git.oblomov.eu Git - linux-2.6/blob - 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
  27 #include <asm/uaccess.h>
  28
  29 #define ROUND_UP(x,y) (((x)+(y)-1)/(y))
  30 #define DEFAULT_POLLMASK (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM)
  31
  32 struct poll_table_page {
  33         struct poll_table_page * next;
  34         struct poll_table_entry * entry;
  35         struct poll_table_entry entries[0];
  36 };
  37
  38 #define POLL_TABLE_FULL(table) \
  39         ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
  40
  41 /*
  42  * Ok, Peter made a complicated, but straightforward multiple_wait() function.
  43  * I have rewritten this, taking some shortcuts: This code may not be easy to
  44  * follow, but it should be free of race-conditions, and it's practical. If you
  45  * understand what I'm doing here, then you understand how the linux
  46  * sleep/wakeup mechanism works.
  47  *
  48  * Two very simple procedures, poll_wait() and poll_freewait() make all the
  49  * work.  poll_wait() is an inline-function defined in <linux/poll.h>,
  50  * as all select/poll functions have to call it to add an entry to the
  51  * poll table.
  52  */
  53 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
  54                        poll_table *p);
  55
  56 void poll_initwait(struct poll_wqueues *pwq)
  57 {
  58         init_poll_funcptr(&pwq->pt, __pollwait);
  59         pwq->error = 0;
  60         pwq->table = NULL;
  61         pwq->inline_index = 0;
  62 }
  63
  64 EXPORT_SYMBOL(poll_initwait);
  65
  66 static void free_poll_entry(struct poll_table_entry *entry)
  67 {
  68         remove_wait_queue(entry->wait_address,&entry->wait);
  69         fput(entry->filp);
  70 }
  71
  72 void poll_freewait(struct poll_wqueues *pwq)
  73 {
  74         struct poll_table_page * p = pwq->table;
  75         int i;
  76         for (i = 0; i < pwq->inline_index; i++)
  77                 free_poll_entry(pwq->inline_entries + i);
  78         while (p) {
  79                 struct poll_table_entry * entry;
  80                 struct poll_table_page *old;
  81
  82                 entry = p->entry;
  83                 do {
  84                         entry--;
  85                         free_poll_entry(entry);
  86                 } while (entry > p->entries);
  87                 old = p;
  88                 p = p->next;
  89                 free_page((unsigned long) old);
  90         }
  91 }
  92
  93 EXPORT_SYMBOL(poll_freewait);
  94
  95 static struct poll_table_entry *poll_get_entry(poll_table *_p)
  96 {
  97         struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt);
  98         struct poll_table_page *table = p->table;
  99
 100         if (p->inline_index < N_INLINE_POLL_ENTRIES)
 101                 return p->inline_entries + p->inline_index++;
 102
 103         if (!table || POLL_TABLE_FULL(table)) {
 104                 struct poll_table_page *new_table;
 105
 106                 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
 107                 if (!new_table) {
 108                         p->error = -ENOMEM;
 109                         __set_current_state(TASK_RUNNING);
 110                         return NULL;
 111                 }
 112                 new_table->entry = new_table->entries;
 113                 new_table->next = table;
 114                 p->table = new_table;
 115                 table = new_table;
 116         }
 117
 118         return table->entry++;
 119 }
 120
 121 /* Add a new entry */
 122 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
 123                                 poll_table *p)
 124 {
 125         struct poll_table_entry *entry = poll_get_entry(p);
 126         if (!entry)
 127                 return;
 128         get_file(filp);
 129         entry->filp = filp;
 130         entry->wait_address = wait_address;
 131         init_waitqueue_entry(&entry->wait, current);
 132         add_wait_queue(wait_address,&entry->wait);
 133 }
 134
 135 #define FDS_IN(fds, n)          (fds->in + n)
 136 #define FDS_OUT(fds, n)         (fds->out + n)
 137 #define FDS_EX(fds, n)          (fds->ex + n)
 138
 139 #define BITS(fds, n)    (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
 140
 141 static int max_select_fd(unsigned long n, fd_set_bits *fds)
 142 {
 143         unsigned long *open_fds;
 144         unsigned long set;
 145         int max;
 146         struct fdtable *fdt;
 147
 148         /* handle last in-complete long-word first */
 149         set = ~(~0UL << (n & (__NFDBITS-1)));
 150         n /= __NFDBITS;
 151         fdt = files_fdtable(current->files);
 152         open_fds = fdt->open_fds->fds_bits+n;
 153         max = 0;
 154         if (set) {
 155                 set &= BITS(fds, n);
 156                 if (set) {
 157                         if (!(set & ~*open_fds))
 158                                 goto get_max;
 159                         return -EBADF;
 160                 }
 161         }
 162         while (n) {
 163                 open_fds--;
 164                 n--;
 165                 set = BITS(fds, n);
 166                 if (!set)
 167                         continue;
 168                 if (set & ~*open_fds)
 169                         return -EBADF;
 170                 if (max)
 171                         continue;
 172 get_max:
 173                 do {
 174                         max++;
 175                         set >>= 1;
 176                 } while (set);
 177                 max += n * __NFDBITS;
 178         }
 179
 180         return max;
 181 }
 182
 183 #define BIT(i)          (1UL << ((i)&(__NFDBITS-1)))
 184 #define MEM(i,m)        ((m)+(unsigned)(i)/__NFDBITS)
 185 #define ISSET(i,m)      (((i)&*(m)) != 0)
 186 #define SET(i,m)        (*(m) |= (i))
 187
 188 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
 189 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
 190 #define POLLEX_SET (POLLPRI)
 191
 192 int do_select(int n, fd_set_bits *fds, s64 *timeout)
 193 {
 194         struct poll_wqueues table;
 195         poll_table *wait;
 196         int retval, i;
 197
 198         rcu_read_lock();
 199         retval = max_select_fd(n, fds);
 200         rcu_read_unlock();
 201
 202         if (retval < 0)
 203                 return retval;
 204         n = retval;
 205
 206         poll_initwait(&table);
 207         wait = &table.pt;
 208         if (!*timeout)
 209                 wait = NULL;
 210         retval = 0;
 211         for (;;) {
 212                 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
 213                 long __timeout;
 214
 215                 set_current_state(TASK_INTERRUPTIBLE);
 216
 217                 inp = fds->in; outp = fds->out; exp = fds->ex;
 218                 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
 219
 220                 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
 221                         unsigned long in, out, ex, all_bits, bit = 1, mask, j;
 222                         unsigned long res_in = 0, res_out = 0, res_ex = 0;
 223                         const struct file_operations *f_op = NULL;
 224                         struct file *file = NULL;
 225
 226                         in = *inp++; out = *outp++; ex = *exp++;
 227                         all_bits = in | out | ex;
 228                         if (all_bits == 0) {
 229                                 i += __NFDBITS;
 230                                 continue;
 231                         }
 232
 233                         for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
 234                                 int fput_needed;
 235                                 if (i >= n)
 236                                         break;
 237                                 if (!(bit & all_bits))
 238                                         continue;
 239                                 file = fget_light(i, &fput_needed);
 240                                 if (file) {
 241                                         f_op = file->f_op;
 242                                         mask = DEFAULT_POLLMASK;
 243                                         if (f_op && f_op->poll)
 244                                                 mask = (*f_op->poll)(file, retval ? NULL : wait);
 245                                         fput_light(file, fput_needed);
 246                                         if ((mask & POLLIN_SET) && (in & bit)) {
 247                                                 res_in |= bit;
 248                                                 retval++;
 249                                         }
 250                                         if ((mask & POLLOUT_SET) && (out & bit)) {
 251                                                 res_out |= bit;
 252                                                 retval++;
 253                                         }
 254                                         if ((mask & POLLEX_SET) && (ex & bit)) {
 255                                                 res_ex |= bit;
 256                                                 retval++;
 257                                         }
 258                                 }
 259                                 cond_resched();
 260                         }
 261                         if (res_in)
 262                                 *rinp = res_in;
 263                         if (res_out)
 264                                 *routp = res_out;
 265                         if (res_ex)
 266                                 *rexp = res_ex;
 267                 }
 268                 wait = NULL;
 269                 if (retval || !*timeout || signal_pending(current))
 270                         break;
 271                 if(table.error) {
 272                         retval = table.error;
 273                         break;
 274                 }
 275
 276                 if (*timeout < 0) {
 277                         /* Wait indefinitely */
 278                         __timeout = MAX_SCHEDULE_TIMEOUT;
 279                 } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT - 1)) {
 280                         /* Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in a loop */
 281                         __timeout = MAX_SCHEDULE_TIMEOUT - 1;
 282                         *timeout -= __timeout;
 283                 } else {
 284                         __timeout = *timeout;
 285                         *timeout = 0;
 286                 }
 287                 __timeout = schedule_timeout(__timeout);
 288                 if (*timeout >= 0)
 289                         *timeout += __timeout;
 290         }
 291         __set_current_state(TASK_RUNNING);
 292
 293         poll_freewait(&table);
 294
 295         return retval;
 296 }
 297
 298 /*
 299  * We can actually return ERESTARTSYS instead of EINTR, but I'd
 300  * like to be certain this leads to no problems. So I return
 301  * EINTR just for safety.
 302  *
 303  * Update: ERESTARTSYS breaks at least the xview clock binary, so
 304  * I'm trying ERESTARTNOHAND which restart only when you want to.
 305  */
 306 #define MAX_SELECT_SECONDS \
 307         ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
 308
 309 static int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 310                            fd_set __user *exp, s64 *timeout)
 311 {
 312         fd_set_bits fds;
 313         char *bits;
 314         int ret, size, max_fdset;
 315         struct fdtable *fdt;
 316         /* Allocate small arguments on the stack to save memory and be faster */
 317         char stack_fds[SELECT_STACK_ALLOC];
 318
 319         ret = -EINVAL;
 320         if (n < 0)
 321                 goto out_nofds;
 322
 323         /* max_fdset can increase, so grab it once to avoid race */
 324         rcu_read_lock();
 325         fdt = files_fdtable(current->files);
 326         max_fdset = fdt->max_fdset;
 327         rcu_read_unlock();
 328         if (n > max_fdset)
 329                 n = max_fdset;
 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         ret = -ENOMEM;
 337         size = FDS_BYTES(n);
 338         if (6*size < SELECT_STACK_ALLOC)
 339                 bits = stack_fds;
 340         else
 341                 bits = kmalloc(6 * size, GFP_KERNEL);
 342         if (!bits)
 343                 goto out_nofds;
 344         fds.in      = (unsigned long *)  bits;
 345         fds.out     = (unsigned long *) (bits +   size);
 346         fds.ex      = (unsigned long *) (bits + 2*size);
 347         fds.res_in  = (unsigned long *) (bits + 3*size);
 348         fds.res_out = (unsigned long *) (bits + 4*size);
 349         fds.res_ex  = (unsigned long *) (bits + 5*size);
 350
 351         if ((ret = get_fd_set(n, inp, fds.in)) ||
 352             (ret = get_fd_set(n, outp, fds.out)) ||
 353             (ret = get_fd_set(n, exp, fds.ex)))
 354                 goto out;
 355         zero_fd_set(n, fds.res_in);
 356         zero_fd_set(n, fds.res_out);
 357         zero_fd_set(n, fds.res_ex);
 358
 359         ret = do_select(n, &fds, timeout);
 360
 361         if (ret < 0)
 362                 goto out;
 363         if (!ret) {
 364                 ret = -ERESTARTNOHAND;
 365                 if (signal_pending(current))
 366                         goto out;
 367                 ret = 0;
 368         }
 369
 370         if (set_fd_set(n, inp, fds.res_in) ||
 371             set_fd_set(n, outp, fds.res_out) ||
 372             set_fd_set(n, exp, fds.res_ex))
 373                 ret = -EFAULT;
 374
 375 out:
 376         if (bits != stack_fds)
 377                 kfree(bits);
 378 out_nofds:
 379         return ret;
 380 }
 381
 382 asmlinkage long sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 383                         fd_set __user *exp, struct timeval __user *tvp)
 384 {
 385         s64 timeout = -1;
 386         struct timeval tv;
 387         int ret;
 388
 389         if (tvp) {
 390                 if (copy_from_user(&tv, tvp, sizeof(tv)))
 391                         return -EFAULT;
 392
 393                 if (tv.tv_sec < 0 || tv.tv_usec < 0)
 394                         return -EINVAL;
 395
 396                 /* Cast to u64 to make GCC stop complaining */
 397                 if ((u64)tv.tv_sec >= (u64)MAX_INT64_SECONDS)
 398                         timeout = -1;   /* infinite */
 399                 else {
 400                         timeout = ROUND_UP(tv.tv_usec, USEC_PER_SEC/HZ);
 401                         timeout += tv.tv_sec * HZ;
 402                 }
 403         }
 404
 405         ret = core_sys_select(n, inp, outp, exp, &timeout);
 406
 407         if (tvp) {
 408                 struct timeval rtv;
 409
 410                 if (current->personality & STICKY_TIMEOUTS)
 411                         goto sticky;
 412                 rtv.tv_usec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ));
 413                 rtv.tv_sec = timeout;
 414                 if (timeval_compare(&rtv, &tv) >= 0)
 415                         rtv = tv;
 416                 if (copy_to_user(tvp, &rtv, sizeof(rtv))) {
 417 sticky:
 418                         /*
 419                          * If an application puts its timeval in read-only
 420                          * memory, we don't want the Linux-specific update to
 421                          * the timeval to cause a fault after the select has
 422                          * completed successfully. However, because we're not
 423                          * updating the timeval, we can't restart the system
 424                          * call.
 425                          */
 426                         if (ret == -ERESTARTNOHAND)
 427                                 ret = -EINTR;
 428                 }
 429         }
 430
 431         return ret;
 432 }
 433
 434 #ifdef TIF_RESTORE_SIGMASK
 435 asmlinkage long sys_pselect7(int n, fd_set __user *inp, fd_set __user *outp,
 436                 fd_set __user *exp, struct timespec __user *tsp,
 437                 const sigset_t __user *sigmask, size_t sigsetsize)
 438 {
 439         s64 timeout = MAX_SCHEDULE_TIMEOUT;
 440         sigset_t ksigmask, sigsaved;
 441         struct timespec ts;
 442         int ret;
 443
 444         if (tsp) {
 445                 if (copy_from_user(&ts, tsp, sizeof(ts)))
 446                         return -EFAULT;
 447
 448                 if (ts.tv_sec < 0 || ts.tv_nsec < 0)
 449                         return -EINVAL;
 450
 451                 /* Cast to u64 to make GCC stop complaining */
 452                 if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
 453                         timeout = -1;   /* infinite */
 454                 else {
 455                         timeout = ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
 456                         timeout += ts.tv_sec * HZ;
 457                 }
 458         }
 459
 460         if (sigmask) {
 461                 /* XXX: Don't preclude handling different sized sigset_t's.  */
 462                 if (sigsetsize != sizeof(sigset_t))
 463                         return -EINVAL;
 464                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 465                         return -EFAULT;
 466
 467                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 468                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 469         }
 470
 471         ret = core_sys_select(n, inp, outp, exp, &timeout);
 472
 473         if (tsp) {
 474                 struct timespec rts;
 475
 476                 if (current->personality & STICKY_TIMEOUTS)
 477                         goto sticky;
 478                 rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
 479                                                 1000;
 480                 rts.tv_sec = timeout;
 481                 if (timespec_compare(&rts, &ts) >= 0)
 482                         rts = ts;
 483                 if (copy_to_user(tsp, &rts, sizeof(rts))) {
 484 sticky:
 485                         /*
 486                          * If an application puts its timeval in read-only
 487                          * memory, we don't want the Linux-specific update to
 488                          * the timeval to cause a fault after the select has
 489                          * completed successfully. However, because we're not
 490                          * updating the timeval, we can't restart the system
 491                          * call.
 492                          */
 493                         if (ret == -ERESTARTNOHAND)
 494                                 ret = -EINTR;
 495                 }
 496         }
 497
 498         if (ret == -ERESTARTNOHAND) {
 499                 /*
 500                  * Don't restore the signal mask yet. Let do_signal() deliver
 501                  * the signal on the way back to userspace, before the signal
 502                  * mask is restored.
 503                  */
 504                 if (sigmask) {
 505                         memcpy(&current->saved_sigmask, &sigsaved,
 506                                         sizeof(sigsaved));
 507                         set_thread_flag(TIF_RESTORE_SIGMASK);
 508                 }
 509         } else if (sigmask)
 510                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 511
 512         return ret;
 513 }
 514
 515 /*
 516  * Most architectures can't handle 7-argument syscalls. So we provide a
 517  * 6-argument version where the sixth argument is a pointer to a structure
 518  * which has a pointer to the sigset_t itself followed by a size_t containing
 519  * the sigset size.
 520  */
 521 asmlinkage long sys_pselect6(int n, fd_set __user *inp, fd_set __user *outp,
 522         fd_set __user *exp, struct timespec __user *tsp, void __user *sig)
 523 {
 524         size_t sigsetsize = 0;
 525         sigset_t __user *up = NULL;
 526
 527         if (sig) {
 528                 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
 529                     || __get_user(up, (sigset_t __user * __user *)sig)
 530                     || __get_user(sigsetsize,
 531                                 (size_t __user *)(sig+sizeof(void *))))
 532                         return -EFAULT;
 533         }
 534
 535         return sys_pselect7(n, inp, outp, exp, tsp, up, sigsetsize);
 536 }
 537 #endif /* TIF_RESTORE_SIGMASK */
 538
 539 struct poll_list {
 540         struct poll_list *next;
 541         int len;
 542         struct pollfd entries[0];
 543 };
 544
 545 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 546
 547 static void do_pollfd(unsigned int num, struct pollfd * fdpage,
 548         poll_table ** pwait, int *count)
 549 {
 550         int i;
 551
 552         for (i = 0; i < num; i++) {
 553                 int fd;
 554                 unsigned int mask;
 555                 struct pollfd *fdp;
 556
 557                 mask = 0;
 558                 fdp = fdpage+i;
 559                 fd = fdp->fd;
 560                 if (fd >= 0) {
 561                         int fput_needed;
 562                         struct file * file = fget_light(fd, &fput_needed);
 563                         mask = POLLNVAL;
 564                         if (file != NULL) {
 565                                 mask = DEFAULT_POLLMASK;
 566                                 if (file->f_op && file->f_op->poll)
 567                                         mask = file->f_op->poll(file, *pwait);
 568                                 mask &= fdp->events | POLLERR | POLLHUP;
 569                                 fput_light(file, fput_needed);
 570                         }
 571                         if (mask) {
 572                                 *pwait = NULL;
 573                                 (*count)++;
 574                         }
 575                 }
 576                 fdp->revents = mask;
 577         }
 578 }
 579
 580 static int do_poll(unsigned int nfds,  struct poll_list *list,
 581                    struct poll_wqueues *wait, s64 *timeout)
 582 {
 583         int count = 0;
 584         poll_table* pt = &wait->pt;
 585
 586         /* Optimise the no-wait case */
 587         if (!(*timeout))
 588                 pt = NULL;
 589
 590         for (;;) {
 591                 struct poll_list *walk;
 592                 long __timeout;
 593
 594                 set_current_state(TASK_INTERRUPTIBLE);
 595                 walk = list;
 596                 while(walk != NULL) {
 597                         do_pollfd( walk->len, walk->entries, &pt, &count);
 598                         walk = walk->next;
 599                 }
 600                 pt = NULL;
 601                 if (count || !*timeout || signal_pending(current))
 602                         break;
 603                 count = wait->error;
 604                 if (count)
 605                         break;
 606
 607                 if (*timeout < 0) {
 608                         /* Wait indefinitely */
 609                         __timeout = MAX_SCHEDULE_TIMEOUT;
 610                 } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT-1)) {
 611                         /*
 612                          * Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in
 613                          * a loop
 614                          */
 615                         __timeout = MAX_SCHEDULE_TIMEOUT - 1;
 616                         *timeout -= __timeout;
 617                 } else {
 618                         __timeout = *timeout;
 619                         *timeout = 0;
 620                 }
 621
 622                 __timeout = schedule_timeout(__timeout);
 623                 if (*timeout >= 0)
 624                         *timeout += __timeout;
 625         }
 626         __set_current_state(TASK_RUNNING);
 627         return count;
 628 }
 629
 630 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
 631                         sizeof(struct pollfd))
 632
 633 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds, s64 *timeout)
 634 {
 635         struct poll_wqueues table;
 636         int fdcount, err;
 637         unsigned int i;
 638         struct poll_list *head;
 639         struct poll_list *walk;
 640         struct fdtable *fdt;
 641         int max_fdset;
 642         /* Allocate small arguments on the stack to save memory and be faster */
 643         char stack_pps[POLL_STACK_ALLOC];
 644         struct poll_list *stack_pp = NULL;
 645
 646         /* Do a sanity check on nfds ... */
 647         rcu_read_lock();
 648         fdt = files_fdtable(current->files);
 649         max_fdset = fdt->max_fdset;
 650         rcu_read_unlock();
 651         if (nfds > max_fdset && nfds > OPEN_MAX)
 652                 return -EINVAL;
 653
 654         poll_initwait(&table);
 655
 656         head = NULL;
 657         walk = NULL;
 658         i = nfds;
 659         err = -ENOMEM;
 660         while(i!=0) {
 661                 struct poll_list *pp;
 662                 int num, size;
 663                 if (stack_pp == NULL)
 664                         num = N_STACK_PPS;
 665                 else
 666                         num = POLLFD_PER_PAGE;
 667                 if (num > i)
 668                         num = i;
 669                 size = sizeof(struct poll_list) + sizeof(struct pollfd)*num;
 670                 if (!stack_pp)
 671                         stack_pp = pp = (struct poll_list *)stack_pps;
 672                 else {
 673                         pp = kmalloc(size, GFP_KERNEL);
 674                         if (!pp)
 675                                 goto out_fds;
 676                 }
 677                 pp->next=NULL;
 678                 pp->len = num;
 679                 if (head == NULL)
 680                         head = pp;
 681                 else
 682                         walk->next = pp;
 683
 684                 walk = pp;
 685                 if (copy_from_user(pp->entries, ufds + nfds-i,
 686                                 sizeof(struct pollfd)*num)) {
 687                         err = -EFAULT;
 688                         goto out_fds;
 689                 }
 690                 i -= pp->len;
 691         }
 692
 693         fdcount = do_poll(nfds, head, &table, timeout);
 694
 695         /* OK, now copy the revents fields back to user space. */
 696         walk = head;
 697         err = -EFAULT;
 698         while(walk != NULL) {
 699                 struct pollfd *fds = walk->entries;
 700                 int j;
 701
 702                 for (j=0; j < walk->len; j++, ufds++) {
 703                         if(__put_user(fds[j].revents, &ufds->revents))
 704                                 goto out_fds;
 705                 }
 706                 walk = walk->next;
 707         }
 708         err = fdcount;
 709         if (!fdcount && signal_pending(current))
 710                 err = -EINTR;
 711 out_fds:
 712         walk = head;
 713         while(walk!=NULL) {
 714                 struct poll_list *pp = walk->next;
 715                 if (walk != stack_pp)
 716                         kfree(walk);
 717                 walk = pp;
 718         }
 719         poll_freewait(&table);
 720         return err;
 721 }
 722
 723 asmlinkage long sys_poll(struct pollfd __user *ufds, unsigned int nfds,
 724                         long timeout_msecs)
 725 {
 726         s64 timeout_jiffies = 0;
 727
 728         if (timeout_msecs) {
 729 #if HZ > 1000
 730                 /* We can only overflow if HZ > 1000 */
 731                 if (timeout_msecs / 1000 > (s64)0x7fffffffffffffffULL / (s64)HZ)
 732                         timeout_jiffies = -1;
 733                 else
 734 #endif
 735                         timeout_jiffies = msecs_to_jiffies(timeout_msecs);
 736         }
 737
 738         return do_sys_poll(ufds, nfds, &timeout_jiffies);
 739 }
 740
 741 #ifdef TIF_RESTORE_SIGMASK
 742 asmlinkage long sys_ppoll(struct pollfd __user *ufds, unsigned int nfds,
 743         struct timespec __user *tsp, const sigset_t __user *sigmask,
 744         size_t sigsetsize)
 745 {
 746         sigset_t ksigmask, sigsaved;
 747         struct timespec ts;
 748         s64 timeout = -1;
 749         int ret;
 750
 751         if (tsp) {
 752                 if (copy_from_user(&ts, tsp, sizeof(ts)))
 753                         return -EFAULT;
 754
 755                 /* Cast to u64 to make GCC stop complaining */
 756                 if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
 757                         timeout = -1;   /* infinite */
 758                 else {
 759                         timeout = ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
 760                         timeout += ts.tv_sec * HZ;
 761                 }
 762         }
 763
 764         if (sigmask) {
 765                 /* XXX: Don't preclude handling different sized sigset_t's.  */
 766                 if (sigsetsize != sizeof(sigset_t))
 767                         return -EINVAL;
 768                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 769                         return -EFAULT;
 770
 771                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 772                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 773         }
 774
 775         ret = do_sys_poll(ufds, nfds, &timeout);
 776
 777         /* We can restart this syscall, usually */
 778         if (ret == -EINTR) {
 779                 /*
 780                  * Don't restore the signal mask yet. Let do_signal() deliver
 781                  * the signal on the way back to userspace, before the signal
 782                  * mask is restored.
 783                  */
 784                 if (sigmask) {
 785                         memcpy(&current->saved_sigmask, &sigsaved,
 786                                         sizeof(sigsaved));
 787                         set_thread_flag(TIF_RESTORE_SIGMASK);
 788                 }
 789                 ret = -ERESTARTNOHAND;
 790         } else if (sigmask)
 791                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 792
 793         if (tsp && timeout >= 0) {
 794                 struct timespec rts;
 795
 796                 if (current->personality & STICKY_TIMEOUTS)
 797                         goto sticky;
 798                 /* Yes, we know it's actually an s64, but it's also positive. */
 799                 rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
 800                                                 1000;
 801                 rts.tv_sec = timeout;
 802                 if (timespec_compare(&rts, &ts) >= 0)
 803                         rts = ts;
 804                 if (copy_to_user(tsp, &rts, sizeof(rts))) {
 805                 sticky:
 806                         /*
 807                          * If an application puts its timeval in read-only
 808                          * memory, we don't want the Linux-specific update to
 809                          * the timeval to cause a fault after the select has
 810                          * completed successfully. However, because we're not
 811                          * updating the timeval, we can't restart the system
 812                          * call.
 813                          */
 814                         if (ret == -ERESTARTNOHAND && timeout >= 0)
 815                                 ret = -EINTR;
 816                 }
 817         }
 818
 819         return ret;
 820 }
 821 #endif /* TIF_RESTORE_SIGMASK */