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