programmed again in the following timerfd_setup() call.
521 */
522 if (ctx->expired && ctx->tintv) {
523 if (isalarm(ctx))
524 alarm_forward_now(&ctx->t.alarm, ctx->tintv);
525 else
526 hrtimer_forward_now(&ctx->t.tmr, ctx->tintv);
527 }
528
529 old->it_value = ktime_to_timespec64(timerfd_get_remaining(ctx));
530 old->it_interval = ktime_to_timespec64(ctx->tintv);
531
532 /*
533 * Re-program the timer to the new value ...
534 */
535 ret = timerfd_setup(ctx, flags, new);
536
537 spin_unlock_irq(&ctx->wqh.lock);
538 fdput(f);
539 return ret;
540}
541
542static int do_timerfd_gettime(int ufd, struct itimerspec64 *t)
543{
544 struct fd f;
545 struct timerfd_ctx *ctx;
546 int ret = timerfd_fget(ufd, &f);
547 if (ret)
548 return ret;
549 ctx = fd_file(f)->private_data;
550
551 spin_lock_irq(&ctx->wqh.lock);
552 if (ctx->expired && ctx->tintv) {
553 ctx->expired = 0;
554
555 if (isalarm(ctx)) {
556 ctx->ticks +=
557 alarm_forward_now(
558 &ctx->t.alarm, ctx->tintv) - 1;
559 alarm_restart(&ctx->t.alarm);
560 } else {
561 ctx->ticks +=
562 hrtimer_forward_now(&ctx->t.tmr, ctx->tintv)
563 - 1;
564 hrtimer_restart(&ctx->t.tmr);
565 }
566 }
567 t->it_value = ktime_to_timespec64(timerfd_get_remaining(ctx));
568 t->it_interval = ktime_to_timespec64(ctx->tintv);
569 spin_unlock_irq(&ctx->wqh.lock);
570 fdput(f);
571 return 0;
572}
573
574SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
575 const struct __kernel_itimerspec __user *, utmr,
576 struct __kernel_itimerspec __user *, otmr)
577{
578 struct itimerspec64 new, old;
579 int ret;
580
581 if (get_itimerspec64(&new, utmr))
582 return -EFAULT;
583 ret = do_timerfd_settime(ufd, flags, &new, &old);
584 if (ret)
585 return ret;
586 if (otmr && put_itimerspec64(&old, otmr))
587 return -EFAULT;
588
589 return ret;
590}
591
592SYSCALL_DEFINE2(timerfd_gettime, int, ufd, struct __kernel_itimerspec __user *, otmr)
593{
594 struct itimerspec64 kotmr;
595 int ret = do_timerfd_gettime(ufd, &kotmr);
596 if (ret)
597 return ret;
598 return put_itimerspec64(&kotmr, otmr) ? -EFAULT : 0;
599}
600
601#ifdef CONFIG_COMPAT_32BIT_TIME
602SYSCALL_DEFINE4(timerfd_settime32, int, ufd, int, flags,
603 const struct old_itimerspec32 __user *, utmr,
604 struct old_itimerspec32 __user *, otmr)
605{
606 struct itimerspec64 new, old;
607 int ret;
608
609 if (get_old_itimerspec32(&new, utmr))
610 return -EFAULT;
611 ret = do_timerfd_settime(ufd, flags, &new, &old);
612 if (ret)
613 return ret;
614 if (otmr && put_old_itimerspec32(&old, otmr))
615 return -EFAULT;
616 return ret;
617}
618
619SYSCALL_DEFINE2(timerfd_gettime32, int, ufd,
620 struct old_itimerspec32 __user *, otmr)
621{
622 struct itimerspec64 kotmr;
623 int ret = do_timerfd_gettime(ufd, &kotmr);
624 if (ret)
625 return ret;
626 return put_old_itimerspec32(&kotmr, otmr) ? -EFAULT : 0;
627}
628#endif