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/fdtable.h>
  25 #include <linux/fs.h>
  26 #include <linux/rcupdate.h>
  27
  28 #include <asm/uaccess.h>
  29
  30 struct poll_table_page {
  31         struct poll_table_page * next;
  32         struct poll_table_entry * entry;
  33         struct poll_table_entry entries[0];
  34 };
  35
  36 #define POLL_TABLE_FULL(table) \
  37         ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
  38
  39 /*
  40  * Ok, Peter made a complicated, but straightforward multiple_wait() function.
  41  * I have rewritten this, taking some shortcuts: This code may not be easy to
  42  * follow, but it should be free of race-conditions, and it's practical. If you
  43  * understand what I'm doing here, then you understand how the linux
  44  * sleep/wakeup mechanism works.
  45  *
  46  * Two very simple procedures, poll_wait() and poll_freewait() make all the
  47  * work.  poll_wait() is an inline-function defined in <linux/poll.h>,
  48  * as all select/poll functions have to call it to add an entry to the
  49  * poll table.
  50  */
  51 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
  52                        poll_table *p);
  53
  54 void poll_initwait(struct poll_wqueues *pwq)
  55 {
  56         init_poll_funcptr(&pwq->pt, __pollwait);
  57         pwq->error = 0;
  58         pwq->table = NULL;
  59         pwq->inline_index = 0;
  60 }
  61
  62 EXPORT_SYMBOL(poll_initwait);
  63
  64 static void free_poll_entry(struct poll_table_entry *entry)
  65 {
  66         remove_wait_queue(entry->wait_address, &entry->wait);
  67         fput(entry->filp);
  68 }
  69
  70 void poll_freewait(struct poll_wqueues *pwq)
  71 {
  72         struct poll_table_page * p = pwq->table;
  73         int i;
  74         for (i = 0; i < pwq->inline_index; i++)
  75                 free_poll_entry(pwq->inline_entries + i);
  76         while (p) {
  77                 struct poll_table_entry * entry;
  78                 struct poll_table_page *old;
  79
  80                 entry = p->entry;
  81                 do {
  82                         entry--;
  83                         free_poll_entry(entry);
  84                 } while (entry > p->entries);
  85                 old = p;
  86                 p = p->next;
  87                 free_page((unsigned long) old);
  88         }
  89 }
  90
  91 EXPORT_SYMBOL(poll_freewait);
  92
  93 static struct poll_table_entry *poll_get_entry(poll_table *_p)
  94 {
  95         struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt);
  96         struct poll_table_page *table = p->table;
  97
  98         if (p->inline_index < N_INLINE_POLL_ENTRIES)
  99                 return p->inline_entries + p->inline_index++;
 100
 101         if (!table || POLL_TABLE_FULL(table)) {
 102                 struct poll_table_page *new_table;
 103
 104                 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
 105                 if (!new_table) {
 106                         p->error = -ENOMEM;
 107                         __set_current_state(TASK_RUNNING);
 108                         return NULL;
 109                 }
 110                 new_table->entry = new_table->entries;
 111                 new_table->next = table;
 112                 p->table = new_table;
 113                 table = new_table;
 114         }
 115
 116         return table->entry++;
 117 }
 118
 119 /* Add a new entry */
 120 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
 121                                 poll_table *p)
 122 {
 123         struct poll_table_entry *entry = poll_get_entry(p);
 124         if (!entry)
 125                 return;
 126         get_file(filp);
 127         entry->filp = filp;
 128         entry->wait_address = wait_address;
 129         init_waitqueue_entry(&entry->wait, current);
 130         add_wait_queue(wait_address, &entry->wait);
 131 }
 132
 133 #define FDS_IN(fds, n)          (fds->in + n)
 134 #define FDS_OUT(fds, n)         (fds->out + n)
 135 #define FDS_EX(fds, n)          (fds->ex + n)
 136
 137 #define BITS(fds, n)    (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
 138
 139 static int max_select_fd(unsigned long n, fd_set_bits *fds)
 140 {
 141         unsigned long *open_fds;
 142         unsigned long set;
 143         int max;
 144         struct fdtable *fdt;
 145
 146         /* handle last in-complete long-word first */
 147         set = ~(~0UL << (n & (__NFDBITS-1)));
 148         n /= __NFDBITS;
 149         fdt = files_fdtable(current->files);
 150         open_fds = fdt->open_fds->fds_bits+n;
 151         max = 0;
 152         if (set) {
 153                 set &= BITS(fds, n);
 154                 if (set) {
 155                         if (!(set & ~*open_fds))
 156                                 goto get_max;
 157                         return -EBADF;
 158                 }
 159         }
 160         while (n) {
 161                 open_fds--;
 162                 n--;
 163                 set = BITS(fds, n);
 164                 if (!set)
 165                         continue;
 166                 if (set & ~*open_fds)
 167                         return -EBADF;
 168                 if (max)
 169                         continue;
 170 get_max:
 171                 do {
 172                         max++;
 173                         set >>= 1;
 174                 } while (set);
 175                 max += n * __NFDBITS;
 176         }
 177
 178         return max;
 179 }
 180
 181 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
 182 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
 183 #define POLLEX_SET (POLLPRI)
 184
 185 int do_select(int n, fd_set_bits *fds, s64 *timeout)
 186 {
 187         struct poll_wqueues table;
 188         poll_table *wait;
 189         int retval, i;
 190
 191         rcu_read_lock();
 192         retval = max_select_fd(n, fds);
 193         rcu_read_unlock();
 194
 195         if (retval < 0)
 196                 return retval;
 197         n = retval;
 198
 199         poll_initwait(&table);
 200         wait = &table.pt;
 201         if (!*timeout)
 202                 wait = NULL;
 203         retval = 0;
 204         for (;;) {
 205                 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
 206                 long __timeout;
 207
 208                 set_current_state(TASK_INTERRUPTIBLE);
 209
 210                 inp = fds->in; outp = fds->out; exp = fds->ex;
 211                 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
 212
 213                 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
 214                         unsigned long in, out, ex, all_bits, bit = 1, mask, j;
 215                         unsigned long res_in = 0, res_out = 0, res_ex = 0;
 216                         const struct file_operations *f_op = NULL;
 217                         struct file *file = NULL;
 218
 219                         in = *inp++; out = *outp++; ex = *exp++;
 220                         all_bits = in | out | ex;
 221                         if (all_bits == 0) {
 222                                 i += __NFDBITS;
 223                                 continue;
 224                         }
 225
 226                         for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
 227                                 int fput_needed;
 228                                 if (i >= n)
 229                                         break;
 230                                 if (!(bit & all_bits))
 231                                         continue;
 232                                 file = fget_light(i, &fput_needed);
 233                                 if (file) {
 234                                         f_op = file->f_op;
 235                                         mask = DEFAULT_POLLMASK;
 236                                         if (f_op && f_op->poll)
 237                                                 mask = (*f_op->poll)(file, retval ? NULL : wait);
 238                                         fput_light(file, fput_needed);
 239                                         if ((mask & POLLIN_SET) && (in & bit)) {
 240                                                 res_in |= bit;
 241                                                 retval++;
 242                                         }
 243                                         if ((mask & POLLOUT_SET) && (out & bit)) {
 244                                                 res_out |= bit;
 245                                                 retval++;
 246                                         }
 247                                         if ((mask & POLLEX_SET) && (ex & bit)) {
 248                                                 res_ex |= bit;
 249                                                 retval++;
 250                                         }
 251                                 }
 252                         }
 253                         if (res_in)
 254                                 *rinp = res_in;
 255                         if (res_out)
 256                                 *routp = res_out;
 257                         if (res_ex)
 258                                 *rexp = res_ex;
 259                         cond_resched();
 260                 }
 261                 wait = NULL;
 262                 if (retval || !*timeout || signal_pending(current))
 263                         break;
 264                 if (table.error) {
 265                         retval = table.error;
 266                         break;
 267                 }
 268
 269                 if (*timeout < 0) {
 270                         /* Wait indefinitely */
 271                         __timeout = MAX_SCHEDULE_TIMEOUT;
 272                 } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT - 1)) {
 273                         /* Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in a loop */
 274                         __timeout = MAX_SCHEDULE_TIMEOUT - 1;
 275                         *timeout -= __timeout;
 276                 } else {
 277                         __timeout = *timeout;
 278                         *timeout = 0;
 279                 }
 280                 __timeout = schedule_timeout(__timeout);
 281                 if (*timeout >= 0)
 282                         *timeout += __timeout;
 283         }
 284         __set_current_state(TASK_RUNNING);
 285
 286         poll_freewait(&table);
 287
 288         return retval;
 289 }
 290
 291 /*
 292  * We can actually return ERESTARTSYS instead of EINTR, but I'd
 293  * like to be certain this leads to no problems. So I return
 294  * EINTR just for safety.
 295  *
 296  * Update: ERESTARTSYS breaks at least the xview clock binary, so
 297  * I'm trying ERESTARTNOHAND which restart only when you want to.
 298  */
 299 #define MAX_SELECT_SECONDS \
 300         ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
 301
 302 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 303                            fd_set __user *exp, s64 *timeout)
 304 {
 305         fd_set_bits fds;
 306         void *bits;
 307         int ret, max_fds;
 308         unsigned int size;
 309         struct fdtable *fdt;
 310         /* Allocate small arguments on the stack to save memory and be faster */
 311         long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
 312
 313         ret = -EINVAL;
 314         if (n < 0)
 315                 goto out_nofds;
 316
 317         /* max_fds can increase, so grab it once to avoid race */
 318         rcu_read_lock();
 319         fdt = files_fdtable(current->files);
 320         max_fds = fdt->max_fds;
 321         rcu_read_unlock();
 322         if (n > max_fds)
 323                 n = max_fds;
 324
 325         /*
 326          * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
 327          * since we used fdset we need to allocate memory in units of
 328          * long-words.
 329          */
 330         size = FDS_BYTES(n);
 331         bits = stack_fds;
 332         if (size > sizeof(stack_fds) / 6) {
 333                 /* Not enough space in on-stack array; must use kmalloc */
 334                 ret = -ENOMEM;
 335                 bits = kmalloc(6 * size, GFP_KERNEL);
 336                 if (!bits)
 337                         goto out_nofds;
 338         }
 339         fds.in      = bits;
 340         fds.out     = bits +   size;
 341         fds.ex      = bits + 2*size;
 342         fds.res_in  = bits + 3*size;
 343         fds.res_out = bits + 4*size;
 344         fds.res_ex  = bits + 5*size;
 345
 346         if ((ret = get_fd_set(n, inp, fds.in)) ||
 347             (ret = get_fd_set(n, outp, fds.out)) ||
 348             (ret = get_fd_set(n, exp, fds.ex)))
 349                 goto out;
 350         zero_fd_set(n, fds.res_in);
 351         zero_fd_set(n, fds.res_out);
 352         zero_fd_set(n, fds.res_ex);
 353
 354         ret = do_select(n, &fds, timeout);
 355
 356         if (ret < 0)
 357                 goto out;
 358         if (!ret) {
 359                 ret = -ERESTARTNOHAND;
 360                 if (signal_pending(current))
 361                         goto out;
 362                 ret = 0;
 363         }
 364
 365         if (set_fd_set(n, inp, fds.res_in) ||
 366             set_fd_set(n, outp, fds.res_out) ||
 367             set_fd_set(n, exp, fds.res_ex))
 368                 ret = -EFAULT;
 369
 370 out:
 371         if (bits != stack_fds)
 372                 kfree(bits);
 373 out_nofds:
 374         return ret;
 375 }
 376
 377 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
 378                 fd_set __user *, exp, struct timeval __user *, tvp)
 379 {
 380         s64 timeout = -1;
 381         struct timeval tv;
 382         int ret;
 383
 384         if (tvp) {
 385                 if (copy_from_user(&tv, tvp, sizeof(tv)))
 386                         return -EFAULT;
 387
 388                 if (tv.tv_sec < 0 || tv.tv_usec < 0)
 389                         return -EINVAL;
 390
 391                 /* Cast to u64 to make GCC stop complaining */
 392                 if ((u64)tv.tv_sec >= (u64)MAX_INT64_SECONDS)
 393                         timeout = -1;   /* infinite */
 394                 else {
 395                         timeout = DIV_ROUND_UP(tv.tv_usec, USEC_PER_SEC/HZ);
 396                         timeout += tv.tv_sec * HZ;
 397                 }
 398         }
 399
 400         ret = core_sys_select(n, inp, outp, exp, &timeout);
 401
 402         if (tvp) {
 403                 struct timeval rtv;
 404
 405                 if (current->personality & STICKY_TIMEOUTS)
 406                         goto sticky;
 407                 rtv.tv_usec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ));
 408                 rtv.tv_sec = timeout;
 409                 if (timeval_compare(&rtv, &tv) >= 0)
 410                         rtv = tv;
 411                 if (copy_to_user(tvp, &rtv, sizeof(rtv))) {
 412 sticky:
 413                         /*
 414                          * If an application puts its timeval in read-only
 415                          * memory, we don't want the Linux-specific update to
 416                          * the timeval to cause a fault after the select has
 417                          * completed successfully. However, because we're not
 418                          * updating the timeval, we can't restart the system
 419                          * call.
 420                          */
 421                         if (ret == -ERESTARTNOHAND)
 422                                 ret = -EINTR;
 423                 }
 424         }
 425
 426         return ret;
 427 }
 428
 429 #ifdef HAVE_SET_RESTORE_SIGMASK
 430 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
 431                        fd_set __user *exp, struct timespec __user *tsp,
 432                        const sigset_t __user *sigmask, size_t sigsetsize)
 433 {
 434         s64 timeout = MAX_SCHEDULE_TIMEOUT;
 435         sigset_t ksigmask, sigsaved;
 436         struct timespec ts;
 437         int ret;
 438
 439         if (tsp) {
 440                 if (copy_from_user(&ts, tsp, sizeof(ts)))
 441                         return -EFAULT;
 442
 443                 if (ts.tv_sec < 0 || ts.tv_nsec < 0)
 444                         return -EINVAL;
 445
 446                 /* Cast to u64 to make GCC stop complaining */
 447                 if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
 448                         timeout = -1;   /* infinite */
 449                 else {
 450                         timeout = DIV_ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
 451                         timeout += ts.tv_sec * HZ;
 452                 }
 453         }
 454
 455         if (sigmask) {
 456                 /* XXX: Don't preclude handling different sized sigset_t's.  */
 457                 if (sigsetsize != sizeof(sigset_t))
 458                         return -EINVAL;
 459                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 460                         return -EFAULT;
 461
 462                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 463                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 464         }
 465
 466         ret = core_sys_select(n, inp, outp, exp, &timeout);
 467
 468         if (tsp) {
 469                 struct timespec rts;
 470
 471                 if (current->personality & STICKY_TIMEOUTS)
 472                         goto sticky;
 473                 rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
 474                                                 1000;
 475                 rts.tv_sec = timeout;
 476                 if (timespec_compare(&rts, &ts) >= 0)
 477                         rts = ts;
 478                 if (copy_to_user(tsp, &rts, sizeof(rts))) {
 479 sticky:
 480                         /*
 481                          * If an application puts its timeval in read-only
 482                          * memory, we don't want the Linux-specific update to
 483                          * the timeval to cause a fault after the select has
 484                          * completed successfully. However, because we're not
 485                          * updating the timeval, we can't restart the system
 486                          * call.
 487                          */
 488                         if (ret == -ERESTARTNOHAND)
 489                                 ret = -EINTR;
 490                 }
 491         }
 492
 493         if (ret == -ERESTARTNOHAND) {
 494                 /*
 495                  * Don't restore the signal mask yet. Let do_signal() deliver
 496                  * the signal on the way back to userspace, before the signal
 497                  * mask is restored.
 498                  */
 499                 if (sigmask) {
 500                         memcpy(&current->saved_sigmask, &sigsaved,
 501                                         sizeof(sigsaved));
 502                         set_restore_sigmask();
 503                 }
 504         } else if (sigmask)
 505                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 506
 507         return ret;
 508 }
 509
 510 /*
 511  * Most architectures can't handle 7-argument syscalls. So we provide a
 512  * 6-argument version where the sixth argument is a pointer to a structure
 513  * which has a pointer to the sigset_t itself followed by a size_t containing
 514  * the sigset size.
 515  */
 516 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
 517                 fd_set __user *, exp, struct timespec __user *, tsp,
 518                 void __user *, sig)
 519 {
 520         size_t sigsetsize = 0;
 521         sigset_t __user *up = NULL;
 522
 523         if (sig) {
 524                 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
 525                     || __get_user(up, (sigset_t __user * __user *)sig)
 526                     || __get_user(sigsetsize,
 527                                 (size_t __user *)(sig+sizeof(void *))))
 528                         return -EFAULT;
 529         }
 530
 531         return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
 532 }
 533 #endif /* HAVE_SET_RESTORE_SIGMASK */
 534
 535 struct poll_list {
 536         struct poll_list *next;
 537         int len;
 538         struct pollfd entries[0];
 539 };
 540
 541 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 542
 543 /*
 544  * Fish for pollable events on the pollfd->fd file descriptor. We're only
 545  * interested in events matching the pollfd->events mask, and the result
 546  * matching that mask is both recorded in pollfd->revents and returned. The
 547  * pwait poll_table will be used by the fd-provided poll handler for waiting,
 548  * if non-NULL.
 549  */
 550 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
 551 {
 552         unsigned int mask;
 553         int fd;
 554
 555         mask = 0;
 556         fd = pollfd->fd;
 557         if (fd >= 0) {
 558                 int fput_needed;
 559                 struct file * file;
 560
 561                 file = fget_light(fd, &fput_needed);
 562                 mask = POLLNVAL;
 563                 if (file != NULL) {
 564                         mask = DEFAULT_POLLMASK;
 565                         if (file->f_op && file->f_op->poll)
 566                                 mask = file->f_op->poll(file, pwait);
 567                         /* Mask out unneeded events. */
 568                         mask &= pollfd->events | POLLERR | POLLHUP;
 569                         fput_light(file, fput_needed);
 570                 }
 571         }
 572         pollfd->revents = mask;
 573
 574         return mask;
 575 }
 576
 577 static int do_poll(unsigned int nfds,  struct poll_list *list,
 578                    struct poll_wqueues *wait, s64 *timeout)
 579 {
 580         int count = 0;
 581         poll_table* pt = &wait->pt;
 582
 583         /* Optimise the no-wait case */
 584         if (!(*timeout))
 585                 pt = NULL;
 586
 587         for (;;) {
 588                 struct poll_list *walk;
 589                 long __timeout;
 590
 591                 set_current_state(TASK_INTERRUPTIBLE);
 592                 for (walk = list; walk != NULL; walk = walk->next) {
 593                         struct pollfd * pfd, * pfd_end;
 594
 595                         pfd = walk->entries;
 596                         pfd_end = pfd + walk->len;
 597                         for (; pfd != pfd_end; pfd++) {
 598                                 /*
 599                                  * Fish for events. If we found one, record it
 600                                  * and kill the poll_table, so we don't
 601                                  * needlessly register any other waiters after
 602                                  * this. They'll get immediately deregistered
 603                                  * when we break out and return.
 604                                  */
 605                                 if (do_pollfd(pfd, pt)) {
 606                                         count++;
 607                                         pt = NULL;
 608                                 }
 609                         }
 610                 }
 611                 /*
 612                  * All waiters have already been registered, so don't provide
 613                  * a poll_table to them on the next loop iteration.
 614                  */
 615                 pt = NULL;
 616                 if (!count) {
 617                         count = wait->error;
 618                         if (signal_pending(current))
 619                                 count = -EINTR;
 620                 }
 621                 if (count || !*timeout)
 622                         break;
 623
 624                 if (*timeout < 0) {
 625                         /* Wait indefinitely */
 626                         __timeout = MAX_SCHEDULE_TIMEOUT;
 627                 } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT-1)) {
 628                         /*
 629                          * Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in
 630                          * a loop
 631                          */
 632                         __timeout = MAX_SCHEDULE_TIMEOUT - 1;
 633                         *timeout -= __timeout;
 634                 } else {
 635                         __timeout = *timeout;
 636                         *timeout = 0;
 637                 }
 638
 639                 __timeout = schedule_timeout(__timeout);
 640                 if (*timeout >= 0)
 641                         *timeout += __timeout;
 642         }
 643         __set_current_state(TASK_RUNNING);
 644         return count;
 645 }
 646
 647 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
 648                         sizeof(struct pollfd))
 649
 650 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds, s64 *timeout)
 651 {
 652         struct poll_wqueues table;
 653         int err = -EFAULT, fdcount, len, size;
 654         /* Allocate small arguments on the stack to save memory and be
 655            faster - use long to make sure the buffer is aligned properly
 656            on 64 bit archs to avoid unaligned access */
 657         long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
 658         struct poll_list *const head = (struct poll_list *)stack_pps;
 659         struct poll_list *walk = head;
 660         unsigned long todo = nfds;
 661
 662         if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
 663                 return -EINVAL;
 664
 665         len = min_t(unsigned int, nfds, N_STACK_PPS);
 666         for (;;) {
 667                 walk->next = NULL;
 668                 walk->len = len;
 669                 if (!len)
 670                         break;
 671
 672                 if (copy_from_user(walk->entries, ufds + nfds-todo,
 673                                         sizeof(struct pollfd) * walk->len))
 674                         goto out_fds;
 675
 676                 todo -= walk->len;
 677                 if (!todo)
 678                         break;
 679
 680                 len = min(todo, POLLFD_PER_PAGE);
 681                 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
 682                 walk = walk->next = kmalloc(size, GFP_KERNEL);
 683                 if (!walk) {
 684                         err = -ENOMEM;
 685                         goto out_fds;
 686                 }
 687         }
 688
 689         poll_initwait(&table);
 690         fdcount = do_poll(nfds, head, &table, timeout);
 691         poll_freewait(&table);
 692
 693         for (walk = head; walk; walk = walk->next) {
 694                 struct pollfd *fds = walk->entries;
 695                 int j;
 696
 697                 for (j = 0; j < walk->len; j++, ufds++)
 698                         if (__put_user(fds[j].revents, &ufds->revents))
 699                                 goto out_fds;
 700         }
 701
 702         err = fdcount;
 703 out_fds:
 704         walk = head->next;
 705         while (walk) {
 706                 struct poll_list *pos = walk;
 707                 walk = walk->next;
 708                 kfree(pos);
 709         }
 710
 711         return err;
 712 }
 713
 714 static long do_restart_poll(struct restart_block *restart_block)
 715 {
 716         struct pollfd __user *ufds = (struct pollfd __user*)restart_block->arg0;
 717         int nfds = restart_block->arg1;
 718         s64 timeout = ((s64)restart_block->arg3<<32) | (s64)restart_block->arg2;
 719         int ret;
 720
 721         ret = do_sys_poll(ufds, nfds, &timeout);
 722         if (ret == -EINTR) {
 723                 restart_block->fn = do_restart_poll;
 724                 restart_block->arg2 = timeout & 0xFFFFFFFF;
 725                 restart_block->arg3 = (u64)timeout >> 32;
 726                 ret = -ERESTART_RESTARTBLOCK;
 727         }
 728         return ret;
 729 }
 730
 731 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
 732                 long, timeout_msecs)
 733 {
 734         s64 timeout_jiffies;
 735         int ret;
 736
 737         if (timeout_msecs > 0) {
 738 #if HZ > 1000
 739                 /* We can only overflow if HZ > 1000 */
 740                 if (timeout_msecs / 1000 > (s64)0x7fffffffffffffffULL / (s64)HZ)
 741                         timeout_jiffies = -1;
 742                 else
 743 #endif
 744                         timeout_jiffies = msecs_to_jiffies(timeout_msecs) + 1;
 745         } else {
 746                 /* Infinite (< 0) or no (0) timeout */
 747                 timeout_jiffies = timeout_msecs;
 748         }
 749
 750         ret = do_sys_poll(ufds, nfds, &timeout_jiffies);
 751         if (ret == -EINTR) {
 752                 struct restart_block *restart_block;
 753                 restart_block = &current_thread_info()->restart_block;
 754                 restart_block->fn = do_restart_poll;
 755                 restart_block->arg0 = (unsigned long)ufds;
 756                 restart_block->arg1 = nfds;
 757                 restart_block->arg2 = timeout_jiffies & 0xFFFFFFFF;
 758                 restart_block->arg3 = (u64)timeout_jiffies >> 32;
 759                 ret = -ERESTART_RESTARTBLOCK;
 760         }
 761         return ret;
 762 }
 763
 764 #ifdef HAVE_SET_RESTORE_SIGMASK
 765 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
 766                 struct timespec __user *, tsp, const sigset_t __user *, sigmask,
 767                 size_t, sigsetsize)
 768 {
 769         sigset_t ksigmask, sigsaved;
 770         struct timespec ts;
 771         s64 timeout = -1;
 772         int ret;
 773
 774         if (tsp) {
 775                 if (copy_from_user(&ts, tsp, sizeof(ts)))
 776                         return -EFAULT;
 777
 778                 /* Cast to u64 to make GCC stop complaining */
 779                 if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
 780                         timeout = -1;   /* infinite */
 781                 else {
 782                         timeout = DIV_ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
 783                         timeout += ts.tv_sec * HZ;
 784                 }
 785         }
 786
 787         if (sigmask) {
 788                 /* XXX: Don't preclude handling different sized sigset_t's.  */
 789                 if (sigsetsize != sizeof(sigset_t))
 790                         return -EINVAL;
 791                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 792                         return -EFAULT;
 793
 794                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 795                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 796         }
 797
 798         ret = do_sys_poll(ufds, nfds, &timeout);
 799
 800         /* We can restart this syscall, usually */
 801         if (ret == -EINTR) {
 802                 /*
 803                  * Don't restore the signal mask yet. Let do_signal() deliver
 804                  * the signal on the way back to userspace, before the signal
 805                  * mask is restored.
 806                  */
 807                 if (sigmask) {
 808                         memcpy(&current->saved_sigmask, &sigsaved,
 809                                         sizeof(sigsaved));
 810                         set_restore_sigmask();
 811                 }
 812                 ret = -ERESTARTNOHAND;
 813         } else if (sigmask)
 814                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 815
 816         if (tsp && timeout >= 0) {
 817                 struct timespec rts;
 818
 819                 if (current->personality & STICKY_TIMEOUTS)
 820                         goto sticky;
 821                 /* Yes, we know it's actually an s64, but it's also positive. */
 822                 rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
 823                                                 1000;
 824                 rts.tv_sec = timeout;
 825                 if (timespec_compare(&rts, &ts) >= 0)
 826                         rts = ts;
 827                 if (copy_to_user(tsp, &rts, sizeof(rts))) {
 828                 sticky:
 829                         /*
 830                          * If an application puts its timeval in read-only
 831                          * memory, we don't want the Linux-specific update to
 832                          * the timeval to cause a fault after the select has
 833                          * completed successfully. However, because we're not
 834                          * updating the timeval, we can't restart the system
 835                          * call.
 836                          */
 837                         if (ret == -ERESTARTNOHAND && timeout >= 0)
 838                                 ret = -EINTR;
 839                 }
 840         }
 841
 842         return ret;
 843 }
 844 #endif /* HAVE_SET_RESTORE_SIGMASK */