]>
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/file.h> | |
12 | #include <linux/poll.h> | |
13 | #include <linux/init.h> | |
14 | #include <linux/fs.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> | |
26 | ||
27 | struct timerfd_ctx { | |
28 | struct hrtimer tmr; | |
29 | ktime_t tintv; | |
30 | ktime_t moffs; | |
31 | wait_queue_head_t wqh; | |
32 | u64 ticks; | |
33 | int expired; | |
34 | int clockid; | |
35 | struct rcu_head rcu; | |
36 | struct list_head clist; | |
37 | bool might_cancel; | |
38 | }; | |
39 | ||
40 | static LIST_HEAD(cancel_list); | |
41 | static DEFINE_SPINLOCK(cancel_lock); | |
42 | ||
43 | /* | |
44 | * This gets called when the timer event triggers. We set the "expired" | |
45 | * flag, but we do not re-arm the timer (in case it's necessary, | |
46 | * tintv.tv64 != 0) until the timer is accessed. | |
47 | */ | |
48 | static enum hrtimer_restart timerfd_tmrproc(struct hrtimer *htmr) | |
49 | { | |
50 | struct timerfd_ctx *ctx = container_of(htmr, struct timerfd_ctx, tmr); | |
51 | unsigned long flags; | |
52 | ||
53 | spin_lock_irqsave(&ctx->wqh.lock, flags); | |
54 | ctx->expired = 1; | |
55 | ctx->ticks++; | |
56 | wake_up_locked(&ctx->wqh); | |
57 | spin_unlock_irqrestore(&ctx->wqh.lock, flags); | |
58 | ||
59 | return HRTIMER_NORESTART; | |
60 | } | |
61 | ||
62 | /* | |
63 | * Called when the clock was set to cancel the timers in the cancel | |
64 | * list. This will wake up processes waiting on these timers. The | |
65 | * wake-up requires ctx->ticks to be non zero, therefore we increment | |
66 | * it before calling wake_up_locked(). | |
67 | */ | |
68 | void timerfd_clock_was_set(void) | |
69 | { | |
70 | ktime_t moffs = ktime_get_monotonic_offset(); | |
71 | struct timerfd_ctx *ctx; | |
72 | unsigned long flags; | |
73 | ||
74 | rcu_read_lock(); | |
75 | list_for_each_entry_rcu(ctx, &cancel_list, clist) { | |
76 | if (!ctx->might_cancel) | |
77 | continue; | |
78 | spin_lock_irqsave(&ctx->wqh.lock, flags); | |
79 | if (ctx->moffs.tv64 != moffs.tv64) { | |
80 | ctx->moffs.tv64 = KTIME_MAX; | |
81 | ctx->ticks++; | |
82 | wake_up_locked(&ctx->wqh); | |
83 | } | |
84 | spin_unlock_irqrestore(&ctx->wqh.lock, flags); | |
85 | } | |
86 | rcu_read_unlock(); | |
87 | } | |
88 | ||
89 | static void timerfd_remove_cancel(struct timerfd_ctx *ctx) | |
90 | { | |
91 | if (ctx->might_cancel) { | |
92 | ctx->might_cancel = false; | |
93 | spin_lock(&cancel_lock); | |
94 | list_del_rcu(&ctx->clist); | |
95 | spin_unlock(&cancel_lock); | |
96 | } | |
97 | } | |
98 | ||
99 | static bool timerfd_canceled(struct timerfd_ctx *ctx) | |
100 | { | |
101 | if (!ctx->might_cancel || ctx->moffs.tv64 != KTIME_MAX) | |
102 | return false; | |
103 | ctx->moffs = ktime_get_monotonic_offset(); | |
104 | return true; | |
105 | } | |
106 | ||
107 | static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags) | |
108 | { | |
109 | if (ctx->clockid == CLOCK_REALTIME && (flags & TFD_TIMER_ABSTIME) && | |
110 | (flags & TFD_TIMER_CANCEL_ON_SET)) { | |
111 | if (!ctx->might_cancel) { | |
112 | ctx->might_cancel = true; | |
113 | spin_lock(&cancel_lock); | |
114 | list_add_rcu(&ctx->clist, &cancel_list); | |
115 | spin_unlock(&cancel_lock); | |
116 | } | |
117 | } else if (ctx->might_cancel) { | |
118 | timerfd_remove_cancel(ctx); | |
119 | } | |
120 | } | |
121 | ||
122 | static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx) | |
123 | { | |
124 | ktime_t remaining; | |
125 | ||
126 | remaining = hrtimer_expires_remaining(&ctx->tmr); | |
127 | return remaining.tv64 < 0 ? ktime_set(0, 0): remaining; | |
128 | } | |
129 | ||
130 | static int timerfd_setup(struct timerfd_ctx *ctx, int flags, | |
131 | const struct itimerspec *ktmr) | |
132 | { | |
133 | enum hrtimer_mode htmode; | |
134 | ktime_t texp; | |
135 | int clockid = ctx->clockid; | |
136 | ||
137 | htmode = (flags & TFD_TIMER_ABSTIME) ? | |
138 | HRTIMER_MODE_ABS: HRTIMER_MODE_REL; | |
139 | ||
140 | texp = timespec_to_ktime(ktmr->it_value); | |
141 | ctx->expired = 0; | |
142 | ctx->ticks = 0; | |
143 | ctx->tintv = timespec_to_ktime(ktmr->it_interval); | |
144 | hrtimer_init(&ctx->tmr, clockid, htmode); | |
145 | hrtimer_set_expires(&ctx->tmr, texp); | |
146 | ctx->tmr.function = timerfd_tmrproc; | |
147 | if (texp.tv64 != 0) { | |
148 | hrtimer_start(&ctx->tmr, texp, htmode); | |
149 | if (timerfd_canceled(ctx)) | |
150 | return -ECANCELED; | |
151 | } | |
152 | return 0; | |
153 | } | |
154 | ||
155 | static int timerfd_release(struct inode *inode, struct file *file) | |
156 | { | |
157 | struct timerfd_ctx *ctx = file->private_data; | |
158 | ||
159 | timerfd_remove_cancel(ctx); | |
160 | hrtimer_cancel(&ctx->tmr); | |
161 | kfree_rcu(ctx, rcu); | |
162 | return 0; | |
163 | } | |
164 | ||
165 | static unsigned int timerfd_poll(struct file *file, poll_table *wait) | |
166 | { | |
167 | struct timerfd_ctx *ctx = file->private_data; | |
168 | unsigned int events = 0; | |
169 | unsigned long flags; | |
170 | ||
171 | poll_wait(file, &ctx->wqh, wait); | |
172 | ||
173 | spin_lock_irqsave(&ctx->wqh.lock, flags); | |
174 | if (ctx->ticks) | |
175 | events |= POLLIN; | |
176 | spin_unlock_irqrestore(&ctx->wqh.lock, flags); | |
177 | ||
178 | return events; | |
179 | } | |
180 | ||
181 | static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count, | |
182 | loff_t *ppos) | |
183 | { | |
184 | struct timerfd_ctx *ctx = file->private_data; | |
185 | ssize_t res; | |
186 | u64 ticks = 0; | |
187 | ||
188 | if (count < sizeof(ticks)) | |
189 | return -EINVAL; | |
190 | spin_lock_irq(&ctx->wqh.lock); | |
191 | if (file->f_flags & O_NONBLOCK) | |
192 | res = -EAGAIN; | |
193 | else | |
194 | res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks); | |
195 | ||
196 | /* | |
197 | * If clock has changed, we do not care about the | |
198 | * ticks and we do not rearm the timer. Userspace must | |
199 | * reevaluate anyway. | |
200 | */ | |
201 | if (timerfd_canceled(ctx)) { | |
202 | ctx->ticks = 0; | |
203 | ctx->expired = 0; | |
204 | res = -ECANCELED; | |
205 | } | |
206 | ||
207 | if (ctx->ticks) { | |
208 | ticks = ctx->ticks; | |
209 | ||
210 | if (ctx->expired && ctx->tintv.tv64) { | |
211 | /* | |
212 | * If tintv.tv64 != 0, this is a periodic timer that | |
213 | * needs to be re-armed. We avoid doing it in the timer | |
214 | * callback to avoid DoS attacks specifying a very | |
215 | * short timer period. | |
216 | */ | |
217 | ticks += hrtimer_forward_now(&ctx->tmr, | |
218 | ctx->tintv) - 1; | |
219 | hrtimer_restart(&ctx->tmr); | |
220 | } | |
221 | ctx->expired = 0; | |
222 | ctx->ticks = 0; | |
223 | } | |
224 | spin_unlock_irq(&ctx->wqh.lock); | |
225 | if (ticks) | |
226 | res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks); | |
227 | return res; | |
228 | } | |
229 | ||
230 | static const struct file_operations timerfd_fops = { | |
231 | .release = timerfd_release, | |
232 | .poll = timerfd_poll, | |
233 | .read = timerfd_read, | |
234 | .llseek = noop_llseek, | |
235 | }; | |
236 | ||
237 | static int timerfd_fget(int fd, struct fd *p) | |
238 | { | |
239 | struct fd f = fdget(fd); | |
240 | if (!f.file) | |
241 | return -EBADF; | |
242 | if (f.file->f_op != &timerfd_fops) { | |
243 | fdput(f); | |
244 | return -EINVAL; | |
245 | } | |
246 | *p = f; | |
247 | return 0; | |
248 | } | |
249 | ||
250 | SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags) | |
251 | { | |
252 | int ufd; | |
253 | struct timerfd_ctx *ctx; | |
254 | ||
255 | /* Check the TFD_* constants for consistency. */ | |
256 | BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC); | |
257 | BUILD_BUG_ON(TFD_NONBLOCK != O_NONBLOCK); | |
258 | ||
259 | if ((flags & ~TFD_CREATE_FLAGS) || | |
260 | (clockid != CLOCK_MONOTONIC && | |
261 | clockid != CLOCK_REALTIME)) | |
262 | return -EINVAL; | |
263 | ||
264 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
265 | if (!ctx) | |
266 | return -ENOMEM; | |
267 | ||
268 | init_waitqueue_head(&ctx->wqh); | |
269 | ctx->clockid = clockid; | |
270 | hrtimer_init(&ctx->tmr, clockid, HRTIMER_MODE_ABS); | |
271 | ctx->moffs = ktime_get_monotonic_offset(); | |
272 | ||
273 | ufd = anon_inode_getfd("[timerfd]", &timerfd_fops, ctx, | |
274 | O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS)); | |
275 | if (ufd < 0) | |
276 | kfree(ctx); | |
277 | ||
278 | return ufd; | |
279 | } | |
280 | ||
281 | SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags, | |
282 | const struct itimerspec __user *, utmr, | |
283 | struct itimerspec __user *, otmr) | |
284 | { | |
285 | struct fd f; | |
286 | struct timerfd_ctx *ctx; | |
287 | struct itimerspec ktmr, kotmr; | |
288 | int ret; | |
289 | ||
290 | if (copy_from_user(&ktmr, utmr, sizeof(ktmr))) | |
291 | return -EFAULT; | |
292 | ||
293 | if ((flags & ~TFD_SETTIME_FLAGS) || | |
294 | !timespec_valid(&ktmr.it_value) || | |
295 | !timespec_valid(&ktmr.it_interval)) | |
296 | return -EINVAL; | |
297 | ||
298 | ret = timerfd_fget(ufd, &f); | |
299 | if (ret) | |
300 | return ret; | |
301 | ctx = f.file->private_data; | |
302 | ||
303 | timerfd_setup_cancel(ctx, flags); | |
304 | ||
305 | /* | |
306 | * We need to stop the existing timer before reprogramming | |
307 | * it to the new values. | |
308 | */ | |
309 | for (;;) { | |
310 | spin_lock_irq(&ctx->wqh.lock); | |
311 | if (hrtimer_try_to_cancel(&ctx->tmr) >= 0) | |
312 | break; | |
313 | spin_unlock_irq(&ctx->wqh.lock); | |
314 | cpu_relax(); | |
315 | } | |
316 | ||
317 | /* | |
318 | * If the timer is expired and it's periodic, we need to advance it | |
319 | * because the caller may want to know the previous expiration time. | |
320 | * We do not update "ticks" and "expired" since the timer will be | |
321 | * re-programmed again in the following timerfd_setup() call. | |
322 | */ | |
323 | if (ctx->expired && ctx->tintv.tv64) | |
324 | hrtimer_forward_now(&ctx->tmr, ctx->tintv); | |
325 | ||
326 | kotmr.it_value = ktime_to_timespec(timerfd_get_remaining(ctx)); | |
327 | kotmr.it_interval = ktime_to_timespec(ctx->tintv); | |
328 | ||
329 | /* | |
330 | * Re-program the timer to the new value ... | |
331 | */ | |
332 | ret = timerfd_setup(ctx, flags, &ktmr); | |
333 | ||
334 | spin_unlock_irq(&ctx->wqh.lock); | |
335 | fdput(f); | |
336 | if (otmr && copy_to_user(otmr, &kotmr, sizeof(kotmr))) | |
337 | return -EFAULT; | |
338 | ||
339 | return ret; | |
340 | } | |
341 | ||
342 | SYSCALL_DEFINE2(timerfd_gettime, int, ufd, struct itimerspec __user *, otmr) | |
343 | { | |
344 | struct fd f; | |
345 | struct timerfd_ctx *ctx; | |
346 | struct itimerspec kotmr; | |
347 | int ret = timerfd_fget(ufd, &f); | |
348 | if (ret) | |
349 | return ret; | |
350 | ctx = f.file->private_data; | |
351 | ||
352 | spin_lock_irq(&ctx->wqh.lock); | |
353 | if (ctx->expired && ctx->tintv.tv64) { | |
354 | ctx->expired = 0; | |
355 | ctx->ticks += | |
356 | hrtimer_forward_now(&ctx->tmr, ctx->tintv) - 1; | |
357 | hrtimer_restart(&ctx->tmr); | |
358 | } | |
359 | kotmr.it_value = ktime_to_timespec(timerfd_get_remaining(ctx)); | |
360 | kotmr.it_interval = ktime_to_timespec(ctx->tintv); | |
361 | spin_unlock_irq(&ctx->wqh.lock); | |
362 | fdput(f); | |
363 | ||
364 | return copy_to_user(otmr, &kotmr, sizeof(kotmr)) ? -EFAULT: 0; | |
365 | } | |
366 |