fs/select.c

   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/fs.h>
  25 #include <linux/rcupdate.h>
  26
  27 #include <asm/uaccess.h>
  28
  29 #define DEFAULT_POLLMASK (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM)
  30
  31 struct poll_table_page {
  32         struct poll_table_page * next;
  33         struct poll_table_entry * entry;
  34         struct poll_table_entry entries[0];
  35 };
  36
  37 #define POLL_TABLE_FULL(table) \
  38         ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
  39
  40 /*
  41  * Ok, Peter made a complicated, but straightforward multiple_wait() function.
  42  * I have rewritten this, taking some shortcuts: This code may not be easy to
  43  * follow, but it should be free of race-conditions, and it's practical. If you
  44  * understand what I'm doing here, then you understand how the linux
  45  * sleep/wakeup mechanism works.
  46  *
  47  * Two very simple procedures, poll_wait() and poll_freewait() make all the
  48  * work.  poll_wait() is an inline-function defined in <linux/poll.h>,
  49  * as all select/poll functions have to call it to add an entry to the
  50  * poll table.
  51  */
  52 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
  53                        poll_table *p);
  54
  55 void poll_initwait(struct poll_wqueues *pwq)
  56 {
  57         init_poll_funcptr(&pwq->pt, __pollwait);
  58         pwq->error = 0;
  59         pwq->table = NULL;
  60         pwq->inline_index = 0;
  61 }
  62
  63 EXPORT_SYMBOL(poll_initwait);
  64
  65 static void free_poll_entry(struct poll_table_entry *entry)
  66 {
  67         remove_wait_queue(entry->wait_address, &entry->wait);
  68         fput(entry->filp);
  69 }
  70
  71 void poll_freewait(struct poll_wqueues *pwq)
  72 {
  73         struct poll_table_page * p = pwq->table;
  74         int i;
  75         for (i = 0; i < pwq->inline_index; i++)
  76                 free_poll_entry(pwq->inline_entries + i);
  77         while (p) {
  78                 struct poll_table_entry * entry;
  79                 struct poll_table_page *old;
  80
  81                 entry = p->entry;
  82                 do {
  83                         entry--;
  84                         free_poll_entry(entry);
  85                 } while (entry > p->entries);
  86                 old = p;
  87                 p = p->next;
  88                 free_page((unsigned long) old);
  89         }
  90 }
  91
  92 EXPORT_SYMBOL(poll_freewait);
  93
  94 static struct poll_table_entry *poll_get_entry(poll_table *_p)
  95 {
  96         struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt);
  97         struct poll_table_page *table = p->table;
  98
  99         if (p->inline_index < N_INLINE_POLL_ENTRIES)
 100                 return p->inline_entries + p->inline_index++;
 101
 102         if (!table || POLL_TABLE_FULL(table)) {
 103                 struct poll_table_page *new_table;
 104
 105                 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
 106                 if (!new_table) {
 107                         p->error = -ENOMEM;
 108                         __set_current_state(TASK_RUNNING);
 109                         return NULL;
 110                 }
 111                 new_table->entry = new_table->entries;
 112                 new_table->next = table;
 113                 p->table = new_table;
 114                 table = new_table;
 115         }
 116
 117         return table->entry++;
 118 }
 119
 120 /* Add a new entry */
 121 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
 122                                 poll_table *p)
 123 {
 124         struct poll_table_entry *entry = poll_get_entry(p);
 125         if (!entry)
 126                 return;
 127         get_file(filp);
 128         entry->filp = filp;
 129         entry->wait_address = wait_address;
 130         init_waitqueue_entry(&entry->wait, current);
 131         add_wait_queue(wait_address, &entry->wait);
 132 }
 133
 134 #define FDS_IN(fds, n)          (fds->in + n)
 135 #define FDS_OUT(fds, n)         (fds->out + n)
 136 #define FDS_EX(fds, n)          (fds->ex + n)
 137
 138 #define BITS(fds, n)    (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
 139
 140 static int max_select_fd(unsigned long n, fd_set_bits *fds)
 141 {
 142         unsigned long *open_fds;
 143         unsigned long set;
 144         int max;
 145         struct fdtable *fdt;
 146
 147         /* handle last in-complete long-word first */
 148         set = ~(~0UL << (n & (__NFDBITS-1)));
 149         n /= __NFDBITS;
 150         fdt = files_fdtable(current->files);
 151         open_fds = fdt->open_fds->fds_bits+n;
 152         max = 0;
 153         if (set) {
 154                 set &= BITS(fds, n);
 155                 if (set) {
 156                         if (!(set & ~*open_fds))
 157                                 goto get_max;
 158                         return -EBADF;
 159                 }
 160         }
 161         while (n) {
 162                 open_fds--;
 163                 n--;
 164                 set = BITS(fds, n);
 165                 if (!set)
 166                         continue;
 167                 if (set & ~*open_fds)
 168                         return -EBADF;
 169                 if (max)
 170                         continue;
 171 get_max:
 172                 do {
 173                         max++;
 174                         set >>= 1;
 175                 } while (set);
 176                 max += n * __NFDBITS;
 177         }
 178
 179         return max;
 180 }
 181
 182 #define BIT(i)          (1UL << ((i)&(__NFDBITS-1)))
 183 #define MEM(i,m)        ((m)+(unsigned)(i)/__NFDBITS)
 184 #define ISSET(i,m)      (((i)&*(m)) != 0)
 185 #define SET(i,m)        (*(m) |= (i))
 186
 187 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
 188 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
 189 #define POLLEX_SET (POLLPRI)
 190
 191 int do_select(int n, fd_set_bits *fds, s64 *timeout)
 192 {
 193         struct poll_wqueues table;
 194         poll_table *wait;
 195         int retval, i;
 196
 197         rcu_read_lock();
 198         retval = max_select_fd(n, fds);
 199         rcu_read_unlock();
 200
 201         if (retval < 0)
 202                 return retval;
 203         n = retval;
 204
 205         poll_initwait(&table);
 206         wait = &table.pt;
 207         if (!*timeout)
 208                 wait = NULL;
 209         retval = 0;
 210         for (;;) {
 211                 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
 212                 long __timeout;
 213
 214                 set_current_state(TASK_INTERRUPTIBLE);
 215
 216                 inp = fds->in; outp = fds->out; exp = fds->ex;
 217                 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
 218
 219                 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
 220                         unsigned long in, out, ex, all_bits, bit = 1, mask, j;
 221                         unsigned long res_in = 0, res_out = 0, res_ex = 0;
 222                         const struct file_operations *f_op = NULL;
 223                         struct file *file = NULL;
 224
 225                         in = *inp++; out = *outp++; ex = *exp++;
 226                         all_bits = in | out | ex;
 227                         if (all_bits == 0) {
 228                                 i += __NFDBITS;
 229                                 continue;
 230                         }
 231
 232                         for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
 233                                 int fput_needed;
 234                                 if (i >= n)
 235                                         break;
 236                                 if (!(bit & all_bits))
 237                                         continue;
 238                                 file = fget_light(i, &fput_needed);
 239                                 if (file) {
 240                                         f_op = file->f_op;
 241                                         mask = DEFAULT_POLLMASK;
 242                                         if (f_op && f_op->poll)
 243                                                 mask = (*f_op->poll)(file, retval ? NULL : wait);
 244                                         fput_light(file, fput_needed);
 245                                         if ((mask & POLLIN_SET) && (in & bit)) {
 246                                                 res_in |= bit;
 247                                                 retval++;
 248                                         }
 249                                         if ((mask & POLLOUT_SET) && (out & bit)) {
 250                                                 res_out |= bit;
 251                                                 retval++;
 252                                         }
 253                                         if ((mask & POLLEX_SET) && (ex & bit)) {
 254                                                 res_ex |= bit;
 255                                                 retval++;
 256                                         }
 257                                 }
 258                                 cond_resched();
 259                         }
 260                         if (res_in)
 261                                 *rinp = res_in;
 262                         if (res_out)
 263                                 *routp = res_out;
 264                         if (res_ex)
 265                                 *rexp = res_ex;
 266                 }
 267                 wait = NULL;
 268                 if (retval || !*timeout || signal_pending(current))
 269                         break;
 270                 if(table.error) {
 271                         retval = table.error;
 272                         break;
 273                 }
 274
 275                 if (*timeout < 0) {
 276                         /* Wait indefinitely */
 277                         __timeout = MAX_SCHEDULE_TIMEOUT;
 278                 } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT - 1)) {
 279                         /* Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in a loop */
 280                         __timeout = MAX_SCHEDULE_TIMEOUT - 1;
 281                         *timeout -= __timeout;
 282                 } else {
 283                         __timeout = *timeout;
 284                         *timeout = 0;
 285                 }
 286                 __timeout = schedule_timeout(__timeout);
 287                 if (*timeout >= 0)
 288                         *timeout += __timeout;
 289         }
 290         __set_current_state(TASK_RUNNING);
 291
 292         poll_freewait(&table);
 293
 294         return retval;
 295 }
 296
 297 /*
 298  * We can actually return ERESTARTSYS instead of EINTR, but I'd
 299  * like to be certain this leads to no problems. So I return
 300  * EINTR just for safety.
 301  *
 302  * Update: ERESTARTSYS breaks at least the xview clock binary, so
 303  * I'm trying ERESTARTNOHAND which restart only when you want to.
 304  */
 305 #define MAX_SELECT_SECONDS \
 306         ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
 307
 308 static int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 309                            fd_set __user *exp, s64 *timeout)
 310 {
 311         fd_set_bits fds;
 312         void *bits;
 313         int ret, max_fds;
 314         unsigned int size;
 315         struct fdtable *fdt;
 316         /* Allocate small arguments on the stack to save memory and be faster */
 317         long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
 318
 319         ret = -EINVAL;
 320         if (n < 0)
 321                 goto out_nofds;
 322
 323         /* max_fds can increase, so grab it once to avoid race */
 324         rcu_read_lock();
 325         fdt = files_fdtable(current->files);
 326         max_fds = fdt->max_fds;
 327         rcu_read_unlock();
 328         if (n > max_fds)
 329                 n = max_fds;
 330
 331         /*
 332          * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
 333          * since we used fdset we need to allocate memory in units of
 334          * long-words.
 335          */
 336         size = FDS_BYTES(n);
 337         bits = stack_fds;
 338         if (size > sizeof(stack_fds) / 6) {
 339                 /* Not enough space in on-stack array; must use kmalloc */
 340                 ret = -ENOMEM;
 341                 bits = kmalloc(6 * size, GFP_KERNEL);
 342                 if (!bits)
 343                         goto out_nofds;
 344         }
 345         fds.in      = bits;
 346         fds.out     = bits +   size;
 347         fds.ex      = bits + 2*size;
 348         fds.res_in  = bits + 3*size;
 349         fds.res_out = bits + 4*size;
 350         fds.res_ex  = bits + 5*size;
 351
 352         if ((ret = get_fd_set(n, inp, fds.in)) ||
 353             (ret = get_fd_set(n, outp, fds.out)) ||
 354             (ret = get_fd_set(n, exp, fds.ex)))
 355                 goto out;
 356         zero_fd_set(n, fds.res_in);
 357         zero_fd_set(n, fds.res_out);
 358         zero_fd_set(n, fds.res_ex);
 359
 360         ret = do_select(n, &fds, timeout);
 361
 362         if (ret < 0)
 363                 goto out;
 364         if (!ret) {
 365                 ret = -ERESTARTNOHAND;
 366                 if (signal_pending(current))
 367                         goto out;
 368                 ret = 0;
 369         }
 370
 371         if (set_fd_set(n, inp, fds.res_in) ||
 372             set_fd_set(n, outp, fds.res_out) ||
 373             set_fd_set(n, exp, fds.res_ex))
 374                 ret = -EFAULT;
 375
 376 out:
 377         if (bits != stack_fds)
 378                 kfree(bits);
 379 out_nofds:
 380         return ret;
 381 }
 382
 383 asmlinkage long sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 384                         fd_set __user *exp, struct timeval __user *tvp)
 385 {
 386         s64 timeout = -1;
 387         struct timeval tv;
 388         int ret;
 389
 390         if (tvp) {
 391                 if (copy_from_user(&tv, tvp, sizeof(tv)))
 392                         return -EFAULT;
 393
 394                 if (tv.tv_sec < 0 || tv.tv_usec < 0)
 395                         return -EINVAL;
 396
 397                 /* Cast to u64 to make GCC stop complaining */
 398                 if ((u64)tv.tv_sec >= (u64)MAX_INT64_SECONDS)
 399                         timeout = -1;   /* infinite */
 400                 else {
 401                         timeout = DIV_ROUND_UP(tv.tv_usec, USEC_PER_SEC/HZ);
 402                         timeout += tv.tv_sec * HZ;
 403                 }
 404         }
 405
 406         ret = core_sys_select(n, inp, outp, exp, &timeout);
 407
 408         if (tvp) {
 409                 struct timeval rtv;
 410
 411                 if (current->personality & STICKY_TIMEOUTS)
 412                         goto sticky;
 413                 rtv.tv_usec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ));
 414                 rtv.tv_sec = timeout;
 415                 if (timeval_compare(&rtv, &tv) >= 0)
 416                         rtv = tv;
 417                 if (copy_to_user(tvp, &rtv, sizeof(rtv))) {
 418 sticky:
 419                         /*
 420                          * If an application puts its timeval in read-only
 421                          * memory, we don't want the Linux-specific update to
 422                          * the timeval to cause a fault after the select has
 423                          * completed successfully. However, because we're not
 424                          * updating the timeval, we can't restart the system
 425                          * call.
 426                          */
 427                         if (ret == -ERESTARTNOHAND)
 428                                 ret = -EINTR;
 429                 }
 430         }
 431
 432         return ret;
 433 }
 434
 435 #ifdef TIF_RESTORE_SIGMASK
 436 asmlinkage long sys_pselect7(int n, fd_set __user *inp, fd_set __user *outp,
 437                 fd_set __user *exp, struct timespec __user *tsp,
 438                 const sigset_t __user *sigmask, size_t sigsetsize)
 439 {
 440         s64 timeout = MAX_SCHEDULE_TIMEOUT;
 441         sigset_t ksigmask, sigsaved;
 442         struct timespec ts;
 443         int ret;
 444
 445         if (tsp) {
 446                 if (copy_from_user(&ts, tsp, sizeof(ts)))
 447                         return -EFAULT;
 448
 449                 if (ts.tv_sec < 0 || ts.tv_nsec < 0)
 450                         return -EINVAL;
 451
 452                 /* Cast to u64 to make GCC stop complaining */
 453                 if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
 454                         timeout = -1;   /* infinite */
 455                 else {
 456                         timeout = DIV_ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
 457                         timeout += ts.tv_sec * HZ;
 458                 }
 459         }
 460
 461         if (sigmask) {
 462                 /* XXX: Don't preclude handling different sized sigset_t's.  */
 463                 if (sigsetsize != sizeof(sigset_t))
 464                         return -EINVAL;
 465                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 466                         return -EFAULT;
 467
 468                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 469                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 470         }
 471
 472         ret = core_sys_select(n, inp, outp, exp, &timeout);
 473
 474         if (tsp) {
 475                 struct timespec rts;
 476
 477                 if (current->personality & STICKY_TIMEOUTS)
 478                         goto sticky;
 479                 rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
 480                                                 1000;
 481                 rts.tv_sec = timeout;
 482                 if (timespec_compare(&rts, &ts) >= 0)
 483                         rts = ts;
 484                 if (copy_to_user(tsp, &rts, sizeof(rts))) {
 485 sticky:
 486                         /*
 487                          * If an application puts its timeval in read-only
 488                          * memory, we don't want the Linux-specific update to
 489                          * the timeval to cause a fault after the select has
 490                          * completed successfully. However, because we're not
 491                          * updating the timeval, we can't restart the system
 492                          * call.
 493                          */
 494                         if (ret == -ERESTARTNOHAND)
 495                                 ret = -EINTR;
 496                 }
 497         }
 498
 499         if (ret == -ERESTARTNOHAND) {
 500                 /*
 501                  * Don't restore the signal mask yet. Let do_signal() deliver
 502                  * the signal on the way back to userspace, before the signal
 503                  * mask is restored.
 504                  */
 505                 if (sigmask) {
 506                         memcpy(&current->saved_sigmask, &sigsaved,
 507                                         sizeof(sigsaved));
 508                         set_thread_flag(TIF_RESTORE_SIGMASK);
 509                 }
 510         } else if (sigmask)
 511                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 512
 513         return ret;
 514 }
 515
 516 /*
 517  * Most architectures can't handle 7-argument syscalls. So we provide a
 518  * 6-argument version where the sixth argument is a pointer to a structure
 519  * which has a pointer to the sigset_t itself followed by a size_t containing
 520  * the sigset size.
 521  */
 522 asmlinkage long sys_pselect6(int n, fd_set __user *inp, fd_set __user *outp,
 523         fd_set __user *exp, struct timespec __user *tsp, void __user *sig)
 524 {
 525         size_t sigsetsize = 0;
 526         sigset_t __user *up = NULL;
 527
 528         if (sig) {
 529                 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
 530                     || __get_user(up, (sigset_t __user * __user *)sig)
 531                     || __get_user(sigsetsize,
 532                                 (size_t __user *)(sig+sizeof(void *))))
 533                         return -EFAULT;
 534         }
 535
 536         return sys_pselect7(n, inp, outp, exp, tsp, up, sigsetsize);
 537 }
 538 #endif /* TIF_RESTORE_SIGMASK */
 539
 540 struct poll_list {
 541         struct poll_list *next;
 542         int len;
 543         struct pollfd entries[0];
 544 };
 545
 546 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 547
 548 /*
 549  * Fish for pollable events on the pollfd->fd file descriptor. We're only
 550  * interested in events matching the pollfd->events mask, and the result
 551  * matching that mask is both recorded in pollfd->revents and returned. The
 552  * pwait poll_table will be used by the fd-provided poll handler for waiting,
 553  * if non-NULL.
 554  */
 555 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
 556 {
 557         unsigned int mask;
 558         int fd;
 559
 560         mask = 0;
 561         fd = pollfd->fd;
 562         if (fd >= 0) {
 563                 int fput_needed;
 564                 struct file * file;
 565
 566                 file = fget_light(fd, &fput_needed);
 567                 mask = POLLNVAL;
 568                 if (file != NULL) {
 569                         mask = DEFAULT_POLLMASK;
 570                         if (file->f_op && file->f_op->poll)
 571                                 mask = file->f_op->poll(file, pwait);
 572                         /* Mask out unneeded events. */
 573                         mask &= pollfd->events | POLLERR | POLLHUP;
 574                         fput_light(file, fput_needed);
 575                 }
 576         }
 577         pollfd->revents = mask;
 578
 579         return mask;
 580 }
 581
 582 static int do_poll(unsigned int nfds,  struct poll_list *list,
 583                    struct poll_wqueues *wait, s64 *timeout)
 584 {
 585         int count = 0;
 586         poll_table* pt = &wait->pt;
 587
 588         /* Optimise the no-wait case */
 589         if (!(*timeout))
 590                 pt = NULL;
 591
 592         for (;;) {
 593                 struct poll_list *walk;
 594                 long __timeout;
 595
 596                 set_current_state(TASK_INTERRUPTIBLE);
 597                 for (walk = list; walk != NULL; walk = walk->next) {
 598                         struct pollfd * pfd, * pfd_end;
 599
 600                         pfd = walk->entries;
 601                         pfd_end = pfd + walk->len;
 602                         for (; pfd != pfd_end; pfd++) {
 603                                 /*
 604                                  * Fish for events. If we found one, record it
 605                                  * and kill the poll_table, so we don't
 606                                  * needlessly register any other waiters after
 607                                  * this. They'll get immediately deregistered
 608                                  * when we break out and return.
 609                                  */
 610                                 if (do_pollfd(pfd, pt)) {
 611                                         count++;
 612                                         pt = NULL;
 613                                 }
 614                         }
 615                 }
 616                 /*
 617                  * All waiters have already been registered, so don't provide
 618                  * a poll_table to them on the next loop iteration.
 619                  */
 620                 pt = NULL;
 621                 if (count || !*timeout || signal_pending(current))
 622                         break;
 623                 count = wait->error;
 624                 if (count)
 625                         break;
 626
 627                 if (*timeout < 0) {
 628                         /* Wait indefinitely */
 629                         __timeout = MAX_SCHEDULE_TIMEOUT;
 630                 } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT-1)) {
 631                         /*
 632                          * Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in
 633                          * a loop
 634                          */
 635                         __timeout = MAX_SCHEDULE_TIMEOUT - 1;
 636                         *timeout -= __timeout;
 637                 } else {
 638                         __timeout = *timeout;
 639                         *timeout = 0;
 640                 }
 641
 642                 __timeout = schedule_timeout(__timeout);
 643                 if (*timeout >= 0)
 644                         *timeout += __timeout;
 645         }
 646         __set_current_state(TASK_RUNNING);
 647         return count;
 648 }
 649
 650 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
 651                         sizeof(struct pollfd))
 652
 653 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds, s64 *timeout)
 654 {
 655         struct poll_wqueues table;
 656         int fdcount, err;
 657         unsigned int i;
 658         struct poll_list *head;
 659         struct poll_list *walk;
 660         /* Allocate small arguments on the stack to save memory and be
 661            faster - use long to make sure the buffer is aligned properly
 662            on 64 bit archs to avoid unaligned access */
 663         long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
 664         struct poll_list *stack_pp = NULL;
 665
 666         /* Do a sanity check on nfds ... */
 667         if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
 668                 return -EINVAL;
 669
 670         poll_initwait(&table);
 671
 672         head = NULL;
 673         walk = NULL;
 674         i = nfds;
 675         err = -ENOMEM;
 676         while(i!=0) {
 677                 struct poll_list *pp;
 678                 int num, size;
 679                 if (stack_pp == NULL)
 680                         num = N_STACK_PPS;
 681                 else
 682                         num = POLLFD_PER_PAGE;
 683                 if (num > i)
 684                         num = i;
 685                 size = sizeof(struct poll_list) + sizeof(struct pollfd)*num;
 686                 if (!stack_pp)
 687                         stack_pp = pp = (struct poll_list *)stack_pps;
 688                 else {
 689                         pp = kmalloc(size, GFP_KERNEL);
 690                         if (!pp)
 691                                 goto out_fds;
 692                 }
 693                 pp->next=NULL;
 694                 pp->len = num;
 695                 if (head == NULL)
 696                         head = pp;
 697                 else
 698                         walk->next = pp;
 699
 700                 walk = pp;
 701                 if (copy_from_user(pp->entries, ufds + nfds-i,
 702                                 sizeof(struct pollfd)*num)) {
 703                         err = -EFAULT;
 704                         goto out_fds;
 705                 }
 706                 i -= pp->len;
 707         }
 708
 709         fdcount = do_poll(nfds, head, &table, timeout);
 710
 711         /* OK, now copy the revents fields back to user space. */
 712         walk = head;
 713         err = -EFAULT;
 714         while(walk != NULL) {
 715                 struct pollfd *fds = walk->entries;
 716                 int j;
 717
 718                 for (j=0; j < walk->len; j++, ufds++) {
 719                         if(__put_user(fds[j].revents, &ufds->revents))
 720                                 goto out_fds;
 721                 }
 722                 walk = walk->next;
 723         }
 724         err = fdcount;
 725         if (!fdcount && signal_pending(current))
 726                 err = -EINTR;
 727 out_fds:
 728         walk = head;
 729         while(walk!=NULL) {
 730                 struct poll_list *pp = walk->next;
 731                 if (walk != stack_pp)
 732                         kfree(walk);
 733                 walk = pp;
 734         }
 735         poll_freewait(&table);
 736         return err;
 737 }
 738
 739 asmlinkage long sys_poll(struct pollfd __user *ufds, unsigned int nfds,
 740                         long timeout_msecs)
 741 {
 742         s64 timeout_jiffies;
 743
 744         if (timeout_msecs > 0) {
 745 #if HZ > 1000
 746                 /* We can only overflow if HZ > 1000 */
 747                 if (timeout_msecs / 1000 > (s64)0x7fffffffffffffffULL / (s64)HZ)
 748                         timeout_jiffies = -1;
 749                 else
 750 #endif
 751                         timeout_jiffies = msecs_to_jiffies(timeout_msecs);
 752         } else {
 753                 /* Infinite (< 0) or no (0) timeout */
 754                 timeout_jiffies = timeout_msecs;
 755         }
 756
 757         return do_sys_poll(ufds, nfds, &timeout_jiffies);
 758 }
 759
 760 #ifdef TIF_RESTORE_SIGMASK
 761 asmlinkage long sys_ppoll(struct pollfd __user *ufds, unsigned int nfds,
 762         struct timespec __user *tsp, const sigset_t __user *sigmask,
 763         size_t sigsetsize)
 764 {
 765         sigset_t ksigmask, sigsaved;
 766         struct timespec ts;
 767         s64 timeout = -1;
 768         int ret;
 769
 770         if (tsp) {
 771                 if (copy_from_user(&ts, tsp, sizeof(ts)))
 772                         return -EFAULT;
 773
 774                 /* Cast to u64 to make GCC stop complaining */
 775                 if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
 776                         timeout = -1;   /* infinite */
 777                 else {
 778                         timeout = DIV_ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
 779                         timeout += ts.tv_sec * HZ;
 780                 }
 781         }
 782
 783         if (sigmask) {
 784                 /* XXX: Don't preclude handling different sized sigset_t's.  */
 785                 if (sigsetsize != sizeof(sigset_t))
 786                         return -EINVAL;
 787                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 788                         return -EFAULT;
 789
 790                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 791                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 792         }
 793
 794         ret = do_sys_poll(ufds, nfds, &timeout);
 795
 796         /* We can restart this syscall, usually */
 797         if (ret == -EINTR) {
 798                 /*
 799                  * Don't restore the signal mask yet. Let do_signal() deliver
 800                  * the signal on the way back to userspace, before the signal
 801                  * mask is restored.
 802                  */
 803                 if (sigmask) {
 804                         memcpy(&current->saved_sigmask, &sigsaved,
 805                                         sizeof(sigsaved));
 806                         set_thread_flag(TIF_RESTORE_SIGMASK);
 807                 }
 808                 ret = -ERESTARTNOHAND;
 809         } else if (sigmask)
 810                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 811
 812         if (tsp && timeout >= 0) {
 813                 struct timespec rts;
 814
 815                 if (current->personality & STICKY_TIMEOUTS)
 816                         goto sticky;
 817                 /* Yes, we know it's actually an s64, but it's also positive. */
 818                 rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
 819                                                 1000;
 820                 rts.tv_sec = timeout;
 821                 if (timespec_compare(&rts, &ts) >= 0)
 822                         rts = ts;
 823                 if (copy_to_user(tsp, &rts, sizeof(rts))) {
 824                 sticky:
 825                         /*
 826                          * If an application puts its timeval in read-only
 827                          * memory, we don't want the Linux-specific update to
 828                          * the timeval to cause a fault after the select has
 829                          * completed successfully. However, because we're not
 830                          * updating the timeval, we can't restart the system
 831                          * call.
 832                          */
 833                         if (ret == -ERESTARTNOHAND && timeout >= 0)
 834                                 ret = -EINTR;
 835                 }
 836         }
 837
 838         return ret;
 839 }
 840 #endif /* TIF_RESTORE_SIGMASK */