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).
16 * Copyright (c) 2013, NVIDIA CORPORATION. All rights reserved.
19 #include <linux/kernel.h>
20 #include <linux/sched.h>
21 #include <linux/syscalls.h>
22 #include <linux/export.h>
23 #include <linux/slab.h>
24 #include <linux/poll.h>
25 #include <linux/personality.h> /* for STICKY_TIMEOUTS */
26 #include <linux/file.h>
27 #include <linux/fdtable.h>
29 #include <linux/rcupdate.h>
30 #include <linux/hrtimer.h>
31 #include <linux/sched/rt.h>
32 #include <linux/freezer.h>
34 #include <asm/uaccess.h>
38 * Estimate expected accuracy in ns from a timeval.
40 * After quite a bit of churning around, we've settled on
41 * a simple thing of taking 0.1% of the timeout as the
42 * slack, with a cap of 100 msec.
43 * "nice" tasks get a 0.5% slack instead.
45 * Consider this comment an open invitation to come up with even
49 #define MAX_SLACK (100 * NSEC_PER_MSEC)
51 static long __estimate_accuracy(struct timespec *tv)
59 if (task_nice(current) > 0)
60 divfactor = divfactor / 5;
62 if (tv->tv_sec > MAX_SLACK / (NSEC_PER_SEC/divfactor))
65 slack = tv->tv_nsec / divfactor;
66 slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
68 if (slack > MAX_SLACK)
74 long select_estimate_accuracy(struct timespec *tv)
80 * Realtime tasks get a slack of 0 for obvious reasons.
87 now = timespec_sub(*tv, now);
88 ret = __estimate_accuracy(&now);
89 if (ret < current->timer_slack_ns)
90 return current->timer_slack_ns;
96 struct poll_table_page {
97 struct poll_table_page * next;
98 struct poll_table_entry * entry;
99 struct poll_table_entry entries[0];
102 #define POLL_TABLE_FULL(table) \
103 ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
106 * Ok, Peter made a complicated, but straightforward multiple_wait() function.
107 * I have rewritten this, taking some shortcuts: This code may not be easy to
108 * follow, but it should be free of race-conditions, and it's practical. If you
109 * understand what I'm doing here, then you understand how the linux
110 * sleep/wakeup mechanism works.
112 * Two very simple procedures, poll_wait() and poll_freewait() make all the
113 * work. poll_wait() is an inline-function defined in <linux/poll.h>,
114 * as all select/poll functions have to call it to add an entry to the
117 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
120 void poll_initwait(struct poll_wqueues *pwq)
122 init_poll_funcptr(&pwq->pt, __pollwait);
123 pwq->polling_task = current;
127 pwq->inline_index = 0;
129 EXPORT_SYMBOL(poll_initwait);
131 static void free_poll_entry(struct poll_table_entry *entry)
133 remove_wait_queue(entry->wait_address, &entry->wait);
137 void poll_freewait(struct poll_wqueues *pwq)
139 struct poll_table_page * p = pwq->table;
141 for (i = 0; i < pwq->inline_index; i++)
142 free_poll_entry(pwq->inline_entries + i);
144 struct poll_table_entry * entry;
145 struct poll_table_page *old;
150 free_poll_entry(entry);
151 } while (entry > p->entries);
154 free_page((unsigned long) old);
157 EXPORT_SYMBOL(poll_freewait);
159 static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
161 struct poll_table_page *table = p->table;
163 if (p->inline_index < N_INLINE_POLL_ENTRIES)
164 return p->inline_entries + p->inline_index++;
166 if (!table || POLL_TABLE_FULL(table)) {
167 struct poll_table_page *new_table;
169 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
174 new_table->entry = new_table->entries;
175 new_table->next = table;
176 p->table = new_table;
180 return table->entry++;
183 static int __pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
185 struct poll_wqueues *pwq = wait->private;
186 DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
189 * Although this function is called under waitqueue lock, LOCK
190 * doesn't imply write barrier and the users expect write
191 * barrier semantics on wakeup functions. The following
192 * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
193 * and is paired with set_mb() in poll_schedule_timeout.
199 * Perform the default wake up operation using a dummy
202 * TODO: This is hacky but there currently is no interface to
203 * pass in @sync. @sync is scheduled to be removed and once
204 * that happens, wake_up_process() can be used directly.
206 return default_wake_function(&dummy_wait, mode, sync, key);
209 int pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
211 struct poll_table_entry *entry;
213 entry = container_of(wait, struct poll_table_entry, wait);
214 if (key && !((unsigned long)key & entry->key))
216 return __pollwake(wait, mode, sync, key);
218 EXPORT_SYMBOL(pollwake);
220 /* Add a new entry */
221 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
224 struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
225 struct poll_table_entry *entry = poll_get_entry(pwq);
228 entry->filp = get_file(filp);
229 entry->wait_address = wait_address;
230 entry->key = p->_key;
231 init_waitqueue_func_entry(&entry->wait, pollwake);
232 entry->wait.private = pwq;
233 add_wait_queue(wait_address, &entry->wait);
236 int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
237 ktime_t *expires, unsigned long slack)
241 set_current_state(state);
243 rc = freezable_schedule_hrtimeout_range(expires, slack,
245 __set_current_state(TASK_RUNNING);
248 * Prepare for the next iteration.
250 * The following set_mb() serves two purposes. First, it's
251 * the counterpart rmb of the wmb in pollwake() such that data
252 * written before wake up is always visible after wake up.
253 * Second, the full barrier guarantees that triggered clearing
254 * doesn't pass event check of the next iteration. Note that
255 * this problem doesn't exist for the first iteration as
256 * add_wait_queue() has full barrier semantics.
258 set_mb(pwq->triggered, 0);
262 EXPORT_SYMBOL(poll_schedule_timeout);
265 * poll_select_set_timeout - helper function to setup the timeout value
266 * @to: pointer to timespec variable for the final timeout
267 * @sec: seconds (from user space)
268 * @nsec: nanoseconds (from user space)
270 * Note, we do not use a timespec for the user space value here, That
271 * way we can use the function for timeval and compat interfaces as well.
273 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
275 int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
277 struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
279 if (!timespec_valid(&ts))
282 /* Optimize for the zero timeout value here */
284 to->tv_sec = to->tv_nsec = 0;
287 *to = timespec_add_safe(*to, ts);
292 static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
293 int timeval, int ret)
301 if (current->personality & STICKY_TIMEOUTS)
304 /* No update for zero timeout */
305 if (!end_time->tv_sec && !end_time->tv_nsec)
309 rts = timespec_sub(*end_time, rts);
311 rts.tv_sec = rts.tv_nsec = 0;
314 if (sizeof(rtv) > sizeof(rtv.tv_sec) + sizeof(rtv.tv_usec))
315 memset(&rtv, 0, sizeof(rtv));
316 rtv.tv_sec = rts.tv_sec;
317 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
319 if (!copy_to_user(p, &rtv, sizeof(rtv)))
322 } else if (!copy_to_user(p, &rts, sizeof(rts)))
326 * If an application puts its timeval in read-only memory, we
327 * don't want the Linux-specific update to the timeval to
328 * cause a fault after the select has completed
329 * successfully. However, because we're not updating the
330 * timeval, we can't restart the system call.
334 if (ret == -ERESTARTNOHAND)
339 #define FDS_IN(fds, n) (fds->in + n)
340 #define FDS_OUT(fds, n) (fds->out + n)
341 #define FDS_EX(fds, n) (fds->ex + n)
343 #define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
345 static int max_select_fd(unsigned long n, fd_set_bits *fds)
347 unsigned long *open_fds;
352 /* handle last in-complete long-word first */
353 set = ~(~0UL << (n & (BITS_PER_LONG-1)));
355 fdt = files_fdtable(current->files);
356 open_fds = fdt->open_fds + n;
361 if (!(set & ~*open_fds))
372 if (set & ~*open_fds)
381 max += n * BITS_PER_LONG;
387 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
388 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
389 #define POLLEX_SET (POLLPRI)
391 static inline void wait_key_set(poll_table *wait, unsigned long in,
392 unsigned long out, unsigned long bit)
394 wait->_key = POLLEX_SET;
396 wait->_key |= POLLIN_SET;
398 wait->_key |= POLLOUT_SET;
401 int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
403 ktime_t expire, *to = NULL;
404 struct poll_wqueues table;
406 int retval, i, timed_out = 0;
407 unsigned long slack = 0;
410 retval = max_select_fd(n, fds);
417 poll_initwait(&table);
419 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
424 if (end_time && !timed_out)
425 slack = select_estimate_accuracy(end_time);
429 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
431 inp = fds->in; outp = fds->out; exp = fds->ex;
432 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
434 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
435 unsigned long in, out, ex, all_bits, bit = 1, mask, j;
436 unsigned long res_in = 0, res_out = 0, res_ex = 0;
438 in = *inp++; out = *outp++; ex = *exp++;
439 all_bits = in | out | ex;
445 for (j = 0; j < BITS_PER_LONG; ++j, ++i, bit <<= 1) {
449 if (!(bit & all_bits))
453 const struct file_operations *f_op;
455 mask = DEFAULT_POLLMASK;
456 if (f_op && f_op->poll) {
457 wait_key_set(wait, in, out, bit);
458 mask = (*f_op->poll)(f.file, wait);
461 if ((mask & POLLIN_SET) && (in & bit)) {
466 if ((mask & POLLOUT_SET) && (out & bit)) {
471 if ((mask & POLLEX_SET) && (ex & bit)) {
487 if (retval || timed_out || signal_pending(current))
490 retval = table.error;
495 * If this is the first loop and we have a timeout
496 * given, then we convert to ktime_t and set the to
497 * pointer to the expiry value.
499 if (end_time && !to) {
500 expire = timespec_to_ktime(*end_time);
504 if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
509 poll_freewait(&table);
515 * We can actually return ERESTARTSYS instead of EINTR, but I'd
516 * like to be certain this leads to no problems. So I return
517 * EINTR just for safety.
519 * Update: ERESTARTSYS breaks at least the xview clock binary, so
520 * I'm trying ERESTARTNOHAND which restart only when you want to.
522 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
523 fd_set __user *exp, struct timespec *end_time)
530 /* Allocate small arguments on the stack to save memory and be faster */
531 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
537 /* max_fds can increase, so grab it once to avoid race */
539 fdt = files_fdtable(current->files);
540 max_fds = fdt->max_fds;
546 * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
547 * since we used fdset we need to allocate memory in units of
552 if (size > sizeof(stack_fds) / 6) {
553 /* Not enough space in on-stack array; must use kmalloc */
555 bits = kmalloc(6 * size, GFP_KERNEL);
560 fds.out = bits + size;
561 fds.ex = bits + 2*size;
562 fds.res_in = bits + 3*size;
563 fds.res_out = bits + 4*size;
564 fds.res_ex = bits + 5*size;
566 if ((ret = get_fd_set(n, inp, fds.in)) ||
567 (ret = get_fd_set(n, outp, fds.out)) ||
568 (ret = get_fd_set(n, exp, fds.ex)))
570 zero_fd_set(n, fds.res_in);
571 zero_fd_set(n, fds.res_out);
572 zero_fd_set(n, fds.res_ex);
574 ret = do_select(n, &fds, end_time);
579 ret = -ERESTARTNOHAND;
580 if (signal_pending(current))
585 if (set_fd_set(n, inp, fds.res_in) ||
586 set_fd_set(n, outp, fds.res_out) ||
587 set_fd_set(n, exp, fds.res_ex))
591 if (bits != stack_fds)
597 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
598 fd_set __user *, exp, struct timeval __user *, tvp)
600 struct timespec end_time, *to = NULL;
605 if (copy_from_user(&tv, tvp, sizeof(tv)))
609 if (poll_select_set_timeout(to,
610 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
611 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
615 ret = core_sys_select(n, inp, outp, exp, to);
616 ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
621 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
622 fd_set __user *exp, struct timespec __user *tsp,
623 const sigset_t __user *sigmask, size_t sigsetsize)
625 sigset_t ksigmask, sigsaved;
626 struct timespec ts, end_time, *to = NULL;
630 if (copy_from_user(&ts, tsp, sizeof(ts)))
634 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
639 /* XXX: Don't preclude handling different sized sigset_t's. */
640 if (sigsetsize != sizeof(sigset_t))
642 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
645 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
646 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
649 ret = core_sys_select(n, inp, outp, exp, to);
650 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
652 if (ret == -ERESTARTNOHAND) {
654 * Don't restore the signal mask yet. Let do_signal() deliver
655 * the signal on the way back to userspace, before the signal
659 memcpy(¤t->saved_sigmask, &sigsaved,
661 set_restore_sigmask();
664 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
670 * Most architectures can't handle 7-argument syscalls. So we provide a
671 * 6-argument version where the sixth argument is a pointer to a structure
672 * which has a pointer to the sigset_t itself followed by a size_t containing
675 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
676 fd_set __user *, exp, struct timespec __user *, tsp,
679 size_t sigsetsize = 0;
680 sigset_t __user *up = NULL;
683 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
684 || __get_user(up, (sigset_t __user * __user *)sig)
685 || __get_user(sigsetsize,
686 (size_t __user *)(sig+sizeof(void *))))
690 return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
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,
723 * if pwait->_qproc is non-NULL.
725 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
733 struct fd f = fdget(fd);
736 mask = DEFAULT_POLLMASK;
737 if (f.file->f_op && f.file->f_op->poll) {
738 pwait->_key = pollfd->events|POLLERR|POLLHUP;
739 mask = f.file->f_op->poll(f.file, pwait);
741 /* Mask out unneeded events. */
742 mask &= pollfd->events | POLLERR | POLLHUP;
746 pollfd->revents = mask;
751 static int do_poll(unsigned int nfds, struct poll_list *list,
752 struct poll_wqueues *wait, struct timespec *end_time)
754 poll_table* pt = &wait->pt;
755 ktime_t expire, *to = NULL;
756 int timed_out = 0, count = 0;
757 unsigned long slack = 0;
759 /* Optimise the no-wait case */
760 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
765 if (end_time && !timed_out)
766 slack = select_estimate_accuracy(end_time);
769 struct poll_list *walk;
771 for (walk = list; walk != NULL; walk = walk->next) {
772 struct pollfd * pfd, * pfd_end;
775 pfd_end = pfd + walk->len;
776 for (; pfd != pfd_end; pfd++) {
778 * Fish for events. If we found one, record it
779 * and kill poll_table->_qproc, so we don't
780 * needlessly register any other waiters after
781 * this. They'll get immediately deregistered
782 * when we break out and return.
784 if (do_pollfd(pfd, pt)) {
791 * All waiters have already been registered, so don't provide
792 * a poll_table->_qproc to them on the next loop iteration.
797 if (signal_pending(current))
800 if (count || timed_out)
804 * If this is the first loop and we have a timeout
805 * given, then we convert to ktime_t and set the to
806 * pointer to the expiry value.
808 if (end_time && !to) {
809 expire = timespec_to_ktime(*end_time);
813 if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
819 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \
820 sizeof(struct pollfd))
822 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
823 struct timespec *end_time)
825 struct poll_wqueues table;
826 int err = -EFAULT, fdcount, len, size;
827 /* Allocate small arguments on the stack to save memory and be
828 faster - use long to make sure the buffer is aligned properly
829 on 64 bit archs to avoid unaligned access */
830 long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
831 struct poll_list *const head = (struct poll_list *)stack_pps;
832 struct poll_list *walk = head;
833 unsigned long todo = nfds;
835 if (nfds > rlimit(RLIMIT_NOFILE))
838 len = min_t(unsigned int, nfds, N_STACK_PPS);
845 if (copy_from_user(walk->entries, ufds + nfds-todo,
846 sizeof(struct pollfd) * walk->len))
853 len = min(todo, POLLFD_PER_PAGE);
854 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
855 walk = walk->next = kmalloc(size, GFP_KERNEL);
862 poll_initwait(&table);
863 fdcount = do_poll(nfds, head, &table, end_time);
864 poll_freewait(&table);
866 for (walk = head; walk; walk = walk->next) {
867 struct pollfd *fds = walk->entries;
870 for (j = 0; j < walk->len; j++, ufds++)
871 if (__put_user(fds[j].revents, &ufds->revents))
879 struct poll_list *pos = walk;
887 static long do_restart_poll(struct restart_block *restart_block)
889 struct pollfd __user *ufds = restart_block->poll.ufds;
890 int nfds = restart_block->poll.nfds;
891 struct timespec *to = NULL, end_time;
894 if (restart_block->poll.has_timeout) {
895 end_time.tv_sec = restart_block->poll.tv_sec;
896 end_time.tv_nsec = restart_block->poll.tv_nsec;
900 ret = do_sys_poll(ufds, nfds, to);
903 restart_block->fn = do_restart_poll;
904 ret = -ERESTART_RESTARTBLOCK;
909 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
912 struct timespec end_time, *to = NULL;
915 if (timeout_msecs >= 0) {
917 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
918 NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
921 ret = do_sys_poll(ufds, nfds, to);
924 struct restart_block *restart_block;
926 restart_block = ¤t_thread_info()->restart_block;
927 restart_block->fn = do_restart_poll;
928 restart_block->poll.ufds = ufds;
929 restart_block->poll.nfds = nfds;
931 if (timeout_msecs >= 0) {
932 restart_block->poll.tv_sec = end_time.tv_sec;
933 restart_block->poll.tv_nsec = end_time.tv_nsec;
934 restart_block->poll.has_timeout = 1;
936 restart_block->poll.has_timeout = 0;
938 ret = -ERESTART_RESTARTBLOCK;
943 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
944 struct timespec __user *, tsp, const sigset_t __user *, sigmask,
947 sigset_t ksigmask, sigsaved;
948 struct timespec ts, end_time, *to = NULL;
952 if (copy_from_user(&ts, tsp, sizeof(ts)))
956 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
961 /* XXX: Don't preclude handling different sized sigset_t's. */
962 if (sigsetsize != sizeof(sigset_t))
964 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
967 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
968 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
971 ret = do_sys_poll(ufds, nfds, to);
973 /* We can restart this syscall, usually */
976 * Don't restore the signal mask yet. Let do_signal() deliver
977 * the signal on the way back to userspace, before the signal
981 memcpy(¤t->saved_sigmask, &sigsaved,
983 set_restore_sigmask();
985 ret = -ERESTARTNOHAND;
987 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
989 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);