xref: /freebsd/sys/kern/vfs_vnops.c (revision 71d104536b513298902be65342afe6f3792f29e4)
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/sx.h>
74 #include <sys/sleepqueue.h>
75 #include <sys/sysctl.h>
76 #include <sys/ttycom.h>
77 #include <sys/conf.h>
78 #include <sys/syslog.h>
79 #include <sys/unistd.h>
80 #include <sys/user.h>
81 
82 #include <security/audit/audit.h>
83 #include <security/mac/mac_framework.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_extern.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pager.h>
92 
93 #ifdef HWPMC_HOOKS
94 #include <sys/pmckern.h>
95 #endif
96 
97 static fo_rdwr_t	vn_read;
98 static fo_rdwr_t	vn_write;
99 static fo_rdwr_t	vn_io_fault;
100 static fo_truncate_t	vn_truncate;
101 static fo_ioctl_t	vn_ioctl;
102 static fo_poll_t	vn_poll;
103 static fo_kqfilter_t	vn_kqfilter;
104 static fo_stat_t	vn_statfile;
105 static fo_close_t	vn_closefile;
106 static fo_mmap_t	vn_mmap;
107 static fo_fallocate_t	vn_fallocate;
108 
109 struct 	fileops vnops = {
110 	.fo_read = vn_io_fault,
111 	.fo_write = vn_io_fault,
112 	.fo_truncate = vn_truncate,
113 	.fo_ioctl = vn_ioctl,
114 	.fo_poll = vn_poll,
115 	.fo_kqfilter = vn_kqfilter,
116 	.fo_stat = vn_statfile,
117 	.fo_close = vn_closefile,
118 	.fo_chmod = vn_chmod,
119 	.fo_chown = vn_chown,
120 	.fo_sendfile = vn_sendfile,
121 	.fo_seek = vn_seek,
122 	.fo_fill_kinfo = vn_fill_kinfo,
123 	.fo_mmap = vn_mmap,
124 	.fo_fallocate = vn_fallocate,
125 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
126 };
127 
128 static const int io_hold_cnt = 16;
129 static int vn_io_fault_enable = 1;
130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
131     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
132 static int vn_io_fault_prefault = 0;
133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW,
134     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
135 static u_long vn_io_faults_cnt;
136 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
137     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
138 
139 /*
140  * Returns true if vn_io_fault mode of handling the i/o request should
141  * be used.
142  */
143 static bool
144 do_vn_io_fault(struct vnode *vp, struct uio *uio)
145 {
146 	struct mount *mp;
147 
148 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
149 	    (mp = vp->v_mount) != NULL &&
150 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
151 }
152 
153 /*
154  * Structure used to pass arguments to vn_io_fault1(), to do either
155  * file- or vnode-based I/O calls.
156  */
157 struct vn_io_fault_args {
158 	enum {
159 		VN_IO_FAULT_FOP,
160 		VN_IO_FAULT_VOP
161 	} kind;
162 	struct ucred *cred;
163 	int flags;
164 	union {
165 		struct fop_args_tag {
166 			struct file *fp;
167 			fo_rdwr_t *doio;
168 		} fop_args;
169 		struct vop_args_tag {
170 			struct vnode *vp;
171 		} vop_args;
172 	} args;
173 };
174 
175 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
176     struct vn_io_fault_args *args, struct thread *td);
177 
178 int
179 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
180 {
181 	struct thread *td = ndp->ni_cnd.cn_thread;
182 
183 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
184 }
185 
186 /*
187  * Common code for vnode open operations via a name lookup.
188  * Lookup the vnode and invoke VOP_CREATE if needed.
189  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
190  *
191  * Note that this does NOT free nameidata for the successful case,
192  * due to the NDINIT being done elsewhere.
193  */
194 int
195 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
196     struct ucred *cred, struct file *fp)
197 {
198 	struct vnode *vp;
199 	struct mount *mp;
200 	struct thread *td = ndp->ni_cnd.cn_thread;
201 	struct vattr vat;
202 	struct vattr *vap = &vat;
203 	int fmode, error;
204 
205 restart:
206 	fmode = *flagp;
207 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
208 	    O_EXCL | O_DIRECTORY))
209 		return (EINVAL);
210 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
211 		ndp->ni_cnd.cn_nameiop = CREATE;
212 		/*
213 		 * Set NOCACHE to avoid flushing the cache when
214 		 * rolling in many files at once.
215 		*/
216 		ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
217 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
218 			ndp->ni_cnd.cn_flags |= FOLLOW;
219 		if ((fmode & O_BENEATH) != 0)
220 			ndp->ni_cnd.cn_flags |= BENEATH;
221 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
222 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
223 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
224 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
225 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
226 			bwillwrite();
227 		if ((error = namei(ndp)) != 0)
228 			return (error);
229 		if (ndp->ni_vp == NULL) {
230 			VATTR_NULL(vap);
231 			vap->va_type = VREG;
232 			vap->va_mode = cmode;
233 			if (fmode & O_EXCL)
234 				vap->va_vaflags |= VA_EXCLUSIVE;
235 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
236 				NDFREE(ndp, NDF_ONLY_PNBUF);
237 				vput(ndp->ni_dvp);
238 				if ((error = vn_start_write(NULL, &mp,
239 				    V_XSLEEP | PCATCH)) != 0)
240 					return (error);
241 				goto restart;
242 			}
243 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
244 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
245 #ifdef MAC
246 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
247 			    &ndp->ni_cnd, vap);
248 			if (error == 0)
249 #endif
250 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
251 						   &ndp->ni_cnd, vap);
252 			vput(ndp->ni_dvp);
253 			vn_finished_write(mp);
254 			if (error) {
255 				NDFREE(ndp, NDF_ONLY_PNBUF);
256 				return (error);
257 			}
258 			fmode &= ~O_TRUNC;
259 			vp = ndp->ni_vp;
260 		} else {
261 			if (ndp->ni_dvp == ndp->ni_vp)
262 				vrele(ndp->ni_dvp);
263 			else
264 				vput(ndp->ni_dvp);
265 			ndp->ni_dvp = NULL;
266 			vp = ndp->ni_vp;
267 			if (fmode & O_EXCL) {
268 				error = EEXIST;
269 				goto bad;
270 			}
271 			if (vp->v_type == VDIR) {
272 				error = EISDIR;
273 				goto bad;
274 			}
275 			fmode &= ~O_CREAT;
276 		}
277 	} else {
278 		ndp->ni_cnd.cn_nameiop = LOOKUP;
279 		ndp->ni_cnd.cn_flags = ISOPEN |
280 		    ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
281 		if (!(fmode & FWRITE))
282 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
283 		if ((fmode & O_BENEATH) != 0)
284 			ndp->ni_cnd.cn_flags |= BENEATH;
285 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
286 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
287 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
288 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
289 		if ((error = namei(ndp)) != 0)
290 			return (error);
291 		vp = ndp->ni_vp;
292 	}
293 	error = vn_open_vnode(vp, fmode, cred, td, fp);
294 	if (error)
295 		goto bad;
296 	*flagp = fmode;
297 	return (0);
298 bad:
299 	NDFREE(ndp, NDF_ONLY_PNBUF);
300 	vput(vp);
301 	*flagp = fmode;
302 	ndp->ni_vp = NULL;
303 	return (error);
304 }
305 
306 static int
307 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
308 {
309 	struct flock lf;
310 	int error, lock_flags, type;
311 
312 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
313 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
314 		return (0);
315 	KASSERT(fp != NULL, ("open with flock requires fp"));
316 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
317 		return (EOPNOTSUPP);
318 
319 	lock_flags = VOP_ISLOCKED(vp);
320 	VOP_UNLOCK(vp);
321 
322 	lf.l_whence = SEEK_SET;
323 	lf.l_start = 0;
324 	lf.l_len = 0;
325 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
326 	type = F_FLOCK;
327 	if ((fmode & FNONBLOCK) == 0)
328 		type |= F_WAIT;
329 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
330 	if (error == 0)
331 		fp->f_flag |= FHASLOCK;
332 
333 	vn_lock(vp, lock_flags | LK_RETRY);
334 	if (error == 0 && VN_IS_DOOMED(vp))
335 		error = ENOENT;
336 	return (error);
337 }
338 
339 /*
340  * Common code for vnode open operations once a vnode is located.
341  * Check permissions, and call the VOP_OPEN routine.
342  */
343 int
344 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
345     struct thread *td, struct file *fp)
346 {
347 	accmode_t accmode;
348 	int error;
349 
350 	if (vp->v_type == VLNK)
351 		return (EMLINK);
352 	if (vp->v_type == VSOCK)
353 		return (EOPNOTSUPP);
354 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
355 		return (ENOTDIR);
356 	accmode = 0;
357 	if (fmode & (FWRITE | O_TRUNC)) {
358 		if (vp->v_type == VDIR)
359 			return (EISDIR);
360 		accmode |= VWRITE;
361 	}
362 	if (fmode & FREAD)
363 		accmode |= VREAD;
364 	if (fmode & FEXEC)
365 		accmode |= VEXEC;
366 	if ((fmode & O_APPEND) && (fmode & FWRITE))
367 		accmode |= VAPPEND;
368 #ifdef MAC
369 	if (fmode & O_CREAT)
370 		accmode |= VCREAT;
371 	if (fmode & O_VERIFY)
372 		accmode |= VVERIFY;
373 	error = mac_vnode_check_open(cred, vp, accmode);
374 	if (error)
375 		return (error);
376 
377 	accmode &= ~(VCREAT | VVERIFY);
378 #endif
379 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
380 		error = VOP_ACCESS(vp, accmode, cred, td);
381 		if (error != 0)
382 			return (error);
383 	}
384 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
385 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
386 	error = VOP_OPEN(vp, fmode, cred, td, fp);
387 	if (error != 0)
388 		return (error);
389 
390 	error = vn_open_vnode_advlock(vp, fmode, fp);
391 	if (error == 0 && (fmode & FWRITE) != 0) {
392 		error = VOP_ADD_WRITECOUNT(vp, 1);
393 		if (error == 0) {
394 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
395 			     __func__, vp, vp->v_writecount);
396 		}
397 	}
398 
399 	/*
400 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
401 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
402 	 * Arrange for that by having fdrop() to use vn_closefile().
403 	 */
404 	if (error != 0) {
405 		fp->f_flag |= FOPENFAILED;
406 		fp->f_vnode = vp;
407 		if (fp->f_ops == &badfileops) {
408 			fp->f_type = DTYPE_VNODE;
409 			fp->f_ops = &vnops;
410 		}
411 		vref(vp);
412 	}
413 
414 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
415 	return (error);
416 
417 }
418 
419 /*
420  * Check for write permissions on the specified vnode.
421  * Prototype text segments cannot be written.
422  * It is racy.
423  */
424 int
425 vn_writechk(struct vnode *vp)
426 {
427 
428 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
429 	/*
430 	 * If there's shared text associated with
431 	 * the vnode, try to free it up once.  If
432 	 * we fail, we can't allow writing.
433 	 */
434 	if (VOP_IS_TEXT(vp))
435 		return (ETXTBSY);
436 
437 	return (0);
438 }
439 
440 /*
441  * Vnode close call
442  */
443 static int
444 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
445     struct thread *td, bool keep_ref)
446 {
447 	struct mount *mp;
448 	int error, lock_flags;
449 
450 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
451 	    MNT_EXTENDED_SHARED(vp->v_mount))
452 		lock_flags = LK_SHARED;
453 	else
454 		lock_flags = LK_EXCLUSIVE;
455 
456 	vn_start_write(vp, &mp, V_WAIT);
457 	vn_lock(vp, lock_flags | LK_RETRY);
458 	AUDIT_ARG_VNODE1(vp);
459 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
460 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
461 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
462 		    __func__, vp, vp->v_writecount);
463 	}
464 	error = VOP_CLOSE(vp, flags, file_cred, td);
465 	if (keep_ref)
466 		VOP_UNLOCK(vp);
467 	else
468 		vput(vp);
469 	vn_finished_write(mp);
470 	return (error);
471 }
472 
473 int
474 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
475     struct thread *td)
476 {
477 
478 	return (vn_close1(vp, flags, file_cred, td, false));
479 }
480 
481 /*
482  * Heuristic to detect sequential operation.
483  */
484 static int
485 sequential_heuristic(struct uio *uio, struct file *fp)
486 {
487 
488 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
489 	if (fp->f_flag & FRDAHEAD)
490 		return (fp->f_seqcount << IO_SEQSHIFT);
491 
492 	/*
493 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
494 	 * that the first I/O is normally considered to be slightly
495 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
496 	 * unless previous seeks have reduced f_seqcount to 0, in which
497 	 * case offset 0 is not special.
498 	 */
499 	if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
500 	    uio->uio_offset == fp->f_nextoff) {
501 		/*
502 		 * f_seqcount is in units of fixed-size blocks so that it
503 		 * depends mainly on the amount of sequential I/O and not
504 		 * much on the number of sequential I/O's.  The fixed size
505 		 * of 16384 is hard-coded here since it is (not quite) just
506 		 * a magic size that works well here.  This size is more
507 		 * closely related to the best I/O size for real disks than
508 		 * to any block size used by software.
509 		 */
510 		if (uio->uio_resid >= IO_SEQMAX * 16384)
511 			fp->f_seqcount = IO_SEQMAX;
512 		else {
513 			fp->f_seqcount += howmany(uio->uio_resid, 16384);
514 			if (fp->f_seqcount > IO_SEQMAX)
515 				fp->f_seqcount = IO_SEQMAX;
516 		}
517 		return (fp->f_seqcount << IO_SEQSHIFT);
518 	}
519 
520 	/* Not sequential.  Quickly draw-down sequentiality. */
521 	if (fp->f_seqcount > 1)
522 		fp->f_seqcount = 1;
523 	else
524 		fp->f_seqcount = 0;
525 	return (0);
526 }
527 
528 /*
529  * Package up an I/O request on a vnode into a uio and do it.
530  */
531 int
532 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
533     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
534     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
535 {
536 	struct uio auio;
537 	struct iovec aiov;
538 	struct mount *mp;
539 	struct ucred *cred;
540 	void *rl_cookie;
541 	struct vn_io_fault_args args;
542 	int error, lock_flags;
543 
544 	if (offset < 0 && vp->v_type != VCHR)
545 		return (EINVAL);
546 	auio.uio_iov = &aiov;
547 	auio.uio_iovcnt = 1;
548 	aiov.iov_base = base;
549 	aiov.iov_len = len;
550 	auio.uio_resid = len;
551 	auio.uio_offset = offset;
552 	auio.uio_segflg = segflg;
553 	auio.uio_rw = rw;
554 	auio.uio_td = td;
555 	error = 0;
556 
557 	if ((ioflg & IO_NODELOCKED) == 0) {
558 		if ((ioflg & IO_RANGELOCKED) == 0) {
559 			if (rw == UIO_READ) {
560 				rl_cookie = vn_rangelock_rlock(vp, offset,
561 				    offset + len);
562 			} else {
563 				rl_cookie = vn_rangelock_wlock(vp, offset,
564 				    offset + len);
565 			}
566 		} else
567 			rl_cookie = NULL;
568 		mp = NULL;
569 		if (rw == UIO_WRITE) {
570 			if (vp->v_type != VCHR &&
571 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
572 			    != 0)
573 				goto out;
574 			if (MNT_SHARED_WRITES(mp) ||
575 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
576 				lock_flags = LK_SHARED;
577 			else
578 				lock_flags = LK_EXCLUSIVE;
579 		} else
580 			lock_flags = LK_SHARED;
581 		vn_lock(vp, lock_flags | LK_RETRY);
582 	} else
583 		rl_cookie = NULL;
584 
585 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
586 #ifdef MAC
587 	if ((ioflg & IO_NOMACCHECK) == 0) {
588 		if (rw == UIO_READ)
589 			error = mac_vnode_check_read(active_cred, file_cred,
590 			    vp);
591 		else
592 			error = mac_vnode_check_write(active_cred, file_cred,
593 			    vp);
594 	}
595 #endif
596 	if (error == 0) {
597 		if (file_cred != NULL)
598 			cred = file_cred;
599 		else
600 			cred = active_cred;
601 		if (do_vn_io_fault(vp, &auio)) {
602 			args.kind = VN_IO_FAULT_VOP;
603 			args.cred = cred;
604 			args.flags = ioflg;
605 			args.args.vop_args.vp = vp;
606 			error = vn_io_fault1(vp, &auio, &args, td);
607 		} else if (rw == UIO_READ) {
608 			error = VOP_READ(vp, &auio, ioflg, cred);
609 		} else /* if (rw == UIO_WRITE) */ {
610 			error = VOP_WRITE(vp, &auio, ioflg, cred);
611 		}
612 	}
613 	if (aresid)
614 		*aresid = auio.uio_resid;
615 	else
616 		if (auio.uio_resid && error == 0)
617 			error = EIO;
618 	if ((ioflg & IO_NODELOCKED) == 0) {
619 		VOP_UNLOCK(vp);
620 		if (mp != NULL)
621 			vn_finished_write(mp);
622 	}
623  out:
624 	if (rl_cookie != NULL)
625 		vn_rangelock_unlock(vp, rl_cookie);
626 	return (error);
627 }
628 
629 /*
630  * Package up an I/O request on a vnode into a uio and do it.  The I/O
631  * request is split up into smaller chunks and we try to avoid saturating
632  * the buffer cache while potentially holding a vnode locked, so we
633  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
634  * to give other processes a chance to lock the vnode (either other processes
635  * core'ing the same binary, or unrelated processes scanning the directory).
636  */
637 int
638 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
639     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
640     struct ucred *file_cred, size_t *aresid, struct thread *td)
641 {
642 	int error = 0;
643 	ssize_t iaresid;
644 
645 	do {
646 		int chunk;
647 
648 		/*
649 		 * Force `offset' to a multiple of MAXBSIZE except possibly
650 		 * for the first chunk, so that filesystems only need to
651 		 * write full blocks except possibly for the first and last
652 		 * chunks.
653 		 */
654 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
655 
656 		if (chunk > len)
657 			chunk = len;
658 		if (rw != UIO_READ && vp->v_type == VREG)
659 			bwillwrite();
660 		iaresid = 0;
661 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
662 		    ioflg, active_cred, file_cred, &iaresid, td);
663 		len -= chunk;	/* aresid calc already includes length */
664 		if (error)
665 			break;
666 		offset += chunk;
667 		base = (char *)base + chunk;
668 		kern_yield(PRI_USER);
669 	} while (len);
670 	if (aresid)
671 		*aresid = len + iaresid;
672 	return (error);
673 }
674 
675 #if OFF_MAX <= LONG_MAX
676 off_t
677 foffset_lock(struct file *fp, int flags)
678 {
679 	volatile short *flagsp;
680 	off_t res;
681 	short state;
682 
683 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
684 
685 	if ((flags & FOF_NOLOCK) != 0)
686 		return (atomic_load_long(&fp->f_offset));
687 
688 	/*
689 	 * According to McKusick the vn lock was protecting f_offset here.
690 	 * It is now protected by the FOFFSET_LOCKED flag.
691 	 */
692 	flagsp = &fp->f_vnread_flags;
693 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
694 		return (atomic_load_long(&fp->f_offset));
695 
696 	sleepq_lock(&fp->f_vnread_flags);
697 	state = atomic_load_16(flagsp);
698 	for (;;) {
699 		if ((state & FOFFSET_LOCKED) == 0) {
700 			if (!atomic_fcmpset_acq_16(flagsp, &state,
701 			    FOFFSET_LOCKED))
702 				continue;
703 			break;
704 		}
705 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
706 			if (!atomic_fcmpset_acq_16(flagsp, &state,
707 			    state | FOFFSET_LOCK_WAITING))
708 				continue;
709 		}
710 		DROP_GIANT();
711 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
712 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
713 		PICKUP_GIANT();
714 		sleepq_lock(&fp->f_vnread_flags);
715 		state = atomic_load_16(flagsp);
716 	}
717 	res = atomic_load_long(&fp->f_offset);
718 	sleepq_release(&fp->f_vnread_flags);
719 	return (res);
720 }
721 
722 void
723 foffset_unlock(struct file *fp, off_t val, int flags)
724 {
725 	volatile short *flagsp;
726 	short state;
727 
728 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
729 
730 	if ((flags & FOF_NOUPDATE) == 0)
731 		atomic_store_long(&fp->f_offset, val);
732 	if ((flags & FOF_NEXTOFF) != 0)
733 		fp->f_nextoff = val;
734 
735 	if ((flags & FOF_NOLOCK) != 0)
736 		return;
737 
738 	flagsp = &fp->f_vnread_flags;
739 	state = atomic_load_16(flagsp);
740 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
741 	    atomic_cmpset_rel_16(flagsp, state, 0))
742 		return;
743 
744 	sleepq_lock(&fp->f_vnread_flags);
745 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
746 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
747 	fp->f_vnread_flags = 0;
748 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
749 	sleepq_release(&fp->f_vnread_flags);
750 }
751 #else
752 off_t
753 foffset_lock(struct file *fp, int flags)
754 {
755 	struct mtx *mtxp;
756 	off_t res;
757 
758 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
759 
760 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
761 	mtx_lock(mtxp);
762 	if ((flags & FOF_NOLOCK) == 0) {
763 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
764 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
765 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
766 			    "vofflock", 0);
767 		}
768 		fp->f_vnread_flags |= FOFFSET_LOCKED;
769 	}
770 	res = fp->f_offset;
771 	mtx_unlock(mtxp);
772 	return (res);
773 }
774 
775 void
776 foffset_unlock(struct file *fp, off_t val, int flags)
777 {
778 	struct mtx *mtxp;
779 
780 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
781 
782 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
783 	mtx_lock(mtxp);
784 	if ((flags & FOF_NOUPDATE) == 0)
785 		fp->f_offset = val;
786 	if ((flags & FOF_NEXTOFF) != 0)
787 		fp->f_nextoff = val;
788 	if ((flags & FOF_NOLOCK) == 0) {
789 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
790 		    ("Lost FOFFSET_LOCKED"));
791 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
792 			wakeup(&fp->f_vnread_flags);
793 		fp->f_vnread_flags = 0;
794 	}
795 	mtx_unlock(mtxp);
796 }
797 #endif
798 
799 void
800 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
801 {
802 
803 	if ((flags & FOF_OFFSET) == 0)
804 		uio->uio_offset = foffset_lock(fp, flags);
805 }
806 
807 void
808 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
809 {
810 
811 	if ((flags & FOF_OFFSET) == 0)
812 		foffset_unlock(fp, uio->uio_offset, flags);
813 }
814 
815 static int
816 get_advice(struct file *fp, struct uio *uio)
817 {
818 	struct mtx *mtxp;
819 	int ret;
820 
821 	ret = POSIX_FADV_NORMAL;
822 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
823 		return (ret);
824 
825 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
826 	mtx_lock(mtxp);
827 	if (fp->f_advice != NULL &&
828 	    uio->uio_offset >= fp->f_advice->fa_start &&
829 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
830 		ret = fp->f_advice->fa_advice;
831 	mtx_unlock(mtxp);
832 	return (ret);
833 }
834 
835 /*
836  * File table vnode read routine.
837  */
838 static int
839 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
840     struct thread *td)
841 {
842 	struct vnode *vp;
843 	off_t orig_offset;
844 	int error, ioflag;
845 	int advice;
846 
847 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
848 	    uio->uio_td, td));
849 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
850 	vp = fp->f_vnode;
851 	ioflag = 0;
852 	if (fp->f_flag & FNONBLOCK)
853 		ioflag |= IO_NDELAY;
854 	if (fp->f_flag & O_DIRECT)
855 		ioflag |= IO_DIRECT;
856 	advice = get_advice(fp, uio);
857 	vn_lock(vp, LK_SHARED | LK_RETRY);
858 
859 	switch (advice) {
860 	case POSIX_FADV_NORMAL:
861 	case POSIX_FADV_SEQUENTIAL:
862 	case POSIX_FADV_NOREUSE:
863 		ioflag |= sequential_heuristic(uio, fp);
864 		break;
865 	case POSIX_FADV_RANDOM:
866 		/* Disable read-ahead for random I/O. */
867 		break;
868 	}
869 	orig_offset = uio->uio_offset;
870 
871 #ifdef MAC
872 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
873 	if (error == 0)
874 #endif
875 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
876 	fp->f_nextoff = uio->uio_offset;
877 	VOP_UNLOCK(vp);
878 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
879 	    orig_offset != uio->uio_offset)
880 		/*
881 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
882 		 * for the backing file after a POSIX_FADV_NOREUSE
883 		 * read(2).
884 		 */
885 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
886 		    POSIX_FADV_DONTNEED);
887 	return (error);
888 }
889 
890 /*
891  * File table vnode write routine.
892  */
893 static int
894 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
895     struct thread *td)
896 {
897 	struct vnode *vp;
898 	struct mount *mp;
899 	off_t orig_offset;
900 	int error, ioflag, lock_flags;
901 	int advice;
902 
903 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
904 	    uio->uio_td, td));
905 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
906 	vp = fp->f_vnode;
907 	if (vp->v_type == VREG)
908 		bwillwrite();
909 	ioflag = IO_UNIT;
910 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
911 		ioflag |= IO_APPEND;
912 	if (fp->f_flag & FNONBLOCK)
913 		ioflag |= IO_NDELAY;
914 	if (fp->f_flag & O_DIRECT)
915 		ioflag |= IO_DIRECT;
916 	if ((fp->f_flag & O_FSYNC) ||
917 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
918 		ioflag |= IO_SYNC;
919 	mp = NULL;
920 	if (vp->v_type != VCHR &&
921 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
922 		goto unlock;
923 
924 	advice = get_advice(fp, uio);
925 
926 	if (MNT_SHARED_WRITES(mp) ||
927 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
928 		lock_flags = LK_SHARED;
929 	} else {
930 		lock_flags = LK_EXCLUSIVE;
931 	}
932 
933 	vn_lock(vp, lock_flags | LK_RETRY);
934 	switch (advice) {
935 	case POSIX_FADV_NORMAL:
936 	case POSIX_FADV_SEQUENTIAL:
937 	case POSIX_FADV_NOREUSE:
938 		ioflag |= sequential_heuristic(uio, fp);
939 		break;
940 	case POSIX_FADV_RANDOM:
941 		/* XXX: Is this correct? */
942 		break;
943 	}
944 	orig_offset = uio->uio_offset;
945 
946 #ifdef MAC
947 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
948 	if (error == 0)
949 #endif
950 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
951 	fp->f_nextoff = uio->uio_offset;
952 	VOP_UNLOCK(vp);
953 	if (vp->v_type != VCHR)
954 		vn_finished_write(mp);
955 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
956 	    orig_offset != uio->uio_offset)
957 		/*
958 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
959 		 * for the backing file after a POSIX_FADV_NOREUSE
960 		 * write(2).
961 		 */
962 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
963 		    POSIX_FADV_DONTNEED);
964 unlock:
965 	return (error);
966 }
967 
968 /*
969  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
970  * prevent the following deadlock:
971  *
972  * Assume that the thread A reads from the vnode vp1 into userspace
973  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
974  * currently not resident, then system ends up with the call chain
975  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
976  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
977  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
978  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
979  * backed by the pages of vnode vp1, and some page in buf2 is not
980  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
981  *
982  * To prevent the lock order reversal and deadlock, vn_io_fault() does
983  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
984  * Instead, it first tries to do the whole range i/o with pagefaults
985  * disabled. If all pages in the i/o buffer are resident and mapped,
986  * VOP will succeed (ignoring the genuine filesystem errors).
987  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
988  * i/o in chunks, with all pages in the chunk prefaulted and held
989  * using vm_fault_quick_hold_pages().
990  *
991  * Filesystems using this deadlock avoidance scheme should use the
992  * array of the held pages from uio, saved in the curthread->td_ma,
993  * instead of doing uiomove().  A helper function
994  * vn_io_fault_uiomove() converts uiomove request into
995  * uiomove_fromphys() over td_ma array.
996  *
997  * Since vnode locks do not cover the whole i/o anymore, rangelocks
998  * make the current i/o request atomic with respect to other i/os and
999  * truncations.
1000  */
1001 
1002 /*
1003  * Decode vn_io_fault_args and perform the corresponding i/o.
1004  */
1005 static int
1006 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1007     struct thread *td)
1008 {
1009 	int error, save;
1010 
1011 	error = 0;
1012 	save = vm_fault_disable_pagefaults();
1013 	switch (args->kind) {
1014 	case VN_IO_FAULT_FOP:
1015 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1016 		    uio, args->cred, args->flags, td);
1017 		break;
1018 	case VN_IO_FAULT_VOP:
1019 		if (uio->uio_rw == UIO_READ) {
1020 			error = VOP_READ(args->args.vop_args.vp, uio,
1021 			    args->flags, args->cred);
1022 		} else if (uio->uio_rw == UIO_WRITE) {
1023 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1024 			    args->flags, args->cred);
1025 		}
1026 		break;
1027 	default:
1028 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1029 		    args->kind, uio->uio_rw);
1030 	}
1031 	vm_fault_enable_pagefaults(save);
1032 	return (error);
1033 }
1034 
1035 static int
1036 vn_io_fault_touch(char *base, const struct uio *uio)
1037 {
1038 	int r;
1039 
1040 	r = fubyte(base);
1041 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1042 		return (EFAULT);
1043 	return (0);
1044 }
1045 
1046 static int
1047 vn_io_fault_prefault_user(const struct uio *uio)
1048 {
1049 	char *base;
1050 	const struct iovec *iov;
1051 	size_t len;
1052 	ssize_t resid;
1053 	int error, i;
1054 
1055 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1056 	    ("vn_io_fault_prefault userspace"));
1057 
1058 	error = i = 0;
1059 	iov = uio->uio_iov;
1060 	resid = uio->uio_resid;
1061 	base = iov->iov_base;
1062 	len = iov->iov_len;
1063 	while (resid > 0) {
1064 		error = vn_io_fault_touch(base, uio);
1065 		if (error != 0)
1066 			break;
1067 		if (len < PAGE_SIZE) {
1068 			if (len != 0) {
1069 				error = vn_io_fault_touch(base + len - 1, uio);
1070 				if (error != 0)
1071 					break;
1072 				resid -= len;
1073 			}
1074 			if (++i >= uio->uio_iovcnt)
1075 				break;
1076 			iov = uio->uio_iov + i;
1077 			base = iov->iov_base;
1078 			len = iov->iov_len;
1079 		} else {
1080 			len -= PAGE_SIZE;
1081 			base += PAGE_SIZE;
1082 			resid -= PAGE_SIZE;
1083 		}
1084 	}
1085 	return (error);
1086 }
1087 
1088 /*
1089  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1090  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1091  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1092  * into args and call vn_io_fault1() to handle faults during the user
1093  * mode buffer accesses.
1094  */
1095 static int
1096 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1097     struct thread *td)
1098 {
1099 	vm_page_t ma[io_hold_cnt + 2];
1100 	struct uio *uio_clone, short_uio;
1101 	struct iovec short_iovec[1];
1102 	vm_page_t *prev_td_ma;
1103 	vm_prot_t prot;
1104 	vm_offset_t addr, end;
1105 	size_t len, resid;
1106 	ssize_t adv;
1107 	int error, cnt, saveheld, prev_td_ma_cnt;
1108 
1109 	if (vn_io_fault_prefault) {
1110 		error = vn_io_fault_prefault_user(uio);
1111 		if (error != 0)
1112 			return (error); /* Or ignore ? */
1113 	}
1114 
1115 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1116 
1117 	/*
1118 	 * The UFS follows IO_UNIT directive and replays back both
1119 	 * uio_offset and uio_resid if an error is encountered during the
1120 	 * operation.  But, since the iovec may be already advanced,
1121 	 * uio is still in an inconsistent state.
1122 	 *
1123 	 * Cache a copy of the original uio, which is advanced to the redo
1124 	 * point using UIO_NOCOPY below.
1125 	 */
1126 	uio_clone = cloneuio(uio);
1127 	resid = uio->uio_resid;
1128 
1129 	short_uio.uio_segflg = UIO_USERSPACE;
1130 	short_uio.uio_rw = uio->uio_rw;
1131 	short_uio.uio_td = uio->uio_td;
1132 
1133 	error = vn_io_fault_doio(args, uio, td);
1134 	if (error != EFAULT)
1135 		goto out;
1136 
1137 	atomic_add_long(&vn_io_faults_cnt, 1);
1138 	uio_clone->uio_segflg = UIO_NOCOPY;
1139 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1140 	uio_clone->uio_segflg = uio->uio_segflg;
1141 
1142 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1143 	prev_td_ma = td->td_ma;
1144 	prev_td_ma_cnt = td->td_ma_cnt;
1145 
1146 	while (uio_clone->uio_resid != 0) {
1147 		len = uio_clone->uio_iov->iov_len;
1148 		if (len == 0) {
1149 			KASSERT(uio_clone->uio_iovcnt >= 1,
1150 			    ("iovcnt underflow"));
1151 			uio_clone->uio_iov++;
1152 			uio_clone->uio_iovcnt--;
1153 			continue;
1154 		}
1155 		if (len > io_hold_cnt * PAGE_SIZE)
1156 			len = io_hold_cnt * PAGE_SIZE;
1157 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1158 		end = round_page(addr + len);
1159 		if (end < addr) {
1160 			error = EFAULT;
1161 			break;
1162 		}
1163 		cnt = atop(end - trunc_page(addr));
1164 		/*
1165 		 * A perfectly misaligned address and length could cause
1166 		 * both the start and the end of the chunk to use partial
1167 		 * page.  +2 accounts for such a situation.
1168 		 */
1169 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1170 		    addr, len, prot, ma, io_hold_cnt + 2);
1171 		if (cnt == -1) {
1172 			error = EFAULT;
1173 			break;
1174 		}
1175 		short_uio.uio_iov = &short_iovec[0];
1176 		short_iovec[0].iov_base = (void *)addr;
1177 		short_uio.uio_iovcnt = 1;
1178 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1179 		short_uio.uio_offset = uio_clone->uio_offset;
1180 		td->td_ma = ma;
1181 		td->td_ma_cnt = cnt;
1182 
1183 		error = vn_io_fault_doio(args, &short_uio, td);
1184 		vm_page_unhold_pages(ma, cnt);
1185 		adv = len - short_uio.uio_resid;
1186 
1187 		uio_clone->uio_iov->iov_base =
1188 		    (char *)uio_clone->uio_iov->iov_base + adv;
1189 		uio_clone->uio_iov->iov_len -= adv;
1190 		uio_clone->uio_resid -= adv;
1191 		uio_clone->uio_offset += adv;
1192 
1193 		uio->uio_resid -= adv;
1194 		uio->uio_offset += adv;
1195 
1196 		if (error != 0 || adv == 0)
1197 			break;
1198 	}
1199 	td->td_ma = prev_td_ma;
1200 	td->td_ma_cnt = prev_td_ma_cnt;
1201 	curthread_pflags_restore(saveheld);
1202 out:
1203 	free(uio_clone, M_IOV);
1204 	return (error);
1205 }
1206 
1207 static int
1208 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1209     int flags, struct thread *td)
1210 {
1211 	fo_rdwr_t *doio;
1212 	struct vnode *vp;
1213 	void *rl_cookie;
1214 	struct vn_io_fault_args args;
1215 	int error;
1216 
1217 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1218 	vp = fp->f_vnode;
1219 	foffset_lock_uio(fp, uio, flags);
1220 	if (do_vn_io_fault(vp, uio)) {
1221 		args.kind = VN_IO_FAULT_FOP;
1222 		args.args.fop_args.fp = fp;
1223 		args.args.fop_args.doio = doio;
1224 		args.cred = active_cred;
1225 		args.flags = flags | FOF_OFFSET;
1226 		if (uio->uio_rw == UIO_READ) {
1227 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1228 			    uio->uio_offset + uio->uio_resid);
1229 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1230 		    (flags & FOF_OFFSET) == 0) {
1231 			/* For appenders, punt and lock the whole range. */
1232 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1233 		} else {
1234 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1235 			    uio->uio_offset + uio->uio_resid);
1236 		}
1237 		error = vn_io_fault1(vp, uio, &args, td);
1238 		vn_rangelock_unlock(vp, rl_cookie);
1239 	} else {
1240 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1241 	}
1242 	foffset_unlock_uio(fp, uio, flags);
1243 	return (error);
1244 }
1245 
1246 /*
1247  * Helper function to perform the requested uiomove operation using
1248  * the held pages for io->uio_iov[0].iov_base buffer instead of
1249  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1250  * instead of iov_base prevents page faults that could occur due to
1251  * pmap_collect() invalidating the mapping created by
1252  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1253  * object cleanup revoking the write access from page mappings.
1254  *
1255  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1256  * instead of plain uiomove().
1257  */
1258 int
1259 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1260 {
1261 	struct uio transp_uio;
1262 	struct iovec transp_iov[1];
1263 	struct thread *td;
1264 	size_t adv;
1265 	int error, pgadv;
1266 
1267 	td = curthread;
1268 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1269 	    uio->uio_segflg != UIO_USERSPACE)
1270 		return (uiomove(data, xfersize, uio));
1271 
1272 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1273 	transp_iov[0].iov_base = data;
1274 	transp_uio.uio_iov = &transp_iov[0];
1275 	transp_uio.uio_iovcnt = 1;
1276 	if (xfersize > uio->uio_resid)
1277 		xfersize = uio->uio_resid;
1278 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1279 	transp_uio.uio_offset = 0;
1280 	transp_uio.uio_segflg = UIO_SYSSPACE;
1281 	/*
1282 	 * Since transp_iov points to data, and td_ma page array
1283 	 * corresponds to original uio->uio_iov, we need to invert the
1284 	 * direction of the i/o operation as passed to
1285 	 * uiomove_fromphys().
1286 	 */
1287 	switch (uio->uio_rw) {
1288 	case UIO_WRITE:
1289 		transp_uio.uio_rw = UIO_READ;
1290 		break;
1291 	case UIO_READ:
1292 		transp_uio.uio_rw = UIO_WRITE;
1293 		break;
1294 	}
1295 	transp_uio.uio_td = uio->uio_td;
1296 	error = uiomove_fromphys(td->td_ma,
1297 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1298 	    xfersize, &transp_uio);
1299 	adv = xfersize - transp_uio.uio_resid;
1300 	pgadv =
1301 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1302 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1303 	td->td_ma += pgadv;
1304 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1305 	    pgadv));
1306 	td->td_ma_cnt -= pgadv;
1307 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1308 	uio->uio_iov->iov_len -= adv;
1309 	uio->uio_resid -= adv;
1310 	uio->uio_offset += adv;
1311 	return (error);
1312 }
1313 
1314 int
1315 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1316     struct uio *uio)
1317 {
1318 	struct thread *td;
1319 	vm_offset_t iov_base;
1320 	int cnt, pgadv;
1321 
1322 	td = curthread;
1323 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1324 	    uio->uio_segflg != UIO_USERSPACE)
1325 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1326 
1327 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1328 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1329 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1330 	switch (uio->uio_rw) {
1331 	case UIO_WRITE:
1332 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1333 		    offset, cnt);
1334 		break;
1335 	case UIO_READ:
1336 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1337 		    cnt);
1338 		break;
1339 	}
1340 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1341 	td->td_ma += pgadv;
1342 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1343 	    pgadv));
1344 	td->td_ma_cnt -= pgadv;
1345 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1346 	uio->uio_iov->iov_len -= cnt;
1347 	uio->uio_resid -= cnt;
1348 	uio->uio_offset += cnt;
1349 	return (0);
1350 }
1351 
1352 /*
1353  * File table truncate routine.
1354  */
1355 static int
1356 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1357     struct thread *td)
1358 {
1359 	struct mount *mp;
1360 	struct vnode *vp;
1361 	void *rl_cookie;
1362 	int error;
1363 
1364 	vp = fp->f_vnode;
1365 
1366 	/*
1367 	 * Lock the whole range for truncation.  Otherwise split i/o
1368 	 * might happen partly before and partly after the truncation.
1369 	 */
1370 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1371 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1372 	if (error)
1373 		goto out1;
1374 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1375 	AUDIT_ARG_VNODE1(vp);
1376 	if (vp->v_type == VDIR) {
1377 		error = EISDIR;
1378 		goto out;
1379 	}
1380 #ifdef MAC
1381 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1382 	if (error)
1383 		goto out;
1384 #endif
1385 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1386 	    fp->f_cred);
1387 out:
1388 	VOP_UNLOCK(vp);
1389 	vn_finished_write(mp);
1390 out1:
1391 	vn_rangelock_unlock(vp, rl_cookie);
1392 	return (error);
1393 }
1394 
1395 /*
1396  * Truncate a file that is already locked.
1397  */
1398 int
1399 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1400     struct ucred *cred)
1401 {
1402 	struct vattr vattr;
1403 	int error;
1404 
1405 	error = VOP_ADD_WRITECOUNT(vp, 1);
1406 	if (error == 0) {
1407 		VATTR_NULL(&vattr);
1408 		vattr.va_size = length;
1409 		if (sync)
1410 			vattr.va_vaflags |= VA_SYNC;
1411 		error = VOP_SETATTR(vp, &vattr, cred);
1412 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1413 	}
1414 	return (error);
1415 }
1416 
1417 /*
1418  * File table vnode stat routine.
1419  */
1420 static int
1421 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1422     struct thread *td)
1423 {
1424 	struct vnode *vp = fp->f_vnode;
1425 	int error;
1426 
1427 	vn_lock(vp, LK_SHARED | LK_RETRY);
1428 	error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1429 	VOP_UNLOCK(vp);
1430 
1431 	return (error);
1432 }
1433 
1434 /*
1435  * Stat a vnode; implementation for the stat syscall
1436  */
1437 int
1438 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred,
1439     struct ucred *file_cred, struct thread *td)
1440 {
1441 	struct vattr vattr;
1442 	struct vattr *vap;
1443 	int error;
1444 	u_short mode;
1445 
1446 	AUDIT_ARG_VNODE1(vp);
1447 #ifdef MAC
1448 	error = mac_vnode_check_stat(active_cred, file_cred, vp);
1449 	if (error)
1450 		return (error);
1451 #endif
1452 
1453 	vap = &vattr;
1454 
1455 	/*
1456 	 * Initialize defaults for new and unusual fields, so that file
1457 	 * systems which don't support these fields don't need to know
1458 	 * about them.
1459 	 */
1460 	vap->va_birthtime.tv_sec = -1;
1461 	vap->va_birthtime.tv_nsec = 0;
1462 	vap->va_fsid = VNOVAL;
1463 	vap->va_rdev = NODEV;
1464 
1465 	error = VOP_GETATTR(vp, vap, active_cred);
1466 	if (error)
1467 		return (error);
1468 
1469 	/*
1470 	 * Zero the spare stat fields
1471 	 */
1472 	bzero(sb, sizeof *sb);
1473 
1474 	/*
1475 	 * Copy from vattr table
1476 	 */
1477 	if (vap->va_fsid != VNOVAL)
1478 		sb->st_dev = vap->va_fsid;
1479 	else
1480 		sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1481 	sb->st_ino = vap->va_fileid;
1482 	mode = vap->va_mode;
1483 	switch (vap->va_type) {
1484 	case VREG:
1485 		mode |= S_IFREG;
1486 		break;
1487 	case VDIR:
1488 		mode |= S_IFDIR;
1489 		break;
1490 	case VBLK:
1491 		mode |= S_IFBLK;
1492 		break;
1493 	case VCHR:
1494 		mode |= S_IFCHR;
1495 		break;
1496 	case VLNK:
1497 		mode |= S_IFLNK;
1498 		break;
1499 	case VSOCK:
1500 		mode |= S_IFSOCK;
1501 		break;
1502 	case VFIFO:
1503 		mode |= S_IFIFO;
1504 		break;
1505 	default:
1506 		return (EBADF);
1507 	}
1508 	sb->st_mode = mode;
1509 	sb->st_nlink = vap->va_nlink;
1510 	sb->st_uid = vap->va_uid;
1511 	sb->st_gid = vap->va_gid;
1512 	sb->st_rdev = vap->va_rdev;
1513 	if (vap->va_size > OFF_MAX)
1514 		return (EOVERFLOW);
1515 	sb->st_size = vap->va_size;
1516 	sb->st_atim.tv_sec = vap->va_atime.tv_sec;
1517 	sb->st_atim.tv_nsec = vap->va_atime.tv_nsec;
1518 	sb->st_mtim.tv_sec = vap->va_mtime.tv_sec;
1519 	sb->st_mtim.tv_nsec = vap->va_mtime.tv_nsec;
1520 	sb->st_ctim.tv_sec = vap->va_ctime.tv_sec;
1521 	sb->st_ctim.tv_nsec = vap->va_ctime.tv_nsec;
1522 	sb->st_birthtim.tv_sec = vap->va_birthtime.tv_sec;
1523 	sb->st_birthtim.tv_nsec = vap->va_birthtime.tv_nsec;
1524 
1525         /*
1526 	 * According to www.opengroup.org, the meaning of st_blksize is
1527 	 *   "a filesystem-specific preferred I/O block size for this
1528 	 *    object.  In some filesystem types, this may vary from file
1529 	 *    to file"
1530 	 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1531 	 */
1532 
1533 	sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1534 
1535 	sb->st_flags = vap->va_flags;
1536 	if (priv_check_cred_vfs_generation(td->td_ucred))
1537 		sb->st_gen = 0;
1538 	else
1539 		sb->st_gen = vap->va_gen;
1540 
1541 	sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1542 	return (0);
1543 }
1544 
1545 /*
1546  * File table vnode ioctl routine.
1547  */
1548 static int
1549 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1550     struct thread *td)
1551 {
1552 	struct vattr vattr;
1553 	struct vnode *vp;
1554 	struct fiobmap2_arg *bmarg;
1555 	int error;
1556 
1557 	vp = fp->f_vnode;
1558 	switch (vp->v_type) {
1559 	case VDIR:
1560 	case VREG:
1561 		switch (com) {
1562 		case FIONREAD:
1563 			vn_lock(vp, LK_SHARED | LK_RETRY);
1564 			error = VOP_GETATTR(vp, &vattr, active_cred);
1565 			VOP_UNLOCK(vp);
1566 			if (error == 0)
1567 				*(int *)data = vattr.va_size - fp->f_offset;
1568 			return (error);
1569 		case FIOBMAP2:
1570 			bmarg = (struct fiobmap2_arg *)data;
1571 			vn_lock(vp, LK_SHARED | LK_RETRY);
1572 #ifdef MAC
1573 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1574 			    vp);
1575 			if (error == 0)
1576 #endif
1577 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1578 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1579 			VOP_UNLOCK(vp);
1580 			return (error);
1581 		case FIONBIO:
1582 		case FIOASYNC:
1583 			return (0);
1584 		default:
1585 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1586 			    active_cred, td));
1587 		}
1588 		break;
1589 	case VCHR:
1590 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1591 		    active_cred, td));
1592 	default:
1593 		return (ENOTTY);
1594 	}
1595 }
1596 
1597 /*
1598  * File table vnode poll routine.
1599  */
1600 static int
1601 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1602     struct thread *td)
1603 {
1604 	struct vnode *vp;
1605 	int error;
1606 
1607 	vp = fp->f_vnode;
1608 #ifdef MAC
1609 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1610 	AUDIT_ARG_VNODE1(vp);
1611 	error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1612 	VOP_UNLOCK(vp);
1613 	if (!error)
1614 #endif
1615 
1616 	error = VOP_POLL(vp, events, fp->f_cred, td);
1617 	return (error);
1618 }
1619 
1620 /*
1621  * Acquire the requested lock and then check for validity.  LK_RETRY
1622  * permits vn_lock to return doomed vnodes.
1623  */
1624 static int __noinline
1625 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1626     int error)
1627 {
1628 
1629 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1630 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1631 
1632 	if (error == 0)
1633 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1634 
1635 	if ((flags & LK_RETRY) == 0) {
1636 		if (error == 0) {
1637 			VOP_UNLOCK(vp);
1638 			error = ENOENT;
1639 		}
1640 		return (error);
1641 	}
1642 
1643 	/*
1644 	 * LK_RETRY case.
1645 	 *
1646 	 * Nothing to do if we got the lock.
1647 	 */
1648 	if (error == 0)
1649 		return (0);
1650 
1651 	/*
1652 	 * Interlock was dropped by the call in _vn_lock.
1653 	 */
1654 	flags &= ~LK_INTERLOCK;
1655 	do {
1656 		error = VOP_LOCK1(vp, flags, file, line);
1657 	} while (error != 0);
1658 	return (0);
1659 }
1660 
1661 int
1662 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1663 {
1664 	int error;
1665 
1666 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1667 	    ("vn_lock: no locktype (%d passed)", flags));
1668 	VNPASS(vp->v_holdcnt > 0, vp);
1669 	error = VOP_LOCK1(vp, flags, file, line);
1670 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1671 		return (_vn_lock_fallback(vp, flags, file, line, error));
1672 	return (0);
1673 }
1674 
1675 /*
1676  * File table vnode close routine.
1677  */
1678 static int
1679 vn_closefile(struct file *fp, struct thread *td)
1680 {
1681 	struct vnode *vp;
1682 	struct flock lf;
1683 	int error;
1684 	bool ref;
1685 
1686 	vp = fp->f_vnode;
1687 	fp->f_ops = &badfileops;
1688 	ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1689 
1690 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1691 
1692 	if (__predict_false(ref)) {
1693 		lf.l_whence = SEEK_SET;
1694 		lf.l_start = 0;
1695 		lf.l_len = 0;
1696 		lf.l_type = F_UNLCK;
1697 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1698 		vrele(vp);
1699 	}
1700 	return (error);
1701 }
1702 
1703 /*
1704  * Preparing to start a filesystem write operation. If the operation is
1705  * permitted, then we bump the count of operations in progress and
1706  * proceed. If a suspend request is in progress, we wait until the
1707  * suspension is over, and then proceed.
1708  */
1709 static int
1710 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1711 {
1712 	int error, mflags;
1713 
1714 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1715 	    vfs_op_thread_enter(mp)) {
1716 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1717 		vfs_mp_count_add_pcpu(mp, writeopcount, 1);
1718 		vfs_op_thread_exit(mp);
1719 		return (0);
1720 	}
1721 
1722 	if (mplocked)
1723 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1724 	else
1725 		MNT_ILOCK(mp);
1726 
1727 	error = 0;
1728 
1729 	/*
1730 	 * Check on status of suspension.
1731 	 */
1732 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1733 	    mp->mnt_susp_owner != curthread) {
1734 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1735 		    (flags & PCATCH) : 0) | (PUSER - 1);
1736 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1737 			if (flags & V_NOWAIT) {
1738 				error = EWOULDBLOCK;
1739 				goto unlock;
1740 			}
1741 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1742 			    "suspfs", 0);
1743 			if (error)
1744 				goto unlock;
1745 		}
1746 	}
1747 	if (flags & V_XSLEEP)
1748 		goto unlock;
1749 	mp->mnt_writeopcount++;
1750 unlock:
1751 	if (error != 0 || (flags & V_XSLEEP) != 0)
1752 		MNT_REL(mp);
1753 	MNT_IUNLOCK(mp);
1754 	return (error);
1755 }
1756 
1757 int
1758 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1759 {
1760 	struct mount *mp;
1761 	int error;
1762 
1763 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1764 	    ("V_MNTREF requires mp"));
1765 
1766 	error = 0;
1767 	/*
1768 	 * If a vnode is provided, get and return the mount point that
1769 	 * to which it will write.
1770 	 */
1771 	if (vp != NULL) {
1772 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1773 			*mpp = NULL;
1774 			if (error != EOPNOTSUPP)
1775 				return (error);
1776 			return (0);
1777 		}
1778 	}
1779 	if ((mp = *mpp) == NULL)
1780 		return (0);
1781 
1782 	/*
1783 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1784 	 * a vfs_ref().
1785 	 * As long as a vnode is not provided we need to acquire a
1786 	 * refcount for the provided mountpoint too, in order to
1787 	 * emulate a vfs_ref().
1788 	 */
1789 	if (vp == NULL && (flags & V_MNTREF) == 0)
1790 		vfs_ref(mp);
1791 
1792 	return (vn_start_write_refed(mp, flags, false));
1793 }
1794 
1795 /*
1796  * Secondary suspension. Used by operations such as vop_inactive
1797  * routines that are needed by the higher level functions. These
1798  * are allowed to proceed until all the higher level functions have
1799  * completed (indicated by mnt_writeopcount dropping to zero). At that
1800  * time, these operations are halted until the suspension is over.
1801  */
1802 int
1803 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1804 {
1805 	struct mount *mp;
1806 	int error;
1807 
1808 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1809 	    ("V_MNTREF requires mp"));
1810 
1811  retry:
1812 	if (vp != NULL) {
1813 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1814 			*mpp = NULL;
1815 			if (error != EOPNOTSUPP)
1816 				return (error);
1817 			return (0);
1818 		}
1819 	}
1820 	/*
1821 	 * If we are not suspended or have not yet reached suspended
1822 	 * mode, then let the operation proceed.
1823 	 */
1824 	if ((mp = *mpp) == NULL)
1825 		return (0);
1826 
1827 	/*
1828 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1829 	 * a vfs_ref().
1830 	 * As long as a vnode is not provided we need to acquire a
1831 	 * refcount for the provided mountpoint too, in order to
1832 	 * emulate a vfs_ref().
1833 	 */
1834 	MNT_ILOCK(mp);
1835 	if (vp == NULL && (flags & V_MNTREF) == 0)
1836 		MNT_REF(mp);
1837 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1838 		mp->mnt_secondary_writes++;
1839 		mp->mnt_secondary_accwrites++;
1840 		MNT_IUNLOCK(mp);
1841 		return (0);
1842 	}
1843 	if (flags & V_NOWAIT) {
1844 		MNT_REL(mp);
1845 		MNT_IUNLOCK(mp);
1846 		return (EWOULDBLOCK);
1847 	}
1848 	/*
1849 	 * Wait for the suspension to finish.
1850 	 */
1851 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1852 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1853 	    "suspfs", 0);
1854 	vfs_rel(mp);
1855 	if (error == 0)
1856 		goto retry;
1857 	return (error);
1858 }
1859 
1860 /*
1861  * Filesystem write operation has completed. If we are suspending and this
1862  * operation is the last one, notify the suspender that the suspension is
1863  * now in effect.
1864  */
1865 void
1866 vn_finished_write(struct mount *mp)
1867 {
1868 	int c;
1869 
1870 	if (mp == NULL)
1871 		return;
1872 
1873 	if (vfs_op_thread_enter(mp)) {
1874 		vfs_mp_count_sub_pcpu(mp, writeopcount, 1);
1875 		vfs_mp_count_sub_pcpu(mp, ref, 1);
1876 		vfs_op_thread_exit(mp);
1877 		return;
1878 	}
1879 
1880 	MNT_ILOCK(mp);
1881 	vfs_assert_mount_counters(mp);
1882 	MNT_REL(mp);
1883 	c = --mp->mnt_writeopcount;
1884 	if (mp->mnt_vfs_ops == 0) {
1885 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1886 		MNT_IUNLOCK(mp);
1887 		return;
1888 	}
1889 	if (c < 0)
1890 		vfs_dump_mount_counters(mp);
1891 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1892 		wakeup(&mp->mnt_writeopcount);
1893 	MNT_IUNLOCK(mp);
1894 }
1895 
1896 /*
1897  * Filesystem secondary write operation has completed. If we are
1898  * suspending and this operation is the last one, notify the suspender
1899  * that the suspension is now in effect.
1900  */
1901 void
1902 vn_finished_secondary_write(struct mount *mp)
1903 {
1904 	if (mp == NULL)
1905 		return;
1906 	MNT_ILOCK(mp);
1907 	MNT_REL(mp);
1908 	mp->mnt_secondary_writes--;
1909 	if (mp->mnt_secondary_writes < 0)
1910 		panic("vn_finished_secondary_write: neg cnt");
1911 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1912 	    mp->mnt_secondary_writes <= 0)
1913 		wakeup(&mp->mnt_secondary_writes);
1914 	MNT_IUNLOCK(mp);
1915 }
1916 
1917 /*
1918  * Request a filesystem to suspend write operations.
1919  */
1920 int
1921 vfs_write_suspend(struct mount *mp, int flags)
1922 {
1923 	int error;
1924 
1925 	vfs_op_enter(mp);
1926 
1927 	MNT_ILOCK(mp);
1928 	vfs_assert_mount_counters(mp);
1929 	if (mp->mnt_susp_owner == curthread) {
1930 		vfs_op_exit_locked(mp);
1931 		MNT_IUNLOCK(mp);
1932 		return (EALREADY);
1933 	}
1934 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
1935 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1936 
1937 	/*
1938 	 * Unmount holds a write reference on the mount point.  If we
1939 	 * own busy reference and drain for writers, we deadlock with
1940 	 * the reference draining in the unmount path.  Callers of
1941 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1942 	 * vfs_busy() reference is owned and caller is not in the
1943 	 * unmount context.
1944 	 */
1945 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1946 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1947 		vfs_op_exit_locked(mp);
1948 		MNT_IUNLOCK(mp);
1949 		return (EBUSY);
1950 	}
1951 
1952 	mp->mnt_kern_flag |= MNTK_SUSPEND;
1953 	mp->mnt_susp_owner = curthread;
1954 	if (mp->mnt_writeopcount > 0)
1955 		(void) msleep(&mp->mnt_writeopcount,
1956 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1957 	else
1958 		MNT_IUNLOCK(mp);
1959 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
1960 		vfs_write_resume(mp, 0);
1961 		/* vfs_write_resume does vfs_op_exit() for us */
1962 	}
1963 	return (error);
1964 }
1965 
1966 /*
1967  * Request a filesystem to resume write operations.
1968  */
1969 void
1970 vfs_write_resume(struct mount *mp, int flags)
1971 {
1972 
1973 	MNT_ILOCK(mp);
1974 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1975 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1976 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1977 				       MNTK_SUSPENDED);
1978 		mp->mnt_susp_owner = NULL;
1979 		wakeup(&mp->mnt_writeopcount);
1980 		wakeup(&mp->mnt_flag);
1981 		curthread->td_pflags &= ~TDP_IGNSUSP;
1982 		if ((flags & VR_START_WRITE) != 0) {
1983 			MNT_REF(mp);
1984 			mp->mnt_writeopcount++;
1985 		}
1986 		MNT_IUNLOCK(mp);
1987 		if ((flags & VR_NO_SUSPCLR) == 0)
1988 			VFS_SUSP_CLEAN(mp);
1989 		vfs_op_exit(mp);
1990 	} else if ((flags & VR_START_WRITE) != 0) {
1991 		MNT_REF(mp);
1992 		vn_start_write_refed(mp, 0, true);
1993 	} else {
1994 		MNT_IUNLOCK(mp);
1995 	}
1996 }
1997 
1998 /*
1999  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2000  * methods.
2001  */
2002 int
2003 vfs_write_suspend_umnt(struct mount *mp)
2004 {
2005 	int error;
2006 
2007 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2008 	    ("vfs_write_suspend_umnt: recursed"));
2009 
2010 	/* dounmount() already called vn_start_write(). */
2011 	for (;;) {
2012 		vn_finished_write(mp);
2013 		error = vfs_write_suspend(mp, 0);
2014 		if (error != 0) {
2015 			vn_start_write(NULL, &mp, V_WAIT);
2016 			return (error);
2017 		}
2018 		MNT_ILOCK(mp);
2019 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2020 			break;
2021 		MNT_IUNLOCK(mp);
2022 		vn_start_write(NULL, &mp, V_WAIT);
2023 	}
2024 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2025 	wakeup(&mp->mnt_flag);
2026 	MNT_IUNLOCK(mp);
2027 	curthread->td_pflags |= TDP_IGNSUSP;
2028 	return (0);
2029 }
2030 
2031 /*
2032  * Implement kqueues for files by translating it to vnode operation.
2033  */
2034 static int
2035 vn_kqfilter(struct file *fp, struct knote *kn)
2036 {
2037 
2038 	return (VOP_KQFILTER(fp->f_vnode, kn));
2039 }
2040 
2041 /*
2042  * Simplified in-kernel wrapper calls for extended attribute access.
2043  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2044  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2045  */
2046 int
2047 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2048     const char *attrname, int *buflen, char *buf, struct thread *td)
2049 {
2050 	struct uio	auio;
2051 	struct iovec	iov;
2052 	int	error;
2053 
2054 	iov.iov_len = *buflen;
2055 	iov.iov_base = buf;
2056 
2057 	auio.uio_iov = &iov;
2058 	auio.uio_iovcnt = 1;
2059 	auio.uio_rw = UIO_READ;
2060 	auio.uio_segflg = UIO_SYSSPACE;
2061 	auio.uio_td = td;
2062 	auio.uio_offset = 0;
2063 	auio.uio_resid = *buflen;
2064 
2065 	if ((ioflg & IO_NODELOCKED) == 0)
2066 		vn_lock(vp, LK_SHARED | LK_RETRY);
2067 
2068 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2069 
2070 	/* authorize attribute retrieval as kernel */
2071 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2072 	    td);
2073 
2074 	if ((ioflg & IO_NODELOCKED) == 0)
2075 		VOP_UNLOCK(vp);
2076 
2077 	if (error == 0) {
2078 		*buflen = *buflen - auio.uio_resid;
2079 	}
2080 
2081 	return (error);
2082 }
2083 
2084 /*
2085  * XXX failure mode if partially written?
2086  */
2087 int
2088 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2089     const char *attrname, int buflen, char *buf, struct thread *td)
2090 {
2091 	struct uio	auio;
2092 	struct iovec	iov;
2093 	struct mount	*mp;
2094 	int	error;
2095 
2096 	iov.iov_len = buflen;
2097 	iov.iov_base = buf;
2098 
2099 	auio.uio_iov = &iov;
2100 	auio.uio_iovcnt = 1;
2101 	auio.uio_rw = UIO_WRITE;
2102 	auio.uio_segflg = UIO_SYSSPACE;
2103 	auio.uio_td = td;
2104 	auio.uio_offset = 0;
2105 	auio.uio_resid = buflen;
2106 
2107 	if ((ioflg & IO_NODELOCKED) == 0) {
2108 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2109 			return (error);
2110 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2111 	}
2112 
2113 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2114 
2115 	/* authorize attribute setting as kernel */
2116 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2117 
2118 	if ((ioflg & IO_NODELOCKED) == 0) {
2119 		vn_finished_write(mp);
2120 		VOP_UNLOCK(vp);
2121 	}
2122 
2123 	return (error);
2124 }
2125 
2126 int
2127 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2128     const char *attrname, struct thread *td)
2129 {
2130 	struct mount	*mp;
2131 	int	error;
2132 
2133 	if ((ioflg & IO_NODELOCKED) == 0) {
2134 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2135 			return (error);
2136 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2137 	}
2138 
2139 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2140 
2141 	/* authorize attribute removal as kernel */
2142 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2143 	if (error == EOPNOTSUPP)
2144 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2145 		    NULL, td);
2146 
2147 	if ((ioflg & IO_NODELOCKED) == 0) {
2148 		vn_finished_write(mp);
2149 		VOP_UNLOCK(vp);
2150 	}
2151 
2152 	return (error);
2153 }
2154 
2155 static int
2156 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2157     struct vnode **rvp)
2158 {
2159 
2160 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2161 }
2162 
2163 int
2164 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2165 {
2166 
2167 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2168 	    lkflags, rvp));
2169 }
2170 
2171 int
2172 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2173     int lkflags, struct vnode **rvp)
2174 {
2175 	struct mount *mp;
2176 	int ltype, error;
2177 
2178 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2179 	mp = vp->v_mount;
2180 	ltype = VOP_ISLOCKED(vp);
2181 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2182 	    ("vn_vget_ino: vp not locked"));
2183 	error = vfs_busy(mp, MBF_NOWAIT);
2184 	if (error != 0) {
2185 		vfs_ref(mp);
2186 		VOP_UNLOCK(vp);
2187 		error = vfs_busy(mp, 0);
2188 		vn_lock(vp, ltype | LK_RETRY);
2189 		vfs_rel(mp);
2190 		if (error != 0)
2191 			return (ENOENT);
2192 		if (VN_IS_DOOMED(vp)) {
2193 			vfs_unbusy(mp);
2194 			return (ENOENT);
2195 		}
2196 	}
2197 	VOP_UNLOCK(vp);
2198 	error = alloc(mp, alloc_arg, lkflags, rvp);
2199 	vfs_unbusy(mp);
2200 	if (error != 0 || *rvp != vp)
2201 		vn_lock(vp, ltype | LK_RETRY);
2202 	if (VN_IS_DOOMED(vp)) {
2203 		if (error == 0) {
2204 			if (*rvp == vp)
2205 				vunref(vp);
2206 			else
2207 				vput(*rvp);
2208 		}
2209 		error = ENOENT;
2210 	}
2211 	return (error);
2212 }
2213 
2214 int
2215 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2216     struct thread *td)
2217 {
2218 
2219 	if (vp->v_type != VREG || td == NULL)
2220 		return (0);
2221 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2222 	    lim_cur(td, RLIMIT_FSIZE)) {
2223 		PROC_LOCK(td->td_proc);
2224 		kern_psignal(td->td_proc, SIGXFSZ);
2225 		PROC_UNLOCK(td->td_proc);
2226 		return (EFBIG);
2227 	}
2228 	return (0);
2229 }
2230 
2231 int
2232 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2233     struct thread *td)
2234 {
2235 	struct vnode *vp;
2236 
2237 	vp = fp->f_vnode;
2238 #ifdef AUDIT
2239 	vn_lock(vp, LK_SHARED | LK_RETRY);
2240 	AUDIT_ARG_VNODE1(vp);
2241 	VOP_UNLOCK(vp);
2242 #endif
2243 	return (setfmode(td, active_cred, vp, mode));
2244 }
2245 
2246 int
2247 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2248     struct thread *td)
2249 {
2250 	struct vnode *vp;
2251 
2252 	vp = fp->f_vnode;
2253 #ifdef AUDIT
2254 	vn_lock(vp, LK_SHARED | LK_RETRY);
2255 	AUDIT_ARG_VNODE1(vp);
2256 	VOP_UNLOCK(vp);
2257 #endif
2258 	return (setfown(td, active_cred, vp, uid, gid));
2259 }
2260 
2261 void
2262 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2263 {
2264 	vm_object_t object;
2265 
2266 	if ((object = vp->v_object) == NULL)
2267 		return;
2268 	VM_OBJECT_WLOCK(object);
2269 	vm_object_page_remove(object, start, end, 0);
2270 	VM_OBJECT_WUNLOCK(object);
2271 }
2272 
2273 int
2274 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2275 {
2276 	struct vattr va;
2277 	daddr_t bn, bnp;
2278 	uint64_t bsize;
2279 	off_t noff;
2280 	int error;
2281 
2282 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2283 	    ("Wrong command %lu", cmd));
2284 
2285 	if (vn_lock(vp, LK_SHARED) != 0)
2286 		return (EBADF);
2287 	if (vp->v_type != VREG) {
2288 		error = ENOTTY;
2289 		goto unlock;
2290 	}
2291 	error = VOP_GETATTR(vp, &va, cred);
2292 	if (error != 0)
2293 		goto unlock;
2294 	noff = *off;
2295 	if (noff >= va.va_size) {
2296 		error = ENXIO;
2297 		goto unlock;
2298 	}
2299 	bsize = vp->v_mount->mnt_stat.f_iosize;
2300 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2301 	    noff % bsize) {
2302 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2303 		if (error == EOPNOTSUPP) {
2304 			error = ENOTTY;
2305 			goto unlock;
2306 		}
2307 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2308 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2309 			noff = bn * bsize;
2310 			if (noff < *off)
2311 				noff = *off;
2312 			goto unlock;
2313 		}
2314 	}
2315 	if (noff > va.va_size)
2316 		noff = va.va_size;
2317 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2318 	if (cmd == FIOSEEKDATA)
2319 		error = ENXIO;
2320 unlock:
2321 	VOP_UNLOCK(vp);
2322 	if (error == 0)
2323 		*off = noff;
2324 	return (error);
2325 }
2326 
2327 int
2328 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2329 {
2330 	struct ucred *cred;
2331 	struct vnode *vp;
2332 	struct vattr vattr;
2333 	off_t foffset, size;
2334 	int error, noneg;
2335 
2336 	cred = td->td_ucred;
2337 	vp = fp->f_vnode;
2338 	foffset = foffset_lock(fp, 0);
2339 	noneg = (vp->v_type != VCHR);
2340 	error = 0;
2341 	switch (whence) {
2342 	case L_INCR:
2343 		if (noneg &&
2344 		    (foffset < 0 ||
2345 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2346 			error = EOVERFLOW;
2347 			break;
2348 		}
2349 		offset += foffset;
2350 		break;
2351 	case L_XTND:
2352 		vn_lock(vp, LK_SHARED | LK_RETRY);
2353 		error = VOP_GETATTR(vp, &vattr, cred);
2354 		VOP_UNLOCK(vp);
2355 		if (error)
2356 			break;
2357 
2358 		/*
2359 		 * If the file references a disk device, then fetch
2360 		 * the media size and use that to determine the ending
2361 		 * offset.
2362 		 */
2363 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2364 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2365 			vattr.va_size = size;
2366 		if (noneg &&
2367 		    (vattr.va_size > OFF_MAX ||
2368 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2369 			error = EOVERFLOW;
2370 			break;
2371 		}
2372 		offset += vattr.va_size;
2373 		break;
2374 	case L_SET:
2375 		break;
2376 	case SEEK_DATA:
2377 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2378 		if (error == ENOTTY)
2379 			error = EINVAL;
2380 		break;
2381 	case SEEK_HOLE:
2382 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2383 		if (error == ENOTTY)
2384 			error = EINVAL;
2385 		break;
2386 	default:
2387 		error = EINVAL;
2388 	}
2389 	if (error == 0 && noneg && offset < 0)
2390 		error = EINVAL;
2391 	if (error != 0)
2392 		goto drop;
2393 	VFS_KNOTE_UNLOCKED(vp, 0);
2394 	td->td_uretoff.tdu_off = offset;
2395 drop:
2396 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2397 	return (error);
2398 }
2399 
2400 int
2401 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2402     struct thread *td)
2403 {
2404 	int error;
2405 
2406 	/*
2407 	 * Grant permission if the caller is the owner of the file, or
2408 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2409 	 * on the file.  If the time pointer is null, then write
2410 	 * permission on the file is also sufficient.
2411 	 *
2412 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2413 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2414 	 * will be allowed to set the times [..] to the current
2415 	 * server time.
2416 	 */
2417 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2418 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2419 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2420 	return (error);
2421 }
2422 
2423 int
2424 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2425 {
2426 	struct vnode *vp;
2427 	int error;
2428 
2429 	if (fp->f_type == DTYPE_FIFO)
2430 		kif->kf_type = KF_TYPE_FIFO;
2431 	else
2432 		kif->kf_type = KF_TYPE_VNODE;
2433 	vp = fp->f_vnode;
2434 	vref(vp);
2435 	FILEDESC_SUNLOCK(fdp);
2436 	error = vn_fill_kinfo_vnode(vp, kif);
2437 	vrele(vp);
2438 	FILEDESC_SLOCK(fdp);
2439 	return (error);
2440 }
2441 
2442 static inline void
2443 vn_fill_junk(struct kinfo_file *kif)
2444 {
2445 	size_t len, olen;
2446 
2447 	/*
2448 	 * Simulate vn_fullpath returning changing values for a given
2449 	 * vp during e.g. coredump.
2450 	 */
2451 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2452 	olen = strlen(kif->kf_path);
2453 	if (len < olen)
2454 		strcpy(&kif->kf_path[len - 1], "$");
2455 	else
2456 		for (; olen < len; olen++)
2457 			strcpy(&kif->kf_path[olen], "A");
2458 }
2459 
2460 int
2461 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2462 {
2463 	struct vattr va;
2464 	char *fullpath, *freepath;
2465 	int error;
2466 
2467 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2468 	freepath = NULL;
2469 	fullpath = "-";
2470 	error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2471 	if (error == 0) {
2472 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2473 	}
2474 	if (freepath != NULL)
2475 		free(freepath, M_TEMP);
2476 
2477 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2478 		vn_fill_junk(kif);
2479 	);
2480 
2481 	/*
2482 	 * Retrieve vnode attributes.
2483 	 */
2484 	va.va_fsid = VNOVAL;
2485 	va.va_rdev = NODEV;
2486 	vn_lock(vp, LK_SHARED | LK_RETRY);
2487 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2488 	VOP_UNLOCK(vp);
2489 	if (error != 0)
2490 		return (error);
2491 	if (va.va_fsid != VNOVAL)
2492 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2493 	else
2494 		kif->kf_un.kf_file.kf_file_fsid =
2495 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2496 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2497 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2498 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2499 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2500 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2501 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2502 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2503 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2504 	return (0);
2505 }
2506 
2507 int
2508 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2509     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2510     struct thread *td)
2511 {
2512 #ifdef HWPMC_HOOKS
2513 	struct pmckern_map_in pkm;
2514 #endif
2515 	struct mount *mp;
2516 	struct vnode *vp;
2517 	vm_object_t object;
2518 	vm_prot_t maxprot;
2519 	boolean_t writecounted;
2520 	int error;
2521 
2522 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2523     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2524 	/*
2525 	 * POSIX shared-memory objects are defined to have
2526 	 * kernel persistence, and are not defined to support
2527 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2528 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2529 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2530 	 * flag to request this behavior.
2531 	 */
2532 	if ((fp->f_flag & FPOSIXSHM) != 0)
2533 		flags |= MAP_NOSYNC;
2534 #endif
2535 	vp = fp->f_vnode;
2536 
2537 	/*
2538 	 * Ensure that file and memory protections are
2539 	 * compatible.  Note that we only worry about
2540 	 * writability if mapping is shared; in this case,
2541 	 * current and max prot are dictated by the open file.
2542 	 * XXX use the vnode instead?  Problem is: what
2543 	 * credentials do we use for determination? What if
2544 	 * proc does a setuid?
2545 	 */
2546 	mp = vp->v_mount;
2547 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2548 		maxprot = VM_PROT_NONE;
2549 		if ((prot & VM_PROT_EXECUTE) != 0)
2550 			return (EACCES);
2551 	} else
2552 		maxprot = VM_PROT_EXECUTE;
2553 	if ((fp->f_flag & FREAD) != 0)
2554 		maxprot |= VM_PROT_READ;
2555 	else if ((prot & VM_PROT_READ) != 0)
2556 		return (EACCES);
2557 
2558 	/*
2559 	 * If we are sharing potential changes via MAP_SHARED and we
2560 	 * are trying to get write permission although we opened it
2561 	 * without asking for it, bail out.
2562 	 */
2563 	if ((flags & MAP_SHARED) != 0) {
2564 		if ((fp->f_flag & FWRITE) != 0)
2565 			maxprot |= VM_PROT_WRITE;
2566 		else if ((prot & VM_PROT_WRITE) != 0)
2567 			return (EACCES);
2568 	} else {
2569 		maxprot |= VM_PROT_WRITE;
2570 		cap_maxprot |= VM_PROT_WRITE;
2571 	}
2572 	maxprot &= cap_maxprot;
2573 
2574 	/*
2575 	 * For regular files and shared memory, POSIX requires that
2576 	 * the value of foff be a legitimate offset within the data
2577 	 * object.  In particular, negative offsets are invalid.
2578 	 * Blocking negative offsets and overflows here avoids
2579 	 * possible wraparound or user-level access into reserved
2580 	 * ranges of the data object later.  In contrast, POSIX does
2581 	 * not dictate how offsets are used by device drivers, so in
2582 	 * the case of a device mapping a negative offset is passed
2583 	 * on.
2584 	 */
2585 	if (
2586 #ifdef _LP64
2587 	    size > OFF_MAX ||
2588 #endif
2589 	    foff < 0 || foff > OFF_MAX - size)
2590 		return (EINVAL);
2591 
2592 	writecounted = FALSE;
2593 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2594 	    &foff, &object, &writecounted);
2595 	if (error != 0)
2596 		return (error);
2597 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2598 	    foff, writecounted, td);
2599 	if (error != 0) {
2600 		/*
2601 		 * If this mapping was accounted for in the vnode's
2602 		 * writecount, then undo that now.
2603 		 */
2604 		if (writecounted)
2605 			vm_pager_release_writecount(object, 0, size);
2606 		vm_object_deallocate(object);
2607 	}
2608 #ifdef HWPMC_HOOKS
2609 	/* Inform hwpmc(4) if an executable is being mapped. */
2610 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2611 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2612 			pkm.pm_file = vp;
2613 			pkm.pm_address = (uintptr_t) *addr;
2614 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2615 		}
2616 	}
2617 #endif
2618 	return (error);
2619 }
2620 
2621 void
2622 vn_fsid(struct vnode *vp, struct vattr *va)
2623 {
2624 	fsid_t *f;
2625 
2626 	f = &vp->v_mount->mnt_stat.f_fsid;
2627 	va->va_fsid = (uint32_t)f->val[1];
2628 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2629 	va->va_fsid += (uint32_t)f->val[0];
2630 }
2631 
2632 int
2633 vn_fsync_buf(struct vnode *vp, int waitfor)
2634 {
2635 	struct buf *bp, *nbp;
2636 	struct bufobj *bo;
2637 	struct mount *mp;
2638 	int error, maxretry;
2639 
2640 	error = 0;
2641 	maxretry = 10000;     /* large, arbitrarily chosen */
2642 	mp = NULL;
2643 	if (vp->v_type == VCHR) {
2644 		VI_LOCK(vp);
2645 		mp = vp->v_rdev->si_mountpt;
2646 		VI_UNLOCK(vp);
2647 	}
2648 	bo = &vp->v_bufobj;
2649 	BO_LOCK(bo);
2650 loop1:
2651 	/*
2652 	 * MARK/SCAN initialization to avoid infinite loops.
2653 	 */
2654         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2655 		bp->b_vflags &= ~BV_SCANNED;
2656 		bp->b_error = 0;
2657 	}
2658 
2659 	/*
2660 	 * Flush all dirty buffers associated with a vnode.
2661 	 */
2662 loop2:
2663 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2664 		if ((bp->b_vflags & BV_SCANNED) != 0)
2665 			continue;
2666 		bp->b_vflags |= BV_SCANNED;
2667 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2668 			if (waitfor != MNT_WAIT)
2669 				continue;
2670 			if (BUF_LOCK(bp,
2671 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2672 			    BO_LOCKPTR(bo)) != 0) {
2673 				BO_LOCK(bo);
2674 				goto loop1;
2675 			}
2676 			BO_LOCK(bo);
2677 		}
2678 		BO_UNLOCK(bo);
2679 		KASSERT(bp->b_bufobj == bo,
2680 		    ("bp %p wrong b_bufobj %p should be %p",
2681 		    bp, bp->b_bufobj, bo));
2682 		if ((bp->b_flags & B_DELWRI) == 0)
2683 			panic("fsync: not dirty");
2684 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2685 			vfs_bio_awrite(bp);
2686 		} else {
2687 			bremfree(bp);
2688 			bawrite(bp);
2689 		}
2690 		if (maxretry < 1000)
2691 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2692 		BO_LOCK(bo);
2693 		goto loop2;
2694 	}
2695 
2696 	/*
2697 	 * If synchronous the caller expects us to completely resolve all
2698 	 * dirty buffers in the system.  Wait for in-progress I/O to
2699 	 * complete (which could include background bitmap writes), then
2700 	 * retry if dirty blocks still exist.
2701 	 */
2702 	if (waitfor == MNT_WAIT) {
2703 		bufobj_wwait(bo, 0, 0);
2704 		if (bo->bo_dirty.bv_cnt > 0) {
2705 			/*
2706 			 * If we are unable to write any of these buffers
2707 			 * then we fail now rather than trying endlessly
2708 			 * to write them out.
2709 			 */
2710 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2711 				if ((error = bp->b_error) != 0)
2712 					break;
2713 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2714 			    (error == 0 && --maxretry >= 0))
2715 				goto loop1;
2716 			if (error == 0)
2717 				error = EAGAIN;
2718 		}
2719 	}
2720 	BO_UNLOCK(bo);
2721 	if (error != 0)
2722 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2723 
2724 	return (error);
2725 }
2726 
2727 /*
2728  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2729  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2730  * to do the actual copy.
2731  * vn_generic_copy_file_range() is factored out, so it can be called
2732  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2733  * different file systems.
2734  */
2735 int
2736 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2737     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2738     struct ucred *outcred, struct thread *fsize_td)
2739 {
2740 	int error;
2741 	size_t len;
2742 	uint64_t uvalin, uvalout;
2743 
2744 	len = *lenp;
2745 	*lenp = 0;		/* For error returns. */
2746 	error = 0;
2747 
2748 	/* Do some sanity checks on the arguments. */
2749 	uvalin = *inoffp;
2750 	uvalin += len;
2751 	uvalout = *outoffp;
2752 	uvalout += len;
2753 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2754 		error = EISDIR;
2755 	else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin <
2756 	    (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX ||
2757 	    uvalout < (uint64_t)*outoffp || invp->v_type != VREG ||
2758 	    outvp->v_type != VREG)
2759 		error = EINVAL;
2760 	if (error != 0)
2761 		goto out;
2762 
2763 	/*
2764 	 * If the two vnode are for the same file system, call
2765 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2766 	 * which can handle copies across multiple file systems.
2767 	 */
2768 	*lenp = len;
2769 	if (invp->v_mount == outvp->v_mount)
2770 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2771 		    lenp, flags, incred, outcred, fsize_td);
2772 	else
2773 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2774 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2775 out:
2776 	return (error);
2777 }
2778 
2779 /*
2780  * Test len bytes of data starting at dat for all bytes == 0.
2781  * Return true if all bytes are zero, false otherwise.
2782  * Expects dat to be well aligned.
2783  */
2784 static bool
2785 mem_iszero(void *dat, int len)
2786 {
2787 	int i;
2788 	const u_int *p;
2789 	const char *cp;
2790 
2791 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2792 		if (len >= sizeof(*p)) {
2793 			if (*p != 0)
2794 				return (false);
2795 		} else {
2796 			cp = (const char *)p;
2797 			for (i = 0; i < len; i++, cp++)
2798 				if (*cp != '\0')
2799 					return (false);
2800 		}
2801 	}
2802 	return (true);
2803 }
2804 
2805 /*
2806  * Look for a hole in the output file and, if found, adjust *outoffp
2807  * and *xferp to skip past the hole.
2808  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2809  * to be written as 0's upon return.
2810  */
2811 static off_t
2812 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2813     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2814 {
2815 	int error;
2816 	off_t delta;
2817 
2818 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2819 		*dataoffp = *outoffp;
2820 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2821 		    curthread);
2822 		if (error == 0) {
2823 			*holeoffp = *dataoffp;
2824 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2825 			    curthread);
2826 		}
2827 		if (error != 0 || *holeoffp == *dataoffp) {
2828 			/*
2829 			 * Since outvp is unlocked, it may be possible for
2830 			 * another thread to do a truncate(), lseek(), write()
2831 			 * creating a hole at startoff between the above
2832 			 * VOP_IOCTL() calls, if the other thread does not do
2833 			 * rangelocking.
2834 			 * If that happens, *holeoffp == *dataoffp and finding
2835 			 * the hole has failed, so disable vn_skip_hole().
2836 			 */
2837 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2838 			return (xfer2);
2839 		}
2840 		KASSERT(*dataoffp >= *outoffp,
2841 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2842 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2843 		KASSERT(*holeoffp > *dataoffp,
2844 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2845 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2846 	}
2847 
2848 	/*
2849 	 * If there is a hole before the data starts, advance *outoffp and
2850 	 * *xferp past the hole.
2851 	 */
2852 	if (*dataoffp > *outoffp) {
2853 		delta = *dataoffp - *outoffp;
2854 		if (delta >= *xferp) {
2855 			/* Entire *xferp is a hole. */
2856 			*outoffp += *xferp;
2857 			*xferp = 0;
2858 			return (0);
2859 		}
2860 		*xferp -= delta;
2861 		*outoffp += delta;
2862 		xfer2 = MIN(xfer2, *xferp);
2863 	}
2864 
2865 	/*
2866 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2867 	 * that the write ends at the start of the hole.
2868 	 * *holeoffp should always be greater than *outoffp, but for the
2869 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2870 	 * value.
2871 	 */
2872 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2873 		xfer2 = *holeoffp - *outoffp;
2874 	return (xfer2);
2875 }
2876 
2877 /*
2878  * Write an xfer sized chunk to outvp in blksize blocks from dat.
2879  * dat is a maximum of blksize in length and can be written repeatedly in
2880  * the chunk.
2881  * If growfile == true, just grow the file via vn_truncate_locked() instead
2882  * of doing actual writes.
2883  * If checkhole == true, a hole is being punched, so skip over any hole
2884  * already in the output file.
2885  */
2886 static int
2887 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2888     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2889 {
2890 	struct mount *mp;
2891 	off_t dataoff, holeoff, xfer2;
2892 	int error, lckf;
2893 
2894 	/*
2895 	 * Loop around doing writes of blksize until write has been completed.
2896 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2897 	 * done for each iteration, since the xfer argument can be very
2898 	 * large if there is a large hole to punch in the output file.
2899 	 */
2900 	error = 0;
2901 	holeoff = 0;
2902 	do {
2903 		xfer2 = MIN(xfer, blksize);
2904 		if (checkhole) {
2905 			/*
2906 			 * Punching a hole.  Skip writing if there is
2907 			 * already a hole in the output file.
2908 			 */
2909 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2910 			    &dataoff, &holeoff, cred);
2911 			if (xfer == 0)
2912 				break;
2913 			if (holeoff < 0)
2914 				checkhole = false;
2915 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2916 			    (intmax_t)xfer2));
2917 		}
2918 		bwillwrite();
2919 		mp = NULL;
2920 		error = vn_start_write(outvp, &mp, V_WAIT);
2921 		if (error == 0) {
2922 			if (MNT_SHARED_WRITES(mp))
2923 				lckf = LK_SHARED;
2924 			else
2925 				lckf = LK_EXCLUSIVE;
2926 			error = vn_lock(outvp, lckf);
2927 		}
2928 		if (error == 0) {
2929 			if (growfile)
2930 				error = vn_truncate_locked(outvp, outoff + xfer,
2931 				    false, cred);
2932 			else {
2933 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
2934 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
2935 				    curthread->td_ucred, cred, NULL, curthread);
2936 				outoff += xfer2;
2937 				xfer -= xfer2;
2938 			}
2939 			VOP_UNLOCK(outvp);
2940 		}
2941 		if (mp != NULL)
2942 			vn_finished_write(mp);
2943 	} while (!growfile && xfer > 0 && error == 0);
2944 	return (error);
2945 }
2946 
2947 /*
2948  * Copy a byte range of one file to another.  This function can handle the
2949  * case where invp and outvp are on different file systems.
2950  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
2951  * is no better file system specific way to do it.
2952  */
2953 int
2954 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
2955     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
2956     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
2957 {
2958 	struct vattr va;
2959 	struct mount *mp;
2960 	struct uio io;
2961 	off_t startoff, endoff, xfer, xfer2;
2962 	u_long blksize;
2963 	int error;
2964 	bool cantseek, readzeros, eof, lastblock;
2965 	ssize_t aresid;
2966 	size_t copylen, len, savlen;
2967 	char *dat;
2968 	long holein, holeout;
2969 
2970 	holein = holeout = 0;
2971 	savlen = len = *lenp;
2972 	error = 0;
2973 	dat = NULL;
2974 
2975 	error = vn_lock(invp, LK_SHARED);
2976 	if (error != 0)
2977 		goto out;
2978 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
2979 		holein = 0;
2980 	VOP_UNLOCK(invp);
2981 
2982 	mp = NULL;
2983 	error = vn_start_write(outvp, &mp, V_WAIT);
2984 	if (error == 0)
2985 		error = vn_lock(outvp, LK_EXCLUSIVE);
2986 	if (error == 0) {
2987 		/*
2988 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
2989 		 * now that outvp is locked.
2990 		 */
2991 		if (fsize_td != NULL) {
2992 			io.uio_offset = *outoffp;
2993 			io.uio_resid = len;
2994 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
2995 			if (error != 0)
2996 				error = EFBIG;
2997 		}
2998 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
2999 			holeout = 0;
3000 		/*
3001 		 * Holes that are past EOF do not need to be written as a block
3002 		 * of zero bytes.  So, truncate the output file as far as
3003 		 * possible and then use va.va_size to decide if writing 0
3004 		 * bytes is necessary in the loop below.
3005 		 */
3006 		if (error == 0)
3007 			error = VOP_GETATTR(outvp, &va, outcred);
3008 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3009 		    *outoffp + len) {
3010 #ifdef MAC
3011 			error = mac_vnode_check_write(curthread->td_ucred,
3012 			    outcred, outvp);
3013 			if (error == 0)
3014 #endif
3015 				error = vn_truncate_locked(outvp, *outoffp,
3016 				    false, outcred);
3017 			if (error == 0)
3018 				va.va_size = *outoffp;
3019 		}
3020 		VOP_UNLOCK(outvp);
3021 	}
3022 	if (mp != NULL)
3023 		vn_finished_write(mp);
3024 	if (error != 0)
3025 		goto out;
3026 
3027 	/*
3028 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3029 	 * If hole sizes aren't available, set the blksize to the larger
3030 	 * f_iosize of invp and outvp.
3031 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3032 	 * This value is clipped at 4Kbytes and 1Mbyte.
3033 	 */
3034 	blksize = MAX(holein, holeout);
3035 	if (blksize == 0)
3036 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3037 		    outvp->v_mount->mnt_stat.f_iosize);
3038 	if (blksize < 4096)
3039 		blksize = 4096;
3040 	else if (blksize > 1024 * 1024)
3041 		blksize = 1024 * 1024;
3042 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3043 
3044 	/*
3045 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3046 	 * to find holes.  Otherwise, just scan the read block for all 0s
3047 	 * in the inner loop where the data copying is done.
3048 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3049 	 * support holes on the server, but do not support FIOSEEKHOLE.
3050 	 */
3051 	eof = false;
3052 	while (len > 0 && error == 0 && !eof) {
3053 		endoff = 0;			/* To shut up compilers. */
3054 		cantseek = true;
3055 		startoff = *inoffp;
3056 		copylen = len;
3057 
3058 		/*
3059 		 * Find the next data area.  If there is just a hole to EOF,
3060 		 * FIOSEEKDATA should fail and then we drop down into the
3061 		 * inner loop and create the hole on the outvp file.
3062 		 * (I do not know if any file system will report a hole to
3063 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3064 		 *  will fail for those file systems.)
3065 		 *
3066 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3067 		 * the code just falls through to the inner copy loop.
3068 		 */
3069 		error = EINVAL;
3070 		if (holein > 0)
3071 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3072 			    incred, curthread);
3073 		if (error == 0) {
3074 			endoff = startoff;
3075 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3076 			    incred, curthread);
3077 			/*
3078 			 * Since invp is unlocked, it may be possible for
3079 			 * another thread to do a truncate(), lseek(), write()
3080 			 * creating a hole at startoff between the above
3081 			 * VOP_IOCTL() calls, if the other thread does not do
3082 			 * rangelocking.
3083 			 * If that happens, startoff == endoff and finding
3084 			 * the hole has failed, so set an error.
3085 			 */
3086 			if (error == 0 && startoff == endoff)
3087 				error = EINVAL; /* Any error. Reset to 0. */
3088 		}
3089 		if (error == 0) {
3090 			if (startoff > *inoffp) {
3091 				/* Found hole before data block. */
3092 				xfer = MIN(startoff - *inoffp, len);
3093 				if (*outoffp < va.va_size) {
3094 					/* Must write 0s to punch hole. */
3095 					xfer2 = MIN(va.va_size - *outoffp,
3096 					    xfer);
3097 					memset(dat, 0, MIN(xfer2, blksize));
3098 					error = vn_write_outvp(outvp, dat,
3099 					    *outoffp, xfer2, blksize, false,
3100 					    holeout > 0, outcred);
3101 				}
3102 
3103 				if (error == 0 && *outoffp + xfer >
3104 				    va.va_size && xfer == len)
3105 					/* Grow last block. */
3106 					error = vn_write_outvp(outvp, dat,
3107 					    *outoffp, xfer, blksize, true,
3108 					    false, outcred);
3109 				if (error == 0) {
3110 					*inoffp += xfer;
3111 					*outoffp += xfer;
3112 					len -= xfer;
3113 				}
3114 			}
3115 			copylen = MIN(len, endoff - startoff);
3116 			cantseek = false;
3117 		} else {
3118 			cantseek = true;
3119 			startoff = *inoffp;
3120 			copylen = len;
3121 			error = 0;
3122 		}
3123 
3124 		xfer = blksize;
3125 		if (cantseek) {
3126 			/*
3127 			 * Set first xfer to end at a block boundary, so that
3128 			 * holes are more likely detected in the loop below via
3129 			 * the for all bytes 0 method.
3130 			 */
3131 			xfer -= (*inoffp % blksize);
3132 		}
3133 		/* Loop copying the data block. */
3134 		while (copylen > 0 && error == 0 && !eof) {
3135 			if (copylen < xfer)
3136 				xfer = copylen;
3137 			error = vn_lock(invp, LK_SHARED);
3138 			if (error != 0)
3139 				goto out;
3140 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3141 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3142 			    curthread->td_ucred, incred, &aresid,
3143 			    curthread);
3144 			VOP_UNLOCK(invp);
3145 			lastblock = false;
3146 			if (error == 0 && aresid > 0) {
3147 				/* Stop the copy at EOF on the input file. */
3148 				xfer -= aresid;
3149 				eof = true;
3150 				lastblock = true;
3151 			}
3152 			if (error == 0) {
3153 				/*
3154 				 * Skip the write for holes past the initial EOF
3155 				 * of the output file, unless this is the last
3156 				 * write of the output file at EOF.
3157 				 */
3158 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3159 				    false;
3160 				if (xfer == len)
3161 					lastblock = true;
3162 				if (!cantseek || *outoffp < va.va_size ||
3163 				    lastblock || !readzeros)
3164 					error = vn_write_outvp(outvp, dat,
3165 					    *outoffp, xfer, blksize,
3166 					    readzeros && lastblock &&
3167 					    *outoffp >= va.va_size, false,
3168 					    outcred);
3169 				if (error == 0) {
3170 					*inoffp += xfer;
3171 					startoff += xfer;
3172 					*outoffp += xfer;
3173 					copylen -= xfer;
3174 					len -= xfer;
3175 				}
3176 			}
3177 			xfer = blksize;
3178 		}
3179 	}
3180 out:
3181 	*lenp = savlen - len;
3182 	free(dat, M_TEMP);
3183 	return (error);
3184 }
3185 
3186 static int
3187 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3188 {
3189 	struct mount *mp;
3190 	struct vnode *vp;
3191 	off_t olen, ooffset;
3192 	int error;
3193 #ifdef AUDIT
3194 	int audited_vnode1 = 0;
3195 #endif
3196 
3197 	vp = fp->f_vnode;
3198 	if (vp->v_type != VREG)
3199 		return (ENODEV);
3200 
3201 	/* Allocating blocks may take a long time, so iterate. */
3202 	for (;;) {
3203 		olen = len;
3204 		ooffset = offset;
3205 
3206 		bwillwrite();
3207 		mp = NULL;
3208 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3209 		if (error != 0)
3210 			break;
3211 		error = vn_lock(vp, LK_EXCLUSIVE);
3212 		if (error != 0) {
3213 			vn_finished_write(mp);
3214 			break;
3215 		}
3216 #ifdef AUDIT
3217 		if (!audited_vnode1) {
3218 			AUDIT_ARG_VNODE1(vp);
3219 			audited_vnode1 = 1;
3220 		}
3221 #endif
3222 #ifdef MAC
3223 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3224 		if (error == 0)
3225 #endif
3226 			error = VOP_ALLOCATE(vp, &offset, &len);
3227 		VOP_UNLOCK(vp);
3228 		vn_finished_write(mp);
3229 
3230 		if (olen + ooffset != offset + len) {
3231 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3232 			    ooffset, olen, offset, len);
3233 		}
3234 		if (error != 0 || len == 0)
3235 			break;
3236 		KASSERT(olen > len, ("Iteration did not make progress?"));
3237 		maybe_yield();
3238 	}
3239 
3240 	return (error);
3241 }
3242