xref: /openbmc/linux/fs/select.c (revision 7dd65feb)
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/sched.h>
19 #include <linux/syscalls.h>
20 #include <linux/module.h>
21 #include <linux/slab.h>
22 #include <linux/poll.h>
23 #include <linux/personality.h> /* for STICKY_TIMEOUTS */
24 #include <linux/file.h>
25 #include <linux/fdtable.h>
26 #include <linux/fs.h>
27 #include <linux/rcupdate.h>
28 #include <linux/hrtimer.h>
29 
30 #include <asm/uaccess.h>
31 
32 
33 /*
34  * Estimate expected accuracy in ns from a timeval.
35  *
36  * After quite a bit of churning around, we've settled on
37  * a simple thing of taking 0.1% of the timeout as the
38  * slack, with a cap of 100 msec.
39  * "nice" tasks get a 0.5% slack instead.
40  *
41  * Consider this comment an open invitation to come up with even
42  * better solutions..
43  */
44 
45 #define MAX_SLACK	(100 * NSEC_PER_MSEC)
46 
47 static long __estimate_accuracy(struct timespec *tv)
48 {
49 	long slack;
50 	int divfactor = 1000;
51 
52 	if (tv->tv_sec < 0)
53 		return 0;
54 
55 	if (task_nice(current) > 0)
56 		divfactor = divfactor / 5;
57 
58 	if (tv->tv_sec > MAX_SLACK / (NSEC_PER_SEC/divfactor))
59 		return MAX_SLACK;
60 
61 	slack = tv->tv_nsec / divfactor;
62 	slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
63 
64 	if (slack > MAX_SLACK)
65 		return MAX_SLACK;
66 
67 	return slack;
68 }
69 
70 static long estimate_accuracy(struct timespec *tv)
71 {
72 	unsigned long ret;
73 	struct timespec now;
74 
75 	/*
76 	 * Realtime tasks get a slack of 0 for obvious reasons.
77 	 */
78 
79 	if (rt_task(current))
80 		return 0;
81 
82 	ktime_get_ts(&now);
83 	now = timespec_sub(*tv, now);
84 	ret = __estimate_accuracy(&now);
85 	if (ret < current->timer_slack_ns)
86 		return current->timer_slack_ns;
87 	return ret;
88 }
89 
90 
91 
92 struct poll_table_page {
93 	struct poll_table_page * next;
94 	struct poll_table_entry * entry;
95 	struct poll_table_entry entries[0];
96 };
97 
98 #define POLL_TABLE_FULL(table) \
99 	((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
100 
101 /*
102  * Ok, Peter made a complicated, but straightforward multiple_wait() function.
103  * I have rewritten this, taking some shortcuts: This code may not be easy to
104  * follow, but it should be free of race-conditions, and it's practical. If you
105  * understand what I'm doing here, then you understand how the linux
106  * sleep/wakeup mechanism works.
107  *
108  * Two very simple procedures, poll_wait() and poll_freewait() make all the
109  * work.  poll_wait() is an inline-function defined in <linux/poll.h>,
110  * as all select/poll functions have to call it to add an entry to the
111  * poll table.
112  */
113 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
114 		       poll_table *p);
115 
116 void poll_initwait(struct poll_wqueues *pwq)
117 {
118 	init_poll_funcptr(&pwq->pt, __pollwait);
119 	pwq->polling_task = current;
120 	pwq->triggered = 0;
121 	pwq->error = 0;
122 	pwq->table = NULL;
123 	pwq->inline_index = 0;
124 }
125 EXPORT_SYMBOL(poll_initwait);
126 
127 static void free_poll_entry(struct poll_table_entry *entry)
128 {
129 	remove_wait_queue(entry->wait_address, &entry->wait);
130 	fput(entry->filp);
131 }
132 
133 void poll_freewait(struct poll_wqueues *pwq)
134 {
135 	struct poll_table_page * p = pwq->table;
136 	int i;
137 	for (i = 0; i < pwq->inline_index; i++)
138 		free_poll_entry(pwq->inline_entries + i);
139 	while (p) {
140 		struct poll_table_entry * entry;
141 		struct poll_table_page *old;
142 
143 		entry = p->entry;
144 		do {
145 			entry--;
146 			free_poll_entry(entry);
147 		} while (entry > p->entries);
148 		old = p;
149 		p = p->next;
150 		free_page((unsigned long) old);
151 	}
152 }
153 EXPORT_SYMBOL(poll_freewait);
154 
155 static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
156 {
157 	struct poll_table_page *table = p->table;
158 
159 	if (p->inline_index < N_INLINE_POLL_ENTRIES)
160 		return p->inline_entries + p->inline_index++;
161 
162 	if (!table || POLL_TABLE_FULL(table)) {
163 		struct poll_table_page *new_table;
164 
165 		new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
166 		if (!new_table) {
167 			p->error = -ENOMEM;
168 			return NULL;
169 		}
170 		new_table->entry = new_table->entries;
171 		new_table->next = table;
172 		p->table = new_table;
173 		table = new_table;
174 	}
175 
176 	return table->entry++;
177 }
178 
179 static int __pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
180 {
181 	struct poll_wqueues *pwq = wait->private;
182 	DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
183 
184 	/*
185 	 * Although this function is called under waitqueue lock, LOCK
186 	 * doesn't imply write barrier and the users expect write
187 	 * barrier semantics on wakeup functions.  The following
188 	 * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
189 	 * and is paired with set_mb() in poll_schedule_timeout.
190 	 */
191 	smp_wmb();
192 	pwq->triggered = 1;
193 
194 	/*
195 	 * Perform the default wake up operation using a dummy
196 	 * waitqueue.
197 	 *
198 	 * TODO: This is hacky but there currently is no interface to
199 	 * pass in @sync.  @sync is scheduled to be removed and once
200 	 * that happens, wake_up_process() can be used directly.
201 	 */
202 	return default_wake_function(&dummy_wait, mode, sync, key);
203 }
204 
205 static int pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
206 {
207 	struct poll_table_entry *entry;
208 
209 	entry = container_of(wait, struct poll_table_entry, wait);
210 	if (key && !((unsigned long)key & entry->key))
211 		return 0;
212 	return __pollwake(wait, mode, sync, key);
213 }
214 
215 /* Add a new entry */
216 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
217 				poll_table *p)
218 {
219 	struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
220 	struct poll_table_entry *entry = poll_get_entry(pwq);
221 	if (!entry)
222 		return;
223 	get_file(filp);
224 	entry->filp = filp;
225 	entry->wait_address = wait_address;
226 	entry->key = p->key;
227 	init_waitqueue_func_entry(&entry->wait, pollwake);
228 	entry->wait.private = pwq;
229 	add_wait_queue(wait_address, &entry->wait);
230 }
231 
232 int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
233 			  ktime_t *expires, unsigned long slack)
234 {
235 	int rc = -EINTR;
236 
237 	set_current_state(state);
238 	if (!pwq->triggered)
239 		rc = schedule_hrtimeout_range(expires, slack, HRTIMER_MODE_ABS);
240 	__set_current_state(TASK_RUNNING);
241 
242 	/*
243 	 * Prepare for the next iteration.
244 	 *
245 	 * The following set_mb() serves two purposes.  First, it's
246 	 * the counterpart rmb of the wmb in pollwake() such that data
247 	 * written before wake up is always visible after wake up.
248 	 * Second, the full barrier guarantees that triggered clearing
249 	 * doesn't pass event check of the next iteration.  Note that
250 	 * this problem doesn't exist for the first iteration as
251 	 * add_wait_queue() has full barrier semantics.
252 	 */
253 	set_mb(pwq->triggered, 0);
254 
255 	return rc;
256 }
257 EXPORT_SYMBOL(poll_schedule_timeout);
258 
259 /**
260  * poll_select_set_timeout - helper function to setup the timeout value
261  * @to:		pointer to timespec variable for the final timeout
262  * @sec:	seconds (from user space)
263  * @nsec:	nanoseconds (from user space)
264  *
265  * Note, we do not use a timespec for the user space value here, That
266  * way we can use the function for timeval and compat interfaces as well.
267  *
268  * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
269  */
270 int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
271 {
272 	struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
273 
274 	if (!timespec_valid(&ts))
275 		return -EINVAL;
276 
277 	/* Optimize for the zero timeout value here */
278 	if (!sec && !nsec) {
279 		to->tv_sec = to->tv_nsec = 0;
280 	} else {
281 		ktime_get_ts(to);
282 		*to = timespec_add_safe(*to, ts);
283 	}
284 	return 0;
285 }
286 
287 static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
288 				      int timeval, int ret)
289 {
290 	struct timespec rts;
291 	struct timeval rtv;
292 
293 	if (!p)
294 		return ret;
295 
296 	if (current->personality & STICKY_TIMEOUTS)
297 		goto sticky;
298 
299 	/* No update for zero timeout */
300 	if (!end_time->tv_sec && !end_time->tv_nsec)
301 		return ret;
302 
303 	ktime_get_ts(&rts);
304 	rts = timespec_sub(*end_time, rts);
305 	if (rts.tv_sec < 0)
306 		rts.tv_sec = rts.tv_nsec = 0;
307 
308 	if (timeval) {
309 		rtv.tv_sec = rts.tv_sec;
310 		rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
311 
312 		if (!copy_to_user(p, &rtv, sizeof(rtv)))
313 			return ret;
314 
315 	} else if (!copy_to_user(p, &rts, sizeof(rts)))
316 		return ret;
317 
318 	/*
319 	 * If an application puts its timeval in read-only memory, we
320 	 * don't want the Linux-specific update to the timeval to
321 	 * cause a fault after the select has completed
322 	 * successfully. However, because we're not updating the
323 	 * timeval, we can't restart the system call.
324 	 */
325 
326 sticky:
327 	if (ret == -ERESTARTNOHAND)
328 		ret = -EINTR;
329 	return ret;
330 }
331 
332 #define FDS_IN(fds, n)		(fds->in + n)
333 #define FDS_OUT(fds, n)		(fds->out + n)
334 #define FDS_EX(fds, n)		(fds->ex + n)
335 
336 #define BITS(fds, n)	(*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
337 
338 static int max_select_fd(unsigned long n, fd_set_bits *fds)
339 {
340 	unsigned long *open_fds;
341 	unsigned long set;
342 	int max;
343 	struct fdtable *fdt;
344 
345 	/* handle last in-complete long-word first */
346 	set = ~(~0UL << (n & (__NFDBITS-1)));
347 	n /= __NFDBITS;
348 	fdt = files_fdtable(current->files);
349 	open_fds = fdt->open_fds->fds_bits+n;
350 	max = 0;
351 	if (set) {
352 		set &= BITS(fds, n);
353 		if (set) {
354 			if (!(set & ~*open_fds))
355 				goto get_max;
356 			return -EBADF;
357 		}
358 	}
359 	while (n) {
360 		open_fds--;
361 		n--;
362 		set = BITS(fds, n);
363 		if (!set)
364 			continue;
365 		if (set & ~*open_fds)
366 			return -EBADF;
367 		if (max)
368 			continue;
369 get_max:
370 		do {
371 			max++;
372 			set >>= 1;
373 		} while (set);
374 		max += n * __NFDBITS;
375 	}
376 
377 	return max;
378 }
379 
380 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
381 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
382 #define POLLEX_SET (POLLPRI)
383 
384 static inline void wait_key_set(poll_table *wait, unsigned long in,
385 				unsigned long out, unsigned long bit)
386 {
387 	if (wait) {
388 		wait->key = POLLEX_SET;
389 		if (in & bit)
390 			wait->key |= POLLIN_SET;
391 		if (out & bit)
392 			wait->key |= POLLOUT_SET;
393 	}
394 }
395 
396 int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
397 {
398 	ktime_t expire, *to = NULL;
399 	struct poll_wqueues table;
400 	poll_table *wait;
401 	int retval, i, timed_out = 0;
402 	unsigned long slack = 0;
403 
404 	rcu_read_lock();
405 	retval = max_select_fd(n, fds);
406 	rcu_read_unlock();
407 
408 	if (retval < 0)
409 		return retval;
410 	n = retval;
411 
412 	poll_initwait(&table);
413 	wait = &table.pt;
414 	if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
415 		wait = NULL;
416 		timed_out = 1;
417 	}
418 
419 	if (end_time && !timed_out)
420 		slack = estimate_accuracy(end_time);
421 
422 	retval = 0;
423 	for (;;) {
424 		unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
425 
426 		inp = fds->in; outp = fds->out; exp = fds->ex;
427 		rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
428 
429 		for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
430 			unsigned long in, out, ex, all_bits, bit = 1, mask, j;
431 			unsigned long res_in = 0, res_out = 0, res_ex = 0;
432 			const struct file_operations *f_op = NULL;
433 			struct file *file = NULL;
434 
435 			in = *inp++; out = *outp++; ex = *exp++;
436 			all_bits = in | out | ex;
437 			if (all_bits == 0) {
438 				i += __NFDBITS;
439 				continue;
440 			}
441 
442 			for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
443 				int fput_needed;
444 				if (i >= n)
445 					break;
446 				if (!(bit & all_bits))
447 					continue;
448 				file = fget_light(i, &fput_needed);
449 				if (file) {
450 					f_op = file->f_op;
451 					mask = DEFAULT_POLLMASK;
452 					if (f_op && f_op->poll) {
453 						wait_key_set(wait, in, out, bit);
454 						mask = (*f_op->poll)(file, wait);
455 					}
456 					fput_light(file, fput_needed);
457 					if ((mask & POLLIN_SET) && (in & bit)) {
458 						res_in |= bit;
459 						retval++;
460 						wait = NULL;
461 					}
462 					if ((mask & POLLOUT_SET) && (out & bit)) {
463 						res_out |= bit;
464 						retval++;
465 						wait = NULL;
466 					}
467 					if ((mask & POLLEX_SET) && (ex & bit)) {
468 						res_ex |= bit;
469 						retval++;
470 						wait = NULL;
471 					}
472 				}
473 			}
474 			if (res_in)
475 				*rinp = res_in;
476 			if (res_out)
477 				*routp = res_out;
478 			if (res_ex)
479 				*rexp = res_ex;
480 			cond_resched();
481 		}
482 		wait = NULL;
483 		if (retval || timed_out || signal_pending(current))
484 			break;
485 		if (table.error) {
486 			retval = table.error;
487 			break;
488 		}
489 
490 		/*
491 		 * If this is the first loop and we have a timeout
492 		 * given, then we convert to ktime_t and set the to
493 		 * pointer to the expiry value.
494 		 */
495 		if (end_time && !to) {
496 			expire = timespec_to_ktime(*end_time);
497 			to = &expire;
498 		}
499 
500 		if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
501 					   to, slack))
502 			timed_out = 1;
503 	}
504 
505 	poll_freewait(&table);
506 
507 	return retval;
508 }
509 
510 /*
511  * We can actually return ERESTARTSYS instead of EINTR, but I'd
512  * like to be certain this leads to no problems. So I return
513  * EINTR just for safety.
514  *
515  * Update: ERESTARTSYS breaks at least the xview clock binary, so
516  * I'm trying ERESTARTNOHAND which restart only when you want to.
517  */
518 #define MAX_SELECT_SECONDS \
519 	((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
520 
521 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
522 			   fd_set __user *exp, struct timespec *end_time)
523 {
524 	fd_set_bits fds;
525 	void *bits;
526 	int ret, max_fds;
527 	unsigned int size;
528 	struct fdtable *fdt;
529 	/* Allocate small arguments on the stack to save memory and be faster */
530 	long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
531 
532 	ret = -EINVAL;
533 	if (n < 0)
534 		goto out_nofds;
535 
536 	/* max_fds can increase, so grab it once to avoid race */
537 	rcu_read_lock();
538 	fdt = files_fdtable(current->files);
539 	max_fds = fdt->max_fds;
540 	rcu_read_unlock();
541 	if (n > max_fds)
542 		n = max_fds;
543 
544 	/*
545 	 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
546 	 * since we used fdset we need to allocate memory in units of
547 	 * long-words.
548 	 */
549 	size = FDS_BYTES(n);
550 	bits = stack_fds;
551 	if (size > sizeof(stack_fds) / 6) {
552 		/* Not enough space in on-stack array; must use kmalloc */
553 		ret = -ENOMEM;
554 		bits = kmalloc(6 * size, GFP_KERNEL);
555 		if (!bits)
556 			goto out_nofds;
557 	}
558 	fds.in      = bits;
559 	fds.out     = bits +   size;
560 	fds.ex      = bits + 2*size;
561 	fds.res_in  = bits + 3*size;
562 	fds.res_out = bits + 4*size;
563 	fds.res_ex  = bits + 5*size;
564 
565 	if ((ret = get_fd_set(n, inp, fds.in)) ||
566 	    (ret = get_fd_set(n, outp, fds.out)) ||
567 	    (ret = get_fd_set(n, exp, fds.ex)))
568 		goto out;
569 	zero_fd_set(n, fds.res_in);
570 	zero_fd_set(n, fds.res_out);
571 	zero_fd_set(n, fds.res_ex);
572 
573 	ret = do_select(n, &fds, end_time);
574 
575 	if (ret < 0)
576 		goto out;
577 	if (!ret) {
578 		ret = -ERESTARTNOHAND;
579 		if (signal_pending(current))
580 			goto out;
581 		ret = 0;
582 	}
583 
584 	if (set_fd_set(n, inp, fds.res_in) ||
585 	    set_fd_set(n, outp, fds.res_out) ||
586 	    set_fd_set(n, exp, fds.res_ex))
587 		ret = -EFAULT;
588 
589 out:
590 	if (bits != stack_fds)
591 		kfree(bits);
592 out_nofds:
593 	return ret;
594 }
595 
596 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
597 		fd_set __user *, exp, struct timeval __user *, tvp)
598 {
599 	struct timespec end_time, *to = NULL;
600 	struct timeval tv;
601 	int ret;
602 
603 	if (tvp) {
604 		if (copy_from_user(&tv, tvp, sizeof(tv)))
605 			return -EFAULT;
606 
607 		to = &end_time;
608 		if (poll_select_set_timeout(to,
609 				tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
610 				(tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
611 			return -EINVAL;
612 	}
613 
614 	ret = core_sys_select(n, inp, outp, exp, to);
615 	ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
616 
617 	return ret;
618 }
619 
620 #ifdef HAVE_SET_RESTORE_SIGMASK
621 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
622 		       fd_set __user *exp, struct timespec __user *tsp,
623 		       const sigset_t __user *sigmask, size_t sigsetsize)
624 {
625 	sigset_t ksigmask, sigsaved;
626 	struct timespec ts, end_time, *to = NULL;
627 	int ret;
628 
629 	if (tsp) {
630 		if (copy_from_user(&ts, tsp, sizeof(ts)))
631 			return -EFAULT;
632 
633 		to = &end_time;
634 		if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
635 			return -EINVAL;
636 	}
637 
638 	if (sigmask) {
639 		/* XXX: Don't preclude handling different sized sigset_t's.  */
640 		if (sigsetsize != sizeof(sigset_t))
641 			return -EINVAL;
642 		if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
643 			return -EFAULT;
644 
645 		sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
646 		sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
647 	}
648 
649 	ret = core_sys_select(n, inp, outp, exp, to);
650 	ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
651 
652 	if (ret == -ERESTARTNOHAND) {
653 		/*
654 		 * Don't restore the signal mask yet. Let do_signal() deliver
655 		 * the signal on the way back to userspace, before the signal
656 		 * mask is restored.
657 		 */
658 		if (sigmask) {
659 			memcpy(&current->saved_sigmask, &sigsaved,
660 					sizeof(sigsaved));
661 			set_restore_sigmask();
662 		}
663 	} else if (sigmask)
664 		sigprocmask(SIG_SETMASK, &sigsaved, NULL);
665 
666 	return ret;
667 }
668 
669 /*
670  * Most architectures can't handle 7-argument syscalls. So we provide a
671  * 6-argument version where the sixth argument is a pointer to a structure
672  * which has a pointer to the sigset_t itself followed by a size_t containing
673  * the sigset size.
674  */
675 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
676 		fd_set __user *, exp, struct timespec __user *, tsp,
677 		void __user *, sig)
678 {
679 	size_t sigsetsize = 0;
680 	sigset_t __user *up = NULL;
681 
682 	if (sig) {
683 		if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
684 		    || __get_user(up, (sigset_t __user * __user *)sig)
685 		    || __get_user(sigsetsize,
686 				(size_t __user *)(sig+sizeof(void *))))
687 			return -EFAULT;
688 	}
689 
690 	return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
691 }
692 #endif /* HAVE_SET_RESTORE_SIGMASK */
693 
694 struct poll_list {
695 	struct poll_list *next;
696 	int len;
697 	struct pollfd entries[0];
698 };
699 
700 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
701 
702 /*
703  * Fish for pollable events on the pollfd->fd file descriptor. We're only
704  * interested in events matching the pollfd->events mask, and the result
705  * matching that mask is both recorded in pollfd->revents and returned. The
706  * pwait poll_table will be used by the fd-provided poll handler for waiting,
707  * if non-NULL.
708  */
709 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
710 {
711 	unsigned int mask;
712 	int fd;
713 
714 	mask = 0;
715 	fd = pollfd->fd;
716 	if (fd >= 0) {
717 		int fput_needed;
718 		struct file * file;
719 
720 		file = fget_light(fd, &fput_needed);
721 		mask = POLLNVAL;
722 		if (file != NULL) {
723 			mask = DEFAULT_POLLMASK;
724 			if (file->f_op && file->f_op->poll) {
725 				if (pwait)
726 					pwait->key = pollfd->events |
727 							POLLERR | POLLHUP;
728 				mask = file->f_op->poll(file, pwait);
729 			}
730 			/* Mask out unneeded events. */
731 			mask &= pollfd->events | POLLERR | POLLHUP;
732 			fput_light(file, fput_needed);
733 		}
734 	}
735 	pollfd->revents = mask;
736 
737 	return mask;
738 }
739 
740 static int do_poll(unsigned int nfds,  struct poll_list *list,
741 		   struct poll_wqueues *wait, struct timespec *end_time)
742 {
743 	poll_table* pt = &wait->pt;
744 	ktime_t expire, *to = NULL;
745 	int timed_out = 0, count = 0;
746 	unsigned long slack = 0;
747 
748 	/* Optimise the no-wait case */
749 	if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
750 		pt = NULL;
751 		timed_out = 1;
752 	}
753 
754 	if (end_time && !timed_out)
755 		slack = estimate_accuracy(end_time);
756 
757 	for (;;) {
758 		struct poll_list *walk;
759 
760 		for (walk = list; walk != NULL; walk = walk->next) {
761 			struct pollfd * pfd, * pfd_end;
762 
763 			pfd = walk->entries;
764 			pfd_end = pfd + walk->len;
765 			for (; pfd != pfd_end; pfd++) {
766 				/*
767 				 * Fish for events. If we found one, record it
768 				 * and kill the poll_table, so we don't
769 				 * needlessly register any other waiters after
770 				 * this. They'll get immediately deregistered
771 				 * when we break out and return.
772 				 */
773 				if (do_pollfd(pfd, pt)) {
774 					count++;
775 					pt = NULL;
776 				}
777 			}
778 		}
779 		/*
780 		 * All waiters have already been registered, so don't provide
781 		 * a poll_table to them on the next loop iteration.
782 		 */
783 		pt = NULL;
784 		if (!count) {
785 			count = wait->error;
786 			if (signal_pending(current))
787 				count = -EINTR;
788 		}
789 		if (count || timed_out)
790 			break;
791 
792 		/*
793 		 * If this is the first loop and we have a timeout
794 		 * given, then we convert to ktime_t and set the to
795 		 * pointer to the expiry value.
796 		 */
797 		if (end_time && !to) {
798 			expire = timespec_to_ktime(*end_time);
799 			to = &expire;
800 		}
801 
802 		if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
803 			timed_out = 1;
804 	}
805 	return count;
806 }
807 
808 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
809 			sizeof(struct pollfd))
810 
811 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
812 		struct timespec *end_time)
813 {
814 	struct poll_wqueues table;
815  	int err = -EFAULT, fdcount, len, size;
816 	/* Allocate small arguments on the stack to save memory and be
817 	   faster - use long to make sure the buffer is aligned properly
818 	   on 64 bit archs to avoid unaligned access */
819 	long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
820 	struct poll_list *const head = (struct poll_list *)stack_pps;
821  	struct poll_list *walk = head;
822  	unsigned long todo = nfds;
823 
824 	if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
825 		return -EINVAL;
826 
827 	len = min_t(unsigned int, nfds, N_STACK_PPS);
828 	for (;;) {
829 		walk->next = NULL;
830 		walk->len = len;
831 		if (!len)
832 			break;
833 
834 		if (copy_from_user(walk->entries, ufds + nfds-todo,
835 					sizeof(struct pollfd) * walk->len))
836 			goto out_fds;
837 
838 		todo -= walk->len;
839 		if (!todo)
840 			break;
841 
842 		len = min(todo, POLLFD_PER_PAGE);
843 		size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
844 		walk = walk->next = kmalloc(size, GFP_KERNEL);
845 		if (!walk) {
846 			err = -ENOMEM;
847 			goto out_fds;
848 		}
849 	}
850 
851 	poll_initwait(&table);
852 	fdcount = do_poll(nfds, head, &table, end_time);
853 	poll_freewait(&table);
854 
855 	for (walk = head; walk; walk = walk->next) {
856 		struct pollfd *fds = walk->entries;
857 		int j;
858 
859 		for (j = 0; j < walk->len; j++, ufds++)
860 			if (__put_user(fds[j].revents, &ufds->revents))
861 				goto out_fds;
862   	}
863 
864 	err = fdcount;
865 out_fds:
866 	walk = head->next;
867 	while (walk) {
868 		struct poll_list *pos = walk;
869 		walk = walk->next;
870 		kfree(pos);
871 	}
872 
873 	return err;
874 }
875 
876 static long do_restart_poll(struct restart_block *restart_block)
877 {
878 	struct pollfd __user *ufds = restart_block->poll.ufds;
879 	int nfds = restart_block->poll.nfds;
880 	struct timespec *to = NULL, end_time;
881 	int ret;
882 
883 	if (restart_block->poll.has_timeout) {
884 		end_time.tv_sec = restart_block->poll.tv_sec;
885 		end_time.tv_nsec = restart_block->poll.tv_nsec;
886 		to = &end_time;
887 	}
888 
889 	ret = do_sys_poll(ufds, nfds, to);
890 
891 	if (ret == -EINTR) {
892 		restart_block->fn = do_restart_poll;
893 		ret = -ERESTART_RESTARTBLOCK;
894 	}
895 	return ret;
896 }
897 
898 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
899 		long, timeout_msecs)
900 {
901 	struct timespec end_time, *to = NULL;
902 	int ret;
903 
904 	if (timeout_msecs >= 0) {
905 		to = &end_time;
906 		poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
907 			NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
908 	}
909 
910 	ret = do_sys_poll(ufds, nfds, to);
911 
912 	if (ret == -EINTR) {
913 		struct restart_block *restart_block;
914 
915 		restart_block = &current_thread_info()->restart_block;
916 		restart_block->fn = do_restart_poll;
917 		restart_block->poll.ufds = ufds;
918 		restart_block->poll.nfds = nfds;
919 
920 		if (timeout_msecs >= 0) {
921 			restart_block->poll.tv_sec = end_time.tv_sec;
922 			restart_block->poll.tv_nsec = end_time.tv_nsec;
923 			restart_block->poll.has_timeout = 1;
924 		} else
925 			restart_block->poll.has_timeout = 0;
926 
927 		ret = -ERESTART_RESTARTBLOCK;
928 	}
929 	return ret;
930 }
931 
932 #ifdef HAVE_SET_RESTORE_SIGMASK
933 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
934 		struct timespec __user *, tsp, const sigset_t __user *, sigmask,
935 		size_t, sigsetsize)
936 {
937 	sigset_t ksigmask, sigsaved;
938 	struct timespec ts, end_time, *to = NULL;
939 	int ret;
940 
941 	if (tsp) {
942 		if (copy_from_user(&ts, tsp, sizeof(ts)))
943 			return -EFAULT;
944 
945 		to = &end_time;
946 		if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
947 			return -EINVAL;
948 	}
949 
950 	if (sigmask) {
951 		/* XXX: Don't preclude handling different sized sigset_t's.  */
952 		if (sigsetsize != sizeof(sigset_t))
953 			return -EINVAL;
954 		if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
955 			return -EFAULT;
956 
957 		sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
958 		sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
959 	}
960 
961 	ret = do_sys_poll(ufds, nfds, to);
962 
963 	/* We can restart this syscall, usually */
964 	if (ret == -EINTR) {
965 		/*
966 		 * Don't restore the signal mask yet. Let do_signal() deliver
967 		 * the signal on the way back to userspace, before the signal
968 		 * mask is restored.
969 		 */
970 		if (sigmask) {
971 			memcpy(&current->saved_sigmask, &sigsaved,
972 					sizeof(sigsaved));
973 			set_restore_sigmask();
974 		}
975 		ret = -ERESTARTNOHAND;
976 	} else if (sigmask)
977 		sigprocmask(SIG_SETMASK, &sigsaved, NULL);
978 
979 	ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
980 
981 	return ret;
982 }
983 #endif /* HAVE_SET_RESTORE_SIGMASK */
984