1 /* $NetBSD: kern_descrip.c,v 1.265 2024/12/21 19:02:31 riastradh Exp $ */
2
3 /*-
4 * Copyright (c) 2008, 2009, 2023 The NetBSD Foundation, Inc.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The NetBSD Foundation
8 * by Andrew Doran.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29 * POSSIBILITY OF SUCH DAMAGE.
30 */
31
32 /*
33 * Copyright (c) 1982, 1986, 1989, 1991, 1993
34 * The Regents of the University of California. All rights reserved.
35 * (c) UNIX System Laboratories, Inc.
36 * All or some portions of this file are derived from material licensed
37 * to the University of California by American Telephone and Telegraph
38 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
39 * the permission of UNIX System Laboratories, Inc.
40 *
41 * Redistribution and use in source and binary forms, with or without
42 * modification, are permitted provided that the following conditions
43 * are met:
44 * 1. Redistributions of source code must retain the above copyright
45 * notice, this list of conditions and the following disclaimer.
46 * 2. Redistributions in binary form must reproduce the above copyright
47 * notice, this list of conditions and the following disclaimer in the
48 * documentation and/or other materials provided with the distribution.
49 * 3. Neither the name of the University nor the names of its contributors
50 * may be used to endorse or promote products derived from this software
51 * without specific prior written permission.
52 *
53 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
54 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
55 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
56 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
57 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
58 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
59 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
60 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
61 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
62 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
63 * SUCH DAMAGE.
64 *
65 * @(#)kern_descrip.c 8.8 (Berkeley) 2/14/95
66 */
67
68 /*
69 * File descriptor management.
70 */
71
72 #include <sys/cdefs.h>
73 __KERNEL_RCSID(0, "$NetBSD: kern_descrip.c,v 1.265 2024/12/21 19:02:31 riastradh Exp $");
74
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/filedesc.h>
78 #include <sys/kernel.h>
79 #include <sys/proc.h>
80 #include <sys/file.h>
81 #include <sys/socket.h>
82 #include <sys/socketvar.h>
83 #include <sys/stat.h>
84 #include <sys/ioctl.h>
85 #include <sys/fcntl.h>
86 #include <sys/pool.h>
87 #include <sys/unistd.h>
88 #include <sys/resourcevar.h>
89 #include <sys/conf.h>
90 #include <sys/event.h>
91 #include <sys/kauth.h>
92 #include <sys/atomic.h>
93 #include <sys/syscallargs.h>
94 #include <sys/cpu.h>
95 #include <sys/kmem.h>
96 #include <sys/vnode.h>
97 #include <sys/sysctl.h>
98 #include <sys/ktrace.h>
99
100 /*
101 * A list (head) of open files, counter, and lock protecting them.
102 */
103 struct filelist filehead __cacheline_aligned;
104 static u_int nfiles __cacheline_aligned;
105 kmutex_t filelist_lock __cacheline_aligned;
106
107 static pool_cache_t filedesc_cache __read_mostly;
108 static pool_cache_t file_cache __read_mostly;
109
110 static int file_ctor(void *, void *, int);
111 static void file_dtor(void *, void *);
112 static void fdfile_ctor(fdfile_t *);
113 static void fdfile_dtor(fdfile_t *);
114 static int filedesc_ctor(void *, void *, int);
115 static void filedesc_dtor(void *, void *);
116 static int filedescopen(dev_t, int, int, lwp_t *);
117
118 static int sysctl_kern_file(SYSCTLFN_PROTO);
119 static int sysctl_kern_file2(SYSCTLFN_PROTO);
120 static void fill_file(struct file *, const struct file *);
121 static void fill_file2(struct kinfo_file *, const file_t *, const fdfile_t *,
122 int, pid_t);
123
124 const struct cdevsw filedesc_cdevsw = {
125 .d_open = filedescopen,
126 .d_close = noclose,
127 .d_read = noread,
128 .d_write = nowrite,
129 .d_ioctl = noioctl,
130 .d_stop = nostop,
131 .d_tty = notty,
132 .d_poll = nopoll,
133 .d_mmap = nommap,
134 .d_kqfilter = nokqfilter,
135 .d_discard = nodiscard,
136 .d_flag = D_OTHER | D_MPSAFE
137 };
138
139 /* For ease of reading. */
__strong_alias(fd_putvnode,fd_putfile)140 __strong_alias(fd_putvnode,fd_putfile)
141 __strong_alias(fd_putsock,fd_putfile)
142
143 /*
144 * Initialize the descriptor system.
145 */
146 void
147 fd_sys_init(void)
148 {
149 static struct sysctllog *clog;
150
151 mutex_init(&filelist_lock, MUTEX_DEFAULT, IPL_NONE);
152
153 LIST_INIT(&filehead);
154
155 file_cache = pool_cache_init(sizeof(file_t), coherency_unit, 0,
156 0, "file", NULL, IPL_NONE, file_ctor, file_dtor, NULL);
157 KASSERT(file_cache != NULL);
158
159 filedesc_cache = pool_cache_init(sizeof(filedesc_t), coherency_unit,
160 0, 0, "filedesc", NULL, IPL_NONE, filedesc_ctor, filedesc_dtor,
161 NULL);
162 KASSERT(filedesc_cache != NULL);
163
164 sysctl_createv(&clog, 0, NULL, NULL,
165 CTLFLAG_PERMANENT,
166 CTLTYPE_STRUCT, "file",
167 SYSCTL_DESCR("System open file table"),
168 sysctl_kern_file, 0, NULL, 0,
169 CTL_KERN, KERN_FILE, CTL_EOL);
170 sysctl_createv(&clog, 0, NULL, NULL,
171 CTLFLAG_PERMANENT,
172 CTLTYPE_STRUCT, "file2",
173 SYSCTL_DESCR("System open file table"),
174 sysctl_kern_file2, 0, NULL, 0,
175 CTL_KERN, KERN_FILE2, CTL_EOL);
176 }
177
178 static bool
fd_isused(filedesc_t * fdp,unsigned fd)179 fd_isused(filedesc_t *fdp, unsigned fd)
180 {
181 u_int off = fd >> NDENTRYSHIFT;
182
183 KASSERT(fd < atomic_load_consume(&fdp->fd_dt)->dt_nfiles);
184
185 return (fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) != 0;
186 }
187
188 /*
189 * Verify that the bitmaps match the descriptor table.
190 */
191 static inline void
fd_checkmaps(filedesc_t * fdp)192 fd_checkmaps(filedesc_t *fdp)
193 {
194 #ifdef DEBUG
195 fdtab_t *dt;
196 u_int fd;
197
198 KASSERT(fdp->fd_refcnt <= 1 || mutex_owned(&fdp->fd_lock));
199
200 dt = fdp->fd_dt;
201 if (fdp->fd_refcnt == -1) {
202 /*
203 * fd_free tears down the table without maintaining its bitmap.
204 */
205 return;
206 }
207 for (fd = 0; fd < dt->dt_nfiles; fd++) {
208 if (fd < NDFDFILE) {
209 KASSERT(dt->dt_ff[fd] ==
210 (fdfile_t *)fdp->fd_dfdfile[fd]);
211 }
212 if (dt->dt_ff[fd] == NULL) {
213 KASSERT(!fd_isused(fdp, fd));
214 } else if (dt->dt_ff[fd]->ff_file != NULL) {
215 KASSERT(fd_isused(fdp, fd));
216 }
217 }
218 #endif
219 }
220
221 static int
fd_next_zero(filedesc_t * fdp,uint32_t * bitmap,int want,u_int bits)222 fd_next_zero(filedesc_t *fdp, uint32_t *bitmap, int want, u_int bits)
223 {
224 int i, off, maxoff;
225 uint32_t sub;
226
227 KASSERT(mutex_owned(&fdp->fd_lock));
228
229 fd_checkmaps(fdp);
230
231 if (want > bits)
232 return -1;
233
234 off = want >> NDENTRYSHIFT;
235 i = want & NDENTRYMASK;
236 if (i) {
237 sub = bitmap[off] | ((u_int)~0 >> (NDENTRIES - i));
238 if (sub != ~0)
239 goto found;
240 off++;
241 }
242
243 maxoff = NDLOSLOTS(bits);
244 while (off < maxoff) {
245 if ((sub = bitmap[off]) != ~0)
246 goto found;
247 off++;
248 }
249
250 return -1;
251
252 found:
253 return (off << NDENTRYSHIFT) + ffs(~sub) - 1;
254 }
255
256 static int
fd_last_set(filedesc_t * fd,int last)257 fd_last_set(filedesc_t *fd, int last)
258 {
259 int off, i;
260 fdfile_t **ff = fd->fd_dt->dt_ff;
261 uint32_t *bitmap = fd->fd_lomap;
262
263 KASSERT(mutex_owned(&fd->fd_lock));
264
265 fd_checkmaps(fd);
266
267 off = (last - 1) >> NDENTRYSHIFT;
268
269 while (off >= 0 && !bitmap[off])
270 off--;
271
272 if (off < 0)
273 return -1;
274
275 i = ((off + 1) << NDENTRYSHIFT) - 1;
276 if (i >= last)
277 i = last - 1;
278
279 /* XXX should use bitmap */
280 while (i > 0 && (ff[i] == NULL || !ff[i]->ff_allocated))
281 i--;
282
283 return i;
284 }
285
286 static inline void
fd_used(filedesc_t * fdp,unsigned fd)287 fd_used(filedesc_t *fdp, unsigned fd)
288 {
289 u_int off = fd >> NDENTRYSHIFT;
290 fdfile_t *ff;
291
292 ff = fdp->fd_dt->dt_ff[fd];
293
294 KASSERT(mutex_owned(&fdp->fd_lock));
295 KASSERT((fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) == 0);
296 KASSERT(ff != NULL);
297 KASSERT(ff->ff_file == NULL);
298 KASSERT(!ff->ff_allocated);
299
300 ff->ff_allocated = true;
301 fdp->fd_lomap[off] |= 1U << (fd & NDENTRYMASK);
302 if (__predict_false(fdp->fd_lomap[off] == ~0)) {
303 KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] &
304 (1U << (off & NDENTRYMASK))) == 0);
305 fdp->fd_himap[off >> NDENTRYSHIFT] |= 1U << (off & NDENTRYMASK);
306 }
307
308 if ((int)fd > fdp->fd_lastfile) {
309 fdp->fd_lastfile = fd;
310 }
311
312 fd_checkmaps(fdp);
313 }
314
315 static inline void
fd_unused(filedesc_t * fdp,unsigned fd)316 fd_unused(filedesc_t *fdp, unsigned fd)
317 {
318 u_int off = fd >> NDENTRYSHIFT;
319 fdfile_t *ff;
320
321 ff = fdp->fd_dt->dt_ff[fd];
322
323 KASSERT(mutex_owned(&fdp->fd_lock));
324 KASSERT(ff != NULL);
325 KASSERT(ff->ff_file == NULL);
326 KASSERT(ff->ff_allocated);
327
328 if (fd < fdp->fd_freefile) {
329 fdp->fd_freefile = fd;
330 }
331
332 if (fdp->fd_lomap[off] == ~0) {
333 KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] &
334 (1U << (off & NDENTRYMASK))) != 0);
335 fdp->fd_himap[off >> NDENTRYSHIFT] &=
336 ~(1U << (off & NDENTRYMASK));
337 }
338 KASSERT((fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) != 0);
339 fdp->fd_lomap[off] &= ~(1U << (fd & NDENTRYMASK));
340 ff->ff_allocated = false;
341
342 KASSERT(fd <= fdp->fd_lastfile);
343 if (fd == fdp->fd_lastfile) {
344 fdp->fd_lastfile = fd_last_set(fdp, fd);
345 }
346 fd_checkmaps(fdp);
347 }
348
349 /*
350 * Look up the file structure corresponding to a file descriptor
351 * and return the file, holding a reference on the descriptor.
352 */
353 file_t *
fd_getfile(unsigned fd)354 fd_getfile(unsigned fd)
355 {
356 filedesc_t *fdp;
357 fdfile_t *ff;
358 file_t *fp;
359 fdtab_t *dt;
360
361 /*
362 * Look up the fdfile structure representing this descriptor.
363 * We are doing this unlocked. See fd_tryexpand().
364 */
365 fdp = curlwp->l_fd;
366 dt = atomic_load_consume(&fdp->fd_dt);
367 if (__predict_false(fd >= dt->dt_nfiles)) {
368 return NULL;
369 }
370 ff = dt->dt_ff[fd];
371 KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
372 if (__predict_false(ff == NULL)) {
373 return NULL;
374 }
375
376 /* Now get a reference to the descriptor. */
377 if (fdp->fd_refcnt == 1) {
378 /*
379 * Single threaded: don't need to worry about concurrent
380 * access (other than earlier calls to kqueue, which may
381 * hold a reference to the descriptor).
382 */
383 ff->ff_refcnt++;
384 } else {
385 /*
386 * Multi threaded: issue a memory barrier to ensure that we
387 * acquire the file pointer _after_ adding a reference. If
388 * no memory barrier, we could fetch a stale pointer.
389 *
390 * In particular, we must coordinate the following four
391 * memory operations:
392 *
393 * A. fd_close store ff->ff_file = NULL
394 * B. fd_close refcnt = atomic_dec_uint_nv(&ff->ff_refcnt)
395 * C. fd_getfile atomic_inc_uint(&ff->ff_refcnt)
396 * D. fd_getfile load fp = ff->ff_file
397 *
398 * If the order is D;A;B;C:
399 *
400 * 1. D: fp = ff->ff_file
401 * 2. A: ff->ff_file = NULL
402 * 3. B: refcnt = atomic_dec_uint_nv(&ff->ff_refcnt)
403 * 4. C: atomic_inc_uint(&ff->ff_refcnt)
404 *
405 * then fd_close determines that there are no more
406 * references and decides to free fp immediately, at
407 * the same that fd_getfile ends up with an fp that's
408 * about to be freed. *boom*
409 *
410 * By making B a release operation in fd_close, and by
411 * making C an acquire operation in fd_getfile, since
412 * they are atomic operations on the same object, which
413 * has a total modification order, we guarantee either:
414 *
415 * - B happens before C. Then since A is
416 * sequenced before B in fd_close, and C is
417 * sequenced before D in fd_getfile, we
418 * guarantee A happens before D, so fd_getfile
419 * reads a null fp and safely fails.
420 *
421 * - C happens before B. Then fd_getfile may read
422 * null or nonnull, but either way, fd_close
423 * will safely wait for references to drain.
424 */
425 atomic_inc_uint(&ff->ff_refcnt);
426 membar_acquire();
427 }
428
429 /*
430 * If the file is not open or is being closed then put the
431 * reference back.
432 */
433 fp = atomic_load_consume(&ff->ff_file);
434 if (__predict_true(fp != NULL)) {
435 return fp;
436 }
437 fd_putfile(fd);
438 return NULL;
439 }
440
441 /*
442 * Release a reference to a file descriptor acquired with fd_getfile().
443 */
444 void
fd_putfile(unsigned fd)445 fd_putfile(unsigned fd)
446 {
447 filedesc_t *fdp;
448 fdfile_t *ff;
449 u_int u, v;
450
451 fdp = curlwp->l_fd;
452 KASSERT(fd < atomic_load_consume(&fdp->fd_dt)->dt_nfiles);
453 ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
454
455 KASSERT(ff != NULL);
456 KASSERT((ff->ff_refcnt & FR_MASK) > 0);
457 KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
458
459 if (fdp->fd_refcnt == 1) {
460 /*
461 * Single threaded: don't need to worry about concurrent
462 * access (other than earlier calls to kqueue, which may
463 * hold a reference to the descriptor).
464 */
465 if (__predict_false((ff->ff_refcnt & FR_CLOSING) != 0)) {
466 fd_close(fd);
467 return;
468 }
469 ff->ff_refcnt--;
470 return;
471 }
472
473 /*
474 * Ensure that any use of the file is complete and globally
475 * visible before dropping the final reference. If no membar,
476 * the current CPU could still access memory associated with
477 * the file after it has been freed or recycled by another
478 * CPU.
479 */
480 membar_release();
481
482 /*
483 * Be optimistic and start out with the assumption that no other
484 * threads are trying to close the descriptor. If the CAS fails,
485 * we lost a race and/or it's being closed.
486 */
487 for (u = ff->ff_refcnt & FR_MASK;; u = v) {
488 v = atomic_cas_uint(&ff->ff_refcnt, u, u - 1);
489 if (__predict_true(u == v)) {
490 return;
491 }
492 if (__predict_false((v & FR_CLOSING) != 0)) {
493 break;
494 }
495 }
496
497 /* Another thread is waiting to close the file: join it. */
498 (void)fd_close(fd);
499 }
500
501 /*
502 * Convenience wrapper around fd_getfile() that returns reference
503 * to a vnode.
504 */
505 int
fd_getvnode(unsigned fd,file_t ** fpp)506 fd_getvnode(unsigned fd, file_t **fpp)
507 {
508 vnode_t *vp;
509 file_t *fp;
510
511 fp = fd_getfile(fd);
512 if (__predict_false(fp == NULL)) {
513 return EBADF;
514 }
515 if (__predict_false(fp->f_type != DTYPE_VNODE)) {
516 fd_putfile(fd);
517 return EINVAL;
518 }
519 vp = fp->f_vnode;
520 if (__predict_false(vp->v_type == VBAD)) {
521 /* XXX Is this case really necessary? */
522 fd_putfile(fd);
523 return EBADF;
524 }
525 *fpp = fp;
526 return 0;
527 }
528
529 /*
530 * Convenience wrapper around fd_getfile() that returns reference
531 * to a socket.
532 */
533 int
fd_getsock1(unsigned fd,struct socket ** sop,file_t ** fp)534 fd_getsock1(unsigned fd, struct socket **sop, file_t **fp)
535 {
536 *fp = fd_getfile(fd);
537 if (__predict_false(*fp == NULL)) {
538 return EBADF;
539 }
540 if (__predict_false((*fp)->f_type != DTYPE_SOCKET)) {
541 fd_putfile(fd);
542 return ENOTSOCK;
543 }
544 *sop = (*fp)->f_socket;
545 return 0;
546 }
547
548 int
fd_getsock(unsigned fd,struct socket ** sop)549 fd_getsock(unsigned fd, struct socket **sop)
550 {
551 file_t *fp;
552 return fd_getsock1(fd, sop, &fp);
553 }
554
555 /*
556 * Look up the file structure corresponding to a file descriptor
557 * and return it with a reference held on the file, not the
558 * descriptor.
559 *
560 * This is heavyweight and only used when accessing descriptors
561 * from a foreign process. The caller must ensure that `p' does
562 * not exit or fork across this call.
563 *
564 * To release the file (not descriptor) reference, use closef().
565 */
566 file_t *
fd_getfile2(proc_t * p,unsigned fd)567 fd_getfile2(proc_t *p, unsigned fd)
568 {
569 filedesc_t *fdp;
570 fdfile_t *ff;
571 file_t *fp;
572 fdtab_t *dt;
573
574 fdp = p->p_fd;
575 mutex_enter(&fdp->fd_lock);
576 dt = fdp->fd_dt;
577 if (fd >= dt->dt_nfiles) {
578 mutex_exit(&fdp->fd_lock);
579 return NULL;
580 }
581 if ((ff = dt->dt_ff[fd]) == NULL) {
582 mutex_exit(&fdp->fd_lock);
583 return NULL;
584 }
585 if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
586 mutex_exit(&fdp->fd_lock);
587 return NULL;
588 }
589 mutex_enter(&fp->f_lock);
590 fp->f_count++;
591 mutex_exit(&fp->f_lock);
592 mutex_exit(&fdp->fd_lock);
593
594 return fp;
595 }
596
597 /*
598 * Internal form of close. Must be called with a reference to the
599 * descriptor, and will drop the reference. When all descriptor
600 * references are dropped, releases the descriptor slot and a single
601 * reference to the file structure.
602 */
603 int
fd_close(unsigned fd)604 fd_close(unsigned fd)
605 {
606 struct flock lf;
607 filedesc_t *fdp;
608 fdfile_t *ff;
609 file_t *fp;
610 proc_t *p;
611 lwp_t *l;
612 u_int refcnt;
613
614 l = curlwp;
615 p = l->l_proc;
616 fdp = l->l_fd;
617 ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
618
619 KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
620
621 mutex_enter(&fdp->fd_lock);
622 KASSERT((ff->ff_refcnt & FR_MASK) > 0);
623 fp = atomic_load_consume(&ff->ff_file);
624 if (__predict_false(fp == NULL)) {
625 /*
626 * Another user of the file is already closing, and is
627 * waiting for other users of the file to drain. Release
628 * our reference, and wake up the closer.
629 */
630 membar_release();
631 atomic_dec_uint(&ff->ff_refcnt);
632 cv_broadcast(&ff->ff_closing);
633 mutex_exit(&fdp->fd_lock);
634
635 /*
636 * An application error, so pretend that the descriptor
637 * was already closed. We can't safely wait for it to
638 * be closed without potentially deadlocking.
639 */
640 return (EBADF);
641 }
642 KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
643
644 /*
645 * There may be multiple users of this file within the process.
646 * Notify existing and new users that the file is closing. This
647 * will prevent them from adding additional uses to this file
648 * while we are closing it.
649 */
650 atomic_store_relaxed(&ff->ff_file, NULL);
651 ff->ff_exclose = false;
652
653 /*
654 * We expect the caller to hold a descriptor reference - drop it.
655 * The reference count may increase beyond zero at this point due
656 * to an erroneous descriptor reference by an application, but
657 * fd_getfile() will notice that the file is being closed and drop
658 * the reference again.
659 */
660 if (fdp->fd_refcnt == 1) {
661 /* Single threaded. */
662 refcnt = --(ff->ff_refcnt);
663 } else {
664 /* Multi threaded. */
665 membar_release();
666 refcnt = atomic_dec_uint_nv(&ff->ff_refcnt);
667 membar_acquire();
668 }
669 if (__predict_false(refcnt != 0)) {
670 /*
671 * Wait for other references to drain. This is typically
672 * an application error - the descriptor is being closed
673 * while still in use.
674 * (Or just a threaded application trying to unblock its
675 * thread that sleeps in (say) accept()).
676 */
677 atomic_or_uint(&ff->ff_refcnt, FR_CLOSING);
678
679 /*
680 * Remove any knotes attached to the file. A knote
681 * attached to the descriptor can hold references on it.
682 */
683 mutex_exit(&fdp->fd_lock);
684 if (!SLIST_EMPTY(&ff->ff_knlist)) {
685 knote_fdclose(fd);
686 }
687
688 /*
689 * Since the file system code doesn't know which fd
690 * each request came from (think dup()), we have to
691 * ask it to return ERESTART for any long-term blocks.
692 * The re-entry through read/write/etc will detect the
693 * closed fd and return EBAFD.
694 * Blocked partial writes may return a short length.
695 */
696 (*fp->f_ops->fo_restart)(fp);
697 mutex_enter(&fdp->fd_lock);
698
699 /*
700 * We need to see the count drop to zero at least once,
701 * in order to ensure that all pre-existing references
702 * have been drained. New references past this point are
703 * of no interest.
704 * XXX (dsl) this may need to call fo_restart() after a
705 * timeout to guarantee that all the system calls exit.
706 */
707 while ((ff->ff_refcnt & FR_MASK) != 0) {
708 cv_wait(&ff->ff_closing, &fdp->fd_lock);
709 }
710 atomic_and_uint(&ff->ff_refcnt, ~FR_CLOSING);
711 } else {
712 /* If no references, there must be no knotes. */
713 KASSERT(SLIST_EMPTY(&ff->ff_knlist));
714 }
715
716 /*
717 * POSIX record locking dictates that any close releases ALL
718 * locks owned by this process. This is handled by setting
719 * a flag in the unlock to free ONLY locks obeying POSIX
720 * semantics, and not to free BSD-style file locks.
721 * If the descriptor was in a message, POSIX-style locks
722 * aren't passed with the descriptor.
723 */
724 if (__predict_false((p->p_flag & PK_ADVLOCK) != 0) &&
725 fp->f_ops->fo_advlock != NULL) {
726 lf.l_whence = SEEK_SET;
727 lf.l_start = 0;
728 lf.l_len = 0;
729 lf.l_type = F_UNLCK;
730 mutex_exit(&fdp->fd_lock);
731 (void)(*fp->f_ops->fo_advlock)(fp, p, F_UNLCK, &lf, F_POSIX);
732 mutex_enter(&fdp->fd_lock);
733 }
734
735 /* Free descriptor slot. */
736 fd_unused(fdp, fd);
737 mutex_exit(&fdp->fd_lock);
738
739 /* Now drop reference to the file itself. */
740 return closef(fp);
741 }
742
743 /*
744 * Duplicate a file descriptor.
745 */
746 int
fd_dup(file_t * fp,int minfd,int * newp,bool exclose)747 fd_dup(file_t *fp, int minfd, int *newp, bool exclose)
748 {
749 proc_t *p = curproc;
750 int error;
751
752 while ((error = fd_alloc(p, minfd, newp)) != 0) {
753 if (error != ENOSPC) {
754 return error;
755 }
756 fd_tryexpand(p);
757 }
758
759 fd_set_exclose(curlwp, *newp, exclose);
760 fd_affix(p, fp, *newp);
761 return 0;
762 }
763
764 /*
765 * dup2 operation.
766 */
767 int
fd_dup2(file_t * fp,unsigned newfd,int flags)768 fd_dup2(file_t *fp, unsigned newfd, int flags)
769 {
770 filedesc_t *fdp = curlwp->l_fd;
771 fdfile_t *ff;
772 fdtab_t *dt;
773
774 if (flags & ~(O_CLOEXEC|O_NONBLOCK|O_NOSIGPIPE))
775 return EINVAL;
776 /*
777 * Ensure there are enough slots in the descriptor table,
778 * and allocate an fdfile_t up front in case we need it.
779 */
780 while (newfd >= atomic_load_consume(&fdp->fd_dt)->dt_nfiles) {
781 fd_tryexpand(curproc);
782 }
783 ff = kmem_alloc(sizeof(*ff), KM_SLEEP);
784 fdfile_ctor(ff);
785
786 /*
787 * If there is already a file open, close it. If the file is
788 * half open, wait for it to be constructed before closing it.
789 * XXX Potential for deadlock here?
790 */
791 mutex_enter(&fdp->fd_lock);
792 while (fd_isused(fdp, newfd)) {
793 mutex_exit(&fdp->fd_lock);
794 if (fd_getfile(newfd) != NULL) {
795 (void)fd_close(newfd);
796 } else {
797 /*
798 * Crummy, but unlikely to happen.
799 * Can occur if we interrupt another
800 * thread while it is opening a file.
801 */
802 kpause("dup2", false, 1, NULL);
803 }
804 mutex_enter(&fdp->fd_lock);
805 }
806 dt = fdp->fd_dt;
807 if (dt->dt_ff[newfd] == NULL) {
808 KASSERT(newfd >= NDFDFILE);
809 dt->dt_ff[newfd] = ff;
810 ff = NULL;
811 }
812 fd_used(fdp, newfd);
813 mutex_exit(&fdp->fd_lock);
814
815 fd_set_exclose(curlwp, newfd, (flags & O_CLOEXEC) != 0);
816 fp->f_flag |= flags & (FNONBLOCK|FNOSIGPIPE);
817 /* Slot is now allocated. Insert copy of the file. */
818 fd_affix(curproc, fp, newfd);
819 if (ff != NULL) {
820 cv_destroy(&ff->ff_closing);
821 kmem_free(ff, sizeof(*ff));
822 }
823 return 0;
824 }
825
826 /*
827 * Drop reference to a file structure.
828 */
829 int
closef(file_t * fp)830 closef(file_t *fp)
831 {
832 struct flock lf;
833 int error;
834
835 /*
836 * Drop reference. If referenced elsewhere it's still open
837 * and we have nothing more to do.
838 */
839 mutex_enter(&fp->f_lock);
840 KASSERT(fp->f_count > 0);
841 if (--fp->f_count > 0) {
842 mutex_exit(&fp->f_lock);
843 return 0;
844 }
845 KASSERT(fp->f_count == 0);
846 mutex_exit(&fp->f_lock);
847
848 /* We held the last reference - release locks, close and free. */
849 if (fp->f_ops->fo_advlock == NULL) {
850 KASSERT((fp->f_flag & FHASLOCK) == 0);
851 } else if (fp->f_flag & FHASLOCK) {
852 lf.l_whence = SEEK_SET;
853 lf.l_start = 0;
854 lf.l_len = 0;
855 lf.l_type = F_UNLCK;
856 (void)(*fp->f_ops->fo_advlock)(fp, fp, F_UNLCK, &lf, F_FLOCK);
857 }
858 if (fp->f_ops != NULL) {
859 error = (*fp->f_ops->fo_close)(fp);
860
861 /*
862 * .fo_close is final, so real errors are frowned on
863 * (but allowed and passed on to close(2)), and
864 * ERESTART is absolutely forbidden because the file
865 * descriptor is gone and there is no chance to retry.
866 */
867 KASSERTMSG(error != ERESTART,
868 "file %p f_ops %p fo_close %p returned ERESTART",
869 fp, fp->f_ops, fp->f_ops->fo_close);
870 } else {
871 error = 0;
872 }
873 KASSERT(fp->f_count == 0);
874 KASSERT(fp->f_cred != NULL);
875 pool_cache_put(file_cache, fp);
876
877 return error;
878 }
879
880 /*
881 * Allocate a file descriptor for the process.
882 *
883 * Future idea for experimentation: replace all of this with radixtree.
884 */
885 int
fd_alloc(proc_t * p,int want,int * result)886 fd_alloc(proc_t *p, int want, int *result)
887 {
888 filedesc_t *fdp = p->p_fd;
889 int i, lim, last, error, hi;
890 u_int off;
891 fdtab_t *dt;
892
893 KASSERT(p == curproc || p == &proc0);
894
895 /*
896 * Search for a free descriptor starting at the higher
897 * of want or fd_freefile.
898 */
899 mutex_enter(&fdp->fd_lock);
900 fd_checkmaps(fdp);
901 dt = fdp->fd_dt;
902 KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
903 lim = uimin((int)p->p_rlimit[RLIMIT_NOFILE].rlim_cur, maxfiles);
904 last = uimin(dt->dt_nfiles, lim);
905
906 for (;;) {
907 if ((i = want) < fdp->fd_freefile)
908 i = fdp->fd_freefile;
909 off = i >> NDENTRYSHIFT;
910 hi = fd_next_zero(fdp, fdp->fd_himap, off,
911 (last + NDENTRIES - 1) >> NDENTRYSHIFT);
912 if (hi == -1)
913 break;
914 i = fd_next_zero(fdp, &fdp->fd_lomap[hi],
915 hi > off ? 0 : i & NDENTRYMASK, NDENTRIES);
916 if (i == -1) {
917 /*
918 * Free file descriptor in this block was
919 * below want, try again with higher want.
920 */
921 want = (hi + 1) << NDENTRYSHIFT;
922 continue;
923 }
924 i += (hi << NDENTRYSHIFT);
925 if (i >= last) {
926 break;
927 }
928 if (dt->dt_ff[i] == NULL) {
929 KASSERT(i >= NDFDFILE);
930 dt->dt_ff[i] = kmem_alloc(sizeof(fdfile_t), KM_SLEEP);
931 fdfile_ctor(dt->dt_ff[i]);
932 }
933 KASSERT(dt->dt_ff[i]->ff_file == NULL);
934 fd_used(fdp, i);
935 if (want <= fdp->fd_freefile) {
936 fdp->fd_freefile = i;
937 }
938 *result = i;
939 KASSERT(i >= NDFDFILE ||
940 dt->dt_ff[i] == (fdfile_t *)fdp->fd_dfdfile[i]);
941 fd_checkmaps(fdp);
942 mutex_exit(&fdp->fd_lock);
943 return 0;
944 }
945
946 /* No space in current array. Let the caller expand and retry. */
947 error = (dt->dt_nfiles >= lim) ? EMFILE : ENOSPC;
948 mutex_exit(&fdp->fd_lock);
949 return error;
950 }
951
952 /*
953 * Allocate memory for a descriptor table.
954 */
955 static fdtab_t *
fd_dtab_alloc(int n)956 fd_dtab_alloc(int n)
957 {
958 fdtab_t *dt;
959 size_t sz;
960
961 KASSERT(n > NDFILE);
962
963 sz = sizeof(*dt) + (n - NDFILE) * sizeof(dt->dt_ff[0]);
964 dt = kmem_alloc(sz, KM_SLEEP);
965 #ifdef DIAGNOSTIC
966 memset(dt, 0xff, sz);
967 #endif
968 dt->dt_nfiles = n;
969 dt->dt_link = NULL;
970 return dt;
971 }
972
973 /*
974 * Free a descriptor table, and all tables linked for deferred free.
975 */
976 static void
fd_dtab_free(fdtab_t * dt)977 fd_dtab_free(fdtab_t *dt)
978 {
979 fdtab_t *next;
980 size_t sz;
981
982 do {
983 next = dt->dt_link;
984 KASSERT(dt->dt_nfiles > NDFILE);
985 sz = sizeof(*dt) +
986 (dt->dt_nfiles - NDFILE) * sizeof(dt->dt_ff[0]);
987 #ifdef DIAGNOSTIC
988 memset(dt, 0xff, sz);
989 #endif
990 kmem_free(dt, sz);
991 dt = next;
992 } while (dt != NULL);
993 }
994
995 /*
996 * Allocate descriptor bitmap.
997 */
998 static void
fd_map_alloc(int n,uint32_t ** lo,uint32_t ** hi)999 fd_map_alloc(int n, uint32_t **lo, uint32_t **hi)
1000 {
1001 uint8_t *ptr;
1002 size_t szlo, szhi;
1003
1004 KASSERT(n > NDENTRIES);
1005
1006 szlo = NDLOSLOTS(n) * sizeof(uint32_t);
1007 szhi = NDHISLOTS(n) * sizeof(uint32_t);
1008 ptr = kmem_alloc(szlo + szhi, KM_SLEEP);
1009 *lo = (uint32_t *)ptr;
1010 *hi = (uint32_t *)(ptr + szlo);
1011 }
1012
1013 /*
1014 * Free descriptor bitmap.
1015 */
1016 static void
fd_map_free(int n,uint32_t * lo,uint32_t * hi)1017 fd_map_free(int n, uint32_t *lo, uint32_t *hi)
1018 {
1019 size_t szlo, szhi;
1020
1021 KASSERT(n > NDENTRIES);
1022
1023 szlo = NDLOSLOTS(n) * sizeof(uint32_t);
1024 szhi = NDHISLOTS(n) * sizeof(uint32_t);
1025 KASSERT(hi == (uint32_t *)((uint8_t *)lo + szlo));
1026 kmem_free(lo, szlo + szhi);
1027 }
1028
1029 /*
1030 * Expand a process' descriptor table.
1031 */
1032 void
fd_tryexpand(proc_t * p)1033 fd_tryexpand(proc_t *p)
1034 {
1035 filedesc_t *fdp;
1036 int i, numfiles, oldnfiles;
1037 fdtab_t *newdt, *dt;
1038 uint32_t *newhimap, *newlomap;
1039
1040 KASSERT(p == curproc || p == &proc0);
1041
1042 fdp = p->p_fd;
1043 newhimap = NULL;
1044 newlomap = NULL;
1045 oldnfiles = atomic_load_consume(&fdp->fd_dt)->dt_nfiles;
1046
1047 if (oldnfiles < NDEXTENT)
1048 numfiles = NDEXTENT;
1049 else
1050 numfiles = 2 * oldnfiles;
1051
1052 newdt = fd_dtab_alloc(numfiles);
1053 if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1054 fd_map_alloc(numfiles, &newlomap, &newhimap);
1055 }
1056
1057 mutex_enter(&fdp->fd_lock);
1058 dt = fdp->fd_dt;
1059 KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1060 if (dt->dt_nfiles != oldnfiles) {
1061 /* fdp changed; caller must retry */
1062 mutex_exit(&fdp->fd_lock);
1063 fd_dtab_free(newdt);
1064 if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1065 fd_map_free(numfiles, newlomap, newhimap);
1066 }
1067 return;
1068 }
1069
1070 /* Copy the existing descriptor table and zero the new portion. */
1071 i = sizeof(fdfile_t *) * oldnfiles;
1072 memcpy(newdt->dt_ff, dt->dt_ff, i);
1073 memset((uint8_t *)newdt->dt_ff + i, 0,
1074 numfiles * sizeof(fdfile_t *) - i);
1075
1076 /*
1077 * Link old descriptor array into list to be discarded. We defer
1078 * freeing until the last reference to the descriptor table goes
1079 * away (usually process exit). This allows us to do lockless
1080 * lookups in fd_getfile().
1081 */
1082 if (oldnfiles > NDFILE) {
1083 if (fdp->fd_refcnt > 1) {
1084 newdt->dt_link = dt;
1085 } else {
1086 fd_dtab_free(dt);
1087 }
1088 }
1089
1090 if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1091 i = NDHISLOTS(oldnfiles) * sizeof(uint32_t);
1092 memcpy(newhimap, fdp->fd_himap, i);
1093 memset((uint8_t *)newhimap + i, 0,
1094 NDHISLOTS(numfiles) * sizeof(uint32_t) - i);
1095
1096 i = NDLOSLOTS(oldnfiles) * sizeof(uint32_t);
1097 memcpy(newlomap, fdp->fd_lomap, i);
1098 memset((uint8_t *)newlomap + i, 0,
1099 NDLOSLOTS(numfiles) * sizeof(uint32_t) - i);
1100
1101 if (NDHISLOTS(oldnfiles) > NDHISLOTS(NDFILE)) {
1102 fd_map_free(oldnfiles, fdp->fd_lomap, fdp->fd_himap);
1103 }
1104 fdp->fd_himap = newhimap;
1105 fdp->fd_lomap = newlomap;
1106 }
1107
1108 /*
1109 * All other modifications must become globally visible before
1110 * the change to fd_dt. See fd_getfile().
1111 */
1112 atomic_store_release(&fdp->fd_dt, newdt);
1113 KASSERT(newdt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1114 fd_checkmaps(fdp);
1115 mutex_exit(&fdp->fd_lock);
1116 }
1117
1118 /*
1119 * Create a new open file structure and allocate a file descriptor
1120 * for the current process.
1121 */
1122 int
fd_allocfile(file_t ** resultfp,int * resultfd)1123 fd_allocfile(file_t **resultfp, int *resultfd)
1124 {
1125 proc_t *p = curproc;
1126 kauth_cred_t cred;
1127 file_t *fp;
1128 int error;
1129
1130 while ((error = fd_alloc(p, 0, resultfd)) != 0) {
1131 if (error != ENOSPC) {
1132 return error;
1133 }
1134 fd_tryexpand(p);
1135 }
1136
1137 fp = pool_cache_get(file_cache, PR_WAITOK);
1138 if (fp == NULL) {
1139 fd_abort(p, NULL, *resultfd);
1140 return ENFILE;
1141 }
1142 KASSERT(fp->f_count == 0);
1143 KASSERT(fp->f_msgcount == 0);
1144 KASSERT(fp->f_unpcount == 0);
1145
1146 /* Replace cached credentials if not what we need. */
1147 cred = curlwp->l_cred;
1148 if (__predict_false(cred != fp->f_cred)) {
1149 kauth_cred_free(fp->f_cred);
1150 fp->f_cred = kauth_cred_hold(cred);
1151 }
1152
1153 /*
1154 * Don't allow recycled files to be scanned.
1155 * See uipc_usrreq.c.
1156 */
1157 if (__predict_false((fp->f_flag & FSCAN) != 0)) {
1158 mutex_enter(&fp->f_lock);
1159 atomic_and_uint(&fp->f_flag, ~FSCAN);
1160 mutex_exit(&fp->f_lock);
1161 }
1162
1163 fp->f_advice = 0;
1164 fp->f_offset = 0;
1165 *resultfp = fp;
1166
1167 return 0;
1168 }
1169
1170 /*
1171 * Successful creation of a new descriptor: make visible to the process.
1172 */
1173 void
fd_affix(proc_t * p,file_t * fp,unsigned fd)1174 fd_affix(proc_t *p, file_t *fp, unsigned fd)
1175 {
1176 fdfile_t *ff;
1177 filedesc_t *fdp;
1178 fdtab_t *dt;
1179
1180 KASSERT(p == curproc || p == &proc0);
1181
1182 /* Add a reference to the file structure. */
1183 mutex_enter(&fp->f_lock);
1184 fp->f_count++;
1185 mutex_exit(&fp->f_lock);
1186
1187 /*
1188 * Insert the new file into the descriptor slot.
1189 */
1190 fdp = p->p_fd;
1191 dt = atomic_load_consume(&fdp->fd_dt);
1192 ff = dt->dt_ff[fd];
1193
1194 KASSERT(ff != NULL);
1195 KASSERT(ff->ff_file == NULL);
1196 KASSERT(ff->ff_allocated);
1197 KASSERT(fd_isused(fdp, fd));
1198 KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1199
1200 /* No need to lock in order to make file initially visible. */
1201 atomic_store_release(&ff->ff_file, fp);
1202 }
1203
1204 /*
1205 * Abort creation of a new descriptor: free descriptor slot and file.
1206 */
1207 void
fd_abort(proc_t * p,file_t * fp,unsigned fd)1208 fd_abort(proc_t *p, file_t *fp, unsigned fd)
1209 {
1210 filedesc_t *fdp;
1211 fdfile_t *ff;
1212
1213 KASSERT(p == curproc || p == &proc0);
1214
1215 fdp = p->p_fd;
1216 ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1217 ff->ff_exclose = false;
1218
1219 KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1220
1221 mutex_enter(&fdp->fd_lock);
1222 KASSERT(fd_isused(fdp, fd));
1223 fd_unused(fdp, fd);
1224 mutex_exit(&fdp->fd_lock);
1225
1226 if (fp != NULL) {
1227 KASSERT(fp->f_count == 0);
1228 KASSERT(fp->f_cred != NULL);
1229 pool_cache_put(file_cache, fp);
1230 }
1231 }
1232
1233 static int
file_ctor(void * arg,void * obj,int flags)1234 file_ctor(void *arg, void *obj, int flags)
1235 {
1236 /*
1237 * It's easy to exhaust the open file limit on a system with many
1238 * CPUs due to caching. Allow a bit of leeway to reduce the element
1239 * of surprise.
1240 */
1241 u_int slop = PCG_NOBJECTS_NORMAL * (ncpu - 1);
1242 file_t *fp = obj;
1243
1244 memset(fp, 0, sizeof(*fp));
1245
1246 mutex_enter(&filelist_lock);
1247 if (__predict_false(nfiles >= slop + maxfiles)) {
1248 mutex_exit(&filelist_lock);
1249 tablefull("file", "increase kern.maxfiles or MAXFILES");
1250 return ENFILE;
1251 }
1252 nfiles++;
1253 LIST_INSERT_HEAD(&filehead, fp, f_list);
1254 mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1255 fp->f_cred = kauth_cred_hold(curlwp->l_cred);
1256 mutex_exit(&filelist_lock);
1257
1258 return 0;
1259 }
1260
1261 static void
file_dtor(void * arg,void * obj)1262 file_dtor(void *arg, void *obj)
1263 {
1264 file_t *fp = obj;
1265
1266 mutex_enter(&filelist_lock);
1267 nfiles--;
1268 LIST_REMOVE(fp, f_list);
1269 mutex_exit(&filelist_lock);
1270
1271 KASSERT(fp->f_count == 0);
1272 kauth_cred_free(fp->f_cred);
1273 mutex_destroy(&fp->f_lock);
1274 }
1275
1276 static void
fdfile_ctor(fdfile_t * ff)1277 fdfile_ctor(fdfile_t *ff)
1278 {
1279
1280 memset(ff, 0, sizeof(*ff));
1281 cv_init(&ff->ff_closing, "fdclose");
1282 }
1283
1284 static void
fdfile_dtor(fdfile_t * ff)1285 fdfile_dtor(fdfile_t *ff)
1286 {
1287
1288 cv_destroy(&ff->ff_closing);
1289 }
1290
1291 file_t *
fgetdummy(void)1292 fgetdummy(void)
1293 {
1294 file_t *fp;
1295
1296 fp = kmem_zalloc(sizeof(*fp), KM_SLEEP);
1297 mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1298 return fp;
1299 }
1300
1301 void
fputdummy(file_t * fp)1302 fputdummy(file_t *fp)
1303 {
1304
1305 mutex_destroy(&fp->f_lock);
1306 kmem_free(fp, sizeof(*fp));
1307 }
1308
1309 /*
1310 * Create an initial filedesc structure.
1311 */
1312 filedesc_t *
fd_init(filedesc_t * fdp)1313 fd_init(filedesc_t *fdp)
1314 {
1315 #ifdef DIAGNOSTIC
1316 unsigned fd;
1317 #endif
1318
1319 if (__predict_true(fdp == NULL)) {
1320 fdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1321 } else {
1322 KASSERT(fdp == &filedesc0);
1323 filedesc_ctor(NULL, fdp, PR_WAITOK);
1324 }
1325
1326 #ifdef DIAGNOSTIC
1327 KASSERT(fdp->fd_lastfile == -1);
1328 KASSERT(fdp->fd_lastkqfile == -1);
1329 KASSERT(fdp->fd_knhash == NULL);
1330 KASSERT(fdp->fd_freefile == 0);
1331 KASSERT(fdp->fd_exclose == false);
1332 KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1333 KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1334 for (fd = 0; fd < NDFDFILE; fd++) {
1335 KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] ==
1336 (fdfile_t *)fdp->fd_dfdfile[fd]);
1337 }
1338 for (fd = NDFDFILE; fd < NDFILE; fd++) {
1339 KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] == NULL);
1340 }
1341 KASSERT(fdp->fd_himap == fdp->fd_dhimap);
1342 KASSERT(fdp->fd_lomap == fdp->fd_dlomap);
1343 #endif /* DIAGNOSTIC */
1344
1345 fdp->fd_refcnt = 1;
1346 fd_checkmaps(fdp);
1347
1348 return fdp;
1349 }
1350
1351 /*
1352 * Initialize a file descriptor table.
1353 */
1354 static int
filedesc_ctor(void * arg,void * obj,int flag)1355 filedesc_ctor(void *arg, void *obj, int flag)
1356 {
1357 filedesc_t *fdp = obj;
1358 fdfile_t **ffp;
1359 int i;
1360
1361 memset(fdp, 0, sizeof(*fdp));
1362 mutex_init(&fdp->fd_lock, MUTEX_DEFAULT, IPL_NONE);
1363 fdp->fd_lastfile = -1;
1364 fdp->fd_lastkqfile = -1;
1365 fdp->fd_dt = &fdp->fd_dtbuiltin;
1366 fdp->fd_dtbuiltin.dt_nfiles = NDFILE;
1367 fdp->fd_himap = fdp->fd_dhimap;
1368 fdp->fd_lomap = fdp->fd_dlomap;
1369
1370 CTASSERT(sizeof(fdp->fd_dfdfile[0]) >= sizeof(fdfile_t));
1371 for (i = 0, ffp = fdp->fd_dt->dt_ff; i < NDFDFILE; i++, ffp++) {
1372 fdfile_ctor(*ffp = (fdfile_t *)fdp->fd_dfdfile[i]);
1373 }
1374
1375 return 0;
1376 }
1377
1378 static void
filedesc_dtor(void * arg,void * obj)1379 filedesc_dtor(void *arg, void *obj)
1380 {
1381 filedesc_t *fdp = obj;
1382 int i;
1383
1384 for (i = 0; i < NDFDFILE; i++) {
1385 fdfile_dtor((fdfile_t *)fdp->fd_dfdfile[i]);
1386 }
1387
1388 mutex_destroy(&fdp->fd_lock);
1389 }
1390
1391 /*
1392 * Make p share curproc's filedesc structure.
1393 */
1394 void
fd_share(struct proc * p)1395 fd_share(struct proc *p)
1396 {
1397 filedesc_t *fdp;
1398
1399 fdp = curlwp->l_fd;
1400 p->p_fd = fdp;
1401 atomic_inc_uint(&fdp->fd_refcnt);
1402 }
1403
1404 /*
1405 * Acquire a hold on a filedesc structure.
1406 */
1407 void
fd_hold(lwp_t * l)1408 fd_hold(lwp_t *l)
1409 {
1410 filedesc_t *fdp = l->l_fd;
1411
1412 atomic_inc_uint(&fdp->fd_refcnt);
1413 }
1414
1415 /*
1416 * Copy a filedesc structure.
1417 */
1418 filedesc_t *
fd_copy(void)1419 fd_copy(void)
1420 {
1421 filedesc_t *newfdp, *fdp;
1422 fdfile_t *ff, **ffp, **nffp, *ff2;
1423 int i, j, numfiles, lastfile, newlast;
1424 file_t *fp;
1425 fdtab_t *newdt;
1426
1427 fdp = curproc->p_fd;
1428 newfdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1429 newfdp->fd_refcnt = 1;
1430
1431 #ifdef DIAGNOSTIC
1432 KASSERT(newfdp->fd_lastfile == -1);
1433 KASSERT(newfdp->fd_lastkqfile == -1);
1434 KASSERT(newfdp->fd_knhash == NULL);
1435 KASSERT(newfdp->fd_freefile == 0);
1436 KASSERT(newfdp->fd_exclose == false);
1437 KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1438 KASSERT(newfdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1439 for (i = 0; i < NDFDFILE; i++) {
1440 KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] ==
1441 (fdfile_t *)&newfdp->fd_dfdfile[i]);
1442 }
1443 for (i = NDFDFILE; i < NDFILE; i++) {
1444 KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] == NULL);
1445 }
1446 #endif /* DIAGNOSTIC */
1447
1448 mutex_enter(&fdp->fd_lock);
1449 fd_checkmaps(fdp);
1450 numfiles = fdp->fd_dt->dt_nfiles;
1451 lastfile = fdp->fd_lastfile;
1452
1453 /*
1454 * If the number of open files fits in the internal arrays
1455 * of the open file structure, use them, otherwise allocate
1456 * additional memory for the number of descriptors currently
1457 * in use.
1458 */
1459 if (lastfile < NDFILE) {
1460 i = NDFILE;
1461 newdt = newfdp->fd_dt;
1462 KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1463 } else {
1464 /*
1465 * Compute the smallest multiple of NDEXTENT needed
1466 * for the file descriptors currently in use,
1467 * allowing the table to shrink.
1468 */
1469 i = numfiles;
1470 while (i >= 2 * NDEXTENT && i > lastfile * 2) {
1471 i /= 2;
1472 }
1473 KASSERT(i > NDFILE);
1474 newdt = fd_dtab_alloc(i);
1475 newfdp->fd_dt = newdt;
1476 memcpy(newdt->dt_ff, newfdp->fd_dtbuiltin.dt_ff,
1477 NDFDFILE * sizeof(fdfile_t **));
1478 memset(newdt->dt_ff + NDFDFILE, 0,
1479 (i - NDFDFILE) * sizeof(fdfile_t **));
1480 }
1481 if (NDHISLOTS(i) <= NDHISLOTS(NDFILE)) {
1482 newfdp->fd_himap = newfdp->fd_dhimap;
1483 newfdp->fd_lomap = newfdp->fd_dlomap;
1484 } else {
1485 fd_map_alloc(i, &newfdp->fd_lomap, &newfdp->fd_himap);
1486 KASSERT(i >= NDENTRIES * NDENTRIES);
1487 memset(newfdp->fd_himap, 0, NDHISLOTS(i)*sizeof(uint32_t));
1488 memset(newfdp->fd_lomap, 0, NDLOSLOTS(i)*sizeof(uint32_t));
1489 }
1490 newfdp->fd_freefile = fdp->fd_freefile;
1491 newfdp->fd_exclose = fdp->fd_exclose;
1492
1493 ffp = fdp->fd_dt->dt_ff;
1494 nffp = newdt->dt_ff;
1495 newlast = -1;
1496 for (i = 0; i <= lastfile; i++, ffp++, nffp++) {
1497 KASSERT(i >= NDFDFILE ||
1498 *nffp == (fdfile_t *)newfdp->fd_dfdfile[i]);
1499 ff = *ffp;
1500 if (ff == NULL ||
1501 (fp = atomic_load_consume(&ff->ff_file)) == NULL) {
1502 /* Descriptor unused, or descriptor half open. */
1503 KASSERT(!fd_isused(newfdp, i));
1504 continue;
1505 }
1506 if (__predict_false(fp->f_type == DTYPE_KQUEUE)) {
1507 /* kqueue descriptors cannot be copied. */
1508 if (i < newfdp->fd_freefile) {
1509 newfdp->fd_freefile = i;
1510 }
1511 continue;
1512 }
1513 /* It's active: add a reference to the file. */
1514 mutex_enter(&fp->f_lock);
1515 fp->f_count++;
1516 mutex_exit(&fp->f_lock);
1517
1518 /* Allocate an fdfile_t to represent it. */
1519 if (i >= NDFDFILE) {
1520 ff2 = kmem_alloc(sizeof(*ff2), KM_SLEEP);
1521 fdfile_ctor(ff2);
1522 *nffp = ff2;
1523 } else {
1524 ff2 = newdt->dt_ff[i];
1525 }
1526 ff2->ff_file = fp;
1527 ff2->ff_exclose = ff->ff_exclose;
1528 ff2->ff_allocated = true;
1529
1530 /* Fix up bitmaps. */
1531 j = i >> NDENTRYSHIFT;
1532 KASSERT((newfdp->fd_lomap[j] & (1U << (i & NDENTRYMASK))) == 0);
1533 newfdp->fd_lomap[j] |= 1U << (i & NDENTRYMASK);
1534 if (__predict_false(newfdp->fd_lomap[j] == ~0)) {
1535 KASSERT((newfdp->fd_himap[j >> NDENTRYSHIFT] &
1536 (1U << (j & NDENTRYMASK))) == 0);
1537 newfdp->fd_himap[j >> NDENTRYSHIFT] |=
1538 1U << (j & NDENTRYMASK);
1539 }
1540 newlast = i;
1541 }
1542 KASSERT(newdt->dt_ff[0] == (fdfile_t *)newfdp->fd_dfdfile[0]);
1543 newfdp->fd_lastfile = newlast;
1544 fd_checkmaps(newfdp);
1545 mutex_exit(&fdp->fd_lock);
1546
1547 return newfdp;
1548 }
1549
1550 /*
1551 * Release a filedesc structure.
1552 */
1553 void
fd_free(void)1554 fd_free(void)
1555 {
1556 fdfile_t *ff;
1557 file_t *fp;
1558 int fd, nf;
1559 fdtab_t *dt;
1560 lwp_t * const l = curlwp;
1561 filedesc_t * const fdp = l->l_fd;
1562 const bool noadvlock = (l->l_proc->p_flag & PK_ADVLOCK) == 0;
1563
1564 KASSERT(atomic_load_consume(&fdp->fd_dt)->dt_ff[0] ==
1565 (fdfile_t *)fdp->fd_dfdfile[0]);
1566 KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1567 KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1568
1569 membar_release();
1570 if (atomic_dec_uint_nv(&fdp->fd_refcnt) > 0)
1571 return;
1572 membar_acquire();
1573
1574 /*
1575 * Close any files that the process holds open.
1576 */
1577 dt = fdp->fd_dt;
1578 fd_checkmaps(fdp);
1579 #ifdef DEBUG
1580 fdp->fd_refcnt = -1; /* see fd_checkmaps */
1581 #endif
1582 for (fd = 0, nf = dt->dt_nfiles; fd < nf; fd++) {
1583 ff = dt->dt_ff[fd];
1584 KASSERT(fd >= NDFDFILE ||
1585 ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1586 if (ff == NULL)
1587 continue;
1588 if ((fp = atomic_load_consume(&ff->ff_file)) != NULL) {
1589 /*
1590 * Must use fd_close() here if there is
1591 * a reference from kqueue or we might have posix
1592 * advisory locks.
1593 */
1594 if (__predict_true(ff->ff_refcnt == 0) &&
1595 (noadvlock || fp->f_type != DTYPE_VNODE)) {
1596 ff->ff_file = NULL;
1597 ff->ff_exclose = false;
1598 ff->ff_allocated = false;
1599 closef(fp);
1600 } else {
1601 ff->ff_refcnt++;
1602 fd_close(fd);
1603 }
1604 }
1605 KASSERT(ff->ff_refcnt == 0);
1606 KASSERT(ff->ff_file == NULL);
1607 KASSERT(!ff->ff_exclose);
1608 KASSERT(!ff->ff_allocated);
1609 if (fd >= NDFDFILE) {
1610 cv_destroy(&ff->ff_closing);
1611 kmem_free(ff, sizeof(*ff));
1612 dt->dt_ff[fd] = NULL;
1613 }
1614 }
1615
1616 /*
1617 * Clean out the descriptor table for the next user and return
1618 * to the cache.
1619 */
1620 if (__predict_false(dt != &fdp->fd_dtbuiltin)) {
1621 fd_dtab_free(fdp->fd_dt);
1622 /* Otherwise, done above. */
1623 memset(&fdp->fd_dtbuiltin.dt_ff[NDFDFILE], 0,
1624 (NDFILE - NDFDFILE) * sizeof(fdp->fd_dtbuiltin.dt_ff[0]));
1625 fdp->fd_dt = &fdp->fd_dtbuiltin;
1626 }
1627 if (__predict_false(NDHISLOTS(nf) > NDHISLOTS(NDFILE))) {
1628 KASSERT(fdp->fd_himap != fdp->fd_dhimap);
1629 KASSERT(fdp->fd_lomap != fdp->fd_dlomap);
1630 fd_map_free(nf, fdp->fd_lomap, fdp->fd_himap);
1631 }
1632 if (__predict_false(fdp->fd_knhash != NULL)) {
1633 hashdone(fdp->fd_knhash, HASH_LIST, fdp->fd_knhashmask);
1634 fdp->fd_knhash = NULL;
1635 fdp->fd_knhashmask = 0;
1636 } else {
1637 KASSERT(fdp->fd_knhashmask == 0);
1638 }
1639 fdp->fd_dt = &fdp->fd_dtbuiltin;
1640 fdp->fd_lastkqfile = -1;
1641 fdp->fd_lastfile = -1;
1642 fdp->fd_freefile = 0;
1643 fdp->fd_exclose = false;
1644 memset(&fdp->fd_startzero, 0, sizeof(*fdp) -
1645 offsetof(filedesc_t, fd_startzero));
1646 fdp->fd_himap = fdp->fd_dhimap;
1647 fdp->fd_lomap = fdp->fd_dlomap;
1648 KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1649 KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1650 KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1651 #ifdef DEBUG
1652 fdp->fd_refcnt = 0; /* see fd_checkmaps */
1653 #endif
1654 fd_checkmaps(fdp);
1655 pool_cache_put(filedesc_cache, fdp);
1656 }
1657
1658 /*
1659 * File Descriptor pseudo-device driver (/dev/fd/).
1660 *
1661 * Opening minor device N dup()s the file (if any) connected to file
1662 * descriptor N belonging to the calling process. Note that this driver
1663 * consists of only the ``open()'' routine, because all subsequent
1664 * references to this file will be direct to the other driver.
1665 */
1666 static int
filedescopen(dev_t dev,int mode,int type,lwp_t * l)1667 filedescopen(dev_t dev, int mode, int type, lwp_t *l)
1668 {
1669
1670 /*
1671 * XXX Kludge: set dupfd to contain the value of the
1672 * the file descriptor being sought for duplication. The error
1673 * return ensures that the vnode for this device will be released
1674 * by vn_open. Open will detect this special error and take the
1675 * actions in fd_dupopen below. Other callers of vn_open or VOP_OPEN
1676 * will simply report the error.
1677 */
1678 l->l_dupfd = minor(dev); /* XXX */
1679 return EDUPFD;
1680 }
1681
1682 /*
1683 * Duplicate the specified descriptor to a free descriptor.
1684 *
1685 * old is the original fd.
1686 * moveit is true if we should move rather than duplicate.
1687 * flags are the open flags (converted from O_* to F*).
1688 * newp returns the new fd on success.
1689 *
1690 * These two cases are produced by the EDUPFD and EMOVEFD magic
1691 * errnos, but in the interest of removing that regrettable interface,
1692 * vn_open has been changed to intercept them. Now vn_open returns
1693 * either a vnode or a filehandle, and the filehandle is accompanied
1694 * by a boolean that says whether we should dup (moveit == false) or
1695 * move (moveit == true) the fd.
1696 *
1697 * The dup case is used by /dev/stderr, /proc/self/fd, and such. The
1698 * move case is used by cloner devices that allocate a fd of their
1699 * own (a layering violation that should go away eventually) that
1700 * then needs to be put in the place open() expects it.
1701 */
1702 int
fd_dupopen(int old,bool moveit,int flags,int * newp)1703 fd_dupopen(int old, bool moveit, int flags, int *newp)
1704 {
1705 filedesc_t *fdp;
1706 fdfile_t *ff;
1707 file_t *fp;
1708 fdtab_t *dt;
1709 int error;
1710
1711 if ((fp = fd_getfile(old)) == NULL) {
1712 return EBADF;
1713 }
1714 fdp = curlwp->l_fd;
1715 dt = atomic_load_consume(&fdp->fd_dt);
1716 ff = dt->dt_ff[old];
1717
1718 /*
1719 * There are two cases of interest here.
1720 *
1721 * 1. moveit == false (used to be the EDUPFD magic errno):
1722 * simply dup (old) to file descriptor (new) and return.
1723 *
1724 * 2. moveit == true (used to be the EMOVEFD magic errno):
1725 * steal away the file structure from (old) and store it in
1726 * (new). (old) is effectively closed by this operation.
1727 */
1728 if (moveit == false) {
1729 /*
1730 * Check that the mode the file is being opened for is a
1731 * subset of the mode of the existing descriptor.
1732 */
1733 if (((flags & (FREAD|FWRITE)) | fp->f_flag) != fp->f_flag) {
1734 error = EACCES;
1735 goto out;
1736 }
1737
1738 /* Copy it. */
1739 error = fd_dup(fp, 0, newp, ff->ff_exclose);
1740 } else {
1741 /* Copy it. */
1742 error = fd_dup(fp, 0, newp, ff->ff_exclose);
1743 if (error != 0) {
1744 goto out;
1745 }
1746
1747 /* Steal away the file pointer from 'old'. */
1748 (void)fd_close(old);
1749 return 0;
1750 }
1751
1752 out:
1753 fd_putfile(old);
1754 return error;
1755 }
1756
1757 /*
1758 * Close open files on exec.
1759 */
1760 void
fd_closeexec(void)1761 fd_closeexec(void)
1762 {
1763 proc_t *p;
1764 filedesc_t *fdp;
1765 fdfile_t *ff;
1766 lwp_t *l;
1767 fdtab_t *dt;
1768 int fd;
1769
1770 l = curlwp;
1771 p = l->l_proc;
1772 fdp = p->p_fd;
1773
1774 if (fdp->fd_refcnt > 1) {
1775 fdp = fd_copy();
1776 fd_free();
1777 p->p_fd = fdp;
1778 l->l_fd = fdp;
1779 }
1780 if (!fdp->fd_exclose) {
1781 return;
1782 }
1783 fdp->fd_exclose = false;
1784 dt = atomic_load_consume(&fdp->fd_dt);
1785
1786 for (fd = 0; fd <= fdp->fd_lastfile; fd++) {
1787 if ((ff = dt->dt_ff[fd]) == NULL) {
1788 KASSERT(fd >= NDFDFILE);
1789 continue;
1790 }
1791 KASSERT(fd >= NDFDFILE ||
1792 ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1793 if (ff->ff_file == NULL)
1794 continue;
1795 if (ff->ff_exclose) {
1796 /*
1797 * We need a reference to close the file.
1798 * No other threads can see the fdfile_t at
1799 * this point, so don't bother locking.
1800 */
1801 KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
1802 ff->ff_refcnt++;
1803 fd_close(fd);
1804 }
1805 }
1806 }
1807
1808 /*
1809 * Sets descriptor owner. If the owner is a process, 'pgid'
1810 * is set to positive value, process ID. If the owner is process group,
1811 * 'pgid' is set to -pg_id.
1812 */
1813 int
fsetown(pid_t * pgid,u_long cmd,const void * data)1814 fsetown(pid_t *pgid, u_long cmd, const void *data)
1815 {
1816 pid_t id = *(const pid_t *)data;
1817 int error;
1818
1819 if (id <= INT_MIN)
1820 return EINVAL;
1821
1822 switch (cmd) {
1823 case TIOCSPGRP:
1824 if (id < 0)
1825 return EINVAL;
1826 id = -id;
1827 break;
1828 default:
1829 break;
1830 }
1831 if (id > 0) {
1832 mutex_enter(&proc_lock);
1833 error = proc_find(id) ? 0 : ESRCH;
1834 mutex_exit(&proc_lock);
1835 } else if (id < 0) {
1836 error = pgid_in_session(curproc, -id);
1837 } else {
1838 error = 0;
1839 }
1840 if (!error) {
1841 *pgid = id;
1842 }
1843 return error;
1844 }
1845
1846 void
fd_set_exclose(struct lwp * l,int fd,bool exclose)1847 fd_set_exclose(struct lwp *l, int fd, bool exclose)
1848 {
1849 filedesc_t *fdp = l->l_fd;
1850 fdfile_t *ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1851
1852 ff->ff_exclose = exclose;
1853 if (exclose)
1854 fdp->fd_exclose = true;
1855 }
1856
1857 /*
1858 * Return descriptor owner information. If the value is positive,
1859 * it's process ID. If it's negative, it's process group ID and
1860 * needs the sign removed before use.
1861 */
1862 int
fgetown(pid_t pgid,u_long cmd,void * data)1863 fgetown(pid_t pgid, u_long cmd, void *data)
1864 {
1865
1866 switch (cmd) {
1867 case TIOCGPGRP:
1868 KASSERT(pgid > INT_MIN);
1869 *(int *)data = -pgid;
1870 break;
1871 default:
1872 *(int *)data = pgid;
1873 break;
1874 }
1875 return 0;
1876 }
1877
1878 /*
1879 * Send signal to descriptor owner, either process or process group.
1880 */
1881 void
fownsignal(pid_t pgid,int signo,int code,int band,void * fdescdata)1882 fownsignal(pid_t pgid, int signo, int code, int band, void *fdescdata)
1883 {
1884 ksiginfo_t ksi;
1885
1886 KASSERT(!cpu_intr_p());
1887
1888 if (pgid == 0) {
1889 return;
1890 }
1891
1892 KSI_INIT(&ksi);
1893 ksi.ksi_signo = signo;
1894 ksi.ksi_code = code;
1895 ksi.ksi_band = band;
1896
1897 mutex_enter(&proc_lock);
1898 if (pgid > 0) {
1899 struct proc *p1;
1900
1901 p1 = proc_find(pgid);
1902 if (p1 != NULL) {
1903 kpsignal(p1, &ksi, fdescdata);
1904 }
1905 } else {
1906 struct pgrp *pgrp;
1907
1908 KASSERT(pgid < 0 && pgid > INT_MIN);
1909 pgrp = pgrp_find(-pgid);
1910 if (pgrp != NULL) {
1911 kpgsignal(pgrp, &ksi, fdescdata, 0);
1912 }
1913 }
1914 mutex_exit(&proc_lock);
1915 }
1916
1917 int
fd_clone(file_t * fp,unsigned fd,int flag,const struct fileops * fops,void * data)1918 fd_clone(file_t *fp, unsigned fd, int flag, const struct fileops *fops,
1919 void *data)
1920 {
1921
1922 fp->f_flag = flag & FMASK;
1923 fd_set_exclose(curlwp, fd, (flag & O_CLOEXEC) != 0);
1924 fp->f_type = DTYPE_MISC;
1925 fp->f_ops = fops;
1926 fp->f_data = data;
1927 curlwp->l_dupfd = fd;
1928 fd_affix(curproc, fp, fd);
1929
1930 return EMOVEFD;
1931 }
1932
1933 int
fnullop_fcntl(file_t * fp,u_int cmd,void * data)1934 fnullop_fcntl(file_t *fp, u_int cmd, void *data)
1935 {
1936
1937 if (cmd == F_SETFL)
1938 return 0;
1939
1940 return EOPNOTSUPP;
1941 }
1942
1943 int
fnullop_poll(file_t * fp,int which)1944 fnullop_poll(file_t *fp, int which)
1945 {
1946
1947 return 0;
1948 }
1949
1950 int
fnullop_kqfilter(file_t * fp,struct knote * kn)1951 fnullop_kqfilter(file_t *fp, struct knote *kn)
1952 {
1953
1954 return EOPNOTSUPP;
1955 }
1956
1957 void
fnullop_restart(file_t * fp)1958 fnullop_restart(file_t *fp)
1959 {
1960
1961 }
1962
1963 int
fbadop_read(file_t * fp,off_t * offset,struct uio * uio,kauth_cred_t cred,int flags)1964 fbadop_read(file_t *fp, off_t *offset, struct uio *uio,
1965 kauth_cred_t cred, int flags)
1966 {
1967
1968 return EOPNOTSUPP;
1969 }
1970
1971 int
fbadop_write(file_t * fp,off_t * offset,struct uio * uio,kauth_cred_t cred,int flags)1972 fbadop_write(file_t *fp, off_t *offset, struct uio *uio,
1973 kauth_cred_t cred, int flags)
1974 {
1975
1976 return EOPNOTSUPP;
1977 }
1978
1979 int
fbadop_ioctl(file_t * fp,u_long com,void * data)1980 fbadop_ioctl(file_t *fp, u_long com, void *data)
1981 {
1982
1983 return EOPNOTSUPP;
1984 }
1985
1986 int
fbadop_stat(file_t * fp,struct stat * sb)1987 fbadop_stat(file_t *fp, struct stat *sb)
1988 {
1989
1990 return EOPNOTSUPP;
1991 }
1992
1993 int
fbadop_close(file_t * fp)1994 fbadop_close(file_t *fp)
1995 {
1996
1997 return EOPNOTSUPP;
1998 }
1999
2000 /*
2001 * sysctl routines pertaining to file descriptors
2002 */
2003
2004 /* Initialized in sysctl_init() for now... */
2005 extern kmutex_t sysctl_file_marker_lock;
2006 static u_int sysctl_file_marker = 1;
2007
2008 /*
2009 * Expects to be called with proc_lock and sysctl_file_marker_lock locked.
2010 */
2011 static void
sysctl_file_marker_reset(void)2012 sysctl_file_marker_reset(void)
2013 {
2014 struct proc *p;
2015
2016 PROCLIST_FOREACH(p, &allproc) {
2017 struct filedesc *fd = p->p_fd;
2018 fdtab_t *dt;
2019 u_int i;
2020
2021 mutex_enter(&fd->fd_lock);
2022 dt = fd->fd_dt;
2023 for (i = 0; i < dt->dt_nfiles; i++) {
2024 struct file *fp;
2025 fdfile_t *ff;
2026
2027 if ((ff = dt->dt_ff[i]) == NULL) {
2028 continue;
2029 }
2030 if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
2031 continue;
2032 }
2033 fp->f_marker = 0;
2034 }
2035 mutex_exit(&fd->fd_lock);
2036 }
2037 }
2038
2039 /*
2040 * sysctl helper routine for kern.file pseudo-subtree.
2041 */
2042 static int
sysctl_kern_file(SYSCTLFN_ARGS)2043 sysctl_kern_file(SYSCTLFN_ARGS)
2044 {
2045 const bool allowaddr = get_expose_address(curproc);
2046 struct filelist flist;
2047 int error;
2048 size_t buflen;
2049 struct file *fp, fbuf;
2050 char *start, *where;
2051 struct proc *p;
2052
2053 start = where = oldp;
2054 buflen = *oldlenp;
2055
2056 if (where == NULL) {
2057 /*
2058 * overestimate by 10 files
2059 */
2060 *oldlenp = sizeof(filehead) + (nfiles + 10) *
2061 sizeof(struct file);
2062 return 0;
2063 }
2064
2065 /*
2066 * first sysctl_copyout filehead
2067 */
2068 if (buflen < sizeof(filehead)) {
2069 *oldlenp = 0;
2070 return 0;
2071 }
2072 sysctl_unlock();
2073 if (allowaddr) {
2074 memcpy(&flist, &filehead, sizeof(flist));
2075 } else {
2076 memset(&flist, 0, sizeof(flist));
2077 }
2078 error = sysctl_copyout(l, &flist, where, sizeof(flist));
2079 if (error) {
2080 sysctl_relock();
2081 return error;
2082 }
2083 buflen -= sizeof(flist);
2084 where += sizeof(flist);
2085
2086 /*
2087 * followed by an array of file structures
2088 */
2089 mutex_enter(&sysctl_file_marker_lock);
2090 mutex_enter(&proc_lock);
2091 PROCLIST_FOREACH(p, &allproc) {
2092 struct filedesc *fd;
2093 fdtab_t *dt;
2094 u_int i;
2095
2096 if (p->p_stat == SIDL) {
2097 /* skip embryonic processes */
2098 continue;
2099 }
2100 mutex_enter(p->p_lock);
2101 error = kauth_authorize_process(l->l_cred,
2102 KAUTH_PROCESS_CANSEE, p,
2103 KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2104 NULL, NULL);
2105 mutex_exit(p->p_lock);
2106 if (error != 0) {
2107 /*
2108 * Don't leak kauth retval if we're silently
2109 * skipping this entry.
2110 */
2111 error = 0;
2112 continue;
2113 }
2114
2115 /*
2116 * Grab a hold on the process.
2117 */
2118 if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2119 continue;
2120 }
2121 mutex_exit(&proc_lock);
2122
2123 fd = p->p_fd;
2124 mutex_enter(&fd->fd_lock);
2125 dt = fd->fd_dt;
2126 for (i = 0; i < dt->dt_nfiles; i++) {
2127 fdfile_t *ff;
2128
2129 if ((ff = dt->dt_ff[i]) == NULL) {
2130 continue;
2131 }
2132 if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
2133 continue;
2134 }
2135
2136 mutex_enter(&fp->f_lock);
2137
2138 if ((fp->f_count == 0) ||
2139 (fp->f_marker == sysctl_file_marker)) {
2140 mutex_exit(&fp->f_lock);
2141 continue;
2142 }
2143
2144 /* Check that we have enough space. */
2145 if (buflen < sizeof(struct file)) {
2146 *oldlenp = where - start;
2147 mutex_exit(&fp->f_lock);
2148 error = ENOMEM;
2149 break;
2150 }
2151
2152 fill_file(&fbuf, fp);
2153 mutex_exit(&fp->f_lock);
2154 error = sysctl_copyout(l, &fbuf, where, sizeof(fbuf));
2155 if (error) {
2156 break;
2157 }
2158 buflen -= sizeof(struct file);
2159 where += sizeof(struct file);
2160
2161 fp->f_marker = sysctl_file_marker;
2162 }
2163 mutex_exit(&fd->fd_lock);
2164
2165 /*
2166 * Release reference to process.
2167 */
2168 mutex_enter(&proc_lock);
2169 rw_exit(&p->p_reflock);
2170
2171 if (error)
2172 break;
2173 }
2174
2175 sysctl_file_marker++;
2176 /* Reset all markers if wrapped. */
2177 if (sysctl_file_marker == 0) {
2178 sysctl_file_marker_reset();
2179 sysctl_file_marker++;
2180 }
2181
2182 mutex_exit(&proc_lock);
2183 mutex_exit(&sysctl_file_marker_lock);
2184
2185 *oldlenp = where - start;
2186 sysctl_relock();
2187 return error;
2188 }
2189
2190 /*
2191 * sysctl helper function for kern.file2
2192 */
2193 static int
sysctl_kern_file2(SYSCTLFN_ARGS)2194 sysctl_kern_file2(SYSCTLFN_ARGS)
2195 {
2196 struct proc *p;
2197 struct file *fp;
2198 struct filedesc *fd;
2199 struct kinfo_file kf;
2200 char *dp;
2201 u_int i, op;
2202 size_t len, needed, elem_size, out_size;
2203 int error, arg, elem_count;
2204 fdfile_t *ff;
2205 fdtab_t *dt;
2206
2207 if (namelen == 1 && name[0] == CTL_QUERY)
2208 return sysctl_query(SYSCTLFN_CALL(rnode));
2209
2210 if (namelen != 4)
2211 return EINVAL;
2212
2213 error = 0;
2214 dp = oldp;
2215 len = (oldp != NULL) ? *oldlenp : 0;
2216 op = name[0];
2217 arg = name[1];
2218 elem_size = name[2];
2219 elem_count = name[3];
2220 out_size = MIN(sizeof(kf), elem_size);
2221 needed = 0;
2222
2223 if (elem_size < 1 || elem_count < 0)
2224 return EINVAL;
2225
2226 switch (op) {
2227 case KERN_FILE_BYFILE:
2228 case KERN_FILE_BYPID:
2229 /*
2230 * We're traversing the process list in both cases; the BYFILE
2231 * case does additional work of keeping track of files already
2232 * looked at.
2233 */
2234
2235 /* doesn't use arg so it must be zero */
2236 if ((op == KERN_FILE_BYFILE) && (arg != 0))
2237 return EINVAL;
2238
2239 if ((op == KERN_FILE_BYPID) && (arg < -1))
2240 /* -1 means all processes */
2241 return EINVAL;
2242
2243 sysctl_unlock();
2244 if (op == KERN_FILE_BYFILE)
2245 mutex_enter(&sysctl_file_marker_lock);
2246 mutex_enter(&proc_lock);
2247 PROCLIST_FOREACH(p, &allproc) {
2248 if (p->p_stat == SIDL) {
2249 /* skip embryonic processes */
2250 continue;
2251 }
2252 if (arg > 0 && p->p_pid != arg) {
2253 /* pick only the one we want */
2254 /* XXX want 0 to mean "kernel files" */
2255 continue;
2256 }
2257 mutex_enter(p->p_lock);
2258 error = kauth_authorize_process(l->l_cred,
2259 KAUTH_PROCESS_CANSEE, p,
2260 KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2261 NULL, NULL);
2262 mutex_exit(p->p_lock);
2263 if (error != 0) {
2264 /*
2265 * Don't leak kauth retval if we're silently
2266 * skipping this entry.
2267 */
2268 error = 0;
2269 continue;
2270 }
2271
2272 /*
2273 * Grab a hold on the process.
2274 */
2275 if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2276 continue;
2277 }
2278 mutex_exit(&proc_lock);
2279
2280 fd = p->p_fd;
2281 mutex_enter(&fd->fd_lock);
2282 dt = fd->fd_dt;
2283 for (i = 0; i < dt->dt_nfiles; i++) {
2284 if ((ff = dt->dt_ff[i]) == NULL) {
2285 continue;
2286 }
2287 if ((fp = atomic_load_consume(&ff->ff_file)) ==
2288 NULL) {
2289 continue;
2290 }
2291
2292 if ((op == KERN_FILE_BYFILE) &&
2293 (fp->f_marker == sysctl_file_marker)) {
2294 continue;
2295 }
2296 if (len >= elem_size && elem_count > 0) {
2297 mutex_enter(&fp->f_lock);
2298 fill_file2(&kf, fp, ff, i, p->p_pid);
2299 mutex_exit(&fp->f_lock);
2300 mutex_exit(&fd->fd_lock);
2301 error = sysctl_copyout(l,
2302 &kf, dp, out_size);
2303 mutex_enter(&fd->fd_lock);
2304 if (error)
2305 break;
2306 dp += elem_size;
2307 len -= elem_size;
2308 }
2309 if (op == KERN_FILE_BYFILE)
2310 fp->f_marker = sysctl_file_marker;
2311 needed += elem_size;
2312 if (elem_count > 0 && elem_count != INT_MAX)
2313 elem_count--;
2314 }
2315 mutex_exit(&fd->fd_lock);
2316
2317 /*
2318 * Release reference to process.
2319 */
2320 mutex_enter(&proc_lock);
2321 rw_exit(&p->p_reflock);
2322 }
2323 if (op == KERN_FILE_BYFILE) {
2324 sysctl_file_marker++;
2325
2326 /* Reset all markers if wrapped. */
2327 if (sysctl_file_marker == 0) {
2328 sysctl_file_marker_reset();
2329 sysctl_file_marker++;
2330 }
2331 }
2332 mutex_exit(&proc_lock);
2333 if (op == KERN_FILE_BYFILE)
2334 mutex_exit(&sysctl_file_marker_lock);
2335 sysctl_relock();
2336 break;
2337 default:
2338 return EINVAL;
2339 }
2340
2341 if (oldp == NULL)
2342 needed += KERN_FILESLOP * elem_size;
2343 *oldlenp = needed;
2344
2345 return error;
2346 }
2347
2348 static void
fill_file(struct file * fp,const struct file * fpsrc)2349 fill_file(struct file *fp, const struct file *fpsrc)
2350 {
2351 const bool allowaddr = get_expose_address(curproc);
2352
2353 memset(fp, 0, sizeof(*fp));
2354
2355 fp->f_offset = fpsrc->f_offset;
2356 COND_SET_PTR(fp->f_cred, fpsrc->f_cred, allowaddr);
2357 COND_SET_CPTR(fp->f_ops, fpsrc->f_ops, allowaddr);
2358 COND_SET_STRUCT(fp->f_undata, fpsrc->f_undata, allowaddr);
2359 COND_SET_STRUCT(fp->f_list, fpsrc->f_list, allowaddr);
2360 fp->f_flag = fpsrc->f_flag;
2361 fp->f_marker = fpsrc->f_marker;
2362 fp->f_type = fpsrc->f_type;
2363 fp->f_advice = fpsrc->f_advice;
2364 fp->f_count = fpsrc->f_count;
2365 fp->f_msgcount = fpsrc->f_msgcount;
2366 fp->f_unpcount = fpsrc->f_unpcount;
2367 COND_SET_STRUCT(fp->f_unplist, fpsrc->f_unplist, allowaddr);
2368 }
2369
2370 static void
fill_file2(struct kinfo_file * kp,const file_t * fp,const fdfile_t * ff,int i,pid_t pid)2371 fill_file2(struct kinfo_file *kp, const file_t *fp, const fdfile_t *ff,
2372 int i, pid_t pid)
2373 {
2374 const bool allowaddr = get_expose_address(curproc);
2375
2376 memset(kp, 0, sizeof(*kp));
2377
2378 COND_SET_VALUE(kp->ki_fileaddr, PTRTOUINT64(fp), allowaddr);
2379 kp->ki_flag = fp->f_flag;
2380 kp->ki_iflags = 0;
2381 kp->ki_ftype = fp->f_type;
2382 kp->ki_count = fp->f_count;
2383 kp->ki_msgcount = fp->f_msgcount;
2384 COND_SET_VALUE(kp->ki_fucred, PTRTOUINT64(fp->f_cred), allowaddr);
2385 kp->ki_fuid = kauth_cred_geteuid(fp->f_cred);
2386 kp->ki_fgid = kauth_cred_getegid(fp->f_cred);
2387 COND_SET_VALUE(kp->ki_fops, PTRTOUINT64(fp->f_ops), allowaddr);
2388 kp->ki_foffset = fp->f_offset;
2389 COND_SET_VALUE(kp->ki_fdata, PTRTOUINT64(fp->f_data), allowaddr);
2390
2391 /* vnode information to glue this file to something */
2392 if (fp->f_type == DTYPE_VNODE) {
2393 struct vnode *vp = fp->f_vnode;
2394
2395 COND_SET_VALUE(kp->ki_vun, PTRTOUINT64(vp->v_un.vu_socket),
2396 allowaddr);
2397 kp->ki_vsize = vp->v_size;
2398 kp->ki_vtype = vp->v_type;
2399 kp->ki_vtag = vp->v_tag;
2400 COND_SET_VALUE(kp->ki_vdata, PTRTOUINT64(vp->v_data),
2401 allowaddr);
2402 }
2403
2404 /* process information when retrieved via KERN_FILE_BYPID */
2405 if (ff != NULL) {
2406 kp->ki_pid = pid;
2407 kp->ki_fd = i;
2408 kp->ki_ofileflags = ff->ff_exclose;
2409 kp->ki_usecount = ff->ff_refcnt;
2410 }
2411 }
2412