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 unsigned 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;
60 static unsigned long estimate_accuracy(struct timespec *tv)
66 * Realtime tasks get a slack of 0 for obvious reasons.
73 now = timespec_sub(*tv, now);
74 ret = __estimate_accuracy(&now);
75 if (ret < current->timer_slack_ns)
76 return current->timer_slack_ns;
82 struct poll_table_page {
83 struct poll_table_page * next;
84 struct poll_table_entry * entry;
85 struct poll_table_entry entries[0];
88 #define POLL_TABLE_FULL(table) \
89 ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
92 * Ok, Peter made a complicated, but straightforward multiple_wait() function.
93 * I have rewritten this, taking some shortcuts: This code may not be easy to
94 * follow, but it should be free of race-conditions, and it's practical. If you
95 * understand what I'm doing here, then you understand how the linux
96 * sleep/wakeup mechanism works.
98 * Two very simple procedures, poll_wait() and poll_freewait() make all the
99 * work. poll_wait() is an inline-function defined in <linux/poll.h>,
100 * as all select/poll functions have to call it to add an entry to the
103 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
106 void poll_initwait(struct poll_wqueues *pwq)
108 init_poll_funcptr(&pwq->pt, __pollwait);
111 pwq->inline_index = 0;
114 EXPORT_SYMBOL(poll_initwait);
116 static void free_poll_entry(struct poll_table_entry *entry)
118 remove_wait_queue(entry->wait_address, &entry->wait);
122 void poll_freewait(struct poll_wqueues *pwq)
124 struct poll_table_page * p = pwq->table;
126 for (i = 0; i < pwq->inline_index; i++)
127 free_poll_entry(pwq->inline_entries + i);
129 struct poll_table_entry * entry;
130 struct poll_table_page *old;
135 free_poll_entry(entry);
136 } while (entry > p->entries);
139 free_page((unsigned long) old);
143 EXPORT_SYMBOL(poll_freewait);
145 static struct poll_table_entry *poll_get_entry(poll_table *_p)
147 struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt);
148 struct poll_table_page *table = p->table;
150 if (p->inline_index < N_INLINE_POLL_ENTRIES)
151 return p->inline_entries + p->inline_index++;
153 if (!table || POLL_TABLE_FULL(table)) {
154 struct poll_table_page *new_table;
156 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
159 __set_current_state(TASK_RUNNING);
162 new_table->entry = new_table->entries;
163 new_table->next = table;
164 p->table = new_table;
168 return table->entry++;
171 /* Add a new entry */
172 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
175 struct poll_table_entry *entry = poll_get_entry(p);
180 entry->wait_address = wait_address;
181 init_waitqueue_entry(&entry->wait, current);
182 add_wait_queue(wait_address, &entry->wait);
186 * poll_select_set_timeout - helper function to setup the timeout value
187 * @to: pointer to timespec variable for the final timeout
188 * @sec: seconds (from user space)
189 * @nsec: nanoseconds (from user space)
191 * Note, we do not use a timespec for the user space value here, That
192 * way we can use the function for timeval and compat interfaces as well.
194 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
196 int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
198 struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
200 if (!timespec_valid(&ts))
203 /* Optimize for the zero timeout value here */
205 to->tv_sec = to->tv_nsec = 0;
208 *to = timespec_add_safe(*to, ts);
213 static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
214 int timeval, int ret)
222 if (current->personality & STICKY_TIMEOUTS)
225 /* No update for zero timeout */
226 if (!end_time->tv_sec && !end_time->tv_nsec)
230 rts = timespec_sub(*end_time, rts);
232 rts.tv_sec = rts.tv_nsec = 0;
235 rtv.tv_sec = rts.tv_sec;
236 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
238 if (!copy_to_user(p, &rtv, sizeof(rtv)))
241 } else if (!copy_to_user(p, &rts, sizeof(rts)))
245 * If an application puts its timeval in read-only memory, we
246 * don't want the Linux-specific update to the timeval to
247 * cause a fault after the select has completed
248 * successfully. However, because we're not updating the
249 * timeval, we can't restart the system call.
253 if (ret == -ERESTARTNOHAND)
258 #define FDS_IN(fds, n) (fds->in + n)
259 #define FDS_OUT(fds, n) (fds->out + n)
260 #define FDS_EX(fds, n) (fds->ex + n)
262 #define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
264 static int max_select_fd(unsigned long n, fd_set_bits *fds)
266 unsigned long *open_fds;
271 /* handle last in-complete long-word first */
272 set = ~(~0UL << (n & (__NFDBITS-1)));
274 fdt = files_fdtable(current->files);
275 open_fds = fdt->open_fds->fds_bits+n;
280 if (!(set & ~*open_fds))
291 if (set & ~*open_fds)
300 max += n * __NFDBITS;
306 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
307 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
308 #define POLLEX_SET (POLLPRI)
310 int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
312 ktime_t expire, *to = NULL;
313 struct poll_wqueues table;
315 int retval, i, timed_out = 0;
316 unsigned long slack = 0;
319 retval = max_select_fd(n, fds);
326 poll_initwait(&table);
328 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
334 slack = estimate_accuracy(end_time);
338 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
340 set_current_state(TASK_INTERRUPTIBLE);
342 inp = fds->in; outp = fds->out; exp = fds->ex;
343 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
345 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
346 unsigned long in, out, ex, all_bits, bit = 1, mask, j;
347 unsigned long res_in = 0, res_out = 0, res_ex = 0;
348 const struct file_operations *f_op = NULL;
349 struct file *file = NULL;
351 in = *inp++; out = *outp++; ex = *exp++;
352 all_bits = in | out | ex;
358 for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
362 if (!(bit & all_bits))
364 file = fget_light(i, &fput_needed);
367 mask = DEFAULT_POLLMASK;
368 if (f_op && f_op->poll)
369 mask = (*f_op->poll)(file, retval ? NULL : wait);
370 fput_light(file, fput_needed);
371 if ((mask & POLLIN_SET) && (in & bit)) {
375 if ((mask & POLLOUT_SET) && (out & bit)) {
379 if ((mask & POLLEX_SET) && (ex & bit)) {
394 if (retval || timed_out || signal_pending(current))
397 retval = table.error;
402 * If this is the first loop and we have a timeout
403 * given, then we convert to ktime_t and set the to
404 * pointer to the expiry value.
406 if (end_time && !to) {
407 expire = timespec_to_ktime(*end_time);
411 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
414 __set_current_state(TASK_RUNNING);
416 poll_freewait(&table);
422 * We can actually return ERESTARTSYS instead of EINTR, but I'd
423 * like to be certain this leads to no problems. So I return
424 * EINTR just for safety.
426 * Update: ERESTARTSYS breaks at least the xview clock binary, so
427 * I'm trying ERESTARTNOHAND which restart only when you want to.
429 #define MAX_SELECT_SECONDS \
430 ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
432 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
433 fd_set __user *exp, struct timespec *end_time)
440 /* Allocate small arguments on the stack to save memory and be faster */
441 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
447 /* max_fds can increase, so grab it once to avoid race */
449 fdt = files_fdtable(current->files);
450 max_fds = fdt->max_fds;
456 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
457 * since we used fdset we need to allocate memory in units of
462 if (size > sizeof(stack_fds) / 6) {
463 /* Not enough space in on-stack array; must use kmalloc */
465 bits = kmalloc(6 * size, GFP_KERNEL);
470 fds.out = bits + size;
471 fds.ex = bits + 2*size;
472 fds.res_in = bits + 3*size;
473 fds.res_out = bits + 4*size;
474 fds.res_ex = bits + 5*size;
476 if ((ret = get_fd_set(n, inp, fds.in)) ||
477 (ret = get_fd_set(n, outp, fds.out)) ||
478 (ret = get_fd_set(n, exp, fds.ex)))
480 zero_fd_set(n, fds.res_in);
481 zero_fd_set(n, fds.res_out);
482 zero_fd_set(n, fds.res_ex);
484 ret = do_select(n, &fds, end_time);
489 ret = -ERESTARTNOHAND;
490 if (signal_pending(current))
495 if (set_fd_set(n, inp, fds.res_in) ||
496 set_fd_set(n, outp, fds.res_out) ||
497 set_fd_set(n, exp, fds.res_ex))
501 if (bits != stack_fds)
507 asmlinkage long sys_select(int n, fd_set __user *inp, fd_set __user *outp,
508 fd_set __user *exp, struct timeval __user *tvp)
510 struct timespec end_time, *to = NULL;
515 if (copy_from_user(&tv, tvp, sizeof(tv)))
519 if (poll_select_set_timeout(to, tv.tv_sec,
520 tv.tv_usec * NSEC_PER_USEC))
524 ret = core_sys_select(n, inp, outp, exp, to);
525 ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
530 #ifdef HAVE_SET_RESTORE_SIGMASK
531 asmlinkage long sys_pselect7(int n, fd_set __user *inp, fd_set __user *outp,
532 fd_set __user *exp, struct timespec __user *tsp,
533 const sigset_t __user *sigmask, size_t sigsetsize)
535 sigset_t ksigmask, sigsaved;
536 struct timespec ts, end_time, *to = NULL;
540 if (copy_from_user(&ts, tsp, sizeof(ts)))
544 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
549 /* XXX: Don't preclude handling different sized sigset_t's. */
550 if (sigsetsize != sizeof(sigset_t))
552 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
555 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
556 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
559 ret = core_sys_select(n, inp, outp, exp, &end_time);
560 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
562 if (ret == -ERESTARTNOHAND) {
564 * Don't restore the signal mask yet. Let do_signal() deliver
565 * the signal on the way back to userspace, before the signal
569 memcpy(¤t->saved_sigmask, &sigsaved,
571 set_restore_sigmask();
574 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
580 * Most architectures can't handle 7-argument syscalls. So we provide a
581 * 6-argument version where the sixth argument is a pointer to a structure
582 * which has a pointer to the sigset_t itself followed by a size_t containing
585 asmlinkage long sys_pselect6(int n, fd_set __user *inp, fd_set __user *outp,
586 fd_set __user *exp, struct timespec __user *tsp, void __user *sig)
588 size_t sigsetsize = 0;
589 sigset_t __user *up = NULL;
592 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
593 || __get_user(up, (sigset_t __user * __user *)sig)
594 || __get_user(sigsetsize,
595 (size_t __user *)(sig+sizeof(void *))))
599 return sys_pselect7(n, inp, outp, exp, tsp, up, sigsetsize);
601 #endif /* HAVE_SET_RESTORE_SIGMASK */
604 struct poll_list *next;
606 struct pollfd entries[0];
609 #define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
612 * Fish for pollable events on the pollfd->fd file descriptor. We're only
613 * interested in events matching the pollfd->events mask, and the result
614 * matching that mask is both recorded in pollfd->revents and returned. The
615 * pwait poll_table will be used by the fd-provided poll handler for waiting,
618 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
629 file = fget_light(fd, &fput_needed);
632 mask = DEFAULT_POLLMASK;
633 if (file->f_op && file->f_op->poll)
634 mask = file->f_op->poll(file, pwait);
635 /* Mask out unneeded events. */
636 mask &= pollfd->events | POLLERR | POLLHUP;
637 fput_light(file, fput_needed);
640 pollfd->revents = mask;
645 static int do_poll(unsigned int nfds, struct poll_list *list,
646 struct poll_wqueues *wait, struct timespec *end_time)
648 poll_table* pt = &wait->pt;
649 ktime_t expire, *to = NULL;
650 int timed_out = 0, count = 0;
651 unsigned long slack = 0;
653 /* Optimise the no-wait case */
654 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
660 slack = estimate_accuracy(end_time);
663 struct poll_list *walk;
665 set_current_state(TASK_INTERRUPTIBLE);
666 for (walk = list; walk != NULL; walk = walk->next) {
667 struct pollfd * pfd, * pfd_end;
670 pfd_end = pfd + walk->len;
671 for (; pfd != pfd_end; pfd++) {
673 * Fish for events. If we found one, record it
674 * and kill the poll_table, so we don't
675 * needlessly register any other waiters after
676 * this. They'll get immediately deregistered
677 * when we break out and return.
679 if (do_pollfd(pfd, pt)) {
686 * All waiters have already been registered, so don't provide
687 * a poll_table to them on the next loop iteration.
692 if (signal_pending(current))
695 if (count || timed_out)
699 * If this is the first loop and we have a timeout
700 * given, then we convert to ktime_t and set the to
701 * pointer to the expiry value.
703 if (end_time && !to) {
704 expire = timespec_to_ktime(*end_time);
708 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
711 __set_current_state(TASK_RUNNING);
715 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \
716 sizeof(struct pollfd))
718 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
719 struct timespec *end_time)
721 struct poll_wqueues table;
722 int err = -EFAULT, fdcount, len, size;
723 /* Allocate small arguments on the stack to save memory and be
724 faster - use long to make sure the buffer is aligned properly
725 on 64 bit archs to avoid unaligned access */
726 long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
727 struct poll_list *const head = (struct poll_list *)stack_pps;
728 struct poll_list *walk = head;
729 unsigned long todo = nfds;
731 if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
734 len = min_t(unsigned int, nfds, N_STACK_PPS);
741 if (copy_from_user(walk->entries, ufds + nfds-todo,
742 sizeof(struct pollfd) * walk->len))
749 len = min(todo, POLLFD_PER_PAGE);
750 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
751 walk = walk->next = kmalloc(size, GFP_KERNEL);
758 poll_initwait(&table);
759 fdcount = do_poll(nfds, head, &table, end_time);
760 poll_freewait(&table);
762 for (walk = head; walk; walk = walk->next) {
763 struct pollfd *fds = walk->entries;
766 for (j = 0; j < walk->len; j++, ufds++)
767 if (__put_user(fds[j].revents, &ufds->revents))
775 struct poll_list *pos = walk;
783 static long do_restart_poll(struct restart_block *restart_block)
785 struct pollfd __user *ufds = restart_block->poll.ufds;
786 int nfds = restart_block->poll.nfds;
787 struct timespec *to = NULL, end_time;
790 if (restart_block->poll.has_timeout) {
791 end_time.tv_sec = restart_block->poll.tv_sec;
792 end_time.tv_nsec = restart_block->poll.tv_nsec;
796 ret = do_sys_poll(ufds, nfds, to);
799 restart_block->fn = do_restart_poll;
800 ret = -ERESTART_RESTARTBLOCK;
805 asmlinkage long sys_poll(struct pollfd __user *ufds, unsigned int nfds,
808 struct timespec end_time, *to = NULL;
811 if (timeout_msecs >= 0) {
813 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
814 NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
817 ret = do_sys_poll(ufds, nfds, to);
820 struct restart_block *restart_block;
822 restart_block = ¤t_thread_info()->restart_block;
823 restart_block->fn = do_restart_poll;
824 restart_block->poll.ufds = ufds;
825 restart_block->poll.nfds = nfds;
827 if (timeout_msecs >= 0) {
828 restart_block->poll.tv_sec = end_time.tv_sec;
829 restart_block->poll.tv_nsec = end_time.tv_nsec;
830 restart_block->poll.has_timeout = 1;
832 restart_block->poll.has_timeout = 0;
834 ret = -ERESTART_RESTARTBLOCK;
839 #ifdef HAVE_SET_RESTORE_SIGMASK
840 asmlinkage long sys_ppoll(struct pollfd __user *ufds, unsigned int nfds,
841 struct timespec __user *tsp, const sigset_t __user *sigmask,
844 sigset_t ksigmask, sigsaved;
845 struct timespec ts, end_time, *to = NULL;
849 if (copy_from_user(&ts, tsp, sizeof(ts)))
853 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
858 /* XXX: Don't preclude handling different sized sigset_t's. */
859 if (sigsetsize != sizeof(sigset_t))
861 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
864 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
865 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
868 ret = do_sys_poll(ufds, nfds, to);
870 /* We can restart this syscall, usually */
873 * Don't restore the signal mask yet. Let do_signal() deliver
874 * the signal on the way back to userspace, before the signal
878 memcpy(¤t->saved_sigmask, &sigsaved,
880 set_restore_sigmask();
882 ret = -ERESTARTNOHAND;
884 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
886 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
890 #endif /* HAVE_SET_RESTORE_SIGMASK */