xref: /freebsd/sys/kern/vfs_vnops.c (revision b542c9e42ba404b12facf2a51b4c2504bb14eab9)
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 = ndp->ni_cnd.cn_thread;
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 ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
209 		res |= AUDITVNODE1;
210 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
211 		res |= NOCAPCHECK;
212 	return (res);
213 }
214 
215 /*
216  * Common code for vnode open operations via a name lookup.
217  * Lookup the vnode and invoke VOP_CREATE if needed.
218  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
219  *
220  * Note that this does NOT free nameidata for the successful case,
221  * due to the NDINIT being done elsewhere.
222  */
223 int
224 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
225     struct ucred *cred, struct file *fp)
226 {
227 	struct vnode *vp;
228 	struct mount *mp;
229 	struct thread *td = ndp->ni_cnd.cn_thread;
230 	struct vattr vat;
231 	struct vattr *vap = &vat;
232 	int fmode, error;
233 	bool first_open;
234 
235 restart:
236 	first_open = false;
237 	fmode = *flagp;
238 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
239 	    O_EXCL | O_DIRECTORY) ||
240 	    (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
241 		return (EINVAL);
242 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
243 		ndp->ni_cnd.cn_nameiop = CREATE;
244 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
245 		/*
246 		 * Set NOCACHE to avoid flushing the cache when
247 		 * rolling in many files at once.
248 		 *
249 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
250 		 * exist despite NOCACHE.
251 		 */
252 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
253 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
254 			ndp->ni_cnd.cn_flags |= FOLLOW;
255 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
256 			bwillwrite();
257 		if ((error = namei(ndp)) != 0)
258 			return (error);
259 		if (ndp->ni_vp == NULL) {
260 			VATTR_NULL(vap);
261 			vap->va_type = VREG;
262 			vap->va_mode = cmode;
263 			if (fmode & O_EXCL)
264 				vap->va_vaflags |= VA_EXCLUSIVE;
265 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
266 				NDFREE(ndp, NDF_ONLY_PNBUF);
267 				vput(ndp->ni_dvp);
268 				if ((error = vn_start_write(NULL, &mp,
269 				    V_XSLEEP | PCATCH)) != 0)
270 					return (error);
271 				NDREINIT(ndp);
272 				goto restart;
273 			}
274 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
275 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
276 #ifdef MAC
277 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
278 			    &ndp->ni_cnd, vap);
279 			if (error == 0)
280 #endif
281 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
282 				    &ndp->ni_cnd, vap);
283 			vp = ndp->ni_vp;
284 			if (error == 0 && (fmode & O_EXCL) != 0 &&
285 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
286 				VI_LOCK(vp);
287 				vp->v_iflag |= VI_FOPENING;
288 				VI_UNLOCK(vp);
289 				first_open = true;
290 			}
291 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
292 			    false);
293 			vn_finished_write(mp);
294 			if (error) {
295 				NDFREE(ndp, NDF_ONLY_PNBUF);
296 				if (error == ERELOOKUP) {
297 					NDREINIT(ndp);
298 					goto restart;
299 				}
300 				return (error);
301 			}
302 			fmode &= ~O_TRUNC;
303 		} else {
304 			if (ndp->ni_dvp == ndp->ni_vp)
305 				vrele(ndp->ni_dvp);
306 			else
307 				vput(ndp->ni_dvp);
308 			ndp->ni_dvp = NULL;
309 			vp = ndp->ni_vp;
310 			if (fmode & O_EXCL) {
311 				error = EEXIST;
312 				goto bad;
313 			}
314 			if (vp->v_type == VDIR) {
315 				error = EISDIR;
316 				goto bad;
317 			}
318 			fmode &= ~O_CREAT;
319 		}
320 	} else {
321 		ndp->ni_cnd.cn_nameiop = LOOKUP;
322 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
323 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
324 		    FOLLOW;
325 		if ((fmode & FWRITE) == 0)
326 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
327 		if ((error = namei(ndp)) != 0)
328 			return (error);
329 		vp = ndp->ni_vp;
330 	}
331 	error = vn_open_vnode(vp, fmode, cred, td, fp);
332 	if (first_open) {
333 		VI_LOCK(vp);
334 		vp->v_iflag &= ~VI_FOPENING;
335 		wakeup(vp);
336 		VI_UNLOCK(vp);
337 	}
338 	if (error)
339 		goto bad;
340 	*flagp = fmode;
341 	return (0);
342 bad:
343 	NDFREE(ndp, NDF_ONLY_PNBUF);
344 	vput(vp);
345 	*flagp = fmode;
346 	ndp->ni_vp = NULL;
347 	return (error);
348 }
349 
350 static int
351 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
352 {
353 	struct flock lf;
354 	int error, lock_flags, type;
355 
356 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
357 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
358 		return (0);
359 	KASSERT(fp != NULL, ("open with flock requires fp"));
360 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
361 		return (EOPNOTSUPP);
362 
363 	lock_flags = VOP_ISLOCKED(vp);
364 	VOP_UNLOCK(vp);
365 
366 	lf.l_whence = SEEK_SET;
367 	lf.l_start = 0;
368 	lf.l_len = 0;
369 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
370 	type = F_FLOCK;
371 	if ((fmode & FNONBLOCK) == 0)
372 		type |= F_WAIT;
373 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
374 		type |= F_FIRSTOPEN;
375 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
376 	if (error == 0)
377 		fp->f_flag |= FHASLOCK;
378 
379 	vn_lock(vp, lock_flags | LK_RETRY);
380 	return (error);
381 }
382 
383 /*
384  * Common code for vnode open operations once a vnode is located.
385  * Check permissions, and call the VOP_OPEN routine.
386  */
387 int
388 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
389     struct thread *td, struct file *fp)
390 {
391 	accmode_t accmode;
392 	int error;
393 
394 	if (vp->v_type == VLNK) {
395 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
396 			return (EMLINK);
397 	}
398 	if (vp->v_type == VSOCK)
399 		return (EOPNOTSUPP);
400 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
401 		return (ENOTDIR);
402 
403 	accmode = 0;
404 	if ((fmode & O_PATH) == 0) {
405 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
406 			if (vp->v_type == VDIR)
407 				return (EISDIR);
408 			accmode |= VWRITE;
409 		}
410 		if ((fmode & FREAD) != 0)
411 			accmode |= VREAD;
412 		if ((fmode & O_APPEND) && (fmode & FWRITE))
413 			accmode |= VAPPEND;
414 #ifdef MAC
415 		if ((fmode & O_CREAT) != 0)
416 			accmode |= VCREAT;
417 #endif
418 	}
419 	if ((fmode & FEXEC) != 0)
420 		accmode |= VEXEC;
421 #ifdef MAC
422 	if ((fmode & O_VERIFY) != 0)
423 		accmode |= VVERIFY;
424 	error = mac_vnode_check_open(cred, vp, accmode);
425 	if (error != 0)
426 		return (error);
427 
428 	accmode &= ~(VCREAT | VVERIFY);
429 #endif
430 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
431 		error = VOP_ACCESS(vp, accmode, cred, td);
432 		if (error != 0)
433 			return (error);
434 	}
435 	if ((fmode & O_PATH) != 0) {
436 		if (vp->v_type == VFIFO)
437 			error = EPIPE;
438 		else
439 			error = VOP_ACCESS(vp, VREAD, cred, td);
440 		if (error == 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 int
920 vn_read_from_obj(struct vnode *vp, struct uio *uio)
921 {
922 	vm_object_t obj;
923 	vm_page_t ma[io_hold_cnt + 2];
924 	off_t off, vsz;
925 	ssize_t resid;
926 	int error, i, j;
927 
928 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
929 	obj = atomic_load_ptr(&vp->v_object);
930 	if (obj == NULL)
931 		return (EJUSTRETURN);
932 
933 	/*
934 	 * Depends on type stability of vm_objects.
935 	 */
936 	vm_object_pip_add(obj, 1);
937 	if ((obj->flags & OBJ_DEAD) != 0) {
938 		/*
939 		 * Note that object might be already reused from the
940 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
941 		 * we recheck for DOOMED vnode state after all pages
942 		 * are busied, and retract then.
943 		 *
944 		 * But we check for OBJ_DEAD to ensure that we do not
945 		 * busy pages while vm_object_terminate_pages()
946 		 * processes the queue.
947 		 */
948 		error = EJUSTRETURN;
949 		goto out_pip;
950 	}
951 
952 	resid = uio->uio_resid;
953 	off = uio->uio_offset;
954 	for (i = 0; resid > 0; i++) {
955 		MPASS(i < io_hold_cnt + 2);
956 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
957 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
958 		    VM_ALLOC_NOWAIT);
959 		if (ma[i] == NULL)
960 			break;
961 
962 		/*
963 		 * Skip invalid pages.  Valid mask can be partial only
964 		 * at EOF, and we clip later.
965 		 */
966 		if (vm_page_none_valid(ma[i])) {
967 			vm_page_sunbusy(ma[i]);
968 			break;
969 		}
970 
971 		resid -= PAGE_SIZE;
972 		off += PAGE_SIZE;
973 	}
974 	if (i == 0) {
975 		error = EJUSTRETURN;
976 		goto out_pip;
977 	}
978 
979 	/*
980 	 * Check VIRF_DOOMED after we busied our pages.  Since
981 	 * vgonel() terminates the vnode' vm_object, it cannot
982 	 * process past pages busied by us.
983 	 */
984 	if (VN_IS_DOOMED(vp)) {
985 		error = EJUSTRETURN;
986 		goto out;
987 	}
988 
989 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
990 	if (resid > uio->uio_resid)
991 		resid = uio->uio_resid;
992 
993 	/*
994 	 * Unlocked read of vnp_size is safe because truncation cannot
995 	 * pass busied page.  But we load vnp_size into a local
996 	 * variable so that possible concurrent extension does not
997 	 * break calculation.
998 	 */
999 #if defined(__powerpc__) && !defined(__powerpc64__)
1000 	vsz = obj->un_pager.vnp.vnp_size;
1001 #else
1002 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1003 #endif
1004 	if (uio->uio_offset >= vsz) {
1005 		error = EJUSTRETURN;
1006 		goto out;
1007 	}
1008 	if (uio->uio_offset + resid > vsz)
1009 		resid = vsz - uio->uio_offset;
1010 
1011 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1012 
1013 out:
1014 	for (j = 0; j < i; j++) {
1015 		if (error == 0)
1016 			vm_page_reference(ma[j]);
1017 		vm_page_sunbusy(ma[j]);
1018 	}
1019 out_pip:
1020 	vm_object_pip_wakeup(obj);
1021 	if (error != 0)
1022 		return (error);
1023 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1024 }
1025 
1026 /*
1027  * File table vnode read routine.
1028  */
1029 static int
1030 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1031     struct thread *td)
1032 {
1033 	struct vnode *vp;
1034 	off_t orig_offset;
1035 	int error, ioflag;
1036 	int advice;
1037 
1038 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1039 	    uio->uio_td, td));
1040 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1041 	vp = fp->f_vnode;
1042 	ioflag = 0;
1043 	if (fp->f_flag & FNONBLOCK)
1044 		ioflag |= IO_NDELAY;
1045 	if (fp->f_flag & O_DIRECT)
1046 		ioflag |= IO_DIRECT;
1047 
1048 	/*
1049 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1050 	 * allows us to avoid unneeded work outright.
1051 	 */
1052 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1053 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1054 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1055 		if (error == 0) {
1056 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1057 			return (0);
1058 		}
1059 		if (error != EJUSTRETURN)
1060 			return (error);
1061 	}
1062 
1063 	advice = get_advice(fp, uio);
1064 	vn_lock(vp, LK_SHARED | LK_RETRY);
1065 
1066 	switch (advice) {
1067 	case POSIX_FADV_NORMAL:
1068 	case POSIX_FADV_SEQUENTIAL:
1069 	case POSIX_FADV_NOREUSE:
1070 		ioflag |= sequential_heuristic(uio, fp);
1071 		break;
1072 	case POSIX_FADV_RANDOM:
1073 		/* Disable read-ahead for random I/O. */
1074 		break;
1075 	}
1076 	orig_offset = uio->uio_offset;
1077 
1078 #ifdef MAC
1079 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1080 	if (error == 0)
1081 #endif
1082 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1083 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1084 	VOP_UNLOCK(vp);
1085 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1086 	    orig_offset != uio->uio_offset)
1087 		/*
1088 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1089 		 * for the backing file after a POSIX_FADV_NOREUSE
1090 		 * read(2).
1091 		 */
1092 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1093 		    POSIX_FADV_DONTNEED);
1094 	return (error);
1095 }
1096 
1097 /*
1098  * File table vnode write routine.
1099  */
1100 static int
1101 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1102     struct thread *td)
1103 {
1104 	struct vnode *vp;
1105 	struct mount *mp;
1106 	off_t orig_offset;
1107 	int error, ioflag;
1108 	int advice;
1109 	bool need_finished_write;
1110 
1111 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1112 	    uio->uio_td, td));
1113 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1114 	vp = fp->f_vnode;
1115 	if (vp->v_type == VREG)
1116 		bwillwrite();
1117 	ioflag = IO_UNIT;
1118 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1119 		ioflag |= IO_APPEND;
1120 	if (fp->f_flag & FNONBLOCK)
1121 		ioflag |= IO_NDELAY;
1122 	if (fp->f_flag & O_DIRECT)
1123 		ioflag |= IO_DIRECT;
1124 
1125 	mp = atomic_load_ptr(&vp->v_mount);
1126 	if ((fp->f_flag & O_FSYNC) ||
1127 	    (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS)))
1128 		ioflag |= IO_SYNC;
1129 
1130 	/*
1131 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
1132 	 * implementations that don't understand IO_DATASYNC fall back to full
1133 	 * O_SYNC behavior.
1134 	 */
1135 	if (fp->f_flag & O_DSYNC)
1136 		ioflag |= IO_SYNC | IO_DATASYNC;
1137 	mp = NULL;
1138 	need_finished_write = false;
1139 	if (vp->v_type != VCHR) {
1140 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1141 		if (error != 0)
1142 			goto unlock;
1143 		need_finished_write = true;
1144 	}
1145 
1146 	advice = get_advice(fp, uio);
1147 
1148 	vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1149 	switch (advice) {
1150 	case POSIX_FADV_NORMAL:
1151 	case POSIX_FADV_SEQUENTIAL:
1152 	case POSIX_FADV_NOREUSE:
1153 		ioflag |= sequential_heuristic(uio, fp);
1154 		break;
1155 	case POSIX_FADV_RANDOM:
1156 		/* XXX: Is this correct? */
1157 		break;
1158 	}
1159 	orig_offset = uio->uio_offset;
1160 
1161 #ifdef MAC
1162 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1163 	if (error == 0)
1164 #endif
1165 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1166 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1167 	VOP_UNLOCK(vp);
1168 	if (need_finished_write)
1169 		vn_finished_write(mp);
1170 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1171 	    orig_offset != uio->uio_offset)
1172 		/*
1173 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1174 		 * for the backing file after a POSIX_FADV_NOREUSE
1175 		 * write(2).
1176 		 */
1177 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1178 		    POSIX_FADV_DONTNEED);
1179 unlock:
1180 	return (error);
1181 }
1182 
1183 /*
1184  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1185  * prevent the following deadlock:
1186  *
1187  * Assume that the thread A reads from the vnode vp1 into userspace
1188  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1189  * currently not resident, then system ends up with the call chain
1190  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1191  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1192  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1193  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1194  * backed by the pages of vnode vp1, and some page in buf2 is not
1195  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1196  *
1197  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1198  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1199  * Instead, it first tries to do the whole range i/o with pagefaults
1200  * disabled. If all pages in the i/o buffer are resident and mapped,
1201  * VOP will succeed (ignoring the genuine filesystem errors).
1202  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1203  * i/o in chunks, with all pages in the chunk prefaulted and held
1204  * using vm_fault_quick_hold_pages().
1205  *
1206  * Filesystems using this deadlock avoidance scheme should use the
1207  * array of the held pages from uio, saved in the curthread->td_ma,
1208  * instead of doing uiomove().  A helper function
1209  * vn_io_fault_uiomove() converts uiomove request into
1210  * uiomove_fromphys() over td_ma array.
1211  *
1212  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1213  * make the current i/o request atomic with respect to other i/os and
1214  * truncations.
1215  */
1216 
1217 /*
1218  * Decode vn_io_fault_args and perform the corresponding i/o.
1219  */
1220 static int
1221 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1222     struct thread *td)
1223 {
1224 	int error, save;
1225 
1226 	error = 0;
1227 	save = vm_fault_disable_pagefaults();
1228 	switch (args->kind) {
1229 	case VN_IO_FAULT_FOP:
1230 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1231 		    uio, args->cred, args->flags, td);
1232 		break;
1233 	case VN_IO_FAULT_VOP:
1234 		if (uio->uio_rw == UIO_READ) {
1235 			error = VOP_READ(args->args.vop_args.vp, uio,
1236 			    args->flags, args->cred);
1237 		} else if (uio->uio_rw == UIO_WRITE) {
1238 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1239 			    args->flags, args->cred);
1240 		}
1241 		break;
1242 	default:
1243 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1244 		    args->kind, uio->uio_rw);
1245 	}
1246 	vm_fault_enable_pagefaults(save);
1247 	return (error);
1248 }
1249 
1250 static int
1251 vn_io_fault_touch(char *base, const struct uio *uio)
1252 {
1253 	int r;
1254 
1255 	r = fubyte(base);
1256 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1257 		return (EFAULT);
1258 	return (0);
1259 }
1260 
1261 static int
1262 vn_io_fault_prefault_user(const struct uio *uio)
1263 {
1264 	char *base;
1265 	const struct iovec *iov;
1266 	size_t len;
1267 	ssize_t resid;
1268 	int error, i;
1269 
1270 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1271 	    ("vn_io_fault_prefault userspace"));
1272 
1273 	error = i = 0;
1274 	iov = uio->uio_iov;
1275 	resid = uio->uio_resid;
1276 	base = iov->iov_base;
1277 	len = iov->iov_len;
1278 	while (resid > 0) {
1279 		error = vn_io_fault_touch(base, uio);
1280 		if (error != 0)
1281 			break;
1282 		if (len < PAGE_SIZE) {
1283 			if (len != 0) {
1284 				error = vn_io_fault_touch(base + len - 1, uio);
1285 				if (error != 0)
1286 					break;
1287 				resid -= len;
1288 			}
1289 			if (++i >= uio->uio_iovcnt)
1290 				break;
1291 			iov = uio->uio_iov + i;
1292 			base = iov->iov_base;
1293 			len = iov->iov_len;
1294 		} else {
1295 			len -= PAGE_SIZE;
1296 			base += PAGE_SIZE;
1297 			resid -= PAGE_SIZE;
1298 		}
1299 	}
1300 	return (error);
1301 }
1302 
1303 /*
1304  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1305  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1306  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1307  * into args and call vn_io_fault1() to handle faults during the user
1308  * mode buffer accesses.
1309  */
1310 static int
1311 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1312     struct thread *td)
1313 {
1314 	vm_page_t ma[io_hold_cnt + 2];
1315 	struct uio *uio_clone, short_uio;
1316 	struct iovec short_iovec[1];
1317 	vm_page_t *prev_td_ma;
1318 	vm_prot_t prot;
1319 	vm_offset_t addr, end;
1320 	size_t len, resid;
1321 	ssize_t adv;
1322 	int error, cnt, saveheld, prev_td_ma_cnt;
1323 
1324 	if (vn_io_fault_prefault) {
1325 		error = vn_io_fault_prefault_user(uio);
1326 		if (error != 0)
1327 			return (error); /* Or ignore ? */
1328 	}
1329 
1330 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1331 
1332 	/*
1333 	 * The UFS follows IO_UNIT directive and replays back both
1334 	 * uio_offset and uio_resid if an error is encountered during the
1335 	 * operation.  But, since the iovec may be already advanced,
1336 	 * uio is still in an inconsistent state.
1337 	 *
1338 	 * Cache a copy of the original uio, which is advanced to the redo
1339 	 * point using UIO_NOCOPY below.
1340 	 */
1341 	uio_clone = cloneuio(uio);
1342 	resid = uio->uio_resid;
1343 
1344 	short_uio.uio_segflg = UIO_USERSPACE;
1345 	short_uio.uio_rw = uio->uio_rw;
1346 	short_uio.uio_td = uio->uio_td;
1347 
1348 	error = vn_io_fault_doio(args, uio, td);
1349 	if (error != EFAULT)
1350 		goto out;
1351 
1352 	atomic_add_long(&vn_io_faults_cnt, 1);
1353 	uio_clone->uio_segflg = UIO_NOCOPY;
1354 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1355 	uio_clone->uio_segflg = uio->uio_segflg;
1356 
1357 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1358 	prev_td_ma = td->td_ma;
1359 	prev_td_ma_cnt = td->td_ma_cnt;
1360 
1361 	while (uio_clone->uio_resid != 0) {
1362 		len = uio_clone->uio_iov->iov_len;
1363 		if (len == 0) {
1364 			KASSERT(uio_clone->uio_iovcnt >= 1,
1365 			    ("iovcnt underflow"));
1366 			uio_clone->uio_iov++;
1367 			uio_clone->uio_iovcnt--;
1368 			continue;
1369 		}
1370 		if (len > ptoa(io_hold_cnt))
1371 			len = ptoa(io_hold_cnt);
1372 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1373 		end = round_page(addr + len);
1374 		if (end < addr) {
1375 			error = EFAULT;
1376 			break;
1377 		}
1378 		cnt = atop(end - trunc_page(addr));
1379 		/*
1380 		 * A perfectly misaligned address and length could cause
1381 		 * both the start and the end of the chunk to use partial
1382 		 * page.  +2 accounts for such a situation.
1383 		 */
1384 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1385 		    addr, len, prot, ma, io_hold_cnt + 2);
1386 		if (cnt == -1) {
1387 			error = EFAULT;
1388 			break;
1389 		}
1390 		short_uio.uio_iov = &short_iovec[0];
1391 		short_iovec[0].iov_base = (void *)addr;
1392 		short_uio.uio_iovcnt = 1;
1393 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1394 		short_uio.uio_offset = uio_clone->uio_offset;
1395 		td->td_ma = ma;
1396 		td->td_ma_cnt = cnt;
1397 
1398 		error = vn_io_fault_doio(args, &short_uio, td);
1399 		vm_page_unhold_pages(ma, cnt);
1400 		adv = len - short_uio.uio_resid;
1401 
1402 		uio_clone->uio_iov->iov_base =
1403 		    (char *)uio_clone->uio_iov->iov_base + adv;
1404 		uio_clone->uio_iov->iov_len -= adv;
1405 		uio_clone->uio_resid -= adv;
1406 		uio_clone->uio_offset += adv;
1407 
1408 		uio->uio_resid -= adv;
1409 		uio->uio_offset += adv;
1410 
1411 		if (error != 0 || adv == 0)
1412 			break;
1413 	}
1414 	td->td_ma = prev_td_ma;
1415 	td->td_ma_cnt = prev_td_ma_cnt;
1416 	curthread_pflags_restore(saveheld);
1417 out:
1418 	free(uio_clone, M_IOV);
1419 	return (error);
1420 }
1421 
1422 static int
1423 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1424     int flags, struct thread *td)
1425 {
1426 	fo_rdwr_t *doio;
1427 	struct vnode *vp;
1428 	void *rl_cookie;
1429 	struct vn_io_fault_args args;
1430 	int error;
1431 
1432 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1433 	vp = fp->f_vnode;
1434 
1435 	/*
1436 	 * The ability to read(2) on a directory has historically been
1437 	 * allowed for all users, but this can and has been the source of
1438 	 * at least one security issue in the past.  As such, it is now hidden
1439 	 * away behind a sysctl for those that actually need it to use it, and
1440 	 * restricted to root when it's turned on to make it relatively safe to
1441 	 * leave on for longer sessions of need.
1442 	 */
1443 	if (vp->v_type == VDIR) {
1444 		KASSERT(uio->uio_rw == UIO_READ,
1445 		    ("illegal write attempted on a directory"));
1446 		if (!vfs_allow_read_dir)
1447 			return (EISDIR);
1448 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1449 			return (EISDIR);
1450 	}
1451 
1452 	foffset_lock_uio(fp, uio, flags);
1453 	if (do_vn_io_fault(vp, uio)) {
1454 		args.kind = VN_IO_FAULT_FOP;
1455 		args.args.fop_args.fp = fp;
1456 		args.args.fop_args.doio = doio;
1457 		args.cred = active_cred;
1458 		args.flags = flags | FOF_OFFSET;
1459 		if (uio->uio_rw == UIO_READ) {
1460 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1461 			    uio->uio_offset + uio->uio_resid);
1462 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1463 		    (flags & FOF_OFFSET) == 0) {
1464 			/* For appenders, punt and lock the whole range. */
1465 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1466 		} else {
1467 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1468 			    uio->uio_offset + uio->uio_resid);
1469 		}
1470 		error = vn_io_fault1(vp, uio, &args, td);
1471 		vn_rangelock_unlock(vp, rl_cookie);
1472 	} else {
1473 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1474 	}
1475 	foffset_unlock_uio(fp, uio, flags);
1476 	return (error);
1477 }
1478 
1479 /*
1480  * Helper function to perform the requested uiomove operation using
1481  * the held pages for io->uio_iov[0].iov_base buffer instead of
1482  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1483  * instead of iov_base prevents page faults that could occur due to
1484  * pmap_collect() invalidating the mapping created by
1485  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1486  * object cleanup revoking the write access from page mappings.
1487  *
1488  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1489  * instead of plain uiomove().
1490  */
1491 int
1492 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1493 {
1494 	struct uio transp_uio;
1495 	struct iovec transp_iov[1];
1496 	struct thread *td;
1497 	size_t adv;
1498 	int error, pgadv;
1499 
1500 	td = curthread;
1501 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1502 	    uio->uio_segflg != UIO_USERSPACE)
1503 		return (uiomove(data, xfersize, uio));
1504 
1505 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1506 	transp_iov[0].iov_base = data;
1507 	transp_uio.uio_iov = &transp_iov[0];
1508 	transp_uio.uio_iovcnt = 1;
1509 	if (xfersize > uio->uio_resid)
1510 		xfersize = uio->uio_resid;
1511 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1512 	transp_uio.uio_offset = 0;
1513 	transp_uio.uio_segflg = UIO_SYSSPACE;
1514 	/*
1515 	 * Since transp_iov points to data, and td_ma page array
1516 	 * corresponds to original uio->uio_iov, we need to invert the
1517 	 * direction of the i/o operation as passed to
1518 	 * uiomove_fromphys().
1519 	 */
1520 	switch (uio->uio_rw) {
1521 	case UIO_WRITE:
1522 		transp_uio.uio_rw = UIO_READ;
1523 		break;
1524 	case UIO_READ:
1525 		transp_uio.uio_rw = UIO_WRITE;
1526 		break;
1527 	}
1528 	transp_uio.uio_td = uio->uio_td;
1529 	error = uiomove_fromphys(td->td_ma,
1530 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1531 	    xfersize, &transp_uio);
1532 	adv = xfersize - transp_uio.uio_resid;
1533 	pgadv =
1534 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1535 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1536 	td->td_ma += pgadv;
1537 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1538 	    pgadv));
1539 	td->td_ma_cnt -= pgadv;
1540 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1541 	uio->uio_iov->iov_len -= adv;
1542 	uio->uio_resid -= adv;
1543 	uio->uio_offset += adv;
1544 	return (error);
1545 }
1546 
1547 int
1548 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1549     struct uio *uio)
1550 {
1551 	struct thread *td;
1552 	vm_offset_t iov_base;
1553 	int cnt, pgadv;
1554 
1555 	td = curthread;
1556 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1557 	    uio->uio_segflg != UIO_USERSPACE)
1558 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1559 
1560 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1561 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1562 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1563 	switch (uio->uio_rw) {
1564 	case UIO_WRITE:
1565 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1566 		    offset, cnt);
1567 		break;
1568 	case UIO_READ:
1569 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1570 		    cnt);
1571 		break;
1572 	}
1573 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1574 	td->td_ma += pgadv;
1575 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1576 	    pgadv));
1577 	td->td_ma_cnt -= pgadv;
1578 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1579 	uio->uio_iov->iov_len -= cnt;
1580 	uio->uio_resid -= cnt;
1581 	uio->uio_offset += cnt;
1582 	return (0);
1583 }
1584 
1585 /*
1586  * File table truncate routine.
1587  */
1588 static int
1589 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1590     struct thread *td)
1591 {
1592 	struct mount *mp;
1593 	struct vnode *vp;
1594 	void *rl_cookie;
1595 	int error;
1596 
1597 	vp = fp->f_vnode;
1598 
1599 retry:
1600 	/*
1601 	 * Lock the whole range for truncation.  Otherwise split i/o
1602 	 * might happen partly before and partly after the truncation.
1603 	 */
1604 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1605 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1606 	if (error)
1607 		goto out1;
1608 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1609 	AUDIT_ARG_VNODE1(vp);
1610 	if (vp->v_type == VDIR) {
1611 		error = EISDIR;
1612 		goto out;
1613 	}
1614 #ifdef MAC
1615 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1616 	if (error)
1617 		goto out;
1618 #endif
1619 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1620 	    fp->f_cred);
1621 out:
1622 	VOP_UNLOCK(vp);
1623 	vn_finished_write(mp);
1624 out1:
1625 	vn_rangelock_unlock(vp, rl_cookie);
1626 	if (error == ERELOOKUP)
1627 		goto retry;
1628 	return (error);
1629 }
1630 
1631 /*
1632  * Truncate a file that is already locked.
1633  */
1634 int
1635 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1636     struct ucred *cred)
1637 {
1638 	struct vattr vattr;
1639 	int error;
1640 
1641 	error = VOP_ADD_WRITECOUNT(vp, 1);
1642 	if (error == 0) {
1643 		VATTR_NULL(&vattr);
1644 		vattr.va_size = length;
1645 		if (sync)
1646 			vattr.va_vaflags |= VA_SYNC;
1647 		error = VOP_SETATTR(vp, &vattr, cred);
1648 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1649 	}
1650 	return (error);
1651 }
1652 
1653 /*
1654  * File table vnode stat routine.
1655  */
1656 int
1657 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1658     struct thread *td)
1659 {
1660 	struct vnode *vp = fp->f_vnode;
1661 	int error;
1662 
1663 	vn_lock(vp, LK_SHARED | LK_RETRY);
1664 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1665 	VOP_UNLOCK(vp);
1666 
1667 	return (error);
1668 }
1669 
1670 /*
1671  * File table vnode ioctl routine.
1672  */
1673 static int
1674 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1675     struct thread *td)
1676 {
1677 	struct vattr vattr;
1678 	struct vnode *vp;
1679 	struct fiobmap2_arg *bmarg;
1680 	int error;
1681 
1682 	vp = fp->f_vnode;
1683 	switch (vp->v_type) {
1684 	case VDIR:
1685 	case VREG:
1686 		switch (com) {
1687 		case FIONREAD:
1688 			vn_lock(vp, LK_SHARED | LK_RETRY);
1689 			error = VOP_GETATTR(vp, &vattr, active_cred);
1690 			VOP_UNLOCK(vp);
1691 			if (error == 0)
1692 				*(int *)data = vattr.va_size - fp->f_offset;
1693 			return (error);
1694 		case FIOBMAP2:
1695 			bmarg = (struct fiobmap2_arg *)data;
1696 			vn_lock(vp, LK_SHARED | LK_RETRY);
1697 #ifdef MAC
1698 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1699 			    vp);
1700 			if (error == 0)
1701 #endif
1702 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1703 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1704 			VOP_UNLOCK(vp);
1705 			return (error);
1706 		case FIONBIO:
1707 		case FIOASYNC:
1708 			return (0);
1709 		default:
1710 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1711 			    active_cred, td));
1712 		}
1713 		break;
1714 	case VCHR:
1715 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1716 		    active_cred, td));
1717 	default:
1718 		return (ENOTTY);
1719 	}
1720 }
1721 
1722 /*
1723  * File table vnode poll routine.
1724  */
1725 static int
1726 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1727     struct thread *td)
1728 {
1729 	struct vnode *vp;
1730 	int error;
1731 
1732 	vp = fp->f_vnode;
1733 #if defined(MAC) || defined(AUDIT)
1734 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1735 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1736 		AUDIT_ARG_VNODE1(vp);
1737 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1738 		VOP_UNLOCK(vp);
1739 		if (error != 0)
1740 			return (error);
1741 	}
1742 #endif
1743 	error = VOP_POLL(vp, events, fp->f_cred, td);
1744 	return (error);
1745 }
1746 
1747 /*
1748  * Acquire the requested lock and then check for validity.  LK_RETRY
1749  * permits vn_lock to return doomed vnodes.
1750  */
1751 static int __noinline
1752 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1753     int error)
1754 {
1755 
1756 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1757 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1758 
1759 	if (error == 0)
1760 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1761 
1762 	if ((flags & LK_RETRY) == 0) {
1763 		if (error == 0) {
1764 			VOP_UNLOCK(vp);
1765 			error = ENOENT;
1766 		}
1767 		return (error);
1768 	}
1769 
1770 	/*
1771 	 * LK_RETRY case.
1772 	 *
1773 	 * Nothing to do if we got the lock.
1774 	 */
1775 	if (error == 0)
1776 		return (0);
1777 
1778 	/*
1779 	 * Interlock was dropped by the call in _vn_lock.
1780 	 */
1781 	flags &= ~LK_INTERLOCK;
1782 	do {
1783 		error = VOP_LOCK1(vp, flags, file, line);
1784 	} while (error != 0);
1785 	return (0);
1786 }
1787 
1788 int
1789 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1790 {
1791 	int error;
1792 
1793 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1794 	    ("vn_lock: no locktype (%d passed)", flags));
1795 	VNPASS(vp->v_holdcnt > 0, vp);
1796 	error = VOP_LOCK1(vp, flags, file, line);
1797 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1798 		return (_vn_lock_fallback(vp, flags, file, line, error));
1799 	return (0);
1800 }
1801 
1802 /*
1803  * File table vnode close routine.
1804  */
1805 static int
1806 vn_closefile(struct file *fp, struct thread *td)
1807 {
1808 	struct vnode *vp;
1809 	struct flock lf;
1810 	int error;
1811 	bool ref;
1812 
1813 	vp = fp->f_vnode;
1814 	fp->f_ops = &badfileops;
1815 	ref = (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1816 
1817 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1818 
1819 	if (__predict_false(ref)) {
1820 		lf.l_whence = SEEK_SET;
1821 		lf.l_start = 0;
1822 		lf.l_len = 0;
1823 		lf.l_type = F_UNLCK;
1824 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1825 		vrele(vp);
1826 	}
1827 	return (error);
1828 }
1829 
1830 /*
1831  * Preparing to start a filesystem write operation. If the operation is
1832  * permitted, then we bump the count of operations in progress and
1833  * proceed. If a suspend request is in progress, we wait until the
1834  * suspension is over, and then proceed.
1835  */
1836 static int
1837 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1838 {
1839 	struct mount_pcpu *mpcpu;
1840 	int error, mflags;
1841 
1842 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1843 	    vfs_op_thread_enter(mp, mpcpu)) {
1844 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1845 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1846 		vfs_op_thread_exit(mp, mpcpu);
1847 		return (0);
1848 	}
1849 
1850 	if (mplocked)
1851 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1852 	else
1853 		MNT_ILOCK(mp);
1854 
1855 	error = 0;
1856 
1857 	/*
1858 	 * Check on status of suspension.
1859 	 */
1860 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1861 	    mp->mnt_susp_owner != curthread) {
1862 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1863 		    (flags & PCATCH) : 0) | (PUSER - 1);
1864 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1865 			if (flags & V_NOWAIT) {
1866 				error = EWOULDBLOCK;
1867 				goto unlock;
1868 			}
1869 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1870 			    "suspfs", 0);
1871 			if (error)
1872 				goto unlock;
1873 		}
1874 	}
1875 	if (flags & V_XSLEEP)
1876 		goto unlock;
1877 	mp->mnt_writeopcount++;
1878 unlock:
1879 	if (error != 0 || (flags & V_XSLEEP) != 0)
1880 		MNT_REL(mp);
1881 	MNT_IUNLOCK(mp);
1882 	return (error);
1883 }
1884 
1885 int
1886 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1887 {
1888 	struct mount *mp;
1889 	int error;
1890 
1891 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1892 	    ("V_MNTREF requires mp"));
1893 
1894 	error = 0;
1895 	/*
1896 	 * If a vnode is provided, get and return the mount point that
1897 	 * to which it will write.
1898 	 */
1899 	if (vp != NULL) {
1900 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1901 			*mpp = NULL;
1902 			if (error != EOPNOTSUPP)
1903 				return (error);
1904 			return (0);
1905 		}
1906 	}
1907 	if ((mp = *mpp) == NULL)
1908 		return (0);
1909 
1910 	/*
1911 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1912 	 * a vfs_ref().
1913 	 * As long as a vnode is not provided we need to acquire a
1914 	 * refcount for the provided mountpoint too, in order to
1915 	 * emulate a vfs_ref().
1916 	 */
1917 	if (vp == NULL && (flags & V_MNTREF) == 0)
1918 		vfs_ref(mp);
1919 
1920 	return (vn_start_write_refed(mp, flags, false));
1921 }
1922 
1923 /*
1924  * Secondary suspension. Used by operations such as vop_inactive
1925  * routines that are needed by the higher level functions. These
1926  * are allowed to proceed until all the higher level functions have
1927  * completed (indicated by mnt_writeopcount dropping to zero). At that
1928  * time, these operations are halted until the suspension is over.
1929  */
1930 int
1931 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1932 {
1933 	struct mount *mp;
1934 	int error;
1935 
1936 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1937 	    ("V_MNTREF requires mp"));
1938 
1939  retry:
1940 	if (vp != NULL) {
1941 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1942 			*mpp = NULL;
1943 			if (error != EOPNOTSUPP)
1944 				return (error);
1945 			return (0);
1946 		}
1947 	}
1948 	/*
1949 	 * If we are not suspended or have not yet reached suspended
1950 	 * mode, then let the operation proceed.
1951 	 */
1952 	if ((mp = *mpp) == NULL)
1953 		return (0);
1954 
1955 	/*
1956 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1957 	 * a vfs_ref().
1958 	 * As long as a vnode is not provided we need to acquire a
1959 	 * refcount for the provided mountpoint too, in order to
1960 	 * emulate a vfs_ref().
1961 	 */
1962 	MNT_ILOCK(mp);
1963 	if (vp == NULL && (flags & V_MNTREF) == 0)
1964 		MNT_REF(mp);
1965 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1966 		mp->mnt_secondary_writes++;
1967 		mp->mnt_secondary_accwrites++;
1968 		MNT_IUNLOCK(mp);
1969 		return (0);
1970 	}
1971 	if (flags & V_NOWAIT) {
1972 		MNT_REL(mp);
1973 		MNT_IUNLOCK(mp);
1974 		return (EWOULDBLOCK);
1975 	}
1976 	/*
1977 	 * Wait for the suspension to finish.
1978 	 */
1979 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1980 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1981 	    "suspfs", 0);
1982 	vfs_rel(mp);
1983 	if (error == 0)
1984 		goto retry;
1985 	return (error);
1986 }
1987 
1988 /*
1989  * Filesystem write operation has completed. If we are suspending and this
1990  * operation is the last one, notify the suspender that the suspension is
1991  * now in effect.
1992  */
1993 void
1994 vn_finished_write(struct mount *mp)
1995 {
1996 	struct mount_pcpu *mpcpu;
1997 	int c;
1998 
1999 	if (mp == NULL)
2000 		return;
2001 
2002 	if (vfs_op_thread_enter(mp, mpcpu)) {
2003 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2004 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2005 		vfs_op_thread_exit(mp, mpcpu);
2006 		return;
2007 	}
2008 
2009 	MNT_ILOCK(mp);
2010 	vfs_assert_mount_counters(mp);
2011 	MNT_REL(mp);
2012 	c = --mp->mnt_writeopcount;
2013 	if (mp->mnt_vfs_ops == 0) {
2014 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2015 		MNT_IUNLOCK(mp);
2016 		return;
2017 	}
2018 	if (c < 0)
2019 		vfs_dump_mount_counters(mp);
2020 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2021 		wakeup(&mp->mnt_writeopcount);
2022 	MNT_IUNLOCK(mp);
2023 }
2024 
2025 /*
2026  * Filesystem secondary write operation has completed. If we are
2027  * suspending and this operation is the last one, notify the suspender
2028  * that the suspension is now in effect.
2029  */
2030 void
2031 vn_finished_secondary_write(struct mount *mp)
2032 {
2033 	if (mp == NULL)
2034 		return;
2035 	MNT_ILOCK(mp);
2036 	MNT_REL(mp);
2037 	mp->mnt_secondary_writes--;
2038 	if (mp->mnt_secondary_writes < 0)
2039 		panic("vn_finished_secondary_write: neg cnt");
2040 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2041 	    mp->mnt_secondary_writes <= 0)
2042 		wakeup(&mp->mnt_secondary_writes);
2043 	MNT_IUNLOCK(mp);
2044 }
2045 
2046 /*
2047  * Request a filesystem to suspend write operations.
2048  */
2049 int
2050 vfs_write_suspend(struct mount *mp, int flags)
2051 {
2052 	int error;
2053 
2054 	vfs_op_enter(mp);
2055 
2056 	MNT_ILOCK(mp);
2057 	vfs_assert_mount_counters(mp);
2058 	if (mp->mnt_susp_owner == curthread) {
2059 		vfs_op_exit_locked(mp);
2060 		MNT_IUNLOCK(mp);
2061 		return (EALREADY);
2062 	}
2063 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2064 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2065 
2066 	/*
2067 	 * Unmount holds a write reference on the mount point.  If we
2068 	 * own busy reference and drain for writers, we deadlock with
2069 	 * the reference draining in the unmount path.  Callers of
2070 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2071 	 * vfs_busy() reference is owned and caller is not in the
2072 	 * unmount context.
2073 	 */
2074 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2075 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2076 		vfs_op_exit_locked(mp);
2077 		MNT_IUNLOCK(mp);
2078 		return (EBUSY);
2079 	}
2080 
2081 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2082 	mp->mnt_susp_owner = curthread;
2083 	if (mp->mnt_writeopcount > 0)
2084 		(void) msleep(&mp->mnt_writeopcount,
2085 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2086 	else
2087 		MNT_IUNLOCK(mp);
2088 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2089 		vfs_write_resume(mp, 0);
2090 		/* vfs_write_resume does vfs_op_exit() for us */
2091 	}
2092 	return (error);
2093 }
2094 
2095 /*
2096  * Request a filesystem to resume write operations.
2097  */
2098 void
2099 vfs_write_resume(struct mount *mp, int flags)
2100 {
2101 
2102 	MNT_ILOCK(mp);
2103 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2104 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2105 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2106 				       MNTK_SUSPENDED);
2107 		mp->mnt_susp_owner = NULL;
2108 		wakeup(&mp->mnt_writeopcount);
2109 		wakeup(&mp->mnt_flag);
2110 		curthread->td_pflags &= ~TDP_IGNSUSP;
2111 		if ((flags & VR_START_WRITE) != 0) {
2112 			MNT_REF(mp);
2113 			mp->mnt_writeopcount++;
2114 		}
2115 		MNT_IUNLOCK(mp);
2116 		if ((flags & VR_NO_SUSPCLR) == 0)
2117 			VFS_SUSP_CLEAN(mp);
2118 		vfs_op_exit(mp);
2119 	} else if ((flags & VR_START_WRITE) != 0) {
2120 		MNT_REF(mp);
2121 		vn_start_write_refed(mp, 0, true);
2122 	} else {
2123 		MNT_IUNLOCK(mp);
2124 	}
2125 }
2126 
2127 /*
2128  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2129  * methods.
2130  */
2131 int
2132 vfs_write_suspend_umnt(struct mount *mp)
2133 {
2134 	int error;
2135 
2136 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2137 	    ("vfs_write_suspend_umnt: recursed"));
2138 
2139 	/* dounmount() already called vn_start_write(). */
2140 	for (;;) {
2141 		vn_finished_write(mp);
2142 		error = vfs_write_suspend(mp, 0);
2143 		if (error != 0) {
2144 			vn_start_write(NULL, &mp, V_WAIT);
2145 			return (error);
2146 		}
2147 		MNT_ILOCK(mp);
2148 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2149 			break;
2150 		MNT_IUNLOCK(mp);
2151 		vn_start_write(NULL, &mp, V_WAIT);
2152 	}
2153 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2154 	wakeup(&mp->mnt_flag);
2155 	MNT_IUNLOCK(mp);
2156 	curthread->td_pflags |= TDP_IGNSUSP;
2157 	return (0);
2158 }
2159 
2160 /*
2161  * Implement kqueues for files by translating it to vnode operation.
2162  */
2163 static int
2164 vn_kqfilter(struct file *fp, struct knote *kn)
2165 {
2166 
2167 	return (VOP_KQFILTER(fp->f_vnode, kn));
2168 }
2169 
2170 int
2171 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2172 {
2173 	if ((fp->f_flag & FKQALLOWED) == 0)
2174 		return (EBADF);
2175 	return (vn_kqfilter(fp, kn));
2176 }
2177 
2178 /*
2179  * Simplified in-kernel wrapper calls for extended attribute access.
2180  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2181  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2182  */
2183 int
2184 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2185     const char *attrname, int *buflen, char *buf, struct thread *td)
2186 {
2187 	struct uio	auio;
2188 	struct iovec	iov;
2189 	int	error;
2190 
2191 	iov.iov_len = *buflen;
2192 	iov.iov_base = buf;
2193 
2194 	auio.uio_iov = &iov;
2195 	auio.uio_iovcnt = 1;
2196 	auio.uio_rw = UIO_READ;
2197 	auio.uio_segflg = UIO_SYSSPACE;
2198 	auio.uio_td = td;
2199 	auio.uio_offset = 0;
2200 	auio.uio_resid = *buflen;
2201 
2202 	if ((ioflg & IO_NODELOCKED) == 0)
2203 		vn_lock(vp, LK_SHARED | LK_RETRY);
2204 
2205 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2206 
2207 	/* authorize attribute retrieval as kernel */
2208 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2209 	    td);
2210 
2211 	if ((ioflg & IO_NODELOCKED) == 0)
2212 		VOP_UNLOCK(vp);
2213 
2214 	if (error == 0) {
2215 		*buflen = *buflen - auio.uio_resid;
2216 	}
2217 
2218 	return (error);
2219 }
2220 
2221 /*
2222  * XXX failure mode if partially written?
2223  */
2224 int
2225 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2226     const char *attrname, int buflen, char *buf, struct thread *td)
2227 {
2228 	struct uio	auio;
2229 	struct iovec	iov;
2230 	struct mount	*mp;
2231 	int	error;
2232 
2233 	iov.iov_len = buflen;
2234 	iov.iov_base = buf;
2235 
2236 	auio.uio_iov = &iov;
2237 	auio.uio_iovcnt = 1;
2238 	auio.uio_rw = UIO_WRITE;
2239 	auio.uio_segflg = UIO_SYSSPACE;
2240 	auio.uio_td = td;
2241 	auio.uio_offset = 0;
2242 	auio.uio_resid = buflen;
2243 
2244 	if ((ioflg & IO_NODELOCKED) == 0) {
2245 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2246 			return (error);
2247 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2248 	}
2249 
2250 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2251 
2252 	/* authorize attribute setting as kernel */
2253 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2254 
2255 	if ((ioflg & IO_NODELOCKED) == 0) {
2256 		vn_finished_write(mp);
2257 		VOP_UNLOCK(vp);
2258 	}
2259 
2260 	return (error);
2261 }
2262 
2263 int
2264 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2265     const char *attrname, struct thread *td)
2266 {
2267 	struct mount	*mp;
2268 	int	error;
2269 
2270 	if ((ioflg & IO_NODELOCKED) == 0) {
2271 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2272 			return (error);
2273 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2274 	}
2275 
2276 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2277 
2278 	/* authorize attribute removal as kernel */
2279 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2280 	if (error == EOPNOTSUPP)
2281 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2282 		    NULL, td);
2283 
2284 	if ((ioflg & IO_NODELOCKED) == 0) {
2285 		vn_finished_write(mp);
2286 		VOP_UNLOCK(vp);
2287 	}
2288 
2289 	return (error);
2290 }
2291 
2292 static int
2293 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2294     struct vnode **rvp)
2295 {
2296 
2297 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2298 }
2299 
2300 int
2301 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2302 {
2303 
2304 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2305 	    lkflags, rvp));
2306 }
2307 
2308 int
2309 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2310     int lkflags, struct vnode **rvp)
2311 {
2312 	struct mount *mp;
2313 	int ltype, error;
2314 
2315 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2316 	mp = vp->v_mount;
2317 	ltype = VOP_ISLOCKED(vp);
2318 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2319 	    ("vn_vget_ino: vp not locked"));
2320 	error = vfs_busy(mp, MBF_NOWAIT);
2321 	if (error != 0) {
2322 		vfs_ref(mp);
2323 		VOP_UNLOCK(vp);
2324 		error = vfs_busy(mp, 0);
2325 		vn_lock(vp, ltype | LK_RETRY);
2326 		vfs_rel(mp);
2327 		if (error != 0)
2328 			return (ENOENT);
2329 		if (VN_IS_DOOMED(vp)) {
2330 			vfs_unbusy(mp);
2331 			return (ENOENT);
2332 		}
2333 	}
2334 	VOP_UNLOCK(vp);
2335 	error = alloc(mp, alloc_arg, lkflags, rvp);
2336 	vfs_unbusy(mp);
2337 	if (error != 0 || *rvp != vp)
2338 		vn_lock(vp, ltype | LK_RETRY);
2339 	if (VN_IS_DOOMED(vp)) {
2340 		if (error == 0) {
2341 			if (*rvp == vp)
2342 				vunref(vp);
2343 			else
2344 				vput(*rvp);
2345 		}
2346 		error = ENOENT;
2347 	}
2348 	return (error);
2349 }
2350 
2351 int
2352 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2353     struct thread *td)
2354 {
2355 	off_t lim;
2356 	bool ktr_write;
2357 
2358 	if (td == NULL)
2359 		return (0);
2360 
2361 	/*
2362 	 * There are conditions where the limit is to be ignored.
2363 	 * However, since it is almost never reached, check it first.
2364 	 */
2365 	ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2366 	lim = lim_cur(td, RLIMIT_FSIZE);
2367 	if (__predict_false(ktr_write))
2368 		lim = td->td_ktr_io_lim;
2369 	if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2370 		return (0);
2371 
2372 	/*
2373 	 * The limit is reached.
2374 	 */
2375 	if (vp->v_type != VREG ||
2376 	    (td->td_pflags2 & TDP2_ACCT) != 0)
2377 		return (0);
2378 
2379 	if (!ktr_write || ktr_filesize_limit_signal) {
2380 		PROC_LOCK(td->td_proc);
2381 		kern_psignal(td->td_proc, SIGXFSZ);
2382 		PROC_UNLOCK(td->td_proc);
2383 	}
2384 	return (EFBIG);
2385 }
2386 
2387 int
2388 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2389     struct thread *td)
2390 {
2391 	struct vnode *vp;
2392 
2393 	vp = fp->f_vnode;
2394 #ifdef AUDIT
2395 	vn_lock(vp, LK_SHARED | LK_RETRY);
2396 	AUDIT_ARG_VNODE1(vp);
2397 	VOP_UNLOCK(vp);
2398 #endif
2399 	return (setfmode(td, active_cred, vp, mode));
2400 }
2401 
2402 int
2403 vn_chown(struct file *fp, uid_t uid, gid_t gid, 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 (setfown(td, active_cred, vp, uid, gid));
2415 }
2416 
2417 void
2418 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2419 {
2420 	vm_object_t object;
2421 
2422 	if ((object = vp->v_object) == NULL)
2423 		return;
2424 	VM_OBJECT_WLOCK(object);
2425 	vm_object_page_remove(object, start, end, 0);
2426 	VM_OBJECT_WUNLOCK(object);
2427 }
2428 
2429 int
2430 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2431     struct ucred *cred)
2432 {
2433 	struct vattr va;
2434 	daddr_t bn, bnp;
2435 	uint64_t bsize;
2436 	off_t noff;
2437 	int error;
2438 
2439 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2440 	    ("%s: Wrong command %lu", __func__, cmd));
2441 	ASSERT_VOP_LOCKED(vp, "vn_bmap_seekhole_locked");
2442 
2443 	if (vp->v_type != VREG) {
2444 		error = ENOTTY;
2445 		goto out;
2446 	}
2447 	error = VOP_GETATTR(vp, &va, cred);
2448 	if (error != 0)
2449 		goto out;
2450 	noff = *off;
2451 	if (noff >= va.va_size) {
2452 		error = ENXIO;
2453 		goto out;
2454 	}
2455 	bsize = vp->v_mount->mnt_stat.f_iosize;
2456 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2457 	    noff % bsize) {
2458 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2459 		if (error == EOPNOTSUPP) {
2460 			error = ENOTTY;
2461 			goto out;
2462 		}
2463 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2464 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2465 			noff = bn * bsize;
2466 			if (noff < *off)
2467 				noff = *off;
2468 			goto out;
2469 		}
2470 	}
2471 	if (noff > va.va_size)
2472 		noff = va.va_size;
2473 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2474 	if (cmd == FIOSEEKDATA)
2475 		error = ENXIO;
2476 out:
2477 	if (error == 0)
2478 		*off = noff;
2479 	return (error);
2480 }
2481 
2482 int
2483 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2484 {
2485 	int error;
2486 
2487 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2488 	    ("%s: Wrong command %lu", __func__, cmd));
2489 
2490 	if (vn_lock(vp, LK_SHARED) != 0)
2491 		return (EBADF);
2492 	error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2493 	VOP_UNLOCK(vp);
2494 	return (error);
2495 }
2496 
2497 int
2498 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2499 {
2500 	struct ucred *cred;
2501 	struct vnode *vp;
2502 	struct vattr vattr;
2503 	off_t foffset, size;
2504 	int error, noneg;
2505 
2506 	cred = td->td_ucred;
2507 	vp = fp->f_vnode;
2508 	foffset = foffset_lock(fp, 0);
2509 	noneg = (vp->v_type != VCHR);
2510 	error = 0;
2511 	switch (whence) {
2512 	case L_INCR:
2513 		if (noneg &&
2514 		    (foffset < 0 ||
2515 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2516 			error = EOVERFLOW;
2517 			break;
2518 		}
2519 		offset += foffset;
2520 		break;
2521 	case L_XTND:
2522 		vn_lock(vp, LK_SHARED | LK_RETRY);
2523 		error = VOP_GETATTR(vp, &vattr, cred);
2524 		VOP_UNLOCK(vp);
2525 		if (error)
2526 			break;
2527 
2528 		/*
2529 		 * If the file references a disk device, then fetch
2530 		 * the media size and use that to determine the ending
2531 		 * offset.
2532 		 */
2533 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2534 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2535 			vattr.va_size = size;
2536 		if (noneg &&
2537 		    (vattr.va_size > OFF_MAX ||
2538 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2539 			error = EOVERFLOW;
2540 			break;
2541 		}
2542 		offset += vattr.va_size;
2543 		break;
2544 	case L_SET:
2545 		break;
2546 	case SEEK_DATA:
2547 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2548 		if (error == ENOTTY)
2549 			error = EINVAL;
2550 		break;
2551 	case SEEK_HOLE:
2552 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2553 		if (error == ENOTTY)
2554 			error = EINVAL;
2555 		break;
2556 	default:
2557 		error = EINVAL;
2558 	}
2559 	if (error == 0 && noneg && offset < 0)
2560 		error = EINVAL;
2561 	if (error != 0)
2562 		goto drop;
2563 	VFS_KNOTE_UNLOCKED(vp, 0);
2564 	td->td_uretoff.tdu_off = offset;
2565 drop:
2566 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2567 	return (error);
2568 }
2569 
2570 int
2571 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2572     struct thread *td)
2573 {
2574 	int error;
2575 
2576 	/*
2577 	 * Grant permission if the caller is the owner of the file, or
2578 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2579 	 * on the file.  If the time pointer is null, then write
2580 	 * permission on the file is also sufficient.
2581 	 *
2582 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2583 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2584 	 * will be allowed to set the times [..] to the current
2585 	 * server time.
2586 	 */
2587 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2588 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2589 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2590 	return (error);
2591 }
2592 
2593 int
2594 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2595 {
2596 	struct vnode *vp;
2597 	int error;
2598 
2599 	if (fp->f_type == DTYPE_FIFO)
2600 		kif->kf_type = KF_TYPE_FIFO;
2601 	else
2602 		kif->kf_type = KF_TYPE_VNODE;
2603 	vp = fp->f_vnode;
2604 	vref(vp);
2605 	FILEDESC_SUNLOCK(fdp);
2606 	error = vn_fill_kinfo_vnode(vp, kif);
2607 	vrele(vp);
2608 	FILEDESC_SLOCK(fdp);
2609 	return (error);
2610 }
2611 
2612 static inline void
2613 vn_fill_junk(struct kinfo_file *kif)
2614 {
2615 	size_t len, olen;
2616 
2617 	/*
2618 	 * Simulate vn_fullpath returning changing values for a given
2619 	 * vp during e.g. coredump.
2620 	 */
2621 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2622 	olen = strlen(kif->kf_path);
2623 	if (len < olen)
2624 		strcpy(&kif->kf_path[len - 1], "$");
2625 	else
2626 		for (; olen < len; olen++)
2627 			strcpy(&kif->kf_path[olen], "A");
2628 }
2629 
2630 int
2631 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2632 {
2633 	struct vattr va;
2634 	char *fullpath, *freepath;
2635 	int error;
2636 
2637 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2638 	freepath = NULL;
2639 	fullpath = "-";
2640 	error = vn_fullpath(vp, &fullpath, &freepath);
2641 	if (error == 0) {
2642 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2643 	}
2644 	if (freepath != NULL)
2645 		free(freepath, M_TEMP);
2646 
2647 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2648 		vn_fill_junk(kif);
2649 	);
2650 
2651 	/*
2652 	 * Retrieve vnode attributes.
2653 	 */
2654 	va.va_fsid = VNOVAL;
2655 	va.va_rdev = NODEV;
2656 	vn_lock(vp, LK_SHARED | LK_RETRY);
2657 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2658 	VOP_UNLOCK(vp);
2659 	if (error != 0)
2660 		return (error);
2661 	if (va.va_fsid != VNOVAL)
2662 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2663 	else
2664 		kif->kf_un.kf_file.kf_file_fsid =
2665 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2666 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2667 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2668 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2669 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2670 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2671 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2672 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2673 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2674 	return (0);
2675 }
2676 
2677 int
2678 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2679     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2680     struct thread *td)
2681 {
2682 #ifdef HWPMC_HOOKS
2683 	struct pmckern_map_in pkm;
2684 #endif
2685 	struct mount *mp;
2686 	struct vnode *vp;
2687 	vm_object_t object;
2688 	vm_prot_t maxprot;
2689 	boolean_t writecounted;
2690 	int error;
2691 
2692 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2693     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2694 	/*
2695 	 * POSIX shared-memory objects are defined to have
2696 	 * kernel persistence, and are not defined to support
2697 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2698 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2699 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2700 	 * flag to request this behavior.
2701 	 */
2702 	if ((fp->f_flag & FPOSIXSHM) != 0)
2703 		flags |= MAP_NOSYNC;
2704 #endif
2705 	vp = fp->f_vnode;
2706 
2707 	/*
2708 	 * Ensure that file and memory protections are
2709 	 * compatible.  Note that we only worry about
2710 	 * writability if mapping is shared; in this case,
2711 	 * current and max prot are dictated by the open file.
2712 	 * XXX use the vnode instead?  Problem is: what
2713 	 * credentials do we use for determination? What if
2714 	 * proc does a setuid?
2715 	 */
2716 	mp = vp->v_mount;
2717 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2718 		maxprot = VM_PROT_NONE;
2719 		if ((prot & VM_PROT_EXECUTE) != 0)
2720 			return (EACCES);
2721 	} else
2722 		maxprot = VM_PROT_EXECUTE;
2723 	if ((fp->f_flag & FREAD) != 0)
2724 		maxprot |= VM_PROT_READ;
2725 	else if ((prot & VM_PROT_READ) != 0)
2726 		return (EACCES);
2727 
2728 	/*
2729 	 * If we are sharing potential changes via MAP_SHARED and we
2730 	 * are trying to get write permission although we opened it
2731 	 * without asking for it, bail out.
2732 	 */
2733 	if ((flags & MAP_SHARED) != 0) {
2734 		if ((fp->f_flag & FWRITE) != 0)
2735 			maxprot |= VM_PROT_WRITE;
2736 		else if ((prot & VM_PROT_WRITE) != 0)
2737 			return (EACCES);
2738 	} else {
2739 		maxprot |= VM_PROT_WRITE;
2740 		cap_maxprot |= VM_PROT_WRITE;
2741 	}
2742 	maxprot &= cap_maxprot;
2743 
2744 	/*
2745 	 * For regular files and shared memory, POSIX requires that
2746 	 * the value of foff be a legitimate offset within the data
2747 	 * object.  In particular, negative offsets are invalid.
2748 	 * Blocking negative offsets and overflows here avoids
2749 	 * possible wraparound or user-level access into reserved
2750 	 * ranges of the data object later.  In contrast, POSIX does
2751 	 * not dictate how offsets are used by device drivers, so in
2752 	 * the case of a device mapping a negative offset is passed
2753 	 * on.
2754 	 */
2755 	if (
2756 #ifdef _LP64
2757 	    size > OFF_MAX ||
2758 #endif
2759 	    foff > OFF_MAX - size)
2760 		return (EINVAL);
2761 
2762 	writecounted = FALSE;
2763 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2764 	    &foff, &object, &writecounted);
2765 	if (error != 0)
2766 		return (error);
2767 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2768 	    foff, writecounted, td);
2769 	if (error != 0) {
2770 		/*
2771 		 * If this mapping was accounted for in the vnode's
2772 		 * writecount, then undo that now.
2773 		 */
2774 		if (writecounted)
2775 			vm_pager_release_writecount(object, 0, size);
2776 		vm_object_deallocate(object);
2777 	}
2778 #ifdef HWPMC_HOOKS
2779 	/* Inform hwpmc(4) if an executable is being mapped. */
2780 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2781 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2782 			pkm.pm_file = vp;
2783 			pkm.pm_address = (uintptr_t) *addr;
2784 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2785 		}
2786 	}
2787 #endif
2788 	return (error);
2789 }
2790 
2791 void
2792 vn_fsid(struct vnode *vp, struct vattr *va)
2793 {
2794 	fsid_t *f;
2795 
2796 	f = &vp->v_mount->mnt_stat.f_fsid;
2797 	va->va_fsid = (uint32_t)f->val[1];
2798 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2799 	va->va_fsid += (uint32_t)f->val[0];
2800 }
2801 
2802 int
2803 vn_fsync_buf(struct vnode *vp, int waitfor)
2804 {
2805 	struct buf *bp, *nbp;
2806 	struct bufobj *bo;
2807 	struct mount *mp;
2808 	int error, maxretry;
2809 
2810 	error = 0;
2811 	maxretry = 10000;     /* large, arbitrarily chosen */
2812 	mp = NULL;
2813 	if (vp->v_type == VCHR) {
2814 		VI_LOCK(vp);
2815 		mp = vp->v_rdev->si_mountpt;
2816 		VI_UNLOCK(vp);
2817 	}
2818 	bo = &vp->v_bufobj;
2819 	BO_LOCK(bo);
2820 loop1:
2821 	/*
2822 	 * MARK/SCAN initialization to avoid infinite loops.
2823 	 */
2824         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2825 		bp->b_vflags &= ~BV_SCANNED;
2826 		bp->b_error = 0;
2827 	}
2828 
2829 	/*
2830 	 * Flush all dirty buffers associated with a vnode.
2831 	 */
2832 loop2:
2833 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2834 		if ((bp->b_vflags & BV_SCANNED) != 0)
2835 			continue;
2836 		bp->b_vflags |= BV_SCANNED;
2837 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2838 			if (waitfor != MNT_WAIT)
2839 				continue;
2840 			if (BUF_LOCK(bp,
2841 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2842 			    BO_LOCKPTR(bo)) != 0) {
2843 				BO_LOCK(bo);
2844 				goto loop1;
2845 			}
2846 			BO_LOCK(bo);
2847 		}
2848 		BO_UNLOCK(bo);
2849 		KASSERT(bp->b_bufobj == bo,
2850 		    ("bp %p wrong b_bufobj %p should be %p",
2851 		    bp, bp->b_bufobj, bo));
2852 		if ((bp->b_flags & B_DELWRI) == 0)
2853 			panic("fsync: not dirty");
2854 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2855 			vfs_bio_awrite(bp);
2856 		} else {
2857 			bremfree(bp);
2858 			bawrite(bp);
2859 		}
2860 		if (maxretry < 1000)
2861 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2862 		BO_LOCK(bo);
2863 		goto loop2;
2864 	}
2865 
2866 	/*
2867 	 * If synchronous the caller expects us to completely resolve all
2868 	 * dirty buffers in the system.  Wait for in-progress I/O to
2869 	 * complete (which could include background bitmap writes), then
2870 	 * retry if dirty blocks still exist.
2871 	 */
2872 	if (waitfor == MNT_WAIT) {
2873 		bufobj_wwait(bo, 0, 0);
2874 		if (bo->bo_dirty.bv_cnt > 0) {
2875 			/*
2876 			 * If we are unable to write any of these buffers
2877 			 * then we fail now rather than trying endlessly
2878 			 * to write them out.
2879 			 */
2880 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2881 				if ((error = bp->b_error) != 0)
2882 					break;
2883 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2884 			    (error == 0 && --maxretry >= 0))
2885 				goto loop1;
2886 			if (error == 0)
2887 				error = EAGAIN;
2888 		}
2889 	}
2890 	BO_UNLOCK(bo);
2891 	if (error != 0)
2892 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2893 
2894 	return (error);
2895 }
2896 
2897 /*
2898  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2899  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2900  * to do the actual copy.
2901  * vn_generic_copy_file_range() is factored out, so it can be called
2902  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2903  * different file systems.
2904  */
2905 int
2906 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2907     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2908     struct ucred *outcred, struct thread *fsize_td)
2909 {
2910 	int error;
2911 	size_t len;
2912 	uint64_t uval;
2913 
2914 	len = *lenp;
2915 	*lenp = 0;		/* For error returns. */
2916 	error = 0;
2917 
2918 	/* Do some sanity checks on the arguments. */
2919 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2920 		error = EISDIR;
2921 	else if (*inoffp < 0 || *outoffp < 0 ||
2922 	    invp->v_type != VREG || outvp->v_type != VREG)
2923 		error = EINVAL;
2924 	if (error != 0)
2925 		goto out;
2926 
2927 	/* Ensure offset + len does not wrap around. */
2928 	uval = *inoffp;
2929 	uval += len;
2930 	if (uval > INT64_MAX)
2931 		len = INT64_MAX - *inoffp;
2932 	uval = *outoffp;
2933 	uval += len;
2934 	if (uval > INT64_MAX)
2935 		len = INT64_MAX - *outoffp;
2936 	if (len == 0)
2937 		goto out;
2938 
2939 	/*
2940 	 * If the two vnode are for the same file system, call
2941 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2942 	 * which can handle copies across multiple file systems.
2943 	 */
2944 	*lenp = len;
2945 	if (invp->v_mount == outvp->v_mount)
2946 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2947 		    lenp, flags, incred, outcred, fsize_td);
2948 	else
2949 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2950 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2951 out:
2952 	return (error);
2953 }
2954 
2955 /*
2956  * Test len bytes of data starting at dat for all bytes == 0.
2957  * Return true if all bytes are zero, false otherwise.
2958  * Expects dat to be well aligned.
2959  */
2960 static bool
2961 mem_iszero(void *dat, int len)
2962 {
2963 	int i;
2964 	const u_int *p;
2965 	const char *cp;
2966 
2967 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2968 		if (len >= sizeof(*p)) {
2969 			if (*p != 0)
2970 				return (false);
2971 		} else {
2972 			cp = (const char *)p;
2973 			for (i = 0; i < len; i++, cp++)
2974 				if (*cp != '\0')
2975 					return (false);
2976 		}
2977 	}
2978 	return (true);
2979 }
2980 
2981 /*
2982  * Look for a hole in the output file and, if found, adjust *outoffp
2983  * and *xferp to skip past the hole.
2984  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2985  * to be written as 0's upon return.
2986  */
2987 static off_t
2988 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2989     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2990 {
2991 	int error;
2992 	off_t delta;
2993 
2994 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2995 		*dataoffp = *outoffp;
2996 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2997 		    curthread);
2998 		if (error == 0) {
2999 			*holeoffp = *dataoffp;
3000 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3001 			    curthread);
3002 		}
3003 		if (error != 0 || *holeoffp == *dataoffp) {
3004 			/*
3005 			 * Since outvp is unlocked, it may be possible for
3006 			 * another thread to do a truncate(), lseek(), write()
3007 			 * creating a hole at startoff between the above
3008 			 * VOP_IOCTL() calls, if the other thread does not do
3009 			 * rangelocking.
3010 			 * If that happens, *holeoffp == *dataoffp and finding
3011 			 * the hole has failed, so disable vn_skip_hole().
3012 			 */
3013 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
3014 			return (xfer2);
3015 		}
3016 		KASSERT(*dataoffp >= *outoffp,
3017 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3018 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
3019 		KASSERT(*holeoffp > *dataoffp,
3020 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3021 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3022 	}
3023 
3024 	/*
3025 	 * If there is a hole before the data starts, advance *outoffp and
3026 	 * *xferp past the hole.
3027 	 */
3028 	if (*dataoffp > *outoffp) {
3029 		delta = *dataoffp - *outoffp;
3030 		if (delta >= *xferp) {
3031 			/* Entire *xferp is a hole. */
3032 			*outoffp += *xferp;
3033 			*xferp = 0;
3034 			return (0);
3035 		}
3036 		*xferp -= delta;
3037 		*outoffp += delta;
3038 		xfer2 = MIN(xfer2, *xferp);
3039 	}
3040 
3041 	/*
3042 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3043 	 * that the write ends at the start of the hole.
3044 	 * *holeoffp should always be greater than *outoffp, but for the
3045 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3046 	 * value.
3047 	 */
3048 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3049 		xfer2 = *holeoffp - *outoffp;
3050 	return (xfer2);
3051 }
3052 
3053 /*
3054  * Write an xfer sized chunk to outvp in blksize blocks from dat.
3055  * dat is a maximum of blksize in length and can be written repeatedly in
3056  * the chunk.
3057  * If growfile == true, just grow the file via vn_truncate_locked() instead
3058  * of doing actual writes.
3059  * If checkhole == true, a hole is being punched, so skip over any hole
3060  * already in the output file.
3061  */
3062 static int
3063 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3064     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3065 {
3066 	struct mount *mp;
3067 	off_t dataoff, holeoff, xfer2;
3068 	int error;
3069 
3070 	/*
3071 	 * Loop around doing writes of blksize until write has been completed.
3072 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3073 	 * done for each iteration, since the xfer argument can be very
3074 	 * large if there is a large hole to punch in the output file.
3075 	 */
3076 	error = 0;
3077 	holeoff = 0;
3078 	do {
3079 		xfer2 = MIN(xfer, blksize);
3080 		if (checkhole) {
3081 			/*
3082 			 * Punching a hole.  Skip writing if there is
3083 			 * already a hole in the output file.
3084 			 */
3085 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3086 			    &dataoff, &holeoff, cred);
3087 			if (xfer == 0)
3088 				break;
3089 			if (holeoff < 0)
3090 				checkhole = false;
3091 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3092 			    (intmax_t)xfer2));
3093 		}
3094 		bwillwrite();
3095 		mp = NULL;
3096 		error = vn_start_write(outvp, &mp, V_WAIT);
3097 		if (error != 0)
3098 			break;
3099 		if (growfile) {
3100 			error = vn_lock(outvp, LK_EXCLUSIVE);
3101 			if (error == 0) {
3102 				error = vn_truncate_locked(outvp, outoff + xfer,
3103 				    false, cred);
3104 				VOP_UNLOCK(outvp);
3105 			}
3106 		} else {
3107 			error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3108 			if (error == 0) {
3109 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3110 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3111 				    curthread->td_ucred, cred, NULL, curthread);
3112 				outoff += xfer2;
3113 				xfer -= xfer2;
3114 				VOP_UNLOCK(outvp);
3115 			}
3116 		}
3117 		if (mp != NULL)
3118 			vn_finished_write(mp);
3119 	} while (!growfile && xfer > 0 && error == 0);
3120 	return (error);
3121 }
3122 
3123 /*
3124  * Copy a byte range of one file to another.  This function can handle the
3125  * case where invp and outvp are on different file systems.
3126  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3127  * is no better file system specific way to do it.
3128  */
3129 int
3130 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3131     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3132     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3133 {
3134 	struct vattr va, inva;
3135 	struct mount *mp;
3136 	struct uio io;
3137 	off_t startoff, endoff, xfer, xfer2;
3138 	u_long blksize;
3139 	int error, interrupted;
3140 	bool cantseek, readzeros, eof, lastblock, holetoeof;
3141 	ssize_t aresid;
3142 	size_t copylen, len, rem, savlen;
3143 	char *dat;
3144 	long holein, holeout;
3145 
3146 	holein = holeout = 0;
3147 	savlen = len = *lenp;
3148 	error = 0;
3149 	interrupted = 0;
3150 	dat = NULL;
3151 
3152 	error = vn_lock(invp, LK_SHARED);
3153 	if (error != 0)
3154 		goto out;
3155 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3156 		holein = 0;
3157 	if (holein > 0)
3158 		error = VOP_GETATTR(invp, &inva, incred);
3159 	VOP_UNLOCK(invp);
3160 	if (error != 0)
3161 		goto out;
3162 
3163 	mp = NULL;
3164 	error = vn_start_write(outvp, &mp, V_WAIT);
3165 	if (error == 0)
3166 		error = vn_lock(outvp, LK_EXCLUSIVE);
3167 	if (error == 0) {
3168 		/*
3169 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3170 		 * now that outvp is locked.
3171 		 */
3172 		if (fsize_td != NULL) {
3173 			io.uio_offset = *outoffp;
3174 			io.uio_resid = len;
3175 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3176 			if (error != 0)
3177 				error = EFBIG;
3178 		}
3179 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3180 			holeout = 0;
3181 		/*
3182 		 * Holes that are past EOF do not need to be written as a block
3183 		 * of zero bytes.  So, truncate the output file as far as
3184 		 * possible and then use va.va_size to decide if writing 0
3185 		 * bytes is necessary in the loop below.
3186 		 */
3187 		if (error == 0)
3188 			error = VOP_GETATTR(outvp, &va, outcred);
3189 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3190 		    *outoffp + len) {
3191 #ifdef MAC
3192 			error = mac_vnode_check_write(curthread->td_ucred,
3193 			    outcred, outvp);
3194 			if (error == 0)
3195 #endif
3196 				error = vn_truncate_locked(outvp, *outoffp,
3197 				    false, outcred);
3198 			if (error == 0)
3199 				va.va_size = *outoffp;
3200 		}
3201 		VOP_UNLOCK(outvp);
3202 	}
3203 	if (mp != NULL)
3204 		vn_finished_write(mp);
3205 	if (error != 0)
3206 		goto out;
3207 
3208 	/*
3209 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3210 	 * If hole sizes aren't available, set the blksize to the larger
3211 	 * f_iosize of invp and outvp.
3212 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3213 	 * This value is clipped at 4Kbytes and 1Mbyte.
3214 	 */
3215 	blksize = MAX(holein, holeout);
3216 
3217 	/* Clip len to end at an exact multiple of hole size. */
3218 	if (blksize > 1) {
3219 		rem = *inoffp % blksize;
3220 		if (rem > 0)
3221 			rem = blksize - rem;
3222 		if (len > rem && len - rem > blksize)
3223 			len = savlen = rounddown(len - rem, blksize) + rem;
3224 	}
3225 
3226 	if (blksize <= 1)
3227 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3228 		    outvp->v_mount->mnt_stat.f_iosize);
3229 	if (blksize < 4096)
3230 		blksize = 4096;
3231 	else if (blksize > 1024 * 1024)
3232 		blksize = 1024 * 1024;
3233 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3234 
3235 	/*
3236 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3237 	 * to find holes.  Otherwise, just scan the read block for all 0s
3238 	 * in the inner loop where the data copying is done.
3239 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3240 	 * support holes on the server, but do not support FIOSEEKHOLE.
3241 	 */
3242 	holetoeof = eof = false;
3243 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3244 		endoff = 0;			/* To shut up compilers. */
3245 		cantseek = true;
3246 		startoff = *inoffp;
3247 		copylen = len;
3248 
3249 		/*
3250 		 * Find the next data area.  If there is just a hole to EOF,
3251 		 * FIOSEEKDATA should fail with ENXIO.
3252 		 * (I do not know if any file system will report a hole to
3253 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3254 		 *  will fail for those file systems.)
3255 		 *
3256 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3257 		 * the code just falls through to the inner copy loop.
3258 		 */
3259 		error = EINVAL;
3260 		if (holein > 0) {
3261 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3262 			    incred, curthread);
3263 			if (error == ENXIO) {
3264 				startoff = endoff = inva.va_size;
3265 				eof = holetoeof = true;
3266 				error = 0;
3267 			}
3268 		}
3269 		if (error == 0 && !holetoeof) {
3270 			endoff = startoff;
3271 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3272 			    incred, curthread);
3273 			/*
3274 			 * Since invp is unlocked, it may be possible for
3275 			 * another thread to do a truncate(), lseek(), write()
3276 			 * creating a hole at startoff between the above
3277 			 * VOP_IOCTL() calls, if the other thread does not do
3278 			 * rangelocking.
3279 			 * If that happens, startoff == endoff and finding
3280 			 * the hole has failed, so set an error.
3281 			 */
3282 			if (error == 0 && startoff == endoff)
3283 				error = EINVAL; /* Any error. Reset to 0. */
3284 		}
3285 		if (error == 0) {
3286 			if (startoff > *inoffp) {
3287 				/* Found hole before data block. */
3288 				xfer = MIN(startoff - *inoffp, len);
3289 				if (*outoffp < va.va_size) {
3290 					/* Must write 0s to punch hole. */
3291 					xfer2 = MIN(va.va_size - *outoffp,
3292 					    xfer);
3293 					memset(dat, 0, MIN(xfer2, blksize));
3294 					error = vn_write_outvp(outvp, dat,
3295 					    *outoffp, xfer2, blksize, false,
3296 					    holeout > 0, outcred);
3297 				}
3298 
3299 				if (error == 0 && *outoffp + xfer >
3300 				    va.va_size && (xfer == len || holetoeof)) {
3301 					/* Grow output file (hole at end). */
3302 					error = vn_write_outvp(outvp, dat,
3303 					    *outoffp, xfer, blksize, true,
3304 					    false, outcred);
3305 				}
3306 				if (error == 0) {
3307 					*inoffp += xfer;
3308 					*outoffp += xfer;
3309 					len -= xfer;
3310 					if (len < savlen)
3311 						interrupted = sig_intr();
3312 				}
3313 			}
3314 			copylen = MIN(len, endoff - startoff);
3315 			cantseek = false;
3316 		} else {
3317 			cantseek = true;
3318 			startoff = *inoffp;
3319 			copylen = len;
3320 			error = 0;
3321 		}
3322 
3323 		xfer = blksize;
3324 		if (cantseek) {
3325 			/*
3326 			 * Set first xfer to end at a block boundary, so that
3327 			 * holes are more likely detected in the loop below via
3328 			 * the for all bytes 0 method.
3329 			 */
3330 			xfer -= (*inoffp % blksize);
3331 		}
3332 		/* Loop copying the data block. */
3333 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3334 			if (copylen < xfer)
3335 				xfer = copylen;
3336 			error = vn_lock(invp, LK_SHARED);
3337 			if (error != 0)
3338 				goto out;
3339 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3340 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3341 			    curthread->td_ucred, incred, &aresid,
3342 			    curthread);
3343 			VOP_UNLOCK(invp);
3344 			lastblock = false;
3345 			if (error == 0 && aresid > 0) {
3346 				/* Stop the copy at EOF on the input file. */
3347 				xfer -= aresid;
3348 				eof = true;
3349 				lastblock = true;
3350 			}
3351 			if (error == 0) {
3352 				/*
3353 				 * Skip the write for holes past the initial EOF
3354 				 * of the output file, unless this is the last
3355 				 * write of the output file at EOF.
3356 				 */
3357 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3358 				    false;
3359 				if (xfer == len)
3360 					lastblock = true;
3361 				if (!cantseek || *outoffp < va.va_size ||
3362 				    lastblock || !readzeros)
3363 					error = vn_write_outvp(outvp, dat,
3364 					    *outoffp, xfer, blksize,
3365 					    readzeros && lastblock &&
3366 					    *outoffp >= va.va_size, false,
3367 					    outcred);
3368 				if (error == 0) {
3369 					*inoffp += xfer;
3370 					startoff += xfer;
3371 					*outoffp += xfer;
3372 					copylen -= xfer;
3373 					len -= xfer;
3374 					if (len < savlen)
3375 						interrupted = sig_intr();
3376 				}
3377 			}
3378 			xfer = blksize;
3379 		}
3380 	}
3381 out:
3382 	*lenp = savlen - len;
3383 	free(dat, M_TEMP);
3384 	return (error);
3385 }
3386 
3387 static int
3388 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3389 {
3390 	struct mount *mp;
3391 	struct vnode *vp;
3392 	off_t olen, ooffset;
3393 	int error;
3394 #ifdef AUDIT
3395 	int audited_vnode1 = 0;
3396 #endif
3397 
3398 	vp = fp->f_vnode;
3399 	if (vp->v_type != VREG)
3400 		return (ENODEV);
3401 
3402 	/* Allocating blocks may take a long time, so iterate. */
3403 	for (;;) {
3404 		olen = len;
3405 		ooffset = offset;
3406 
3407 		bwillwrite();
3408 		mp = NULL;
3409 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3410 		if (error != 0)
3411 			break;
3412 		error = vn_lock(vp, LK_EXCLUSIVE);
3413 		if (error != 0) {
3414 			vn_finished_write(mp);
3415 			break;
3416 		}
3417 #ifdef AUDIT
3418 		if (!audited_vnode1) {
3419 			AUDIT_ARG_VNODE1(vp);
3420 			audited_vnode1 = 1;
3421 		}
3422 #endif
3423 #ifdef MAC
3424 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3425 		if (error == 0)
3426 #endif
3427 			error = VOP_ALLOCATE(vp, &offset, &len);
3428 		VOP_UNLOCK(vp);
3429 		vn_finished_write(mp);
3430 
3431 		if (olen + ooffset != offset + len) {
3432 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3433 			    ooffset, olen, offset, len);
3434 		}
3435 		if (error != 0 || len == 0)
3436 			break;
3437 		KASSERT(olen > len, ("Iteration did not make progress?"));
3438 		maybe_yield();
3439 	}
3440 
3441 	return (error);
3442 }
3443 
3444 static int
3445 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3446     int ioflg, struct ucred *active_cred, struct ucred *file_cred)
3447 {
3448 	struct mount *mp;
3449 	void *rl_cookie;
3450 	off_t off, len;
3451 	int error;
3452 #ifdef AUDIT
3453 	bool audited_vnode1 = false;
3454 #endif
3455 
3456 	rl_cookie = NULL;
3457 	error = 0;
3458 	mp = NULL;
3459 	off = *offset;
3460 	len = *length;
3461 
3462 	if ((ioflg & (IO_NODELOCKED|IO_RANGELOCKED)) == 0)
3463 		rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3464 	while (len > 0 && error == 0) {
3465 		/*
3466 		 * Try to deallocate the longest range in one pass.
3467 		 * In case a pass takes too long to be executed, it returns
3468 		 * partial result. The residue will be proceeded in the next
3469 		 * pass.
3470 		 */
3471 
3472 		if ((ioflg & IO_NODELOCKED) == 0) {
3473 			bwillwrite();
3474 			if ((error = vn_start_write(vp, &mp,
3475 			    V_WAIT | PCATCH)) != 0)
3476 				goto out;
3477 			vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3478 		}
3479 #ifdef AUDIT
3480 		if (!audited_vnode1) {
3481 			AUDIT_ARG_VNODE1(vp);
3482 			audited_vnode1 = true;
3483 		}
3484 #endif
3485 
3486 #ifdef MAC
3487 		if ((ioflg & IO_NOMACCHECK) == 0)
3488 			error = mac_vnode_check_write(active_cred, file_cred,
3489 			    vp);
3490 #endif
3491 		if (error == 0)
3492 			error = VOP_DEALLOCATE(vp, &off, &len, flags,
3493 			    active_cred);
3494 
3495 		if ((ioflg & IO_NODELOCKED) == 0) {
3496 			VOP_UNLOCK(vp);
3497 			if (mp != NULL) {
3498 				vn_finished_write(mp);
3499 				mp = NULL;
3500 			}
3501 		}
3502 	}
3503 out:
3504 	if (rl_cookie != NULL)
3505 		vn_rangelock_unlock(vp, rl_cookie);
3506 	*offset = off;
3507 	*length = len;
3508 	return (error);
3509 }
3510 
3511 int
3512 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3513     int ioflg, struct ucred *active_cred, struct ucred *file_cred)
3514 {
3515 	if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3516 	    flags != 0)
3517 		return (EINVAL);
3518 	if (vp->v_type != VREG)
3519 		return (ENODEV);
3520 
3521 	return (vn_deallocate_impl(vp, offset, length, flags, ioflg,
3522 	    active_cred, file_cred));
3523 }
3524 
3525 static int
3526 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3527     struct ucred *active_cred, struct thread *td)
3528 {
3529 	int error;
3530 	struct vnode *vp;
3531 
3532 	vp = fp->f_vnode;
3533 
3534 	if (cmd != SPACECTL_DEALLOC || *offset < 0 || *length <= 0 ||
3535 	    *length > OFF_MAX - *offset || flags != 0)
3536 		return (EINVAL);
3537 	if (vp->v_type != VREG)
3538 		return (ENODEV);
3539 
3540 	switch (cmd) {
3541 	case SPACECTL_DEALLOC:
3542 		error = vn_deallocate_impl(vp, offset, length, flags, 0,
3543 		    active_cred, fp->f_cred);
3544 		break;
3545 	default:
3546 		panic("vn_fspacectl: unknown cmd %d", cmd);
3547 	}
3548 
3549 	return (error);
3550 }
3551 
3552 static u_long vn_lock_pair_pause_cnt;
3553 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3554     &vn_lock_pair_pause_cnt, 0,
3555     "Count of vn_lock_pair deadlocks");
3556 
3557 u_int vn_lock_pair_pause_max;
3558 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3559     &vn_lock_pair_pause_max, 0,
3560     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3561 
3562 static void
3563 vn_lock_pair_pause(const char *wmesg)
3564 {
3565 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3566 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3567 }
3568 
3569 /*
3570  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3571  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3572  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3573  * can be NULL.
3574  *
3575  * The function returns with both vnodes exclusively locked, and
3576  * guarantees that it does not create lock order reversal with other
3577  * threads during its execution.  Both vnodes could be unlocked
3578  * temporary (and reclaimed).
3579  */
3580 void
3581 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3582     bool vp2_locked)
3583 {
3584 	int error;
3585 
3586 	if (vp1 == NULL && vp2 == NULL)
3587 		return;
3588 	if (vp1 != NULL) {
3589 		if (vp1_locked)
3590 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3591 		else
3592 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3593 	} else {
3594 		vp1_locked = true;
3595 	}
3596 	if (vp2 != NULL) {
3597 		if (vp2_locked)
3598 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3599 		else
3600 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3601 	} else {
3602 		vp2_locked = true;
3603 	}
3604 	if (!vp1_locked && !vp2_locked) {
3605 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3606 		vp1_locked = true;
3607 	}
3608 
3609 	for (;;) {
3610 		if (vp1_locked && vp2_locked)
3611 			break;
3612 		if (vp1_locked && vp2 != NULL) {
3613 			if (vp1 != NULL) {
3614 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3615 				    __FILE__, __LINE__);
3616 				if (error == 0)
3617 					break;
3618 				VOP_UNLOCK(vp1);
3619 				vp1_locked = false;
3620 				vn_lock_pair_pause("vlp1");
3621 			}
3622 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3623 			vp2_locked = true;
3624 		}
3625 		if (vp2_locked && vp1 != NULL) {
3626 			if (vp2 != NULL) {
3627 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3628 				    __FILE__, __LINE__);
3629 				if (error == 0)
3630 					break;
3631 				VOP_UNLOCK(vp2);
3632 				vp2_locked = false;
3633 				vn_lock_pair_pause("vlp2");
3634 			}
3635 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3636 			vp1_locked = true;
3637 		}
3638 	}
3639 	if (vp1 != NULL)
3640 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3641 	if (vp2 != NULL)
3642 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3643 }
3644 
3645 int
3646 vn_lktype_write(struct mount *mp, struct vnode *vp)
3647 {
3648 	if (MNT_SHARED_WRITES(mp) ||
3649 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3650 		return (LK_SHARED);
3651 	return (LK_EXCLUSIVE);
3652 }
3653