xref: /freebsd/sys/kern/vfs_vnops.c (revision bc7512cc58af2e8bbe5bbf5ca0059b1daa1da897)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13  * Copyright (c) 2013, 2014 The FreeBSD Foundation
14  *
15  * Portions of this software were developed by Konstantin Belousov
16  * under sponsorship from the FreeBSD Foundation.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_hwpmc_hooks.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/disk.h>
53 #include <sys/fail.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/kdb.h>
57 #include <sys/ktr.h>
58 #include <sys/stat.h>
59 #include <sys/priv.h>
60 #include <sys/proc.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/mman.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/bio.h>
69 #include <sys/buf.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/prng.h>
74 #include <sys/sx.h>
75 #include <sys/sleepqueue.h>
76 #include <sys/sysctl.h>
77 #include <sys/ttycom.h>
78 #include <sys/conf.h>
79 #include <sys/syslog.h>
80 #include <sys/unistd.h>
81 #include <sys/user.h>
82 #include <sys/ktrace.h>
83 
84 #include <security/audit/audit.h>
85 #include <security/mac/mac_framework.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_extern.h>
89 #include <vm/pmap.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_pager.h>
94 
95 #ifdef HWPMC_HOOKS
96 #include <sys/pmckern.h>
97 #endif
98 
99 static fo_rdwr_t	vn_read;
100 static fo_rdwr_t	vn_write;
101 static fo_rdwr_t	vn_io_fault;
102 static fo_truncate_t	vn_truncate;
103 static fo_ioctl_t	vn_ioctl;
104 static fo_poll_t	vn_poll;
105 static fo_kqfilter_t	vn_kqfilter;
106 static fo_close_t	vn_closefile;
107 static fo_mmap_t	vn_mmap;
108 static fo_fallocate_t	vn_fallocate;
109 static fo_fspacectl_t	vn_fspacectl;
110 
111 struct 	fileops vnops = {
112 	.fo_read = vn_io_fault,
113 	.fo_write = vn_io_fault,
114 	.fo_truncate = vn_truncate,
115 	.fo_ioctl = vn_ioctl,
116 	.fo_poll = vn_poll,
117 	.fo_kqfilter = vn_kqfilter,
118 	.fo_stat = vn_statfile,
119 	.fo_close = vn_closefile,
120 	.fo_chmod = vn_chmod,
121 	.fo_chown = vn_chown,
122 	.fo_sendfile = vn_sendfile,
123 	.fo_seek = vn_seek,
124 	.fo_fill_kinfo = vn_fill_kinfo,
125 	.fo_mmap = vn_mmap,
126 	.fo_fallocate = vn_fallocate,
127 	.fo_fspacectl = vn_fspacectl,
128 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
129 };
130 
131 const u_int io_hold_cnt = 16;
132 static int vn_io_fault_enable = 1;
133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
134     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
135 static int vn_io_fault_prefault = 0;
136 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
137     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
138 static int vn_io_pgcache_read_enable = 1;
139 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
140     &vn_io_pgcache_read_enable, 0,
141     "Enable copying from page cache for reads, avoiding fs");
142 static u_long vn_io_faults_cnt;
143 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
144     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
145 
146 static int vfs_allow_read_dir = 0;
147 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
148     &vfs_allow_read_dir, 0,
149     "Enable read(2) of directory by root for filesystems that support it");
150 
151 /*
152  * Returns true if vn_io_fault mode of handling the i/o request should
153  * be used.
154  */
155 static bool
156 do_vn_io_fault(struct vnode *vp, struct uio *uio)
157 {
158 	struct mount *mp;
159 
160 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
161 	    (mp = vp->v_mount) != NULL &&
162 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
163 }
164 
165 /*
166  * Structure used to pass arguments to vn_io_fault1(), to do either
167  * file- or vnode-based I/O calls.
168  */
169 struct vn_io_fault_args {
170 	enum {
171 		VN_IO_FAULT_FOP,
172 		VN_IO_FAULT_VOP
173 	} kind;
174 	struct ucred *cred;
175 	int flags;
176 	union {
177 		struct fop_args_tag {
178 			struct file *fp;
179 			fo_rdwr_t *doio;
180 		} fop_args;
181 		struct vop_args_tag {
182 			struct vnode *vp;
183 		} vop_args;
184 	} args;
185 };
186 
187 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
188     struct vn_io_fault_args *args, struct thread *td);
189 
190 int
191 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
192 {
193 	struct thread *td = curthread;
194 
195 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
196 }
197 
198 static uint64_t
199 open2nameif(int fmode, u_int vn_open_flags)
200 {
201 	uint64_t res;
202 
203 	res = ISOPEN | LOCKLEAF;
204 	if ((fmode & O_RESOLVE_BENEATH) != 0)
205 		res |= RBENEATH;
206 	if ((fmode & O_EMPTY_PATH) != 0)
207 		res |= EMPTYPATH;
208 	if ((fmode & FREAD) != 0)
209 		res |= OPENREAD;
210 	if ((fmode & FWRITE) != 0)
211 		res |= OPENWRITE;
212 	if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
213 		res |= AUDITVNODE1;
214 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
215 		res |= NOCAPCHECK;
216 	if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0)
217 		res |= WANTIOCTLCAPS;
218 	return (res);
219 }
220 
221 /*
222  * Common code for vnode open operations via a name lookup.
223  * Lookup the vnode and invoke VOP_CREATE if needed.
224  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
225  *
226  * Note that this does NOT free nameidata for the successful case,
227  * due to the NDINIT being done elsewhere.
228  */
229 int
230 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
231     struct ucred *cred, struct file *fp)
232 {
233 	struct vnode *vp;
234 	struct mount *mp;
235 	struct vattr vat;
236 	struct vattr *vap = &vat;
237 	int fmode, error;
238 	bool first_open;
239 
240 restart:
241 	first_open = false;
242 	fmode = *flagp;
243 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
244 	    O_EXCL | O_DIRECTORY) ||
245 	    (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
246 		return (EINVAL);
247 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
248 		ndp->ni_cnd.cn_nameiop = CREATE;
249 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
250 		/*
251 		 * Set NOCACHE to avoid flushing the cache when
252 		 * rolling in many files at once.
253 		 *
254 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
255 		 * exist despite NOCACHE.
256 		 */
257 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
258 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
259 			ndp->ni_cnd.cn_flags |= FOLLOW;
260 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
261 			bwillwrite();
262 		if ((error = namei(ndp)) != 0)
263 			return (error);
264 		if (ndp->ni_vp == NULL) {
265 			VATTR_NULL(vap);
266 			vap->va_type = VREG;
267 			vap->va_mode = cmode;
268 			if (fmode & O_EXCL)
269 				vap->va_vaflags |= VA_EXCLUSIVE;
270 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
271 				NDFREE_PNBUF(ndp);
272 				vput(ndp->ni_dvp);
273 				if ((error = vn_start_write(NULL, &mp,
274 				    V_XSLEEP | PCATCH)) != 0)
275 					return (error);
276 				NDREINIT(ndp);
277 				goto restart;
278 			}
279 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
280 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
281 #ifdef MAC
282 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
283 			    &ndp->ni_cnd, vap);
284 			if (error == 0)
285 #endif
286 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
287 				    &ndp->ni_cnd, vap);
288 			vp = ndp->ni_vp;
289 			if (error == 0 && (fmode & O_EXCL) != 0 &&
290 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
291 				VI_LOCK(vp);
292 				vp->v_iflag |= VI_FOPENING;
293 				VI_UNLOCK(vp);
294 				first_open = true;
295 			}
296 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
297 			    false);
298 			vn_finished_write(mp);
299 			if (error) {
300 				NDFREE_PNBUF(ndp);
301 				if (error == ERELOOKUP) {
302 					NDREINIT(ndp);
303 					goto restart;
304 				}
305 				return (error);
306 			}
307 			fmode &= ~O_TRUNC;
308 		} else {
309 			if (ndp->ni_dvp == ndp->ni_vp)
310 				vrele(ndp->ni_dvp);
311 			else
312 				vput(ndp->ni_dvp);
313 			ndp->ni_dvp = NULL;
314 			vp = ndp->ni_vp;
315 			if (fmode & O_EXCL) {
316 				error = EEXIST;
317 				goto bad;
318 			}
319 			if (vp->v_type == VDIR) {
320 				error = EISDIR;
321 				goto bad;
322 			}
323 			fmode &= ~O_CREAT;
324 		}
325 	} else {
326 		ndp->ni_cnd.cn_nameiop = LOOKUP;
327 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
328 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
329 		    FOLLOW;
330 		if ((fmode & FWRITE) == 0)
331 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
332 		if ((error = namei(ndp)) != 0)
333 			return (error);
334 		vp = ndp->ni_vp;
335 	}
336 	error = vn_open_vnode(vp, fmode, cred, curthread, fp);
337 	if (first_open) {
338 		VI_LOCK(vp);
339 		vp->v_iflag &= ~VI_FOPENING;
340 		wakeup(vp);
341 		VI_UNLOCK(vp);
342 	}
343 	if (error)
344 		goto bad;
345 	*flagp = fmode;
346 	return (0);
347 bad:
348 	NDFREE_PNBUF(ndp);
349 	vput(vp);
350 	*flagp = fmode;
351 	ndp->ni_vp = NULL;
352 	return (error);
353 }
354 
355 static int
356 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
357 {
358 	struct flock lf;
359 	int error, lock_flags, type;
360 
361 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
362 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
363 		return (0);
364 	KASSERT(fp != NULL, ("open with flock requires fp"));
365 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
366 		return (EOPNOTSUPP);
367 
368 	lock_flags = VOP_ISLOCKED(vp);
369 	VOP_UNLOCK(vp);
370 
371 	lf.l_whence = SEEK_SET;
372 	lf.l_start = 0;
373 	lf.l_len = 0;
374 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
375 	type = F_FLOCK;
376 	if ((fmode & FNONBLOCK) == 0)
377 		type |= F_WAIT;
378 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
379 		type |= F_FIRSTOPEN;
380 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
381 	if (error == 0)
382 		fp->f_flag |= FHASLOCK;
383 
384 	vn_lock(vp, lock_flags | LK_RETRY);
385 	return (error);
386 }
387 
388 /*
389  * Common code for vnode open operations once a vnode is located.
390  * Check permissions, and call the VOP_OPEN routine.
391  */
392 int
393 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
394     struct thread *td, struct file *fp)
395 {
396 	accmode_t accmode;
397 	int error;
398 
399 	if (vp->v_type == VLNK) {
400 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
401 			return (EMLINK);
402 	}
403 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
404 		return (ENOTDIR);
405 
406 	accmode = 0;
407 	if ((fmode & O_PATH) == 0) {
408 		if (vp->v_type == VSOCK)
409 			return (EOPNOTSUPP);
410 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
411 			if (vp->v_type == VDIR)
412 				return (EISDIR);
413 			accmode |= VWRITE;
414 		}
415 		if ((fmode & FREAD) != 0)
416 			accmode |= VREAD;
417 		if ((fmode & O_APPEND) && (fmode & FWRITE))
418 			accmode |= VAPPEND;
419 #ifdef MAC
420 		if ((fmode & O_CREAT) != 0)
421 			accmode |= VCREAT;
422 #endif
423 	}
424 	if ((fmode & FEXEC) != 0)
425 		accmode |= VEXEC;
426 #ifdef MAC
427 	if ((fmode & O_VERIFY) != 0)
428 		accmode |= VVERIFY;
429 	error = mac_vnode_check_open(cred, vp, accmode);
430 	if (error != 0)
431 		return (error);
432 
433 	accmode &= ~(VCREAT | VVERIFY);
434 #endif
435 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
436 		error = VOP_ACCESS(vp, accmode, cred, td);
437 		if (error != 0)
438 			return (error);
439 	}
440 	if ((fmode & O_PATH) != 0) {
441 		if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
442 		    VOP_ACCESS(vp, VREAD, cred, td) == 0)
443 			fp->f_flag |= FKQALLOWED;
444 		return (0);
445 	}
446 
447 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
448 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
449 	error = VOP_OPEN(vp, fmode, cred, td, fp);
450 	if (error != 0)
451 		return (error);
452 
453 	error = vn_open_vnode_advlock(vp, fmode, fp);
454 	if (error == 0 && (fmode & FWRITE) != 0) {
455 		error = VOP_ADD_WRITECOUNT(vp, 1);
456 		if (error == 0) {
457 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
458 			     __func__, vp, vp->v_writecount);
459 		}
460 	}
461 
462 	/*
463 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
464 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
465 	 */
466 	if (error != 0) {
467 		if (fp != NULL) {
468 			/*
469 			 * Arrange the call by having fdrop() to use
470 			 * vn_closefile().  This is to satisfy
471 			 * filesystems like devfs or tmpfs, which
472 			 * override fo_close().
473 			 */
474 			fp->f_flag |= FOPENFAILED;
475 			fp->f_vnode = vp;
476 			if (fp->f_ops == &badfileops) {
477 				fp->f_type = DTYPE_VNODE;
478 				fp->f_ops = &vnops;
479 			}
480 			vref(vp);
481 		} else {
482 			/*
483 			 * If there is no fp, due to kernel-mode open,
484 			 * we can call VOP_CLOSE() now.
485 			 */
486 			if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
487 			    !MNT_EXTENDED_SHARED(vp->v_mount) &&
488 			    VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
489 				vn_lock(vp, LK_UPGRADE | LK_RETRY);
490 			(void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
491 			    cred, td);
492 		}
493 	}
494 
495 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
496 	return (error);
497 
498 }
499 
500 /*
501  * Check for write permissions on the specified vnode.
502  * Prototype text segments cannot be written.
503  * It is racy.
504  */
505 int
506 vn_writechk(struct vnode *vp)
507 {
508 
509 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
510 	/*
511 	 * If there's shared text associated with
512 	 * the vnode, try to free it up once.  If
513 	 * we fail, we can't allow writing.
514 	 */
515 	if (VOP_IS_TEXT(vp))
516 		return (ETXTBSY);
517 
518 	return (0);
519 }
520 
521 /*
522  * Vnode close call
523  */
524 static int
525 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
526     struct thread *td, bool keep_ref)
527 {
528 	struct mount *mp;
529 	int error, lock_flags;
530 
531 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
532 	    MNT_EXTENDED_SHARED(vp->v_mount))
533 		lock_flags = LK_SHARED;
534 	else
535 		lock_flags = LK_EXCLUSIVE;
536 
537 	vn_start_write(vp, &mp, V_WAIT);
538 	vn_lock(vp, lock_flags | LK_RETRY);
539 	AUDIT_ARG_VNODE1(vp);
540 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
541 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
542 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
543 		    __func__, vp, vp->v_writecount);
544 	}
545 	error = VOP_CLOSE(vp, flags, file_cred, td);
546 	if (keep_ref)
547 		VOP_UNLOCK(vp);
548 	else
549 		vput(vp);
550 	vn_finished_write(mp);
551 	return (error);
552 }
553 
554 int
555 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
556     struct thread *td)
557 {
558 
559 	return (vn_close1(vp, flags, file_cred, td, false));
560 }
561 
562 /*
563  * Heuristic to detect sequential operation.
564  */
565 static int
566 sequential_heuristic(struct uio *uio, struct file *fp)
567 {
568 	enum uio_rw rw;
569 
570 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
571 
572 	rw = uio->uio_rw;
573 	if (fp->f_flag & FRDAHEAD)
574 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
575 
576 	/*
577 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
578 	 * that the first I/O is normally considered to be slightly
579 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
580 	 * unless previous seeks have reduced f_seqcount to 0, in which
581 	 * case offset 0 is not special.
582 	 */
583 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
584 	    uio->uio_offset == fp->f_nextoff[rw]) {
585 		/*
586 		 * f_seqcount is in units of fixed-size blocks so that it
587 		 * depends mainly on the amount of sequential I/O and not
588 		 * much on the number of sequential I/O's.  The fixed size
589 		 * of 16384 is hard-coded here since it is (not quite) just
590 		 * a magic size that works well here.  This size is more
591 		 * closely related to the best I/O size for real disks than
592 		 * to any block size used by software.
593 		 */
594 		if (uio->uio_resid >= IO_SEQMAX * 16384)
595 			fp->f_seqcount[rw] = IO_SEQMAX;
596 		else {
597 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
598 			if (fp->f_seqcount[rw] > IO_SEQMAX)
599 				fp->f_seqcount[rw] = IO_SEQMAX;
600 		}
601 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
602 	}
603 
604 	/* Not sequential.  Quickly draw-down sequentiality. */
605 	if (fp->f_seqcount[rw] > 1)
606 		fp->f_seqcount[rw] = 1;
607 	else
608 		fp->f_seqcount[rw] = 0;
609 	return (0);
610 }
611 
612 /*
613  * Package up an I/O request on a vnode into a uio and do it.
614  */
615 int
616 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
617     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
618     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
619 {
620 	struct uio auio;
621 	struct iovec aiov;
622 	struct mount *mp;
623 	struct ucred *cred;
624 	void *rl_cookie;
625 	struct vn_io_fault_args args;
626 	int error, lock_flags;
627 
628 	if (offset < 0 && vp->v_type != VCHR)
629 		return (EINVAL);
630 	auio.uio_iov = &aiov;
631 	auio.uio_iovcnt = 1;
632 	aiov.iov_base = base;
633 	aiov.iov_len = len;
634 	auio.uio_resid = len;
635 	auio.uio_offset = offset;
636 	auio.uio_segflg = segflg;
637 	auio.uio_rw = rw;
638 	auio.uio_td = td;
639 	error = 0;
640 
641 	if ((ioflg & IO_NODELOCKED) == 0) {
642 		if ((ioflg & IO_RANGELOCKED) == 0) {
643 			if (rw == UIO_READ) {
644 				rl_cookie = vn_rangelock_rlock(vp, offset,
645 				    offset + len);
646 			} else if ((ioflg & IO_APPEND) != 0) {
647 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
648 			} else {
649 				rl_cookie = vn_rangelock_wlock(vp, offset,
650 				    offset + len);
651 			}
652 		} else
653 			rl_cookie = NULL;
654 		mp = NULL;
655 		if (rw == UIO_WRITE) {
656 			if (vp->v_type != VCHR &&
657 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
658 			    != 0)
659 				goto out;
660 			lock_flags = vn_lktype_write(mp, vp);
661 		} else
662 			lock_flags = LK_SHARED;
663 		vn_lock(vp, lock_flags | LK_RETRY);
664 	} else
665 		rl_cookie = NULL;
666 
667 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
668 #ifdef MAC
669 	if ((ioflg & IO_NOMACCHECK) == 0) {
670 		if (rw == UIO_READ)
671 			error = mac_vnode_check_read(active_cred, file_cred,
672 			    vp);
673 		else
674 			error = mac_vnode_check_write(active_cred, file_cred,
675 			    vp);
676 	}
677 #endif
678 	if (error == 0) {
679 		if (file_cred != NULL)
680 			cred = file_cred;
681 		else
682 			cred = active_cred;
683 		if (do_vn_io_fault(vp, &auio)) {
684 			args.kind = VN_IO_FAULT_VOP;
685 			args.cred = cred;
686 			args.flags = ioflg;
687 			args.args.vop_args.vp = vp;
688 			error = vn_io_fault1(vp, &auio, &args, td);
689 		} else if (rw == UIO_READ) {
690 			error = VOP_READ(vp, &auio, ioflg, cred);
691 		} else /* if (rw == UIO_WRITE) */ {
692 			error = VOP_WRITE(vp, &auio, ioflg, cred);
693 		}
694 	}
695 	if (aresid)
696 		*aresid = auio.uio_resid;
697 	else
698 		if (auio.uio_resid && error == 0)
699 			error = EIO;
700 	if ((ioflg & IO_NODELOCKED) == 0) {
701 		VOP_UNLOCK(vp);
702 		if (mp != NULL)
703 			vn_finished_write(mp);
704 	}
705  out:
706 	if (rl_cookie != NULL)
707 		vn_rangelock_unlock(vp, rl_cookie);
708 	return (error);
709 }
710 
711 /*
712  * Package up an I/O request on a vnode into a uio and do it.  The I/O
713  * request is split up into smaller chunks and we try to avoid saturating
714  * the buffer cache while potentially holding a vnode locked, so we
715  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
716  * to give other processes a chance to lock the vnode (either other processes
717  * core'ing the same binary, or unrelated processes scanning the directory).
718  */
719 int
720 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
721     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
722     struct ucred *file_cred, size_t *aresid, struct thread *td)
723 {
724 	int error = 0;
725 	ssize_t iaresid;
726 
727 	do {
728 		int chunk;
729 
730 		/*
731 		 * Force `offset' to a multiple of MAXBSIZE except possibly
732 		 * for the first chunk, so that filesystems only need to
733 		 * write full blocks except possibly for the first and last
734 		 * chunks.
735 		 */
736 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
737 
738 		if (chunk > len)
739 			chunk = len;
740 		if (rw != UIO_READ && vp->v_type == VREG)
741 			bwillwrite();
742 		iaresid = 0;
743 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
744 		    ioflg, active_cred, file_cred, &iaresid, td);
745 		len -= chunk;	/* aresid calc already includes length */
746 		if (error)
747 			break;
748 		offset += chunk;
749 		base = (char *)base + chunk;
750 		kern_yield(PRI_USER);
751 	} while (len);
752 	if (aresid)
753 		*aresid = len + iaresid;
754 	return (error);
755 }
756 
757 #if OFF_MAX <= LONG_MAX
758 off_t
759 foffset_lock(struct file *fp, int flags)
760 {
761 	volatile short *flagsp;
762 	off_t res;
763 	short state;
764 
765 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
766 
767 	if ((flags & FOF_NOLOCK) != 0)
768 		return (atomic_load_long(&fp->f_offset));
769 
770 	/*
771 	 * According to McKusick the vn lock was protecting f_offset here.
772 	 * It is now protected by the FOFFSET_LOCKED flag.
773 	 */
774 	flagsp = &fp->f_vnread_flags;
775 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
776 		return (atomic_load_long(&fp->f_offset));
777 
778 	sleepq_lock(&fp->f_vnread_flags);
779 	state = atomic_load_16(flagsp);
780 	for (;;) {
781 		if ((state & FOFFSET_LOCKED) == 0) {
782 			if (!atomic_fcmpset_acq_16(flagsp, &state,
783 			    FOFFSET_LOCKED))
784 				continue;
785 			break;
786 		}
787 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
788 			if (!atomic_fcmpset_acq_16(flagsp, &state,
789 			    state | FOFFSET_LOCK_WAITING))
790 				continue;
791 		}
792 		DROP_GIANT();
793 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
794 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
795 		PICKUP_GIANT();
796 		sleepq_lock(&fp->f_vnread_flags);
797 		state = atomic_load_16(flagsp);
798 	}
799 	res = atomic_load_long(&fp->f_offset);
800 	sleepq_release(&fp->f_vnread_flags);
801 	return (res);
802 }
803 
804 void
805 foffset_unlock(struct file *fp, off_t val, int flags)
806 {
807 	volatile short *flagsp;
808 	short state;
809 
810 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
811 
812 	if ((flags & FOF_NOUPDATE) == 0)
813 		atomic_store_long(&fp->f_offset, val);
814 	if ((flags & FOF_NEXTOFF_R) != 0)
815 		fp->f_nextoff[UIO_READ] = val;
816 	if ((flags & FOF_NEXTOFF_W) != 0)
817 		fp->f_nextoff[UIO_WRITE] = val;
818 
819 	if ((flags & FOF_NOLOCK) != 0)
820 		return;
821 
822 	flagsp = &fp->f_vnread_flags;
823 	state = atomic_load_16(flagsp);
824 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
825 	    atomic_cmpset_rel_16(flagsp, state, 0))
826 		return;
827 
828 	sleepq_lock(&fp->f_vnread_flags);
829 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
830 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
831 	fp->f_vnread_flags = 0;
832 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
833 	sleepq_release(&fp->f_vnread_flags);
834 }
835 #else
836 off_t
837 foffset_lock(struct file *fp, int flags)
838 {
839 	struct mtx *mtxp;
840 	off_t res;
841 
842 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
843 
844 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
845 	mtx_lock(mtxp);
846 	if ((flags & FOF_NOLOCK) == 0) {
847 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
848 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
849 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
850 			    "vofflock", 0);
851 		}
852 		fp->f_vnread_flags |= FOFFSET_LOCKED;
853 	}
854 	res = fp->f_offset;
855 	mtx_unlock(mtxp);
856 	return (res);
857 }
858 
859 void
860 foffset_unlock(struct file *fp, off_t val, int flags)
861 {
862 	struct mtx *mtxp;
863 
864 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
865 
866 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
867 	mtx_lock(mtxp);
868 	if ((flags & FOF_NOUPDATE) == 0)
869 		fp->f_offset = val;
870 	if ((flags & FOF_NEXTOFF_R) != 0)
871 		fp->f_nextoff[UIO_READ] = val;
872 	if ((flags & FOF_NEXTOFF_W) != 0)
873 		fp->f_nextoff[UIO_WRITE] = val;
874 	if ((flags & FOF_NOLOCK) == 0) {
875 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
876 		    ("Lost FOFFSET_LOCKED"));
877 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
878 			wakeup(&fp->f_vnread_flags);
879 		fp->f_vnread_flags = 0;
880 	}
881 	mtx_unlock(mtxp);
882 }
883 #endif
884 
885 void
886 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
887 {
888 
889 	if ((flags & FOF_OFFSET) == 0)
890 		uio->uio_offset = foffset_lock(fp, flags);
891 }
892 
893 void
894 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
895 {
896 
897 	if ((flags & FOF_OFFSET) == 0)
898 		foffset_unlock(fp, uio->uio_offset, flags);
899 }
900 
901 static int
902 get_advice(struct file *fp, struct uio *uio)
903 {
904 	struct mtx *mtxp;
905 	int ret;
906 
907 	ret = POSIX_FADV_NORMAL;
908 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
909 		return (ret);
910 
911 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
912 	mtx_lock(mtxp);
913 	if (fp->f_advice != NULL &&
914 	    uio->uio_offset >= fp->f_advice->fa_start &&
915 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
916 		ret = fp->f_advice->fa_advice;
917 	mtx_unlock(mtxp);
918 	return (ret);
919 }
920 
921 static int
922 get_write_ioflag(struct file *fp)
923 {
924 	int ioflag;
925 	struct mount *mp;
926 	struct vnode *vp;
927 
928 	ioflag = 0;
929 	vp = fp->f_vnode;
930 	mp = atomic_load_ptr(&vp->v_mount);
931 
932 	if ((fp->f_flag & O_DIRECT) != 0)
933 		ioflag |= IO_DIRECT;
934 
935 	if ((fp->f_flag & O_FSYNC) != 0 ||
936 	    (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
937 		ioflag |= IO_SYNC;
938 
939 	/*
940 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
941 	 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
942 	 * fall back to full O_SYNC behavior.
943 	 */
944 	if ((fp->f_flag & O_DSYNC) != 0)
945 		ioflag |= IO_SYNC | IO_DATASYNC;
946 
947 	return (ioflag);
948 }
949 
950 int
951 vn_read_from_obj(struct vnode *vp, struct uio *uio)
952 {
953 	vm_object_t obj;
954 	vm_page_t ma[io_hold_cnt + 2];
955 	off_t off, vsz;
956 	ssize_t resid;
957 	int error, i, j;
958 
959 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
960 	obj = atomic_load_ptr(&vp->v_object);
961 	if (obj == NULL)
962 		return (EJUSTRETURN);
963 
964 	/*
965 	 * Depends on type stability of vm_objects.
966 	 */
967 	vm_object_pip_add(obj, 1);
968 	if ((obj->flags & OBJ_DEAD) != 0) {
969 		/*
970 		 * Note that object might be already reused from the
971 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
972 		 * we recheck for DOOMED vnode state after all pages
973 		 * are busied, and retract then.
974 		 *
975 		 * But we check for OBJ_DEAD to ensure that we do not
976 		 * busy pages while vm_object_terminate_pages()
977 		 * processes the queue.
978 		 */
979 		error = EJUSTRETURN;
980 		goto out_pip;
981 	}
982 
983 	resid = uio->uio_resid;
984 	off = uio->uio_offset;
985 	for (i = 0; resid > 0; i++) {
986 		MPASS(i < io_hold_cnt + 2);
987 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
988 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
989 		    VM_ALLOC_NOWAIT);
990 		if (ma[i] == NULL)
991 			break;
992 
993 		/*
994 		 * Skip invalid pages.  Valid mask can be partial only
995 		 * at EOF, and we clip later.
996 		 */
997 		if (vm_page_none_valid(ma[i])) {
998 			vm_page_sunbusy(ma[i]);
999 			break;
1000 		}
1001 
1002 		resid -= PAGE_SIZE;
1003 		off += PAGE_SIZE;
1004 	}
1005 	if (i == 0) {
1006 		error = EJUSTRETURN;
1007 		goto out_pip;
1008 	}
1009 
1010 	/*
1011 	 * Check VIRF_DOOMED after we busied our pages.  Since
1012 	 * vgonel() terminates the vnode' vm_object, it cannot
1013 	 * process past pages busied by us.
1014 	 */
1015 	if (VN_IS_DOOMED(vp)) {
1016 		error = EJUSTRETURN;
1017 		goto out;
1018 	}
1019 
1020 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1021 	if (resid > uio->uio_resid)
1022 		resid = uio->uio_resid;
1023 
1024 	/*
1025 	 * Unlocked read of vnp_size is safe because truncation cannot
1026 	 * pass busied page.  But we load vnp_size into a local
1027 	 * variable so that possible concurrent extension does not
1028 	 * break calculation.
1029 	 */
1030 #if defined(__powerpc__) && !defined(__powerpc64__)
1031 	vsz = obj->un_pager.vnp.vnp_size;
1032 #else
1033 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1034 #endif
1035 	if (uio->uio_offset >= vsz) {
1036 		error = EJUSTRETURN;
1037 		goto out;
1038 	}
1039 	if (uio->uio_offset + resid > vsz)
1040 		resid = vsz - uio->uio_offset;
1041 
1042 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1043 
1044 out:
1045 	for (j = 0; j < i; j++) {
1046 		if (error == 0)
1047 			vm_page_reference(ma[j]);
1048 		vm_page_sunbusy(ma[j]);
1049 	}
1050 out_pip:
1051 	vm_object_pip_wakeup(obj);
1052 	if (error != 0)
1053 		return (error);
1054 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1055 }
1056 
1057 /*
1058  * File table vnode read routine.
1059  */
1060 static int
1061 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1062     struct thread *td)
1063 {
1064 	struct vnode *vp;
1065 	off_t orig_offset;
1066 	int error, ioflag;
1067 	int advice;
1068 
1069 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1070 	    uio->uio_td, td));
1071 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1072 	vp = fp->f_vnode;
1073 	ioflag = 0;
1074 	if (fp->f_flag & FNONBLOCK)
1075 		ioflag |= IO_NDELAY;
1076 	if (fp->f_flag & O_DIRECT)
1077 		ioflag |= IO_DIRECT;
1078 
1079 	/*
1080 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1081 	 * allows us to avoid unneeded work outright.
1082 	 */
1083 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1084 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1085 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1086 		if (error == 0) {
1087 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1088 			return (0);
1089 		}
1090 		if (error != EJUSTRETURN)
1091 			return (error);
1092 	}
1093 
1094 	advice = get_advice(fp, uio);
1095 	vn_lock(vp, LK_SHARED | LK_RETRY);
1096 
1097 	switch (advice) {
1098 	case POSIX_FADV_NORMAL:
1099 	case POSIX_FADV_SEQUENTIAL:
1100 	case POSIX_FADV_NOREUSE:
1101 		ioflag |= sequential_heuristic(uio, fp);
1102 		break;
1103 	case POSIX_FADV_RANDOM:
1104 		/* Disable read-ahead for random I/O. */
1105 		break;
1106 	}
1107 	orig_offset = uio->uio_offset;
1108 
1109 #ifdef MAC
1110 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1111 	if (error == 0)
1112 #endif
1113 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1114 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1115 	VOP_UNLOCK(vp);
1116 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1117 	    orig_offset != uio->uio_offset)
1118 		/*
1119 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1120 		 * for the backing file after a POSIX_FADV_NOREUSE
1121 		 * read(2).
1122 		 */
1123 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1124 		    POSIX_FADV_DONTNEED);
1125 	return (error);
1126 }
1127 
1128 /*
1129  * File table vnode write routine.
1130  */
1131 static int
1132 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1133     struct thread *td)
1134 {
1135 	struct vnode *vp;
1136 	struct mount *mp;
1137 	off_t orig_offset;
1138 	int error, ioflag;
1139 	int advice;
1140 	bool need_finished_write;
1141 
1142 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1143 	    uio->uio_td, td));
1144 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1145 	vp = fp->f_vnode;
1146 	if (vp->v_type == VREG)
1147 		bwillwrite();
1148 	ioflag = IO_UNIT;
1149 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1150 		ioflag |= IO_APPEND;
1151 	if ((fp->f_flag & FNONBLOCK) != 0)
1152 		ioflag |= IO_NDELAY;
1153 	ioflag |= get_write_ioflag(fp);
1154 
1155 	mp = NULL;
1156 	need_finished_write = false;
1157 	if (vp->v_type != VCHR) {
1158 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1159 		if (error != 0)
1160 			goto unlock;
1161 		need_finished_write = true;
1162 	}
1163 
1164 	advice = get_advice(fp, uio);
1165 
1166 	vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1167 	switch (advice) {
1168 	case POSIX_FADV_NORMAL:
1169 	case POSIX_FADV_SEQUENTIAL:
1170 	case POSIX_FADV_NOREUSE:
1171 		ioflag |= sequential_heuristic(uio, fp);
1172 		break;
1173 	case POSIX_FADV_RANDOM:
1174 		/* XXX: Is this correct? */
1175 		break;
1176 	}
1177 	orig_offset = uio->uio_offset;
1178 
1179 #ifdef MAC
1180 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1181 	if (error == 0)
1182 #endif
1183 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1184 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1185 	VOP_UNLOCK(vp);
1186 	if (need_finished_write)
1187 		vn_finished_write(mp);
1188 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1189 	    orig_offset != uio->uio_offset)
1190 		/*
1191 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1192 		 * for the backing file after a POSIX_FADV_NOREUSE
1193 		 * write(2).
1194 		 */
1195 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1196 		    POSIX_FADV_DONTNEED);
1197 unlock:
1198 	return (error);
1199 }
1200 
1201 /*
1202  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1203  * prevent the following deadlock:
1204  *
1205  * Assume that the thread A reads from the vnode vp1 into userspace
1206  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1207  * currently not resident, then system ends up with the call chain
1208  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1209  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1210  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1211  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1212  * backed by the pages of vnode vp1, and some page in buf2 is not
1213  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1214  *
1215  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1216  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1217  * Instead, it first tries to do the whole range i/o with pagefaults
1218  * disabled. If all pages in the i/o buffer are resident and mapped,
1219  * VOP will succeed (ignoring the genuine filesystem errors).
1220  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1221  * i/o in chunks, with all pages in the chunk prefaulted and held
1222  * using vm_fault_quick_hold_pages().
1223  *
1224  * Filesystems using this deadlock avoidance scheme should use the
1225  * array of the held pages from uio, saved in the curthread->td_ma,
1226  * instead of doing uiomove().  A helper function
1227  * vn_io_fault_uiomove() converts uiomove request into
1228  * uiomove_fromphys() over td_ma array.
1229  *
1230  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1231  * make the current i/o request atomic with respect to other i/os and
1232  * truncations.
1233  */
1234 
1235 /*
1236  * Decode vn_io_fault_args and perform the corresponding i/o.
1237  */
1238 static int
1239 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1240     struct thread *td)
1241 {
1242 	int error, save;
1243 
1244 	error = 0;
1245 	save = vm_fault_disable_pagefaults();
1246 	switch (args->kind) {
1247 	case VN_IO_FAULT_FOP:
1248 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1249 		    uio, args->cred, args->flags, td);
1250 		break;
1251 	case VN_IO_FAULT_VOP:
1252 		if (uio->uio_rw == UIO_READ) {
1253 			error = VOP_READ(args->args.vop_args.vp, uio,
1254 			    args->flags, args->cred);
1255 		} else if (uio->uio_rw == UIO_WRITE) {
1256 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1257 			    args->flags, args->cred);
1258 		}
1259 		break;
1260 	default:
1261 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1262 		    args->kind, uio->uio_rw);
1263 	}
1264 	vm_fault_enable_pagefaults(save);
1265 	return (error);
1266 }
1267 
1268 static int
1269 vn_io_fault_touch(char *base, const struct uio *uio)
1270 {
1271 	int r;
1272 
1273 	r = fubyte(base);
1274 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1275 		return (EFAULT);
1276 	return (0);
1277 }
1278 
1279 static int
1280 vn_io_fault_prefault_user(const struct uio *uio)
1281 {
1282 	char *base;
1283 	const struct iovec *iov;
1284 	size_t len;
1285 	ssize_t resid;
1286 	int error, i;
1287 
1288 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1289 	    ("vn_io_fault_prefault userspace"));
1290 
1291 	error = i = 0;
1292 	iov = uio->uio_iov;
1293 	resid = uio->uio_resid;
1294 	base = iov->iov_base;
1295 	len = iov->iov_len;
1296 	while (resid > 0) {
1297 		error = vn_io_fault_touch(base, uio);
1298 		if (error != 0)
1299 			break;
1300 		if (len < PAGE_SIZE) {
1301 			if (len != 0) {
1302 				error = vn_io_fault_touch(base + len - 1, uio);
1303 				if (error != 0)
1304 					break;
1305 				resid -= len;
1306 			}
1307 			if (++i >= uio->uio_iovcnt)
1308 				break;
1309 			iov = uio->uio_iov + i;
1310 			base = iov->iov_base;
1311 			len = iov->iov_len;
1312 		} else {
1313 			len -= PAGE_SIZE;
1314 			base += PAGE_SIZE;
1315 			resid -= PAGE_SIZE;
1316 		}
1317 	}
1318 	return (error);
1319 }
1320 
1321 /*
1322  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1323  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1324  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1325  * into args and call vn_io_fault1() to handle faults during the user
1326  * mode buffer accesses.
1327  */
1328 static int
1329 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1330     struct thread *td)
1331 {
1332 	vm_page_t ma[io_hold_cnt + 2];
1333 	struct uio *uio_clone, short_uio;
1334 	struct iovec short_iovec[1];
1335 	vm_page_t *prev_td_ma;
1336 	vm_prot_t prot;
1337 	vm_offset_t addr, end;
1338 	size_t len, resid;
1339 	ssize_t adv;
1340 	int error, cnt, saveheld, prev_td_ma_cnt;
1341 
1342 	if (vn_io_fault_prefault) {
1343 		error = vn_io_fault_prefault_user(uio);
1344 		if (error != 0)
1345 			return (error); /* Or ignore ? */
1346 	}
1347 
1348 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1349 
1350 	/*
1351 	 * The UFS follows IO_UNIT directive and replays back both
1352 	 * uio_offset and uio_resid if an error is encountered during the
1353 	 * operation.  But, since the iovec may be already advanced,
1354 	 * uio is still in an inconsistent state.
1355 	 *
1356 	 * Cache a copy of the original uio, which is advanced to the redo
1357 	 * point using UIO_NOCOPY below.
1358 	 */
1359 	uio_clone = cloneuio(uio);
1360 	resid = uio->uio_resid;
1361 
1362 	short_uio.uio_segflg = UIO_USERSPACE;
1363 	short_uio.uio_rw = uio->uio_rw;
1364 	short_uio.uio_td = uio->uio_td;
1365 
1366 	error = vn_io_fault_doio(args, uio, td);
1367 	if (error != EFAULT)
1368 		goto out;
1369 
1370 	atomic_add_long(&vn_io_faults_cnt, 1);
1371 	uio_clone->uio_segflg = UIO_NOCOPY;
1372 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1373 	uio_clone->uio_segflg = uio->uio_segflg;
1374 
1375 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1376 	prev_td_ma = td->td_ma;
1377 	prev_td_ma_cnt = td->td_ma_cnt;
1378 
1379 	while (uio_clone->uio_resid != 0) {
1380 		len = uio_clone->uio_iov->iov_len;
1381 		if (len == 0) {
1382 			KASSERT(uio_clone->uio_iovcnt >= 1,
1383 			    ("iovcnt underflow"));
1384 			uio_clone->uio_iov++;
1385 			uio_clone->uio_iovcnt--;
1386 			continue;
1387 		}
1388 		if (len > ptoa(io_hold_cnt))
1389 			len = ptoa(io_hold_cnt);
1390 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1391 		end = round_page(addr + len);
1392 		if (end < addr) {
1393 			error = EFAULT;
1394 			break;
1395 		}
1396 		cnt = atop(end - trunc_page(addr));
1397 		/*
1398 		 * A perfectly misaligned address and length could cause
1399 		 * both the start and the end of the chunk to use partial
1400 		 * page.  +2 accounts for such a situation.
1401 		 */
1402 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1403 		    addr, len, prot, ma, io_hold_cnt + 2);
1404 		if (cnt == -1) {
1405 			error = EFAULT;
1406 			break;
1407 		}
1408 		short_uio.uio_iov = &short_iovec[0];
1409 		short_iovec[0].iov_base = (void *)addr;
1410 		short_uio.uio_iovcnt = 1;
1411 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1412 		short_uio.uio_offset = uio_clone->uio_offset;
1413 		td->td_ma = ma;
1414 		td->td_ma_cnt = cnt;
1415 
1416 		error = vn_io_fault_doio(args, &short_uio, td);
1417 		vm_page_unhold_pages(ma, cnt);
1418 		adv = len - short_uio.uio_resid;
1419 
1420 		uio_clone->uio_iov->iov_base =
1421 		    (char *)uio_clone->uio_iov->iov_base + adv;
1422 		uio_clone->uio_iov->iov_len -= adv;
1423 		uio_clone->uio_resid -= adv;
1424 		uio_clone->uio_offset += adv;
1425 
1426 		uio->uio_resid -= adv;
1427 		uio->uio_offset += adv;
1428 
1429 		if (error != 0 || adv == 0)
1430 			break;
1431 	}
1432 	td->td_ma = prev_td_ma;
1433 	td->td_ma_cnt = prev_td_ma_cnt;
1434 	curthread_pflags_restore(saveheld);
1435 out:
1436 	free(uio_clone, M_IOV);
1437 	return (error);
1438 }
1439 
1440 static int
1441 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1442     int flags, struct thread *td)
1443 {
1444 	fo_rdwr_t *doio;
1445 	struct vnode *vp;
1446 	void *rl_cookie;
1447 	struct vn_io_fault_args args;
1448 	int error;
1449 
1450 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1451 	vp = fp->f_vnode;
1452 
1453 	/*
1454 	 * The ability to read(2) on a directory has historically been
1455 	 * allowed for all users, but this can and has been the source of
1456 	 * at least one security issue in the past.  As such, it is now hidden
1457 	 * away behind a sysctl for those that actually need it to use it, and
1458 	 * restricted to root when it's turned on to make it relatively safe to
1459 	 * leave on for longer sessions of need.
1460 	 */
1461 	if (vp->v_type == VDIR) {
1462 		KASSERT(uio->uio_rw == UIO_READ,
1463 		    ("illegal write attempted on a directory"));
1464 		if (!vfs_allow_read_dir)
1465 			return (EISDIR);
1466 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1467 			return (EISDIR);
1468 	}
1469 
1470 	foffset_lock_uio(fp, uio, flags);
1471 	if (do_vn_io_fault(vp, uio)) {
1472 		args.kind = VN_IO_FAULT_FOP;
1473 		args.args.fop_args.fp = fp;
1474 		args.args.fop_args.doio = doio;
1475 		args.cred = active_cred;
1476 		args.flags = flags | FOF_OFFSET;
1477 		if (uio->uio_rw == UIO_READ) {
1478 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1479 			    uio->uio_offset + uio->uio_resid);
1480 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1481 		    (flags & FOF_OFFSET) == 0) {
1482 			/* For appenders, punt and lock the whole range. */
1483 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1484 		} else {
1485 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1486 			    uio->uio_offset + uio->uio_resid);
1487 		}
1488 		error = vn_io_fault1(vp, uio, &args, td);
1489 		vn_rangelock_unlock(vp, rl_cookie);
1490 	} else {
1491 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1492 	}
1493 	foffset_unlock_uio(fp, uio, flags);
1494 	return (error);
1495 }
1496 
1497 /*
1498  * Helper function to perform the requested uiomove operation using
1499  * the held pages for io->uio_iov[0].iov_base buffer instead of
1500  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1501  * instead of iov_base prevents page faults that could occur due to
1502  * pmap_collect() invalidating the mapping created by
1503  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1504  * object cleanup revoking the write access from page mappings.
1505  *
1506  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1507  * instead of plain uiomove().
1508  */
1509 int
1510 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1511 {
1512 	struct uio transp_uio;
1513 	struct iovec transp_iov[1];
1514 	struct thread *td;
1515 	size_t adv;
1516 	int error, pgadv;
1517 
1518 	td = curthread;
1519 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1520 	    uio->uio_segflg != UIO_USERSPACE)
1521 		return (uiomove(data, xfersize, uio));
1522 
1523 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1524 	transp_iov[0].iov_base = data;
1525 	transp_uio.uio_iov = &transp_iov[0];
1526 	transp_uio.uio_iovcnt = 1;
1527 	if (xfersize > uio->uio_resid)
1528 		xfersize = uio->uio_resid;
1529 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1530 	transp_uio.uio_offset = 0;
1531 	transp_uio.uio_segflg = UIO_SYSSPACE;
1532 	/*
1533 	 * Since transp_iov points to data, and td_ma page array
1534 	 * corresponds to original uio->uio_iov, we need to invert the
1535 	 * direction of the i/o operation as passed to
1536 	 * uiomove_fromphys().
1537 	 */
1538 	switch (uio->uio_rw) {
1539 	case UIO_WRITE:
1540 		transp_uio.uio_rw = UIO_READ;
1541 		break;
1542 	case UIO_READ:
1543 		transp_uio.uio_rw = UIO_WRITE;
1544 		break;
1545 	}
1546 	transp_uio.uio_td = uio->uio_td;
1547 	error = uiomove_fromphys(td->td_ma,
1548 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1549 	    xfersize, &transp_uio);
1550 	adv = xfersize - transp_uio.uio_resid;
1551 	pgadv =
1552 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1553 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1554 	td->td_ma += pgadv;
1555 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1556 	    pgadv));
1557 	td->td_ma_cnt -= pgadv;
1558 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1559 	uio->uio_iov->iov_len -= adv;
1560 	uio->uio_resid -= adv;
1561 	uio->uio_offset += adv;
1562 	return (error);
1563 }
1564 
1565 int
1566 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1567     struct uio *uio)
1568 {
1569 	struct thread *td;
1570 	vm_offset_t iov_base;
1571 	int cnt, pgadv;
1572 
1573 	td = curthread;
1574 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1575 	    uio->uio_segflg != UIO_USERSPACE)
1576 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1577 
1578 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1579 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1580 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1581 	switch (uio->uio_rw) {
1582 	case UIO_WRITE:
1583 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1584 		    offset, cnt);
1585 		break;
1586 	case UIO_READ:
1587 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1588 		    cnt);
1589 		break;
1590 	}
1591 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1592 	td->td_ma += pgadv;
1593 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1594 	    pgadv));
1595 	td->td_ma_cnt -= pgadv;
1596 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1597 	uio->uio_iov->iov_len -= cnt;
1598 	uio->uio_resid -= cnt;
1599 	uio->uio_offset += cnt;
1600 	return (0);
1601 }
1602 
1603 /*
1604  * File table truncate routine.
1605  */
1606 static int
1607 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1608     struct thread *td)
1609 {
1610 	struct mount *mp;
1611 	struct vnode *vp;
1612 	void *rl_cookie;
1613 	int error;
1614 
1615 	vp = fp->f_vnode;
1616 
1617 retry:
1618 	/*
1619 	 * Lock the whole range for truncation.  Otherwise split i/o
1620 	 * might happen partly before and partly after the truncation.
1621 	 */
1622 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1623 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1624 	if (error)
1625 		goto out1;
1626 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1627 	AUDIT_ARG_VNODE1(vp);
1628 	if (vp->v_type == VDIR) {
1629 		error = EISDIR;
1630 		goto out;
1631 	}
1632 #ifdef MAC
1633 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1634 	if (error)
1635 		goto out;
1636 #endif
1637 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1638 	    fp->f_cred);
1639 out:
1640 	VOP_UNLOCK(vp);
1641 	vn_finished_write(mp);
1642 out1:
1643 	vn_rangelock_unlock(vp, rl_cookie);
1644 	if (error == ERELOOKUP)
1645 		goto retry;
1646 	return (error);
1647 }
1648 
1649 /*
1650  * Truncate a file that is already locked.
1651  */
1652 int
1653 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1654     struct ucred *cred)
1655 {
1656 	struct vattr vattr;
1657 	int error;
1658 
1659 	error = VOP_ADD_WRITECOUNT(vp, 1);
1660 	if (error == 0) {
1661 		VATTR_NULL(&vattr);
1662 		vattr.va_size = length;
1663 		if (sync)
1664 			vattr.va_vaflags |= VA_SYNC;
1665 		error = VOP_SETATTR(vp, &vattr, cred);
1666 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1667 	}
1668 	return (error);
1669 }
1670 
1671 /*
1672  * File table vnode stat routine.
1673  */
1674 int
1675 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred)
1676 {
1677 	struct vnode *vp = fp->f_vnode;
1678 	int error;
1679 
1680 	vn_lock(vp, LK_SHARED | LK_RETRY);
1681 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred);
1682 	VOP_UNLOCK(vp);
1683 
1684 	return (error);
1685 }
1686 
1687 /*
1688  * File table vnode ioctl routine.
1689  */
1690 static int
1691 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1692     struct thread *td)
1693 {
1694 	struct vattr vattr;
1695 	struct vnode *vp;
1696 	struct fiobmap2_arg *bmarg;
1697 	int error;
1698 
1699 	vp = fp->f_vnode;
1700 	switch (vp->v_type) {
1701 	case VDIR:
1702 	case VREG:
1703 		switch (com) {
1704 		case FIONREAD:
1705 			vn_lock(vp, LK_SHARED | LK_RETRY);
1706 			error = VOP_GETATTR(vp, &vattr, active_cred);
1707 			VOP_UNLOCK(vp);
1708 			if (error == 0)
1709 				*(int *)data = vattr.va_size - fp->f_offset;
1710 			return (error);
1711 		case FIOBMAP2:
1712 			bmarg = (struct fiobmap2_arg *)data;
1713 			vn_lock(vp, LK_SHARED | LK_RETRY);
1714 #ifdef MAC
1715 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1716 			    vp);
1717 			if (error == 0)
1718 #endif
1719 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1720 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1721 			VOP_UNLOCK(vp);
1722 			return (error);
1723 		case FIONBIO:
1724 		case FIOASYNC:
1725 			return (0);
1726 		default:
1727 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1728 			    active_cred, td));
1729 		}
1730 		break;
1731 	case VCHR:
1732 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1733 		    active_cred, td));
1734 	default:
1735 		return (ENOTTY);
1736 	}
1737 }
1738 
1739 /*
1740  * File table vnode poll routine.
1741  */
1742 static int
1743 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1744     struct thread *td)
1745 {
1746 	struct vnode *vp;
1747 	int error;
1748 
1749 	vp = fp->f_vnode;
1750 #if defined(MAC) || defined(AUDIT)
1751 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1752 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1753 		AUDIT_ARG_VNODE1(vp);
1754 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1755 		VOP_UNLOCK(vp);
1756 		if (error != 0)
1757 			return (error);
1758 	}
1759 #endif
1760 	error = VOP_POLL(vp, events, fp->f_cred, td);
1761 	return (error);
1762 }
1763 
1764 /*
1765  * Acquire the requested lock and then check for validity.  LK_RETRY
1766  * permits vn_lock to return doomed vnodes.
1767  */
1768 static int __noinline
1769 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1770     int error)
1771 {
1772 
1773 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1774 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1775 
1776 	if (error == 0)
1777 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1778 
1779 	if ((flags & LK_RETRY) == 0) {
1780 		if (error == 0) {
1781 			VOP_UNLOCK(vp);
1782 			error = ENOENT;
1783 		}
1784 		return (error);
1785 	}
1786 
1787 	/*
1788 	 * LK_RETRY case.
1789 	 *
1790 	 * Nothing to do if we got the lock.
1791 	 */
1792 	if (error == 0)
1793 		return (0);
1794 
1795 	/*
1796 	 * Interlock was dropped by the call in _vn_lock.
1797 	 */
1798 	flags &= ~LK_INTERLOCK;
1799 	do {
1800 		error = VOP_LOCK1(vp, flags, file, line);
1801 	} while (error != 0);
1802 	return (0);
1803 }
1804 
1805 int
1806 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1807 {
1808 	int error;
1809 
1810 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1811 	    ("vn_lock: no locktype (%d passed)", flags));
1812 	VNPASS(vp->v_holdcnt > 0, vp);
1813 	error = VOP_LOCK1(vp, flags, file, line);
1814 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1815 		return (_vn_lock_fallback(vp, flags, file, line, error));
1816 	return (0);
1817 }
1818 
1819 /*
1820  * File table vnode close routine.
1821  */
1822 static int
1823 vn_closefile(struct file *fp, struct thread *td)
1824 {
1825 	struct vnode *vp;
1826 	struct flock lf;
1827 	int error;
1828 	bool ref;
1829 
1830 	vp = fp->f_vnode;
1831 	fp->f_ops = &badfileops;
1832 	ref = (fp->f_flag & FHASLOCK) != 0;
1833 
1834 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1835 
1836 	if (__predict_false(ref)) {
1837 		lf.l_whence = SEEK_SET;
1838 		lf.l_start = 0;
1839 		lf.l_len = 0;
1840 		lf.l_type = F_UNLCK;
1841 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1842 		vrele(vp);
1843 	}
1844 	return (error);
1845 }
1846 
1847 /*
1848  * Preparing to start a filesystem write operation. If the operation is
1849  * permitted, then we bump the count of operations in progress and
1850  * proceed. If a suspend request is in progress, we wait until the
1851  * suspension is over, and then proceed.
1852  */
1853 static int
1854 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1855 {
1856 	struct mount_pcpu *mpcpu;
1857 	int error, mflags;
1858 
1859 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1860 	    vfs_op_thread_enter(mp, mpcpu)) {
1861 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1862 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1863 		vfs_op_thread_exit(mp, mpcpu);
1864 		return (0);
1865 	}
1866 
1867 	if (mplocked)
1868 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1869 	else
1870 		MNT_ILOCK(mp);
1871 
1872 	error = 0;
1873 
1874 	/*
1875 	 * Check on status of suspension.
1876 	 */
1877 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1878 	    mp->mnt_susp_owner != curthread) {
1879 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1880 		    (flags & PCATCH) : 0) | (PUSER - 1);
1881 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1882 			if (flags & V_NOWAIT) {
1883 				error = EWOULDBLOCK;
1884 				goto unlock;
1885 			}
1886 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1887 			    "suspfs", 0);
1888 			if (error)
1889 				goto unlock;
1890 		}
1891 	}
1892 	if (flags & V_XSLEEP)
1893 		goto unlock;
1894 	mp->mnt_writeopcount++;
1895 unlock:
1896 	if (error != 0 || (flags & V_XSLEEP) != 0)
1897 		MNT_REL(mp);
1898 	MNT_IUNLOCK(mp);
1899 	return (error);
1900 }
1901 
1902 int
1903 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1904 {
1905 	struct mount *mp;
1906 	int error;
1907 
1908 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1909 	    ("V_MNTREF requires mp"));
1910 
1911 	error = 0;
1912 	/*
1913 	 * If a vnode is provided, get and return the mount point that
1914 	 * to which it will write.
1915 	 */
1916 	if (vp != NULL) {
1917 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1918 			*mpp = NULL;
1919 			if (error != EOPNOTSUPP)
1920 				return (error);
1921 			return (0);
1922 		}
1923 	}
1924 	if ((mp = *mpp) == NULL)
1925 		return (0);
1926 
1927 	/*
1928 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1929 	 * a vfs_ref().
1930 	 * As long as a vnode is not provided we need to acquire a
1931 	 * refcount for the provided mountpoint too, in order to
1932 	 * emulate a vfs_ref().
1933 	 */
1934 	if (vp == NULL && (flags & V_MNTREF) == 0)
1935 		vfs_ref(mp);
1936 
1937 	return (vn_start_write_refed(mp, flags, false));
1938 }
1939 
1940 /*
1941  * Secondary suspension. Used by operations such as vop_inactive
1942  * routines that are needed by the higher level functions. These
1943  * are allowed to proceed until all the higher level functions have
1944  * completed (indicated by mnt_writeopcount dropping to zero). At that
1945  * time, these operations are halted until the suspension is over.
1946  */
1947 int
1948 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1949 {
1950 	struct mount *mp;
1951 	int error;
1952 
1953 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1954 	    ("V_MNTREF requires mp"));
1955 
1956  retry:
1957 	if (vp != NULL) {
1958 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1959 			*mpp = NULL;
1960 			if (error != EOPNOTSUPP)
1961 				return (error);
1962 			return (0);
1963 		}
1964 	}
1965 	/*
1966 	 * If we are not suspended or have not yet reached suspended
1967 	 * mode, then let the operation proceed.
1968 	 */
1969 	if ((mp = *mpp) == NULL)
1970 		return (0);
1971 
1972 	/*
1973 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1974 	 * a vfs_ref().
1975 	 * As long as a vnode is not provided we need to acquire a
1976 	 * refcount for the provided mountpoint too, in order to
1977 	 * emulate a vfs_ref().
1978 	 */
1979 	MNT_ILOCK(mp);
1980 	if (vp == NULL && (flags & V_MNTREF) == 0)
1981 		MNT_REF(mp);
1982 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1983 		mp->mnt_secondary_writes++;
1984 		mp->mnt_secondary_accwrites++;
1985 		MNT_IUNLOCK(mp);
1986 		return (0);
1987 	}
1988 	if (flags & V_NOWAIT) {
1989 		MNT_REL(mp);
1990 		MNT_IUNLOCK(mp);
1991 		return (EWOULDBLOCK);
1992 	}
1993 	/*
1994 	 * Wait for the suspension to finish.
1995 	 */
1996 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1997 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1998 	    "suspfs", 0);
1999 	vfs_rel(mp);
2000 	if (error == 0)
2001 		goto retry;
2002 	return (error);
2003 }
2004 
2005 /*
2006  * Filesystem write operation has completed. If we are suspending and this
2007  * operation is the last one, notify the suspender that the suspension is
2008  * now in effect.
2009  */
2010 void
2011 vn_finished_write(struct mount *mp)
2012 {
2013 	struct mount_pcpu *mpcpu;
2014 	int c;
2015 
2016 	if (mp == NULL)
2017 		return;
2018 
2019 	if (vfs_op_thread_enter(mp, mpcpu)) {
2020 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2021 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2022 		vfs_op_thread_exit(mp, mpcpu);
2023 		return;
2024 	}
2025 
2026 	MNT_ILOCK(mp);
2027 	vfs_assert_mount_counters(mp);
2028 	MNT_REL(mp);
2029 	c = --mp->mnt_writeopcount;
2030 	if (mp->mnt_vfs_ops == 0) {
2031 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2032 		MNT_IUNLOCK(mp);
2033 		return;
2034 	}
2035 	if (c < 0)
2036 		vfs_dump_mount_counters(mp);
2037 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2038 		wakeup(&mp->mnt_writeopcount);
2039 	MNT_IUNLOCK(mp);
2040 }
2041 
2042 /*
2043  * Filesystem secondary write operation has completed. If we are
2044  * suspending and this operation is the last one, notify the suspender
2045  * that the suspension is now in effect.
2046  */
2047 void
2048 vn_finished_secondary_write(struct mount *mp)
2049 {
2050 	if (mp == NULL)
2051 		return;
2052 	MNT_ILOCK(mp);
2053 	MNT_REL(mp);
2054 	mp->mnt_secondary_writes--;
2055 	if (mp->mnt_secondary_writes < 0)
2056 		panic("vn_finished_secondary_write: neg cnt");
2057 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2058 	    mp->mnt_secondary_writes <= 0)
2059 		wakeup(&mp->mnt_secondary_writes);
2060 	MNT_IUNLOCK(mp);
2061 }
2062 
2063 /*
2064  * Request a filesystem to suspend write operations.
2065  */
2066 int
2067 vfs_write_suspend(struct mount *mp, int flags)
2068 {
2069 	int error;
2070 
2071 	vfs_op_enter(mp);
2072 
2073 	MNT_ILOCK(mp);
2074 	vfs_assert_mount_counters(mp);
2075 	if (mp->mnt_susp_owner == curthread) {
2076 		vfs_op_exit_locked(mp);
2077 		MNT_IUNLOCK(mp);
2078 		return (EALREADY);
2079 	}
2080 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2081 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2082 
2083 	/*
2084 	 * Unmount holds a write reference on the mount point.  If we
2085 	 * own busy reference and drain for writers, we deadlock with
2086 	 * the reference draining in the unmount path.  Callers of
2087 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2088 	 * vfs_busy() reference is owned and caller is not in the
2089 	 * unmount context.
2090 	 */
2091 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2092 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2093 		vfs_op_exit_locked(mp);
2094 		MNT_IUNLOCK(mp);
2095 		return (EBUSY);
2096 	}
2097 
2098 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2099 	mp->mnt_susp_owner = curthread;
2100 	if (mp->mnt_writeopcount > 0)
2101 		(void) msleep(&mp->mnt_writeopcount,
2102 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2103 	else
2104 		MNT_IUNLOCK(mp);
2105 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2106 		vfs_write_resume(mp, 0);
2107 		/* vfs_write_resume does vfs_op_exit() for us */
2108 	}
2109 	return (error);
2110 }
2111 
2112 /*
2113  * Request a filesystem to resume write operations.
2114  */
2115 void
2116 vfs_write_resume(struct mount *mp, int flags)
2117 {
2118 
2119 	MNT_ILOCK(mp);
2120 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2121 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2122 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2123 				       MNTK_SUSPENDED);
2124 		mp->mnt_susp_owner = NULL;
2125 		wakeup(&mp->mnt_writeopcount);
2126 		wakeup(&mp->mnt_flag);
2127 		curthread->td_pflags &= ~TDP_IGNSUSP;
2128 		if ((flags & VR_START_WRITE) != 0) {
2129 			MNT_REF(mp);
2130 			mp->mnt_writeopcount++;
2131 		}
2132 		MNT_IUNLOCK(mp);
2133 		if ((flags & VR_NO_SUSPCLR) == 0)
2134 			VFS_SUSP_CLEAN(mp);
2135 		vfs_op_exit(mp);
2136 	} else if ((flags & VR_START_WRITE) != 0) {
2137 		MNT_REF(mp);
2138 		vn_start_write_refed(mp, 0, true);
2139 	} else {
2140 		MNT_IUNLOCK(mp);
2141 	}
2142 }
2143 
2144 /*
2145  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2146  * methods.
2147  */
2148 int
2149 vfs_write_suspend_umnt(struct mount *mp)
2150 {
2151 	int error;
2152 
2153 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2154 	    ("vfs_write_suspend_umnt: recursed"));
2155 
2156 	/* dounmount() already called vn_start_write(). */
2157 	for (;;) {
2158 		vn_finished_write(mp);
2159 		error = vfs_write_suspend(mp, 0);
2160 		if (error != 0) {
2161 			vn_start_write(NULL, &mp, V_WAIT);
2162 			return (error);
2163 		}
2164 		MNT_ILOCK(mp);
2165 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2166 			break;
2167 		MNT_IUNLOCK(mp);
2168 		vn_start_write(NULL, &mp, V_WAIT);
2169 	}
2170 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2171 	wakeup(&mp->mnt_flag);
2172 	MNT_IUNLOCK(mp);
2173 	curthread->td_pflags |= TDP_IGNSUSP;
2174 	return (0);
2175 }
2176 
2177 /*
2178  * Implement kqueues for files by translating it to vnode operation.
2179  */
2180 static int
2181 vn_kqfilter(struct file *fp, struct knote *kn)
2182 {
2183 
2184 	return (VOP_KQFILTER(fp->f_vnode, kn));
2185 }
2186 
2187 int
2188 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2189 {
2190 	if ((fp->f_flag & FKQALLOWED) == 0)
2191 		return (EBADF);
2192 	return (vn_kqfilter(fp, kn));
2193 }
2194 
2195 /*
2196  * Simplified in-kernel wrapper calls for extended attribute access.
2197  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2198  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2199  */
2200 int
2201 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2202     const char *attrname, int *buflen, char *buf, struct thread *td)
2203 {
2204 	struct uio	auio;
2205 	struct iovec	iov;
2206 	int	error;
2207 
2208 	iov.iov_len = *buflen;
2209 	iov.iov_base = buf;
2210 
2211 	auio.uio_iov = &iov;
2212 	auio.uio_iovcnt = 1;
2213 	auio.uio_rw = UIO_READ;
2214 	auio.uio_segflg = UIO_SYSSPACE;
2215 	auio.uio_td = td;
2216 	auio.uio_offset = 0;
2217 	auio.uio_resid = *buflen;
2218 
2219 	if ((ioflg & IO_NODELOCKED) == 0)
2220 		vn_lock(vp, LK_SHARED | LK_RETRY);
2221 
2222 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2223 
2224 	/* authorize attribute retrieval as kernel */
2225 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2226 	    td);
2227 
2228 	if ((ioflg & IO_NODELOCKED) == 0)
2229 		VOP_UNLOCK(vp);
2230 
2231 	if (error == 0) {
2232 		*buflen = *buflen - auio.uio_resid;
2233 	}
2234 
2235 	return (error);
2236 }
2237 
2238 /*
2239  * XXX failure mode if partially written?
2240  */
2241 int
2242 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2243     const char *attrname, int buflen, char *buf, struct thread *td)
2244 {
2245 	struct uio	auio;
2246 	struct iovec	iov;
2247 	struct mount	*mp;
2248 	int	error;
2249 
2250 	iov.iov_len = buflen;
2251 	iov.iov_base = buf;
2252 
2253 	auio.uio_iov = &iov;
2254 	auio.uio_iovcnt = 1;
2255 	auio.uio_rw = UIO_WRITE;
2256 	auio.uio_segflg = UIO_SYSSPACE;
2257 	auio.uio_td = td;
2258 	auio.uio_offset = 0;
2259 	auio.uio_resid = buflen;
2260 
2261 	if ((ioflg & IO_NODELOCKED) == 0) {
2262 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2263 			return (error);
2264 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2265 	}
2266 
2267 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2268 
2269 	/* authorize attribute setting as kernel */
2270 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2271 
2272 	if ((ioflg & IO_NODELOCKED) == 0) {
2273 		vn_finished_write(mp);
2274 		VOP_UNLOCK(vp);
2275 	}
2276 
2277 	return (error);
2278 }
2279 
2280 int
2281 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2282     const char *attrname, struct thread *td)
2283 {
2284 	struct mount	*mp;
2285 	int	error;
2286 
2287 	if ((ioflg & IO_NODELOCKED) == 0) {
2288 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2289 			return (error);
2290 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2291 	}
2292 
2293 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2294 
2295 	/* authorize attribute removal as kernel */
2296 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2297 	if (error == EOPNOTSUPP)
2298 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2299 		    NULL, td);
2300 
2301 	if ((ioflg & IO_NODELOCKED) == 0) {
2302 		vn_finished_write(mp);
2303 		VOP_UNLOCK(vp);
2304 	}
2305 
2306 	return (error);
2307 }
2308 
2309 static int
2310 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2311     struct vnode **rvp)
2312 {
2313 
2314 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2315 }
2316 
2317 int
2318 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2319 {
2320 
2321 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2322 	    lkflags, rvp));
2323 }
2324 
2325 int
2326 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2327     int lkflags, struct vnode **rvp)
2328 {
2329 	struct mount *mp;
2330 	int ltype, error;
2331 
2332 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2333 	mp = vp->v_mount;
2334 	ltype = VOP_ISLOCKED(vp);
2335 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2336 	    ("vn_vget_ino: vp not locked"));
2337 	error = vfs_busy(mp, MBF_NOWAIT);
2338 	if (error != 0) {
2339 		vfs_ref(mp);
2340 		VOP_UNLOCK(vp);
2341 		error = vfs_busy(mp, 0);
2342 		vn_lock(vp, ltype | LK_RETRY);
2343 		vfs_rel(mp);
2344 		if (error != 0)
2345 			return (ENOENT);
2346 		if (VN_IS_DOOMED(vp)) {
2347 			vfs_unbusy(mp);
2348 			return (ENOENT);
2349 		}
2350 	}
2351 	VOP_UNLOCK(vp);
2352 	error = alloc(mp, alloc_arg, lkflags, rvp);
2353 	vfs_unbusy(mp);
2354 	if (error != 0 || *rvp != vp)
2355 		vn_lock(vp, ltype | LK_RETRY);
2356 	if (VN_IS_DOOMED(vp)) {
2357 		if (error == 0) {
2358 			if (*rvp == vp)
2359 				vunref(vp);
2360 			else
2361 				vput(*rvp);
2362 		}
2363 		error = ENOENT;
2364 	}
2365 	return (error);
2366 }
2367 
2368 int
2369 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2370     struct thread *td)
2371 {
2372 	off_t lim;
2373 	bool ktr_write;
2374 
2375 	if (td == NULL)
2376 		return (0);
2377 
2378 	/*
2379 	 * There are conditions where the limit is to be ignored.
2380 	 * However, since it is almost never reached, check it first.
2381 	 */
2382 	ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2383 	lim = lim_cur(td, RLIMIT_FSIZE);
2384 	if (__predict_false(ktr_write))
2385 		lim = td->td_ktr_io_lim;
2386 	if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2387 		return (0);
2388 
2389 	/*
2390 	 * The limit is reached.
2391 	 */
2392 	if (vp->v_type != VREG ||
2393 	    (td->td_pflags2 & TDP2_ACCT) != 0)
2394 		return (0);
2395 
2396 	if (!ktr_write || ktr_filesize_limit_signal) {
2397 		PROC_LOCK(td->td_proc);
2398 		kern_psignal(td->td_proc, SIGXFSZ);
2399 		PROC_UNLOCK(td->td_proc);
2400 	}
2401 	return (EFBIG);
2402 }
2403 
2404 int
2405 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2406     struct thread *td)
2407 {
2408 	struct vnode *vp;
2409 
2410 	vp = fp->f_vnode;
2411 #ifdef AUDIT
2412 	vn_lock(vp, LK_SHARED | LK_RETRY);
2413 	AUDIT_ARG_VNODE1(vp);
2414 	VOP_UNLOCK(vp);
2415 #endif
2416 	return (setfmode(td, active_cred, vp, mode));
2417 }
2418 
2419 int
2420 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2421     struct thread *td)
2422 {
2423 	struct vnode *vp;
2424 
2425 	vp = fp->f_vnode;
2426 #ifdef AUDIT
2427 	vn_lock(vp, LK_SHARED | LK_RETRY);
2428 	AUDIT_ARG_VNODE1(vp);
2429 	VOP_UNLOCK(vp);
2430 #endif
2431 	return (setfown(td, active_cred, vp, uid, gid));
2432 }
2433 
2434 /*
2435  * Remove pages in the range ["start", "end") from the vnode's VM object.  If
2436  * "end" is 0, then the range extends to the end of the object.
2437  */
2438 void
2439 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2440 {
2441 	vm_object_t object;
2442 
2443 	if ((object = vp->v_object) == NULL)
2444 		return;
2445 	VM_OBJECT_WLOCK(object);
2446 	vm_object_page_remove(object, start, end, 0);
2447 	VM_OBJECT_WUNLOCK(object);
2448 }
2449 
2450 /*
2451  * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2452  * mapped into any process' address space.  Filesystems may use this in
2453  * preference to vn_pages_remove() to avoid blocking on pages busied in
2454  * preparation for a VOP_GETPAGES.
2455  */
2456 void
2457 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2458 {
2459 	vm_object_t object;
2460 
2461 	if ((object = vp->v_object) == NULL)
2462 		return;
2463 	VM_OBJECT_WLOCK(object);
2464 	vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2465 	VM_OBJECT_WUNLOCK(object);
2466 }
2467 
2468 int
2469 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2470     struct ucred *cred)
2471 {
2472 	struct vattr va;
2473 	daddr_t bn, bnp;
2474 	uint64_t bsize;
2475 	off_t noff;
2476 	int error;
2477 
2478 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2479 	    ("%s: Wrong command %lu", __func__, cmd));
2480 	ASSERT_VOP_LOCKED(vp, "vn_bmap_seekhole_locked");
2481 
2482 	if (vp->v_type != VREG) {
2483 		error = ENOTTY;
2484 		goto out;
2485 	}
2486 	error = VOP_GETATTR(vp, &va, cred);
2487 	if (error != 0)
2488 		goto out;
2489 	noff = *off;
2490 	if (noff >= va.va_size) {
2491 		error = ENXIO;
2492 		goto out;
2493 	}
2494 	bsize = vp->v_mount->mnt_stat.f_iosize;
2495 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2496 	    noff % bsize) {
2497 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2498 		if (error == EOPNOTSUPP) {
2499 			error = ENOTTY;
2500 			goto out;
2501 		}
2502 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2503 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2504 			noff = bn * bsize;
2505 			if (noff < *off)
2506 				noff = *off;
2507 			goto out;
2508 		}
2509 	}
2510 	if (noff > va.va_size)
2511 		noff = va.va_size;
2512 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2513 	if (cmd == FIOSEEKDATA)
2514 		error = ENXIO;
2515 out:
2516 	if (error == 0)
2517 		*off = noff;
2518 	return (error);
2519 }
2520 
2521 int
2522 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2523 {
2524 	int error;
2525 
2526 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2527 	    ("%s: Wrong command %lu", __func__, cmd));
2528 
2529 	if (vn_lock(vp, LK_SHARED) != 0)
2530 		return (EBADF);
2531 	error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2532 	VOP_UNLOCK(vp);
2533 	return (error);
2534 }
2535 
2536 int
2537 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2538 {
2539 	struct ucred *cred;
2540 	struct vnode *vp;
2541 	struct vattr vattr;
2542 	off_t foffset, size;
2543 	int error, noneg;
2544 
2545 	cred = td->td_ucred;
2546 	vp = fp->f_vnode;
2547 	foffset = foffset_lock(fp, 0);
2548 	noneg = (vp->v_type != VCHR);
2549 	error = 0;
2550 	switch (whence) {
2551 	case L_INCR:
2552 		if (noneg &&
2553 		    (foffset < 0 ||
2554 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2555 			error = EOVERFLOW;
2556 			break;
2557 		}
2558 		offset += foffset;
2559 		break;
2560 	case L_XTND:
2561 		vn_lock(vp, LK_SHARED | LK_RETRY);
2562 		error = VOP_GETATTR(vp, &vattr, cred);
2563 		VOP_UNLOCK(vp);
2564 		if (error)
2565 			break;
2566 
2567 		/*
2568 		 * If the file references a disk device, then fetch
2569 		 * the media size and use that to determine the ending
2570 		 * offset.
2571 		 */
2572 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2573 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2574 			vattr.va_size = size;
2575 		if (noneg &&
2576 		    (vattr.va_size > OFF_MAX ||
2577 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2578 			error = EOVERFLOW;
2579 			break;
2580 		}
2581 		offset += vattr.va_size;
2582 		break;
2583 	case L_SET:
2584 		break;
2585 	case SEEK_DATA:
2586 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2587 		if (error == ENOTTY)
2588 			error = EINVAL;
2589 		break;
2590 	case SEEK_HOLE:
2591 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2592 		if (error == ENOTTY)
2593 			error = EINVAL;
2594 		break;
2595 	default:
2596 		error = EINVAL;
2597 	}
2598 	if (error == 0 && noneg && offset < 0)
2599 		error = EINVAL;
2600 	if (error != 0)
2601 		goto drop;
2602 	VFS_KNOTE_UNLOCKED(vp, 0);
2603 	td->td_uretoff.tdu_off = offset;
2604 drop:
2605 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2606 	return (error);
2607 }
2608 
2609 int
2610 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2611     struct thread *td)
2612 {
2613 	int error;
2614 
2615 	/*
2616 	 * Grant permission if the caller is the owner of the file, or
2617 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2618 	 * on the file.  If the time pointer is null, then write
2619 	 * permission on the file is also sufficient.
2620 	 *
2621 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2622 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2623 	 * will be allowed to set the times [..] to the current
2624 	 * server time.
2625 	 */
2626 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2627 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2628 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2629 	return (error);
2630 }
2631 
2632 int
2633 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2634 {
2635 	struct vnode *vp;
2636 	int error;
2637 
2638 	if (fp->f_type == DTYPE_FIFO)
2639 		kif->kf_type = KF_TYPE_FIFO;
2640 	else
2641 		kif->kf_type = KF_TYPE_VNODE;
2642 	vp = fp->f_vnode;
2643 	vref(vp);
2644 	FILEDESC_SUNLOCK(fdp);
2645 	error = vn_fill_kinfo_vnode(vp, kif);
2646 	vrele(vp);
2647 	FILEDESC_SLOCK(fdp);
2648 	return (error);
2649 }
2650 
2651 static inline void
2652 vn_fill_junk(struct kinfo_file *kif)
2653 {
2654 	size_t len, olen;
2655 
2656 	/*
2657 	 * Simulate vn_fullpath returning changing values for a given
2658 	 * vp during e.g. coredump.
2659 	 */
2660 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2661 	olen = strlen(kif->kf_path);
2662 	if (len < olen)
2663 		strcpy(&kif->kf_path[len - 1], "$");
2664 	else
2665 		for (; olen < len; olen++)
2666 			strcpy(&kif->kf_path[olen], "A");
2667 }
2668 
2669 int
2670 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2671 {
2672 	struct vattr va;
2673 	char *fullpath, *freepath;
2674 	int error;
2675 
2676 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2677 	freepath = NULL;
2678 	fullpath = "-";
2679 	error = vn_fullpath(vp, &fullpath, &freepath);
2680 	if (error == 0) {
2681 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2682 	}
2683 	if (freepath != NULL)
2684 		free(freepath, M_TEMP);
2685 
2686 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2687 		vn_fill_junk(kif);
2688 	);
2689 
2690 	/*
2691 	 * Retrieve vnode attributes.
2692 	 */
2693 	va.va_fsid = VNOVAL;
2694 	va.va_rdev = NODEV;
2695 	vn_lock(vp, LK_SHARED | LK_RETRY);
2696 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2697 	VOP_UNLOCK(vp);
2698 	if (error != 0)
2699 		return (error);
2700 	if (va.va_fsid != VNOVAL)
2701 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2702 	else
2703 		kif->kf_un.kf_file.kf_file_fsid =
2704 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2705 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2706 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2707 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2708 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2709 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2710 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2711 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2712 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2713 	return (0);
2714 }
2715 
2716 int
2717 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2718     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2719     struct thread *td)
2720 {
2721 #ifdef HWPMC_HOOKS
2722 	struct pmckern_map_in pkm;
2723 #endif
2724 	struct mount *mp;
2725 	struct vnode *vp;
2726 	vm_object_t object;
2727 	vm_prot_t maxprot;
2728 	boolean_t writecounted;
2729 	int error;
2730 
2731 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2732     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2733 	/*
2734 	 * POSIX shared-memory objects are defined to have
2735 	 * kernel persistence, and are not defined to support
2736 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2737 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2738 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2739 	 * flag to request this behavior.
2740 	 */
2741 	if ((fp->f_flag & FPOSIXSHM) != 0)
2742 		flags |= MAP_NOSYNC;
2743 #endif
2744 	vp = fp->f_vnode;
2745 
2746 	/*
2747 	 * Ensure that file and memory protections are
2748 	 * compatible.  Note that we only worry about
2749 	 * writability if mapping is shared; in this case,
2750 	 * current and max prot are dictated by the open file.
2751 	 * XXX use the vnode instead?  Problem is: what
2752 	 * credentials do we use for determination? What if
2753 	 * proc does a setuid?
2754 	 */
2755 	mp = vp->v_mount;
2756 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2757 		maxprot = VM_PROT_NONE;
2758 		if ((prot & VM_PROT_EXECUTE) != 0)
2759 			return (EACCES);
2760 	} else
2761 		maxprot = VM_PROT_EXECUTE;
2762 	if ((fp->f_flag & FREAD) != 0)
2763 		maxprot |= VM_PROT_READ;
2764 	else if ((prot & VM_PROT_READ) != 0)
2765 		return (EACCES);
2766 
2767 	/*
2768 	 * If we are sharing potential changes via MAP_SHARED and we
2769 	 * are trying to get write permission although we opened it
2770 	 * without asking for it, bail out.
2771 	 */
2772 	if ((flags & MAP_SHARED) != 0) {
2773 		if ((fp->f_flag & FWRITE) != 0)
2774 			maxprot |= VM_PROT_WRITE;
2775 		else if ((prot & VM_PROT_WRITE) != 0)
2776 			return (EACCES);
2777 	} else {
2778 		maxprot |= VM_PROT_WRITE;
2779 		cap_maxprot |= VM_PROT_WRITE;
2780 	}
2781 	maxprot &= cap_maxprot;
2782 
2783 	/*
2784 	 * For regular files and shared memory, POSIX requires that
2785 	 * the value of foff be a legitimate offset within the data
2786 	 * object.  In particular, negative offsets are invalid.
2787 	 * Blocking negative offsets and overflows here avoids
2788 	 * possible wraparound or user-level access into reserved
2789 	 * ranges of the data object later.  In contrast, POSIX does
2790 	 * not dictate how offsets are used by device drivers, so in
2791 	 * the case of a device mapping a negative offset is passed
2792 	 * on.
2793 	 */
2794 	if (
2795 #ifdef _LP64
2796 	    size > OFF_MAX ||
2797 #endif
2798 	    foff > OFF_MAX - size)
2799 		return (EINVAL);
2800 
2801 	writecounted = FALSE;
2802 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2803 	    &foff, &object, &writecounted);
2804 	if (error != 0)
2805 		return (error);
2806 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2807 	    foff, writecounted, td);
2808 	if (error != 0) {
2809 		/*
2810 		 * If this mapping was accounted for in the vnode's
2811 		 * writecount, then undo that now.
2812 		 */
2813 		if (writecounted)
2814 			vm_pager_release_writecount(object, 0, size);
2815 		vm_object_deallocate(object);
2816 	}
2817 #ifdef HWPMC_HOOKS
2818 	/* Inform hwpmc(4) if an executable is being mapped. */
2819 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2820 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2821 			pkm.pm_file = vp;
2822 			pkm.pm_address = (uintptr_t) *addr;
2823 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2824 		}
2825 	}
2826 #endif
2827 	return (error);
2828 }
2829 
2830 void
2831 vn_fsid(struct vnode *vp, struct vattr *va)
2832 {
2833 	fsid_t *f;
2834 
2835 	f = &vp->v_mount->mnt_stat.f_fsid;
2836 	va->va_fsid = (uint32_t)f->val[1];
2837 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2838 	va->va_fsid += (uint32_t)f->val[0];
2839 }
2840 
2841 int
2842 vn_fsync_buf(struct vnode *vp, int waitfor)
2843 {
2844 	struct buf *bp, *nbp;
2845 	struct bufobj *bo;
2846 	struct mount *mp;
2847 	int error, maxretry;
2848 
2849 	error = 0;
2850 	maxretry = 10000;     /* large, arbitrarily chosen */
2851 	mp = NULL;
2852 	if (vp->v_type == VCHR) {
2853 		VI_LOCK(vp);
2854 		mp = vp->v_rdev->si_mountpt;
2855 		VI_UNLOCK(vp);
2856 	}
2857 	bo = &vp->v_bufobj;
2858 	BO_LOCK(bo);
2859 loop1:
2860 	/*
2861 	 * MARK/SCAN initialization to avoid infinite loops.
2862 	 */
2863         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2864 		bp->b_vflags &= ~BV_SCANNED;
2865 		bp->b_error = 0;
2866 	}
2867 
2868 	/*
2869 	 * Flush all dirty buffers associated with a vnode.
2870 	 */
2871 loop2:
2872 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2873 		if ((bp->b_vflags & BV_SCANNED) != 0)
2874 			continue;
2875 		bp->b_vflags |= BV_SCANNED;
2876 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2877 			if (waitfor != MNT_WAIT)
2878 				continue;
2879 			if (BUF_LOCK(bp,
2880 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2881 			    BO_LOCKPTR(bo)) != 0) {
2882 				BO_LOCK(bo);
2883 				goto loop1;
2884 			}
2885 			BO_LOCK(bo);
2886 		}
2887 		BO_UNLOCK(bo);
2888 		KASSERT(bp->b_bufobj == bo,
2889 		    ("bp %p wrong b_bufobj %p should be %p",
2890 		    bp, bp->b_bufobj, bo));
2891 		if ((bp->b_flags & B_DELWRI) == 0)
2892 			panic("fsync: not dirty");
2893 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2894 			vfs_bio_awrite(bp);
2895 		} else {
2896 			bremfree(bp);
2897 			bawrite(bp);
2898 		}
2899 		if (maxretry < 1000)
2900 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2901 		BO_LOCK(bo);
2902 		goto loop2;
2903 	}
2904 
2905 	/*
2906 	 * If synchronous the caller expects us to completely resolve all
2907 	 * dirty buffers in the system.  Wait for in-progress I/O to
2908 	 * complete (which could include background bitmap writes), then
2909 	 * retry if dirty blocks still exist.
2910 	 */
2911 	if (waitfor == MNT_WAIT) {
2912 		bufobj_wwait(bo, 0, 0);
2913 		if (bo->bo_dirty.bv_cnt > 0) {
2914 			/*
2915 			 * If we are unable to write any of these buffers
2916 			 * then we fail now rather than trying endlessly
2917 			 * to write them out.
2918 			 */
2919 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2920 				if ((error = bp->b_error) != 0)
2921 					break;
2922 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2923 			    (error == 0 && --maxretry >= 0))
2924 				goto loop1;
2925 			if (error == 0)
2926 				error = EAGAIN;
2927 		}
2928 	}
2929 	BO_UNLOCK(bo);
2930 	if (error != 0)
2931 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2932 
2933 	return (error);
2934 }
2935 
2936 /*
2937  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2938  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2939  * to do the actual copy.
2940  * vn_generic_copy_file_range() is factored out, so it can be called
2941  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2942  * different file systems.
2943  */
2944 int
2945 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2946     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2947     struct ucred *outcred, struct thread *fsize_td)
2948 {
2949 	int error;
2950 	size_t len;
2951 	uint64_t uval;
2952 
2953 	len = *lenp;
2954 	*lenp = 0;		/* For error returns. */
2955 	error = 0;
2956 
2957 	/* Do some sanity checks on the arguments. */
2958 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2959 		error = EISDIR;
2960 	else if (*inoffp < 0 || *outoffp < 0 ||
2961 	    invp->v_type != VREG || outvp->v_type != VREG)
2962 		error = EINVAL;
2963 	if (error != 0)
2964 		goto out;
2965 
2966 	/* Ensure offset + len does not wrap around. */
2967 	uval = *inoffp;
2968 	uval += len;
2969 	if (uval > INT64_MAX)
2970 		len = INT64_MAX - *inoffp;
2971 	uval = *outoffp;
2972 	uval += len;
2973 	if (uval > INT64_MAX)
2974 		len = INT64_MAX - *outoffp;
2975 	if (len == 0)
2976 		goto out;
2977 
2978 	/*
2979 	 * If the two vnode are for the same file system, call
2980 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2981 	 * which can handle copies across multiple file systems.
2982 	 */
2983 	*lenp = len;
2984 	if (invp->v_mount == outvp->v_mount)
2985 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2986 		    lenp, flags, incred, outcred, fsize_td);
2987 	else
2988 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2989 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2990 out:
2991 	return (error);
2992 }
2993 
2994 /*
2995  * Test len bytes of data starting at dat for all bytes == 0.
2996  * Return true if all bytes are zero, false otherwise.
2997  * Expects dat to be well aligned.
2998  */
2999 static bool
3000 mem_iszero(void *dat, int len)
3001 {
3002 	int i;
3003 	const u_int *p;
3004 	const char *cp;
3005 
3006 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
3007 		if (len >= sizeof(*p)) {
3008 			if (*p != 0)
3009 				return (false);
3010 		} else {
3011 			cp = (const char *)p;
3012 			for (i = 0; i < len; i++, cp++)
3013 				if (*cp != '\0')
3014 					return (false);
3015 		}
3016 	}
3017 	return (true);
3018 }
3019 
3020 /*
3021  * Look for a hole in the output file and, if found, adjust *outoffp
3022  * and *xferp to skip past the hole.
3023  * *xferp is the entire hole length to be written and xfer2 is how many bytes
3024  * to be written as 0's upon return.
3025  */
3026 static off_t
3027 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3028     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3029 {
3030 	int error;
3031 	off_t delta;
3032 
3033 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3034 		*dataoffp = *outoffp;
3035 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3036 		    curthread);
3037 		if (error == 0) {
3038 			*holeoffp = *dataoffp;
3039 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3040 			    curthread);
3041 		}
3042 		if (error != 0 || *holeoffp == *dataoffp) {
3043 			/*
3044 			 * Since outvp is unlocked, it may be possible for
3045 			 * another thread to do a truncate(), lseek(), write()
3046 			 * creating a hole at startoff between the above
3047 			 * VOP_IOCTL() calls, if the other thread does not do
3048 			 * rangelocking.
3049 			 * If that happens, *holeoffp == *dataoffp and finding
3050 			 * the hole has failed, so disable vn_skip_hole().
3051 			 */
3052 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
3053 			return (xfer2);
3054 		}
3055 		KASSERT(*dataoffp >= *outoffp,
3056 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3057 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
3058 		KASSERT(*holeoffp > *dataoffp,
3059 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3060 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3061 	}
3062 
3063 	/*
3064 	 * If there is a hole before the data starts, advance *outoffp and
3065 	 * *xferp past the hole.
3066 	 */
3067 	if (*dataoffp > *outoffp) {
3068 		delta = *dataoffp - *outoffp;
3069 		if (delta >= *xferp) {
3070 			/* Entire *xferp is a hole. */
3071 			*outoffp += *xferp;
3072 			*xferp = 0;
3073 			return (0);
3074 		}
3075 		*xferp -= delta;
3076 		*outoffp += delta;
3077 		xfer2 = MIN(xfer2, *xferp);
3078 	}
3079 
3080 	/*
3081 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3082 	 * that the write ends at the start of the hole.
3083 	 * *holeoffp should always be greater than *outoffp, but for the
3084 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3085 	 * value.
3086 	 */
3087 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3088 		xfer2 = *holeoffp - *outoffp;
3089 	return (xfer2);
3090 }
3091 
3092 /*
3093  * Write an xfer sized chunk to outvp in blksize blocks from dat.
3094  * dat is a maximum of blksize in length and can be written repeatedly in
3095  * the chunk.
3096  * If growfile == true, just grow the file via vn_truncate_locked() instead
3097  * of doing actual writes.
3098  * If checkhole == true, a hole is being punched, so skip over any hole
3099  * already in the output file.
3100  */
3101 static int
3102 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3103     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3104 {
3105 	struct mount *mp;
3106 	off_t dataoff, holeoff, xfer2;
3107 	int error;
3108 
3109 	/*
3110 	 * Loop around doing writes of blksize until write has been completed.
3111 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3112 	 * done for each iteration, since the xfer argument can be very
3113 	 * large if there is a large hole to punch in the output file.
3114 	 */
3115 	error = 0;
3116 	holeoff = 0;
3117 	do {
3118 		xfer2 = MIN(xfer, blksize);
3119 		if (checkhole) {
3120 			/*
3121 			 * Punching a hole.  Skip writing if there is
3122 			 * already a hole in the output file.
3123 			 */
3124 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3125 			    &dataoff, &holeoff, cred);
3126 			if (xfer == 0)
3127 				break;
3128 			if (holeoff < 0)
3129 				checkhole = false;
3130 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3131 			    (intmax_t)xfer2));
3132 		}
3133 		bwillwrite();
3134 		mp = NULL;
3135 		error = vn_start_write(outvp, &mp, V_WAIT);
3136 		if (error != 0)
3137 			break;
3138 		if (growfile) {
3139 			error = vn_lock(outvp, LK_EXCLUSIVE);
3140 			if (error == 0) {
3141 				error = vn_truncate_locked(outvp, outoff + xfer,
3142 				    false, cred);
3143 				VOP_UNLOCK(outvp);
3144 			}
3145 		} else {
3146 			error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3147 			if (error == 0) {
3148 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3149 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3150 				    curthread->td_ucred, cred, NULL, curthread);
3151 				outoff += xfer2;
3152 				xfer -= xfer2;
3153 				VOP_UNLOCK(outvp);
3154 			}
3155 		}
3156 		if (mp != NULL)
3157 			vn_finished_write(mp);
3158 	} while (!growfile && xfer > 0 && error == 0);
3159 	return (error);
3160 }
3161 
3162 /*
3163  * Copy a byte range of one file to another.  This function can handle the
3164  * case where invp and outvp are on different file systems.
3165  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3166  * is no better file system specific way to do it.
3167  */
3168 int
3169 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3170     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3171     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3172 {
3173 	struct vattr va, inva;
3174 	struct mount *mp;
3175 	struct uio io;
3176 	off_t startoff, endoff, xfer, xfer2;
3177 	u_long blksize;
3178 	int error, interrupted;
3179 	bool cantseek, readzeros, eof, lastblock, holetoeof;
3180 	ssize_t aresid;
3181 	size_t copylen, len, rem, savlen;
3182 	char *dat;
3183 	long holein, holeout;
3184 	struct timespec curts, endts;
3185 
3186 	holein = holeout = 0;
3187 	savlen = len = *lenp;
3188 	error = 0;
3189 	interrupted = 0;
3190 	dat = NULL;
3191 
3192 	error = vn_lock(invp, LK_SHARED);
3193 	if (error != 0)
3194 		goto out;
3195 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3196 		holein = 0;
3197 	if (holein > 0)
3198 		error = VOP_GETATTR(invp, &inva, incred);
3199 	VOP_UNLOCK(invp);
3200 	if (error != 0)
3201 		goto out;
3202 
3203 	mp = NULL;
3204 	error = vn_start_write(outvp, &mp, V_WAIT);
3205 	if (error == 0)
3206 		error = vn_lock(outvp, LK_EXCLUSIVE);
3207 	if (error == 0) {
3208 		/*
3209 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3210 		 * now that outvp is locked.
3211 		 */
3212 		if (fsize_td != NULL) {
3213 			io.uio_offset = *outoffp;
3214 			io.uio_resid = len;
3215 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3216 			if (error != 0)
3217 				error = EFBIG;
3218 		}
3219 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3220 			holeout = 0;
3221 		/*
3222 		 * Holes that are past EOF do not need to be written as a block
3223 		 * of zero bytes.  So, truncate the output file as far as
3224 		 * possible and then use va.va_size to decide if writing 0
3225 		 * bytes is necessary in the loop below.
3226 		 */
3227 		if (error == 0)
3228 			error = VOP_GETATTR(outvp, &va, outcred);
3229 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3230 		    *outoffp + len) {
3231 #ifdef MAC
3232 			error = mac_vnode_check_write(curthread->td_ucred,
3233 			    outcred, outvp);
3234 			if (error == 0)
3235 #endif
3236 				error = vn_truncate_locked(outvp, *outoffp,
3237 				    false, outcred);
3238 			if (error == 0)
3239 				va.va_size = *outoffp;
3240 		}
3241 		VOP_UNLOCK(outvp);
3242 	}
3243 	if (mp != NULL)
3244 		vn_finished_write(mp);
3245 	if (error != 0)
3246 		goto out;
3247 
3248 	/*
3249 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3250 	 * If hole sizes aren't available, set the blksize to the larger
3251 	 * f_iosize of invp and outvp.
3252 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3253 	 * This value is clipped at 4Kbytes and 1Mbyte.
3254 	 */
3255 	blksize = MAX(holein, holeout);
3256 
3257 	/* Clip len to end at an exact multiple of hole size. */
3258 	if (blksize > 1) {
3259 		rem = *inoffp % blksize;
3260 		if (rem > 0)
3261 			rem = blksize - rem;
3262 		if (len > rem && len - rem > blksize)
3263 			len = savlen = rounddown(len - rem, blksize) + rem;
3264 	}
3265 
3266 	if (blksize <= 1)
3267 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3268 		    outvp->v_mount->mnt_stat.f_iosize);
3269 	if (blksize < 4096)
3270 		blksize = 4096;
3271 	else if (blksize > 1024 * 1024)
3272 		blksize = 1024 * 1024;
3273 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3274 
3275 	/*
3276 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3277 	 * to find holes.  Otherwise, just scan the read block for all 0s
3278 	 * in the inner loop where the data copying is done.
3279 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3280 	 * support holes on the server, but do not support FIOSEEKHOLE.
3281 	 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3282 	 * that this function should return after 1second with a partial
3283 	 * completion.
3284 	 */
3285 	if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3286 		getnanouptime(&endts);
3287 		endts.tv_sec++;
3288 	} else
3289 		timespecclear(&endts);
3290 	holetoeof = eof = false;
3291 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3292 		endoff = 0;			/* To shut up compilers. */
3293 		cantseek = true;
3294 		startoff = *inoffp;
3295 		copylen = len;
3296 
3297 		/*
3298 		 * Find the next data area.  If there is just a hole to EOF,
3299 		 * FIOSEEKDATA should fail with ENXIO.
3300 		 * (I do not know if any file system will report a hole to
3301 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3302 		 *  will fail for those file systems.)
3303 		 *
3304 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3305 		 * the code just falls through to the inner copy loop.
3306 		 */
3307 		error = EINVAL;
3308 		if (holein > 0) {
3309 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3310 			    incred, curthread);
3311 			if (error == ENXIO) {
3312 				startoff = endoff = inva.va_size;
3313 				eof = holetoeof = true;
3314 				error = 0;
3315 			}
3316 		}
3317 		if (error == 0 && !holetoeof) {
3318 			endoff = startoff;
3319 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3320 			    incred, curthread);
3321 			/*
3322 			 * Since invp is unlocked, it may be possible for
3323 			 * another thread to do a truncate(), lseek(), write()
3324 			 * creating a hole at startoff between the above
3325 			 * VOP_IOCTL() calls, if the other thread does not do
3326 			 * rangelocking.
3327 			 * If that happens, startoff == endoff and finding
3328 			 * the hole has failed, so set an error.
3329 			 */
3330 			if (error == 0 && startoff == endoff)
3331 				error = EINVAL; /* Any error. Reset to 0. */
3332 		}
3333 		if (error == 0) {
3334 			if (startoff > *inoffp) {
3335 				/* Found hole before data block. */
3336 				xfer = MIN(startoff - *inoffp, len);
3337 				if (*outoffp < va.va_size) {
3338 					/* Must write 0s to punch hole. */
3339 					xfer2 = MIN(va.va_size - *outoffp,
3340 					    xfer);
3341 					memset(dat, 0, MIN(xfer2, blksize));
3342 					error = vn_write_outvp(outvp, dat,
3343 					    *outoffp, xfer2, blksize, false,
3344 					    holeout > 0, outcred);
3345 				}
3346 
3347 				if (error == 0 && *outoffp + xfer >
3348 				    va.va_size && (xfer == len || holetoeof)) {
3349 					/* Grow output file (hole at end). */
3350 					error = vn_write_outvp(outvp, dat,
3351 					    *outoffp, xfer, blksize, true,
3352 					    false, outcred);
3353 				}
3354 				if (error == 0) {
3355 					*inoffp += xfer;
3356 					*outoffp += xfer;
3357 					len -= xfer;
3358 					if (len < savlen) {
3359 						interrupted = sig_intr();
3360 						if (timespecisset(&endts) &&
3361 						    interrupted == 0) {
3362 							getnanouptime(&curts);
3363 							if (timespeccmp(&curts,
3364 							    &endts, >=))
3365 								interrupted =
3366 								    EINTR;
3367 						}
3368 					}
3369 				}
3370 			}
3371 			copylen = MIN(len, endoff - startoff);
3372 			cantseek = false;
3373 		} else {
3374 			cantseek = true;
3375 			startoff = *inoffp;
3376 			copylen = len;
3377 			error = 0;
3378 		}
3379 
3380 		xfer = blksize;
3381 		if (cantseek) {
3382 			/*
3383 			 * Set first xfer to end at a block boundary, so that
3384 			 * holes are more likely detected in the loop below via
3385 			 * the for all bytes 0 method.
3386 			 */
3387 			xfer -= (*inoffp % blksize);
3388 		}
3389 		/* Loop copying the data block. */
3390 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3391 			if (copylen < xfer)
3392 				xfer = copylen;
3393 			error = vn_lock(invp, LK_SHARED);
3394 			if (error != 0)
3395 				goto out;
3396 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3397 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3398 			    curthread->td_ucred, incred, &aresid,
3399 			    curthread);
3400 			VOP_UNLOCK(invp);
3401 			lastblock = false;
3402 			if (error == 0 && aresid > 0) {
3403 				/* Stop the copy at EOF on the input file. */
3404 				xfer -= aresid;
3405 				eof = true;
3406 				lastblock = true;
3407 			}
3408 			if (error == 0) {
3409 				/*
3410 				 * Skip the write for holes past the initial EOF
3411 				 * of the output file, unless this is the last
3412 				 * write of the output file at EOF.
3413 				 */
3414 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3415 				    false;
3416 				if (xfer == len)
3417 					lastblock = true;
3418 				if (!cantseek || *outoffp < va.va_size ||
3419 				    lastblock || !readzeros)
3420 					error = vn_write_outvp(outvp, dat,
3421 					    *outoffp, xfer, blksize,
3422 					    readzeros && lastblock &&
3423 					    *outoffp >= va.va_size, false,
3424 					    outcred);
3425 				if (error == 0) {
3426 					*inoffp += xfer;
3427 					startoff += xfer;
3428 					*outoffp += xfer;
3429 					copylen -= xfer;
3430 					len -= xfer;
3431 					if (len < savlen) {
3432 						interrupted = sig_intr();
3433 						if (timespecisset(&endts) &&
3434 						    interrupted == 0) {
3435 							getnanouptime(&curts);
3436 							if (timespeccmp(&curts,
3437 							    &endts, >=))
3438 								interrupted =
3439 								    EINTR;
3440 						}
3441 					}
3442 				}
3443 			}
3444 			xfer = blksize;
3445 		}
3446 	}
3447 out:
3448 	*lenp = savlen - len;
3449 	free(dat, M_TEMP);
3450 	return (error);
3451 }
3452 
3453 static int
3454 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3455 {
3456 	struct mount *mp;
3457 	struct vnode *vp;
3458 	off_t olen, ooffset;
3459 	int error;
3460 #ifdef AUDIT
3461 	int audited_vnode1 = 0;
3462 #endif
3463 
3464 	vp = fp->f_vnode;
3465 	if (vp->v_type != VREG)
3466 		return (ENODEV);
3467 
3468 	/* Allocating blocks may take a long time, so iterate. */
3469 	for (;;) {
3470 		olen = len;
3471 		ooffset = offset;
3472 
3473 		bwillwrite();
3474 		mp = NULL;
3475 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3476 		if (error != 0)
3477 			break;
3478 		error = vn_lock(vp, LK_EXCLUSIVE);
3479 		if (error != 0) {
3480 			vn_finished_write(mp);
3481 			break;
3482 		}
3483 #ifdef AUDIT
3484 		if (!audited_vnode1) {
3485 			AUDIT_ARG_VNODE1(vp);
3486 			audited_vnode1 = 1;
3487 		}
3488 #endif
3489 #ifdef MAC
3490 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3491 		if (error == 0)
3492 #endif
3493 			error = VOP_ALLOCATE(vp, &offset, &len, 0,
3494 			    td->td_ucred);
3495 		VOP_UNLOCK(vp);
3496 		vn_finished_write(mp);
3497 
3498 		if (olen + ooffset != offset + len) {
3499 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3500 			    ooffset, olen, offset, len);
3501 		}
3502 		if (error != 0 || len == 0)
3503 			break;
3504 		KASSERT(olen > len, ("Iteration did not make progress?"));
3505 		maybe_yield();
3506 	}
3507 
3508 	return (error);
3509 }
3510 
3511 static int
3512 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3513     int ioflag, struct ucred *cred, struct ucred *active_cred,
3514     struct ucred *file_cred)
3515 {
3516 	struct mount *mp;
3517 	void *rl_cookie;
3518 	off_t off, len;
3519 	int error;
3520 #ifdef AUDIT
3521 	bool audited_vnode1 = false;
3522 #endif
3523 
3524 	rl_cookie = NULL;
3525 	error = 0;
3526 	mp = NULL;
3527 	off = *offset;
3528 	len = *length;
3529 
3530 	if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3531 		rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3532 	while (len > 0 && error == 0) {
3533 		/*
3534 		 * Try to deallocate the longest range in one pass.
3535 		 * In case a pass takes too long to be executed, it returns
3536 		 * partial result. The residue will be proceeded in the next
3537 		 * pass.
3538 		 */
3539 
3540 		if ((ioflag & IO_NODELOCKED) == 0) {
3541 			bwillwrite();
3542 			if ((error = vn_start_write(vp, &mp,
3543 			    V_WAIT | PCATCH)) != 0)
3544 				goto out;
3545 			vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3546 		}
3547 #ifdef AUDIT
3548 		if (!audited_vnode1) {
3549 			AUDIT_ARG_VNODE1(vp);
3550 			audited_vnode1 = true;
3551 		}
3552 #endif
3553 
3554 #ifdef MAC
3555 		if ((ioflag & IO_NOMACCHECK) == 0)
3556 			error = mac_vnode_check_write(active_cred, file_cred,
3557 			    vp);
3558 #endif
3559 		if (error == 0)
3560 			error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3561 			    cred);
3562 
3563 		if ((ioflag & IO_NODELOCKED) == 0) {
3564 			VOP_UNLOCK(vp);
3565 			if (mp != NULL) {
3566 				vn_finished_write(mp);
3567 				mp = NULL;
3568 			}
3569 		}
3570 		if (error == 0 && len != 0)
3571 			maybe_yield();
3572 	}
3573 out:
3574 	if (rl_cookie != NULL)
3575 		vn_rangelock_unlock(vp, rl_cookie);
3576 	*offset = off;
3577 	*length = len;
3578 	return (error);
3579 }
3580 
3581 /*
3582  * This function is supposed to be used in the situations where the deallocation
3583  * is not triggered by a user request.
3584  */
3585 int
3586 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3587     int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3588 {
3589 	struct ucred *cred;
3590 
3591 	if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3592 	    flags != 0)
3593 		return (EINVAL);
3594 	if (vp->v_type != VREG)
3595 		return (ENODEV);
3596 
3597 	cred = file_cred != NOCRED ? file_cred : active_cred;
3598 	return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3599 	    active_cred, file_cred));
3600 }
3601 
3602 static int
3603 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3604     struct ucred *active_cred, struct thread *td)
3605 {
3606 	int error;
3607 	struct vnode *vp;
3608 	int ioflag;
3609 
3610 	vp = fp->f_vnode;
3611 
3612 	if (cmd != SPACECTL_DEALLOC || *offset < 0 || *length <= 0 ||
3613 	    *length > OFF_MAX - *offset || flags != 0)
3614 		return (EINVAL);
3615 	if (vp->v_type != VREG)
3616 		return (ENODEV);
3617 
3618 	ioflag = get_write_ioflag(fp);
3619 
3620 	switch (cmd) {
3621 	case SPACECTL_DEALLOC:
3622 		error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3623 		    active_cred, active_cred, fp->f_cred);
3624 		break;
3625 	default:
3626 		panic("vn_fspacectl: unknown cmd %d", cmd);
3627 	}
3628 
3629 	return (error);
3630 }
3631 
3632 static u_long vn_lock_pair_pause_cnt;
3633 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3634     &vn_lock_pair_pause_cnt, 0,
3635     "Count of vn_lock_pair deadlocks");
3636 
3637 u_int vn_lock_pair_pause_max;
3638 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3639     &vn_lock_pair_pause_max, 0,
3640     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3641 
3642 static void
3643 vn_lock_pair_pause(const char *wmesg)
3644 {
3645 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3646 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3647 }
3648 
3649 /*
3650  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3651  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3652  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3653  * can be NULL.
3654  *
3655  * The function returns with both vnodes exclusively locked, and
3656  * guarantees that it does not create lock order reversal with other
3657  * threads during its execution.  Both vnodes could be unlocked
3658  * temporary (and reclaimed).
3659  */
3660 void
3661 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3662     bool vp2_locked)
3663 {
3664 	int error;
3665 
3666 	if (vp1 == NULL && vp2 == NULL)
3667 		return;
3668 	if (vp1 != NULL) {
3669 		if (vp1_locked)
3670 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3671 		else
3672 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3673 	} else {
3674 		vp1_locked = true;
3675 	}
3676 	if (vp2 != NULL) {
3677 		if (vp2_locked)
3678 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3679 		else
3680 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3681 	} else {
3682 		vp2_locked = true;
3683 	}
3684 	if (!vp1_locked && !vp2_locked) {
3685 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3686 		vp1_locked = true;
3687 	}
3688 
3689 	for (;;) {
3690 		if (vp1_locked && vp2_locked)
3691 			break;
3692 		if (vp1_locked && vp2 != NULL) {
3693 			if (vp1 != NULL) {
3694 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3695 				    __FILE__, __LINE__);
3696 				if (error == 0)
3697 					break;
3698 				VOP_UNLOCK(vp1);
3699 				vp1_locked = false;
3700 				vn_lock_pair_pause("vlp1");
3701 			}
3702 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3703 			vp2_locked = true;
3704 		}
3705 		if (vp2_locked && vp1 != NULL) {
3706 			if (vp2 != NULL) {
3707 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3708 				    __FILE__, __LINE__);
3709 				if (error == 0)
3710 					break;
3711 				VOP_UNLOCK(vp2);
3712 				vp2_locked = false;
3713 				vn_lock_pair_pause("vlp2");
3714 			}
3715 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3716 			vp1_locked = true;
3717 		}
3718 	}
3719 	if (vp1 != NULL)
3720 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3721 	if (vp2 != NULL)
3722 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3723 }
3724 
3725 int
3726 vn_lktype_write(struct mount *mp, struct vnode *vp)
3727 {
3728 	if (MNT_SHARED_WRITES(mp) ||
3729 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3730 		return (LK_SHARED);
3731 	return (LK_EXCLUSIVE);
3732 }
3733