SAFE public projects git trees. - safe/jmp/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                         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                                 if (i >= n)
 235                                         break;
 236                                 if (!(bit & all_bits))
 237                                         continue;
 238                                 file = fget(i);
 239                                 if (file) {
 240                                         f_op = file->f_op;
 241                                         mask = DEFAULT_POLLMASK;
 242                                         if (f_op && f_op->poll)
 243                                                 mask = (*f_op->poll)(file, retval ? NULL : wait);
 244                                         fput(file);
 245                                         if ((mask & POLLIN_SET) && (in & bit)) {
 246                                                 res_in |= bit;
 247                                                 retval++;
 248                                         }
 249                                         if ((mask & POLLOUT_SET) && (out & bit)) {
 250                                                 res_out |= bit;
 251                                                 retval++;
 252                                         }
 253                                         if ((mask & POLLEX_SET) && (ex & bit)) {
 254                                                 res_ex |= bit;
 255                                                 retval++;
 256                                         }
 257                                 }
 258                                 cond_resched();
 259                         }
 260                         if (res_in)
 261                                 *rinp = res_in;
 262                         if (res_out)
 263                                 *routp = res_out;
 264                         if (res_ex)
 265                                 *rexp = res_ex;
 266                 }
 267                 wait = NULL;
 268                 if (retval || !*timeout || signal_pending(current))
 269                         break;
 270                 if(table.error) {
 271                         retval = table.error;
 272                         break;
 273                 }
 274
 275                 if (*timeout < 0) {
 276                         /* Wait indefinitely */
 277                         __timeout = MAX_SCHEDULE_TIMEOUT;
 278                 } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT - 1)) {
 279                         /* Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in a loop */
 280                         __timeout = MAX_SCHEDULE_TIMEOUT - 1;
 281                         *timeout -= __timeout;
 282                 } else {
 283                         __timeout = *timeout;
 284                         *timeout = 0;
 285                 }
 286                 __timeout = schedule_timeout(__timeout);
 287                 if (*timeout >= 0)
 288                         *timeout += __timeout;
 289         }
 290         __set_current_state(TASK_RUNNING);
 291
 292         poll_freewait(&table);
 293
 294         return retval;
 295 }
 296
 297 /*
 298  * We can actually return ERESTARTSYS instead of EINTR, but I'd
 299  * like to be certain this leads to no problems. So I return
 300  * EINTR just for safety.
 301  *
 302  * Update: ERESTARTSYS breaks at least the xview clock binary, so
 303  * I'm trying ERESTARTNOHAND which restart only when you want to.
 304  */
 305 #define MAX_SELECT_SECONDS \
 306         ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
 307
 308 static int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 309                            fd_set __user *exp, s64 *timeout)
 310 {
 311         fd_set_bits fds;
 312         char *bits;
 313         int ret, size, max_fdset;
 314         struct fdtable *fdt;
 315         /* Allocate small arguments on the stack to save memory and be faster */
 316         char stack_fds[SELECT_STACK_ALLOC];
 317
 318         ret = -EINVAL;
 319         if (n < 0)
 320                 goto out_nofds;
 321
 322         /* max_fdset can increase, so grab it once to avoid race */
 323         rcu_read_lock();
 324         fdt = files_fdtable(current->files);
 325         max_fdset = fdt->max_fdset;
 326         rcu_read_unlock();
 327         if (n > max_fdset)
 328                 n = max_fdset;
 329
 330         /*
 331          * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
 332          * since we used fdset we need to allocate memory in units of
 333          * long-words.
 334          */
 335         ret = -ENOMEM;
 336         size = FDS_BYTES(n);
 337         if (6*size < SELECT_STACK_ALLOC)
 338                 bits = stack_fds;
 339         else
 340                 bits = kmalloc(6 * size, GFP_KERNEL);
 341         if (!bits)
 342                 goto out_nofds;
 343         fds.in      = (unsigned long *)  bits;
 344         fds.out     = (unsigned long *) (bits +   size);
 345         fds.ex      = (unsigned long *) (bits + 2*size);
 346         fds.res_in  = (unsigned long *) (bits + 3*size);
 347         fds.res_out = (unsigned long *) (bits + 4*size);
 348         fds.res_ex  = (unsigned long *) (bits + 5*size);
 349
 350         if ((ret = get_fd_set(n, inp, fds.in)) ||
 351             (ret = get_fd_set(n, outp, fds.out)) ||
 352             (ret = get_fd_set(n, exp, fds.ex)))
 353                 goto out;
 354         zero_fd_set(n, fds.res_in);
 355         zero_fd_set(n, fds.res_out);
 356         zero_fd_set(n, fds.res_ex);
 357
 358         ret = do_select(n, &fds, timeout);
 359
 360         if (ret < 0)
 361                 goto out;
 362         if (!ret) {
 363                 ret = -ERESTARTNOHAND;
 364                 if (signal_pending(current))
 365                         goto out;
 366                 ret = 0;
 367         }
 368
 369         if (set_fd_set(n, inp, fds.res_in) ||
 370             set_fd_set(n, outp, fds.res_out) ||
 371             set_fd_set(n, exp, fds.res_ex))
 372                 ret = -EFAULT;
 373
 374 out:
 375         if (bits != stack_fds)
 376                 kfree(bits);
 377 out_nofds:
 378         return ret;
 379 }
 380
 381 asmlinkage long sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 382                         fd_set __user *exp, struct timeval __user *tvp)
 383 {
 384         s64 timeout = -1;
 385         struct timeval tv;
 386         int ret;
 387
 388         if (tvp) {
 389                 if (copy_from_user(&tv, tvp, sizeof(tv)))
 390                         return -EFAULT;
 391
 392                 if (tv.tv_sec < 0 || tv.tv_usec < 0)
 393                         return -EINVAL;
 394
 395                 /* Cast to u64 to make GCC stop complaining */
 396                 if ((u64)tv.tv_sec >= (u64)MAX_INT64_SECONDS)
 397                         timeout = -1;   /* infinite */
 398                 else {
 399                         timeout = ROUND_UP(tv.tv_usec, USEC_PER_SEC/HZ);
 400                         timeout += tv.tv_sec * HZ;
 401                 }
 402         }
 403
 404         ret = core_sys_select(n, inp, outp, exp, &timeout);
 405
 406         if (tvp) {
 407                 struct timeval rtv;
 408
 409                 if (current->personality & STICKY_TIMEOUTS)
 410                         goto sticky;
 411                 rtv.tv_usec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ));
 412                 rtv.tv_sec = timeout;
 413                 if (timeval_compare(&rtv, &tv) >= 0)
 414                         rtv = tv;
 415                 if (copy_to_user(tvp, &rtv, sizeof(rtv))) {
 416 sticky:
 417                         /*
 418                          * If an application puts its timeval in read-only
 419                          * memory, we don't want the Linux-specific update to
 420                          * the timeval to cause a fault after the select has
 421                          * completed successfully. However, because we're not
 422                          * updating the timeval, we can't restart the system
 423                          * call.
 424                          */
 425                         if (ret == -ERESTARTNOHAND)
 426                                 ret = -EINTR;
 427                 }
 428         }
 429
 430         return ret;
 431 }
 432
 433 #ifdef TIF_RESTORE_SIGMASK
 434 asmlinkage long sys_pselect7(int n, fd_set __user *inp, fd_set __user *outp,
 435                 fd_set __user *exp, struct timespec __user *tsp,
 436                 const sigset_t __user *sigmask, size_t sigsetsize)
 437 {
 438         s64 timeout = MAX_SCHEDULE_TIMEOUT;
 439         sigset_t ksigmask, sigsaved;
 440         struct timespec ts;
 441         int ret;
 442
 443         if (tsp) {
 444                 if (copy_from_user(&ts, tsp, sizeof(ts)))
 445                         return -EFAULT;
 446
 447                 if (ts.tv_sec < 0 || ts.tv_nsec < 0)
 448                         return -EINVAL;
 449
 450                 /* Cast to u64 to make GCC stop complaining */
 451                 if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
 452                         timeout = -1;   /* infinite */
 453                 else {
 454                         timeout = ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
 455                         timeout += ts.tv_sec * HZ;
 456                 }
 457         }
 458
 459         if (sigmask) {
 460                 /* XXX: Don't preclude handling different sized sigset_t's.  */
 461                 if (sigsetsize != sizeof(sigset_t))
 462                         return -EINVAL;
 463                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 464                         return -EFAULT;
 465
 466                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 467                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 468         }
 469
 470         ret = core_sys_select(n, inp, outp, exp, &timeout);
 471
 472         if (tsp) {
 473                 struct timespec rts;
 474
 475                 if (current->personality & STICKY_TIMEOUTS)
 476                         goto sticky;
 477                 rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
 478                                                 1000;
 479                 rts.tv_sec = timeout;
 480                 if (timespec_compare(&rts, &ts) >= 0)
 481                         rts = ts;
 482                 if (copy_to_user(tsp, &rts, sizeof(rts))) {
 483 sticky:
 484                         /*
 485                          * If an application puts its timeval in read-only
 486                          * memory, we don't want the Linux-specific update to
 487                          * the timeval to cause a fault after the select has
 488                          * completed successfully. However, because we're not
 489                          * updating the timeval, we can't restart the system
 490                          * call.
 491                          */
 492                         if (ret == -ERESTARTNOHAND)
 493                                 ret = -EINTR;
 494                 }
 495         }
 496
 497         if (ret == -ERESTARTNOHAND) {
 498                 /*
 499                  * Don't restore the signal mask yet. Let do_signal() deliver
 500                  * the signal on the way back to userspace, before the signal
 501                  * mask is restored.
 502                  */
 503                 if (sigmask) {
 504                         memcpy(&current->saved_sigmask, &sigsaved,
 505                                         sizeof(sigsaved));
 506                         set_thread_flag(TIF_RESTORE_SIGMASK);
 507                 }
 508         } else if (sigmask)
 509                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 510
 511         return ret;
 512 }
 513
 514 /*
 515  * Most architectures can't handle 7-argument syscalls. So we provide a
 516  * 6-argument version where the sixth argument is a pointer to a structure
 517  * which has a pointer to the sigset_t itself followed by a size_t containing
 518  * the sigset size.
 519  */
 520 asmlinkage long sys_pselect6(int n, fd_set __user *inp, fd_set __user *outp,
 521         fd_set __user *exp, struct timespec __user *tsp, void __user *sig)
 522 {
 523         size_t sigsetsize = 0;
 524         sigset_t __user *up = NULL;
 525
 526         if (sig) {
 527                 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
 528                     || __get_user(up, (sigset_t __user * __user *)sig)
 529                     || __get_user(sigsetsize,
 530                                 (size_t __user *)(sig+sizeof(void *))))
 531                         return -EFAULT;
 532         }
 533
 534         return sys_pselect7(n, inp, outp, exp, tsp, up, sigsetsize);
 535 }
 536 #endif /* TIF_RESTORE_SIGMASK */
 537
 538 struct poll_list {
 539         struct poll_list *next;
 540         int len;
 541         struct pollfd entries[0];
 542 };
 543
 544 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 545
 546 static void do_pollfd(unsigned int num, struct pollfd * fdpage,
 547         poll_table ** pwait, int *count)
 548 {
 549         int i;
 550
 551         for (i = 0; i < num; i++) {
 552                 int fd;
 553                 unsigned int mask;
 554                 struct pollfd *fdp;
 555
 556                 mask = 0;
 557                 fdp = fdpage+i;
 558                 fd = fdp->fd;
 559                 if (fd >= 0) {
 560                         struct file * file = fget(fd);
 561                         mask = POLLNVAL;
 562                         if (file != NULL) {
 563                                 mask = DEFAULT_POLLMASK;
 564                                 if (file->f_op && file->f_op->poll)
 565                                         mask = file->f_op->poll(file, *pwait);
 566                                 mask &= fdp->events | POLLERR | POLLHUP;
 567                                 fput(file);
 568                         }
 569                         if (mask) {
 570                                 *pwait = NULL;
 571                                 (*count)++;
 572                         }
 573                 }
 574                 fdp->revents = mask;
 575         }
 576 }
 577
 578 static int do_poll(unsigned int nfds,  struct poll_list *list,
 579                    struct poll_wqueues *wait, s64 *timeout)
 580 {
 581         int count = 0;
 582         poll_table* pt = &wait->pt;
 583
 584         /* Optimise the no-wait case */
 585         if (!(*timeout))
 586                 pt = NULL;
 587
 588         for (;;) {
 589                 struct poll_list *walk;
 590                 long __timeout;
 591
 592                 set_current_state(TASK_INTERRUPTIBLE);
 593                 walk = list;
 594                 while(walk != NULL) {
 595                         do_pollfd( walk->len, walk->entries, &pt, &count);
 596                         walk = walk->next;
 597                 }
 598                 pt = NULL;
 599                 if (count || !*timeout || signal_pending(current))
 600                         break;
 601                 count = wait->error;
 602                 if (count)
 603                         break;
 604
 605                 if (*timeout < 0) {
 606                         /* Wait indefinitely */
 607                         __timeout = MAX_SCHEDULE_TIMEOUT;
 608                 } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT-1)) {
 609                         /*
 610                          * Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in
 611                          * a loop
 612                          */
 613                         __timeout = MAX_SCHEDULE_TIMEOUT - 1;
 614                         *timeout -= __timeout;
 615                 } else {
 616                         __timeout = *timeout;
 617                         *timeout = 0;
 618                 }
 619
 620                 __timeout = schedule_timeout(__timeout);
 621                 if (*timeout >= 0)
 622                         *timeout += __timeout;
 623         }
 624         __set_current_state(TASK_RUNNING);
 625         return count;
 626 }
 627
 628 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
 629                         sizeof(struct pollfd))
 630
 631 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds, s64 *timeout)
 632 {
 633         struct poll_wqueues table;
 634         int fdcount, err;
 635         unsigned int i;
 636         struct poll_list *head;
 637         struct poll_list *walk;
 638         struct fdtable *fdt;
 639         int max_fdset;
 640         /* Allocate small arguments on the stack to save memory and be faster */
 641         char stack_pps[POLL_STACK_ALLOC];
 642         struct poll_list *stack_pp = NULL;
 643
 644         /* Do a sanity check on nfds ... */
 645         rcu_read_lock();
 646         fdt = files_fdtable(current->files);
 647         max_fdset = fdt->max_fdset;
 648         rcu_read_unlock();
 649         if (nfds > max_fdset && nfds > OPEN_MAX)
 650                 return -EINVAL;
 651
 652         poll_initwait(&table);
 653
 654         head = NULL;
 655         walk = NULL;
 656         i = nfds;
 657         err = -ENOMEM;
 658         while(i!=0) {
 659                 struct poll_list *pp;
 660                 int num, size;
 661                 if (stack_pp == NULL)
 662                         num = N_STACK_PPS;
 663                 else
 664                         num = POLLFD_PER_PAGE;
 665                 if (num > i)
 666                         num = i;
 667                 size = sizeof(struct poll_list) + sizeof(struct pollfd)*num;
 668                 if (!stack_pp)
 669                         stack_pp = pp = (struct poll_list *)stack_pps;
 670                 else {
 671                         pp = kmalloc(size, GFP_KERNEL);
 672                         if (!pp)
 673                                 goto out_fds;
 674                 }
 675                 pp->next=NULL;
 676                 pp->len = num;
 677                 if (head == NULL)
 678                         head = pp;
 679                 else
 680                         walk->next = pp;
 681
 682                 walk = pp;
 683                 if (copy_from_user(pp->entries, ufds + nfds-i,
 684                                 sizeof(struct pollfd)*num)) {
 685                         err = -EFAULT;
 686                         goto out_fds;
 687                 }
 688                 i -= pp->len;
 689         }
 690
 691         fdcount = do_poll(nfds, head, &table, timeout);
 692
 693         /* OK, now copy the revents fields back to user space. */
 694         walk = head;
 695         err = -EFAULT;
 696         while(walk != NULL) {
 697                 struct pollfd *fds = walk->entries;
 698                 int j;
 699
 700                 for (j=0; j < walk->len; j++, ufds++) {
 701                         if(__put_user(fds[j].revents, &ufds->revents))
 702                                 goto out_fds;
 703                 }
 704                 walk = walk->next;
 705         }
 706         err = fdcount;
 707         if (!fdcount && signal_pending(current))
 708                 err = -EINTR;
 709 out_fds:
 710         walk = head;
 711         while(walk!=NULL) {
 712                 struct poll_list *pp = walk->next;
 713                 if (walk != stack_pp)
 714                         kfree(walk);
 715                 walk = pp;
 716         }
 717         poll_freewait(&table);
 718         return err;
 719 }
 720
 721 asmlinkage long sys_poll(struct pollfd __user *ufds, unsigned int nfds,
 722                         long timeout_msecs)
 723 {
 724         s64 timeout_jiffies = 0;
 725
 726         if (timeout_msecs) {
 727 #if HZ > 1000
 728                 /* We can only overflow if HZ > 1000 */
 729                 if (timeout_msecs / 1000 > (s64)0x7fffffffffffffffULL / (s64)HZ)
 730                         timeout_jiffies = -1;
 731                 else
 732 #endif
 733                         timeout_jiffies = msecs_to_jiffies(timeout_msecs);
 734         }
 735
 736         return do_sys_poll(ufds, nfds, &timeout_jiffies);
 737 }
 738
 739 #ifdef TIF_RESTORE_SIGMASK
 740 asmlinkage long sys_ppoll(struct pollfd __user *ufds, unsigned int nfds,
 741         struct timespec __user *tsp, const sigset_t __user *sigmask,
 742         size_t sigsetsize)
 743 {
 744         sigset_t ksigmask, sigsaved;
 745         struct timespec ts;
 746         s64 timeout = -1;
 747         int ret;
 748
 749         if (tsp) {
 750                 if (copy_from_user(&ts, tsp, sizeof(ts)))
 751                         return -EFAULT;
 752
 753                 /* Cast to u64 to make GCC stop complaining */
 754                 if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
 755                         timeout = -1;   /* infinite */
 756                 else {
 757                         timeout = ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
 758                         timeout += ts.tv_sec * HZ;
 759                 }
 760         }
 761
 762         if (sigmask) {
 763                 /* XXX: Don't preclude handling different sized sigset_t's.  */
 764                 if (sigsetsize != sizeof(sigset_t))
 765                         return -EINVAL;
 766                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 767                         return -EFAULT;
 768
 769                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 770                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 771         }
 772
 773         ret = do_sys_poll(ufds, nfds, &timeout);
 774
 775         /* We can restart this syscall, usually */
 776         if (ret == -EINTR) {
 777                 /*
 778                  * Don't restore the signal mask yet. Let do_signal() deliver
 779                  * the signal on the way back to userspace, before the signal
 780                  * mask is restored.
 781                  */
 782                 if (sigmask) {
 783                         memcpy(&current->saved_sigmask, &sigsaved,
 784                                         sizeof(sigsaved));
 785                         set_thread_flag(TIF_RESTORE_SIGMASK);
 786                 }
 787                 ret = -ERESTARTNOHAND;
 788         } else if (sigmask)
 789                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 790
 791         if (tsp && timeout >= 0) {
 792                 struct timespec rts;
 793
 794                 if (current->personality & STICKY_TIMEOUTS)
 795                         goto sticky;
 796                 /* Yes, we know it's actually an s64, but it's also positive. */
 797                 rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
 798                                                 1000;
 799                 rts.tv_sec = timeout;
 800                 if (timespec_compare(&rts, &ts) >= 0)
 801                         rts = ts;
 802                 if (copy_to_user(tsp, &rts, sizeof(rts))) {
 803                 sticky:
 804                         /*
 805                          * If an application puts its timeval in read-only
 806                          * memory, we don't want the Linux-specific update to
 807                          * the timeval to cause a fault after the select has
 808                          * completed successfully. However, because we're not
 809                          * updating the timeval, we can't restart the system
 810                          * call.
 811                          */
 812                         if (ret == -ERESTARTNOHAND && timeout >= 0)
 813                                 ret = -EINTR;
 814                 }
 815         }
 816
 817         return ret;
 818 }
 819 #endif /* TIF_RESTORE_SIGMASK */