linux/fs/select.c
<<
>>
Prefs
   1/*
   2 * This file contains the procedures for the handling of select and poll
   3 *
   4 * Created for Linux based loosely upon Mathius Lattner's minix
   5 * patches by Peter MacDonald. Heavily edited by Linus.
   6 *
   7 *  4 February 1994
   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.
  11 *
  12 *  24 January 2000
  13 *     Changed sys_poll()/do_poll() to use PAGE_SIZE chunk-based allocation 
  14 *     of fds to overcome nfds < 16390 descriptors limit (Tigran Aivazian).
  15 */
  16
  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>
  25#include <linux/fs.h>
  26#include <linux/rcupdate.h>
  27#include <linux/hrtimer.h>
  28
  29#include <asm/uaccess.h>
  30
  31
  32/*
  33 * Estimate expected accuracy in ns from a timeval.
  34 *
  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.
  39 *
  40 * Consider this comment an open invitation to come up with even
  41 * better solutions..
  42 */
  43
  44static long __estimate_accuracy(struct timespec *tv)
  45{
  46        long slack;
  47        int divfactor = 1000;
  48
  49        if (task_nice(current) > 0)
  50                divfactor = divfactor / 5;
  51
  52        slack = tv->tv_nsec / divfactor;
  53        slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
  54
  55        if (slack > 100 * NSEC_PER_MSEC)
  56                slack =  100 * NSEC_PER_MSEC;
  57
  58        if (slack < 0)
  59                slack = 0;
  60        return slack;
  61}
  62
  63static long estimate_accuracy(struct timespec *tv)
  64{
  65        unsigned long ret;
  66        struct timespec now;
  67
  68        /*
  69         * Realtime tasks get a slack of 0 for obvious reasons.
  70         */
  71
  72        if (rt_task(current))
  73                return 0;
  74
  75        ktime_get_ts(&now);
  76        now = timespec_sub(*tv, now);
  77        ret = __estimate_accuracy(&now);
  78        if (ret < current->timer_slack_ns)
  79                return current->timer_slack_ns;
  80        return ret;
  81}
  82
  83
  84
  85struct poll_table_page {
  86        struct poll_table_page * next;
  87        struct poll_table_entry * entry;
  88        struct poll_table_entry entries[0];
  89};
  90
  91#define POLL_TABLE_FULL(table) \
  92        ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
  93
  94/*
  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.
 100 *
 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
 104 * poll table.
 105 */
 106static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
 107                       poll_table *p);
 108
 109void poll_initwait(struct poll_wqueues *pwq)
 110{
 111        init_poll_funcptr(&pwq->pt, __pollwait);
 112        pwq->error = 0;
 113        pwq->table = NULL;
 114        pwq->inline_index = 0;
 115}
 116
 117EXPORT_SYMBOL(poll_initwait);
 118
 119static void free_poll_entry(struct poll_table_entry *entry)
 120{
 121        remove_wait_queue(entry->wait_address, &entry->wait);
 122        fput(entry->filp);
 123}
 124
 125void poll_freewait(struct poll_wqueues *pwq)
 126{
 127        struct poll_table_page * p = pwq->table;
 128        int i;
 129        for (i = 0; i < pwq->inline_index; i++)
 130                free_poll_entry(pwq->inline_entries + i);
 131        while (p) {
 132                struct poll_table_entry * entry;
 133                struct poll_table_page *old;
 134
 135                entry = p->entry;
 136                do {
 137                        entry--;
 138                        free_poll_entry(entry);
 139                } while (entry > p->entries);
 140                old = p;
 141                p = p->next;
 142                free_page((unsigned long) old);
 143        }
 144}
 145
 146EXPORT_SYMBOL(poll_freewait);
 147
 148static struct poll_table_entry *poll_get_entry(poll_table *_p)
 149{
 150        struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt);
 151        struct poll_table_page *table = p->table;
 152
 153        if (p->inline_index < N_INLINE_POLL_ENTRIES)
 154                return p->inline_entries + p->inline_index++;
 155
 156        if (!table || POLL_TABLE_FULL(table)) {
 157                struct poll_table_page *new_table;
 158
 159                new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
 160                if (!new_table) {
 161                        p->error = -ENOMEM;
 162                        __set_current_state(TASK_RUNNING);
 163                        return NULL;
 164                }
 165                new_table->entry = new_table->entries;
 166                new_table->next = table;
 167                p->table = new_table;
 168                table = new_table;
 169        }
 170
 171        return table->entry++;
 172}
 173
 174/* Add a new entry */
 175static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
 176                                poll_table *p)
 177{
 178        struct poll_table_entry *entry = poll_get_entry(p);
 179        if (!entry)
 180                return;
 181        get_file(filp);
 182        entry->filp = filp;
 183        entry->wait_address = wait_address;
 184        init_waitqueue_entry(&entry->wait, current);
 185        add_wait_queue(wait_address, &entry->wait);
 186}
 187
 188/**
 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)
 193 *
 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.
 196 *
 197 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
 198 */
 199int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
 200{
 201        struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
 202
 203        if (!timespec_valid(&ts))
 204                return -EINVAL;
 205
 206        /* Optimize for the zero timeout value here */
 207        if (!sec && !nsec) {
 208                to->tv_sec = to->tv_nsec = 0;
 209        } else {
 210                ktime_get_ts(to);
 211                *to = timespec_add_safe(*to, ts);
 212        }
 213        return 0;
 214}
 215
 216static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
 217                                      int timeval, int ret)
 218{
 219        struct timespec rts;
 220        struct timeval rtv;
 221
 222        if (!p)
 223                return ret;
 224
 225        if (current->personality & STICKY_TIMEOUTS)
 226                goto sticky;
 227
 228        /* No update for zero timeout */
 229        if (!end_time->tv_sec && !end_time->tv_nsec)
 230                return ret;
 231
 232        ktime_get_ts(&rts);
 233        rts = timespec_sub(*end_time, rts);
 234        if (rts.tv_sec < 0)
 235                rts.tv_sec = rts.tv_nsec = 0;
 236
 237        if (timeval) {
 238                rtv.tv_sec = rts.tv_sec;
 239                rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
 240
 241                if (!copy_to_user(p, &rtv, sizeof(rtv)))
 242                        return ret;
 243
 244        } else if (!copy_to_user(p, &rts, sizeof(rts)))
 245                return ret;
 246
 247        /*
 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.
 253         */
 254
 255sticky:
 256        if (ret == -ERESTARTNOHAND)
 257                ret = -EINTR;
 258        return ret;
 259}
 260
 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)
 264
 265#define BITS(fds, n)    (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
 266
 267static int max_select_fd(unsigned long n, fd_set_bits *fds)
 268{
 269        unsigned long *open_fds;
 270        unsigned long set;
 271        int max;
 272        struct fdtable *fdt;
 273
 274        /* handle last in-complete long-word first */
 275        set = ~(~0UL << (n & (__NFDBITS-1)));
 276        n /= __NFDBITS;
 277        fdt = files_fdtable(current->files);
 278        open_fds = fdt->open_fds->fds_bits+n;
 279        max = 0;
 280        if (set) {
 281                set &= BITS(fds, n);
 282                if (set) {
 283                        if (!(set & ~*open_fds))
 284                                goto get_max;
 285                        return -EBADF;
 286                }
 287        }
 288        while (n) {
 289                open_fds--;
 290                n--;
 291                set = BITS(fds, n);
 292                if (!set)
 293                        continue;
 294                if (set & ~*open_fds)
 295                        return -EBADF;
 296                if (max)
 297                        continue;
 298get_max:
 299                do {
 300                        max++;
 301                        set >>= 1;
 302                } while (set);
 303                max += n * __NFDBITS;
 304        }
 305
 306        return max;
 307}
 308
 309#define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
 310#define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
 311#define POLLEX_SET (POLLPRI)
 312
 313int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
 314{
 315        ktime_t expire, *to = NULL;
 316        struct poll_wqueues table;
 317        poll_table *wait;
 318        int retval, i, timed_out = 0;
 319        unsigned long slack = 0;
 320
 321        rcu_read_lock();
 322        retval = max_select_fd(n, fds);
 323        rcu_read_unlock();
 324
 325        if (retval < 0)
 326                return retval;
 327        n = retval;
 328
 329        poll_initwait(&table);
 330        wait = &table.pt;
 331        if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
 332                wait = NULL;
 333                timed_out = 1;
 334        }
 335
 336        if (end_time && !timed_out)
 337                slack = estimate_accuracy(end_time);
 338
 339        retval = 0;
 340        for (;;) {
 341                unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
 342
 343                set_current_state(TASK_INTERRUPTIBLE);
 344
 345                inp = fds->in; outp = fds->out; exp = fds->ex;
 346                rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
 347
 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;
 353
 354                        in = *inp++; out = *outp++; ex = *exp++;
 355                        all_bits = in | out | ex;
 356                        if (all_bits == 0) {
 357                                i += __NFDBITS;
 358                                continue;
 359                        }
 360
 361                        for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
 362                                int fput_needed;
 363                                if (i >= n)
 364                                        break;
 365                                if (!(bit & all_bits))
 366                                        continue;
 367                                file = fget_light(i, &fput_needed);
 368                                if (file) {
 369                                        f_op = file->f_op;
 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)) {
 375                                                res_in |= bit;
 376                                                retval++;
 377                                        }
 378                                        if ((mask & POLLOUT_SET) && (out & bit)) {
 379                                                res_out |= bit;
 380                                                retval++;
 381                                        }
 382                                        if ((mask & POLLEX_SET) && (ex & bit)) {
 383                                                res_ex |= bit;
 384                                                retval++;
 385                                        }
 386                                }
 387                        }
 388                        if (res_in)
 389                                *rinp = res_in;
 390                        if (res_out)
 391                                *routp = res_out;
 392                        if (res_ex)
 393                                *rexp = res_ex;
 394                        cond_resched();
 395                }
 396                wait = NULL;
 397                if (retval || timed_out || signal_pending(current))
 398                        break;
 399                if (table.error) {
 400                        retval = table.error;
 401                        break;
 402                }
 403
 404                /*
 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.
 408                 */
 409                if (end_time && !to) {
 410                        expire = timespec_to_ktime(*end_time);
 411                        to = &expire;
 412                }
 413
 414                if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
 415                        timed_out = 1;
 416        }
 417        __set_current_state(TASK_RUNNING);
 418
 419        poll_freewait(&table);
 420
 421        return retval;
 422}
 423
 424/*
 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.
 428 *
 429 * Update: ERESTARTSYS breaks at least the xview clock binary, so
 430 * I'm trying ERESTARTNOHAND which restart only when you want to.
 431 */
 432#define MAX_SELECT_SECONDS \
 433        ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
 434
 435int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 436                           fd_set __user *exp, struct timespec *end_time)
 437{
 438        fd_set_bits fds;
 439        void *bits;
 440        int ret, max_fds;
 441        unsigned int size;
 442        struct fdtable *fdt;
 443        /* Allocate small arguments on the stack to save memory and be faster */
 444        long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
 445
 446        ret = -EINVAL;
 447        if (n < 0)
 448                goto out_nofds;
 449
 450        /* max_fds can increase, so grab it once to avoid race */
 451        rcu_read_lock();
 452        fdt = files_fdtable(current->files);
 453        max_fds = fdt->max_fds;
 454        rcu_read_unlock();
 455        if (n > max_fds)
 456                n = max_fds;
 457
 458        /*
 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
 461         * long-words. 
 462         */
 463        size = FDS_BYTES(n);
 464        bits = stack_fds;
 465        if (size > sizeof(stack_fds) / 6) {
 466                /* Not enough space in on-stack array; must use kmalloc */
 467                ret = -ENOMEM;
 468                bits = kmalloc(6 * size, GFP_KERNEL);
 469                if (!bits)
 470                        goto out_nofds;
 471        }
 472        fds.in      = bits;
 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;
 478
 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)))
 482                goto out;
 483        zero_fd_set(n, fds.res_in);
 484        zero_fd_set(n, fds.res_out);
 485        zero_fd_set(n, fds.res_ex);
 486
 487        ret = do_select(n, &fds, end_time);
 488
 489        if (ret < 0)
 490                goto out;
 491        if (!ret) {
 492                ret = -ERESTARTNOHAND;
 493                if (signal_pending(current))
 494                        goto out;
 495                ret = 0;
 496        }
 497
 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))
 501                ret = -EFAULT;
 502
 503out:
 504        if (bits != stack_fds)
 505                kfree(bits);
 506out_nofds:
 507        return ret;
 508}
 509
 510SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
 511                fd_set __user *, exp, struct timeval __user *, tvp)
 512{
 513        struct timespec end_time, *to = NULL;
 514        struct timeval tv;
 515        int ret;
 516
 517        if (tvp) {
 518                if (copy_from_user(&tv, tvp, sizeof(tv)))
 519                        return -EFAULT;
 520
 521                to = &end_time;
 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))
 525                        return -EINVAL;
 526        }
 527
 528        ret = core_sys_select(n, inp, outp, exp, to);
 529        ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
 530
 531        return ret;
 532}
 533
 534#ifdef HAVE_SET_RESTORE_SIGMASK
 535static long do_pselect(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)
 538{
 539        sigset_t ksigmask, sigsaved;
 540        struct timespec ts, end_time, *to = NULL;
 541        int ret;
 542
 543        if (tsp) {
 544                if (copy_from_user(&ts, tsp, sizeof(ts)))
 545                        return -EFAULT;
 546
 547                to = &end_time;
 548                if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
 549                        return -EINVAL;
 550        }
 551
 552        if (sigmask) {
 553                /* XXX: Don't preclude handling different sized sigset_t's.  */
 554                if (sigsetsize != sizeof(sigset_t))
 555                        return -EINVAL;
 556                if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 557                        return -EFAULT;
 558
 559                sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 560                sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 561        }
 562
 563        ret = core_sys_select(n, inp, outp, exp, to);
 564        ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
 565
 566        if (ret == -ERESTARTNOHAND) {
 567                /*
 568                 * Don't restore the signal mask yet. Let do_signal() deliver
 569                 * the signal on the way back to userspace, before the signal
 570                 * mask is restored.
 571                 */
 572                if (sigmask) {
 573                        memcpy(&current->saved_sigmask, &sigsaved,
 574                                        sizeof(sigsaved));
 575                        set_restore_sigmask();
 576                }
 577        } else if (sigmask)
 578                sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 579
 580        return ret;
 581}
 582
 583/*
 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
 587 * the sigset size.
 588 */
 589SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
 590                fd_set __user *, exp, struct timespec __user *, tsp,
 591                void __user *, sig)
 592{
 593        size_t sigsetsize = 0;
 594        sigset_t __user *up = NULL;
 595
 596        if (sig) {
 597                if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
 598                    || __get_user(up, (sigset_t __user * __user *)sig)
 599                    || __get_user(sigsetsize,
 600                                (size_t __user *)(sig+sizeof(void *))))
 601                        return -EFAULT;
 602        }
 603
 604        return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
 605}
 606#endif /* HAVE_SET_RESTORE_SIGMASK */
 607
 608struct poll_list {
 609        struct poll_list *next;
 610        int len;
 611        struct pollfd entries[0];
 612};
 613
 614#define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 615
 616/*
 617 * Fish for pollable events on the pollfd->fd file descriptor. We're only
 618 * interested in events matching the pollfd->events mask, and the result
 619 * matching that mask is both recorded in pollfd->revents and returned. The
 620 * pwait poll_table will be used by the fd-provided poll handler for waiting,
 621 * if non-NULL.
 622 */
 623static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
 624{
 625        unsigned int mask;
 626        int fd;
 627
 628        mask = 0;
 629        fd = pollfd->fd;
 630        if (fd >= 0) {
 631                int fput_needed;
 632                struct file * file;
 633
 634                file = fget_light(fd, &fput_needed);
 635                mask = POLLNVAL;
 636                if (file != NULL) {
 637                        mask = DEFAULT_POLLMASK;
 638                        if (file->f_op && file->f_op->poll)
 639                                mask = file->f_op->poll(file, pwait);
 640                        /* Mask out unneeded events. */
 641                        mask &= pollfd->events | POLLERR | POLLHUP;
 642                        fput_light(file, fput_needed);
 643                }
 644        }
 645        pollfd->revents = mask;
 646
 647        return mask;
 648}
 649
 650static int do_poll(unsigned int nfds,  struct poll_list *list,
 651                   struct poll_wqueues *wait, struct timespec *end_time)
 652{
 653        poll_table* pt = &wait->pt;
 654        ktime_t expire, *to = NULL;
 655        int timed_out = 0, count = 0;
 656        unsigned long slack = 0;
 657
 658        /* Optimise the no-wait case */
 659        if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
 660                pt = NULL;
 661                timed_out = 1;
 662        }
 663
 664        if (end_time && !timed_out)
 665                slack = estimate_accuracy(end_time);
 666
 667        for (;;) {
 668                struct poll_list *walk;
 669
 670                set_current_state(TASK_INTERRUPTIBLE);
 671                for (walk = list; walk != NULL; walk = walk->next) {
 672                        struct pollfd * pfd, * pfd_end;
 673
 674                        pfd = walk->entries;
 675                        pfd_end = pfd + walk->len;
 676                        for (; pfd != pfd_end; pfd++) {
 677                                /*
 678                                 * Fish for events. If we found one, record it
 679                                 * and kill the poll_table, so we don't
 680                                 * needlessly register any other waiters after
 681                                 * this. They'll get immediately deregistered
 682                                 * when we break out and return.
 683                                 */
 684                                if (do_pollfd(pfd, pt)) {
 685                                        count++;
 686                                        pt = NULL;
 687                                }
 688                        }
 689                }
 690                /*
 691                 * All waiters have already been registered, so don't provide
 692                 * a poll_table to them on the next loop iteration.
 693                 */
 694                pt = NULL;
 695                if (!count) {
 696                        count = wait->error;
 697                        if (signal_pending(current))
 698                                count = -EINTR;
 699                }
 700                if (count || timed_out)
 701                        break;
 702
 703                /*
 704                 * If this is the first loop and we have a timeout
 705                 * given, then we convert to ktime_t and set the to
 706                 * pointer to the expiry value.
 707                 */
 708                if (end_time && !to) {
 709                        expire = timespec_to_ktime(*end_time);
 710                        to = &expire;
 711                }
 712
 713                if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
 714                        timed_out = 1;
 715        }
 716        __set_current_state(TASK_RUNNING);
 717        return count;
 718}
 719
 720#define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
 721                        sizeof(struct pollfd))
 722
 723int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
 724                struct timespec *end_time)
 725{
 726        struct poll_wqueues table;
 727        int err = -EFAULT, fdcount, len, size;
 728        /* Allocate small arguments on the stack to save memory and be
 729           faster - use long to make sure the buffer is aligned properly
 730           on 64 bit archs to avoid unaligned access */
 731        long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
 732        struct poll_list *const head = (struct poll_list *)stack_pps;
 733        struct poll_list *walk = head;
 734        unsigned long todo = nfds;
 735
 736        if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
 737                return -EINVAL;
 738
 739        len = min_t(unsigned int, nfds, N_STACK_PPS);
 740        for (;;) {
 741                walk->next = NULL;
 742                walk->len = len;
 743                if (!len)
 744                        break;
 745
 746                if (copy_from_user(walk->entries, ufds + nfds-todo,
 747                                        sizeof(struct pollfd) * walk->len))
 748                        goto out_fds;
 749
 750                todo -= walk->len;
 751                if (!todo)
 752                        break;
 753
 754                len = min(todo, POLLFD_PER_PAGE);
 755                size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
 756                walk = walk->next = kmalloc(size, GFP_KERNEL);
 757                if (!walk) {
 758                        err = -ENOMEM;
 759                        goto out_fds;
 760                }
 761        }
 762
 763        poll_initwait(&table);
 764        fdcount = do_poll(nfds, head, &table, end_time);
 765        poll_freewait(&table);
 766
 767        for (walk = head; walk; walk = walk->next) {
 768                struct pollfd *fds = walk->entries;
 769                int j;
 770
 771                for (j = 0; j < walk->len; j++, ufds++)
 772                        if (__put_user(fds[j].revents, &ufds->revents))
 773                                goto out_fds;
 774        }
 775
 776        err = fdcount;
 777out_fds:
 778        walk = head->next;
 779        while (walk) {
 780                struct poll_list *pos = walk;
 781                walk = walk->next;
 782                kfree(pos);
 783        }
 784
 785        return err;
 786}
 787
 788static long do_restart_poll(struct restart_block *restart_block)
 789{
 790        struct pollfd __user *ufds = restart_block->poll.ufds;
 791        int nfds = restart_block->poll.nfds;
 792        struct timespec *to = NULL, end_time;
 793        int ret;
 794
 795        if (restart_block->poll.has_timeout) {
 796                end_time.tv_sec = restart_block->poll.tv_sec;
 797                end_time.tv_nsec = restart_block->poll.tv_nsec;
 798                to = &end_time;
 799        }
 800
 801        ret = do_sys_poll(ufds, nfds, to);
 802
 803        if (ret == -EINTR) {
 804                restart_block->fn = do_restart_poll;
 805                ret = -ERESTART_RESTARTBLOCK;
 806        }
 807        return ret;
 808}
 809
 810SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
 811                long, timeout_msecs)
 812{
 813        struct timespec end_time, *to = NULL;
 814        int ret;
 815
 816        if (timeout_msecs >= 0) {
 817                to = &end_time;
 818                poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
 819                        NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
 820        }
 821
 822        ret = do_sys_poll(ufds, nfds, to);
 823
 824        if (ret == -EINTR) {
 825                struct restart_block *restart_block;
 826
 827                restart_block = &current_thread_info()->restart_block;
 828                restart_block->fn = do_restart_poll;
 829                restart_block->poll.ufds = ufds;
 830                restart_block->poll.nfds = nfds;
 831
 832                if (timeout_msecs >= 0) {
 833                        restart_block->poll.tv_sec = end_time.tv_sec;
 834                        restart_block->poll.tv_nsec = end_time.tv_nsec;
 835                        restart_block->poll.has_timeout = 1;
 836                } else
 837                        restart_block->poll.has_timeout = 0;
 838
 839                ret = -ERESTART_RESTARTBLOCK;
 840        }
 841        return ret;
 842}
 843
 844#ifdef HAVE_SET_RESTORE_SIGMASK
 845SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
 846                struct timespec __user *, tsp, const sigset_t __user *, sigmask,
 847                size_t, sigsetsize)
 848{
 849        sigset_t ksigmask, sigsaved;
 850        struct timespec ts, end_time, *to = NULL;
 851        int ret;
 852
 853        if (tsp) {
 854                if (copy_from_user(&ts, tsp, sizeof(ts)))
 855                        return -EFAULT;
 856
 857                to = &end_time;
 858                if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
 859                        return -EINVAL;
 860        }
 861
 862        if (sigmask) {
 863                /* XXX: Don't preclude handling different sized sigset_t's.  */
 864                if (sigsetsize != sizeof(sigset_t))
 865                        return -EINVAL;
 866                if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 867                        return -EFAULT;
 868
 869                sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 870                sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 871        }
 872
 873        ret = do_sys_poll(ufds, nfds, to);
 874
 875        /* We can restart this syscall, usually */
 876        if (ret == -EINTR) {
 877                /*
 878                 * Don't restore the signal mask yet. Let do_signal() deliver
 879                 * the signal on the way back to userspace, before the signal
 880                 * mask is restored.
 881                 */
 882                if (sigmask) {
 883                        memcpy(&current->saved_sigmask, &sigsaved,
 884                                        sizeof(sigsaved));
 885                        set_restore_sigmask();
 886                }
 887                ret = -ERESTARTNOHAND;
 888        } else if (sigmask)
 889                sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 890
 891        ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
 892
 893        return ret;
 894}
 895#endif /* HAVE_SET_RESTORE_SIGMASK */
 896