fs/select.c

   1 /*
   2  * This file contains the procedures for the handling of select and poll
   3  *
   4  * Created for Linux based loosely upon Mathius Lattner's minix
   5  * patches by Peter MacDonald. Heavily edited by Linus.
   6  *
   7  *  4 February 1994
   8  *     COFF/ELF binary emulation. If the process has the STICKY_TIMEOUTS
   9  *     flag set in its personality we do *not* modify the given timeout
  10  *     parameter to reflect time remaining.
  11  *
  12  *  24 January 2000
  13  *     Changed sys_poll()/do_poll() to use PAGE_SIZE chunk-based allocation
  14  *     of fds to overcome nfds < 16390 descriptors limit (Tigran Aivazian).
  15  */
  16
  17 #include <linux/kernel.h>
  18 #include <linux/sched.h>
  19 #include <linux/syscalls.h>
  20 #include <linux/export.h>
  21 #include <linux/slab.h>
  22 #include <linux/poll.h>
  23 #include <linux/personality.h> /* for STICKY_TIMEOUTS */
  24 #include <linux/file.h>
  25 #include <linux/fdtable.h>
  26 #include <linux/fs.h>
  27 #include <linux/rcupdate.h>
  28 #include <linux/hrtimer.h>
  29 #include <linux/sched/rt.h>
  30 #include <linux/freezer.h>
  31 #include <net/busy_poll.h>
  32 #include <linux/vmalloc.h>
  33
  34 #include <asm/uaccess.h>
  35
  36
  37 /*
  38  * Estimate expected accuracy in ns from a timeval.
  39  *
  40  * After quite a bit of churning around, we've settled on
  41  * a simple thing of taking 0.1% of the timeout as the
  42  * slack, with a cap of 100 msec.
  43  * "nice" tasks get a 0.5% slack instead.
  44  *
  45  * Consider this comment an open invitation to come up with even
  46  * better solutions..
  47  */
  48
  49 #define MAX_SLACK       (100 * NSEC_PER_MSEC)
  50
  51 static long __estimate_accuracy(struct timespec *tv)
  52 {
  53         long slack;
  54         int divfactor = 1000;
  55
  56         if (tv->tv_sec < 0)
  57                 return 0;
  58
  59         if (task_nice(current) > 0)
  60                 divfactor = divfactor / 5;
  61
  62         if (tv->tv_sec > MAX_SLACK / (NSEC_PER_SEC/divfactor))
  63                 return MAX_SLACK;
  64
  65         slack = tv->tv_nsec / divfactor;
  66         slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
  67
  68         if (slack > MAX_SLACK)
  69                 return MAX_SLACK;
  70
  71         return slack;
  72 }
  73
  74 long select_estimate_accuracy(struct timespec *tv)
  75 {
  76         unsigned long ret;
  77         struct timespec now;
  78
  79         /*
  80          * Realtime tasks get a slack of 0 for obvious reasons.
  81          */
  82
  83         if (rt_task(current))
  84                 return 0;
  85
  86         ktime_get_ts(&now);
  87         now = timespec_sub(*tv, now);
  88         ret = __estimate_accuracy(&now);
  89         if (ret < current->timer_slack_ns)
  90                 return current->timer_slack_ns;
  91         return ret;
  92 }
  93
  94
  95
  96 struct poll_table_page {
  97         struct poll_table_page * next;
  98         struct poll_table_entry * entry;
  99         struct poll_table_entry entries[0];
 100 };
 101
 102 #define POLL_TABLE_FULL(table) \
 103         ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
 104
 105 /*
 106  * Ok, Peter made a complicated, but straightforward multiple_wait() function.
 107  * I have rewritten this, taking some shortcuts: This code may not be easy to
 108  * follow, but it should be free of race-conditions, and it's practical. If you
 109  * understand what I'm doing here, then you understand how the linux
 110  * sleep/wakeup mechanism works.
 111  *
 112  * Two very simple procedures, poll_wait() and poll_freewait() make all the
 113  * work.  poll_wait() is an inline-function defined in <linux/poll.h>,
 114  * as all select/poll functions have to call it to add an entry to the
 115  * poll table.
 116  */
 117 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
 118                        poll_table *p);
 119
 120 void poll_initwait(struct poll_wqueues *pwq)
 121 {
 122         init_poll_funcptr(&pwq->pt, __pollwait);
 123         pwq->polling_task = current;
 124         pwq->triggered = 0;
 125         pwq->error = 0;
 126         pwq->table = NULL;
 127         pwq->inline_index = 0;
 128 }
 129 EXPORT_SYMBOL(poll_initwait);
 130
 131 static void free_poll_entry(struct poll_table_entry *entry)
 132 {
 133         remove_wait_queue(entry->wait_address, &entry->wait);
 134         fput(entry->filp);
 135 }
 136
 137 void poll_freewait(struct poll_wqueues *pwq)
 138 {
 139         struct poll_table_page * p = pwq->table;
 140         int i;
 141         for (i = 0; i < pwq->inline_index; i++)
 142                 free_poll_entry(pwq->inline_entries + i);
 143         while (p) {
 144                 struct poll_table_entry * entry;
 145                 struct poll_table_page *old;
 146
 147                 entry = p->entry;
 148                 do {
 149                         entry--;
 150                         free_poll_entry(entry);
 151                 } while (entry > p->entries);
 152                 old = p;
 153                 p = p->next;
 154                 free_page((unsigned long) old);
 155         }
 156 }
 157 EXPORT_SYMBOL(poll_freewait);
 158
 159 static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
 160 {
 161         struct poll_table_page *table = p->table;
 162
 163         if (p->inline_index < N_INLINE_POLL_ENTRIES)
 164                 return p->inline_entries + p->inline_index++;
 165
 166         if (!table || POLL_TABLE_FULL(table)) {
 167                 struct poll_table_page *new_table;
 168
 169                 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
 170                 if (!new_table) {
 171                         p->error = -ENOMEM;
 172                         return NULL;
 173                 }
 174                 new_table->entry = new_table->entries;
 175                 new_table->next = table;
 176                 p->table = new_table;
 177                 table = new_table;
 178         }
 179
 180         return table->entry++;
 181 }
 182
 183 static int __pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
 184 {
 185         struct poll_wqueues *pwq = wait->private;
 186         DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
 187
 188         /*
 189          * Although this function is called under waitqueue lock, LOCK
 190          * doesn't imply write barrier and the users expect write
 191          * barrier semantics on wakeup functions.  The following
 192          * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
 193          * and is paired with smp_store_mb() in poll_schedule_timeout.
 194          */
 195         smp_wmb();
 196         pwq->triggered = 1;
 197
 198         /*
 199          * Perform the default wake up operation using a dummy
 200          * waitqueue.
 201          *
 202          * TODO: This is hacky but there currently is no interface to
 203          * pass in @sync.  @sync is scheduled to be removed and once
 204          * that happens, wake_up_process() can be used directly.
 205          */
 206         return default_wake_function(&dummy_wait, mode, sync, key);
 207 }
 208
 209 static int pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
 210 {
 211         struct poll_table_entry *entry;
 212
 213         entry = container_of(wait, struct poll_table_entry, wait);
 214         if (key && !((unsigned long)key & entry->key))
 215                 return 0;
 216         return __pollwake(wait, mode, sync, key);
 217 }
 218
 219 /* Add a new entry */
 220 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
 221                                 poll_table *p)
 222 {
 223         struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
 224         struct poll_table_entry *entry = poll_get_entry(pwq);
 225         if (!entry)
 226                 return;
 227         entry->filp = get_file(filp);
 228         entry->wait_address = wait_address;
 229         entry->key = p->_key;
 230         init_waitqueue_func_entry(&entry->wait, pollwake);
 231         entry->wait.private = pwq;
 232         add_wait_queue(wait_address, &entry->wait);
 233 }
 234
 235 int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
 236                           ktime_t *expires, unsigned long slack)
 237 {
 238         int rc = -EINTR;
 239
 240         set_current_state(state);
 241         if (!pwq->triggered)
 242                 rc = schedule_hrtimeout_range(expires, slack, HRTIMER_MODE_ABS);
 243         __set_current_state(TASK_RUNNING);
 244
 245         /*
 246          * Prepare for the next iteration.
 247          *
 248          * The following smp_store_mb() serves two purposes.  First, it's
 249          * the counterpart rmb of the wmb in pollwake() such that data
 250          * written before wake up is always visible after wake up.
 251          * Second, the full barrier guarantees that triggered clearing
 252          * doesn't pass event check of the next iteration.  Note that
 253          * this problem doesn't exist for the first iteration as
 254          * add_wait_queue() has full barrier semantics.
 255          */
 256         smp_store_mb(pwq->triggered, 0);
 257
 258         return rc;
 259 }
 260 EXPORT_SYMBOL(poll_schedule_timeout);
 261
 262 /**
 263  * poll_select_set_timeout - helper function to setup the timeout value
 264  * @to:         pointer to timespec variable for the final timeout
 265  * @sec:        seconds (from user space)
 266  * @nsec:       nanoseconds (from user space)
 267  *
 268  * Note, we do not use a timespec for the user space value here, That
 269  * way we can use the function for timeval and compat interfaces as well.
 270  *
 271  * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
 272  */
 273 int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
 274 {
 275         struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
 276
 277         if (!timespec_valid(&ts))
 278                 return -EINVAL;
 279
 280         /* Optimize for the zero timeout value here */
 281         if (!sec && !nsec) {
 282                 to->tv_sec = to->tv_nsec = 0;
 283         } else {
 284                 ktime_get_ts(to);
 285                 *to = timespec_add_safe(*to, ts);
 286         }
 287         return 0;
 288 }
 289
 290 static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
 291                                       int timeval, int ret)
 292 {
 293         struct timespec rts;
 294         struct timeval rtv;
 295
 296         if (!p)
 297                 return ret;
 298
 299         if (current->personality & STICKY_TIMEOUTS)
 300                 goto sticky;
 301
 302         /* No update for zero timeout */
 303         if (!end_time->tv_sec && !end_time->tv_nsec)
 304                 return ret;
 305
 306         ktime_get_ts(&rts);
 307         rts = timespec_sub(*end_time, rts);
 308         if (rts.tv_sec < 0)
 309                 rts.tv_sec = rts.tv_nsec = 0;
 310
 311         if (timeval) {
 312                 if (sizeof(rtv) > sizeof(rtv.tv_sec) + sizeof(rtv.tv_usec))
 313                         memset(&rtv, 0, sizeof(rtv));
 314                 rtv.tv_sec = rts.tv_sec;
 315                 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
 316
 317                 if (!copy_to_user(p, &rtv, sizeof(rtv)))
 318                         return ret;
 319
 320         } else if (!copy_to_user(p, &rts, sizeof(rts)))
 321                 return ret;
 322
 323         /*
 324          * If an application puts its timeval in read-only memory, we
 325          * don't want the Linux-specific update to the timeval to
 326          * cause a fault after the select has completed
 327          * successfully. However, because we're not updating the
 328          * timeval, we can't restart the system call.
 329          */
 330
 331 sticky:
 332         if (ret == -ERESTARTNOHAND)
 333                 ret = -EINTR;
 334         return ret;
 335 }
 336
 337 #define FDS_IN(fds, n)          (fds->in + n)
 338 #define FDS_OUT(fds, n)         (fds->out + n)
 339 #define FDS_EX(fds, n)          (fds->ex + n)
 340
 341 #define BITS(fds, n)    (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
 342
 343 static int max_select_fd(unsigned long n, fd_set_bits *fds)
 344 {
 345         unsigned long *open_fds;
 346         unsigned long set;
 347         int max;
 348         struct fdtable *fdt;
 349
 350         /* handle last in-complete long-word first */
 351         set = ~(~0UL << (n & (BITS_PER_LONG-1)));
 352         n /= BITS_PER_LONG;
 353         fdt = files_fdtable(current->files);
 354         open_fds = fdt->open_fds + n;
 355         max = 0;
 356         if (set) {
 357                 set &= BITS(fds, n);
 358                 if (set) {
 359                         if (!(set & ~*open_fds))
 360                                 goto get_max;
 361                         return -EBADF;
 362                 }
 363         }
 364         while (n) {
 365                 open_fds--;
 366                 n--;
 367                 set = BITS(fds, n);
 368                 if (!set)
 369                         continue;
 370                 if (set & ~*open_fds)
 371                         return -EBADF;
 372                 if (max)
 373                         continue;
 374 get_max:
 375                 do {
 376                         max++;
 377                         set >>= 1;
 378                 } while (set);
 379                 max += n * BITS_PER_LONG;
 380         }
 381
 382         return max;
 383 }
 384
 385 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
 386 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
 387 #define POLLEX_SET (POLLPRI)
 388
 389 static inline void wait_key_set(poll_table *wait, unsigned long in,
 390                                 unsigned long out, unsigned long bit,
 391                                 unsigned int ll_flag)
 392 {
 393         wait->_key = POLLEX_SET | ll_flag;
 394         if (in & bit)
 395                 wait->_key |= POLLIN_SET;
 396         if (out & bit)
 397                 wait->_key |= POLLOUT_SET;
 398 }
 399
 400 int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
 401 {
 402         ktime_t expire, *to = NULL;
 403         struct poll_wqueues table;
 404         poll_table *wait;
 405         int retval, i, timed_out = 0;
 406         unsigned long slack = 0;
 407         unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
 408         unsigned long busy_end = 0;
 409
 410         rcu_read_lock();
 411         retval = max_select_fd(n, fds);
 412         rcu_read_unlock();
 413
 414         if (retval < 0)
 415                 return retval;
 416         n = retval;
 417
 418         poll_initwait(&table);
 419         wait = &table.pt;
 420         if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
 421                 wait->_qproc = NULL;
 422                 timed_out = 1;
 423         }
 424
 425         if (end_time && !timed_out)
 426                 slack = select_estimate_accuracy(end_time);
 427
 428         retval = 0;
 429         for (;;) {
 430                 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
 431                 bool can_busy_loop = false;
 432
 433                 inp = fds->in; outp = fds->out; exp = fds->ex;
 434                 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
 435
 436                 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
 437                         unsigned long in, out, ex, all_bits, bit = 1, mask, j;
 438                         unsigned long res_in = 0, res_out = 0, res_ex = 0;
 439
 440                         in = *inp++; out = *outp++; ex = *exp++;
 441                         all_bits = in | out | ex;
 442                         if (all_bits == 0) {
 443                                 i += BITS_PER_LONG;
 444                                 continue;
 445                         }
 446
 447                         for (j = 0; j < BITS_PER_LONG; ++j, ++i, bit <<= 1) {
 448                                 struct fd f;
 449                                 if (i >= n)
 450                                         break;
 451                                 if (!(bit & all_bits))
 452                                         continue;
 453                                 f = fdget(i);
 454                                 if (f.file) {
 455                                         const struct file_operations *f_op;
 456                                         f_op = f.file->f_op;
 457                                         mask = DEFAULT_POLLMASK;
 458                                         if (f_op->poll) {
 459                                                 wait_key_set(wait, in, out,
 460                                                              bit, busy_flag);
 461                                                 mask = (*f_op->poll)(f.file, wait);
 462                                         }
 463                                         fdput(f);
 464                                         if ((mask & POLLIN_SET) && (in & bit)) {
 465                                                 res_in |= bit;
 466                                                 retval++;
 467                                                 wait->_qproc = NULL;
 468                                         }
 469                                         if ((mask & POLLOUT_SET) && (out & bit)) {
 470                                                 res_out |= bit;
 471                                                 retval++;
 472                                                 wait->_qproc = NULL;
 473                                         }
 474                                         if ((mask & POLLEX_SET) && (ex & bit)) {
 475                                                 res_ex |= bit;
 476                                                 retval++;
 477                                                 wait->_qproc = NULL;
 478                                         }
 479                                         /* got something, stop busy polling */
 480                                         if (retval) {
 481                                                 can_busy_loop = false;
 482                                                 busy_flag = 0;
 483
 484                                         /*
 485                                          * only remember a returned
 486                                          * POLL_BUSY_LOOP if we asked for it
 487                                          */
 488                                         } else if (busy_flag & mask)
 489                                                 can_busy_loop = true;
 490
 491                                 }
 492                         }
 493                         if (res_in)
 494                                 *rinp = res_in;
 495                         if (res_out)
 496                                 *routp = res_out;
 497                         if (res_ex)
 498                                 *rexp = res_ex;
 499                         cond_resched();
 500                 }
 501                 wait->_qproc = NULL;
 502                 if (retval || timed_out || signal_pending(current))
 503                         break;
 504                 if (table.error) {
 505                         retval = table.error;
 506                         break;
 507                 }
 508
 509                 /* only if found POLL_BUSY_LOOP sockets && not out of time */
 510                 if (can_busy_loop && !need_resched()) {
 511                         if (!busy_end) {
 512                                 busy_end = busy_loop_end_time();
 513                                 continue;
 514                         }
 515                         if (!busy_loop_timeout(busy_end))
 516                                 continue;
 517                 }
 518                 busy_flag = 0;
 519
 520                 /*
 521                  * If this is the first loop and we have a timeout
 522                  * given, then we convert to ktime_t and set the to
 523                  * pointer to the expiry value.
 524                  */
 525                 if (end_time && !to) {
 526                         expire = timespec_to_ktime(*end_time);
 527                         to = &expire;
 528                 }
 529
 530                 if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
 531                                            to, slack))
 532                         timed_out = 1;
 533         }
 534
 535         poll_freewait(&table);
 536
 537         return retval;
 538 }
 539
 540 /*
 541  * We can actually return ERESTARTSYS instead of EINTR, but I'd
 542  * like to be certain this leads to no problems. So I return
 543  * EINTR just for safety.
 544  *
 545  * Update: ERESTARTSYS breaks at least the xview clock binary, so
 546  * I'm trying ERESTARTNOHAND which restart only when you want to.
 547  */
 548 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 549                            fd_set __user *exp, struct timespec *end_time)
 550 {
 551         fd_set_bits fds;
 552         void *bits;
 553         int ret, max_fds;
 554         size_t size, alloc_size;
 555         struct fdtable *fdt;
 556         /* Allocate small arguments on the stack to save memory and be faster */
 557         long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
 558
 559         ret = -EINVAL;
 560         if (n < 0)
 561                 goto out_nofds;
 562
 563         /* max_fds can increase, so grab it once to avoid race */
 564         rcu_read_lock();
 565         fdt = files_fdtable(current->files);
 566         max_fds = fdt->max_fds;
 567         rcu_read_unlock();
 568         if (n > max_fds)
 569                 n = max_fds;
 570
 571         /*
 572          * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
 573          * since we used fdset we need to allocate memory in units of
 574          * long-words.
 575          */
 576         size = FDS_BYTES(n);
 577         bits = stack_fds;
 578         if (size > sizeof(stack_fds) / 6) {
 579                 /* Not enough space in on-stack array; must use kmalloc */
 580                 ret = -ENOMEM;
 581                 if (size > (SIZE_MAX / 6))
 582                         goto out_nofds;
 583
 584                 alloc_size = 6 * size;
 585                 bits = kmalloc(alloc_size, GFP_KERNEL|__GFP_NOWARN);
 586                 if (!bits && alloc_size > PAGE_SIZE)
 587                         bits = vmalloc(alloc_size);
 588
 589                 if (!bits)
 590                         goto out_nofds;
 591         }
 592         fds.in      = bits;
 593         fds.out     = bits +   size;
 594         fds.ex      = bits + 2*size;
 595         fds.res_in  = bits + 3*size;
 596         fds.res_out = bits + 4*size;
 597         fds.res_ex  = bits + 5*size;
 598
 599         if ((ret = get_fd_set(n, inp, fds.in)) ||
 600             (ret = get_fd_set(n, outp, fds.out)) ||
 601             (ret = get_fd_set(n, exp, fds.ex)))
 602                 goto out;
 603         zero_fd_set(n, fds.res_in);
 604         zero_fd_set(n, fds.res_out);
 605         zero_fd_set(n, fds.res_ex);
 606
 607         ret = do_select(n, &fds, end_time);
 608
 609         if (ret < 0)
 610                 goto out;
 611         if (!ret) {
 612                 ret = -ERESTARTNOHAND;
 613                 if (signal_pending(current))
 614                         goto out;
 615                 ret = 0;
 616         }
 617
 618         if (set_fd_set(n, inp, fds.res_in) ||
 619             set_fd_set(n, outp, fds.res_out) ||
 620             set_fd_set(n, exp, fds.res_ex))
 621                 ret = -EFAULT;
 622
 623 out:
 624         if (bits != stack_fds)
 625                 kvfree(bits);
 626 out_nofds:
 627         return ret;
 628 }
 629
 630 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
 631                 fd_set __user *, exp, struct timeval __user *, tvp)
 632 {
 633         struct timespec end_time, *to = NULL;
 634         struct timeval tv;
 635         int ret;
 636
 637         if (tvp) {
 638                 if (copy_from_user(&tv, tvp, sizeof(tv)))
 639                         return -EFAULT;
 640
 641                 to = &end_time;
 642                 if (poll_select_set_timeout(to,
 643                                 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
 644                                 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
 645                         return -EINVAL;
 646         }
 647
 648         ret = core_sys_select(n, inp, outp, exp, to);
 649         ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
 650
 651         return ret;
 652 }
 653
 654 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
 655                        fd_set __user *exp, struct timespec __user *tsp,
 656                        const sigset_t __user *sigmask, size_t sigsetsize)
 657 {
 658         sigset_t ksigmask, sigsaved;
 659         struct timespec ts, end_time, *to = NULL;
 660         int ret;
 661
 662         if (tsp) {
 663                 if (copy_from_user(&ts, tsp, sizeof(ts)))
 664                         return -EFAULT;
 665
 666                 to = &end_time;
 667                 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
 668                         return -EINVAL;
 669         }
 670
 671         if (sigmask) {
 672                 /* XXX: Don't preclude handling different sized sigset_t's.  */
 673                 if (sigsetsize != sizeof(sigset_t))
 674                         return -EINVAL;
 675                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 676                         return -EFAULT;
 677
 678                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 679                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 680         }
 681
 682         ret = core_sys_select(n, inp, outp, exp, to);
 683         ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
 684
 685         if (ret == -ERESTARTNOHAND) {
 686                 /*
 687                  * Don't restore the signal mask yet. Let do_signal() deliver
 688                  * the signal on the way back to userspace, before the signal
 689                  * mask is restored.
 690                  */
 691                 if (sigmask) {
 692                         memcpy(&current->saved_sigmask, &sigsaved,
 693                                         sizeof(sigsaved));
 694                         set_restore_sigmask();
 695                 }
 696         } else if (sigmask)
 697                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 698
 699         return ret;
 700 }
 701
 702 /*
 703  * Most architectures can't handle 7-argument syscalls. So we provide a
 704  * 6-argument version where the sixth argument is a pointer to a structure
 705  * which has a pointer to the sigset_t itself followed by a size_t containing
 706  * the sigset size.
 707  */
 708 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
 709                 fd_set __user *, exp, struct timespec __user *, tsp,
 710                 void __user *, sig)
 711 {
 712         size_t sigsetsize = 0;
 713         sigset_t __user *up = NULL;
 714
 715         if (sig) {
 716                 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
 717                     || __get_user(up, (sigset_t __user * __user *)sig)
 718                     || __get_user(sigsetsize,
 719                                 (size_t __user *)(sig+sizeof(void *))))
 720                         return -EFAULT;
 721         }
 722
 723         return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
 724 }
 725
 726 #ifdef __ARCH_WANT_SYS_OLD_SELECT
 727 struct sel_arg_struct {
 728         unsigned long n;
 729         fd_set __user *inp, *outp, *exp;
 730         struct timeval __user *tvp;
 731 };
 732
 733 SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg)
 734 {
 735         struct sel_arg_struct a;
 736
 737         if (copy_from_user(&a, arg, sizeof(a)))
 738                 return -EFAULT;
 739         return sys_select(a.n, a.inp, a.outp, a.exp, a.tvp);
 740 }
 741 #endif
 742
 743 struct poll_list {
 744         struct poll_list *next;
 745         int len;
 746         struct pollfd entries[0];
 747 };
 748
 749 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 750
 751 /*
 752  * Fish for pollable events on the pollfd->fd file descriptor. We're only
 753  * interested in events matching the pollfd->events mask, and the result
 754  * matching that mask is both recorded in pollfd->revents and returned. The
 755  * pwait poll_table will be used by the fd-provided poll handler for waiting,
 756  * if pwait->_qproc is non-NULL.
 757  */
 758 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait,
 759                                      bool *can_busy_poll,
 760                                      unsigned int busy_flag)
 761 {
 762         unsigned int mask;
 763         int fd;
 764
 765         mask = 0;
 766         fd = pollfd->fd;
 767         if (fd >= 0) {
 768                 struct fd f = fdget(fd);
 769                 mask = POLLNVAL;
 770                 if (f.file) {
 771                         mask = DEFAULT_POLLMASK;
 772                         if (f.file->f_op->poll) {
 773                                 pwait->_key = pollfd->events|POLLERR|POLLHUP;
 774                                 pwait->_key |= busy_flag;
 775                                 mask = f.file->f_op->poll(f.file, pwait);
 776                                 if (mask & busy_flag)
 777                                         *can_busy_poll = true;
 778                         }
 779                         /* Mask out unneeded events. */
 780                         mask &= pollfd->events | POLLERR | POLLHUP;
 781                         fdput(f);
 782                 }
 783         }
 784         pollfd->revents = mask;
 785
 786         return mask;
 787 }
 788
 789 static int do_poll(unsigned int nfds,  struct poll_list *list,
 790                    struct poll_wqueues *wait, struct timespec *end_time)
 791 {
 792         poll_table* pt = &wait->pt;
 793         ktime_t expire, *to = NULL;
 794         int timed_out = 0, count = 0;
 795         unsigned long slack = 0;
 796         unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
 797         unsigned long busy_end = 0;
 798
 799         /* Optimise the no-wait case */
 800         if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
 801                 pt->_qproc = NULL;
 802                 timed_out = 1;
 803         }
 804
 805         if (end_time && !timed_out)
 806                 slack = select_estimate_accuracy(end_time);
 807
 808         for (;;) {
 809                 struct poll_list *walk;
 810                 bool can_busy_loop = false;
 811
 812                 for (walk = list; walk != NULL; walk = walk->next) {
 813                         struct pollfd * pfd, * pfd_end;
 814
 815                         pfd = walk->entries;
 816                         pfd_end = pfd + walk->len;
 817                         for (; pfd != pfd_end; pfd++) {
 818                                 /*
 819                                  * Fish for events. If we found one, record it
 820                                  * and kill poll_table->_qproc, so we don't
 821                                  * needlessly register any other waiters after
 822                                  * this. They'll get immediately deregistered
 823                                  * when we break out and return.
 824                                  */
 825                                 if (do_pollfd(pfd, pt, &can_busy_loop,
 826                                               busy_flag)) {
 827                                         count++;
 828                                         pt->_qproc = NULL;
 829                                         /* found something, stop busy polling */
 830                                         busy_flag = 0;
 831                                         can_busy_loop = false;
 832                                 }
 833                         }
 834                 }
 835                 /*
 836                  * All waiters have already been registered, so don't provide
 837                  * a poll_table->_qproc to them on the next loop iteration.
 838                  */
 839                 pt->_qproc = NULL;
 840                 if (!count) {
 841                         count = wait->error;
 842                         if (signal_pending(current))
 843                                 count = -EINTR;
 844                 }
 845                 if (count || timed_out)
 846                         break;
 847
 848                 /* only if found POLL_BUSY_LOOP sockets && not out of time */
 849                 if (can_busy_loop && !need_resched()) {
 850                         if (!busy_end) {
 851                                 busy_end = busy_loop_end_time();
 852                                 continue;
 853                         }
 854                         if (!busy_loop_timeout(busy_end))
 855                                 continue;
 856                 }
 857                 busy_flag = 0;
 858
 859                 /*
 860                  * If this is the first loop and we have a timeout
 861                  * given, then we convert to ktime_t and set the to
 862                  * pointer to the expiry value.
 863                  */
 864                 if (end_time && !to) {
 865                         expire = timespec_to_ktime(*end_time);
 866                         to = &expire;
 867                 }
 868
 869                 if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
 870                         timed_out = 1;
 871         }
 872         return count;
 873 }
 874
 875 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
 876                         sizeof(struct pollfd))
 877
 878 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
 879                 struct timespec *end_time)
 880 {
 881         struct poll_wqueues table;
 882         int err = -EFAULT, fdcount, len, size;
 883         /* Allocate small arguments on the stack to save memory and be
 884            faster - use long to make sure the buffer is aligned properly
 885            on 64 bit archs to avoid unaligned access */
 886         long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
 887         struct poll_list *const head = (struct poll_list *)stack_pps;
 888         struct poll_list *walk = head;
 889         unsigned long todo = nfds;
 890
 891         if (nfds > rlimit(RLIMIT_NOFILE))
 892                 return -EINVAL;
 893
 894         len = min_t(unsigned int, nfds, N_STACK_PPS);
 895         for (;;) {
 896                 walk->next = NULL;
 897                 walk->len = len;
 898                 if (!len)
 899                         break;
 900
 901                 if (copy_from_user(walk->entries, ufds + nfds-todo,
 902                                         sizeof(struct pollfd) * walk->len))
 903                         goto out_fds;
 904
 905                 todo -= walk->len;
 906                 if (!todo)
 907                         break;
 908
 909                 len = min(todo, POLLFD_PER_PAGE);
 910                 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
 911                 walk = walk->next = kmalloc(size, GFP_KERNEL);
 912                 if (!walk) {
 913                         err = -ENOMEM;
 914                         goto out_fds;
 915                 }
 916         }
 917
 918         poll_initwait(&table);
 919         fdcount = do_poll(nfds, head, &table, end_time);
 920         poll_freewait(&table);
 921
 922         for (walk = head; walk; walk = walk->next) {
 923                 struct pollfd *fds = walk->entries;
 924                 int j;
 925
 926                 for (j = 0; j < walk->len; j++, ufds++)
 927                         if (__put_user(fds[j].revents, &ufds->revents))
 928                                 goto out_fds;
 929         }
 930
 931         err = fdcount;
 932 out_fds:
 933         walk = head->next;
 934         while (walk) {
 935                 struct poll_list *pos = walk;
 936                 walk = walk->next;
 937                 kfree(pos);
 938         }
 939
 940         return err;
 941 }
 942
 943 static long do_restart_poll(struct restart_block *restart_block)
 944 {
 945         struct pollfd __user *ufds = restart_block->poll.ufds;
 946         int nfds = restart_block->poll.nfds;
 947         struct timespec *to = NULL, end_time;
 948         int ret;
 949
 950         if (restart_block->poll.has_timeout) {
 951                 end_time.tv_sec = restart_block->poll.tv_sec;
 952                 end_time.tv_nsec = restart_block->poll.tv_nsec;
 953                 to = &end_time;
 954         }
 955
 956         ret = do_sys_poll(ufds, nfds, to);
 957
 958         if (ret == -EINTR) {
 959                 restart_block->fn = do_restart_poll;
 960                 ret = -ERESTART_RESTARTBLOCK;
 961         }
 962         return ret;
 963 }
 964
 965 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
 966                 int, timeout_msecs)
 967 {
 968         struct timespec end_time, *to = NULL;
 969         int ret;
 970
 971         if (timeout_msecs >= 0) {
 972                 to = &end_time;
 973                 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
 974                         NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
 975         }
 976
 977         ret = do_sys_poll(ufds, nfds, to);
 978
 979         if (ret == -EINTR) {
 980                 struct restart_block *restart_block;
 981
 982                 restart_block = &current->restart_block;
 983                 restart_block->fn = do_restart_poll;
 984                 restart_block->poll.ufds = ufds;
 985                 restart_block->poll.nfds = nfds;
 986
 987                 if (timeout_msecs >= 0) {
 988                         restart_block->poll.tv_sec = end_time.tv_sec;
 989                         restart_block->poll.tv_nsec = end_time.tv_nsec;
 990                         restart_block->poll.has_timeout = 1;
 991                 } else
 992                         restart_block->poll.has_timeout = 0;
 993
 994                 ret = -ERESTART_RESTARTBLOCK;
 995         }
 996         return ret;
 997 }
 998
 999 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
1000                 struct timespec __user *, tsp, const sigset_t __user *, sigmask,
1001                 size_t, sigsetsize)
1002 {
1003         sigset_t ksigmask, sigsaved;
1004         struct timespec ts, end_time, *to = NULL;
1005         int ret;
1006
1007         if (tsp) {
1008                 if (copy_from_user(&ts, tsp, sizeof(ts)))
1009                         return -EFAULT;
1010
1011                 to = &end_time;
1012                 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1013                         return -EINVAL;
1014         }
1015
1016         if (sigmask) {
1017                 /* XXX: Don't preclude handling different sized sigset_t's.  */
1018                 if (sigsetsize != sizeof(sigset_t))
1019                         return -EINVAL;
1020                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1021                         return -EFAULT;
1022
1023                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
1024                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1025         }
1026
1027         ret = do_sys_poll(ufds, nfds, to);
1028
1029         /* We can restart this syscall, usually */
1030         if (ret == -EINTR) {
1031                 /*
1032                  * Don't restore the signal mask yet. Let do_signal() deliver
1033                  * the signal on the way back to userspace, before the signal
1034                  * mask is restored.
1035                  */
1036                 if (sigmask) {
1037                         memcpy(&current->saved_sigmask, &sigsaved,
1038                                         sizeof(sigsaved));
1039                         set_restore_sigmask();
1040                 }
1041                 ret = -ERESTARTNOHAND;
1042         } else if (sigmask)
1043                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1044
1045         ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
1046
1047         return ret;
1048 }