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>
27 #include <linux/hrtimer.h>
29 #include <asm/uaccess.h>
33 * Estimate expected accuracy in ns from a timeval.
35 * After quite a bit of churning around, we've settled on
36 * a simple thing of taking 0.1% of the timeout as the
37 * slack, with a cap of 100 msec.
38 * "nice" tasks get a 0.5% slack instead.
40 * Consider this comment an open invitation to come up with even
44 static long __estimate_accuracy(struct timespec *tv)
49 if (task_nice(current) > 0)
50 divfactor = divfactor / 5;
52 slack = tv->tv_nsec / divfactor;
53 slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
55 if (slack > 100 * NSEC_PER_MSEC)
56 slack = 100 * NSEC_PER_MSEC;
63 static long estimate_accuracy(struct timespec *tv)
69 * Realtime tasks get a slack of 0 for obvious reasons.
76 now = timespec_sub(*tv, now);
77 ret = __estimate_accuracy(&now);
78 if (ret < current->timer_slack_ns)
79 return current->timer_slack_ns;
85 struct poll_table_page {
86 struct poll_table_page * next;
87 struct poll_table_entry * entry;
88 struct poll_table_entry entries[0];
91 #define POLL_TABLE_FULL(table) \
92 ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
95 * Ok, Peter made a complicated, but straightforward multiple_wait() function.
96 * I have rewritten this, taking some shortcuts: This code may not be easy to
97 * follow, but it should be free of race-conditions, and it's practical. If you
98 * understand what I'm doing here, then you understand how the linux
99 * sleep/wakeup mechanism works.
101 * Two very simple procedures, poll_wait() and poll_freewait() make all the
102 * work. poll_wait() is an inline-function defined in <linux/poll.h>,
103 * as all select/poll functions have to call it to add an entry to the
106 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
109 void poll_initwait(struct poll_wqueues *pwq)
111 init_poll_funcptr(&pwq->pt, __pollwait);
112 pwq->polling_task = current;
115 pwq->inline_index = 0;
117 EXPORT_SYMBOL(poll_initwait);
119 static void free_poll_entry(struct poll_table_entry *entry)
121 remove_wait_queue(entry->wait_address, &entry->wait);
125 void poll_freewait(struct poll_wqueues *pwq)
127 struct poll_table_page * p = pwq->table;
129 for (i = 0; i < pwq->inline_index; i++)
130 free_poll_entry(pwq->inline_entries + i);
132 struct poll_table_entry * entry;
133 struct poll_table_page *old;
138 free_poll_entry(entry);
139 } while (entry > p->entries);
142 free_page((unsigned long) old);
145 EXPORT_SYMBOL(poll_freewait);
147 static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
149 struct poll_table_page *table = p->table;
151 if (p->inline_index < N_INLINE_POLL_ENTRIES)
152 return p->inline_entries + p->inline_index++;
154 if (!table || POLL_TABLE_FULL(table)) {
155 struct poll_table_page *new_table;
157 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
162 new_table->entry = new_table->entries;
163 new_table->next = table;
164 p->table = new_table;
168 return table->entry++;
171 static int pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
173 struct poll_wqueues *pwq = wait->private;
174 DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
177 * Although this function is called under waitqueue lock, LOCK
178 * doesn't imply write barrier and the users expect write
179 * barrier semantics on wakeup functions. The following
180 * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
181 * and is paired with set_mb() in poll_schedule_timeout.
187 * Perform the default wake up operation using a dummy
190 * TODO: This is hacky but there currently is no interface to
191 * pass in @sync. @sync is scheduled to be removed and once
192 * that happens, wake_up_process() can be used directly.
194 return default_wake_function(&dummy_wait, mode, sync, key);
197 /* Add a new entry */
198 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
201 struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
202 struct poll_table_entry *entry = poll_get_entry(pwq);
207 entry->wait_address = wait_address;
208 init_waitqueue_func_entry(&entry->wait, pollwake);
209 entry->wait.private = pwq;
210 add_wait_queue(wait_address, &entry->wait);
213 int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
214 ktime_t *expires, unsigned long slack)
218 set_current_state(state);
220 rc = schedule_hrtimeout_range(expires, slack, HRTIMER_MODE_ABS);
221 __set_current_state(TASK_RUNNING);
224 * Prepare for the next iteration.
226 * The following set_mb() serves two purposes. First, it's
227 * the counterpart rmb of the wmb in pollwake() such that data
228 * written before wake up is always visible after wake up.
229 * Second, the full barrier guarantees that triggered clearing
230 * doesn't pass event check of the next iteration. Note that
231 * this problem doesn't exist for the first iteration as
232 * add_wait_queue() has full barrier semantics.
234 set_mb(pwq->triggered, 0);
238 EXPORT_SYMBOL(poll_schedule_timeout);
241 * poll_select_set_timeout - helper function to setup the timeout value
242 * @to: pointer to timespec variable for the final timeout
243 * @sec: seconds (from user space)
244 * @nsec: nanoseconds (from user space)
246 * Note, we do not use a timespec for the user space value here, That
247 * way we can use the function for timeval and compat interfaces as well.
249 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
251 int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
253 struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
255 if (!timespec_valid(&ts))
258 /* Optimize for the zero timeout value here */
260 to->tv_sec = to->tv_nsec = 0;
263 *to = timespec_add_safe(*to, ts);
268 static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
269 int timeval, int ret)
277 if (current->personality & STICKY_TIMEOUTS)
280 /* No update for zero timeout */
281 if (!end_time->tv_sec && !end_time->tv_nsec)
285 rts = timespec_sub(*end_time, rts);
287 rts.tv_sec = rts.tv_nsec = 0;
290 rtv.tv_sec = rts.tv_sec;
291 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
293 if (!copy_to_user(p, &rtv, sizeof(rtv)))
296 } else if (!copy_to_user(p, &rts, sizeof(rts)))
300 * If an application puts its timeval in read-only memory, we
301 * don't want the Linux-specific update to the timeval to
302 * cause a fault after the select has completed
303 * successfully. However, because we're not updating the
304 * timeval, we can't restart the system call.
308 if (ret == -ERESTARTNOHAND)
313 #define FDS_IN(fds, n) (fds->in + n)
314 #define FDS_OUT(fds, n) (fds->out + n)
315 #define FDS_EX(fds, n) (fds->ex + n)
317 #define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
319 static int max_select_fd(unsigned long n, fd_set_bits *fds)
321 unsigned long *open_fds;
326 /* handle last in-complete long-word first */
327 set = ~(~0UL << (n & (__NFDBITS-1)));
329 fdt = files_fdtable(current->files);
330 open_fds = fdt->open_fds->fds_bits+n;
335 if (!(set & ~*open_fds))
346 if (set & ~*open_fds)
355 max += n * __NFDBITS;
361 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
362 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
363 #define POLLEX_SET (POLLPRI)
365 int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
367 ktime_t expire, *to = NULL;
368 struct poll_wqueues table;
370 int retval, i, timed_out = 0;
371 unsigned long slack = 0;
374 retval = max_select_fd(n, fds);
381 poll_initwait(&table);
383 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
388 if (end_time && !timed_out)
389 slack = estimate_accuracy(end_time);
393 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
395 inp = fds->in; outp = fds->out; exp = fds->ex;
396 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
398 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
399 unsigned long in, out, ex, all_bits, bit = 1, mask, j;
400 unsigned long res_in = 0, res_out = 0, res_ex = 0;
401 const struct file_operations *f_op = NULL;
402 struct file *file = NULL;
404 in = *inp++; out = *outp++; ex = *exp++;
405 all_bits = in | out | ex;
411 for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
415 if (!(bit & all_bits))
417 file = fget_light(i, &fput_needed);
420 mask = DEFAULT_POLLMASK;
421 if (f_op && f_op->poll)
422 mask = (*f_op->poll)(file, retval ? NULL : wait);
423 fput_light(file, fput_needed);
424 if ((mask & POLLIN_SET) && (in & bit)) {
428 if ((mask & POLLOUT_SET) && (out & bit)) {
432 if ((mask & POLLEX_SET) && (ex & bit)) {
447 if (retval || timed_out || signal_pending(current))
450 retval = table.error;
455 * If this is the first loop and we have a timeout
456 * given, then we convert to ktime_t and set the to
457 * pointer to the expiry value.
459 if (end_time && !to) {
460 expire = timespec_to_ktime(*end_time);
464 if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
469 poll_freewait(&table);
475 * We can actually return ERESTARTSYS instead of EINTR, but I'd
476 * like to be certain this leads to no problems. So I return
477 * EINTR just for safety.
479 * Update: ERESTARTSYS breaks at least the xview clock binary, so
480 * I'm trying ERESTARTNOHAND which restart only when you want to.
482 #define MAX_SELECT_SECONDS \
483 ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
485 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
486 fd_set __user *exp, struct timespec *end_time)
493 /* Allocate small arguments on the stack to save memory and be faster */
494 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
500 /* max_fds can increase, so grab it once to avoid race */
502 fdt = files_fdtable(current->files);
503 max_fds = fdt->max_fds;
509 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
510 * since we used fdset we need to allocate memory in units of
515 if (size > sizeof(stack_fds) / 6) {
516 /* Not enough space in on-stack array; must use kmalloc */
518 bits = kmalloc(6 * size, GFP_KERNEL);
523 fds.out = bits + size;
524 fds.ex = bits + 2*size;
525 fds.res_in = bits + 3*size;
526 fds.res_out = bits + 4*size;
527 fds.res_ex = bits + 5*size;
529 if ((ret = get_fd_set(n, inp, fds.in)) ||
530 (ret = get_fd_set(n, outp, fds.out)) ||
531 (ret = get_fd_set(n, exp, fds.ex)))
533 zero_fd_set(n, fds.res_in);
534 zero_fd_set(n, fds.res_out);
535 zero_fd_set(n, fds.res_ex);
537 ret = do_select(n, &fds, end_time);
542 ret = -ERESTARTNOHAND;
543 if (signal_pending(current))
548 if (set_fd_set(n, inp, fds.res_in) ||
549 set_fd_set(n, outp, fds.res_out) ||
550 set_fd_set(n, exp, fds.res_ex))
554 if (bits != stack_fds)
560 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
561 fd_set __user *, exp, struct timeval __user *, tvp)
563 struct timespec end_time, *to = NULL;
568 if (copy_from_user(&tv, tvp, sizeof(tv)))
572 if (poll_select_set_timeout(to,
573 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
574 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
578 ret = core_sys_select(n, inp, outp, exp, to);
579 ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
584 #ifdef HAVE_SET_RESTORE_SIGMASK
585 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
586 fd_set __user *exp, struct timespec __user *tsp,
587 const sigset_t __user *sigmask, size_t sigsetsize)
589 sigset_t ksigmask, sigsaved;
590 struct timespec ts, end_time, *to = NULL;
594 if (copy_from_user(&ts, tsp, sizeof(ts)))
598 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
603 /* XXX: Don't preclude handling different sized sigset_t's. */
604 if (sigsetsize != sizeof(sigset_t))
606 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
609 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
610 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
613 ret = core_sys_select(n, inp, outp, exp, to);
614 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
616 if (ret == -ERESTARTNOHAND) {
618 * Don't restore the signal mask yet. Let do_signal() deliver
619 * the signal on the way back to userspace, before the signal
623 memcpy(¤t->saved_sigmask, &sigsaved,
625 set_restore_sigmask();
628 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
634 * Most architectures can't handle 7-argument syscalls. So we provide a
635 * 6-argument version where the sixth argument is a pointer to a structure
636 * which has a pointer to the sigset_t itself followed by a size_t containing
639 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
640 fd_set __user *, exp, struct timespec __user *, tsp,
643 size_t sigsetsize = 0;
644 sigset_t __user *up = NULL;
647 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
648 || __get_user(up, (sigset_t __user * __user *)sig)
649 || __get_user(sigsetsize,
650 (size_t __user *)(sig+sizeof(void *))))
654 return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
656 #endif /* HAVE_SET_RESTORE_SIGMASK */
659 struct poll_list *next;
661 struct pollfd entries[0];
664 #define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
667 * Fish for pollable events on the pollfd->fd file descriptor. We're only
668 * interested in events matching the pollfd->events mask, and the result
669 * matching that mask is both recorded in pollfd->revents and returned. The
670 * pwait poll_table will be used by the fd-provided poll handler for waiting,
673 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
684 file = fget_light(fd, &fput_needed);
687 mask = DEFAULT_POLLMASK;
688 if (file->f_op && file->f_op->poll)
689 mask = file->f_op->poll(file, pwait);
690 /* Mask out unneeded events. */
691 mask &= pollfd->events | POLLERR | POLLHUP;
692 fput_light(file, fput_needed);
695 pollfd->revents = mask;
700 static int do_poll(unsigned int nfds, struct poll_list *list,
701 struct poll_wqueues *wait, struct timespec *end_time)
703 poll_table* pt = &wait->pt;
704 ktime_t expire, *to = NULL;
705 int timed_out = 0, count = 0;
706 unsigned long slack = 0;
708 /* Optimise the no-wait case */
709 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
714 if (end_time && !timed_out)
715 slack = estimate_accuracy(end_time);
718 struct poll_list *walk;
720 for (walk = list; walk != NULL; walk = walk->next) {
721 struct pollfd * pfd, * pfd_end;
724 pfd_end = pfd + walk->len;
725 for (; pfd != pfd_end; pfd++) {
727 * Fish for events. If we found one, record it
728 * and kill the poll_table, so we don't
729 * needlessly register any other waiters after
730 * this. They'll get immediately deregistered
731 * when we break out and return.
733 if (do_pollfd(pfd, pt)) {
740 * All waiters have already been registered, so don't provide
741 * a poll_table to them on the next loop iteration.
746 if (signal_pending(current))
749 if (count || timed_out)
753 * If this is the first loop and we have a timeout
754 * given, then we convert to ktime_t and set the to
755 * pointer to the expiry value.
757 if (end_time && !to) {
758 expire = timespec_to_ktime(*end_time);
762 if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
768 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \
769 sizeof(struct pollfd))
771 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
772 struct timespec *end_time)
774 struct poll_wqueues table;
775 int err = -EFAULT, fdcount, len, size;
776 /* Allocate small arguments on the stack to save memory and be
777 faster - use long to make sure the buffer is aligned properly
778 on 64 bit archs to avoid unaligned access */
779 long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
780 struct poll_list *const head = (struct poll_list *)stack_pps;
781 struct poll_list *walk = head;
782 unsigned long todo = nfds;
784 if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
787 len = min_t(unsigned int, nfds, N_STACK_PPS);
794 if (copy_from_user(walk->entries, ufds + nfds-todo,
795 sizeof(struct pollfd) * walk->len))
802 len = min(todo, POLLFD_PER_PAGE);
803 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
804 walk = walk->next = kmalloc(size, GFP_KERNEL);
811 poll_initwait(&table);
812 fdcount = do_poll(nfds, head, &table, end_time);
813 poll_freewait(&table);
815 for (walk = head; walk; walk = walk->next) {
816 struct pollfd *fds = walk->entries;
819 for (j = 0; j < walk->len; j++, ufds++)
820 if (__put_user(fds[j].revents, &ufds->revents))
828 struct poll_list *pos = walk;
836 static long do_restart_poll(struct restart_block *restart_block)
838 struct pollfd __user *ufds = restart_block->poll.ufds;
839 int nfds = restart_block->poll.nfds;
840 struct timespec *to = NULL, end_time;
843 if (restart_block->poll.has_timeout) {
844 end_time.tv_sec = restart_block->poll.tv_sec;
845 end_time.tv_nsec = restart_block->poll.tv_nsec;
849 ret = do_sys_poll(ufds, nfds, to);
852 restart_block->fn = do_restart_poll;
853 ret = -ERESTART_RESTARTBLOCK;
858 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
861 struct timespec end_time, *to = NULL;
864 if (timeout_msecs >= 0) {
866 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
867 NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
870 ret = do_sys_poll(ufds, nfds, to);
873 struct restart_block *restart_block;
875 restart_block = ¤t_thread_info()->restart_block;
876 restart_block->fn = do_restart_poll;
877 restart_block->poll.ufds = ufds;
878 restart_block->poll.nfds = nfds;
880 if (timeout_msecs >= 0) {
881 restart_block->poll.tv_sec = end_time.tv_sec;
882 restart_block->poll.tv_nsec = end_time.tv_nsec;
883 restart_block->poll.has_timeout = 1;
885 restart_block->poll.has_timeout = 0;
887 ret = -ERESTART_RESTARTBLOCK;
892 #ifdef HAVE_SET_RESTORE_SIGMASK
893 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
894 struct timespec __user *, tsp, const sigset_t __user *, sigmask,
897 sigset_t ksigmask, sigsaved;
898 struct timespec ts, end_time, *to = NULL;
902 if (copy_from_user(&ts, tsp, sizeof(ts)))
906 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
911 /* XXX: Don't preclude handling different sized sigset_t's. */
912 if (sigsetsize != sizeof(sigset_t))
914 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
917 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
918 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
921 ret = do_sys_poll(ufds, nfds, to);
923 /* We can restart this syscall, usually */
926 * Don't restore the signal mask yet. Let do_signal() deliver
927 * the signal on the way back to userspace, before the signal
931 memcpy(¤t->saved_sigmask, &sigsaved,
933 set_restore_sigmask();
935 ret = -ERESTARTNOHAND;
937 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
939 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
943 #endif /* HAVE_SET_RESTORE_SIGMASK */