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/sched.h>
19 #include <linux/syscalls.h>
20 #include <linux/module.h>
21 #include <linux/slab.h>
22 #include <linux/poll.h>
23 #include <linux/personality.h> /* for STICKY_TIMEOUTS */
24 #include <linux/file.h>
25 #include <linux/fdtable.h>
27 #include <linux/rcupdate.h>
28 #include <linux/hrtimer.h>
30 #include <asm/uaccess.h>
34 * Estimate expected accuracy in ns from a timeval.
36 * After quite a bit of churning around, we've settled on
37 * a simple thing of taking 0.1% of the timeout as the
38 * slack, with a cap of 100 msec.
39 * "nice" tasks get a 0.5% slack instead.
41 * Consider this comment an open invitation to come up with even
45 #define MAX_SLACK (100 * NSEC_PER_MSEC)
47 static long __estimate_accuracy(struct timespec *tv)
55 if (task_nice(current) > 0)
56 divfactor = divfactor / 5;
58 if (tv->tv_sec > MAX_SLACK / (NSEC_PER_SEC/divfactor))
61 slack = tv->tv_nsec / divfactor;
62 slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
64 if (slack > MAX_SLACK)
70 long select_estimate_accuracy(struct timespec *tv)
76 * Realtime tasks get a slack of 0 for obvious reasons.
83 now = timespec_sub(*tv, now);
84 ret = __estimate_accuracy(&now);
85 if (ret < current->timer_slack_ns)
86 return current->timer_slack_ns;
92 struct poll_table_page {
93 struct poll_table_page * next;
94 struct poll_table_entry * entry;
95 struct poll_table_entry entries[0];
98 #define POLL_TABLE_FULL(table) \
99 ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
102 * Ok, Peter made a complicated, but straightforward multiple_wait() function.
103 * I have rewritten this, taking some shortcuts: This code may not be easy to
104 * follow, but it should be free of race-conditions, and it's practical. If you
105 * understand what I'm doing here, then you understand how the linux
106 * sleep/wakeup mechanism works.
108 * Two very simple procedures, poll_wait() and poll_freewait() make all the
109 * work. poll_wait() is an inline-function defined in <linux/poll.h>,
110 * as all select/poll functions have to call it to add an entry to the
113 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
116 void poll_initwait(struct poll_wqueues *pwq)
118 init_poll_funcptr(&pwq->pt, __pollwait);
119 pwq->polling_task = current;
123 pwq->inline_index = 0;
125 EXPORT_SYMBOL(poll_initwait);
127 static void free_poll_entry(struct poll_table_entry *entry)
129 remove_wait_queue(entry->wait_address, &entry->wait);
133 void poll_freewait(struct poll_wqueues *pwq)
135 struct poll_table_page * p = pwq->table;
137 for (i = 0; i < pwq->inline_index; i++)
138 free_poll_entry(pwq->inline_entries + i);
140 struct poll_table_entry * entry;
141 struct poll_table_page *old;
146 free_poll_entry(entry);
147 } while (entry > p->entries);
150 free_page((unsigned long) old);
153 EXPORT_SYMBOL(poll_freewait);
155 static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
157 struct poll_table_page *table = p->table;
159 if (p->inline_index < N_INLINE_POLL_ENTRIES)
160 return p->inline_entries + p->inline_index++;
162 if (!table || POLL_TABLE_FULL(table)) {
163 struct poll_table_page *new_table;
165 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
170 new_table->entry = new_table->entries;
171 new_table->next = table;
172 p->table = new_table;
176 return table->entry++;
179 static int __pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
181 struct poll_wqueues *pwq = wait->private;
182 DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
185 * Although this function is called under waitqueue lock, LOCK
186 * doesn't imply write barrier and the users expect write
187 * barrier semantics on wakeup functions. The following
188 * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
189 * and is paired with set_mb() in poll_schedule_timeout.
195 * Perform the default wake up operation using a dummy
198 * TODO: This is hacky but there currently is no interface to
199 * pass in @sync. @sync is scheduled to be removed and once
200 * that happens, wake_up_process() can be used directly.
202 return default_wake_function(&dummy_wait, mode, sync, key);
205 static int pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
207 struct poll_table_entry *entry;
209 entry = container_of(wait, struct poll_table_entry, wait);
210 if (key && !((unsigned long)key & entry->key))
212 return __pollwake(wait, mode, sync, key);
215 /* Add a new entry */
216 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
219 struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
220 struct poll_table_entry *entry = poll_get_entry(pwq);
225 entry->wait_address = wait_address;
227 init_waitqueue_func_entry(&entry->wait, pollwake);
228 entry->wait.private = pwq;
229 add_wait_queue(wait_address, &entry->wait);
232 int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
233 ktime_t *expires, unsigned long slack)
237 set_current_state(state);
239 rc = schedule_hrtimeout_range(expires, slack, HRTIMER_MODE_ABS);
240 __set_current_state(TASK_RUNNING);
243 * Prepare for the next iteration.
245 * The following set_mb() serves two purposes. First, it's
246 * the counterpart rmb of the wmb in pollwake() such that data
247 * written before wake up is always visible after wake up.
248 * Second, the full barrier guarantees that triggered clearing
249 * doesn't pass event check of the next iteration. Note that
250 * this problem doesn't exist for the first iteration as
251 * add_wait_queue() has full barrier semantics.
253 set_mb(pwq->triggered, 0);
257 EXPORT_SYMBOL(poll_schedule_timeout);
260 * poll_select_set_timeout - helper function to setup the timeout value
261 * @to: pointer to timespec variable for the final timeout
262 * @sec: seconds (from user space)
263 * @nsec: nanoseconds (from user space)
265 * Note, we do not use a timespec for the user space value here, That
266 * way we can use the function for timeval and compat interfaces as well.
268 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
270 int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
272 struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
274 if (!timespec_valid(&ts))
277 /* Optimize for the zero timeout value here */
279 to->tv_sec = to->tv_nsec = 0;
282 *to = timespec_add_safe(*to, ts);
287 static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
288 int timeval, int ret)
296 if (current->personality & STICKY_TIMEOUTS)
299 /* No update for zero timeout */
300 if (!end_time->tv_sec && !end_time->tv_nsec)
304 rts = timespec_sub(*end_time, rts);
306 rts.tv_sec = rts.tv_nsec = 0;
309 if (sizeof(rtv) > sizeof(rtv.tv_sec) + sizeof(rtv.tv_usec))
310 memset(&rtv, 0, sizeof(rtv));
311 rtv.tv_sec = rts.tv_sec;
312 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
314 if (!copy_to_user(p, &rtv, sizeof(rtv)))
317 } else if (!copy_to_user(p, &rts, sizeof(rts)))
321 * If an application puts its timeval in read-only memory, we
322 * don't want the Linux-specific update to the timeval to
323 * cause a fault after the select has completed
324 * successfully. However, because we're not updating the
325 * timeval, we can't restart the system call.
329 if (ret == -ERESTARTNOHAND)
334 #define FDS_IN(fds, n) (fds->in + n)
335 #define FDS_OUT(fds, n) (fds->out + n)
336 #define FDS_EX(fds, n) (fds->ex + n)
338 #define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
340 static int max_select_fd(unsigned long n, fd_set_bits *fds)
342 unsigned long *open_fds;
347 /* handle last in-complete long-word first */
348 set = ~(~0UL << (n & (__NFDBITS-1)));
350 fdt = files_fdtable(current->files);
351 open_fds = fdt->open_fds->fds_bits+n;
356 if (!(set & ~*open_fds))
367 if (set & ~*open_fds)
376 max += n * __NFDBITS;
382 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
383 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
384 #define POLLEX_SET (POLLPRI)
386 static inline void wait_key_set(poll_table *wait, unsigned long in,
387 unsigned long out, unsigned long bit)
390 wait->key = POLLEX_SET;
392 wait->key |= POLLIN_SET;
394 wait->key |= POLLOUT_SET;
398 int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
400 ktime_t expire, *to = NULL;
401 struct poll_wqueues table;
403 int retval, i, timed_out = 0;
404 unsigned long slack = 0;
407 retval = max_select_fd(n, fds);
414 poll_initwait(&table);
416 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
421 if (end_time && !timed_out)
422 slack = select_estimate_accuracy(end_time);
426 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
428 inp = fds->in; outp = fds->out; exp = fds->ex;
429 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
431 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
432 unsigned long in, out, ex, all_bits, bit = 1, mask, j;
433 unsigned long res_in = 0, res_out = 0, res_ex = 0;
434 const struct file_operations *f_op = NULL;
435 struct file *file = NULL;
437 in = *inp++; out = *outp++; ex = *exp++;
438 all_bits = in | out | ex;
444 for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
448 if (!(bit & all_bits))
450 file = fget_light(i, &fput_needed);
453 mask = DEFAULT_POLLMASK;
454 if (f_op && f_op->poll) {
455 wait_key_set(wait, in, out, bit);
456 mask = (*f_op->poll)(file, wait);
458 fput_light(file, fput_needed);
459 if ((mask & POLLIN_SET) && (in & bit)) {
464 if ((mask & POLLOUT_SET) && (out & bit)) {
469 if ((mask & POLLEX_SET) && (ex & bit)) {
485 if (retval || timed_out || signal_pending(current))
488 retval = table.error;
493 * If this is the first loop and we have a timeout
494 * given, then we convert to ktime_t and set the to
495 * pointer to the expiry value.
497 if (end_time && !to) {
498 expire = timespec_to_ktime(*end_time);
502 if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
507 poll_freewait(&table);
513 * We can actually return ERESTARTSYS instead of EINTR, but I'd
514 * like to be certain this leads to no problems. So I return
515 * EINTR just for safety.
517 * Update: ERESTARTSYS breaks at least the xview clock binary, so
518 * I'm trying ERESTARTNOHAND which restart only when you want to.
520 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
521 fd_set __user *exp, struct timespec *end_time)
528 /* Allocate small arguments on the stack to save memory and be faster */
529 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
535 /* max_fds can increase, so grab it once to avoid race */
537 fdt = files_fdtable(current->files);
538 max_fds = fdt->max_fds;
544 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
545 * since we used fdset we need to allocate memory in units of
550 if (size > sizeof(stack_fds) / 6) {
551 /* Not enough space in on-stack array; must use kmalloc */
553 bits = kmalloc(6 * size, GFP_KERNEL);
558 fds.out = bits + size;
559 fds.ex = bits + 2*size;
560 fds.res_in = bits + 3*size;
561 fds.res_out = bits + 4*size;
562 fds.res_ex = bits + 5*size;
564 if ((ret = get_fd_set(n, inp, fds.in)) ||
565 (ret = get_fd_set(n, outp, fds.out)) ||
566 (ret = get_fd_set(n, exp, fds.ex)))
568 zero_fd_set(n, fds.res_in);
569 zero_fd_set(n, fds.res_out);
570 zero_fd_set(n, fds.res_ex);
572 ret = do_select(n, &fds, end_time);
577 ret = -ERESTARTNOHAND;
578 if (signal_pending(current))
583 if (set_fd_set(n, inp, fds.res_in) ||
584 set_fd_set(n, outp, fds.res_out) ||
585 set_fd_set(n, exp, fds.res_ex))
589 if (bits != stack_fds)
595 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
596 fd_set __user *, exp, struct timeval __user *, tvp)
598 struct timespec end_time, *to = NULL;
603 if (copy_from_user(&tv, tvp, sizeof(tv)))
607 if (poll_select_set_timeout(to,
608 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
609 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
613 ret = core_sys_select(n, inp, outp, exp, to);
614 ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
619 #ifdef HAVE_SET_RESTORE_SIGMASK
620 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
621 fd_set __user *exp, struct timespec __user *tsp,
622 const sigset_t __user *sigmask, size_t sigsetsize)
624 sigset_t ksigmask, sigsaved;
625 struct timespec ts, end_time, *to = NULL;
629 if (copy_from_user(&ts, tsp, sizeof(ts)))
633 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
638 /* XXX: Don't preclude handling different sized sigset_t's. */
639 if (sigsetsize != sizeof(sigset_t))
641 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
644 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
645 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
648 ret = core_sys_select(n, inp, outp, exp, to);
649 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
651 if (ret == -ERESTARTNOHAND) {
653 * Don't restore the signal mask yet. Let do_signal() deliver
654 * the signal on the way back to userspace, before the signal
658 memcpy(¤t->saved_sigmask, &sigsaved,
660 set_restore_sigmask();
663 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
669 * Most architectures can't handle 7-argument syscalls. So we provide a
670 * 6-argument version where the sixth argument is a pointer to a structure
671 * which has a pointer to the sigset_t itself followed by a size_t containing
674 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
675 fd_set __user *, exp, struct timespec __user *, tsp,
678 size_t sigsetsize = 0;
679 sigset_t __user *up = NULL;
682 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
683 || __get_user(up, (sigset_t __user * __user *)sig)
684 || __get_user(sigsetsize,
685 (size_t __user *)(sig+sizeof(void *))))
689 return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
691 #endif /* HAVE_SET_RESTORE_SIGMASK */
693 #ifdef __ARCH_WANT_SYS_OLD_SELECT
694 struct sel_arg_struct {
696 fd_set __user *inp, *outp, *exp;
697 struct timeval __user *tvp;
700 SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg)
702 struct sel_arg_struct a;
704 if (copy_from_user(&a, arg, sizeof(a)))
706 return sys_select(a.n, a.inp, a.outp, a.exp, a.tvp);
711 struct poll_list *next;
713 struct pollfd entries[0];
716 #define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
719 * Fish for pollable events on the pollfd->fd file descriptor. We're only
720 * interested in events matching the pollfd->events mask, and the result
721 * matching that mask is both recorded in pollfd->revents and returned. The
722 * pwait poll_table will be used by the fd-provided poll handler for waiting,
725 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
736 file = fget_light(fd, &fput_needed);
739 mask = DEFAULT_POLLMASK;
740 if (file->f_op && file->f_op->poll) {
742 pwait->key = pollfd->events |
744 mask = file->f_op->poll(file, pwait);
746 /* Mask out unneeded events. */
747 mask &= pollfd->events | POLLERR | POLLHUP;
748 fput_light(file, fput_needed);
751 pollfd->revents = mask;
756 static int do_poll(unsigned int nfds, struct poll_list *list,
757 struct poll_wqueues *wait, struct timespec *end_time)
759 poll_table* pt = &wait->pt;
760 ktime_t expire, *to = NULL;
761 int timed_out = 0, count = 0;
762 unsigned long slack = 0;
764 /* Optimise the no-wait case */
765 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
770 if (end_time && !timed_out)
771 slack = select_estimate_accuracy(end_time);
774 struct poll_list *walk;
776 for (walk = list; walk != NULL; walk = walk->next) {
777 struct pollfd * pfd, * pfd_end;
780 pfd_end = pfd + walk->len;
781 for (; pfd != pfd_end; pfd++) {
783 * Fish for events. If we found one, record it
784 * and kill the poll_table, so we don't
785 * needlessly register any other waiters after
786 * this. They'll get immediately deregistered
787 * when we break out and return.
789 if (do_pollfd(pfd, pt)) {
796 * All waiters have already been registered, so don't provide
797 * a poll_table to them on the next loop iteration.
802 if (signal_pending(current))
805 if (count || timed_out)
809 * If this is the first loop and we have a timeout
810 * given, then we convert to ktime_t and set the to
811 * pointer to the expiry value.
813 if (end_time && !to) {
814 expire = timespec_to_ktime(*end_time);
818 if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
824 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \
825 sizeof(struct pollfd))
827 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
828 struct timespec *end_time)
830 struct poll_wqueues table;
831 int err = -EFAULT, fdcount, len, size;
832 /* Allocate small arguments on the stack to save memory and be
833 faster - use long to make sure the buffer is aligned properly
834 on 64 bit archs to avoid unaligned access */
835 long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
836 struct poll_list *const head = (struct poll_list *)stack_pps;
837 struct poll_list *walk = head;
838 unsigned long todo = nfds;
840 if (nfds > rlimit(RLIMIT_NOFILE))
843 len = min_t(unsigned int, nfds, N_STACK_PPS);
850 if (copy_from_user(walk->entries, ufds + nfds-todo,
851 sizeof(struct pollfd) * walk->len))
858 len = min(todo, POLLFD_PER_PAGE);
859 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
860 walk = walk->next = kmalloc(size, GFP_KERNEL);
867 poll_initwait(&table);
868 fdcount = do_poll(nfds, head, &table, end_time);
869 poll_freewait(&table);
871 for (walk = head; walk; walk = walk->next) {
872 struct pollfd *fds = walk->entries;
875 for (j = 0; j < walk->len; j++, ufds++)
876 if (__put_user(fds[j].revents, &ufds->revents))
884 struct poll_list *pos = walk;
892 static long do_restart_poll(struct restart_block *restart_block)
894 struct pollfd __user *ufds = restart_block->poll.ufds;
895 int nfds = restart_block->poll.nfds;
896 struct timespec *to = NULL, end_time;
899 if (restart_block->poll.has_timeout) {
900 end_time.tv_sec = restart_block->poll.tv_sec;
901 end_time.tv_nsec = restart_block->poll.tv_nsec;
905 ret = do_sys_poll(ufds, nfds, to);
908 restart_block->fn = do_restart_poll;
909 ret = -ERESTART_RESTARTBLOCK;
914 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
917 struct timespec end_time, *to = NULL;
920 if (timeout_msecs >= 0) {
922 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
923 NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
926 ret = do_sys_poll(ufds, nfds, to);
929 struct restart_block *restart_block;
931 restart_block = ¤t_thread_info()->restart_block;
932 restart_block->fn = do_restart_poll;
933 restart_block->poll.ufds = ufds;
934 restart_block->poll.nfds = nfds;
936 if (timeout_msecs >= 0) {
937 restart_block->poll.tv_sec = end_time.tv_sec;
938 restart_block->poll.tv_nsec = end_time.tv_nsec;
939 restart_block->poll.has_timeout = 1;
941 restart_block->poll.has_timeout = 0;
943 ret = -ERESTART_RESTARTBLOCK;
948 #ifdef HAVE_SET_RESTORE_SIGMASK
949 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
950 struct timespec __user *, tsp, const sigset_t __user *, sigmask,
953 sigset_t ksigmask, sigsaved;
954 struct timespec ts, end_time, *to = NULL;
958 if (copy_from_user(&ts, tsp, sizeof(ts)))
962 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
967 /* XXX: Don't preclude handling different sized sigset_t's. */
968 if (sigsetsize != sizeof(sigset_t))
970 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
973 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
974 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
977 ret = do_sys_poll(ufds, nfds, to);
979 /* We can restart this syscall, usually */
982 * Don't restore the signal mask yet. Let do_signal() deliver
983 * the signal on the way back to userspace, before the signal
987 memcpy(¤t->saved_sigmask, &sigsaved,
989 set_restore_sigmask();
991 ret = -ERESTARTNOHAND;
993 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
995 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
999 #endif /* HAVE_SET_RESTORE_SIGMASK */