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