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 = nfiles == 0 ? NULL :
856 kmem_zalloc(nfiles * sizeof (uf_entry_t), KM_SLEEP);
857
858 for (fd = 0, pufp = pfip->fi_list, cufp = cfip->fi_list; fd < nfiles;
859 fd++, pufp++, cufp++) {
860 cufp->uf_file = pufp->uf_file;
861 cufp->uf_alloc = pufp->uf_alloc;
862 cufp->uf_flag = pufp->uf_flag;
863 cufp->uf_busy = pufp->uf_busy;
864 if (pufp->uf_file == NULL) {
865 ASSERT(pufp->uf_flag == 0);
866 if (pufp->uf_busy) {
867 /*
868 * Grab locks to appease ASSERTs in fd_reserve
869 */
870 mutex_enter(&cfip->fi_lock);
871 mutex_enter(&cufp->uf_lock);
872 fd_reserve(cfip, fd, -1);
873 mutex_exit(&cufp->uf_lock);
874 mutex_exit(&cfip->fi_lock);
875 }
876 }
877 }
878 }
879
880 /*
881 * Close all open file descriptors for the current process.
882 * This is only called from exit(), which is single-threaded,
883 * so we don't need any locking.
884 */
885 void
886 closeall(uf_info_t *fip)
887 {
888 int fd;
889 file_t *fp;
890 uf_entry_t *ufp;
891
892 ufp = fip->fi_list;
893 for (fd = 0; fd < fip->fi_nfiles; fd++, ufp++) {
894 if ((fp = ufp->uf_file) != NULL) {
895 ufp->uf_file = NULL;
896 if (ufp->uf_portfd != NULL) {
897 portfd_t *pfd;
898 /* remove event port association */
899 pfd = ufp->uf_portfd;
900 ufp->uf_portfd = NULL;
901 port_close_fd(pfd);
902 }
903 ASSERT(ufp->uf_fpollinfo == NULL);
904 (void) closef(fp);
905 }
906 }
907
908 kmem_free(fip->fi_list, fip->fi_nfiles * sizeof (uf_entry_t));
909 fip->fi_list = NULL;
910 fip->fi_nfiles = 0;
911 while (fip->fi_rlist != NULL) {
912 uf_rlist_t *urp = fip->fi_rlist;
913 fip->fi_rlist = urp->ur_next;
914 kmem_free(urp->ur_list, urp->ur_nfiles * sizeof (uf_entry_t));
915 kmem_free(urp, sizeof (uf_rlist_t));
916 }
917 }
918
919 /*
920 * Internal form of close. Decrement reference count on file
921 * structure. Decrement reference count on the vnode following
922 * removal of the referencing file structure.
923 */
924 int
925 closef(file_t *fp)
926 {
927 vnode_t *vp;
928 int error;
929 int count;
930 int flag;
931 offset_t offset;
932
933 /*
934 * audit close of file (may be exit)
935 */
936 if (AU_AUDITING())
937 audit_closef(fp);
938 ASSERT(MUTEX_NOT_HELD(&P_FINFO(curproc)->fi_lock));
939
940 mutex_enter(&fp->f_tlock);
941
942 ASSERT(fp->f_count > 0);
943
944 count = fp->f_count--;
945 flag = fp->f_flag;
946 offset = fp->f_offset;
947
948 vp = fp->f_vnode;
949
950 error = VOP_CLOSE(vp, flag, count, offset, fp->f_cred, NULL);
951
952 if (count > 1) {
953 mutex_exit(&fp->f_tlock);
954 return (error);
955 }
956 ASSERT(fp->f_count == 0);
957 /* Last reference, remove any OFD style lock for the file_t */
958 ofdcleanlock(fp);
959 mutex_exit(&fp->f_tlock);
960
961 /*
962 * If DTrace has getf() subroutines active, it will set dtrace_closef
963 * to point to code that implements a barrier with respect to probe
964 * context. This must be called before the file_t is freed (and the
965 * vnode that it refers to is released) -- but it must be after the
966 * file_t has been removed from the uf_entry_t. That is, there must
967 * be no way for a racing getf() in probe context to yield the fp that
968 * we're operating upon.
969 */
970 if (dtrace_closef != NULL)
971 (*dtrace_closef)();
972
973 VN_RELE(vp);
974 /*
975 * deallocate resources to audit_data
976 */
977 if (audit_active)
978 audit_unfalloc(fp);
979 crfree(fp->f_cred);
980 kmem_cache_free(file_cache, fp);
981 return (error);
982 }
983
984 /*
985 * This is a combination of ufalloc() and setf().
986 */
987 int
988 ufalloc_file(int start, file_t *fp)
989 {
990 proc_t *p = curproc;
991 uf_info_t *fip = P_FINFO(p);
992 int filelimit;
993 uf_entry_t *ufp;
994 int nfiles;
995 int fd;
996
997 /*
998 * Assertion is to convince the correctness of the following
999 * assignment for filelimit after casting to int.
1000 */
1001 ASSERT(p->p_fno_ctl <= INT_MAX);
1002 filelimit = (int)p->p_fno_ctl;
1003
1004 for (;;) {
1005 mutex_enter(&fip->fi_lock);
1006 fd = fd_find(fip, start);
1007 if (fd >= 0 && fd == fip->fi_badfd) {
1008 start = fd + 1;
1009 mutex_exit(&fip->fi_lock);
1010 continue;
1011 }
1012 if ((uint_t)fd < filelimit)
1013 break;
1014 if (fd >= filelimit) {
1015 mutex_exit(&fip->fi_lock);
1016 mutex_enter(&p->p_lock);
1017 (void) rctl_action(rctlproc_legacy[RLIMIT_NOFILE],
1018 p->p_rctls, p, RCA_SAFE);
1019 mutex_exit(&p->p_lock);
1020 return (-1);
1021 }
1022 /* fd_find() returned -1 */
1023 nfiles = fip->fi_nfiles;
1024 mutex_exit(&fip->fi_lock);
1025 flist_grow(MAX(start, nfiles));
1026 }
1027
1028 UF_ENTER(ufp, fip, fd);
1029 fd_reserve(fip, fd, 1);
1030 ASSERT(ufp->uf_file == NULL);
1031 ufp->uf_file = fp;
1032 UF_EXIT(ufp);
1033 mutex_exit(&fip->fi_lock);
1034 return (fd);
1035 }
1036
1037 /*
1038 * Allocate a user file descriptor greater than or equal to "start".
1039 */
1040 int
1041 ufalloc(int start)
1042 {
1043 return (ufalloc_file(start, NULL));
1044 }
1045
1046 /*
1047 * Check that a future allocation of count fds on proc p has a good
1048 * chance of succeeding. If not, do rctl processing as if we'd failed
1049 * the allocation.
1050 *
1051 * Our caller must guarantee that p cannot disappear underneath us.
1052 */
1053 int
1054 ufcanalloc(proc_t *p, uint_t count)
1055 {
1056 uf_info_t *fip = P_FINFO(p);
1057 int filelimit;
1058 int current;
1059
1060 if (count == 0)
1061 return (1);
1062
1063 ASSERT(p->p_fno_ctl <= INT_MAX);
1064 filelimit = (int)p->p_fno_ctl;
1065
1066 mutex_enter(&fip->fi_lock);
1067 current = flist_nalloc(fip); /* # of in-use descriptors */
1068 mutex_exit(&fip->fi_lock);
1069
1070 /*
1071 * If count is a positive integer, the worst that can happen is
1072 * an overflow to a negative value, which is caught by the >= 0 check.
1073 */
1074 current += count;
1075 if (count <= INT_MAX && current >= 0 && current <= filelimit)
1076 return (1);
1077
1078 mutex_enter(&p->p_lock);
1079 (void) rctl_action(rctlproc_legacy[RLIMIT_NOFILE],
1080 p->p_rctls, p, RCA_SAFE);
1081 mutex_exit(&p->p_lock);
1082 return (0);
1083 }
1084
1085 /*
1086 * Allocate a user file descriptor and a file structure.
1087 * Initialize the descriptor to point at the file structure.
1088 * If fdp is NULL, the user file descriptor will not be allocated.
1089 */
1090 int
1091 falloc(vnode_t *vp, int flag, file_t **fpp, int *fdp)
1092 {
1093 file_t *fp;
1094 int fd;
1095
1096 if (fdp) {
1097 if ((fd = ufalloc(0)) == -1)
1098 return (EMFILE);
1099 }
1100 fp = kmem_cache_alloc(file_cache, KM_SLEEP);
1101 /*
1102 * Note: falloc returns the fp locked
1103 */
1104 mutex_enter(&fp->f_tlock);
1105 fp->f_count = 1;
1106 fp->f_flag = (ushort_t)flag;
1107 fp->f_flag2 = (flag & (FSEARCH|FEXEC)) >> 16;
1108 fp->f_vnode = vp;
1109 fp->f_offset = 0;
1110 fp->f_audit_data = 0;
1111 crhold(fp->f_cred = CRED());
1112 /*
1113 * allocate resources to audit_data
1114 */
1115 if (audit_active)
1116 audit_falloc(fp);
1117 *fpp = fp;
1118 if (fdp)
1119 *fdp = fd;
1120 return (0);
1121 }
1122
1123 /*ARGSUSED*/
1124 static int
1125 file_cache_constructor(void *buf, void *cdrarg, int kmflags)
1126 {
1127 file_t *fp = buf;
1128
1129 mutex_init(&fp->f_tlock, NULL, MUTEX_DEFAULT, NULL);
1130 return (0);
1131 }
1132
1133 /*ARGSUSED*/
1134 static void
1135 file_cache_destructor(void *buf, void *cdrarg)
1136 {
1137 file_t *fp = buf;
1138
1139 mutex_destroy(&fp->f_tlock);
1140 }
1141
1142 void
1143 finit()
1144 {
1145 file_cache = kmem_cache_create("file_cache", sizeof (file_t), 0,
1146 file_cache_constructor, file_cache_destructor, NULL, NULL, NULL, 0);
1147 }
1148
1149 void
1150 unfalloc(file_t *fp)
1151 {
1152 ASSERT(MUTEX_HELD(&fp->f_tlock));
1153 if (--fp->f_count <= 0) {
1154 /*
1155 * deallocate resources to audit_data
1156 */
1157 if (audit_active)
1158 audit_unfalloc(fp);
1159 crfree(fp->f_cred);
1160 mutex_exit(&fp->f_tlock);
1161 kmem_cache_free(file_cache, fp);
1162 } else
1163 mutex_exit(&fp->f_tlock);
1164 }
1165
1166 /*
1167 * Given a file descriptor, set the user's
1168 * file pointer to the given parameter.
1169 */
1170 void
1171 setf(int fd, file_t *fp)
1172 {
1173 uf_info_t *fip = P_FINFO(curproc);
1174 uf_entry_t *ufp;
1175
1176 if (AU_AUDITING())
1177 audit_setf(fp, fd);
1178
1179 if (fp == NULL) {
1180 mutex_enter(&fip->fi_lock);
1181 UF_ENTER(ufp, fip, fd);
1182 fd_reserve(fip, fd, -1);
1183 mutex_exit(&fip->fi_lock);
1184 } else {
1185 UF_ENTER(ufp, fip, fd);
1186 ASSERT(ufp->uf_busy);
1187 }
1188 ASSERT(ufp->uf_fpollinfo == NULL);
1189 ASSERT(ufp->uf_flag == 0);
1190 ufp->uf_file = fp;
1191 cv_broadcast(&ufp->uf_wanted_cv);
1192 UF_EXIT(ufp);
1193 }
1194
1195 /*
1196 * Given a file descriptor, return the file table flags, plus,
1197 * if this is a socket in asynchronous mode, the FASYNC flag.
1198 * getf() may or may not have been called before calling f_getfl().
1199 */
1200 int
1201 f_getfl(int fd, int *flagp)
1202 {
1203 uf_info_t *fip = P_FINFO(curproc);
1204 uf_entry_t *ufp;
1205 file_t *fp;
1206 int error;
1207
1208 if ((uint_t)fd >= fip->fi_nfiles)
1209 error = EBADF;
1210 else {
1211 UF_ENTER(ufp, fip, fd);
1212 if ((fp = ufp->uf_file) == NULL)
1213 error = EBADF;
1214 else {
1215 vnode_t *vp = fp->f_vnode;
1216 int flag = fp->f_flag |
1217 ((fp->f_flag2 & ~FEPOLLED) << 16);
1218
1219 /*
1220 * BSD fcntl() FASYNC compatibility.
1221 */
1222 if (vp->v_type == VSOCK)
1223 flag |= sock_getfasync(vp);
1224 *flagp = flag;
1225 error = 0;
1226 }
1227 UF_EXIT(ufp);
1228 }
1229
1230 return (error);
1231 }
1232
1233 /*
1234 * Given a file descriptor, return the user's file flags.
1235 * Force the FD_CLOEXEC flag for writable self-open /proc files.
1236 * getf() may or may not have been called before calling f_getfd_error().
1237 */
1238 int
1239 f_getfd_error(int fd, int *flagp)
1240 {
1241 uf_info_t *fip = P_FINFO(curproc);
1242 uf_entry_t *ufp;
1243 file_t *fp;
1244 int flag;
1245 int error;
1246
1247 if ((uint_t)fd >= fip->fi_nfiles)
1248 error = EBADF;
1249 else {
1250 UF_ENTER(ufp, fip, fd);
1251 if ((fp = ufp->uf_file) == NULL)
1252 error = EBADF;
1253 else {
1254 flag = ufp->uf_flag;
1255 if ((fp->f_flag & FWRITE) && pr_isself(fp->f_vnode))
1256 flag |= FD_CLOEXEC;
1257 *flagp = flag;
1258 error = 0;
1259 }
1260 UF_EXIT(ufp);
1261 }
1262
1263 return (error);
1264 }
1265
1266 /*
1267 * getf() must have been called before calling f_getfd().
1268 */
1269 char
1270 f_getfd(int fd)
1271 {
1272 int flag = 0;
1273 (void) f_getfd_error(fd, &flag);
1274 return ((char)flag);
1275 }
1276
1277 /*
1278 * Given a file descriptor and file flags, set the user's file flags.
1279 * At present, the only valid flag is FD_CLOEXEC.
1280 * getf() may or may not have been called before calling f_setfd_error().
1281 */
1282 int
1283 f_setfd_error(int fd, int flags)
1284 {
1285 uf_info_t *fip = P_FINFO(curproc);
1286 uf_entry_t *ufp;
1287 int error;
1288
1289 if ((uint_t)fd >= fip->fi_nfiles)
1290 error = EBADF;
1291 else {
1292 UF_ENTER(ufp, fip, fd);
1293 if (ufp->uf_file == NULL)
1294 error = EBADF;
1295 else {
1296 ufp->uf_flag = flags & FD_CLOEXEC;
1297 error = 0;
1298 }
1299 UF_EXIT(ufp);
1300 }
1301 return (error);
1302 }
1303
1304 void
1305 f_setfd(int fd, char flags)
1306 {
1307 (void) f_setfd_error(fd, flags);
1308 }
1309
1310 #define BADFD_MIN 3
1311 #define BADFD_MAX 255
1312
1313 /*
1314 * Attempt to allocate a file descriptor which is bad and which
1315 * is "poison" to the application. It cannot be closed (except
1316 * on exec), allocated for a different use, etc.
1317 */
1318 int
1319 f_badfd(int start, int *fdp, int action)
1320 {
1321 int fdr;
1322 int badfd;
1323 uf_info_t *fip = P_FINFO(curproc);
1324
1325 #ifdef _LP64
1326 /* No restrictions on 64 bit _file */
1327 if (get_udatamodel() != DATAMODEL_ILP32)
1328 return (EINVAL);
1329 #endif
1330
1331 if (start > BADFD_MAX || start < BADFD_MIN)
1332 return (EINVAL);
1333
1334 if (action >= NSIG || action < 0)
1335 return (EINVAL);
1336
1337 mutex_enter(&fip->fi_lock);
1338 badfd = fip->fi_badfd;
1339 mutex_exit(&fip->fi_lock);
1340
1341 if (badfd != -1)
1342 return (EAGAIN);
1343
1344 fdr = ufalloc(start);
1345
1346 if (fdr > BADFD_MAX) {
1347 setf(fdr, NULL);
1348 return (EMFILE);
1349 }
1350 if (fdr < 0)
1351 return (EMFILE);
1352
1353 mutex_enter(&fip->fi_lock);
1354 if (fip->fi_badfd != -1) {
1355 /* Lost race */
1356 mutex_exit(&fip->fi_lock);
1357 setf(fdr, NULL);
1358 return (EAGAIN);
1359 }
1360 fip->fi_action = action;
1361 fip->fi_badfd = fdr;
1362 mutex_exit(&fip->fi_lock);
1363 setf(fdr, NULL);
1364
1365 *fdp = fdr;
1366
1367 return (0);
1368 }
1369
1370 /*
1371 * Allocate a file descriptor and assign it to the vnode "*vpp",
1372 * performing the usual open protocol upon it and returning the
1373 * file descriptor allocated. It is the responsibility of the
1374 * caller to dispose of "*vpp" if any error occurs.
1375 */
1376 int
1377 fassign(vnode_t **vpp, int mode, int *fdp)
1378 {
1379 file_t *fp;
1380 int error;
1381 int fd;
1382
1383 if (error = falloc((vnode_t *)NULL, mode, &fp, &fd))
1384 return (error);
1385 if (error = VOP_OPEN(vpp, mode, fp->f_cred, NULL)) {
1386 setf(fd, NULL);
1387 unfalloc(fp);
1388 return (error);
1389 }
1390 fp->f_vnode = *vpp;
1391 mutex_exit(&fp->f_tlock);
1392 /*
1393 * Fill in the slot falloc reserved.
1394 */
1395 setf(fd, fp);
1396 *fdp = fd;
1397 return (0);
1398 }
1399
1400 /*
1401 * When a process forks it must increment the f_count of all file pointers
1402 * since there is a new process pointing at them. fcnt_add(fip, 1) does this.
1403 * Since we are called when there is only 1 active lwp we don't need to
1404 * hold fi_lock or any uf_lock. If the fork fails, fork_fail() calls
1405 * fcnt_add(fip, -1) to restore the counts.
1406 */
1407 void
1408 fcnt_add(uf_info_t *fip, int incr)
1409 {
1410 int i;
1411 uf_entry_t *ufp;
1412 file_t *fp;
1413
1414 ufp = fip->fi_list;
1415 for (i = 0; i < fip->fi_nfiles; i++, ufp++) {
1416 if ((fp = ufp->uf_file) != NULL) {
1417 mutex_enter(&fp->f_tlock);
1418 ASSERT((incr == 1 && fp->f_count >= 1) ||
1419 (incr == -1 && fp->f_count >= 2));
1420 fp->f_count += incr;
1421 mutex_exit(&fp->f_tlock);
1422 }
1423 }
1424 }
1425
1426 /*
1427 * This is called from exec to close all fd's that have the FD_CLOEXEC flag
1428 * set and also to close all self-open for write /proc file descriptors.
1429 */
1430 void
1431 close_exec(uf_info_t *fip)
1432 {
1433 int fd;
1434 file_t *fp;
1435 fpollinfo_t *fpip;
1436 uf_entry_t *ufp;
1437 portfd_t *pfd;
1438
1439 ufp = fip->fi_list;
1440 for (fd = 0; fd < fip->fi_nfiles; fd++, ufp++) {
1441 if ((fp = ufp->uf_file) != NULL &&
1442 ((ufp->uf_flag & FD_CLOEXEC) ||
1443 ((fp->f_flag & FWRITE) && pr_isself(fp->f_vnode)))) {
1444 fpip = ufp->uf_fpollinfo;
1445 mutex_enter(&fip->fi_lock);
1446 mutex_enter(&ufp->uf_lock);
1447 fd_reserve(fip, fd, -1);
1448 mutex_exit(&fip->fi_lock);
1449 ufp->uf_file = NULL;
1450 ufp->uf_fpollinfo = NULL;
1451 ufp->uf_flag = 0;
1452 /*
1453 * We may need to cleanup some cached poll states
1454 * in t_pollstate before the fd can be reused. It
1455 * is important that we don't access a stale thread
1456 * structure. We will do the cleanup in two
1457 * phases to avoid deadlock and holding uf_lock for
1458 * too long. In phase 1, hold the uf_lock and call
1459 * pollblockexit() to set state in t_pollstate struct
1460 * so that a thread does not exit on us. In phase 2,
1461 * we drop the uf_lock and call pollcacheclean().
1462 */
1463 pfd = ufp->uf_portfd;
1464 ufp->uf_portfd = NULL;
1465 if (fpip != NULL)
1466 pollblockexit(fpip);
1467 mutex_exit(&ufp->uf_lock);
1468 if (fpip != NULL)
1469 pollcacheclean(fpip, fd);
1470 if (pfd)
1471 port_close_fd(pfd);
1472 (void) closef(fp);
1473 }
1474 }
1475
1476 /* Reset bad fd */
1477 fip->fi_badfd = -1;
1478 fip->fi_action = -1;
1479 }
1480
1481 /*
1482 * Utility function called by most of the *at() system call interfaces.
1483 *
1484 * Generate a starting vnode pointer for an (fd, path) pair where 'fd'
1485 * is an open file descriptor for a directory to be used as the starting
1486 * point for the lookup of the relative pathname 'path' (or, if path is
1487 * NULL, generate a vnode pointer for the direct target of the operation).
1488 *
1489 * If we successfully return a non-NULL startvp, it has been the target
1490 * of VN_HOLD() and the caller must call VN_RELE() on it.
1491 */
1492 int
1493 fgetstartvp(int fd, char *path, vnode_t **startvpp)
1494 {
1495 vnode_t *startvp;
1496 file_t *startfp;
1497 char startchar;
1498
1499 if (fd == AT_FDCWD && path == NULL)
1500 return (EFAULT);
1501
1502 if (fd == AT_FDCWD) {
1503 /*
1504 * Start from the current working directory.
1505 */
1506 startvp = NULL;
1507 } else {
1508 if (path == NULL)
1509 startchar = '\0';
1510 else if (copyin(path, &startchar, sizeof (char)))
1511 return (EFAULT);
1512
1513 if (startchar == '/') {
1514 /*
1515 * 'path' is an absolute pathname.
1516 */
1517 startvp = NULL;
1518 } else {
1519 /*
1520 * 'path' is a relative pathname or we will
1521 * be applying the operation to 'fd' itself.
1522 */
1523 if ((startfp = getf(fd)) == NULL)
1524 return (EBADF);
1525 startvp = startfp->f_vnode;
1526 VN_HOLD(startvp);
1527 releasef(fd);
1528 }
1529 }
1530 *startvpp = startvp;
1531 return (0);
1532 }
1533
1534 /*
1535 * Called from fchownat() and fchmodat() to set ownership and mode.
1536 * The contents of *vap must be set before calling here.
1537 */
1538 int
1539 fsetattrat(int fd, char *path, int flags, struct vattr *vap)
1540 {
1541 vnode_t *startvp;
1542 vnode_t *vp;
1543 int error;
1544
1545 /*
1546 * Since we are never called to set the size of a file, we don't
1547 * need to check for non-blocking locks (via nbl_need_check(vp)).
1548 */
1549 ASSERT(!(vap->va_mask & AT_SIZE));
1550
1551 if ((error = fgetstartvp(fd, path, &startvp)) != 0)
1552 return (error);
1553 if (AU_AUDITING() && startvp != NULL)
1554 audit_setfsat_path(1);
1555
1556 /*
1557 * Do lookup for fchownat/fchmodat when path not NULL
1558 */
1559 if (path != NULL) {
1560 if (error = lookupnameat(path, UIO_USERSPACE,
1561 (flags == AT_SYMLINK_NOFOLLOW) ?
1562 NO_FOLLOW : FOLLOW,
1563 NULLVPP, &vp, startvp)) {
1564 if (startvp != NULL)
1565 VN_RELE(startvp);
1566 return (error);
1567 }
1568 } else {
1569 vp = startvp;
1570 ASSERT(vp);
1571 VN_HOLD(vp);
1572 }
1573
1574 if (vn_is_readonly(vp)) {
1575 error = EROFS;
1576 } else {
1577 error = VOP_SETATTR(vp, vap, 0, CRED(), NULL);
1578 }
1579
1580 if (startvp != NULL)
1581 VN_RELE(startvp);
1582 VN_RELE(vp);
1583
1584 return (error);
1585 }
1586
1587 /*
1588 * Return true if the given vnode is referenced by any
1589 * entry in the current process's file descriptor table.
1590 */
1591 int
1592 fisopen(vnode_t *vp)
1593 {
1594 int fd;
1595 file_t *fp;
1596 vnode_t *ovp;
1597 uf_info_t *fip = P_FINFO(curproc);
1598 uf_entry_t *ufp;
1599
1600 mutex_enter(&fip->fi_lock);
1601 for (fd = 0; fd < fip->fi_nfiles; fd++) {
1602 UF_ENTER(ufp, fip, fd);
1603 if ((fp = ufp->uf_file) != NULL &&
1604 (ovp = fp->f_vnode) != NULL && VN_CMP(vp, ovp)) {
1605 UF_EXIT(ufp);
1606 mutex_exit(&fip->fi_lock);
1607 return (1);
1608 }
1609 UF_EXIT(ufp);
1610 }
1611 mutex_exit(&fip->fi_lock);
1612 return (0);
1613 }
1614
1615 /*
1616 * Return zero if at least one file currently open (by curproc) shouldn't be
1617 * allowed to change zones.
1618 */
1619 int
1620 files_can_change_zones(void)
1621 {
1622 int fd;
1623 file_t *fp;
1624 uf_info_t *fip = P_FINFO(curproc);
1625 uf_entry_t *ufp;
1626
1627 mutex_enter(&fip->fi_lock);
1628 for (fd = 0; fd < fip->fi_nfiles; fd++) {
1629 UF_ENTER(ufp, fip, fd);
1630 if ((fp = ufp->uf_file) != NULL &&
1631 !vn_can_change_zones(fp->f_vnode)) {
1632 UF_EXIT(ufp);
1633 mutex_exit(&fip->fi_lock);
1634 return (0);
1635 }
1636 UF_EXIT(ufp);
1637 }
1638 mutex_exit(&fip->fi_lock);
1639 return (1);
1640 }
1641
1642 #ifdef DEBUG
1643
1644 /*
1645 * The following functions are only used in ASSERT()s elsewhere.
1646 * They do not modify the state of the system.
1647 */
1648
1649 /*
1650 * Return true (1) if the current thread is in the fpollinfo
1651 * list for this file descriptor, else false (0).
1652 */
1653 static int
1654 curthread_in_plist(uf_entry_t *ufp)
1655 {
1656 fpollinfo_t *fpip;
1657
1658 ASSERT(MUTEX_HELD(&ufp->uf_lock));
1659 for (fpip = ufp->uf_fpollinfo; fpip; fpip = fpip->fp_next)
1660 if (fpip->fp_thread == curthread)
1661 return (1);
1662 return (0);
1663 }
1664
1665 /*
1666 * Sanity check to make sure that after lwp_exit(),
1667 * curthread does not appear on any fd's fpollinfo list.
1668 */
1669 void
1670 checkfpollinfo(void)
1671 {
1672 int fd;
1673 uf_info_t *fip = P_FINFO(curproc);
1674 uf_entry_t *ufp;
1675
1676 mutex_enter(&fip->fi_lock);
1677 for (fd = 0; fd < fip->fi_nfiles; fd++) {
1678 UF_ENTER(ufp, fip, fd);
1679 ASSERT(!curthread_in_plist(ufp));
1680 UF_EXIT(ufp);
1681 }
1682 mutex_exit(&fip->fi_lock);
1683 }
1684
1685 /*
1686 * Return true (1) if the current thread is in the fpollinfo
1687 * list for this file descriptor, else false (0).
1688 * This is the same as curthread_in_plist(),
1689 * but is called w/o holding uf_lock.
1690 */
1691 int
1692 infpollinfo(int fd)
1693 {
1694 uf_info_t *fip = P_FINFO(curproc);
1695 uf_entry_t *ufp;
1696 int rc;
1697
1698 UF_ENTER(ufp, fip, fd);
1699 rc = curthread_in_plist(ufp);
1700 UF_EXIT(ufp);
1701 return (rc);
1702 }
1703
1704 #endif /* DEBUG */
1705
1706 /*
1707 * Add the curthread to fpollinfo list, meaning this fd is currently in the
1708 * thread's poll cache. Each lwp polling this file descriptor should call
1709 * this routine once.
1710 */
1711 void
1712 addfpollinfo(int fd)
1713 {
1714 struct uf_entry *ufp;
1715 fpollinfo_t *fpip;
1716 uf_info_t *fip = P_FINFO(curproc);
1717
1718 fpip = kmem_zalloc(sizeof (fpollinfo_t), KM_SLEEP);
1719 fpip->fp_thread = curthread;
1720 UF_ENTER(ufp, fip, fd);
1721 /*
1722 * Assert we are not already on the list, that is, that
1723 * this lwp did not call addfpollinfo twice for the same fd.
1724 */
1725 ASSERT(!curthread_in_plist(ufp));
1726 /*
1727 * addfpollinfo is always done inside the getf/releasef pair.
1728 */
1729 ASSERT(ufp->uf_refcnt >= 1);
1730 fpip->fp_next = ufp->uf_fpollinfo;
1731 ufp->uf_fpollinfo = fpip;
1732 UF_EXIT(ufp);
1733 }
1734
1735 /*
1736 * Delete curthread from fpollinfo list if it is there.
1737 */
1738 void
1739 delfpollinfo(int fd)
1740 {
1741 struct uf_entry *ufp;
1742 struct fpollinfo *fpip;
1743 struct fpollinfo **fpipp;
1744 uf_info_t *fip = P_FINFO(curproc);
1745
1746 UF_ENTER(ufp, fip, fd);
1747 for (fpipp = &ufp->uf_fpollinfo;
1748 (fpip = *fpipp) != NULL;
1749 fpipp = &fpip->fp_next) {
1750 if (fpip->fp_thread == curthread) {
1751 *fpipp = fpip->fp_next;
1752 kmem_free(fpip, sizeof (fpollinfo_t));
1753 break;
1754 }
1755 }
1756 /*
1757 * Assert that we are not still on the list, that is, that
1758 * this lwp did not call addfpollinfo twice for the same fd.
1759 */
1760 ASSERT(!curthread_in_plist(ufp));
1761 UF_EXIT(ufp);
1762 }
1763
1764 /*
1765 * fd is associated with a port. pfd is a pointer to the fd entry in the
1766 * cache of the port.
1767 */
1768
1769 void
1770 addfd_port(int fd, portfd_t *pfd)
1771 {
1772 struct uf_entry *ufp;
1773 uf_info_t *fip = P_FINFO(curproc);
1774
1775 UF_ENTER(ufp, fip, fd);
1776 /*
1777 * addfd_port is always done inside the getf/releasef pair.
1778 */
1779 ASSERT(ufp->uf_refcnt >= 1);
1780 if (ufp->uf_portfd == NULL) {
1781 /* first entry */
1782 ufp->uf_portfd = pfd;
1783 pfd->pfd_next = NULL;
1784 } else {
1785 pfd->pfd_next = ufp->uf_portfd;
1786 ufp->uf_portfd = pfd;
1787 pfd->pfd_next->pfd_prev = pfd;
1788 }
1789 UF_EXIT(ufp);
1790 }
1791
1792 void
1793 delfd_port(int fd, portfd_t *pfd)
1794 {
1795 struct uf_entry *ufp;
1796 uf_info_t *fip = P_FINFO(curproc);
1797
1798 UF_ENTER(ufp, fip, fd);
1799 /*
1800 * delfd_port is always done inside the getf/releasef pair.
1801 */
1802 ASSERT(ufp->uf_refcnt >= 1);
1803 if (ufp->uf_portfd == pfd) {
1804 /* remove first entry */
1805 ufp->uf_portfd = pfd->pfd_next;
1806 } else {
1807 pfd->pfd_prev->pfd_next = pfd->pfd_next;
1808 if (pfd->pfd_next != NULL)
1809 pfd->pfd_next->pfd_prev = pfd->pfd_prev;
1810 }
1811 UF_EXIT(ufp);
1812 }
1813
1814 static void
1815 port_close_fd(portfd_t *pfd)
1816 {
1817 portfd_t *pfdn;
1818
1819 /*
1820 * At this point, no other thread should access
1821 * the portfd_t list for this fd. The uf_file, uf_portfd
1822 * pointers in the uf_entry_t struct for this fd would
1823 * be set to NULL.
1824 */
1825 for (; pfd != NULL; pfd = pfdn) {
1826 pfdn = pfd->pfd_next;
1827 port_close_pfd(pfd);
1828 }
1829 }