xref: /openbmc/linux/fs/select.c (revision 8730046c)
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/export.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 #include <linux/sched/rt.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_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_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 #define FDS_IN(fds, n)		(fds->in + n)
342 #define FDS_OUT(fds, n)		(fds->out + n)
343 #define FDS_EX(fds, n)		(fds->ex + n)
344 
345 #define BITS(fds, n)	(*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
346 
347 static int max_select_fd(unsigned long n, fd_set_bits *fds)
348 {
349 	unsigned long *open_fds;
350 	unsigned long set;
351 	int max;
352 	struct fdtable *fdt;
353 
354 	/* handle last in-complete long-word first */
355 	set = ~(~0UL << (n & (BITS_PER_LONG-1)));
356 	n /= BITS_PER_LONG;
357 	fdt = files_fdtable(current->files);
358 	open_fds = fdt->open_fds + n;
359 	max = 0;
360 	if (set) {
361 		set &= BITS(fds, n);
362 		if (set) {
363 			if (!(set & ~*open_fds))
364 				goto get_max;
365 			return -EBADF;
366 		}
367 	}
368 	while (n) {
369 		open_fds--;
370 		n--;
371 		set = BITS(fds, n);
372 		if (!set)
373 			continue;
374 		if (set & ~*open_fds)
375 			return -EBADF;
376 		if (max)
377 			continue;
378 get_max:
379 		do {
380 			max++;
381 			set >>= 1;
382 		} while (set);
383 		max += n * BITS_PER_LONG;
384 	}
385 
386 	return max;
387 }
388 
389 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
390 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
391 #define POLLEX_SET (POLLPRI)
392 
393 static inline void wait_key_set(poll_table *wait, unsigned long in,
394 				unsigned long out, unsigned long bit,
395 				unsigned int ll_flag)
396 {
397 	wait->_key = POLLEX_SET | ll_flag;
398 	if (in & bit)
399 		wait->_key |= POLLIN_SET;
400 	if (out & bit)
401 		wait->_key |= POLLOUT_SET;
402 }
403 
404 int do_select(int n, fd_set_bits *fds, struct timespec64 *end_time)
405 {
406 	ktime_t expire, *to = NULL;
407 	struct poll_wqueues table;
408 	poll_table *wait;
409 	int retval, i, timed_out = 0;
410 	u64 slack = 0;
411 	unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
412 	unsigned long busy_end = 0;
413 
414 	rcu_read_lock();
415 	retval = max_select_fd(n, fds);
416 	rcu_read_unlock();
417 
418 	if (retval < 0)
419 		return retval;
420 	n = retval;
421 
422 	poll_initwait(&table);
423 	wait = &table.pt;
424 	if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
425 		wait->_qproc = NULL;
426 		timed_out = 1;
427 	}
428 
429 	if (end_time && !timed_out)
430 		slack = select_estimate_accuracy(end_time);
431 
432 	retval = 0;
433 	for (;;) {
434 		unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
435 		bool can_busy_loop = false;
436 
437 		inp = fds->in; outp = fds->out; exp = fds->ex;
438 		rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
439 
440 		for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
441 			unsigned long in, out, ex, all_bits, bit = 1, mask, j;
442 			unsigned long res_in = 0, res_out = 0, res_ex = 0;
443 
444 			in = *inp++; out = *outp++; ex = *exp++;
445 			all_bits = in | out | ex;
446 			if (all_bits == 0) {
447 				i += BITS_PER_LONG;
448 				continue;
449 			}
450 
451 			for (j = 0; j < BITS_PER_LONG; ++j, ++i, bit <<= 1) {
452 				struct fd f;
453 				if (i >= n)
454 					break;
455 				if (!(bit & all_bits))
456 					continue;
457 				f = fdget(i);
458 				if (f.file) {
459 					const struct file_operations *f_op;
460 					f_op = f.file->f_op;
461 					mask = DEFAULT_POLLMASK;
462 					if (f_op->poll) {
463 						wait_key_set(wait, in, out,
464 							     bit, busy_flag);
465 						mask = (*f_op->poll)(f.file, wait);
466 					}
467 					fdput(f);
468 					if ((mask & POLLIN_SET) && (in & bit)) {
469 						res_in |= bit;
470 						retval++;
471 						wait->_qproc = NULL;
472 					}
473 					if ((mask & POLLOUT_SET) && (out & bit)) {
474 						res_out |= bit;
475 						retval++;
476 						wait->_qproc = NULL;
477 					}
478 					if ((mask & POLLEX_SET) && (ex & bit)) {
479 						res_ex |= bit;
480 						retval++;
481 						wait->_qproc = NULL;
482 					}
483 					/* got something, stop busy polling */
484 					if (retval) {
485 						can_busy_loop = false;
486 						busy_flag = 0;
487 
488 					/*
489 					 * only remember a returned
490 					 * POLL_BUSY_LOOP if we asked for it
491 					 */
492 					} else if (busy_flag & mask)
493 						can_busy_loop = true;
494 
495 				}
496 			}
497 			if (res_in)
498 				*rinp = res_in;
499 			if (res_out)
500 				*routp = res_out;
501 			if (res_ex)
502 				*rexp = res_ex;
503 			cond_resched();
504 		}
505 		wait->_qproc = NULL;
506 		if (retval || timed_out || signal_pending(current))
507 			break;
508 		if (table.error) {
509 			retval = table.error;
510 			break;
511 		}
512 
513 		/* only if found POLL_BUSY_LOOP sockets && not out of time */
514 		if (can_busy_loop && !need_resched()) {
515 			if (!busy_end) {
516 				busy_end = busy_loop_end_time();
517 				continue;
518 			}
519 			if (!busy_loop_timeout(busy_end))
520 				continue;
521 		}
522 		busy_flag = 0;
523 
524 		/*
525 		 * If this is the first loop and we have a timeout
526 		 * given, then we convert to ktime_t and set the to
527 		 * pointer to the expiry value.
528 		 */
529 		if (end_time && !to) {
530 			expire = timespec64_to_ktime(*end_time);
531 			to = &expire;
532 		}
533 
534 		if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
535 					   to, slack))
536 			timed_out = 1;
537 	}
538 
539 	poll_freewait(&table);
540 
541 	return retval;
542 }
543 
544 /*
545  * We can actually return ERESTARTSYS instead of EINTR, but I'd
546  * like to be certain this leads to no problems. So I return
547  * EINTR just for safety.
548  *
549  * Update: ERESTARTSYS breaks at least the xview clock binary, so
550  * I'm trying ERESTARTNOHAND which restart only when you want to.
551  */
552 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
553 			   fd_set __user *exp, struct timespec64 *end_time)
554 {
555 	fd_set_bits fds;
556 	void *bits;
557 	int ret, max_fds;
558 	size_t size, alloc_size;
559 	struct fdtable *fdt;
560 	/* Allocate small arguments on the stack to save memory and be faster */
561 	long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
562 
563 	ret = -EINVAL;
564 	if (n < 0)
565 		goto out_nofds;
566 
567 	/* max_fds can increase, so grab it once to avoid race */
568 	rcu_read_lock();
569 	fdt = files_fdtable(current->files);
570 	max_fds = fdt->max_fds;
571 	rcu_read_unlock();
572 	if (n > max_fds)
573 		n = max_fds;
574 
575 	/*
576 	 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
577 	 * since we used fdset we need to allocate memory in units of
578 	 * long-words.
579 	 */
580 	size = FDS_BYTES(n);
581 	bits = stack_fds;
582 	if (size > sizeof(stack_fds) / 6) {
583 		/* Not enough space in on-stack array; must use kmalloc */
584 		ret = -ENOMEM;
585 		if (size > (SIZE_MAX / 6))
586 			goto out_nofds;
587 
588 		alloc_size = 6 * size;
589 		bits = kmalloc(alloc_size, GFP_KERNEL|__GFP_NOWARN);
590 		if (!bits && alloc_size > PAGE_SIZE)
591 			bits = vmalloc(alloc_size);
592 
593 		if (!bits)
594 			goto out_nofds;
595 	}
596 	fds.in      = bits;
597 	fds.out     = bits +   size;
598 	fds.ex      = bits + 2*size;
599 	fds.res_in  = bits + 3*size;
600 	fds.res_out = bits + 4*size;
601 	fds.res_ex  = bits + 5*size;
602 
603 	if ((ret = get_fd_set(n, inp, fds.in)) ||
604 	    (ret = get_fd_set(n, outp, fds.out)) ||
605 	    (ret = get_fd_set(n, exp, fds.ex)))
606 		goto out;
607 	zero_fd_set(n, fds.res_in);
608 	zero_fd_set(n, fds.res_out);
609 	zero_fd_set(n, fds.res_ex);
610 
611 	ret = do_select(n, &fds, end_time);
612 
613 	if (ret < 0)
614 		goto out;
615 	if (!ret) {
616 		ret = -ERESTARTNOHAND;
617 		if (signal_pending(current))
618 			goto out;
619 		ret = 0;
620 	}
621 
622 	if (set_fd_set(n, inp, fds.res_in) ||
623 	    set_fd_set(n, outp, fds.res_out) ||
624 	    set_fd_set(n, exp, fds.res_ex))
625 		ret = -EFAULT;
626 
627 out:
628 	if (bits != stack_fds)
629 		kvfree(bits);
630 out_nofds:
631 	return ret;
632 }
633 
634 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
635 		fd_set __user *, exp, struct timeval __user *, tvp)
636 {
637 	struct timespec64 end_time, *to = NULL;
638 	struct timeval tv;
639 	int ret;
640 
641 	if (tvp) {
642 		if (copy_from_user(&tv, tvp, sizeof(tv)))
643 			return -EFAULT;
644 
645 		to = &end_time;
646 		if (poll_select_set_timeout(to,
647 				tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
648 				(tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
649 			return -EINVAL;
650 	}
651 
652 	ret = core_sys_select(n, inp, outp, exp, to);
653 	ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
654 
655 	return ret;
656 }
657 
658 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
659 		       fd_set __user *exp, struct timespec __user *tsp,
660 		       const sigset_t __user *sigmask, size_t sigsetsize)
661 {
662 	sigset_t ksigmask, sigsaved;
663 	struct timespec ts;
664 	struct timespec64 ts64, end_time, *to = NULL;
665 	int ret;
666 
667 	if (tsp) {
668 		if (copy_from_user(&ts, tsp, sizeof(ts)))
669 			return -EFAULT;
670 		ts64 = timespec_to_timespec64(ts);
671 
672 		to = &end_time;
673 		if (poll_select_set_timeout(to, ts64.tv_sec, ts64.tv_nsec))
674 			return -EINVAL;
675 	}
676 
677 	if (sigmask) {
678 		/* XXX: Don't preclude handling different sized sigset_t's.  */
679 		if (sigsetsize != sizeof(sigset_t))
680 			return -EINVAL;
681 		if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
682 			return -EFAULT;
683 
684 		sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
685 		sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
686 	}
687 
688 	ret = core_sys_select(n, inp, outp, exp, to);
689 	ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
690 
691 	if (ret == -ERESTARTNOHAND) {
692 		/*
693 		 * Don't restore the signal mask yet. Let do_signal() deliver
694 		 * the signal on the way back to userspace, before the signal
695 		 * mask is restored.
696 		 */
697 		if (sigmask) {
698 			memcpy(&current->saved_sigmask, &sigsaved,
699 					sizeof(sigsaved));
700 			set_restore_sigmask();
701 		}
702 	} else if (sigmask)
703 		sigprocmask(SIG_SETMASK, &sigsaved, NULL);
704 
705 	return ret;
706 }
707 
708 /*
709  * Most architectures can't handle 7-argument syscalls. So we provide a
710  * 6-argument version where the sixth argument is a pointer to a structure
711  * which has a pointer to the sigset_t itself followed by a size_t containing
712  * the sigset size.
713  */
714 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
715 		fd_set __user *, exp, struct timespec __user *, tsp,
716 		void __user *, sig)
717 {
718 	size_t sigsetsize = 0;
719 	sigset_t __user *up = NULL;
720 
721 	if (sig) {
722 		if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
723 		    || __get_user(up, (sigset_t __user * __user *)sig)
724 		    || __get_user(sigsetsize,
725 				(size_t __user *)(sig+sizeof(void *))))
726 			return -EFAULT;
727 	}
728 
729 	return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
730 }
731 
732 #ifdef __ARCH_WANT_SYS_OLD_SELECT
733 struct sel_arg_struct {
734 	unsigned long n;
735 	fd_set __user *inp, *outp, *exp;
736 	struct timeval __user *tvp;
737 };
738 
739 SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg)
740 {
741 	struct sel_arg_struct a;
742 
743 	if (copy_from_user(&a, arg, sizeof(a)))
744 		return -EFAULT;
745 	return sys_select(a.n, a.inp, a.outp, a.exp, a.tvp);
746 }
747 #endif
748 
749 struct poll_list {
750 	struct poll_list *next;
751 	int len;
752 	struct pollfd entries[0];
753 };
754 
755 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
756 
757 /*
758  * Fish for pollable events on the pollfd->fd file descriptor. We're only
759  * interested in events matching the pollfd->events mask, and the result
760  * matching that mask is both recorded in pollfd->revents and returned. The
761  * pwait poll_table will be used by the fd-provided poll handler for waiting,
762  * if pwait->_qproc is non-NULL.
763  */
764 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait,
765 				     bool *can_busy_poll,
766 				     unsigned int busy_flag)
767 {
768 	unsigned int mask;
769 	int fd;
770 
771 	mask = 0;
772 	fd = pollfd->fd;
773 	if (fd >= 0) {
774 		struct fd f = fdget(fd);
775 		mask = POLLNVAL;
776 		if (f.file) {
777 			mask = DEFAULT_POLLMASK;
778 			if (f.file->f_op->poll) {
779 				pwait->_key = pollfd->events|POLLERR|POLLHUP;
780 				pwait->_key |= busy_flag;
781 				mask = f.file->f_op->poll(f.file, pwait);
782 				if (mask & busy_flag)
783 					*can_busy_poll = true;
784 			}
785 			/* Mask out unneeded events. */
786 			mask &= pollfd->events | POLLERR | POLLHUP;
787 			fdput(f);
788 		}
789 	}
790 	pollfd->revents = mask;
791 
792 	return mask;
793 }
794 
795 static int do_poll(struct poll_list *list, struct poll_wqueues *wait,
796 		   struct timespec64 *end_time)
797 {
798 	poll_table* pt = &wait->pt;
799 	ktime_t expire, *to = NULL;
800 	int timed_out = 0, count = 0;
801 	u64 slack = 0;
802 	unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
803 	unsigned long busy_end = 0;
804 
805 	/* Optimise the no-wait case */
806 	if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
807 		pt->_qproc = NULL;
808 		timed_out = 1;
809 	}
810 
811 	if (end_time && !timed_out)
812 		slack = select_estimate_accuracy(end_time);
813 
814 	for (;;) {
815 		struct poll_list *walk;
816 		bool can_busy_loop = false;
817 
818 		for (walk = list; walk != NULL; walk = walk->next) {
819 			struct pollfd * pfd, * pfd_end;
820 
821 			pfd = walk->entries;
822 			pfd_end = pfd + walk->len;
823 			for (; pfd != pfd_end; pfd++) {
824 				/*
825 				 * Fish for events. If we found one, record it
826 				 * and kill poll_table->_qproc, so we don't
827 				 * needlessly register any other waiters after
828 				 * this. They'll get immediately deregistered
829 				 * when we break out and return.
830 				 */
831 				if (do_pollfd(pfd, pt, &can_busy_loop,
832 					      busy_flag)) {
833 					count++;
834 					pt->_qproc = NULL;
835 					/* found something, stop busy polling */
836 					busy_flag = 0;
837 					can_busy_loop = false;
838 				}
839 			}
840 		}
841 		/*
842 		 * All waiters have already been registered, so don't provide
843 		 * a poll_table->_qproc to them on the next loop iteration.
844 		 */
845 		pt->_qproc = NULL;
846 		if (!count) {
847 			count = wait->error;
848 			if (signal_pending(current))
849 				count = -EINTR;
850 		}
851 		if (count || timed_out)
852 			break;
853 
854 		/* only if found POLL_BUSY_LOOP sockets && not out of time */
855 		if (can_busy_loop && !need_resched()) {
856 			if (!busy_end) {
857 				busy_end = busy_loop_end_time();
858 				continue;
859 			}
860 			if (!busy_loop_timeout(busy_end))
861 				continue;
862 		}
863 		busy_flag = 0;
864 
865 		/*
866 		 * If this is the first loop and we have a timeout
867 		 * given, then we convert to ktime_t and set the to
868 		 * pointer to the expiry value.
869 		 */
870 		if (end_time && !to) {
871 			expire = timespec64_to_ktime(*end_time);
872 			to = &expire;
873 		}
874 
875 		if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
876 			timed_out = 1;
877 	}
878 	return count;
879 }
880 
881 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
882 			sizeof(struct pollfd))
883 
884 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
885 		struct timespec64 *end_time)
886 {
887 	struct poll_wqueues table;
888  	int err = -EFAULT, fdcount, len, size;
889 	/* Allocate small arguments on the stack to save memory and be
890 	   faster - use long to make sure the buffer is aligned properly
891 	   on 64 bit archs to avoid unaligned access */
892 	long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
893 	struct poll_list *const head = (struct poll_list *)stack_pps;
894  	struct poll_list *walk = head;
895  	unsigned long todo = nfds;
896 
897 	if (nfds > rlimit(RLIMIT_NOFILE))
898 		return -EINVAL;
899 
900 	len = min_t(unsigned int, nfds, N_STACK_PPS);
901 	for (;;) {
902 		walk->next = NULL;
903 		walk->len = len;
904 		if (!len)
905 			break;
906 
907 		if (copy_from_user(walk->entries, ufds + nfds-todo,
908 					sizeof(struct pollfd) * walk->len))
909 			goto out_fds;
910 
911 		todo -= walk->len;
912 		if (!todo)
913 			break;
914 
915 		len = min(todo, POLLFD_PER_PAGE);
916 		size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
917 		walk = walk->next = kmalloc(size, GFP_KERNEL);
918 		if (!walk) {
919 			err = -ENOMEM;
920 			goto out_fds;
921 		}
922 	}
923 
924 	poll_initwait(&table);
925 	fdcount = do_poll(head, &table, end_time);
926 	poll_freewait(&table);
927 
928 	for (walk = head; walk; walk = walk->next) {
929 		struct pollfd *fds = walk->entries;
930 		int j;
931 
932 		for (j = 0; j < walk->len; j++, ufds++)
933 			if (__put_user(fds[j].revents, &ufds->revents))
934 				goto out_fds;
935   	}
936 
937 	err = fdcount;
938 out_fds:
939 	walk = head->next;
940 	while (walk) {
941 		struct poll_list *pos = walk;
942 		walk = walk->next;
943 		kfree(pos);
944 	}
945 
946 	return err;
947 }
948 
949 static long do_restart_poll(struct restart_block *restart_block)
950 {
951 	struct pollfd __user *ufds = restart_block->poll.ufds;
952 	int nfds = restart_block->poll.nfds;
953 	struct timespec64 *to = NULL, end_time;
954 	int ret;
955 
956 	if (restart_block->poll.has_timeout) {
957 		end_time.tv_sec = restart_block->poll.tv_sec;
958 		end_time.tv_nsec = restart_block->poll.tv_nsec;
959 		to = &end_time;
960 	}
961 
962 	ret = do_sys_poll(ufds, nfds, to);
963 
964 	if (ret == -EINTR) {
965 		restart_block->fn = do_restart_poll;
966 		ret = -ERESTART_RESTARTBLOCK;
967 	}
968 	return ret;
969 }
970 
971 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
972 		int, timeout_msecs)
973 {
974 	struct timespec64 end_time, *to = NULL;
975 	int ret;
976 
977 	if (timeout_msecs >= 0) {
978 		to = &end_time;
979 		poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
980 			NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
981 	}
982 
983 	ret = do_sys_poll(ufds, nfds, to);
984 
985 	if (ret == -EINTR) {
986 		struct restart_block *restart_block;
987 
988 		restart_block = &current->restart_block;
989 		restart_block->fn = do_restart_poll;
990 		restart_block->poll.ufds = ufds;
991 		restart_block->poll.nfds = nfds;
992 
993 		if (timeout_msecs >= 0) {
994 			restart_block->poll.tv_sec = end_time.tv_sec;
995 			restart_block->poll.tv_nsec = end_time.tv_nsec;
996 			restart_block->poll.has_timeout = 1;
997 		} else
998 			restart_block->poll.has_timeout = 0;
999 
1000 		ret = -ERESTART_RESTARTBLOCK;
1001 	}
1002 	return ret;
1003 }
1004 
1005 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
1006 		struct timespec __user *, tsp, const sigset_t __user *, sigmask,
1007 		size_t, sigsetsize)
1008 {
1009 	sigset_t ksigmask, sigsaved;
1010 	struct timespec ts;
1011 	struct timespec64 end_time, *to = NULL;
1012 	int ret;
1013 
1014 	if (tsp) {
1015 		if (copy_from_user(&ts, tsp, sizeof(ts)))
1016 			return -EFAULT;
1017 
1018 		to = &end_time;
1019 		if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1020 			return -EINVAL;
1021 	}
1022 
1023 	if (sigmask) {
1024 		/* XXX: Don't preclude handling different sized sigset_t's.  */
1025 		if (sigsetsize != sizeof(sigset_t))
1026 			return -EINVAL;
1027 		if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1028 			return -EFAULT;
1029 
1030 		sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
1031 		sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1032 	}
1033 
1034 	ret = do_sys_poll(ufds, nfds, to);
1035 
1036 	/* We can restart this syscall, usually */
1037 	if (ret == -EINTR) {
1038 		/*
1039 		 * Don't restore the signal mask yet. Let do_signal() deliver
1040 		 * the signal on the way back to userspace, before the signal
1041 		 * mask is restored.
1042 		 */
1043 		if (sigmask) {
1044 			memcpy(&current->saved_sigmask, &sigsaved,
1045 					sizeof(sigsaved));
1046 			set_restore_sigmask();
1047 		}
1048 		ret = -ERESTARTNOHAND;
1049 	} else if (sigmask)
1050 		sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1051 
1052 	ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
1053 
1054 	return ret;
1055 }
1056