net/packet: fix use-after-free
[linux/fpc-iii.git] / fs / select.c
blob317891ff8165ba19b775fcfaa8f6deccb58ba18f
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * This file contains the procedures for the handling of select and poll
5 * Created for Linux based loosely upon Mathius Lattner's minix
6 * patches by Peter MacDonald. Heavily edited by Linus.
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.
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).
18 #include <linux/kernel.h>
19 #include <linux/sched/signal.h>
20 #include <linux/sched/rt.h>
21 #include <linux/syscalls.h>
22 #include <linux/export.h>
23 #include <linux/slab.h>
24 #include <linux/poll.h>
25 #include <linux/personality.h> /* for STICKY_TIMEOUTS */
26 #include <linux/file.h>
27 #include <linux/fdtable.h>
28 #include <linux/fs.h>
29 #include <linux/rcupdate.h>
30 #include <linux/hrtimer.h>
31 #include <linux/freezer.h>
32 #include <net/busy_poll.h>
33 #include <linux/vmalloc.h>
35 #include <linux/uaccess.h>
37 __poll_t vfs_poll(struct file *file, struct poll_table_struct *pt)
39 if (file->f_op->poll) {
40 return file->f_op->poll(file, pt);
41 } else if (file_has_poll_mask(file)) {
42 unsigned int events = poll_requested_events(pt);
43 struct wait_queue_head *head;
45 if (pt && pt->_qproc) {
46 head = file->f_op->get_poll_head(file, events);
47 if (!head)
48 return DEFAULT_POLLMASK;
49 if (IS_ERR(head))
50 return EPOLLERR;
51 pt->_qproc(file, head, pt);
54 return file->f_op->poll_mask(file, events);
55 } else {
56 return DEFAULT_POLLMASK;
59 EXPORT_SYMBOL_GPL(vfs_poll);
62 * Estimate expected accuracy in ns from a timeval.
64 * After quite a bit of churning around, we've settled on
65 * a simple thing of taking 0.1% of the timeout as the
66 * slack, with a cap of 100 msec.
67 * "nice" tasks get a 0.5% slack instead.
69 * Consider this comment an open invitation to come up with even
70 * better solutions..
73 #define MAX_SLACK (100 * NSEC_PER_MSEC)
75 static long __estimate_accuracy(struct timespec64 *tv)
77 long slack;
78 int divfactor = 1000;
80 if (tv->tv_sec < 0)
81 return 0;
83 if (task_nice(current) > 0)
84 divfactor = divfactor / 5;
86 if (tv->tv_sec > MAX_SLACK / (NSEC_PER_SEC/divfactor))
87 return MAX_SLACK;
89 slack = tv->tv_nsec / divfactor;
90 slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
92 if (slack > MAX_SLACK)
93 return MAX_SLACK;
95 return slack;
98 u64 select_estimate_accuracy(struct timespec64 *tv)
100 u64 ret;
101 struct timespec64 now;
104 * Realtime tasks get a slack of 0 for obvious reasons.
107 if (rt_task(current))
108 return 0;
110 ktime_get_ts64(&now);
111 now = timespec64_sub(*tv, now);
112 ret = __estimate_accuracy(&now);
113 if (ret < current->timer_slack_ns)
114 return current->timer_slack_ns;
115 return ret;
120 struct poll_table_page {
121 struct poll_table_page * next;
122 struct poll_table_entry * entry;
123 struct poll_table_entry entries[0];
126 #define POLL_TABLE_FULL(table) \
127 ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
130 * Ok, Peter made a complicated, but straightforward multiple_wait() function.
131 * I have rewritten this, taking some shortcuts: This code may not be easy to
132 * follow, but it should be free of race-conditions, and it's practical. If you
133 * understand what I'm doing here, then you understand how the linux
134 * sleep/wakeup mechanism works.
136 * Two very simple procedures, poll_wait() and poll_freewait() make all the
137 * work. poll_wait() is an inline-function defined in <linux/poll.h>,
138 * as all select/poll functions have to call it to add an entry to the
139 * poll table.
141 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
142 poll_table *p);
144 void poll_initwait(struct poll_wqueues *pwq)
146 init_poll_funcptr(&pwq->pt, __pollwait);
147 pwq->polling_task = current;
148 pwq->triggered = 0;
149 pwq->error = 0;
150 pwq->table = NULL;
151 pwq->inline_index = 0;
153 EXPORT_SYMBOL(poll_initwait);
155 static void free_poll_entry(struct poll_table_entry *entry)
157 remove_wait_queue(entry->wait_address, &entry->wait);
158 fput(entry->filp);
161 void poll_freewait(struct poll_wqueues *pwq)
163 struct poll_table_page * p = pwq->table;
164 int i;
165 for (i = 0; i < pwq->inline_index; i++)
166 free_poll_entry(pwq->inline_entries + i);
167 while (p) {
168 struct poll_table_entry * entry;
169 struct poll_table_page *old;
171 entry = p->entry;
172 do {
173 entry--;
174 free_poll_entry(entry);
175 } while (entry > p->entries);
176 old = p;
177 p = p->next;
178 free_page((unsigned long) old);
181 EXPORT_SYMBOL(poll_freewait);
183 static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
185 struct poll_table_page *table = p->table;
187 if (p->inline_index < N_INLINE_POLL_ENTRIES)
188 return p->inline_entries + p->inline_index++;
190 if (!table || POLL_TABLE_FULL(table)) {
191 struct poll_table_page *new_table;
193 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
194 if (!new_table) {
195 p->error = -ENOMEM;
196 return NULL;
198 new_table->entry = new_table->entries;
199 new_table->next = table;
200 p->table = new_table;
201 table = new_table;
204 return table->entry++;
207 static int __pollwake(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
209 struct poll_wqueues *pwq = wait->private;
210 DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
213 * Although this function is called under waitqueue lock, LOCK
214 * doesn't imply write barrier and the users expect write
215 * barrier semantics on wakeup functions. The following
216 * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
217 * and is paired with smp_store_mb() in poll_schedule_timeout.
219 smp_wmb();
220 pwq->triggered = 1;
223 * Perform the default wake up operation using a dummy
224 * waitqueue.
226 * TODO: This is hacky but there currently is no interface to
227 * pass in @sync. @sync is scheduled to be removed and once
228 * that happens, wake_up_process() can be used directly.
230 return default_wake_function(&dummy_wait, mode, sync, key);
233 static int pollwake(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
235 struct poll_table_entry *entry;
237 entry = container_of(wait, struct poll_table_entry, wait);
238 if (key && !(key_to_poll(key) & entry->key))
239 return 0;
240 return __pollwake(wait, mode, sync, key);
243 /* Add a new entry */
244 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
245 poll_table *p)
247 struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
248 struct poll_table_entry *entry = poll_get_entry(pwq);
249 if (!entry)
250 return;
251 entry->filp = get_file(filp);
252 entry->wait_address = wait_address;
253 entry->key = p->_key;
254 init_waitqueue_func_entry(&entry->wait, pollwake);
255 entry->wait.private = pwq;
256 add_wait_queue(wait_address, &entry->wait);
259 static int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
260 ktime_t *expires, unsigned long slack)
262 int rc = -EINTR;
264 set_current_state(state);
265 if (!pwq->triggered)
266 rc = schedule_hrtimeout_range(expires, slack, HRTIMER_MODE_ABS);
267 __set_current_state(TASK_RUNNING);
270 * Prepare for the next iteration.
272 * The following smp_store_mb() serves two purposes. First, it's
273 * the counterpart rmb of the wmb in pollwake() such that data
274 * written before wake up is always visible after wake up.
275 * Second, the full barrier guarantees that triggered clearing
276 * doesn't pass event check of the next iteration. Note that
277 * this problem doesn't exist for the first iteration as
278 * add_wait_queue() has full barrier semantics.
280 smp_store_mb(pwq->triggered, 0);
282 return rc;
286 * poll_select_set_timeout - helper function to setup the timeout value
287 * @to: pointer to timespec64 variable for the final timeout
288 * @sec: seconds (from user space)
289 * @nsec: nanoseconds (from user space)
291 * Note, we do not use a timespec for the user space value here, That
292 * way we can use the function for timeval and compat interfaces as well.
294 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
296 int poll_select_set_timeout(struct timespec64 *to, time64_t sec, long nsec)
298 struct timespec64 ts = {.tv_sec = sec, .tv_nsec = nsec};
300 if (!timespec64_valid(&ts))
301 return -EINVAL;
303 /* Optimize for the zero timeout value here */
304 if (!sec && !nsec) {
305 to->tv_sec = to->tv_nsec = 0;
306 } else {
307 ktime_get_ts64(to);
308 *to = timespec64_add_safe(*to, ts);
310 return 0;
313 static int poll_select_copy_remaining(struct timespec64 *end_time,
314 void __user *p,
315 int timeval, int ret)
317 struct timespec64 rts;
318 struct timeval rtv;
320 if (!p)
321 return ret;
323 if (current->personality & STICKY_TIMEOUTS)
324 goto sticky;
326 /* No update for zero timeout */
327 if (!end_time->tv_sec && !end_time->tv_nsec)
328 return ret;
330 ktime_get_ts64(&rts);
331 rts = timespec64_sub(*end_time, rts);
332 if (rts.tv_sec < 0)
333 rts.tv_sec = rts.tv_nsec = 0;
336 if (timeval) {
337 if (sizeof(rtv) > sizeof(rtv.tv_sec) + sizeof(rtv.tv_usec))
338 memset(&rtv, 0, sizeof(rtv));
339 rtv.tv_sec = rts.tv_sec;
340 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
342 if (!copy_to_user(p, &rtv, sizeof(rtv)))
343 return ret;
345 } else if (!put_timespec64(&rts, p))
346 return ret;
349 * If an application puts its timeval in read-only memory, we
350 * don't want the Linux-specific update to the timeval to
351 * cause a fault after the select has completed
352 * successfully. However, because we're not updating the
353 * timeval, we can't restart the system call.
356 sticky:
357 if (ret == -ERESTARTNOHAND)
358 ret = -EINTR;
359 return ret;
363 * Scalable version of the fd_set.
366 typedef struct {
367 unsigned long *in, *out, *ex;
368 unsigned long *res_in, *res_out, *res_ex;
369 } fd_set_bits;
372 * How many longwords for "nr" bits?
374 #define FDS_BITPERLONG (8*sizeof(long))
375 #define FDS_LONGS(nr) (((nr)+FDS_BITPERLONG-1)/FDS_BITPERLONG)
376 #define FDS_BYTES(nr) (FDS_LONGS(nr)*sizeof(long))
379 * We do a VERIFY_WRITE here even though we are only reading this time:
380 * we'll write to it eventually..
382 * Use "unsigned long" accesses to let user-mode fd_set's be long-aligned.
384 static inline
385 int get_fd_set(unsigned long nr, void __user *ufdset, unsigned long *fdset)
387 nr = FDS_BYTES(nr);
388 if (ufdset)
389 return copy_from_user(fdset, ufdset, nr) ? -EFAULT : 0;
391 memset(fdset, 0, nr);
392 return 0;
395 static inline unsigned long __must_check
396 set_fd_set(unsigned long nr, void __user *ufdset, unsigned long *fdset)
398 if (ufdset)
399 return __copy_to_user(ufdset, fdset, FDS_BYTES(nr));
400 return 0;
403 static inline
404 void zero_fd_set(unsigned long nr, unsigned long *fdset)
406 memset(fdset, 0, FDS_BYTES(nr));
409 #define FDS_IN(fds, n) (fds->in + n)
410 #define FDS_OUT(fds, n) (fds->out + n)
411 #define FDS_EX(fds, n) (fds->ex + n)
413 #define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
415 static int max_select_fd(unsigned long n, fd_set_bits *fds)
417 unsigned long *open_fds;
418 unsigned long set;
419 int max;
420 struct fdtable *fdt;
422 /* handle last in-complete long-word first */
423 set = ~(~0UL << (n & (BITS_PER_LONG-1)));
424 n /= BITS_PER_LONG;
425 fdt = files_fdtable(current->files);
426 open_fds = fdt->open_fds + n;
427 max = 0;
428 if (set) {
429 set &= BITS(fds, n);
430 if (set) {
431 if (!(set & ~*open_fds))
432 goto get_max;
433 return -EBADF;
436 while (n) {
437 open_fds--;
438 n--;
439 set = BITS(fds, n);
440 if (!set)
441 continue;
442 if (set & ~*open_fds)
443 return -EBADF;
444 if (max)
445 continue;
446 get_max:
447 do {
448 max++;
449 set >>= 1;
450 } while (set);
451 max += n * BITS_PER_LONG;
454 return max;
457 #define POLLIN_SET (EPOLLRDNORM | EPOLLRDBAND | EPOLLIN | EPOLLHUP | EPOLLERR)
458 #define POLLOUT_SET (EPOLLWRBAND | EPOLLWRNORM | EPOLLOUT | EPOLLERR)
459 #define POLLEX_SET (EPOLLPRI)
461 static inline void wait_key_set(poll_table *wait, unsigned long in,
462 unsigned long out, unsigned long bit,
463 __poll_t ll_flag)
465 wait->_key = POLLEX_SET | ll_flag;
466 if (in & bit)
467 wait->_key |= POLLIN_SET;
468 if (out & bit)
469 wait->_key |= POLLOUT_SET;
472 static int do_select(int n, fd_set_bits *fds, struct timespec64 *end_time)
474 ktime_t expire, *to = NULL;
475 struct poll_wqueues table;
476 poll_table *wait;
477 int retval, i, timed_out = 0;
478 u64 slack = 0;
479 __poll_t busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
480 unsigned long busy_start = 0;
482 rcu_read_lock();
483 retval = max_select_fd(n, fds);
484 rcu_read_unlock();
486 if (retval < 0)
487 return retval;
488 n = retval;
490 poll_initwait(&table);
491 wait = &table.pt;
492 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
493 wait->_qproc = NULL;
494 timed_out = 1;
497 if (end_time && !timed_out)
498 slack = select_estimate_accuracy(end_time);
500 retval = 0;
501 for (;;) {
502 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
503 bool can_busy_loop = false;
505 inp = fds->in; outp = fds->out; exp = fds->ex;
506 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
508 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
509 unsigned long in, out, ex, all_bits, bit = 1, j;
510 unsigned long res_in = 0, res_out = 0, res_ex = 0;
511 __poll_t mask;
513 in = *inp++; out = *outp++; ex = *exp++;
514 all_bits = in | out | ex;
515 if (all_bits == 0) {
516 i += BITS_PER_LONG;
517 continue;
520 for (j = 0; j < BITS_PER_LONG; ++j, ++i, bit <<= 1) {
521 struct fd f;
522 if (i >= n)
523 break;
524 if (!(bit & all_bits))
525 continue;
526 f = fdget(i);
527 if (f.file) {
528 wait_key_set(wait, in, out, bit,
529 busy_flag);
530 mask = vfs_poll(f.file, wait);
532 fdput(f);
533 if ((mask & POLLIN_SET) && (in & bit)) {
534 res_in |= bit;
535 retval++;
536 wait->_qproc = NULL;
538 if ((mask & POLLOUT_SET) && (out & bit)) {
539 res_out |= bit;
540 retval++;
541 wait->_qproc = NULL;
543 if ((mask & POLLEX_SET) && (ex & bit)) {
544 res_ex |= bit;
545 retval++;
546 wait->_qproc = NULL;
548 /* got something, stop busy polling */
549 if (retval) {
550 can_busy_loop = false;
551 busy_flag = 0;
554 * only remember a returned
555 * POLL_BUSY_LOOP if we asked for it
557 } else if (busy_flag & mask)
558 can_busy_loop = true;
562 if (res_in)
563 *rinp = res_in;
564 if (res_out)
565 *routp = res_out;
566 if (res_ex)
567 *rexp = res_ex;
568 cond_resched();
570 wait->_qproc = NULL;
571 if (retval || timed_out || signal_pending(current))
572 break;
573 if (table.error) {
574 retval = table.error;
575 break;
578 /* only if found POLL_BUSY_LOOP sockets && not out of time */
579 if (can_busy_loop && !need_resched()) {
580 if (!busy_start) {
581 busy_start = busy_loop_current_time();
582 continue;
584 if (!busy_loop_timeout(busy_start))
585 continue;
587 busy_flag = 0;
590 * If this is the first loop and we have a timeout
591 * given, then we convert to ktime_t and set the to
592 * pointer to the expiry value.
594 if (end_time && !to) {
595 expire = timespec64_to_ktime(*end_time);
596 to = &expire;
599 if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
600 to, slack))
601 timed_out = 1;
604 poll_freewait(&table);
606 return retval;
610 * We can actually return ERESTARTSYS instead of EINTR, but I'd
611 * like to be certain this leads to no problems. So I return
612 * EINTR just for safety.
614 * Update: ERESTARTSYS breaks at least the xview clock binary, so
615 * I'm trying ERESTARTNOHAND which restart only when you want to.
617 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
618 fd_set __user *exp, struct timespec64 *end_time)
620 fd_set_bits fds;
621 void *bits;
622 int ret, max_fds;
623 size_t size, alloc_size;
624 struct fdtable *fdt;
625 /* Allocate small arguments on the stack to save memory and be faster */
626 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
628 ret = -EINVAL;
629 if (n < 0)
630 goto out_nofds;
632 /* max_fds can increase, so grab it once to avoid race */
633 rcu_read_lock();
634 fdt = files_fdtable(current->files);
635 max_fds = fdt->max_fds;
636 rcu_read_unlock();
637 if (n > max_fds)
638 n = max_fds;
641 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
642 * since we used fdset we need to allocate memory in units of
643 * long-words.
645 size = FDS_BYTES(n);
646 bits = stack_fds;
647 if (size > sizeof(stack_fds) / 6) {
648 /* Not enough space in on-stack array; must use kmalloc */
649 ret = -ENOMEM;
650 if (size > (SIZE_MAX / 6))
651 goto out_nofds;
653 alloc_size = 6 * size;
654 bits = kvmalloc(alloc_size, GFP_KERNEL);
655 if (!bits)
656 goto out_nofds;
658 fds.in = bits;
659 fds.out = bits + size;
660 fds.ex = bits + 2*size;
661 fds.res_in = bits + 3*size;
662 fds.res_out = bits + 4*size;
663 fds.res_ex = bits + 5*size;
665 if ((ret = get_fd_set(n, inp, fds.in)) ||
666 (ret = get_fd_set(n, outp, fds.out)) ||
667 (ret = get_fd_set(n, exp, fds.ex)))
668 goto out;
669 zero_fd_set(n, fds.res_in);
670 zero_fd_set(n, fds.res_out);
671 zero_fd_set(n, fds.res_ex);
673 ret = do_select(n, &fds, end_time);
675 if (ret < 0)
676 goto out;
677 if (!ret) {
678 ret = -ERESTARTNOHAND;
679 if (signal_pending(current))
680 goto out;
681 ret = 0;
684 if (set_fd_set(n, inp, fds.res_in) ||
685 set_fd_set(n, outp, fds.res_out) ||
686 set_fd_set(n, exp, fds.res_ex))
687 ret = -EFAULT;
689 out:
690 if (bits != stack_fds)
691 kvfree(bits);
692 out_nofds:
693 return ret;
696 static int kern_select(int n, fd_set __user *inp, fd_set __user *outp,
697 fd_set __user *exp, struct timeval __user *tvp)
699 struct timespec64 end_time, *to = NULL;
700 struct timeval tv;
701 int ret;
703 if (tvp) {
704 if (copy_from_user(&tv, tvp, sizeof(tv)))
705 return -EFAULT;
707 to = &end_time;
708 if (poll_select_set_timeout(to,
709 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
710 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
711 return -EINVAL;
714 ret = core_sys_select(n, inp, outp, exp, to);
715 ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
717 return ret;
720 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
721 fd_set __user *, exp, struct timeval __user *, tvp)
723 return kern_select(n, inp, outp, exp, tvp);
726 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
727 fd_set __user *exp, struct timespec __user *tsp,
728 const sigset_t __user *sigmask, size_t sigsetsize)
730 sigset_t ksigmask, sigsaved;
731 struct timespec64 ts, end_time, *to = NULL;
732 int ret;
734 if (tsp) {
735 if (get_timespec64(&ts, tsp))
736 return -EFAULT;
738 to = &end_time;
739 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
740 return -EINVAL;
743 if (sigmask) {
744 /* XXX: Don't preclude handling different sized sigset_t's. */
745 if (sigsetsize != sizeof(sigset_t))
746 return -EINVAL;
747 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
748 return -EFAULT;
750 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
751 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
754 ret = core_sys_select(n, inp, outp, exp, to);
755 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
757 if (ret == -ERESTARTNOHAND) {
759 * Don't restore the signal mask yet. Let do_signal() deliver
760 * the signal on the way back to userspace, before the signal
761 * mask is restored.
763 if (sigmask) {
764 memcpy(&current->saved_sigmask, &sigsaved,
765 sizeof(sigsaved));
766 set_restore_sigmask();
768 } else if (sigmask)
769 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
771 return ret;
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.
780 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
781 fd_set __user *, exp, struct timespec __user *, tsp,
782 void __user *, sig)
784 size_t sigsetsize = 0;
785 sigset_t __user *up = NULL;
787 if (sig) {
788 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
789 || __get_user(up, (sigset_t __user * __user *)sig)
790 || __get_user(sigsetsize,
791 (size_t __user *)(sig+sizeof(void *))))
792 return -EFAULT;
795 return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
798 #ifdef __ARCH_WANT_SYS_OLD_SELECT
799 struct sel_arg_struct {
800 unsigned long n;
801 fd_set __user *inp, *outp, *exp;
802 struct timeval __user *tvp;
805 SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg)
807 struct sel_arg_struct a;
809 if (copy_from_user(&a, arg, sizeof(a)))
810 return -EFAULT;
811 return kern_select(a.n, a.inp, a.outp, a.exp, a.tvp);
813 #endif
815 struct poll_list {
816 struct poll_list *next;
817 int len;
818 struct pollfd entries[0];
821 #define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
824 * Fish for pollable events on the pollfd->fd file descriptor. We're only
825 * interested in events matching the pollfd->events mask, and the result
826 * matching that mask is both recorded in pollfd->revents and returned. The
827 * pwait poll_table will be used by the fd-provided poll handler for waiting,
828 * if pwait->_qproc is non-NULL.
830 static inline __poll_t do_pollfd(struct pollfd *pollfd, poll_table *pwait,
831 bool *can_busy_poll,
832 __poll_t busy_flag)
834 int fd = pollfd->fd;
835 __poll_t mask = 0, filter;
836 struct fd f;
838 if (fd < 0)
839 goto out;
840 mask = EPOLLNVAL;
841 f = fdget(fd);
842 if (!f.file)
843 goto out;
845 /* userland u16 ->events contains POLL... bitmap */
846 filter = demangle_poll(pollfd->events) | EPOLLERR | EPOLLHUP;
847 pwait->_key = filter | busy_flag;
848 mask = vfs_poll(f.file, pwait);
849 if (mask & busy_flag)
850 *can_busy_poll = true;
851 mask &= filter; /* Mask out unneeded events. */
852 fdput(f);
854 out:
855 /* ... and so does ->revents */
856 pollfd->revents = mangle_poll(mask);
857 return mask;
860 static int do_poll(struct poll_list *list, struct poll_wqueues *wait,
861 struct timespec64 *end_time)
863 poll_table* pt = &wait->pt;
864 ktime_t expire, *to = NULL;
865 int timed_out = 0, count = 0;
866 u64 slack = 0;
867 __poll_t busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
868 unsigned long busy_start = 0;
870 /* Optimise the no-wait case */
871 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
872 pt->_qproc = NULL;
873 timed_out = 1;
876 if (end_time && !timed_out)
877 slack = select_estimate_accuracy(end_time);
879 for (;;) {
880 struct poll_list *walk;
881 bool can_busy_loop = false;
883 for (walk = list; walk != NULL; walk = walk->next) {
884 struct pollfd * pfd, * pfd_end;
886 pfd = walk->entries;
887 pfd_end = pfd + walk->len;
888 for (; pfd != pfd_end; pfd++) {
890 * Fish for events. If we found one, record it
891 * and kill poll_table->_qproc, so we don't
892 * needlessly register any other waiters after
893 * this. They'll get immediately deregistered
894 * when we break out and return.
896 if (do_pollfd(pfd, pt, &can_busy_loop,
897 busy_flag)) {
898 count++;
899 pt->_qproc = NULL;
900 /* found something, stop busy polling */
901 busy_flag = 0;
902 can_busy_loop = false;
907 * All waiters have already been registered, so don't provide
908 * a poll_table->_qproc to them on the next loop iteration.
910 pt->_qproc = NULL;
911 if (!count) {
912 count = wait->error;
913 if (signal_pending(current))
914 count = -EINTR;
916 if (count || timed_out)
917 break;
919 /* only if found POLL_BUSY_LOOP sockets && not out of time */
920 if (can_busy_loop && !need_resched()) {
921 if (!busy_start) {
922 busy_start = busy_loop_current_time();
923 continue;
925 if (!busy_loop_timeout(busy_start))
926 continue;
928 busy_flag = 0;
931 * If this is the first loop and we have a timeout
932 * given, then we convert to ktime_t and set the to
933 * pointer to the expiry value.
935 if (end_time && !to) {
936 expire = timespec64_to_ktime(*end_time);
937 to = &expire;
940 if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
941 timed_out = 1;
943 return count;
946 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \
947 sizeof(struct pollfd))
949 static int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
950 struct timespec64 *end_time)
952 struct poll_wqueues table;
953 int err = -EFAULT, fdcount, len, size;
954 /* Allocate small arguments on the stack to save memory and be
955 faster - use long to make sure the buffer is aligned properly
956 on 64 bit archs to avoid unaligned access */
957 long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
958 struct poll_list *const head = (struct poll_list *)stack_pps;
959 struct poll_list *walk = head;
960 unsigned long todo = nfds;
962 if (nfds > rlimit(RLIMIT_NOFILE))
963 return -EINVAL;
965 len = min_t(unsigned int, nfds, N_STACK_PPS);
966 for (;;) {
967 walk->next = NULL;
968 walk->len = len;
969 if (!len)
970 break;
972 if (copy_from_user(walk->entries, ufds + nfds-todo,
973 sizeof(struct pollfd) * walk->len))
974 goto out_fds;
976 todo -= walk->len;
977 if (!todo)
978 break;
980 len = min(todo, POLLFD_PER_PAGE);
981 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
982 walk = walk->next = kmalloc(size, GFP_KERNEL);
983 if (!walk) {
984 err = -ENOMEM;
985 goto out_fds;
989 poll_initwait(&table);
990 fdcount = do_poll(head, &table, end_time);
991 poll_freewait(&table);
993 for (walk = head; walk; walk = walk->next) {
994 struct pollfd *fds = walk->entries;
995 int j;
997 for (j = 0; j < walk->len; j++, ufds++)
998 if (__put_user(fds[j].revents, &ufds->revents))
999 goto out_fds;
1002 err = fdcount;
1003 out_fds:
1004 walk = head->next;
1005 while (walk) {
1006 struct poll_list *pos = walk;
1007 walk = walk->next;
1008 kfree(pos);
1011 return err;
1014 static long do_restart_poll(struct restart_block *restart_block)
1016 struct pollfd __user *ufds = restart_block->poll.ufds;
1017 int nfds = restart_block->poll.nfds;
1018 struct timespec64 *to = NULL, end_time;
1019 int ret;
1021 if (restart_block->poll.has_timeout) {
1022 end_time.tv_sec = restart_block->poll.tv_sec;
1023 end_time.tv_nsec = restart_block->poll.tv_nsec;
1024 to = &end_time;
1027 ret = do_sys_poll(ufds, nfds, to);
1029 if (ret == -EINTR) {
1030 restart_block->fn = do_restart_poll;
1031 ret = -ERESTART_RESTARTBLOCK;
1033 return ret;
1036 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
1037 int, timeout_msecs)
1039 struct timespec64 end_time, *to = NULL;
1040 int ret;
1042 if (timeout_msecs >= 0) {
1043 to = &end_time;
1044 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
1045 NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
1048 ret = do_sys_poll(ufds, nfds, to);
1050 if (ret == -EINTR) {
1051 struct restart_block *restart_block;
1053 restart_block = &current->restart_block;
1054 restart_block->fn = do_restart_poll;
1055 restart_block->poll.ufds = ufds;
1056 restart_block->poll.nfds = nfds;
1058 if (timeout_msecs >= 0) {
1059 restart_block->poll.tv_sec = end_time.tv_sec;
1060 restart_block->poll.tv_nsec = end_time.tv_nsec;
1061 restart_block->poll.has_timeout = 1;
1062 } else
1063 restart_block->poll.has_timeout = 0;
1065 ret = -ERESTART_RESTARTBLOCK;
1067 return ret;
1070 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
1071 struct timespec __user *, tsp, const sigset_t __user *, sigmask,
1072 size_t, sigsetsize)
1074 sigset_t ksigmask, sigsaved;
1075 struct timespec64 ts, end_time, *to = NULL;
1076 int ret;
1078 if (tsp) {
1079 if (get_timespec64(&ts, tsp))
1080 return -EFAULT;
1082 to = &end_time;
1083 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1084 return -EINVAL;
1087 if (sigmask) {
1088 /* XXX: Don't preclude handling different sized sigset_t's. */
1089 if (sigsetsize != sizeof(sigset_t))
1090 return -EINVAL;
1091 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1092 return -EFAULT;
1094 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
1095 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1098 ret = do_sys_poll(ufds, nfds, to);
1100 /* We can restart this syscall, usually */
1101 if (ret == -EINTR) {
1103 * Don't restore the signal mask yet. Let do_signal() deliver
1104 * the signal on the way back to userspace, before the signal
1105 * mask is restored.
1107 if (sigmask) {
1108 memcpy(&current->saved_sigmask, &sigsaved,
1109 sizeof(sigsaved));
1110 set_restore_sigmask();
1112 ret = -ERESTARTNOHAND;
1113 } else if (sigmask)
1114 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1116 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
1118 return ret;
1121 #ifdef CONFIG_COMPAT
1122 #define __COMPAT_NFDBITS (8 * sizeof(compat_ulong_t))
1124 static
1125 int compat_poll_select_copy_remaining(struct timespec64 *end_time, void __user *p,
1126 int timeval, int ret)
1128 struct timespec64 ts;
1130 if (!p)
1131 return ret;
1133 if (current->personality & STICKY_TIMEOUTS)
1134 goto sticky;
1136 /* No update for zero timeout */
1137 if (!end_time->tv_sec && !end_time->tv_nsec)
1138 return ret;
1140 ktime_get_ts64(&ts);
1141 ts = timespec64_sub(*end_time, ts);
1142 if (ts.tv_sec < 0)
1143 ts.tv_sec = ts.tv_nsec = 0;
1145 if (timeval) {
1146 struct compat_timeval rtv;
1148 rtv.tv_sec = ts.tv_sec;
1149 rtv.tv_usec = ts.tv_nsec / NSEC_PER_USEC;
1151 if (!copy_to_user(p, &rtv, sizeof(rtv)))
1152 return ret;
1153 } else {
1154 if (!compat_put_timespec64(&ts, p))
1155 return ret;
1158 * If an application puts its timeval in read-only memory, we
1159 * don't want the Linux-specific update to the timeval to
1160 * cause a fault after the select has completed
1161 * successfully. However, because we're not updating the
1162 * timeval, we can't restart the system call.
1165 sticky:
1166 if (ret == -ERESTARTNOHAND)
1167 ret = -EINTR;
1168 return ret;
1172 * Ooo, nasty. We need here to frob 32-bit unsigned longs to
1173 * 64-bit unsigned longs.
1175 static
1176 int compat_get_fd_set(unsigned long nr, compat_ulong_t __user *ufdset,
1177 unsigned long *fdset)
1179 if (ufdset) {
1180 return compat_get_bitmap(fdset, ufdset, nr);
1181 } else {
1182 zero_fd_set(nr, fdset);
1183 return 0;
1187 static
1188 int compat_set_fd_set(unsigned long nr, compat_ulong_t __user *ufdset,
1189 unsigned long *fdset)
1191 if (!ufdset)
1192 return 0;
1193 return compat_put_bitmap(ufdset, fdset, nr);
1198 * This is a virtual copy of sys_select from fs/select.c and probably
1199 * should be compared to it from time to time
1203 * We can actually return ERESTARTSYS instead of EINTR, but I'd
1204 * like to be certain this leads to no problems. So I return
1205 * EINTR just for safety.
1207 * Update: ERESTARTSYS breaks at least the xview clock binary, so
1208 * I'm trying ERESTARTNOHAND which restart only when you want to.
1210 static int compat_core_sys_select(int n, compat_ulong_t __user *inp,
1211 compat_ulong_t __user *outp, compat_ulong_t __user *exp,
1212 struct timespec64 *end_time)
1214 fd_set_bits fds;
1215 void *bits;
1216 int size, max_fds, ret = -EINVAL;
1217 struct fdtable *fdt;
1218 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
1220 if (n < 0)
1221 goto out_nofds;
1223 /* max_fds can increase, so grab it once to avoid race */
1224 rcu_read_lock();
1225 fdt = files_fdtable(current->files);
1226 max_fds = fdt->max_fds;
1227 rcu_read_unlock();
1228 if (n > max_fds)
1229 n = max_fds;
1232 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
1233 * since we used fdset we need to allocate memory in units of
1234 * long-words.
1236 size = FDS_BYTES(n);
1237 bits = stack_fds;
1238 if (size > sizeof(stack_fds) / 6) {
1239 bits = kmalloc_array(6, size, GFP_KERNEL);
1240 ret = -ENOMEM;
1241 if (!bits)
1242 goto out_nofds;
1244 fds.in = (unsigned long *) bits;
1245 fds.out = (unsigned long *) (bits + size);
1246 fds.ex = (unsigned long *) (bits + 2*size);
1247 fds.res_in = (unsigned long *) (bits + 3*size);
1248 fds.res_out = (unsigned long *) (bits + 4*size);
1249 fds.res_ex = (unsigned long *) (bits + 5*size);
1251 if ((ret = compat_get_fd_set(n, inp, fds.in)) ||
1252 (ret = compat_get_fd_set(n, outp, fds.out)) ||
1253 (ret = compat_get_fd_set(n, exp, fds.ex)))
1254 goto out;
1255 zero_fd_set(n, fds.res_in);
1256 zero_fd_set(n, fds.res_out);
1257 zero_fd_set(n, fds.res_ex);
1259 ret = do_select(n, &fds, end_time);
1261 if (ret < 0)
1262 goto out;
1263 if (!ret) {
1264 ret = -ERESTARTNOHAND;
1265 if (signal_pending(current))
1266 goto out;
1267 ret = 0;
1270 if (compat_set_fd_set(n, inp, fds.res_in) ||
1271 compat_set_fd_set(n, outp, fds.res_out) ||
1272 compat_set_fd_set(n, exp, fds.res_ex))
1273 ret = -EFAULT;
1274 out:
1275 if (bits != stack_fds)
1276 kfree(bits);
1277 out_nofds:
1278 return ret;
1281 static int do_compat_select(int n, compat_ulong_t __user *inp,
1282 compat_ulong_t __user *outp, compat_ulong_t __user *exp,
1283 struct compat_timeval __user *tvp)
1285 struct timespec64 end_time, *to = NULL;
1286 struct compat_timeval tv;
1287 int ret;
1289 if (tvp) {
1290 if (copy_from_user(&tv, tvp, sizeof(tv)))
1291 return -EFAULT;
1293 to = &end_time;
1294 if (poll_select_set_timeout(to,
1295 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
1296 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
1297 return -EINVAL;
1300 ret = compat_core_sys_select(n, inp, outp, exp, to);
1301 ret = compat_poll_select_copy_remaining(&end_time, tvp, 1, ret);
1303 return ret;
1306 COMPAT_SYSCALL_DEFINE5(select, int, n, compat_ulong_t __user *, inp,
1307 compat_ulong_t __user *, outp, compat_ulong_t __user *, exp,
1308 struct compat_timeval __user *, tvp)
1310 return do_compat_select(n, inp, outp, exp, tvp);
1313 struct compat_sel_arg_struct {
1314 compat_ulong_t n;
1315 compat_uptr_t inp;
1316 compat_uptr_t outp;
1317 compat_uptr_t exp;
1318 compat_uptr_t tvp;
1321 COMPAT_SYSCALL_DEFINE1(old_select, struct compat_sel_arg_struct __user *, arg)
1323 struct compat_sel_arg_struct a;
1325 if (copy_from_user(&a, arg, sizeof(a)))
1326 return -EFAULT;
1327 return do_compat_select(a.n, compat_ptr(a.inp), compat_ptr(a.outp),
1328 compat_ptr(a.exp), compat_ptr(a.tvp));
1331 static long do_compat_pselect(int n, compat_ulong_t __user *inp,
1332 compat_ulong_t __user *outp, compat_ulong_t __user *exp,
1333 struct compat_timespec __user *tsp, compat_sigset_t __user *sigmask,
1334 compat_size_t sigsetsize)
1336 sigset_t ksigmask, sigsaved;
1337 struct timespec64 ts, end_time, *to = NULL;
1338 int ret;
1340 if (tsp) {
1341 if (compat_get_timespec64(&ts, tsp))
1342 return -EFAULT;
1344 to = &end_time;
1345 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1346 return -EINVAL;
1349 if (sigmask) {
1350 if (sigsetsize != sizeof(compat_sigset_t))
1351 return -EINVAL;
1352 if (get_compat_sigset(&ksigmask, sigmask))
1353 return -EFAULT;
1355 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
1356 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1359 ret = compat_core_sys_select(n, inp, outp, exp, to);
1360 ret = compat_poll_select_copy_remaining(&end_time, tsp, 0, ret);
1362 if (ret == -ERESTARTNOHAND) {
1364 * Don't restore the signal mask yet. Let do_signal() deliver
1365 * the signal on the way back to userspace, before the signal
1366 * mask is restored.
1368 if (sigmask) {
1369 memcpy(&current->saved_sigmask, &sigsaved,
1370 sizeof(sigsaved));
1371 set_restore_sigmask();
1373 } else if (sigmask)
1374 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1376 return ret;
1379 COMPAT_SYSCALL_DEFINE6(pselect6, int, n, compat_ulong_t __user *, inp,
1380 compat_ulong_t __user *, outp, compat_ulong_t __user *, exp,
1381 struct compat_timespec __user *, tsp, void __user *, sig)
1383 compat_size_t sigsetsize = 0;
1384 compat_uptr_t up = 0;
1386 if (sig) {
1387 if (!access_ok(VERIFY_READ, sig,
1388 sizeof(compat_uptr_t)+sizeof(compat_size_t)) ||
1389 __get_user(up, (compat_uptr_t __user *)sig) ||
1390 __get_user(sigsetsize,
1391 (compat_size_t __user *)(sig+sizeof(up))))
1392 return -EFAULT;
1394 return do_compat_pselect(n, inp, outp, exp, tsp, compat_ptr(up),
1395 sigsetsize);
1398 COMPAT_SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds,
1399 unsigned int, nfds, struct compat_timespec __user *, tsp,
1400 const compat_sigset_t __user *, sigmask, compat_size_t, sigsetsize)
1402 sigset_t ksigmask, sigsaved;
1403 struct timespec64 ts, end_time, *to = NULL;
1404 int ret;
1406 if (tsp) {
1407 if (compat_get_timespec64(&ts, tsp))
1408 return -EFAULT;
1410 to = &end_time;
1411 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1412 return -EINVAL;
1415 if (sigmask) {
1416 if (sigsetsize != sizeof(compat_sigset_t))
1417 return -EINVAL;
1418 if (get_compat_sigset(&ksigmask, sigmask))
1419 return -EFAULT;
1421 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
1422 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1425 ret = do_sys_poll(ufds, nfds, to);
1427 /* We can restart this syscall, usually */
1428 if (ret == -EINTR) {
1430 * Don't restore the signal mask yet. Let do_signal() deliver
1431 * the signal on the way back to userspace, before the signal
1432 * mask is restored.
1434 if (sigmask) {
1435 memcpy(&current->saved_sigmask, &sigsaved,
1436 sizeof(sigsaved));
1437 set_restore_sigmask();
1439 ret = -ERESTARTNOHAND;
1440 } else if (sigmask)
1441 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1443 ret = compat_poll_select_copy_remaining(&end_time, tsp, 0, ret);
1445 return ret;
1447 #endif