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 ll_flag = ll_get_flag();
 406         u64 ll_start = ll_start_time(ll_flag);
 407         u64 ll_time = ll_run_time();
 408
 409         rcu_read_lock();
 410         retval = max_select_fd(n, fds);
 411         rcu_read_unlock();
 412
 413         if (retval < 0)
 414                 return retval;
 415         n = retval;
 416
 417         poll_initwait(&table);
 418         wait = &table.pt;
 419         if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
 420                 wait->_qproc = NULL;
 421                 timed_out = 1;
 422         }
 423
 424         if (end_time && !timed_out)
 425                 slack = select_estimate_accuracy(end_time);
 426
 427         retval = 0;
 428         for (;;) {
 429                 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
 430                 bool can_ll = false;
 431
 432                 inp = fds->in; outp = fds->out; exp = fds->ex;
 433                 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
 434
 435                 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
 436                         unsigned long in, out, ex, all_bits, bit = 1, mask, j;
 437                         unsigned long res_in = 0, res_out = 0, res_ex = 0;
 438
 439                         in = *inp++; out = *outp++; ex = *exp++;
 440                         all_bits = in | out | ex;
 441                         if (all_bits == 0) {
 442                                 i += BITS_PER_LONG;
 443                                 continue;
 444                         }
 445
 446                         for (j = 0; j < BITS_PER_LONG; ++j, ++i, bit <<= 1) {
 447                                 struct fd f;
 448                                 if (i >= n)
 449                                         break;
 450                                 if (!(bit & all_bits))
 451                                         continue;
 452                                 f = fdget(i);
 453                                 if (f.file) {
 454                                         const struct file_operations *f_op;
 455                                         f_op = f.file->f_op;
 456                                         mask = DEFAULT_POLLMASK;
 457                                         if (f_op && f_op->poll) {
 458                                                 wait_key_set(wait, in, out,
 459                                                              bit, ll_flag);
 460                                                 mask = (*f_op->poll)(f.file, wait);
 461                                         }
 462                                         fdput(f);
 463                                         if ((mask & POLLIN_SET) && (in & bit)) {
 464                                                 res_in |= bit;
 465                                                 retval++;
 466                                                 wait->_qproc = NULL;
 467                                         }
 468                                         if ((mask & POLLOUT_SET) && (out & bit)) {
 469                                                 res_out |= bit;
 470                                                 retval++;
 471                                                 wait->_qproc = NULL;
 472                                         }
 473                                         if ((mask & POLLEX_SET) && (ex & bit)) {
 474                                                 res_ex |= bit;
 475                                                 retval++;
 476                                                 wait->_qproc = NULL;
 477                                         }
 478                                         if (mask & POLL_LL)
 479                                                 can_ll = true;
 480                                         /* got something, stop busy polling */
 481                                         if (retval)
 482                                                 ll_flag = 0;
 483                                 }
 484                         }
 485                         if (res_in)
 486                                 *rinp = res_in;
 487                         if (res_out)
 488                                 *routp = res_out;
 489                         if (res_ex)
 490                                 *rexp = res_ex;
 491                         cond_resched();
 492                 }
 493                 wait->_qproc = NULL;
 494                 if (retval || timed_out || signal_pending(current))
 495                         break;
 496                 if (table.error) {
 497                         retval = table.error;
 498                         break;
 499                 }
 500
 501                 /* only if on, have sockets with POLL_LL and not out of time */
 502                 if (ll_flag && can_ll && can_poll_ll(ll_start, ll_time))
 503                         continue;
 504
 505                 /*
 506                  * If this is the first loop and we have a timeout
 507                  * given, then we convert to ktime_t and set the to
 508                  * pointer to the expiry value.
 509                  */
 510                 if (end_time && !to) {
 511                         expire = timespec_to_ktime(*end_time);
 512                         to = &expire;
 513                 }
 514
 515                 if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
 516                                            to, slack))
 517                         timed_out = 1;
 518         }
 519
 520         poll_freewait(&table);
 521
 522         return retval;
 523 }
 524
 525 /*
 526  * We can actually return ERESTARTSYS instead of EINTR, but I'd
 527  * like to be certain this leads to no problems. So I return
 528  * EINTR just for safety.
 529  *
 530  * Update: ERESTARTSYS breaks at least the xview clock binary, so
 531  * I'm trying ERESTARTNOHAND which restart only when you want to.
 532  */
 533 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 534                            fd_set __user *exp, struct timespec *end_time)
 535 {
 536         fd_set_bits fds;
 537         void *bits;
 538         int ret, max_fds;
 539         unsigned int size;
 540         struct fdtable *fdt;
 541         /* Allocate small arguments on the stack to save memory and be faster */
 542         long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
 543
 544         ret = -EINVAL;
 545         if (n < 0)
 546                 goto out_nofds;
 547
 548         /* max_fds can increase, so grab it once to avoid race */
 549         rcu_read_lock();
 550         fdt = files_fdtable(current->files);
 551         max_fds = fdt->max_fds;
 552         rcu_read_unlock();
 553         if (n > max_fds)
 554                 n = max_fds;
 555
 556         /*
 557          * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
 558          * since we used fdset we need to allocate memory in units of
 559          * long-words.
 560          */
 561         size = FDS_BYTES(n);
 562         bits = stack_fds;
 563         if (size > sizeof(stack_fds) / 6) {
 564                 /* Not enough space in on-stack array; must use kmalloc */
 565                 ret = -ENOMEM;
 566                 bits = kmalloc(6 * size, GFP_KERNEL);
 567                 if (!bits)
 568                         goto out_nofds;
 569         }
 570         fds.in      = bits;
 571         fds.out     = bits +   size;
 572         fds.ex      = bits + 2*size;
 573         fds.res_in  = bits + 3*size;
 574         fds.res_out = bits + 4*size;
 575         fds.res_ex  = bits + 5*size;
 576
 577         if ((ret = get_fd_set(n, inp, fds.in)) ||
 578             (ret = get_fd_set(n, outp, fds.out)) ||
 579             (ret = get_fd_set(n, exp, fds.ex)))
 580                 goto out;
 581         zero_fd_set(n, fds.res_in);
 582         zero_fd_set(n, fds.res_out);
 583         zero_fd_set(n, fds.res_ex);
 584
 585         ret = do_select(n, &fds, end_time);
 586
 587         if (ret < 0)
 588                 goto out;
 589         if (!ret) {
 590                 ret = -ERESTARTNOHAND;
 591                 if (signal_pending(current))
 592                         goto out;
 593                 ret = 0;
 594         }
 595
 596         if (set_fd_set(n, inp, fds.res_in) ||
 597             set_fd_set(n, outp, fds.res_out) ||
 598             set_fd_set(n, exp, fds.res_ex))
 599                 ret = -EFAULT;
 600
 601 out:
 602         if (bits != stack_fds)
 603                 kfree(bits);
 604 out_nofds:
 605         return ret;
 606 }
 607
 608 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp,
 609                 fd_set __user *, exp, struct timeval __user *, tvp)
 610 {
 611         struct timespec end_time, *to = NULL;
 612         struct timeval tv;
 613         int ret;
 614
 615         if (tvp) {
 616                 if (copy_from_user(&tv, tvp, sizeof(tv)))
 617                         return -EFAULT;
 618
 619                 to = &end_time;
 620                 if (poll_select_set_timeout(to,
 621                                 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
 622                                 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
 623                         return -EINVAL;
 624         }
 625
 626         ret = core_sys_select(n, inp, outp, exp, to);
 627         ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
 628
 629         return ret;
 630 }
 631
 632 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp,
 633                        fd_set __user *exp, struct timespec __user *tsp,
 634                        const sigset_t __user *sigmask, size_t sigsetsize)
 635 {
 636         sigset_t ksigmask, sigsaved;
 637         struct timespec ts, end_time, *to = NULL;
 638         int ret;
 639
 640         if (tsp) {
 641                 if (copy_from_user(&ts, tsp, sizeof(ts)))
 642                         return -EFAULT;
 643
 644                 to = &end_time;
 645                 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
 646                         return -EINVAL;
 647         }
 648
 649         if (sigmask) {
 650                 /* XXX: Don't preclude handling different sized sigset_t's.  */
 651                 if (sigsetsize != sizeof(sigset_t))
 652                         return -EINVAL;
 653                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 654                         return -EFAULT;
 655
 656                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 657                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 658         }
 659
 660         ret = core_sys_select(n, inp, outp, exp, to);
 661         ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
 662
 663         if (ret == -ERESTARTNOHAND) {
 664                 /*
 665                  * Don't restore the signal mask yet. Let do_signal() deliver
 666                  * the signal on the way back to userspace, before the signal
 667                  * mask is restored.
 668                  */
 669                 if (sigmask) {
 670                         memcpy(&current->saved_sigmask, &sigsaved,
 671                                         sizeof(sigsaved));
 672                         set_restore_sigmask();
 673                 }
 674         } else if (sigmask)
 675                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 676
 677         return ret;
 678 }
 679
 680 /*
 681  * Most architectures can't handle 7-argument syscalls. So we provide a
 682  * 6-argument version where the sixth argument is a pointer to a structure
 683  * which has a pointer to the sigset_t itself followed by a size_t containing
 684  * the sigset size.
 685  */
 686 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp,
 687                 fd_set __user *, exp, struct timespec __user *, tsp,
 688                 void __user *, sig)
 689 {
 690         size_t sigsetsize = 0;
 691         sigset_t __user *up = NULL;
 692
 693         if (sig) {
 694                 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
 695                     || __get_user(up, (sigset_t __user * __user *)sig)
 696                     || __get_user(sigsetsize,
 697                                 (size_t __user *)(sig+sizeof(void *))))
 698                         return -EFAULT;
 699         }
 700
 701         return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize);
 702 }
 703
 704 #ifdef __ARCH_WANT_SYS_OLD_SELECT
 705 struct sel_arg_struct {
 706         unsigned long n;
 707         fd_set __user *inp, *outp, *exp;
 708         struct timeval __user *tvp;
 709 };
 710
 711 SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg)
 712 {
 713         struct sel_arg_struct a;
 714
 715         if (copy_from_user(&a, arg, sizeof(a)))
 716                 return -EFAULT;
 717         return sys_select(a.n, a.inp, a.outp, a.exp, a.tvp);
 718 }
 719 #endif
 720
 721 struct poll_list {
 722         struct poll_list *next;
 723         int len;
 724         struct pollfd entries[0];
 725 };
 726
 727 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 728
 729 /*
 730  * Fish for pollable events on the pollfd->fd file descriptor. We're only
 731  * interested in events matching the pollfd->events mask, and the result
 732  * matching that mask is both recorded in pollfd->revents and returned. The
 733  * pwait poll_table will be used by the fd-provided poll handler for waiting,
 734  * if pwait->_qproc is non-NULL.
 735  */
 736 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait,
 737                                      bool *can_ll, unsigned int ll_flag)
 738 {
 739         unsigned int mask;
 740         int fd;
 741
 742         mask = 0;
 743         fd = pollfd->fd;
 744         if (fd >= 0) {
 745                 struct fd f = fdget(fd);
 746                 mask = POLLNVAL;
 747                 if (f.file) {
 748                         mask = DEFAULT_POLLMASK;
 749                         if (f.file->f_op && f.file->f_op->poll) {
 750                                 pwait->_key = pollfd->events|POLLERR|POLLHUP;
 751                                 pwait->_key |= ll_flag;
 752                                 mask = f.file->f_op->poll(f.file, pwait);
 753                                 if (mask & POLL_LL)
 754                                         *can_ll = true;
 755                         }
 756                         /* Mask out unneeded events. */
 757                         mask &= pollfd->events | POLLERR | POLLHUP;
 758                         fdput(f);
 759                 }
 760         }
 761         pollfd->revents = mask;
 762
 763         return mask;
 764 }
 765
 766 static int do_poll(unsigned int nfds,  struct poll_list *list,
 767                    struct poll_wqueues *wait, struct timespec *end_time)
 768 {
 769         poll_table* pt = &wait->pt;
 770         ktime_t expire, *to = NULL;
 771         int timed_out = 0, count = 0;
 772         unsigned long slack = 0;
 773         unsigned int ll_flag = ll_get_flag();
 774         u64 ll_start = ll_start_time(ll_flag);
 775         u64 ll_time = ll_run_time();
 776
 777         /* Optimise the no-wait case */
 778         if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
 779                 pt->_qproc = NULL;
 780                 timed_out = 1;
 781         }
 782
 783         if (end_time && !timed_out)
 784                 slack = select_estimate_accuracy(end_time);
 785
 786         for (;;) {
 787                 struct poll_list *walk;
 788                 bool can_ll = false;
 789
 790                 for (walk = list; walk != NULL; walk = walk->next) {
 791                         struct pollfd * pfd, * pfd_end;
 792
 793                         pfd = walk->entries;
 794                         pfd_end = pfd + walk->len;
 795                         for (; pfd != pfd_end; pfd++) {
 796                                 /*
 797                                  * Fish for events. If we found one, record it
 798                                  * and kill poll_table->_qproc, so we don't
 799                                  * needlessly register any other waiters after
 800                                  * this. They'll get immediately deregistered
 801                                  * when we break out and return.
 802                                  */
 803                                 if (do_pollfd(pfd, pt, &can_ll, ll_flag)) {
 804                                         count++;
 805                                         pt->_qproc = NULL;
 806                                         ll_flag = 0;
 807                                 }
 808                         }
 809                 }
 810                 /*
 811                  * All waiters have already been registered, so don't provide
 812                  * a poll_table->_qproc to them on the next loop iteration.
 813                  */
 814                 pt->_qproc = NULL;
 815                 if (!count) {
 816                         count = wait->error;
 817                         if (signal_pending(current))
 818                                 count = -EINTR;
 819                 }
 820                 if (count || timed_out)
 821                         break;
 822
 823                 /* only if on, have sockets with POLL_LL and not out of time */
 824                 if (ll_flag && can_ll && can_poll_ll(ll_start, ll_time))
 825                         continue;
 826
 827                 /*
 828                  * If this is the first loop and we have a timeout
 829                  * given, then we convert to ktime_t and set the to
 830                  * pointer to the expiry value.
 831                  */
 832                 if (end_time && !to) {
 833                         expire = timespec_to_ktime(*end_time);
 834                         to = &expire;
 835                 }
 836
 837                 if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
 838                         timed_out = 1;
 839         }
 840         return count;
 841 }
 842
 843 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
 844                         sizeof(struct pollfd))
 845
 846 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
 847                 struct timespec *end_time)
 848 {
 849         struct poll_wqueues table;
 850         int err = -EFAULT, fdcount, len, size;
 851         /* Allocate small arguments on the stack to save memory and be
 852            faster - use long to make sure the buffer is aligned properly
 853            on 64 bit archs to avoid unaligned access */
 854         long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
 855         struct poll_list *const head = (struct poll_list *)stack_pps;
 856         struct poll_list *walk = head;
 857         unsigned long todo = nfds;
 858
 859         if (nfds > rlimit(RLIMIT_NOFILE))
 860                 return -EINVAL;
 861
 862         len = min_t(unsigned int, nfds, N_STACK_PPS);
 863         for (;;) {
 864                 walk->next = NULL;
 865                 walk->len = len;
 866                 if (!len)
 867                         break;
 868
 869                 if (copy_from_user(walk->entries, ufds + nfds-todo,
 870                                         sizeof(struct pollfd) * walk->len))
 871                         goto out_fds;
 872
 873                 todo -= walk->len;
 874                 if (!todo)
 875                         break;
 876
 877                 len = min(todo, POLLFD_PER_PAGE);
 878                 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
 879                 walk = walk->next = kmalloc(size, GFP_KERNEL);
 880                 if (!walk) {
 881                         err = -ENOMEM;
 882                         goto out_fds;
 883                 }
 884         }
 885
 886         poll_initwait(&table);
 887         fdcount = do_poll(nfds, head, &table, end_time);
 888         poll_freewait(&table);
 889
 890         for (walk = head; walk; walk = walk->next) {
 891                 struct pollfd *fds = walk->entries;
 892                 int j;
 893
 894                 for (j = 0; j < walk->len; j++, ufds++)
 895                         if (__put_user(fds[j].revents, &ufds->revents))
 896                                 goto out_fds;
 897         }
 898
 899         err = fdcount;
 900 out_fds:
 901         walk = head->next;
 902         while (walk) {
 903                 struct poll_list *pos = walk;
 904                 walk = walk->next;
 905                 kfree(pos);
 906         }
 907
 908         return err;
 909 }
 910
 911 static long do_restart_poll(struct restart_block *restart_block)
 912 {
 913         struct pollfd __user *ufds = restart_block->poll.ufds;
 914         int nfds = restart_block->poll.nfds;
 915         struct timespec *to = NULL, end_time;
 916         int ret;
 917
 918         if (restart_block->poll.has_timeout) {
 919                 end_time.tv_sec = restart_block->poll.tv_sec;
 920                 end_time.tv_nsec = restart_block->poll.tv_nsec;
 921                 to = &end_time;
 922         }
 923
 924         ret = do_sys_poll(ufds, nfds, to);
 925
 926         if (ret == -EINTR) {
 927                 restart_block->fn = do_restart_poll;
 928                 ret = -ERESTART_RESTARTBLOCK;
 929         }
 930         return ret;
 931 }
 932
 933 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
 934                 int, timeout_msecs)
 935 {
 936         struct timespec end_time, *to = NULL;
 937         int ret;
 938
 939         if (timeout_msecs >= 0) {
 940                 to = &end_time;
 941                 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
 942                         NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
 943         }
 944
 945         ret = do_sys_poll(ufds, nfds, to);
 946
 947         if (ret == -EINTR) {
 948                 struct restart_block *restart_block;
 949
 950                 restart_block = &current_thread_info()->restart_block;
 951                 restart_block->fn = do_restart_poll;
 952                 restart_block->poll.ufds = ufds;
 953                 restart_block->poll.nfds = nfds;
 954
 955                 if (timeout_msecs >= 0) {
 956                         restart_block->poll.tv_sec = end_time.tv_sec;
 957                         restart_block->poll.tv_nsec = end_time.tv_nsec;
 958                         restart_block->poll.has_timeout = 1;
 959                 } else
 960                         restart_block->poll.has_timeout = 0;
 961
 962                 ret = -ERESTART_RESTARTBLOCK;
 963         }
 964         return ret;
 965 }
 966
 967 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds,
 968                 struct timespec __user *, tsp, const sigset_t __user *, sigmask,
 969                 size_t, sigsetsize)
 970 {
 971         sigset_t ksigmask, sigsaved;
 972         struct timespec ts, end_time, *to = NULL;
 973         int ret;
 974
 975         if (tsp) {
 976                 if (copy_from_user(&ts, tsp, sizeof(ts)))
 977                         return -EFAULT;
 978
 979                 to = &end_time;
 980                 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
 981                         return -EINVAL;
 982         }
 983
 984         if (sigmask) {
 985                 /* XXX: Don't preclude handling different sized sigset_t's.  */
 986                 if (sigsetsize != sizeof(sigset_t))
 987                         return -EINVAL;
 988                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
 989                         return -EFAULT;
 990
 991                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
 992                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
 993         }
 994
 995         ret = do_sys_poll(ufds, nfds, to);
 996
 997         /* We can restart this syscall, usually */
 998         if (ret == -EINTR) {
 999                 /*
1000                  * Don't restore the signal mask yet. Let do_signal() deliver
1001                  * the signal on the way back to userspace, before the signal
1002                  * mask is restored.
1003                  */
1004                 if (sigmask) {
1005                         memcpy(&current->saved_sigmask, &sigsaved,
1006                                         sizeof(sigsaved));
1007                         set_restore_sigmask();
1008                 }
1009                 ret = -ERESTARTNOHAND;
1010         } else if (sigmask)
1011                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1012
1013         ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
1014
1015         return ret;
1016 }