xref: /freebsd/sys/kern/vfs_vnops.c (revision 1f88aa09417f1cfb3929fd37531b1ab51213c2d6)
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 ((fmode & FREAD) != 0)
209 		res |= OPENREAD;
210 	if ((fmode & FWRITE) != 0)
211 		res |= OPENWRITE;
212 	if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
213 		res |= AUDITVNODE1;
214 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
215 		res |= NOCAPCHECK;
216 	return (res);
217 }
218 
219 /*
220  * Common code for vnode open operations via a name lookup.
221  * Lookup the vnode and invoke VOP_CREATE if needed.
222  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
223  *
224  * Note that this does NOT free nameidata for the successful case,
225  * due to the NDINIT being done elsewhere.
226  */
227 int
228 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
229     struct ucred *cred, struct file *fp)
230 {
231 	struct vnode *vp;
232 	struct mount *mp;
233 	struct thread *td = ndp->ni_cnd.cn_thread;
234 	struct vattr vat;
235 	struct vattr *vap = &vat;
236 	int fmode, error;
237 	bool first_open;
238 
239 restart:
240 	first_open = false;
241 	fmode = *flagp;
242 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
243 	    O_EXCL | O_DIRECTORY) ||
244 	    (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
245 		return (EINVAL);
246 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
247 		ndp->ni_cnd.cn_nameiop = CREATE;
248 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
249 		/*
250 		 * Set NOCACHE to avoid flushing the cache when
251 		 * rolling in many files at once.
252 		 *
253 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
254 		 * exist despite NOCACHE.
255 		 */
256 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
257 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
258 			ndp->ni_cnd.cn_flags |= FOLLOW;
259 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
260 			bwillwrite();
261 		if ((error = namei(ndp)) != 0)
262 			return (error);
263 		if (ndp->ni_vp == NULL) {
264 			VATTR_NULL(vap);
265 			vap->va_type = VREG;
266 			vap->va_mode = cmode;
267 			if (fmode & O_EXCL)
268 				vap->va_vaflags |= VA_EXCLUSIVE;
269 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
270 				NDFREE(ndp, NDF_ONLY_PNBUF);
271 				vput(ndp->ni_dvp);
272 				if ((error = vn_start_write(NULL, &mp,
273 				    V_XSLEEP | PCATCH)) != 0)
274 					return (error);
275 				NDREINIT(ndp);
276 				goto restart;
277 			}
278 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
279 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
280 #ifdef MAC
281 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
282 			    &ndp->ni_cnd, vap);
283 			if (error == 0)
284 #endif
285 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
286 				    &ndp->ni_cnd, vap);
287 			vp = ndp->ni_vp;
288 			if (error == 0 && (fmode & O_EXCL) != 0 &&
289 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
290 				VI_LOCK(vp);
291 				vp->v_iflag |= VI_FOPENING;
292 				VI_UNLOCK(vp);
293 				first_open = true;
294 			}
295 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
296 			    false);
297 			vn_finished_write(mp);
298 			if (error) {
299 				NDFREE(ndp, NDF_ONLY_PNBUF);
300 				if (error == ERELOOKUP) {
301 					NDREINIT(ndp);
302 					goto restart;
303 				}
304 				return (error);
305 			}
306 			fmode &= ~O_TRUNC;
307 		} else {
308 			if (ndp->ni_dvp == ndp->ni_vp)
309 				vrele(ndp->ni_dvp);
310 			else
311 				vput(ndp->ni_dvp);
312 			ndp->ni_dvp = NULL;
313 			vp = ndp->ni_vp;
314 			if (fmode & O_EXCL) {
315 				error = EEXIST;
316 				goto bad;
317 			}
318 			if (vp->v_type == VDIR) {
319 				error = EISDIR;
320 				goto bad;
321 			}
322 			fmode &= ~O_CREAT;
323 		}
324 	} else {
325 		ndp->ni_cnd.cn_nameiop = LOOKUP;
326 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
327 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
328 		    FOLLOW;
329 		if ((fmode & FWRITE) == 0)
330 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
331 		if ((error = namei(ndp)) != 0)
332 			return (error);
333 		vp = ndp->ni_vp;
334 	}
335 	error = vn_open_vnode(vp, fmode, cred, td, fp);
336 	if (first_open) {
337 		VI_LOCK(vp);
338 		vp->v_iflag &= ~VI_FOPENING;
339 		wakeup(vp);
340 		VI_UNLOCK(vp);
341 	}
342 	if (error)
343 		goto bad;
344 	*flagp = fmode;
345 	return (0);
346 bad:
347 	NDFREE(ndp, NDF_ONLY_PNBUF);
348 	vput(vp);
349 	*flagp = fmode;
350 	ndp->ni_vp = NULL;
351 	return (error);
352 }
353 
354 static int
355 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
356 {
357 	struct flock lf;
358 	int error, lock_flags, type;
359 
360 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
361 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
362 		return (0);
363 	KASSERT(fp != NULL, ("open with flock requires fp"));
364 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
365 		return (EOPNOTSUPP);
366 
367 	lock_flags = VOP_ISLOCKED(vp);
368 	VOP_UNLOCK(vp);
369 
370 	lf.l_whence = SEEK_SET;
371 	lf.l_start = 0;
372 	lf.l_len = 0;
373 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
374 	type = F_FLOCK;
375 	if ((fmode & FNONBLOCK) == 0)
376 		type |= F_WAIT;
377 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
378 		type |= F_FIRSTOPEN;
379 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
380 	if (error == 0)
381 		fp->f_flag |= FHASLOCK;
382 
383 	vn_lock(vp, lock_flags | LK_RETRY);
384 	return (error);
385 }
386 
387 /*
388  * Common code for vnode open operations once a vnode is located.
389  * Check permissions, and call the VOP_OPEN routine.
390  */
391 int
392 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
393     struct thread *td, struct file *fp)
394 {
395 	accmode_t accmode;
396 	int error;
397 
398 	if (vp->v_type == VLNK) {
399 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
400 			return (EMLINK);
401 	}
402 	if (vp->v_type == VSOCK)
403 		return (EOPNOTSUPP);
404 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
405 		return (ENOTDIR);
406 
407 	accmode = 0;
408 	if ((fmode & O_PATH) == 0) {
409 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
410 			if (vp->v_type == VDIR)
411 				return (EISDIR);
412 			accmode |= VWRITE;
413 		}
414 		if ((fmode & FREAD) != 0)
415 			accmode |= VREAD;
416 		if ((fmode & O_APPEND) && (fmode & FWRITE))
417 			accmode |= VAPPEND;
418 #ifdef MAC
419 		if ((fmode & O_CREAT) != 0)
420 			accmode |= VCREAT;
421 #endif
422 	}
423 	if ((fmode & FEXEC) != 0)
424 		accmode |= VEXEC;
425 #ifdef MAC
426 	if ((fmode & O_VERIFY) != 0)
427 		accmode |= VVERIFY;
428 	error = mac_vnode_check_open(cred, vp, accmode);
429 	if (error != 0)
430 		return (error);
431 
432 	accmode &= ~(VCREAT | VVERIFY);
433 #endif
434 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
435 		error = VOP_ACCESS(vp, accmode, cred, td);
436 		if (error != 0)
437 			return (error);
438 	}
439 	if ((fmode & O_PATH) != 0) {
440 		if (vp->v_type == VFIFO)
441 			error = EPIPE;
442 		else
443 			error = VOP_ACCESS(vp, VREAD, cred, td);
444 		if (error == 0)
445 			fp->f_flag |= FKQALLOWED;
446 		return (0);
447 	}
448 
449 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
450 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
451 	error = VOP_OPEN(vp, fmode, cred, td, fp);
452 	if (error != 0)
453 		return (error);
454 
455 	error = vn_open_vnode_advlock(vp, fmode, fp);
456 	if (error == 0 && (fmode & FWRITE) != 0) {
457 		error = VOP_ADD_WRITECOUNT(vp, 1);
458 		if (error == 0) {
459 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
460 			     __func__, vp, vp->v_writecount);
461 		}
462 	}
463 
464 	/*
465 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
466 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
467 	 */
468 	if (error != 0) {
469 		if (fp != NULL) {
470 			/*
471 			 * Arrange the call by having fdrop() to use
472 			 * vn_closefile().  This is to satisfy
473 			 * filesystems like devfs or tmpfs, which
474 			 * override fo_close().
475 			 */
476 			fp->f_flag |= FOPENFAILED;
477 			fp->f_vnode = vp;
478 			if (fp->f_ops == &badfileops) {
479 				fp->f_type = DTYPE_VNODE;
480 				fp->f_ops = &vnops;
481 			}
482 			vref(vp);
483 		} else {
484 			/*
485 			 * If there is no fp, due to kernel-mode open,
486 			 * we can call VOP_CLOSE() now.
487 			 */
488 			if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
489 			    !MNT_EXTENDED_SHARED(vp->v_mount) &&
490 			    VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
491 				vn_lock(vp, LK_UPGRADE | LK_RETRY);
492 			(void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
493 			    cred, td);
494 		}
495 	}
496 
497 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
498 	return (error);
499 
500 }
501 
502 /*
503  * Check for write permissions on the specified vnode.
504  * Prototype text segments cannot be written.
505  * It is racy.
506  */
507 int
508 vn_writechk(struct vnode *vp)
509 {
510 
511 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
512 	/*
513 	 * If there's shared text associated with
514 	 * the vnode, try to free it up once.  If
515 	 * we fail, we can't allow writing.
516 	 */
517 	if (VOP_IS_TEXT(vp))
518 		return (ETXTBSY);
519 
520 	return (0);
521 }
522 
523 /*
524  * Vnode close call
525  */
526 static int
527 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
528     struct thread *td, bool keep_ref)
529 {
530 	struct mount *mp;
531 	int error, lock_flags;
532 
533 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
534 	    MNT_EXTENDED_SHARED(vp->v_mount))
535 		lock_flags = LK_SHARED;
536 	else
537 		lock_flags = LK_EXCLUSIVE;
538 
539 	vn_start_write(vp, &mp, V_WAIT);
540 	vn_lock(vp, lock_flags | LK_RETRY);
541 	AUDIT_ARG_VNODE1(vp);
542 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
543 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
544 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
545 		    __func__, vp, vp->v_writecount);
546 	}
547 	error = VOP_CLOSE(vp, flags, file_cred, td);
548 	if (keep_ref)
549 		VOP_UNLOCK(vp);
550 	else
551 		vput(vp);
552 	vn_finished_write(mp);
553 	return (error);
554 }
555 
556 int
557 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
558     struct thread *td)
559 {
560 
561 	return (vn_close1(vp, flags, file_cred, td, false));
562 }
563 
564 /*
565  * Heuristic to detect sequential operation.
566  */
567 static int
568 sequential_heuristic(struct uio *uio, struct file *fp)
569 {
570 	enum uio_rw rw;
571 
572 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
573 
574 	rw = uio->uio_rw;
575 	if (fp->f_flag & FRDAHEAD)
576 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
577 
578 	/*
579 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
580 	 * that the first I/O is normally considered to be slightly
581 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
582 	 * unless previous seeks have reduced f_seqcount to 0, in which
583 	 * case offset 0 is not special.
584 	 */
585 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
586 	    uio->uio_offset == fp->f_nextoff[rw]) {
587 		/*
588 		 * f_seqcount is in units of fixed-size blocks so that it
589 		 * depends mainly on the amount of sequential I/O and not
590 		 * much on the number of sequential I/O's.  The fixed size
591 		 * of 16384 is hard-coded here since it is (not quite) just
592 		 * a magic size that works well here.  This size is more
593 		 * closely related to the best I/O size for real disks than
594 		 * to any block size used by software.
595 		 */
596 		if (uio->uio_resid >= IO_SEQMAX * 16384)
597 			fp->f_seqcount[rw] = IO_SEQMAX;
598 		else {
599 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
600 			if (fp->f_seqcount[rw] > IO_SEQMAX)
601 				fp->f_seqcount[rw] = IO_SEQMAX;
602 		}
603 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
604 	}
605 
606 	/* Not sequential.  Quickly draw-down sequentiality. */
607 	if (fp->f_seqcount[rw] > 1)
608 		fp->f_seqcount[rw] = 1;
609 	else
610 		fp->f_seqcount[rw] = 0;
611 	return (0);
612 }
613 
614 /*
615  * Package up an I/O request on a vnode into a uio and do it.
616  */
617 int
618 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
619     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
620     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
621 {
622 	struct uio auio;
623 	struct iovec aiov;
624 	struct mount *mp;
625 	struct ucred *cred;
626 	void *rl_cookie;
627 	struct vn_io_fault_args args;
628 	int error, lock_flags;
629 
630 	if (offset < 0 && vp->v_type != VCHR)
631 		return (EINVAL);
632 	auio.uio_iov = &aiov;
633 	auio.uio_iovcnt = 1;
634 	aiov.iov_base = base;
635 	aiov.iov_len = len;
636 	auio.uio_resid = len;
637 	auio.uio_offset = offset;
638 	auio.uio_segflg = segflg;
639 	auio.uio_rw = rw;
640 	auio.uio_td = td;
641 	error = 0;
642 
643 	if ((ioflg & IO_NODELOCKED) == 0) {
644 		if ((ioflg & IO_RANGELOCKED) == 0) {
645 			if (rw == UIO_READ) {
646 				rl_cookie = vn_rangelock_rlock(vp, offset,
647 				    offset + len);
648 			} else if ((ioflg & IO_APPEND) != 0) {
649 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
650 			} else {
651 				rl_cookie = vn_rangelock_wlock(vp, offset,
652 				    offset + len);
653 			}
654 		} else
655 			rl_cookie = NULL;
656 		mp = NULL;
657 		if (rw == UIO_WRITE) {
658 			if (vp->v_type != VCHR &&
659 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
660 			    != 0)
661 				goto out;
662 			lock_flags = vn_lktype_write(mp, vp);
663 		} else
664 			lock_flags = LK_SHARED;
665 		vn_lock(vp, lock_flags | LK_RETRY);
666 	} else
667 		rl_cookie = NULL;
668 
669 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
670 #ifdef MAC
671 	if ((ioflg & IO_NOMACCHECK) == 0) {
672 		if (rw == UIO_READ)
673 			error = mac_vnode_check_read(active_cred, file_cred,
674 			    vp);
675 		else
676 			error = mac_vnode_check_write(active_cred, file_cred,
677 			    vp);
678 	}
679 #endif
680 	if (error == 0) {
681 		if (file_cred != NULL)
682 			cred = file_cred;
683 		else
684 			cred = active_cred;
685 		if (do_vn_io_fault(vp, &auio)) {
686 			args.kind = VN_IO_FAULT_VOP;
687 			args.cred = cred;
688 			args.flags = ioflg;
689 			args.args.vop_args.vp = vp;
690 			error = vn_io_fault1(vp, &auio, &args, td);
691 		} else if (rw == UIO_READ) {
692 			error = VOP_READ(vp, &auio, ioflg, cred);
693 		} else /* if (rw == UIO_WRITE) */ {
694 			error = VOP_WRITE(vp, &auio, ioflg, cred);
695 		}
696 	}
697 	if (aresid)
698 		*aresid = auio.uio_resid;
699 	else
700 		if (auio.uio_resid && error == 0)
701 			error = EIO;
702 	if ((ioflg & IO_NODELOCKED) == 0) {
703 		VOP_UNLOCK(vp);
704 		if (mp != NULL)
705 			vn_finished_write(mp);
706 	}
707  out:
708 	if (rl_cookie != NULL)
709 		vn_rangelock_unlock(vp, rl_cookie);
710 	return (error);
711 }
712 
713 /*
714  * Package up an I/O request on a vnode into a uio and do it.  The I/O
715  * request is split up into smaller chunks and we try to avoid saturating
716  * the buffer cache while potentially holding a vnode locked, so we
717  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
718  * to give other processes a chance to lock the vnode (either other processes
719  * core'ing the same binary, or unrelated processes scanning the directory).
720  */
721 int
722 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
723     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
724     struct ucred *file_cred, size_t *aresid, struct thread *td)
725 {
726 	int error = 0;
727 	ssize_t iaresid;
728 
729 	do {
730 		int chunk;
731 
732 		/*
733 		 * Force `offset' to a multiple of MAXBSIZE except possibly
734 		 * for the first chunk, so that filesystems only need to
735 		 * write full blocks except possibly for the first and last
736 		 * chunks.
737 		 */
738 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
739 
740 		if (chunk > len)
741 			chunk = len;
742 		if (rw != UIO_READ && vp->v_type == VREG)
743 			bwillwrite();
744 		iaresid = 0;
745 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
746 		    ioflg, active_cred, file_cred, &iaresid, td);
747 		len -= chunk;	/* aresid calc already includes length */
748 		if (error)
749 			break;
750 		offset += chunk;
751 		base = (char *)base + chunk;
752 		kern_yield(PRI_USER);
753 	} while (len);
754 	if (aresid)
755 		*aresid = len + iaresid;
756 	return (error);
757 }
758 
759 #if OFF_MAX <= LONG_MAX
760 off_t
761 foffset_lock(struct file *fp, int flags)
762 {
763 	volatile short *flagsp;
764 	off_t res;
765 	short state;
766 
767 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
768 
769 	if ((flags & FOF_NOLOCK) != 0)
770 		return (atomic_load_long(&fp->f_offset));
771 
772 	/*
773 	 * According to McKusick the vn lock was protecting f_offset here.
774 	 * It is now protected by the FOFFSET_LOCKED flag.
775 	 */
776 	flagsp = &fp->f_vnread_flags;
777 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
778 		return (atomic_load_long(&fp->f_offset));
779 
780 	sleepq_lock(&fp->f_vnread_flags);
781 	state = atomic_load_16(flagsp);
782 	for (;;) {
783 		if ((state & FOFFSET_LOCKED) == 0) {
784 			if (!atomic_fcmpset_acq_16(flagsp, &state,
785 			    FOFFSET_LOCKED))
786 				continue;
787 			break;
788 		}
789 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
790 			if (!atomic_fcmpset_acq_16(flagsp, &state,
791 			    state | FOFFSET_LOCK_WAITING))
792 				continue;
793 		}
794 		DROP_GIANT();
795 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
796 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
797 		PICKUP_GIANT();
798 		sleepq_lock(&fp->f_vnread_flags);
799 		state = atomic_load_16(flagsp);
800 	}
801 	res = atomic_load_long(&fp->f_offset);
802 	sleepq_release(&fp->f_vnread_flags);
803 	return (res);
804 }
805 
806 void
807 foffset_unlock(struct file *fp, off_t val, int flags)
808 {
809 	volatile short *flagsp;
810 	short state;
811 
812 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
813 
814 	if ((flags & FOF_NOUPDATE) == 0)
815 		atomic_store_long(&fp->f_offset, val);
816 	if ((flags & FOF_NEXTOFF_R) != 0)
817 		fp->f_nextoff[UIO_READ] = val;
818 	if ((flags & FOF_NEXTOFF_W) != 0)
819 		fp->f_nextoff[UIO_WRITE] = val;
820 
821 	if ((flags & FOF_NOLOCK) != 0)
822 		return;
823 
824 	flagsp = &fp->f_vnread_flags;
825 	state = atomic_load_16(flagsp);
826 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
827 	    atomic_cmpset_rel_16(flagsp, state, 0))
828 		return;
829 
830 	sleepq_lock(&fp->f_vnread_flags);
831 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
832 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
833 	fp->f_vnread_flags = 0;
834 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
835 	sleepq_release(&fp->f_vnread_flags);
836 }
837 #else
838 off_t
839 foffset_lock(struct file *fp, int flags)
840 {
841 	struct mtx *mtxp;
842 	off_t res;
843 
844 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
845 
846 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
847 	mtx_lock(mtxp);
848 	if ((flags & FOF_NOLOCK) == 0) {
849 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
850 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
851 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
852 			    "vofflock", 0);
853 		}
854 		fp->f_vnread_flags |= FOFFSET_LOCKED;
855 	}
856 	res = fp->f_offset;
857 	mtx_unlock(mtxp);
858 	return (res);
859 }
860 
861 void
862 foffset_unlock(struct file *fp, off_t val, int flags)
863 {
864 	struct mtx *mtxp;
865 
866 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
867 
868 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
869 	mtx_lock(mtxp);
870 	if ((flags & FOF_NOUPDATE) == 0)
871 		fp->f_offset = val;
872 	if ((flags & FOF_NEXTOFF_R) != 0)
873 		fp->f_nextoff[UIO_READ] = val;
874 	if ((flags & FOF_NEXTOFF_W) != 0)
875 		fp->f_nextoff[UIO_WRITE] = val;
876 	if ((flags & FOF_NOLOCK) == 0) {
877 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
878 		    ("Lost FOFFSET_LOCKED"));
879 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
880 			wakeup(&fp->f_vnread_flags);
881 		fp->f_vnread_flags = 0;
882 	}
883 	mtx_unlock(mtxp);
884 }
885 #endif
886 
887 void
888 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
889 {
890 
891 	if ((flags & FOF_OFFSET) == 0)
892 		uio->uio_offset = foffset_lock(fp, flags);
893 }
894 
895 void
896 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
897 {
898 
899 	if ((flags & FOF_OFFSET) == 0)
900 		foffset_unlock(fp, uio->uio_offset, flags);
901 }
902 
903 static int
904 get_advice(struct file *fp, struct uio *uio)
905 {
906 	struct mtx *mtxp;
907 	int ret;
908 
909 	ret = POSIX_FADV_NORMAL;
910 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
911 		return (ret);
912 
913 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
914 	mtx_lock(mtxp);
915 	if (fp->f_advice != NULL &&
916 	    uio->uio_offset >= fp->f_advice->fa_start &&
917 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
918 		ret = fp->f_advice->fa_advice;
919 	mtx_unlock(mtxp);
920 	return (ret);
921 }
922 
923 int
924 vn_read_from_obj(struct vnode *vp, struct uio *uio)
925 {
926 	vm_object_t obj;
927 	vm_page_t ma[io_hold_cnt + 2];
928 	off_t off, vsz;
929 	ssize_t resid;
930 	int error, i, j;
931 
932 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
933 	obj = atomic_load_ptr(&vp->v_object);
934 	if (obj == NULL)
935 		return (EJUSTRETURN);
936 
937 	/*
938 	 * Depends on type stability of vm_objects.
939 	 */
940 	vm_object_pip_add(obj, 1);
941 	if ((obj->flags & OBJ_DEAD) != 0) {
942 		/*
943 		 * Note that object might be already reused from the
944 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
945 		 * we recheck for DOOMED vnode state after all pages
946 		 * are busied, and retract then.
947 		 *
948 		 * But we check for OBJ_DEAD to ensure that we do not
949 		 * busy pages while vm_object_terminate_pages()
950 		 * processes the queue.
951 		 */
952 		error = EJUSTRETURN;
953 		goto out_pip;
954 	}
955 
956 	resid = uio->uio_resid;
957 	off = uio->uio_offset;
958 	for (i = 0; resid > 0; i++) {
959 		MPASS(i < io_hold_cnt + 2);
960 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
961 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
962 		    VM_ALLOC_NOWAIT);
963 		if (ma[i] == NULL)
964 			break;
965 
966 		/*
967 		 * Skip invalid pages.  Valid mask can be partial only
968 		 * at EOF, and we clip later.
969 		 */
970 		if (vm_page_none_valid(ma[i])) {
971 			vm_page_sunbusy(ma[i]);
972 			break;
973 		}
974 
975 		resid -= PAGE_SIZE;
976 		off += PAGE_SIZE;
977 	}
978 	if (i == 0) {
979 		error = EJUSTRETURN;
980 		goto out_pip;
981 	}
982 
983 	/*
984 	 * Check VIRF_DOOMED after we busied our pages.  Since
985 	 * vgonel() terminates the vnode' vm_object, it cannot
986 	 * process past pages busied by us.
987 	 */
988 	if (VN_IS_DOOMED(vp)) {
989 		error = EJUSTRETURN;
990 		goto out;
991 	}
992 
993 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
994 	if (resid > uio->uio_resid)
995 		resid = uio->uio_resid;
996 
997 	/*
998 	 * Unlocked read of vnp_size is safe because truncation cannot
999 	 * pass busied page.  But we load vnp_size into a local
1000 	 * variable so that possible concurrent extension does not
1001 	 * break calculation.
1002 	 */
1003 #if defined(__powerpc__) && !defined(__powerpc64__)
1004 	vsz = obj->un_pager.vnp.vnp_size;
1005 #else
1006 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1007 #endif
1008 	if (uio->uio_offset >= vsz) {
1009 		error = EJUSTRETURN;
1010 		goto out;
1011 	}
1012 	if (uio->uio_offset + resid > vsz)
1013 		resid = vsz - uio->uio_offset;
1014 
1015 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1016 
1017 out:
1018 	for (j = 0; j < i; j++) {
1019 		if (error == 0)
1020 			vm_page_reference(ma[j]);
1021 		vm_page_sunbusy(ma[j]);
1022 	}
1023 out_pip:
1024 	vm_object_pip_wakeup(obj);
1025 	if (error != 0)
1026 		return (error);
1027 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1028 }
1029 
1030 /*
1031  * File table vnode read routine.
1032  */
1033 static int
1034 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1035     struct thread *td)
1036 {
1037 	struct vnode *vp;
1038 	off_t orig_offset;
1039 	int error, ioflag;
1040 	int advice;
1041 
1042 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1043 	    uio->uio_td, td));
1044 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1045 	vp = fp->f_vnode;
1046 	ioflag = 0;
1047 	if (fp->f_flag & FNONBLOCK)
1048 		ioflag |= IO_NDELAY;
1049 	if (fp->f_flag & O_DIRECT)
1050 		ioflag |= IO_DIRECT;
1051 
1052 	/*
1053 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1054 	 * allows us to avoid unneeded work outright.
1055 	 */
1056 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1057 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1058 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1059 		if (error == 0) {
1060 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1061 			return (0);
1062 		}
1063 		if (error != EJUSTRETURN)
1064 			return (error);
1065 	}
1066 
1067 	advice = get_advice(fp, uio);
1068 	vn_lock(vp, LK_SHARED | LK_RETRY);
1069 
1070 	switch (advice) {
1071 	case POSIX_FADV_NORMAL:
1072 	case POSIX_FADV_SEQUENTIAL:
1073 	case POSIX_FADV_NOREUSE:
1074 		ioflag |= sequential_heuristic(uio, fp);
1075 		break;
1076 	case POSIX_FADV_RANDOM:
1077 		/* Disable read-ahead for random I/O. */
1078 		break;
1079 	}
1080 	orig_offset = uio->uio_offset;
1081 
1082 #ifdef MAC
1083 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1084 	if (error == 0)
1085 #endif
1086 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1087 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1088 	VOP_UNLOCK(vp);
1089 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1090 	    orig_offset != uio->uio_offset)
1091 		/*
1092 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1093 		 * for the backing file after a POSIX_FADV_NOREUSE
1094 		 * read(2).
1095 		 */
1096 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1097 		    POSIX_FADV_DONTNEED);
1098 	return (error);
1099 }
1100 
1101 /*
1102  * File table vnode write routine.
1103  */
1104 static int
1105 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1106     struct thread *td)
1107 {
1108 	struct vnode *vp;
1109 	struct mount *mp;
1110 	off_t orig_offset;
1111 	int error, ioflag;
1112 	int advice;
1113 	bool need_finished_write;
1114 
1115 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1116 	    uio->uio_td, td));
1117 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1118 	vp = fp->f_vnode;
1119 	if (vp->v_type == VREG)
1120 		bwillwrite();
1121 	ioflag = IO_UNIT;
1122 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1123 		ioflag |= IO_APPEND;
1124 	if (fp->f_flag & FNONBLOCK)
1125 		ioflag |= IO_NDELAY;
1126 	if (fp->f_flag & O_DIRECT)
1127 		ioflag |= IO_DIRECT;
1128 
1129 	mp = atomic_load_ptr(&vp->v_mount);
1130 	if ((fp->f_flag & O_FSYNC) ||
1131 	    (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS)))
1132 		ioflag |= IO_SYNC;
1133 
1134 	/*
1135 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
1136 	 * implementations that don't understand IO_DATASYNC fall back to full
1137 	 * O_SYNC behavior.
1138 	 */
1139 	if (fp->f_flag & O_DSYNC)
1140 		ioflag |= IO_SYNC | IO_DATASYNC;
1141 	mp = NULL;
1142 	need_finished_write = false;
1143 	if (vp->v_type != VCHR) {
1144 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1145 		if (error != 0)
1146 			goto unlock;
1147 		need_finished_write = true;
1148 	}
1149 
1150 	advice = get_advice(fp, uio);
1151 
1152 	vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1153 	switch (advice) {
1154 	case POSIX_FADV_NORMAL:
1155 	case POSIX_FADV_SEQUENTIAL:
1156 	case POSIX_FADV_NOREUSE:
1157 		ioflag |= sequential_heuristic(uio, fp);
1158 		break;
1159 	case POSIX_FADV_RANDOM:
1160 		/* XXX: Is this correct? */
1161 		break;
1162 	}
1163 	orig_offset = uio->uio_offset;
1164 
1165 #ifdef MAC
1166 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1167 	if (error == 0)
1168 #endif
1169 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1170 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1171 	VOP_UNLOCK(vp);
1172 	if (need_finished_write)
1173 		vn_finished_write(mp);
1174 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1175 	    orig_offset != uio->uio_offset)
1176 		/*
1177 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1178 		 * for the backing file after a POSIX_FADV_NOREUSE
1179 		 * write(2).
1180 		 */
1181 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1182 		    POSIX_FADV_DONTNEED);
1183 unlock:
1184 	return (error);
1185 }
1186 
1187 /*
1188  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1189  * prevent the following deadlock:
1190  *
1191  * Assume that the thread A reads from the vnode vp1 into userspace
1192  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1193  * currently not resident, then system ends up with the call chain
1194  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1195  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1196  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1197  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1198  * backed by the pages of vnode vp1, and some page in buf2 is not
1199  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1200  *
1201  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1202  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1203  * Instead, it first tries to do the whole range i/o with pagefaults
1204  * disabled. If all pages in the i/o buffer are resident and mapped,
1205  * VOP will succeed (ignoring the genuine filesystem errors).
1206  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1207  * i/o in chunks, with all pages in the chunk prefaulted and held
1208  * using vm_fault_quick_hold_pages().
1209  *
1210  * Filesystems using this deadlock avoidance scheme should use the
1211  * array of the held pages from uio, saved in the curthread->td_ma,
1212  * instead of doing uiomove().  A helper function
1213  * vn_io_fault_uiomove() converts uiomove request into
1214  * uiomove_fromphys() over td_ma array.
1215  *
1216  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1217  * make the current i/o request atomic with respect to other i/os and
1218  * truncations.
1219  */
1220 
1221 /*
1222  * Decode vn_io_fault_args and perform the corresponding i/o.
1223  */
1224 static int
1225 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1226     struct thread *td)
1227 {
1228 	int error, save;
1229 
1230 	error = 0;
1231 	save = vm_fault_disable_pagefaults();
1232 	switch (args->kind) {
1233 	case VN_IO_FAULT_FOP:
1234 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1235 		    uio, args->cred, args->flags, td);
1236 		break;
1237 	case VN_IO_FAULT_VOP:
1238 		if (uio->uio_rw == UIO_READ) {
1239 			error = VOP_READ(args->args.vop_args.vp, uio,
1240 			    args->flags, args->cred);
1241 		} else if (uio->uio_rw == UIO_WRITE) {
1242 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1243 			    args->flags, args->cred);
1244 		}
1245 		break;
1246 	default:
1247 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1248 		    args->kind, uio->uio_rw);
1249 	}
1250 	vm_fault_enable_pagefaults(save);
1251 	return (error);
1252 }
1253 
1254 static int
1255 vn_io_fault_touch(char *base, const struct uio *uio)
1256 {
1257 	int r;
1258 
1259 	r = fubyte(base);
1260 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1261 		return (EFAULT);
1262 	return (0);
1263 }
1264 
1265 static int
1266 vn_io_fault_prefault_user(const struct uio *uio)
1267 {
1268 	char *base;
1269 	const struct iovec *iov;
1270 	size_t len;
1271 	ssize_t resid;
1272 	int error, i;
1273 
1274 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1275 	    ("vn_io_fault_prefault userspace"));
1276 
1277 	error = i = 0;
1278 	iov = uio->uio_iov;
1279 	resid = uio->uio_resid;
1280 	base = iov->iov_base;
1281 	len = iov->iov_len;
1282 	while (resid > 0) {
1283 		error = vn_io_fault_touch(base, uio);
1284 		if (error != 0)
1285 			break;
1286 		if (len < PAGE_SIZE) {
1287 			if (len != 0) {
1288 				error = vn_io_fault_touch(base + len - 1, uio);
1289 				if (error != 0)
1290 					break;
1291 				resid -= len;
1292 			}
1293 			if (++i >= uio->uio_iovcnt)
1294 				break;
1295 			iov = uio->uio_iov + i;
1296 			base = iov->iov_base;
1297 			len = iov->iov_len;
1298 		} else {
1299 			len -= PAGE_SIZE;
1300 			base += PAGE_SIZE;
1301 			resid -= PAGE_SIZE;
1302 		}
1303 	}
1304 	return (error);
1305 }
1306 
1307 /*
1308  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1309  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1310  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1311  * into args and call vn_io_fault1() to handle faults during the user
1312  * mode buffer accesses.
1313  */
1314 static int
1315 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1316     struct thread *td)
1317 {
1318 	vm_page_t ma[io_hold_cnt + 2];
1319 	struct uio *uio_clone, short_uio;
1320 	struct iovec short_iovec[1];
1321 	vm_page_t *prev_td_ma;
1322 	vm_prot_t prot;
1323 	vm_offset_t addr, end;
1324 	size_t len, resid;
1325 	ssize_t adv;
1326 	int error, cnt, saveheld, prev_td_ma_cnt;
1327 
1328 	if (vn_io_fault_prefault) {
1329 		error = vn_io_fault_prefault_user(uio);
1330 		if (error != 0)
1331 			return (error); /* Or ignore ? */
1332 	}
1333 
1334 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1335 
1336 	/*
1337 	 * The UFS follows IO_UNIT directive and replays back both
1338 	 * uio_offset and uio_resid if an error is encountered during the
1339 	 * operation.  But, since the iovec may be already advanced,
1340 	 * uio is still in an inconsistent state.
1341 	 *
1342 	 * Cache a copy of the original uio, which is advanced to the redo
1343 	 * point using UIO_NOCOPY below.
1344 	 */
1345 	uio_clone = cloneuio(uio);
1346 	resid = uio->uio_resid;
1347 
1348 	short_uio.uio_segflg = UIO_USERSPACE;
1349 	short_uio.uio_rw = uio->uio_rw;
1350 	short_uio.uio_td = uio->uio_td;
1351 
1352 	error = vn_io_fault_doio(args, uio, td);
1353 	if (error != EFAULT)
1354 		goto out;
1355 
1356 	atomic_add_long(&vn_io_faults_cnt, 1);
1357 	uio_clone->uio_segflg = UIO_NOCOPY;
1358 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1359 	uio_clone->uio_segflg = uio->uio_segflg;
1360 
1361 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1362 	prev_td_ma = td->td_ma;
1363 	prev_td_ma_cnt = td->td_ma_cnt;
1364 
1365 	while (uio_clone->uio_resid != 0) {
1366 		len = uio_clone->uio_iov->iov_len;
1367 		if (len == 0) {
1368 			KASSERT(uio_clone->uio_iovcnt >= 1,
1369 			    ("iovcnt underflow"));
1370 			uio_clone->uio_iov++;
1371 			uio_clone->uio_iovcnt--;
1372 			continue;
1373 		}
1374 		if (len > ptoa(io_hold_cnt))
1375 			len = ptoa(io_hold_cnt);
1376 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1377 		end = round_page(addr + len);
1378 		if (end < addr) {
1379 			error = EFAULT;
1380 			break;
1381 		}
1382 		cnt = atop(end - trunc_page(addr));
1383 		/*
1384 		 * A perfectly misaligned address and length could cause
1385 		 * both the start and the end of the chunk to use partial
1386 		 * page.  +2 accounts for such a situation.
1387 		 */
1388 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1389 		    addr, len, prot, ma, io_hold_cnt + 2);
1390 		if (cnt == -1) {
1391 			error = EFAULT;
1392 			break;
1393 		}
1394 		short_uio.uio_iov = &short_iovec[0];
1395 		short_iovec[0].iov_base = (void *)addr;
1396 		short_uio.uio_iovcnt = 1;
1397 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1398 		short_uio.uio_offset = uio_clone->uio_offset;
1399 		td->td_ma = ma;
1400 		td->td_ma_cnt = cnt;
1401 
1402 		error = vn_io_fault_doio(args, &short_uio, td);
1403 		vm_page_unhold_pages(ma, cnt);
1404 		adv = len - short_uio.uio_resid;
1405 
1406 		uio_clone->uio_iov->iov_base =
1407 		    (char *)uio_clone->uio_iov->iov_base + adv;
1408 		uio_clone->uio_iov->iov_len -= adv;
1409 		uio_clone->uio_resid -= adv;
1410 		uio_clone->uio_offset += adv;
1411 
1412 		uio->uio_resid -= adv;
1413 		uio->uio_offset += adv;
1414 
1415 		if (error != 0 || adv == 0)
1416 			break;
1417 	}
1418 	td->td_ma = prev_td_ma;
1419 	td->td_ma_cnt = prev_td_ma_cnt;
1420 	curthread_pflags_restore(saveheld);
1421 out:
1422 	free(uio_clone, M_IOV);
1423 	return (error);
1424 }
1425 
1426 static int
1427 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1428     int flags, struct thread *td)
1429 {
1430 	fo_rdwr_t *doio;
1431 	struct vnode *vp;
1432 	void *rl_cookie;
1433 	struct vn_io_fault_args args;
1434 	int error;
1435 
1436 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1437 	vp = fp->f_vnode;
1438 
1439 	/*
1440 	 * The ability to read(2) on a directory has historically been
1441 	 * allowed for all users, but this can and has been the source of
1442 	 * at least one security issue in the past.  As such, it is now hidden
1443 	 * away behind a sysctl for those that actually need it to use it, and
1444 	 * restricted to root when it's turned on to make it relatively safe to
1445 	 * leave on for longer sessions of need.
1446 	 */
1447 	if (vp->v_type == VDIR) {
1448 		KASSERT(uio->uio_rw == UIO_READ,
1449 		    ("illegal write attempted on a directory"));
1450 		if (!vfs_allow_read_dir)
1451 			return (EISDIR);
1452 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1453 			return (EISDIR);
1454 	}
1455 
1456 	foffset_lock_uio(fp, uio, flags);
1457 	if (do_vn_io_fault(vp, uio)) {
1458 		args.kind = VN_IO_FAULT_FOP;
1459 		args.args.fop_args.fp = fp;
1460 		args.args.fop_args.doio = doio;
1461 		args.cred = active_cred;
1462 		args.flags = flags | FOF_OFFSET;
1463 		if (uio->uio_rw == UIO_READ) {
1464 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1465 			    uio->uio_offset + uio->uio_resid);
1466 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1467 		    (flags & FOF_OFFSET) == 0) {
1468 			/* For appenders, punt and lock the whole range. */
1469 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1470 		} else {
1471 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1472 			    uio->uio_offset + uio->uio_resid);
1473 		}
1474 		error = vn_io_fault1(vp, uio, &args, td);
1475 		vn_rangelock_unlock(vp, rl_cookie);
1476 	} else {
1477 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1478 	}
1479 	foffset_unlock_uio(fp, uio, flags);
1480 	return (error);
1481 }
1482 
1483 /*
1484  * Helper function to perform the requested uiomove operation using
1485  * the held pages for io->uio_iov[0].iov_base buffer instead of
1486  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1487  * instead of iov_base prevents page faults that could occur due to
1488  * pmap_collect() invalidating the mapping created by
1489  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1490  * object cleanup revoking the write access from page mappings.
1491  *
1492  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1493  * instead of plain uiomove().
1494  */
1495 int
1496 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1497 {
1498 	struct uio transp_uio;
1499 	struct iovec transp_iov[1];
1500 	struct thread *td;
1501 	size_t adv;
1502 	int error, pgadv;
1503 
1504 	td = curthread;
1505 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1506 	    uio->uio_segflg != UIO_USERSPACE)
1507 		return (uiomove(data, xfersize, uio));
1508 
1509 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1510 	transp_iov[0].iov_base = data;
1511 	transp_uio.uio_iov = &transp_iov[0];
1512 	transp_uio.uio_iovcnt = 1;
1513 	if (xfersize > uio->uio_resid)
1514 		xfersize = uio->uio_resid;
1515 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1516 	transp_uio.uio_offset = 0;
1517 	transp_uio.uio_segflg = UIO_SYSSPACE;
1518 	/*
1519 	 * Since transp_iov points to data, and td_ma page array
1520 	 * corresponds to original uio->uio_iov, we need to invert the
1521 	 * direction of the i/o operation as passed to
1522 	 * uiomove_fromphys().
1523 	 */
1524 	switch (uio->uio_rw) {
1525 	case UIO_WRITE:
1526 		transp_uio.uio_rw = UIO_READ;
1527 		break;
1528 	case UIO_READ:
1529 		transp_uio.uio_rw = UIO_WRITE;
1530 		break;
1531 	}
1532 	transp_uio.uio_td = uio->uio_td;
1533 	error = uiomove_fromphys(td->td_ma,
1534 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1535 	    xfersize, &transp_uio);
1536 	adv = xfersize - transp_uio.uio_resid;
1537 	pgadv =
1538 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1539 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1540 	td->td_ma += pgadv;
1541 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1542 	    pgadv));
1543 	td->td_ma_cnt -= pgadv;
1544 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1545 	uio->uio_iov->iov_len -= adv;
1546 	uio->uio_resid -= adv;
1547 	uio->uio_offset += adv;
1548 	return (error);
1549 }
1550 
1551 int
1552 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1553     struct uio *uio)
1554 {
1555 	struct thread *td;
1556 	vm_offset_t iov_base;
1557 	int cnt, pgadv;
1558 
1559 	td = curthread;
1560 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1561 	    uio->uio_segflg != UIO_USERSPACE)
1562 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1563 
1564 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1565 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1566 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1567 	switch (uio->uio_rw) {
1568 	case UIO_WRITE:
1569 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1570 		    offset, cnt);
1571 		break;
1572 	case UIO_READ:
1573 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1574 		    cnt);
1575 		break;
1576 	}
1577 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1578 	td->td_ma += pgadv;
1579 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1580 	    pgadv));
1581 	td->td_ma_cnt -= pgadv;
1582 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1583 	uio->uio_iov->iov_len -= cnt;
1584 	uio->uio_resid -= cnt;
1585 	uio->uio_offset += cnt;
1586 	return (0);
1587 }
1588 
1589 /*
1590  * File table truncate routine.
1591  */
1592 static int
1593 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1594     struct thread *td)
1595 {
1596 	struct mount *mp;
1597 	struct vnode *vp;
1598 	void *rl_cookie;
1599 	int error;
1600 
1601 	vp = fp->f_vnode;
1602 
1603 retry:
1604 	/*
1605 	 * Lock the whole range for truncation.  Otherwise split i/o
1606 	 * might happen partly before and partly after the truncation.
1607 	 */
1608 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1609 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1610 	if (error)
1611 		goto out1;
1612 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1613 	AUDIT_ARG_VNODE1(vp);
1614 	if (vp->v_type == VDIR) {
1615 		error = EISDIR;
1616 		goto out;
1617 	}
1618 #ifdef MAC
1619 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1620 	if (error)
1621 		goto out;
1622 #endif
1623 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1624 	    fp->f_cred);
1625 out:
1626 	VOP_UNLOCK(vp);
1627 	vn_finished_write(mp);
1628 out1:
1629 	vn_rangelock_unlock(vp, rl_cookie);
1630 	if (error == ERELOOKUP)
1631 		goto retry;
1632 	return (error);
1633 }
1634 
1635 /*
1636  * Truncate a file that is already locked.
1637  */
1638 int
1639 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1640     struct ucred *cred)
1641 {
1642 	struct vattr vattr;
1643 	int error;
1644 
1645 	error = VOP_ADD_WRITECOUNT(vp, 1);
1646 	if (error == 0) {
1647 		VATTR_NULL(&vattr);
1648 		vattr.va_size = length;
1649 		if (sync)
1650 			vattr.va_vaflags |= VA_SYNC;
1651 		error = VOP_SETATTR(vp, &vattr, cred);
1652 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1653 	}
1654 	return (error);
1655 }
1656 
1657 /*
1658  * File table vnode stat routine.
1659  */
1660 int
1661 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1662     struct thread *td)
1663 {
1664 	struct vnode *vp = fp->f_vnode;
1665 	int error;
1666 
1667 	vn_lock(vp, LK_SHARED | LK_RETRY);
1668 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1669 	VOP_UNLOCK(vp);
1670 
1671 	return (error);
1672 }
1673 
1674 /*
1675  * File table vnode ioctl routine.
1676  */
1677 static int
1678 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1679     struct thread *td)
1680 {
1681 	struct vattr vattr;
1682 	struct vnode *vp;
1683 	struct fiobmap2_arg *bmarg;
1684 	int error;
1685 
1686 	vp = fp->f_vnode;
1687 	switch (vp->v_type) {
1688 	case VDIR:
1689 	case VREG:
1690 		switch (com) {
1691 		case FIONREAD:
1692 			vn_lock(vp, LK_SHARED | LK_RETRY);
1693 			error = VOP_GETATTR(vp, &vattr, active_cred);
1694 			VOP_UNLOCK(vp);
1695 			if (error == 0)
1696 				*(int *)data = vattr.va_size - fp->f_offset;
1697 			return (error);
1698 		case FIOBMAP2:
1699 			bmarg = (struct fiobmap2_arg *)data;
1700 			vn_lock(vp, LK_SHARED | LK_RETRY);
1701 #ifdef MAC
1702 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1703 			    vp);
1704 			if (error == 0)
1705 #endif
1706 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1707 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1708 			VOP_UNLOCK(vp);
1709 			return (error);
1710 		case FIONBIO:
1711 		case FIOASYNC:
1712 			return (0);
1713 		default:
1714 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1715 			    active_cred, td));
1716 		}
1717 		break;
1718 	case VCHR:
1719 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1720 		    active_cred, td));
1721 	default:
1722 		return (ENOTTY);
1723 	}
1724 }
1725 
1726 /*
1727  * File table vnode poll routine.
1728  */
1729 static int
1730 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1731     struct thread *td)
1732 {
1733 	struct vnode *vp;
1734 	int error;
1735 
1736 	vp = fp->f_vnode;
1737 #if defined(MAC) || defined(AUDIT)
1738 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1739 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1740 		AUDIT_ARG_VNODE1(vp);
1741 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1742 		VOP_UNLOCK(vp);
1743 		if (error != 0)
1744 			return (error);
1745 	}
1746 #endif
1747 	error = VOP_POLL(vp, events, fp->f_cred, td);
1748 	return (error);
1749 }
1750 
1751 /*
1752  * Acquire the requested lock and then check for validity.  LK_RETRY
1753  * permits vn_lock to return doomed vnodes.
1754  */
1755 static int __noinline
1756 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1757     int error)
1758 {
1759 
1760 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1761 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1762 
1763 	if (error == 0)
1764 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1765 
1766 	if ((flags & LK_RETRY) == 0) {
1767 		if (error == 0) {
1768 			VOP_UNLOCK(vp);
1769 			error = ENOENT;
1770 		}
1771 		return (error);
1772 	}
1773 
1774 	/*
1775 	 * LK_RETRY case.
1776 	 *
1777 	 * Nothing to do if we got the lock.
1778 	 */
1779 	if (error == 0)
1780 		return (0);
1781 
1782 	/*
1783 	 * Interlock was dropped by the call in _vn_lock.
1784 	 */
1785 	flags &= ~LK_INTERLOCK;
1786 	do {
1787 		error = VOP_LOCK1(vp, flags, file, line);
1788 	} while (error != 0);
1789 	return (0);
1790 }
1791 
1792 int
1793 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1794 {
1795 	int error;
1796 
1797 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1798 	    ("vn_lock: no locktype (%d passed)", flags));
1799 	VNPASS(vp->v_holdcnt > 0, vp);
1800 	error = VOP_LOCK1(vp, flags, file, line);
1801 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1802 		return (_vn_lock_fallback(vp, flags, file, line, error));
1803 	return (0);
1804 }
1805 
1806 /*
1807  * File table vnode close routine.
1808  */
1809 static int
1810 vn_closefile(struct file *fp, struct thread *td)
1811 {
1812 	struct vnode *vp;
1813 	struct flock lf;
1814 	int error;
1815 	bool ref;
1816 
1817 	vp = fp->f_vnode;
1818 	fp->f_ops = &badfileops;
1819 	ref = (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1820 
1821 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1822 
1823 	if (__predict_false(ref)) {
1824 		lf.l_whence = SEEK_SET;
1825 		lf.l_start = 0;
1826 		lf.l_len = 0;
1827 		lf.l_type = F_UNLCK;
1828 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1829 		vrele(vp);
1830 	}
1831 	return (error);
1832 }
1833 
1834 /*
1835  * Preparing to start a filesystem write operation. If the operation is
1836  * permitted, then we bump the count of operations in progress and
1837  * proceed. If a suspend request is in progress, we wait until the
1838  * suspension is over, and then proceed.
1839  */
1840 static int
1841 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1842 {
1843 	struct mount_pcpu *mpcpu;
1844 	int error, mflags;
1845 
1846 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1847 	    vfs_op_thread_enter(mp, mpcpu)) {
1848 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1849 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1850 		vfs_op_thread_exit(mp, mpcpu);
1851 		return (0);
1852 	}
1853 
1854 	if (mplocked)
1855 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1856 	else
1857 		MNT_ILOCK(mp);
1858 
1859 	error = 0;
1860 
1861 	/*
1862 	 * Check on status of suspension.
1863 	 */
1864 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1865 	    mp->mnt_susp_owner != curthread) {
1866 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1867 		    (flags & PCATCH) : 0) | (PUSER - 1);
1868 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1869 			if (flags & V_NOWAIT) {
1870 				error = EWOULDBLOCK;
1871 				goto unlock;
1872 			}
1873 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1874 			    "suspfs", 0);
1875 			if (error)
1876 				goto unlock;
1877 		}
1878 	}
1879 	if (flags & V_XSLEEP)
1880 		goto unlock;
1881 	mp->mnt_writeopcount++;
1882 unlock:
1883 	if (error != 0 || (flags & V_XSLEEP) != 0)
1884 		MNT_REL(mp);
1885 	MNT_IUNLOCK(mp);
1886 	return (error);
1887 }
1888 
1889 int
1890 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1891 {
1892 	struct mount *mp;
1893 	int error;
1894 
1895 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1896 	    ("V_MNTREF requires mp"));
1897 
1898 	error = 0;
1899 	/*
1900 	 * If a vnode is provided, get and return the mount point that
1901 	 * to which it will write.
1902 	 */
1903 	if (vp != NULL) {
1904 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1905 			*mpp = NULL;
1906 			if (error != EOPNOTSUPP)
1907 				return (error);
1908 			return (0);
1909 		}
1910 	}
1911 	if ((mp = *mpp) == NULL)
1912 		return (0);
1913 
1914 	/*
1915 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1916 	 * a vfs_ref().
1917 	 * As long as a vnode is not provided we need to acquire a
1918 	 * refcount for the provided mountpoint too, in order to
1919 	 * emulate a vfs_ref().
1920 	 */
1921 	if (vp == NULL && (flags & V_MNTREF) == 0)
1922 		vfs_ref(mp);
1923 
1924 	return (vn_start_write_refed(mp, flags, false));
1925 }
1926 
1927 /*
1928  * Secondary suspension. Used by operations such as vop_inactive
1929  * routines that are needed by the higher level functions. These
1930  * are allowed to proceed until all the higher level functions have
1931  * completed (indicated by mnt_writeopcount dropping to zero). At that
1932  * time, these operations are halted until the suspension is over.
1933  */
1934 int
1935 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1936 {
1937 	struct mount *mp;
1938 	int error;
1939 
1940 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1941 	    ("V_MNTREF requires mp"));
1942 
1943  retry:
1944 	if (vp != NULL) {
1945 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1946 			*mpp = NULL;
1947 			if (error != EOPNOTSUPP)
1948 				return (error);
1949 			return (0);
1950 		}
1951 	}
1952 	/*
1953 	 * If we are not suspended or have not yet reached suspended
1954 	 * mode, then let the operation proceed.
1955 	 */
1956 	if ((mp = *mpp) == NULL)
1957 		return (0);
1958 
1959 	/*
1960 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1961 	 * a vfs_ref().
1962 	 * As long as a vnode is not provided we need to acquire a
1963 	 * refcount for the provided mountpoint too, in order to
1964 	 * emulate a vfs_ref().
1965 	 */
1966 	MNT_ILOCK(mp);
1967 	if (vp == NULL && (flags & V_MNTREF) == 0)
1968 		MNT_REF(mp);
1969 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1970 		mp->mnt_secondary_writes++;
1971 		mp->mnt_secondary_accwrites++;
1972 		MNT_IUNLOCK(mp);
1973 		return (0);
1974 	}
1975 	if (flags & V_NOWAIT) {
1976 		MNT_REL(mp);
1977 		MNT_IUNLOCK(mp);
1978 		return (EWOULDBLOCK);
1979 	}
1980 	/*
1981 	 * Wait for the suspension to finish.
1982 	 */
1983 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1984 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1985 	    "suspfs", 0);
1986 	vfs_rel(mp);
1987 	if (error == 0)
1988 		goto retry;
1989 	return (error);
1990 }
1991 
1992 /*
1993  * Filesystem write operation has completed. If we are suspending and this
1994  * operation is the last one, notify the suspender that the suspension is
1995  * now in effect.
1996  */
1997 void
1998 vn_finished_write(struct mount *mp)
1999 {
2000 	struct mount_pcpu *mpcpu;
2001 	int c;
2002 
2003 	if (mp == NULL)
2004 		return;
2005 
2006 	if (vfs_op_thread_enter(mp, mpcpu)) {
2007 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2008 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2009 		vfs_op_thread_exit(mp, mpcpu);
2010 		return;
2011 	}
2012 
2013 	MNT_ILOCK(mp);
2014 	vfs_assert_mount_counters(mp);
2015 	MNT_REL(mp);
2016 	c = --mp->mnt_writeopcount;
2017 	if (mp->mnt_vfs_ops == 0) {
2018 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2019 		MNT_IUNLOCK(mp);
2020 		return;
2021 	}
2022 	if (c < 0)
2023 		vfs_dump_mount_counters(mp);
2024 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2025 		wakeup(&mp->mnt_writeopcount);
2026 	MNT_IUNLOCK(mp);
2027 }
2028 
2029 /*
2030  * Filesystem secondary write operation has completed. If we are
2031  * suspending and this operation is the last one, notify the suspender
2032  * that the suspension is now in effect.
2033  */
2034 void
2035 vn_finished_secondary_write(struct mount *mp)
2036 {
2037 	if (mp == NULL)
2038 		return;
2039 	MNT_ILOCK(mp);
2040 	MNT_REL(mp);
2041 	mp->mnt_secondary_writes--;
2042 	if (mp->mnt_secondary_writes < 0)
2043 		panic("vn_finished_secondary_write: neg cnt");
2044 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2045 	    mp->mnt_secondary_writes <= 0)
2046 		wakeup(&mp->mnt_secondary_writes);
2047 	MNT_IUNLOCK(mp);
2048 }
2049 
2050 /*
2051  * Request a filesystem to suspend write operations.
2052  */
2053 int
2054 vfs_write_suspend(struct mount *mp, int flags)
2055 {
2056 	int error;
2057 
2058 	vfs_op_enter(mp);
2059 
2060 	MNT_ILOCK(mp);
2061 	vfs_assert_mount_counters(mp);
2062 	if (mp->mnt_susp_owner == curthread) {
2063 		vfs_op_exit_locked(mp);
2064 		MNT_IUNLOCK(mp);
2065 		return (EALREADY);
2066 	}
2067 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2068 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2069 
2070 	/*
2071 	 * Unmount holds a write reference on the mount point.  If we
2072 	 * own busy reference and drain for writers, we deadlock with
2073 	 * the reference draining in the unmount path.  Callers of
2074 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2075 	 * vfs_busy() reference is owned and caller is not in the
2076 	 * unmount context.
2077 	 */
2078 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2079 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2080 		vfs_op_exit_locked(mp);
2081 		MNT_IUNLOCK(mp);
2082 		return (EBUSY);
2083 	}
2084 
2085 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2086 	mp->mnt_susp_owner = curthread;
2087 	if (mp->mnt_writeopcount > 0)
2088 		(void) msleep(&mp->mnt_writeopcount,
2089 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2090 	else
2091 		MNT_IUNLOCK(mp);
2092 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2093 		vfs_write_resume(mp, 0);
2094 		/* vfs_write_resume does vfs_op_exit() for us */
2095 	}
2096 	return (error);
2097 }
2098 
2099 /*
2100  * Request a filesystem to resume write operations.
2101  */
2102 void
2103 vfs_write_resume(struct mount *mp, int flags)
2104 {
2105 
2106 	MNT_ILOCK(mp);
2107 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2108 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2109 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2110 				       MNTK_SUSPENDED);
2111 		mp->mnt_susp_owner = NULL;
2112 		wakeup(&mp->mnt_writeopcount);
2113 		wakeup(&mp->mnt_flag);
2114 		curthread->td_pflags &= ~TDP_IGNSUSP;
2115 		if ((flags & VR_START_WRITE) != 0) {
2116 			MNT_REF(mp);
2117 			mp->mnt_writeopcount++;
2118 		}
2119 		MNT_IUNLOCK(mp);
2120 		if ((flags & VR_NO_SUSPCLR) == 0)
2121 			VFS_SUSP_CLEAN(mp);
2122 		vfs_op_exit(mp);
2123 	} else if ((flags & VR_START_WRITE) != 0) {
2124 		MNT_REF(mp);
2125 		vn_start_write_refed(mp, 0, true);
2126 	} else {
2127 		MNT_IUNLOCK(mp);
2128 	}
2129 }
2130 
2131 /*
2132  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2133  * methods.
2134  */
2135 int
2136 vfs_write_suspend_umnt(struct mount *mp)
2137 {
2138 	int error;
2139 
2140 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2141 	    ("vfs_write_suspend_umnt: recursed"));
2142 
2143 	/* dounmount() already called vn_start_write(). */
2144 	for (;;) {
2145 		vn_finished_write(mp);
2146 		error = vfs_write_suspend(mp, 0);
2147 		if (error != 0) {
2148 			vn_start_write(NULL, &mp, V_WAIT);
2149 			return (error);
2150 		}
2151 		MNT_ILOCK(mp);
2152 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2153 			break;
2154 		MNT_IUNLOCK(mp);
2155 		vn_start_write(NULL, &mp, V_WAIT);
2156 	}
2157 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2158 	wakeup(&mp->mnt_flag);
2159 	MNT_IUNLOCK(mp);
2160 	curthread->td_pflags |= TDP_IGNSUSP;
2161 	return (0);
2162 }
2163 
2164 /*
2165  * Implement kqueues for files by translating it to vnode operation.
2166  */
2167 static int
2168 vn_kqfilter(struct file *fp, struct knote *kn)
2169 {
2170 
2171 	return (VOP_KQFILTER(fp->f_vnode, kn));
2172 }
2173 
2174 int
2175 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2176 {
2177 	if ((fp->f_flag & FKQALLOWED) == 0)
2178 		return (EBADF);
2179 	return (vn_kqfilter(fp, kn));
2180 }
2181 
2182 /*
2183  * Simplified in-kernel wrapper calls for extended attribute access.
2184  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2185  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2186  */
2187 int
2188 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2189     const char *attrname, int *buflen, char *buf, struct thread *td)
2190 {
2191 	struct uio	auio;
2192 	struct iovec	iov;
2193 	int	error;
2194 
2195 	iov.iov_len = *buflen;
2196 	iov.iov_base = buf;
2197 
2198 	auio.uio_iov = &iov;
2199 	auio.uio_iovcnt = 1;
2200 	auio.uio_rw = UIO_READ;
2201 	auio.uio_segflg = UIO_SYSSPACE;
2202 	auio.uio_td = td;
2203 	auio.uio_offset = 0;
2204 	auio.uio_resid = *buflen;
2205 
2206 	if ((ioflg & IO_NODELOCKED) == 0)
2207 		vn_lock(vp, LK_SHARED | LK_RETRY);
2208 
2209 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2210 
2211 	/* authorize attribute retrieval as kernel */
2212 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2213 	    td);
2214 
2215 	if ((ioflg & IO_NODELOCKED) == 0)
2216 		VOP_UNLOCK(vp);
2217 
2218 	if (error == 0) {
2219 		*buflen = *buflen - auio.uio_resid;
2220 	}
2221 
2222 	return (error);
2223 }
2224 
2225 /*
2226  * XXX failure mode if partially written?
2227  */
2228 int
2229 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2230     const char *attrname, int buflen, char *buf, struct thread *td)
2231 {
2232 	struct uio	auio;
2233 	struct iovec	iov;
2234 	struct mount	*mp;
2235 	int	error;
2236 
2237 	iov.iov_len = buflen;
2238 	iov.iov_base = buf;
2239 
2240 	auio.uio_iov = &iov;
2241 	auio.uio_iovcnt = 1;
2242 	auio.uio_rw = UIO_WRITE;
2243 	auio.uio_segflg = UIO_SYSSPACE;
2244 	auio.uio_td = td;
2245 	auio.uio_offset = 0;
2246 	auio.uio_resid = buflen;
2247 
2248 	if ((ioflg & IO_NODELOCKED) == 0) {
2249 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2250 			return (error);
2251 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2252 	}
2253 
2254 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2255 
2256 	/* authorize attribute setting as kernel */
2257 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2258 
2259 	if ((ioflg & IO_NODELOCKED) == 0) {
2260 		vn_finished_write(mp);
2261 		VOP_UNLOCK(vp);
2262 	}
2263 
2264 	return (error);
2265 }
2266 
2267 int
2268 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2269     const char *attrname, struct thread *td)
2270 {
2271 	struct mount	*mp;
2272 	int	error;
2273 
2274 	if ((ioflg & IO_NODELOCKED) == 0) {
2275 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2276 			return (error);
2277 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2278 	}
2279 
2280 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2281 
2282 	/* authorize attribute removal as kernel */
2283 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2284 	if (error == EOPNOTSUPP)
2285 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2286 		    NULL, td);
2287 
2288 	if ((ioflg & IO_NODELOCKED) == 0) {
2289 		vn_finished_write(mp);
2290 		VOP_UNLOCK(vp);
2291 	}
2292 
2293 	return (error);
2294 }
2295 
2296 static int
2297 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2298     struct vnode **rvp)
2299 {
2300 
2301 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2302 }
2303 
2304 int
2305 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2306 {
2307 
2308 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2309 	    lkflags, rvp));
2310 }
2311 
2312 int
2313 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2314     int lkflags, struct vnode **rvp)
2315 {
2316 	struct mount *mp;
2317 	int ltype, error;
2318 
2319 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2320 	mp = vp->v_mount;
2321 	ltype = VOP_ISLOCKED(vp);
2322 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2323 	    ("vn_vget_ino: vp not locked"));
2324 	error = vfs_busy(mp, MBF_NOWAIT);
2325 	if (error != 0) {
2326 		vfs_ref(mp);
2327 		VOP_UNLOCK(vp);
2328 		error = vfs_busy(mp, 0);
2329 		vn_lock(vp, ltype | LK_RETRY);
2330 		vfs_rel(mp);
2331 		if (error != 0)
2332 			return (ENOENT);
2333 		if (VN_IS_DOOMED(vp)) {
2334 			vfs_unbusy(mp);
2335 			return (ENOENT);
2336 		}
2337 	}
2338 	VOP_UNLOCK(vp);
2339 	error = alloc(mp, alloc_arg, lkflags, rvp);
2340 	vfs_unbusy(mp);
2341 	if (error != 0 || *rvp != vp)
2342 		vn_lock(vp, ltype | LK_RETRY);
2343 	if (VN_IS_DOOMED(vp)) {
2344 		if (error == 0) {
2345 			if (*rvp == vp)
2346 				vunref(vp);
2347 			else
2348 				vput(*rvp);
2349 		}
2350 		error = ENOENT;
2351 	}
2352 	return (error);
2353 }
2354 
2355 int
2356 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2357     struct thread *td)
2358 {
2359 	off_t lim;
2360 	bool ktr_write;
2361 
2362 	if (td == NULL)
2363 		return (0);
2364 
2365 	/*
2366 	 * There are conditions where the limit is to be ignored.
2367 	 * However, since it is almost never reached, check it first.
2368 	 */
2369 	ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2370 	lim = lim_cur(td, RLIMIT_FSIZE);
2371 	if (__predict_false(ktr_write))
2372 		lim = td->td_ktr_io_lim;
2373 	if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2374 		return (0);
2375 
2376 	/*
2377 	 * The limit is reached.
2378 	 */
2379 	if (vp->v_type != VREG ||
2380 	    (td->td_pflags2 & TDP2_ACCT) != 0)
2381 		return (0);
2382 
2383 	if (!ktr_write || ktr_filesize_limit_signal) {
2384 		PROC_LOCK(td->td_proc);
2385 		kern_psignal(td->td_proc, SIGXFSZ);
2386 		PROC_UNLOCK(td->td_proc);
2387 	}
2388 	return (EFBIG);
2389 }
2390 
2391 int
2392 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2393     struct thread *td)
2394 {
2395 	struct vnode *vp;
2396 
2397 	vp = fp->f_vnode;
2398 #ifdef AUDIT
2399 	vn_lock(vp, LK_SHARED | LK_RETRY);
2400 	AUDIT_ARG_VNODE1(vp);
2401 	VOP_UNLOCK(vp);
2402 #endif
2403 	return (setfmode(td, active_cred, vp, mode));
2404 }
2405 
2406 int
2407 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2408     struct thread *td)
2409 {
2410 	struct vnode *vp;
2411 
2412 	vp = fp->f_vnode;
2413 #ifdef AUDIT
2414 	vn_lock(vp, LK_SHARED | LK_RETRY);
2415 	AUDIT_ARG_VNODE1(vp);
2416 	VOP_UNLOCK(vp);
2417 #endif
2418 	return (setfown(td, active_cred, vp, uid, gid));
2419 }
2420 
2421 void
2422 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2423 {
2424 	vm_object_t object;
2425 
2426 	if ((object = vp->v_object) == NULL)
2427 		return;
2428 	VM_OBJECT_WLOCK(object);
2429 	vm_object_page_remove(object, start, end, 0);
2430 	VM_OBJECT_WUNLOCK(object);
2431 }
2432 
2433 int
2434 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2435     struct ucred *cred)
2436 {
2437 	struct vattr va;
2438 	daddr_t bn, bnp;
2439 	uint64_t bsize;
2440 	off_t noff;
2441 	int error;
2442 
2443 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2444 	    ("%s: Wrong command %lu", __func__, cmd));
2445 	ASSERT_VOP_LOCKED(vp, "vn_bmap_seekhole_locked");
2446 
2447 	if (vp->v_type != VREG) {
2448 		error = ENOTTY;
2449 		goto out;
2450 	}
2451 	error = VOP_GETATTR(vp, &va, cred);
2452 	if (error != 0)
2453 		goto out;
2454 	noff = *off;
2455 	if (noff >= va.va_size) {
2456 		error = ENXIO;
2457 		goto out;
2458 	}
2459 	bsize = vp->v_mount->mnt_stat.f_iosize;
2460 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2461 	    noff % bsize) {
2462 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2463 		if (error == EOPNOTSUPP) {
2464 			error = ENOTTY;
2465 			goto out;
2466 		}
2467 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2468 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2469 			noff = bn * bsize;
2470 			if (noff < *off)
2471 				noff = *off;
2472 			goto out;
2473 		}
2474 	}
2475 	if (noff > va.va_size)
2476 		noff = va.va_size;
2477 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2478 	if (cmd == FIOSEEKDATA)
2479 		error = ENXIO;
2480 out:
2481 	if (error == 0)
2482 		*off = noff;
2483 	return (error);
2484 }
2485 
2486 int
2487 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2488 {
2489 	int error;
2490 
2491 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2492 	    ("%s: Wrong command %lu", __func__, cmd));
2493 
2494 	if (vn_lock(vp, LK_SHARED) != 0)
2495 		return (EBADF);
2496 	error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2497 	VOP_UNLOCK(vp);
2498 	return (error);
2499 }
2500 
2501 int
2502 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2503 {
2504 	struct ucred *cred;
2505 	struct vnode *vp;
2506 	struct vattr vattr;
2507 	off_t foffset, size;
2508 	int error, noneg;
2509 
2510 	cred = td->td_ucred;
2511 	vp = fp->f_vnode;
2512 	foffset = foffset_lock(fp, 0);
2513 	noneg = (vp->v_type != VCHR);
2514 	error = 0;
2515 	switch (whence) {
2516 	case L_INCR:
2517 		if (noneg &&
2518 		    (foffset < 0 ||
2519 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2520 			error = EOVERFLOW;
2521 			break;
2522 		}
2523 		offset += foffset;
2524 		break;
2525 	case L_XTND:
2526 		vn_lock(vp, LK_SHARED | LK_RETRY);
2527 		error = VOP_GETATTR(vp, &vattr, cred);
2528 		VOP_UNLOCK(vp);
2529 		if (error)
2530 			break;
2531 
2532 		/*
2533 		 * If the file references a disk device, then fetch
2534 		 * the media size and use that to determine the ending
2535 		 * offset.
2536 		 */
2537 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2538 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2539 			vattr.va_size = size;
2540 		if (noneg &&
2541 		    (vattr.va_size > OFF_MAX ||
2542 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2543 			error = EOVERFLOW;
2544 			break;
2545 		}
2546 		offset += vattr.va_size;
2547 		break;
2548 	case L_SET:
2549 		break;
2550 	case SEEK_DATA:
2551 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2552 		if (error == ENOTTY)
2553 			error = EINVAL;
2554 		break;
2555 	case SEEK_HOLE:
2556 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2557 		if (error == ENOTTY)
2558 			error = EINVAL;
2559 		break;
2560 	default:
2561 		error = EINVAL;
2562 	}
2563 	if (error == 0 && noneg && offset < 0)
2564 		error = EINVAL;
2565 	if (error != 0)
2566 		goto drop;
2567 	VFS_KNOTE_UNLOCKED(vp, 0);
2568 	td->td_uretoff.tdu_off = offset;
2569 drop:
2570 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2571 	return (error);
2572 }
2573 
2574 int
2575 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2576     struct thread *td)
2577 {
2578 	int error;
2579 
2580 	/*
2581 	 * Grant permission if the caller is the owner of the file, or
2582 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2583 	 * on the file.  If the time pointer is null, then write
2584 	 * permission on the file is also sufficient.
2585 	 *
2586 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2587 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2588 	 * will be allowed to set the times [..] to the current
2589 	 * server time.
2590 	 */
2591 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2592 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2593 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2594 	return (error);
2595 }
2596 
2597 int
2598 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2599 {
2600 	struct vnode *vp;
2601 	int error;
2602 
2603 	if (fp->f_type == DTYPE_FIFO)
2604 		kif->kf_type = KF_TYPE_FIFO;
2605 	else
2606 		kif->kf_type = KF_TYPE_VNODE;
2607 	vp = fp->f_vnode;
2608 	vref(vp);
2609 	FILEDESC_SUNLOCK(fdp);
2610 	error = vn_fill_kinfo_vnode(vp, kif);
2611 	vrele(vp);
2612 	FILEDESC_SLOCK(fdp);
2613 	return (error);
2614 }
2615 
2616 static inline void
2617 vn_fill_junk(struct kinfo_file *kif)
2618 {
2619 	size_t len, olen;
2620 
2621 	/*
2622 	 * Simulate vn_fullpath returning changing values for a given
2623 	 * vp during e.g. coredump.
2624 	 */
2625 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2626 	olen = strlen(kif->kf_path);
2627 	if (len < olen)
2628 		strcpy(&kif->kf_path[len - 1], "$");
2629 	else
2630 		for (; olen < len; olen++)
2631 			strcpy(&kif->kf_path[olen], "A");
2632 }
2633 
2634 int
2635 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2636 {
2637 	struct vattr va;
2638 	char *fullpath, *freepath;
2639 	int error;
2640 
2641 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2642 	freepath = NULL;
2643 	fullpath = "-";
2644 	error = vn_fullpath(vp, &fullpath, &freepath);
2645 	if (error == 0) {
2646 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2647 	}
2648 	if (freepath != NULL)
2649 		free(freepath, M_TEMP);
2650 
2651 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2652 		vn_fill_junk(kif);
2653 	);
2654 
2655 	/*
2656 	 * Retrieve vnode attributes.
2657 	 */
2658 	va.va_fsid = VNOVAL;
2659 	va.va_rdev = NODEV;
2660 	vn_lock(vp, LK_SHARED | LK_RETRY);
2661 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2662 	VOP_UNLOCK(vp);
2663 	if (error != 0)
2664 		return (error);
2665 	if (va.va_fsid != VNOVAL)
2666 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2667 	else
2668 		kif->kf_un.kf_file.kf_file_fsid =
2669 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2670 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2671 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2672 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2673 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2674 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2675 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2676 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2677 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2678 	return (0);
2679 }
2680 
2681 int
2682 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2683     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2684     struct thread *td)
2685 {
2686 #ifdef HWPMC_HOOKS
2687 	struct pmckern_map_in pkm;
2688 #endif
2689 	struct mount *mp;
2690 	struct vnode *vp;
2691 	vm_object_t object;
2692 	vm_prot_t maxprot;
2693 	boolean_t writecounted;
2694 	int error;
2695 
2696 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2697     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2698 	/*
2699 	 * POSIX shared-memory objects are defined to have
2700 	 * kernel persistence, and are not defined to support
2701 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2702 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2703 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2704 	 * flag to request this behavior.
2705 	 */
2706 	if ((fp->f_flag & FPOSIXSHM) != 0)
2707 		flags |= MAP_NOSYNC;
2708 #endif
2709 	vp = fp->f_vnode;
2710 
2711 	/*
2712 	 * Ensure that file and memory protections are
2713 	 * compatible.  Note that we only worry about
2714 	 * writability if mapping is shared; in this case,
2715 	 * current and max prot are dictated by the open file.
2716 	 * XXX use the vnode instead?  Problem is: what
2717 	 * credentials do we use for determination? What if
2718 	 * proc does a setuid?
2719 	 */
2720 	mp = vp->v_mount;
2721 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2722 		maxprot = VM_PROT_NONE;
2723 		if ((prot & VM_PROT_EXECUTE) != 0)
2724 			return (EACCES);
2725 	} else
2726 		maxprot = VM_PROT_EXECUTE;
2727 	if ((fp->f_flag & FREAD) != 0)
2728 		maxprot |= VM_PROT_READ;
2729 	else if ((prot & VM_PROT_READ) != 0)
2730 		return (EACCES);
2731 
2732 	/*
2733 	 * If we are sharing potential changes via MAP_SHARED and we
2734 	 * are trying to get write permission although we opened it
2735 	 * without asking for it, bail out.
2736 	 */
2737 	if ((flags & MAP_SHARED) != 0) {
2738 		if ((fp->f_flag & FWRITE) != 0)
2739 			maxprot |= VM_PROT_WRITE;
2740 		else if ((prot & VM_PROT_WRITE) != 0)
2741 			return (EACCES);
2742 	} else {
2743 		maxprot |= VM_PROT_WRITE;
2744 		cap_maxprot |= VM_PROT_WRITE;
2745 	}
2746 	maxprot &= cap_maxprot;
2747 
2748 	/*
2749 	 * For regular files and shared memory, POSIX requires that
2750 	 * the value of foff be a legitimate offset within the data
2751 	 * object.  In particular, negative offsets are invalid.
2752 	 * Blocking negative offsets and overflows here avoids
2753 	 * possible wraparound or user-level access into reserved
2754 	 * ranges of the data object later.  In contrast, POSIX does
2755 	 * not dictate how offsets are used by device drivers, so in
2756 	 * the case of a device mapping a negative offset is passed
2757 	 * on.
2758 	 */
2759 	if (
2760 #ifdef _LP64
2761 	    size > OFF_MAX ||
2762 #endif
2763 	    foff > OFF_MAX - size)
2764 		return (EINVAL);
2765 
2766 	writecounted = FALSE;
2767 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2768 	    &foff, &object, &writecounted);
2769 	if (error != 0)
2770 		return (error);
2771 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2772 	    foff, writecounted, td);
2773 	if (error != 0) {
2774 		/*
2775 		 * If this mapping was accounted for in the vnode's
2776 		 * writecount, then undo that now.
2777 		 */
2778 		if (writecounted)
2779 			vm_pager_release_writecount(object, 0, size);
2780 		vm_object_deallocate(object);
2781 	}
2782 #ifdef HWPMC_HOOKS
2783 	/* Inform hwpmc(4) if an executable is being mapped. */
2784 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2785 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2786 			pkm.pm_file = vp;
2787 			pkm.pm_address = (uintptr_t) *addr;
2788 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2789 		}
2790 	}
2791 #endif
2792 	return (error);
2793 }
2794 
2795 void
2796 vn_fsid(struct vnode *vp, struct vattr *va)
2797 {
2798 	fsid_t *f;
2799 
2800 	f = &vp->v_mount->mnt_stat.f_fsid;
2801 	va->va_fsid = (uint32_t)f->val[1];
2802 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2803 	va->va_fsid += (uint32_t)f->val[0];
2804 }
2805 
2806 int
2807 vn_fsync_buf(struct vnode *vp, int waitfor)
2808 {
2809 	struct buf *bp, *nbp;
2810 	struct bufobj *bo;
2811 	struct mount *mp;
2812 	int error, maxretry;
2813 
2814 	error = 0;
2815 	maxretry = 10000;     /* large, arbitrarily chosen */
2816 	mp = NULL;
2817 	if (vp->v_type == VCHR) {
2818 		VI_LOCK(vp);
2819 		mp = vp->v_rdev->si_mountpt;
2820 		VI_UNLOCK(vp);
2821 	}
2822 	bo = &vp->v_bufobj;
2823 	BO_LOCK(bo);
2824 loop1:
2825 	/*
2826 	 * MARK/SCAN initialization to avoid infinite loops.
2827 	 */
2828         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2829 		bp->b_vflags &= ~BV_SCANNED;
2830 		bp->b_error = 0;
2831 	}
2832 
2833 	/*
2834 	 * Flush all dirty buffers associated with a vnode.
2835 	 */
2836 loop2:
2837 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2838 		if ((bp->b_vflags & BV_SCANNED) != 0)
2839 			continue;
2840 		bp->b_vflags |= BV_SCANNED;
2841 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2842 			if (waitfor != MNT_WAIT)
2843 				continue;
2844 			if (BUF_LOCK(bp,
2845 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2846 			    BO_LOCKPTR(bo)) != 0) {
2847 				BO_LOCK(bo);
2848 				goto loop1;
2849 			}
2850 			BO_LOCK(bo);
2851 		}
2852 		BO_UNLOCK(bo);
2853 		KASSERT(bp->b_bufobj == bo,
2854 		    ("bp %p wrong b_bufobj %p should be %p",
2855 		    bp, bp->b_bufobj, bo));
2856 		if ((bp->b_flags & B_DELWRI) == 0)
2857 			panic("fsync: not dirty");
2858 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2859 			vfs_bio_awrite(bp);
2860 		} else {
2861 			bremfree(bp);
2862 			bawrite(bp);
2863 		}
2864 		if (maxretry < 1000)
2865 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2866 		BO_LOCK(bo);
2867 		goto loop2;
2868 	}
2869 
2870 	/*
2871 	 * If synchronous the caller expects us to completely resolve all
2872 	 * dirty buffers in the system.  Wait for in-progress I/O to
2873 	 * complete (which could include background bitmap writes), then
2874 	 * retry if dirty blocks still exist.
2875 	 */
2876 	if (waitfor == MNT_WAIT) {
2877 		bufobj_wwait(bo, 0, 0);
2878 		if (bo->bo_dirty.bv_cnt > 0) {
2879 			/*
2880 			 * If we are unable to write any of these buffers
2881 			 * then we fail now rather than trying endlessly
2882 			 * to write them out.
2883 			 */
2884 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2885 				if ((error = bp->b_error) != 0)
2886 					break;
2887 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2888 			    (error == 0 && --maxretry >= 0))
2889 				goto loop1;
2890 			if (error == 0)
2891 				error = EAGAIN;
2892 		}
2893 	}
2894 	BO_UNLOCK(bo);
2895 	if (error != 0)
2896 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2897 
2898 	return (error);
2899 }
2900 
2901 /*
2902  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2903  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2904  * to do the actual copy.
2905  * vn_generic_copy_file_range() is factored out, so it can be called
2906  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2907  * different file systems.
2908  */
2909 int
2910 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2911     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2912     struct ucred *outcred, struct thread *fsize_td)
2913 {
2914 	int error;
2915 	size_t len;
2916 	uint64_t uval;
2917 
2918 	len = *lenp;
2919 	*lenp = 0;		/* For error returns. */
2920 	error = 0;
2921 
2922 	/* Do some sanity checks on the arguments. */
2923 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2924 		error = EISDIR;
2925 	else if (*inoffp < 0 || *outoffp < 0 ||
2926 	    invp->v_type != VREG || outvp->v_type != VREG)
2927 		error = EINVAL;
2928 	if (error != 0)
2929 		goto out;
2930 
2931 	/* Ensure offset + len does not wrap around. */
2932 	uval = *inoffp;
2933 	uval += len;
2934 	if (uval > INT64_MAX)
2935 		len = INT64_MAX - *inoffp;
2936 	uval = *outoffp;
2937 	uval += len;
2938 	if (uval > INT64_MAX)
2939 		len = INT64_MAX - *outoffp;
2940 	if (len == 0)
2941 		goto out;
2942 
2943 	/*
2944 	 * If the two vnode are for the same file system, call
2945 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2946 	 * which can handle copies across multiple file systems.
2947 	 */
2948 	*lenp = len;
2949 	if (invp->v_mount == outvp->v_mount)
2950 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2951 		    lenp, flags, incred, outcred, fsize_td);
2952 	else
2953 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2954 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2955 out:
2956 	return (error);
2957 }
2958 
2959 /*
2960  * Test len bytes of data starting at dat for all bytes == 0.
2961  * Return true if all bytes are zero, false otherwise.
2962  * Expects dat to be well aligned.
2963  */
2964 static bool
2965 mem_iszero(void *dat, int len)
2966 {
2967 	int i;
2968 	const u_int *p;
2969 	const char *cp;
2970 
2971 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2972 		if (len >= sizeof(*p)) {
2973 			if (*p != 0)
2974 				return (false);
2975 		} else {
2976 			cp = (const char *)p;
2977 			for (i = 0; i < len; i++, cp++)
2978 				if (*cp != '\0')
2979 					return (false);
2980 		}
2981 	}
2982 	return (true);
2983 }
2984 
2985 /*
2986  * Look for a hole in the output file and, if found, adjust *outoffp
2987  * and *xferp to skip past the hole.
2988  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2989  * to be written as 0's upon return.
2990  */
2991 static off_t
2992 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2993     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2994 {
2995 	int error;
2996 	off_t delta;
2997 
2998 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2999 		*dataoffp = *outoffp;
3000 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3001 		    curthread);
3002 		if (error == 0) {
3003 			*holeoffp = *dataoffp;
3004 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3005 			    curthread);
3006 		}
3007 		if (error != 0 || *holeoffp == *dataoffp) {
3008 			/*
3009 			 * Since outvp is unlocked, it may be possible for
3010 			 * another thread to do a truncate(), lseek(), write()
3011 			 * creating a hole at startoff between the above
3012 			 * VOP_IOCTL() calls, if the other thread does not do
3013 			 * rangelocking.
3014 			 * If that happens, *holeoffp == *dataoffp and finding
3015 			 * the hole has failed, so disable vn_skip_hole().
3016 			 */
3017 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
3018 			return (xfer2);
3019 		}
3020 		KASSERT(*dataoffp >= *outoffp,
3021 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3022 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
3023 		KASSERT(*holeoffp > *dataoffp,
3024 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3025 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3026 	}
3027 
3028 	/*
3029 	 * If there is a hole before the data starts, advance *outoffp and
3030 	 * *xferp past the hole.
3031 	 */
3032 	if (*dataoffp > *outoffp) {
3033 		delta = *dataoffp - *outoffp;
3034 		if (delta >= *xferp) {
3035 			/* Entire *xferp is a hole. */
3036 			*outoffp += *xferp;
3037 			*xferp = 0;
3038 			return (0);
3039 		}
3040 		*xferp -= delta;
3041 		*outoffp += delta;
3042 		xfer2 = MIN(xfer2, *xferp);
3043 	}
3044 
3045 	/*
3046 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3047 	 * that the write ends at the start of the hole.
3048 	 * *holeoffp should always be greater than *outoffp, but for the
3049 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3050 	 * value.
3051 	 */
3052 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3053 		xfer2 = *holeoffp - *outoffp;
3054 	return (xfer2);
3055 }
3056 
3057 /*
3058  * Write an xfer sized chunk to outvp in blksize blocks from dat.
3059  * dat is a maximum of blksize in length and can be written repeatedly in
3060  * the chunk.
3061  * If growfile == true, just grow the file via vn_truncate_locked() instead
3062  * of doing actual writes.
3063  * If checkhole == true, a hole is being punched, so skip over any hole
3064  * already in the output file.
3065  */
3066 static int
3067 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3068     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3069 {
3070 	struct mount *mp;
3071 	off_t dataoff, holeoff, xfer2;
3072 	int error;
3073 
3074 	/*
3075 	 * Loop around doing writes of blksize until write has been completed.
3076 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3077 	 * done for each iteration, since the xfer argument can be very
3078 	 * large if there is a large hole to punch in the output file.
3079 	 */
3080 	error = 0;
3081 	holeoff = 0;
3082 	do {
3083 		xfer2 = MIN(xfer, blksize);
3084 		if (checkhole) {
3085 			/*
3086 			 * Punching a hole.  Skip writing if there is
3087 			 * already a hole in the output file.
3088 			 */
3089 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3090 			    &dataoff, &holeoff, cred);
3091 			if (xfer == 0)
3092 				break;
3093 			if (holeoff < 0)
3094 				checkhole = false;
3095 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3096 			    (intmax_t)xfer2));
3097 		}
3098 		bwillwrite();
3099 		mp = NULL;
3100 		error = vn_start_write(outvp, &mp, V_WAIT);
3101 		if (error != 0)
3102 			break;
3103 		if (growfile) {
3104 			error = vn_lock(outvp, LK_EXCLUSIVE);
3105 			if (error == 0) {
3106 				error = vn_truncate_locked(outvp, outoff + xfer,
3107 				    false, cred);
3108 				VOP_UNLOCK(outvp);
3109 			}
3110 		} else {
3111 			error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3112 			if (error == 0) {
3113 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3114 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3115 				    curthread->td_ucred, cred, NULL, curthread);
3116 				outoff += xfer2;
3117 				xfer -= xfer2;
3118 				VOP_UNLOCK(outvp);
3119 			}
3120 		}
3121 		if (mp != NULL)
3122 			vn_finished_write(mp);
3123 	} while (!growfile && xfer > 0 && error == 0);
3124 	return (error);
3125 }
3126 
3127 /*
3128  * Copy a byte range of one file to another.  This function can handle the
3129  * case where invp and outvp are on different file systems.
3130  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3131  * is no better file system specific way to do it.
3132  */
3133 int
3134 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3135     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3136     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3137 {
3138 	struct vattr va, inva;
3139 	struct mount *mp;
3140 	struct uio io;
3141 	off_t startoff, endoff, xfer, xfer2;
3142 	u_long blksize;
3143 	int error, interrupted;
3144 	bool cantseek, readzeros, eof, lastblock, holetoeof;
3145 	ssize_t aresid;
3146 	size_t copylen, len, rem, savlen;
3147 	char *dat;
3148 	long holein, holeout;
3149 
3150 	holein = holeout = 0;
3151 	savlen = len = *lenp;
3152 	error = 0;
3153 	interrupted = 0;
3154 	dat = NULL;
3155 
3156 	error = vn_lock(invp, LK_SHARED);
3157 	if (error != 0)
3158 		goto out;
3159 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3160 		holein = 0;
3161 	if (holein > 0)
3162 		error = VOP_GETATTR(invp, &inva, incred);
3163 	VOP_UNLOCK(invp);
3164 	if (error != 0)
3165 		goto out;
3166 
3167 	mp = NULL;
3168 	error = vn_start_write(outvp, &mp, V_WAIT);
3169 	if (error == 0)
3170 		error = vn_lock(outvp, LK_EXCLUSIVE);
3171 	if (error == 0) {
3172 		/*
3173 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3174 		 * now that outvp is locked.
3175 		 */
3176 		if (fsize_td != NULL) {
3177 			io.uio_offset = *outoffp;
3178 			io.uio_resid = len;
3179 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3180 			if (error != 0)
3181 				error = EFBIG;
3182 		}
3183 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3184 			holeout = 0;
3185 		/*
3186 		 * Holes that are past EOF do not need to be written as a block
3187 		 * of zero bytes.  So, truncate the output file as far as
3188 		 * possible and then use va.va_size to decide if writing 0
3189 		 * bytes is necessary in the loop below.
3190 		 */
3191 		if (error == 0)
3192 			error = VOP_GETATTR(outvp, &va, outcred);
3193 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3194 		    *outoffp + len) {
3195 #ifdef MAC
3196 			error = mac_vnode_check_write(curthread->td_ucred,
3197 			    outcred, outvp);
3198 			if (error == 0)
3199 #endif
3200 				error = vn_truncate_locked(outvp, *outoffp,
3201 				    false, outcred);
3202 			if (error == 0)
3203 				va.va_size = *outoffp;
3204 		}
3205 		VOP_UNLOCK(outvp);
3206 	}
3207 	if (mp != NULL)
3208 		vn_finished_write(mp);
3209 	if (error != 0)
3210 		goto out;
3211 
3212 	/*
3213 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3214 	 * If hole sizes aren't available, set the blksize to the larger
3215 	 * f_iosize of invp and outvp.
3216 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3217 	 * This value is clipped at 4Kbytes and 1Mbyte.
3218 	 */
3219 	blksize = MAX(holein, holeout);
3220 
3221 	/* Clip len to end at an exact multiple of hole size. */
3222 	if (blksize > 1) {
3223 		rem = *inoffp % blksize;
3224 		if (rem > 0)
3225 			rem = blksize - rem;
3226 		if (len > rem && len - rem > blksize)
3227 			len = savlen = rounddown(len - rem, blksize) + rem;
3228 	}
3229 
3230 	if (blksize <= 1)
3231 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3232 		    outvp->v_mount->mnt_stat.f_iosize);
3233 	if (blksize < 4096)
3234 		blksize = 4096;
3235 	else if (blksize > 1024 * 1024)
3236 		blksize = 1024 * 1024;
3237 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3238 
3239 	/*
3240 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3241 	 * to find holes.  Otherwise, just scan the read block for all 0s
3242 	 * in the inner loop where the data copying is done.
3243 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3244 	 * support holes on the server, but do not support FIOSEEKHOLE.
3245 	 */
3246 	holetoeof = eof = false;
3247 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3248 		endoff = 0;			/* To shut up compilers. */
3249 		cantseek = true;
3250 		startoff = *inoffp;
3251 		copylen = len;
3252 
3253 		/*
3254 		 * Find the next data area.  If there is just a hole to EOF,
3255 		 * FIOSEEKDATA should fail with ENXIO.
3256 		 * (I do not know if any file system will report a hole to
3257 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3258 		 *  will fail for those file systems.)
3259 		 *
3260 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3261 		 * the code just falls through to the inner copy loop.
3262 		 */
3263 		error = EINVAL;
3264 		if (holein > 0) {
3265 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3266 			    incred, curthread);
3267 			if (error == ENXIO) {
3268 				startoff = endoff = inva.va_size;
3269 				eof = holetoeof = true;
3270 				error = 0;
3271 			}
3272 		}
3273 		if (error == 0 && !holetoeof) {
3274 			endoff = startoff;
3275 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3276 			    incred, curthread);
3277 			/*
3278 			 * Since invp is unlocked, it may be possible for
3279 			 * another thread to do a truncate(), lseek(), write()
3280 			 * creating a hole at startoff between the above
3281 			 * VOP_IOCTL() calls, if the other thread does not do
3282 			 * rangelocking.
3283 			 * If that happens, startoff == endoff and finding
3284 			 * the hole has failed, so set an error.
3285 			 */
3286 			if (error == 0 && startoff == endoff)
3287 				error = EINVAL; /* Any error. Reset to 0. */
3288 		}
3289 		if (error == 0) {
3290 			if (startoff > *inoffp) {
3291 				/* Found hole before data block. */
3292 				xfer = MIN(startoff - *inoffp, len);
3293 				if (*outoffp < va.va_size) {
3294 					/* Must write 0s to punch hole. */
3295 					xfer2 = MIN(va.va_size - *outoffp,
3296 					    xfer);
3297 					memset(dat, 0, MIN(xfer2, blksize));
3298 					error = vn_write_outvp(outvp, dat,
3299 					    *outoffp, xfer2, blksize, false,
3300 					    holeout > 0, outcred);
3301 				}
3302 
3303 				if (error == 0 && *outoffp + xfer >
3304 				    va.va_size && (xfer == len || holetoeof)) {
3305 					/* Grow output file (hole at end). */
3306 					error = vn_write_outvp(outvp, dat,
3307 					    *outoffp, xfer, blksize, true,
3308 					    false, outcred);
3309 				}
3310 				if (error == 0) {
3311 					*inoffp += xfer;
3312 					*outoffp += xfer;
3313 					len -= xfer;
3314 					if (len < savlen)
3315 						interrupted = sig_intr();
3316 				}
3317 			}
3318 			copylen = MIN(len, endoff - startoff);
3319 			cantseek = false;
3320 		} else {
3321 			cantseek = true;
3322 			startoff = *inoffp;
3323 			copylen = len;
3324 			error = 0;
3325 		}
3326 
3327 		xfer = blksize;
3328 		if (cantseek) {
3329 			/*
3330 			 * Set first xfer to end at a block boundary, so that
3331 			 * holes are more likely detected in the loop below via
3332 			 * the for all bytes 0 method.
3333 			 */
3334 			xfer -= (*inoffp % blksize);
3335 		}
3336 		/* Loop copying the data block. */
3337 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3338 			if (copylen < xfer)
3339 				xfer = copylen;
3340 			error = vn_lock(invp, LK_SHARED);
3341 			if (error != 0)
3342 				goto out;
3343 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3344 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3345 			    curthread->td_ucred, incred, &aresid,
3346 			    curthread);
3347 			VOP_UNLOCK(invp);
3348 			lastblock = false;
3349 			if (error == 0 && aresid > 0) {
3350 				/* Stop the copy at EOF on the input file. */
3351 				xfer -= aresid;
3352 				eof = true;
3353 				lastblock = true;
3354 			}
3355 			if (error == 0) {
3356 				/*
3357 				 * Skip the write for holes past the initial EOF
3358 				 * of the output file, unless this is the last
3359 				 * write of the output file at EOF.
3360 				 */
3361 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3362 				    false;
3363 				if (xfer == len)
3364 					lastblock = true;
3365 				if (!cantseek || *outoffp < va.va_size ||
3366 				    lastblock || !readzeros)
3367 					error = vn_write_outvp(outvp, dat,
3368 					    *outoffp, xfer, blksize,
3369 					    readzeros && lastblock &&
3370 					    *outoffp >= va.va_size, false,
3371 					    outcred);
3372 				if (error == 0) {
3373 					*inoffp += xfer;
3374 					startoff += xfer;
3375 					*outoffp += xfer;
3376 					copylen -= xfer;
3377 					len -= xfer;
3378 					if (len < savlen)
3379 						interrupted = sig_intr();
3380 				}
3381 			}
3382 			xfer = blksize;
3383 		}
3384 	}
3385 out:
3386 	*lenp = savlen - len;
3387 	free(dat, M_TEMP);
3388 	return (error);
3389 }
3390 
3391 static int
3392 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3393 {
3394 	struct mount *mp;
3395 	struct vnode *vp;
3396 	off_t olen, ooffset;
3397 	int error;
3398 #ifdef AUDIT
3399 	int audited_vnode1 = 0;
3400 #endif
3401 
3402 	vp = fp->f_vnode;
3403 	if (vp->v_type != VREG)
3404 		return (ENODEV);
3405 
3406 	/* Allocating blocks may take a long time, so iterate. */
3407 	for (;;) {
3408 		olen = len;
3409 		ooffset = offset;
3410 
3411 		bwillwrite();
3412 		mp = NULL;
3413 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3414 		if (error != 0)
3415 			break;
3416 		error = vn_lock(vp, LK_EXCLUSIVE);
3417 		if (error != 0) {
3418 			vn_finished_write(mp);
3419 			break;
3420 		}
3421 #ifdef AUDIT
3422 		if (!audited_vnode1) {
3423 			AUDIT_ARG_VNODE1(vp);
3424 			audited_vnode1 = 1;
3425 		}
3426 #endif
3427 #ifdef MAC
3428 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3429 		if (error == 0)
3430 #endif
3431 			error = VOP_ALLOCATE(vp, &offset, &len);
3432 		VOP_UNLOCK(vp);
3433 		vn_finished_write(mp);
3434 
3435 		if (olen + ooffset != offset + len) {
3436 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3437 			    ooffset, olen, offset, len);
3438 		}
3439 		if (error != 0 || len == 0)
3440 			break;
3441 		KASSERT(olen > len, ("Iteration did not make progress?"));
3442 		maybe_yield();
3443 	}
3444 
3445 	return (error);
3446 }
3447 
3448 static int
3449 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3450     int ioflg, struct ucred *active_cred, struct ucred *file_cred)
3451 {
3452 	struct mount *mp;
3453 	void *rl_cookie;
3454 	off_t off, len;
3455 	int error;
3456 #ifdef AUDIT
3457 	bool audited_vnode1 = false;
3458 #endif
3459 
3460 	rl_cookie = NULL;
3461 	error = 0;
3462 	mp = NULL;
3463 	off = *offset;
3464 	len = *length;
3465 
3466 	if ((ioflg & (IO_NODELOCKED|IO_RANGELOCKED)) == 0)
3467 		rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3468 	while (len > 0 && error == 0) {
3469 		/*
3470 		 * Try to deallocate the longest range in one pass.
3471 		 * In case a pass takes too long to be executed, it returns
3472 		 * partial result. The residue will be proceeded in the next
3473 		 * pass.
3474 		 */
3475 
3476 		if ((ioflg & IO_NODELOCKED) == 0) {
3477 			bwillwrite();
3478 			if ((error = vn_start_write(vp, &mp,
3479 			    V_WAIT | PCATCH)) != 0)
3480 				goto out;
3481 			vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3482 		}
3483 #ifdef AUDIT
3484 		if (!audited_vnode1) {
3485 			AUDIT_ARG_VNODE1(vp);
3486 			audited_vnode1 = true;
3487 		}
3488 #endif
3489 
3490 #ifdef MAC
3491 		if ((ioflg & IO_NOMACCHECK) == 0)
3492 			error = mac_vnode_check_write(active_cred, file_cred,
3493 			    vp);
3494 #endif
3495 		if (error == 0)
3496 			error = VOP_DEALLOCATE(vp, &off, &len, flags,
3497 			    active_cred);
3498 
3499 		if ((ioflg & IO_NODELOCKED) == 0) {
3500 			VOP_UNLOCK(vp);
3501 			if (mp != NULL) {
3502 				vn_finished_write(mp);
3503 				mp = NULL;
3504 			}
3505 		}
3506 	}
3507 out:
3508 	if (rl_cookie != NULL)
3509 		vn_rangelock_unlock(vp, rl_cookie);
3510 	*offset = off;
3511 	*length = len;
3512 	return (error);
3513 }
3514 
3515 int
3516 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3517     int ioflg, struct ucred *active_cred, struct ucred *file_cred)
3518 {
3519 	if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3520 	    flags != 0)
3521 		return (EINVAL);
3522 	if (vp->v_type != VREG)
3523 		return (ENODEV);
3524 
3525 	return (vn_deallocate_impl(vp, offset, length, flags, ioflg,
3526 	    active_cred, file_cred));
3527 }
3528 
3529 static int
3530 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3531     struct ucred *active_cred, struct thread *td)
3532 {
3533 	int error;
3534 	struct vnode *vp;
3535 
3536 	vp = fp->f_vnode;
3537 
3538 	if (cmd != SPACECTL_DEALLOC || *offset < 0 || *length <= 0 ||
3539 	    *length > OFF_MAX - *offset || flags != 0)
3540 		return (EINVAL);
3541 	if (vp->v_type != VREG)
3542 		return (ENODEV);
3543 
3544 	switch (cmd) {
3545 	case SPACECTL_DEALLOC:
3546 		error = vn_deallocate_impl(vp, offset, length, flags, 0,
3547 		    active_cred, fp->f_cred);
3548 		break;
3549 	default:
3550 		panic("vn_fspacectl: unknown cmd %d", cmd);
3551 	}
3552 
3553 	return (error);
3554 }
3555 
3556 static u_long vn_lock_pair_pause_cnt;
3557 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3558     &vn_lock_pair_pause_cnt, 0,
3559     "Count of vn_lock_pair deadlocks");
3560 
3561 u_int vn_lock_pair_pause_max;
3562 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3563     &vn_lock_pair_pause_max, 0,
3564     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3565 
3566 static void
3567 vn_lock_pair_pause(const char *wmesg)
3568 {
3569 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3570 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3571 }
3572 
3573 /*
3574  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3575  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3576  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3577  * can be NULL.
3578  *
3579  * The function returns with both vnodes exclusively locked, and
3580  * guarantees that it does not create lock order reversal with other
3581  * threads during its execution.  Both vnodes could be unlocked
3582  * temporary (and reclaimed).
3583  */
3584 void
3585 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3586     bool vp2_locked)
3587 {
3588 	int error;
3589 
3590 	if (vp1 == NULL && vp2 == NULL)
3591 		return;
3592 	if (vp1 != NULL) {
3593 		if (vp1_locked)
3594 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3595 		else
3596 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3597 	} else {
3598 		vp1_locked = true;
3599 	}
3600 	if (vp2 != NULL) {
3601 		if (vp2_locked)
3602 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3603 		else
3604 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3605 	} else {
3606 		vp2_locked = true;
3607 	}
3608 	if (!vp1_locked && !vp2_locked) {
3609 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3610 		vp1_locked = true;
3611 	}
3612 
3613 	for (;;) {
3614 		if (vp1_locked && vp2_locked)
3615 			break;
3616 		if (vp1_locked && vp2 != NULL) {
3617 			if (vp1 != NULL) {
3618 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3619 				    __FILE__, __LINE__);
3620 				if (error == 0)
3621 					break;
3622 				VOP_UNLOCK(vp1);
3623 				vp1_locked = false;
3624 				vn_lock_pair_pause("vlp1");
3625 			}
3626 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3627 			vp2_locked = true;
3628 		}
3629 		if (vp2_locked && vp1 != NULL) {
3630 			if (vp2 != NULL) {
3631 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3632 				    __FILE__, __LINE__);
3633 				if (error == 0)
3634 					break;
3635 				VOP_UNLOCK(vp2);
3636 				vp2_locked = false;
3637 				vn_lock_pair_pause("vlp2");
3638 			}
3639 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3640 			vp1_locked = true;
3641 		}
3642 	}
3643 	if (vp1 != NULL)
3644 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3645 	if (vp2 != NULL)
3646 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3647 }
3648 
3649 int
3650 vn_lktype_write(struct mount *mp, struct vnode *vp)
3651 {
3652 	if (MNT_SHARED_WRITES(mp) ||
3653 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3654 		return (LK_SHARED);
3655 	return (LK_EXCLUSIVE);
3656 }
3657