1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 
  22 /*
  23  * Copyright (c) 1989, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Copyright 2015, Joyent Inc.
  25  */
  26 
  27 /*      Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T     */
  28 /*      All Rights Reserved */
  29 
  30 #include <sys/types.h>
  31 #include <sys/sysmacros.h>
  32 #include <sys/param.h>
  33 #include <sys/systm.h>
  34 #include <sys/errno.h>
  35 #include <sys/signal.h>
  36 #include <sys/cred.h>
  37 #include <sys/user.h>
  38 #include <sys/conf.h>
  39 #include <sys/vfs.h>
  40 #include <sys/vnode.h>
  41 #include <sys/pathname.h>
  42 #include <sys/file.h>
  43 #include <sys/flock.h>
  44 #include <sys/proc.h>
  45 #include <sys/var.h>
  46 #include <sys/cpuvar.h>
  47 #include <sys/open.h>
  48 #include <sys/cmn_err.h>
  49 #include <sys/priocntl.h>
  50 #include <sys/procset.h>
  51 #include <sys/prsystm.h>
  52 #include <sys/debug.h>
  53 #include <sys/kmem.h>
  54 #include <sys/atomic.h>
  55 #include <sys/fcntl.h>
  56 #include <sys/poll.h>
  57 #include <sys/rctl.h>
  58 #include <sys/port_impl.h>
  59 #include <sys/dtrace.h>
  60 
  61 #include <c2/audit.h>
  62 #include <sys/nbmlock.h>
  63 
  64 #ifdef DEBUG
  65 
  66 static uint32_t afd_maxfd;      /* # of entries in maximum allocated array */
  67 static uint32_t afd_alloc;      /* count of kmem_alloc()s */
  68 static uint32_t afd_free;       /* count of kmem_free()s */
  69 static uint32_t afd_wait;       /* count of waits on non-zero ref count */
  70 #define MAXFD(x)        (afd_maxfd = ((afd_maxfd >= (x))? afd_maxfd : (x)))
  71 #define COUNT(x)        atomic_inc_32(&x)
  72 
  73 #else   /* DEBUG */
  74 
  75 #define MAXFD(x)
  76 #define COUNT(x)
  77 
  78 #endif  /* DEBUG */
  79 
  80 kmem_cache_t *file_cache;
  81 
  82 static void port_close_fd(portfd_t *);
  83 
  84 /*
  85  * File descriptor allocation.
  86  *
  87  * fd_find(fip, minfd) finds the first available descriptor >= minfd.
  88  * The most common case is open(2), in which minfd = 0, but we must also
  89  * support fcntl(fd, F_DUPFD, minfd).
  90  *
  91  * The algorithm is as follows: we keep all file descriptors in an infix
  92  * binary tree in which each node records the number of descriptors
  93  * allocated in its right subtree, including itself.  Starting at minfd,
  94  * we ascend the tree until we find a non-fully allocated right subtree.
  95  * We then descend that subtree in a binary search for the smallest fd.
  96  * Finally, we ascend the tree again to increment the allocation count
  97  * of every subtree containing the newly-allocated fd.  Freeing an fd
  98  * requires only the last step: we ascend the tree to decrement allocation
  99  * counts.  Each of these three steps (ascent to find non-full subtree,
 100  * descent to find lowest fd, ascent to update allocation counts) is
 101  * O(log n), thus the algorithm as a whole is O(log n).
 102  *
 103  * We don't implement the fd tree using the customary left/right/parent
 104  * pointers, but instead take advantage of the glorious mathematics of
 105  * full infix binary trees.  For reference, here's an illustration of the
 106  * logical structure of such a tree, rooted at 4 (binary 100), covering
 107  * the range 1-7 (binary 001-111).  Our canonical trees do not include
 108  * fd 0; we'll deal with that later.
 109  *
 110  *            100
 111  *           /   \
 112  *          /     \
 113  *        010     110
 114  *        / \     / \
 115  *      001 011 101 111
 116  *
 117  * We make the following observations, all of which are easily proven by
 118  * induction on the depth of the tree:
 119  *
 120  * (T1) The least-significant bit (LSB) of any node is equal to its level
 121  *      in the tree.  In our example, nodes 001, 011, 101 and 111 are at
 122  *      level 0; nodes 010 and 110 are at level 1; and node 100 is at level 2.
 123  *
 124  * (T2) The child size (CSIZE) of node N -- that is, the total number of
 125  *      right-branch descendants in a child of node N, including itself -- is
 126  *      given by clearing all but the least significant bit of N.  This
 127  *      follows immediately from (T1).  Applying this rule to our example, we
 128  *      see that CSIZE(100) = 100, CSIZE(x10) = 10, and CSIZE(xx1) = 1.
 129  *
 130  * (T3) The nearest left ancestor (LPARENT) of node N -- that is, the nearest
 131  *      ancestor containing node N in its right child -- is given by clearing
 132  *      the LSB of N.  For example, LPARENT(111) = 110 and LPARENT(110) = 100.
 133  *      Clearing the LSB of nodes 001, 010 or 100 yields zero, reflecting
 134  *      the fact that these are leftmost nodes.  Note that this algorithm
 135  *      automatically skips generations as necessary.  For example, the parent
 136  *      of node 101 is 110, which is a *right* ancestor (not what we want);
 137  *      but its grandparent is 100, which is a left ancestor. Clearing the LSB
 138  *      of 101 gets us to 100 directly, skipping right past the uninteresting
 139  *      generation (110).
 140  *
 141  *      Note that since LPARENT clears the LSB, whereas CSIZE clears all *but*
 142  *      the LSB, we can express LPARENT() nicely in terms of CSIZE():
 143  *
 144  *      LPARENT(N) = N - CSIZE(N)
 145  *
 146  * (T4) The nearest right ancestor (RPARENT) of node N is given by:
 147  *
 148  *      RPARENT(N) = N + CSIZE(N)
 149  *
 150  * (T5) For every interior node, the children differ from their parent by
 151  *      CSIZE(parent) / 2.  In our example, CSIZE(100) / 2 = 2 = 10 binary,
 152  *      and indeed, the children of 100 are 100 +/- 10 = 010 and 110.
 153  *
 154  * Next, we'll need a few two's-complement math tricks.  Suppose a number,
 155  * N, has the following form:
 156  *
 157  *              N = xxxx10...0
 158  *
 159  * That is, the binary representation of N consists of some string of bits,
 160  * then a 1, then all zeroes.  This amounts to nothing more than saying that
 161  * N has a least-significant bit, which is true for any N != 0.  If we look
 162  * at N and N - 1 together, we see that we can combine them in useful ways:
 163  *
 164  *                N = xxxx10...0
 165  *            N - 1 = xxxx01...1
 166  *      ------------------------
 167  *      N & (N - 1) = xxxx000000
 168  *      N | (N - 1) = xxxx111111
 169  *      N ^ (N - 1) =     111111
 170  *
 171  * In particular, this suggests several easy ways to clear all but the LSB,
 172  * which by (T2) is exactly what we need to determine CSIZE(N) = 10...0.
 173  * We'll opt for this formulation:
 174  *
 175  *      (C1) CSIZE(N) = (N - 1) ^ (N | (N - 1))
 176  *
 177  * Similarly, we have an easy way to determine LPARENT(N), which requires
 178  * that we clear the LSB of N:
 179  *
 180  *      (L1) LPARENT(N) = N & (N - 1)
 181  *
 182  * We note in the above relations that (N | (N - 1)) - N = CSIZE(N) - 1.
 183  * When combined with (T4), this yields an easy way to compute RPARENT(N):
 184  *
 185  *      (R1) RPARENT(N) = (N | (N - 1)) + 1
 186  *
 187  * Finally, to accommodate fd 0 we must adjust all of our results by +/-1 to
 188  * move the fd range from [1, 2^n) to [0, 2^n - 1).  This is straightforward,
 189  * so there's no need to belabor the algebra; the revised relations become:
 190  *
 191  *      (C1a) CSIZE(N) = N ^ (N | (N + 1))
 192  *
 193  *      (L1a) LPARENT(N) = (N & (N + 1)) - 1
 194  *
 195  *      (R1a) RPARENT(N) = N | (N + 1)
 196  *
 197  * This completes the mathematical framework.  We now have all the tools
 198  * we need to implement fd_find() and fd_reserve().
 199  *
 200  * fd_find(fip, minfd) finds the smallest available file descriptor >= minfd.
 201  * It does not actually allocate the descriptor; that's done by fd_reserve().
 202  * fd_find() proceeds in two steps:
 203  *
 204  * (1) Find the leftmost subtree that contains a descriptor >= minfd.
 205  *     We start at the right subtree rooted at minfd.  If this subtree is
 206  *     not full -- if fip->fi_list[minfd].uf_alloc != CSIZE(minfd) -- then
 207  *     step 1 is done.  Otherwise, we know that all fds in this subtree
 208  *     are taken, so we ascend to RPARENT(minfd) using (R1a).  We repeat
 209  *     this process until we either find a candidate subtree or exceed
 210  *     fip->fi_nfiles.  We use (C1a) to compute CSIZE().
 211  *
 212  * (2) Find the smallest fd in the subtree discovered by step 1.
 213  *     Starting at the root of this subtree, we descend to find the
 214  *     smallest available fd.  Since the left children have the smaller
 215  *     fds, we will descend rightward only when the left child is full.
 216  *
 217  *     We begin by comparing the number of allocated fds in the root
 218  *     to the number of allocated fds in its right child; if they differ
 219  *     by exactly CSIZE(child), we know the left subtree is full, so we
 220  *     descend right; that is, the right child becomes the search root.
 221  *     Otherwise we leave the root alone and start following the right
 222  *     child's left children.  As fortune would have it, this is very
 223  *     simple computationally: by (T5), the right child of fd is just
 224  *     fd + size, where size = CSIZE(fd) / 2.  Applying (T5) again,
 225  *     we find that the right child's left child is fd + size - (size / 2) =
 226  *     fd + (size / 2); *its* left child is fd + (size / 2) - (size / 4) =
 227  *     fd + (size / 4), and so on.  In general, fd's right child's
 228  *     leftmost nth descendant is fd + (size >> n).  Thus, to follow
 229  *     the right child's left descendants, we just halve the size in
 230  *     each iteration of the search.
 231  *
 232  *     When we descend leftward, we must keep track of the number of fds
 233  *     that were allocated in all the right subtrees we rejected, so we
 234  *     know how many of the root fd's allocations are in the remaining
 235  *     (as yet unexplored) leftmost part of its right subtree.  When we
 236  *     encounter a fully-allocated left child -- that is, when we find
 237  *     that fip->fi_list[fd].uf_alloc == ralloc + size -- we descend right
 238  *     (as described earlier), resetting ralloc to zero.
 239  *
 240  * fd_reserve(fip, fd, incr) either allocates or frees fd, depending
 241  * on whether incr is 1 or -1.  Starting at fd, fd_reserve() ascends
 242  * the leftmost ancestors (see (T3)) and updates the allocation counts.
 243  * At each step we use (L1a) to compute LPARENT(), the next left ancestor.
 244  *
 245  * flist_minsize() finds the minimal tree that still covers all
 246  * used fds; as long as the allocation count of a root node is zero, we
 247  * don't need that node or its right subtree.
 248  *
 249  * flist_nalloc() counts the number of allocated fds in the tree, by starting
 250  * at the top of the tree and summing the right-subtree allocation counts as
 251  * it descends leftwards.
 252  *
 253  * Note: we assume that flist_grow() will keep fip->fi_nfiles of the form
 254  * 2^n - 1.  This ensures that the fd trees are always full, which saves
 255  * quite a bit of boundary checking.
 256  */
 257 static int
 258 fd_find(uf_info_t *fip, int minfd)
 259 {
 260         int size, ralloc, fd;
 261 
 262         ASSERT(MUTEX_HELD(&fip->fi_lock));
 263         ASSERT((fip->fi_nfiles & (fip->fi_nfiles + 1)) == 0);
 264 
 265         for (fd = minfd; (uint_t)fd < fip->fi_nfiles; fd |= fd + 1) {
 266                 size = fd ^ (fd | (fd + 1));
 267                 if (fip->fi_list[fd].uf_alloc == size)
 268                         continue;
 269                 for (ralloc = 0, size >>= 1; size != 0; size >>= 1) {
 270                         ralloc += fip->fi_list[fd + size].uf_alloc;
 271                         if (fip->fi_list[fd].uf_alloc == ralloc + size) {
 272                                 fd += size;
 273                                 ralloc = 0;
 274                         }
 275                 }
 276                 return (fd);
 277         }
 278         return (-1);
 279 }
 280 
 281 static void
 282 fd_reserve(uf_info_t *fip, int fd, int incr)
 283 {
 284         int pfd;
 285         uf_entry_t *ufp = &fip->fi_list[fd];
 286 
 287         ASSERT((uint_t)fd < fip->fi_nfiles);
 288         ASSERT((ufp->uf_busy == 0 && incr == 1) ||
 289             (ufp->uf_busy == 1 && incr == -1));
 290         ASSERT(MUTEX_HELD(&ufp->uf_lock));
 291         ASSERT(MUTEX_HELD(&fip->fi_lock));
 292 
 293         for (pfd = fd; pfd >= 0; pfd = (pfd & (pfd + 1)) - 1)
 294                 fip->fi_list[pfd].uf_alloc += incr;
 295 
 296         ufp->uf_busy += incr;
 297 }
 298 
 299 static int
 300 flist_minsize(uf_info_t *fip)
 301 {
 302         int fd;
 303 
 304         /*
 305          * We'd like to ASSERT(MUTEX_HELD(&fip->fi_lock)), but we're called
 306          * by flist_fork(), which relies on other mechanisms for mutual
 307          * exclusion.
 308          */
 309         ASSERT((fip->fi_nfiles & (fip->fi_nfiles + 1)) == 0);
 310 
 311         for (fd = fip->fi_nfiles; fd != 0; fd >>= 1)
 312                 if (fip->fi_list[fd >> 1].uf_alloc != 0)
 313                         break;
 314 
 315         return (fd);
 316 }
 317 
 318 static int
 319 flist_nalloc(uf_info_t *fip)
 320 {
 321         int fd;
 322         int nalloc = 0;
 323 
 324         ASSERT(MUTEX_HELD(&fip->fi_lock));
 325         ASSERT((fip->fi_nfiles & (fip->fi_nfiles + 1)) == 0);
 326 
 327         for (fd = fip->fi_nfiles; fd != 0; fd >>= 1)
 328                 nalloc += fip->fi_list[fd >> 1].uf_alloc;
 329 
 330         return (nalloc);
 331 }
 332 
 333 /*
 334  * Increase size of the fi_list array to accommodate at least maxfd.
 335  * We keep the size of the form 2^n - 1 for benefit of fd_find().
 336  */
 337 static void
 338 flist_grow(int maxfd)
 339 {
 340         uf_info_t *fip = P_FINFO(curproc);
 341         int newcnt, oldcnt;
 342         uf_entry_t *src, *dst, *newlist, *oldlist, *newend, *oldend;
 343         uf_rlist_t *urp;
 344 
 345         for (newcnt = 1; newcnt <= maxfd; newcnt = (newcnt << 1) | 1)
 346                 continue;
 347 
 348         newlist = kmem_zalloc(newcnt * sizeof (uf_entry_t), KM_SLEEP);
 349 
 350         mutex_enter(&fip->fi_lock);
 351         oldcnt = fip->fi_nfiles;
 352         if (newcnt <= oldcnt) {
 353                 mutex_exit(&fip->fi_lock);
 354                 kmem_free(newlist, newcnt * sizeof (uf_entry_t));
 355                 return;
 356         }
 357         ASSERT((newcnt & (newcnt + 1)) == 0);
 358         oldlist = fip->fi_list;
 359         oldend = oldlist + oldcnt;
 360         newend = newlist + oldcnt;      /* no need to lock beyond old end */
 361 
 362         /*
 363          * fi_list and fi_nfiles cannot change while any uf_lock is held,
 364          * so we must grab all the old locks *and* the new locks up to oldcnt.
 365          * (Locks beyond the end of oldcnt aren't visible until we store
 366          * the new fi_nfiles, which is the last thing we do before dropping
 367          * all the locks, so there's no need to acquire these locks).
 368          * Holding the new locks is necessary because when fi_list changes
 369          * to point to the new list, fi_nfiles won't have been stored yet.
 370          * If we *didn't* hold the new locks, someone doing a UF_ENTER()
 371          * could see the new fi_list, grab the new uf_lock, and then see
 372          * fi_nfiles change while the lock is held -- in violation of
 373          * UF_ENTER() semantics.
 374          */
 375         for (src = oldlist; src < oldend; src++)
 376                 mutex_enter(&src->uf_lock);
 377 
 378         for (dst = newlist; dst < newend; dst++)
 379                 mutex_enter(&dst->uf_lock);
 380 
 381         for (src = oldlist, dst = newlist; src < oldend; src++, dst++) {
 382                 dst->uf_file = src->uf_file;
 383                 dst->uf_fpollinfo = src->uf_fpollinfo;
 384                 dst->uf_refcnt = src->uf_refcnt;
 385                 dst->uf_alloc = src->uf_alloc;
 386                 dst->uf_flag = src->uf_flag;
 387                 dst->uf_busy = src->uf_busy;
 388                 dst->uf_portfd = src->uf_portfd;
 389         }
 390 
 391         /*
 392          * As soon as we store the new flist, future locking operations
 393          * will use it.  Therefore, we must ensure that all the state
 394          * we've just established reaches global visibility before the
 395          * new flist does.
 396          */
 397         membar_producer();
 398         fip->fi_list = newlist;
 399 
 400         /*
 401          * Routines like getf() make an optimistic check on the validity
 402          * of the supplied file descriptor: if it's less than the current
 403          * value of fi_nfiles -- examined without any locks -- then it's
 404          * safe to attempt a UF_ENTER() on that fd (which is a valid
 405          * assumption because fi_nfiles only increases).  Therefore, it
 406          * is critical that the new value of fi_nfiles not reach global
 407          * visibility until after the new fi_list: if it happened the
 408          * other way around, getf() could see the new fi_nfiles and attempt
 409          * a UF_ENTER() on the old fi_list, which would write beyond its
 410          * end if the fd exceeded the old fi_nfiles.
 411          */
 412         membar_producer();
 413         fip->fi_nfiles = newcnt;
 414 
 415         /*
 416          * The new state is consistent now, so we can drop all the locks.
 417          */
 418         for (dst = newlist; dst < newend; dst++)
 419                 mutex_exit(&dst->uf_lock);
 420 
 421         for (src = oldlist; src < oldend; src++) {
 422                 /*
 423                  * If any threads are blocked on the old cvs, wake them.
 424                  * This will force them to wake up, discover that fi_list
 425                  * has changed, and go back to sleep on the new cvs.
 426                  */
 427                 cv_broadcast(&src->uf_wanted_cv);
 428                 cv_broadcast(&src->uf_closing_cv);
 429                 mutex_exit(&src->uf_lock);
 430         }
 431 
 432         mutex_exit(&fip->fi_lock);
 433 
 434         /*
 435          * Retire the old flist.  We can't actually kmem_free() it now
 436          * because someone may still have a pointer to it.  Instead,
 437          * we link it onto a list of retired flists.  The new flist
 438          * is at least double the size of the previous flist, so the
 439          * total size of all retired flists will be less than the size
 440          * of the current one (to prove, consider the sum of a geometric
 441          * series in powers of 2).  exit() frees the retired flists.
 442          */
 443         urp = kmem_zalloc(sizeof (uf_rlist_t), KM_SLEEP);
 444         urp->ur_list = oldlist;
 445         urp->ur_nfiles = oldcnt;
 446 
 447         mutex_enter(&fip->fi_lock);
 448         urp->ur_next = fip->fi_rlist;
 449         fip->fi_rlist = urp;
 450         mutex_exit(&fip->fi_lock);
 451 }
 452 
 453 /*
 454  * Utility functions for keeping track of the active file descriptors.
 455  */
 456 void
 457 clear_stale_fd()                /* called from post_syscall() */
 458 {
 459         afd_t *afd = &curthread->t_activefd;
 460         int i;
 461 
 462         /* uninitialized is ok here, a_nfd is then zero */
 463         for (i = 0; i < afd->a_nfd; i++) {
 464                 /* assert that this should not be necessary */
 465                 ASSERT(afd->a_fd[i] == -1);
 466                 afd->a_fd[i] = -1;
 467         }
 468         afd->a_stale = 0;
 469 }
 470 
 471 void
 472 free_afd(afd_t *afd)            /* called below and from thread_free() */
 473 {
 474         int i;
 475 
 476         /* free the buffer if it was kmem_alloc()ed */
 477         if (afd->a_nfd > sizeof (afd->a_buf) / sizeof (afd->a_buf[0])) {
 478                 COUNT(afd_free);
 479                 kmem_free(afd->a_fd, afd->a_nfd * sizeof (afd->a_fd[0]));
 480         }
 481 
 482         /* (re)initialize the structure */
 483         afd->a_fd = &afd->a_buf[0];
 484         afd->a_nfd = sizeof (afd->a_buf) / sizeof (afd->a_buf[0]);
 485         afd->a_stale = 0;
 486         for (i = 0; i < afd->a_nfd; i++)
 487                 afd->a_fd[i] = -1;
 488 }
 489 
 490 static void
 491 set_active_fd(int fd)
 492 {
 493         afd_t *afd = &curthread->t_activefd;
 494         int i;
 495         int *old_fd;
 496         int old_nfd;
 497         int *new_fd;
 498         int new_nfd;
 499 
 500         if (afd->a_nfd == 0) {       /* first time initialization */
 501                 ASSERT(fd == -1);
 502                 mutex_enter(&afd->a_fdlock);
 503                 free_afd(afd);
 504                 mutex_exit(&afd->a_fdlock);
 505         }
 506 
 507         /* insert fd into vacant slot, if any */
 508         for (i = 0; i < afd->a_nfd; i++) {
 509                 if (afd->a_fd[i] == -1) {
 510                         afd->a_fd[i] = fd;
 511                         return;
 512                 }
 513         }
 514 
 515         /*
 516          * Reallocate the a_fd[] array to add one more slot.
 517          */
 518         ASSERT(fd == -1);
 519         old_nfd = afd->a_nfd;
 520         old_fd = afd->a_fd;
 521         new_nfd = old_nfd + 1;
 522         new_fd = kmem_alloc(new_nfd * sizeof (afd->a_fd[0]), KM_SLEEP);
 523         MAXFD(new_nfd);
 524         COUNT(afd_alloc);
 525 
 526         mutex_enter(&afd->a_fdlock);
 527         afd->a_fd = new_fd;
 528         afd->a_nfd = new_nfd;
 529         for (i = 0; i < old_nfd; i++)
 530                 afd->a_fd[i] = old_fd[i];
 531         afd->a_fd[i] = fd;
 532         mutex_exit(&afd->a_fdlock);
 533 
 534         if (old_nfd > sizeof (afd->a_buf) / sizeof (afd->a_buf[0])) {
 535                 COUNT(afd_free);
 536                 kmem_free(old_fd, old_nfd * sizeof (afd->a_fd[0]));
 537         }
 538 }
 539 
 540 void
 541 clear_active_fd(int fd)         /* called below and from aio.c */
 542 {
 543         afd_t *afd = &curthread->t_activefd;
 544         int i;
 545 
 546         for (i = 0; i < afd->a_nfd; i++) {
 547                 if (afd->a_fd[i] == fd) {
 548                         afd->a_fd[i] = -1;
 549                         break;
 550                 }
 551         }
 552         ASSERT(i < afd->a_nfd);           /* not found is not ok */
 553 }
 554 
 555 /*
 556  * Does this thread have this fd active?
 557  */
 558 static int
 559 is_active_fd(kthread_t *t, int fd)
 560 {
 561         afd_t *afd = &t->t_activefd;
 562         int i;
 563 
 564         ASSERT(t != curthread);
 565         mutex_enter(&afd->a_fdlock);
 566         /* uninitialized is ok here, a_nfd is then zero */
 567         for (i = 0; i < afd->a_nfd; i++) {
 568                 if (afd->a_fd[i] == fd) {
 569                         mutex_exit(&afd->a_fdlock);
 570                         return (1);
 571                 }
 572         }
 573         mutex_exit(&afd->a_fdlock);
 574         return (0);
 575 }
 576 
 577 /*
 578  * Convert a user supplied file descriptor into a pointer to a file
 579  * structure.  Only task is to check range of the descriptor (soft
 580  * resource limit was enforced at open time and shouldn't be checked
 581  * here).
 582  */
 583 file_t *
 584 getf(int fd)
 585 {
 586         uf_info_t *fip = P_FINFO(curproc);
 587         uf_entry_t *ufp;
 588         file_t *fp;
 589 
 590         if ((uint_t)fd >= fip->fi_nfiles)
 591                 return (NULL);
 592 
 593         /*
 594          * Reserve a slot in the active fd array now so we can call
 595          * set_active_fd(fd) for real below, while still inside UF_ENTER().
 596          */
 597         set_active_fd(-1);
 598 
 599         UF_ENTER(ufp, fip, fd);
 600 
 601         if ((fp = ufp->uf_file) == NULL) {
 602                 UF_EXIT(ufp);
 603 
 604                 if (fd == fip->fi_badfd && fip->fi_action > 0)
 605                         tsignal(curthread, fip->fi_action);
 606 
 607                 return (NULL);
 608         }
 609         ufp->uf_refcnt++;
 610 
 611         set_active_fd(fd);      /* record the active file descriptor */
 612 
 613         UF_EXIT(ufp);
 614 
 615         return (fp);
 616 }
 617 
 618 /*
 619  * Close whatever file currently occupies the file descriptor slot
 620  * and install the new file, usually NULL, in the file descriptor slot.
 621  * The close must complete before we release the file descriptor slot.
 622  * If newfp != NULL we only return an error if we can't allocate the
 623  * slot so the caller knows that it needs to free the filep;
 624  * in the other cases we return the error number from closef().
 625  */
 626 int
 627 closeandsetf(int fd, file_t *newfp)
 628 {
 629         proc_t *p = curproc;
 630         uf_info_t *fip = P_FINFO(p);
 631         uf_entry_t *ufp;
 632         file_t *fp;
 633         fpollinfo_t *fpip;
 634         portfd_t *pfd;
 635         int error;
 636 
 637         if ((uint_t)fd >= fip->fi_nfiles) {
 638                 if (newfp == NULL)
 639                         return (EBADF);
 640                 flist_grow(fd);
 641         }
 642 
 643         if (newfp != NULL) {
 644                 /*
 645                  * If ufp is reserved but has no file pointer, it's in the
 646                  * transition between ufalloc() and setf().  We must wait
 647                  * for this transition to complete before assigning the
 648                  * new non-NULL file pointer.
 649                  */
 650                 mutex_enter(&fip->fi_lock);
 651                 if (fd == fip->fi_badfd) {
 652                         mutex_exit(&fip->fi_lock);
 653                         if (fip->fi_action > 0)
 654                                 tsignal(curthread, fip->fi_action);
 655                         return (EBADF);
 656                 }
 657                 UF_ENTER(ufp, fip, fd);
 658                 while (ufp->uf_busy && ufp->uf_file == NULL) {
 659                         mutex_exit(&fip->fi_lock);
 660                         cv_wait_stop(&ufp->uf_wanted_cv, &ufp->uf_lock, 250);
 661                         UF_EXIT(ufp);
 662                         mutex_enter(&fip->fi_lock);
 663                         UF_ENTER(ufp, fip, fd);
 664                 }
 665                 if ((fp = ufp->uf_file) == NULL) {
 666                         ASSERT(ufp->uf_fpollinfo == NULL);
 667                         ASSERT(ufp->uf_flag == 0);
 668                         fd_reserve(fip, fd, 1);
 669                         ufp->uf_file = newfp;
 670                         UF_EXIT(ufp);
 671                         mutex_exit(&fip->fi_lock);
 672                         return (0);
 673                 }
 674                 mutex_exit(&fip->fi_lock);
 675         } else {
 676                 UF_ENTER(ufp, fip, fd);
 677                 if ((fp = ufp->uf_file) == NULL) {
 678                         UF_EXIT(ufp);
 679                         return (EBADF);
 680                 }
 681         }
 682 
 683         ASSERT(ufp->uf_busy);
 684         ufp->uf_file = NULL;
 685         ufp->uf_flag = 0;
 686 
 687         /*
 688          * If the file descriptor reference count is non-zero, then
 689          * some other lwp in the process is performing system call
 690          * activity on the file.  To avoid blocking here for a long
 691          * time (the other lwp might be in a long term sleep in its
 692          * system call), we scan all other lwps in the process to
 693          * find the ones with this fd as one of their active fds,
 694          * set their a_stale flag, and set them running if they
 695          * are in an interruptible sleep so they will emerge from
 696          * their system calls immediately.  post_syscall() will
 697          * test the a_stale flag and set errno to EBADF.
 698          */
 699         ASSERT(ufp->uf_refcnt == 0 || p->p_lwpcnt > 1);
 700         if (ufp->uf_refcnt > 0) {
 701                 kthread_t *t;
 702 
 703                 /*
 704                  * We call sprlock_proc(p) to ensure that the thread
 705                  * list will not change while we are scanning it.
 706                  * To do this, we must drop ufp->uf_lock and then
 707                  * reacquire it (so we are not holding both p->p_lock
 708                  * and ufp->uf_lock at the same time).  ufp->uf_lock
 709                  * must be held for is_active_fd() to be correct
 710                  * (set_active_fd() is called while holding ufp->uf_lock).
 711                  *
 712                  * This is a convoluted dance, but it is better than
 713                  * the old brute-force method of stopping every thread
 714                  * in the process by calling holdlwps(SHOLDFORK1).
 715                  */
 716 
 717                 UF_EXIT(ufp);
 718                 COUNT(afd_wait);
 719 
 720                 mutex_enter(&p->p_lock);
 721                 sprlock_proc(p);
 722                 mutex_exit(&p->p_lock);
 723 
 724                 UF_ENTER(ufp, fip, fd);
 725                 ASSERT(ufp->uf_file == NULL);
 726 
 727                 if (ufp->uf_refcnt > 0) {
 728                         for (t = curthread->t_forw;
 729                             t != curthread;
 730                             t = t->t_forw) {
 731                                 if (is_active_fd(t, fd)) {
 732                                         thread_lock(t);
 733                                         t->t_activefd.a_stale = 1;
 734                                         t->t_post_sys = 1;
 735                                         if (ISWAKEABLE(t))
 736                                                 setrun_locked(t);
 737                                         thread_unlock(t);
 738                                 }
 739                         }
 740                 }
 741 
 742                 UF_EXIT(ufp);
 743 
 744                 mutex_enter(&p->p_lock);
 745                 sprunlock(p);
 746 
 747                 UF_ENTER(ufp, fip, fd);
 748                 ASSERT(ufp->uf_file == NULL);
 749         }
 750 
 751         /*
 752          * Wait for other lwps to stop using this file descriptor.
 753          */
 754         while (ufp->uf_refcnt > 0) {
 755                 cv_wait_stop(&ufp->uf_closing_cv, &ufp->uf_lock, 250);
 756                 /*
 757                  * cv_wait_stop() drops ufp->uf_lock, so the file list
 758                  * can change.  Drop the lock on our (possibly) stale
 759                  * ufp and let UF_ENTER() find and lock the current ufp.
 760                  */
 761                 UF_EXIT(ufp);
 762                 UF_ENTER(ufp, fip, fd);
 763         }
 764 
 765 #ifdef DEBUG
 766         /*
 767          * catch a watchfd on device's pollhead list but not on fpollinfo list
 768          */
 769         if (ufp->uf_fpollinfo != NULL)
 770                 checkwfdlist(fp->f_vnode, ufp->uf_fpollinfo);
 771 #endif  /* DEBUG */
 772 
 773         /*
 774          * We may need to cleanup some cached poll states in t_pollstate
 775          * before the fd can be reused. It is important that we don't
 776          * access a stale thread structure. We will do the cleanup in two
 777          * phases to avoid deadlock and holding uf_lock for too long.
 778          * In phase 1, hold the uf_lock and call pollblockexit() to set
 779          * state in t_pollstate struct so that a thread does not exit on
 780          * us. In phase 2, we drop the uf_lock and call pollcacheclean().
 781          */
 782         pfd = ufp->uf_portfd;
 783         ufp->uf_portfd = NULL;
 784         fpip = ufp->uf_fpollinfo;
 785         ufp->uf_fpollinfo = NULL;
 786         if (fpip != NULL)
 787                 pollblockexit(fpip);
 788         UF_EXIT(ufp);
 789         if (fpip != NULL)
 790                 pollcacheclean(fpip, fd);
 791         if (pfd)
 792                 port_close_fd(pfd);
 793 
 794         /*
 795          * Keep the file descriptor entry reserved across the closef().
 796          */
 797         error = closef(fp);
 798 
 799         setf(fd, newfp);
 800 
 801         /* Only return closef() error when closing is all we do */
 802         return (newfp == NULL ? error : 0);
 803 }
 804 
 805 /*
 806  * Decrement uf_refcnt; wakeup anyone waiting to close the file.
 807  */
 808 void
 809 releasef(int fd)
 810 {
 811         uf_info_t *fip = P_FINFO(curproc);
 812         uf_entry_t *ufp;
 813 
 814         UF_ENTER(ufp, fip, fd);
 815         ASSERT(ufp->uf_refcnt > 0);
 816         clear_active_fd(fd);    /* clear the active file descriptor */
 817         if (--ufp->uf_refcnt == 0)
 818                 cv_broadcast(&ufp->uf_closing_cv);
 819         UF_EXIT(ufp);
 820 }
 821 
 822 /*
 823  * Identical to releasef() but can be called from another process.
 824  */
 825 void
 826 areleasef(int fd, uf_info_t *fip)
 827 {
 828         uf_entry_t *ufp;
 829 
 830         UF_ENTER(ufp, fip, fd);
 831         ASSERT(ufp->uf_refcnt > 0);
 832         if (--ufp->uf_refcnt == 0)
 833                 cv_broadcast(&ufp->uf_closing_cv);
 834         UF_EXIT(ufp);
 835 }
 836 
 837 /*
 838  * Duplicate all file descriptors across a fork.
 839  */
 840 void
 841 flist_fork(uf_info_t *pfip, uf_info_t *cfip)
 842 {
 843         int fd, nfiles;
 844         uf_entry_t *pufp, *cufp;
 845 
 846         mutex_init(&cfip->fi_lock, NULL, MUTEX_DEFAULT, NULL);
 847         cfip->fi_rlist = NULL;
 848 
 849         /*
 850          * We don't need to hold fi_lock because all other lwp's in the
 851          * parent have been held.
 852          */
 853         cfip->fi_nfiles = nfiles = flist_minsize(pfip);
 854 
 855         cfip->fi_list = kmem_zalloc(nfiles * sizeof (uf_entry_t), KM_SLEEP);
 856 
 857         for (fd = 0, pufp = pfip->fi_list, cufp = cfip->fi_list; fd < nfiles;
 858             fd++, pufp++, cufp++) {
 859                 cufp->uf_file = pufp->uf_file;
 860                 cufp->uf_alloc = pufp->uf_alloc;
 861                 cufp->uf_flag = pufp->uf_flag;
 862                 cufp->uf_busy = pufp->uf_busy;
 863                 if (pufp->uf_file == NULL) {
 864                         ASSERT(pufp->uf_flag == 0);
 865                         if (pufp->uf_busy) {
 866                                 /*
 867                                  * Grab locks to appease ASSERTs in fd_reserve
 868                                  */
 869                                 mutex_enter(&cfip->fi_lock);
 870                                 mutex_enter(&cufp->uf_lock);
 871                                 fd_reserve(cfip, fd, -1);
 872                                 mutex_exit(&cufp->uf_lock);
 873                                 mutex_exit(&cfip->fi_lock);
 874                         }
 875                 }
 876         }
 877 }
 878 
 879 /*
 880  * Close all open file descriptors for the current process.
 881  * This is only called from exit(), which is single-threaded,
 882  * so we don't need any locking.
 883  */
 884 void
 885 closeall(uf_info_t *fip)
 886 {
 887         int fd;
 888         file_t *fp;
 889         uf_entry_t *ufp;
 890 
 891         ufp = fip->fi_list;
 892         for (fd = 0; fd < fip->fi_nfiles; fd++, ufp++) {
 893                 if ((fp = ufp->uf_file) != NULL) {
 894                         ufp->uf_file = NULL;
 895                         if (ufp->uf_portfd != NULL) {
 896                                 portfd_t *pfd;
 897                                 /* remove event port association */
 898                                 pfd = ufp->uf_portfd;
 899                                 ufp->uf_portfd = NULL;
 900                                 port_close_fd(pfd);
 901                         }
 902                         ASSERT(ufp->uf_fpollinfo == NULL);
 903                         (void) closef(fp);
 904                 }
 905         }
 906 
 907         kmem_free(fip->fi_list, fip->fi_nfiles * sizeof (uf_entry_t));
 908         fip->fi_list = NULL;
 909         fip->fi_nfiles = 0;
 910         while (fip->fi_rlist != NULL) {
 911                 uf_rlist_t *urp = fip->fi_rlist;
 912                 fip->fi_rlist = urp->ur_next;
 913                 kmem_free(urp->ur_list, urp->ur_nfiles * sizeof (uf_entry_t));
 914                 kmem_free(urp, sizeof (uf_rlist_t));
 915         }
 916 }
 917 
 918 /*
 919  * Internal form of close.  Decrement reference count on file
 920  * structure.  Decrement reference count on the vnode following
 921  * removal of the referencing file structure.
 922  */
 923 int
 924 closef(file_t *fp)
 925 {
 926         vnode_t *vp;
 927         int error;
 928         int count;
 929         int flag;
 930         offset_t offset;
 931 
 932         /*
 933          * audit close of file (may be exit)
 934          */
 935         if (AU_AUDITING())
 936                 audit_closef(fp);
 937         ASSERT(MUTEX_NOT_HELD(&P_FINFO(curproc)->fi_lock));
 938 
 939         mutex_enter(&fp->f_tlock);
 940 
 941         ASSERT(fp->f_count > 0);
 942 
 943         count = fp->f_count--;
 944         flag = fp->f_flag;
 945         offset = fp->f_offset;
 946 
 947         vp = fp->f_vnode;
 948 
 949         error = VOP_CLOSE(vp, flag, count, offset, fp->f_cred, NULL);
 950 
 951         if (count > 1) {
 952                 mutex_exit(&fp->f_tlock);
 953                 return (error);
 954         }
 955         ASSERT(fp->f_count == 0);
 956         /* Last reference, remove any OFD style lock for the file_t */
 957         ofdcleanlock(fp);
 958         mutex_exit(&fp->f_tlock);
 959 
 960         /*
 961          * If DTrace has getf() subroutines active, it will set dtrace_closef
 962          * to point to code that implements a barrier with respect to probe
 963          * context.  This must be called before the file_t is freed (and the
 964          * vnode that it refers to is released) -- but it must be after the
 965          * file_t has been removed from the uf_entry_t.  That is, there must
 966          * be no way for a racing getf() in probe context to yield the fp that
 967          * we're operating upon.
 968          */
 969         if (dtrace_closef != NULL)
 970                 (*dtrace_closef)();
 971 
 972         VN_RELE(vp);
 973         /*
 974          * deallocate resources to audit_data
 975          */
 976         if (audit_active)
 977                 audit_unfalloc(fp);
 978         crfree(fp->f_cred);
 979         kmem_cache_free(file_cache, fp);
 980         return (error);
 981 }
 982 
 983 /*
 984  * This is a combination of ufalloc() and setf().
 985  */
 986 int
 987 ufalloc_file(int start, file_t *fp)
 988 {
 989         proc_t *p = curproc;
 990         uf_info_t *fip = P_FINFO(p);
 991         int filelimit;
 992         uf_entry_t *ufp;
 993         int nfiles;
 994         int fd;
 995 
 996         /*
 997          * Assertion is to convince the correctness of the following
 998          * assignment for filelimit after casting to int.
 999          */
1000         ASSERT(p->p_fno_ctl <= INT_MAX);
1001         filelimit = (int)p->p_fno_ctl;
1002 
1003         for (;;) {
1004                 mutex_enter(&fip->fi_lock);
1005                 fd = fd_find(fip, start);
1006                 if (fd >= 0 && fd == fip->fi_badfd) {
1007                         start = fd + 1;
1008                         mutex_exit(&fip->fi_lock);
1009                         continue;
1010                 }
1011                 if ((uint_t)fd < filelimit)
1012                         break;
1013                 if (fd >= filelimit) {
1014                         mutex_exit(&fip->fi_lock);
1015                         mutex_enter(&p->p_lock);
1016                         (void) rctl_action(rctlproc_legacy[RLIMIT_NOFILE],
1017                             p->p_rctls, p, RCA_SAFE);
1018                         mutex_exit(&p->p_lock);
1019                         return (-1);
1020                 }
1021                 /* fd_find() returned -1 */
1022                 nfiles = fip->fi_nfiles;
1023                 mutex_exit(&fip->fi_lock);
1024                 flist_grow(MAX(start, nfiles));
1025         }
1026 
1027         UF_ENTER(ufp, fip, fd);
1028         fd_reserve(fip, fd, 1);
1029         ASSERT(ufp->uf_file == NULL);
1030         ufp->uf_file = fp;
1031         UF_EXIT(ufp);
1032         mutex_exit(&fip->fi_lock);
1033         return (fd);
1034 }
1035 
1036 /*
1037  * Allocate a user file descriptor greater than or equal to "start".
1038  */
1039 int
1040 ufalloc(int start)
1041 {
1042         return (ufalloc_file(start, NULL));
1043 }
1044 
1045 /*
1046  * Check that a future allocation of count fds on proc p has a good
1047  * chance of succeeding.  If not, do rctl processing as if we'd failed
1048  * the allocation.
1049  *
1050  * Our caller must guarantee that p cannot disappear underneath us.
1051  */
1052 int
1053 ufcanalloc(proc_t *p, uint_t count)
1054 {
1055         uf_info_t *fip = P_FINFO(p);
1056         int filelimit;
1057         int current;
1058 
1059         if (count == 0)
1060                 return (1);
1061 
1062         ASSERT(p->p_fno_ctl <= INT_MAX);
1063         filelimit = (int)p->p_fno_ctl;
1064 
1065         mutex_enter(&fip->fi_lock);
1066         current = flist_nalloc(fip);            /* # of in-use descriptors */
1067         mutex_exit(&fip->fi_lock);
1068 
1069         /*
1070          * If count is a positive integer, the worst that can happen is
1071          * an overflow to a negative value, which is caught by the >= 0 check.
1072          */
1073         current += count;
1074         if (count <= INT_MAX && current >= 0 && current <= filelimit)
1075                 return (1);
1076 
1077         mutex_enter(&p->p_lock);
1078         (void) rctl_action(rctlproc_legacy[RLIMIT_NOFILE],
1079             p->p_rctls, p, RCA_SAFE);
1080         mutex_exit(&p->p_lock);
1081         return (0);
1082 }
1083 
1084 /*
1085  * Allocate a user file descriptor and a file structure.
1086  * Initialize the descriptor to point at the file structure.
1087  * If fdp is NULL, the user file descriptor will not be allocated.
1088  */
1089 int
1090 falloc(vnode_t *vp, int flag, file_t **fpp, int *fdp)
1091 {
1092         file_t *fp;
1093         int fd;
1094 
1095         if (fdp) {
1096                 if ((fd = ufalloc(0)) == -1)
1097                         return (EMFILE);
1098         }
1099         fp = kmem_cache_alloc(file_cache, KM_SLEEP);
1100         /*
1101          * Note: falloc returns the fp locked
1102          */
1103         mutex_enter(&fp->f_tlock);
1104         fp->f_count = 1;
1105         fp->f_flag = (ushort_t)flag;
1106         fp->f_flag2 = (flag & (FSEARCH|FEXEC)) >> 16;
1107         fp->f_vnode = vp;
1108         fp->f_offset = 0;
1109         fp->f_audit_data = 0;
1110         crhold(fp->f_cred = CRED());
1111         /*
1112          * allocate resources to audit_data
1113          */
1114         if (audit_active)
1115                 audit_falloc(fp);
1116         *fpp = fp;
1117         if (fdp)
1118                 *fdp = fd;
1119         return (0);
1120 }
1121 
1122 /*ARGSUSED*/
1123 static int
1124 file_cache_constructor(void *buf, void *cdrarg, int kmflags)
1125 {
1126         file_t *fp = buf;
1127 
1128         mutex_init(&fp->f_tlock, NULL, MUTEX_DEFAULT, NULL);
1129         return (0);
1130 }
1131 
1132 /*ARGSUSED*/
1133 static void
1134 file_cache_destructor(void *buf, void *cdrarg)
1135 {
1136         file_t *fp = buf;
1137 
1138         mutex_destroy(&fp->f_tlock);
1139 }
1140 
1141 void
1142 finit()
1143 {
1144         file_cache = kmem_cache_create("file_cache", sizeof (file_t), 0,
1145             file_cache_constructor, file_cache_destructor, NULL, NULL, NULL, 0);
1146 }
1147 
1148 void
1149 unfalloc(file_t *fp)
1150 {
1151         ASSERT(MUTEX_HELD(&fp->f_tlock));
1152         if (--fp->f_count <= 0) {
1153                 /*
1154                  * deallocate resources to audit_data
1155                  */
1156                 if (audit_active)
1157                         audit_unfalloc(fp);
1158                 crfree(fp->f_cred);
1159                 mutex_exit(&fp->f_tlock);
1160                 kmem_cache_free(file_cache, fp);
1161         } else
1162                 mutex_exit(&fp->f_tlock);
1163 }
1164 
1165 /*
1166  * Given a file descriptor, set the user's
1167  * file pointer to the given parameter.
1168  */
1169 void
1170 setf(int fd, file_t *fp)
1171 {
1172         uf_info_t *fip = P_FINFO(curproc);
1173         uf_entry_t *ufp;
1174 
1175         if (AU_AUDITING())
1176                 audit_setf(fp, fd);
1177 
1178         if (fp == NULL) {
1179                 mutex_enter(&fip->fi_lock);
1180                 UF_ENTER(ufp, fip, fd);
1181                 fd_reserve(fip, fd, -1);
1182                 mutex_exit(&fip->fi_lock);
1183         } else {
1184                 UF_ENTER(ufp, fip, fd);
1185                 ASSERT(ufp->uf_busy);
1186         }
1187         ASSERT(ufp->uf_fpollinfo == NULL);
1188         ASSERT(ufp->uf_flag == 0);
1189         ufp->uf_file = fp;
1190         cv_broadcast(&ufp->uf_wanted_cv);
1191         UF_EXIT(ufp);
1192 }
1193 
1194 /*
1195  * Given a file descriptor, return the file table flags, plus,
1196  * if this is a socket in asynchronous mode, the FASYNC flag.
1197  * getf() may or may not have been called before calling f_getfl().
1198  */
1199 int
1200 f_getfl(int fd, int *flagp)
1201 {
1202         uf_info_t *fip = P_FINFO(curproc);
1203         uf_entry_t *ufp;
1204         file_t *fp;
1205         int error;
1206 
1207         if ((uint_t)fd >= fip->fi_nfiles)
1208                 error = EBADF;
1209         else {
1210                 UF_ENTER(ufp, fip, fd);
1211                 if ((fp = ufp->uf_file) == NULL)
1212                         error = EBADF;
1213                 else {
1214                         vnode_t *vp = fp->f_vnode;
1215                         int flag = fp->f_flag |
1216                             ((fp->f_flag2 & ~FEPOLLED) << 16);
1217 
1218                         /*
1219                          * BSD fcntl() FASYNC compatibility.
1220                          */
1221                         if (vp->v_type == VSOCK)
1222                                 flag |= sock_getfasync(vp);
1223                         *flagp = flag;
1224                         error = 0;
1225                 }
1226                 UF_EXIT(ufp);
1227         }
1228 
1229         return (error);
1230 }
1231 
1232 /*
1233  * Given a file descriptor, return the user's file flags.
1234  * Force the FD_CLOEXEC flag for writable self-open /proc files.
1235  * getf() may or may not have been called before calling f_getfd_error().
1236  */
1237 int
1238 f_getfd_error(int fd, int *flagp)
1239 {
1240         uf_info_t *fip = P_FINFO(curproc);
1241         uf_entry_t *ufp;
1242         file_t *fp;
1243         int flag;
1244         int error;
1245 
1246         if ((uint_t)fd >= fip->fi_nfiles)
1247                 error = EBADF;
1248         else {
1249                 UF_ENTER(ufp, fip, fd);
1250                 if ((fp = ufp->uf_file) == NULL)
1251                         error = EBADF;
1252                 else {
1253                         flag = ufp->uf_flag;
1254                         if ((fp->f_flag & FWRITE) && pr_isself(fp->f_vnode))
1255                                 flag |= FD_CLOEXEC;
1256                         *flagp = flag;
1257                         error = 0;
1258                 }
1259                 UF_EXIT(ufp);
1260         }
1261 
1262         return (error);
1263 }
1264 
1265 /*
1266  * getf() must have been called before calling f_getfd().
1267  */
1268 char
1269 f_getfd(int fd)
1270 {
1271         int flag = 0;
1272         (void) f_getfd_error(fd, &flag);
1273         return ((char)flag);
1274 }
1275 
1276 /*
1277  * Given a file descriptor and file flags, set the user's file flags.
1278  * At present, the only valid flag is FD_CLOEXEC.
1279  * getf() may or may not have been called before calling f_setfd_error().
1280  */
1281 int
1282 f_setfd_error(int fd, int flags)
1283 {
1284         uf_info_t *fip = P_FINFO(curproc);
1285         uf_entry_t *ufp;
1286         int error;
1287 
1288         if ((uint_t)fd >= fip->fi_nfiles)
1289                 error = EBADF;
1290         else {
1291                 UF_ENTER(ufp, fip, fd);
1292                 if (ufp->uf_file == NULL)
1293                         error = EBADF;
1294                 else {
1295                         ufp->uf_flag = flags & FD_CLOEXEC;
1296                         error = 0;
1297                 }
1298                 UF_EXIT(ufp);
1299         }
1300         return (error);
1301 }
1302 
1303 void
1304 f_setfd(int fd, char flags)
1305 {
1306         (void) f_setfd_error(fd, flags);
1307 }
1308 
1309 #define BADFD_MIN       3
1310 #define BADFD_MAX       255
1311 
1312 /*
1313  * Attempt to allocate a file descriptor which is bad and which
1314  * is "poison" to the application.  It cannot be closed (except
1315  * on exec), allocated for a different use, etc.
1316  */
1317 int
1318 f_badfd(int start, int *fdp, int action)
1319 {
1320         int fdr;
1321         int badfd;
1322         uf_info_t *fip = P_FINFO(curproc);
1323 
1324 #ifdef _LP64
1325         /* No restrictions on 64 bit _file */
1326         if (get_udatamodel() != DATAMODEL_ILP32)
1327                 return (EINVAL);
1328 #endif
1329 
1330         if (start > BADFD_MAX || start < BADFD_MIN)
1331                 return (EINVAL);
1332 
1333         if (action >= NSIG || action < 0)
1334                 return (EINVAL);
1335 
1336         mutex_enter(&fip->fi_lock);
1337         badfd = fip->fi_badfd;
1338         mutex_exit(&fip->fi_lock);
1339 
1340         if (badfd != -1)
1341                 return (EAGAIN);
1342 
1343         fdr = ufalloc(start);
1344 
1345         if (fdr > BADFD_MAX) {
1346                 setf(fdr, NULL);
1347                 return (EMFILE);
1348         }
1349         if (fdr < 0)
1350                 return (EMFILE);
1351 
1352         mutex_enter(&fip->fi_lock);
1353         if (fip->fi_badfd != -1) {
1354                 /* Lost race */
1355                 mutex_exit(&fip->fi_lock);
1356                 setf(fdr, NULL);
1357                 return (EAGAIN);
1358         }
1359         fip->fi_action = action;
1360         fip->fi_badfd = fdr;
1361         mutex_exit(&fip->fi_lock);
1362         setf(fdr, NULL);
1363 
1364         *fdp = fdr;
1365 
1366         return (0);
1367 }
1368 
1369 /*
1370  * Allocate a file descriptor and assign it to the vnode "*vpp",
1371  * performing the usual open protocol upon it and returning the
1372  * file descriptor allocated.  It is the responsibility of the
1373  * caller to dispose of "*vpp" if any error occurs.
1374  */
1375 int
1376 fassign(vnode_t **vpp, int mode, int *fdp)
1377 {
1378         file_t *fp;
1379         int error;
1380         int fd;
1381 
1382         if (error = falloc((vnode_t *)NULL, mode, &fp, &fd))
1383                 return (error);
1384         if (error = VOP_OPEN(vpp, mode, fp->f_cred, NULL)) {
1385                 setf(fd, NULL);
1386                 unfalloc(fp);
1387                 return (error);
1388         }
1389         fp->f_vnode = *vpp;
1390         mutex_exit(&fp->f_tlock);
1391         /*
1392          * Fill in the slot falloc reserved.
1393          */
1394         setf(fd, fp);
1395         *fdp = fd;
1396         return (0);
1397 }
1398 
1399 /*
1400  * When a process forks it must increment the f_count of all file pointers
1401  * since there is a new process pointing at them.  fcnt_add(fip, 1) does this.
1402  * Since we are called when there is only 1 active lwp we don't need to
1403  * hold fi_lock or any uf_lock.  If the fork fails, fork_fail() calls
1404  * fcnt_add(fip, -1) to restore the counts.
1405  */
1406 void
1407 fcnt_add(uf_info_t *fip, int incr)
1408 {
1409         int i;
1410         uf_entry_t *ufp;
1411         file_t *fp;
1412 
1413         ufp = fip->fi_list;
1414         for (i = 0; i < fip->fi_nfiles; i++, ufp++) {
1415                 if ((fp = ufp->uf_file) != NULL) {
1416                         mutex_enter(&fp->f_tlock);
1417                         ASSERT((incr == 1 && fp->f_count >= 1) ||
1418                             (incr == -1 && fp->f_count >= 2));
1419                         fp->f_count += incr;
1420                         mutex_exit(&fp->f_tlock);
1421                 }
1422         }
1423 }
1424 
1425 /*
1426  * This is called from exec to close all fd's that have the FD_CLOEXEC flag
1427  * set and also to close all self-open for write /proc file descriptors.
1428  */
1429 void
1430 close_exec(uf_info_t *fip)
1431 {
1432         int fd;
1433         file_t *fp;
1434         fpollinfo_t *fpip;
1435         uf_entry_t *ufp;
1436         portfd_t *pfd;
1437 
1438         ufp = fip->fi_list;
1439         for (fd = 0; fd < fip->fi_nfiles; fd++, ufp++) {
1440                 if ((fp = ufp->uf_file) != NULL &&
1441                     ((ufp->uf_flag & FD_CLOEXEC) ||
1442                     ((fp->f_flag & FWRITE) && pr_isself(fp->f_vnode)))) {
1443                         fpip = ufp->uf_fpollinfo;
1444                         mutex_enter(&fip->fi_lock);
1445                         mutex_enter(&ufp->uf_lock);
1446                         fd_reserve(fip, fd, -1);
1447                         mutex_exit(&fip->fi_lock);
1448                         ufp->uf_file = NULL;
1449                         ufp->uf_fpollinfo = NULL;
1450                         ufp->uf_flag = 0;
1451                         /*
1452                          * We may need to cleanup some cached poll states
1453                          * in t_pollstate before the fd can be reused. It
1454                          * is important that we don't access a stale thread
1455                          * structure. We will do the cleanup in two
1456                          * phases to avoid deadlock and holding uf_lock for
1457                          * too long. In phase 1, hold the uf_lock and call
1458                          * pollblockexit() to set state in t_pollstate struct
1459                          * so that a thread does not exit on us. In phase 2,
1460                          * we drop the uf_lock and call pollcacheclean().
1461                          */
1462                         pfd = ufp->uf_portfd;
1463                         ufp->uf_portfd = NULL;
1464                         if (fpip != NULL)
1465                                 pollblockexit(fpip);
1466                         mutex_exit(&ufp->uf_lock);
1467                         if (fpip != NULL)
1468                                 pollcacheclean(fpip, fd);
1469                         if (pfd)
1470                                 port_close_fd(pfd);
1471                         (void) closef(fp);
1472                 }
1473         }
1474 
1475         /* Reset bad fd */
1476         fip->fi_badfd = -1;
1477         fip->fi_action = -1;
1478 }
1479 
1480 /*
1481  * Utility function called by most of the *at() system call interfaces.
1482  *
1483  * Generate a starting vnode pointer for an (fd, path) pair where 'fd'
1484  * is an open file descriptor for a directory to be used as the starting
1485  * point for the lookup of the relative pathname 'path' (or, if path is
1486  * NULL, generate a vnode pointer for the direct target of the operation).
1487  *
1488  * If we successfully return a non-NULL startvp, it has been the target
1489  * of VN_HOLD() and the caller must call VN_RELE() on it.
1490  */
1491 int
1492 fgetstartvp(int fd, char *path, vnode_t **startvpp)
1493 {
1494         vnode_t         *startvp;
1495         file_t          *startfp;
1496         char            startchar;
1497 
1498         if (fd == AT_FDCWD && path == NULL)
1499                 return (EFAULT);
1500 
1501         if (fd == AT_FDCWD) {
1502                 /*
1503                  * Start from the current working directory.
1504                  */
1505                 startvp = NULL;
1506         } else {
1507                 if (path == NULL)
1508                         startchar = '\0';
1509                 else if (copyin(path, &startchar, sizeof (char)))
1510                         return (EFAULT);
1511 
1512                 if (startchar == '/') {
1513                         /*
1514                          * 'path' is an absolute pathname.
1515                          */
1516                         startvp = NULL;
1517                 } else {
1518                         /*
1519                          * 'path' is a relative pathname or we will
1520                          * be applying the operation to 'fd' itself.
1521                          */
1522                         if ((startfp = getf(fd)) == NULL)
1523                                 return (EBADF);
1524                         startvp = startfp->f_vnode;
1525                         VN_HOLD(startvp);
1526                         releasef(fd);
1527                 }
1528         }
1529         *startvpp = startvp;
1530         return (0);
1531 }
1532 
1533 /*
1534  * Called from fchownat() and fchmodat() to set ownership and mode.
1535  * The contents of *vap must be set before calling here.
1536  */
1537 int
1538 fsetattrat(int fd, char *path, int flags, struct vattr *vap)
1539 {
1540         vnode_t         *startvp;
1541         vnode_t         *vp;
1542         int             error;
1543 
1544         /*
1545          * Since we are never called to set the size of a file, we don't
1546          * need to check for non-blocking locks (via nbl_need_check(vp)).
1547          */
1548         ASSERT(!(vap->va_mask & AT_SIZE));
1549 
1550         if ((error = fgetstartvp(fd, path, &startvp)) != 0)
1551                 return (error);
1552         if (AU_AUDITING() && startvp != NULL)
1553                 audit_setfsat_path(1);
1554 
1555         /*
1556          * Do lookup for fchownat/fchmodat when path not NULL
1557          */
1558         if (path != NULL) {
1559                 if (error = lookupnameat(path, UIO_USERSPACE,
1560                     (flags == AT_SYMLINK_NOFOLLOW) ?
1561                     NO_FOLLOW : FOLLOW,
1562                     NULLVPP, &vp, startvp)) {
1563                         if (startvp != NULL)
1564                                 VN_RELE(startvp);
1565                         return (error);
1566                 }
1567         } else {
1568                 vp = startvp;
1569                 ASSERT(vp);
1570                 VN_HOLD(vp);
1571         }
1572 
1573         if (vn_is_readonly(vp)) {
1574                 error = EROFS;
1575         } else {
1576                 error = VOP_SETATTR(vp, vap, 0, CRED(), NULL);
1577         }
1578 
1579         if (startvp != NULL)
1580                 VN_RELE(startvp);
1581         VN_RELE(vp);
1582 
1583         return (error);
1584 }
1585 
1586 /*
1587  * Return true if the given vnode is referenced by any
1588  * entry in the current process's file descriptor table.
1589  */
1590 int
1591 fisopen(vnode_t *vp)
1592 {
1593         int fd;
1594         file_t *fp;
1595         vnode_t *ovp;
1596         uf_info_t *fip = P_FINFO(curproc);
1597         uf_entry_t *ufp;
1598 
1599         mutex_enter(&fip->fi_lock);
1600         for (fd = 0; fd < fip->fi_nfiles; fd++) {
1601                 UF_ENTER(ufp, fip, fd);
1602                 if ((fp = ufp->uf_file) != NULL &&
1603                     (ovp = fp->f_vnode) != NULL && VN_CMP(vp, ovp)) {
1604                         UF_EXIT(ufp);
1605                         mutex_exit(&fip->fi_lock);
1606                         return (1);
1607                 }
1608                 UF_EXIT(ufp);
1609         }
1610         mutex_exit(&fip->fi_lock);
1611         return (0);
1612 }
1613 
1614 /*
1615  * Return zero if at least one file currently open (by curproc) shouldn't be
1616  * allowed to change zones.
1617  */
1618 int
1619 files_can_change_zones(void)
1620 {
1621         int fd;
1622         file_t *fp;
1623         uf_info_t *fip = P_FINFO(curproc);
1624         uf_entry_t *ufp;
1625 
1626         mutex_enter(&fip->fi_lock);
1627         for (fd = 0; fd < fip->fi_nfiles; fd++) {
1628                 UF_ENTER(ufp, fip, fd);
1629                 if ((fp = ufp->uf_file) != NULL &&
1630                     !vn_can_change_zones(fp->f_vnode)) {
1631                         UF_EXIT(ufp);
1632                         mutex_exit(&fip->fi_lock);
1633                         return (0);
1634                 }
1635                 UF_EXIT(ufp);
1636         }
1637         mutex_exit(&fip->fi_lock);
1638         return (1);
1639 }
1640 
1641 #ifdef DEBUG
1642 
1643 /*
1644  * The following functions are only used in ASSERT()s elsewhere.
1645  * They do not modify the state of the system.
1646  */
1647 
1648 /*
1649  * Return true (1) if the current thread is in the fpollinfo
1650  * list for this file descriptor, else false (0).
1651  */
1652 static int
1653 curthread_in_plist(uf_entry_t *ufp)
1654 {
1655         fpollinfo_t *fpip;
1656 
1657         ASSERT(MUTEX_HELD(&ufp->uf_lock));
1658         for (fpip = ufp->uf_fpollinfo; fpip; fpip = fpip->fp_next)
1659                 if (fpip->fp_thread == curthread)
1660                         return (1);
1661         return (0);
1662 }
1663 
1664 /*
1665  * Sanity check to make sure that after lwp_exit(),
1666  * curthread does not appear on any fd's fpollinfo list.
1667  */
1668 void
1669 checkfpollinfo(void)
1670 {
1671         int fd;
1672         uf_info_t *fip = P_FINFO(curproc);
1673         uf_entry_t *ufp;
1674 
1675         mutex_enter(&fip->fi_lock);
1676         for (fd = 0; fd < fip->fi_nfiles; fd++) {
1677                 UF_ENTER(ufp, fip, fd);
1678                 ASSERT(!curthread_in_plist(ufp));
1679                 UF_EXIT(ufp);
1680         }
1681         mutex_exit(&fip->fi_lock);
1682 }
1683 
1684 /*
1685  * Return true (1) if the current thread is in the fpollinfo
1686  * list for this file descriptor, else false (0).
1687  * This is the same as curthread_in_plist(),
1688  * but is called w/o holding uf_lock.
1689  */
1690 int
1691 infpollinfo(int fd)
1692 {
1693         uf_info_t *fip = P_FINFO(curproc);
1694         uf_entry_t *ufp;
1695         int rc;
1696 
1697         UF_ENTER(ufp, fip, fd);
1698         rc = curthread_in_plist(ufp);
1699         UF_EXIT(ufp);
1700         return (rc);
1701 }
1702 
1703 #endif  /* DEBUG */
1704 
1705 /*
1706  * Add the curthread to fpollinfo list, meaning this fd is currently in the
1707  * thread's poll cache. Each lwp polling this file descriptor should call
1708  * this routine once.
1709  */
1710 void
1711 addfpollinfo(int fd)
1712 {
1713         struct uf_entry *ufp;
1714         fpollinfo_t *fpip;
1715         uf_info_t *fip = P_FINFO(curproc);
1716 
1717         fpip = kmem_zalloc(sizeof (fpollinfo_t), KM_SLEEP);
1718         fpip->fp_thread = curthread;
1719         UF_ENTER(ufp, fip, fd);
1720         /*
1721          * Assert we are not already on the list, that is, that
1722          * this lwp did not call addfpollinfo twice for the same fd.
1723          */
1724         ASSERT(!curthread_in_plist(ufp));
1725         /*
1726          * addfpollinfo is always done inside the getf/releasef pair.
1727          */
1728         ASSERT(ufp->uf_refcnt >= 1);
1729         fpip->fp_next = ufp->uf_fpollinfo;
1730         ufp->uf_fpollinfo = fpip;
1731         UF_EXIT(ufp);
1732 }
1733 
1734 /*
1735  * Delete curthread from fpollinfo list if it is there.
1736  */
1737 void
1738 delfpollinfo(int fd)
1739 {
1740         struct uf_entry *ufp;
1741         struct fpollinfo *fpip;
1742         struct fpollinfo **fpipp;
1743         uf_info_t *fip = P_FINFO(curproc);
1744 
1745         UF_ENTER(ufp, fip, fd);
1746         for (fpipp = &ufp->uf_fpollinfo;
1747             (fpip = *fpipp) != NULL;
1748             fpipp = &fpip->fp_next) {
1749                 if (fpip->fp_thread == curthread) {
1750                         *fpipp = fpip->fp_next;
1751                         kmem_free(fpip, sizeof (fpollinfo_t));
1752                         break;
1753                 }
1754         }
1755         /*
1756          * Assert that we are not still on the list, that is, that
1757          * this lwp did not call addfpollinfo twice for the same fd.
1758          */
1759         ASSERT(!curthread_in_plist(ufp));
1760         UF_EXIT(ufp);
1761 }
1762 
1763 /*
1764  * fd is associated with a port. pfd is a pointer to the fd entry in the
1765  * cache of the port.
1766  */
1767 
1768 void
1769 addfd_port(int fd, portfd_t *pfd)
1770 {
1771         struct uf_entry *ufp;
1772         uf_info_t *fip = P_FINFO(curproc);
1773 
1774         UF_ENTER(ufp, fip, fd);
1775         /*
1776          * addfd_port is always done inside the getf/releasef pair.
1777          */
1778         ASSERT(ufp->uf_refcnt >= 1);
1779         if (ufp->uf_portfd == NULL) {
1780                 /* first entry */
1781                 ufp->uf_portfd = pfd;
1782                 pfd->pfd_next = NULL;
1783         } else {
1784                 pfd->pfd_next = ufp->uf_portfd;
1785                 ufp->uf_portfd = pfd;
1786                 pfd->pfd_next->pfd_prev = pfd;
1787         }
1788         UF_EXIT(ufp);
1789 }
1790 
1791 void
1792 delfd_port(int fd, portfd_t *pfd)
1793 {
1794         struct uf_entry *ufp;
1795         uf_info_t *fip = P_FINFO(curproc);
1796 
1797         UF_ENTER(ufp, fip, fd);
1798         /*
1799          * delfd_port is always done inside the getf/releasef pair.
1800          */
1801         ASSERT(ufp->uf_refcnt >= 1);
1802         if (ufp->uf_portfd == pfd) {
1803                 /* remove first entry */
1804                 ufp->uf_portfd = pfd->pfd_next;
1805         } else {
1806                 pfd->pfd_prev->pfd_next = pfd->pfd_next;
1807                 if (pfd->pfd_next != NULL)
1808                         pfd->pfd_next->pfd_prev = pfd->pfd_prev;
1809         }
1810         UF_EXIT(ufp);
1811 }
1812 
1813 static void
1814 port_close_fd(portfd_t *pfd)
1815 {
1816         portfd_t        *pfdn;
1817 
1818         /*
1819          * At this point, no other thread should access
1820          * the portfd_t list for this fd. The uf_file, uf_portfd
1821          * pointers in the uf_entry_t struct for this fd would
1822          * be set to NULL.
1823          */
1824         for (; pfd != NULL; pfd = pfdn) {
1825                 pfdn = pfd->pfd_next;
1826                 port_close_pfd(pfd);
1827         }
1828 }