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);
114 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);
146 EXPORT_SYMBOL(poll_freewait);
148 static struct poll_table_entry *poll_get_entry(poll_table *_p)
150 struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt);
151 struct poll_table_page *table = p->table;
153 if (p->inline_index < N_INLINE_POLL_ENTRIES)
154 return p->inline_entries + p->inline_index++;
156 if (!table || POLL_TABLE_FULL(table)) {
157 struct poll_table_page *new_table;
159 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
162 __set_current_state(TASK_RUNNING);
165 new_table->entry = new_table->entries;
166 new_table->next = table;
167 p->table = new_table;
171 return table->entry++;
174 /* Add a new entry */
175 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
178 struct poll_table_entry *entry = poll_get_entry(p);
183 entry->wait_address = wait_address;
184 init_waitqueue_entry(&entry->wait, current);
185 add_wait_queue(wait_address, &entry->wait);
189 * poll_select_set_timeout - helper function to setup the timeout value
190 * @to: pointer to timespec variable for the final timeout
191 * @sec: seconds (from user space)
192 * @nsec: nanoseconds (from user space)
194 * Note, we do not use a timespec for the user space value here, That
195 * way we can use the function for timeval and compat interfaces as well.
197 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
199 int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
201 struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
203 if (!timespec_valid(&ts))
206 /* Optimize for the zero timeout value here */
208 to->tv_sec = to->tv_nsec = 0;
211 *to = timespec_add_safe(*to, ts);
216 static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
217 int timeval, int ret)
225 if (current->personality & STICKY_TIMEOUTS)
228 /* No update for zero timeout */
229 if (!end_time->tv_sec && !end_time->tv_nsec)
233 rts = timespec_sub(*end_time, rts);
235 rts.tv_sec = rts.tv_nsec = 0;
238 rtv.tv_sec = rts.tv_sec;
239 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
241 if (!copy_to_user(p, &rtv, sizeof(rtv)))
244 } else if (!copy_to_user(p, &rts, sizeof(rts)))
248 * If an application puts its timeval in read-only memory, we
249 * don't want the Linux-specific update to the timeval to
250 * cause a fault after the select has completed
251 * successfully. However, because we're not updating the
252 * timeval, we can't restart the system call.
256 if (ret == -ERESTARTNOHAND)
261 #define FDS_IN(fds, n) (fds->in + n)
262 #define FDS_OUT(fds, n) (fds->out + n)
263 #define FDS_EX(fds, n) (fds->ex + n)
265 #define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
267 static int max_select_fd(unsigned long n, fd_set_bits *fds)
269 unsigned long *open_fds;
274 /* handle last in-complete long-word first */
275 set = ~(~0UL << (n & (__NFDBITS-1)));
277 fdt = files_fdtable(current->files);
278 open_fds = fdt->open_fds->fds_bits+n;
283 if (!(set & ~*open_fds))
294 if (set & ~*open_fds)
303 max += n * __NFDBITS;
309 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
310 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
311 #define POLLEX_SET (POLLPRI)
313 int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
315 ktime_t expire, *to = NULL;
316 struct poll_wqueues table;
318 int retval, i, timed_out = 0;
319 unsigned long slack = 0;
322 retval = max_select_fd(n, fds);
329 poll_initwait(&table);
331 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
336 if (end_time && !timed_out)
337 slack = estimate_accuracy(end_time);
341 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
343 set_current_state(TASK_INTERRUPTIBLE);
345 inp = fds->in; outp = fds->out; exp = fds->ex;
346 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
348 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
349 unsigned long in, out, ex, all_bits, bit = 1, mask, j;
350 unsigned long res_in = 0, res_out = 0, res_ex = 0;
351 const struct file_operations *f_op = NULL;
352 struct file *file = NULL;
354 in = *inp++; out = *outp++; ex = *exp++;
355 all_bits = in | out | ex;
361 for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
365 if (!(bit & all_bits))
367 file = fget_light(i, &fput_needed);
370 mask = DEFAULT_POLLMASK;
371 if (f_op && f_op->poll)
372 mask = (*f_op->poll)(file, retval ? NULL : wait);
373 fput_light(file, fput_needed);
374 if ((mask & POLLIN_SET) && (in & bit)) {
378 if ((mask & POLLOUT_SET) && (out & bit)) {
382 if ((mask & POLLEX_SET) && (ex & bit)) {
397 if (retval || timed_out || signal_pending(current))
400 retval = table.error;
405 * If this is the first loop and we have a timeout
406 * given, then we convert to ktime_t and set the to
407 * pointer to the expiry value.
409 if (end_time && !to) {
410 expire = timespec_to_ktime(*end_time);
414 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
417 __set_current_state(TASK_RUNNING);
419 poll_freewait(&table);
425 * We can actually return ERESTARTSYS instead of EINTR, but I'd
426 * like to be certain this leads to no problems. So I return
427 * EINTR just for safety.
429 * Update: ERESTARTSYS breaks at least the xview clock binary, so
430 * I'm trying ERESTARTNOHAND which restart only when you want to.
432 #define MAX_SELECT_SECONDS \
433 ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
435 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
436 fd_set __user *exp, struct timespec *end_time)
443 /* Allocate small arguments on the stack to save memory and be faster */
444 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
450 /* max_fds can increase, so grab it once to avoid race */
452 fdt = files_fdtable(current->files);
453 max_fds = fdt->max_fds;
459 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
460 * since we used fdset we need to allocate memory in units of
465 if (size > sizeof(stack_fds) / 6) {
466 /* Not enough space in on-stack array; must use kmalloc */
468 bits = kmalloc(6 * size, GFP_KERNEL);
473 fds.out = bits + size;
474 fds.ex = bits + 2*size;
475 fds.res_in = bits + 3*size;
476 fds.res_out = bits + 4*size;
477 fds.res_ex = bits + 5*size;
479 if ((ret = get_fd_set(n, inp, fds.in)) ||
480 (ret = get_fd_set(n, outp, fds.out)) ||
481 (ret = get_fd_set(n, exp, fds.ex)))
483 zero_fd_set(n, fds.res_in);
484 zero_fd_set(n, fds.res_out);
485 zero_fd_set(n, fds.res_ex);
487 ret = do_select(n, &fds, end_time);
492 ret = -ERESTARTNOHAND;
493 if (signal_pending(current))
498 if (set_fd_set(n, inp, fds.res_in) ||
499 set_fd_set(n, outp, fds.res_out) ||
500 set_fd_set(n, exp, fds.res_ex))
504 if (bits != stack_fds)
510 asmlinkage long sys_select(int n, fd_set __user *inp, fd_set __user *outp,
511 fd_set __user *exp, struct timeval __user *tvp)
513 struct timespec end_time, *to = NULL;
518 if (copy_from_user(&tv, tvp, sizeof(tv)))
522 if (poll_select_set_timeout(to,
523 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
524 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
528 ret = core_sys_select(n, inp, outp, exp, to);
529 ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
534 #ifdef HAVE_SET_RESTORE_SIGMASK
535 asmlinkage long sys_pselect7(int n, fd_set __user *inp, fd_set __user *outp,
536 fd_set __user *exp, struct timespec __user *tsp,
537 const sigset_t __user *sigmask, size_t sigsetsize)
539 sigset_t ksigmask, sigsaved;
540 struct timespec ts, end_time, *to = NULL;
544 if (copy_from_user(&ts, tsp, sizeof(ts)))
548 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
553 /* XXX: Don't preclude handling different sized sigset_t's. */
554 if (sigsetsize != sizeof(sigset_t))
556 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
559 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
560 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
563 ret = core_sys_select(n, inp, outp, exp, &end_time);
564 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
566 if (ret == -ERESTARTNOHAND) {
568 * Don't restore the signal mask yet. Let do_signal() deliver
569 * the signal on the way back to userspace, before the signal
573 memcpy(¤t->saved_sigmask, &sigsaved,
575 set_restore_sigmask();
578 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
584 * Most architectures can't handle 7-argument syscalls. So we provide a
585 * 6-argument version where the sixth argument is a pointer to a structure
586 * which has a pointer to the sigset_t itself followed by a size_t containing
589 asmlinkage long sys_pselect6(int n, fd_set __user *inp, fd_set __user *outp,
590 fd_set __user *exp, struct timespec __user *tsp, void __user *sig)
592 size_t sigsetsize = 0;
593 sigset_t __user *up = NULL;
596 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
597 || __get_user(up, (sigset_t __user * __user *)sig)
598 || __get_user(sigsetsize,
599 (size_t __user *)(sig+sizeof(void *))))
603 return sys_pselect7(n, inp, outp, exp, tsp, up, sigsetsize);
605 #endif /* HAVE_SET_RESTORE_SIGMASK */
608 struct poll_list *next;
610 struct pollfd entries[0];
613 #define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
616 * Fish for pollable events on the pollfd->fd file descriptor. We're only
617 * interested in events matching the pollfd->events mask, and the result
618 * matching that mask is both recorded in pollfd->revents and returned. The
619 * pwait poll_table will be used by the fd-provided poll handler for waiting,
622 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
633 file = fget_light(fd, &fput_needed);
636 mask = DEFAULT_POLLMASK;
637 if (file->f_op && file->f_op->poll)
638 mask = file->f_op->poll(file, pwait);
639 /* Mask out unneeded events. */
640 mask &= pollfd->events | POLLERR | POLLHUP;
641 fput_light(file, fput_needed);
644 pollfd->revents = mask;
649 static int do_poll(unsigned int nfds, struct poll_list *list,
650 struct poll_wqueues *wait, struct timespec *end_time)
652 poll_table* pt = &wait->pt;
653 ktime_t expire, *to = NULL;
654 int timed_out = 0, count = 0;
655 unsigned long slack = 0;
657 /* Optimise the no-wait case */
658 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
663 if (end_time && !timed_out)
664 slack = estimate_accuracy(end_time);
667 struct poll_list *walk;
669 set_current_state(TASK_INTERRUPTIBLE);
670 for (walk = list; walk != NULL; walk = walk->next) {
671 struct pollfd * pfd, * pfd_end;
674 pfd_end = pfd + walk->len;
675 for (; pfd != pfd_end; pfd++) {
677 * Fish for events. If we found one, record it
678 * and kill the poll_table, so we don't
679 * needlessly register any other waiters after
680 * this. They'll get immediately deregistered
681 * when we break out and return.
683 if (do_pollfd(pfd, pt)) {
690 * All waiters have already been registered, so don't provide
691 * a poll_table to them on the next loop iteration.
696 if (signal_pending(current))
699 if (count || timed_out)
703 * If this is the first loop and we have a timeout
704 * given, then we convert to ktime_t and set the to
705 * pointer to the expiry value.
707 if (end_time && !to) {
708 expire = timespec_to_ktime(*end_time);
712 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
715 __set_current_state(TASK_RUNNING);
719 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \
720 sizeof(struct pollfd))
722 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
723 struct timespec *end_time)
725 struct poll_wqueues table;
726 int err = -EFAULT, fdcount, len, size;
727 /* Allocate small arguments on the stack to save memory and be
728 faster - use long to make sure the buffer is aligned properly
729 on 64 bit archs to avoid unaligned access */
730 long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
731 struct poll_list *const head = (struct poll_list *)stack_pps;
732 struct poll_list *walk = head;
733 unsigned long todo = nfds;
735 if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
738 len = min_t(unsigned int, nfds, N_STACK_PPS);
745 if (copy_from_user(walk->entries, ufds + nfds-todo,
746 sizeof(struct pollfd) * walk->len))
753 len = min(todo, POLLFD_PER_PAGE);
754 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
755 walk = walk->next = kmalloc(size, GFP_KERNEL);
762 poll_initwait(&table);
763 fdcount = do_poll(nfds, head, &table, end_time);
764 poll_freewait(&table);
766 for (walk = head; walk; walk = walk->next) {
767 struct pollfd *fds = walk->entries;
770 for (j = 0; j < walk->len; j++, ufds++)
771 if (__put_user(fds[j].revents, &ufds->revents))
779 struct poll_list *pos = walk;
787 static long do_restart_poll(struct restart_block *restart_block)
789 struct pollfd __user *ufds = restart_block->poll.ufds;
790 int nfds = restart_block->poll.nfds;
791 struct timespec *to = NULL, end_time;
794 if (restart_block->poll.has_timeout) {
795 end_time.tv_sec = restart_block->poll.tv_sec;
796 end_time.tv_nsec = restart_block->poll.tv_nsec;
800 ret = do_sys_poll(ufds, nfds, to);
803 restart_block->fn = do_restart_poll;
804 ret = -ERESTART_RESTARTBLOCK;
809 asmlinkage long sys_poll(struct pollfd __user *ufds, unsigned int nfds,
812 struct timespec end_time, *to = NULL;
815 if (timeout_msecs >= 0) {
817 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
818 NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
821 ret = do_sys_poll(ufds, nfds, to);
824 struct restart_block *restart_block;
826 restart_block = ¤t_thread_info()->restart_block;
827 restart_block->fn = do_restart_poll;
828 restart_block->poll.ufds = ufds;
829 restart_block->poll.nfds = nfds;
831 if (timeout_msecs >= 0) {
832 restart_block->poll.tv_sec = end_time.tv_sec;
833 restart_block->poll.tv_nsec = end_time.tv_nsec;
834 restart_block->poll.has_timeout = 1;
836 restart_block->poll.has_timeout = 0;
838 ret = -ERESTART_RESTARTBLOCK;
843 #ifdef HAVE_SET_RESTORE_SIGMASK
844 asmlinkage long sys_ppoll(struct pollfd __user *ufds, unsigned int nfds,
845 struct timespec __user *tsp, const sigset_t __user *sigmask,
848 sigset_t ksigmask, sigsaved;
849 struct timespec ts, end_time, *to = NULL;
853 if (copy_from_user(&ts, tsp, sizeof(ts)))
857 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
862 /* XXX: Don't preclude handling different sized sigset_t's. */
863 if (sigsetsize != sizeof(sigset_t))
865 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
868 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
869 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
872 ret = do_sys_poll(ufds, nfds, to);
874 /* We can restart this syscall, usually */
877 * Don't restore the signal mask yet. Let do_signal() deliver
878 * the signal on the way back to userspace, before the signal
882 memcpy(¤t->saved_sigmask, &sigsaved,
884 set_restore_sigmask();
886 ret = -ERESTARTNOHAND;
888 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
890 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
894 #endif /* HAVE_SET_RESTORE_SIGMASK */