Loading...
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
51static 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
74u64 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
96struct 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 */
117static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
118 poll_table *p);
119
120void 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}
129EXPORT_SYMBOL(poll_initwait);
130
131static 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
137void 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}
157EXPORT_SYMBOL(poll_freewait);
158
159static 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
183static 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
209static 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 */
220static 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
235int 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}
260EXPORT_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 */
273int 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
290static 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
335sticky:
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
347static 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;
378get_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
393static 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
404int 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 */
552int 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
627out:
628 if (bits != stack_fds)
629 kvfree(bits);
630out_nofds:
631 return ret;
632}
633
634SYSCALL_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
658static 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(¤t->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 */
714SYSCALL_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
733struct sel_arg_struct {
734 unsigned long n;
735 fd_set __user *inp, *outp, *exp;
736 struct timeval __user *tvp;
737};
738
739SYSCALL_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
749struct 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 */
764static 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
795static 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
884int 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;
938out_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
949static 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
971SYSCALL_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 = ¤t->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
1005SYSCALL_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(¤t->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}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * This file contains the procedures for the handling of select and poll
4 *
5 * Created for Linux based loosely upon Mathius Lattner's minix
6 * patches by Peter MacDonald. Heavily edited by Linus.
7 *
8 * 4 February 1994
9 * COFF/ELF binary emulation. If the process has the STICKY_TIMEOUTS
10 * flag set in its personality we do *not* modify the given timeout
11 * parameter to reflect time remaining.
12 *
13 * 24 January 2000
14 * Changed sys_poll()/do_poll() to use PAGE_SIZE chunk-based allocation
15 * of fds to overcome nfds < 16390 descriptors limit (Tigran Aivazian).
16 */
17
18#include <linux/compat.h>
19#include <linux/kernel.h>
20#include <linux/sched/signal.h>
21#include <linux/sched/rt.h>
22#include <linux/syscalls.h>
23#include <linux/export.h>
24#include <linux/slab.h>
25#include <linux/poll.h>
26#include <linux/personality.h> /* for STICKY_TIMEOUTS */
27#include <linux/file.h>
28#include <linux/fdtable.h>
29#include <linux/fs.h>
30#include <linux/rcupdate.h>
31#include <linux/hrtimer.h>
32#include <linux/freezer.h>
33#include <net/busy_poll.h>
34#include <linux/vmalloc.h>
35
36#include <linux/uaccess.h>
37
38
39/*
40 * Estimate expected accuracy in ns from a timeval.
41 *
42 * After quite a bit of churning around, we've settled on
43 * a simple thing of taking 0.1% of the timeout as the
44 * slack, with a cap of 100 msec.
45 * "nice" tasks get a 0.5% slack instead.
46 *
47 * Consider this comment an open invitation to come up with even
48 * better solutions..
49 */
50
51#define MAX_SLACK (100 * NSEC_PER_MSEC)
52
53static long __estimate_accuracy(struct timespec64 *tv)
54{
55 long slack;
56 int divfactor = 1000;
57
58 if (tv->tv_sec < 0)
59 return 0;
60
61 if (task_nice(current) > 0)
62 divfactor = divfactor / 5;
63
64 if (tv->tv_sec > MAX_SLACK / (NSEC_PER_SEC/divfactor))
65 return MAX_SLACK;
66
67 slack = tv->tv_nsec / divfactor;
68 slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
69
70 if (slack > MAX_SLACK)
71 return MAX_SLACK;
72
73 return slack;
74}
75
76u64 select_estimate_accuracy(struct timespec64 *tv)
77{
78 u64 ret;
79 struct timespec64 now;
80
81 /*
82 * Realtime tasks get a slack of 0 for obvious reasons.
83 */
84
85 if (rt_task(current))
86 return 0;
87
88 ktime_get_ts64(&now);
89 now = timespec64_sub(*tv, now);
90 ret = __estimate_accuracy(&now);
91 if (ret < current->timer_slack_ns)
92 return current->timer_slack_ns;
93 return ret;
94}
95
96
97
98struct poll_table_page {
99 struct poll_table_page * next;
100 struct poll_table_entry * entry;
101 struct poll_table_entry entries[];
102};
103
104#define POLL_TABLE_FULL(table) \
105 ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
106
107/*
108 * Ok, Peter made a complicated, but straightforward multiple_wait() function.
109 * I have rewritten this, taking some shortcuts: This code may not be easy to
110 * follow, but it should be free of race-conditions, and it's practical. If you
111 * understand what I'm doing here, then you understand how the linux
112 * sleep/wakeup mechanism works.
113 *
114 * Two very simple procedures, poll_wait() and poll_freewait() make all the
115 * work. poll_wait() is an inline-function defined in <linux/poll.h>,
116 * as all select/poll functions have to call it to add an entry to the
117 * poll table.
118 */
119static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
120 poll_table *p);
121
122void poll_initwait(struct poll_wqueues *pwq)
123{
124 init_poll_funcptr(&pwq->pt, __pollwait);
125 pwq->polling_task = current;
126 pwq->triggered = 0;
127 pwq->error = 0;
128 pwq->table = NULL;
129 pwq->inline_index = 0;
130}
131EXPORT_SYMBOL(poll_initwait);
132
133static void free_poll_entry(struct poll_table_entry *entry)
134{
135 remove_wait_queue(entry->wait_address, &entry->wait);
136 fput(entry->filp);
137}
138
139void poll_freewait(struct poll_wqueues *pwq)
140{
141 struct poll_table_page * p = pwq->table;
142 int i;
143 for (i = 0; i < pwq->inline_index; i++)
144 free_poll_entry(pwq->inline_entries + i);
145 while (p) {
146 struct poll_table_entry * entry;
147 struct poll_table_page *old;
148
149 entry = p->entry;
150 do {
151 entry--;
152 free_poll_entry(entry);
153 } while (entry > p->entries);
154 old = p;
155 p = p->next;
156 free_page((unsigned long) old);
157 }
158}
159EXPORT_SYMBOL(poll_freewait);
160
161static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
162{
163 struct poll_table_page *table = p->table;
164
165 if (p->inline_index < N_INLINE_POLL_ENTRIES)
166 return p->inline_entries + p->inline_index++;
167
168 if (!table || POLL_TABLE_FULL(table)) {
169 struct poll_table_page *new_table;
170
171 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
172 if (!new_table) {
173 p->error = -ENOMEM;
174 return NULL;
175 }
176 new_table->entry = new_table->entries;
177 new_table->next = table;
178 p->table = new_table;
179 table = new_table;
180 }
181
182 return table->entry++;
183}
184
185static int __pollwake(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
186{
187 struct poll_wqueues *pwq = wait->private;
188 DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
189
190 /*
191 * Although this function is called under waitqueue lock, LOCK
192 * doesn't imply write barrier and the users expect write
193 * barrier semantics on wakeup functions. The following
194 * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
195 * and is paired with smp_store_mb() in poll_schedule_timeout.
196 */
197 smp_wmb();
198 pwq->triggered = 1;
199
200 /*
201 * Perform the default wake up operation using a dummy
202 * waitqueue.
203 *
204 * TODO: This is hacky but there currently is no interface to
205 * pass in @sync. @sync is scheduled to be removed and once
206 * that happens, wake_up_process() can be used directly.
207 */
208 return default_wake_function(&dummy_wait, mode, sync, key);
209}
210
211static int pollwake(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
212{
213 struct poll_table_entry *entry;
214
215 entry = container_of(wait, struct poll_table_entry, wait);
216 if (key && !(key_to_poll(key) & entry->key))
217 return 0;
218 return __pollwake(wait, mode, sync, key);
219}
220
221/* Add a new entry */
222static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
223 poll_table *p)
224{
225 struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
226 struct poll_table_entry *entry = poll_get_entry(pwq);
227 if (!entry)
228 return;
229 entry->filp = get_file(filp);
230 entry->wait_address = wait_address;
231 entry->key = p->_key;
232 init_waitqueue_func_entry(&entry->wait, pollwake);
233 entry->wait.private = pwq;
234 add_wait_queue(wait_address, &entry->wait);
235}
236
237static int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
238 ktime_t *expires, unsigned long slack)
239{
240 int rc = -EINTR;
241
242 set_current_state(state);
243 if (!pwq->triggered)
244 rc = schedule_hrtimeout_range(expires, slack, HRTIMER_MODE_ABS);
245 __set_current_state(TASK_RUNNING);
246
247 /*
248 * Prepare for the next iteration.
249 *
250 * The following smp_store_mb() serves two purposes. First, it's
251 * the counterpart rmb of the wmb in pollwake() such that data
252 * written before wake up is always visible after wake up.
253 * Second, the full barrier guarantees that triggered clearing
254 * doesn't pass event check of the next iteration. Note that
255 * this problem doesn't exist for the first iteration as
256 * add_wait_queue() has full barrier semantics.
257 */
258 smp_store_mb(pwq->triggered, 0);
259
260 return rc;
261}
262
263/**
264 * poll_select_set_timeout - helper function to setup the timeout value
265 * @to: pointer to timespec64 variable for the final timeout
266 * @sec: seconds (from user space)
267 * @nsec: nanoseconds (from user space)
268 *
269 * Note, we do not use a timespec for the user space value here, That
270 * way we can use the function for timeval and compat interfaces as well.
271 *
272 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
273 */
274int poll_select_set_timeout(struct timespec64 *to, time64_t sec, long nsec)
275{
276 struct timespec64 ts = {.tv_sec = sec, .tv_nsec = nsec};
277
278 if (!timespec64_valid(&ts))
279 return -EINVAL;
280
281 /* Optimize for the zero timeout value here */
282 if (!sec && !nsec) {
283 to->tv_sec = to->tv_nsec = 0;
284 } else {
285 ktime_get_ts64(to);
286 *to = timespec64_add_safe(*to, ts);
287 }
288 return 0;
289}
290
291enum poll_time_type {
292 PT_TIMEVAL = 0,
293 PT_OLD_TIMEVAL = 1,
294 PT_TIMESPEC = 2,
295 PT_OLD_TIMESPEC = 3,
296};
297
298static int poll_select_finish(struct timespec64 *end_time,
299 void __user *p,
300 enum poll_time_type pt_type, int ret)
301{
302 struct timespec64 rts;
303
304 restore_saved_sigmask_unless(ret == -ERESTARTNOHAND);
305
306 if (!p)
307 return ret;
308
309 if (current->personality & STICKY_TIMEOUTS)
310 goto sticky;
311
312 /* No update for zero timeout */
313 if (!end_time->tv_sec && !end_time->tv_nsec)
314 return ret;
315
316 ktime_get_ts64(&rts);
317 rts = timespec64_sub(*end_time, rts);
318 if (rts.tv_sec < 0)
319 rts.tv_sec = rts.tv_nsec = 0;
320
321
322 switch (pt_type) {
323 case PT_TIMEVAL:
324 {
325 struct __kernel_old_timeval rtv;
326
327 if (sizeof(rtv) > sizeof(rtv.tv_sec) + sizeof(rtv.tv_usec))
328 memset(&rtv, 0, sizeof(rtv));
329 rtv.tv_sec = rts.tv_sec;
330 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
331 if (!copy_to_user(p, &rtv, sizeof(rtv)))
332 return ret;
333 }
334 break;
335 case PT_OLD_TIMEVAL:
336 {
337 struct old_timeval32 rtv;
338
339 rtv.tv_sec = rts.tv_sec;
340 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
341 if (!copy_to_user(p, &rtv, sizeof(rtv)))
342 return ret;
343 }
344 break;
345 case PT_TIMESPEC:
346 if (!put_timespec64(&rts, p))
347 return ret;
348 break;
349 case PT_OLD_TIMESPEC:
350 if (!put_old_timespec32(&rts, p))
351 return ret;
352 break;
353 default:
354 BUG();
355 }
356 /*
357 * If an application puts its timeval in read-only memory, we
358 * don't want the Linux-specific update to the timeval to
359 * cause a fault after the select has completed
360 * successfully. However, because we're not updating the
361 * timeval, we can't restart the system call.
362 */
363
364sticky:
365 if (ret == -ERESTARTNOHAND)
366 ret = -EINTR;
367 return ret;
368}
369
370/*
371 * Scalable version of the fd_set.
372 */
373
374typedef struct {
375 unsigned long *in, *out, *ex;
376 unsigned long *res_in, *res_out, *res_ex;
377} fd_set_bits;
378
379/*
380 * How many longwords for "nr" bits?
381 */
382#define FDS_BITPERLONG (8*sizeof(long))
383#define FDS_LONGS(nr) (((nr)+FDS_BITPERLONG-1)/FDS_BITPERLONG)
384#define FDS_BYTES(nr) (FDS_LONGS(nr)*sizeof(long))
385
386/*
387 * Use "unsigned long" accesses to let user-mode fd_set's be long-aligned.
388 */
389static inline
390int get_fd_set(unsigned long nr, void __user *ufdset, unsigned long *fdset)
391{
392 nr = FDS_BYTES(nr);
393 if (ufdset)
394 return copy_from_user(fdset, ufdset, nr) ? -EFAULT : 0;
395
396 memset(fdset, 0, nr);
397 return 0;
398}
399
400static inline unsigned long __must_check
401set_fd_set(unsigned long nr, void __user *ufdset, unsigned long *fdset)
402{
403 if (ufdset)
404 return __copy_to_user(ufdset, fdset, FDS_BYTES(nr));
405 return 0;
406}
407
408static inline
409void zero_fd_set(unsigned long nr, unsigned long *fdset)
410{
411 memset(fdset, 0, FDS_BYTES(nr));
412}
413
414#define FDS_IN(fds, n) (fds->in + n)
415#define FDS_OUT(fds, n) (fds->out + n)
416#define FDS_EX(fds, n) (fds->ex + n)
417
418#define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
419
420static int max_select_fd(unsigned long n, fd_set_bits *fds)
421{
422 unsigned long *open_fds;
423 unsigned long set;
424 int max;
425 struct fdtable *fdt;
426
427 /* handle last in-complete long-word first */
428 set = ~(~0UL << (n & (BITS_PER_LONG-1)));
429 n /= BITS_PER_LONG;
430 fdt = files_fdtable(current->files);
431 open_fds = fdt->open_fds + n;
432 max = 0;
433 if (set) {
434 set &= BITS(fds, n);
435 if (set) {
436 if (!(set & ~*open_fds))
437 goto get_max;
438 return -EBADF;
439 }
440 }
441 while (n) {
442 open_fds--;
443 n--;
444 set = BITS(fds, n);
445 if (!set)
446 continue;
447 if (set & ~*open_fds)
448 return -EBADF;
449 if (max)
450 continue;
451get_max:
452 do {
453 max++;
454 set >>= 1;
455 } while (set);
456 max += n * BITS_PER_LONG;
457 }
458
459 return max;
460}
461
462#define POLLIN_SET (EPOLLRDNORM | EPOLLRDBAND | EPOLLIN | EPOLLHUP | EPOLLERR |\
463 EPOLLNVAL)
464#define POLLOUT_SET (EPOLLWRBAND | EPOLLWRNORM | EPOLLOUT | EPOLLERR |\
465 EPOLLNVAL)
466#define POLLEX_SET (EPOLLPRI | EPOLLNVAL)
467
468static inline void wait_key_set(poll_table *wait, unsigned long in,
469 unsigned long out, unsigned long bit,
470 __poll_t ll_flag)
471{
472 wait->_key = POLLEX_SET | ll_flag;
473 if (in & bit)
474 wait->_key |= POLLIN_SET;
475 if (out & bit)
476 wait->_key |= POLLOUT_SET;
477}
478
479static int do_select(int n, fd_set_bits *fds, struct timespec64 *end_time)
480{
481 ktime_t expire, *to = NULL;
482 struct poll_wqueues table;
483 poll_table *wait;
484 int retval, i, timed_out = 0;
485 u64 slack = 0;
486 __poll_t busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
487 unsigned long busy_start = 0;
488
489 rcu_read_lock();
490 retval = max_select_fd(n, fds);
491 rcu_read_unlock();
492
493 if (retval < 0)
494 return retval;
495 n = retval;
496
497 poll_initwait(&table);
498 wait = &table.pt;
499 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
500 wait->_qproc = NULL;
501 timed_out = 1;
502 }
503
504 if (end_time && !timed_out)
505 slack = select_estimate_accuracy(end_time);
506
507 retval = 0;
508 for (;;) {
509 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
510 bool can_busy_loop = false;
511
512 inp = fds->in; outp = fds->out; exp = fds->ex;
513 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
514
515 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
516 unsigned long in, out, ex, all_bits, bit = 1, j;
517 unsigned long res_in = 0, res_out = 0, res_ex = 0;
518 __poll_t mask;
519
520 in = *inp++; out = *outp++; ex = *exp++;
521 all_bits = in | out | ex;
522 if (all_bits == 0) {
523 i += BITS_PER_LONG;
524 continue;
525 }
526
527 for (j = 0; j < BITS_PER_LONG; ++j, ++i, bit <<= 1) {
528 struct fd f;
529 if (i >= n)
530 break;
531 if (!(bit & all_bits))
532 continue;
533 mask = EPOLLNVAL;
534 f = fdget(i);
535 if (f.file) {
536 wait_key_set(wait, in, out, bit,
537 busy_flag);
538 mask = vfs_poll(f.file, wait);
539
540 fdput(f);
541 }
542 if ((mask & POLLIN_SET) && (in & bit)) {
543 res_in |= bit;
544 retval++;
545 wait->_qproc = NULL;
546 }
547 if ((mask & POLLOUT_SET) && (out & bit)) {
548 res_out |= bit;
549 retval++;
550 wait->_qproc = NULL;
551 }
552 if ((mask & POLLEX_SET) && (ex & bit)) {
553 res_ex |= bit;
554 retval++;
555 wait->_qproc = NULL;
556 }
557 /* got something, stop busy polling */
558 if (retval) {
559 can_busy_loop = false;
560 busy_flag = 0;
561
562 /*
563 * only remember a returned
564 * POLL_BUSY_LOOP if we asked for it
565 */
566 } else if (busy_flag & mask)
567 can_busy_loop = true;
568
569 }
570 if (res_in)
571 *rinp = res_in;
572 if (res_out)
573 *routp = res_out;
574 if (res_ex)
575 *rexp = res_ex;
576 cond_resched();
577 }
578 wait->_qproc = NULL;
579 if (retval || timed_out || signal_pending(current))
580 break;
581 if (table.error) {
582 retval = table.error;
583 break;
584 }
585
586 /* only if found POLL_BUSY_LOOP sockets && not out of time */
587 if (can_busy_loop && !need_resched()) {
588 if (!busy_start) {
589 busy_start = busy_loop_current_time();
590 continue;
591 }
592 if (!busy_loop_timeout(busy_start))
593 continue;
594 }
595 busy_flag = 0;
596
597 /*
598 * If this is the first loop and we have a timeout
599 * given, then we convert to ktime_t and set the to
600 * pointer to the expiry value.
601 */
602 if (end_time && !to) {
603 expire = timespec64_to_ktime(*end_time);
604 to = &expire;
605 }
606
607 if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
608 to, slack))
609 timed_out = 1;
610 }
611
612 poll_freewait(&table);
613
614 return retval;
615}
616
617/*
618 * We can actually return ERESTARTSYS instead of EINTR, but I'd
619 * like to be certain this leads to no problems. So I return
620 * EINTR just for safety.
621 *
622 * Update: ERESTARTSYS breaks at least the xview clock binary, so
623 * I'm trying ERESTARTNOHAND which restart only when you want to.
624 */
625int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
626 fd_set __user *exp, struct timespec64 *end_time)
627{
628 fd_set_bits fds;
629 void *bits;
630 int ret, max_fds;
631 size_t size, alloc_size;
632 struct fdtable *fdt;
633 /* Allocate small arguments on the stack to save memory and be faster */
634 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
635
636 ret = -EINVAL;
637 if (n < 0)
638 goto out_nofds;
639
640 /* max_fds can increase, so grab it once to avoid race */
641 rcu_read_lock();
642 fdt = files_fdtable(current->files);
643 max_fds = fdt->max_fds;
644 rcu_read_unlock();
645 if (n > max_fds)
646 n = max_fds;
647
648 /*
649 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
650 * since we used fdset we need to allocate memory in units of
651 * long-words.
652 */
653 size = FDS_BYTES(n);
654 bits = stack_fds;
655 if (size > sizeof(stack_fds) / 6) {
656 /* Not enough space in on-stack array; must use kmalloc */
657 ret = -ENOMEM;
658 if (size > (SIZE_MAX / 6))
659 goto out_nofds;
660
661 alloc_size = 6 * size;
662 bits = kvmalloc(alloc_size, GFP_KERNEL);
663 if (!bits)
664 goto out_nofds;
665 }
666 fds.in = bits;
667 fds.out = bits + size;
668 fds.ex = bits + 2*size;
669 fds.res_in = bits + 3*size;
670 fds.res_out = bits + 4*size;
671 fds.res_ex = bits + 5*size;
672
673 if ((ret = get_fd_set(n, inp, fds.in)) ||
674 (ret = get_fd_set(n, outp, fds.out)) ||
675 (ret = get_fd_set(n, exp, fds.ex)))
676 goto out;
677 zero_fd_set(n, fds.res_in);
678 zero_fd_set(n, fds.res_out);
679 zero_fd_set(n, fds.res_ex);
680
681 ret = do_select(n, &fds, end_time);
682
683 if (ret < 0)
684 goto out;
685 if (!ret) {
686 ret = -ERESTARTNOHAND;
687 if (signal_pending(current))
688 goto out;
689 ret = 0;
690 }
691
692 if (set_fd_set(n, inp, fds.res_in) ||
693 set_fd_set(n, outp, fds.res_out) ||
694 set_fd_set(n, exp, fds.res_ex))
695 ret = -EFAULT;
696
697out:
698 if (bits != stack_fds)
699 kvfree(bits);
700out_nofds:
701 return ret;
702}
703
704static int kern_select(int n, fd_set __user *inp, fd_set __user *outp,
705 fd_set __user *exp, struct __kernel_old_timeval __user *tvp)
706{
707 struct timespec64 end_time, *to = NULL;
708 struct __kernel_old_timeval tv;
709 int ret;
710
711 if (tvp) {
712 if (copy_from_user(&tv, tvp, sizeof(tv)))
713 return -EFAULT;
714
715 to = &end_time;
716 if (poll_select_set_timeout(to,
717 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
718 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
719 return -EINVAL;
720 }
721
722 ret = core_sys_select(n, inp, outp, exp, to);
723 return poll_select_finish(&end_time, tvp, PT_TIMEVAL, ret);
724}
725
726SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
727 fd_set __user *, exp, struct __kernel_old_timeval __user *, tvp)
728{
729 return kern_select(n, inp, outp, exp, tvp);
730}
731
732static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
733 fd_set __user *exp, void __user *tsp,
734 const sigset_t __user *sigmask, size_t sigsetsize,
735 enum poll_time_type type)
736{
737 struct timespec64 ts, end_time, *to = NULL;
738 int ret;
739
740 if (tsp) {
741 switch (type) {
742 case PT_TIMESPEC:
743 if (get_timespec64(&ts, tsp))
744 return -EFAULT;
745 break;
746 case PT_OLD_TIMESPEC:
747 if (get_old_timespec32(&ts, tsp))
748 return -EFAULT;
749 break;
750 default:
751 BUG();
752 }
753
754 to = &end_time;
755 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
756 return -EINVAL;
757 }
758
759 ret = set_user_sigmask(sigmask, sigsetsize);
760 if (ret)
761 return ret;
762
763 ret = core_sys_select(n, inp, outp, exp, to);
764 return poll_select_finish(&end_time, tsp, type, ret);
765}
766
767/*
768 * Most architectures can't handle 7-argument syscalls. So we provide a
769 * 6-argument version where the sixth argument is a pointer to a structure
770 * which has a pointer to the sigset_t itself followed by a size_t containing
771 * the sigset size.
772 */
773struct sigset_argpack {
774 sigset_t __user *p;
775 size_t size;
776};
777
778static inline int get_sigset_argpack(struct sigset_argpack *to,
779 struct sigset_argpack __user *from)
780{
781 // the path is hot enough for overhead of copy_from_user() to matter
782 if (from) {
783 if (!user_read_access_begin(from, sizeof(*from)))
784 return -EFAULT;
785 unsafe_get_user(to->p, &from->p, Efault);
786 unsafe_get_user(to->size, &from->size, Efault);
787 user_read_access_end();
788 }
789 return 0;
790Efault:
791 user_access_end();
792 return -EFAULT;
793}
794
795SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
796 fd_set __user *, exp, struct __kernel_timespec __user *, tsp,
797 void __user *, sig)
798{
799 struct sigset_argpack x = {NULL, 0};
800
801 if (get_sigset_argpack(&x, sig))
802 return -EFAULT;
803
804 return do_pselect(n, inp, outp, exp, tsp, x.p, x.size, PT_TIMESPEC);
805}
806
807#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
808
809SYSCALL_DEFINE6(pselect6_time32, int, n, fd_set __user *, inp, fd_set __user *, outp,
810 fd_set __user *, exp, struct old_timespec32 __user *, tsp,
811 void __user *, sig)
812{
813 struct sigset_argpack x = {NULL, 0};
814
815 if (get_sigset_argpack(&x, sig))
816 return -EFAULT;
817
818 return do_pselect(n, inp, outp, exp, tsp, x.p, x.size, PT_OLD_TIMESPEC);
819}
820
821#endif
822
823#ifdef __ARCH_WANT_SYS_OLD_SELECT
824struct sel_arg_struct {
825 unsigned long n;
826 fd_set __user *inp, *outp, *exp;
827 struct __kernel_old_timeval __user *tvp;
828};
829
830SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg)
831{
832 struct sel_arg_struct a;
833
834 if (copy_from_user(&a, arg, sizeof(a)))
835 return -EFAULT;
836 return kern_select(a.n, a.inp, a.outp, a.exp, a.tvp);
837}
838#endif
839
840struct poll_list {
841 struct poll_list *next;
842 int len;
843 struct pollfd entries[];
844};
845
846#define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
847
848/*
849 * Fish for pollable events on the pollfd->fd file descriptor. We're only
850 * interested in events matching the pollfd->events mask, and the result
851 * matching that mask is both recorded in pollfd->revents and returned. The
852 * pwait poll_table will be used by the fd-provided poll handler for waiting,
853 * if pwait->_qproc is non-NULL.
854 */
855static inline __poll_t do_pollfd(struct pollfd *pollfd, poll_table *pwait,
856 bool *can_busy_poll,
857 __poll_t busy_flag)
858{
859 int fd = pollfd->fd;
860 __poll_t mask = 0, filter;
861 struct fd f;
862
863 if (fd < 0)
864 goto out;
865 mask = EPOLLNVAL;
866 f = fdget(fd);
867 if (!f.file)
868 goto out;
869
870 /* userland u16 ->events contains POLL... bitmap */
871 filter = demangle_poll(pollfd->events) | EPOLLERR | EPOLLHUP;
872 pwait->_key = filter | busy_flag;
873 mask = vfs_poll(f.file, pwait);
874 if (mask & busy_flag)
875 *can_busy_poll = true;
876 mask &= filter; /* Mask out unneeded events. */
877 fdput(f);
878
879out:
880 /* ... and so does ->revents */
881 pollfd->revents = mangle_poll(mask);
882 return mask;
883}
884
885static int do_poll(struct poll_list *list, struct poll_wqueues *wait,
886 struct timespec64 *end_time)
887{
888 poll_table* pt = &wait->pt;
889 ktime_t expire, *to = NULL;
890 int timed_out = 0, count = 0;
891 u64 slack = 0;
892 __poll_t busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
893 unsigned long busy_start = 0;
894
895 /* Optimise the no-wait case */
896 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
897 pt->_qproc = NULL;
898 timed_out = 1;
899 }
900
901 if (end_time && !timed_out)
902 slack = select_estimate_accuracy(end_time);
903
904 for (;;) {
905 struct poll_list *walk;
906 bool can_busy_loop = false;
907
908 for (walk = list; walk != NULL; walk = walk->next) {
909 struct pollfd * pfd, * pfd_end;
910
911 pfd = walk->entries;
912 pfd_end = pfd + walk->len;
913 for (; pfd != pfd_end; pfd++) {
914 /*
915 * Fish for events. If we found one, record it
916 * and kill poll_table->_qproc, so we don't
917 * needlessly register any other waiters after
918 * this. They'll get immediately deregistered
919 * when we break out and return.
920 */
921 if (do_pollfd(pfd, pt, &can_busy_loop,
922 busy_flag)) {
923 count++;
924 pt->_qproc = NULL;
925 /* found something, stop busy polling */
926 busy_flag = 0;
927 can_busy_loop = false;
928 }
929 }
930 }
931 /*
932 * All waiters have already been registered, so don't provide
933 * a poll_table->_qproc to them on the next loop iteration.
934 */
935 pt->_qproc = NULL;
936 if (!count) {
937 count = wait->error;
938 if (signal_pending(current))
939 count = -ERESTARTNOHAND;
940 }
941 if (count || timed_out)
942 break;
943
944 /* only if found POLL_BUSY_LOOP sockets && not out of time */
945 if (can_busy_loop && !need_resched()) {
946 if (!busy_start) {
947 busy_start = busy_loop_current_time();
948 continue;
949 }
950 if (!busy_loop_timeout(busy_start))
951 continue;
952 }
953 busy_flag = 0;
954
955 /*
956 * If this is the first loop and we have a timeout
957 * given, then we convert to ktime_t and set the to
958 * pointer to the expiry value.
959 */
960 if (end_time && !to) {
961 expire = timespec64_to_ktime(*end_time);
962 to = &expire;
963 }
964
965 if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
966 timed_out = 1;
967 }
968 return count;
969}
970
971#define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \
972 sizeof(struct pollfd))
973
974static int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
975 struct timespec64 *end_time)
976{
977 struct poll_wqueues table;
978 int err = -EFAULT, fdcount, len;
979 /* Allocate small arguments on the stack to save memory and be
980 faster - use long to make sure the buffer is aligned properly
981 on 64 bit archs to avoid unaligned access */
982 long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
983 struct poll_list *const head = (struct poll_list *)stack_pps;
984 struct poll_list *walk = head;
985 unsigned long todo = nfds;
986
987 if (nfds > rlimit(RLIMIT_NOFILE))
988 return -EINVAL;
989
990 len = min_t(unsigned int, nfds, N_STACK_PPS);
991 for (;;) {
992 walk->next = NULL;
993 walk->len = len;
994 if (!len)
995 break;
996
997 if (copy_from_user(walk->entries, ufds + nfds-todo,
998 sizeof(struct pollfd) * walk->len))
999 goto out_fds;
1000
1001 todo -= walk->len;
1002 if (!todo)
1003 break;
1004
1005 len = min(todo, POLLFD_PER_PAGE);
1006 walk = walk->next = kmalloc(struct_size(walk, entries, len),
1007 GFP_KERNEL);
1008 if (!walk) {
1009 err = -ENOMEM;
1010 goto out_fds;
1011 }
1012 }
1013
1014 poll_initwait(&table);
1015 fdcount = do_poll(head, &table, end_time);
1016 poll_freewait(&table);
1017
1018 if (!user_write_access_begin(ufds, nfds * sizeof(*ufds)))
1019 goto out_fds;
1020
1021 for (walk = head; walk; walk = walk->next) {
1022 struct pollfd *fds = walk->entries;
1023 int j;
1024
1025 for (j = walk->len; j; fds++, ufds++, j--)
1026 unsafe_put_user(fds->revents, &ufds->revents, Efault);
1027 }
1028 user_write_access_end();
1029
1030 err = fdcount;
1031out_fds:
1032 walk = head->next;
1033 while (walk) {
1034 struct poll_list *pos = walk;
1035 walk = walk->next;
1036 kfree(pos);
1037 }
1038
1039 return err;
1040
1041Efault:
1042 user_write_access_end();
1043 err = -EFAULT;
1044 goto out_fds;
1045}
1046
1047static long do_restart_poll(struct restart_block *restart_block)
1048{
1049 struct pollfd __user *ufds = restart_block->poll.ufds;
1050 int nfds = restart_block->poll.nfds;
1051 struct timespec64 *to = NULL, end_time;
1052 int ret;
1053
1054 if (restart_block->poll.has_timeout) {
1055 end_time.tv_sec = restart_block->poll.tv_sec;
1056 end_time.tv_nsec = restart_block->poll.tv_nsec;
1057 to = &end_time;
1058 }
1059
1060 ret = do_sys_poll(ufds, nfds, to);
1061
1062 if (ret == -ERESTARTNOHAND)
1063 ret = set_restart_fn(restart_block, do_restart_poll);
1064
1065 return ret;
1066}
1067
1068SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
1069 int, timeout_msecs)
1070{
1071 struct timespec64 end_time, *to = NULL;
1072 int ret;
1073
1074 if (timeout_msecs >= 0) {
1075 to = &end_time;
1076 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
1077 NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
1078 }
1079
1080 ret = do_sys_poll(ufds, nfds, to);
1081
1082 if (ret == -ERESTARTNOHAND) {
1083 struct restart_block *restart_block;
1084
1085 restart_block = ¤t->restart_block;
1086 restart_block->poll.ufds = ufds;
1087 restart_block->poll.nfds = nfds;
1088
1089 if (timeout_msecs >= 0) {
1090 restart_block->poll.tv_sec = end_time.tv_sec;
1091 restart_block->poll.tv_nsec = end_time.tv_nsec;
1092 restart_block->poll.has_timeout = 1;
1093 } else
1094 restart_block->poll.has_timeout = 0;
1095
1096 ret = set_restart_fn(restart_block, do_restart_poll);
1097 }
1098 return ret;
1099}
1100
1101SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
1102 struct __kernel_timespec __user *, tsp, const sigset_t __user *, sigmask,
1103 size_t, sigsetsize)
1104{
1105 struct timespec64 ts, end_time, *to = NULL;
1106 int ret;
1107
1108 if (tsp) {
1109 if (get_timespec64(&ts, tsp))
1110 return -EFAULT;
1111
1112 to = &end_time;
1113 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1114 return -EINVAL;
1115 }
1116
1117 ret = set_user_sigmask(sigmask, sigsetsize);
1118 if (ret)
1119 return ret;
1120
1121 ret = do_sys_poll(ufds, nfds, to);
1122 return poll_select_finish(&end_time, tsp, PT_TIMESPEC, ret);
1123}
1124
1125#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
1126
1127SYSCALL_DEFINE5(ppoll_time32, struct pollfd __user *, ufds, unsigned int, nfds,
1128 struct old_timespec32 __user *, tsp, const sigset_t __user *, sigmask,
1129 size_t, sigsetsize)
1130{
1131 struct timespec64 ts, end_time, *to = NULL;
1132 int ret;
1133
1134 if (tsp) {
1135 if (get_old_timespec32(&ts, tsp))
1136 return -EFAULT;
1137
1138 to = &end_time;
1139 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1140 return -EINVAL;
1141 }
1142
1143 ret = set_user_sigmask(sigmask, sigsetsize);
1144 if (ret)
1145 return ret;
1146
1147 ret = do_sys_poll(ufds, nfds, to);
1148 return poll_select_finish(&end_time, tsp, PT_OLD_TIMESPEC, ret);
1149}
1150#endif
1151
1152#ifdef CONFIG_COMPAT
1153#define __COMPAT_NFDBITS (8 * sizeof(compat_ulong_t))
1154
1155/*
1156 * Ooo, nasty. We need here to frob 32-bit unsigned longs to
1157 * 64-bit unsigned longs.
1158 */
1159static
1160int compat_get_fd_set(unsigned long nr, compat_ulong_t __user *ufdset,
1161 unsigned long *fdset)
1162{
1163 if (ufdset) {
1164 return compat_get_bitmap(fdset, ufdset, nr);
1165 } else {
1166 zero_fd_set(nr, fdset);
1167 return 0;
1168 }
1169}
1170
1171static
1172int compat_set_fd_set(unsigned long nr, compat_ulong_t __user *ufdset,
1173 unsigned long *fdset)
1174{
1175 if (!ufdset)
1176 return 0;
1177 return compat_put_bitmap(ufdset, fdset, nr);
1178}
1179
1180
1181/*
1182 * This is a virtual copy of sys_select from fs/select.c and probably
1183 * should be compared to it from time to time
1184 */
1185
1186/*
1187 * We can actually return ERESTARTSYS instead of EINTR, but I'd
1188 * like to be certain this leads to no problems. So I return
1189 * EINTR just for safety.
1190 *
1191 * Update: ERESTARTSYS breaks at least the xview clock binary, so
1192 * I'm trying ERESTARTNOHAND which restart only when you want to.
1193 */
1194static int compat_core_sys_select(int n, compat_ulong_t __user *inp,
1195 compat_ulong_t __user *outp, compat_ulong_t __user *exp,
1196 struct timespec64 *end_time)
1197{
1198 fd_set_bits fds;
1199 void *bits;
1200 int size, max_fds, ret = -EINVAL;
1201 struct fdtable *fdt;
1202 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
1203
1204 if (n < 0)
1205 goto out_nofds;
1206
1207 /* max_fds can increase, so grab it once to avoid race */
1208 rcu_read_lock();
1209 fdt = files_fdtable(current->files);
1210 max_fds = fdt->max_fds;
1211 rcu_read_unlock();
1212 if (n > max_fds)
1213 n = max_fds;
1214
1215 /*
1216 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
1217 * since we used fdset we need to allocate memory in units of
1218 * long-words.
1219 */
1220 size = FDS_BYTES(n);
1221 bits = stack_fds;
1222 if (size > sizeof(stack_fds) / 6) {
1223 bits = kmalloc_array(6, size, GFP_KERNEL);
1224 ret = -ENOMEM;
1225 if (!bits)
1226 goto out_nofds;
1227 }
1228 fds.in = (unsigned long *) bits;
1229 fds.out = (unsigned long *) (bits + size);
1230 fds.ex = (unsigned long *) (bits + 2*size);
1231 fds.res_in = (unsigned long *) (bits + 3*size);
1232 fds.res_out = (unsigned long *) (bits + 4*size);
1233 fds.res_ex = (unsigned long *) (bits + 5*size);
1234
1235 if ((ret = compat_get_fd_set(n, inp, fds.in)) ||
1236 (ret = compat_get_fd_set(n, outp, fds.out)) ||
1237 (ret = compat_get_fd_set(n, exp, fds.ex)))
1238 goto out;
1239 zero_fd_set(n, fds.res_in);
1240 zero_fd_set(n, fds.res_out);
1241 zero_fd_set(n, fds.res_ex);
1242
1243 ret = do_select(n, &fds, end_time);
1244
1245 if (ret < 0)
1246 goto out;
1247 if (!ret) {
1248 ret = -ERESTARTNOHAND;
1249 if (signal_pending(current))
1250 goto out;
1251 ret = 0;
1252 }
1253
1254 if (compat_set_fd_set(n, inp, fds.res_in) ||
1255 compat_set_fd_set(n, outp, fds.res_out) ||
1256 compat_set_fd_set(n, exp, fds.res_ex))
1257 ret = -EFAULT;
1258out:
1259 if (bits != stack_fds)
1260 kfree(bits);
1261out_nofds:
1262 return ret;
1263}
1264
1265static int do_compat_select(int n, compat_ulong_t __user *inp,
1266 compat_ulong_t __user *outp, compat_ulong_t __user *exp,
1267 struct old_timeval32 __user *tvp)
1268{
1269 struct timespec64 end_time, *to = NULL;
1270 struct old_timeval32 tv;
1271 int ret;
1272
1273 if (tvp) {
1274 if (copy_from_user(&tv, tvp, sizeof(tv)))
1275 return -EFAULT;
1276
1277 to = &end_time;
1278 if (poll_select_set_timeout(to,
1279 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
1280 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
1281 return -EINVAL;
1282 }
1283
1284 ret = compat_core_sys_select(n, inp, outp, exp, to);
1285 return poll_select_finish(&end_time, tvp, PT_OLD_TIMEVAL, ret);
1286}
1287
1288COMPAT_SYSCALL_DEFINE5(select, int, n, compat_ulong_t __user *, inp,
1289 compat_ulong_t __user *, outp, compat_ulong_t __user *, exp,
1290 struct old_timeval32 __user *, tvp)
1291{
1292 return do_compat_select(n, inp, outp, exp, tvp);
1293}
1294
1295struct compat_sel_arg_struct {
1296 compat_ulong_t n;
1297 compat_uptr_t inp;
1298 compat_uptr_t outp;
1299 compat_uptr_t exp;
1300 compat_uptr_t tvp;
1301};
1302
1303COMPAT_SYSCALL_DEFINE1(old_select, struct compat_sel_arg_struct __user *, arg)
1304{
1305 struct compat_sel_arg_struct a;
1306
1307 if (copy_from_user(&a, arg, sizeof(a)))
1308 return -EFAULT;
1309 return do_compat_select(a.n, compat_ptr(a.inp), compat_ptr(a.outp),
1310 compat_ptr(a.exp), compat_ptr(a.tvp));
1311}
1312
1313static long do_compat_pselect(int n, compat_ulong_t __user *inp,
1314 compat_ulong_t __user *outp, compat_ulong_t __user *exp,
1315 void __user *tsp, compat_sigset_t __user *sigmask,
1316 compat_size_t sigsetsize, enum poll_time_type type)
1317{
1318 struct timespec64 ts, end_time, *to = NULL;
1319 int ret;
1320
1321 if (tsp) {
1322 switch (type) {
1323 case PT_OLD_TIMESPEC:
1324 if (get_old_timespec32(&ts, tsp))
1325 return -EFAULT;
1326 break;
1327 case PT_TIMESPEC:
1328 if (get_timespec64(&ts, tsp))
1329 return -EFAULT;
1330 break;
1331 default:
1332 BUG();
1333 }
1334
1335 to = &end_time;
1336 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1337 return -EINVAL;
1338 }
1339
1340 ret = set_compat_user_sigmask(sigmask, sigsetsize);
1341 if (ret)
1342 return ret;
1343
1344 ret = compat_core_sys_select(n, inp, outp, exp, to);
1345 return poll_select_finish(&end_time, tsp, type, ret);
1346}
1347
1348struct compat_sigset_argpack {
1349 compat_uptr_t p;
1350 compat_size_t size;
1351};
1352static inline int get_compat_sigset_argpack(struct compat_sigset_argpack *to,
1353 struct compat_sigset_argpack __user *from)
1354{
1355 if (from) {
1356 if (!user_read_access_begin(from, sizeof(*from)))
1357 return -EFAULT;
1358 unsafe_get_user(to->p, &from->p, Efault);
1359 unsafe_get_user(to->size, &from->size, Efault);
1360 user_read_access_end();
1361 }
1362 return 0;
1363Efault:
1364 user_access_end();
1365 return -EFAULT;
1366}
1367
1368COMPAT_SYSCALL_DEFINE6(pselect6_time64, int, n, compat_ulong_t __user *, inp,
1369 compat_ulong_t __user *, outp, compat_ulong_t __user *, exp,
1370 struct __kernel_timespec __user *, tsp, void __user *, sig)
1371{
1372 struct compat_sigset_argpack x = {0, 0};
1373
1374 if (get_compat_sigset_argpack(&x, sig))
1375 return -EFAULT;
1376
1377 return do_compat_pselect(n, inp, outp, exp, tsp, compat_ptr(x.p),
1378 x.size, PT_TIMESPEC);
1379}
1380
1381#if defined(CONFIG_COMPAT_32BIT_TIME)
1382
1383COMPAT_SYSCALL_DEFINE6(pselect6_time32, int, n, compat_ulong_t __user *, inp,
1384 compat_ulong_t __user *, outp, compat_ulong_t __user *, exp,
1385 struct old_timespec32 __user *, tsp, void __user *, sig)
1386{
1387 struct compat_sigset_argpack x = {0, 0};
1388
1389 if (get_compat_sigset_argpack(&x, sig))
1390 return -EFAULT;
1391
1392 return do_compat_pselect(n, inp, outp, exp, tsp, compat_ptr(x.p),
1393 x.size, PT_OLD_TIMESPEC);
1394}
1395
1396#endif
1397
1398#if defined(CONFIG_COMPAT_32BIT_TIME)
1399COMPAT_SYSCALL_DEFINE5(ppoll_time32, struct pollfd __user *, ufds,
1400 unsigned int, nfds, struct old_timespec32 __user *, tsp,
1401 const compat_sigset_t __user *, sigmask, compat_size_t, sigsetsize)
1402{
1403 struct timespec64 ts, end_time, *to = NULL;
1404 int ret;
1405
1406 if (tsp) {
1407 if (get_old_timespec32(&ts, tsp))
1408 return -EFAULT;
1409
1410 to = &end_time;
1411 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1412 return -EINVAL;
1413 }
1414
1415 ret = set_compat_user_sigmask(sigmask, sigsetsize);
1416 if (ret)
1417 return ret;
1418
1419 ret = do_sys_poll(ufds, nfds, to);
1420 return poll_select_finish(&end_time, tsp, PT_OLD_TIMESPEC, ret);
1421}
1422#endif
1423
1424/* New compat syscall for 64 bit time_t*/
1425COMPAT_SYSCALL_DEFINE5(ppoll_time64, struct pollfd __user *, ufds,
1426 unsigned int, nfds, struct __kernel_timespec __user *, tsp,
1427 const compat_sigset_t __user *, sigmask, compat_size_t, sigsetsize)
1428{
1429 struct timespec64 ts, end_time, *to = NULL;
1430 int ret;
1431
1432 if (tsp) {
1433 if (get_timespec64(&ts, tsp))
1434 return -EFAULT;
1435
1436 to = &end_time;
1437 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1438 return -EINVAL;
1439 }
1440
1441 ret = set_compat_user_sigmask(sigmask, sigsetsize);
1442 if (ret)
1443 return ret;
1444
1445 ret = do_sys_poll(ufds, nfds, to);
1446 return poll_select_finish(&end_time, tsp, PT_TIMESPEC, ret);
1447}
1448
1449#endif