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/export.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>
29 #include <linux/sched/rt.h>
30 #include <linux/freezer.h>
32 #include <asm/uaccess.h>
36 * Estimate expected accuracy in ns from a timeval.
38 * After quite a bit of churning around, we've settled on
39 * a simple thing of taking 0.1% of the timeout as the
40 * slack, with a cap of 100 msec.
41 * "nice" tasks get a 0.5% slack instead.
43 * Consider this comment an open invitation to come up with even
47 #define MAX_SLACK (100 * NSEC_PER_MSEC)
49 static long __estimate_accuracy(struct timespec *tv)
57 if (task_nice(current) > 0)
58 divfactor = divfactor / 5;
60 if (tv->tv_sec > MAX_SLACK / (NSEC_PER_SEC/divfactor))
63 slack = tv->tv_nsec / divfactor;
64 slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
66 if (slack > MAX_SLACK)
72 long select_estimate_accuracy(struct timespec *tv)
78 * Realtime tasks get a slack of 0 for obvious reasons.
85 now = timespec_sub(*tv, now);
86 ret = __estimate_accuracy(&now);
87 if (ret < current->timer_slack_ns)
88 return current->timer_slack_ns;
94 struct poll_table_page {
95 struct poll_table_page * next;
96 struct poll_table_entry * entry;
97 struct poll_table_entry entries[0];
100 #define POLL_TABLE_FULL(table) \
101 ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
104 * Ok, Peter made a complicated, but straightforward multiple_wait() function.
105 * I have rewritten this, taking some shortcuts: This code may not be easy to
106 * follow, but it should be free of race-conditions, and it's practical. If you
107 * understand what I'm doing here, then you understand how the linux
108 * sleep/wakeup mechanism works.
110 * Two very simple procedures, poll_wait() and poll_freewait() make all the
111 * work. poll_wait() is an inline-function defined in <linux/poll.h>,
112 * as all select/poll functions have to call it to add an entry to the
115 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
118 void poll_initwait(struct poll_wqueues *pwq)
120 init_poll_funcptr(&pwq->pt, __pollwait);
121 pwq->polling_task = current;
125 pwq->inline_index = 0;
127 EXPORT_SYMBOL(poll_initwait);
129 static void free_poll_entry(struct poll_table_entry *entry)
131 remove_wait_queue(entry->wait_address, &entry->wait);
135 void poll_freewait(struct poll_wqueues *pwq)
137 struct poll_table_page * p = pwq->table;
139 for (i = 0; i < pwq->inline_index; i++)
140 free_poll_entry(pwq->inline_entries + i);
142 struct poll_table_entry * entry;
143 struct poll_table_page *old;
148 free_poll_entry(entry);
149 } while (entry > p->entries);
152 free_page((unsigned long) old);
155 EXPORT_SYMBOL(poll_freewait);
157 static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
159 struct poll_table_page *table = p->table;
161 if (p->inline_index < N_INLINE_POLL_ENTRIES)
162 return p->inline_entries + p->inline_index++;
164 if (!table || POLL_TABLE_FULL(table)) {
165 struct poll_table_page *new_table;
167 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
172 new_table->entry = new_table->entries;
173 new_table->next = table;
174 p->table = new_table;
178 return table->entry++;
181 static int __pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
183 struct poll_wqueues *pwq = wait->private;
184 DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
187 * Although this function is called under waitqueue lock, LOCK
188 * doesn't imply write barrier and the users expect write
189 * barrier semantics on wakeup functions. The following
190 * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
191 * and is paired with set_mb() in poll_schedule_timeout.
197 * Perform the default wake up operation using a dummy
200 * TODO: This is hacky but there currently is no interface to
201 * pass in @sync. @sync is scheduled to be removed and once
202 * that happens, wake_up_process() can be used directly.
204 return default_wake_function(&dummy_wait, mode, sync, key);
207 static int pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
209 struct poll_table_entry *entry;
211 entry = container_of(wait, struct poll_table_entry, wait);
212 if (key && !((unsigned long)key & entry->key))
214 return __pollwake(wait, mode, sync, key);
217 /* Add a new entry */
218 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
221 struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
222 struct poll_table_entry *entry = poll_get_entry(pwq);
225 entry->filp = get_file(filp);
226 entry->wait_address = wait_address;
227 entry->key = p->_key;
228 init_waitqueue_func_entry(&entry->wait, pollwake);
229 entry->wait.private = pwq;
230 add_wait_queue(wait_address, &entry->wait);
233 int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
234 ktime_t *expires, unsigned long slack)
238 set_current_state(state);
240 rc = freezable_schedule_hrtimeout_range(expires, slack,
242 __set_current_state(TASK_RUNNING);
245 * Prepare for the next iteration.
247 * The following set_mb() serves two purposes. First, it's
248 * the counterpart rmb of the wmb in pollwake() such that data
249 * written before wake up is always visible after wake up.
250 * Second, the full barrier guarantees that triggered clearing
251 * doesn't pass event check of the next iteration. Note that
252 * this problem doesn't exist for the first iteration as
253 * add_wait_queue() has full barrier semantics.
255 set_mb(pwq->triggered, 0);
259 EXPORT_SYMBOL(poll_schedule_timeout);
262 * poll_select_set_timeout - helper function to setup the timeout value
263 * @to: pointer to timespec variable for the final timeout
264 * @sec: seconds (from user space)
265 * @nsec: nanoseconds (from user space)
267 * Note, we do not use a timespec for the user space value here, That
268 * way we can use the function for timeval and compat interfaces as well.
270 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
272 int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
274 struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
276 if (!timespec_valid(&ts))
279 /* Optimize for the zero timeout value here */
281 to->tv_sec = to->tv_nsec = 0;
284 *to = timespec_add_safe(*to, ts);
289 static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
290 int timeval, int ret)
298 if (current->personality & STICKY_TIMEOUTS)
301 /* No update for zero timeout */
302 if (!end_time->tv_sec && !end_time->tv_nsec)
306 rts = timespec_sub(*end_time, rts);
308 rts.tv_sec = rts.tv_nsec = 0;
311 if (sizeof(rtv) > sizeof(rtv.tv_sec) + sizeof(rtv.tv_usec))
312 memset(&rtv, 0, sizeof(rtv));
313 rtv.tv_sec = rts.tv_sec;
314 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
316 if (!copy_to_user(p, &rtv, sizeof(rtv)))
319 } else if (!copy_to_user(p, &rts, sizeof(rts)))
323 * If an application puts its timeval in read-only memory, we
324 * don't want the Linux-specific update to the timeval to
325 * cause a fault after the select has completed
326 * successfully. However, because we're not updating the
327 * timeval, we can't restart the system call.
331 if (ret == -ERESTARTNOHAND)
336 #define FDS_IN(fds, n) (fds->in + n)
337 #define FDS_OUT(fds, n) (fds->out + n)
338 #define FDS_EX(fds, n) (fds->ex + n)
340 #define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
342 static int max_select_fd(unsigned long n, fd_set_bits *fds)
344 unsigned long *open_fds;
349 /* handle last in-complete long-word first */
350 set = ~(~0UL << (n & (BITS_PER_LONG-1)));
352 fdt = files_fdtable(current->files);
353 open_fds = fdt->open_fds + n;
358 if (!(set & ~*open_fds))
369 if (set & ~*open_fds)
378 max += n * BITS_PER_LONG;
384 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
385 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
386 #define POLLEX_SET (POLLPRI)
388 static inline void wait_key_set(poll_table *wait, unsigned long in,
389 unsigned long out, unsigned long bit)
391 wait->_key = POLLEX_SET;
393 wait->_key |= POLLIN_SET;
395 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;
435 in = *inp++; out = *outp++; ex = *exp++;
436 all_bits = in | out | ex;
442 for (j = 0; j < BITS_PER_LONG; ++j, ++i, bit <<= 1) {
446 if (!(bit & all_bits))
450 const struct file_operations *f_op;
452 mask = DEFAULT_POLLMASK;
453 if (f_op && f_op->poll) {
454 wait_key_set(wait, in, out, bit);
455 mask = (*f_op->poll)(f.file, wait);
458 if ((mask & POLLIN_SET) && (in & bit)) {
463 if ((mask & POLLOUT_SET) && (out & bit)) {
468 if ((mask & POLLEX_SET) && (ex & bit)) {
484 if (retval || timed_out || signal_pending(current))
487 retval = table.error;
492 * If this is the first loop and we have a timeout
493 * given, then we convert to ktime_t and set the to
494 * pointer to the expiry value.
496 if (end_time && !to) {
497 expire = timespec_to_ktime(*end_time);
501 if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
506 poll_freewait(&table);
512 * We can actually return ERESTARTSYS instead of EINTR, but I'd
513 * like to be certain this leads to no problems. So I return
514 * EINTR just for safety.
516 * Update: ERESTARTSYS breaks at least the xview clock binary, so
517 * I'm trying ERESTARTNOHAND which restart only when you want to.
519 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
520 fd_set __user *exp, struct timespec *end_time)
527 /* Allocate small arguments on the stack to save memory and be faster */
528 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
534 /* max_fds can increase, so grab it once to avoid race */
536 fdt = files_fdtable(current->files);
537 max_fds = fdt->max_fds;
543 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
544 * since we used fdset we need to allocate memory in units of
549 if (size > sizeof(stack_fds) / 6) {
550 /* Not enough space in on-stack array; must use kmalloc */
552 bits = kmalloc(6 * size, GFP_KERNEL);
557 fds.out = bits + size;
558 fds.ex = bits + 2*size;
559 fds.res_in = bits + 3*size;
560 fds.res_out = bits + 4*size;
561 fds.res_ex = bits + 5*size;
563 if ((ret = get_fd_set(n, inp, fds.in)) ||
564 (ret = get_fd_set(n, outp, fds.out)) ||
565 (ret = get_fd_set(n, exp, fds.ex)))
567 zero_fd_set(n, fds.res_in);
568 zero_fd_set(n, fds.res_out);
569 zero_fd_set(n, fds.res_ex);
571 ret = do_select(n, &fds, end_time);
576 ret = -ERESTARTNOHAND;
577 if (signal_pending(current))
582 if (set_fd_set(n, inp, fds.res_in) ||
583 set_fd_set(n, outp, fds.res_out) ||
584 set_fd_set(n, exp, fds.res_ex))
588 if (bits != stack_fds)
594 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
595 fd_set __user *, exp, struct timeval __user *, tvp)
597 struct timespec end_time, *to = NULL;
602 if (copy_from_user(&tv, tvp, sizeof(tv)))
606 if (poll_select_set_timeout(to,
607 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
608 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
612 ret = core_sys_select(n, inp, outp, exp, to);
613 ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
618 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
619 fd_set __user *exp, struct timespec __user *tsp,
620 const sigset_t __user *sigmask, size_t sigsetsize)
622 sigset_t ksigmask, sigsaved;
623 struct timespec ts, end_time, *to = NULL;
627 if (copy_from_user(&ts, tsp, sizeof(ts)))
631 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
636 /* XXX: Don't preclude handling different sized sigset_t's. */
637 if (sigsetsize != sizeof(sigset_t))
639 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
642 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
643 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
646 ret = core_sys_select(n, inp, outp, exp, to);
647 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
649 if (ret == -ERESTARTNOHAND) {
651 * Don't restore the signal mask yet. Let do_signal() deliver
652 * the signal on the way back to userspace, before the signal
656 memcpy(¤t->saved_sigmask, &sigsaved,
658 set_restore_sigmask();
661 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
667 * Most architectures can't handle 7-argument syscalls. So we provide a
668 * 6-argument version where the sixth argument is a pointer to a structure
669 * which has a pointer to the sigset_t itself followed by a size_t containing
672 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
673 fd_set __user *, exp, struct timespec __user *, tsp,
676 size_t sigsetsize = 0;
677 sigset_t __user *up = NULL;
680 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
681 || __get_user(up, (sigset_t __user * __user *)sig)
682 || __get_user(sigsetsize,
683 (size_t __user *)(sig+sizeof(void *))))
687 return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
690 #ifdef __ARCH_WANT_SYS_OLD_SELECT
691 struct sel_arg_struct {
693 fd_set __user *inp, *outp, *exp;
694 struct timeval __user *tvp;
697 SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg)
699 struct sel_arg_struct a;
701 if (copy_from_user(&a, arg, sizeof(a)))
703 return sys_select(a.n, a.inp, a.outp, a.exp, a.tvp);
708 struct poll_list *next;
710 struct pollfd entries[0];
713 #define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
716 * Fish for pollable events on the pollfd->fd file descriptor. We're only
717 * interested in events matching the pollfd->events mask, and the result
718 * matching that mask is both recorded in pollfd->revents and returned. The
719 * pwait poll_table will be used by the fd-provided poll handler for waiting,
720 * if pwait->_qproc is non-NULL.
722 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
730 struct fd f = fdget(fd);
733 mask = DEFAULT_POLLMASK;
734 if (f.file->f_op && f.file->f_op->poll) {
735 pwait->_key = pollfd->events|POLLERR|POLLHUP;
736 mask = f.file->f_op->poll(f.file, pwait);
738 /* Mask out unneeded events. */
739 mask &= pollfd->events | POLLERR | POLLHUP;
743 pollfd->revents = mask;
748 static int do_poll(unsigned int nfds, struct poll_list *list,
749 struct poll_wqueues *wait, struct timespec *end_time)
751 poll_table* pt = &wait->pt;
752 ktime_t expire, *to = NULL;
753 int timed_out = 0, count = 0;
754 unsigned long slack = 0;
756 /* Optimise the no-wait case */
757 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
762 if (end_time && !timed_out)
763 slack = select_estimate_accuracy(end_time);
766 struct poll_list *walk;
768 for (walk = list; walk != NULL; walk = walk->next) {
769 struct pollfd * pfd, * pfd_end;
772 pfd_end = pfd + walk->len;
773 for (; pfd != pfd_end; pfd++) {
775 * Fish for events. If we found one, record it
776 * and kill poll_table->_qproc, so we don't
777 * needlessly register any other waiters after
778 * this. They'll get immediately deregistered
779 * when we break out and return.
781 if (do_pollfd(pfd, pt)) {
788 * All waiters have already been registered, so don't provide
789 * a poll_table->_qproc to them on the next loop iteration.
794 if (signal_pending(current))
797 if (count || timed_out)
801 * If this is the first loop and we have a timeout
802 * given, then we convert to ktime_t and set the to
803 * pointer to the expiry value.
805 if (end_time && !to) {
806 expire = timespec_to_ktime(*end_time);
810 if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
816 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \
817 sizeof(struct pollfd))
819 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
820 struct timespec *end_time)
822 struct poll_wqueues table;
823 int err = -EFAULT, fdcount, len, size;
824 /* Allocate small arguments on the stack to save memory and be
825 faster - use long to make sure the buffer is aligned properly
826 on 64 bit archs to avoid unaligned access */
827 long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
828 struct poll_list *const head = (struct poll_list *)stack_pps;
829 struct poll_list *walk = head;
830 unsigned long todo = nfds;
832 if (nfds > rlimit(RLIMIT_NOFILE))
835 len = min_t(unsigned int, nfds, N_STACK_PPS);
842 if (copy_from_user(walk->entries, ufds + nfds-todo,
843 sizeof(struct pollfd) * walk->len))
850 len = min(todo, POLLFD_PER_PAGE);
851 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
852 walk = walk->next = kmalloc(size, GFP_KERNEL);
859 poll_initwait(&table);
860 fdcount = do_poll(nfds, head, &table, end_time);
861 poll_freewait(&table);
863 for (walk = head; walk; walk = walk->next) {
864 struct pollfd *fds = walk->entries;
867 for (j = 0; j < walk->len; j++, ufds++)
868 if (__put_user(fds[j].revents, &ufds->revents))
876 struct poll_list *pos = walk;
884 static long do_restart_poll(struct restart_block *restart_block)
886 struct pollfd __user *ufds = restart_block->poll.ufds;
887 int nfds = restart_block->poll.nfds;
888 struct timespec *to = NULL, end_time;
891 if (restart_block->poll.has_timeout) {
892 end_time.tv_sec = restart_block->poll.tv_sec;
893 end_time.tv_nsec = restart_block->poll.tv_nsec;
897 ret = do_sys_poll(ufds, nfds, to);
900 restart_block->fn = do_restart_poll;
901 ret = -ERESTART_RESTARTBLOCK;
906 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
909 struct timespec end_time, *to = NULL;
912 if (timeout_msecs >= 0) {
914 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
915 NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
918 ret = do_sys_poll(ufds, nfds, to);
921 struct restart_block *restart_block;
923 restart_block = ¤t_thread_info()->restart_block;
924 restart_block->fn = do_restart_poll;
925 restart_block->poll.ufds = ufds;
926 restart_block->poll.nfds = nfds;
928 if (timeout_msecs >= 0) {
929 restart_block->poll.tv_sec = end_time.tv_sec;
930 restart_block->poll.tv_nsec = end_time.tv_nsec;
931 restart_block->poll.has_timeout = 1;
933 restart_block->poll.has_timeout = 0;
935 ret = -ERESTART_RESTARTBLOCK;
940 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
941 struct timespec __user *, tsp, const sigset_t __user *, sigmask,
944 sigset_t ksigmask, sigsaved;
945 struct timespec ts, end_time, *to = NULL;
949 if (copy_from_user(&ts, tsp, sizeof(ts)))
953 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
958 /* XXX: Don't preclude handling different sized sigset_t's. */
959 if (sigsetsize != sizeof(sigset_t))
961 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
964 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
965 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
968 ret = do_sys_poll(ufds, nfds, to);
970 /* We can restart this syscall, usually */
973 * Don't restore the signal mask yet. Let do_signal() deliver
974 * the signal on the way back to userspace, before the signal
978 memcpy(¤t->saved_sigmask, &sigsaved,
980 set_restore_sigmask();
982 ret = -ERESTARTNOHAND;
984 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
986 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);