xref: /freebsd/sys/kern/vfs_vnops.c (revision a7dea1671b87c07d2d266f836bfa8b58efc7c134)
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/sysctl.h>
75 #include <sys/ttycom.h>
76 #include <sys/conf.h>
77 #include <sys/syslog.h>
78 #include <sys/unistd.h>
79 #include <sys/user.h>
80 
81 #include <security/audit/audit.h>
82 #include <security/mac/mac_framework.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_extern.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pager.h>
91 
92 #ifdef HWPMC_HOOKS
93 #include <sys/pmckern.h>
94 #endif
95 
96 static fo_rdwr_t	vn_read;
97 static fo_rdwr_t	vn_write;
98 static fo_rdwr_t	vn_io_fault;
99 static fo_truncate_t	vn_truncate;
100 static fo_ioctl_t	vn_ioctl;
101 static fo_poll_t	vn_poll;
102 static fo_kqfilter_t	vn_kqfilter;
103 static fo_stat_t	vn_statfile;
104 static fo_close_t	vn_closefile;
105 static fo_mmap_t	vn_mmap;
106 static fo_fallocate_t	vn_fallocate;
107 
108 struct 	fileops vnops = {
109 	.fo_read = vn_io_fault,
110 	.fo_write = vn_io_fault,
111 	.fo_truncate = vn_truncate,
112 	.fo_ioctl = vn_ioctl,
113 	.fo_poll = vn_poll,
114 	.fo_kqfilter = vn_kqfilter,
115 	.fo_stat = vn_statfile,
116 	.fo_close = vn_closefile,
117 	.fo_chmod = vn_chmod,
118 	.fo_chown = vn_chown,
119 	.fo_sendfile = vn_sendfile,
120 	.fo_seek = vn_seek,
121 	.fo_fill_kinfo = vn_fill_kinfo,
122 	.fo_mmap = vn_mmap,
123 	.fo_fallocate = vn_fallocate,
124 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
125 };
126 
127 static const int io_hold_cnt = 16;
128 static int vn_io_fault_enable = 1;
129 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
130     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
131 static int vn_io_fault_prefault = 0;
132 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW,
133     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
134 static u_long vn_io_faults_cnt;
135 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
136     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
137 
138 /*
139  * Returns true if vn_io_fault mode of handling the i/o request should
140  * be used.
141  */
142 static bool
143 do_vn_io_fault(struct vnode *vp, struct uio *uio)
144 {
145 	struct mount *mp;
146 
147 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
148 	    (mp = vp->v_mount) != NULL &&
149 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
150 }
151 
152 /*
153  * Structure used to pass arguments to vn_io_fault1(), to do either
154  * file- or vnode-based I/O calls.
155  */
156 struct vn_io_fault_args {
157 	enum {
158 		VN_IO_FAULT_FOP,
159 		VN_IO_FAULT_VOP
160 	} kind;
161 	struct ucred *cred;
162 	int flags;
163 	union {
164 		struct fop_args_tag {
165 			struct file *fp;
166 			fo_rdwr_t *doio;
167 		} fop_args;
168 		struct vop_args_tag {
169 			struct vnode *vp;
170 		} vop_args;
171 	} args;
172 };
173 
174 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
175     struct vn_io_fault_args *args, struct thread *td);
176 
177 int
178 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
179 {
180 	struct thread *td = ndp->ni_cnd.cn_thread;
181 
182 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
183 }
184 
185 /*
186  * Common code for vnode open operations via a name lookup.
187  * Lookup the vnode and invoke VOP_CREATE if needed.
188  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
189  *
190  * Note that this does NOT free nameidata for the successful case,
191  * due to the NDINIT being done elsewhere.
192  */
193 int
194 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
195     struct ucred *cred, struct file *fp)
196 {
197 	struct vnode *vp;
198 	struct mount *mp;
199 	struct thread *td = ndp->ni_cnd.cn_thread;
200 	struct vattr vat;
201 	struct vattr *vap = &vat;
202 	int fmode, error;
203 
204 restart:
205 	fmode = *flagp;
206 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
207 	    O_EXCL | O_DIRECTORY))
208 		return (EINVAL);
209 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
210 		ndp->ni_cnd.cn_nameiop = CREATE;
211 		/*
212 		 * Set NOCACHE to avoid flushing the cache when
213 		 * rolling in many files at once.
214 		*/
215 		ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
216 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
217 			ndp->ni_cnd.cn_flags |= FOLLOW;
218 		if ((fmode & O_BENEATH) != 0)
219 			ndp->ni_cnd.cn_flags |= BENEATH;
220 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
221 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
222 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
223 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
224 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
225 			bwillwrite();
226 		if ((error = namei(ndp)) != 0)
227 			return (error);
228 		if (ndp->ni_vp == NULL) {
229 			VATTR_NULL(vap);
230 			vap->va_type = VREG;
231 			vap->va_mode = cmode;
232 			if (fmode & O_EXCL)
233 				vap->va_vaflags |= VA_EXCLUSIVE;
234 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
235 				NDFREE(ndp, NDF_ONLY_PNBUF);
236 				vput(ndp->ni_dvp);
237 				if ((error = vn_start_write(NULL, &mp,
238 				    V_XSLEEP | PCATCH)) != 0)
239 					return (error);
240 				goto restart;
241 			}
242 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
243 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
244 #ifdef MAC
245 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
246 			    &ndp->ni_cnd, vap);
247 			if (error == 0)
248 #endif
249 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
250 						   &ndp->ni_cnd, vap);
251 			vput(ndp->ni_dvp);
252 			vn_finished_write(mp);
253 			if (error) {
254 				NDFREE(ndp, NDF_ONLY_PNBUF);
255 				return (error);
256 			}
257 			fmode &= ~O_TRUNC;
258 			vp = ndp->ni_vp;
259 		} else {
260 			if (ndp->ni_dvp == ndp->ni_vp)
261 				vrele(ndp->ni_dvp);
262 			else
263 				vput(ndp->ni_dvp);
264 			ndp->ni_dvp = NULL;
265 			vp = ndp->ni_vp;
266 			if (fmode & O_EXCL) {
267 				error = EEXIST;
268 				goto bad;
269 			}
270 			if (vp->v_type == VDIR) {
271 				error = EISDIR;
272 				goto bad;
273 			}
274 			fmode &= ~O_CREAT;
275 		}
276 	} else {
277 		ndp->ni_cnd.cn_nameiop = LOOKUP;
278 		ndp->ni_cnd.cn_flags = ISOPEN |
279 		    ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
280 		if (!(fmode & FWRITE))
281 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
282 		if ((fmode & O_BENEATH) != 0)
283 			ndp->ni_cnd.cn_flags |= BENEATH;
284 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
285 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
286 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
287 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
288 		if ((error = namei(ndp)) != 0)
289 			return (error);
290 		vp = ndp->ni_vp;
291 	}
292 	error = vn_open_vnode(vp, fmode, cred, td, fp);
293 	if (error)
294 		goto bad;
295 	*flagp = fmode;
296 	return (0);
297 bad:
298 	NDFREE(ndp, NDF_ONLY_PNBUF);
299 	vput(vp);
300 	*flagp = fmode;
301 	ndp->ni_vp = NULL;
302 	return (error);
303 }
304 
305 static int
306 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
307 {
308 	struct flock lf;
309 	int error, lock_flags, type;
310 
311 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
312 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
313 		return (0);
314 	KASSERT(fp != NULL, ("open with flock requires fp"));
315 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
316 		return (EOPNOTSUPP);
317 
318 	lock_flags = VOP_ISLOCKED(vp);
319 	VOP_UNLOCK(vp);
320 
321 	lf.l_whence = SEEK_SET;
322 	lf.l_start = 0;
323 	lf.l_len = 0;
324 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
325 	type = F_FLOCK;
326 	if ((fmode & FNONBLOCK) == 0)
327 		type |= F_WAIT;
328 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
329 	if (error == 0)
330 		fp->f_flag |= FHASLOCK;
331 
332 	vn_lock(vp, lock_flags | LK_RETRY);
333 	if (error == 0 && VN_IS_DOOMED(vp))
334 		error = ENOENT;
335 	return (error);
336 }
337 
338 /*
339  * Common code for vnode open operations once a vnode is located.
340  * Check permissions, and call the VOP_OPEN routine.
341  */
342 int
343 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
344     struct thread *td, struct file *fp)
345 {
346 	accmode_t accmode;
347 	int error;
348 
349 	if (vp->v_type == VLNK)
350 		return (EMLINK);
351 	if (vp->v_type == VSOCK)
352 		return (EOPNOTSUPP);
353 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
354 		return (ENOTDIR);
355 	accmode = 0;
356 	if (fmode & (FWRITE | O_TRUNC)) {
357 		if (vp->v_type == VDIR)
358 			return (EISDIR);
359 		accmode |= VWRITE;
360 	}
361 	if (fmode & FREAD)
362 		accmode |= VREAD;
363 	if (fmode & FEXEC)
364 		accmode |= VEXEC;
365 	if ((fmode & O_APPEND) && (fmode & FWRITE))
366 		accmode |= VAPPEND;
367 #ifdef MAC
368 	if (fmode & O_CREAT)
369 		accmode |= VCREAT;
370 	if (fmode & O_VERIFY)
371 		accmode |= VVERIFY;
372 	error = mac_vnode_check_open(cred, vp, accmode);
373 	if (error)
374 		return (error);
375 
376 	accmode &= ~(VCREAT | VVERIFY);
377 #endif
378 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
379 		error = VOP_ACCESS(vp, accmode, cred, td);
380 		if (error != 0)
381 			return (error);
382 	}
383 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
384 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
385 	error = VOP_OPEN(vp, fmode, cred, td, fp);
386 	if (error != 0)
387 		return (error);
388 
389 	error = vn_open_vnode_advlock(vp, fmode, fp);
390 	if (error == 0 && (fmode & FWRITE) != 0) {
391 		error = VOP_ADD_WRITECOUNT(vp, 1);
392 		if (error == 0) {
393 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
394 			     __func__, vp, vp->v_writecount);
395 		}
396 	}
397 
398 	/*
399 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
400 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
401 	 * Arrange for that by having fdrop() to use vn_closefile().
402 	 */
403 	if (error != 0) {
404 		fp->f_flag |= FOPENFAILED;
405 		fp->f_vnode = vp;
406 		if (fp->f_ops == &badfileops) {
407 			fp->f_type = DTYPE_VNODE;
408 			fp->f_ops = &vnops;
409 		}
410 		vref(vp);
411 	}
412 
413 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
414 	return (error);
415 
416 }
417 
418 /*
419  * Check for write permissions on the specified vnode.
420  * Prototype text segments cannot be written.
421  * It is racy.
422  */
423 int
424 vn_writechk(struct vnode *vp)
425 {
426 
427 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
428 	/*
429 	 * If there's shared text associated with
430 	 * the vnode, try to free it up once.  If
431 	 * we fail, we can't allow writing.
432 	 */
433 	if (VOP_IS_TEXT(vp))
434 		return (ETXTBSY);
435 
436 	return (0);
437 }
438 
439 /*
440  * Vnode close call
441  */
442 static int
443 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
444     struct thread *td, bool keep_ref)
445 {
446 	struct mount *mp;
447 	int error, lock_flags;
448 
449 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
450 	    MNT_EXTENDED_SHARED(vp->v_mount))
451 		lock_flags = LK_SHARED;
452 	else
453 		lock_flags = LK_EXCLUSIVE;
454 
455 	vn_start_write(vp, &mp, V_WAIT);
456 	vn_lock(vp, lock_flags | LK_RETRY);
457 	AUDIT_ARG_VNODE1(vp);
458 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
459 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
460 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
461 		    __func__, vp, vp->v_writecount);
462 	}
463 	error = VOP_CLOSE(vp, flags, file_cred, td);
464 	if (keep_ref)
465 		VOP_UNLOCK(vp);
466 	else
467 		vput(vp);
468 	vn_finished_write(mp);
469 	return (error);
470 }
471 
472 int
473 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
474     struct thread *td)
475 {
476 
477 	return (vn_close1(vp, flags, file_cred, td, false));
478 }
479 
480 /*
481  * Heuristic to detect sequential operation.
482  */
483 static int
484 sequential_heuristic(struct uio *uio, struct file *fp)
485 {
486 
487 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
488 	if (fp->f_flag & FRDAHEAD)
489 		return (fp->f_seqcount << IO_SEQSHIFT);
490 
491 	/*
492 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
493 	 * that the first I/O is normally considered to be slightly
494 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
495 	 * unless previous seeks have reduced f_seqcount to 0, in which
496 	 * case offset 0 is not special.
497 	 */
498 	if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
499 	    uio->uio_offset == fp->f_nextoff) {
500 		/*
501 		 * f_seqcount is in units of fixed-size blocks so that it
502 		 * depends mainly on the amount of sequential I/O and not
503 		 * much on the number of sequential I/O's.  The fixed size
504 		 * of 16384 is hard-coded here since it is (not quite) just
505 		 * a magic size that works well here.  This size is more
506 		 * closely related to the best I/O size for real disks than
507 		 * to any block size used by software.
508 		 */
509 		if (uio->uio_resid >= IO_SEQMAX * 16384)
510 			fp->f_seqcount = IO_SEQMAX;
511 		else {
512 			fp->f_seqcount += howmany(uio->uio_resid, 16384);
513 			if (fp->f_seqcount > IO_SEQMAX)
514 				fp->f_seqcount = IO_SEQMAX;
515 		}
516 		return (fp->f_seqcount << IO_SEQSHIFT);
517 	}
518 
519 	/* Not sequential.  Quickly draw-down sequentiality. */
520 	if (fp->f_seqcount > 1)
521 		fp->f_seqcount = 1;
522 	else
523 		fp->f_seqcount = 0;
524 	return (0);
525 }
526 
527 /*
528  * Package up an I/O request on a vnode into a uio and do it.
529  */
530 int
531 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
532     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
533     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
534 {
535 	struct uio auio;
536 	struct iovec aiov;
537 	struct mount *mp;
538 	struct ucred *cred;
539 	void *rl_cookie;
540 	struct vn_io_fault_args args;
541 	int error, lock_flags;
542 
543 	if (offset < 0 && vp->v_type != VCHR)
544 		return (EINVAL);
545 	auio.uio_iov = &aiov;
546 	auio.uio_iovcnt = 1;
547 	aiov.iov_base = base;
548 	aiov.iov_len = len;
549 	auio.uio_resid = len;
550 	auio.uio_offset = offset;
551 	auio.uio_segflg = segflg;
552 	auio.uio_rw = rw;
553 	auio.uio_td = td;
554 	error = 0;
555 
556 	if ((ioflg & IO_NODELOCKED) == 0) {
557 		if ((ioflg & IO_RANGELOCKED) == 0) {
558 			if (rw == UIO_READ) {
559 				rl_cookie = vn_rangelock_rlock(vp, offset,
560 				    offset + len);
561 			} else {
562 				rl_cookie = vn_rangelock_wlock(vp, offset,
563 				    offset + len);
564 			}
565 		} else
566 			rl_cookie = NULL;
567 		mp = NULL;
568 		if (rw == UIO_WRITE) {
569 			if (vp->v_type != VCHR &&
570 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
571 			    != 0)
572 				goto out;
573 			if (MNT_SHARED_WRITES(mp) ||
574 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
575 				lock_flags = LK_SHARED;
576 			else
577 				lock_flags = LK_EXCLUSIVE;
578 		} else
579 			lock_flags = LK_SHARED;
580 		vn_lock(vp, lock_flags | LK_RETRY);
581 	} else
582 		rl_cookie = NULL;
583 
584 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
585 #ifdef MAC
586 	if ((ioflg & IO_NOMACCHECK) == 0) {
587 		if (rw == UIO_READ)
588 			error = mac_vnode_check_read(active_cred, file_cred,
589 			    vp);
590 		else
591 			error = mac_vnode_check_write(active_cred, file_cred,
592 			    vp);
593 	}
594 #endif
595 	if (error == 0) {
596 		if (file_cred != NULL)
597 			cred = file_cred;
598 		else
599 			cred = active_cred;
600 		if (do_vn_io_fault(vp, &auio)) {
601 			args.kind = VN_IO_FAULT_VOP;
602 			args.cred = cred;
603 			args.flags = ioflg;
604 			args.args.vop_args.vp = vp;
605 			error = vn_io_fault1(vp, &auio, &args, td);
606 		} else if (rw == UIO_READ) {
607 			error = VOP_READ(vp, &auio, ioflg, cred);
608 		} else /* if (rw == UIO_WRITE) */ {
609 			error = VOP_WRITE(vp, &auio, ioflg, cred);
610 		}
611 	}
612 	if (aresid)
613 		*aresid = auio.uio_resid;
614 	else
615 		if (auio.uio_resid && error == 0)
616 			error = EIO;
617 	if ((ioflg & IO_NODELOCKED) == 0) {
618 		VOP_UNLOCK(vp);
619 		if (mp != NULL)
620 			vn_finished_write(mp);
621 	}
622  out:
623 	if (rl_cookie != NULL)
624 		vn_rangelock_unlock(vp, rl_cookie);
625 	return (error);
626 }
627 
628 /*
629  * Package up an I/O request on a vnode into a uio and do it.  The I/O
630  * request is split up into smaller chunks and we try to avoid saturating
631  * the buffer cache while potentially holding a vnode locked, so we
632  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
633  * to give other processes a chance to lock the vnode (either other processes
634  * core'ing the same binary, or unrelated processes scanning the directory).
635  */
636 int
637 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
638     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
639     struct ucred *file_cred, size_t *aresid, struct thread *td)
640 {
641 	int error = 0;
642 	ssize_t iaresid;
643 
644 	do {
645 		int chunk;
646 
647 		/*
648 		 * Force `offset' to a multiple of MAXBSIZE except possibly
649 		 * for the first chunk, so that filesystems only need to
650 		 * write full blocks except possibly for the first and last
651 		 * chunks.
652 		 */
653 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
654 
655 		if (chunk > len)
656 			chunk = len;
657 		if (rw != UIO_READ && vp->v_type == VREG)
658 			bwillwrite();
659 		iaresid = 0;
660 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
661 		    ioflg, active_cred, file_cred, &iaresid, td);
662 		len -= chunk;	/* aresid calc already includes length */
663 		if (error)
664 			break;
665 		offset += chunk;
666 		base = (char *)base + chunk;
667 		kern_yield(PRI_USER);
668 	} while (len);
669 	if (aresid)
670 		*aresid = len + iaresid;
671 	return (error);
672 }
673 
674 off_t
675 foffset_lock(struct file *fp, int flags)
676 {
677 	struct mtx *mtxp;
678 	off_t res;
679 
680 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
681 
682 #if OFF_MAX <= LONG_MAX
683 	/*
684 	 * Caller only wants the current f_offset value.  Assume that
685 	 * the long and shorter integer types reads are atomic.
686 	 */
687 	if ((flags & FOF_NOLOCK) != 0)
688 		return (fp->f_offset);
689 #endif
690 
691 	/*
692 	 * According to McKusick the vn lock was protecting f_offset here.
693 	 * It is now protected by the FOFFSET_LOCKED flag.
694 	 */
695 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
696 	mtx_lock(mtxp);
697 	if ((flags & FOF_NOLOCK) == 0) {
698 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
699 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
700 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
701 			    "vofflock", 0);
702 		}
703 		fp->f_vnread_flags |= FOFFSET_LOCKED;
704 	}
705 	res = fp->f_offset;
706 	mtx_unlock(mtxp);
707 	return (res);
708 }
709 
710 void
711 foffset_unlock(struct file *fp, off_t val, int flags)
712 {
713 	struct mtx *mtxp;
714 
715 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
716 
717 #if OFF_MAX <= LONG_MAX
718 	if ((flags & FOF_NOLOCK) != 0) {
719 		if ((flags & FOF_NOUPDATE) == 0)
720 			fp->f_offset = val;
721 		if ((flags & FOF_NEXTOFF) != 0)
722 			fp->f_nextoff = val;
723 		return;
724 	}
725 #endif
726 
727 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
728 	mtx_lock(mtxp);
729 	if ((flags & FOF_NOUPDATE) == 0)
730 		fp->f_offset = val;
731 	if ((flags & FOF_NEXTOFF) != 0)
732 		fp->f_nextoff = val;
733 	if ((flags & FOF_NOLOCK) == 0) {
734 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
735 		    ("Lost FOFFSET_LOCKED"));
736 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
737 			wakeup(&fp->f_vnread_flags);
738 		fp->f_vnread_flags = 0;
739 	}
740 	mtx_unlock(mtxp);
741 }
742 
743 void
744 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
745 {
746 
747 	if ((flags & FOF_OFFSET) == 0)
748 		uio->uio_offset = foffset_lock(fp, flags);
749 }
750 
751 void
752 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
753 {
754 
755 	if ((flags & FOF_OFFSET) == 0)
756 		foffset_unlock(fp, uio->uio_offset, flags);
757 }
758 
759 static int
760 get_advice(struct file *fp, struct uio *uio)
761 {
762 	struct mtx *mtxp;
763 	int ret;
764 
765 	ret = POSIX_FADV_NORMAL;
766 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
767 		return (ret);
768 
769 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
770 	mtx_lock(mtxp);
771 	if (fp->f_advice != NULL &&
772 	    uio->uio_offset >= fp->f_advice->fa_start &&
773 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
774 		ret = fp->f_advice->fa_advice;
775 	mtx_unlock(mtxp);
776 	return (ret);
777 }
778 
779 /*
780  * File table vnode read routine.
781  */
782 static int
783 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
784     struct thread *td)
785 {
786 	struct vnode *vp;
787 	off_t orig_offset;
788 	int error, ioflag;
789 	int advice;
790 
791 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
792 	    uio->uio_td, td));
793 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
794 	vp = fp->f_vnode;
795 	ioflag = 0;
796 	if (fp->f_flag & FNONBLOCK)
797 		ioflag |= IO_NDELAY;
798 	if (fp->f_flag & O_DIRECT)
799 		ioflag |= IO_DIRECT;
800 	advice = get_advice(fp, uio);
801 	vn_lock(vp, LK_SHARED | LK_RETRY);
802 
803 	switch (advice) {
804 	case POSIX_FADV_NORMAL:
805 	case POSIX_FADV_SEQUENTIAL:
806 	case POSIX_FADV_NOREUSE:
807 		ioflag |= sequential_heuristic(uio, fp);
808 		break;
809 	case POSIX_FADV_RANDOM:
810 		/* Disable read-ahead for random I/O. */
811 		break;
812 	}
813 	orig_offset = uio->uio_offset;
814 
815 #ifdef MAC
816 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
817 	if (error == 0)
818 #endif
819 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
820 	fp->f_nextoff = uio->uio_offset;
821 	VOP_UNLOCK(vp);
822 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
823 	    orig_offset != uio->uio_offset)
824 		/*
825 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
826 		 * for the backing file after a POSIX_FADV_NOREUSE
827 		 * read(2).
828 		 */
829 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
830 		    POSIX_FADV_DONTNEED);
831 	return (error);
832 }
833 
834 /*
835  * File table vnode write routine.
836  */
837 static int
838 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
839     struct thread *td)
840 {
841 	struct vnode *vp;
842 	struct mount *mp;
843 	off_t orig_offset;
844 	int error, ioflag, lock_flags;
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 	if (vp->v_type == VREG)
852 		bwillwrite();
853 	ioflag = IO_UNIT;
854 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
855 		ioflag |= IO_APPEND;
856 	if (fp->f_flag & FNONBLOCK)
857 		ioflag |= IO_NDELAY;
858 	if (fp->f_flag & O_DIRECT)
859 		ioflag |= IO_DIRECT;
860 	if ((fp->f_flag & O_FSYNC) ||
861 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
862 		ioflag |= IO_SYNC;
863 	mp = NULL;
864 	if (vp->v_type != VCHR &&
865 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
866 		goto unlock;
867 
868 	advice = get_advice(fp, uio);
869 
870 	if (MNT_SHARED_WRITES(mp) ||
871 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
872 		lock_flags = LK_SHARED;
873 	} else {
874 		lock_flags = LK_EXCLUSIVE;
875 	}
876 
877 	vn_lock(vp, lock_flags | LK_RETRY);
878 	switch (advice) {
879 	case POSIX_FADV_NORMAL:
880 	case POSIX_FADV_SEQUENTIAL:
881 	case POSIX_FADV_NOREUSE:
882 		ioflag |= sequential_heuristic(uio, fp);
883 		break;
884 	case POSIX_FADV_RANDOM:
885 		/* XXX: Is this correct? */
886 		break;
887 	}
888 	orig_offset = uio->uio_offset;
889 
890 #ifdef MAC
891 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
892 	if (error == 0)
893 #endif
894 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
895 	fp->f_nextoff = uio->uio_offset;
896 	VOP_UNLOCK(vp);
897 	if (vp->v_type != VCHR)
898 		vn_finished_write(mp);
899 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
900 	    orig_offset != uio->uio_offset)
901 		/*
902 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
903 		 * for the backing file after a POSIX_FADV_NOREUSE
904 		 * write(2).
905 		 */
906 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
907 		    POSIX_FADV_DONTNEED);
908 unlock:
909 	return (error);
910 }
911 
912 /*
913  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
914  * prevent the following deadlock:
915  *
916  * Assume that the thread A reads from the vnode vp1 into userspace
917  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
918  * currently not resident, then system ends up with the call chain
919  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
920  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
921  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
922  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
923  * backed by the pages of vnode vp1, and some page in buf2 is not
924  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
925  *
926  * To prevent the lock order reversal and deadlock, vn_io_fault() does
927  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
928  * Instead, it first tries to do the whole range i/o with pagefaults
929  * disabled. If all pages in the i/o buffer are resident and mapped,
930  * VOP will succeed (ignoring the genuine filesystem errors).
931  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
932  * i/o in chunks, with all pages in the chunk prefaulted and held
933  * using vm_fault_quick_hold_pages().
934  *
935  * Filesystems using this deadlock avoidance scheme should use the
936  * array of the held pages from uio, saved in the curthread->td_ma,
937  * instead of doing uiomove().  A helper function
938  * vn_io_fault_uiomove() converts uiomove request into
939  * uiomove_fromphys() over td_ma array.
940  *
941  * Since vnode locks do not cover the whole i/o anymore, rangelocks
942  * make the current i/o request atomic with respect to other i/os and
943  * truncations.
944  */
945 
946 /*
947  * Decode vn_io_fault_args and perform the corresponding i/o.
948  */
949 static int
950 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
951     struct thread *td)
952 {
953 	int error, save;
954 
955 	error = 0;
956 	save = vm_fault_disable_pagefaults();
957 	switch (args->kind) {
958 	case VN_IO_FAULT_FOP:
959 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
960 		    uio, args->cred, args->flags, td);
961 		break;
962 	case VN_IO_FAULT_VOP:
963 		if (uio->uio_rw == UIO_READ) {
964 			error = VOP_READ(args->args.vop_args.vp, uio,
965 			    args->flags, args->cred);
966 		} else if (uio->uio_rw == UIO_WRITE) {
967 			error = VOP_WRITE(args->args.vop_args.vp, uio,
968 			    args->flags, args->cred);
969 		}
970 		break;
971 	default:
972 		panic("vn_io_fault_doio: unknown kind of io %d %d",
973 		    args->kind, uio->uio_rw);
974 	}
975 	vm_fault_enable_pagefaults(save);
976 	return (error);
977 }
978 
979 static int
980 vn_io_fault_touch(char *base, const struct uio *uio)
981 {
982 	int r;
983 
984 	r = fubyte(base);
985 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
986 		return (EFAULT);
987 	return (0);
988 }
989 
990 static int
991 vn_io_fault_prefault_user(const struct uio *uio)
992 {
993 	char *base;
994 	const struct iovec *iov;
995 	size_t len;
996 	ssize_t resid;
997 	int error, i;
998 
999 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1000 	    ("vn_io_fault_prefault userspace"));
1001 
1002 	error = i = 0;
1003 	iov = uio->uio_iov;
1004 	resid = uio->uio_resid;
1005 	base = iov->iov_base;
1006 	len = iov->iov_len;
1007 	while (resid > 0) {
1008 		error = vn_io_fault_touch(base, uio);
1009 		if (error != 0)
1010 			break;
1011 		if (len < PAGE_SIZE) {
1012 			if (len != 0) {
1013 				error = vn_io_fault_touch(base + len - 1, uio);
1014 				if (error != 0)
1015 					break;
1016 				resid -= len;
1017 			}
1018 			if (++i >= uio->uio_iovcnt)
1019 				break;
1020 			iov = uio->uio_iov + i;
1021 			base = iov->iov_base;
1022 			len = iov->iov_len;
1023 		} else {
1024 			len -= PAGE_SIZE;
1025 			base += PAGE_SIZE;
1026 			resid -= PAGE_SIZE;
1027 		}
1028 	}
1029 	return (error);
1030 }
1031 
1032 /*
1033  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1034  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1035  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1036  * into args and call vn_io_fault1() to handle faults during the user
1037  * mode buffer accesses.
1038  */
1039 static int
1040 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1041     struct thread *td)
1042 {
1043 	vm_page_t ma[io_hold_cnt + 2];
1044 	struct uio *uio_clone, short_uio;
1045 	struct iovec short_iovec[1];
1046 	vm_page_t *prev_td_ma;
1047 	vm_prot_t prot;
1048 	vm_offset_t addr, end;
1049 	size_t len, resid;
1050 	ssize_t adv;
1051 	int error, cnt, saveheld, prev_td_ma_cnt;
1052 
1053 	if (vn_io_fault_prefault) {
1054 		error = vn_io_fault_prefault_user(uio);
1055 		if (error != 0)
1056 			return (error); /* Or ignore ? */
1057 	}
1058 
1059 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1060 
1061 	/*
1062 	 * The UFS follows IO_UNIT directive and replays back both
1063 	 * uio_offset and uio_resid if an error is encountered during the
1064 	 * operation.  But, since the iovec may be already advanced,
1065 	 * uio is still in an inconsistent state.
1066 	 *
1067 	 * Cache a copy of the original uio, which is advanced to the redo
1068 	 * point using UIO_NOCOPY below.
1069 	 */
1070 	uio_clone = cloneuio(uio);
1071 	resid = uio->uio_resid;
1072 
1073 	short_uio.uio_segflg = UIO_USERSPACE;
1074 	short_uio.uio_rw = uio->uio_rw;
1075 	short_uio.uio_td = uio->uio_td;
1076 
1077 	error = vn_io_fault_doio(args, uio, td);
1078 	if (error != EFAULT)
1079 		goto out;
1080 
1081 	atomic_add_long(&vn_io_faults_cnt, 1);
1082 	uio_clone->uio_segflg = UIO_NOCOPY;
1083 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1084 	uio_clone->uio_segflg = uio->uio_segflg;
1085 
1086 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1087 	prev_td_ma = td->td_ma;
1088 	prev_td_ma_cnt = td->td_ma_cnt;
1089 
1090 	while (uio_clone->uio_resid != 0) {
1091 		len = uio_clone->uio_iov->iov_len;
1092 		if (len == 0) {
1093 			KASSERT(uio_clone->uio_iovcnt >= 1,
1094 			    ("iovcnt underflow"));
1095 			uio_clone->uio_iov++;
1096 			uio_clone->uio_iovcnt--;
1097 			continue;
1098 		}
1099 		if (len > io_hold_cnt * PAGE_SIZE)
1100 			len = io_hold_cnt * PAGE_SIZE;
1101 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1102 		end = round_page(addr + len);
1103 		if (end < addr) {
1104 			error = EFAULT;
1105 			break;
1106 		}
1107 		cnt = atop(end - trunc_page(addr));
1108 		/*
1109 		 * A perfectly misaligned address and length could cause
1110 		 * both the start and the end of the chunk to use partial
1111 		 * page.  +2 accounts for such a situation.
1112 		 */
1113 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1114 		    addr, len, prot, ma, io_hold_cnt + 2);
1115 		if (cnt == -1) {
1116 			error = EFAULT;
1117 			break;
1118 		}
1119 		short_uio.uio_iov = &short_iovec[0];
1120 		short_iovec[0].iov_base = (void *)addr;
1121 		short_uio.uio_iovcnt = 1;
1122 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1123 		short_uio.uio_offset = uio_clone->uio_offset;
1124 		td->td_ma = ma;
1125 		td->td_ma_cnt = cnt;
1126 
1127 		error = vn_io_fault_doio(args, &short_uio, td);
1128 		vm_page_unhold_pages(ma, cnt);
1129 		adv = len - short_uio.uio_resid;
1130 
1131 		uio_clone->uio_iov->iov_base =
1132 		    (char *)uio_clone->uio_iov->iov_base + adv;
1133 		uio_clone->uio_iov->iov_len -= adv;
1134 		uio_clone->uio_resid -= adv;
1135 		uio_clone->uio_offset += adv;
1136 
1137 		uio->uio_resid -= adv;
1138 		uio->uio_offset += adv;
1139 
1140 		if (error != 0 || adv == 0)
1141 			break;
1142 	}
1143 	td->td_ma = prev_td_ma;
1144 	td->td_ma_cnt = prev_td_ma_cnt;
1145 	curthread_pflags_restore(saveheld);
1146 out:
1147 	free(uio_clone, M_IOV);
1148 	return (error);
1149 }
1150 
1151 static int
1152 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1153     int flags, struct thread *td)
1154 {
1155 	fo_rdwr_t *doio;
1156 	struct vnode *vp;
1157 	void *rl_cookie;
1158 	struct vn_io_fault_args args;
1159 	int error;
1160 
1161 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1162 	vp = fp->f_vnode;
1163 	foffset_lock_uio(fp, uio, flags);
1164 	if (do_vn_io_fault(vp, uio)) {
1165 		args.kind = VN_IO_FAULT_FOP;
1166 		args.args.fop_args.fp = fp;
1167 		args.args.fop_args.doio = doio;
1168 		args.cred = active_cred;
1169 		args.flags = flags | FOF_OFFSET;
1170 		if (uio->uio_rw == UIO_READ) {
1171 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1172 			    uio->uio_offset + uio->uio_resid);
1173 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1174 		    (flags & FOF_OFFSET) == 0) {
1175 			/* For appenders, punt and lock the whole range. */
1176 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1177 		} else {
1178 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1179 			    uio->uio_offset + uio->uio_resid);
1180 		}
1181 		error = vn_io_fault1(vp, uio, &args, td);
1182 		vn_rangelock_unlock(vp, rl_cookie);
1183 	} else {
1184 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1185 	}
1186 	foffset_unlock_uio(fp, uio, flags);
1187 	return (error);
1188 }
1189 
1190 /*
1191  * Helper function to perform the requested uiomove operation using
1192  * the held pages for io->uio_iov[0].iov_base buffer instead of
1193  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1194  * instead of iov_base prevents page faults that could occur due to
1195  * pmap_collect() invalidating the mapping created by
1196  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1197  * object cleanup revoking the write access from page mappings.
1198  *
1199  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1200  * instead of plain uiomove().
1201  */
1202 int
1203 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1204 {
1205 	struct uio transp_uio;
1206 	struct iovec transp_iov[1];
1207 	struct thread *td;
1208 	size_t adv;
1209 	int error, pgadv;
1210 
1211 	td = curthread;
1212 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1213 	    uio->uio_segflg != UIO_USERSPACE)
1214 		return (uiomove(data, xfersize, uio));
1215 
1216 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1217 	transp_iov[0].iov_base = data;
1218 	transp_uio.uio_iov = &transp_iov[0];
1219 	transp_uio.uio_iovcnt = 1;
1220 	if (xfersize > uio->uio_resid)
1221 		xfersize = uio->uio_resid;
1222 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1223 	transp_uio.uio_offset = 0;
1224 	transp_uio.uio_segflg = UIO_SYSSPACE;
1225 	/*
1226 	 * Since transp_iov points to data, and td_ma page array
1227 	 * corresponds to original uio->uio_iov, we need to invert the
1228 	 * direction of the i/o operation as passed to
1229 	 * uiomove_fromphys().
1230 	 */
1231 	switch (uio->uio_rw) {
1232 	case UIO_WRITE:
1233 		transp_uio.uio_rw = UIO_READ;
1234 		break;
1235 	case UIO_READ:
1236 		transp_uio.uio_rw = UIO_WRITE;
1237 		break;
1238 	}
1239 	transp_uio.uio_td = uio->uio_td;
1240 	error = uiomove_fromphys(td->td_ma,
1241 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1242 	    xfersize, &transp_uio);
1243 	adv = xfersize - transp_uio.uio_resid;
1244 	pgadv =
1245 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1246 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1247 	td->td_ma += pgadv;
1248 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1249 	    pgadv));
1250 	td->td_ma_cnt -= pgadv;
1251 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1252 	uio->uio_iov->iov_len -= adv;
1253 	uio->uio_resid -= adv;
1254 	uio->uio_offset += adv;
1255 	return (error);
1256 }
1257 
1258 int
1259 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1260     struct uio *uio)
1261 {
1262 	struct thread *td;
1263 	vm_offset_t iov_base;
1264 	int cnt, pgadv;
1265 
1266 	td = curthread;
1267 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1268 	    uio->uio_segflg != UIO_USERSPACE)
1269 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1270 
1271 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1272 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1273 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1274 	switch (uio->uio_rw) {
1275 	case UIO_WRITE:
1276 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1277 		    offset, cnt);
1278 		break;
1279 	case UIO_READ:
1280 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1281 		    cnt);
1282 		break;
1283 	}
1284 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1285 	td->td_ma += pgadv;
1286 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1287 	    pgadv));
1288 	td->td_ma_cnt -= pgadv;
1289 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1290 	uio->uio_iov->iov_len -= cnt;
1291 	uio->uio_resid -= cnt;
1292 	uio->uio_offset += cnt;
1293 	return (0);
1294 }
1295 
1296 
1297 /*
1298  * File table truncate routine.
1299  */
1300 static int
1301 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1302     struct thread *td)
1303 {
1304 	struct mount *mp;
1305 	struct vnode *vp;
1306 	void *rl_cookie;
1307 	int error;
1308 
1309 	vp = fp->f_vnode;
1310 
1311 	/*
1312 	 * Lock the whole range for truncation.  Otherwise split i/o
1313 	 * might happen partly before and partly after the truncation.
1314 	 */
1315 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1316 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1317 	if (error)
1318 		goto out1;
1319 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1320 	AUDIT_ARG_VNODE1(vp);
1321 	if (vp->v_type == VDIR) {
1322 		error = EISDIR;
1323 		goto out;
1324 	}
1325 #ifdef MAC
1326 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1327 	if (error)
1328 		goto out;
1329 #endif
1330 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1331 	    fp->f_cred);
1332 out:
1333 	VOP_UNLOCK(vp);
1334 	vn_finished_write(mp);
1335 out1:
1336 	vn_rangelock_unlock(vp, rl_cookie);
1337 	return (error);
1338 }
1339 
1340 /*
1341  * Truncate a file that is already locked.
1342  */
1343 int
1344 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1345     struct ucred *cred)
1346 {
1347 	struct vattr vattr;
1348 	int error;
1349 
1350 	error = VOP_ADD_WRITECOUNT(vp, 1);
1351 	if (error == 0) {
1352 		VATTR_NULL(&vattr);
1353 		vattr.va_size = length;
1354 		if (sync)
1355 			vattr.va_vaflags |= VA_SYNC;
1356 		error = VOP_SETATTR(vp, &vattr, cred);
1357 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1358 	}
1359 	return (error);
1360 }
1361 
1362 /*
1363  * File table vnode stat routine.
1364  */
1365 static int
1366 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1367     struct thread *td)
1368 {
1369 	struct vnode *vp = fp->f_vnode;
1370 	int error;
1371 
1372 	vn_lock(vp, LK_SHARED | LK_RETRY);
1373 	error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1374 	VOP_UNLOCK(vp);
1375 
1376 	return (error);
1377 }
1378 
1379 /*
1380  * Stat a vnode; implementation for the stat syscall
1381  */
1382 int
1383 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred,
1384     struct ucred *file_cred, struct thread *td)
1385 {
1386 	struct vattr vattr;
1387 	struct vattr *vap;
1388 	int error;
1389 	u_short mode;
1390 
1391 	AUDIT_ARG_VNODE1(vp);
1392 #ifdef MAC
1393 	error = mac_vnode_check_stat(active_cred, file_cred, vp);
1394 	if (error)
1395 		return (error);
1396 #endif
1397 
1398 	vap = &vattr;
1399 
1400 	/*
1401 	 * Initialize defaults for new and unusual fields, so that file
1402 	 * systems which don't support these fields don't need to know
1403 	 * about them.
1404 	 */
1405 	vap->va_birthtime.tv_sec = -1;
1406 	vap->va_birthtime.tv_nsec = 0;
1407 	vap->va_fsid = VNOVAL;
1408 	vap->va_rdev = NODEV;
1409 
1410 	error = VOP_GETATTR(vp, vap, active_cred);
1411 	if (error)
1412 		return (error);
1413 
1414 	/*
1415 	 * Zero the spare stat fields
1416 	 */
1417 	bzero(sb, sizeof *sb);
1418 
1419 	/*
1420 	 * Copy from vattr table
1421 	 */
1422 	if (vap->va_fsid != VNOVAL)
1423 		sb->st_dev = vap->va_fsid;
1424 	else
1425 		sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1426 	sb->st_ino = vap->va_fileid;
1427 	mode = vap->va_mode;
1428 	switch (vap->va_type) {
1429 	case VREG:
1430 		mode |= S_IFREG;
1431 		break;
1432 	case VDIR:
1433 		mode |= S_IFDIR;
1434 		break;
1435 	case VBLK:
1436 		mode |= S_IFBLK;
1437 		break;
1438 	case VCHR:
1439 		mode |= S_IFCHR;
1440 		break;
1441 	case VLNK:
1442 		mode |= S_IFLNK;
1443 		break;
1444 	case VSOCK:
1445 		mode |= S_IFSOCK;
1446 		break;
1447 	case VFIFO:
1448 		mode |= S_IFIFO;
1449 		break;
1450 	default:
1451 		return (EBADF);
1452 	}
1453 	sb->st_mode = mode;
1454 	sb->st_nlink = vap->va_nlink;
1455 	sb->st_uid = vap->va_uid;
1456 	sb->st_gid = vap->va_gid;
1457 	sb->st_rdev = vap->va_rdev;
1458 	if (vap->va_size > OFF_MAX)
1459 		return (EOVERFLOW);
1460 	sb->st_size = vap->va_size;
1461 	sb->st_atim.tv_sec = vap->va_atime.tv_sec;
1462 	sb->st_atim.tv_nsec = vap->va_atime.tv_nsec;
1463 	sb->st_mtim.tv_sec = vap->va_mtime.tv_sec;
1464 	sb->st_mtim.tv_nsec = vap->va_mtime.tv_nsec;
1465 	sb->st_ctim.tv_sec = vap->va_ctime.tv_sec;
1466 	sb->st_ctim.tv_nsec = vap->va_ctime.tv_nsec;
1467 	sb->st_birthtim.tv_sec = vap->va_birthtime.tv_sec;
1468 	sb->st_birthtim.tv_nsec = vap->va_birthtime.tv_nsec;
1469 
1470         /*
1471 	 * According to www.opengroup.org, the meaning of st_blksize is
1472 	 *   "a filesystem-specific preferred I/O block size for this
1473 	 *    object.  In some filesystem types, this may vary from file
1474 	 *    to file"
1475 	 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1476 	 */
1477 
1478 	sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1479 
1480 	sb->st_flags = vap->va_flags;
1481 	if (priv_check(td, PRIV_VFS_GENERATION))
1482 		sb->st_gen = 0;
1483 	else
1484 		sb->st_gen = vap->va_gen;
1485 
1486 	sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1487 	return (0);
1488 }
1489 
1490 /*
1491  * File table vnode ioctl routine.
1492  */
1493 static int
1494 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1495     struct thread *td)
1496 {
1497 	struct vattr vattr;
1498 	struct vnode *vp;
1499 	struct fiobmap2_arg *bmarg;
1500 	int error;
1501 
1502 	vp = fp->f_vnode;
1503 	switch (vp->v_type) {
1504 	case VDIR:
1505 	case VREG:
1506 		switch (com) {
1507 		case FIONREAD:
1508 			vn_lock(vp, LK_SHARED | LK_RETRY);
1509 			error = VOP_GETATTR(vp, &vattr, active_cred);
1510 			VOP_UNLOCK(vp);
1511 			if (error == 0)
1512 				*(int *)data = vattr.va_size - fp->f_offset;
1513 			return (error);
1514 		case FIOBMAP2:
1515 			bmarg = (struct fiobmap2_arg *)data;
1516 			vn_lock(vp, LK_SHARED | LK_RETRY);
1517 #ifdef MAC
1518 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1519 			    vp);
1520 			if (error == 0)
1521 #endif
1522 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1523 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1524 			VOP_UNLOCK(vp);
1525 			return (error);
1526 		case FIONBIO:
1527 		case FIOASYNC:
1528 			return (0);
1529 		default:
1530 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1531 			    active_cred, td));
1532 		}
1533 		break;
1534 	case VCHR:
1535 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1536 		    active_cred, td));
1537 	default:
1538 		return (ENOTTY);
1539 	}
1540 }
1541 
1542 /*
1543  * File table vnode poll routine.
1544  */
1545 static int
1546 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1547     struct thread *td)
1548 {
1549 	struct vnode *vp;
1550 	int error;
1551 
1552 	vp = fp->f_vnode;
1553 #ifdef MAC
1554 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1555 	AUDIT_ARG_VNODE1(vp);
1556 	error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1557 	VOP_UNLOCK(vp);
1558 	if (!error)
1559 #endif
1560 
1561 	error = VOP_POLL(vp, events, fp->f_cred, td);
1562 	return (error);
1563 }
1564 
1565 /*
1566  * Acquire the requested lock and then check for validity.  LK_RETRY
1567  * permits vn_lock to return doomed vnodes.
1568  */
1569 static int __noinline
1570 _vn_lock_fallback(struct vnode *vp, int flags, char *file, int line, int error)
1571 {
1572 
1573 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1574 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1575 
1576 	if (error == 0)
1577 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1578 
1579 	if ((flags & LK_RETRY) == 0) {
1580 		if (error == 0) {
1581 			VOP_UNLOCK(vp);
1582 			error = ENOENT;
1583 		}
1584 		return (error);
1585 	}
1586 
1587 	/*
1588 	 * LK_RETRY case.
1589 	 *
1590 	 * Nothing to do if we got the lock.
1591 	 */
1592 	if (error == 0)
1593 		return (0);
1594 
1595 	/*
1596 	 * Interlock was dropped by the call in _vn_lock.
1597 	 */
1598 	flags &= ~LK_INTERLOCK;
1599 	do {
1600 		error = VOP_LOCK1(vp, flags, file, line);
1601 	} while (error != 0);
1602 	return (0);
1603 }
1604 
1605 int
1606 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1607 {
1608 	int error;
1609 
1610 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp, ("vn_lock: no locktype"));
1611 	VNASSERT(vp->v_holdcnt != 0, vp, ("vn_lock: zero hold count"));
1612 	error = VOP_LOCK1(vp, flags, file, line);
1613 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1614 		return (_vn_lock_fallback(vp, flags, file, line, error));
1615 	return (0);
1616 }
1617 
1618 /*
1619  * File table vnode close routine.
1620  */
1621 static int
1622 vn_closefile(struct file *fp, struct thread *td)
1623 {
1624 	struct vnode *vp;
1625 	struct flock lf;
1626 	int error;
1627 	bool ref;
1628 
1629 	vp = fp->f_vnode;
1630 	fp->f_ops = &badfileops;
1631 	ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1632 
1633 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1634 
1635 	if (__predict_false(ref)) {
1636 		lf.l_whence = SEEK_SET;
1637 		lf.l_start = 0;
1638 		lf.l_len = 0;
1639 		lf.l_type = F_UNLCK;
1640 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1641 		vrele(vp);
1642 	}
1643 	return (error);
1644 }
1645 
1646 static bool
1647 vn_suspendable(struct mount *mp)
1648 {
1649 
1650 	return (mp->mnt_op->vfs_susp_clean != NULL);
1651 }
1652 
1653 /*
1654  * Preparing to start a filesystem write operation. If the operation is
1655  * permitted, then we bump the count of operations in progress and
1656  * proceed. If a suspend request is in progress, we wait until the
1657  * suspension is over, and then proceed.
1658  */
1659 static int
1660 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1661 {
1662 	int error, mflags;
1663 
1664 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1665 	    vfs_op_thread_enter(mp)) {
1666 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1667 		vfs_mp_count_add_pcpu(mp, writeopcount, 1);
1668 		vfs_op_thread_exit(mp);
1669 		return (0);
1670 	}
1671 
1672 	if (mplocked)
1673 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1674 	else
1675 		MNT_ILOCK(mp);
1676 
1677 	error = 0;
1678 
1679 	/*
1680 	 * Check on status of suspension.
1681 	 */
1682 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1683 	    mp->mnt_susp_owner != curthread) {
1684 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1685 		    (flags & PCATCH) : 0) | (PUSER - 1);
1686 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1687 			if (flags & V_NOWAIT) {
1688 				error = EWOULDBLOCK;
1689 				goto unlock;
1690 			}
1691 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1692 			    "suspfs", 0);
1693 			if (error)
1694 				goto unlock;
1695 		}
1696 	}
1697 	if (flags & V_XSLEEP)
1698 		goto unlock;
1699 	mp->mnt_writeopcount++;
1700 unlock:
1701 	if (error != 0 || (flags & V_XSLEEP) != 0)
1702 		MNT_REL(mp);
1703 	MNT_IUNLOCK(mp);
1704 	return (error);
1705 }
1706 
1707 int
1708 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1709 {
1710 	struct mount *mp;
1711 	int error;
1712 
1713 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1714 	    ("V_MNTREF requires mp"));
1715 
1716 	error = 0;
1717 	/*
1718 	 * If a vnode is provided, get and return the mount point that
1719 	 * to which it will write.
1720 	 */
1721 	if (vp != NULL) {
1722 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1723 			*mpp = NULL;
1724 			if (error != EOPNOTSUPP)
1725 				return (error);
1726 			return (0);
1727 		}
1728 	}
1729 	if ((mp = *mpp) == NULL)
1730 		return (0);
1731 
1732 	if (!vn_suspendable(mp)) {
1733 		if (vp != NULL || (flags & V_MNTREF) != 0)
1734 			vfs_rel(mp);
1735 		return (0);
1736 	}
1737 
1738 	/*
1739 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1740 	 * a vfs_ref().
1741 	 * As long as a vnode is not provided we need to acquire a
1742 	 * refcount for the provided mountpoint too, in order to
1743 	 * emulate a vfs_ref().
1744 	 */
1745 	if (vp == NULL && (flags & V_MNTREF) == 0)
1746 		vfs_ref(mp);
1747 
1748 	return (vn_start_write_refed(mp, flags, false));
1749 }
1750 
1751 /*
1752  * Secondary suspension. Used by operations such as vop_inactive
1753  * routines that are needed by the higher level functions. These
1754  * are allowed to proceed until all the higher level functions have
1755  * completed (indicated by mnt_writeopcount dropping to zero). At that
1756  * time, these operations are halted until the suspension is over.
1757  */
1758 int
1759 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1760 {
1761 	struct mount *mp;
1762 	int error;
1763 
1764 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1765 	    ("V_MNTREF requires mp"));
1766 
1767  retry:
1768 	if (vp != NULL) {
1769 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1770 			*mpp = NULL;
1771 			if (error != EOPNOTSUPP)
1772 				return (error);
1773 			return (0);
1774 		}
1775 	}
1776 	/*
1777 	 * If we are not suspended or have not yet reached suspended
1778 	 * mode, then let the operation proceed.
1779 	 */
1780 	if ((mp = *mpp) == NULL)
1781 		return (0);
1782 
1783 	if (!vn_suspendable(mp)) {
1784 		if (vp != NULL || (flags & V_MNTREF) != 0)
1785 			vfs_rel(mp);
1786 		return (0);
1787 	}
1788 
1789 	/*
1790 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1791 	 * a vfs_ref().
1792 	 * As long as a vnode is not provided we need to acquire a
1793 	 * refcount for the provided mountpoint too, in order to
1794 	 * emulate a vfs_ref().
1795 	 */
1796 	MNT_ILOCK(mp);
1797 	if (vp == NULL && (flags & V_MNTREF) == 0)
1798 		MNT_REF(mp);
1799 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1800 		mp->mnt_secondary_writes++;
1801 		mp->mnt_secondary_accwrites++;
1802 		MNT_IUNLOCK(mp);
1803 		return (0);
1804 	}
1805 	if (flags & V_NOWAIT) {
1806 		MNT_REL(mp);
1807 		MNT_IUNLOCK(mp);
1808 		return (EWOULDBLOCK);
1809 	}
1810 	/*
1811 	 * Wait for the suspension to finish.
1812 	 */
1813 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1814 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1815 	    "suspfs", 0);
1816 	vfs_rel(mp);
1817 	if (error == 0)
1818 		goto retry;
1819 	return (error);
1820 }
1821 
1822 /*
1823  * Filesystem write operation has completed. If we are suspending and this
1824  * operation is the last one, notify the suspender that the suspension is
1825  * now in effect.
1826  */
1827 void
1828 vn_finished_write(struct mount *mp)
1829 {
1830 	int c;
1831 
1832 	if (mp == NULL || !vn_suspendable(mp))
1833 		return;
1834 
1835 	if (vfs_op_thread_enter(mp)) {
1836 		vfs_mp_count_sub_pcpu(mp, writeopcount, 1);
1837 		vfs_mp_count_sub_pcpu(mp, ref, 1);
1838 		vfs_op_thread_exit(mp);
1839 		return;
1840 	}
1841 
1842 	MNT_ILOCK(mp);
1843 	vfs_assert_mount_counters(mp);
1844 	MNT_REL(mp);
1845 	c = --mp->mnt_writeopcount;
1846 	if (mp->mnt_vfs_ops == 0) {
1847 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1848 		MNT_IUNLOCK(mp);
1849 		return;
1850 	}
1851 	if (c < 0)
1852 		vfs_dump_mount_counters(mp);
1853 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1854 		wakeup(&mp->mnt_writeopcount);
1855 	MNT_IUNLOCK(mp);
1856 }
1857 
1858 
1859 /*
1860  * Filesystem secondary write operation has completed. If we are
1861  * suspending and this operation is the last one, notify the suspender
1862  * that the suspension is now in effect.
1863  */
1864 void
1865 vn_finished_secondary_write(struct mount *mp)
1866 {
1867 	if (mp == NULL || !vn_suspendable(mp))
1868 		return;
1869 	MNT_ILOCK(mp);
1870 	MNT_REL(mp);
1871 	mp->mnt_secondary_writes--;
1872 	if (mp->mnt_secondary_writes < 0)
1873 		panic("vn_finished_secondary_write: neg cnt");
1874 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1875 	    mp->mnt_secondary_writes <= 0)
1876 		wakeup(&mp->mnt_secondary_writes);
1877 	MNT_IUNLOCK(mp);
1878 }
1879 
1880 
1881 
1882 /*
1883  * Request a filesystem to suspend write operations.
1884  */
1885 int
1886 vfs_write_suspend(struct mount *mp, int flags)
1887 {
1888 	int error;
1889 
1890 	MPASS(vn_suspendable(mp));
1891 
1892 	vfs_op_enter(mp);
1893 
1894 	MNT_ILOCK(mp);
1895 	vfs_assert_mount_counters(mp);
1896 	if (mp->mnt_susp_owner == curthread) {
1897 		vfs_op_exit_locked(mp);
1898 		MNT_IUNLOCK(mp);
1899 		return (EALREADY);
1900 	}
1901 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
1902 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1903 
1904 	/*
1905 	 * Unmount holds a write reference on the mount point.  If we
1906 	 * own busy reference and drain for writers, we deadlock with
1907 	 * the reference draining in the unmount path.  Callers of
1908 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1909 	 * vfs_busy() reference is owned and caller is not in the
1910 	 * unmount context.
1911 	 */
1912 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1913 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1914 		vfs_op_exit_locked(mp);
1915 		MNT_IUNLOCK(mp);
1916 		return (EBUSY);
1917 	}
1918 
1919 	mp->mnt_kern_flag |= MNTK_SUSPEND;
1920 	mp->mnt_susp_owner = curthread;
1921 	if (mp->mnt_writeopcount > 0)
1922 		(void) msleep(&mp->mnt_writeopcount,
1923 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1924 	else
1925 		MNT_IUNLOCK(mp);
1926 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
1927 		vfs_write_resume(mp, 0);
1928 		vfs_op_exit(mp);
1929 	}
1930 	return (error);
1931 }
1932 
1933 /*
1934  * Request a filesystem to resume write operations.
1935  */
1936 void
1937 vfs_write_resume(struct mount *mp, int flags)
1938 {
1939 
1940 	MPASS(vn_suspendable(mp));
1941 
1942 	MNT_ILOCK(mp);
1943 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1944 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1945 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1946 				       MNTK_SUSPENDED);
1947 		mp->mnt_susp_owner = NULL;
1948 		wakeup(&mp->mnt_writeopcount);
1949 		wakeup(&mp->mnt_flag);
1950 		curthread->td_pflags &= ~TDP_IGNSUSP;
1951 		if ((flags & VR_START_WRITE) != 0) {
1952 			MNT_REF(mp);
1953 			mp->mnt_writeopcount++;
1954 		}
1955 		MNT_IUNLOCK(mp);
1956 		if ((flags & VR_NO_SUSPCLR) == 0)
1957 			VFS_SUSP_CLEAN(mp);
1958 		vfs_op_exit(mp);
1959 	} else if ((flags & VR_START_WRITE) != 0) {
1960 		MNT_REF(mp);
1961 		vn_start_write_refed(mp, 0, true);
1962 	} else {
1963 		MNT_IUNLOCK(mp);
1964 	}
1965 }
1966 
1967 /*
1968  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1969  * methods.
1970  */
1971 int
1972 vfs_write_suspend_umnt(struct mount *mp)
1973 {
1974 	int error;
1975 
1976 	MPASS(vn_suspendable(mp));
1977 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1978 	    ("vfs_write_suspend_umnt: recursed"));
1979 
1980 	/* dounmount() already called vn_start_write(). */
1981 	for (;;) {
1982 		vn_finished_write(mp);
1983 		error = vfs_write_suspend(mp, 0);
1984 		if (error != 0) {
1985 			vn_start_write(NULL, &mp, V_WAIT);
1986 			return (error);
1987 		}
1988 		MNT_ILOCK(mp);
1989 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1990 			break;
1991 		MNT_IUNLOCK(mp);
1992 		vn_start_write(NULL, &mp, V_WAIT);
1993 	}
1994 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1995 	wakeup(&mp->mnt_flag);
1996 	MNT_IUNLOCK(mp);
1997 	curthread->td_pflags |= TDP_IGNSUSP;
1998 	return (0);
1999 }
2000 
2001 /*
2002  * Implement kqueues for files by translating it to vnode operation.
2003  */
2004 static int
2005 vn_kqfilter(struct file *fp, struct knote *kn)
2006 {
2007 
2008 	return (VOP_KQFILTER(fp->f_vnode, kn));
2009 }
2010 
2011 /*
2012  * Simplified in-kernel wrapper calls for extended attribute access.
2013  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2014  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2015  */
2016 int
2017 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2018     const char *attrname, int *buflen, char *buf, struct thread *td)
2019 {
2020 	struct uio	auio;
2021 	struct iovec	iov;
2022 	int	error;
2023 
2024 	iov.iov_len = *buflen;
2025 	iov.iov_base = buf;
2026 
2027 	auio.uio_iov = &iov;
2028 	auio.uio_iovcnt = 1;
2029 	auio.uio_rw = UIO_READ;
2030 	auio.uio_segflg = UIO_SYSSPACE;
2031 	auio.uio_td = td;
2032 	auio.uio_offset = 0;
2033 	auio.uio_resid = *buflen;
2034 
2035 	if ((ioflg & IO_NODELOCKED) == 0)
2036 		vn_lock(vp, LK_SHARED | LK_RETRY);
2037 
2038 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2039 
2040 	/* authorize attribute retrieval as kernel */
2041 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2042 	    td);
2043 
2044 	if ((ioflg & IO_NODELOCKED) == 0)
2045 		VOP_UNLOCK(vp);
2046 
2047 	if (error == 0) {
2048 		*buflen = *buflen - auio.uio_resid;
2049 	}
2050 
2051 	return (error);
2052 }
2053 
2054 /*
2055  * XXX failure mode if partially written?
2056  */
2057 int
2058 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2059     const char *attrname, int buflen, char *buf, struct thread *td)
2060 {
2061 	struct uio	auio;
2062 	struct iovec	iov;
2063 	struct mount	*mp;
2064 	int	error;
2065 
2066 	iov.iov_len = buflen;
2067 	iov.iov_base = buf;
2068 
2069 	auio.uio_iov = &iov;
2070 	auio.uio_iovcnt = 1;
2071 	auio.uio_rw = UIO_WRITE;
2072 	auio.uio_segflg = UIO_SYSSPACE;
2073 	auio.uio_td = td;
2074 	auio.uio_offset = 0;
2075 	auio.uio_resid = buflen;
2076 
2077 	if ((ioflg & IO_NODELOCKED) == 0) {
2078 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2079 			return (error);
2080 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2081 	}
2082 
2083 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2084 
2085 	/* authorize attribute setting as kernel */
2086 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2087 
2088 	if ((ioflg & IO_NODELOCKED) == 0) {
2089 		vn_finished_write(mp);
2090 		VOP_UNLOCK(vp);
2091 	}
2092 
2093 	return (error);
2094 }
2095 
2096 int
2097 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2098     const char *attrname, struct thread *td)
2099 {
2100 	struct mount	*mp;
2101 	int	error;
2102 
2103 	if ((ioflg & IO_NODELOCKED) == 0) {
2104 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2105 			return (error);
2106 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2107 	}
2108 
2109 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2110 
2111 	/* authorize attribute removal as kernel */
2112 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2113 	if (error == EOPNOTSUPP)
2114 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2115 		    NULL, td);
2116 
2117 	if ((ioflg & IO_NODELOCKED) == 0) {
2118 		vn_finished_write(mp);
2119 		VOP_UNLOCK(vp);
2120 	}
2121 
2122 	return (error);
2123 }
2124 
2125 static int
2126 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2127     struct vnode **rvp)
2128 {
2129 
2130 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2131 }
2132 
2133 int
2134 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2135 {
2136 
2137 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2138 	    lkflags, rvp));
2139 }
2140 
2141 int
2142 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2143     int lkflags, struct vnode **rvp)
2144 {
2145 	struct mount *mp;
2146 	int ltype, error;
2147 
2148 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2149 	mp = vp->v_mount;
2150 	ltype = VOP_ISLOCKED(vp);
2151 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2152 	    ("vn_vget_ino: vp not locked"));
2153 	error = vfs_busy(mp, MBF_NOWAIT);
2154 	if (error != 0) {
2155 		vfs_ref(mp);
2156 		VOP_UNLOCK(vp);
2157 		error = vfs_busy(mp, 0);
2158 		vn_lock(vp, ltype | LK_RETRY);
2159 		vfs_rel(mp);
2160 		if (error != 0)
2161 			return (ENOENT);
2162 		if (VN_IS_DOOMED(vp)) {
2163 			vfs_unbusy(mp);
2164 			return (ENOENT);
2165 		}
2166 	}
2167 	VOP_UNLOCK(vp);
2168 	error = alloc(mp, alloc_arg, lkflags, rvp);
2169 	vfs_unbusy(mp);
2170 	if (error != 0 || *rvp != vp)
2171 		vn_lock(vp, ltype | LK_RETRY);
2172 	if (VN_IS_DOOMED(vp)) {
2173 		if (error == 0) {
2174 			if (*rvp == vp)
2175 				vunref(vp);
2176 			else
2177 				vput(*rvp);
2178 		}
2179 		error = ENOENT;
2180 	}
2181 	return (error);
2182 }
2183 
2184 int
2185 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2186     struct thread *td)
2187 {
2188 
2189 	if (vp->v_type != VREG || td == NULL)
2190 		return (0);
2191 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2192 	    lim_cur(td, RLIMIT_FSIZE)) {
2193 		PROC_LOCK(td->td_proc);
2194 		kern_psignal(td->td_proc, SIGXFSZ);
2195 		PROC_UNLOCK(td->td_proc);
2196 		return (EFBIG);
2197 	}
2198 	return (0);
2199 }
2200 
2201 int
2202 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2203     struct thread *td)
2204 {
2205 	struct vnode *vp;
2206 
2207 	vp = fp->f_vnode;
2208 #ifdef AUDIT
2209 	vn_lock(vp, LK_SHARED | LK_RETRY);
2210 	AUDIT_ARG_VNODE1(vp);
2211 	VOP_UNLOCK(vp);
2212 #endif
2213 	return (setfmode(td, active_cred, vp, mode));
2214 }
2215 
2216 int
2217 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2218     struct thread *td)
2219 {
2220 	struct vnode *vp;
2221 
2222 	vp = fp->f_vnode;
2223 #ifdef AUDIT
2224 	vn_lock(vp, LK_SHARED | LK_RETRY);
2225 	AUDIT_ARG_VNODE1(vp);
2226 	VOP_UNLOCK(vp);
2227 #endif
2228 	return (setfown(td, active_cred, vp, uid, gid));
2229 }
2230 
2231 void
2232 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2233 {
2234 	vm_object_t object;
2235 
2236 	if ((object = vp->v_object) == NULL)
2237 		return;
2238 	VM_OBJECT_WLOCK(object);
2239 	vm_object_page_remove(object, start, end, 0);
2240 	VM_OBJECT_WUNLOCK(object);
2241 }
2242 
2243 int
2244 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2245 {
2246 	struct vattr va;
2247 	daddr_t bn, bnp;
2248 	uint64_t bsize;
2249 	off_t noff;
2250 	int error;
2251 
2252 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2253 	    ("Wrong command %lu", cmd));
2254 
2255 	if (vn_lock(vp, LK_SHARED) != 0)
2256 		return (EBADF);
2257 	if (vp->v_type != VREG) {
2258 		error = ENOTTY;
2259 		goto unlock;
2260 	}
2261 	error = VOP_GETATTR(vp, &va, cred);
2262 	if (error != 0)
2263 		goto unlock;
2264 	noff = *off;
2265 	if (noff >= va.va_size) {
2266 		error = ENXIO;
2267 		goto unlock;
2268 	}
2269 	bsize = vp->v_mount->mnt_stat.f_iosize;
2270 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2271 	    noff % bsize) {
2272 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2273 		if (error == EOPNOTSUPP) {
2274 			error = ENOTTY;
2275 			goto unlock;
2276 		}
2277 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2278 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2279 			noff = bn * bsize;
2280 			if (noff < *off)
2281 				noff = *off;
2282 			goto unlock;
2283 		}
2284 	}
2285 	if (noff > va.va_size)
2286 		noff = va.va_size;
2287 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2288 	if (cmd == FIOSEEKDATA)
2289 		error = ENXIO;
2290 unlock:
2291 	VOP_UNLOCK(vp);
2292 	if (error == 0)
2293 		*off = noff;
2294 	return (error);
2295 }
2296 
2297 int
2298 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2299 {
2300 	struct ucred *cred;
2301 	struct vnode *vp;
2302 	struct vattr vattr;
2303 	off_t foffset, size;
2304 	int error, noneg;
2305 
2306 	cred = td->td_ucred;
2307 	vp = fp->f_vnode;
2308 	foffset = foffset_lock(fp, 0);
2309 	noneg = (vp->v_type != VCHR);
2310 	error = 0;
2311 	switch (whence) {
2312 	case L_INCR:
2313 		if (noneg &&
2314 		    (foffset < 0 ||
2315 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2316 			error = EOVERFLOW;
2317 			break;
2318 		}
2319 		offset += foffset;
2320 		break;
2321 	case L_XTND:
2322 		vn_lock(vp, LK_SHARED | LK_RETRY);
2323 		error = VOP_GETATTR(vp, &vattr, cred);
2324 		VOP_UNLOCK(vp);
2325 		if (error)
2326 			break;
2327 
2328 		/*
2329 		 * If the file references a disk device, then fetch
2330 		 * the media size and use that to determine the ending
2331 		 * offset.
2332 		 */
2333 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2334 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2335 			vattr.va_size = size;
2336 		if (noneg &&
2337 		    (vattr.va_size > OFF_MAX ||
2338 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2339 			error = EOVERFLOW;
2340 			break;
2341 		}
2342 		offset += vattr.va_size;
2343 		break;
2344 	case L_SET:
2345 		break;
2346 	case SEEK_DATA:
2347 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2348 		if (error == ENOTTY)
2349 			error = EINVAL;
2350 		break;
2351 	case SEEK_HOLE:
2352 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2353 		if (error == ENOTTY)
2354 			error = EINVAL;
2355 		break;
2356 	default:
2357 		error = EINVAL;
2358 	}
2359 	if (error == 0 && noneg && offset < 0)
2360 		error = EINVAL;
2361 	if (error != 0)
2362 		goto drop;
2363 	VFS_KNOTE_UNLOCKED(vp, 0);
2364 	td->td_uretoff.tdu_off = offset;
2365 drop:
2366 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2367 	return (error);
2368 }
2369 
2370 int
2371 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2372     struct thread *td)
2373 {
2374 	int error;
2375 
2376 	/*
2377 	 * Grant permission if the caller is the owner of the file, or
2378 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2379 	 * on the file.  If the time pointer is null, then write
2380 	 * permission on the file is also sufficient.
2381 	 *
2382 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2383 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2384 	 * will be allowed to set the times [..] to the current
2385 	 * server time.
2386 	 */
2387 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2388 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2389 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2390 	return (error);
2391 }
2392 
2393 int
2394 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2395 {
2396 	struct vnode *vp;
2397 	int error;
2398 
2399 	if (fp->f_type == DTYPE_FIFO)
2400 		kif->kf_type = KF_TYPE_FIFO;
2401 	else
2402 		kif->kf_type = KF_TYPE_VNODE;
2403 	vp = fp->f_vnode;
2404 	vref(vp);
2405 	FILEDESC_SUNLOCK(fdp);
2406 	error = vn_fill_kinfo_vnode(vp, kif);
2407 	vrele(vp);
2408 	FILEDESC_SLOCK(fdp);
2409 	return (error);
2410 }
2411 
2412 static inline void
2413 vn_fill_junk(struct kinfo_file *kif)
2414 {
2415 	size_t len, olen;
2416 
2417 	/*
2418 	 * Simulate vn_fullpath returning changing values for a given
2419 	 * vp during e.g. coredump.
2420 	 */
2421 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2422 	olen = strlen(kif->kf_path);
2423 	if (len < olen)
2424 		strcpy(&kif->kf_path[len - 1], "$");
2425 	else
2426 		for (; olen < len; olen++)
2427 			strcpy(&kif->kf_path[olen], "A");
2428 }
2429 
2430 int
2431 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2432 {
2433 	struct vattr va;
2434 	char *fullpath, *freepath;
2435 	int error;
2436 
2437 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2438 	freepath = NULL;
2439 	fullpath = "-";
2440 	error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2441 	if (error == 0) {
2442 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2443 	}
2444 	if (freepath != NULL)
2445 		free(freepath, M_TEMP);
2446 
2447 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2448 		vn_fill_junk(kif);
2449 	);
2450 
2451 	/*
2452 	 * Retrieve vnode attributes.
2453 	 */
2454 	va.va_fsid = VNOVAL;
2455 	va.va_rdev = NODEV;
2456 	vn_lock(vp, LK_SHARED | LK_RETRY);
2457 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2458 	VOP_UNLOCK(vp);
2459 	if (error != 0)
2460 		return (error);
2461 	if (va.va_fsid != VNOVAL)
2462 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2463 	else
2464 		kif->kf_un.kf_file.kf_file_fsid =
2465 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2466 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2467 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2468 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2469 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2470 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2471 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2472 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2473 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2474 	return (0);
2475 }
2476 
2477 int
2478 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2479     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2480     struct thread *td)
2481 {
2482 #ifdef HWPMC_HOOKS
2483 	struct pmckern_map_in pkm;
2484 #endif
2485 	struct mount *mp;
2486 	struct vnode *vp;
2487 	vm_object_t object;
2488 	vm_prot_t maxprot;
2489 	boolean_t writecounted;
2490 	int error;
2491 
2492 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2493     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2494 	/*
2495 	 * POSIX shared-memory objects are defined to have
2496 	 * kernel persistence, and are not defined to support
2497 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2498 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2499 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2500 	 * flag to request this behavior.
2501 	 */
2502 	if ((fp->f_flag & FPOSIXSHM) != 0)
2503 		flags |= MAP_NOSYNC;
2504 #endif
2505 	vp = fp->f_vnode;
2506 
2507 	/*
2508 	 * Ensure that file and memory protections are
2509 	 * compatible.  Note that we only worry about
2510 	 * writability if mapping is shared; in this case,
2511 	 * current and max prot are dictated by the open file.
2512 	 * XXX use the vnode instead?  Problem is: what
2513 	 * credentials do we use for determination? What if
2514 	 * proc does a setuid?
2515 	 */
2516 	mp = vp->v_mount;
2517 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2518 		maxprot = VM_PROT_NONE;
2519 		if ((prot & VM_PROT_EXECUTE) != 0)
2520 			return (EACCES);
2521 	} else
2522 		maxprot = VM_PROT_EXECUTE;
2523 	if ((fp->f_flag & FREAD) != 0)
2524 		maxprot |= VM_PROT_READ;
2525 	else if ((prot & VM_PROT_READ) != 0)
2526 		return (EACCES);
2527 
2528 	/*
2529 	 * If we are sharing potential changes via MAP_SHARED and we
2530 	 * are trying to get write permission although we opened it
2531 	 * without asking for it, bail out.
2532 	 */
2533 	if ((flags & MAP_SHARED) != 0) {
2534 		if ((fp->f_flag & FWRITE) != 0)
2535 			maxprot |= VM_PROT_WRITE;
2536 		else if ((prot & VM_PROT_WRITE) != 0)
2537 			return (EACCES);
2538 	} else {
2539 		maxprot |= VM_PROT_WRITE;
2540 		cap_maxprot |= VM_PROT_WRITE;
2541 	}
2542 	maxprot &= cap_maxprot;
2543 
2544 	/*
2545 	 * For regular files and shared memory, POSIX requires that
2546 	 * the value of foff be a legitimate offset within the data
2547 	 * object.  In particular, negative offsets are invalid.
2548 	 * Blocking negative offsets and overflows here avoids
2549 	 * possible wraparound or user-level access into reserved
2550 	 * ranges of the data object later.  In contrast, POSIX does
2551 	 * not dictate how offsets are used by device drivers, so in
2552 	 * the case of a device mapping a negative offset is passed
2553 	 * on.
2554 	 */
2555 	if (
2556 #ifdef _LP64
2557 	    size > OFF_MAX ||
2558 #endif
2559 	    foff < 0 || foff > OFF_MAX - size)
2560 		return (EINVAL);
2561 
2562 	writecounted = FALSE;
2563 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2564 	    &foff, &object, &writecounted);
2565 	if (error != 0)
2566 		return (error);
2567 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2568 	    foff, writecounted, td);
2569 	if (error != 0) {
2570 		/*
2571 		 * If this mapping was accounted for in the vnode's
2572 		 * writecount, then undo that now.
2573 		 */
2574 		if (writecounted)
2575 			vm_pager_release_writecount(object, 0, size);
2576 		vm_object_deallocate(object);
2577 	}
2578 #ifdef HWPMC_HOOKS
2579 	/* Inform hwpmc(4) if an executable is being mapped. */
2580 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2581 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2582 			pkm.pm_file = vp;
2583 			pkm.pm_address = (uintptr_t) *addr;
2584 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2585 		}
2586 	}
2587 #endif
2588 	return (error);
2589 }
2590 
2591 void
2592 vn_fsid(struct vnode *vp, struct vattr *va)
2593 {
2594 	fsid_t *f;
2595 
2596 	f = &vp->v_mount->mnt_stat.f_fsid;
2597 	va->va_fsid = (uint32_t)f->val[1];
2598 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2599 	va->va_fsid += (uint32_t)f->val[0];
2600 }
2601 
2602 int
2603 vn_fsync_buf(struct vnode *vp, int waitfor)
2604 {
2605 	struct buf *bp, *nbp;
2606 	struct bufobj *bo;
2607 	struct mount *mp;
2608 	int error, maxretry;
2609 
2610 	error = 0;
2611 	maxretry = 10000;     /* large, arbitrarily chosen */
2612 	mp = NULL;
2613 	if (vp->v_type == VCHR) {
2614 		VI_LOCK(vp);
2615 		mp = vp->v_rdev->si_mountpt;
2616 		VI_UNLOCK(vp);
2617 	}
2618 	bo = &vp->v_bufobj;
2619 	BO_LOCK(bo);
2620 loop1:
2621 	/*
2622 	 * MARK/SCAN initialization to avoid infinite loops.
2623 	 */
2624         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2625 		bp->b_vflags &= ~BV_SCANNED;
2626 		bp->b_error = 0;
2627 	}
2628 
2629 	/*
2630 	 * Flush all dirty buffers associated with a vnode.
2631 	 */
2632 loop2:
2633 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2634 		if ((bp->b_vflags & BV_SCANNED) != 0)
2635 			continue;
2636 		bp->b_vflags |= BV_SCANNED;
2637 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2638 			if (waitfor != MNT_WAIT)
2639 				continue;
2640 			if (BUF_LOCK(bp,
2641 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2642 			    BO_LOCKPTR(bo)) != 0) {
2643 				BO_LOCK(bo);
2644 				goto loop1;
2645 			}
2646 			BO_LOCK(bo);
2647 		}
2648 		BO_UNLOCK(bo);
2649 		KASSERT(bp->b_bufobj == bo,
2650 		    ("bp %p wrong b_bufobj %p should be %p",
2651 		    bp, bp->b_bufobj, bo));
2652 		if ((bp->b_flags & B_DELWRI) == 0)
2653 			panic("fsync: not dirty");
2654 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2655 			vfs_bio_awrite(bp);
2656 		} else {
2657 			bremfree(bp);
2658 			bawrite(bp);
2659 		}
2660 		if (maxretry < 1000)
2661 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2662 		BO_LOCK(bo);
2663 		goto loop2;
2664 	}
2665 
2666 	/*
2667 	 * If synchronous the caller expects us to completely resolve all
2668 	 * dirty buffers in the system.  Wait for in-progress I/O to
2669 	 * complete (which could include background bitmap writes), then
2670 	 * retry if dirty blocks still exist.
2671 	 */
2672 	if (waitfor == MNT_WAIT) {
2673 		bufobj_wwait(bo, 0, 0);
2674 		if (bo->bo_dirty.bv_cnt > 0) {
2675 			/*
2676 			 * If we are unable to write any of these buffers
2677 			 * then we fail now rather than trying endlessly
2678 			 * to write them out.
2679 			 */
2680 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2681 				if ((error = bp->b_error) != 0)
2682 					break;
2683 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2684 			    (error == 0 && --maxretry >= 0))
2685 				goto loop1;
2686 			if (error == 0)
2687 				error = EAGAIN;
2688 		}
2689 	}
2690 	BO_UNLOCK(bo);
2691 	if (error != 0)
2692 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2693 
2694 	return (error);
2695 }
2696 
2697 /*
2698  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2699  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2700  * to do the actual copy.
2701  * vn_generic_copy_file_range() is factored out, so it can be called
2702  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2703  * different file systems.
2704  */
2705 int
2706 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2707     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2708     struct ucred *outcred, struct thread *fsize_td)
2709 {
2710 	int error;
2711 	size_t len;
2712 	uint64_t uvalin, uvalout;
2713 
2714 	len = *lenp;
2715 	*lenp = 0;		/* For error returns. */
2716 	error = 0;
2717 
2718 	/* Do some sanity checks on the arguments. */
2719 	uvalin = *inoffp;
2720 	uvalin += len;
2721 	uvalout = *outoffp;
2722 	uvalout += len;
2723 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2724 		error = EISDIR;
2725 	else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin <
2726 	    (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX ||
2727 	    uvalout < (uint64_t)*outoffp || invp->v_type != VREG ||
2728 	    outvp->v_type != VREG)
2729 		error = EINVAL;
2730 	if (error != 0)
2731 		goto out;
2732 
2733 	/*
2734 	 * If the two vnode are for the same file system, call
2735 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2736 	 * which can handle copies across multiple file systems.
2737 	 */
2738 	*lenp = len;
2739 	if (invp->v_mount == outvp->v_mount)
2740 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2741 		    lenp, flags, incred, outcred, fsize_td);
2742 	else
2743 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2744 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2745 out:
2746 	return (error);
2747 }
2748 
2749 /*
2750  * Test len bytes of data starting at dat for all bytes == 0.
2751  * Return true if all bytes are zero, false otherwise.
2752  * Expects dat to be well aligned.
2753  */
2754 static bool
2755 mem_iszero(void *dat, int len)
2756 {
2757 	int i;
2758 	const u_int *p;
2759 	const char *cp;
2760 
2761 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2762 		if (len >= sizeof(*p)) {
2763 			if (*p != 0)
2764 				return (false);
2765 		} else {
2766 			cp = (const char *)p;
2767 			for (i = 0; i < len; i++, cp++)
2768 				if (*cp != '\0')
2769 					return (false);
2770 		}
2771 	}
2772 	return (true);
2773 }
2774 
2775 /*
2776  * Look for a hole in the output file and, if found, adjust *outoffp
2777  * and *xferp to skip past the hole.
2778  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2779  * to be written as 0's upon return.
2780  */
2781 static off_t
2782 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2783     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2784 {
2785 	int error;
2786 	off_t delta;
2787 
2788 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2789 		*dataoffp = *outoffp;
2790 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2791 		    curthread);
2792 		if (error == 0) {
2793 			*holeoffp = *dataoffp;
2794 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2795 			    curthread);
2796 		}
2797 		if (error != 0 || *holeoffp == *dataoffp) {
2798 			/*
2799 			 * Since outvp is unlocked, it may be possible for
2800 			 * another thread to do a truncate(), lseek(), write()
2801 			 * creating a hole at startoff between the above
2802 			 * VOP_IOCTL() calls, if the other thread does not do
2803 			 * rangelocking.
2804 			 * If that happens, *holeoffp == *dataoffp and finding
2805 			 * the hole has failed, so disable vn_skip_hole().
2806 			 */
2807 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2808 			return (xfer2);
2809 		}
2810 		KASSERT(*dataoffp >= *outoffp,
2811 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2812 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2813 		KASSERT(*holeoffp > *dataoffp,
2814 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2815 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2816 	}
2817 
2818 	/*
2819 	 * If there is a hole before the data starts, advance *outoffp and
2820 	 * *xferp past the hole.
2821 	 */
2822 	if (*dataoffp > *outoffp) {
2823 		delta = *dataoffp - *outoffp;
2824 		if (delta >= *xferp) {
2825 			/* Entire *xferp is a hole. */
2826 			*outoffp += *xferp;
2827 			*xferp = 0;
2828 			return (0);
2829 		}
2830 		*xferp -= delta;
2831 		*outoffp += delta;
2832 		xfer2 = MIN(xfer2, *xferp);
2833 	}
2834 
2835 	/*
2836 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2837 	 * that the write ends at the start of the hole.
2838 	 * *holeoffp should always be greater than *outoffp, but for the
2839 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2840 	 * value.
2841 	 */
2842 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2843 		xfer2 = *holeoffp - *outoffp;
2844 	return (xfer2);
2845 }
2846 
2847 /*
2848  * Write an xfer sized chunk to outvp in blksize blocks from dat.
2849  * dat is a maximum of blksize in length and can be written repeatedly in
2850  * the chunk.
2851  * If growfile == true, just grow the file via vn_truncate_locked() instead
2852  * of doing actual writes.
2853  * If checkhole == true, a hole is being punched, so skip over any hole
2854  * already in the output file.
2855  */
2856 static int
2857 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2858     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2859 {
2860 	struct mount *mp;
2861 	off_t dataoff, holeoff, xfer2;
2862 	int error, lckf;
2863 
2864 	/*
2865 	 * Loop around doing writes of blksize until write has been completed.
2866 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2867 	 * done for each iteration, since the xfer argument can be very
2868 	 * large if there is a large hole to punch in the output file.
2869 	 */
2870 	error = 0;
2871 	holeoff = 0;
2872 	do {
2873 		xfer2 = MIN(xfer, blksize);
2874 		if (checkhole) {
2875 			/*
2876 			 * Punching a hole.  Skip writing if there is
2877 			 * already a hole in the output file.
2878 			 */
2879 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2880 			    &dataoff, &holeoff, cred);
2881 			if (xfer == 0)
2882 				break;
2883 			if (holeoff < 0)
2884 				checkhole = false;
2885 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2886 			    (intmax_t)xfer2));
2887 		}
2888 		bwillwrite();
2889 		mp = NULL;
2890 		error = vn_start_write(outvp, &mp, V_WAIT);
2891 		if (error == 0) {
2892 			if (MNT_SHARED_WRITES(mp))
2893 				lckf = LK_SHARED;
2894 			else
2895 				lckf = LK_EXCLUSIVE;
2896 			error = vn_lock(outvp, lckf);
2897 		}
2898 		if (error == 0) {
2899 			if (growfile)
2900 				error = vn_truncate_locked(outvp, outoff + xfer,
2901 				    false, cred);
2902 			else {
2903 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
2904 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
2905 				    curthread->td_ucred, cred, NULL, curthread);
2906 				outoff += xfer2;
2907 				xfer -= xfer2;
2908 			}
2909 			VOP_UNLOCK(outvp);
2910 		}
2911 		if (mp != NULL)
2912 			vn_finished_write(mp);
2913 	} while (!growfile && xfer > 0 && error == 0);
2914 	return (error);
2915 }
2916 
2917 /*
2918  * Copy a byte range of one file to another.  This function can handle the
2919  * case where invp and outvp are on different file systems.
2920  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
2921  * is no better file system specific way to do it.
2922  */
2923 int
2924 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
2925     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
2926     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
2927 {
2928 	struct vattr va;
2929 	struct mount *mp;
2930 	struct uio io;
2931 	off_t startoff, endoff, xfer, xfer2;
2932 	u_long blksize;
2933 	int error;
2934 	bool cantseek, readzeros, eof, lastblock;
2935 	ssize_t aresid;
2936 	size_t copylen, len, savlen;
2937 	char *dat;
2938 	long holein, holeout;
2939 
2940 	holein = holeout = 0;
2941 	savlen = len = *lenp;
2942 	error = 0;
2943 	dat = NULL;
2944 
2945 	error = vn_lock(invp, LK_SHARED);
2946 	if (error != 0)
2947 		goto out;
2948 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
2949 		holein = 0;
2950 	VOP_UNLOCK(invp);
2951 
2952 	mp = NULL;
2953 	error = vn_start_write(outvp, &mp, V_WAIT);
2954 	if (error == 0)
2955 		error = vn_lock(outvp, LK_EXCLUSIVE);
2956 	if (error == 0) {
2957 		/*
2958 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
2959 		 * now that outvp is locked.
2960 		 */
2961 		if (fsize_td != NULL) {
2962 			io.uio_offset = *outoffp;
2963 			io.uio_resid = len;
2964 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
2965 			if (error != 0)
2966 				error = EFBIG;
2967 		}
2968 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
2969 			holeout = 0;
2970 		/*
2971 		 * Holes that are past EOF do not need to be written as a block
2972 		 * of zero bytes.  So, truncate the output file as far as
2973 		 * possible and then use va.va_size to decide if writing 0
2974 		 * bytes is necessary in the loop below.
2975 		 */
2976 		if (error == 0)
2977 			error = VOP_GETATTR(outvp, &va, outcred);
2978 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
2979 		    *outoffp + len) {
2980 #ifdef MAC
2981 			error = mac_vnode_check_write(curthread->td_ucred,
2982 			    outcred, outvp);
2983 			if (error == 0)
2984 #endif
2985 				error = vn_truncate_locked(outvp, *outoffp,
2986 				    false, outcred);
2987 			if (error == 0)
2988 				va.va_size = *outoffp;
2989 		}
2990 		VOP_UNLOCK(outvp);
2991 	}
2992 	if (mp != NULL)
2993 		vn_finished_write(mp);
2994 	if (error != 0)
2995 		goto out;
2996 
2997 	/*
2998 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
2999 	 * If hole sizes aren't available, set the blksize to the larger
3000 	 * f_iosize of invp and outvp.
3001 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3002 	 * This value is clipped at 4Kbytes and 1Mbyte.
3003 	 */
3004 	blksize = MAX(holein, holeout);
3005 	if (blksize == 0)
3006 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3007 		    outvp->v_mount->mnt_stat.f_iosize);
3008 	if (blksize < 4096)
3009 		blksize = 4096;
3010 	else if (blksize > 1024 * 1024)
3011 		blksize = 1024 * 1024;
3012 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3013 
3014 	/*
3015 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3016 	 * to find holes.  Otherwise, just scan the read block for all 0s
3017 	 * in the inner loop where the data copying is done.
3018 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3019 	 * support holes on the server, but do not support FIOSEEKHOLE.
3020 	 */
3021 	eof = false;
3022 	while (len > 0 && error == 0 && !eof) {
3023 		endoff = 0;			/* To shut up compilers. */
3024 		cantseek = true;
3025 		startoff = *inoffp;
3026 		copylen = len;
3027 
3028 		/*
3029 		 * Find the next data area.  If there is just a hole to EOF,
3030 		 * FIOSEEKDATA should fail and then we drop down into the
3031 		 * inner loop and create the hole on the outvp file.
3032 		 * (I do not know if any file system will report a hole to
3033 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3034 		 *  will fail for those file systems.)
3035 		 *
3036 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3037 		 * the code just falls through to the inner copy loop.
3038 		 */
3039 		error = EINVAL;
3040 		if (holein > 0)
3041 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3042 			    incred, curthread);
3043 		if (error == 0) {
3044 			endoff = startoff;
3045 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3046 			    incred, curthread);
3047 			/*
3048 			 * Since invp is unlocked, it may be possible for
3049 			 * another thread to do a truncate(), lseek(), write()
3050 			 * creating a hole at startoff between the above
3051 			 * VOP_IOCTL() calls, if the other thread does not do
3052 			 * rangelocking.
3053 			 * If that happens, startoff == endoff and finding
3054 			 * the hole has failed, so set an error.
3055 			 */
3056 			if (error == 0 && startoff == endoff)
3057 				error = EINVAL; /* Any error. Reset to 0. */
3058 		}
3059 		if (error == 0) {
3060 			if (startoff > *inoffp) {
3061 				/* Found hole before data block. */
3062 				xfer = MIN(startoff - *inoffp, len);
3063 				if (*outoffp < va.va_size) {
3064 					/* Must write 0s to punch hole. */
3065 					xfer2 = MIN(va.va_size - *outoffp,
3066 					    xfer);
3067 					memset(dat, 0, MIN(xfer2, blksize));
3068 					error = vn_write_outvp(outvp, dat,
3069 					    *outoffp, xfer2, blksize, false,
3070 					    holeout > 0, outcred);
3071 				}
3072 
3073 				if (error == 0 && *outoffp + xfer >
3074 				    va.va_size && xfer == len)
3075 					/* Grow last block. */
3076 					error = vn_write_outvp(outvp, dat,
3077 					    *outoffp, xfer, blksize, true,
3078 					    false, outcred);
3079 				if (error == 0) {
3080 					*inoffp += xfer;
3081 					*outoffp += xfer;
3082 					len -= xfer;
3083 				}
3084 			}
3085 			copylen = MIN(len, endoff - startoff);
3086 			cantseek = false;
3087 		} else {
3088 			cantseek = true;
3089 			startoff = *inoffp;
3090 			copylen = len;
3091 			error = 0;
3092 		}
3093 
3094 		xfer = blksize;
3095 		if (cantseek) {
3096 			/*
3097 			 * Set first xfer to end at a block boundary, so that
3098 			 * holes are more likely detected in the loop below via
3099 			 * the for all bytes 0 method.
3100 			 */
3101 			xfer -= (*inoffp % blksize);
3102 		}
3103 		/* Loop copying the data block. */
3104 		while (copylen > 0 && error == 0 && !eof) {
3105 			if (copylen < xfer)
3106 				xfer = copylen;
3107 			error = vn_lock(invp, LK_SHARED);
3108 			if (error != 0)
3109 				goto out;
3110 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3111 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3112 			    curthread->td_ucred, incred, &aresid,
3113 			    curthread);
3114 			VOP_UNLOCK(invp);
3115 			lastblock = false;
3116 			if (error == 0 && aresid > 0) {
3117 				/* Stop the copy at EOF on the input file. */
3118 				xfer -= aresid;
3119 				eof = true;
3120 				lastblock = true;
3121 			}
3122 			if (error == 0) {
3123 				/*
3124 				 * Skip the write for holes past the initial EOF
3125 				 * of the output file, unless this is the last
3126 				 * write of the output file at EOF.
3127 				 */
3128 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3129 				    false;
3130 				if (xfer == len)
3131 					lastblock = true;
3132 				if (!cantseek || *outoffp < va.va_size ||
3133 				    lastblock || !readzeros)
3134 					error = vn_write_outvp(outvp, dat,
3135 					    *outoffp, xfer, blksize,
3136 					    readzeros && lastblock &&
3137 					    *outoffp >= va.va_size, false,
3138 					    outcred);
3139 				if (error == 0) {
3140 					*inoffp += xfer;
3141 					startoff += xfer;
3142 					*outoffp += xfer;
3143 					copylen -= xfer;
3144 					len -= xfer;
3145 				}
3146 			}
3147 			xfer = blksize;
3148 		}
3149 	}
3150 out:
3151 	*lenp = savlen - len;
3152 	free(dat, M_TEMP);
3153 	return (error);
3154 }
3155 
3156 static int
3157 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3158 {
3159 	struct mount *mp;
3160 	struct vnode *vp;
3161 	off_t olen, ooffset;
3162 	int error;
3163 #ifdef AUDIT
3164 	int audited_vnode1 = 0;
3165 #endif
3166 
3167 	vp = fp->f_vnode;
3168 	if (vp->v_type != VREG)
3169 		return (ENODEV);
3170 
3171 	/* Allocating blocks may take a long time, so iterate. */
3172 	for (;;) {
3173 		olen = len;
3174 		ooffset = offset;
3175 
3176 		bwillwrite();
3177 		mp = NULL;
3178 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3179 		if (error != 0)
3180 			break;
3181 		error = vn_lock(vp, LK_EXCLUSIVE);
3182 		if (error != 0) {
3183 			vn_finished_write(mp);
3184 			break;
3185 		}
3186 #ifdef AUDIT
3187 		if (!audited_vnode1) {
3188 			AUDIT_ARG_VNODE1(vp);
3189 			audited_vnode1 = 1;
3190 		}
3191 #endif
3192 #ifdef MAC
3193 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3194 		if (error == 0)
3195 #endif
3196 			error = VOP_ALLOCATE(vp, &offset, &len);
3197 		VOP_UNLOCK(vp);
3198 		vn_finished_write(mp);
3199 
3200 		if (olen + ooffset != offset + len) {
3201 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3202 			    ooffset, olen, offset, len);
3203 		}
3204 		if (error != 0 || len == 0)
3205 			break;
3206 		KASSERT(olen > len, ("Iteration did not make progress?"));
3207 		maybe_yield();
3208 	}
3209 
3210 	return (error);
3211 }
3212