projects/linux·blob v6.12 · refs

1449 lines · 34.8 KiB

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 u64 slack = current->timer_slack_ns;
81
82 if (slack == 0)
83 return 0;
84
85 ktime_get_ts64(&now);
86 now = timespec64_sub(*tv, now);
87 ret = __estimate_accuracy(&now);
88 if (ret < slack)
89 return slack;
90 return ret;
91}
92
93
94
95struct poll_table_page {
96 struct poll_table_page * next;
97 struct poll_table_entry * entry;
98 struct poll_table_entry entries[];
99};
100
101#define POLL_TABLE_FULL(table) \
102 ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
103
104/*
105 * Ok, Peter made a complicated, but straightforward multiple_wait() function.
106 * I have rewritten this, taking some shortcuts: This code may not be easy to
107 * follow, but it should be free of race-conditions, and it's practical. If you
108 * understand what I'm doing here, then you understand how the linux
109 * sleep/wakeup mechanism works.
110 *
111 * Two very simple procedures, poll_wait() and poll_freewait() make all the
112 * work. poll_wait() is an inline-function defined in <linux/poll.h>,
113 * as all select/poll functions have to call it to add an entry to the
114 * poll table.
115 */
116static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
117 poll_table *p);
118
119void poll_initwait(struct poll_wqueues *pwq)
120{
121 init_poll_funcptr(&pwq->pt, __pollwait);
122 pwq->polling_task = current;
123 pwq->triggered = 0;
124 pwq->error = 0;
125 pwq->table = NULL;
126 pwq->inline_index = 0;
127}
128EXPORT_SYMBOL(poll_initwait);
129
130static void free_poll_entry(struct poll_table_entry *entry)
131{
132 remove_wait_queue(entry->wait_address, &entry->wait);
133 fput(entry->filp);
134}
135
136void poll_freewait(struct poll_wqueues *pwq)
137{
138 struct poll_table_page * p = pwq->table;
139 int i;
140 for (i = 0; i < pwq->inline_index; i++)
141 free_poll_entry(pwq->inline_entries + i);
142 while (p) {
143 struct poll_table_entry * entry;
144 struct poll_table_page *old;
145
146 entry = p->entry;
147 do {
148 entry--;
149 free_poll_entry(entry);
150 } while (entry > p->entries);
151 old = p;
152 p = p->next;
153 free_page((unsigned long) old);
154 }
155}
156EXPORT_SYMBOL(poll_freewait);
157
158static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
159{
160 struct poll_table_page *table = p->table;
161
162 if (p->inline_index < N_INLINE_POLL_ENTRIES)
163 return p->inline_entries + p->inline_index++;
164
165 if (!table || POLL_TABLE_FULL(table)) {
166 struct poll_table_page *new_table;
167
168 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
169 if (!new_table) {
170 p->error = -ENOMEM;
171 return NULL;
172 }
173 new_table->entry = new_table->entries;
174 new_table->next = table;
175 p->table = new_table;
176 table = new_table;
177 }
178
179 return table->entry++;
180}
181
182static int __pollwake(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
183{
184 struct poll_wqueues *pwq = wait->private;
185 DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
186
187 /*
188 * Although this function is called under waitqueue lock, LOCK
189 * doesn't imply write barrier and the users expect write
190 * barrier semantics on wakeup functions. The following
191 * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
192 * and is paired with smp_store_mb() in poll_schedule_timeout.
193 */
194 smp_wmb();
195 pwq->triggered = 1;
196
197 /*
198 * Perform the default wake up operation using a dummy
199 * waitqueue.
200 *
201 * TODO: This is hacky but there currently is no interface to
202 * pass in @sync. @sync is scheduled to be removed and once
203 * that happens, wake_up_process() can be used directly.
204 */
205 return default_wake_function(&dummy_wait, mode, sync, key);
206}
207
208static int pollwake(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
209{
210 struct poll_table_entry *entry;
211
212 entry = container_of(wait, struct poll_table_entry, wait);
213 if (key && !(key_to_poll(key) & entry->key))
214 return 0;
215 return __pollwake(wait, mode, sync, key);
216}
217
218/* Add a new entry */
219static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
220 poll_table *p)
221{
222 struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
223 struct poll_table_entry *entry = poll_get_entry(pwq);
224 if (!entry)
225 return;
226 entry->filp = get_file(filp);
227 entry->wait_address = wait_address;
228 entry->key = p->_key;
229 init_waitqueue_func_entry(&entry->wait, pollwake);
230 entry->wait.private = pwq;
231 add_wait_queue(wait_address, &entry->wait);
232}
233
234static int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
235 ktime_t *expires, unsigned long slack)
236{
237 int rc = -EINTR;
238
239 set_current_state(state);
240 if (!pwq->triggered)
241 rc = schedule_hrtimeout_range(expires, slack, HRTIMER_MODE_ABS);
242 __set_current_state(TASK_RUNNING);
243
244 /*
245 * Prepare for the next iteration.
246 *
247 * The following smp_store_mb() serves two purposes. First, it's
248 * the counterpart rmb of the wmb in pollwake() such that data
249 * written before wake up is always visible after wake up.
250 * Second, the full barrier guarantees that triggered clearing
251 * doesn't pass event check of the next iteration. Note that
252 * this problem doesn't exist for the first iteration as
253 * add_wait_queue() has full barrier semantics.
254 */
255 smp_store_mb(pwq->triggered, 0);
256
257 return rc;
258}
259
260/**
261 * poll_select_set_timeout - helper function to setup the timeout value
262 * @to: pointer to timespec64 variable for the final timeout
263 * @sec: seconds (from user space)
264 * @nsec: nanoseconds (from user space)
265 *
266 * Note, we do not use a timespec for the user space value here, That
267 * way we can use the function for timeval and compat interfaces as well.
268 *
269 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
270 */
271int poll_select_set_timeout(struct timespec64 *to, time64_t sec, long nsec)
272{
273 struct timespec64 ts = {.tv_sec = sec, .tv_nsec = nsec};
274
275 if (!timespec64_valid(&ts))
276 return -EINVAL;
277
278 /* Optimize for the zero timeout value here */
279 if (!sec && !nsec) {
280 to->tv_sec = to->tv_nsec = 0;
281 } else {
282 ktime_get_ts64(to);
283 *to = timespec64_add_safe(*to, ts);
284 }
285 return 0;
286}
287
288enum poll_time_type {
289 PT_TIMEVAL = 0,
290 PT_OLD_TIMEVAL = 1,
291 PT_TIMESPEC = 2,
292 PT_OLD_TIMESPEC = 3,
293};
294
295static int poll_select_finish(struct timespec64 *end_time,
296 void __user *p,
297 enum poll_time_type pt_type, int ret)
298{
299 struct timespec64 rts;
300
301 restore_saved_sigmask_unless(ret == -ERESTARTNOHAND);
302
303 if (!p)
304 return ret;
305
306 if (current->personality & STICKY_TIMEOUTS)
307 goto sticky;
308
309 /* No update for zero timeout */
310 if (!end_time->tv_sec && !end_time->tv_nsec)
311 return ret;
312
313 ktime_get_ts64(&rts);
314 rts = timespec64_sub(*end_time, rts);
315 if (rts.tv_sec < 0)
316 rts.tv_sec = rts.tv_nsec = 0;
317
318
319 switch (pt_type) {
320 case PT_TIMEVAL:
321 {
322 struct __kernel_old_timeval rtv;
323
324 if (sizeof(rtv) > sizeof(rtv.tv_sec) + sizeof(rtv.tv_usec))
325 memset(&rtv, 0, sizeof(rtv));
326 rtv.tv_sec = rts.tv_sec;
327 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
328 if (!copy_to_user(p, &rtv, sizeof(rtv)))
329 return ret;
330 }
331 break;
332 case PT_OLD_TIMEVAL:
333 {
334 struct old_timeval32 rtv;
335
336 rtv.tv_sec = rts.tv_sec;
337 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
338 if (!copy_to_user(p, &rtv, sizeof(rtv)))
339 return ret;
340 }
341 break;
342 case PT_TIMESPEC:
343 if (!put_timespec64(&rts, p))
344 return ret;
345 break;
346 case PT_OLD_TIMESPEC:
347 if (!put_old_timespec32(&rts, p))
348 return ret;
349 break;
350 default:
351 BUG();
352 }
353 /*
354 * If an application puts its timeval in read-only memory, we
355 * don't want the Linux-specific update to the timeval to
356 * cause a fault after the select has completed
357 * successfully. However, because we're not updating the
358 * timeval, we can't restart the system call.
359 */
360
361sticky:
362 if (ret == -ERESTARTNOHAND)
363 ret = -EINTR;
364 return ret;
365}
366
367/*
368 * Scalable version of the fd_set.
369 */
370
371typedef struct {
372 unsigned long *in, *out, *ex;
373 unsigned long *res_in, *res_out, *res_ex;
374} fd_set_bits;
375
376/*
377 * How many longwords for "nr" bits?
378 */
379#define FDS_BITPERLONG (8*sizeof(long))
380#define FDS_LONGS(nr) (((nr)+FDS_BITPERLONG-1)/FDS_BITPERLONG)
381#define FDS_BYTES(nr) (FDS_LONGS(nr)*sizeof(long))
382
383/*
384 * Use "unsigned long" accesses to let user-mode fd_set's be long-aligned.
385 */
386static inline
387int get_fd_set(unsigned long nr, void __user *ufdset, unsigned long *fdset)
388{
389 nr = FDS_BYTES(nr);
390 if (ufdset)
391 return copy_from_user(fdset, ufdset, nr) ? -EFAULT : 0;
392
393 memset(fdset, 0, nr);
394 return 0;
395}
396
397static inline unsigned long __must_check
398set_fd_set(unsigned long nr, void __user *ufdset, unsigned long *fdset)
399{
400 if (ufdset)
401 return __copy_to_user(ufdset, fdset, FDS_BYTES(nr));
402 return 0;
403}
404
405static inline
406void zero_fd_set(unsigned long nr, unsigned long *fdset)
407{
408 memset(fdset, 0, FDS_BYTES(nr));
409}
410
411#define FDS_IN(fds, n) (fds->in + n)
412#define FDS_OUT(fds, n) (fds->out + n)
413#define FDS_EX(fds, n) (fds->ex + n)
414
415#define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
416
417static int max_select_fd(unsigned long n, fd_set_bits *fds)
418{
419 unsigned long *open_fds;
420 unsigned long set;
421 int max;
422 struct fdtable *fdt;
423
424 /* handle last in-complete long-word first */
425 set = ~(~0UL << (n & (BITS_PER_LONG-1)));
426 n /= BITS_PER_LONG;
427 fdt = files_fdtable(current->files);
428 open_fds = fdt->open_fds + n;
429 max = 0;
430 if (set) {
431 set &= BITS(fds, n);
432 if (set) {
433 if (!(set & ~*open_fds))
434 goto get_max;
435 return -EBADF;
436 }
437 }
438 while (n) {
439 open_fds--;
440 n--;
441 set = BITS(fds, n);
442 if (!set)
443 continue;
444 if (set & ~*open_fds)
445 return -EBADF;
446 if (max)
447 continue;
448get_max:
449 do {
450 max++;
451 set >>= 1;
452 } while (set);
453 max += n * BITS_PER_LONG;
454 }
455
456 return max;
457}
458
459#define POLLIN_SET (EPOLLRDNORM | EPOLLRDBAND | EPOLLIN | EPOLLHUP | EPOLLERR |\
460 EPOLLNVAL)
461#define POLLOUT_SET (EPOLLWRBAND | EPOLLWRNORM | EPOLLOUT | EPOLLERR |\
462 EPOLLNVAL)
463#define POLLEX_SET (EPOLLPRI | EPOLLNVAL)
464
465static inline void wait_key_set(poll_table *wait, unsigned long in,
466 unsigned long out, unsigned long bit,
467 __poll_t ll_flag)
468{
469 wait->_key = POLLEX_SET | ll_flag;
470 if (in & bit)
471 wait->_key |= POLLIN_SET;
472 if (out & bit)
473 wait->_key |= POLLOUT_SET;
474}
475
476static noinline_for_stack int do_select(int n, fd_set_bits *fds, struct timespec64 *end_time)
477{
478 ktime_t expire, *to = NULL;
479 struct poll_wqueues table;
480 poll_table *wait;
481 int retval, i, timed_out = 0;
482 u64 slack = 0;
483 __poll_t busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
484 unsigned long busy_start = 0;
485
486 rcu_read_lock();
487 retval = max_select_fd(n, fds);
488 rcu_read_unlock();
489
490 if (retval < 0)
491 return retval;
492 n = retval;
493
494 poll_initwait(&table);
495 wait = &table.pt;
496 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
497 wait->_qproc = NULL;
498 timed_out = 1;
499 }
500
501 if (end_time && !timed_out)
502 slack = select_estimate_accuracy(end_time);
503
504 retval = 0;
505 for (;;) {
506 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
507 bool can_busy_loop = false;
508
509 inp = fds->in; outp = fds->out; exp = fds->ex;
510 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
511
512 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
513 unsigned long in, out, ex, all_bits, bit = 1, j;
514 unsigned long res_in = 0, res_out = 0, res_ex = 0;
515 __poll_t mask;
516
517 in = *inp++; out = *outp++; ex = *exp++;
518 all_bits = in | out | ex;
519 if (all_bits == 0) {
520 i += BITS_PER_LONG;
521 continue;
522 }
523
524 for (j = 0; j < BITS_PER_LONG; ++j, ++i, bit <<= 1) {
525 struct fd f;
526 if (i >= n)
527 break;
528 if (!(bit & all_bits))
529 continue;
530 mask = EPOLLNVAL;
531 f = fdget(i);
532 if (fd_file(f)) {
533 wait_key_set(wait, in, out, bit,
534 busy_flag);
535 mask = vfs_poll(fd_file(f), wait);
536
537 fdput(f);
538 }
539 if ((mask & POLLIN_SET) && (in & bit)) {
540 res_in |= bit;
541 retval++;
542 wait->_qproc = NULL;
543 }
544 if ((mask & POLLOUT_SET) && (out & bit)) {
545 res_out |= bit;
546 retval++;
547 wait->_qproc = NULL;
548 }
549 if ((mask & POLLEX_SET) && (ex & bit)) {
550 res_ex |= bit;
551 retval++;
552 wait->_qproc = NULL;
553 }
554 /* got something, stop busy polling */
555 if (retval) {
556 can_busy_loop = false;
557 busy_flag = 0;
558
559 /*
560 * only remember a returned
561 * POLL_BUSY_LOOP if we asked for it
562 */
563 } else if (busy_flag & mask)
564 can_busy_loop = true;
565
566 }
567 if (res_in)
568 *rinp = res_in;
569 if (res_out)
570 *routp = res_out;
571 if (res_ex)
572 *rexp = res_ex;
573 cond_resched();
574 }
575 wait->_qproc = NULL;
576 if (retval || timed_out || signal_pending(current))
577 break;
578 if (table.error) {
579 retval = table.error;
580 break;
581 }
582
583 /* only if found POLL_BUSY_LOOP sockets && not out of time */
584 if (can_busy_loop && !need_resched()) {
585 if (!busy_start) {
586 busy_start = busy_loop_current_time();
587 continue;
588 }
589 if (!busy_loop_timeout(busy_start))
590 continue;
591 }
592 busy_flag = 0;
593
594 /*
595 * If this is the first loop and we have a timeout
596 * given, then we convert to ktime_t and set the to
597 * pointer to the expiry value.
598 */
599 if (end_time && !to) {
600 expire = timespec64_to_ktime(*end_time);
601 to = &expire;
602 }
603
604 if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
605 to, slack))
606 timed_out = 1;
607 }
608
609 poll_freewait(&table);
610
611 return retval;
612}
613
614/*
615 * We can actually return ERESTARTSYS instead of EINTR, but I'd
616 * like to be certain this leads to no problems. So I return
617 * EINTR just for safety.
618 *
619 * Update: ERESTARTSYS breaks at least the xview clock binary, so
620 * I'm trying ERESTARTNOHAND which restart only when you want to.
621 */
622int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
623 fd_set __user *exp, struct timespec64 *end_time)
624{
625 fd_set_bits fds;
626 void *bits;
627 int ret, max_fds;
628 size_t size, alloc_size;
629 struct fdtable *fdt;
630 /* Allocate small arguments on the stack to save memory and be faster */
631 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
632
633 ret = -EINVAL;
634 if (n < 0)
635 goto out_nofds;
636
637 /* max_fds can increase, so grab it once to avoid race */
638 rcu_read_lock();
639 fdt = files_fdtable(current->files);
640 max_fds = fdt->max_fds;
641 rcu_read_unlock();
642 if (n > max_fds)
643 n = max_fds;
644
645 /*
646 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
647 * since we used fdset we need to allocate memory in units of
648 * long-words.
649 */
650 size = FDS_BYTES(n);
651 bits = stack_fds;
652 if (size > sizeof(stack_fds) / 6) {
653 /* Not enough space in on-stack array; must use kmalloc */
654 ret = -ENOMEM;
655 if (size > (SIZE_MAX / 6))
656 goto out_nofds;
657
658 alloc_size = 6 * size;
659 bits = kvmalloc(alloc_size, GFP_KERNEL);
660 if (!bits)
661 goto out_nofds;
662 }
663 fds.in = bits;
664 fds.out = bits + size;
665 fds.ex = bits + 2*size;
666 fds.res_in = bits + 3*size;
667 fds.res_out = bits + 4*size;
668 fds.res_ex = bits + 5*size;
669
670 if ((ret = get_fd_set(n, inp, fds.in)) ||
671 (ret = get_fd_set(n, outp, fds.out)) ||
672 (ret = get_fd_set(n, exp, fds.ex)))
673 goto out;
674 zero_fd_set(n, fds.res_in);
675 zero_fd_set(n, fds.res_out);
676 zero_fd_set(n, fds.res_ex);
677
678 ret = do_select(n, &fds, end_time);
679
680 if (ret < 0)
681 goto out;
682 if (!ret) {
683 ret = -ERESTARTNOHAND;
684 if (signal_pending(current))
685 goto out;
686 ret = 0;
687 }
688
689 if (set_fd_set(n, inp, fds.res_in) ||
690 set_fd_set(n, outp, fds.res_out) ||
691 set_fd_set(n, exp, fds.res_ex))
692 ret = -EFAULT;
693
694out:
695 if (bits != stack_fds)
696 kvfree(bits);
697out_nofds:
698 return ret;
699}
700
701static int kern_select(int n, fd_set __user *inp, fd_set __user *outp,
702 fd_set __user *exp, struct __kernel_old_timeval __user *tvp)
703{
704 struct timespec64 end_time, *to = NULL;
705 struct __kernel_old_timeval tv;
706 int ret;
707
708 if (tvp) {
709 if (copy_from_user(&tv, tvp, sizeof(tv)))
710 return -EFAULT;
711
712 to = &end_time;
713 if (poll_select_set_timeout(to,
714 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
715 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
716 return -EINVAL;
717 }
718
719 ret = core_sys_select(n, inp, outp, exp, to);
720 return poll_select_finish(&end_time, tvp, PT_TIMEVAL, ret);
721}
722
723SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
724 fd_set __user *, exp, struct __kernel_old_timeval __user *, tvp)
725{
726 return kern_select(n, inp, outp, exp, tvp);
727}
728
729static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
730 fd_set __user *exp, void __user *tsp,
731 const sigset_t __user *sigmask, size_t sigsetsize,
732 enum poll_time_type type)
733{
734 struct timespec64 ts, end_time, *to = NULL;
735 int ret;
736
737 if (tsp) {
738 switch (type) {
739 case PT_TIMESPEC:
740 if (get_timespec64(&ts, tsp))
741 return -EFAULT;
742 break;
743 case PT_OLD_TIMESPEC:
744 if (get_old_timespec32(&ts, tsp))
745 return -EFAULT;
746 break;
747 default:
748 BUG();
749 }
750
751 to = &end_time;
752 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
753 return -EINVAL;
754 }
755
756 ret = set_user_sigmask(sigmask, sigsetsize);
757 if (ret)
758 return ret;
759
760 ret = core_sys_select(n, inp, outp, exp, to);
761 return poll_select_finish(&end_time, tsp, type, ret);
762}
763
764/*
765 * Most architectures can't handle 7-argument syscalls. So we provide a
766 * 6-argument version where the sixth argument is a pointer to a structure
767 * which has a pointer to the sigset_t itself followed by a size_t containing
768 * the sigset size.
769 */
770struct sigset_argpack {
771 sigset_t __user *p;
772 size_t size;
773};
774
775static inline int get_sigset_argpack(struct sigset_argpack *to,
776 struct sigset_argpack __user *from)
777{
778 // the path is hot enough for overhead of copy_from_user() to matter
779 if (from) {
780 if (can_do_masked_user_access())
781 from = masked_user_access_begin(from);
782 else if (!user_read_access_begin(from, sizeof(*from)))
783 return -EFAULT;
784 unsafe_get_user(to->p, &from->p, Efault);
785 unsafe_get_user(to->size, &from->size, Efault);
786 user_read_access_end();
787 }
788 return 0;
789Efault:
790 user_access_end();
791 return -EFAULT;
792}
793
794SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
795 fd_set __user *, exp, struct __kernel_timespec __user *, tsp,
796 void __user *, sig)
797{
798 struct sigset_argpack x = {NULL, 0};
799
800 if (get_sigset_argpack(&x, sig))
801 return -EFAULT;
802
803 return do_pselect(n, inp, outp, exp, tsp, x.p, x.size, PT_TIMESPEC);
804}
805
806#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
807
808SYSCALL_DEFINE6(pselect6_time32, int, n, fd_set __user *, inp, fd_set __user *, outp,
809 fd_set __user *, exp, struct old_timespec32 __user *, tsp,
810 void __user *, sig)
811{
812 struct sigset_argpack x = {NULL, 0};
813
814 if (get_sigset_argpack(&x, sig))
815 return -EFAULT;
816
817 return do_pselect(n, inp, outp, exp, tsp, x.p, x.size, PT_OLD_TIMESPEC);
818}
819
820#endif
821
822#ifdef __ARCH_WANT_SYS_OLD_SELECT
823struct sel_arg_struct {
824 unsigned long n;
825 fd_set __user *inp, *outp, *exp;
826 struct __kernel_old_timeval __user *tvp;
827};
828
829SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg)
830{
831 struct sel_arg_struct a;
832
833 if (copy_from_user(&a, arg, sizeof(a)))
834 return -EFAULT;
835 return kern_select(a.n, a.inp, a.outp, a.exp, a.tvp);
836}
837#endif
838
839struct poll_list {
840 struct poll_list *next;
841 unsigned int len;
842 struct pollfd entries[] __counted_by(len);
843};
844
845#define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
846
847/*
848 * Fish for pollable events on the pollfd->fd file descriptor. We're only
849 * interested in events matching the pollfd->events mask, and the result
850 * matching that mask is both recorded in pollfd->revents and returned. The
851 * pwait poll_table will be used by the fd-provided poll handler for waiting,
852 * if pwait->_qproc is non-NULL.
853 */
854static inline __poll_t do_pollfd(struct pollfd *pollfd, poll_table *pwait,
855 bool *can_busy_poll,
856 __poll_t busy_flag)
857{
858 int fd = pollfd->fd;
859 __poll_t mask = 0, filter;
860 struct fd f;
861
862 if (fd < 0)
863 goto out;
864 mask = EPOLLNVAL;
865 f = fdget(fd);
866 if (!fd_file(f))
867 goto out;
868
869 /* userland u16 ->events contains POLL... bitmap */
870 filter = demangle_poll(pollfd->events) | EPOLLERR | EPOLLHUP;
871 pwait->_key = filter | busy_flag;
872 mask = vfs_poll(fd_file(f), pwait);
873 if (mask & busy_flag)
874 *can_busy_poll = true;
875 mask &= filter; /* Mask out unneeded events. */
876 fdput(f);
877
878out:
879 /* ... and so does ->revents */
880 pollfd->revents = mangle_poll(mask);
881 return mask;
882}
883
884static int do_poll(struct poll_list *list, struct poll_wqueues *wait,
885 struct timespec64 *end_time)
886{
887 poll_table* pt = &wait->pt;
888 ktime_t expire, *to = NULL;
889 int timed_out = 0, count = 0;
890 u64 slack = 0;
891 __poll_t busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
892 unsigned long busy_start = 0;
893
894 /* Optimise the no-wait case */
895 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
896 pt->_qproc = NULL;
897 timed_out = 1;
898 }
899
900 if (end_time && !timed_out)
901 slack = select_estimate_accuracy(end_time);
902
903 for (;;) {
904 struct poll_list *walk;
905 bool can_busy_loop = false;
906
907 for (walk = list; walk != NULL; walk = walk->next) {
908 struct pollfd * pfd, * pfd_end;
909
910 pfd = walk->entries;
911 pfd_end = pfd + walk->len;
912 for (; pfd != pfd_end; pfd++) {
913 /*
914 * Fish for events. If we found one, record it
915 * and kill poll_table->_qproc, so we don't
916 * needlessly register any other waiters after
917 * this. They'll get immediately deregistered
918 * when we break out and return.
919 */
920 if (do_pollfd(pfd, pt, &can_busy_loop,
921 busy_flag)) {
922 count++;
923 pt->_qproc = NULL;
924 /* found something, stop busy polling */
925 busy_flag = 0;
926 can_busy_loop = false;
927 }
928 }
929 }
930 /*
931 * All waiters have already been registered, so don't provide
932 * a poll_table->_qproc to them on the next loop iteration.
933 */
934 pt->_qproc = NULL;
935 if (!count) {
936 count = wait->error;
937 if (signal_pending(current))
938 count = -ERESTARTNOHAND;
939 }
940 if (count || timed_out)
941 break;
942
943 /* only if found POLL_BUSY_LOOP sockets && not out of time */
944 if (can_busy_loop && !need_resched()) {
945 if (!busy_start) {
946 busy_start = busy_loop_current_time();
947 continue;
948 }
949 if (!busy_loop_timeout(busy_start))
950 continue;
951 }
952 busy_flag = 0;
953
954 /*
955 * If this is the first loop and we have a timeout
956 * given, then we convert to ktime_t and set the to
957 * pointer to the expiry value.
958 */
959 if (end_time && !to) {
960 expire = timespec64_to_ktime(*end_time);
961 to = &expire;
962 }
963
964 if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
965 timed_out = 1;
966 }
967 return count;
968}
969
970#define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \
971 sizeof(struct pollfd))
972
973static int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
974 struct timespec64 *end_time)
975{
976 struct poll_wqueues table;
977 int err = -EFAULT, fdcount;
978 /* Allocate small arguments on the stack to save memory and be
979 faster - use long to make sure the buffer is aligned properly
980 on 64 bit archs to avoid unaligned access */
981 long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
982 struct poll_list *const head = (struct poll_list *)stack_pps;
983 struct poll_list *walk = head;
984 unsigned int todo = nfds;
985 unsigned int len;
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 if (walk->len >= todo)
1002 break;
1003 todo -= walk->len;
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 unsigned 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 = &current->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