projects/linux·blob assoc-array-fixes · refs

1438 lines · 34.7 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 kfree(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 = kmalloc(PAGE_SIZE, 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 WRITE_ONCE(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 (!READ_ONCE(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 __poll_t select_poll_one(int fd, poll_table *wait, unsigned long in,
466 unsigned long out, unsigned long bit,
467 __poll_t ll_flag)
468{
469 CLASS(fd, f)(fd);
470
471 if (fd_empty(f))
472 return EPOLLNVAL;
473
474 wait->_key = POLLEX_SET | ll_flag;
475 if (in & bit)
476 wait->_key |= POLLIN_SET;
477 if (out & bit)
478 wait->_key |= POLLOUT_SET;
479
480 return vfs_poll(fd_file(f), wait);
481}
482
483static noinline_for_stack int do_select(int n, fd_set_bits *fds, struct timespec64 *end_time)
484{
485 ktime_t expire, *to = NULL;
486 struct poll_wqueues table;
487 poll_table *wait;
488 int retval, i, timed_out = 0;
489 u64 slack = 0;
490 __poll_t busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
491 unsigned long busy_start = 0;
492
493 rcu_read_lock();
494 retval = max_select_fd(n, fds);
495 rcu_read_unlock();
496
497 if (retval < 0)
498 return retval;
499 n = retval;
500
501 poll_initwait(&table);
502 wait = &table.pt;
503 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
504 wait->_qproc = NULL;
505 timed_out = 1;
506 }
507
508 if (end_time && !timed_out)
509 slack = select_estimate_accuracy(end_time);
510
511 retval = 0;
512 for (;;) {
513 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
514 bool can_busy_loop = false;
515
516 inp = fds->in; outp = fds->out; exp = fds->ex;
517 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
518
519 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
520 unsigned long in, out, ex, all_bits, bit = 1, j;
521 unsigned long res_in = 0, res_out = 0, res_ex = 0;
522 __poll_t mask;
523
524 in = *inp++; out = *outp++; ex = *exp++;
525 all_bits = in | out | ex;
526 if (all_bits == 0) {
527 i += BITS_PER_LONG;
528 continue;
529 }
530
531 for (j = 0; j < BITS_PER_LONG; ++j, ++i, bit <<= 1) {
532 if (i >= n)
533 break;
534 if (!(bit & all_bits))
535 continue;
536 mask = select_poll_one(i, wait, in, out, bit,
537 busy_flag);
538 if ((mask & POLLIN_SET) && (in & bit)) {
539 res_in |= bit;
540 retval++;
541 wait->_qproc = NULL;
542 }
543 if ((mask & POLLOUT_SET) && (out & bit)) {
544 res_out |= bit;
545 retval++;
546 wait->_qproc = NULL;
547 }
548 if ((mask & POLLEX_SET) && (ex & bit)) {
549 res_ex |= bit;
550 retval++;
551 wait->_qproc = NULL;
552 }
553 /* got something, stop busy polling */
554 if (retval) {
555 can_busy_loop = false;
556 busy_flag = 0;
557
558 /*
559 * only remember a returned
560 * POLL_BUSY_LOOP if we asked for it
561 */
562 } else if (busy_flag & mask)
563 can_busy_loop = true;
564
565 }
566 if (res_in)
567 *rinp = res_in;
568 if (res_out)
569 *routp = res_out;
570 if (res_ex)
571 *rexp = res_ex;
572 cond_resched();
573 }
574 wait->_qproc = NULL;
575 if (retval || timed_out || signal_pending(current))
576 break;
577 if (table.error) {
578 retval = table.error;
579 break;
580 }
581
582 /* only if found POLL_BUSY_LOOP sockets && not out of time */
583 if (can_busy_loop && !need_resched()) {
584 if (!busy_start) {
585 busy_start = busy_loop_current_time();
586 continue;
587 }
588 if (!busy_loop_timeout(busy_start))
589 continue;
590 }
591 busy_flag = 0;
592
593 /*
594 * If this is the first loop and we have a timeout
595 * given, then we convert to ktime_t and set the to
596 * pointer to the expiry value.
597 */
598 if (end_time && !to) {
599 expire = timespec64_to_ktime(*end_time);
600 to = &expire;
601 }
602
603 if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
604 to, slack))
605 timed_out = 1;
606 }
607
608 poll_freewait(&table);
609
610 return retval;
611}
612
613/*
614 * We can actually return ERESTARTSYS instead of EINTR, but I'd
615 * like to be certain this leads to no problems. So I return
616 * EINTR just for safety.
617 *
618 * Update: ERESTARTSYS breaks at least the xview clock binary, so
619 * I'm trying ERESTARTNOHAND which restart only when you want to.
620 */
621int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
622 fd_set __user *exp, struct timespec64 *end_time)
623{
624 fd_set_bits fds;
625 void *bits;
626 int ret, max_fds;
627 size_t size, alloc_size;
628 struct fdtable *fdt;
629 /* Allocate small arguments on the stack to save memory and be faster */
630 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
631
632 ret = -EINVAL;
633 if (unlikely(n < 0))
634 goto out_nofds;
635
636 /* max_fds can increase, so grab it once to avoid race */
637 rcu_read_lock();
638 fdt = files_fdtable(current->files);
639 max_fds = fdt->max_fds;
640 rcu_read_unlock();
641 if (n > max_fds)
642 n = max_fds;
643
644 /*
645 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
646 * since we used fdset we need to allocate memory in units of
647 * long-words.
648 */
649 size = FDS_BYTES(n);
650 bits = stack_fds;
651 if (size > sizeof(stack_fds) / 6) {
652 /* Not enough space in on-stack array; must use kmalloc */
653 ret = -ENOMEM;
654 if (size > (SIZE_MAX / 6))
655 goto out_nofds;
656
657 alloc_size = 6 * size;
658 bits = kvmalloc(alloc_size, GFP_KERNEL);
659 if (!bits)
660 goto out_nofds;
661 }
662 fds.in = bits;
663 fds.out = bits + size;
664 fds.ex = bits + 2*size;
665 fds.res_in = bits + 3*size;
666 fds.res_out = bits + 4*size;
667 fds.res_ex = bits + 5*size;
668
669 if ((ret = get_fd_set(n, inp, fds.in)) ||
670 (ret = get_fd_set(n, outp, fds.out)) ||
671 (ret = get_fd_set(n, exp, fds.ex)))
672 goto out;
673 zero_fd_set(n, fds.res_in);
674 zero_fd_set(n, fds.res_out);
675 zero_fd_set(n, fds.res_ex);
676
677 ret = do_select(n, &fds, end_time);
678
679 if (ret < 0)
680 goto out;
681 if (!ret) {
682 ret = -ERESTARTNOHAND;
683 if (signal_pending(current))
684 goto out;
685 ret = 0;
686 }
687
688 if (set_fd_set(n, inp, fds.res_in) ||
689 set_fd_set(n, outp, fds.res_out) ||
690 set_fd_set(n, exp, fds.res_ex))
691 ret = -EFAULT;
692
693out:
694 if (bits != stack_fds)
695 kvfree(bits);
696out_nofds:
697 return ret;
698}
699
700static int kern_select(int n, fd_set __user *inp, fd_set __user *outp,
701 fd_set __user *exp, struct __kernel_old_timeval __user *tvp)
702{
703 struct timespec64 end_time, *to = NULL;
704 struct __kernel_old_timeval tv;
705 int ret;
706
707 if (tvp) {
708 if (copy_from_user(&tv, tvp, sizeof(tv)))
709 return -EFAULT;
710
711 /*
712 * Reject negative components before normalisation. The seconds
713 * sum below is performed in signed long and a crafted negative
714 * timeval can wrap to a positive value that passes
715 * timespec64_valid() and turns into an effectively-infinite
716 * deadline via timespec64_add_safe()'s saturation, instead of
717 * the -EINVAL POSIX requires for negative timeouts.
718 */
719 if (tv.tv_sec < 0 || tv.tv_usec < 0)
720 return -EINVAL;
721
722 to = &end_time;
723 if (poll_select_set_timeout(to,
724 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
725 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
726 return -EINVAL;
727 }
728
729 ret = core_sys_select(n, inp, outp, exp, to);
730 return poll_select_finish(&end_time, tvp, PT_TIMEVAL, ret);
731}
732
733SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
734 fd_set __user *, exp, struct __kernel_old_timeval __user *, tvp)
735{
736 return kern_select(n, inp, outp, exp, tvp);
737}
738
739static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
740 fd_set __user *exp, void __user *tsp,
741 const sigset_t __user *sigmask, size_t sigsetsize,
742 enum poll_time_type type)
743{
744 struct timespec64 ts, end_time, *to = NULL;
745 int ret;
746
747 if (tsp) {
748 switch (type) {
749 case PT_TIMESPEC:
750 if (get_timespec64(&ts, tsp))
751 return -EFAULT;
752 break;
753 case PT_OLD_TIMESPEC:
754 if (get_old_timespec32(&ts, tsp))
755 return -EFAULT;
756 break;
757 default:
758 BUG();
759 }
760
761 to = &end_time;
762 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
763 return -EINVAL;
764 }
765
766 ret = set_user_sigmask(sigmask, sigsetsize);
767 if (ret)
768 return ret;
769
770 ret = core_sys_select(n, inp, outp, exp, to);
771 return poll_select_finish(&end_time, tsp, type, ret);
772}
773
774/*
775 * Most architectures can't handle 7-argument syscalls. So we provide a
776 * 6-argument version where the sixth argument is a pointer to a structure
777 * which has a pointer to the sigset_t itself followed by a size_t containing
778 * the sigset size.
779 */
780struct sigset_argpack {
781 sigset_t __user *p;
782 size_t size;
783};
784
785static inline int get_sigset_argpack(struct sigset_argpack *to,
786 struct sigset_argpack __user *from)
787{
788 // the path is hot enough for overhead of copy_from_user() to matter
789 if (from) {
790 scoped_user_read_access(from, Efault) {
791 unsafe_get_user(to->p, &from->p, Efault);
792 unsafe_get_user(to->size, &from->size, Efault);
793 }
794 }
795 return 0;
796Efault:
797 return -EFAULT;
798}
799
800SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
801 fd_set __user *, exp, struct __kernel_timespec __user *, tsp,
802 void __user *, sig)
803{
804 struct sigset_argpack x = {NULL, 0};
805
806 if (get_sigset_argpack(&x, sig))
807 return -EFAULT;
808
809 return do_pselect(n, inp, outp, exp, tsp, x.p, x.size, PT_TIMESPEC);
810}
811
812#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
813
814SYSCALL_DEFINE6(pselect6_time32, int, n, fd_set __user *, inp, fd_set __user *, outp,
815 fd_set __user *, exp, struct old_timespec32 __user *, tsp,
816 void __user *, sig)
817{
818 struct sigset_argpack x = {NULL, 0};
819
820 if (get_sigset_argpack(&x, sig))
821 return -EFAULT;
822
823 return do_pselect(n, inp, outp, exp, tsp, x.p, x.size, PT_OLD_TIMESPEC);
824}
825
826#endif
827
828#ifdef __ARCH_WANT_SYS_OLD_SELECT
829struct sel_arg_struct {
830 unsigned long n;
831 fd_set __user *inp, *outp, *exp;
832 struct __kernel_old_timeval __user *tvp;
833};
834
835SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg)
836{
837 struct sel_arg_struct a;
838
839 if (copy_from_user(&a, arg, sizeof(a)))
840 return -EFAULT;
841 return kern_select(a.n, a.inp, a.outp, a.exp, a.tvp);
842}
843#endif
844
845struct poll_list {
846 struct poll_list *next;
847 unsigned int len;
848 struct pollfd entries[] __counted_by(len);
849};
850
851#define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
852
853/*
854 * Fish for pollable events on the pollfd->fd file descriptor. We're only
855 * interested in events matching the pollfd->events mask, and the result
856 * matching that mask is both recorded in pollfd->revents and returned. The
857 * pwait poll_table will be used by the fd-provided poll handler for waiting,
858 * if pwait->_qproc is non-NULL.
859 */
860static inline __poll_t do_pollfd(struct pollfd *pollfd, poll_table *pwait,
861 bool *can_busy_poll,
862 __poll_t busy_flag)
863{
864 int fd = pollfd->fd;
865 __poll_t mask, filter;
866
867 if (unlikely(fd < 0))
868 return 0;
869
870 CLASS(fd, f)(fd);
871 if (fd_empty(f))
872 return EPOLLNVAL;
873
874 /* userland u16 ->events contains POLL... bitmap */
875 filter = demangle_poll(pollfd->events) | EPOLLERR | EPOLLHUP;
876 pwait->_key = filter | busy_flag;
877 mask = vfs_poll(fd_file(f), pwait);
878 if (mask & busy_flag)
879 *can_busy_poll = true;
880 return mask & filter; /* Mask out unneeded events. */
881}
882
883static int do_poll(struct poll_list *list, struct poll_wqueues *wait,
884 struct timespec64 *end_time)
885{
886 poll_table* pt = &wait->pt;
887 ktime_t expire, *to = NULL;
888 int timed_out = 0, count = 0;
889 u64 slack = 0;
890 __poll_t busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
891 unsigned long busy_start = 0;
892
893 /* Optimise the no-wait case */
894 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
895 pt->_qproc = NULL;
896 timed_out = 1;
897 }
898
899 if (end_time && !timed_out)
900 slack = select_estimate_accuracy(end_time);
901
902 for (;;) {
903 struct poll_list *walk;
904 bool can_busy_loop = false;
905
906 for (walk = list; walk != NULL; walk = walk->next) {
907 struct pollfd * pfd, * pfd_end;
908
909 pfd = walk->entries;
910 pfd_end = pfd + walk->len;
911 for (; pfd != pfd_end; pfd++) {
912 __poll_t mask;
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 mask = do_pollfd(pfd, pt, &can_busy_loop, busy_flag);
921 pfd->revents = mangle_poll(mask);
922 if (mask) {
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;
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 int todo = nfds;
986 unsigned int len;
987
988 if (nfds > rlimit(RLIMIT_NOFILE))
989 return -EINVAL;
990
991 len = min_t(unsigned int, nfds, N_STACK_PPS);
992 for (;;) {
993 walk->next = NULL;
994 walk->len = len;
995 if (!len)
996 break;
997
998 if (copy_from_user(walk->entries, ufds + nfds-todo,
999 sizeof(struct pollfd) * walk->len))
1000 goto out_fds;
1001
1002 if (walk->len >= todo)
1003 break;
1004 todo -= walk->len;
1005
1006 len = min(todo, POLLFD_PER_PAGE);
1007 walk = walk->next = kmalloc_flex(*walk, entries, len);
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 scoped_user_write_access_size(ufds, nfds * sizeof(*ufds), out_fds) {
1019 struct pollfd __user *_ufds = ufds;
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, out_fds);
1027 }
1028 }
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}
1041
1042static long do_restart_poll(struct restart_block *restart_block)
1043{
1044 struct pollfd __user *ufds = restart_block->poll.ufds;
1045 int nfds = restart_block->poll.nfds;
1046 struct timespec64 *to = NULL;
1047 int ret;
1048
1049 if (restart_block->poll.has_timeout)
1050 to = &restart_block->poll.end_time;
1051
1052 ret = do_sys_poll(ufds, nfds, to);
1053
1054 if (ret == -ERESTARTNOHAND)
1055 ret = set_restart_fn(restart_block, do_restart_poll);
1056
1057 return ret;
1058}
1059
1060SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
1061 int, timeout_msecs)
1062{
1063 struct timespec64 end_time, *to = NULL;
1064 int ret;
1065
1066 if (timeout_msecs >= 0) {
1067 to = &end_time;
1068 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
1069 NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
1070 }
1071
1072 ret = do_sys_poll(ufds, nfds, to);
1073
1074 if (ret == -ERESTARTNOHAND) {
1075 struct restart_block *restart_block;
1076
1077 restart_block = &current->restart_block;
1078 restart_block->poll.ufds = ufds;
1079 restart_block->poll.nfds = nfds;
1080
1081 if (timeout_msecs >= 0) {
1082 restart_block->poll.end_time = end_time;
1083 restart_block->poll.has_timeout = 1;
1084 } else
1085 restart_block->poll.has_timeout = 0;
1086
1087 ret = set_restart_fn(restart_block, do_restart_poll);
1088 }
1089 return ret;
1090}
1091
1092SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
1093 struct __kernel_timespec __user *, tsp, const sigset_t __user *, sigmask,
1094 size_t, sigsetsize)
1095{
1096 struct timespec64 ts, end_time, *to = NULL;
1097 int ret;
1098
1099 if (tsp) {
1100 if (get_timespec64(&ts, tsp))
1101 return -EFAULT;
1102
1103 to = &end_time;
1104 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1105 return -EINVAL;
1106 }
1107
1108 ret = set_user_sigmask(sigmask, sigsetsize);
1109 if (ret)
1110 return ret;
1111
1112 ret = do_sys_poll(ufds, nfds, to);
1113 return poll_select_finish(&end_time, tsp, PT_TIMESPEC, ret);
1114}
1115
1116#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
1117
1118SYSCALL_DEFINE5(ppoll_time32, struct pollfd __user *, ufds, unsigned int, nfds,
1119 struct old_timespec32 __user *, tsp, const sigset_t __user *, sigmask,
1120 size_t, sigsetsize)
1121{
1122 struct timespec64 ts, end_time, *to = NULL;
1123 int ret;
1124
1125 if (tsp) {
1126 if (get_old_timespec32(&ts, tsp))
1127 return -EFAULT;
1128
1129 to = &end_time;
1130 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1131 return -EINVAL;
1132 }
1133
1134 ret = set_user_sigmask(sigmask, sigsetsize);
1135 if (ret)
1136 return ret;
1137
1138 ret = do_sys_poll(ufds, nfds, to);
1139 return poll_select_finish(&end_time, tsp, PT_OLD_TIMESPEC, ret);
1140}
1141#endif
1142
1143#ifdef CONFIG_COMPAT
1144#define __COMPAT_NFDBITS (8 * sizeof(compat_ulong_t))
1145
1146/*
1147 * Ooo, nasty. We need here to frob 32-bit unsigned longs to
1148 * 64-bit unsigned longs.
1149 */
1150static
1151int compat_get_fd_set(unsigned long nr, compat_ulong_t __user *ufdset,
1152 unsigned long *fdset)
1153{
1154 if (ufdset) {
1155 return compat_get_bitmap(fdset, ufdset, nr);
1156 } else {
1157 zero_fd_set(nr, fdset);
1158 return 0;
1159 }
1160}
1161
1162static
1163int compat_set_fd_set(unsigned long nr, compat_ulong_t __user *ufdset,
1164 unsigned long *fdset)
1165{
1166 if (!ufdset)
1167 return 0;
1168 return compat_put_bitmap(ufdset, fdset, nr);
1169}
1170
1171
1172/*
1173 * This is a virtual copy of sys_select from fs/select.c and probably
1174 * should be compared to it from time to time
1175 */
1176
1177/*
1178 * We can actually return ERESTARTSYS instead of EINTR, but I'd
1179 * like to be certain this leads to no problems. So I return
1180 * EINTR just for safety.
1181 *
1182 * Update: ERESTARTSYS breaks at least the xview clock binary, so
1183 * I'm trying ERESTARTNOHAND which restart only when you want to.
1184 */
1185static int compat_core_sys_select(int n, compat_ulong_t __user *inp,
1186 compat_ulong_t __user *outp, compat_ulong_t __user *exp,
1187 struct timespec64 *end_time)
1188{
1189 fd_set_bits fds;
1190 void *bits;
1191 int size, max_fds, ret = -EINVAL;
1192 struct fdtable *fdt;
1193 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
1194
1195 if (n < 0)
1196 goto out_nofds;
1197
1198 /* max_fds can increase, so grab it once to avoid race */
1199 rcu_read_lock();
1200 fdt = files_fdtable(current->files);
1201 max_fds = fdt->max_fds;
1202 rcu_read_unlock();
1203 if (n > max_fds)
1204 n = max_fds;
1205
1206 /*
1207 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
1208 * since we used fdset we need to allocate memory in units of
1209 * long-words.
1210 */
1211 size = FDS_BYTES(n);
1212 bits = stack_fds;
1213 if (size > sizeof(stack_fds) / 6) {
1214 bits = kmalloc_array(6, size, GFP_KERNEL);
1215 ret = -ENOMEM;
1216 if (!bits)
1217 goto out_nofds;
1218 }
1219 fds.in = (unsigned long *) bits;
1220 fds.out = (unsigned long *) (bits + size);
1221 fds.ex = (unsigned long *) (bits + 2*size);
1222 fds.res_in = (unsigned long *) (bits + 3*size);
1223 fds.res_out = (unsigned long *) (bits + 4*size);
1224 fds.res_ex = (unsigned long *) (bits + 5*size);
1225
1226 if ((ret = compat_get_fd_set(n, inp, fds.in)) ||
1227 (ret = compat_get_fd_set(n, outp, fds.out)) ||
1228 (ret = compat_get_fd_set(n, exp, fds.ex)))
1229 goto out;
1230 zero_fd_set(n, fds.res_in);
1231 zero_fd_set(n, fds.res_out);
1232 zero_fd_set(n, fds.res_ex);
1233
1234 ret = do_select(n, &fds, end_time);
1235
1236 if (ret < 0)
1237 goto out;
1238 if (!ret) {
1239 ret = -ERESTARTNOHAND;
1240 if (signal_pending(current))
1241 goto out;
1242 ret = 0;
1243 }
1244
1245 if (compat_set_fd_set(n, inp, fds.res_in) ||
1246 compat_set_fd_set(n, outp, fds.res_out) ||
1247 compat_set_fd_set(n, exp, fds.res_ex))
1248 ret = -EFAULT;
1249out:
1250 if (bits != stack_fds)
1251 kfree(bits);
1252out_nofds:
1253 return ret;
1254}
1255
1256static int do_compat_select(int n, compat_ulong_t __user *inp,
1257 compat_ulong_t __user *outp, compat_ulong_t __user *exp,
1258 struct old_timeval32 __user *tvp)
1259{
1260 struct timespec64 end_time, *to = NULL;
1261 struct old_timeval32 tv;
1262 int ret;
1263
1264 if (tvp) {
1265 if (copy_from_user(&tv, tvp, sizeof(tv)))
1266 return -EFAULT;
1267
1268 to = &end_time;
1269 if (poll_select_set_timeout(to,
1270 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
1271 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
1272 return -EINVAL;
1273 }
1274
1275 ret = compat_core_sys_select(n, inp, outp, exp, to);
1276 return poll_select_finish(&end_time, tvp, PT_OLD_TIMEVAL, ret);
1277}
1278
1279COMPAT_SYSCALL_DEFINE5(select, int, n, compat_ulong_t __user *, inp,
1280 compat_ulong_t __user *, outp, compat_ulong_t __user *, exp,
1281 struct old_timeval32 __user *, tvp)
1282{
1283 return do_compat_select(n, inp, outp, exp, tvp);
1284}
1285
1286struct compat_sel_arg_struct {
1287 compat_ulong_t n;
1288 compat_uptr_t inp;
1289 compat_uptr_t outp;
1290 compat_uptr_t exp;
1291 compat_uptr_t tvp;
1292};
1293
1294COMPAT_SYSCALL_DEFINE1(old_select, struct compat_sel_arg_struct __user *, arg)
1295{
1296 struct compat_sel_arg_struct a;
1297
1298 if (copy_from_user(&a, arg, sizeof(a)))
1299 return -EFAULT;
1300 return do_compat_select(a.n, compat_ptr(a.inp), compat_ptr(a.outp),
1301 compat_ptr(a.exp), compat_ptr(a.tvp));
1302}
1303
1304static long do_compat_pselect(int n, compat_ulong_t __user *inp,
1305 compat_ulong_t __user *outp, compat_ulong_t __user *exp,
1306 void __user *tsp, compat_sigset_t __user *sigmask,
1307 compat_size_t sigsetsize, enum poll_time_type type)
1308{
1309 struct timespec64 ts, end_time, *to = NULL;
1310 int ret;
1311
1312 if (tsp) {
1313 switch (type) {
1314 case PT_OLD_TIMESPEC:
1315 if (get_old_timespec32(&ts, tsp))
1316 return -EFAULT;
1317 break;
1318 case PT_TIMESPEC:
1319 if (get_timespec64(&ts, tsp))
1320 return -EFAULT;
1321 break;
1322 default:
1323 BUG();
1324 }
1325
1326 to = &end_time;
1327 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1328 return -EINVAL;
1329 }
1330
1331 ret = set_compat_user_sigmask(sigmask, sigsetsize);
1332 if (ret)
1333 return ret;
1334
1335 ret = compat_core_sys_select(n, inp, outp, exp, to);
1336 return poll_select_finish(&end_time, tsp, type, ret);
1337}
1338
1339struct compat_sigset_argpack {
1340 compat_uptr_t p;
1341 compat_size_t size;
1342};
1343static inline int get_compat_sigset_argpack(struct compat_sigset_argpack *to,
1344 struct compat_sigset_argpack __user *from)
1345{
1346 if (from) {
1347 scoped_user_read_access(from, efault) {
1348 unsafe_get_user(to->p, &from->p, efault);
1349 unsafe_get_user(to->size, &from->size, efault);
1350 }
1351 }
1352 return 0;
1353efault:
1354 return -EFAULT;
1355}
1356
1357COMPAT_SYSCALL_DEFINE6(pselect6_time64, int, n, compat_ulong_t __user *, inp,
1358 compat_ulong_t __user *, outp, compat_ulong_t __user *, exp,
1359 struct __kernel_timespec __user *, tsp, void __user *, sig)
1360{
1361 struct compat_sigset_argpack x = {0, 0};
1362
1363 if (get_compat_sigset_argpack(&x, sig))
1364 return -EFAULT;
1365
1366 return do_compat_pselect(n, inp, outp, exp, tsp, compat_ptr(x.p),
1367 x.size, PT_TIMESPEC);
1368}
1369
1370#if defined(CONFIG_COMPAT_32BIT_TIME)
1371
1372COMPAT_SYSCALL_DEFINE6(pselect6_time32, int, n, compat_ulong_t __user *, inp,
1373 compat_ulong_t __user *, outp, compat_ulong_t __user *, exp,
1374 struct old_timespec32 __user *, tsp, void __user *, sig)
1375{
1376 struct compat_sigset_argpack x = {0, 0};
1377
1378 if (get_compat_sigset_argpack(&x, sig))
1379 return -EFAULT;
1380
1381 return do_compat_pselect(n, inp, outp, exp, tsp, compat_ptr(x.p),
1382 x.size, PT_OLD_TIMESPEC);
1383}
1384
1385#endif
1386
1387#if defined(CONFIG_COMPAT_32BIT_TIME)
1388COMPAT_SYSCALL_DEFINE5(ppoll_time32, struct pollfd __user *, ufds,
1389 unsigned int, nfds, struct old_timespec32 __user *, tsp,
1390 const compat_sigset_t __user *, sigmask, compat_size_t, sigsetsize)
1391{
1392 struct timespec64 ts, end_time, *to = NULL;
1393 int ret;
1394
1395 if (tsp) {
1396 if (get_old_timespec32(&ts, tsp))
1397 return -EFAULT;
1398
1399 to = &end_time;
1400 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1401 return -EINVAL;
1402 }
1403
1404 ret = set_compat_user_sigmask(sigmask, sigsetsize);
1405 if (ret)
1406 return ret;
1407
1408 ret = do_sys_poll(ufds, nfds, to);
1409 return poll_select_finish(&end_time, tsp, PT_OLD_TIMESPEC, ret);
1410}
1411#endif
1412
1413/* New compat syscall for 64 bit time_t*/
1414COMPAT_SYSCALL_DEFINE5(ppoll_time64, struct pollfd __user *, ufds,
1415 unsigned int, nfds, struct __kernel_timespec __user *, tsp,
1416 const compat_sigset_t __user *, sigmask, compat_size_t, sigsetsize)
1417{
1418 struct timespec64 ts, end_time, *to = NULL;
1419 int ret;
1420
1421 if (tsp) {
1422 if (get_timespec64(&ts, tsp))
1423 return -EFAULT;
1424
1425 to = &end_time;
1426 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1427 return -EINVAL;
1428 }
1429
1430 ret = set_compat_user_sigmask(sigmask, sigsetsize);
1431 if (ret)
1432 return ret;
1433
1434 ret = do_sys_poll(ufds, nfds, to);
1435 return poll_select_finish(&end_time, tsp, PT_TIMESPEC, ret);
1436}
1437
1438#endif