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))
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.
69 if (current->policy == SCHED_FIFO ||
70 current->policy == SCHED_RR)
74 now = timespec_sub(*tv, now);
75 ret = __estimate_accuracy(&now);
76 if (ret < current->timer_slack_ns)
77 return current->timer_slack_ns;
83 struct poll_table_page {
84 struct poll_table_page * next;
85 struct poll_table_entry * entry;
86 struct poll_table_entry entries[0];
89 #define POLL_TABLE_FULL(table) \
90 ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
93 * Ok, Peter made a complicated, but straightforward multiple_wait() function.
94 * I have rewritten this, taking some shortcuts: This code may not be easy to
95 * follow, but it should be free of race-conditions, and it's practical. If you
96 * understand what I'm doing here, then you understand how the linux
97 * sleep/wakeup mechanism works.
99 * Two very simple procedures, poll_wait() and poll_freewait() make all the
100 * work. poll_wait() is an inline-function defined in <linux/poll.h>,
101 * as all select/poll functions have to call it to add an entry to the
104 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
107 void poll_initwait(struct poll_wqueues *pwq)
109 init_poll_funcptr(&pwq->pt, __pollwait);
112 pwq->inline_index = 0;
115 EXPORT_SYMBOL(poll_initwait);
117 static void free_poll_entry(struct poll_table_entry *entry)
119 remove_wait_queue(entry->wait_address, &entry->wait);
123 void poll_freewait(struct poll_wqueues *pwq)
125 struct poll_table_page * p = pwq->table;
127 for (i = 0; i < pwq->inline_index; i++)
128 free_poll_entry(pwq->inline_entries + i);
130 struct poll_table_entry * entry;
131 struct poll_table_page *old;
136 free_poll_entry(entry);
137 } while (entry > p->entries);
140 free_page((unsigned long) old);
144 EXPORT_SYMBOL(poll_freewait);
146 static struct poll_table_entry *poll_get_entry(poll_table *_p)
148 struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt);
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);
160 __set_current_state(TASK_RUNNING);
163 new_table->entry = new_table->entries;
164 new_table->next = table;
165 p->table = new_table;
169 return table->entry++;
172 /* Add a new entry */
173 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
176 struct poll_table_entry *entry = poll_get_entry(p);
181 entry->wait_address = wait_address;
182 init_waitqueue_entry(&entry->wait, current);
183 add_wait_queue(wait_address, &entry->wait);
187 * poll_select_set_timeout - helper function to setup the timeout value
188 * @to: pointer to timespec variable for the final timeout
189 * @sec: seconds (from user space)
190 * @nsec: nanoseconds (from user space)
192 * Note, we do not use a timespec for the user space value here, That
193 * way we can use the function for timeval and compat interfaces as well.
195 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
197 int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
199 struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
201 if (!timespec_valid(&ts))
204 /* Optimize for the zero timeout value here */
206 to->tv_sec = to->tv_nsec = 0;
209 *to = timespec_add_safe(*to, ts);
214 static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
215 int timeval, int ret)
223 if (current->personality & STICKY_TIMEOUTS)
226 /* No update for zero timeout */
227 if (!end_time->tv_sec && !end_time->tv_nsec)
231 rts = timespec_sub(*end_time, rts);
233 rts.tv_sec = rts.tv_nsec = 0;
236 rtv.tv_sec = rts.tv_sec;
237 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
239 if (!copy_to_user(p, &rtv, sizeof(rtv)))
242 } else if (!copy_to_user(p, &rts, sizeof(rts)))
246 * If an application puts its timeval in read-only memory, we
247 * don't want the Linux-specific update to the timeval to
248 * cause a fault after the select has completed
249 * successfully. However, because we're not updating the
250 * timeval, we can't restart the system call.
254 if (ret == -ERESTARTNOHAND)
259 #define FDS_IN(fds, n) (fds->in + n)
260 #define FDS_OUT(fds, n) (fds->out + n)
261 #define FDS_EX(fds, n) (fds->ex + n)
263 #define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
265 static int max_select_fd(unsigned long n, fd_set_bits *fds)
267 unsigned long *open_fds;
272 /* handle last in-complete long-word first */
273 set = ~(~0UL << (n & (__NFDBITS-1)));
275 fdt = files_fdtable(current->files);
276 open_fds = fdt->open_fds->fds_bits+n;
281 if (!(set & ~*open_fds))
292 if (set & ~*open_fds)
301 max += n * __NFDBITS;
307 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
308 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
309 #define POLLEX_SET (POLLPRI)
311 int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
313 ktime_t expire, *to = NULL;
314 struct poll_wqueues table;
316 int retval, i, timed_out = 0;
317 unsigned long slack = 0;
320 retval = max_select_fd(n, fds);
327 poll_initwait(&table);
329 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
335 slack = estimate_accuracy(end_time);
339 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
341 set_current_state(TASK_INTERRUPTIBLE);
343 inp = fds->in; outp = fds->out; exp = fds->ex;
344 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
346 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
347 unsigned long in, out, ex, all_bits, bit = 1, mask, j;
348 unsigned long res_in = 0, res_out = 0, res_ex = 0;
349 const struct file_operations *f_op = NULL;
350 struct file *file = NULL;
352 in = *inp++; out = *outp++; ex = *exp++;
353 all_bits = in | out | ex;
359 for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
363 if (!(bit & all_bits))
365 file = fget_light(i, &fput_needed);
368 mask = DEFAULT_POLLMASK;
369 if (f_op && f_op->poll)
370 mask = (*f_op->poll)(file, retval ? NULL : wait);
371 fput_light(file, fput_needed);
372 if ((mask & POLLIN_SET) && (in & bit)) {
376 if ((mask & POLLOUT_SET) && (out & bit)) {
380 if ((mask & POLLEX_SET) && (ex & bit)) {
395 if (retval || timed_out || signal_pending(current))
398 retval = table.error;
403 * If this is the first loop and we have a timeout
404 * given, then we convert to ktime_t and set the to
405 * pointer to the expiry value.
407 if (end_time && !to) {
408 expire = timespec_to_ktime(*end_time);
412 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
415 __set_current_state(TASK_RUNNING);
417 poll_freewait(&table);
423 * We can actually return ERESTARTSYS instead of EINTR, but I'd
424 * like to be certain this leads to no problems. So I return
425 * EINTR just for safety.
427 * Update: ERESTARTSYS breaks at least the xview clock binary, so
428 * I'm trying ERESTARTNOHAND which restart only when you want to.
430 #define MAX_SELECT_SECONDS \
431 ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
433 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
434 fd_set __user *exp, struct timespec *end_time)
441 /* Allocate small arguments on the stack to save memory and be faster */
442 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
448 /* max_fds can increase, so grab it once to avoid race */
450 fdt = files_fdtable(current->files);
451 max_fds = fdt->max_fds;
457 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
458 * since we used fdset we need to allocate memory in units of
463 if (size > sizeof(stack_fds) / 6) {
464 /* Not enough space in on-stack array; must use kmalloc */
466 bits = kmalloc(6 * size, GFP_KERNEL);
471 fds.out = bits + size;
472 fds.ex = bits + 2*size;
473 fds.res_in = bits + 3*size;
474 fds.res_out = bits + 4*size;
475 fds.res_ex = bits + 5*size;
477 if ((ret = get_fd_set(n, inp, fds.in)) ||
478 (ret = get_fd_set(n, outp, fds.out)) ||
479 (ret = get_fd_set(n, exp, fds.ex)))
481 zero_fd_set(n, fds.res_in);
482 zero_fd_set(n, fds.res_out);
483 zero_fd_set(n, fds.res_ex);
485 ret = do_select(n, &fds, end_time);
490 ret = -ERESTARTNOHAND;
491 if (signal_pending(current))
496 if (set_fd_set(n, inp, fds.res_in) ||
497 set_fd_set(n, outp, fds.res_out) ||
498 set_fd_set(n, exp, fds.res_ex))
502 if (bits != stack_fds)
508 asmlinkage long sys_select(int n, fd_set __user *inp, fd_set __user *outp,
509 fd_set __user *exp, struct timeval __user *tvp)
511 struct timespec end_time, *to = NULL;
516 if (copy_from_user(&tv, tvp, sizeof(tv)))
520 if (poll_select_set_timeout(to, tv.tv_sec,
521 tv.tv_usec * NSEC_PER_USEC))
525 ret = core_sys_select(n, inp, outp, exp, to);
526 ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
531 #ifdef HAVE_SET_RESTORE_SIGMASK
532 asmlinkage long sys_pselect7(int n, fd_set __user *inp, fd_set __user *outp,
533 fd_set __user *exp, struct timespec __user *tsp,
534 const sigset_t __user *sigmask, size_t sigsetsize)
536 sigset_t ksigmask, sigsaved;
537 struct timespec ts, end_time, *to = NULL;
541 if (copy_from_user(&ts, tsp, sizeof(ts)))
545 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
550 /* XXX: Don't preclude handling different sized sigset_t's. */
551 if (sigsetsize != sizeof(sigset_t))
553 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
556 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
557 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
560 ret = core_sys_select(n, inp, outp, exp, &end_time);
561 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
563 if (ret == -ERESTARTNOHAND) {
565 * Don't restore the signal mask yet. Let do_signal() deliver
566 * the signal on the way back to userspace, before the signal
570 memcpy(¤t->saved_sigmask, &sigsaved,
572 set_restore_sigmask();
575 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
581 * Most architectures can't handle 7-argument syscalls. So we provide a
582 * 6-argument version where the sixth argument is a pointer to a structure
583 * which has a pointer to the sigset_t itself followed by a size_t containing
586 asmlinkage long sys_pselect6(int n, fd_set __user *inp, fd_set __user *outp,
587 fd_set __user *exp, struct timespec __user *tsp, void __user *sig)
589 size_t sigsetsize = 0;
590 sigset_t __user *up = NULL;
593 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
594 || __get_user(up, (sigset_t __user * __user *)sig)
595 || __get_user(sigsetsize,
596 (size_t __user *)(sig+sizeof(void *))))
600 return sys_pselect7(n, inp, outp, exp, tsp, up, sigsetsize);
602 #endif /* HAVE_SET_RESTORE_SIGMASK */
605 struct poll_list *next;
607 struct pollfd entries[0];
610 #define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
613 * Fish for pollable events on the pollfd->fd file descriptor. We're only
614 * interested in events matching the pollfd->events mask, and the result
615 * matching that mask is both recorded in pollfd->revents and returned. The
616 * pwait poll_table will be used by the fd-provided poll handler for waiting,
619 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
630 file = fget_light(fd, &fput_needed);
633 mask = DEFAULT_POLLMASK;
634 if (file->f_op && file->f_op->poll)
635 mask = file->f_op->poll(file, pwait);
636 /* Mask out unneeded events. */
637 mask &= pollfd->events | POLLERR | POLLHUP;
638 fput_light(file, fput_needed);
641 pollfd->revents = mask;
646 static int do_poll(unsigned int nfds, struct poll_list *list,
647 struct poll_wqueues *wait, struct timespec *end_time)
649 poll_table* pt = &wait->pt;
650 ktime_t expire, *to = NULL;
651 int timed_out = 0, count = 0;
652 unsigned long slack = 0;
654 /* Optimise the no-wait case */
655 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
661 slack = estimate_accuracy(end_time);
664 struct poll_list *walk;
666 set_current_state(TASK_INTERRUPTIBLE);
667 for (walk = list; walk != NULL; walk = walk->next) {
668 struct pollfd * pfd, * pfd_end;
671 pfd_end = pfd + walk->len;
672 for (; pfd != pfd_end; pfd++) {
674 * Fish for events. If we found one, record it
675 * and kill the poll_table, so we don't
676 * needlessly register any other waiters after
677 * this. They'll get immediately deregistered
678 * when we break out and return.
680 if (do_pollfd(pfd, pt)) {
687 * All waiters have already been registered, so don't provide
688 * a poll_table to them on the next loop iteration.
693 if (signal_pending(current))
696 if (count || timed_out)
700 * If this is the first loop and we have a timeout
701 * given, then we convert to ktime_t and set the to
702 * pointer to the expiry value.
704 if (end_time && !to) {
705 expire = timespec_to_ktime(*end_time);
709 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
712 __set_current_state(TASK_RUNNING);
716 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \
717 sizeof(struct pollfd))
719 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
720 struct timespec *end_time)
722 struct poll_wqueues table;
723 int err = -EFAULT, fdcount, len, size;
724 /* Allocate small arguments on the stack to save memory and be
725 faster - use long to make sure the buffer is aligned properly
726 on 64 bit archs to avoid unaligned access */
727 long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
728 struct poll_list *const head = (struct poll_list *)stack_pps;
729 struct poll_list *walk = head;
730 unsigned long todo = nfds;
732 if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
735 len = min_t(unsigned int, nfds, N_STACK_PPS);
742 if (copy_from_user(walk->entries, ufds + nfds-todo,
743 sizeof(struct pollfd) * walk->len))
750 len = min(todo, POLLFD_PER_PAGE);
751 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
752 walk = walk->next = kmalloc(size, GFP_KERNEL);
759 poll_initwait(&table);
760 fdcount = do_poll(nfds, head, &table, end_time);
761 poll_freewait(&table);
763 for (walk = head; walk; walk = walk->next) {
764 struct pollfd *fds = walk->entries;
767 for (j = 0; j < walk->len; j++, ufds++)
768 if (__put_user(fds[j].revents, &ufds->revents))
776 struct poll_list *pos = walk;
784 static long do_restart_poll(struct restart_block *restart_block)
786 struct pollfd __user *ufds = restart_block->poll.ufds;
787 int nfds = restart_block->poll.nfds;
788 struct timespec *to = NULL, end_time;
791 if (restart_block->poll.has_timeout) {
792 end_time.tv_sec = restart_block->poll.tv_sec;
793 end_time.tv_nsec = restart_block->poll.tv_nsec;
797 ret = do_sys_poll(ufds, nfds, to);
800 restart_block->fn = do_restart_poll;
801 ret = -ERESTART_RESTARTBLOCK;
806 asmlinkage long sys_poll(struct pollfd __user *ufds, unsigned int nfds,
809 struct timespec end_time, *to = NULL;
812 if (timeout_msecs >= 0) {
814 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
815 NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
818 ret = do_sys_poll(ufds, nfds, to);
821 struct restart_block *restart_block;
823 restart_block = ¤t_thread_info()->restart_block;
824 restart_block->fn = do_restart_poll;
825 restart_block->poll.ufds = ufds;
826 restart_block->poll.nfds = nfds;
828 if (timeout_msecs >= 0) {
829 restart_block->poll.tv_sec = end_time.tv_sec;
830 restart_block->poll.tv_nsec = end_time.tv_nsec;
831 restart_block->poll.has_timeout = 1;
833 restart_block->poll.has_timeout = 0;
835 ret = -ERESTART_RESTARTBLOCK;
840 #ifdef HAVE_SET_RESTORE_SIGMASK
841 asmlinkage long sys_ppoll(struct pollfd __user *ufds, unsigned int nfds,
842 struct timespec __user *tsp, const sigset_t __user *sigmask,
845 sigset_t ksigmask, sigsaved;
846 struct timespec ts, end_time, *to = NULL;
850 if (copy_from_user(&ts, tsp, sizeof(ts)))
854 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
859 /* XXX: Don't preclude handling different sized sigset_t's. */
860 if (sigsetsize != sizeof(sigset_t))
862 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
865 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
866 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
869 ret = do_sys_poll(ufds, nfds, to);
871 /* We can restart this syscall, usually */
874 * Don't restore the signal mask yet. Let do_signal() deliver
875 * the signal on the way back to userspace, before the signal
879 memcpy(¤t->saved_sigmask, &sigsaved,
881 set_restore_sigmask();
883 ret = -ERESTARTNOHAND;
885 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
887 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
891 #endif /* HAVE_SET_RESTORE_SIGMASK */