2 * This file contains the procedures for the handling of select and poll
4 * Created for Linux based loosely upon Mathius Lattner's minix
5 * patches by Peter MacDonald. Heavily edited by Linus.
8 * COFF/ELF binary emulation. If the process has the STICKY_TIMEOUTS
9 * flag set in its personality we do *not* modify the given timeout
10 * parameter to reflect time remaining.
13 * Changed sys_poll()/do_poll() to use PAGE_SIZE chunk-based allocation
14 * of fds to overcome nfds < 16390 descriptors limit (Tigran Aivazian).
17 #include <linux/kernel.h>
18 #include <linux/syscalls.h>
19 #include <linux/module.h>
20 #include <linux/slab.h>
21 #include <linux/poll.h>
22 #include <linux/personality.h> /* for STICKY_TIMEOUTS */
23 #include <linux/file.h>
24 #include <linux/fdtable.h>
26 #include <linux/rcupdate.h>
28 #include <asm/uaccess.h>
30 struct poll_table_page {
31 struct poll_table_page * next;
32 struct poll_table_entry * entry;
33 struct poll_table_entry entries[0];
36 #define POLL_TABLE_FULL(table) \
37 ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
40 * Ok, Peter made a complicated, but straightforward multiple_wait() function.
41 * I have rewritten this, taking some shortcuts: This code may not be easy to
42 * follow, but it should be free of race-conditions, and it's practical. If you
43 * understand what I'm doing here, then you understand how the linux
44 * sleep/wakeup mechanism works.
46 * Two very simple procedures, poll_wait() and poll_freewait() make all the
47 * work. poll_wait() is an inline-function defined in <linux/poll.h>,
48 * as all select/poll functions have to call it to add an entry to the
51 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
54 void poll_initwait(struct poll_wqueues *pwq)
56 init_poll_funcptr(&pwq->pt, __pollwait);
59 pwq->inline_index = 0;
62 EXPORT_SYMBOL(poll_initwait);
64 static void free_poll_entry(struct poll_table_entry *entry)
66 remove_wait_queue(entry->wait_address, &entry->wait);
70 void poll_freewait(struct poll_wqueues *pwq)
72 struct poll_table_page * p = pwq->table;
74 for (i = 0; i < pwq->inline_index; i++)
75 free_poll_entry(pwq->inline_entries + i);
77 struct poll_table_entry * entry;
78 struct poll_table_page *old;
83 free_poll_entry(entry);
84 } while (entry > p->entries);
87 free_page((unsigned long) old);
91 EXPORT_SYMBOL(poll_freewait);
93 static struct poll_table_entry *poll_get_entry(poll_table *_p)
95 struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt);
96 struct poll_table_page *table = p->table;
98 if (p->inline_index < N_INLINE_POLL_ENTRIES)
99 return p->inline_entries + p->inline_index++;
101 if (!table || POLL_TABLE_FULL(table)) {
102 struct poll_table_page *new_table;
104 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
107 __set_current_state(TASK_RUNNING);
110 new_table->entry = new_table->entries;
111 new_table->next = table;
112 p->table = new_table;
116 return table->entry++;
119 /* Add a new entry */
120 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
123 struct poll_table_entry *entry = poll_get_entry(p);
128 entry->wait_address = wait_address;
129 init_waitqueue_entry(&entry->wait, current);
130 add_wait_queue(wait_address, &entry->wait);
133 #define FDS_IN(fds, n) (fds->in + n)
134 #define FDS_OUT(fds, n) (fds->out + n)
135 #define FDS_EX(fds, n) (fds->ex + n)
137 #define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
139 static int max_select_fd(unsigned long n, fd_set_bits *fds)
141 unsigned long *open_fds;
146 /* handle last in-complete long-word first */
147 set = ~(~0UL << (n & (__NFDBITS-1)));
149 fdt = files_fdtable(current->files);
150 open_fds = fdt->open_fds->fds_bits+n;
155 if (!(set & ~*open_fds))
166 if (set & ~*open_fds)
175 max += n * __NFDBITS;
181 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
182 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
183 #define POLLEX_SET (POLLPRI)
185 int do_select(int n, fd_set_bits *fds, s64 *timeout)
187 struct poll_wqueues table;
192 retval = max_select_fd(n, fds);
199 poll_initwait(&table);
205 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
208 set_current_state(TASK_INTERRUPTIBLE);
210 inp = fds->in; outp = fds->out; exp = fds->ex;
211 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
213 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
214 unsigned long in, out, ex, all_bits, bit = 1, mask, j;
215 unsigned long res_in = 0, res_out = 0, res_ex = 0;
216 const struct file_operations *f_op = NULL;
217 struct file *file = NULL;
219 in = *inp++; out = *outp++; ex = *exp++;
220 all_bits = in | out | ex;
226 for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
230 if (!(bit & all_bits))
232 file = fget_light(i, &fput_needed);
235 mask = DEFAULT_POLLMASK;
236 if (f_op && f_op->poll)
237 mask = (*f_op->poll)(file, retval ? NULL : wait);
238 fput_light(file, fput_needed);
239 if ((mask & POLLIN_SET) && (in & bit)) {
243 if ((mask & POLLOUT_SET) && (out & bit)) {
247 if ((mask & POLLEX_SET) && (ex & bit)) {
262 if (retval || !*timeout || signal_pending(current))
265 retval = table.error;
270 /* Wait indefinitely */
271 __timeout = MAX_SCHEDULE_TIMEOUT;
272 } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT - 1)) {
273 /* Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in a loop */
274 __timeout = MAX_SCHEDULE_TIMEOUT - 1;
275 *timeout -= __timeout;
277 __timeout = *timeout;
280 __timeout = schedule_timeout(__timeout);
282 *timeout += __timeout;
284 __set_current_state(TASK_RUNNING);
286 poll_freewait(&table);
292 * We can actually return ERESTARTSYS instead of EINTR, but I'd
293 * like to be certain this leads to no problems. So I return
294 * EINTR just for safety.
296 * Update: ERESTARTSYS breaks at least the xview clock binary, so
297 * I'm trying ERESTARTNOHAND which restart only when you want to.
299 #define MAX_SELECT_SECONDS \
300 ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
302 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
303 fd_set __user *exp, s64 *timeout)
310 /* Allocate small arguments on the stack to save memory and be faster */
311 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
317 /* max_fds can increase, so grab it once to avoid race */
319 fdt = files_fdtable(current->files);
320 max_fds = fdt->max_fds;
326 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
327 * since we used fdset we need to allocate memory in units of
332 if (size > sizeof(stack_fds) / 6) {
333 /* Not enough space in on-stack array; must use kmalloc */
335 bits = kmalloc(6 * size, GFP_KERNEL);
340 fds.out = bits + size;
341 fds.ex = bits + 2*size;
342 fds.res_in = bits + 3*size;
343 fds.res_out = bits + 4*size;
344 fds.res_ex = bits + 5*size;
346 if ((ret = get_fd_set(n, inp, fds.in)) ||
347 (ret = get_fd_set(n, outp, fds.out)) ||
348 (ret = get_fd_set(n, exp, fds.ex)))
350 zero_fd_set(n, fds.res_in);
351 zero_fd_set(n, fds.res_out);
352 zero_fd_set(n, fds.res_ex);
354 ret = do_select(n, &fds, timeout);
359 ret = -ERESTARTNOHAND;
360 if (signal_pending(current))
365 if (set_fd_set(n, inp, fds.res_in) ||
366 set_fd_set(n, outp, fds.res_out) ||
367 set_fd_set(n, exp, fds.res_ex))
371 if (bits != stack_fds)
377 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
378 fd_set __user *, exp, struct timeval __user *, tvp)
385 if (copy_from_user(&tv, tvp, sizeof(tv)))
388 if (tv.tv_sec < 0 || tv.tv_usec < 0)
391 /* Cast to u64 to make GCC stop complaining */
392 if ((u64)tv.tv_sec >= (u64)MAX_INT64_SECONDS)
393 timeout = -1; /* infinite */
395 timeout = DIV_ROUND_UP(tv.tv_usec, USEC_PER_SEC/HZ);
396 timeout += tv.tv_sec * HZ;
400 ret = core_sys_select(n, inp, outp, exp, &timeout);
405 if (current->personality & STICKY_TIMEOUTS)
407 rtv.tv_usec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ));
408 rtv.tv_sec = timeout;
409 if (timeval_compare(&rtv, &tv) >= 0)
411 if (copy_to_user(tvp, &rtv, sizeof(rtv))) {
414 * If an application puts its timeval in read-only
415 * memory, we don't want the Linux-specific update to
416 * the timeval to cause a fault after the select has
417 * completed successfully. However, because we're not
418 * updating the timeval, we can't restart the system
421 if (ret == -ERESTARTNOHAND)
429 #ifdef HAVE_SET_RESTORE_SIGMASK
430 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
431 fd_set __user *exp, struct timespec __user *tsp,
432 const sigset_t __user *sigmask, size_t sigsetsize)
434 s64 timeout = MAX_SCHEDULE_TIMEOUT;
435 sigset_t ksigmask, sigsaved;
440 if (copy_from_user(&ts, tsp, sizeof(ts)))
443 if (ts.tv_sec < 0 || ts.tv_nsec < 0)
446 /* Cast to u64 to make GCC stop complaining */
447 if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
448 timeout = -1; /* infinite */
450 timeout = DIV_ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
451 timeout += ts.tv_sec * HZ;
456 /* XXX: Don't preclude handling different sized sigset_t's. */
457 if (sigsetsize != sizeof(sigset_t))
459 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
462 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
463 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
466 ret = core_sys_select(n, inp, outp, exp, &timeout);
471 if (current->personality & STICKY_TIMEOUTS)
473 rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
475 rts.tv_sec = timeout;
476 if (timespec_compare(&rts, &ts) >= 0)
478 if (copy_to_user(tsp, &rts, sizeof(rts))) {
481 * If an application puts its timeval in read-only
482 * memory, we don't want the Linux-specific update to
483 * the timeval to cause a fault after the select has
484 * completed successfully. However, because we're not
485 * updating the timeval, we can't restart the system
488 if (ret == -ERESTARTNOHAND)
493 if (ret == -ERESTARTNOHAND) {
495 * Don't restore the signal mask yet. Let do_signal() deliver
496 * the signal on the way back to userspace, before the signal
500 memcpy(¤t->saved_sigmask, &sigsaved,
502 set_restore_sigmask();
505 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
511 * Most architectures can't handle 7-argument syscalls. So we provide a
512 * 6-argument version where the sixth argument is a pointer to a structure
513 * which has a pointer to the sigset_t itself followed by a size_t containing
516 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
517 fd_set __user *, exp, struct timespec __user *, tsp,
520 size_t sigsetsize = 0;
521 sigset_t __user *up = NULL;
524 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
525 || __get_user(up, (sigset_t __user * __user *)sig)
526 || __get_user(sigsetsize,
527 (size_t __user *)(sig+sizeof(void *))))
531 return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
533 #endif /* HAVE_SET_RESTORE_SIGMASK */
536 struct poll_list *next;
538 struct pollfd entries[0];
541 #define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
544 * Fish for pollable events on the pollfd->fd file descriptor. We're only
545 * interested in events matching the pollfd->events mask, and the result
546 * matching that mask is both recorded in pollfd->revents and returned. The
547 * pwait poll_table will be used by the fd-provided poll handler for waiting,
550 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
561 file = fget_light(fd, &fput_needed);
564 mask = DEFAULT_POLLMASK;
565 if (file->f_op && file->f_op->poll)
566 mask = file->f_op->poll(file, pwait);
567 /* Mask out unneeded events. */
568 mask &= pollfd->events | POLLERR | POLLHUP;
569 fput_light(file, fput_needed);
572 pollfd->revents = mask;
577 static int do_poll(unsigned int nfds, struct poll_list *list,
578 struct poll_wqueues *wait, s64 *timeout)
581 poll_table* pt = &wait->pt;
583 /* Optimise the no-wait case */
588 struct poll_list *walk;
591 set_current_state(TASK_INTERRUPTIBLE);
592 for (walk = list; walk != NULL; walk = walk->next) {
593 struct pollfd * pfd, * pfd_end;
596 pfd_end = pfd + walk->len;
597 for (; pfd != pfd_end; pfd++) {
599 * Fish for events. If we found one, record it
600 * and kill the poll_table, so we don't
601 * needlessly register any other waiters after
602 * this. They'll get immediately deregistered
603 * when we break out and return.
605 if (do_pollfd(pfd, pt)) {
612 * All waiters have already been registered, so don't provide
613 * a poll_table to them on the next loop iteration.
618 if (signal_pending(current))
621 if (count || !*timeout)
625 /* Wait indefinitely */
626 __timeout = MAX_SCHEDULE_TIMEOUT;
627 } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT-1)) {
629 * Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in
632 __timeout = MAX_SCHEDULE_TIMEOUT - 1;
633 *timeout -= __timeout;
635 __timeout = *timeout;
639 __timeout = schedule_timeout(__timeout);
641 *timeout += __timeout;
643 __set_current_state(TASK_RUNNING);
647 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \
648 sizeof(struct pollfd))
650 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds, s64 *timeout)
652 struct poll_wqueues table;
653 int err = -EFAULT, fdcount, len, size;
654 /* Allocate small arguments on the stack to save memory and be
655 faster - use long to make sure the buffer is aligned properly
656 on 64 bit archs to avoid unaligned access */
657 long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
658 struct poll_list *const head = (struct poll_list *)stack_pps;
659 struct poll_list *walk = head;
660 unsigned long todo = nfds;
662 if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
665 len = min_t(unsigned int, nfds, N_STACK_PPS);
672 if (copy_from_user(walk->entries, ufds + nfds-todo,
673 sizeof(struct pollfd) * walk->len))
680 len = min(todo, POLLFD_PER_PAGE);
681 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
682 walk = walk->next = kmalloc(size, GFP_KERNEL);
689 poll_initwait(&table);
690 fdcount = do_poll(nfds, head, &table, timeout);
691 poll_freewait(&table);
693 for (walk = head; walk; walk = walk->next) {
694 struct pollfd *fds = walk->entries;
697 for (j = 0; j < walk->len; j++, ufds++)
698 if (__put_user(fds[j].revents, &ufds->revents))
706 struct poll_list *pos = walk;
714 static long do_restart_poll(struct restart_block *restart_block)
716 struct pollfd __user *ufds = (struct pollfd __user*)restart_block->arg0;
717 int nfds = restart_block->arg1;
718 s64 timeout = ((s64)restart_block->arg3<<32) | (s64)restart_block->arg2;
721 ret = do_sys_poll(ufds, nfds, &timeout);
723 restart_block->fn = do_restart_poll;
724 restart_block->arg2 = timeout & 0xFFFFFFFF;
725 restart_block->arg3 = (u64)timeout >> 32;
726 ret = -ERESTART_RESTARTBLOCK;
731 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
737 if (timeout_msecs > 0) {
739 /* We can only overflow if HZ > 1000 */
740 if (timeout_msecs / 1000 > (s64)0x7fffffffffffffffULL / (s64)HZ)
741 timeout_jiffies = -1;
744 timeout_jiffies = msecs_to_jiffies(timeout_msecs) + 1;
746 /* Infinite (< 0) or no (0) timeout */
747 timeout_jiffies = timeout_msecs;
750 ret = do_sys_poll(ufds, nfds, &timeout_jiffies);
752 struct restart_block *restart_block;
753 restart_block = ¤t_thread_info()->restart_block;
754 restart_block->fn = do_restart_poll;
755 restart_block->arg0 = (unsigned long)ufds;
756 restart_block->arg1 = nfds;
757 restart_block->arg2 = timeout_jiffies & 0xFFFFFFFF;
758 restart_block->arg3 = (u64)timeout_jiffies >> 32;
759 ret = -ERESTART_RESTARTBLOCK;
764 #ifdef HAVE_SET_RESTORE_SIGMASK
765 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
766 struct timespec __user *, tsp, const sigset_t __user *, sigmask,
769 sigset_t ksigmask, sigsaved;
775 if (copy_from_user(&ts, tsp, sizeof(ts)))
778 /* Cast to u64 to make GCC stop complaining */
779 if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
780 timeout = -1; /* infinite */
782 timeout = DIV_ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
783 timeout += ts.tv_sec * HZ;
788 /* XXX: Don't preclude handling different sized sigset_t's. */
789 if (sigsetsize != sizeof(sigset_t))
791 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
794 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
795 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
798 ret = do_sys_poll(ufds, nfds, &timeout);
800 /* We can restart this syscall, usually */
803 * Don't restore the signal mask yet. Let do_signal() deliver
804 * the signal on the way back to userspace, before the signal
808 memcpy(¤t->saved_sigmask, &sigsaved,
810 set_restore_sigmask();
812 ret = -ERESTARTNOHAND;
814 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
816 if (tsp && timeout >= 0) {
819 if (current->personality & STICKY_TIMEOUTS)
821 /* Yes, we know it's actually an s64, but it's also positive. */
822 rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
824 rts.tv_sec = timeout;
825 if (timespec_compare(&rts, &ts) >= 0)
827 if (copy_to_user(tsp, &rts, sizeof(rts))) {
830 * If an application puts its timeval in read-only
831 * memory, we don't want the Linux-specific update to
832 * the timeval to cause a fault after the select has
833 * completed successfully. However, because we're not
834 * updating the timeval, we can't restart the system
837 if (ret == -ERESTARTNOHAND && timeout >= 0)
844 #endif /* HAVE_SET_RESTORE_SIGMASK */