xref: /openbmc/linux/fs/select.c (revision 8cb5d748)
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/signal.h>
19 #include <linux/sched/rt.h>
20 #include <linux/syscalls.h>
21 #include <linux/export.h>
22 #include <linux/slab.h>
23 #include <linux/poll.h>
24 #include <linux/personality.h> /* for STICKY_TIMEOUTS */
25 #include <linux/file.h>
26 #include <linux/fdtable.h>
27 #include <linux/fs.h>
28 #include <linux/rcupdate.h>
29 #include <linux/hrtimer.h>
30 #include <linux/freezer.h>
31 #include <net/busy_poll.h>
32 #include <linux/vmalloc.h>
33 
34 #include <linux/uaccess.h>
35 
36 
37 /*
38  * Estimate expected accuracy in ns from a timeval.
39  *
40  * After quite a bit of churning around, we've settled on
41  * a simple thing of taking 0.1% of the timeout as the
42  * slack, with a cap of 100 msec.
43  * "nice" tasks get a 0.5% slack instead.
44  *
45  * Consider this comment an open invitation to come up with even
46  * better solutions..
47  */
48 
49 #define MAX_SLACK	(100 * NSEC_PER_MSEC)
50 
51 static long __estimate_accuracy(struct timespec64 *tv)
52 {
53 	long slack;
54 	int divfactor = 1000;
55 
56 	if (tv->tv_sec < 0)
57 		return 0;
58 
59 	if (task_nice(current) > 0)
60 		divfactor = divfactor / 5;
61 
62 	if (tv->tv_sec > MAX_SLACK / (NSEC_PER_SEC/divfactor))
63 		return MAX_SLACK;
64 
65 	slack = tv->tv_nsec / divfactor;
66 	slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
67 
68 	if (slack > MAX_SLACK)
69 		return MAX_SLACK;
70 
71 	return slack;
72 }
73 
74 u64 select_estimate_accuracy(struct timespec64 *tv)
75 {
76 	u64 ret;
77 	struct timespec64 now;
78 
79 	/*
80 	 * Realtime tasks get a slack of 0 for obvious reasons.
81 	 */
82 
83 	if (rt_task(current))
84 		return 0;
85 
86 	ktime_get_ts64(&now);
87 	now = timespec64_sub(*tv, now);
88 	ret = __estimate_accuracy(&now);
89 	if (ret < current->timer_slack_ns)
90 		return current->timer_slack_ns;
91 	return ret;
92 }
93 
94 
95 
96 struct poll_table_page {
97 	struct poll_table_page * next;
98 	struct poll_table_entry * entry;
99 	struct poll_table_entry entries[0];
100 };
101 
102 #define POLL_TABLE_FULL(table) \
103 	((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
104 
105 /*
106  * Ok, Peter made a complicated, but straightforward multiple_wait() function.
107  * I have rewritten this, taking some shortcuts: This code may not be easy to
108  * follow, but it should be free of race-conditions, and it's practical. If you
109  * understand what I'm doing here, then you understand how the linux
110  * sleep/wakeup mechanism works.
111  *
112  * Two very simple procedures, poll_wait() and poll_freewait() make all the
113  * work.  poll_wait() is an inline-function defined in <linux/poll.h>,
114  * as all select/poll functions have to call it to add an entry to the
115  * poll table.
116  */
117 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
118 		       poll_table *p);
119 
120 void poll_initwait(struct poll_wqueues *pwq)
121 {
122 	init_poll_funcptr(&pwq->pt, __pollwait);
123 	pwq->polling_task = current;
124 	pwq->triggered = 0;
125 	pwq->error = 0;
126 	pwq->table = NULL;
127 	pwq->inline_index = 0;
128 }
129 EXPORT_SYMBOL(poll_initwait);
130 
131 static void free_poll_entry(struct poll_table_entry *entry)
132 {
133 	remove_wait_queue(entry->wait_address, &entry->wait);
134 	fput(entry->filp);
135 }
136 
137 void poll_freewait(struct poll_wqueues *pwq)
138 {
139 	struct poll_table_page * p = pwq->table;
140 	int i;
141 	for (i = 0; i < pwq->inline_index; i++)
142 		free_poll_entry(pwq->inline_entries + i);
143 	while (p) {
144 		struct poll_table_entry * entry;
145 		struct poll_table_page *old;
146 
147 		entry = p->entry;
148 		do {
149 			entry--;
150 			free_poll_entry(entry);
151 		} while (entry > p->entries);
152 		old = p;
153 		p = p->next;
154 		free_page((unsigned long) old);
155 	}
156 }
157 EXPORT_SYMBOL(poll_freewait);
158 
159 static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
160 {
161 	struct poll_table_page *table = p->table;
162 
163 	if (p->inline_index < N_INLINE_POLL_ENTRIES)
164 		return p->inline_entries + p->inline_index++;
165 
166 	if (!table || POLL_TABLE_FULL(table)) {
167 		struct poll_table_page *new_table;
168 
169 		new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
170 		if (!new_table) {
171 			p->error = -ENOMEM;
172 			return NULL;
173 		}
174 		new_table->entry = new_table->entries;
175 		new_table->next = table;
176 		p->table = new_table;
177 		table = new_table;
178 	}
179 
180 	return table->entry++;
181 }
182 
183 static int __pollwake(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
184 {
185 	struct poll_wqueues *pwq = wait->private;
186 	DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
187 
188 	/*
189 	 * Although this function is called under waitqueue lock, LOCK
190 	 * doesn't imply write barrier and the users expect write
191 	 * barrier semantics on wakeup functions.  The following
192 	 * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
193 	 * and is paired with smp_store_mb() in poll_schedule_timeout.
194 	 */
195 	smp_wmb();
196 	pwq->triggered = 1;
197 
198 	/*
199 	 * Perform the default wake up operation using a dummy
200 	 * waitqueue.
201 	 *
202 	 * TODO: This is hacky but there currently is no interface to
203 	 * pass in @sync.  @sync is scheduled to be removed and once
204 	 * that happens, wake_up_process() can be used directly.
205 	 */
206 	return default_wake_function(&dummy_wait, mode, sync, key);
207 }
208 
209 static int pollwake(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
210 {
211 	struct poll_table_entry *entry;
212 
213 	entry = container_of(wait, struct poll_table_entry, wait);
214 	if (key && !((unsigned long)key & entry->key))
215 		return 0;
216 	return __pollwake(wait, mode, sync, key);
217 }
218 
219 /* Add a new entry */
220 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
221 				poll_table *p)
222 {
223 	struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
224 	struct poll_table_entry *entry = poll_get_entry(pwq);
225 	if (!entry)
226 		return;
227 	entry->filp = get_file(filp);
228 	entry->wait_address = wait_address;
229 	entry->key = p->_key;
230 	init_waitqueue_func_entry(&entry->wait, pollwake);
231 	entry->wait.private = pwq;
232 	add_wait_queue(wait_address, &entry->wait);
233 }
234 
235 int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
236 			  ktime_t *expires, unsigned long slack)
237 {
238 	int rc = -EINTR;
239 
240 	set_current_state(state);
241 	if (!pwq->triggered)
242 		rc = schedule_hrtimeout_range(expires, slack, HRTIMER_MODE_ABS);
243 	__set_current_state(TASK_RUNNING);
244 
245 	/*
246 	 * Prepare for the next iteration.
247 	 *
248 	 * The following smp_store_mb() serves two purposes.  First, it's
249 	 * the counterpart rmb of the wmb in pollwake() such that data
250 	 * written before wake up is always visible after wake up.
251 	 * Second, the full barrier guarantees that triggered clearing
252 	 * doesn't pass event check of the next iteration.  Note that
253 	 * this problem doesn't exist for the first iteration as
254 	 * add_wait_queue() has full barrier semantics.
255 	 */
256 	smp_store_mb(pwq->triggered, 0);
257 
258 	return rc;
259 }
260 EXPORT_SYMBOL(poll_schedule_timeout);
261 
262 /**
263  * poll_select_set_timeout - helper function to setup the timeout value
264  * @to:		pointer to timespec64 variable for the final timeout
265  * @sec:	seconds (from user space)
266  * @nsec:	nanoseconds (from user space)
267  *
268  * Note, we do not use a timespec for the user space value here, That
269  * way we can use the function for timeval and compat interfaces as well.
270  *
271  * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
272  */
273 int poll_select_set_timeout(struct timespec64 *to, time64_t sec, long nsec)
274 {
275 	struct timespec64 ts = {.tv_sec = sec, .tv_nsec = nsec};
276 
277 	if (!timespec64_valid(&ts))
278 		return -EINVAL;
279 
280 	/* Optimize for the zero timeout value here */
281 	if (!sec && !nsec) {
282 		to->tv_sec = to->tv_nsec = 0;
283 	} else {
284 		ktime_get_ts64(to);
285 		*to = timespec64_add_safe(*to, ts);
286 	}
287 	return 0;
288 }
289 
290 static int poll_select_copy_remaining(struct timespec64 *end_time,
291 				      void __user *p,
292 				      int timeval, int ret)
293 {
294 	struct timespec64 rts64;
295 	struct timespec rts;
296 	struct timeval rtv;
297 
298 	if (!p)
299 		return ret;
300 
301 	if (current->personality & STICKY_TIMEOUTS)
302 		goto sticky;
303 
304 	/* No update for zero timeout */
305 	if (!end_time->tv_sec && !end_time->tv_nsec)
306 		return ret;
307 
308 	ktime_get_ts64(&rts64);
309 	rts64 = timespec64_sub(*end_time, rts64);
310 	if (rts64.tv_sec < 0)
311 		rts64.tv_sec = rts64.tv_nsec = 0;
312 
313 	rts = timespec64_to_timespec(rts64);
314 
315 	if (timeval) {
316 		if (sizeof(rtv) > sizeof(rtv.tv_sec) + sizeof(rtv.tv_usec))
317 			memset(&rtv, 0, sizeof(rtv));
318 		rtv.tv_sec = rts64.tv_sec;
319 		rtv.tv_usec = rts64.tv_nsec / NSEC_PER_USEC;
320 
321 		if (!copy_to_user(p, &rtv, sizeof(rtv)))
322 			return ret;
323 
324 	} else if (!copy_to_user(p, &rts, sizeof(rts)))
325 		return ret;
326 
327 	/*
328 	 * If an application puts its timeval in read-only memory, we
329 	 * don't want the Linux-specific update to the timeval to
330 	 * cause a fault after the select has completed
331 	 * successfully. However, because we're not updating the
332 	 * timeval, we can't restart the system call.
333 	 */
334 
335 sticky:
336 	if (ret == -ERESTARTNOHAND)
337 		ret = -EINTR;
338 	return ret;
339 }
340 
341 /*
342  * Scalable version of the fd_set.
343  */
344 
345 typedef struct {
346 	unsigned long *in, *out, *ex;
347 	unsigned long *res_in, *res_out, *res_ex;
348 } fd_set_bits;
349 
350 /*
351  * How many longwords for "nr" bits?
352  */
353 #define FDS_BITPERLONG	(8*sizeof(long))
354 #define FDS_LONGS(nr)	(((nr)+FDS_BITPERLONG-1)/FDS_BITPERLONG)
355 #define FDS_BYTES(nr)	(FDS_LONGS(nr)*sizeof(long))
356 
357 /*
358  * We do a VERIFY_WRITE here even though we are only reading this time:
359  * we'll write to it eventually..
360  *
361  * Use "unsigned long" accesses to let user-mode fd_set's be long-aligned.
362  */
363 static inline
364 int get_fd_set(unsigned long nr, void __user *ufdset, unsigned long *fdset)
365 {
366 	nr = FDS_BYTES(nr);
367 	if (ufdset)
368 		return copy_from_user(fdset, ufdset, nr) ? -EFAULT : 0;
369 
370 	memset(fdset, 0, nr);
371 	return 0;
372 }
373 
374 static inline unsigned long __must_check
375 set_fd_set(unsigned long nr, void __user *ufdset, unsigned long *fdset)
376 {
377 	if (ufdset)
378 		return __copy_to_user(ufdset, fdset, FDS_BYTES(nr));
379 	return 0;
380 }
381 
382 static inline
383 void zero_fd_set(unsigned long nr, unsigned long *fdset)
384 {
385 	memset(fdset, 0, FDS_BYTES(nr));
386 }
387 
388 #define FDS_IN(fds, n)		(fds->in + n)
389 #define FDS_OUT(fds, n)		(fds->out + n)
390 #define FDS_EX(fds, n)		(fds->ex + n)
391 
392 #define BITS(fds, n)	(*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
393 
394 static int max_select_fd(unsigned long n, fd_set_bits *fds)
395 {
396 	unsigned long *open_fds;
397 	unsigned long set;
398 	int max;
399 	struct fdtable *fdt;
400 
401 	/* handle last in-complete long-word first */
402 	set = ~(~0UL << (n & (BITS_PER_LONG-1)));
403 	n /= BITS_PER_LONG;
404 	fdt = files_fdtable(current->files);
405 	open_fds = fdt->open_fds + n;
406 	max = 0;
407 	if (set) {
408 		set &= BITS(fds, n);
409 		if (set) {
410 			if (!(set & ~*open_fds))
411 				goto get_max;
412 			return -EBADF;
413 		}
414 	}
415 	while (n) {
416 		open_fds--;
417 		n--;
418 		set = BITS(fds, n);
419 		if (!set)
420 			continue;
421 		if (set & ~*open_fds)
422 			return -EBADF;
423 		if (max)
424 			continue;
425 get_max:
426 		do {
427 			max++;
428 			set >>= 1;
429 		} while (set);
430 		max += n * BITS_PER_LONG;
431 	}
432 
433 	return max;
434 }
435 
436 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
437 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
438 #define POLLEX_SET (POLLPRI)
439 
440 static inline void wait_key_set(poll_table *wait, unsigned long in,
441 				unsigned long out, unsigned long bit,
442 				unsigned int ll_flag)
443 {
444 	wait->_key = POLLEX_SET | ll_flag;
445 	if (in & bit)
446 		wait->_key |= POLLIN_SET;
447 	if (out & bit)
448 		wait->_key |= POLLOUT_SET;
449 }
450 
451 static int do_select(int n, fd_set_bits *fds, struct timespec64 *end_time)
452 {
453 	ktime_t expire, *to = NULL;
454 	struct poll_wqueues table;
455 	poll_table *wait;
456 	int retval, i, timed_out = 0;
457 	u64 slack = 0;
458 	unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
459 	unsigned long busy_start = 0;
460 
461 	rcu_read_lock();
462 	retval = max_select_fd(n, fds);
463 	rcu_read_unlock();
464 
465 	if (retval < 0)
466 		return retval;
467 	n = retval;
468 
469 	poll_initwait(&table);
470 	wait = &table.pt;
471 	if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
472 		wait->_qproc = NULL;
473 		timed_out = 1;
474 	}
475 
476 	if (end_time && !timed_out)
477 		slack = select_estimate_accuracy(end_time);
478 
479 	retval = 0;
480 	for (;;) {
481 		unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
482 		bool can_busy_loop = false;
483 
484 		inp = fds->in; outp = fds->out; exp = fds->ex;
485 		rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
486 
487 		for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
488 			unsigned long in, out, ex, all_bits, bit = 1, mask, j;
489 			unsigned long res_in = 0, res_out = 0, res_ex = 0;
490 
491 			in = *inp++; out = *outp++; ex = *exp++;
492 			all_bits = in | out | ex;
493 			if (all_bits == 0) {
494 				i += BITS_PER_LONG;
495 				continue;
496 			}
497 
498 			for (j = 0; j < BITS_PER_LONG; ++j, ++i, bit <<= 1) {
499 				struct fd f;
500 				if (i >= n)
501 					break;
502 				if (!(bit & all_bits))
503 					continue;
504 				f = fdget(i);
505 				if (f.file) {
506 					const struct file_operations *f_op;
507 					f_op = f.file->f_op;
508 					mask = DEFAULT_POLLMASK;
509 					if (f_op->poll) {
510 						wait_key_set(wait, in, out,
511 							     bit, busy_flag);
512 						mask = (*f_op->poll)(f.file, wait);
513 					}
514 					fdput(f);
515 					if ((mask & POLLIN_SET) && (in & bit)) {
516 						res_in |= bit;
517 						retval++;
518 						wait->_qproc = NULL;
519 					}
520 					if ((mask & POLLOUT_SET) && (out & bit)) {
521 						res_out |= bit;
522 						retval++;
523 						wait->_qproc = NULL;
524 					}
525 					if ((mask & POLLEX_SET) && (ex & bit)) {
526 						res_ex |= bit;
527 						retval++;
528 						wait->_qproc = NULL;
529 					}
530 					/* got something, stop busy polling */
531 					if (retval) {
532 						can_busy_loop = false;
533 						busy_flag = 0;
534 
535 					/*
536 					 * only remember a returned
537 					 * POLL_BUSY_LOOP if we asked for it
538 					 */
539 					} else if (busy_flag & mask)
540 						can_busy_loop = true;
541 
542 				}
543 			}
544 			if (res_in)
545 				*rinp = res_in;
546 			if (res_out)
547 				*routp = res_out;
548 			if (res_ex)
549 				*rexp = res_ex;
550 			cond_resched();
551 		}
552 		wait->_qproc = NULL;
553 		if (retval || timed_out || signal_pending(current))
554 			break;
555 		if (table.error) {
556 			retval = table.error;
557 			break;
558 		}
559 
560 		/* only if found POLL_BUSY_LOOP sockets && not out of time */
561 		if (can_busy_loop && !need_resched()) {
562 			if (!busy_start) {
563 				busy_start = busy_loop_current_time();
564 				continue;
565 			}
566 			if (!busy_loop_timeout(busy_start))
567 				continue;
568 		}
569 		busy_flag = 0;
570 
571 		/*
572 		 * If this is the first loop and we have a timeout
573 		 * given, then we convert to ktime_t and set the to
574 		 * pointer to the expiry value.
575 		 */
576 		if (end_time && !to) {
577 			expire = timespec64_to_ktime(*end_time);
578 			to = &expire;
579 		}
580 
581 		if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
582 					   to, slack))
583 			timed_out = 1;
584 	}
585 
586 	poll_freewait(&table);
587 
588 	return retval;
589 }
590 
591 /*
592  * We can actually return ERESTARTSYS instead of EINTR, but I'd
593  * like to be certain this leads to no problems. So I return
594  * EINTR just for safety.
595  *
596  * Update: ERESTARTSYS breaks at least the xview clock binary, so
597  * I'm trying ERESTARTNOHAND which restart only when you want to.
598  */
599 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
600 			   fd_set __user *exp, struct timespec64 *end_time)
601 {
602 	fd_set_bits fds;
603 	void *bits;
604 	int ret, max_fds;
605 	size_t size, alloc_size;
606 	struct fdtable *fdt;
607 	/* Allocate small arguments on the stack to save memory and be faster */
608 	long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
609 
610 	ret = -EINVAL;
611 	if (n < 0)
612 		goto out_nofds;
613 
614 	/* max_fds can increase, so grab it once to avoid race */
615 	rcu_read_lock();
616 	fdt = files_fdtable(current->files);
617 	max_fds = fdt->max_fds;
618 	rcu_read_unlock();
619 	if (n > max_fds)
620 		n = max_fds;
621 
622 	/*
623 	 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
624 	 * since we used fdset we need to allocate memory in units of
625 	 * long-words.
626 	 */
627 	size = FDS_BYTES(n);
628 	bits = stack_fds;
629 	if (size > sizeof(stack_fds) / 6) {
630 		/* Not enough space in on-stack array; must use kmalloc */
631 		ret = -ENOMEM;
632 		if (size > (SIZE_MAX / 6))
633 			goto out_nofds;
634 
635 		alloc_size = 6 * size;
636 		bits = kvmalloc(alloc_size, GFP_KERNEL);
637 		if (!bits)
638 			goto out_nofds;
639 	}
640 	fds.in      = bits;
641 	fds.out     = bits +   size;
642 	fds.ex      = bits + 2*size;
643 	fds.res_in  = bits + 3*size;
644 	fds.res_out = bits + 4*size;
645 	fds.res_ex  = bits + 5*size;
646 
647 	if ((ret = get_fd_set(n, inp, fds.in)) ||
648 	    (ret = get_fd_set(n, outp, fds.out)) ||
649 	    (ret = get_fd_set(n, exp, fds.ex)))
650 		goto out;
651 	zero_fd_set(n, fds.res_in);
652 	zero_fd_set(n, fds.res_out);
653 	zero_fd_set(n, fds.res_ex);
654 
655 	ret = do_select(n, &fds, end_time);
656 
657 	if (ret < 0)
658 		goto out;
659 	if (!ret) {
660 		ret = -ERESTARTNOHAND;
661 		if (signal_pending(current))
662 			goto out;
663 		ret = 0;
664 	}
665 
666 	if (set_fd_set(n, inp, fds.res_in) ||
667 	    set_fd_set(n, outp, fds.res_out) ||
668 	    set_fd_set(n, exp, fds.res_ex))
669 		ret = -EFAULT;
670 
671 out:
672 	if (bits != stack_fds)
673 		kvfree(bits);
674 out_nofds:
675 	return ret;
676 }
677 
678 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
679 		fd_set __user *, exp, struct timeval __user *, tvp)
680 {
681 	struct timespec64 end_time, *to = NULL;
682 	struct timeval tv;
683 	int ret;
684 
685 	if (tvp) {
686 		if (copy_from_user(&tv, tvp, sizeof(tv)))
687 			return -EFAULT;
688 
689 		to = &end_time;
690 		if (poll_select_set_timeout(to,
691 				tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
692 				(tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
693 			return -EINVAL;
694 	}
695 
696 	ret = core_sys_select(n, inp, outp, exp, to);
697 	ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
698 
699 	return ret;
700 }
701 
702 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
703 		       fd_set __user *exp, struct timespec __user *tsp,
704 		       const sigset_t __user *sigmask, size_t sigsetsize)
705 {
706 	sigset_t ksigmask, sigsaved;
707 	struct timespec ts;
708 	struct timespec64 ts64, end_time, *to = NULL;
709 	int ret;
710 
711 	if (tsp) {
712 		if (copy_from_user(&ts, tsp, sizeof(ts)))
713 			return -EFAULT;
714 		ts64 = timespec_to_timespec64(ts);
715 
716 		to = &end_time;
717 		if (poll_select_set_timeout(to, ts64.tv_sec, ts64.tv_nsec))
718 			return -EINVAL;
719 	}
720 
721 	if (sigmask) {
722 		/* XXX: Don't preclude handling different sized sigset_t's.  */
723 		if (sigsetsize != sizeof(sigset_t))
724 			return -EINVAL;
725 		if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
726 			return -EFAULT;
727 
728 		sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
729 		sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
730 	}
731 
732 	ret = core_sys_select(n, inp, outp, exp, to);
733 	ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
734 
735 	if (ret == -ERESTARTNOHAND) {
736 		/*
737 		 * Don't restore the signal mask yet. Let do_signal() deliver
738 		 * the signal on the way back to userspace, before the signal
739 		 * mask is restored.
740 		 */
741 		if (sigmask) {
742 			memcpy(&current->saved_sigmask, &sigsaved,
743 					sizeof(sigsaved));
744 			set_restore_sigmask();
745 		}
746 	} else if (sigmask)
747 		sigprocmask(SIG_SETMASK, &sigsaved, NULL);
748 
749 	return ret;
750 }
751 
752 /*
753  * Most architectures can't handle 7-argument syscalls. So we provide a
754  * 6-argument version where the sixth argument is a pointer to a structure
755  * which has a pointer to the sigset_t itself followed by a size_t containing
756  * the sigset size.
757  */
758 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
759 		fd_set __user *, exp, struct timespec __user *, tsp,
760 		void __user *, sig)
761 {
762 	size_t sigsetsize = 0;
763 	sigset_t __user *up = NULL;
764 
765 	if (sig) {
766 		if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
767 		    || __get_user(up, (sigset_t __user * __user *)sig)
768 		    || __get_user(sigsetsize,
769 				(size_t __user *)(sig+sizeof(void *))))
770 			return -EFAULT;
771 	}
772 
773 	return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
774 }
775 
776 #ifdef __ARCH_WANT_SYS_OLD_SELECT
777 struct sel_arg_struct {
778 	unsigned long n;
779 	fd_set __user *inp, *outp, *exp;
780 	struct timeval __user *tvp;
781 };
782 
783 SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg)
784 {
785 	struct sel_arg_struct a;
786 
787 	if (copy_from_user(&a, arg, sizeof(a)))
788 		return -EFAULT;
789 	return sys_select(a.n, a.inp, a.outp, a.exp, a.tvp);
790 }
791 #endif
792 
793 struct poll_list {
794 	struct poll_list *next;
795 	int len;
796 	struct pollfd entries[0];
797 };
798 
799 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
800 
801 /*
802  * Fish for pollable events on the pollfd->fd file descriptor. We're only
803  * interested in events matching the pollfd->events mask, and the result
804  * matching that mask is both recorded in pollfd->revents and returned. The
805  * pwait poll_table will be used by the fd-provided poll handler for waiting,
806  * if pwait->_qproc is non-NULL.
807  */
808 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait,
809 				     bool *can_busy_poll,
810 				     unsigned int busy_flag)
811 {
812 	unsigned int mask;
813 	int fd;
814 
815 	mask = 0;
816 	fd = pollfd->fd;
817 	if (fd >= 0) {
818 		struct fd f = fdget(fd);
819 		mask = POLLNVAL;
820 		if (f.file) {
821 			mask = DEFAULT_POLLMASK;
822 			if (f.file->f_op->poll) {
823 				pwait->_key = pollfd->events|POLLERR|POLLHUP;
824 				pwait->_key |= busy_flag;
825 				mask = f.file->f_op->poll(f.file, pwait);
826 				if (mask & busy_flag)
827 					*can_busy_poll = true;
828 			}
829 			/* Mask out unneeded events. */
830 			mask &= pollfd->events | POLLERR | POLLHUP;
831 			fdput(f);
832 		}
833 	}
834 	pollfd->revents = mask;
835 
836 	return mask;
837 }
838 
839 static int do_poll(struct poll_list *list, struct poll_wqueues *wait,
840 		   struct timespec64 *end_time)
841 {
842 	poll_table* pt = &wait->pt;
843 	ktime_t expire, *to = NULL;
844 	int timed_out = 0, count = 0;
845 	u64 slack = 0;
846 	unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
847 	unsigned long busy_start = 0;
848 
849 	/* Optimise the no-wait case */
850 	if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
851 		pt->_qproc = NULL;
852 		timed_out = 1;
853 	}
854 
855 	if (end_time && !timed_out)
856 		slack = select_estimate_accuracy(end_time);
857 
858 	for (;;) {
859 		struct poll_list *walk;
860 		bool can_busy_loop = false;
861 
862 		for (walk = list; walk != NULL; walk = walk->next) {
863 			struct pollfd * pfd, * pfd_end;
864 
865 			pfd = walk->entries;
866 			pfd_end = pfd + walk->len;
867 			for (; pfd != pfd_end; pfd++) {
868 				/*
869 				 * Fish for events. If we found one, record it
870 				 * and kill poll_table->_qproc, so we don't
871 				 * needlessly register any other waiters after
872 				 * this. They'll get immediately deregistered
873 				 * when we break out and return.
874 				 */
875 				if (do_pollfd(pfd, pt, &can_busy_loop,
876 					      busy_flag)) {
877 					count++;
878 					pt->_qproc = NULL;
879 					/* found something, stop busy polling */
880 					busy_flag = 0;
881 					can_busy_loop = false;
882 				}
883 			}
884 		}
885 		/*
886 		 * All waiters have already been registered, so don't provide
887 		 * a poll_table->_qproc to them on the next loop iteration.
888 		 */
889 		pt->_qproc = NULL;
890 		if (!count) {
891 			count = wait->error;
892 			if (signal_pending(current))
893 				count = -EINTR;
894 		}
895 		if (count || timed_out)
896 			break;
897 
898 		/* only if found POLL_BUSY_LOOP sockets && not out of time */
899 		if (can_busy_loop && !need_resched()) {
900 			if (!busy_start) {
901 				busy_start = busy_loop_current_time();
902 				continue;
903 			}
904 			if (!busy_loop_timeout(busy_start))
905 				continue;
906 		}
907 		busy_flag = 0;
908 
909 		/*
910 		 * If this is the first loop and we have a timeout
911 		 * given, then we convert to ktime_t and set the to
912 		 * pointer to the expiry value.
913 		 */
914 		if (end_time && !to) {
915 			expire = timespec64_to_ktime(*end_time);
916 			to = &expire;
917 		}
918 
919 		if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
920 			timed_out = 1;
921 	}
922 	return count;
923 }
924 
925 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
926 			sizeof(struct pollfd))
927 
928 static int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
929 		struct timespec64 *end_time)
930 {
931 	struct poll_wqueues table;
932  	int err = -EFAULT, fdcount, len, size;
933 	/* Allocate small arguments on the stack to save memory and be
934 	   faster - use long to make sure the buffer is aligned properly
935 	   on 64 bit archs to avoid unaligned access */
936 	long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
937 	struct poll_list *const head = (struct poll_list *)stack_pps;
938  	struct poll_list *walk = head;
939  	unsigned long todo = nfds;
940 
941 	if (nfds > rlimit(RLIMIT_NOFILE))
942 		return -EINVAL;
943 
944 	len = min_t(unsigned int, nfds, N_STACK_PPS);
945 	for (;;) {
946 		walk->next = NULL;
947 		walk->len = len;
948 		if (!len)
949 			break;
950 
951 		if (copy_from_user(walk->entries, ufds + nfds-todo,
952 					sizeof(struct pollfd) * walk->len))
953 			goto out_fds;
954 
955 		todo -= walk->len;
956 		if (!todo)
957 			break;
958 
959 		len = min(todo, POLLFD_PER_PAGE);
960 		size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
961 		walk = walk->next = kmalloc(size, GFP_KERNEL);
962 		if (!walk) {
963 			err = -ENOMEM;
964 			goto out_fds;
965 		}
966 	}
967 
968 	poll_initwait(&table);
969 	fdcount = do_poll(head, &table, end_time);
970 	poll_freewait(&table);
971 
972 	for (walk = head; walk; walk = walk->next) {
973 		struct pollfd *fds = walk->entries;
974 		int j;
975 
976 		for (j = 0; j < walk->len; j++, ufds++)
977 			if (__put_user(fds[j].revents, &ufds->revents))
978 				goto out_fds;
979   	}
980 
981 	err = fdcount;
982 out_fds:
983 	walk = head->next;
984 	while (walk) {
985 		struct poll_list *pos = walk;
986 		walk = walk->next;
987 		kfree(pos);
988 	}
989 
990 	return err;
991 }
992 
993 static long do_restart_poll(struct restart_block *restart_block)
994 {
995 	struct pollfd __user *ufds = restart_block->poll.ufds;
996 	int nfds = restart_block->poll.nfds;
997 	struct timespec64 *to = NULL, end_time;
998 	int ret;
999 
1000 	if (restart_block->poll.has_timeout) {
1001 		end_time.tv_sec = restart_block->poll.tv_sec;
1002 		end_time.tv_nsec = restart_block->poll.tv_nsec;
1003 		to = &end_time;
1004 	}
1005 
1006 	ret = do_sys_poll(ufds, nfds, to);
1007 
1008 	if (ret == -EINTR) {
1009 		restart_block->fn = do_restart_poll;
1010 		ret = -ERESTART_RESTARTBLOCK;
1011 	}
1012 	return ret;
1013 }
1014 
1015 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
1016 		int, timeout_msecs)
1017 {
1018 	struct timespec64 end_time, *to = NULL;
1019 	int ret;
1020 
1021 	if (timeout_msecs >= 0) {
1022 		to = &end_time;
1023 		poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
1024 			NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
1025 	}
1026 
1027 	ret = do_sys_poll(ufds, nfds, to);
1028 
1029 	if (ret == -EINTR) {
1030 		struct restart_block *restart_block;
1031 
1032 		restart_block = &current->restart_block;
1033 		restart_block->fn = do_restart_poll;
1034 		restart_block->poll.ufds = ufds;
1035 		restart_block->poll.nfds = nfds;
1036 
1037 		if (timeout_msecs >= 0) {
1038 			restart_block->poll.tv_sec = end_time.tv_sec;
1039 			restart_block->poll.tv_nsec = end_time.tv_nsec;
1040 			restart_block->poll.has_timeout = 1;
1041 		} else
1042 			restart_block->poll.has_timeout = 0;
1043 
1044 		ret = -ERESTART_RESTARTBLOCK;
1045 	}
1046 	return ret;
1047 }
1048 
1049 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
1050 		struct timespec __user *, tsp, const sigset_t __user *, sigmask,
1051 		size_t, sigsetsize)
1052 {
1053 	sigset_t ksigmask, sigsaved;
1054 	struct timespec ts;
1055 	struct timespec64 end_time, *to = NULL;
1056 	int ret;
1057 
1058 	if (tsp) {
1059 		if (copy_from_user(&ts, tsp, sizeof(ts)))
1060 			return -EFAULT;
1061 
1062 		to = &end_time;
1063 		if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1064 			return -EINVAL;
1065 	}
1066 
1067 	if (sigmask) {
1068 		/* XXX: Don't preclude handling different sized sigset_t's.  */
1069 		if (sigsetsize != sizeof(sigset_t))
1070 			return -EINVAL;
1071 		if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1072 			return -EFAULT;
1073 
1074 		sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
1075 		sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1076 	}
1077 
1078 	ret = do_sys_poll(ufds, nfds, to);
1079 
1080 	/* We can restart this syscall, usually */
1081 	if (ret == -EINTR) {
1082 		/*
1083 		 * Don't restore the signal mask yet. Let do_signal() deliver
1084 		 * the signal on the way back to userspace, before the signal
1085 		 * mask is restored.
1086 		 */
1087 		if (sigmask) {
1088 			memcpy(&current->saved_sigmask, &sigsaved,
1089 					sizeof(sigsaved));
1090 			set_restore_sigmask();
1091 		}
1092 		ret = -ERESTARTNOHAND;
1093 	} else if (sigmask)
1094 		sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1095 
1096 	ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
1097 
1098 	return ret;
1099 }
1100 
1101 #ifdef CONFIG_COMPAT
1102 #define __COMPAT_NFDBITS       (8 * sizeof(compat_ulong_t))
1103 
1104 static
1105 int compat_poll_select_copy_remaining(struct timespec *end_time, void __user *p,
1106 				      int timeval, int ret)
1107 {
1108 	struct timespec ts;
1109 
1110 	if (!p)
1111 		return ret;
1112 
1113 	if (current->personality & STICKY_TIMEOUTS)
1114 		goto sticky;
1115 
1116 	/* No update for zero timeout */
1117 	if (!end_time->tv_sec && !end_time->tv_nsec)
1118 		return ret;
1119 
1120 	ktime_get_ts(&ts);
1121 	ts = timespec_sub(*end_time, ts);
1122 	if (ts.tv_sec < 0)
1123 		ts.tv_sec = ts.tv_nsec = 0;
1124 
1125 	if (timeval) {
1126 		struct compat_timeval rtv;
1127 
1128 		rtv.tv_sec = ts.tv_sec;
1129 		rtv.tv_usec = ts.tv_nsec / NSEC_PER_USEC;
1130 
1131 		if (!copy_to_user(p, &rtv, sizeof(rtv)))
1132 			return ret;
1133 	} else {
1134 		struct compat_timespec rts;
1135 
1136 		rts.tv_sec = ts.tv_sec;
1137 		rts.tv_nsec = ts.tv_nsec;
1138 
1139 		if (!copy_to_user(p, &rts, sizeof(rts)))
1140 			return ret;
1141 	}
1142 	/*
1143 	 * If an application puts its timeval in read-only memory, we
1144 	 * don't want the Linux-specific update to the timeval to
1145 	 * cause a fault after the select has completed
1146 	 * successfully. However, because we're not updating the
1147 	 * timeval, we can't restart the system call.
1148 	 */
1149 
1150 sticky:
1151 	if (ret == -ERESTARTNOHAND)
1152 		ret = -EINTR;
1153 	return ret;
1154 }
1155 
1156 /*
1157  * Ooo, nasty.  We need here to frob 32-bit unsigned longs to
1158  * 64-bit unsigned longs.
1159  */
1160 static
1161 int compat_get_fd_set(unsigned long nr, compat_ulong_t __user *ufdset,
1162 			unsigned long *fdset)
1163 {
1164 	if (ufdset) {
1165 		return compat_get_bitmap(fdset, ufdset, nr);
1166 	} else {
1167 		zero_fd_set(nr, fdset);
1168 		return 0;
1169 	}
1170 }
1171 
1172 static
1173 int compat_set_fd_set(unsigned long nr, compat_ulong_t __user *ufdset,
1174 		      unsigned long *fdset)
1175 {
1176 	if (!ufdset)
1177 		return 0;
1178 	return compat_put_bitmap(ufdset, fdset, nr);
1179 }
1180 
1181 
1182 /*
1183  * This is a virtual copy of sys_select from fs/select.c and probably
1184  * should be compared to it from time to time
1185  */
1186 
1187 /*
1188  * We can actually return ERESTARTSYS instead of EINTR, but I'd
1189  * like to be certain this leads to no problems. So I return
1190  * EINTR just for safety.
1191  *
1192  * Update: ERESTARTSYS breaks at least the xview clock binary, so
1193  * I'm trying ERESTARTNOHAND which restart only when you want to.
1194  */
1195 static int compat_core_sys_select(int n, compat_ulong_t __user *inp,
1196 	compat_ulong_t __user *outp, compat_ulong_t __user *exp,
1197 	struct timespec *end_time)
1198 {
1199 	fd_set_bits fds;
1200 	void *bits;
1201 	int size, max_fds, ret = -EINVAL;
1202 	struct fdtable *fdt;
1203 	long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
1204 
1205 	if (n < 0)
1206 		goto out_nofds;
1207 
1208 	/* max_fds can increase, so grab it once to avoid race */
1209 	rcu_read_lock();
1210 	fdt = files_fdtable(current->files);
1211 	max_fds = fdt->max_fds;
1212 	rcu_read_unlock();
1213 	if (n > max_fds)
1214 		n = max_fds;
1215 
1216 	/*
1217 	 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
1218 	 * since we used fdset we need to allocate memory in units of
1219 	 * long-words.
1220 	 */
1221 	size = FDS_BYTES(n);
1222 	bits = stack_fds;
1223 	if (size > sizeof(stack_fds) / 6) {
1224 		bits = kmalloc(6 * size, GFP_KERNEL);
1225 		ret = -ENOMEM;
1226 		if (!bits)
1227 			goto out_nofds;
1228 	}
1229 	fds.in      = (unsigned long *)  bits;
1230 	fds.out     = (unsigned long *) (bits +   size);
1231 	fds.ex      = (unsigned long *) (bits + 2*size);
1232 	fds.res_in  = (unsigned long *) (bits + 3*size);
1233 	fds.res_out = (unsigned long *) (bits + 4*size);
1234 	fds.res_ex  = (unsigned long *) (bits + 5*size);
1235 
1236 	if ((ret = compat_get_fd_set(n, inp, fds.in)) ||
1237 	    (ret = compat_get_fd_set(n, outp, fds.out)) ||
1238 	    (ret = compat_get_fd_set(n, exp, fds.ex)))
1239 		goto out;
1240 	zero_fd_set(n, fds.res_in);
1241 	zero_fd_set(n, fds.res_out);
1242 	zero_fd_set(n, fds.res_ex);
1243 
1244 	ret = do_select(n, &fds, end_time);
1245 
1246 	if (ret < 0)
1247 		goto out;
1248 	if (!ret) {
1249 		ret = -ERESTARTNOHAND;
1250 		if (signal_pending(current))
1251 			goto out;
1252 		ret = 0;
1253 	}
1254 
1255 	if (compat_set_fd_set(n, inp, fds.res_in) ||
1256 	    compat_set_fd_set(n, outp, fds.res_out) ||
1257 	    compat_set_fd_set(n, exp, fds.res_ex))
1258 		ret = -EFAULT;
1259 out:
1260 	if (bits != stack_fds)
1261 		kfree(bits);
1262 out_nofds:
1263 	return ret;
1264 }
1265 
1266 COMPAT_SYSCALL_DEFINE5(select, int, n, compat_ulong_t __user *, inp,
1267 	compat_ulong_t __user *, outp, compat_ulong_t __user *, exp,
1268 	struct compat_timeval __user *, tvp)
1269 {
1270 	struct timespec end_time, *to = NULL;
1271 	struct compat_timeval tv;
1272 	int ret;
1273 
1274 	if (tvp) {
1275 		if (copy_from_user(&tv, tvp, sizeof(tv)))
1276 			return -EFAULT;
1277 
1278 		to = &end_time;
1279 		if (poll_select_set_timeout(to,
1280 				tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
1281 				(tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
1282 			return -EINVAL;
1283 	}
1284 
1285 	ret = compat_core_sys_select(n, inp, outp, exp, to);
1286 	ret = compat_poll_select_copy_remaining(&end_time, tvp, 1, ret);
1287 
1288 	return ret;
1289 }
1290 
1291 struct compat_sel_arg_struct {
1292 	compat_ulong_t n;
1293 	compat_uptr_t inp;
1294 	compat_uptr_t outp;
1295 	compat_uptr_t exp;
1296 	compat_uptr_t tvp;
1297 };
1298 
1299 COMPAT_SYSCALL_DEFINE1(old_select, struct compat_sel_arg_struct __user *, arg)
1300 {
1301 	struct compat_sel_arg_struct a;
1302 
1303 	if (copy_from_user(&a, arg, sizeof(a)))
1304 		return -EFAULT;
1305 	return compat_sys_select(a.n, compat_ptr(a.inp), compat_ptr(a.outp),
1306 				 compat_ptr(a.exp), compat_ptr(a.tvp));
1307 }
1308 
1309 static long do_compat_pselect(int n, compat_ulong_t __user *inp,
1310 	compat_ulong_t __user *outp, compat_ulong_t __user *exp,
1311 	struct compat_timespec __user *tsp, compat_sigset_t __user *sigmask,
1312 	compat_size_t sigsetsize)
1313 {
1314 	compat_sigset_t ss32;
1315 	sigset_t ksigmask, sigsaved;
1316 	struct compat_timespec ts;
1317 	struct timespec end_time, *to = NULL;
1318 	int ret;
1319 
1320 	if (tsp) {
1321 		if (copy_from_user(&ts, tsp, sizeof(ts)))
1322 			return -EFAULT;
1323 
1324 		to = &end_time;
1325 		if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1326 			return -EINVAL;
1327 	}
1328 
1329 	if (sigmask) {
1330 		if (sigsetsize != sizeof(compat_sigset_t))
1331 			return -EINVAL;
1332 		if (copy_from_user(&ss32, sigmask, sizeof(ss32)))
1333 			return -EFAULT;
1334 		sigset_from_compat(&ksigmask, &ss32);
1335 
1336 		sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
1337 		sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1338 	}
1339 
1340 	ret = compat_core_sys_select(n, inp, outp, exp, to);
1341 	ret = compat_poll_select_copy_remaining(&end_time, tsp, 0, ret);
1342 
1343 	if (ret == -ERESTARTNOHAND) {
1344 		/*
1345 		 * Don't restore the signal mask yet. Let do_signal() deliver
1346 		 * the signal on the way back to userspace, before the signal
1347 		 * mask is restored.
1348 		 */
1349 		if (sigmask) {
1350 			memcpy(&current->saved_sigmask, &sigsaved,
1351 					sizeof(sigsaved));
1352 			set_restore_sigmask();
1353 		}
1354 	} else if (sigmask)
1355 		sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1356 
1357 	return ret;
1358 }
1359 
1360 COMPAT_SYSCALL_DEFINE6(pselect6, int, n, compat_ulong_t __user *, inp,
1361 	compat_ulong_t __user *, outp, compat_ulong_t __user *, exp,
1362 	struct compat_timespec __user *, tsp, void __user *, sig)
1363 {
1364 	compat_size_t sigsetsize = 0;
1365 	compat_uptr_t up = 0;
1366 
1367 	if (sig) {
1368 		if (!access_ok(VERIFY_READ, sig,
1369 				sizeof(compat_uptr_t)+sizeof(compat_size_t)) ||
1370 		    	__get_user(up, (compat_uptr_t __user *)sig) ||
1371 		    	__get_user(sigsetsize,
1372 				(compat_size_t __user *)(sig+sizeof(up))))
1373 			return -EFAULT;
1374 	}
1375 	return do_compat_pselect(n, inp, outp, exp, tsp, compat_ptr(up),
1376 				 sigsetsize);
1377 }
1378 
1379 COMPAT_SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds,
1380 	unsigned int,  nfds, struct compat_timespec __user *, tsp,
1381 	const compat_sigset_t __user *, sigmask, compat_size_t, sigsetsize)
1382 {
1383 	compat_sigset_t ss32;
1384 	sigset_t ksigmask, sigsaved;
1385 	struct compat_timespec ts;
1386 	struct timespec end_time, *to = NULL;
1387 	int ret;
1388 
1389 	if (tsp) {
1390 		if (copy_from_user(&ts, tsp, sizeof(ts)))
1391 			return -EFAULT;
1392 
1393 		to = &end_time;
1394 		if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1395 			return -EINVAL;
1396 	}
1397 
1398 	if (sigmask) {
1399 		if (sigsetsize != sizeof(compat_sigset_t))
1400 			return -EINVAL;
1401 		if (copy_from_user(&ss32, sigmask, sizeof(ss32)))
1402 			return -EFAULT;
1403 		sigset_from_compat(&ksigmask, &ss32);
1404 
1405 		sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
1406 		sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1407 	}
1408 
1409 	ret = do_sys_poll(ufds, nfds, to);
1410 
1411 	/* We can restart this syscall, usually */
1412 	if (ret == -EINTR) {
1413 		/*
1414 		 * Don't restore the signal mask yet. Let do_signal() deliver
1415 		 * the signal on the way back to userspace, before the signal
1416 		 * mask is restored.
1417 		 */
1418 		if (sigmask) {
1419 			memcpy(&current->saved_sigmask, &sigsaved,
1420 				sizeof(sigsaved));
1421 			set_restore_sigmask();
1422 		}
1423 		ret = -ERESTARTNOHAND;
1424 	} else if (sigmask)
1425 		sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1426 
1427 	ret = compat_poll_select_copy_remaining(&end_time, tsp, 0, ret);
1428 
1429 	return ret;
1430 }
1431 #endif
1432