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->polling_task = current;
 113        pwq->triggered = 0;
 114        pwq->error = 0;
 115        pwq->table = NULL;
 116        pwq->inline_index = 0;
 117}
 118EXPORT_SYMBOL(poll_initwait);
 119
 120static void free_poll_entry(struct poll_table_entry *entry)
 121{
 122        remove_wait_queue(entry->wait_address, &entry->wait);
 123        fput(entry->filp);
 124}
 125
 126void poll_freewait(struct poll_wqueues *pwq)
 127{
 128        struct poll_table_page * p = pwq->table;
 129        int i;
 130        for (i = 0; i < pwq->inline_index; i++)
 131                free_poll_entry(pwq->inline_entries + i);
 132        while (p) {
 133                struct poll_table_entry * entry;
 134                struct poll_table_page *old;
 135
 136                entry = p->entry;
 137                do {
 138                        entry--;
 139                        free_poll_entry(entry);
 140                } while (entry > p->entries);
 141                old = p;
 142                p = p->next;
 143                free_page((unsigned long) old);
 144        }
 145}
 146EXPORT_SYMBOL(poll_freewait);
 147
 148static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
 149{
 150        struct poll_table_page *table = p->table;
 151
 152        if (p->inline_index < N_INLINE_POLL_ENTRIES)
 153                return p->inline_entries + p->inline_index++;
 154
 155        if (!table || POLL_TABLE_FULL(table)) {
 156                struct poll_table_page *new_table;
 157
 158                new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
 159                if (!new_table) {
 160                        p->error = -ENOMEM;
 161                        return NULL;
 162                }
 163                new_table->entry = new_table->entries;
 164                new_table->next = table;
 165                p->table = new_table;
 166                table = new_table;
 167        }
 168
 169        return table->entry++;
 170}
 171
 172static int pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
 173{
 174        struct poll_wqueues *pwq = wait->private;
 175        DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
 176
 177        /*
 178         * Although this function is called under waitqueue lock, LOCK
 179         * doesn't imply write barrier and the users expect write
 180         * barrier semantics on wakeup functions.  The following
 181         * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
 182         * and is paired with set_mb() in poll_schedule_timeout.
 183         */
 184        smp_wmb();
 185        pwq->triggered = 1;
 186
 187        /*
 188         * Perform the default wake up operation using a dummy
 189         * waitqueue.
 190         *
 191         * TODO: This is hacky but there currently is no interface to
 192         * pass in @sync.  @sync is scheduled to be removed and once
 193         * that happens, wake_up_process() can be used directly.
 194         */
 195        return default_wake_function(&dummy_wait, mode, sync, key);
 196}
 197
 198/* Add a new entry */
 199static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
 200                                poll_table *p)
 201{
 202        struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
 203        struct poll_table_entry *entry = poll_get_entry(pwq);
 204        if (!entry)
 205                return;
 206        get_file(filp);
 207        entry->filp = filp;
 208        entry->wait_address = wait_address;
 209        init_waitqueue_func_entry(&entry->wait, pollwake);
 210        entry->wait.private = pwq;
 211        add_wait_queue(wait_address, &entry->wait);
 212}
 213
 214int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
 215                          ktime_t *expires, unsigned long slack)
 216{
 217        int rc = -EINTR;
 218
 219        set_current_state(state);
 220        if (!pwq->triggered)
 221                rc = schedule_hrtimeout_range(expires, slack, HRTIMER_MODE_ABS);
 222        __set_current_state(TASK_RUNNING);
 223
 224        /*
 225         * Prepare for the next iteration.
 226         *
 227         * The following set_mb() serves two purposes.  First, it's
 228         * the counterpart rmb of the wmb in pollwake() such that data
 229         * written before wake up is always visible after wake up.
 230         * Second, the full barrier guarantees that triggered clearing
 231         * doesn't pass event check of the next iteration.  Note that
 232         * this problem doesn't exist for the first iteration as
 233         * add_wait_queue() has full barrier semantics.
 234         */
 235        set_mb(pwq->triggered, 0);
 236
 237        return rc;
 238}
 239EXPORT_SYMBOL(poll_schedule_timeout);
 240
 241/**
 242 * poll_select_set_timeout - helper function to setup the timeout value
 243 * @to:         pointer to timespec variable for the final timeout
 244 * @sec:        seconds (from user space)
 245 * @nsec:       nanoseconds (from user space)
 246 *
 247 * Note, we do not use a timespec for the user space value here, That
 248 * way we can use the function for timeval and compat interfaces as well.
 249 *
 250 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
 251 */
 252int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
 253{
 254        struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
 255
 256        if (!timespec_valid(&ts))
 257                return -EINVAL;
 258
 259        /* Optimize for the zero timeout value here */
 260        if (!sec && !nsec) {
 261                to->tv_sec = to->tv_nsec = 0;
 262        } else {
 263                ktime_get_ts(to);
 264                *to = timespec_add_safe(*to, ts);
 265        }
 266        return 0;
 267}
 268
 269static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
 270                                      int timeval, int ret)
 271{
 272        struct timespec rts;
 273        struct timeval rtv;
 274
 275        if (!p)
 276                return ret;
 277
 278        if (current->personality & STICKY_TIMEOUTS)
 279                goto sticky;
 280
 281        /* No update for zero timeout */
 282        if (!end_time->tv_sec && !end_time->tv_nsec)
 283                return ret;
 284
 285        ktime_get_ts(&rts);
 286        rts = timespec_sub(*end_time, rts);
 287        if (rts.tv_sec < 0)
 288                rts.tv_sec = rts.tv_nsec = 0;
 289
 290        if (timeval) {
 291                rtv.tv_sec = rts.tv_sec;
 292                rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
 293
 294                if (!copy_to_user(p, &rtv, sizeof(rtv)))
 295                        return ret;
 296
 297        } else if (!copy_to_user(p, &rts, sizeof(rts)))
 298                return ret;
 299
 300        /*
 301         * If an application puts its timeval in read-only memory, we
 302         * don't want the Linux-specific update to the timeval to
 303         * cause a fault after the select has completed
 304         * successfully. However, because we're not updating the
 305         * timeval, we can't restart the system call.
 306         */
 307
 308sticky:
 309        if (ret == -ERESTARTNOHAND)
 310                ret = -EINTR;
 311        return ret;
 312}
 313
 314#define FDS_IN(fds, n)          (fds->in + n)
 315#define FDS_OUT(fds, n)         (fds->out + n)
 316#define FDS_EX(fds, n)          (fds->ex + n)
 317
 318#define BITS(fds, n)    (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
 319
 320static int max_select_fd(unsigned long n, fd_set_bits *fds)
 321{
 322        unsigned long *open_fds;
 323        unsigned long set;
 324        int max;
 325        struct fdtable *fdt;
 326
 327        /* handle last in-complete long-word first */
 328        set = ~(~0UL << (n & (__NFDBITS-1)));
 329        n /= __NFDBITS;
 330        fdt = files_fdtable(current->files);
 331        open_fds = fdt->open_fds->fds_bits+n;
 332        max = 0;
 333        if (set) {
 334                set &= BITS(fds, n);
 335                if (set) {
 336                        if (!(set & ~*open_fds))
 337                                goto get_max;
 338                        return -EBADF;
 339                }
 340        }
 341        while (n) {
 342                open_fds--;
 343                n--;
 344                set = BITS(fds, n);
 345                if (!set)
 346                        continue;
 347                if (set & ~*open_fds)
 348                        return -EBADF;
 349                if (max)
 350                        continue;
 351get_max:
 352                do {
 353                        max++;
 354                        set >>= 1;
 355                } while (set);
 356                max += n * __NFDBITS;
 357        }
 358
 359        return max;
 360}
 361
 362#define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
 363#define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
 364#define POLLEX_SET (POLLPRI)
 365
 366int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
 367{
 368        ktime_t expire, *to = NULL;
 369        struct poll_wqueues table;
 370        poll_table *wait;
 371        int retval, i, timed_out = 0;
 372        unsigned long slack = 0;
 373
 374        rcu_read_lock();
 375        retval = max_select_fd(n, fds);
 376        rcu_read_unlock();
 377
 378        if (retval < 0)
 379                return retval;
 380        n = retval;
 381
 382        poll_initwait(&table);
 383        wait = &table.pt;
 384        if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
 385                wait = NULL;
 386                timed_out = 1;
 387        }
 388
 389        if (end_time && !timed_out)
 390                slack = estimate_accuracy(end_time);
 391
 392        retval = 0;
 393        for (;;) {
 394                unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
 395
 396                inp = fds->in; outp = fds->out; exp = fds->ex;
 397                rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
 398
 399                for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
 400                        unsigned long in, out, ex, all_bits, bit = 1, mask, j;
 401                        unsigned long res_in = 0, res_out = 0, res_ex = 0;
 402                        const struct file_operations *f_op = NULL;
 403                        struct file *file = NULL;
 404
 405                        in = *inp++; out = *outp++; ex = *exp++;
 406                        all_bits = in | out | ex;
 407                        if (all_bits == 0) {
 408                                i += __NFDBITS;
 409                                continue;
 410                        }
 411
 412                        for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
 413                                int fput_needed;
 414                                if (i >= n)
 415                                        break;
 416                                if (!(bit & all_bits))
 417                                        continue;
 418                                file = fget_light(i, &fput_needed);
 419                                if (file) {
 420                                        f_op = file->f_op;
 421                                        mask = DEFAULT_POLLMASK;
 422                                        if (f_op && f_op->poll)
 423                                                mask = (*f_op->poll)(file, retval ? NULL : wait);
 424                                        fput_light(file, fput_needed);
 425                                        if ((mask & POLLIN_SET) && (in & bit)) {
 426                                                res_in |= bit;
 427                                                retval++;
 428                                        }
 429                                        if ((mask & POLLOUT_SET) && (out & bit)) {
 430                                                res_out |= bit;
 431                                                retval++;
 432                                        }
 433                                        if ((mask & POLLEX_SET) && (ex & bit)) {
 434                                                res_ex |= bit;
 435                                                retval++;
 436                                        }
 437                                }
 438                        }
 439                        if (res_in)
 440                                *rinp = res_in;
 441                        if (res_out)
 442                                *routp = res_out;
 443                        if (res_ex)
 444                                *rexp = res_ex;
 445                        cond_resched();
 446                }
 447                wait = NULL;
 448                if (retval || timed_out || signal_pending(current))
 449                        break;
 450                if (table.error) {
 451                        retval = table.error;
 452                        break;
 453                }
 454
 455                /*
 456                 * If this is the first loop and we have a timeout
 457                 * given, then we convert to ktime_t and set the to
 458                 * pointer to the expiry value.
 459                 */
 460                if (end_time && !to) {
 461                        expire = timespec_to_ktime(*end_time);
 462                        to = &expire;
 463                }
 464
 465                if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
 466                                           to, slack))
 467                        timed_out = 1;
 468        }
 469
 470        poll_freewait(&table);
 471
 472        return retval;
 473}
 474
 475/*
 476 * We can actually return ERESTARTSYS instead of EINTR, but I'd
 477 * like to be certain this leads to no problems. So I return
 478 * EINTR just for safety.
 479 *
 480 * Update: ERESTARTSYS breaks at least the xview clock binary, so
 481 * I'm trying ERESTARTNOHAND which restart only when you want to.
 482 */
 483#define MAX_SELECT_SECONDS \
 484        ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
 485
 486int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 487                           fd_set __user *exp, struct timespec *end_time)
 488{
 489        fd_set_bits fds;
 490        void *bits;
 491        int ret, max_fds;
 492        unsigned int size;
 493        struct fdtable *fdt;
 494        /* Allocate small arguments on the stack to save memory and be faster */
 495        long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
 496
 497        ret = -EINVAL;
 498        if (n < 0)
 499                goto out_nofds;
 500
 501        /* max_fds can increase, so grab it once to avoid race */
 502        rcu_read_lock();
 503        fdt = files_fdtable(current->files);
 504        max_fds = fdt->max_fds;
 505        rcu_read_unlock();
 506        if (n > max_fds)
 507                n = max_fds;
 508
 509        /*
 510         * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
 511         * since we used fdset we need to allocate memory in units of
 512         * long-words. 
 513         */
 514        size = FDS_BYTES(n);
 515        bits = stack_fds;
 516        if (size > sizeof(stack_fds) / 6) {
 517                /* Not enough space in on-stack array; must use kmalloc */
 518                ret = -ENOMEM;
 519                bits = kmalloc(6 * size, GFP_KERNEL);
 520                if (!bits)
 521                        goto out_nofds;
 522        }
 523        fds.in      = bits;
 524        fds.out     = bits +   size;
 525        fds.ex      = bits + 2*size;
 526        fds.res_in  = bits + 3*size;
 527        fds.res_out = bits + 4*size;
 528        fds.res_ex  = bits + 5*size;
 529
 530        if ((ret = get_fd_set(n, inp, fds.in)) ||
 531            (ret = get_fd_set(n, outp, fds.out)) ||
 532            (ret = get_fd_set(n, exp, fds.ex)))
 533                goto out;
 534        zero_fd_set(n, fds.res_in);
 535        zero_fd_set(n, fds.res_out);
 536        zero_fd_set(n, fds.res_ex);
 537
 538        ret = do_select(n, &fds, end_time);
 539
 540        if (ret < 0)
 541                goto out;
 542        if (!ret) {
 543                ret = -ERESTARTNOHAND;
 544                if (signal_pending(current))
 545                        goto out;
 546                ret = 0;
 547        }
 548
 549        if (set_fd_set(n, inp, fds.res_in) ||
 550            set_fd_set(n, outp, fds.res_out) ||
 551            set_fd_set(n, exp, fds.res_ex))
 552                ret = -EFAULT;
 553
 554out:
 555        if (bits != stack_fds)
 556                kfree(bits);
 557out_nofds:
 558        return ret;
 559}
 560
 561SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
 562                fd_set __user *, exp, struct timeval __user *, tvp)
 563{
 564        struct timespec end_time, *to = NULL;
 565        struct timeval tv;
 566        int ret;
 567
 568        if (tvp) {
 569                if (copy_from_user(&tv, tvp, sizeof(tv)))
 570                        return -EFAULT;
 571
 572                to = &end_time;
 573                if (poll_select_set_timeout(to,
 574                                tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
 575                                (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
 576                        return -EINVAL;
 577        }
 578
 579        ret = core_sys_select(n, inp, outp, exp, to);
 580        ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
 581
 582        return ret;
 583}
 584
 585#ifdef HAVE_SET_RESTORE_SIGMASK
 586static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
 587                       fd_set __user *exp, struct timespec __user *tsp,
 588                       const sigset_t __user *sigmask, size_t sigsetsize)
 589{
 590        sigset_t ksigmask, sigsaved;
 591        struct timespec ts, end_time, *to = NULL;
 592        int ret;
 593
 594        if (tsp) {
 595                if (copy_from_user(&ts, tsp, sizeof(ts)))
 596                        return -EFAULT;
 597
 598                to = &end_time;
 599                if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
 600                        return -EINVAL;
 601        }
 602
 603        if (sigmask) {
 604                /* XXX: Don't preclude handling different sized sigset_t's.  */
 605                if (sigsetsize != sizeof(sigset_t))
 606                        return -EINVAL;
 607                if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 608                        return -EFAULT;
 609
 610                sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 611                sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 612        }
 613
 614        ret = core_sys_select(n, inp, outp, exp, to);
 615        ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
 616
 617        if (ret == -ERESTARTNOHAND) {
 618                /*
 619                 * Don't restore the signal mask yet. Let do_signal() deliver
 620                 * the signal on the way back to userspace, before the signal
 621                 * mask is restored.
 622                 */
 623                if (sigmask) {
 624                        memcpy(&current->saved_sigmask, &sigsaved,
 625                                        sizeof(sigsaved));
 626                        set_restore_sigmask();
 627                }
 628        } else if (sigmask)
 629                sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 630
 631        return ret;
 632}
 633
 634/*
 635 * Most architectures can't handle 7-argument syscalls. So we provide a
 636 * 6-argument version where the sixth argument is a pointer to a structure
 637 * which has a pointer to the sigset_t itself followed by a size_t containing
 638 * the sigset size.
 639 */
 640SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
 641                fd_set __user *, exp, struct timespec __user *, tsp,
 642                void __user *, sig)
 643{
 644        size_t sigsetsize = 0;
 645        sigset_t __user *up = NULL;
 646
 647        if (sig) {
 648                if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
 649                    || __get_user(up, (sigset_t __user * __user *)sig)
 650                    || __get_user(sigsetsize,
 651                                (size_t __user *)(sig+sizeof(void *))))
 652                        return -EFAULT;
 653        }
 654
 655        return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
 656}
 657#endif /* HAVE_SET_RESTORE_SIGMASK */
 658
 659struct poll_list {
 660        struct poll_list *next;
 661        int len;
 662        struct pollfd entries[0];
 663};
 664
 665#define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 666
 667/*
 668 * Fish for pollable events on the pollfd->fd file descriptor. We're only
 669 * interested in events matching the pollfd->events mask, and the result
 670 * matching that mask is both recorded in pollfd->revents and returned. The
 671 * pwait poll_table will be used by the fd-provided poll handler for waiting,
 672 * if non-NULL.
 673 */
 674static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
 675{
 676        unsigned int mask;
 677        int fd;
 678
 679        mask = 0;
 680        fd = pollfd->fd;
 681        if (fd >= 0) {
 682                int fput_needed;
 683                struct file * file;
 684
 685                file = fget_light(fd, &fput_needed);
 686                mask = POLLNVAL;
 687                if (file != NULL) {
 688                        mask = DEFAULT_POLLMASK;
 689                        if (file->f_op && file->f_op->poll)
 690                                mask = file->f_op->poll(file, pwait);
 691                        /* Mask out unneeded events. */
 692                        mask &= pollfd->events | POLLERR | POLLHUP;
 693                        fput_light(file, fput_needed);
 694                }
 695        }
 696        pollfd->revents = mask;
 697
 698        return mask;
 699}
 700
 701static int do_poll(unsigned int nfds,  struct poll_list *list,
 702                   struct poll_wqueues *wait, struct timespec *end_time)
 703{
 704        poll_table* pt = &wait->pt;
 705        ktime_t expire, *to = NULL;
 706        int timed_out = 0, count = 0;
 707        unsigned long slack = 0;
 708
 709        /* Optimise the no-wait case */
 710        if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
 711                pt = NULL;
 712                timed_out = 1;
 713        }
 714
 715        if (end_time && !timed_out)
 716                slack = estimate_accuracy(end_time);
 717
 718        for (;;) {
 719                struct poll_list *walk;
 720
 721                for (walk = list; walk != NULL; walk = walk->next) {
 722                        struct pollfd * pfd, * pfd_end;
 723
 724                        pfd = walk->entries;
 725                        pfd_end = pfd + walk->len;
 726                        for (; pfd != pfd_end; pfd++) {
 727                                /*
 728                                 * Fish for events. If we found one, record it
 729                                 * and kill the poll_table, so we don't
 730                                 * needlessly register any other waiters after
 731                                 * this. They'll get immediately deregistered
 732                                 * when we break out and return.
 733                                 */
 734                                if (do_pollfd(pfd, pt)) {
 735                                        count++;
 736                                        pt = NULL;
 737                                }
 738                        }
 739                }
 740                /*
 741                 * All waiters have already been registered, so don't provide
 742                 * a poll_table to them on the next loop iteration.
 743                 */
 744                pt = NULL;
 745                if (!count) {
 746                        count = wait->error;
 747                        if (signal_pending(current))
 748                                count = -EINTR;
 749                }
 750                if (count || timed_out)
 751                        break;
 752
 753                /*
 754                 * If this is the first loop and we have a timeout
 755                 * given, then we convert to ktime_t and set the to
 756                 * pointer to the expiry value.
 757                 */
 758                if (end_time && !to) {
 759                        expire = timespec_to_ktime(*end_time);
 760                        to = &expire;
 761                }
 762
 763                if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
 764                        timed_out = 1;
 765        }
 766        return count;
 767}
 768
 769#define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
 770                        sizeof(struct pollfd))
 771
 772int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
 773                struct timespec *end_time)
 774{
 775        struct poll_wqueues table;
 776        int err = -EFAULT, fdcount, len, size;
 777        /* Allocate small arguments on the stack to save memory and be
 778           faster - use long to make sure the buffer is aligned properly
 779           on 64 bit archs to avoid unaligned access */
 780        long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
 781        struct poll_list *const head = (struct poll_list *)stack_pps;
 782        struct poll_list *walk = head;
 783        unsigned long todo = nfds;
 784
 785        if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
 786                return -EINVAL;
 787
 788        len = min_t(unsigned int, nfds, N_STACK_PPS);
 789        for (;;) {
 790                walk->next = NULL;
 791                walk->len = len;
 792                if (!len)
 793                        break;
 794
 795                if (copy_from_user(walk->entries, ufds + nfds-todo,
 796                                        sizeof(struct pollfd) * walk->len))
 797                        goto out_fds;
 798
 799                todo -= walk->len;
 800                if (!todo)
 801                        break;
 802
 803                len = min(todo, POLLFD_PER_PAGE);
 804                size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
 805                walk = walk->next = kmalloc(size, GFP_KERNEL);
 806                if (!walk) {
 807                        err = -ENOMEM;
 808                        goto out_fds;
 809                }
 810        }
 811
 812        poll_initwait(&table);
 813        fdcount = do_poll(nfds, head, &table, end_time);
 814        poll_freewait(&table);
 815
 816        for (walk = head; walk; walk = walk->next) {
 817                struct pollfd *fds = walk->entries;
 818                int j;
 819
 820                for (j = 0; j < walk->len; j++, ufds++)
 821                        if (__put_user(fds[j].revents, &ufds->revents))
 822                                goto out_fds;
 823        }
 824
 825        err = fdcount;
 826out_fds:
 827        walk = head->next;
 828        while (walk) {
 829                struct poll_list *pos = walk;
 830                walk = walk->next;
 831                kfree(pos);
 832        }
 833
 834        return err;
 835}
 836
 837static long do_restart_poll(struct restart_block *restart_block)
 838{
 839        struct pollfd __user *ufds = restart_block->poll.ufds;
 840        int nfds = restart_block->poll.nfds;
 841        struct timespec *to = NULL, end_time;
 842        int ret;
 843
 844        if (restart_block->poll.has_timeout) {
 845                end_time.tv_sec = restart_block->poll.tv_sec;
 846                end_time.tv_nsec = restart_block->poll.tv_nsec;
 847                to = &end_time;
 848        }
 849
 850        ret = do_sys_poll(ufds, nfds, to);
 851
 852        if (ret == -EINTR) {
 853                restart_block->fn = do_restart_poll;
 854                ret = -ERESTART_RESTARTBLOCK;
 855        }
 856        return ret;
 857}
 858
 859SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
 860                long, timeout_msecs)
 861{
 862        struct timespec end_time, *to = NULL;
 863        int ret;
 864
 865        if (timeout_msecs >= 0) {
 866                to = &end_time;
 867                poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
 868                        NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
 869        }
 870
 871        ret = do_sys_poll(ufds, nfds, to);
 872
 873        if (ret == -EINTR) {
 874                struct restart_block *restart_block;
 875
 876                restart_block = &current_thread_info()->restart_block;
 877                restart_block->fn = do_restart_poll;
 878                restart_block->poll.ufds = ufds;
 879                restart_block->poll.nfds = nfds;
 880
 881                if (timeout_msecs >= 0) {
 882                        restart_block->poll.tv_sec = end_time.tv_sec;
 883                        restart_block->poll.tv_nsec = end_time.tv_nsec;
 884                        restart_block->poll.has_timeout = 1;
 885                } else
 886                        restart_block->poll.has_timeout = 0;
 887
 888                ret = -ERESTART_RESTARTBLOCK;
 889        }
 890        return ret;
 891}
 892
 893#ifdef HAVE_SET_RESTORE_SIGMASK
 894SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
 895                struct timespec __user *, tsp, const sigset_t __user *, sigmask,
 896                size_t, sigsetsize)
 897{
 898        sigset_t ksigmask, sigsaved;
 899        struct timespec ts, end_time, *to = NULL;
 900        int ret;
 901
 902        if (tsp) {
 903                if (copy_from_user(&ts, tsp, sizeof(ts)))
 904                        return -EFAULT;
 905
 906                to = &end_time;
 907                if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
 908                        return -EINVAL;
 909        }
 910
 911        if (sigmask) {
 912                /* XXX: Don't preclude handling different sized sigset_t's.  */
 913                if (sigsetsize != sizeof(sigset_t))
 914                        return -EINVAL;
 915                if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 916                        return -EFAULT;
 917
 918                sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 919                sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 920        }
 921
 922        ret = do_sys_poll(ufds, nfds, to);
 923
 924        /* We can restart this syscall, usually */
 925        if (ret == -EINTR) {
 926                /*
 927                 * Don't restore the signal mask yet. Let do_signal() deliver
 928                 * the signal on the way back to userspace, before the signal
 929                 * mask is restored.
 930                 */
 931                if (sigmask) {
 932                        memcpy(&current->saved_sigmask, &sigsaved,
 933                                        sizeof(sigsaved));
 934                        set_restore_sigmask();
 935                }
 936                ret = -ERESTARTNOHAND;
 937        } else if (sigmask)
 938                sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 939
 940        ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
 941
 942        return ret;
 943}
 944#endif /* HAVE_SET_RESTORE_SIGMASK */
 945