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
  33 #include <asm/uaccess.h>
  34
  35
  36 /*
  37  * Estimate expected accuracy in ns from a timeval.
  38  *
  39  * After quite a bit of churning around, we've settled on
  40  * a simple thing of taking 0.1% of the timeout as the
  41  * slack, with a cap of 100 msec.
  42  * "nice" tasks get a 0.5% slack instead.
  43  *
  44  * Consider this comment an open invitation to come up with even
  45  * better solutions..
  46  */
  47
  48 #define MAX_SLACK       (100 * NSEC_PER_MSEC)
  49
  50 static long __estimate_accuracy(struct timespec *tv)
  51 {
  52         long slack;
  53         int divfactor = 1000;
  54
  55         if (tv->tv_sec < 0)
  56                 return 0;
  57
  58         if (task_nice(current) > 0)
  59                 divfactor = divfactor / 5;
  60
  61         if (tv->tv_sec > MAX_SLACK / (NSEC_PER_SEC/divfactor))
  62                 return MAX_SLACK;
  63
  64         slack = tv->tv_nsec / divfactor;
  65         slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
  66
  67         if (slack > MAX_SLACK)
  68                 return MAX_SLACK;
  69
  70         return slack;
  71 }
  72
  73 long select_estimate_accuracy(struct timespec *tv)
  74 {
  75         unsigned long ret;
  76         struct timespec now;
  77
  78         /*
  79          * Realtime tasks get a slack of 0 for obvious reasons.
  80          */
  81
  82         if (rt_task(current))
  83                 return 0;
  84
  85         ktime_get_ts(&now);
  86         now = timespec_sub(*tv, now);
  87         ret = __estimate_accuracy(&now);
  88         if (ret < current->timer_slack_ns)
  89                 return current->timer_slack_ns;
  90         return ret;
  91 }
  92
  93
  94
  95 struct poll_table_page {
  96         struct poll_table_page * next;
  97         struct poll_table_entry * entry;
  98         struct poll_table_entry entries[0];
  99 };
 100
 101 #define POLL_TABLE_FULL(table) \
 102         ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
 103
 104 /*
 105  * Ok, Peter made a complicated, but straightforward multiple_wait() function.
 106  * I have rewritten this, taking some shortcuts: This code may not be easy to
 107  * follow, but it should be free of race-conditions, and it's practical. If you
 108  * understand what I'm doing here, then you understand how the linux
 109  * sleep/wakeup mechanism works.
 110  *
 111  * Two very simple procedures, poll_wait() and poll_freewait() make all the
 112  * work.  poll_wait() is an inline-function defined in <linux/poll.h>,
 113  * as all select/poll functions have to call it to add an entry to the
 114  * poll table.
 115  */
 116 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
 117                        poll_table *p);
 118
 119 void poll_initwait(struct poll_wqueues *pwq)
 120 {
 121         init_poll_funcptr(&pwq->pt, __pollwait);
 122         pwq->polling_task = current;
 123         pwq->triggered = 0;
 124         pwq->error = 0;
 125         pwq->table = NULL;
 126         pwq->inline_index = 0;
 127 }
 128 EXPORT_SYMBOL(poll_initwait);
 129
 130 static void free_poll_entry(struct poll_table_entry *entry)
 131 {
 132         remove_wait_queue(entry->wait_address, &entry->wait);
 133         fput(entry->filp);
 134 }
 135
 136 void poll_freewait(struct poll_wqueues *pwq)
 137 {
 138         struct poll_table_page * p = pwq->table;
 139         int i;
 140         for (i = 0; i < pwq->inline_index; i++)
 141                 free_poll_entry(pwq->inline_entries + i);
 142         while (p) {
 143                 struct poll_table_entry * entry;
 144                 struct poll_table_page *old;
 145
 146                 entry = p->entry;
 147                 do {
 148                         entry--;
 149                         free_poll_entry(entry);
 150                 } while (entry > p->entries);
 151                 old = p;
 152                 p = p->next;
 153                 free_page((unsigned long) old);
 154         }
 155 }
 156 EXPORT_SYMBOL(poll_freewait);
 157
 158 static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p)
 159 {
 160         struct poll_table_page *table = p->table;
 161
 162         if (p->inline_index < N_INLINE_POLL_ENTRIES)
 163                 return p->inline_entries + p->inline_index++;
 164
 165         if (!table || POLL_TABLE_FULL(table)) {
 166                 struct poll_table_page *new_table;
 167
 168                 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
 169                 if (!new_table) {
 170                         p->error = -ENOMEM;
 171                         return NULL;
 172                 }
 173                 new_table->entry = new_table->entries;
 174                 new_table->next = table;
 175                 p->table = new_table;
 176                 table = new_table;
 177         }
 178
 179         return table->entry++;
 180 }
 181
 182 static int __pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
 183 {
 184         struct poll_wqueues *pwq = wait->private;
 185         DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task);
 186
 187         /*
 188          * Although this function is called under waitqueue lock, LOCK
 189          * doesn't imply write barrier and the users expect write
 190          * barrier semantics on wakeup functions.  The following
 191          * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
 192          * and is paired with set_mb() in poll_schedule_timeout.
 193          */
 194         smp_wmb();
 195         pwq->triggered = 1;
 196
 197         /*
 198          * Perform the default wake up operation using a dummy
 199          * waitqueue.
 200          *
 201          * TODO: This is hacky but there currently is no interface to
 202          * pass in @sync.  @sync is scheduled to be removed and once
 203          * that happens, wake_up_process() can be used directly.
 204          */
 205         return default_wake_function(&dummy_wait, mode, sync, key);
 206 }
 207
 208 static int pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key)
 209 {
 210         struct poll_table_entry *entry;
 211
 212         entry = container_of(wait, struct poll_table_entry, wait);
 213         if (key && !((unsigned long)key & entry->key))
 214                 return 0;
 215         return __pollwake(wait, mode, sync, key);
 216 }
 217
 218 /* Add a new entry */
 219 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
 220                                 poll_table *p)
 221 {
 222         struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt);
 223         struct poll_table_entry *entry = poll_get_entry(pwq);
 224         if (!entry)
 225                 return;
 226         entry->filp = get_file(filp);
 227         entry->wait_address = wait_address;
 228         entry->key = p->_key;
 229         init_waitqueue_func_entry(&entry->wait, pollwake);
 230         entry->wait.private = pwq;
 231         add_wait_queue(wait_address, &entry->wait);
 232 }
 233
 234 int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
 235                           ktime_t *expires, unsigned long slack)
 236 {
 237         int rc = -EINTR;
 238
 239         set_current_state(state);
 240         if (!pwq->triggered)
 241                 rc = freezable_schedule_hrtimeout_range(expires, slack,
 242                                                         HRTIMER_MODE_ABS);
 243         __set_current_state(TASK_RUNNING);
 244
 245         /*
 246          * Prepare for the next iteration.
 247          *
 248          * The following set_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         set_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 && 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         unsigned int 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                 bits = kmalloc(6 * size, GFP_KERNEL);
 582                 if (!bits)
 583                         goto out_nofds;
 584         }
 585         fds.in      = bits;
 586         fds.out     = bits +   size;
 587         fds.ex      = bits + 2*size;
 588         fds.res_in  = bits + 3*size;
 589         fds.res_out = bits + 4*size;
 590         fds.res_ex  = bits + 5*size;
 591
 592         if ((ret = get_fd_set(n, inp, fds.in)) ||
 593             (ret = get_fd_set(n, outp, fds.out)) ||
 594             (ret = get_fd_set(n, exp, fds.ex)))
 595                 goto out;
 596         zero_fd_set(n, fds.res_in);
 597         zero_fd_set(n, fds.res_out);
 598         zero_fd_set(n, fds.res_ex);
 599
 600         ret = do_select(n, &fds, end_time);
 601
 602         if (ret < 0)
 603                 goto out;
 604         if (!ret) {
 605                 ret = -ERESTARTNOHAND;
 606                 if (signal_pending(current))
 607                         goto out;
 608                 ret = 0;
 609         }
 610
 611         if (set_fd_set(n, inp, fds.res_in) ||
 612             set_fd_set(n, outp, fds.res_out) ||
 613             set_fd_set(n, exp, fds.res_ex))
 614                 ret = -EFAULT;
 615
 616 out:
 617         if (bits != stack_fds)
 618                 kfree(bits);
 619 out_nofds:
 620         return ret;
 621 }
 622
 623 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
 624                 fd_set __user *, exp, struct timeval __user *, tvp)
 625 {
 626         struct timespec end_time, *to = NULL;
 627         struct timeval tv;
 628         int ret;
 629
 630         if (tvp) {
 631                 if (copy_from_user(&tv, tvp, sizeof(tv)))
 632                         return -EFAULT;
 633
 634                 to = &end_time;
 635                 if (poll_select_set_timeout(to,
 636                                 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
 637                                 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
 638                         return -EINVAL;
 639         }
 640
 641         ret = core_sys_select(n, inp, outp, exp, to);
 642         ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
 643
 644         return ret;
 645 }
 646
 647 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
 648                        fd_set __user *exp, struct timespec __user *tsp,
 649                        const sigset_t __user *sigmask, size_t sigsetsize)
 650 {
 651         sigset_t ksigmask, sigsaved;
 652         struct timespec ts, end_time, *to = NULL;
 653         int ret;
 654
 655         if (tsp) {
 656                 if (copy_from_user(&ts, tsp, sizeof(ts)))
 657                         return -EFAULT;
 658
 659                 to = &end_time;
 660                 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
 661                         return -EINVAL;
 662         }
 663
 664         if (sigmask) {
 665                 /* XXX: Don't preclude handling different sized sigset_t's.  */
 666                 if (sigsetsize != sizeof(sigset_t))
 667                         return -EINVAL;
 668                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 669                         return -EFAULT;
 670
 671                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 672                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 673         }
 674
 675         ret = core_sys_select(n, inp, outp, exp, to);
 676         ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
 677
 678         if (ret == -ERESTARTNOHAND) {
 679                 /*
 680                  * Don't restore the signal mask yet. Let do_signal() deliver
 681                  * the signal on the way back to userspace, before the signal
 682                  * mask is restored.
 683                  */
 684                 if (sigmask) {
 685                         memcpy(&current->saved_sigmask, &sigsaved,
 686                                         sizeof(sigsaved));
 687                         set_restore_sigmask();
 688                 }
 689         } else if (sigmask)
 690                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 691
 692         return ret;
 693 }
 694
 695 /*
 696  * Most architectures can't handle 7-argument syscalls. So we provide a
 697  * 6-argument version where the sixth argument is a pointer to a structure
 698  * which has a pointer to the sigset_t itself followed by a size_t containing
 699  * the sigset size.
 700  */
 701 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
 702                 fd_set __user *, exp, struct timespec __user *, tsp,
 703                 void __user *, sig)
 704 {
 705         size_t sigsetsize = 0;
 706         sigset_t __user *up = NULL;
 707
 708         if (sig) {
 709                 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
 710                     || __get_user(up, (sigset_t __user * __user *)sig)
 711                     || __get_user(sigsetsize,
 712                                 (size_t __user *)(sig+sizeof(void *))))
 713                         return -EFAULT;
 714         }
 715
 716         return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
 717 }
 718
 719 #ifdef __ARCH_WANT_SYS_OLD_SELECT
 720 struct sel_arg_struct {
 721         unsigned long n;
 722         fd_set __user *inp, *outp, *exp;
 723         struct timeval __user *tvp;
 724 };
 725
 726 SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg)
 727 {
 728         struct sel_arg_struct a;
 729
 730         if (copy_from_user(&a, arg, sizeof(a)))
 731                 return -EFAULT;
 732         return sys_select(a.n, a.inp, a.outp, a.exp, a.tvp);
 733 }
 734 #endif
 735
 736 struct poll_list {
 737         struct poll_list *next;
 738         int len;
 739         struct pollfd entries[0];
 740 };
 741
 742 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 743
 744 /*
 745  * Fish for pollable events on the pollfd->fd file descriptor. We're only
 746  * interested in events matching the pollfd->events mask, and the result
 747  * matching that mask is both recorded in pollfd->revents and returned. The
 748  * pwait poll_table will be used by the fd-provided poll handler for waiting,
 749  * if pwait->_qproc is non-NULL.
 750  */
 751 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait,
 752                                      bool *can_busy_poll,
 753                                      unsigned int busy_flag)
 754 {
 755         unsigned int mask;
 756         int fd;
 757
 758         mask = 0;
 759         fd = pollfd->fd;
 760         if (fd >= 0) {
 761                 struct fd f = fdget(fd);
 762                 mask = POLLNVAL;
 763                 if (f.file) {
 764                         mask = DEFAULT_POLLMASK;
 765                         if (f.file->f_op && f.file->f_op->poll) {
 766                                 pwait->_key = pollfd->events|POLLERR|POLLHUP;
 767                                 pwait->_key |= busy_flag;
 768                                 mask = f.file->f_op->poll(f.file, pwait);
 769                                 if (mask & busy_flag)
 770                                         *can_busy_poll = true;
 771                         }
 772                         /* Mask out unneeded events. */
 773                         mask &= pollfd->events | POLLERR | POLLHUP;
 774                         fdput(f);
 775                 }
 776         }
 777         pollfd->revents = mask;
 778
 779         return mask;
 780 }
 781
 782 static int do_poll(unsigned int nfds,  struct poll_list *list,
 783                    struct poll_wqueues *wait, struct timespec *end_time)
 784 {
 785         poll_table* pt = &wait->pt;
 786         ktime_t expire, *to = NULL;
 787         int timed_out = 0, count = 0;
 788         unsigned long slack = 0;
 789         unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
 790         unsigned long busy_end = 0;
 791
 792         /* Optimise the no-wait case */
 793         if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
 794                 pt->_qproc = NULL;
 795                 timed_out = 1;
 796         }
 797
 798         if (end_time && !timed_out)
 799                 slack = select_estimate_accuracy(end_time);
 800
 801         for (;;) {
 802                 struct poll_list *walk;
 803                 bool can_busy_loop = false;
 804
 805                 for (walk = list; walk != NULL; walk = walk->next) {
 806                         struct pollfd * pfd, * pfd_end;
 807
 808                         pfd = walk->entries;
 809                         pfd_end = pfd + walk->len;
 810                         for (; pfd != pfd_end; pfd++) {
 811                                 /*
 812                                  * Fish for events. If we found one, record it
 813                                  * and kill poll_table->_qproc, so we don't
 814                                  * needlessly register any other waiters after
 815                                  * this. They'll get immediately deregistered
 816                                  * when we break out and return.
 817                                  */
 818                                 if (do_pollfd(pfd, pt, &can_busy_loop,
 819                                               busy_flag)) {
 820                                         count++;
 821                                         pt->_qproc = NULL;
 822                                         /* found something, stop busy polling */
 823                                         busy_flag = 0;
 824                                         can_busy_loop = false;
 825                                 }
 826                         }
 827                 }
 828                 /*
 829                  * All waiters have already been registered, so don't provide
 830                  * a poll_table->_qproc to them on the next loop iteration.
 831                  */
 832                 pt->_qproc = NULL;
 833                 if (!count) {
 834                         count = wait->error;
 835                         if (signal_pending(current))
 836                                 count = -EINTR;
 837                 }
 838                 if (count || timed_out)
 839                         break;
 840
 841                 /* only if found POLL_BUSY_LOOP sockets && not out of time */
 842                 if (can_busy_loop && !need_resched()) {
 843                         if (!busy_end) {
 844                                 busy_end = busy_loop_end_time();
 845                                 continue;
 846                         }
 847                         if (!busy_loop_timeout(busy_end))
 848                                 continue;
 849                 }
 850                 busy_flag = 0;
 851
 852                 /*
 853                  * If this is the first loop and we have a timeout
 854                  * given, then we convert to ktime_t and set the to
 855                  * pointer to the expiry value.
 856                  */
 857                 if (end_time && !to) {
 858                         expire = timespec_to_ktime(*end_time);
 859                         to = &expire;
 860                 }
 861
 862                 if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
 863                         timed_out = 1;
 864         }
 865         return count;
 866 }
 867
 868 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
 869                         sizeof(struct pollfd))
 870
 871 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
 872                 struct timespec *end_time)
 873 {
 874         struct poll_wqueues table;
 875         int err = -EFAULT, fdcount, len, size;
 876         /* Allocate small arguments on the stack to save memory and be
 877            faster - use long to make sure the buffer is aligned properly
 878            on 64 bit archs to avoid unaligned access */
 879         long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
 880         struct poll_list *const head = (struct poll_list *)stack_pps;
 881         struct poll_list *walk = head;
 882         unsigned long todo = nfds;
 883
 884         if (nfds > rlimit(RLIMIT_NOFILE))
 885                 return -EINVAL;
 886
 887         len = min_t(unsigned int, nfds, N_STACK_PPS);
 888         for (;;) {
 889                 walk->next = NULL;
 890                 walk->len = len;
 891                 if (!len)
 892                         break;
 893
 894                 if (copy_from_user(walk->entries, ufds + nfds-todo,
 895                                         sizeof(struct pollfd) * walk->len))
 896                         goto out_fds;
 897
 898                 todo -= walk->len;
 899                 if (!todo)
 900                         break;
 901
 902                 len = min(todo, POLLFD_PER_PAGE);
 903                 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
 904                 walk = walk->next = kmalloc(size, GFP_KERNEL);
 905                 if (!walk) {
 906                         err = -ENOMEM;
 907                         goto out_fds;
 908                 }
 909         }
 910
 911         poll_initwait(&table);
 912         fdcount = do_poll(nfds, head, &table, end_time);
 913         poll_freewait(&table);
 914
 915         for (walk = head; walk; walk = walk->next) {
 916                 struct pollfd *fds = walk->entries;
 917                 int j;
 918
 919                 for (j = 0; j < walk->len; j++, ufds++)
 920                         if (__put_user(fds[j].revents, &ufds->revents))
 921                                 goto out_fds;
 922         }
 923
 924         err = fdcount;
 925 out_fds:
 926         walk = head->next;
 927         while (walk) {
 928                 struct poll_list *pos = walk;
 929                 walk = walk->next;
 930                 kfree(pos);
 931         }
 932
 933         return err;
 934 }
 935
 936 static long do_restart_poll(struct restart_block *restart_block)
 937 {
 938         struct pollfd __user *ufds = restart_block->poll.ufds;
 939         int nfds = restart_block->poll.nfds;
 940         struct timespec *to = NULL, end_time;
 941         int ret;
 942
 943         if (restart_block->poll.has_timeout) {
 944                 end_time.tv_sec = restart_block->poll.tv_sec;
 945                 end_time.tv_nsec = restart_block->poll.tv_nsec;
 946                 to = &end_time;
 947         }
 948
 949         ret = do_sys_poll(ufds, nfds, to);
 950
 951         if (ret == -EINTR) {
 952                 restart_block->fn = do_restart_poll;
 953                 ret = -ERESTART_RESTARTBLOCK;
 954         }
 955         return ret;
 956 }
 957
 958 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
 959                 int, timeout_msecs)
 960 {
 961         struct timespec end_time, *to = NULL;
 962         int ret;
 963
 964         if (timeout_msecs >= 0) {
 965                 to = &end_time;
 966                 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
 967                         NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
 968         }
 969
 970         ret = do_sys_poll(ufds, nfds, to);
 971
 972         if (ret == -EINTR) {
 973                 struct restart_block *restart_block;
 974
 975                 restart_block = &current_thread_info()->restart_block;
 976                 restart_block->fn = do_restart_poll;
 977                 restart_block->poll.ufds = ufds;
 978                 restart_block->poll.nfds = nfds;
 979
 980                 if (timeout_msecs >= 0) {
 981                         restart_block->poll.tv_sec = end_time.tv_sec;
 982                         restart_block->poll.tv_nsec = end_time.tv_nsec;
 983                         restart_block->poll.has_timeout = 1;
 984                 } else
 985                         restart_block->poll.has_timeout = 0;
 986
 987                 ret = -ERESTART_RESTARTBLOCK;
 988         }
 989         return ret;
 990 }
 991
 992 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
 993                 struct timespec __user *, tsp, const sigset_t __user *, sigmask,
 994                 size_t, sigsetsize)
 995 {
 996         sigset_t ksigmask, sigsaved;
 997         struct timespec ts, end_time, *to = NULL;
 998         int ret;
 999
1000         if (tsp) {
1001                 if (copy_from_user(&ts, tsp, sizeof(ts)))
1002                         return -EFAULT;
1003
1004                 to = &end_time;
1005                 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
1006                         return -EINVAL;
1007         }
1008
1009         if (sigmask) {
1010                 /* XXX: Don't preclude handling different sized sigset_t's.  */
1011                 if (sigsetsize != sizeof(sigset_t))
1012                         return -EINVAL;
1013                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1014                         return -EFAULT;
1015
1016                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
1017                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1018         }
1019
1020         ret = do_sys_poll(ufds, nfds, to);
1021
1022         /* We can restart this syscall, usually */
1023         if (ret == -EINTR) {
1024                 /*
1025                  * Don't restore the signal mask yet. Let do_signal() deliver
1026                  * the signal on the way back to userspace, before the signal
1027                  * mask is restored.
1028                  */
1029                 if (sigmask) {
1030                         memcpy(&current->saved_sigmask, &sigsaved,
1031                                         sizeof(sigsaved));
1032                         set_restore_sigmask();
1033                 }
1034                 ret = -ERESTARTNOHAND;
1035         } else if (sigmask)
1036                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1037
1038         ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
1039
1040         return ret;
1041 }