xref: /linux/fs/timerfd.c (revision 765532c8aaac624b5f8687af6d319c6a1138a257)
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 
26 struct timerfd_ctx {
27 	struct hrtimer tmr;
28 	ktime_t tintv;
29 	wait_queue_head_t wqh;
30 	u64 ticks;
31 	int expired;
32 	int clockid;
33 };
34 
35 /*
36  * This gets called when the timer event triggers. We set the "expired"
37  * flag, but we do not re-arm the timer (in case it's necessary,
38  * tintv.tv64 != 0) until the timer is accessed.
39  */
40 static enum hrtimer_restart timerfd_tmrproc(struct hrtimer *htmr)
41 {
42 	struct timerfd_ctx *ctx = container_of(htmr, struct timerfd_ctx, tmr);
43 	unsigned long flags;
44 
45 	spin_lock_irqsave(&ctx->wqh.lock, flags);
46 	ctx->expired = 1;
47 	ctx->ticks++;
48 	wake_up_locked(&ctx->wqh);
49 	spin_unlock_irqrestore(&ctx->wqh.lock, flags);
50 
51 	return HRTIMER_NORESTART;
52 }
53 
54 static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx)
55 {
56 	ktime_t remaining;
57 
58 	remaining = hrtimer_expires_remaining(&ctx->tmr);
59 	return remaining.tv64 < 0 ? ktime_set(0, 0): remaining;
60 }
61 
62 static void timerfd_setup(struct timerfd_ctx *ctx, int flags,
63 			  const struct itimerspec *ktmr)
64 {
65 	enum hrtimer_mode htmode;
66 	ktime_t texp;
67 
68 	htmode = (flags & TFD_TIMER_ABSTIME) ?
69 		HRTIMER_MODE_ABS: HRTIMER_MODE_REL;
70 
71 	texp = timespec_to_ktime(ktmr->it_value);
72 	ctx->expired = 0;
73 	ctx->ticks = 0;
74 	ctx->tintv = timespec_to_ktime(ktmr->it_interval);
75 	hrtimer_init(&ctx->tmr, ctx->clockid, htmode);
76 	hrtimer_set_expires(&ctx->tmr, texp);
77 	ctx->tmr.function = timerfd_tmrproc;
78 	if (texp.tv64 != 0)
79 		hrtimer_start(&ctx->tmr, texp, htmode);
80 }
81 
82 static int timerfd_release(struct inode *inode, struct file *file)
83 {
84 	struct timerfd_ctx *ctx = file->private_data;
85 
86 	hrtimer_cancel(&ctx->tmr);
87 	kfree(ctx);
88 	return 0;
89 }
90 
91 static unsigned int timerfd_poll(struct file *file, poll_table *wait)
92 {
93 	struct timerfd_ctx *ctx = file->private_data;
94 	unsigned int events = 0;
95 	unsigned long flags;
96 
97 	poll_wait(file, &ctx->wqh, wait);
98 
99 	spin_lock_irqsave(&ctx->wqh.lock, flags);
100 	if (ctx->ticks)
101 		events |= POLLIN;
102 	spin_unlock_irqrestore(&ctx->wqh.lock, flags);
103 
104 	return events;
105 }
106 
107 static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
108 			    loff_t *ppos)
109 {
110 	struct timerfd_ctx *ctx = file->private_data;
111 	ssize_t res;
112 	u64 ticks = 0;
113 
114 	if (count < sizeof(ticks))
115 		return -EINVAL;
116 	spin_lock_irq(&ctx->wqh.lock);
117 	if (file->f_flags & O_NONBLOCK)
118 		res = -EAGAIN;
119 	else
120 		res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks);
121 	if (ctx->ticks) {
122 		ticks = ctx->ticks;
123 		if (ctx->expired && ctx->tintv.tv64) {
124 			/*
125 			 * If tintv.tv64 != 0, this is a periodic timer that
126 			 * needs to be re-armed. We avoid doing it in the timer
127 			 * callback to avoid DoS attacks specifying a very
128 			 * short timer period.
129 			 */
130 			ticks += hrtimer_forward_now(&ctx->tmr,
131 						     ctx->tintv) - 1;
132 			hrtimer_restart(&ctx->tmr);
133 		}
134 		ctx->expired = 0;
135 		ctx->ticks = 0;
136 	}
137 	spin_unlock_irq(&ctx->wqh.lock);
138 	if (ticks)
139 		res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks);
140 	return res;
141 }
142 
143 static const struct file_operations timerfd_fops = {
144 	.release	= timerfd_release,
145 	.poll		= timerfd_poll,
146 	.read		= timerfd_read,
147 	.llseek		= noop_llseek,
148 };
149 
150 static struct file *timerfd_fget(int fd)
151 {
152 	struct file *file;
153 
154 	file = fget(fd);
155 	if (!file)
156 		return ERR_PTR(-EBADF);
157 	if (file->f_op != &timerfd_fops) {
158 		fput(file);
159 		return ERR_PTR(-EINVAL);
160 	}
161 
162 	return file;
163 }
164 
165 SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
166 {
167 	int ufd;
168 	struct timerfd_ctx *ctx;
169 
170 	/* Check the TFD_* constants for consistency.  */
171 	BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC);
172 	BUILD_BUG_ON(TFD_NONBLOCK != O_NONBLOCK);
173 
174 	if ((flags & ~TFD_CREATE_FLAGS) ||
175 	    (clockid != CLOCK_MONOTONIC &&
176 	     clockid != CLOCK_REALTIME))
177 		return -EINVAL;
178 
179 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
180 	if (!ctx)
181 		return -ENOMEM;
182 
183 	init_waitqueue_head(&ctx->wqh);
184 	ctx->clockid = clockid;
185 	hrtimer_init(&ctx->tmr, clockid, HRTIMER_MODE_ABS);
186 
187 	ufd = anon_inode_getfd("[timerfd]", &timerfd_fops, ctx,
188 			       O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
189 	if (ufd < 0)
190 		kfree(ctx);
191 
192 	return ufd;
193 }
194 
195 SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
196 		const struct itimerspec __user *, utmr,
197 		struct itimerspec __user *, otmr)
198 {
199 	struct file *file;
200 	struct timerfd_ctx *ctx;
201 	struct itimerspec ktmr, kotmr;
202 
203 	if (copy_from_user(&ktmr, utmr, sizeof(ktmr)))
204 		return -EFAULT;
205 
206 	if ((flags & ~TFD_SETTIME_FLAGS) ||
207 	    !timespec_valid(&ktmr.it_value) ||
208 	    !timespec_valid(&ktmr.it_interval))
209 		return -EINVAL;
210 
211 	file = timerfd_fget(ufd);
212 	if (IS_ERR(file))
213 		return PTR_ERR(file);
214 	ctx = file->private_data;
215 
216 	/*
217 	 * We need to stop the existing timer before reprogramming
218 	 * it to the new values.
219 	 */
220 	for (;;) {
221 		spin_lock_irq(&ctx->wqh.lock);
222 		if (hrtimer_try_to_cancel(&ctx->tmr) >= 0)
223 			break;
224 		spin_unlock_irq(&ctx->wqh.lock);
225 		cpu_relax();
226 	}
227 
228 	/*
229 	 * If the timer is expired and it's periodic, we need to advance it
230 	 * because the caller may want to know the previous expiration time.
231 	 * We do not update "ticks" and "expired" since the timer will be
232 	 * re-programmed again in the following timerfd_setup() call.
233 	 */
234 	if (ctx->expired && ctx->tintv.tv64)
235 		hrtimer_forward_now(&ctx->tmr, ctx->tintv);
236 
237 	kotmr.it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
238 	kotmr.it_interval = ktime_to_timespec(ctx->tintv);
239 
240 	/*
241 	 * Re-program the timer to the new value ...
242 	 */
243 	timerfd_setup(ctx, flags, &ktmr);
244 
245 	spin_unlock_irq(&ctx->wqh.lock);
246 	fput(file);
247 	if (otmr && copy_to_user(otmr, &kotmr, sizeof(kotmr)))
248 		return -EFAULT;
249 
250 	return 0;
251 }
252 
253 SYSCALL_DEFINE2(timerfd_gettime, int, ufd, struct itimerspec __user *, otmr)
254 {
255 	struct file *file;
256 	struct timerfd_ctx *ctx;
257 	struct itimerspec kotmr;
258 
259 	file = timerfd_fget(ufd);
260 	if (IS_ERR(file))
261 		return PTR_ERR(file);
262 	ctx = file->private_data;
263 
264 	spin_lock_irq(&ctx->wqh.lock);
265 	if (ctx->expired && ctx->tintv.tv64) {
266 		ctx->expired = 0;
267 		ctx->ticks +=
268 			hrtimer_forward_now(&ctx->tmr, ctx->tintv) - 1;
269 		hrtimer_restart(&ctx->tmr);
270 	}
271 	kotmr.it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
272 	kotmr.it_interval = ktime_to_timespec(ctx->tintv);
273 	spin_unlock_irq(&ctx->wqh.lock);
274 	fput(file);
275 
276 	return copy_to_user(otmr, &kotmr, sizeof(kotmr)) ? -EFAULT: 0;
277 }
278 
279