xref: /linux/fs/select.c (revision 5bdef865eb358b6f3760e25e591ae115e9eeddef)
1 /*
2  * This file contains the procedures for the handling of select and poll
3  *
4  * Created for Linux based loosely upon Mathius Lattner's minix
5  * patches by Peter MacDonald. Heavily edited by Linus.
6  *
7  *  4 February 1994
8  *     COFF/ELF binary emulation. If the process has the STICKY_TIMEOUTS
9  *     flag set in its personality we do *not* modify the given timeout
10  *     parameter to reflect time remaining.
11  *
12  *  24 January 2000
13  *     Changed sys_poll()/do_poll() to use PAGE_SIZE chunk-based allocation
14  *     of fds to overcome nfds < 16390 descriptors limit (Tigran Aivazian).
15  */
16 
17 #include <linux/kernel.h>
18 #include <linux/syscalls.h>
19 #include <linux/module.h>
20 #include <linux/slab.h>
21 #include <linux/poll.h>
22 #include <linux/personality.h> /* for STICKY_TIMEOUTS */
23 #include <linux/file.h>
24 #include <linux/fdtable.h>
25 #include <linux/fs.h>
26 #include <linux/rcupdate.h>
27 #include <linux/hrtimer.h>
28 
29 #include <asm/uaccess.h>
30 
31 
32 /*
33  * Estimate expected accuracy in ns from a timeval.
34  *
35  * After quite a bit of churning around, we've settled on
36  * a simple thing of taking 0.1% of the timeout as the
37  * slack, with a cap of 100 msec.
38  * "nice" tasks get a 0.5% slack instead.
39  *
40  * Consider this comment an open invitation to come up with even
41  * better solutions..
42  */
43 
44 static long __estimate_accuracy(struct timespec *tv)
45 {
46 	long slack;
47 	int divfactor = 1000;
48 
49 	if (task_nice(current) > 0)
50 		divfactor = divfactor / 5;
51 
52 	slack = tv->tv_nsec / divfactor;
53 	slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
54 
55 	if (slack > 100 * NSEC_PER_MSEC)
56 		slack =  100 * NSEC_PER_MSEC;
57 
58 	if (slack < 0)
59 		slack = 0;
60 	return slack;
61 }
62 
63 static long estimate_accuracy(struct timespec *tv)
64 {
65 	unsigned long ret;
66 	struct timespec now;
67 
68 	/*
69 	 * Realtime tasks get a slack of 0 for obvious reasons.
70 	 */
71 
72 	if (rt_task(current))
73 		return 0;
74 
75 	ktime_get_ts(&now);
76 	now = timespec_sub(*tv, now);
77 	ret = __estimate_accuracy(&now);
78 	if (ret < current->timer_slack_ns)
79 		return current->timer_slack_ns;
80 	return ret;
81 }
82 
83 
84 
85 struct poll_table_page {
86 	struct poll_table_page * next;
87 	struct poll_table_entry * entry;
88 	struct poll_table_entry entries[0];
89 };
90 
91 #define POLL_TABLE_FULL(table) \
92 	((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
93 
94 /*
95  * Ok, Peter made a complicated, but straightforward multiple_wait() function.
96  * I have rewritten this, taking some shortcuts: This code may not be easy to
97  * follow, but it should be free of race-conditions, and it's practical. If you
98  * understand what I'm doing here, then you understand how the linux
99  * sleep/wakeup mechanism works.
100  *
101  * Two very simple procedures, poll_wait() and poll_freewait() make all the
102  * work.  poll_wait() is an inline-function defined in <linux/poll.h>,
103  * as all select/poll functions have to call it to add an entry to the
104  * poll table.
105  */
106 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
107 		       poll_table *p);
108 
109 void poll_initwait(struct poll_wqueues *pwq)
110 {
111 	init_poll_funcptr(&pwq->pt, __pollwait);
112 	pwq->polling_task = current;
113 	pwq->error = 0;
114 	pwq->table = NULL;
115 	pwq->inline_index = 0;
116 }
117 EXPORT_SYMBOL(poll_initwait);
118 
119 static void free_poll_entry(struct poll_table_entry *entry)
120 {
121 	remove_wait_queue(entry->wait_address, &entry->wait);
122 	fput(entry->filp);
123 }
124 
125 void poll_freewait(struct poll_wqueues *pwq)
126 {
127 	struct poll_table_page * p = pwq->table;
128 	int i;
129 	for (i = 0; i < pwq->inline_index; i++)
130 		free_poll_entry(pwq->inline_entries + i);
131 	while (p) {
132 		struct poll_table_entry * entry;
133 		struct poll_table_page *old;
134 
135 		entry = p->entry;
136 		do {
137 			entry--;
138 			free_poll_entry(entry);
139 		} while (entry > p->entries);
140 		old = p;
141 		p = p->next;
142 		free_page((unsigned long) old);
143 	}
144 }
145 EXPORT_SYMBOL(poll_freewait);
146 
147 static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
148 {
149 	struct poll_table_page *table = p->table;
150 
151 	if (p->inline_index < N_INLINE_POLL_ENTRIES)
152 		return p->inline_entries + p->inline_index++;
153 
154 	if (!table || POLL_TABLE_FULL(table)) {
155 		struct poll_table_page *new_table;
156 
157 		new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
158 		if (!new_table) {
159 			p->error = -ENOMEM;
160 			return NULL;
161 		}
162 		new_table->entry = new_table->entries;
163 		new_table->next = table;
164 		p->table = new_table;
165 		table = new_table;
166 	}
167 
168 	return table->entry++;
169 }
170 
171 static int __pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
172 {
173 	struct poll_wqueues *pwq = wait->private;
174 	DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
175 
176 	/*
177 	 * Although this function is called under waitqueue lock, LOCK
178 	 * doesn't imply write barrier and the users expect write
179 	 * barrier semantics on wakeup functions.  The following
180 	 * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
181 	 * and is paired with set_mb() in poll_schedule_timeout.
182 	 */
183 	smp_wmb();
184 	pwq->triggered = 1;
185 
186 	/*
187 	 * Perform the default wake up operation using a dummy
188 	 * waitqueue.
189 	 *
190 	 * TODO: This is hacky but there currently is no interface to
191 	 * pass in @sync.  @sync is scheduled to be removed and once
192 	 * that happens, wake_up_process() can be used directly.
193 	 */
194 	return default_wake_function(&dummy_wait, mode, sync, key);
195 }
196 
197 static int pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
198 {
199 	struct poll_table_entry *entry;
200 
201 	entry = container_of(wait, struct poll_table_entry, wait);
202 	if (key && !((unsigned long)key & entry->key))
203 		return 0;
204 	return __pollwake(wait, mode, sync, key);
205 }
206 
207 /* Add a new entry */
208 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
209 				poll_table *p)
210 {
211 	struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
212 	struct poll_table_entry *entry = poll_get_entry(pwq);
213 	if (!entry)
214 		return;
215 	get_file(filp);
216 	entry->filp = filp;
217 	entry->wait_address = wait_address;
218 	entry->key = p->key;
219 	init_waitqueue_func_entry(&entry->wait, pollwake);
220 	entry->wait.private = pwq;
221 	add_wait_queue(wait_address, &entry->wait);
222 }
223 
224 int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
225 			  ktime_t *expires, unsigned long slack)
226 {
227 	int rc = -EINTR;
228 
229 	set_current_state(state);
230 	if (!pwq->triggered)
231 		rc = schedule_hrtimeout_range(expires, slack, HRTIMER_MODE_ABS);
232 	__set_current_state(TASK_RUNNING);
233 
234 	/*
235 	 * Prepare for the next iteration.
236 	 *
237 	 * The following set_mb() serves two purposes.  First, it's
238 	 * the counterpart rmb of the wmb in pollwake() such that data
239 	 * written before wake up is always visible after wake up.
240 	 * Second, the full barrier guarantees that triggered clearing
241 	 * doesn't pass event check of the next iteration.  Note that
242 	 * this problem doesn't exist for the first iteration as
243 	 * add_wait_queue() has full barrier semantics.
244 	 */
245 	set_mb(pwq->triggered, 0);
246 
247 	return rc;
248 }
249 EXPORT_SYMBOL(poll_schedule_timeout);
250 
251 /**
252  * poll_select_set_timeout - helper function to setup the timeout value
253  * @to:		pointer to timespec variable for the final timeout
254  * @sec:	seconds (from user space)
255  * @nsec:	nanoseconds (from user space)
256  *
257  * Note, we do not use a timespec for the user space value here, That
258  * way we can use the function for timeval and compat interfaces as well.
259  *
260  * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
261  */
262 int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
263 {
264 	struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
265 
266 	if (!timespec_valid(&ts))
267 		return -EINVAL;
268 
269 	/* Optimize for the zero timeout value here */
270 	if (!sec && !nsec) {
271 		to->tv_sec = to->tv_nsec = 0;
272 	} else {
273 		ktime_get_ts(to);
274 		*to = timespec_add_safe(*to, ts);
275 	}
276 	return 0;
277 }
278 
279 static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
280 				      int timeval, int ret)
281 {
282 	struct timespec rts;
283 	struct timeval rtv;
284 
285 	if (!p)
286 		return ret;
287 
288 	if (current->personality & STICKY_TIMEOUTS)
289 		goto sticky;
290 
291 	/* No update for zero timeout */
292 	if (!end_time->tv_sec && !end_time->tv_nsec)
293 		return ret;
294 
295 	ktime_get_ts(&rts);
296 	rts = timespec_sub(*end_time, rts);
297 	if (rts.tv_sec < 0)
298 		rts.tv_sec = rts.tv_nsec = 0;
299 
300 	if (timeval) {
301 		rtv.tv_sec = rts.tv_sec;
302 		rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
303 
304 		if (!copy_to_user(p, &rtv, sizeof(rtv)))
305 			return ret;
306 
307 	} else if (!copy_to_user(p, &rts, sizeof(rts)))
308 		return ret;
309 
310 	/*
311 	 * If an application puts its timeval in read-only memory, we
312 	 * don't want the Linux-specific update to the timeval to
313 	 * cause a fault after the select has completed
314 	 * successfully. However, because we're not updating the
315 	 * timeval, we can't restart the system call.
316 	 */
317 
318 sticky:
319 	if (ret == -ERESTARTNOHAND)
320 		ret = -EINTR;
321 	return ret;
322 }
323 
324 #define FDS_IN(fds, n)		(fds->in + n)
325 #define FDS_OUT(fds, n)		(fds->out + n)
326 #define FDS_EX(fds, n)		(fds->ex + n)
327 
328 #define BITS(fds, n)	(*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
329 
330 static int max_select_fd(unsigned long n, fd_set_bits *fds)
331 {
332 	unsigned long *open_fds;
333 	unsigned long set;
334 	int max;
335 	struct fdtable *fdt;
336 
337 	/* handle last in-complete long-word first */
338 	set = ~(~0UL << (n & (__NFDBITS-1)));
339 	n /= __NFDBITS;
340 	fdt = files_fdtable(current->files);
341 	open_fds = fdt->open_fds->fds_bits+n;
342 	max = 0;
343 	if (set) {
344 		set &= BITS(fds, n);
345 		if (set) {
346 			if (!(set & ~*open_fds))
347 				goto get_max;
348 			return -EBADF;
349 		}
350 	}
351 	while (n) {
352 		open_fds--;
353 		n--;
354 		set = BITS(fds, n);
355 		if (!set)
356 			continue;
357 		if (set & ~*open_fds)
358 			return -EBADF;
359 		if (max)
360 			continue;
361 get_max:
362 		do {
363 			max++;
364 			set >>= 1;
365 		} while (set);
366 		max += n * __NFDBITS;
367 	}
368 
369 	return max;
370 }
371 
372 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
373 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
374 #define POLLEX_SET (POLLPRI)
375 
376 static inline void wait_key_set(poll_table *wait, unsigned long in,
377 				unsigned long out, unsigned long bit)
378 {
379 	if (wait) {
380 		wait->key = POLLEX_SET;
381 		if (in & bit)
382 			wait->key |= POLLIN_SET;
383 		if (out & bit)
384 			wait->key |= POLLOUT_SET;
385 	}
386 }
387 
388 int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
389 {
390 	ktime_t expire, *to = NULL;
391 	struct poll_wqueues table;
392 	poll_table *wait;
393 	int retval, i, timed_out = 0;
394 	unsigned long slack = 0;
395 
396 	rcu_read_lock();
397 	retval = max_select_fd(n, fds);
398 	rcu_read_unlock();
399 
400 	if (retval < 0)
401 		return retval;
402 	n = retval;
403 
404 	poll_initwait(&table);
405 	wait = &table.pt;
406 	if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
407 		wait = NULL;
408 		timed_out = 1;
409 	}
410 
411 	if (end_time && !timed_out)
412 		slack = estimate_accuracy(end_time);
413 
414 	retval = 0;
415 	for (;;) {
416 		unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
417 
418 		inp = fds->in; outp = fds->out; exp = fds->ex;
419 		rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
420 
421 		for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
422 			unsigned long in, out, ex, all_bits, bit = 1, mask, j;
423 			unsigned long res_in = 0, res_out = 0, res_ex = 0;
424 			const struct file_operations *f_op = NULL;
425 			struct file *file = NULL;
426 
427 			in = *inp++; out = *outp++; ex = *exp++;
428 			all_bits = in | out | ex;
429 			if (all_bits == 0) {
430 				i += __NFDBITS;
431 				continue;
432 			}
433 
434 			for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
435 				int fput_needed;
436 				if (i >= n)
437 					break;
438 				if (!(bit & all_bits))
439 					continue;
440 				file = fget_light(i, &fput_needed);
441 				if (file) {
442 					f_op = file->f_op;
443 					mask = DEFAULT_POLLMASK;
444 					if (f_op && f_op->poll) {
445 						wait_key_set(wait, in, out, bit);
446 						mask = (*f_op->poll)(file, wait);
447 					}
448 					fput_light(file, fput_needed);
449 					if ((mask & POLLIN_SET) && (in & bit)) {
450 						res_in |= bit;
451 						retval++;
452 						wait = NULL;
453 					}
454 					if ((mask & POLLOUT_SET) && (out & bit)) {
455 						res_out |= bit;
456 						retval++;
457 						wait = NULL;
458 					}
459 					if ((mask & POLLEX_SET) && (ex & bit)) {
460 						res_ex |= bit;
461 						retval++;
462 						wait = NULL;
463 					}
464 				}
465 			}
466 			if (res_in)
467 				*rinp = res_in;
468 			if (res_out)
469 				*routp = res_out;
470 			if (res_ex)
471 				*rexp = res_ex;
472 			cond_resched();
473 		}
474 		wait = NULL;
475 		if (retval || timed_out || signal_pending(current))
476 			break;
477 		if (table.error) {
478 			retval = table.error;
479 			break;
480 		}
481 
482 		/*
483 		 * If this is the first loop and we have a timeout
484 		 * given, then we convert to ktime_t and set the to
485 		 * pointer to the expiry value.
486 		 */
487 		if (end_time && !to) {
488 			expire = timespec_to_ktime(*end_time);
489 			to = &expire;
490 		}
491 
492 		if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
493 					   to, slack))
494 			timed_out = 1;
495 	}
496 
497 	poll_freewait(&table);
498 
499 	return retval;
500 }
501 
502 /*
503  * We can actually return ERESTARTSYS instead of EINTR, but I'd
504  * like to be certain this leads to no problems. So I return
505  * EINTR just for safety.
506  *
507  * Update: ERESTARTSYS breaks at least the xview clock binary, so
508  * I'm trying ERESTARTNOHAND which restart only when you want to.
509  */
510 #define MAX_SELECT_SECONDS \
511 	((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
512 
513 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
514 			   fd_set __user *exp, struct timespec *end_time)
515 {
516 	fd_set_bits fds;
517 	void *bits;
518 	int ret, max_fds;
519 	unsigned int size;
520 	struct fdtable *fdt;
521 	/* Allocate small arguments on the stack to save memory and be faster */
522 	long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
523 
524 	ret = -EINVAL;
525 	if (n < 0)
526 		goto out_nofds;
527 
528 	/* max_fds can increase, so grab it once to avoid race */
529 	rcu_read_lock();
530 	fdt = files_fdtable(current->files);
531 	max_fds = fdt->max_fds;
532 	rcu_read_unlock();
533 	if (n > max_fds)
534 		n = max_fds;
535 
536 	/*
537 	 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
538 	 * since we used fdset we need to allocate memory in units of
539 	 * long-words.
540 	 */
541 	size = FDS_BYTES(n);
542 	bits = stack_fds;
543 	if (size > sizeof(stack_fds) / 6) {
544 		/* Not enough space in on-stack array; must use kmalloc */
545 		ret = -ENOMEM;
546 		bits = kmalloc(6 * size, GFP_KERNEL);
547 		if (!bits)
548 			goto out_nofds;
549 	}
550 	fds.in      = bits;
551 	fds.out     = bits +   size;
552 	fds.ex      = bits + 2*size;
553 	fds.res_in  = bits + 3*size;
554 	fds.res_out = bits + 4*size;
555 	fds.res_ex  = bits + 5*size;
556 
557 	if ((ret = get_fd_set(n, inp, fds.in)) ||
558 	    (ret = get_fd_set(n, outp, fds.out)) ||
559 	    (ret = get_fd_set(n, exp, fds.ex)))
560 		goto out;
561 	zero_fd_set(n, fds.res_in);
562 	zero_fd_set(n, fds.res_out);
563 	zero_fd_set(n, fds.res_ex);
564 
565 	ret = do_select(n, &fds, end_time);
566 
567 	if (ret < 0)
568 		goto out;
569 	if (!ret) {
570 		ret = -ERESTARTNOHAND;
571 		if (signal_pending(current))
572 			goto out;
573 		ret = 0;
574 	}
575 
576 	if (set_fd_set(n, inp, fds.res_in) ||
577 	    set_fd_set(n, outp, fds.res_out) ||
578 	    set_fd_set(n, exp, fds.res_ex))
579 		ret = -EFAULT;
580 
581 out:
582 	if (bits != stack_fds)
583 		kfree(bits);
584 out_nofds:
585 	return ret;
586 }
587 
588 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
589 		fd_set __user *, exp, struct timeval __user *, tvp)
590 {
591 	struct timespec end_time, *to = NULL;
592 	struct timeval tv;
593 	int ret;
594 
595 	if (tvp) {
596 		if (copy_from_user(&tv, tvp, sizeof(tv)))
597 			return -EFAULT;
598 
599 		to = &end_time;
600 		if (poll_select_set_timeout(to,
601 				tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
602 				(tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
603 			return -EINVAL;
604 	}
605 
606 	ret = core_sys_select(n, inp, outp, exp, to);
607 	ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
608 
609 	return ret;
610 }
611 
612 #ifdef HAVE_SET_RESTORE_SIGMASK
613 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
614 		       fd_set __user *exp, struct timespec __user *tsp,
615 		       const sigset_t __user *sigmask, size_t sigsetsize)
616 {
617 	sigset_t ksigmask, sigsaved;
618 	struct timespec ts, end_time, *to = NULL;
619 	int ret;
620 
621 	if (tsp) {
622 		if (copy_from_user(&ts, tsp, sizeof(ts)))
623 			return -EFAULT;
624 
625 		to = &end_time;
626 		if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
627 			return -EINVAL;
628 	}
629 
630 	if (sigmask) {
631 		/* XXX: Don't preclude handling different sized sigset_t's.  */
632 		if (sigsetsize != sizeof(sigset_t))
633 			return -EINVAL;
634 		if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
635 			return -EFAULT;
636 
637 		sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
638 		sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
639 	}
640 
641 	ret = core_sys_select(n, inp, outp, exp, to);
642 	ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
643 
644 	if (ret == -ERESTARTNOHAND) {
645 		/*
646 		 * Don't restore the signal mask yet. Let do_signal() deliver
647 		 * the signal on the way back to userspace, before the signal
648 		 * mask is restored.
649 		 */
650 		if (sigmask) {
651 			memcpy(&current->saved_sigmask, &sigsaved,
652 					sizeof(sigsaved));
653 			set_restore_sigmask();
654 		}
655 	} else if (sigmask)
656 		sigprocmask(SIG_SETMASK, &sigsaved, NULL);
657 
658 	return ret;
659 }
660 
661 /*
662  * Most architectures can't handle 7-argument syscalls. So we provide a
663  * 6-argument version where the sixth argument is a pointer to a structure
664  * which has a pointer to the sigset_t itself followed by a size_t containing
665  * the sigset size.
666  */
667 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
668 		fd_set __user *, exp, struct timespec __user *, tsp,
669 		void __user *, sig)
670 {
671 	size_t sigsetsize = 0;
672 	sigset_t __user *up = NULL;
673 
674 	if (sig) {
675 		if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
676 		    || __get_user(up, (sigset_t __user * __user *)sig)
677 		    || __get_user(sigsetsize,
678 				(size_t __user *)(sig+sizeof(void *))))
679 			return -EFAULT;
680 	}
681 
682 	return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
683 }
684 #endif /* HAVE_SET_RESTORE_SIGMASK */
685 
686 struct poll_list {
687 	struct poll_list *next;
688 	int len;
689 	struct pollfd entries[0];
690 };
691 
692 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
693 
694 /*
695  * Fish for pollable events on the pollfd->fd file descriptor. We're only
696  * interested in events matching the pollfd->events mask, and the result
697  * matching that mask is both recorded in pollfd->revents and returned. The
698  * pwait poll_table will be used by the fd-provided poll handler for waiting,
699  * if non-NULL.
700  */
701 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
702 {
703 	unsigned int mask;
704 	int fd;
705 
706 	mask = 0;
707 	fd = pollfd->fd;
708 	if (fd >= 0) {
709 		int fput_needed;
710 		struct file * file;
711 
712 		file = fget_light(fd, &fput_needed);
713 		mask = POLLNVAL;
714 		if (file != NULL) {
715 			mask = DEFAULT_POLLMASK;
716 			if (file->f_op && file->f_op->poll) {
717 				if (pwait)
718 					pwait->key = pollfd->events |
719 							POLLERR | POLLHUP;
720 				mask = file->f_op->poll(file, pwait);
721 			}
722 			/* Mask out unneeded events. */
723 			mask &= pollfd->events | POLLERR | POLLHUP;
724 			fput_light(file, fput_needed);
725 		}
726 	}
727 	pollfd->revents = mask;
728 
729 	return mask;
730 }
731 
732 static int do_poll(unsigned int nfds,  struct poll_list *list,
733 		   struct poll_wqueues *wait, struct timespec *end_time)
734 {
735 	poll_table* pt = &wait->pt;
736 	ktime_t expire, *to = NULL;
737 	int timed_out = 0, count = 0;
738 	unsigned long slack = 0;
739 
740 	/* Optimise the no-wait case */
741 	if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
742 		pt = NULL;
743 		timed_out = 1;
744 	}
745 
746 	if (end_time && !timed_out)
747 		slack = estimate_accuracy(end_time);
748 
749 	for (;;) {
750 		struct poll_list *walk;
751 
752 		for (walk = list; walk != NULL; walk = walk->next) {
753 			struct pollfd * pfd, * pfd_end;
754 
755 			pfd = walk->entries;
756 			pfd_end = pfd + walk->len;
757 			for (; pfd != pfd_end; pfd++) {
758 				/*
759 				 * Fish for events. If we found one, record it
760 				 * and kill the poll_table, so we don't
761 				 * needlessly register any other waiters after
762 				 * this. They'll get immediately deregistered
763 				 * when we break out and return.
764 				 */
765 				if (do_pollfd(pfd, pt)) {
766 					count++;
767 					pt = NULL;
768 				}
769 			}
770 		}
771 		/*
772 		 * All waiters have already been registered, so don't provide
773 		 * a poll_table to them on the next loop iteration.
774 		 */
775 		pt = NULL;
776 		if (!count) {
777 			count = wait->error;
778 			if (signal_pending(current))
779 				count = -EINTR;
780 		}
781 		if (count || timed_out)
782 			break;
783 
784 		/*
785 		 * If this is the first loop and we have a timeout
786 		 * given, then we convert to ktime_t and set the to
787 		 * pointer to the expiry value.
788 		 */
789 		if (end_time && !to) {
790 			expire = timespec_to_ktime(*end_time);
791 			to = &expire;
792 		}
793 
794 		if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
795 			timed_out = 1;
796 	}
797 	return count;
798 }
799 
800 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
801 			sizeof(struct pollfd))
802 
803 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
804 		struct timespec *end_time)
805 {
806 	struct poll_wqueues table;
807  	int err = -EFAULT, fdcount, len, size;
808 	/* Allocate small arguments on the stack to save memory and be
809 	   faster - use long to make sure the buffer is aligned properly
810 	   on 64 bit archs to avoid unaligned access */
811 	long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
812 	struct poll_list *const head = (struct poll_list *)stack_pps;
813  	struct poll_list *walk = head;
814  	unsigned long todo = nfds;
815 
816 	if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
817 		return -EINVAL;
818 
819 	len = min_t(unsigned int, nfds, N_STACK_PPS);
820 	for (;;) {
821 		walk->next = NULL;
822 		walk->len = len;
823 		if (!len)
824 			break;
825 
826 		if (copy_from_user(walk->entries, ufds + nfds-todo,
827 					sizeof(struct pollfd) * walk->len))
828 			goto out_fds;
829 
830 		todo -= walk->len;
831 		if (!todo)
832 			break;
833 
834 		len = min(todo, POLLFD_PER_PAGE);
835 		size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
836 		walk = walk->next = kmalloc(size, GFP_KERNEL);
837 		if (!walk) {
838 			err = -ENOMEM;
839 			goto out_fds;
840 		}
841 	}
842 
843 	poll_initwait(&table);
844 	fdcount = do_poll(nfds, head, &table, end_time);
845 	poll_freewait(&table);
846 
847 	for (walk = head; walk; walk = walk->next) {
848 		struct pollfd *fds = walk->entries;
849 		int j;
850 
851 		for (j = 0; j < walk->len; j++, ufds++)
852 			if (__put_user(fds[j].revents, &ufds->revents))
853 				goto out_fds;
854   	}
855 
856 	err = fdcount;
857 out_fds:
858 	walk = head->next;
859 	while (walk) {
860 		struct poll_list *pos = walk;
861 		walk = walk->next;
862 		kfree(pos);
863 	}
864 
865 	return err;
866 }
867 
868 static long do_restart_poll(struct restart_block *restart_block)
869 {
870 	struct pollfd __user *ufds = restart_block->poll.ufds;
871 	int nfds = restart_block->poll.nfds;
872 	struct timespec *to = NULL, end_time;
873 	int ret;
874 
875 	if (restart_block->poll.has_timeout) {
876 		end_time.tv_sec = restart_block->poll.tv_sec;
877 		end_time.tv_nsec = restart_block->poll.tv_nsec;
878 		to = &end_time;
879 	}
880 
881 	ret = do_sys_poll(ufds, nfds, to);
882 
883 	if (ret == -EINTR) {
884 		restart_block->fn = do_restart_poll;
885 		ret = -ERESTART_RESTARTBLOCK;
886 	}
887 	return ret;
888 }
889 
890 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
891 		long, timeout_msecs)
892 {
893 	struct timespec end_time, *to = NULL;
894 	int ret;
895 
896 	if (timeout_msecs >= 0) {
897 		to = &end_time;
898 		poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
899 			NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
900 	}
901 
902 	ret = do_sys_poll(ufds, nfds, to);
903 
904 	if (ret == -EINTR) {
905 		struct restart_block *restart_block;
906 
907 		restart_block = &current_thread_info()->restart_block;
908 		restart_block->fn = do_restart_poll;
909 		restart_block->poll.ufds = ufds;
910 		restart_block->poll.nfds = nfds;
911 
912 		if (timeout_msecs >= 0) {
913 			restart_block->poll.tv_sec = end_time.tv_sec;
914 			restart_block->poll.tv_nsec = end_time.tv_nsec;
915 			restart_block->poll.has_timeout = 1;
916 		} else
917 			restart_block->poll.has_timeout = 0;
918 
919 		ret = -ERESTART_RESTARTBLOCK;
920 	}
921 	return ret;
922 }
923 
924 #ifdef HAVE_SET_RESTORE_SIGMASK
925 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
926 		struct timespec __user *, tsp, const sigset_t __user *, sigmask,
927 		size_t, sigsetsize)
928 {
929 	sigset_t ksigmask, sigsaved;
930 	struct timespec ts, end_time, *to = NULL;
931 	int ret;
932 
933 	if (tsp) {
934 		if (copy_from_user(&ts, tsp, sizeof(ts)))
935 			return -EFAULT;
936 
937 		to = &end_time;
938 		if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
939 			return -EINVAL;
940 	}
941 
942 	if (sigmask) {
943 		/* XXX: Don't preclude handling different sized sigset_t's.  */
944 		if (sigsetsize != sizeof(sigset_t))
945 			return -EINVAL;
946 		if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
947 			return -EFAULT;
948 
949 		sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
950 		sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
951 	}
952 
953 	ret = do_sys_poll(ufds, nfds, to);
954 
955 	/* We can restart this syscall, usually */
956 	if (ret == -EINTR) {
957 		/*
958 		 * Don't restore the signal mask yet. Let do_signal() deliver
959 		 * the signal on the way back to userspace, before the signal
960 		 * mask is restored.
961 		 */
962 		if (sigmask) {
963 			memcpy(&current->saved_sigmask, &sigsaved,
964 					sizeof(sigsaved));
965 			set_restore_sigmask();
966 		}
967 		ret = -ERESTARTNOHAND;
968 	} else if (sigmask)
969 		sigprocmask(SIG_SETMASK, &sigsaved, NULL);
970 
971 	ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
972 
973 	return ret;
974 }
975 #endif /* HAVE_SET_RESTORE_SIGMASK */
976