xref: /freebsd/sys/kern/vfs_vnops.c (revision f7c32ed617858bcd22f8d1b03199099d50125721)
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 	return (res);
217 }
218 
219 /*
220  * Common code for vnode open operations via a name lookup.
221  * Lookup the vnode and invoke VOP_CREATE if needed.
222  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
223  *
224  * Note that this does NOT free nameidata for the successful case,
225  * due to the NDINIT being done elsewhere.
226  */
227 int
228 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
229     struct ucred *cred, struct file *fp)
230 {
231 	struct vnode *vp;
232 	struct mount *mp;
233 	struct vattr vat;
234 	struct vattr *vap = &vat;
235 	int fmode, error;
236 	bool first_open;
237 
238 restart:
239 	first_open = false;
240 	fmode = *flagp;
241 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
242 	    O_EXCL | O_DIRECTORY) ||
243 	    (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
244 		return (EINVAL);
245 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
246 		ndp->ni_cnd.cn_nameiop = CREATE;
247 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
248 		/*
249 		 * Set NOCACHE to avoid flushing the cache when
250 		 * rolling in many files at once.
251 		 *
252 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
253 		 * exist despite NOCACHE.
254 		 */
255 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
256 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
257 			ndp->ni_cnd.cn_flags |= FOLLOW;
258 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
259 			bwillwrite();
260 		if ((error = namei(ndp)) != 0)
261 			return (error);
262 		if (ndp->ni_vp == NULL) {
263 			VATTR_NULL(vap);
264 			vap->va_type = VREG;
265 			vap->va_mode = cmode;
266 			if (fmode & O_EXCL)
267 				vap->va_vaflags |= VA_EXCLUSIVE;
268 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
269 				NDFREE(ndp, NDF_ONLY_PNBUF);
270 				vput(ndp->ni_dvp);
271 				if ((error = vn_start_write(NULL, &mp,
272 				    V_XSLEEP | PCATCH)) != 0)
273 					return (error);
274 				NDREINIT(ndp);
275 				goto restart;
276 			}
277 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
278 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
279 #ifdef MAC
280 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
281 			    &ndp->ni_cnd, vap);
282 			if (error == 0)
283 #endif
284 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
285 				    &ndp->ni_cnd, vap);
286 			vp = ndp->ni_vp;
287 			if (error == 0 && (fmode & O_EXCL) != 0 &&
288 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
289 				VI_LOCK(vp);
290 				vp->v_iflag |= VI_FOPENING;
291 				VI_UNLOCK(vp);
292 				first_open = true;
293 			}
294 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
295 			    false);
296 			vn_finished_write(mp);
297 			if (error) {
298 				NDFREE(ndp, NDF_ONLY_PNBUF);
299 				if (error == ERELOOKUP) {
300 					NDREINIT(ndp);
301 					goto restart;
302 				}
303 				return (error);
304 			}
305 			fmode &= ~O_TRUNC;
306 		} else {
307 			if (ndp->ni_dvp == ndp->ni_vp)
308 				vrele(ndp->ni_dvp);
309 			else
310 				vput(ndp->ni_dvp);
311 			ndp->ni_dvp = NULL;
312 			vp = ndp->ni_vp;
313 			if (fmode & O_EXCL) {
314 				error = EEXIST;
315 				goto bad;
316 			}
317 			if (vp->v_type == VDIR) {
318 				error = EISDIR;
319 				goto bad;
320 			}
321 			fmode &= ~O_CREAT;
322 		}
323 	} else {
324 		ndp->ni_cnd.cn_nameiop = LOOKUP;
325 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
326 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
327 		    FOLLOW;
328 		if ((fmode & FWRITE) == 0)
329 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
330 		if ((error = namei(ndp)) != 0)
331 			return (error);
332 		vp = ndp->ni_vp;
333 	}
334 	error = vn_open_vnode(vp, fmode, cred, curthread, fp);
335 	if (first_open) {
336 		VI_LOCK(vp);
337 		vp->v_iflag &= ~VI_FOPENING;
338 		wakeup(vp);
339 		VI_UNLOCK(vp);
340 	}
341 	if (error)
342 		goto bad;
343 	*flagp = fmode;
344 	return (0);
345 bad:
346 	NDFREE(ndp, NDF_ONLY_PNBUF);
347 	vput(vp);
348 	*flagp = fmode;
349 	ndp->ni_vp = NULL;
350 	return (error);
351 }
352 
353 static int
354 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
355 {
356 	struct flock lf;
357 	int error, lock_flags, type;
358 
359 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
360 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
361 		return (0);
362 	KASSERT(fp != NULL, ("open with flock requires fp"));
363 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
364 		return (EOPNOTSUPP);
365 
366 	lock_flags = VOP_ISLOCKED(vp);
367 	VOP_UNLOCK(vp);
368 
369 	lf.l_whence = SEEK_SET;
370 	lf.l_start = 0;
371 	lf.l_len = 0;
372 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
373 	type = F_FLOCK;
374 	if ((fmode & FNONBLOCK) == 0)
375 		type |= F_WAIT;
376 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
377 		type |= F_FIRSTOPEN;
378 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
379 	if (error == 0)
380 		fp->f_flag |= FHASLOCK;
381 
382 	vn_lock(vp, lock_flags | LK_RETRY);
383 	return (error);
384 }
385 
386 /*
387  * Common code for vnode open operations once a vnode is located.
388  * Check permissions, and call the VOP_OPEN routine.
389  */
390 int
391 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
392     struct thread *td, struct file *fp)
393 {
394 	accmode_t accmode;
395 	int error;
396 
397 	if (vp->v_type == VLNK) {
398 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
399 			return (EMLINK);
400 	}
401 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
402 		return (ENOTDIR);
403 
404 	accmode = 0;
405 	if ((fmode & O_PATH) == 0) {
406 		if (vp->v_type == VSOCK)
407 			return (EOPNOTSUPP);
408 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
409 			if (vp->v_type == VDIR)
410 				return (EISDIR);
411 			accmode |= VWRITE;
412 		}
413 		if ((fmode & FREAD) != 0)
414 			accmode |= VREAD;
415 		if ((fmode & O_APPEND) && (fmode & FWRITE))
416 			accmode |= VAPPEND;
417 #ifdef MAC
418 		if ((fmode & O_CREAT) != 0)
419 			accmode |= VCREAT;
420 #endif
421 	}
422 	if ((fmode & FEXEC) != 0)
423 		accmode |= VEXEC;
424 #ifdef MAC
425 	if ((fmode & O_VERIFY) != 0)
426 		accmode |= VVERIFY;
427 	error = mac_vnode_check_open(cred, vp, accmode);
428 	if (error != 0)
429 		return (error);
430 
431 	accmode &= ~(VCREAT | VVERIFY);
432 #endif
433 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
434 		error = VOP_ACCESS(vp, accmode, cred, td);
435 		if (error != 0)
436 			return (error);
437 	}
438 	if ((fmode & O_PATH) != 0) {
439 		if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
440 		    VOP_ACCESS(vp, VREAD, cred, td) == 0)
441 			fp->f_flag |= FKQALLOWED;
442 		return (0);
443 	}
444 
445 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
446 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
447 	error = VOP_OPEN(vp, fmode, cred, td, fp);
448 	if (error != 0)
449 		return (error);
450 
451 	error = vn_open_vnode_advlock(vp, fmode, fp);
452 	if (error == 0 && (fmode & FWRITE) != 0) {
453 		error = VOP_ADD_WRITECOUNT(vp, 1);
454 		if (error == 0) {
455 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
456 			     __func__, vp, vp->v_writecount);
457 		}
458 	}
459 
460 	/*
461 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
462 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
463 	 */
464 	if (error != 0) {
465 		if (fp != NULL) {
466 			/*
467 			 * Arrange the call by having fdrop() to use
468 			 * vn_closefile().  This is to satisfy
469 			 * filesystems like devfs or tmpfs, which
470 			 * override fo_close().
471 			 */
472 			fp->f_flag |= FOPENFAILED;
473 			fp->f_vnode = vp;
474 			if (fp->f_ops == &badfileops) {
475 				fp->f_type = DTYPE_VNODE;
476 				fp->f_ops = &vnops;
477 			}
478 			vref(vp);
479 		} else {
480 			/*
481 			 * If there is no fp, due to kernel-mode open,
482 			 * we can call VOP_CLOSE() now.
483 			 */
484 			if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
485 			    !MNT_EXTENDED_SHARED(vp->v_mount) &&
486 			    VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
487 				vn_lock(vp, LK_UPGRADE | LK_RETRY);
488 			(void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
489 			    cred, td);
490 		}
491 	}
492 
493 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
494 	return (error);
495 
496 }
497 
498 /*
499  * Check for write permissions on the specified vnode.
500  * Prototype text segments cannot be written.
501  * It is racy.
502  */
503 int
504 vn_writechk(struct vnode *vp)
505 {
506 
507 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
508 	/*
509 	 * If there's shared text associated with
510 	 * the vnode, try to free it up once.  If
511 	 * we fail, we can't allow writing.
512 	 */
513 	if (VOP_IS_TEXT(vp))
514 		return (ETXTBSY);
515 
516 	return (0);
517 }
518 
519 /*
520  * Vnode close call
521  */
522 static int
523 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
524     struct thread *td, bool keep_ref)
525 {
526 	struct mount *mp;
527 	int error, lock_flags;
528 
529 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
530 	    MNT_EXTENDED_SHARED(vp->v_mount))
531 		lock_flags = LK_SHARED;
532 	else
533 		lock_flags = LK_EXCLUSIVE;
534 
535 	vn_start_write(vp, &mp, V_WAIT);
536 	vn_lock(vp, lock_flags | LK_RETRY);
537 	AUDIT_ARG_VNODE1(vp);
538 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
539 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
540 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
541 		    __func__, vp, vp->v_writecount);
542 	}
543 	error = VOP_CLOSE(vp, flags, file_cred, td);
544 	if (keep_ref)
545 		VOP_UNLOCK(vp);
546 	else
547 		vput(vp);
548 	vn_finished_write(mp);
549 	return (error);
550 }
551 
552 int
553 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
554     struct thread *td)
555 {
556 
557 	return (vn_close1(vp, flags, file_cred, td, false));
558 }
559 
560 /*
561  * Heuristic to detect sequential operation.
562  */
563 static int
564 sequential_heuristic(struct uio *uio, struct file *fp)
565 {
566 	enum uio_rw rw;
567 
568 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
569 
570 	rw = uio->uio_rw;
571 	if (fp->f_flag & FRDAHEAD)
572 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
573 
574 	/*
575 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
576 	 * that the first I/O is normally considered to be slightly
577 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
578 	 * unless previous seeks have reduced f_seqcount to 0, in which
579 	 * case offset 0 is not special.
580 	 */
581 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
582 	    uio->uio_offset == fp->f_nextoff[rw]) {
583 		/*
584 		 * f_seqcount is in units of fixed-size blocks so that it
585 		 * depends mainly on the amount of sequential I/O and not
586 		 * much on the number of sequential I/O's.  The fixed size
587 		 * of 16384 is hard-coded here since it is (not quite) just
588 		 * a magic size that works well here.  This size is more
589 		 * closely related to the best I/O size for real disks than
590 		 * to any block size used by software.
591 		 */
592 		if (uio->uio_resid >= IO_SEQMAX * 16384)
593 			fp->f_seqcount[rw] = IO_SEQMAX;
594 		else {
595 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
596 			if (fp->f_seqcount[rw] > IO_SEQMAX)
597 				fp->f_seqcount[rw] = IO_SEQMAX;
598 		}
599 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
600 	}
601 
602 	/* Not sequential.  Quickly draw-down sequentiality. */
603 	if (fp->f_seqcount[rw] > 1)
604 		fp->f_seqcount[rw] = 1;
605 	else
606 		fp->f_seqcount[rw] = 0;
607 	return (0);
608 }
609 
610 /*
611  * Package up an I/O request on a vnode into a uio and do it.
612  */
613 int
614 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
615     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
616     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
617 {
618 	struct uio auio;
619 	struct iovec aiov;
620 	struct mount *mp;
621 	struct ucred *cred;
622 	void *rl_cookie;
623 	struct vn_io_fault_args args;
624 	int error, lock_flags;
625 
626 	if (offset < 0 && vp->v_type != VCHR)
627 		return (EINVAL);
628 	auio.uio_iov = &aiov;
629 	auio.uio_iovcnt = 1;
630 	aiov.iov_base = base;
631 	aiov.iov_len = len;
632 	auio.uio_resid = len;
633 	auio.uio_offset = offset;
634 	auio.uio_segflg = segflg;
635 	auio.uio_rw = rw;
636 	auio.uio_td = td;
637 	error = 0;
638 
639 	if ((ioflg & IO_NODELOCKED) == 0) {
640 		if ((ioflg & IO_RANGELOCKED) == 0) {
641 			if (rw == UIO_READ) {
642 				rl_cookie = vn_rangelock_rlock(vp, offset,
643 				    offset + len);
644 			} else if ((ioflg & IO_APPEND) != 0) {
645 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
646 			} else {
647 				rl_cookie = vn_rangelock_wlock(vp, offset,
648 				    offset + len);
649 			}
650 		} else
651 			rl_cookie = NULL;
652 		mp = NULL;
653 		if (rw == UIO_WRITE) {
654 			if (vp->v_type != VCHR &&
655 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
656 			    != 0)
657 				goto out;
658 			lock_flags = vn_lktype_write(mp, vp);
659 		} else
660 			lock_flags = LK_SHARED;
661 		vn_lock(vp, lock_flags | LK_RETRY);
662 	} else
663 		rl_cookie = NULL;
664 
665 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
666 #ifdef MAC
667 	if ((ioflg & IO_NOMACCHECK) == 0) {
668 		if (rw == UIO_READ)
669 			error = mac_vnode_check_read(active_cred, file_cred,
670 			    vp);
671 		else
672 			error = mac_vnode_check_write(active_cred, file_cred,
673 			    vp);
674 	}
675 #endif
676 	if (error == 0) {
677 		if (file_cred != NULL)
678 			cred = file_cred;
679 		else
680 			cred = active_cred;
681 		if (do_vn_io_fault(vp, &auio)) {
682 			args.kind = VN_IO_FAULT_VOP;
683 			args.cred = cred;
684 			args.flags = ioflg;
685 			args.args.vop_args.vp = vp;
686 			error = vn_io_fault1(vp, &auio, &args, td);
687 		} else if (rw == UIO_READ) {
688 			error = VOP_READ(vp, &auio, ioflg, cred);
689 		} else /* if (rw == UIO_WRITE) */ {
690 			error = VOP_WRITE(vp, &auio, ioflg, cred);
691 		}
692 	}
693 	if (aresid)
694 		*aresid = auio.uio_resid;
695 	else
696 		if (auio.uio_resid && error == 0)
697 			error = EIO;
698 	if ((ioflg & IO_NODELOCKED) == 0) {
699 		VOP_UNLOCK(vp);
700 		if (mp != NULL)
701 			vn_finished_write(mp);
702 	}
703  out:
704 	if (rl_cookie != NULL)
705 		vn_rangelock_unlock(vp, rl_cookie);
706 	return (error);
707 }
708 
709 /*
710  * Package up an I/O request on a vnode into a uio and do it.  The I/O
711  * request is split up into smaller chunks and we try to avoid saturating
712  * the buffer cache while potentially holding a vnode locked, so we
713  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
714  * to give other processes a chance to lock the vnode (either other processes
715  * core'ing the same binary, or unrelated processes scanning the directory).
716  */
717 int
718 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
719     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
720     struct ucred *file_cred, size_t *aresid, struct thread *td)
721 {
722 	int error = 0;
723 	ssize_t iaresid;
724 
725 	do {
726 		int chunk;
727 
728 		/*
729 		 * Force `offset' to a multiple of MAXBSIZE except possibly
730 		 * for the first chunk, so that filesystems only need to
731 		 * write full blocks except possibly for the first and last
732 		 * chunks.
733 		 */
734 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
735 
736 		if (chunk > len)
737 			chunk = len;
738 		if (rw != UIO_READ && vp->v_type == VREG)
739 			bwillwrite();
740 		iaresid = 0;
741 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
742 		    ioflg, active_cred, file_cred, &iaresid, td);
743 		len -= chunk;	/* aresid calc already includes length */
744 		if (error)
745 			break;
746 		offset += chunk;
747 		base = (char *)base + chunk;
748 		kern_yield(PRI_USER);
749 	} while (len);
750 	if (aresid)
751 		*aresid = len + iaresid;
752 	return (error);
753 }
754 
755 #if OFF_MAX <= LONG_MAX
756 off_t
757 foffset_lock(struct file *fp, int flags)
758 {
759 	volatile short *flagsp;
760 	off_t res;
761 	short state;
762 
763 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
764 
765 	if ((flags & FOF_NOLOCK) != 0)
766 		return (atomic_load_long(&fp->f_offset));
767 
768 	/*
769 	 * According to McKusick the vn lock was protecting f_offset here.
770 	 * It is now protected by the FOFFSET_LOCKED flag.
771 	 */
772 	flagsp = &fp->f_vnread_flags;
773 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
774 		return (atomic_load_long(&fp->f_offset));
775 
776 	sleepq_lock(&fp->f_vnread_flags);
777 	state = atomic_load_16(flagsp);
778 	for (;;) {
779 		if ((state & FOFFSET_LOCKED) == 0) {
780 			if (!atomic_fcmpset_acq_16(flagsp, &state,
781 			    FOFFSET_LOCKED))
782 				continue;
783 			break;
784 		}
785 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
786 			if (!atomic_fcmpset_acq_16(flagsp, &state,
787 			    state | FOFFSET_LOCK_WAITING))
788 				continue;
789 		}
790 		DROP_GIANT();
791 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
792 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
793 		PICKUP_GIANT();
794 		sleepq_lock(&fp->f_vnread_flags);
795 		state = atomic_load_16(flagsp);
796 	}
797 	res = atomic_load_long(&fp->f_offset);
798 	sleepq_release(&fp->f_vnread_flags);
799 	return (res);
800 }
801 
802 void
803 foffset_unlock(struct file *fp, off_t val, int flags)
804 {
805 	volatile short *flagsp;
806 	short state;
807 
808 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
809 
810 	if ((flags & FOF_NOUPDATE) == 0)
811 		atomic_store_long(&fp->f_offset, val);
812 	if ((flags & FOF_NEXTOFF_R) != 0)
813 		fp->f_nextoff[UIO_READ] = val;
814 	if ((flags & FOF_NEXTOFF_W) != 0)
815 		fp->f_nextoff[UIO_WRITE] = val;
816 
817 	if ((flags & FOF_NOLOCK) != 0)
818 		return;
819 
820 	flagsp = &fp->f_vnread_flags;
821 	state = atomic_load_16(flagsp);
822 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
823 	    atomic_cmpset_rel_16(flagsp, state, 0))
824 		return;
825 
826 	sleepq_lock(&fp->f_vnread_flags);
827 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
828 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
829 	fp->f_vnread_flags = 0;
830 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
831 	sleepq_release(&fp->f_vnread_flags);
832 }
833 #else
834 off_t
835 foffset_lock(struct file *fp, int flags)
836 {
837 	struct mtx *mtxp;
838 	off_t res;
839 
840 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
841 
842 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
843 	mtx_lock(mtxp);
844 	if ((flags & FOF_NOLOCK) == 0) {
845 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
846 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
847 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
848 			    "vofflock", 0);
849 		}
850 		fp->f_vnread_flags |= FOFFSET_LOCKED;
851 	}
852 	res = fp->f_offset;
853 	mtx_unlock(mtxp);
854 	return (res);
855 }
856 
857 void
858 foffset_unlock(struct file *fp, off_t val, int flags)
859 {
860 	struct mtx *mtxp;
861 
862 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
863 
864 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
865 	mtx_lock(mtxp);
866 	if ((flags & FOF_NOUPDATE) == 0)
867 		fp->f_offset = val;
868 	if ((flags & FOF_NEXTOFF_R) != 0)
869 		fp->f_nextoff[UIO_READ] = val;
870 	if ((flags & FOF_NEXTOFF_W) != 0)
871 		fp->f_nextoff[UIO_WRITE] = val;
872 	if ((flags & FOF_NOLOCK) == 0) {
873 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
874 		    ("Lost FOFFSET_LOCKED"));
875 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
876 			wakeup(&fp->f_vnread_flags);
877 		fp->f_vnread_flags = 0;
878 	}
879 	mtx_unlock(mtxp);
880 }
881 #endif
882 
883 void
884 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
885 {
886 
887 	if ((flags & FOF_OFFSET) == 0)
888 		uio->uio_offset = foffset_lock(fp, flags);
889 }
890 
891 void
892 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
893 {
894 
895 	if ((flags & FOF_OFFSET) == 0)
896 		foffset_unlock(fp, uio->uio_offset, flags);
897 }
898 
899 static int
900 get_advice(struct file *fp, struct uio *uio)
901 {
902 	struct mtx *mtxp;
903 	int ret;
904 
905 	ret = POSIX_FADV_NORMAL;
906 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
907 		return (ret);
908 
909 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
910 	mtx_lock(mtxp);
911 	if (fp->f_advice != NULL &&
912 	    uio->uio_offset >= fp->f_advice->fa_start &&
913 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
914 		ret = fp->f_advice->fa_advice;
915 	mtx_unlock(mtxp);
916 	return (ret);
917 }
918 
919 static int
920 get_write_ioflag(struct file *fp)
921 {
922 	int ioflag;
923 	struct mount *mp;
924 	struct vnode *vp;
925 
926 	ioflag = 0;
927 	vp = fp->f_vnode;
928 	mp = atomic_load_ptr(&vp->v_mount);
929 
930 	if ((fp->f_flag & O_DIRECT) != 0)
931 		ioflag |= IO_DIRECT;
932 
933 	if ((fp->f_flag & O_FSYNC) != 0 ||
934 	    (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
935 		ioflag |= IO_SYNC;
936 
937 	/*
938 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
939 	 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
940 	 * fall back to full O_SYNC behavior.
941 	 */
942 	if ((fp->f_flag & O_DSYNC) != 0)
943 		ioflag |= IO_SYNC | IO_DATASYNC;
944 
945 	return (ioflag);
946 }
947 
948 int
949 vn_read_from_obj(struct vnode *vp, struct uio *uio)
950 {
951 	vm_object_t obj;
952 	vm_page_t ma[io_hold_cnt + 2];
953 	off_t off, vsz;
954 	ssize_t resid;
955 	int error, i, j;
956 
957 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
958 	obj = atomic_load_ptr(&vp->v_object);
959 	if (obj == NULL)
960 		return (EJUSTRETURN);
961 
962 	/*
963 	 * Depends on type stability of vm_objects.
964 	 */
965 	vm_object_pip_add(obj, 1);
966 	if ((obj->flags & OBJ_DEAD) != 0) {
967 		/*
968 		 * Note that object might be already reused from the
969 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
970 		 * we recheck for DOOMED vnode state after all pages
971 		 * are busied, and retract then.
972 		 *
973 		 * But we check for OBJ_DEAD to ensure that we do not
974 		 * busy pages while vm_object_terminate_pages()
975 		 * processes the queue.
976 		 */
977 		error = EJUSTRETURN;
978 		goto out_pip;
979 	}
980 
981 	resid = uio->uio_resid;
982 	off = uio->uio_offset;
983 	for (i = 0; resid > 0; i++) {
984 		MPASS(i < io_hold_cnt + 2);
985 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
986 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
987 		    VM_ALLOC_NOWAIT);
988 		if (ma[i] == NULL)
989 			break;
990 
991 		/*
992 		 * Skip invalid pages.  Valid mask can be partial only
993 		 * at EOF, and we clip later.
994 		 */
995 		if (vm_page_none_valid(ma[i])) {
996 			vm_page_sunbusy(ma[i]);
997 			break;
998 		}
999 
1000 		resid -= PAGE_SIZE;
1001 		off += PAGE_SIZE;
1002 	}
1003 	if (i == 0) {
1004 		error = EJUSTRETURN;
1005 		goto out_pip;
1006 	}
1007 
1008 	/*
1009 	 * Check VIRF_DOOMED after we busied our pages.  Since
1010 	 * vgonel() terminates the vnode' vm_object, it cannot
1011 	 * process past pages busied by us.
1012 	 */
1013 	if (VN_IS_DOOMED(vp)) {
1014 		error = EJUSTRETURN;
1015 		goto out;
1016 	}
1017 
1018 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1019 	if (resid > uio->uio_resid)
1020 		resid = uio->uio_resid;
1021 
1022 	/*
1023 	 * Unlocked read of vnp_size is safe because truncation cannot
1024 	 * pass busied page.  But we load vnp_size into a local
1025 	 * variable so that possible concurrent extension does not
1026 	 * break calculation.
1027 	 */
1028 #if defined(__powerpc__) && !defined(__powerpc64__)
1029 	vsz = obj->un_pager.vnp.vnp_size;
1030 #else
1031 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1032 #endif
1033 	if (uio->uio_offset >= vsz) {
1034 		error = EJUSTRETURN;
1035 		goto out;
1036 	}
1037 	if (uio->uio_offset + resid > vsz)
1038 		resid = vsz - uio->uio_offset;
1039 
1040 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1041 
1042 out:
1043 	for (j = 0; j < i; j++) {
1044 		if (error == 0)
1045 			vm_page_reference(ma[j]);
1046 		vm_page_sunbusy(ma[j]);
1047 	}
1048 out_pip:
1049 	vm_object_pip_wakeup(obj);
1050 	if (error != 0)
1051 		return (error);
1052 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1053 }
1054 
1055 /*
1056  * File table vnode read routine.
1057  */
1058 static int
1059 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1060     struct thread *td)
1061 {
1062 	struct vnode *vp;
1063 	off_t orig_offset;
1064 	int error, ioflag;
1065 	int advice;
1066 
1067 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1068 	    uio->uio_td, td));
1069 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1070 	vp = fp->f_vnode;
1071 	ioflag = 0;
1072 	if (fp->f_flag & FNONBLOCK)
1073 		ioflag |= IO_NDELAY;
1074 	if (fp->f_flag & O_DIRECT)
1075 		ioflag |= IO_DIRECT;
1076 
1077 	/*
1078 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1079 	 * allows us to avoid unneeded work outright.
1080 	 */
1081 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1082 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1083 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1084 		if (error == 0) {
1085 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1086 			return (0);
1087 		}
1088 		if (error != EJUSTRETURN)
1089 			return (error);
1090 	}
1091 
1092 	advice = get_advice(fp, uio);
1093 	vn_lock(vp, LK_SHARED | LK_RETRY);
1094 
1095 	switch (advice) {
1096 	case POSIX_FADV_NORMAL:
1097 	case POSIX_FADV_SEQUENTIAL:
1098 	case POSIX_FADV_NOREUSE:
1099 		ioflag |= sequential_heuristic(uio, fp);
1100 		break;
1101 	case POSIX_FADV_RANDOM:
1102 		/* Disable read-ahead for random I/O. */
1103 		break;
1104 	}
1105 	orig_offset = uio->uio_offset;
1106 
1107 #ifdef MAC
1108 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1109 	if (error == 0)
1110 #endif
1111 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1112 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1113 	VOP_UNLOCK(vp);
1114 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1115 	    orig_offset != uio->uio_offset)
1116 		/*
1117 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1118 		 * for the backing file after a POSIX_FADV_NOREUSE
1119 		 * read(2).
1120 		 */
1121 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1122 		    POSIX_FADV_DONTNEED);
1123 	return (error);
1124 }
1125 
1126 /*
1127  * File table vnode write routine.
1128  */
1129 static int
1130 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1131     struct thread *td)
1132 {
1133 	struct vnode *vp;
1134 	struct mount *mp;
1135 	off_t orig_offset;
1136 	int error, ioflag;
1137 	int advice;
1138 	bool need_finished_write;
1139 
1140 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1141 	    uio->uio_td, td));
1142 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1143 	vp = fp->f_vnode;
1144 	if (vp->v_type == VREG)
1145 		bwillwrite();
1146 	ioflag = IO_UNIT;
1147 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1148 		ioflag |= IO_APPEND;
1149 	if ((fp->f_flag & FNONBLOCK) != 0)
1150 		ioflag |= IO_NDELAY;
1151 	ioflag |= get_write_ioflag(fp);
1152 
1153 	mp = NULL;
1154 	need_finished_write = false;
1155 	if (vp->v_type != VCHR) {
1156 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1157 		if (error != 0)
1158 			goto unlock;
1159 		need_finished_write = true;
1160 	}
1161 
1162 	advice = get_advice(fp, uio);
1163 
1164 	vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1165 	switch (advice) {
1166 	case POSIX_FADV_NORMAL:
1167 	case POSIX_FADV_SEQUENTIAL:
1168 	case POSIX_FADV_NOREUSE:
1169 		ioflag |= sequential_heuristic(uio, fp);
1170 		break;
1171 	case POSIX_FADV_RANDOM:
1172 		/* XXX: Is this correct? */
1173 		break;
1174 	}
1175 	orig_offset = uio->uio_offset;
1176 
1177 #ifdef MAC
1178 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1179 	if (error == 0)
1180 #endif
1181 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1182 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1183 	VOP_UNLOCK(vp);
1184 	if (need_finished_write)
1185 		vn_finished_write(mp);
1186 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1187 	    orig_offset != uio->uio_offset)
1188 		/*
1189 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1190 		 * for the backing file after a POSIX_FADV_NOREUSE
1191 		 * write(2).
1192 		 */
1193 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1194 		    POSIX_FADV_DONTNEED);
1195 unlock:
1196 	return (error);
1197 }
1198 
1199 /*
1200  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1201  * prevent the following deadlock:
1202  *
1203  * Assume that the thread A reads from the vnode vp1 into userspace
1204  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1205  * currently not resident, then system ends up with the call chain
1206  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1207  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1208  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1209  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1210  * backed by the pages of vnode vp1, and some page in buf2 is not
1211  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1212  *
1213  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1214  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1215  * Instead, it first tries to do the whole range i/o with pagefaults
1216  * disabled. If all pages in the i/o buffer are resident and mapped,
1217  * VOP will succeed (ignoring the genuine filesystem errors).
1218  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1219  * i/o in chunks, with all pages in the chunk prefaulted and held
1220  * using vm_fault_quick_hold_pages().
1221  *
1222  * Filesystems using this deadlock avoidance scheme should use the
1223  * array of the held pages from uio, saved in the curthread->td_ma,
1224  * instead of doing uiomove().  A helper function
1225  * vn_io_fault_uiomove() converts uiomove request into
1226  * uiomove_fromphys() over td_ma array.
1227  *
1228  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1229  * make the current i/o request atomic with respect to other i/os and
1230  * truncations.
1231  */
1232 
1233 /*
1234  * Decode vn_io_fault_args and perform the corresponding i/o.
1235  */
1236 static int
1237 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1238     struct thread *td)
1239 {
1240 	int error, save;
1241 
1242 	error = 0;
1243 	save = vm_fault_disable_pagefaults();
1244 	switch (args->kind) {
1245 	case VN_IO_FAULT_FOP:
1246 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1247 		    uio, args->cred, args->flags, td);
1248 		break;
1249 	case VN_IO_FAULT_VOP:
1250 		if (uio->uio_rw == UIO_READ) {
1251 			error = VOP_READ(args->args.vop_args.vp, uio,
1252 			    args->flags, args->cred);
1253 		} else if (uio->uio_rw == UIO_WRITE) {
1254 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1255 			    args->flags, args->cred);
1256 		}
1257 		break;
1258 	default:
1259 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1260 		    args->kind, uio->uio_rw);
1261 	}
1262 	vm_fault_enable_pagefaults(save);
1263 	return (error);
1264 }
1265 
1266 static int
1267 vn_io_fault_touch(char *base, const struct uio *uio)
1268 {
1269 	int r;
1270 
1271 	r = fubyte(base);
1272 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1273 		return (EFAULT);
1274 	return (0);
1275 }
1276 
1277 static int
1278 vn_io_fault_prefault_user(const struct uio *uio)
1279 {
1280 	char *base;
1281 	const struct iovec *iov;
1282 	size_t len;
1283 	ssize_t resid;
1284 	int error, i;
1285 
1286 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1287 	    ("vn_io_fault_prefault userspace"));
1288 
1289 	error = i = 0;
1290 	iov = uio->uio_iov;
1291 	resid = uio->uio_resid;
1292 	base = iov->iov_base;
1293 	len = iov->iov_len;
1294 	while (resid > 0) {
1295 		error = vn_io_fault_touch(base, uio);
1296 		if (error != 0)
1297 			break;
1298 		if (len < PAGE_SIZE) {
1299 			if (len != 0) {
1300 				error = vn_io_fault_touch(base + len - 1, uio);
1301 				if (error != 0)
1302 					break;
1303 				resid -= len;
1304 			}
1305 			if (++i >= uio->uio_iovcnt)
1306 				break;
1307 			iov = uio->uio_iov + i;
1308 			base = iov->iov_base;
1309 			len = iov->iov_len;
1310 		} else {
1311 			len -= PAGE_SIZE;
1312 			base += PAGE_SIZE;
1313 			resid -= PAGE_SIZE;
1314 		}
1315 	}
1316 	return (error);
1317 }
1318 
1319 /*
1320  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1321  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1322  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1323  * into args and call vn_io_fault1() to handle faults during the user
1324  * mode buffer accesses.
1325  */
1326 static int
1327 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1328     struct thread *td)
1329 {
1330 	vm_page_t ma[io_hold_cnt + 2];
1331 	struct uio *uio_clone, short_uio;
1332 	struct iovec short_iovec[1];
1333 	vm_page_t *prev_td_ma;
1334 	vm_prot_t prot;
1335 	vm_offset_t addr, end;
1336 	size_t len, resid;
1337 	ssize_t adv;
1338 	int error, cnt, saveheld, prev_td_ma_cnt;
1339 
1340 	if (vn_io_fault_prefault) {
1341 		error = vn_io_fault_prefault_user(uio);
1342 		if (error != 0)
1343 			return (error); /* Or ignore ? */
1344 	}
1345 
1346 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1347 
1348 	/*
1349 	 * The UFS follows IO_UNIT directive and replays back both
1350 	 * uio_offset and uio_resid if an error is encountered during the
1351 	 * operation.  But, since the iovec may be already advanced,
1352 	 * uio is still in an inconsistent state.
1353 	 *
1354 	 * Cache a copy of the original uio, which is advanced to the redo
1355 	 * point using UIO_NOCOPY below.
1356 	 */
1357 	uio_clone = cloneuio(uio);
1358 	resid = uio->uio_resid;
1359 
1360 	short_uio.uio_segflg = UIO_USERSPACE;
1361 	short_uio.uio_rw = uio->uio_rw;
1362 	short_uio.uio_td = uio->uio_td;
1363 
1364 	error = vn_io_fault_doio(args, uio, td);
1365 	if (error != EFAULT)
1366 		goto out;
1367 
1368 	atomic_add_long(&vn_io_faults_cnt, 1);
1369 	uio_clone->uio_segflg = UIO_NOCOPY;
1370 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1371 	uio_clone->uio_segflg = uio->uio_segflg;
1372 
1373 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1374 	prev_td_ma = td->td_ma;
1375 	prev_td_ma_cnt = td->td_ma_cnt;
1376 
1377 	while (uio_clone->uio_resid != 0) {
1378 		len = uio_clone->uio_iov->iov_len;
1379 		if (len == 0) {
1380 			KASSERT(uio_clone->uio_iovcnt >= 1,
1381 			    ("iovcnt underflow"));
1382 			uio_clone->uio_iov++;
1383 			uio_clone->uio_iovcnt--;
1384 			continue;
1385 		}
1386 		if (len > ptoa(io_hold_cnt))
1387 			len = ptoa(io_hold_cnt);
1388 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1389 		end = round_page(addr + len);
1390 		if (end < addr) {
1391 			error = EFAULT;
1392 			break;
1393 		}
1394 		cnt = atop(end - trunc_page(addr));
1395 		/*
1396 		 * A perfectly misaligned address and length could cause
1397 		 * both the start and the end of the chunk to use partial
1398 		 * page.  +2 accounts for such a situation.
1399 		 */
1400 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1401 		    addr, len, prot, ma, io_hold_cnt + 2);
1402 		if (cnt == -1) {
1403 			error = EFAULT;
1404 			break;
1405 		}
1406 		short_uio.uio_iov = &short_iovec[0];
1407 		short_iovec[0].iov_base = (void *)addr;
1408 		short_uio.uio_iovcnt = 1;
1409 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1410 		short_uio.uio_offset = uio_clone->uio_offset;
1411 		td->td_ma = ma;
1412 		td->td_ma_cnt = cnt;
1413 
1414 		error = vn_io_fault_doio(args, &short_uio, td);
1415 		vm_page_unhold_pages(ma, cnt);
1416 		adv = len - short_uio.uio_resid;
1417 
1418 		uio_clone->uio_iov->iov_base =
1419 		    (char *)uio_clone->uio_iov->iov_base + adv;
1420 		uio_clone->uio_iov->iov_len -= adv;
1421 		uio_clone->uio_resid -= adv;
1422 		uio_clone->uio_offset += adv;
1423 
1424 		uio->uio_resid -= adv;
1425 		uio->uio_offset += adv;
1426 
1427 		if (error != 0 || adv == 0)
1428 			break;
1429 	}
1430 	td->td_ma = prev_td_ma;
1431 	td->td_ma_cnt = prev_td_ma_cnt;
1432 	curthread_pflags_restore(saveheld);
1433 out:
1434 	free(uio_clone, M_IOV);
1435 	return (error);
1436 }
1437 
1438 static int
1439 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1440     int flags, struct thread *td)
1441 {
1442 	fo_rdwr_t *doio;
1443 	struct vnode *vp;
1444 	void *rl_cookie;
1445 	struct vn_io_fault_args args;
1446 	int error;
1447 
1448 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1449 	vp = fp->f_vnode;
1450 
1451 	/*
1452 	 * The ability to read(2) on a directory has historically been
1453 	 * allowed for all users, but this can and has been the source of
1454 	 * at least one security issue in the past.  As such, it is now hidden
1455 	 * away behind a sysctl for those that actually need it to use it, and
1456 	 * restricted to root when it's turned on to make it relatively safe to
1457 	 * leave on for longer sessions of need.
1458 	 */
1459 	if (vp->v_type == VDIR) {
1460 		KASSERT(uio->uio_rw == UIO_READ,
1461 		    ("illegal write attempted on a directory"));
1462 		if (!vfs_allow_read_dir)
1463 			return (EISDIR);
1464 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1465 			return (EISDIR);
1466 	}
1467 
1468 	foffset_lock_uio(fp, uio, flags);
1469 	if (do_vn_io_fault(vp, uio)) {
1470 		args.kind = VN_IO_FAULT_FOP;
1471 		args.args.fop_args.fp = fp;
1472 		args.args.fop_args.doio = doio;
1473 		args.cred = active_cred;
1474 		args.flags = flags | FOF_OFFSET;
1475 		if (uio->uio_rw == UIO_READ) {
1476 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1477 			    uio->uio_offset + uio->uio_resid);
1478 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1479 		    (flags & FOF_OFFSET) == 0) {
1480 			/* For appenders, punt and lock the whole range. */
1481 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1482 		} else {
1483 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1484 			    uio->uio_offset + uio->uio_resid);
1485 		}
1486 		error = vn_io_fault1(vp, uio, &args, td);
1487 		vn_rangelock_unlock(vp, rl_cookie);
1488 	} else {
1489 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1490 	}
1491 	foffset_unlock_uio(fp, uio, flags);
1492 	return (error);
1493 }
1494 
1495 /*
1496  * Helper function to perform the requested uiomove operation using
1497  * the held pages for io->uio_iov[0].iov_base buffer instead of
1498  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1499  * instead of iov_base prevents page faults that could occur due to
1500  * pmap_collect() invalidating the mapping created by
1501  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1502  * object cleanup revoking the write access from page mappings.
1503  *
1504  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1505  * instead of plain uiomove().
1506  */
1507 int
1508 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1509 {
1510 	struct uio transp_uio;
1511 	struct iovec transp_iov[1];
1512 	struct thread *td;
1513 	size_t adv;
1514 	int error, pgadv;
1515 
1516 	td = curthread;
1517 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1518 	    uio->uio_segflg != UIO_USERSPACE)
1519 		return (uiomove(data, xfersize, uio));
1520 
1521 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1522 	transp_iov[0].iov_base = data;
1523 	transp_uio.uio_iov = &transp_iov[0];
1524 	transp_uio.uio_iovcnt = 1;
1525 	if (xfersize > uio->uio_resid)
1526 		xfersize = uio->uio_resid;
1527 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1528 	transp_uio.uio_offset = 0;
1529 	transp_uio.uio_segflg = UIO_SYSSPACE;
1530 	/*
1531 	 * Since transp_iov points to data, and td_ma page array
1532 	 * corresponds to original uio->uio_iov, we need to invert the
1533 	 * direction of the i/o operation as passed to
1534 	 * uiomove_fromphys().
1535 	 */
1536 	switch (uio->uio_rw) {
1537 	case UIO_WRITE:
1538 		transp_uio.uio_rw = UIO_READ;
1539 		break;
1540 	case UIO_READ:
1541 		transp_uio.uio_rw = UIO_WRITE;
1542 		break;
1543 	}
1544 	transp_uio.uio_td = uio->uio_td;
1545 	error = uiomove_fromphys(td->td_ma,
1546 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1547 	    xfersize, &transp_uio);
1548 	adv = xfersize - transp_uio.uio_resid;
1549 	pgadv =
1550 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1551 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1552 	td->td_ma += pgadv;
1553 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1554 	    pgadv));
1555 	td->td_ma_cnt -= pgadv;
1556 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1557 	uio->uio_iov->iov_len -= adv;
1558 	uio->uio_resid -= adv;
1559 	uio->uio_offset += adv;
1560 	return (error);
1561 }
1562 
1563 int
1564 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1565     struct uio *uio)
1566 {
1567 	struct thread *td;
1568 	vm_offset_t iov_base;
1569 	int cnt, pgadv;
1570 
1571 	td = curthread;
1572 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1573 	    uio->uio_segflg != UIO_USERSPACE)
1574 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1575 
1576 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1577 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1578 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1579 	switch (uio->uio_rw) {
1580 	case UIO_WRITE:
1581 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1582 		    offset, cnt);
1583 		break;
1584 	case UIO_READ:
1585 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1586 		    cnt);
1587 		break;
1588 	}
1589 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1590 	td->td_ma += pgadv;
1591 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1592 	    pgadv));
1593 	td->td_ma_cnt -= pgadv;
1594 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1595 	uio->uio_iov->iov_len -= cnt;
1596 	uio->uio_resid -= cnt;
1597 	uio->uio_offset += cnt;
1598 	return (0);
1599 }
1600 
1601 /*
1602  * File table truncate routine.
1603  */
1604 static int
1605 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1606     struct thread *td)
1607 {
1608 	struct mount *mp;
1609 	struct vnode *vp;
1610 	void *rl_cookie;
1611 	int error;
1612 
1613 	vp = fp->f_vnode;
1614 
1615 retry:
1616 	/*
1617 	 * Lock the whole range for truncation.  Otherwise split i/o
1618 	 * might happen partly before and partly after the truncation.
1619 	 */
1620 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1621 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1622 	if (error)
1623 		goto out1;
1624 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1625 	AUDIT_ARG_VNODE1(vp);
1626 	if (vp->v_type == VDIR) {
1627 		error = EISDIR;
1628 		goto out;
1629 	}
1630 #ifdef MAC
1631 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1632 	if (error)
1633 		goto out;
1634 #endif
1635 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1636 	    fp->f_cred);
1637 out:
1638 	VOP_UNLOCK(vp);
1639 	vn_finished_write(mp);
1640 out1:
1641 	vn_rangelock_unlock(vp, rl_cookie);
1642 	if (error == ERELOOKUP)
1643 		goto retry;
1644 	return (error);
1645 }
1646 
1647 /*
1648  * Truncate a file that is already locked.
1649  */
1650 int
1651 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1652     struct ucred *cred)
1653 {
1654 	struct vattr vattr;
1655 	int error;
1656 
1657 	error = VOP_ADD_WRITECOUNT(vp, 1);
1658 	if (error == 0) {
1659 		VATTR_NULL(&vattr);
1660 		vattr.va_size = length;
1661 		if (sync)
1662 			vattr.va_vaflags |= VA_SYNC;
1663 		error = VOP_SETATTR(vp, &vattr, cred);
1664 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1665 	}
1666 	return (error);
1667 }
1668 
1669 /*
1670  * File table vnode stat routine.
1671  */
1672 int
1673 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred)
1674 {
1675 	struct vnode *vp = fp->f_vnode;
1676 	int error;
1677 
1678 	vn_lock(vp, LK_SHARED | LK_RETRY);
1679 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred);
1680 	VOP_UNLOCK(vp);
1681 
1682 	return (error);
1683 }
1684 
1685 /*
1686  * File table vnode ioctl routine.
1687  */
1688 static int
1689 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1690     struct thread *td)
1691 {
1692 	struct vattr vattr;
1693 	struct vnode *vp;
1694 	struct fiobmap2_arg *bmarg;
1695 	int error;
1696 
1697 	vp = fp->f_vnode;
1698 	switch (vp->v_type) {
1699 	case VDIR:
1700 	case VREG:
1701 		switch (com) {
1702 		case FIONREAD:
1703 			vn_lock(vp, LK_SHARED | LK_RETRY);
1704 			error = VOP_GETATTR(vp, &vattr, active_cred);
1705 			VOP_UNLOCK(vp);
1706 			if (error == 0)
1707 				*(int *)data = vattr.va_size - fp->f_offset;
1708 			return (error);
1709 		case FIOBMAP2:
1710 			bmarg = (struct fiobmap2_arg *)data;
1711 			vn_lock(vp, LK_SHARED | LK_RETRY);
1712 #ifdef MAC
1713 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1714 			    vp);
1715 			if (error == 0)
1716 #endif
1717 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1718 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1719 			VOP_UNLOCK(vp);
1720 			return (error);
1721 		case FIONBIO:
1722 		case FIOASYNC:
1723 			return (0);
1724 		default:
1725 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1726 			    active_cred, td));
1727 		}
1728 		break;
1729 	case VCHR:
1730 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1731 		    active_cred, td));
1732 	default:
1733 		return (ENOTTY);
1734 	}
1735 }
1736 
1737 /*
1738  * File table vnode poll routine.
1739  */
1740 static int
1741 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1742     struct thread *td)
1743 {
1744 	struct vnode *vp;
1745 	int error;
1746 
1747 	vp = fp->f_vnode;
1748 #if defined(MAC) || defined(AUDIT)
1749 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1750 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1751 		AUDIT_ARG_VNODE1(vp);
1752 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1753 		VOP_UNLOCK(vp);
1754 		if (error != 0)
1755 			return (error);
1756 	}
1757 #endif
1758 	error = VOP_POLL(vp, events, fp->f_cred, td);
1759 	return (error);
1760 }
1761 
1762 /*
1763  * Acquire the requested lock and then check for validity.  LK_RETRY
1764  * permits vn_lock to return doomed vnodes.
1765  */
1766 static int __noinline
1767 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1768     int error)
1769 {
1770 
1771 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1772 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1773 
1774 	if (error == 0)
1775 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1776 
1777 	if ((flags & LK_RETRY) == 0) {
1778 		if (error == 0) {
1779 			VOP_UNLOCK(vp);
1780 			error = ENOENT;
1781 		}
1782 		return (error);
1783 	}
1784 
1785 	/*
1786 	 * LK_RETRY case.
1787 	 *
1788 	 * Nothing to do if we got the lock.
1789 	 */
1790 	if (error == 0)
1791 		return (0);
1792 
1793 	/*
1794 	 * Interlock was dropped by the call in _vn_lock.
1795 	 */
1796 	flags &= ~LK_INTERLOCK;
1797 	do {
1798 		error = VOP_LOCK1(vp, flags, file, line);
1799 	} while (error != 0);
1800 	return (0);
1801 }
1802 
1803 int
1804 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1805 {
1806 	int error;
1807 
1808 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1809 	    ("vn_lock: no locktype (%d passed)", flags));
1810 	VNPASS(vp->v_holdcnt > 0, vp);
1811 	error = VOP_LOCK1(vp, flags, file, line);
1812 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1813 		return (_vn_lock_fallback(vp, flags, file, line, error));
1814 	return (0);
1815 }
1816 
1817 /*
1818  * File table vnode close routine.
1819  */
1820 static int
1821 vn_closefile(struct file *fp, struct thread *td)
1822 {
1823 	struct vnode *vp;
1824 	struct flock lf;
1825 	int error;
1826 	bool ref;
1827 
1828 	vp = fp->f_vnode;
1829 	fp->f_ops = &badfileops;
1830 	ref = (fp->f_flag & FHASLOCK) != 0;
1831 
1832 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1833 
1834 	if (__predict_false(ref)) {
1835 		lf.l_whence = SEEK_SET;
1836 		lf.l_start = 0;
1837 		lf.l_len = 0;
1838 		lf.l_type = F_UNLCK;
1839 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1840 		vrele(vp);
1841 	}
1842 	return (error);
1843 }
1844 
1845 /*
1846  * Preparing to start a filesystem write operation. If the operation is
1847  * permitted, then we bump the count of operations in progress and
1848  * proceed. If a suspend request is in progress, we wait until the
1849  * suspension is over, and then proceed.
1850  */
1851 static int
1852 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1853 {
1854 	struct mount_pcpu *mpcpu;
1855 	int error, mflags;
1856 
1857 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1858 	    vfs_op_thread_enter(mp, mpcpu)) {
1859 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1860 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1861 		vfs_op_thread_exit(mp, mpcpu);
1862 		return (0);
1863 	}
1864 
1865 	if (mplocked)
1866 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1867 	else
1868 		MNT_ILOCK(mp);
1869 
1870 	error = 0;
1871 
1872 	/*
1873 	 * Check on status of suspension.
1874 	 */
1875 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1876 	    mp->mnt_susp_owner != curthread) {
1877 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1878 		    (flags & PCATCH) : 0) | (PUSER - 1);
1879 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1880 			if (flags & V_NOWAIT) {
1881 				error = EWOULDBLOCK;
1882 				goto unlock;
1883 			}
1884 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1885 			    "suspfs", 0);
1886 			if (error)
1887 				goto unlock;
1888 		}
1889 	}
1890 	if (flags & V_XSLEEP)
1891 		goto unlock;
1892 	mp->mnt_writeopcount++;
1893 unlock:
1894 	if (error != 0 || (flags & V_XSLEEP) != 0)
1895 		MNT_REL(mp);
1896 	MNT_IUNLOCK(mp);
1897 	return (error);
1898 }
1899 
1900 int
1901 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1902 {
1903 	struct mount *mp;
1904 	int error;
1905 
1906 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1907 	    ("V_MNTREF requires mp"));
1908 
1909 	error = 0;
1910 	/*
1911 	 * If a vnode is provided, get and return the mount point that
1912 	 * to which it will write.
1913 	 */
1914 	if (vp != NULL) {
1915 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1916 			*mpp = NULL;
1917 			if (error != EOPNOTSUPP)
1918 				return (error);
1919 			return (0);
1920 		}
1921 	}
1922 	if ((mp = *mpp) == NULL)
1923 		return (0);
1924 
1925 	/*
1926 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1927 	 * a vfs_ref().
1928 	 * As long as a vnode is not provided we need to acquire a
1929 	 * refcount for the provided mountpoint too, in order to
1930 	 * emulate a vfs_ref().
1931 	 */
1932 	if (vp == NULL && (flags & V_MNTREF) == 0)
1933 		vfs_ref(mp);
1934 
1935 	return (vn_start_write_refed(mp, flags, false));
1936 }
1937 
1938 /*
1939  * Secondary suspension. Used by operations such as vop_inactive
1940  * routines that are needed by the higher level functions. These
1941  * are allowed to proceed until all the higher level functions have
1942  * completed (indicated by mnt_writeopcount dropping to zero). At that
1943  * time, these operations are halted until the suspension is over.
1944  */
1945 int
1946 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1947 {
1948 	struct mount *mp;
1949 	int error;
1950 
1951 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1952 	    ("V_MNTREF requires mp"));
1953 
1954  retry:
1955 	if (vp != NULL) {
1956 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1957 			*mpp = NULL;
1958 			if (error != EOPNOTSUPP)
1959 				return (error);
1960 			return (0);
1961 		}
1962 	}
1963 	/*
1964 	 * If we are not suspended or have not yet reached suspended
1965 	 * mode, then let the operation proceed.
1966 	 */
1967 	if ((mp = *mpp) == NULL)
1968 		return (0);
1969 
1970 	/*
1971 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1972 	 * a vfs_ref().
1973 	 * As long as a vnode is not provided we need to acquire a
1974 	 * refcount for the provided mountpoint too, in order to
1975 	 * emulate a vfs_ref().
1976 	 */
1977 	MNT_ILOCK(mp);
1978 	if (vp == NULL && (flags & V_MNTREF) == 0)
1979 		MNT_REF(mp);
1980 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1981 		mp->mnt_secondary_writes++;
1982 		mp->mnt_secondary_accwrites++;
1983 		MNT_IUNLOCK(mp);
1984 		return (0);
1985 	}
1986 	if (flags & V_NOWAIT) {
1987 		MNT_REL(mp);
1988 		MNT_IUNLOCK(mp);
1989 		return (EWOULDBLOCK);
1990 	}
1991 	/*
1992 	 * Wait for the suspension to finish.
1993 	 */
1994 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1995 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1996 	    "suspfs", 0);
1997 	vfs_rel(mp);
1998 	if (error == 0)
1999 		goto retry;
2000 	return (error);
2001 }
2002 
2003 /*
2004  * Filesystem write operation has completed. If we are suspending and this
2005  * operation is the last one, notify the suspender that the suspension is
2006  * now in effect.
2007  */
2008 void
2009 vn_finished_write(struct mount *mp)
2010 {
2011 	struct mount_pcpu *mpcpu;
2012 	int c;
2013 
2014 	if (mp == NULL)
2015 		return;
2016 
2017 	if (vfs_op_thread_enter(mp, mpcpu)) {
2018 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2019 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2020 		vfs_op_thread_exit(mp, mpcpu);
2021 		return;
2022 	}
2023 
2024 	MNT_ILOCK(mp);
2025 	vfs_assert_mount_counters(mp);
2026 	MNT_REL(mp);
2027 	c = --mp->mnt_writeopcount;
2028 	if (mp->mnt_vfs_ops == 0) {
2029 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2030 		MNT_IUNLOCK(mp);
2031 		return;
2032 	}
2033 	if (c < 0)
2034 		vfs_dump_mount_counters(mp);
2035 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2036 		wakeup(&mp->mnt_writeopcount);
2037 	MNT_IUNLOCK(mp);
2038 }
2039 
2040 /*
2041  * Filesystem secondary write operation has completed. If we are
2042  * suspending and this operation is the last one, notify the suspender
2043  * that the suspension is now in effect.
2044  */
2045 void
2046 vn_finished_secondary_write(struct mount *mp)
2047 {
2048 	if (mp == NULL)
2049 		return;
2050 	MNT_ILOCK(mp);
2051 	MNT_REL(mp);
2052 	mp->mnt_secondary_writes--;
2053 	if (mp->mnt_secondary_writes < 0)
2054 		panic("vn_finished_secondary_write: neg cnt");
2055 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2056 	    mp->mnt_secondary_writes <= 0)
2057 		wakeup(&mp->mnt_secondary_writes);
2058 	MNT_IUNLOCK(mp);
2059 }
2060 
2061 /*
2062  * Request a filesystem to suspend write operations.
2063  */
2064 int
2065 vfs_write_suspend(struct mount *mp, int flags)
2066 {
2067 	int error;
2068 
2069 	vfs_op_enter(mp);
2070 
2071 	MNT_ILOCK(mp);
2072 	vfs_assert_mount_counters(mp);
2073 	if (mp->mnt_susp_owner == curthread) {
2074 		vfs_op_exit_locked(mp);
2075 		MNT_IUNLOCK(mp);
2076 		return (EALREADY);
2077 	}
2078 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2079 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2080 
2081 	/*
2082 	 * Unmount holds a write reference on the mount point.  If we
2083 	 * own busy reference and drain for writers, we deadlock with
2084 	 * the reference draining in the unmount path.  Callers of
2085 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2086 	 * vfs_busy() reference is owned and caller is not in the
2087 	 * unmount context.
2088 	 */
2089 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2090 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2091 		vfs_op_exit_locked(mp);
2092 		MNT_IUNLOCK(mp);
2093 		return (EBUSY);
2094 	}
2095 
2096 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2097 	mp->mnt_susp_owner = curthread;
2098 	if (mp->mnt_writeopcount > 0)
2099 		(void) msleep(&mp->mnt_writeopcount,
2100 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2101 	else
2102 		MNT_IUNLOCK(mp);
2103 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2104 		vfs_write_resume(mp, 0);
2105 		/* vfs_write_resume does vfs_op_exit() for us */
2106 	}
2107 	return (error);
2108 }
2109 
2110 /*
2111  * Request a filesystem to resume write operations.
2112  */
2113 void
2114 vfs_write_resume(struct mount *mp, int flags)
2115 {
2116 
2117 	MNT_ILOCK(mp);
2118 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2119 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2120 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2121 				       MNTK_SUSPENDED);
2122 		mp->mnt_susp_owner = NULL;
2123 		wakeup(&mp->mnt_writeopcount);
2124 		wakeup(&mp->mnt_flag);
2125 		curthread->td_pflags &= ~TDP_IGNSUSP;
2126 		if ((flags & VR_START_WRITE) != 0) {
2127 			MNT_REF(mp);
2128 			mp->mnt_writeopcount++;
2129 		}
2130 		MNT_IUNLOCK(mp);
2131 		if ((flags & VR_NO_SUSPCLR) == 0)
2132 			VFS_SUSP_CLEAN(mp);
2133 		vfs_op_exit(mp);
2134 	} else if ((flags & VR_START_WRITE) != 0) {
2135 		MNT_REF(mp);
2136 		vn_start_write_refed(mp, 0, true);
2137 	} else {
2138 		MNT_IUNLOCK(mp);
2139 	}
2140 }
2141 
2142 /*
2143  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2144  * methods.
2145  */
2146 int
2147 vfs_write_suspend_umnt(struct mount *mp)
2148 {
2149 	int error;
2150 
2151 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2152 	    ("vfs_write_suspend_umnt: recursed"));
2153 
2154 	/* dounmount() already called vn_start_write(). */
2155 	for (;;) {
2156 		vn_finished_write(mp);
2157 		error = vfs_write_suspend(mp, 0);
2158 		if (error != 0) {
2159 			vn_start_write(NULL, &mp, V_WAIT);
2160 			return (error);
2161 		}
2162 		MNT_ILOCK(mp);
2163 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2164 			break;
2165 		MNT_IUNLOCK(mp);
2166 		vn_start_write(NULL, &mp, V_WAIT);
2167 	}
2168 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2169 	wakeup(&mp->mnt_flag);
2170 	MNT_IUNLOCK(mp);
2171 	curthread->td_pflags |= TDP_IGNSUSP;
2172 	return (0);
2173 }
2174 
2175 /*
2176  * Implement kqueues for files by translating it to vnode operation.
2177  */
2178 static int
2179 vn_kqfilter(struct file *fp, struct knote *kn)
2180 {
2181 
2182 	return (VOP_KQFILTER(fp->f_vnode, kn));
2183 }
2184 
2185 int
2186 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2187 {
2188 	if ((fp->f_flag & FKQALLOWED) == 0)
2189 		return (EBADF);
2190 	return (vn_kqfilter(fp, kn));
2191 }
2192 
2193 /*
2194  * Simplified in-kernel wrapper calls for extended attribute access.
2195  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2196  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2197  */
2198 int
2199 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2200     const char *attrname, int *buflen, char *buf, struct thread *td)
2201 {
2202 	struct uio	auio;
2203 	struct iovec	iov;
2204 	int	error;
2205 
2206 	iov.iov_len = *buflen;
2207 	iov.iov_base = buf;
2208 
2209 	auio.uio_iov = &iov;
2210 	auio.uio_iovcnt = 1;
2211 	auio.uio_rw = UIO_READ;
2212 	auio.uio_segflg = UIO_SYSSPACE;
2213 	auio.uio_td = td;
2214 	auio.uio_offset = 0;
2215 	auio.uio_resid = *buflen;
2216 
2217 	if ((ioflg & IO_NODELOCKED) == 0)
2218 		vn_lock(vp, LK_SHARED | LK_RETRY);
2219 
2220 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2221 
2222 	/* authorize attribute retrieval as kernel */
2223 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2224 	    td);
2225 
2226 	if ((ioflg & IO_NODELOCKED) == 0)
2227 		VOP_UNLOCK(vp);
2228 
2229 	if (error == 0) {
2230 		*buflen = *buflen - auio.uio_resid;
2231 	}
2232 
2233 	return (error);
2234 }
2235 
2236 /*
2237  * XXX failure mode if partially written?
2238  */
2239 int
2240 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2241     const char *attrname, int buflen, char *buf, struct thread *td)
2242 {
2243 	struct uio	auio;
2244 	struct iovec	iov;
2245 	struct mount	*mp;
2246 	int	error;
2247 
2248 	iov.iov_len = buflen;
2249 	iov.iov_base = buf;
2250 
2251 	auio.uio_iov = &iov;
2252 	auio.uio_iovcnt = 1;
2253 	auio.uio_rw = UIO_WRITE;
2254 	auio.uio_segflg = UIO_SYSSPACE;
2255 	auio.uio_td = td;
2256 	auio.uio_offset = 0;
2257 	auio.uio_resid = buflen;
2258 
2259 	if ((ioflg & IO_NODELOCKED) == 0) {
2260 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2261 			return (error);
2262 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2263 	}
2264 
2265 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2266 
2267 	/* authorize attribute setting as kernel */
2268 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2269 
2270 	if ((ioflg & IO_NODELOCKED) == 0) {
2271 		vn_finished_write(mp);
2272 		VOP_UNLOCK(vp);
2273 	}
2274 
2275 	return (error);
2276 }
2277 
2278 int
2279 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2280     const char *attrname, struct thread *td)
2281 {
2282 	struct mount	*mp;
2283 	int	error;
2284 
2285 	if ((ioflg & IO_NODELOCKED) == 0) {
2286 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2287 			return (error);
2288 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2289 	}
2290 
2291 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2292 
2293 	/* authorize attribute removal as kernel */
2294 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2295 	if (error == EOPNOTSUPP)
2296 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2297 		    NULL, td);
2298 
2299 	if ((ioflg & IO_NODELOCKED) == 0) {
2300 		vn_finished_write(mp);
2301 		VOP_UNLOCK(vp);
2302 	}
2303 
2304 	return (error);
2305 }
2306 
2307 static int
2308 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2309     struct vnode **rvp)
2310 {
2311 
2312 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2313 }
2314 
2315 int
2316 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2317 {
2318 
2319 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2320 	    lkflags, rvp));
2321 }
2322 
2323 int
2324 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2325     int lkflags, struct vnode **rvp)
2326 {
2327 	struct mount *mp;
2328 	int ltype, error;
2329 
2330 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2331 	mp = vp->v_mount;
2332 	ltype = VOP_ISLOCKED(vp);
2333 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2334 	    ("vn_vget_ino: vp not locked"));
2335 	error = vfs_busy(mp, MBF_NOWAIT);
2336 	if (error != 0) {
2337 		vfs_ref(mp);
2338 		VOP_UNLOCK(vp);
2339 		error = vfs_busy(mp, 0);
2340 		vn_lock(vp, ltype | LK_RETRY);
2341 		vfs_rel(mp);
2342 		if (error != 0)
2343 			return (ENOENT);
2344 		if (VN_IS_DOOMED(vp)) {
2345 			vfs_unbusy(mp);
2346 			return (ENOENT);
2347 		}
2348 	}
2349 	VOP_UNLOCK(vp);
2350 	error = alloc(mp, alloc_arg, lkflags, rvp);
2351 	vfs_unbusy(mp);
2352 	if (error != 0 || *rvp != vp)
2353 		vn_lock(vp, ltype | LK_RETRY);
2354 	if (VN_IS_DOOMED(vp)) {
2355 		if (error == 0) {
2356 			if (*rvp == vp)
2357 				vunref(vp);
2358 			else
2359 				vput(*rvp);
2360 		}
2361 		error = ENOENT;
2362 	}
2363 	return (error);
2364 }
2365 
2366 int
2367 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2368     struct thread *td)
2369 {
2370 	off_t lim;
2371 	bool ktr_write;
2372 
2373 	if (td == NULL)
2374 		return (0);
2375 
2376 	/*
2377 	 * There are conditions where the limit is to be ignored.
2378 	 * However, since it is almost never reached, check it first.
2379 	 */
2380 	ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2381 	lim = lim_cur(td, RLIMIT_FSIZE);
2382 	if (__predict_false(ktr_write))
2383 		lim = td->td_ktr_io_lim;
2384 	if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2385 		return (0);
2386 
2387 	/*
2388 	 * The limit is reached.
2389 	 */
2390 	if (vp->v_type != VREG ||
2391 	    (td->td_pflags2 & TDP2_ACCT) != 0)
2392 		return (0);
2393 
2394 	if (!ktr_write || ktr_filesize_limit_signal) {
2395 		PROC_LOCK(td->td_proc);
2396 		kern_psignal(td->td_proc, SIGXFSZ);
2397 		PROC_UNLOCK(td->td_proc);
2398 	}
2399 	return (EFBIG);
2400 }
2401 
2402 int
2403 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2404     struct thread *td)
2405 {
2406 	struct vnode *vp;
2407 
2408 	vp = fp->f_vnode;
2409 #ifdef AUDIT
2410 	vn_lock(vp, LK_SHARED | LK_RETRY);
2411 	AUDIT_ARG_VNODE1(vp);
2412 	VOP_UNLOCK(vp);
2413 #endif
2414 	return (setfmode(td, active_cred, vp, mode));
2415 }
2416 
2417 int
2418 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2419     struct thread *td)
2420 {
2421 	struct vnode *vp;
2422 
2423 	vp = fp->f_vnode;
2424 #ifdef AUDIT
2425 	vn_lock(vp, LK_SHARED | LK_RETRY);
2426 	AUDIT_ARG_VNODE1(vp);
2427 	VOP_UNLOCK(vp);
2428 #endif
2429 	return (setfown(td, active_cred, vp, uid, gid));
2430 }
2431 
2432 /*
2433  * Remove pages in the range ["start", "end") from the vnode's VM object.  If
2434  * "end" is 0, then the range extends to the end of the object.
2435  */
2436 void
2437 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2438 {
2439 	vm_object_t object;
2440 
2441 	if ((object = vp->v_object) == NULL)
2442 		return;
2443 	VM_OBJECT_WLOCK(object);
2444 	vm_object_page_remove(object, start, end, 0);
2445 	VM_OBJECT_WUNLOCK(object);
2446 }
2447 
2448 /*
2449  * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2450  * mapped into any process' address space.  Filesystems may use this in
2451  * preference to vn_pages_remove() to avoid blocking on pages busied in
2452  * preparation for a VOP_GETPAGES.
2453  */
2454 void
2455 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2456 {
2457 	vm_object_t object;
2458 
2459 	if ((object = vp->v_object) == NULL)
2460 		return;
2461 	VM_OBJECT_WLOCK(object);
2462 	vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2463 	VM_OBJECT_WUNLOCK(object);
2464 }
2465 
2466 int
2467 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2468     struct ucred *cred)
2469 {
2470 	struct vattr va;
2471 	daddr_t bn, bnp;
2472 	uint64_t bsize;
2473 	off_t noff;
2474 	int error;
2475 
2476 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2477 	    ("%s: Wrong command %lu", __func__, cmd));
2478 	ASSERT_VOP_LOCKED(vp, "vn_bmap_seekhole_locked");
2479 
2480 	if (vp->v_type != VREG) {
2481 		error = ENOTTY;
2482 		goto out;
2483 	}
2484 	error = VOP_GETATTR(vp, &va, cred);
2485 	if (error != 0)
2486 		goto out;
2487 	noff = *off;
2488 	if (noff >= va.va_size) {
2489 		error = ENXIO;
2490 		goto out;
2491 	}
2492 	bsize = vp->v_mount->mnt_stat.f_iosize;
2493 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2494 	    noff % bsize) {
2495 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2496 		if (error == EOPNOTSUPP) {
2497 			error = ENOTTY;
2498 			goto out;
2499 		}
2500 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2501 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2502 			noff = bn * bsize;
2503 			if (noff < *off)
2504 				noff = *off;
2505 			goto out;
2506 		}
2507 	}
2508 	if (noff > va.va_size)
2509 		noff = va.va_size;
2510 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2511 	if (cmd == FIOSEEKDATA)
2512 		error = ENXIO;
2513 out:
2514 	if (error == 0)
2515 		*off = noff;
2516 	return (error);
2517 }
2518 
2519 int
2520 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2521 {
2522 	int error;
2523 
2524 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2525 	    ("%s: Wrong command %lu", __func__, cmd));
2526 
2527 	if (vn_lock(vp, LK_SHARED) != 0)
2528 		return (EBADF);
2529 	error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2530 	VOP_UNLOCK(vp);
2531 	return (error);
2532 }
2533 
2534 int
2535 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2536 {
2537 	struct ucred *cred;
2538 	struct vnode *vp;
2539 	struct vattr vattr;
2540 	off_t foffset, size;
2541 	int error, noneg;
2542 
2543 	cred = td->td_ucred;
2544 	vp = fp->f_vnode;
2545 	foffset = foffset_lock(fp, 0);
2546 	noneg = (vp->v_type != VCHR);
2547 	error = 0;
2548 	switch (whence) {
2549 	case L_INCR:
2550 		if (noneg &&
2551 		    (foffset < 0 ||
2552 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2553 			error = EOVERFLOW;
2554 			break;
2555 		}
2556 		offset += foffset;
2557 		break;
2558 	case L_XTND:
2559 		vn_lock(vp, LK_SHARED | LK_RETRY);
2560 		error = VOP_GETATTR(vp, &vattr, cred);
2561 		VOP_UNLOCK(vp);
2562 		if (error)
2563 			break;
2564 
2565 		/*
2566 		 * If the file references a disk device, then fetch
2567 		 * the media size and use that to determine the ending
2568 		 * offset.
2569 		 */
2570 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2571 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2572 			vattr.va_size = size;
2573 		if (noneg &&
2574 		    (vattr.va_size > OFF_MAX ||
2575 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2576 			error = EOVERFLOW;
2577 			break;
2578 		}
2579 		offset += vattr.va_size;
2580 		break;
2581 	case L_SET:
2582 		break;
2583 	case SEEK_DATA:
2584 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2585 		if (error == ENOTTY)
2586 			error = EINVAL;
2587 		break;
2588 	case SEEK_HOLE:
2589 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2590 		if (error == ENOTTY)
2591 			error = EINVAL;
2592 		break;
2593 	default:
2594 		error = EINVAL;
2595 	}
2596 	if (error == 0 && noneg && offset < 0)
2597 		error = EINVAL;
2598 	if (error != 0)
2599 		goto drop;
2600 	VFS_KNOTE_UNLOCKED(vp, 0);
2601 	td->td_uretoff.tdu_off = offset;
2602 drop:
2603 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2604 	return (error);
2605 }
2606 
2607 int
2608 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2609     struct thread *td)
2610 {
2611 	int error;
2612 
2613 	/*
2614 	 * Grant permission if the caller is the owner of the file, or
2615 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2616 	 * on the file.  If the time pointer is null, then write
2617 	 * permission on the file is also sufficient.
2618 	 *
2619 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2620 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2621 	 * will be allowed to set the times [..] to the current
2622 	 * server time.
2623 	 */
2624 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2625 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2626 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2627 	return (error);
2628 }
2629 
2630 int
2631 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2632 {
2633 	struct vnode *vp;
2634 	int error;
2635 
2636 	if (fp->f_type == DTYPE_FIFO)
2637 		kif->kf_type = KF_TYPE_FIFO;
2638 	else
2639 		kif->kf_type = KF_TYPE_VNODE;
2640 	vp = fp->f_vnode;
2641 	vref(vp);
2642 	FILEDESC_SUNLOCK(fdp);
2643 	error = vn_fill_kinfo_vnode(vp, kif);
2644 	vrele(vp);
2645 	FILEDESC_SLOCK(fdp);
2646 	return (error);
2647 }
2648 
2649 static inline void
2650 vn_fill_junk(struct kinfo_file *kif)
2651 {
2652 	size_t len, olen;
2653 
2654 	/*
2655 	 * Simulate vn_fullpath returning changing values for a given
2656 	 * vp during e.g. coredump.
2657 	 */
2658 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2659 	olen = strlen(kif->kf_path);
2660 	if (len < olen)
2661 		strcpy(&kif->kf_path[len - 1], "$");
2662 	else
2663 		for (; olen < len; olen++)
2664 			strcpy(&kif->kf_path[olen], "A");
2665 }
2666 
2667 int
2668 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2669 {
2670 	struct vattr va;
2671 	char *fullpath, *freepath;
2672 	int error;
2673 
2674 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2675 	freepath = NULL;
2676 	fullpath = "-";
2677 	error = vn_fullpath(vp, &fullpath, &freepath);
2678 	if (error == 0) {
2679 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2680 	}
2681 	if (freepath != NULL)
2682 		free(freepath, M_TEMP);
2683 
2684 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2685 		vn_fill_junk(kif);
2686 	);
2687 
2688 	/*
2689 	 * Retrieve vnode attributes.
2690 	 */
2691 	va.va_fsid = VNOVAL;
2692 	va.va_rdev = NODEV;
2693 	vn_lock(vp, LK_SHARED | LK_RETRY);
2694 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2695 	VOP_UNLOCK(vp);
2696 	if (error != 0)
2697 		return (error);
2698 	if (va.va_fsid != VNOVAL)
2699 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2700 	else
2701 		kif->kf_un.kf_file.kf_file_fsid =
2702 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2703 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2704 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2705 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2706 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2707 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2708 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2709 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2710 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2711 	return (0);
2712 }
2713 
2714 int
2715 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2716     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2717     struct thread *td)
2718 {
2719 #ifdef HWPMC_HOOKS
2720 	struct pmckern_map_in pkm;
2721 #endif
2722 	struct mount *mp;
2723 	struct vnode *vp;
2724 	vm_object_t object;
2725 	vm_prot_t maxprot;
2726 	boolean_t writecounted;
2727 	int error;
2728 
2729 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2730     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2731 	/*
2732 	 * POSIX shared-memory objects are defined to have
2733 	 * kernel persistence, and are not defined to support
2734 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2735 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2736 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2737 	 * flag to request this behavior.
2738 	 */
2739 	if ((fp->f_flag & FPOSIXSHM) != 0)
2740 		flags |= MAP_NOSYNC;
2741 #endif
2742 	vp = fp->f_vnode;
2743 
2744 	/*
2745 	 * Ensure that file and memory protections are
2746 	 * compatible.  Note that we only worry about
2747 	 * writability if mapping is shared; in this case,
2748 	 * current and max prot are dictated by the open file.
2749 	 * XXX use the vnode instead?  Problem is: what
2750 	 * credentials do we use for determination? What if
2751 	 * proc does a setuid?
2752 	 */
2753 	mp = vp->v_mount;
2754 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2755 		maxprot = VM_PROT_NONE;
2756 		if ((prot & VM_PROT_EXECUTE) != 0)
2757 			return (EACCES);
2758 	} else
2759 		maxprot = VM_PROT_EXECUTE;
2760 	if ((fp->f_flag & FREAD) != 0)
2761 		maxprot |= VM_PROT_READ;
2762 	else if ((prot & VM_PROT_READ) != 0)
2763 		return (EACCES);
2764 
2765 	/*
2766 	 * If we are sharing potential changes via MAP_SHARED and we
2767 	 * are trying to get write permission although we opened it
2768 	 * without asking for it, bail out.
2769 	 */
2770 	if ((flags & MAP_SHARED) != 0) {
2771 		if ((fp->f_flag & FWRITE) != 0)
2772 			maxprot |= VM_PROT_WRITE;
2773 		else if ((prot & VM_PROT_WRITE) != 0)
2774 			return (EACCES);
2775 	} else {
2776 		maxprot |= VM_PROT_WRITE;
2777 		cap_maxprot |= VM_PROT_WRITE;
2778 	}
2779 	maxprot &= cap_maxprot;
2780 
2781 	/*
2782 	 * For regular files and shared memory, POSIX requires that
2783 	 * the value of foff be a legitimate offset within the data
2784 	 * object.  In particular, negative offsets are invalid.
2785 	 * Blocking negative offsets and overflows here avoids
2786 	 * possible wraparound or user-level access into reserved
2787 	 * ranges of the data object later.  In contrast, POSIX does
2788 	 * not dictate how offsets are used by device drivers, so in
2789 	 * the case of a device mapping a negative offset is passed
2790 	 * on.
2791 	 */
2792 	if (
2793 #ifdef _LP64
2794 	    size > OFF_MAX ||
2795 #endif
2796 	    foff > OFF_MAX - size)
2797 		return (EINVAL);
2798 
2799 	writecounted = FALSE;
2800 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2801 	    &foff, &object, &writecounted);
2802 	if (error != 0)
2803 		return (error);
2804 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2805 	    foff, writecounted, td);
2806 	if (error != 0) {
2807 		/*
2808 		 * If this mapping was accounted for in the vnode's
2809 		 * writecount, then undo that now.
2810 		 */
2811 		if (writecounted)
2812 			vm_pager_release_writecount(object, 0, size);
2813 		vm_object_deallocate(object);
2814 	}
2815 #ifdef HWPMC_HOOKS
2816 	/* Inform hwpmc(4) if an executable is being mapped. */
2817 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2818 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2819 			pkm.pm_file = vp;
2820 			pkm.pm_address = (uintptr_t) *addr;
2821 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2822 		}
2823 	}
2824 #endif
2825 	return (error);
2826 }
2827 
2828 void
2829 vn_fsid(struct vnode *vp, struct vattr *va)
2830 {
2831 	fsid_t *f;
2832 
2833 	f = &vp->v_mount->mnt_stat.f_fsid;
2834 	va->va_fsid = (uint32_t)f->val[1];
2835 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2836 	va->va_fsid += (uint32_t)f->val[0];
2837 }
2838 
2839 int
2840 vn_fsync_buf(struct vnode *vp, int waitfor)
2841 {
2842 	struct buf *bp, *nbp;
2843 	struct bufobj *bo;
2844 	struct mount *mp;
2845 	int error, maxretry;
2846 
2847 	error = 0;
2848 	maxretry = 10000;     /* large, arbitrarily chosen */
2849 	mp = NULL;
2850 	if (vp->v_type == VCHR) {
2851 		VI_LOCK(vp);
2852 		mp = vp->v_rdev->si_mountpt;
2853 		VI_UNLOCK(vp);
2854 	}
2855 	bo = &vp->v_bufobj;
2856 	BO_LOCK(bo);
2857 loop1:
2858 	/*
2859 	 * MARK/SCAN initialization to avoid infinite loops.
2860 	 */
2861         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2862 		bp->b_vflags &= ~BV_SCANNED;
2863 		bp->b_error = 0;
2864 	}
2865 
2866 	/*
2867 	 * Flush all dirty buffers associated with a vnode.
2868 	 */
2869 loop2:
2870 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2871 		if ((bp->b_vflags & BV_SCANNED) != 0)
2872 			continue;
2873 		bp->b_vflags |= BV_SCANNED;
2874 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2875 			if (waitfor != MNT_WAIT)
2876 				continue;
2877 			if (BUF_LOCK(bp,
2878 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2879 			    BO_LOCKPTR(bo)) != 0) {
2880 				BO_LOCK(bo);
2881 				goto loop1;
2882 			}
2883 			BO_LOCK(bo);
2884 		}
2885 		BO_UNLOCK(bo);
2886 		KASSERT(bp->b_bufobj == bo,
2887 		    ("bp %p wrong b_bufobj %p should be %p",
2888 		    bp, bp->b_bufobj, bo));
2889 		if ((bp->b_flags & B_DELWRI) == 0)
2890 			panic("fsync: not dirty");
2891 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2892 			vfs_bio_awrite(bp);
2893 		} else {
2894 			bremfree(bp);
2895 			bawrite(bp);
2896 		}
2897 		if (maxretry < 1000)
2898 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2899 		BO_LOCK(bo);
2900 		goto loop2;
2901 	}
2902 
2903 	/*
2904 	 * If synchronous the caller expects us to completely resolve all
2905 	 * dirty buffers in the system.  Wait for in-progress I/O to
2906 	 * complete (which could include background bitmap writes), then
2907 	 * retry if dirty blocks still exist.
2908 	 */
2909 	if (waitfor == MNT_WAIT) {
2910 		bufobj_wwait(bo, 0, 0);
2911 		if (bo->bo_dirty.bv_cnt > 0) {
2912 			/*
2913 			 * If we are unable to write any of these buffers
2914 			 * then we fail now rather than trying endlessly
2915 			 * to write them out.
2916 			 */
2917 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2918 				if ((error = bp->b_error) != 0)
2919 					break;
2920 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2921 			    (error == 0 && --maxretry >= 0))
2922 				goto loop1;
2923 			if (error == 0)
2924 				error = EAGAIN;
2925 		}
2926 	}
2927 	BO_UNLOCK(bo);
2928 	if (error != 0)
2929 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2930 
2931 	return (error);
2932 }
2933 
2934 /*
2935  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2936  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2937  * to do the actual copy.
2938  * vn_generic_copy_file_range() is factored out, so it can be called
2939  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2940  * different file systems.
2941  */
2942 int
2943 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2944     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2945     struct ucred *outcred, struct thread *fsize_td)
2946 {
2947 	int error;
2948 	size_t len;
2949 	uint64_t uval;
2950 
2951 	len = *lenp;
2952 	*lenp = 0;		/* For error returns. */
2953 	error = 0;
2954 
2955 	/* Do some sanity checks on the arguments. */
2956 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2957 		error = EISDIR;
2958 	else if (*inoffp < 0 || *outoffp < 0 ||
2959 	    invp->v_type != VREG || outvp->v_type != VREG)
2960 		error = EINVAL;
2961 	if (error != 0)
2962 		goto out;
2963 
2964 	/* Ensure offset + len does not wrap around. */
2965 	uval = *inoffp;
2966 	uval += len;
2967 	if (uval > INT64_MAX)
2968 		len = INT64_MAX - *inoffp;
2969 	uval = *outoffp;
2970 	uval += len;
2971 	if (uval > INT64_MAX)
2972 		len = INT64_MAX - *outoffp;
2973 	if (len == 0)
2974 		goto out;
2975 
2976 	/*
2977 	 * If the two vnode are for the same file system, call
2978 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2979 	 * which can handle copies across multiple file systems.
2980 	 */
2981 	*lenp = len;
2982 	if (invp->v_mount == outvp->v_mount)
2983 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2984 		    lenp, flags, incred, outcred, fsize_td);
2985 	else
2986 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2987 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2988 out:
2989 	return (error);
2990 }
2991 
2992 /*
2993  * Test len bytes of data starting at dat for all bytes == 0.
2994  * Return true if all bytes are zero, false otherwise.
2995  * Expects dat to be well aligned.
2996  */
2997 static bool
2998 mem_iszero(void *dat, int len)
2999 {
3000 	int i;
3001 	const u_int *p;
3002 	const char *cp;
3003 
3004 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
3005 		if (len >= sizeof(*p)) {
3006 			if (*p != 0)
3007 				return (false);
3008 		} else {
3009 			cp = (const char *)p;
3010 			for (i = 0; i < len; i++, cp++)
3011 				if (*cp != '\0')
3012 					return (false);
3013 		}
3014 	}
3015 	return (true);
3016 }
3017 
3018 /*
3019  * Look for a hole in the output file and, if found, adjust *outoffp
3020  * and *xferp to skip past the hole.
3021  * *xferp is the entire hole length to be written and xfer2 is how many bytes
3022  * to be written as 0's upon return.
3023  */
3024 static off_t
3025 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3026     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3027 {
3028 	int error;
3029 	off_t delta;
3030 
3031 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3032 		*dataoffp = *outoffp;
3033 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3034 		    curthread);
3035 		if (error == 0) {
3036 			*holeoffp = *dataoffp;
3037 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3038 			    curthread);
3039 		}
3040 		if (error != 0 || *holeoffp == *dataoffp) {
3041 			/*
3042 			 * Since outvp is unlocked, it may be possible for
3043 			 * another thread to do a truncate(), lseek(), write()
3044 			 * creating a hole at startoff between the above
3045 			 * VOP_IOCTL() calls, if the other thread does not do
3046 			 * rangelocking.
3047 			 * If that happens, *holeoffp == *dataoffp and finding
3048 			 * the hole has failed, so disable vn_skip_hole().
3049 			 */
3050 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
3051 			return (xfer2);
3052 		}
3053 		KASSERT(*dataoffp >= *outoffp,
3054 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3055 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
3056 		KASSERT(*holeoffp > *dataoffp,
3057 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3058 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3059 	}
3060 
3061 	/*
3062 	 * If there is a hole before the data starts, advance *outoffp and
3063 	 * *xferp past the hole.
3064 	 */
3065 	if (*dataoffp > *outoffp) {
3066 		delta = *dataoffp - *outoffp;
3067 		if (delta >= *xferp) {
3068 			/* Entire *xferp is a hole. */
3069 			*outoffp += *xferp;
3070 			*xferp = 0;
3071 			return (0);
3072 		}
3073 		*xferp -= delta;
3074 		*outoffp += delta;
3075 		xfer2 = MIN(xfer2, *xferp);
3076 	}
3077 
3078 	/*
3079 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3080 	 * that the write ends at the start of the hole.
3081 	 * *holeoffp should always be greater than *outoffp, but for the
3082 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3083 	 * value.
3084 	 */
3085 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3086 		xfer2 = *holeoffp - *outoffp;
3087 	return (xfer2);
3088 }
3089 
3090 /*
3091  * Write an xfer sized chunk to outvp in blksize blocks from dat.
3092  * dat is a maximum of blksize in length and can be written repeatedly in
3093  * the chunk.
3094  * If growfile == true, just grow the file via vn_truncate_locked() instead
3095  * of doing actual writes.
3096  * If checkhole == true, a hole is being punched, so skip over any hole
3097  * already in the output file.
3098  */
3099 static int
3100 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3101     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3102 {
3103 	struct mount *mp;
3104 	off_t dataoff, holeoff, xfer2;
3105 	int error;
3106 
3107 	/*
3108 	 * Loop around doing writes of blksize until write has been completed.
3109 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3110 	 * done for each iteration, since the xfer argument can be very
3111 	 * large if there is a large hole to punch in the output file.
3112 	 */
3113 	error = 0;
3114 	holeoff = 0;
3115 	do {
3116 		xfer2 = MIN(xfer, blksize);
3117 		if (checkhole) {
3118 			/*
3119 			 * Punching a hole.  Skip writing if there is
3120 			 * already a hole in the output file.
3121 			 */
3122 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3123 			    &dataoff, &holeoff, cred);
3124 			if (xfer == 0)
3125 				break;
3126 			if (holeoff < 0)
3127 				checkhole = false;
3128 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3129 			    (intmax_t)xfer2));
3130 		}
3131 		bwillwrite();
3132 		mp = NULL;
3133 		error = vn_start_write(outvp, &mp, V_WAIT);
3134 		if (error != 0)
3135 			break;
3136 		if (growfile) {
3137 			error = vn_lock(outvp, LK_EXCLUSIVE);
3138 			if (error == 0) {
3139 				error = vn_truncate_locked(outvp, outoff + xfer,
3140 				    false, cred);
3141 				VOP_UNLOCK(outvp);
3142 			}
3143 		} else {
3144 			error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3145 			if (error == 0) {
3146 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3147 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3148 				    curthread->td_ucred, cred, NULL, curthread);
3149 				outoff += xfer2;
3150 				xfer -= xfer2;
3151 				VOP_UNLOCK(outvp);
3152 			}
3153 		}
3154 		if (mp != NULL)
3155 			vn_finished_write(mp);
3156 	} while (!growfile && xfer > 0 && error == 0);
3157 	return (error);
3158 }
3159 
3160 /*
3161  * Copy a byte range of one file to another.  This function can handle the
3162  * case where invp and outvp are on different file systems.
3163  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3164  * is no better file system specific way to do it.
3165  */
3166 int
3167 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3168     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3169     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3170 {
3171 	struct vattr va, inva;
3172 	struct mount *mp;
3173 	struct uio io;
3174 	off_t startoff, endoff, xfer, xfer2;
3175 	u_long blksize;
3176 	int error, interrupted;
3177 	bool cantseek, readzeros, eof, lastblock, holetoeof;
3178 	ssize_t aresid;
3179 	size_t copylen, len, rem, savlen;
3180 	char *dat;
3181 	long holein, holeout;
3182 	struct timespec curts, endts;
3183 
3184 	holein = holeout = 0;
3185 	savlen = len = *lenp;
3186 	error = 0;
3187 	interrupted = 0;
3188 	dat = NULL;
3189 
3190 	error = vn_lock(invp, LK_SHARED);
3191 	if (error != 0)
3192 		goto out;
3193 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3194 		holein = 0;
3195 	if (holein > 0)
3196 		error = VOP_GETATTR(invp, &inva, incred);
3197 	VOP_UNLOCK(invp);
3198 	if (error != 0)
3199 		goto out;
3200 
3201 	mp = NULL;
3202 	error = vn_start_write(outvp, &mp, V_WAIT);
3203 	if (error == 0)
3204 		error = vn_lock(outvp, LK_EXCLUSIVE);
3205 	if (error == 0) {
3206 		/*
3207 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3208 		 * now that outvp is locked.
3209 		 */
3210 		if (fsize_td != NULL) {
3211 			io.uio_offset = *outoffp;
3212 			io.uio_resid = len;
3213 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3214 			if (error != 0)
3215 				error = EFBIG;
3216 		}
3217 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3218 			holeout = 0;
3219 		/*
3220 		 * Holes that are past EOF do not need to be written as a block
3221 		 * of zero bytes.  So, truncate the output file as far as
3222 		 * possible and then use va.va_size to decide if writing 0
3223 		 * bytes is necessary in the loop below.
3224 		 */
3225 		if (error == 0)
3226 			error = VOP_GETATTR(outvp, &va, outcred);
3227 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3228 		    *outoffp + len) {
3229 #ifdef MAC
3230 			error = mac_vnode_check_write(curthread->td_ucred,
3231 			    outcred, outvp);
3232 			if (error == 0)
3233 #endif
3234 				error = vn_truncate_locked(outvp, *outoffp,
3235 				    false, outcred);
3236 			if (error == 0)
3237 				va.va_size = *outoffp;
3238 		}
3239 		VOP_UNLOCK(outvp);
3240 	}
3241 	if (mp != NULL)
3242 		vn_finished_write(mp);
3243 	if (error != 0)
3244 		goto out;
3245 
3246 	/*
3247 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3248 	 * If hole sizes aren't available, set the blksize to the larger
3249 	 * f_iosize of invp and outvp.
3250 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3251 	 * This value is clipped at 4Kbytes and 1Mbyte.
3252 	 */
3253 	blksize = MAX(holein, holeout);
3254 
3255 	/* Clip len to end at an exact multiple of hole size. */
3256 	if (blksize > 1) {
3257 		rem = *inoffp % blksize;
3258 		if (rem > 0)
3259 			rem = blksize - rem;
3260 		if (len > rem && len - rem > blksize)
3261 			len = savlen = rounddown(len - rem, blksize) + rem;
3262 	}
3263 
3264 	if (blksize <= 1)
3265 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3266 		    outvp->v_mount->mnt_stat.f_iosize);
3267 	if (blksize < 4096)
3268 		blksize = 4096;
3269 	else if (blksize > 1024 * 1024)
3270 		blksize = 1024 * 1024;
3271 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3272 
3273 	/*
3274 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3275 	 * to find holes.  Otherwise, just scan the read block for all 0s
3276 	 * in the inner loop where the data copying is done.
3277 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3278 	 * support holes on the server, but do not support FIOSEEKHOLE.
3279 	 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3280 	 * that this function should return after 1second with a partial
3281 	 * completion.
3282 	 */
3283 	if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3284 		getnanouptime(&endts);
3285 		endts.tv_sec++;
3286 	} else
3287 		timespecclear(&endts);
3288 	holetoeof = eof = false;
3289 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3290 		endoff = 0;			/* To shut up compilers. */
3291 		cantseek = true;
3292 		startoff = *inoffp;
3293 		copylen = len;
3294 
3295 		/*
3296 		 * Find the next data area.  If there is just a hole to EOF,
3297 		 * FIOSEEKDATA should fail with ENXIO.
3298 		 * (I do not know if any file system will report a hole to
3299 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3300 		 *  will fail for those file systems.)
3301 		 *
3302 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3303 		 * the code just falls through to the inner copy loop.
3304 		 */
3305 		error = EINVAL;
3306 		if (holein > 0) {
3307 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3308 			    incred, curthread);
3309 			if (error == ENXIO) {
3310 				startoff = endoff = inva.va_size;
3311 				eof = holetoeof = true;
3312 				error = 0;
3313 			}
3314 		}
3315 		if (error == 0 && !holetoeof) {
3316 			endoff = startoff;
3317 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3318 			    incred, curthread);
3319 			/*
3320 			 * Since invp is unlocked, it may be possible for
3321 			 * another thread to do a truncate(), lseek(), write()
3322 			 * creating a hole at startoff between the above
3323 			 * VOP_IOCTL() calls, if the other thread does not do
3324 			 * rangelocking.
3325 			 * If that happens, startoff == endoff and finding
3326 			 * the hole has failed, so set an error.
3327 			 */
3328 			if (error == 0 && startoff == endoff)
3329 				error = EINVAL; /* Any error. Reset to 0. */
3330 		}
3331 		if (error == 0) {
3332 			if (startoff > *inoffp) {
3333 				/* Found hole before data block. */
3334 				xfer = MIN(startoff - *inoffp, len);
3335 				if (*outoffp < va.va_size) {
3336 					/* Must write 0s to punch hole. */
3337 					xfer2 = MIN(va.va_size - *outoffp,
3338 					    xfer);
3339 					memset(dat, 0, MIN(xfer2, blksize));
3340 					error = vn_write_outvp(outvp, dat,
3341 					    *outoffp, xfer2, blksize, false,
3342 					    holeout > 0, outcred);
3343 				}
3344 
3345 				if (error == 0 && *outoffp + xfer >
3346 				    va.va_size && (xfer == len || holetoeof)) {
3347 					/* Grow output file (hole at end). */
3348 					error = vn_write_outvp(outvp, dat,
3349 					    *outoffp, xfer, blksize, true,
3350 					    false, outcred);
3351 				}
3352 				if (error == 0) {
3353 					*inoffp += xfer;
3354 					*outoffp += xfer;
3355 					len -= xfer;
3356 					if (len < savlen) {
3357 						interrupted = sig_intr();
3358 						if (timespecisset(&endts) &&
3359 						    interrupted == 0) {
3360 							getnanouptime(&curts);
3361 							if (timespeccmp(&curts,
3362 							    &endts, >=))
3363 								interrupted =
3364 								    EINTR;
3365 						}
3366 					}
3367 				}
3368 			}
3369 			copylen = MIN(len, endoff - startoff);
3370 			cantseek = false;
3371 		} else {
3372 			cantseek = true;
3373 			startoff = *inoffp;
3374 			copylen = len;
3375 			error = 0;
3376 		}
3377 
3378 		xfer = blksize;
3379 		if (cantseek) {
3380 			/*
3381 			 * Set first xfer to end at a block boundary, so that
3382 			 * holes are more likely detected in the loop below via
3383 			 * the for all bytes 0 method.
3384 			 */
3385 			xfer -= (*inoffp % blksize);
3386 		}
3387 		/* Loop copying the data block. */
3388 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3389 			if (copylen < xfer)
3390 				xfer = copylen;
3391 			error = vn_lock(invp, LK_SHARED);
3392 			if (error != 0)
3393 				goto out;
3394 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3395 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3396 			    curthread->td_ucred, incred, &aresid,
3397 			    curthread);
3398 			VOP_UNLOCK(invp);
3399 			lastblock = false;
3400 			if (error == 0 && aresid > 0) {
3401 				/* Stop the copy at EOF on the input file. */
3402 				xfer -= aresid;
3403 				eof = true;
3404 				lastblock = true;
3405 			}
3406 			if (error == 0) {
3407 				/*
3408 				 * Skip the write for holes past the initial EOF
3409 				 * of the output file, unless this is the last
3410 				 * write of the output file at EOF.
3411 				 */
3412 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3413 				    false;
3414 				if (xfer == len)
3415 					lastblock = true;
3416 				if (!cantseek || *outoffp < va.va_size ||
3417 				    lastblock || !readzeros)
3418 					error = vn_write_outvp(outvp, dat,
3419 					    *outoffp, xfer, blksize,
3420 					    readzeros && lastblock &&
3421 					    *outoffp >= va.va_size, false,
3422 					    outcred);
3423 				if (error == 0) {
3424 					*inoffp += xfer;
3425 					startoff += xfer;
3426 					*outoffp += xfer;
3427 					copylen -= xfer;
3428 					len -= xfer;
3429 					if (len < savlen) {
3430 						interrupted = sig_intr();
3431 						if (timespecisset(&endts) &&
3432 						    interrupted == 0) {
3433 							getnanouptime(&curts);
3434 							if (timespeccmp(&curts,
3435 							    &endts, >=))
3436 								interrupted =
3437 								    EINTR;
3438 						}
3439 					}
3440 				}
3441 			}
3442 			xfer = blksize;
3443 		}
3444 	}
3445 out:
3446 	*lenp = savlen - len;
3447 	free(dat, M_TEMP);
3448 	return (error);
3449 }
3450 
3451 static int
3452 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3453 {
3454 	struct mount *mp;
3455 	struct vnode *vp;
3456 	off_t olen, ooffset;
3457 	int error;
3458 #ifdef AUDIT
3459 	int audited_vnode1 = 0;
3460 #endif
3461 
3462 	vp = fp->f_vnode;
3463 	if (vp->v_type != VREG)
3464 		return (ENODEV);
3465 
3466 	/* Allocating blocks may take a long time, so iterate. */
3467 	for (;;) {
3468 		olen = len;
3469 		ooffset = offset;
3470 
3471 		bwillwrite();
3472 		mp = NULL;
3473 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3474 		if (error != 0)
3475 			break;
3476 		error = vn_lock(vp, LK_EXCLUSIVE);
3477 		if (error != 0) {
3478 			vn_finished_write(mp);
3479 			break;
3480 		}
3481 #ifdef AUDIT
3482 		if (!audited_vnode1) {
3483 			AUDIT_ARG_VNODE1(vp);
3484 			audited_vnode1 = 1;
3485 		}
3486 #endif
3487 #ifdef MAC
3488 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3489 		if (error == 0)
3490 #endif
3491 			error = VOP_ALLOCATE(vp, &offset, &len, 0,
3492 			    td->td_ucred);
3493 		VOP_UNLOCK(vp);
3494 		vn_finished_write(mp);
3495 
3496 		if (olen + ooffset != offset + len) {
3497 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3498 			    ooffset, olen, offset, len);
3499 		}
3500 		if (error != 0 || len == 0)
3501 			break;
3502 		KASSERT(olen > len, ("Iteration did not make progress?"));
3503 		maybe_yield();
3504 	}
3505 
3506 	return (error);
3507 }
3508 
3509 static int
3510 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3511     int ioflag, struct ucred *cred, struct ucred *active_cred,
3512     struct ucred *file_cred)
3513 {
3514 	struct mount *mp;
3515 	void *rl_cookie;
3516 	off_t off, len;
3517 	int error;
3518 #ifdef AUDIT
3519 	bool audited_vnode1 = false;
3520 #endif
3521 
3522 	rl_cookie = NULL;
3523 	error = 0;
3524 	mp = NULL;
3525 	off = *offset;
3526 	len = *length;
3527 
3528 	if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3529 		rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3530 	while (len > 0 && error == 0) {
3531 		/*
3532 		 * Try to deallocate the longest range in one pass.
3533 		 * In case a pass takes too long to be executed, it returns
3534 		 * partial result. The residue will be proceeded in the next
3535 		 * pass.
3536 		 */
3537 
3538 		if ((ioflag & IO_NODELOCKED) == 0) {
3539 			bwillwrite();
3540 			if ((error = vn_start_write(vp, &mp,
3541 			    V_WAIT | PCATCH)) != 0)
3542 				goto out;
3543 			vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3544 		}
3545 #ifdef AUDIT
3546 		if (!audited_vnode1) {
3547 			AUDIT_ARG_VNODE1(vp);
3548 			audited_vnode1 = true;
3549 		}
3550 #endif
3551 
3552 #ifdef MAC
3553 		if ((ioflag & IO_NOMACCHECK) == 0)
3554 			error = mac_vnode_check_write(active_cred, file_cred,
3555 			    vp);
3556 #endif
3557 		if (error == 0)
3558 			error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3559 			    cred);
3560 
3561 		if ((ioflag & IO_NODELOCKED) == 0) {
3562 			VOP_UNLOCK(vp);
3563 			if (mp != NULL) {
3564 				vn_finished_write(mp);
3565 				mp = NULL;
3566 			}
3567 		}
3568 		if (error == 0 && len != 0)
3569 			maybe_yield();
3570 	}
3571 out:
3572 	if (rl_cookie != NULL)
3573 		vn_rangelock_unlock(vp, rl_cookie);
3574 	*offset = off;
3575 	*length = len;
3576 	return (error);
3577 }
3578 
3579 /*
3580  * This function is supposed to be used in the situations where the deallocation
3581  * is not triggered by a user request.
3582  */
3583 int
3584 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3585     int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3586 {
3587 	struct ucred *cred;
3588 
3589 	if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3590 	    flags != 0)
3591 		return (EINVAL);
3592 	if (vp->v_type != VREG)
3593 		return (ENODEV);
3594 
3595 	cred = file_cred != NOCRED ? file_cred : active_cred;
3596 	return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3597 	    active_cred, file_cred));
3598 }
3599 
3600 static int
3601 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3602     struct ucred *active_cred, struct thread *td)
3603 {
3604 	int error;
3605 	struct vnode *vp;
3606 	int ioflag;
3607 
3608 	vp = fp->f_vnode;
3609 
3610 	if (cmd != SPACECTL_DEALLOC || *offset < 0 || *length <= 0 ||
3611 	    *length > OFF_MAX - *offset || flags != 0)
3612 		return (EINVAL);
3613 	if (vp->v_type != VREG)
3614 		return (ENODEV);
3615 
3616 	ioflag = get_write_ioflag(fp);
3617 
3618 	switch (cmd) {
3619 	case SPACECTL_DEALLOC:
3620 		error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3621 		    active_cred, active_cred, fp->f_cred);
3622 		break;
3623 	default:
3624 		panic("vn_fspacectl: unknown cmd %d", cmd);
3625 	}
3626 
3627 	return (error);
3628 }
3629 
3630 static u_long vn_lock_pair_pause_cnt;
3631 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3632     &vn_lock_pair_pause_cnt, 0,
3633     "Count of vn_lock_pair deadlocks");
3634 
3635 u_int vn_lock_pair_pause_max;
3636 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3637     &vn_lock_pair_pause_max, 0,
3638     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3639 
3640 static void
3641 vn_lock_pair_pause(const char *wmesg)
3642 {
3643 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3644 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3645 }
3646 
3647 /*
3648  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3649  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3650  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3651  * can be NULL.
3652  *
3653  * The function returns with both vnodes exclusively locked, and
3654  * guarantees that it does not create lock order reversal with other
3655  * threads during its execution.  Both vnodes could be unlocked
3656  * temporary (and reclaimed).
3657  */
3658 void
3659 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3660     bool vp2_locked)
3661 {
3662 	int error;
3663 
3664 	if (vp1 == NULL && vp2 == NULL)
3665 		return;
3666 	if (vp1 != NULL) {
3667 		if (vp1_locked)
3668 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3669 		else
3670 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3671 	} else {
3672 		vp1_locked = true;
3673 	}
3674 	if (vp2 != NULL) {
3675 		if (vp2_locked)
3676 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3677 		else
3678 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3679 	} else {
3680 		vp2_locked = true;
3681 	}
3682 	if (!vp1_locked && !vp2_locked) {
3683 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3684 		vp1_locked = true;
3685 	}
3686 
3687 	for (;;) {
3688 		if (vp1_locked && vp2_locked)
3689 			break;
3690 		if (vp1_locked && vp2 != NULL) {
3691 			if (vp1 != NULL) {
3692 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3693 				    __FILE__, __LINE__);
3694 				if (error == 0)
3695 					break;
3696 				VOP_UNLOCK(vp1);
3697 				vp1_locked = false;
3698 				vn_lock_pair_pause("vlp1");
3699 			}
3700 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3701 			vp2_locked = true;
3702 		}
3703 		if (vp2_locked && vp1 != NULL) {
3704 			if (vp2 != NULL) {
3705 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3706 				    __FILE__, __LINE__);
3707 				if (error == 0)
3708 					break;
3709 				VOP_UNLOCK(vp2);
3710 				vp2_locked = false;
3711 				vn_lock_pair_pause("vlp2");
3712 			}
3713 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3714 			vp1_locked = true;
3715 		}
3716 	}
3717 	if (vp1 != NULL)
3718 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3719 	if (vp2 != NULL)
3720 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3721 }
3722 
3723 int
3724 vn_lktype_write(struct mount *mp, struct vnode *vp)
3725 {
3726 	if (MNT_SHARED_WRITES(mp) ||
3727 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3728 		return (LK_SHARED);
3729 	return (LK_EXCLUSIVE);
3730 }
3731